Professional Documents
Culture Documents
Analogue Filter Networks, Developments in Theory, Design and Analysis (1995) [J.I. Sewell]
Analogue Filter Networks, Developments in Theory, Design and Analysis (1995) [J.I. Sewell]
How to cite:
Sewell, J.I. (1995) Analogue lter networks: developments in theory, design and analyses, Durham theses,
Durham University. Available at Durham E-Theses Online: http://etheses.dur.ac.uk/5501/
Use policy
The full-text may be used and/or reproduced, and given to third parties in any format or medium, without prior permission or
charge, for personal research or study, educational, or not-for-prot purposes provided that:
• a full bibliographic reference is made to the original source
Academic Support Oce, Durham University, University Oce, Old Elvet, Durham DH1 3HP
e-mail: e-theses.admin@dur.ac.uk Tel: +44 0191 334 6107
http://etheses.dur.ac.uk
ANALOGUE FILTER NETWORKS, DEVELOPMENTS
IN THEORY, DESIGN AND ANALYSIS
by
J.I. SEWELL
f r o m it should be acknowledged.
Doctor of Science
March 1995
Dedication
"Have you learned lessons only of those who admired you, and were tender with you, and
stood aside for you ? Have you not learned great lessons from those who braced themselves
The late Dorothy Sewell - my mother: though denied personal access to scholarship, began
my "disputed passage".
passage".
1. Curriculum Vitae
2. Publications Submitted
lo Curriculum Vitae
CJi.R R 1 C U L U M VI T AE
General Education
Year School
Examinations Taken
University Education
Dais College
4. Employment History
Dates
1968 - 1974)
1975- 1978) Second Year Laboratory - 156 hrs/year
1984- 1985)
1968- 1971)
1975- 1976) First Year Problems Class 24 hrs/year
- M.Sc. 1973
University Duties
Published Work
1. "Table for the voltage transfer functions of single amplifier double ladder feedback systems",
Electronics Letters, May 1965, Vol.1, No.3, pp 70-71, (with A.G.J.Holt).
4. "Table for the voltage transfer functions of single amplifier double ladder parallel feedback
systems", I.E.E.E. Trans. C.T., September 1966, Vol. CT-13, pp.326-328, (with A . G J .
Holt).
5. "The Computer in electronics, Part I " , Design Electronics, Oct. 1967, Vol.5. N o . l , pp. 34-
36. (with J.K. Fidler).
6. "The computer in electronics, Part I I " , Design Electronics. Nov. 1967, Vol.5, No.2, pp. 38-
41, (with J.K. Fidler).
8. "Matrix tables for the generalised 5-terminal amplifier". The Radio and Electronic Engineer,
April 1969, Vol. 37, No.4, pp.247-55, (with F.W.Stephenson).
10. "Synthesis of multi-loop feedback systems, Part I I " , International Journal of Control, May
1969, Vol.9, No.5, pp.499-508. (with A.G.J. Holt).
11. "A simple method for producing floating inductors", Proc. I.E.E.E. December 1969, Vol. 57,
No.12, pp. 2155-2156.
12. "Accumulant theory and applications", International Journal of Control, March 1970, Vol.
l t , N o . 3 , pp.479-488.
13. "Synthesis of active devices". International Journal of Electronics, December 1970, Vol.29,
No.6, pp. 501-511.
14. "A synthetic analysis of general active networks", I.E.E. International Conference on
Computer Aided Design, Southampton, April 1972. I.E.E. Conference Publication No.86,
pp.12-18, (with I.Walker).
15. "A course in computer-aided electronic circuit design for undergraduates", Computer Aided
Design, April 1972, Vol 4, No.3, pp. 135-137, (with A.G. Martin ).
16. "General synthetic analysis of active networks", Proc. Second International Symposium on
Network Theory, Herceg-Novi, Yugoslavia, July 1972, pp. 61-69, (with C. Fridas).
17. "The multiterminal active transformer". International Journal of Control, July 1972, Vol. 16,
N o . l , pp. 171-175.
18. "Active circulator systems". International Journal of Electronics, December 1972, Vol.33,
No.6, pp.677-679.
20. "Symbolic analysis for computer-aided circuit design - The interpolative approach", I.E.E.E.
Trans. C.T., November 1973 Vol. CT-20, No.6, pp. 738-741. (with J.K. Fidler).
21. "A partition method for symbolic network analysis", I.E.E. Colloquium on Computer Aided
Circuit Design, Colloquium Digest 1974 - 4, pp. 7/1 - 7/2 March 1974, (with C. Fridas).
22. "Digital realisation of linear and parametric networks", I.E.R.E. Colloquium on Signal
Processing in Feedback Control Systems, London, April 1974, (with F.G.A. Coupe).
8
24. "Correction to symbolic analysis for computer-aided circuit design - the interpolate
approach", I.E.E.E. Trans. CAS, May 1974, Vol. CAS-21. No.3, P.462, (with J.K.Fidler).
25. "The design of variable immittance amplifiers", I.E.E. Colloquium on Amplifier Design -
Theory and Practice, Colloquium Digest 1974 - 24, pp.11/1 - 11/2, May 1974. (with J.A.
Neill and F.G.A. Coupe).
26. "Interpolative analysis for computer aided design".Proc. Summer School on Circuit Theory
1974, Prague, September 1974, Vol.2, pp. 217 - 223. (with C. Fridas).
27. "An adaptive electronic circulator for use in telephones", I.E.E.E. Trans. COM, July 1975,
Vol. COM-23, No.7, pp. 728-731, (with R. Lodge).
28. "Error estimation and improvement in interpolative symbolic network analysis", Proc. Third
International Symposium on Network Theory, Split, Yugoslavia, September 1975, pp. 121-
129, (with C. Fridas).
29. "Network partition methods for use in symbolic analysis by interpolation", Proc. 1976
I.E.E.E. International Symposium on Circuits and Systems, Munich, April 1976, pp. 45-48
(with C. Fridas).
30. "A theory of equivalent active networks", Proc. 1976 I.E.E.E. International Symposium on
Circuits and Systems, Munich, April 1976, pp. 497-500 (with L.G.Grant).
31. "A theory of equivalent active networks", I.E.E.E. Trans. CAS, June 1976, Vol. CAS-23,
No.6, pp.350-354, (with L.G.Grant).
32. "Network partition methods for use in symbolic analysis by interpolation". Computer Aided
Design, July 1976, Vol. 8, No.3. pp. 157-164, (with C. Fridas).
33. "Symbolic network sensitivities using partition methods", Int. Journal of Electronics, July
1976, Vol.41, No.l, pp.25-32, (with C. Fridas).
34. "Partitioned transformations in equivalent active network theory", Proc. 1976 European
Conference on Circuit Theory and Design, Genoa, September 1976, pp. 232-238, (with L.G.
Grant).
35. "On the evaluation of sensitivity polynomials in active equivalent network theory", Proc.
I.E.E.E. Conference on Computer Aided Design of Electronic and Microwave Circuits and
Systems, Hull, July 1977, pp.84-89, (with L.G. Grant).
37. "On sensitivity theorems in active equivalent network theory", Proc. 1978 I.E.E.E.
International Symposium on Circuits and Systems, New York, May 1978, pp. 1060-1063,
(with J.Dunning-Davies and F.D. Faulkner).
38. "Some further sensitivity theorems in active equivalent network theory", I.E.E. Journal on
Electronic Circuits and Systems, November 1978, Vol.2, No.6, pp. 193-198, (with
J.Dunning-D^vies and F.D. Faulkner).
9
39. "An improved adaptive electronic circulator for telephone applications". I.E.E.E. Trans.
COM, August 1979, Vol. COM-27, No.8, pp. 1218-1224 (with H. Gazioglu and D.A.
Homer).
40. "An adaptive electronic circulator for telephone hybrid applications", Proc. Fourth
International Symposium on Network Theory, Ljubljana, September 1979. pp. 369-376,
(with H. Gazioglu and D.A. Homer).
41. "Analysis of active switched-capacitor networks", Proc. I.E.E.E. February 1980. Vol. 68,
No.2, pp. 292-293.
42. "A degree enhancement scheme in electronic engineering". Proc. Conference on Electronic
Engineering in Degree Courses - Teaching for the 80's, University of Hull, April 1980,
pp.21/1 -21/11.
43. "Inclusion of amplifier finite gain and bandwidth in analysis of switched-capacitor filters",
Electronics Letters, June 1980, Vol.16, No.12, pp.462-463, (with C. Lau).
44. "Digital active network analysis", Proc. I.E.E., Part G, Electronic Circuits and Systems,
August 1980, Vol.127, No. 4, pp. 181-186 (with F.G.A. Coupe).
45. "A degree enhancement scheme in electronic engineering". Int. J. Elect. Enging. Educ,
January 1981, Vol.18, pp. 29-34.
48. "CAD tools for switched capacitor filter design", Proc. I.E.E.E. Workshop on Design and
Fabrication of Integrated Circuit Filters, London, September 1982, pp. B1-B47.
49. "Compact matrix scheme for use in computer analysis of switched capacitor networks".
Electronics Letters, September 1982, Vol. 18, No. 19, pp. 840-841, (with C. Lau).
50. "Software for switched capacitor filter design", I.E.R.E. Colloquium on Analogue Filters,
Colloquium Digest 2pp. November 1983.
52. "Some improvement schemes for the interpolative symbolic analysis of switched capacitor
networks", Proc. I.E.E.E. International Symposium on Circuits and Systems, Montreal,
May 1984, 4pp, (with D.G. Johnson).
55. "On time-sharing and noise minimisation in switched-capacitor filters", I.E.E. Saraga
Colloquium on Electronic Filters, Colloquium Digest, 6pp, May 1985, (with M . Powell and
R.C.J. Taylor).
Electronic Filters, Colloquium Digest, 5pp, May 1985, (with D.J. Gilbert).
57. "Symbolic analysis of ideal and non-ideal switched capacitor networks", Proc. I.E.E.E.
International Symposium on Circuits and Systems, Kyoto, June 1985, pp.1165-1168 (with
C.K. Pun).
58. "On the generation of equivalent switched capacitor networks", Proc. European Conference on
Circuit Theory and Design, Prague, September, 1985 pp.697-700 (with D.J. Gilbert).
59. "Techniques for improving the efficiency of analysis of software for large switched-capacitor
networks, Proc. 28th Midwest Symposium on Circuits and Systems, Louisville, August
1985, pp.390-393 (with A.D. Meakin and L.B. Wolovitz).
60. "Efficient computer techniques for the exact analysis of all non-ideal effects of switched-
capacitor networks in the time domain", Proc. I.E.E.E. International Symposium on Circuits
and Systems, San Jose, May 1986, pp. 373-376, (with L.B. Wolovitz).
61. "Application of equivalent network theory to strays insensitive switched capacitory filters",
I.E.E Saraga Colloquium on Electronic Filters, Colloquium Digest pp. 3/1 - 3/8, May 1986,
(with D.J. Gilbert).
62. "Multiphase equivalence theory for switched capacitor networks", Proc. 29th Midwest
Symposium on Circuits and Systems, Lincoln, Na., August 1986, pp. 801 - 806, (with D.J.
Gilbert).
63. "Noise analysis of switched capacitor networks in symbolic form", Proc. 29th Midwest
Symposium on Circuits and Systems, Lincoln, Na., August 1986, pp 807 - 810, (with C.K.
Pun and A.G. Hall).
64. "The LUD approach to switched-capacitor filter design", I.E.E. Saraga Colloquium on
Electronic Filters, Colloquium Digest pp. 7/1 - 7/7, March 1987, (with L i Ping).
66. "The LUD approach to switched-capacitor filter design", I.E.E.E. Trans. CAS, Dec. 1987,
Vol. CAS-34, No.12, pp.1611 - 1614 (with L i Ping).
67. "The SCNAP series of software for switched capacitor circuit design", I.E.E. Colloquium on
Analogue IC Design, London, November 1987 pp. 10/1 - 10/5.
68. "Matrix methods for switched capacitor filter design", Proc. I.E.E.E. International
Symposium on Circuits and Systems, Helsinki, June 1988, pp. 1021 - 1024 (with L i Ping
and R K Henderson).
69. "General analysis of large linear switched capacitor networks", Proc. I.E.E. Part G, Electronic
Circuits and Systems, June 1988, Vol. 135, No.3, pp. 119 - 124, (with L.B. Wolovitz).
70. "Filter realisation by passive network simulation", Proc. I.E.E. Part G, Electronic Circuits
and Systems, August 1988, Vol. 135, No.4, pp. 167 - 176 (with L i Ping).
71. "Design and Optimisation of GaAs switched capacitor filter", I.E.E. Saraga Colloquium on
Electronic Filters, London, May 1988, pp. 1/1 - 1/10 (with D.G.Haigh, C. Toumazou, S.J.
Harrold).
1
72. "PANDDA : A program for advanced network design: digital and analogue", I.E.E. Saraga
Colloquium on Electronic Filters, London, May 1988, pp. 4/1 - 4/8. (with R.K. Henderson
and L i Ping).
73. "Digital filter realisation by passive network simulation", I.E.E. Saraga Colloquium on
Electronic Filters. London. May 1989. pp. 8/1 - 8/8 (with L i Ping).
74. "On low sensitivity/noise digital filter structures". Proc. I.E.E.E. International Conference on
Acoustics, Speech and Signal Processing, Glasgow. May 1989, pp. 845 - 848 (with L i
Ping).
75. "A new filter approximation and design algorithm", Proc. I.E.E.E. International Symposium
on Circuits and Systems, Portland, May 1989, pp. 1063 - 1066, (with L i Ping and R.K.
Henderson).
76. "Design and optimisation of a GaAs switched capacitor filter", Proc. I.E.E.E. International
Symposium on Circuits and Systems, Portland, May 1989, pp. 1449 - 1454 (with D.G.
Haigh, C. Toumazou, S.J. Harrold and K. Steptoe).
77. "A design algorithm for all-pass delay equalisers," I.E.E. Saraga Colloquium on Electronic
Filters, London, June 1989, pp. 11/1-11/8 (with R.K. Henderson).
78. "The TWINTOR in bandstop switched-capacitor ladder filter realisation," I.E.E.E. Trans.
CAS, July 1989, Vol. CAS-36, No.7, pp. 1041 - 1044 (with L i Ping).
79. "The TWINTOR in bandstop switched-capacitor ladder filter realisation", Proc. European
Conference on Circuit Theory and Design, Brighton, September 1989, pp. 2 7 - 3 1 (with L i
Ping).
80. "Design program for digital and analogue filters: PANDDA", Proc. European Conference on
Circuit Theory and Design, Brighton, September 1989, pp. 289 - 293 (with L i Ping and
R.K. Henderson).
81. "High performance filter networks and symmetric matrix systems", Proc. I.E.E., Part G,
Circuits Devices and Systems, Dec. 1989, Vol. 136, No.6, pp. 327 - 336, (with L i Ping).
82. "A unified approach to the design of canonic integrated ladder filters," Proc. I.E.E.E.
International Symposium on Circuits and Systems, New Orleans, May 1990, pp.2272 -
2275 (with L i Ping and R.K. Henderson).
83. "Switched capacitor and active - RC allpass ladder filters," Proc. I.E.E.E. International
Symposium on Circuits and Systems, New Orleans, May 1990, pp. 2833 - 2836 (with L i
Ping).
84. "Design and testing of a GaAs switched capacitor filter", Proc. I.E.E.E. International
Symposium on Circuits and Systems, New Orleans, May 1990, pp. 2825 - 2828 (with D.G.
Haigh, C. Toumazou, S.J. Harrold, K. Steptoe and R. Bayruns).
85. "400 MHz switching rate GaAs switched capacitor filter" Electronics Letters, March 1990,
No.7, pp. 460 - 461 (with D.G. Haigh, C. Toumazou, S.J. Harrold, K. Steptoe and R.
Bayruns).
87. "Active and digital ladder-based allpass filters", Proc. I.E.E., Part G, Circuits Devices and
Systems, December 1990, Vol.137, No.6. pp.439-445 (with L i Ping).
88. "Extended Remez algorithms for filter amplitude and group delay approximation", Proc.
I.E.E., Part G, Circuits Devices and Systems, June 1991, Vol.138. No.3, pp.289-300 (with
R. K. Henderson and L i Ping).
89. "Canonical design of integrated ladder filters", Proc. I.E.E.. Part G, Circuits Devices and
Systems, April 1991, Vol.138, No.2, pp.222-228 (with R.K. Henderson and L i Ping).
90. "Design, optimisation and testing of a GaAs switched capacitor filter", I.EJE.E.Trans. CAS.
August 1991, CAS-38, No.8, pp.825-837 (with D.G. Haigh, C. Toumazou. S.J. Harrold. K.
Steptoe and R. Bayruns).
91. "Matrix methods for the design of transconductor ladder filters", Proc. I.E.E.E.International
Symposium on Circuits and Systems, Singapore, June 1991, pp.1355-1358 (with N.P.J.
Greer, R.K. Henderson and Li Ping).
92. "Design of a switched-capacitor filter for voice band signals", Proc. I.E.E.E.International
Symposium on Circuits and Systems, Singapore, June 1991, pp.1574-1577 (with L i Ping,
R.C.J. Taylor and R.K Henderson).
93. "A methodology for integrated ladder filter design", I.E.E.E.Trans. CAS, August 1991.
CAS-38, No.8, pp.853-868 (with L i Ping and R.K. Henderson).
94. "A 500 MHz GaAs second-order bandpass switched-capacitor filter", Proc. European
Conference on Circuit Theory and Design, Copenhagen, September 1991, pp.1171-1177
(with S.J Harrold, D.G. Haigh, C Toumazou, K. Steptoe and R. Bayruns).
95. "Software for the design of transconductor-capacitor filters and equalisers", I.E.E. Saraga
Colloquium on Digital and Analogue Filters and Filtering Systems. London, December
1991, pp.7/1-7/6 (with Lu Yue and N.P.J Greer).
96. "High performance front-end signal processing: system design aids", I.E.E. Colloquium on
Circuit Theory and DSP, London, February 1992, pp.2/1-2/6.
97. " A method for the evaluation of multirate sigma delta systems". Proc. I.E.E.E. International
Symposium on Circuits and Systems, San Diego, May 1992, pp. 1328-1331, (with D.M.
Hossack).
98. "Design of a switched-capacitor filter for a mobile telephone receiver", I.E.E.E. Journal of
Solid State Circuits, September 1992, Vol. 27, No.9., pp. 1294-1298 (with L i Ping, R.K.
Henderson and R.C.J. Taylor).
99. "Ladder based transconductor-capacitor filter and equaliser design", I.E.E. Saraga Colloquium
on Digital and Analogue Filters and Filtering Systems, London, November 1992, pp. 7/1 -
7/8, (with Lu Yue and N.P.J. Greer).
100. "Analog integrated filter compilation". Journal of Analog Integrated Circuits and Signal
Processing", March 1993, Vol.3, pp.217-228 (with R.K. Henderson and L i Ping).
101. "A transconductor-capacitor video filter and equaliser design", Proc. I.E.E.E. International
Symposium on Circuits and Systems, Chicago. May 1993, pp986-989 (with Lu Yue and,
N.P.J Greer).
102. "Efficient sensitivity analysis for large non-ideal switched capacitor networks", Proc. I.E.E.E.
International Symposium on Circuits and Systems, Chicago, May 1993, pp. 1405-1407 (with
Z.Q. Shang).
103. "Design of high order sigma-delta modulators with minimum weighted noise", Proc. I.E.E.E.
International Symposium on Circuits and Systems, Chicago. May 1993, pp.180-183 (with
D.M. Hossack).
104. "Computer-aided design of high order sigma-delta ADC's", I.E.E. Colloquium on Advanced
A-D and D-A Conversion Techniques and Applications, London, May 1993, pp.2/1-2/5 (with
D . M . Hossack).
105. "XFILT: an X-window based modern filter and equaliser design system", Proc. European
Conference on Circuit Theory and Design", Davos, Switzerland, August 1993, pp.305-310
(with Lu Yue, R.K. Henderson).
106. "A comparison study of SC biquads in the realisation of SC filters", Proc. I.E.E. Saraga
Colloquium on Digital and Analogue Filters and Filtering Systems. London, November 1993,
pp.10/1-10/9 (with Lu Yue).
107. "Issues in the design of low oversampling ratio single bit sigma-delta modulators", Proc.
I.E.E. Colloquium on Oversampling Techniques and Sigma-Delta Modulation, London,
March 1994, pp.3/1-3/5 (with D.M. Hossack, J. Reid).
108. "Matrix methods for the design of transconductor ladder filters", Proc. I.E.E., Part G, Circuits
Devices and Systems, April 1994, Vol. 141, No. 2, pp.89-100 (with N.P.J. Greer,
R.K. Henderson, L i Ping).
109. "Efficient noise analysis methods for large non-ideal SC and SI circuits", Proc. I.E.E.E.
International Symposium on Circuits and Systems, London, June 1994, pp.5/565-5/568 (with
Z.Q. Shang).
110. "A comparison study of SC biquads in the realisation of SC filters", Proc. I.E.E.E.
International Symposium on Circuits and Systems, London, June 1994, pp.5/711-5/714 (with
Lu Yue).
112. "The application of redundant number systems to digital sigma-delta modulators", Proc.
I.E.E.E. International Symposium on Circuits and Systems, London, June 1994, pp.2/481-
2/484 (with D.M. Hossack).
113. "Efficient analysis of some non-linearities in SC and SI filter networks", Proc. I.E.E. Saraga
Colloquium on Digital and Analogue Filters and Filtering Systems, London, November 1994,
pp. 10/1-10/5 (with Z.Q. Shang).
114. "First or second generation SI cells? A comparison of sensitivity from an SI filter system
viewpoint", Proc. I.E.E. Saraga Colloquium on Digital and Analogue Filters and Filtering
Systems, London, November 1994, pp. 11/1-11/5 (with L u Yue).
Grants
Engineering & Physical Research Council Grant 1995-1997 £ 242306 (with S.P. Beaumont)
Guest lecturer at the same college on "The use of a Digital Computer in the Design and Analysis of
Networks" in a course on Modern Network Analysis and Synthesis.
Guest lecturer at the University of Leeds, Electronics Exhibition 1967, on "Computer Aided Design of
Electrical Networks".
Lectures to Institution of Electrical Engineers Local Centre:
Presented paper at Conference on Teaching Electronics in Degree Courses, University of Hull, April
1973. (Ref. 19)
Presented paper at I.E.E. Colloquium on Amplifier Design - Theory and Practice, London, May 1974.
(Ref. 25)
Presented paper at the Fourth Summer School on Circuit Theory, Prague, Czechoslovakia, September
1974. (Ref. 26)
Presented two papers at I.E.E.E. International Symposium on Circuits and Systems, Munich, April
1976. (Refs. 29, 30)
Gave lecture to I.E.E.E. Circuits and Systems Chapter, London, May 1976.
Presented paper at I.E.E.E. International Symposium on Circuits and Systems, New York, May 1978.
(Ref. 37)
Presented paper at Conference on Electronic Engineering in Degree Courses - Teaching for the 80's,
University of Hull, April 1980. (Ref. 42)
Gave lecture at Conference for Sixth Formers held under the auspices of the Mathematical
Association, Hull, April 1981.
Presented paper at I.E.E.E. International Symposium on Circuits and Systems, Rome, May 1982.
(Ref. 47)
Presented paper at I.E.R.E. Colloquium on Analogue Filters, London, November 1983. (Ref. 50)
Presented paper at I.E.E. Colloquium on Design Software, London, January 1984. (Ref. 51)
Presented paper at I.E.E.E. International Symposium on Circuits and Systems, Montreal, May 1984.
(Ref. 52)
Presented paper at I.E E. Colloquium on Electronic Filters, London, May 1985 (Ref. 55)
Presented paper at I.E.E.E. International Symposium on Circuits and Systems, Kyoto, June 1985
(Ref 57)
Presented paper at I.E.E. Colloquium on Analogue IC Design, London, November 1987 (Ref. 67)
Presented paper at I.E.E.E. International Symposium on Circuits and Systems, Helsinki, June 1988.
(Ref. 68)
Presented paper at I.E.E.E. International Symposium on Circuits and Systems, New Orleans, May
1990 (Ref. 83)
Presented papers at I.E.E.E. International Symposium on Circuits and Systems, Singapore, June 1991
(Refs.91,92)
Presented paper at I.E.E. Colloquium on Circuit Theory and DSP, London, February 1992 (Ref. 96)
Presented poster paper at I.E.E.E. International Symposium on Circuits and Systems, Chicago, May
1993 (Ref. 103)
Refereeing Duties
Referee for Institute of Electrical and Electronic Engineers Inc. (America) - Transactions on Circuits
and Systems; Proceedings of the International Symposium on Circuits and Systems, Munich (1968),
Montreal (1984), Helsinki (1988), Portland (1989), New Orleans (1990), Chicago (1993).
Referee for Institution of Electrical Engineers - Proceedings, Electronics Letters and Journal on
Circuits, Devices and Systems.
External Duties
I.E.E.E. Committee on Circuits and Systems Chapter for the UK and Republic of Ireland - Member
(1975 - present). Vice Chairman (1978 - 1980), Chairman (1980 - 1983).
I.E.E.E. Section Committee for the UK and Republic of Ireland - Member (1980 - 1983).
General Chairman of the Organising Committee, I.E.E.E. Conference on Computer Aided Design of
Electronic and Microwave Circuits and Systems, University of Hull, July 1977.
Joint-organiser of I.E.E.E. Workshop on Design and Fabrication of Integrated Circuit Filters, London,
September 1982.
Member of the Organising Committee, European Conference on Circuit Theory and Design, Brighton,
September 1989.
Member of the Editorial Board of the International Journal of Electronics 1989 - present.
1. "Table for the voltage transfer functions of single amplifier double ladder feedback
systems", Electronics Letters, May 1965, Vol.1, No.3, pp 70-71, (with A.G.J.Holt).
(50%)
3. "Table for the voltage transfer functions of single amplifier double ladder parallel
feedback systems", I.E.E.E. Trans. C.T., September 1966, Vol. CT-13, pp.326-328,
(with A.G.J. Holt). (50%)
5. "Matrix tables for the generalised 5-terminal amplifier", The Radio and Electronic
Engineer, April 1969, Vol. 37, No.4, pp.247-55, (with F.W.Stephenson). (50%)
8. "A simple method for producing floating inductors", Proc. I.E.E.E. December 1969,
Vol. 57, No. 12, pp. 2155-2156. (100%)
11. "A course in computer-aided electronic circuit design for undergraduates", Computer
Aided Design, April 1972, Vol 4, No.3, pp. 135-137, (with A.G. Martin). (50%)
12. "The multiterminal active transformer", International Journal of Control, July 1972,
Vol. 16, N o . l , pp. 171-175. (100%)
14. "Symbolic analysis for computer-aided circuit design - The interpolative approach",
I.E.E.E. Trans. C.T., November 1973, Vol. CT-20, No.6, pp. 738-741. (with J.K.
Fidler). (50%)
15. "Correction to symbolic analysis for computer-aided circuit design - the interpolative
approach", I.E.E.E. Trans. CAS, May 1974, Vol. CAS-21, No.3, P.462, (with
J.K.Fidler). (50%)
17. "An adaptive electronic circulator for use in telephones", I.E.E.E. Trans. COM, July
u
1975, Vol. COM-23, No.7, pp. 728-731, (with R. Lodge ). (70%)
18. "A theory of equivalent active networks", I.E.E.E. Trans. CAS, June 1976, Vol. CAS-
p
23, No.6, pp.350-354, (with L.G.Grant ). (50%)
19. "Network partition methods for use in symbolic analysis by interpolation", Computer
1
Aided Design, July 1976, Vol. 8, No.3, pp. 157-164, (with C. Fridas *). (50%)
20. "Symbolic network sensitivities using partition methods", Int. Journal of Electronics,
p
July 1976, Vol.41, No.l, pp.25-32, (with C. Fridas ). (50%)
22. "Some further sensitivity theorems in active equivalent network theory", I.E.E. Journal
on Electronic Circuits and Systems, November 1978, Vol.2, No.6, pp. 193-198, (with
p
J.Dunning-Davies and F.D. Faulkner ). (40%)
23. "An improved adaptive electronic circulator for telephone applications", I.E.E.E.
u
Trans. COM, August 1979, Vol. COM-27, No.8, pp. 1218-1224 (with H. Gazioglu
and D.A. HomerU). (40%)
24. "Analysis of active switched-capacitor networks", Proc. I.E.E.E. February 1980, Vol.
68, No.2, pp. 292-293. (100%)
26. "Digital active network analysis", Proc. I.E.E., Part G, Electronic Circuits and
p
Systems, August 1980, Vol.127, No. 4, pp. 181-186 (with F.G.A. Coupe ). (50%)
27. "A degree enhancement scheme in electronic engineering", Int. J. Elect. Enging.
Educ, January 1981, Vol.18, pp. 29-34. (100%)
28. "Compact matrix scheme for use in computer analysis of switched capacitor
networks", Electronics Letters, September 1982, Vol. 18, No. 19, pp. 840-841, (with
p
C. Lau ). (50%)
29. "Improved z plane polynomial interpolative analysis of switched capacitor networks",
I.E.E.E. Trans. CAS, July 1984, Vol. CAS-31, No.7, pp. 666-668 (with D.G.
p
Johnson ). (50%)
31. "Symbolic analysis of ideal and non-ideal switched capacitor networks", Proc. I.E.E.E.
International Symposium on Circuits and Systems, Kyoto, June 1985, pp. 1165-1168
p
(with C.K. Pun ). (50%)
32. "On the generation of equivalent switched capacitor networks", Proc. European
Conference on Circuit Theory and Design, Prague, September, 1985 pp.697-700 (with
D.J. Gilbert). (50%)
33. "Efficient computer techniques for the exact analysis of all non-ideal effects of
switched-capacitor networks in the time domain", Proc. I.E.E.E. International
Symposium on Circuits and Systems, San Jose, May 1986, pp. 373-376, (with L.B.
p
Wolovitz ). (50%)
35. "The LUD approach to switched-capacitor filter design", I.E.E.E. Trans. CAS, Dec.
p
1987, Vol. CAS-34, No. 12, pp.1611 - 1614 (with L i Ping ). (50%)
36. "General analysis of large linear switched capacitor networks", Proc. I.E.E. Part G,
Electronic Circuits and Systems, June 1988, Vol. 135, No.3, pp. 119 - 124, (with L.B.
p
Wolovitz ). (50%)
37. "Filter realisation by passive network simulation", Proc. I.E.E. Part G, Electronic
p
Circuits and Systems, August 1988, Vol. 135, No.4, pp. 167 - 176 (with L i Ping ).
(50%)
39. "High performance filter networks and symmetric matrix systems", Proc. I.E.E., Part
G, Circuits Devices and Systems, Dec. 1989, Vol. 136, No.6, pp. 327 - 336, (with L i
p
Ping ). (50%)
40. "400 MHz switching rate GaAs switched capacitor filter" Electronics Letters, March
1990, No.7, pp. 460 - 461 (with D.G. Haigh, C. Toumazou, S.J. Harrold, K. Steptoe
and R. Bayruns). (17%)
41 . "Active and digital ladder-based allpass filters", Proc. I.E.E., Part G, Circuits Devices
and Systems, December 1990, Vol.137, No.6, pp.439-445 (with Li Ping?). (50%)
42*. "Extended Remez algorithms for filter amplitude and group delay approximation",
Proc. I.E.E., Part G, Circuits Devices and Systems, June 1991, Vol.138, No.3, pp.289-
p
300 (with R. K. Henderson? and Li Ping ). (33%)
2 3
43*. "Canonical design of integrated ladder filters", Proc. I.E.E., Part G, Circuits Devices
and Systems, April 1991, Vol.138, No.2, pp.222-228 (with R.K. Henderson? and L i
Ping?). (33%)
44. "Design, optimisation and testing of a GaAs switched capacitor filter", I.E.E.E.Trans.
CAS, August 1991, CAS-38, No.8, pp.825-837 (with D.G. Haigh, C. Toumazou, S.J.
Harrold, K. Steptoe and R. Bayruns). (17%)
45. "A methodology for integrated ladder filter design", I.E.E.E.Trans. CAS, August 1991,
p
CAS-38, No.8, pp.853-868 (with Li PingP and R.K. Henderson ). (33%)
46. "A method for the evaluation of multirate sigma delta systems", Proc. I.E.E.E.
International Symposium on Circuits and Systems, San Diego, May 1992, pp. 1328-
1331, (with D.M. Hossack?). (50%)
48. "Analog integrated filter compilation", Journal of Analog Integrated Circuits and
Signal Processing", March 1993, Vol.3, pp.217-228 (with R.K. Henderson? and L i
PingP). (33%)
49. "Efficient sensitivity analysis for large non-ideal switched capacitor networks", Proc.
I.E.E.E. International Symposium on Circuits and Systems, Chicago, May 1993,
p
pp. 1405-1407 (with Z.Q. Shang ). (50%)
50. "Design of high order sigma-delta modulators with minimum weighted noise", Proc.
I.E.E.E. International Symposium on Circuits and Systems, Chicago, May 1993,
pp. 180-183 (with D.M. Hossack?). (50%)
51. "Issues in the design of low oversampling ratio single bit sigma-delta modulators",
Proc. I.E.E. Colloquium on Oversampling Techniques and Sigma-Delta Modulation,
p
London, March 1994, pp3/l-3/5 (with D.M. Hossack , J. Reid). (33%)
52. "Matrix methods for the design of transconductor ladder filters", Proc. I.E.E., Part G,
Circuits Devices and Systems, April 1994, Vol. 141, No. 2, pp.89-100, (with N.P.J.
Greer, R.K. Henderson?, L i Ping?). (25%)
53. "Efficient noise analysis methods for large non-ideal SC and SI circuits", Proc.
I.E.E.E. International Symposium on Circuits and Systems, London, June 1994,
pp.5/565-5/568 (with Z.Q. Shang?). (50%)
54. "A comparison study of SC biquads in the realisation of SC filters", Proc. I.E.E.E.
International Symposium on Circuits and Systems, London, June 1994, pp.5/711-
p
5/714 (with Lu Yue ). (50%)
58. "A systematic approach for ladder based switched-current filter design", accepted for
Proc. I.E.E.E. International Symposium on Circuits and Sytems, Seattle, May 1995
p
(with Lu Yue ). (50%)
59. "Multirate SC and SI filter system design by XFILT", accepted for Proc. I.E.E.E.
p
International Symposium on Circuits and Sytems, Seattle, May 1995 (with Lu Yue ).
(50%)
25
Page
publications in the learned society literature are quite diverse: some are pure research, some
currently exists; others reflect the quest to close obvious gaps created by the progress of
goal has always been to find solutions to real problems, in the hope that some of these might
translate through to commercial products or facilitate the design of electronic apparatus that
over many years, resulting in publication under joint authorship. Whilst it is impossible to
delineate precisely the balance of contribution, I was responsible in general for the
conception and initiation of the major themes of the work and have carried the task of
maintaining the continuity of topics over an extended period of time. A l l the work
Proportions of contribution have been attributed in summary list. I adopted the principle of
listing authors alphabetically, though occasionally where one student has contributed
(Papers 2, 6,7, 8, 10, 12, 18, 21, 22, 32, 34, 35, 37, 38, 39, 40, 41. 42, 43, 44, 45, 47, 48, 52,
The origins of active filter theory emerge from classified work carried out during
World War I I , though it was not until demands on circuit size were imposed by Space
exploration that the real stimulus for serious work appeared. The whole topic gained new
impetus from the requirements imposed by IC developments, a demand which persists to the
present time. The initial designs were committed to low frequency and audio band operation,
but this expanded over the years so that the range of frequencies being covered now
extends from below 1 Hz to 500 MHz and beyond. In the development of synthesis
techniques, the aim has been towards general applicability rather than specific solutions for a
particular frequency range or building block. However, various examples are given in the
works cited, where application to certain problems with specified fabrication technologies
The early papers [2, 6, 7] address the design of multi-loop feedback active filters, the
development of a general synthesis theory and the production of a number of new circuit
implementations. These realisations employed the voltage operational amplifier as the basic
active element. This has transpired to be a very significant operational element in the whole
of active filter work and v/hereas early work on analysis anticipated the emanation of other
amplifiers, it was many years before some of these were actually utilised. Three more early
papers [8, 10, 12] show further development of synthesis theories and various operational
three papers [18, 21, 22]. Prior to this, such theory had been limited to classical passive
networks and the only attempt at the extension to active topologies proved to be very
restricted. These papers develop the rigorous mathematics for the general equivalence
together with minimisation of component sensitivity, were reported- At a later date these
principles were developed further in the context of switched capacitor networks [32, 34] and
integrated circuit form, a major step forward was apparent in the design of filters for the mass
communication market. This technology has proved to be most reliable and extremely
robust. The initial designs were very closely linked to existing active-RC solutions and could
had an early attraction because of ease of design in a mathematical sense. It had been
established many years earlier, that active filter realisations, derived from a passive ladder
based synthesis was troubled by inexact simulation of termination resistances and by the
requirement for a strays-free realisation, in order to avoid the deleterious effects of stray
papers [ 35, 37, 38, 39, 41, 42, 43, 45, 47, 48, 52, 54, 56, 58] began to address the above
problems in the general synthesis of ladder-based SC filters. It rapidly became apparent that
not only was there a good solution to the difficulties associated with the existing ladder
approach (leapfrog), but a completely general matrix scheme could be formulated to describe
the whole synthesis process. Furthermore, many new realisations were forthcoming merely
by applying standard matrix decomposition techniques. The general theory was equally
applicable to active-RC and digital ladder-based filters [37, 39] and many new realisations
followed. Later, the same general theory was also found to be applicable to
Simultaneously, two other important developments were taking place. The first was the
extension of the realisation of ladder-based allpass filters from purely digital configurations
to active ones [41]. This circumvented the need for lattice type or differential structures and
retained the essential low sensitivity ladder property; a further benefit was that these
realisations were also encompassed within the matrix formulation. A second major
breakthrough was the development of a completely arbitrary approximation theory [42]
which could include all the properties of existing classical approximations, together with a
new high order touch point concept and general template weighting functions. The general
group delay. These developments not only allow a designer the freedom to explore many
unusual approximations whilst maintaining readability of the circuit, but other very
important practical consequences follow. Many low frequency stop-band zeros can be
forced to the origin, thus reducing the number of large time constants required in circuit
realisation. Specific, sharp notches produced by high order zeros can be incorporated to
feedthrough paths can be eliminated, thus breaking long delay free paths that cause restricted
settling times when many amplifiers appear in series connection. High order touch points
can be used within the passband to create "smoother" ripples which are particularly helpful at
the band edges in reducing circuit sensitivity and group-delay equalisation requirements.
Ripple characteristics can be tapered to reduce circuit sensitivities. Weighted passbands can
transmission lines. Most importantly, a new form of optimisation was now available in
which the order of the search space was not dictated by the number of circuit components,
but simply by the original order of approximation. Optimisation could be used to offset the
capacitors and even transistors from combinations of unit-sized elements in integrated circuit
layout. Thus a unified scheme based on the iterative Remez method was applicable to both
filter approximation and optimisation. A combination of various ideas also enabled progress
synthesised and then input circuitry of the resultant filter is adjusted to return the original
response [43, 56]. The climax of this work was a seminal paper on the methodology for
integrated ladder filter design [45] and the application of many of these principles to
industrial problems. One design [47] has been of noted commercial success with large
volume sales. The group of papers [41, 42, 43] were awarded the 1992 I E E J.J. Thompson
Premium.
The philosophy behind the software suite which incorporates the general synthesis
theories was outlined [48] and developed to a commercial level as X F I L T software. The
control of accuracy during the synthesis of high order networks is assured by methods
conventional and novel. Recent extensions include multirate SC and SI systems design [59],
when more attractive solutions can often be obtained for filters with particularly demanding
specifications, for example very narrow band ones. Powerful network analysis software [49,
53] has been incorporated and the suite can deliver full sensitivity and noise analysis to aid
the design task. A wide range of biquadratic sections has been added to provide a
Two other papers on filter design [40, 44] have been included in this section. These
describe the combined work of a team of investigators including the author, which produced,
in 1990, the highest frequency SC filter known to date. GaAs fabrication technology was
employed and many design obstacles needed solution. My contribution was extensive circuit
simulation using software outlined in the following section. The simulations revealed
limitations due to initial amplifier and switch designs and it was possible to incorporate
(Papers 1, 3,4, 5, 9, 13, 14, 15,16, 19, 20, 24, 25, 26, 28, 29, 30, 31, 33, 36,49, 53, 57).
Network analysis is the handmaiden of circuit synthesis and design. Traditionally the
linear analysis methods consisted of general matrix or topological techniques together with a
network and superposition theorems. Certain networks, such as ladders, lent themselves to
closed form solutions. Before the advent of computers, the circuit designer was forced into a
selection of special techniques and approximations for his analysis tools. Over the past 30
years, as computing power has become more readily available, it has been possible to observe
the move away from the analysis magician towards general matrix methods which are
capable of providing analysis to whatever degree of detail and accuracy is required by the
designer. The papers selected herein reflect this general trend quite accurately. The forward
progress has not always been a smooth transition from particular to general, as evidenced in
the case of SC networks. During the peak of activity in their development throughout the
early 1980's, a variant on the traditional approach appeared, when methods for computer
analysis of such networks were divided into ideal and non-ideal. The former assumed ideal
devices (switches and operational amplifiers) within the network and permitted
straightforward and quick solution for the first pass in design. When non-idealities (finite
switch resistance, finite GB values for the amplifiers) were allowed in the device
answers in a finite computing time. And how accurate were these solutions anyway? The
techniques and routines have been refined to such an extent that the user of a modest
workstation or PC can now expect to obtain reliable answers for complex noise or sensitivity
The work represented is mainly concerned with linear networks though, more
recently, extensions have been made into the analysis of dominant nonlinear effects in
switched networks.
The development of closed form solutions [1, 3, 9] for multi-loop feedback networks
class of network when the operational device is assumed ideal. The networks were important
active filters, whose synthesis is outlined in the papers of Section 4. A hybrid ideal/nonideal
solution [4] for these networks when non-ideal marcromodels for the amplifiers were
assumed, presented interesting principles for computer analysis that could be extended to
many other classes of network. Early work on the derivation of admittance matrices for all
four basic types of operational amplifier [5] pre-dated much of the application and design
work that has followed from many authors. An interesting manipulation of matrix
descriptors for networks of active circulators [13] provided not only a simple analysis tool
but also gave a compact method for examining novel interconnection strategies.
Symbolic analysis has been the bete noir of a considerable number of researchers for
many years. It is universally acknowledged that the production of fully symbolic transfer
parameters, could provide very attractive insight for the designer and the generation of
differentials for sensitivity and optimisation studies. But it was realised at an early stage, by
a large majority of investigators, that the exponential Cor worse) growth in size of expression
and the associated storage requirements would render the technique impractical for all but
small circuits. A number of researchers still continue such an approach and even with
"gardening" and truncation, the restriction to circuits of modest size persists. A number of
rather more realistic variations on symbolic techniques have been proposed and I have been
[14, 15, 16, 19, 20, 29, 57]. In this case the results of computer analysis are polynomials in s
or z or both variables, whose coefficients are numeric (though a small number of symbolic
components can be included at the cost of repeating the whole procedure for each element).
The first significant step was to establish the general application of interpolation techniques
to the efficient analysis of active networks [14, 15]. As the size of network was increased,
problems of accuracy arose and major progress towards resolving this issue was produced by
partitioned polynomial interpolation [16]. This utilises sample points located in a circular
manner and involves circles of different radii or sets of samples on the same circle with
different initial set angles or combinations of both. The work was important for two reasons;
it was an elegant solution to the accuracy problem and secondly the choice of circular
samples prompted other investigators to realise that the Fast Fourier Transform (FFT) was
the exact implementation of the interpolation scheme when a circle of unit radius was used.
This meant that all of the highly efficient FFT routines could be used as the basic numerical
engine for semi-literal interpolative analysis, a situation which prevails to the present time.
Other techniques, such as network partition methods [19, 20] were developed to cope with
the problems of large networks. These contributed to further improvements in accuracy and
storage improvements, together with features useful in the computation of symbolic network
modest size by interpolation in the z domain, raised more severe accuracy concerns than had
transformation of the plane and partitioned polynomial interpolation again proved to be very
successful in maintaining accuracy during the analysis of high order SC networks [29]. The
all straightforward. An early attempt using a Kader sequence for matrix inversion [31]
enjoyed some success, but application was limited to circuits of moderate order because of
the computing resources needed. The system was considered as a general one with
polynomials in s and z. The problem has been reconsidered very recently [57], when an
interpolative approach has once again been used to advantage in the general analysis of non-
ideal SC and SI networks. Essentially there is no closed form symbolic expression for the
transfer function of a non-ideal switched network, since the computation of an extended state
transition matrix is implicit to all solutions and at best this can only be approximated in a
of slightly higher order than those expected in the ideal case, are produced to describe exactly
the network behaviour [57]. Accuracy has been maintained during the analysis of large non-
ideal SC and SI networks. It is also shown that significant advantages can result when this
technique is used in noise analysis of switched networks with 2 clock phases. All the aliasing
effects of wide band noise are easily computed by numerical substitution of a large number
of frequency values into the polynomials. Unfortunately such advantages are quickly eroded
As already indicated, the introduction of SC networks during the late 1970's and
early 1980's brought great demands for comprehensive analysis. Initially many schemes
abounded, even for ideal networks, and I was able to make contributions to the early work
[24]. This paper was also chosen as a selected reprint [30] by I E E E Press for a volume on
significant papers in the subject. Translation of analytical techniques into CAD tools was an
obvious development, but the success story associated with SC filters in particular
circuits. This quickly brought about the demise of much of the less durable analysis
software and triggered the search for improved mathematical and algorithmic routines. A
compaction scheme [28] was very effective in coping with multi-phase higher order SC
to simple circuits [25]. A series of papers [33, 36, 49, 53] made a serious attack on the
compendium of numerical techniques were fully exploited, including sparse matrix methods,
of the non-ideal switched network description. The resultant software has gained a reputation
for speed, accuracy and robustness, the current version of SCNAP4 is in use at a number of
industrial and university sites around the world and negotiations for full commercial
exploitation are in hand with two companies. The analysis of the largest known multi-rate
switched networks to give frequency responses, sensitivity behaviour and noise performance
can now be accomplished on modest work stations and PCs in very realistic computational
times.
An investigation into the description and design of digital active networks [26]
produced very interesting results. However these have been largely superceded by the
irrepressible development of digital filters and DSP on the one hand and by SC and SI
These two papers report work on practical circuit design for telephone applications.
In many telephone networks worldwide, problems occur when matching the wide variations
in local subscriber lines caused by different line lengths, line material (Cu or Al), conductor
dimensions and mixtures of all of these in different proportions. In rural areas the length of
local lines can aggravate the problem. In urban districts the runs are shorter and
approximately equivalent in length, but the actual lines may have been modified so often
with odd pieces of cable of such random specifications that the resultant line characteristics
often exhibit considerable variations. The adaptive electronic circulator was developed for
use in these situations, as an effective means for providing isolation between transmit and
receive signals and a constant level of side-tone for user comfort. Basic problems of voice
switching were addressed by extra feedback. The initial work was confined to resistive
matching only. In the second paper, a more general adaptive impedance match was
successfully demonstrated, low and high frequency pilot tones being used to derive error
signals for two term control. The voice switching problem was eliminated and good adaptive
performance was demonstrated over a wide frequency range for a great variety of subscriber
The advantage of these circulator circuits over traditional passive hybrids was their
ease of realisation as part of any integrated circuitry associated with a modern telephone set.
interface problems, they permit shaping of the noise characteristic, provide multi-bit
conversion and above all they can be realised in integrated circuit form without the need for
One of the major problems in this area of work is the complexity of simulating the
behaviour of multirate sigma delta systems. A general behavioural simulator was developed
[46] with an emphasis on providing a user-friendly interface. The software FUNSIM has
been used extensively in one industrial establishment. It has also facilitated studies in the
design of high order sigma-delta modulators with tailored noise characteristics [50] suitable
utilised techniques for adaptive step size to give companding and have developed ideas on
noise gating. It was also shown that when these ideas are applied to the problem of speech
coding/decoding, the resultant system can give significant performance improvement over
More recent work has been concerned with digital sigma-delta modulators [55], the
speeds of which are limited by latency in loop filters. An interesting solution has been
complement number system. There seems to be scope for many further developments and
(Papers 11,27)
These papers describe two course developments in which I was involved. In the early
1970's the introduction of CAD into the teaching of electronic circuits began to receive
attention. Unfortunately many of the schemes were piecemeal and somewhat ad hoc. The
first paper [11] details a scheme to integrate a student's acquisition of theoretical knowledge
from lectures, with his practical experience from laboratory work and attempts to force a
convergence of these skills through the medium of a 4 week design project. Further
motivation derived from the use of analysis software to evaluate designs in a controlled
manner. The whole exercise proved to be very successful with students and ran with early
analysis software for a period of 10 years, until more general software became common
place.
The second paper [27] describes a different educational initiative. In the late 1970's
there were numerous high level reports on how to improve the quality of engineering
graduates. At the University of Hull, I chaired a working party which had the task of
producing a realistic scheme for an enhanced degree in electronic engineering. The scheme
enjoyed considerable success over a number of years. However, it is interesting to note that
many of the premises on which that initiative was based have been eroded over the
intervening years and many of the caveats mentioned at the conclusion of the paper have
been ignored. The standard of education of electronic engineers in UK might well be sinking
This set of collected papers covers a period of 30 years, during which electronics has
developed from an advanced art into a highly sophisticated one. Surely it must be one of the
few experiences of modern existence, when after such a length of time, more functionality
can be obtained at less cost. Furthermore, there is no sign that these trends have run their
And more is in store across the landscape presented in this work. The human form
operates almost exclusively in an analogue mode and relates to a largely analogue, physical
world by analogue means. Yet the most powerful forms of collecting data, processing
information and stimulating response lines reside firmly in the digital domain. The way
forward is not by a frantic search for remaining "analogue gaps" and the development of
electronic solutions for these niche areas only. It must be via fully engineered interfaces
between analogue and digital environments and technologies - completely integrated designs
where the advantages of each domain are utilised to maximum effect. As the possibilities of
truly mixed-mode design unfold, the demands for new simulation facilities increase, which in
Many of the techniques reported in this work and the resultant software packages
have been applied to the design of signal conditioning circuits whose frequencies of
operation lie in the frequency range 1 Hz to 500 MHz. We are already addressing the
problems associated with designs operating at 1 GHz and in the near future 12 GHz should
be within reach. The challenge is to develop proper engineering design techniques and
Publication Page
[22] Some further sensitivity theorems in active equivalent network theory 110
[39] High performance filter networks and symmetric matrix systems 148
39
Page
[40] 400 MHz switching rate GaAs switched capacitor filter 159
[42] Extended Remez algorithms for filter amplitude and group delay approximation 170
[44] Design, optimisation and testing of a GaAs switched capacitor filter 191
[52] Matrix methods for the design of transconductor ladder filters 241
PAPER 2
amplifier to obtain biquadratic responses
A . G . J . Holt, P h . D . , A . M . I . E . E . , and J . I. S e w e l l , B . S c , Graduate I . E . E .
Synopsis
In this paper, networks are presented in tabular form to simplify the design of RC active filters having
magnitude responses characterised by pass-band attenuation ripples and stop-band transmission zeros.
The responses of these networks may be described by biquadratic rational functions. In all, the form and
element values for 11 different networks are tabulated.
The basic network discussed in the paper is an operational amplifier with multiloop feedback in cascade
with a passive RC circuit and having a second-order response. Higher-order responses than the second may
be obtained by cascading second-order stages of the tabulated networks, with suitable isolating stages
when necessary.
Two different forms of passive circuit are given, together with their design formulas. Both low-pass and
high-pass filters are considered; these can be cascaded to give band-pass or band-stop characteristics.
The operational amplifier is assumed to have infinite input and zero output impedance; the effects of
departure from these ideal conditions on the filter response is considered.
ot, = constant attenuation coefficient for net- procedure has been presented by Holt and Stephenson.
work design from Reference 15 ft is known that parallel ladder networks may be somewhat
D — high-pass transfer function difficult to adjust, and a method which reduces their number,
HP
M — amplifier open-loop gain while yielding the same form of response, would appear to be
advantageous.
The methods of design described in this paper are intended
1 Introduction to provide networks having responses described by elliptic
functions of the second order; other biquadratic responses
Filters having responses described by biquadratic
may be obtained by the same method. Responses described by
rational functions are well known in passive-/LC-network
higher-order functions may be obtained by cascading second-
design. These networks may have transmission zeros (often
order stages.
known as poles of attenuation) at finite frequencies which lie
on the imaginary axis of the p plane. The responses of
certain classes of these networks may be described by the
elliptic functions.
2 Principles
The derivation of all transfer functions considered in
Paper 4855 E , firs! received 27th August I964 and in final form 5th this paper is based on the cascaded sections shown in Fig. I.
August 1965 An operational amplifier and multiloop feedback circuit, as
Dr. Holt and Mr. Sewell are with the Department of Electrical Engineer-
ing. University of Newcastle upon Tyne, Newcastle upon Tyne 1 shown in Fig. 2. are included in the active network. It is
PROC I EE. Vol. 112. No. 12. DECEMBER 1965 2227
assumed lhat the input impedance of ..he amplifier is infinite a simple result is with a parallel RC network and )', • 1/R,.
and its output impedance is zero, conditions which are easily- Y -- pC .
2 K, - p C j . > = l / / ? and >' - C\ - l / R .
: 4 4 ? P 5
active passive
R,
2
B »B,p*p
0 a * P
0
2
2
b *b,p+p
0
2
8 *B,p*p
0
(b)
This gives
2 _ bt_
(8)
<?2 b, 2b -
0
Fig. 2
Basic active multiloop feedback section
2
(9)
b0 ^ b l P ->- p
(10)
2
(1)
p ^ b l P
•x.
(N
r-j
30
r-4 C
v.
O .as
U'as Iw- a:
O =s
^ a;
(_i as ^ L, as U ax
O as O a:
a: i
I
Some solutions require the choice of an arbitrary variable to 3.3 Denormalisation
satisfy readability conditions. It is frequently advantageous to denormalise the
Configurations 1 and 5 in Table 1 give simple solutions, active and passive sections to different impedance levels.
having four standard elements using only two resistive This enables the designer to obtain convenient component
elements with nonunity normalised values. values and also avoid distortion due to drawing large currents
from the amplifier. Fig. 3B shows the circuit designed for a
3.2 E x a m p l e u s i n g four s t a n d a r d e l e m e n t s cutoff frequency of lOOc/s: R = l 5 9 - 2 k £ } . C == 0 01 tiF for
Fig. 3C shows the measured response obtained when the active section, and R = 15 -92kQ and C = 0-1 uF for the
solution 5 from Table ! is used to realise the typical example passive section.
of a second-order function:
3.4 E x a m p l e using three standard e l e m e n t s
pi 7.464
D = 2 (12) Consider alternative 4 (Table I). Let C , = 20/^F, and,
p J - 0-6075p - 0-7559 • • • •
for the function of eqn. 12, R, = 0-5902kfl and R = 2
Fig. 3A shows the passive networks. In the active section, the 01067kfi.
calculated element values are R = 3-921, R = 0 - 4 0 1 9 (to
t x The numerator
four significant figures).
2
F(p) = 0-5532 - 10-4203/j - p . . . . (13)
0 0 3 6 8 0 01022 O3770 01682 02420 02195
o VW 1| « |H o-W—«JVWUJ(W o choosing
656kO
F i g . 3B Fig. 4
Complete circuit for example of Section 3.2 Complete circuit for example of Section }.4
20
+10
til Ul
o -10 \
\
20 \
30
\ \f T
+
Y
40
10 100 1000
freauency. c/s
F i g . 3C
Response of active, passive and combined sections of example of
Section 3 2
a Active section
b Resultant response
c Passive section
2230 PROC. I EE, Vol. 112, No. 12, DECEMBER 1965
The transfer function for the passive section is 3.5 Circuits with one series RC element
1 Table 2 shows solutions which are applicable for the
p - 7-464
D 2
(15) range of functions under consideration.
p - 10-4203/) - 0-5532 Some solutions are more complex than others and require
the choice of one or more arbitrary variable components in
With these coefficients, it is convenient to use the synthesis
order to satisfy readability conditions.
technique of Reference 15.
For 93 9jkr>
«o ~ P2
1 16-95 kf) 01 kn
A 2
( ~) 0062
It
Fig. 6
C 2 = 0 5» Complete circuit for the filter using method of Reference 15 for the
passive network
(17)
^3 •= ~
Ry = For example, taking alternative 3, again the required trans-
fer function is
2
o -f 7-464
D (21)
C - ^
GV") 0-6075/j -f 0 7559
7-464
Fig. 5
2-3408
Basic canonical form of the syiillw.sk procedure of Reference 15
; 5•5429
6
The second network produces a response with a slightly •- 1
greater gain; this is due to the different constant multiplier Continuing the synthesis yields the network shown in
produced by each synthesis method. Fig. 7. The passive section is denormalised to R = 15-92kfl,
The passive function produced by these methods has the C = 0 1/xF, and the active section to /? = 159-2 k i l ,
form C =-- 0 01 fiF.
«o ~ P 1
2
(18) 26-BhO OOl^lF
°" ^ B a ^ B l P - p rAW |h
187-4 0 0014 O-0341
For the first (Guillemin) network, 159-2kfl 6366kf) kfl uF
i—WA
i—wHP
„ >}B 0
63-61 kO 28 23kO
(19)
B 0 ~ '•.<'() OOlpF 0O324
where
5595 r 00185
kO
BoK 0
Fig. 7
K B,n K\Bn KnB, K, Bn
Bo
(20) The response of this configuration is identical with previous
x = x-> — -
methods, apart from a constant gain factor.
" "o
The ratio 3t /a, gives the gain difference, for this example
3
3.6 Effect of finite input and nonzero output
impedances of amplifier on filter response
I 148
As stated in Section 2, the theory presented in this
i.e. 1 -208dB. paper was derived on the assumptions that the input impe-
The complete tiller consisting of feedback amplifier in dance of the operational amplifier is infinite, and it's output
cascade with the loaded twin-T network is shown in Fig. 6. 2231
PROC I EE. Vol. 112. So. 12. DECEMBER 1965
O -f
O
a;
0- ft;
- I O +1
+ —I o
O
I
OS as
OS o O ' „ i
O
^ -lo §
1 as II it
a;
O
-la:
a;
o as
o a;
O la;
a O O
O
O o" O
O
aslaT O as la; M
ft,
o ft.
'Si. o ; |os + + +
os let o + o< a ;
as ft; m
+ o O O
1
Olas
Cla; O l a ;
Cs. o + +
Ol +
+ a. + +
Bi a ;
OS las T + + +
z ola; GV G!=c O
Ola;
o
+ ri m aslas
p
< ri ui + + Oia-
as oc ClaT +
Ola; —i 1 + Ola; Ola;
z o Olas
2 OS +
s + a ; a ; Gla? -r
O la;
+ +
o Ola; + ok + osla;
+ Olas
Ola;
o la;
asla;
GV + O Ics o'k a.
las
a.
Olas
Ola?
+ + + +
+ Ola; O ias
ft.
+ +
a.
+ + +
las las
la? -|o?l
i las Ha|
a;
3 C
-1°=
O
o.
O
cs. o
10! a;
o!o II
_ 1+
h 0 - b p -
t p-
reduced input
impedance
substituting the high-pass transformation p --
cn-20
( i - , 0
\an 1 (24)
"(2)-
freauency, c/5
(.25)
Fig. 8 V ' 2
Effect of i ciliu ini! atnpli/lcr input impedance on tiller response Substituting the coefficients gives the numerator
F = 1 -6458 - 4-2916p - p-
measured response of the filter circuit of Fig. 3B when the
amplifier input impedance is 50k£>. together with the curve = (p - 3 - 8 7 ) 0 - 0-4261) (26)
for the unmodified amplifier. It is clear from these curves
that an input impedance as low as 50kU has very little effect using an arbitrary polynomial,
on the response. Another test was carried out to find the effect q = (/> - !>(/> - 5) (27)
of increasing the output impedance. A resistor was connected
in series with the amplifier output at the position marked A and continuing the synthesis gives the network shown in
in Fig. 3B. N O measurable effect on the filter response was Fig. 9.
observed until this resistor was increased to about I0k£2. OOIJJF
It was concluded that departures from the ideal values of
159-2kf) O0351JJF
input and output impedances have little effect on the filter
396 inn 9959RO / 04573IJF
response. W — 1
,-v.v—II-, Ih
The results discussed above were obtained with an amplifier
I M 2kO 89 36kO
circuit using valves. A transistor amplifier may be expected — sjvs—
to have a lower input impedance than a valve amplifier. In 00108.
F
3 N A I - T U I N G T ' E N G : 'Synthesis o f electrical filters w i t h two section RC filters having 2nd- and 4th-order Chebyshev responses in pass and
circuits', ibid., 1963, Pt. 1, (4), p. 47 stop bands', Proc. JEE, 1964, 111, (11), p. 1807
PAPER 6
INT. J. CONTROL, 1969.. VOL. 9, NO. 5, 483-497
A . G. J . H O L T arid J . I . S E W E L L i
Department of Electrical Engineering, University of Newcastle upon Tvne,
England
This paper presents a general synthesis method for multiple loop feedback systems
which is applicable to the realization of all the known configurations together with
othc s resulting from this investigation. The active device may be an amplifier of an;'
type, gain or phase characteristics.
The basic method utilizes an extended node introduction synthesis theory, which
facilitates the generation of many possible realizations for a given function.
Although the theory developed is used to synthesize R C multi-loop feedback
structures, it is equally applicable to the synthesis of R L C multi-loop feedback systems.
Rules for the numbers of active and passive node introductions are given, and a
number of examples illustrate the implementation of the technique and the various
transformations required.
Synthesis of various functions is considered, emphasis being laid upon the
realization of those which provide filter characteristics.
1. Introduction
M u l t i p l e loop R C feedback systems i n w h i c h the constraining device is a single
amplifier have eluded rigorous synthesis f o r some t i m e . A considerable n u m b e r
and v a r i e t y o f circuits belonging t o this class o f system have been reported i n
the l i t e r a t u r e (Aggarwal 1963, B r i d g m a n a n d B r e n n a n 1957, Nichols a n d R a u c h
1954, S h u m a r d 1960, W a d h w a 1962). The m a j o r i t y o f papers so f a r . w i t h the
exception o f ( H a z o n y a n d Joseph 1964. M i t r a 1967, Pande a n d Shukla 1965)
have been directed towards o b t a i n i n g these networks b y a t r i a l a n d error m e t h o d
based on analysis; a n e t w o r k f o r m h a v i n g been assumed a priori a n d t h e n
f o r m u l a e f o r the n e t w o r k elements to simulate various transfer f u n c t i o n s
developed. T h i s approach has m a n y obvious drawbacks a n d a systematic
synthesis procedure f o r these systems w o u l d u n d o u b t e d l y be a great i m p r o v e -
ment,
The methods o u t l i n e d i n H a z o n y and Joseph (1964), M i t r a (1967), Pande
a n d Shukla (1965) present techniques applicable o n l y t o i n d i v i d u a l systems
a n d rely u p o n an analysis step w h i c h reduces the p r o b l e m t o one o f passive
synthesis.
The synthesis procedure to be described is capable o f p r o d u c i n g networks
h a v i n g useful filter responses and provides a general synthesis m e t h o d f o r systems
constrained w i t h one amplifier irrespective of its t y p e , gain, or phase characteristic.
This embraces all the present k n o w n single amplifier feedback type filters and
others produced as a consequence o f this i n v e s t i g a t i o n , w h i c h include systems
cox.
484 A.G.J. Holt and J . I . Sewell
Fig. 1
General structure.
General structure.
(b)
(a) Single ladder active multi-loop structure, (b) Double ladder active m u l t i
structure.
5 2
(1)
1 n \ n
Um
(2)
X{P) D(P)
I Vnl linn
Q(P) Q(P)
Q(p) being some a r b i t r a r y divisor p o l y n o m i a l .
F o r the present systems the t w o p o r t m a t r i x w i l l assume a general f o r m :
1 n
II n
Y = (3)
n L x x j
and the i d e n t i f i c a t i o n o f polynomials and a d m i t t a n c e parameters as i n (1), (2)
cannot be achieved.
To overcome this d i l e m m a , consider another m a t r i x , say } ' * w h i c h
characterizes the performance between nodes 1 a n d k o f an n node n e t w o r k , i.e.
1 k
( 1
1 Un Uu.
y* > (4)
k i Ui.i Ui.k j
2L 2
486 A.G. J.Holt and J . I . Sewell
: ! l
V' kk~JlO £> ~J z( n' 'n)'
F o r single a n d double ladder systems this linear relationship can be s i m p l y
expressed b y a suitable divisor p o l y n o m i a l , i.e. the f o l l o w i n g i d e n t i f i c a t i o n
f r o m ( 4 ) , ( 1 ) is possible :
1 k
1
{ .'/n* .'/,/, '
r
, j v »(P) \ • ( 5 )
n y.n (6)
l
' y n-i, n-i y n-1. H !
J,a V,,i
,_y_n. »—1 y mi j
CO 0]
r= | (?)
5 6
2
and are a r b i t r a r y i n this example ; Y = (p' + b p + b ):Q(p). K 2 x 0
and g i — # 1 2 = Y g i v i n g the m a t r i x :
2 2 l
1 2 3
1 ^ Oil 012-012 0
y , (!?32-023)(!723-023)
021 — &12 023 _
^23
033 + ^ 33
A p p l y i n g t h i s i n t r o d u c t i o n gives :
1 2 3 4
v - (!7» -^.i)((/«-««)
2
f/ i-«i T^(9. ' + 1V)
2 2
1
0-12 1 , '^32
S
4- I ~ 173 3 • ^ .13
3 0 #33 ~T i 33 ^32^23 - » 2 3 '
i 2 3 4
1 ^ on -012 0 0
3 0 -0,2' 033+ ^ 33 0
4 0 - 6" 23 Y' + K33
490 A.G.J. Holt and J . I . Seivell
0 #23
= 9 V2 +6 12 + 0 ".
23 (9)
9iP + %
Various restrictions on component spread and size may now be imposed. For
instance, i f i t is required to have all unity normalized resistors, this will give
Now,
Let 6 ", 6 be frequency dependent terms, and these are equal because of
23 23
p- +p{b 1 + 6 ") + b -
23 0 g 23 = (q lP + q )(d 0 12 + B ' + 0 ")
12 23
2
= VAP +P{h + ? l [ » 1 2 - #23]) + ^12 - #23- (11)
0
Comparing coefficients of (11) give :
?A=1, (12 a)
b + B ," = e + q [e -g l
l i i l i2 23 (12 6)
If
1
(7(p)= 2
1 + l-62;p+p '
a n
Setting #23= — 1 d solving 12 («. — c ) gives the following matrix:
1 2 3 4
1 1 -1 0 0
o -1 2+1-24^ -1 -1-242J
=
•1-4
3 0 -1 14-0-812? 0
Fig. 3
=r 1-24
i i 2
O AAAAA/VV-
E. O.BIO
"out
Second-order Sallen and K e y circuit. For all circuit realizations the element values
are in mhos and farads.
1
1
£/n l/l2
(13)
3 2
^ p + bp t + bp + b
1 0
2/21
Q(P)
r = = :
F o r node i n t r o d u c t i o n s i n (13) 3 n 0 !/ii*> ^ ' i 2 2/i2*> ^ 21 = 2/21*
_ pt + bvpt + btf + bo
022
m •
=
A p p l y i n g a passive i n t r o d u c t i o n w i t h Y l 2 = cy — 0 , F i
12 12 2 (J21 ~ ^12 g i v e s :
1 0 3
^ Oil 0
Y — 0 y ! Uhi-023.(92*-023,)
1
i"21 U23 ~ ^23
33 ~ - #33 + ^ 33
0 (Ii3 - #33 + ^ 33
492 A.G. J. Holt and J . I . Seicell
3 3
^ nn -0 1 2
-- - -[tf,2+}'l2']
L
V '^23^23 , "^23
-012
1
022 x + \ - 0 23 +A'A 23
A A
33 33
1 011 Y l 2
0 0
3 0 - 0 . 2 ' <?33 ~r J 33 ,
^ 3 3
4 0 0 33 ~~T-K\ 3:l
m a t r i x o f e q n . ( 1 4 ) results.
1 2 3 4
nil Yx*. 0 i) 0
1"« (1
-9, =' — y
;/.n - 1
3 3
(14)
u -0, ' 3
?.,3 -r 1 33 ,
'V|3
(I 0
0 — y -A.,
Setting T* = 3~ = J /
12 21 12 and collecting the various parameter relationships:
let
032 = 0,
Xode 4
This is the amplifier input node. i.e. a v i r t u a l earth point and therefore
indefinite dominance is desired, t h u s :
Node 3
D e f i n i t e dominance
a > a
(i u 2 \ u\j (where each a is a m a t r i x element)
3?!
#33 + Y 3 3 - ^ l
= V + 0i " + A 2 23 + 0'. (16)
33
Xode 2
D e f i n i t e dominance is also possible here; introduce a t e r m 8" t o allow f o r
this:
Y 023(^33-023) f. , , n„ -v
i o 2 2
„ , v — =
y
i 2 + y + 2/i2 + • y
('•<)
1
#33 + 33
S u b s t i t u t i n g i n (16) f r o m ( 1 7 ) f o r the t e r m g 33 + T 3 3 gives:
A 2 A ,
v (012^ - 330l2" )/ 23023 a , , , , „ , a» 11 n\
= + + + e
" (A33/A2 023')(0 2' + 012" + 0 ' + A 3 1 2 3 ) l + "» r 'J« ' ^
' \ j 3 = / Q » ) = *'33;P-
(iii) A l l resistances i n the f i n a l n e t w o r k are equal i n value, i.e.
, / a i = 0 ' = 0 " = A = y = 0.
12 12 23
^23^23
l A t/2 3 2 3 A W
23 23 J J
p* + b p* + b p + b = k |V +
2 1 0 {30U-330 + 30(9'A- } 33
^23^23 '^23^23
0 A r j ' "1
2 3 2 3
6'6"k , 3 J •
_ 40A fl ' 2 3 2 3
b = W + 80 2
(22)
1
^'^"33 S'6"
7
F o r the numerical f u n c t i o n given, 6" is chosen as u n i t } . Solving eqns. (15),
(21)-(23) then yields t h e m a t r i x :
1 3 4 0
1 01992 -0-1992 0 0 0
o -01992 p -0-5976 -0-1992 0 -01992
Fig. ±
0 - 1 9 92 01992 0-1992
WJW^- 0 VW\W
n out
4. Conclusion
The synthesis technique o u t l i n e d provides a m e t h o d whereby m u l t i - l o o p
feedback systems m a y be synthesized i n a rigorous manner. A s the m e t h o d
requires the d e t e r m i n a t i o n o f one or more roots o f a p o l y n o m i a l o f order equal
t o the denominator o f the characteristic f u n c t i o n , high-order n e t w o r k f u n c t i o n s
d e m a n d the use o f a d i g i t a l computer. W i t h increase i n order, the number
a n d c o m p l e x i t y o f node i n t r o d u c t i o n s increase a n d the t r a n s f o r m a t i o n w o r k is
also m u l t i p l i e d .
The synthesis technique p e r m i t s the inclusion of a n y t y p e o f amplifier, e.g.
voltage, current t r a n s a d m i t t a n c e . transimpedance, finite or i n f i n i t e gain, as
the constraining device o f a m u l t i - l o o p feedback n e t w o r k . W h e t h e r the device
is ideal or n o t is o f no consequence as i t is possible to incorporate the non-idealities
i n the synthesis.
The m e t h o d is applicable t o the realization o f b o t h transfer ratios and d r i v i n g
p o i n t imittances.
Constraints concerning component ratios have been used here, b u t t h e
i n t r o d u c t i o n o f others d e t e r m i n i n g s e n s i t i v i t y is also a feasible extension.
As w i t h m a n y synthesis techniques there are usually numerous physical
realizations of a specific f u n c t i o n , the examples shown here illustrate some
realizations o f various transfer f u n c t i o n s . I t was indicated at various points
d u r i n g the synthesis o f these networks t h a t , h a d other identifications been made,
d i f f e r e n t networks w o u l d result. The m e t h o d does n o t restrict realization t o one
p a r t i c u l a r f o r m , b u t can be used t o display d i f f e r e n t ones.
The f u n c t i o n s synthesized here a l l have constant numerators, a n d therefore
no emphasis was l a i d u p o n n u m e r a t o r realization. N u m e r a t o r inclusion is the
subject considered i n a f u r t h e r c o m m u n i c a t i o n .
ACKNOWLEDGMENT
Appendix
T h e d e r i v a t i o n o f the required node i n t r o d u c t i o n s is presented here.
F o l l o w i n g the n o r m a l d e r i v a t i o n (Piercey 1962). i.e. commencing w i t h a 2 x 2
matrix:
IIL1 ^ 12
Y,.,=
I Y, 1 022
using a 1-node i n t r o d u c t i o n g i v e s :
^ 33 — }',! + «! J , Y 0 2 2 + a. A, 2 a A,
3 (24)
M s
a,A 3 « 3 ^ 3
f /
3 ^ 3 — 033 + 1 33'
f _
'2^3 = J/32 ^23 !
(25)
fl J
M i = ffia- ia- *i2. !
Ct A 2 x
=
f?21 ^12 —
21 • j
where t h e (/"s are frequency independent terms, ffs a n d Y s can be f u n c t i o n s o f
the complex f r e q u e n c y variable.
B y s u b s t i t u t i o n the r e m a i n i n g u n k n o w n s m a y be d e t e r m i n e d :
(#32 - ^ s X f e - #23)
a A.
2 2
.- )
033 + * 33
g
(gai ~ i 2 ~ y ) ( g 2 1 3 a ~ 0 )(033 + 23 y 3 3 )
(032 ~ ^ 3 ) ( f e - ^ 2 3 )
( g l 2 - ^ 1 2 ~ ^12)(y23-^23)(!733 + ^33)
« 3 ^ 3 = A A + 3' 3 3 !
3 3
«.,.4 ='W^' - 3
aA 3 2 = A.vj/v - ^23-
a,A v = A 1 2 A - #12- r
= A A - 012- Y
2 1
21-
Synthesis of multiple loop feedback systems 491
A f t e r expansion
K A 2 3 A , 3 (Ao3^23 '
a.,A,,=
A
A
33 33
« -4i =
3
A
33l"#21 + ^12]) +
A
*i 3 3
.
A
32
A
32 32"
V A A
^ 21 33 ,
A y
a i - J 3 = \ 1 33L 21-T 1 2lJ) + ••
A
A
23 A
23 23~
A AjoAojAjg A
33
(A 1 2 [0 1 3 +r 2 1 ]+A 2 1 [0 +rK 1 2 ])
A03 A 3 2 A A
23 32
A A. A
1 2 2 1 3 3 / ^23 #23 N
+ +
A
A
23 32 A
V#23 32>
REFERENCES
PAPER 7
I N T . J . C O N T R O L . 1909. V O L . 9. N O . .3. 4!I!I-5US
A. G. J. H O L T and J . I . S E W ELL:!:
Department of Electrical Engineering, University of Newcastle upon Tyne-
England
1. Introduction
The general method of synthesis of m u l t i - l o o p feedback systems has already
been described ( H o l t and Sewell 1.968). w i t h particular emphasis on the synthesis
of functions w i t h constant numerators. M a n y useful network functions have
numerators t h a t are polynomials and i t is therefore necessary to include the
realization of a numerator polynomial i n the synthesis.
cm- '"'
1 •>
3 4
] u U
' mi -f.-.
o v '^23^23 ;,
-".2 -• 23
r ..,=
4
o
O 0 -fl->'
4 0 \
•-2,' '
''2'!
= P + V\» + ' W
2
Setting (A g0 /A\, ) = 3 w i l l give A + 0, ' = 3- i f 0,./ = 0. then A.,., - 3A., + 3/.\ = 0 ;
2 23 3 23 2 rs :i i;i
Fig.
,
0 , = 2 5. I f the numerator co-factor is now evaluated i t appears t o f a i l to
1
Fig. 2
i
I
A/vWW <>
0-3
II
in
a p { ) = y t M .
{ 1 )
' , f , D{p) '
where Xj are single or groups o f numerator terms, r being their number.
2M2
69
Fie
cut
«)
1
• U
1 out
9 I
Double ladder active multi-loop structure, (a) General st ructure. (//) One realization
Synthesis of multiple loop fiidback systems 503
1 k
1
= Dip) '-'
W i t h k = 2 the m a t r i x is :
.'/12*
r*
J L\l') i
C M
If
2
P + bp + b 1 0
; ; :
022 -
and y 0 2 2 = Y 023-A{p),
* | ^ Oil "^12
Y* =
Y Y *
J J
21 022
5
2 / n * = I^oiu 2/i2*= ' i 2 : #21* = Y-21- K u l e 2 ( m = 2 , <7 = 2 and >- = 2), shows t h a t
one passive a n d t w o active introductions are required.
P e r f o r m i n g the passive i n t r o d u c t i o n w i t h j / , . , = g r = 0 and 6.,./ = 0 gives: M
l 2 3
A p p l y i n g a K alignment i n t r o d u c t i o n to the b o t t o m r i g h t - h a n d 2 x 2 m a t r i x ,
with A = A = 6 = 0 :
12 21 V2
3 4
(e -r )fr/:,:,-r„)
I S 1 5
Oil -flu 0
f/ia
fl 'V, r/ 2 r
3 3
r : 3 -A A : i : l
A ; 3
Synthesis of multiple loop feedback systems 505
032 !/23
Aotfe 2
Here definite dominance is possible :
Aoc/e 3
A g a i n definite dominance is possible :
\ 0 '
f/33 + ^33 = 6vl" + A
2 3 + #13 + Z/33- ( ) 5
A
33
Xode 4
T h i s is the v i r t u a l earth node and thus only indefinite dominance is p r a c t i c a l :
A
23
A t the o u t p u t node :
J
V = 7 + A.,3 + 0.,3".
E q u a t i o n ( 4 ) becomes:
A
v f/327 ^23#2:i' 33f/32 . , n , „
1
- (122 - -T x = " I + 7 + #12 + U-Z2-
A
23 '"33 "23
p- + b [ p + h ll _ _ A 2 3 0 2 3 ' = /, 3 3 (/ 3 2
+ y+ 0 1 2 + /y
P ^23 P / l
33 A
23
I f h
o = (</ 2y)
3 ''"W- then :
, 1 ^23^23 ^33?32 , , / ) , , , /•r-x
iJ + fcj : = - r + 7 + 012+#22- (0
K
33 "23
A.G. J. Noll and J . I . Seicell
!
<3S
+
Synthesis of multiple loop fadb'n-/: systems .">i.iT
b
b =e
1 2i + Y + e 1 2 + - ^ . (s)
S u b s t i t u t i n g f o r y i n (8) gives:
b 1 = e . 2 i + +g l 2 +6 -(y 1 M + A ).
ta (10)
& l - ( ! / 3 3 + ^ 3 )
5
^ i 3 ~ ^ i 3 P - ^ a ' ^ ^ n ' i - Thus f r o m (11) and (2) (neglecting any constant m u l t i p l i e r
produced):
a
^12^24
or 0n 1 2 = 0 3^ 1 ^3 ^ L
23
. (12)
e =a d. '(}
v l 0 2 3 -e^ye . u
024 (&l -
2
l>33 + A
2 3 ] ) +M&l - L'/33 + 'VJ)( - [y 33 + A, ])
3
Solutions o f this equation are possible for elliptic functions. Take the numerical
example :
; j 2 + 7 . 4 f i 4
2
p + 0-607.3/; + 0-7559
508 Synthesis of -multiple loop feedback systems
2
0-4075# 24 - 0-08150. 4- 0-003.1 1.41 = 0.
24
d = 0-050190.
l n
0 = 0-999.
13
a
y= 1-2487 x 10" :
k. =
33
33
_o
2-45397 x
—
"
Fig. 4
OOOI3 5O-000245
0.999 3 O-OOI
Multi-loop circuit realizing a second-order elliptic function. For all circuit realizations
the element values are in mhos and farads.
4. Conclusion
I t has been demonstrated how various numerator polynomials can be included
in the synthesis o f m u l t i - l o o p feedback systems. This increases the range o f
functions which may be synthesized using these systems.
The method o f synthesis f o r parallel systems may be extended to include
circuits of order n and rank m ( H o l t and Sewell 1 9 6 6 ) . There is, o f course,
increased complexity associated w i t h the realization o f such networks.
1
A ( K X O W L E I> CO [ E X T S
REFERENCES
HOLT. A. G. •)., and S E W E I . L . .J. I . . 1 9 6 0 . J.E.E.E. Trans. Circuit Thtonj, 13, 326:
1908, Jut. J. Control, 9, 4 8 3 .
S E W E L L . J . 1.. 1966, Ph.D. Thesis. University of Newcastle upon Tyne.
76
PAPER 8
PROCEEDINGS LETTERS 2155
but v (t)
: | vdt. (at General one-port network, tbi input current versus integral
Jo of input voltage for general one-port network.
Manuscript receded Juno 2. 1969: revised August 19. 1969. Manuscript received August I I . 1969.
2156 P R O C E E D I N G S O F T H E I E E E . D E C E M B E R 1969
(I ] A . G . J H o l t a n d J . T a y l o r . " M e t h o d o f r e p l a c i n g u n g r o u n d e d i n d u c t o r s by g r o u n d e d
g y r a l o r s . " Electronics Letter*, v o l . 1. p. 105. 1965.
(2 | D . F S h e a n a n . " G y r a t o r - f l o i a i i o n c i r c u i t . " Electronic Letters, v o l . 3. pp. 3 9 - 4 0 . 1967.
[ ) ) G. i. D e b o o , " A p p l i c a t i o n o f a g y r a tor-type circuit to realize u n g r o u n d e d i n d u c t o r s . "
IEEE Trans. Circuit Theory, v o l . C T - 1 4 . p p . 101-103. M a r c h 1967.
| 4 ] J . M . R o l l e t t . " C i r c u l a t o r - c a p a c i t o r n e t w o r k s . " Electronics Letters, vol 4. pp. 5 9 9 - 6 0 1 .
I96K.
[5] T . N . R a o . " R e a d i l y biased w i d e b a n d g y r a i o r for floating or earthed i n d u c t o r s . "
Electronics Letters, v o l . 5. p p . 3 0 9 - 3 1 0 , 1969.
(6) J . I . Sewelt a n d F . W . S t e p h e n s o n , " M a t r i x tables for the generalized V i e r m i n a l a m p l i -
fier." Radio and Electronic Engrg.. v o l . 37. pp. 2 4 7 - 2 5 5 . 1969
( 7 | A . G . J . Holt a n d R t.inggard. " T h e m u l t i l e r m i n a l g y r a i o r . " Pmc IEEE ( L e t t e r s ) ,
veil. 5n. pp 1354 1155. A u g u s i I96H
| 8 | K . J . S m y i h a n d G . C . H a y w a r d . " T h e design and l i m i t a t i o n s o f practical g y r a t o r s , "
IEE Colloquium Digest, no. I968/J2.
F i g . 2. G y r a i o r circuit p r o d u c i n g a floating i n d u c t o r f r o m a g r o u n d e d c a p a c i t o r .
circuit shown in Fig. 2 has an admittance matrix In the configuration outlined in Fig. 1. visible light (our experiment used
a laser beam) is shot through a pinhole onto the pits and hillocks that have
0 0 \ 9 been etched into the surface of the crystal. From the reflected light pat-
0 0 tern, the orientation of the axis and the degree of incline of the surface
\ ~9
from the axis can be determined.
-9 9 ! o
In our experiment, we tried to control the width of the laser beam by
using slit diameters from 0.5 mm to 1.0 mm. The photographs obtained
A capacitor connected between node 3 and ground will produce a float-
under such conditions are shown in Fig. 2; etched pits appear in the orienta-
ing inductor between nodes 1 and 2. There are two other methods of realiz-
tion of [111 ] and [100] and hillocks appear in the orientation of [110].
ing gyrators which are inherently capable of producing floating inductors.
We concluded that the reflective surface was convex because the re-
The connection of two transimpedance amplifiers (current-controlled ,:
flected ;iht pattern was enlarged more than the diameter of the laser beam.
voltage sources) in series is one of these, but one which uses ungrounded
Besides this, the size of the reflection pattern changed when the distance
amplifiers, This circuit will give a floating inductor from a floating ca-
between sample and screen was changed. We can observe the effects of a
pacitor, or with one amplifier grounded, the circuit will give a floating
convex mirror in the targe etched pits of orientation [ 100 J and [110]. and
inductor from a grounded capacitor. The gyraior produced by a negative
the effects of a concave mirror in [111]. However, we were interested in
impedance convenor iNICl. negative impedance inverior (Nil) cascade
why there were such special pits and hillocks corresponding to each orien-
using an entirely active Nil [8] will produce a floating inductor from a
grounded capacitor. This circuit employs two grounded amplifiers.
The advantage of these methods is the low number of amplifiers re- M a n u s c r i p t received F e b r u a r y 17. 1969.
79
PAPER 10
INT. J . ELECTRONICS. 1970. VOL. 29. >~0. 6. 501-511
1. Introduction
I n recent years there has been considerable interest shown i n the design and
use o f active n e t w o r k devices f a b r i c a t e d f r o m operational amplifiers and
resistors. Devices such as gyrators (Morse a n d H u e l s m a n 1964, A n t o n i o u
1967) negative impedance converters (Morse 1964, A n t o n i o u 1965), active
transformers ( H o l t and Stewart 1968, Keen a n d Glover 1968 b), circulators
(Keen ef al. 196S, R o l l e t t and Greenaway 1968), a n d various others have been
realized i n this f o r m . These devices are finding extensive use i n active
n e t w o r k theory.
Certain m a t h e m a t i c a l techniques have been e m p l o y e d i n the design o f
these elements, such as the technique using a basic active f o r m o f n e t w o r k
(Keen and Glover 1968 a) and t h e n u t i l i z i n g a m a t r i x f a c t o r i z i n g method.
This m e t h o d appears t o be the sole a t t e m p t at approaching the whole problem
m a t h e m a t i c a l l y , a l t h o u g h even here some basic active f o r m is assumed a priori.
I n the references cited, no rigorous mathematical techniques are employed i n
the c i r c u i t realization.
The m e t h o d t o be o u t l i n e d does n o t assume any detailed i n i t i a l active
topology or rely upon any particular design s k i l l . F r o m the i n i t i a l device
specifications, the f i n a l c i r c u i t is derived b y basic m a t r i x operations. The
technique o f node i n t r o d u c t i o n is employed ; t h i s m a t h e m a t i c a l m e t h o d has
been-rigorously defined (Piercey 1962) and successfully a p p l i e d t o the synthesis
of certain active networks e m p l o y i n g operational amplifiers ( H o l t and Sewell
1969).
i i + sa
Y =
1 1
convenience define
£ j -+Kg,
Y=Y +Y ,
A B (2)
Kg 0 0 Kg
Kg 0 L0 Kg.
Synthesis of active devices 503
0 -8 12 -A(0 1 2 +r 1 2 )
A
^ B(3'3) — -d l 2 Kg-2K6 23 - ^ 3
r 3 3 + 2 A
^23
With d 12 = o, 9, = \g and r
3 3 3 = flrt h i s becomes :
0 0
y B(3-3) —
0 0
K
0 =J> K*+ , + -
9 0 1 0
0 1 2
?'i = A . T
2 =
0 0 1 ~9
0 1
K
The t r a n s f o r m e d m a t r i x becomes
0 0 Z l
Y B(3'3) — z i (3)
•~>
2
0 -g + K g+-
2N 2
83
504 J . I . He well
A = A = 0, 0 = O. 0 < \ <z
1 2 2 1 12 13 1. A = - A " .
3 1 = (f?'2) l ' = g. ? m
0 0
0 0 — 2
•>
-9 " 1 0 0"
1 0
K-
T — 0
J
1 0 T — 0 1 0
1A — ' 2A —
0 0 1 -9
0 1
_K
P e r f o r m i n g the t r a n s f o r m a t i o n yields
A(3-3) —
0 0 — 29 (4)
o
1
K
-9-K 0 g+ o
Fig. 1
Y = -9
-(/-A'* -g + K* ig
w h i c h leads t o a k n o w n I N I C c i r c u i t shown i n f i g . 1 (Morse 1964).
0 1 - 1
- 1 0 1
1 I 0
1 -1
7
the phase p a t t e r n o f (? is always t h a t f o r A odd. The phase m a t r i x K is
e
""1 0]
A,= (~)
0 0
where 5, ->0.
S 2 S 2
Si
r..=
-1 1
Si
N o w investigate these terms i n the l i m i t
Hence
g -g
a 2 2
- A A
a is a constant i n t r o d u c e d to aid synthesis, i f a = 2 synthesis proceeds i n a
convenient manner :
g_
Y,= 2 2
(8)
2 2
-A A
Y 22 •
-g
Y, (3-3) — -g A"- A 2
q „ 2
— •A
•? + —•> —o 4 •->- + A
L -0c 0c
and
£
Y
/Z, (021 - #12 - 21) j_, ft/9 1023
(023 ~- 23
^23) P
I.
2
(022 + 5 ' ) 22
' ^ ( 0 , + J'
2 9
,)
?12 = °> 021 = °.- 023 = . 012= - 0 -
0
0 0
•>
2
-0 A
2
— + — A+</ -gr -</ + A
:
i(4'4)
•0c 0 0e
-032
032 032
' 1 -1 1 "1
0 0 0" ~\ 0 2
0 1 0 0 , T = 2 0 1 0 6
0 0 I 0 0 0 1 0
_ I 0 -1 1 0 0 0 1
0 + 0c 0 -g c -0
2
^•u-4>= 0 K + g - A
-0c 0 j/, 0
-0 -0 0
508 J.I. SeweU
Fig. 2
3
I W 1
3'
Circulator section.
Fig. 3
9 3 g
3 3
I — v w - i -
I 2 N
U U 7
L
A l p o r t circulator (A odd).
5. Conclusion
Appendix
F o l l o w i n g Piercey (1962) and H o l t and Sewell (1969) i t is easy t o derive
the node i n t r o d u c t i o n s used i n this w o r k . For a m a t r i x
Y J
Y '
-'Oil 12
v
1
-
—
2-2
Y
1 1
Y
21 022.
a 1-node i n t r o d u c t i o n gives
Y + aA
o n 1 1 Y + aA
x 2 2 x aA 3 1
Y
1
-
—
Y i-haiA
2 2 Y + aA
0 2 2 2 2
a
3A 2
3-3
aA x 3 aA 2 3 Q
3^3J
0 J = r —
3 ^2 9 23 ^23'
a = —
2^3 !732 ^23'
On s u b s t i t u t i o n the i n t r o d u c t i o n becomes :
^ Oil
1 (!7l2-012- *'l2)(l/21-012- 3
'2l) 012-012
( 0 i 2 - 0 i 2 ~ Y )(g 12 23 - B)
23
Y
022+ 22 022+ ^22
(021 ~ #12- )r
2l)(<732 - #2 ) 3
032 - ^23 ( 0 2 3 - 0 2 3 ) ( 0 3 2 - 023)
g 22 + }'.„ 022 + Y-.2
510 J.J. Sewell
a. A. = K +
2 2 Y. .22
aA
2 3 = A / v - 6. .
32 23 K -> x .
o Ao
3 = A / Y — $03.
23
Fig. 4
- 6ij Yon • V H
- [ S J J X J J -tSjjXjj] - Y M X J 5 \ J J
K XJI An k X„ X 5 5
K Xj, X j 2
K X 5 J
K X13 \ 3 3 L
Fig. 5
-9.1 -ft,
- 9,
+ 9,, X 55 Xj,
X,. X , ( J
s
4,
X.JYJS - >.» ( e . ^ Y . ) e,j x „ X„
Y
x:, 5 I
K X31X.3 K X,» K X„
X53
K X, Xi,
t x ! 5
K Xj, X J J
K X11
X,} X), X),
K X,i XJJ
K Xj, K X35
X>*
Fig. 6
'•n
-e, 2
+ 9j3 X , l X55
Xn
y „ _ — (x«o - x „ f l „ ) - y„x„>,,.
5 JS
X 1
5
>51
A
15
-t-
K Xj, K Xj, X ! S
KXJJ
X33
K X „ Xj, K X U
K X S 5
X 5 T
The A c t i v e I I I a ( A ? i 1) i n t r o d u c t i o n has :
33
( A
<3 3 = ^33^- + ^ 33'
« i ' 4 = A /v - 3 31 d 1 3 .
a A A A
3 l = 1 3 ' ~ ^13-
and s i m i l a r l y f o r the A c t i v e I I I b ( A ^ 1) i n t r o d u c t i o n : 3 3
C A
<3 3 = K s ^ + ^ 33-
aA 2 3 = A 3 2 A- 8.23
aA 3 2 = XK - 23 6.23
REFERENCES
A N T O N I O C . A.. 1965. Electron. Lett.. 1, 88 ; 1967. Ibid.. 3, 350 : 1968, Ibid.. 4, 87.
H O L T , A . G . J., and C A R E Y . J. R . . 1966. I . E . E . Conference Publication 23, p. 343.
HOLT A. G . J . , and S E W E L L . .1. I . . 1969. hit. J . Control. 9, 4 8 3 .
HOLT A. G . J . , and S T E W A R T . R . J . . 1968. hit. J . Control. 8, 517.
K E E N . A. \V.. and G L O V E R . J . L . . 1968 a. Proc. Instn elect. Engrs. 1 1 5 , 9 1 8 : 1968 b.
Ibid.. 115, 1114.
K E E N , A. Y V . . G L O V E R . J . L . . and H A R R I S . R . J.. 1968. Electron. Lett., 4, 389.
M I T R A . S. K . . 1969. Analysis and Synthesis of Linear Active Set works (John Wiley).
M O R S E . A. S.. 1964. M.Sc. Thesis. Universitv of Arizona.
M O R S E . A. S.. and H C E L S M A X . L . P.. mi.'l.E.E.E. Trans. Circuit Theory. 11, 277.
P I E R C E v. R. X . G . . 1962. Ph.D. Thesis. University of London.
R O L L E T T . .1. M . . and G R E E N A W A Y . P. E.. 1968. Electron. Lett.. 4, 579.
91
PAPER 12
INT. J . CONTROL. 1972. VOL. Mi. NO. 1. 171-175
J. I . SEW E L L
1. Introduction
O f t h e a c t i v e realizations w h i c h have been p r o d u c e d f o r t r a n s f o r m e r s
(Weiss 1966, Su 1966, B r o d i e a n d Crocker 1966, H o l t a n d S t e w a r t 1968) o n l y
one does n o t have the r e s t r i c t i o n t h a t one t e r m i n a l on the i n p u t side a n d one
on t h e o u t p u t side f o r m a c o m m o n e a r t h . T h i s n a t u r a l l y prevents realizations
o f m u l t i p o r t t r a n s f o r m e r s b y these techniques. H o w e v e r t h e m e t h o d developed
b y Weiss (1966) allows the r e a l i z a t i o n o f the m u l t i p o r t t r a n s f o r m e r , b u t u n -
f o r t u n a t e l y employs a considerable n u m b e r o f negative resistances and requires
t h e m t o be f l o a t i n g . T h e m e t h o d t o be described overcomes the r e s t r i c t i o n s
so f a r encountered.
2. Transformer realization
T h e basic procedure is t o realize a m u l t i t e r m i n a l t r a n s f o r m e r , each t e r m i n a l
of w h i c h represents one end o f a w i n d i n g t h a t has the o t h e r end g r o u n d e d .
Fig. 1
1 1
1.
MULTI- 2 MULTI - 2
2 n
TERMINAL TERMINAL
GY RATOR GYRATOR
n
m G
2
o
172 J. J. Sewell
where G is an n x m g y r a t i o n m a t r i x ,
x
G is an m x I g y r a t i o n m a t r i x ,
2
and Y is an / x I l o a d m a t r i x .
L
tion.
Y AB- Y B~ A. l
L
l
2
J7,
YL =
0 9L. - I I _-b2 6_ X
s u b s t i t u t i o n gives
where
b,-b 2
C:
2 2 2
I n this case the constant 2c ((/ /'</ ) can be considered t o represent a t u r n s
1 2
Fig. 2
o
o
o rx
• — i +
Y = % BB~ YjB- B l l
= — n Yr.
2 2
<7 2 S'2 ^
The action of this circuit differs from the previous one in that although there
is a turns ratio effect represented by (gi!g )~, the form of the input matrix is 2
the same as the load one and hence does not reveal the same isolation effect.
1 1
possible to utilize eqn. ( 2 ) . However, i t is obvious that G Y ~ G ' = constant 2 L 2
in J. I. Sewell
Fie. 3
MULTI - MULTI-
TERMINAL TERMINAL
GYRATOR GYRATOR
G , G„
-1] and G = 9i
1
_ — 1
i.e.
r 0"
~9L
A-- , Y L = 3
- I 0 9L.
then
' 1
0
9L
G*Y,;-*C? = gf[\ - 1 ]
1
0 —
9LJ
S i
Or,
and hence
[1 - 1 ]
1 - 1
-1/2
Fig. 4
3
o
C>
35
3. Conclusions
T h e r e m o v a l o f the r e s t r i c t i o n o f g r o u n d e d t e r m i n a l s w h i c h t h e above
t e c h n i q u e has f a c i l i t a t e d , means t h a t a c t i v e realizations o f t h e i d e a l t r a n s -
f o r m e r can n o w be used t o replace f l o a t i n g 2-port t r a n s f o r m e r s i n n u m e r o u s
synthesis techniques, a l t h o u g h i t m a y be necessary, f o r d.c. s t a b i l i t y con-
siderations, t o g r o u n d one t e r m i n a l o f some o f the structures presented. These
techniques, together w i t h i n d u c t o r replacement b y g y r a t o r s a n d capacitors
w i l l enable complete m i c r o m i n i a t u r i z a t i o n o f the c o n v e n t i o n a l c i r c u i t s . I t also
provides a novel m e t h o d f o r p r o d u c i n g a m u l t i p o r t t r a n s f o r m e r b y a c t i v e means.
REFERE>TCES
PAPER 18
350 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS. VOL. CAS-23. NO. 6, JUNE 1976
networks can be generated for active structures in which important param- V. p r o d u c i n g Y , the constrained admittance m a t r i x of the
eters such as component spread and value, and sensitivity, can be optim- active circuit. I n the case of an embedded single i n p u t
ized. Methods of controlling the number of elements in the equivalent operational a m p l i f i e r w i t h i n 9 i between nodes / and k,
networks are also given. the constraint imposed o n Y consists of r e m o v i n g c o l u m n
/ and r o w k. A l s o it is well k n o w n that i n the l i m i t of
I. INTRODUCTION
i n f i n i t e gain the ideal current a m p l i f i e r has exactly the
same effect as the ideal voltage a m p l i f i e r , hence the
T H E theory of equivalent networks as presented [1], [2], method may be generalized to include current amplifiers.
a n d w i t h later elaborations [3], [5], has been effec- Keeping the n u m b e r i n g of the rows and columns of the
tively applied to passive networks. However the nonsym- constrained m a t r i x Y unchanged, network functions may
r
metry of the nodal admittance m a t r i x due to the presence be calculated f r o m Y i n the usual manner. Y' is a r
of active devices renders the application of the theory to nonsymmetric m a t r i x ; however, due to the symmetry o f Y
active networks d i f f i c u l t . O n transformation, unsymmetric certain pairs of elements are equal. I t is postulated that
terms appear i n the admittance m a t r i x to such an extent matrix transformations can be applied to Y' p r o d u c i n g a
that it is almost impossible to i d e n t i f y any conventional scaled equivalent constrained m a t r i x Y", which has the
active devices. By constraining the n o d a l admittance necessary symmetry in the same places as Y . T h e n the r
matrix of the passive part of the circuit, it is shown that a matrix Y' is rebuilt to produce a symmetric admittance
general scaled H o w i t t t r a n s f o r m a t i o n may be applied to matrix Y' of the passive part of a network 9 1 ' w i t h
circuits incorporating operational amplifiers. The o p t i m i - voltage amplifiers embedded in the same positions as i n
zation of network criteria may then be undertaken using a the original passive network 9 1 . This rebuilt admittance
constrained nonlinear p r o g r a m m i n g method. The type of matrix must obey the usual realizability constraints o f a
active element remains unchanged and its position w i t h i n passive nodal admittance m a t r i x
the network is constant, but the passive elements may
change i n number, k i n d , a n d connection between the Y f < 0 (2a)
nodes of the initial structure.
v i ; - 2 ^ > o , i - i . ( 2 b )
II. CONSTRAINT T H E O R Y
r ; - r ; - o . (2c)
The e m b e d d i n g of a d i f f e r e n t i a l i n p u t operational
a m p l i f i e r w i t h i n a passive network may be considered as a
III. SCALED EQUIVALENCE
constraint imposed u p o n the nodal admittance m a t r i x of
the passive part of the circuit [6]. [7], Consider a passive Definition: T w o networks 9 1 , 9 1 ' are scaled equivalent
network 9 1 w i t h admittance matrix Y. Embed a d i f f e r e n -
w i t h respect to some voltage transfer f u n c t i o n T„(s) if and
tial a m p l i f i e r gain A w i t h i n 9 1 , positive input node i ,
only i f T;.(s)= HT (s),H c ER. N o t e that H= \ gives the
negative input node j , and o u t p u t node k. This introduces
a constraint i n the network f o r c i n g (he potentials V , Vj, usual d e f i n i t i o n of equivalence. Scaled equivalence w i l l be
t
* { V - V j ) = K - (i)
As A-*oo, (1) becomes
a n d the scaling f a c t o r H = l / k .
Proof:
(4)
A ^ ) - d e t [>"(*)]„
I t is required to determine the s u f f i c i e n t conditions on -det y (*)«(*)]„
T | ( J ) , J ( J ) f o r (4) to h o i d . First we prove three lemmas.
Lemma I: = 2 2d«[l(*)] d«[ ^)Ldet h (lemma 1)
det[y«] -J-dei[/»]. fl Hence the elements of the passive part of the scaled
equivalent network may be f o u n d after r e b u i l d i n g Y'(s)
/Voo/.- The result is obtained by expanding det [A] in symbolic f o r m as quadratic equations i n the entries of
along r o w ( c o l u m n ) /'. T | ( J ) and i{s).
Lemma 3: 1) I f the / t h c o l u m n of a m a t r i x A = mu ,u
i i
The scaling f a c t o r H w i l l enable the generation of
the /'th u n i t vector, then equivalent networks w i t h d i f f e r e n t gain factors, w h i c h is
of some significance in active RC network synthesis.
det [ / ) ] „ = (), Vp*/.
F u r t h e r m o r e , the t r a n s f o r m a t i o n matrices T)(J),£(J) can
2) I f the 7th row of a m a t r i x A=mu ,u J
T
j the yth unit assume any type of entry, and hence i t is easy to d e m o n -
vector, then strate the equivalence t r a n s f o r m a t i o n of an RLC network
i n t o an active RC network c o n t a i n i n g FDNR elements.
d e t [ < 4 ] „ = 0, Vp+j.
A n i m p o r t a n t observation is that whereas i n generation
Proof: For 1), r e m o v i n g any c o l u m n p¥=i from A of passive equivalent networks the quadratic t r a n s f o r m a -
makes c o l u m n /' of [A] lp all zeros. Hence dei[A] ip =Q tion is generally required to retain symmetry and hence
Vp¥=i. The proof of 2) f o l l o w s . realizability, no such l i m i t a t i o n s are imposed in the active
10 0
situation. As the constrained matrix is unsymmetric i n i t i - 5) F o r m the scaled equivalent component admittance
ally, a simple l e f t - h a n d or right-hand t r a n s f o r m a t i o n is matrices of the passive part of the network, Y^. Y' by the c
amplifiers 1 • • • A.
If D is the n u m b e r of d i f f e r e n t i a l - i n p u t operational d) Put
amplifiers embedded i n the network, then let 0 , , 0 , - • • ,0 2 D
f o r the pairs
= o u t p u t nodes of differential-input operational amplifiers
1 - - - Z ) ; P P ,- • • ,P , = positive input nodes of d i f f e r e n - YQP, — YQ„ + YQ„ p=M q ]y = P; l
V 2 1
2) F o r m the c o n s t r a i n e d c o m p o n e n t a d m i t t a n c e f ) Put
matrices Y' , Y' by the f o l l o w i n g operations on Y , Y :
a c 0 c
YQI, = YQI, ~ YQI,, •V ( * V • • , / „ ; < > „ • ••,<)„; M, T
a) remove rows / , , • • • ,I ; A
r t n e a r s
M; D
e) remove columns M - • • ,M . YT D
g) Put the resulting matrices equal to Y^, Y^..
3) F o r m the symbolic t r a n s f o r m a t i o n matrices with 6) W e n o w have symbolic scaled equivalent passive
dimensions {N-A — D)X(N — A - D) where part matrices Y' , Y^. w h i c h are symmetrical i n a l l b u t
G
m8y. N -(l+4(A+D))N
7=1
+ (A+D)(2 + A + D)
V •>1
pairs of entries.
where X, is a symbolic variable and t = (N — A — D-
and
VI. OPTIMIZATION
min/°(x),x = (X|,x , 2
1.11(1
H j ( x ) = 0, j=m+l,--,m + q
P(x,r )
k = F(x)-r 2ilog G^x) k t
0 . 0 791
i
I
+ 2l+i[«,(x)] A*+x£ i»W 0.4141 0.4141
j I
0.079]
where [ r ) is a sequence that decreases strictly m o n o t o n i -
k
0.1716 0.4141 0.4)41
cally to zero, the conjecture being that the sequence of
unconstrained m i n i m a {x(r )} k w i l l approach the solution
to the o r i g i n a l p r o g r a m m i n g p r o b l e m as r —>0. K
1.1040
I n general, the inequality constraints G , ( x ) in this prob-
lem f o r m u l a t i o n w i l l never be all satisfied strictly, and so
a small b o u n d a r y term 5 , « m i n ( G , ) , where G^O, was
added to each inequality constraint, enabling an interior
starting p o i n t x ° to be f o u n d . 5, is somewhat dependent Fig. 1. (a)Original Rauch fourth order Butterworth realization. (Ele-
ments in Siemens and farads). (b)Equal resistance realization, (c)
u p o n the computer w o r d length, a n d the particular prob- Combined equal capacitance and resistance realization.
8
lem, f o r the examples included, B = 1 0 ~ proves satisfac-
t o r y ; i f B is m u c h larger, then an excessive number of
i
nonzero elements is introduced i n the solution. For a VII. RESULTS AND CONCLUSIONS
smaller value of B r o u n d i n g errors prevent p r o f i t a b l e
i
REFERENCES
11J Howitt, N . "Group Tneory and the Electric Circuit," Phys. Rev. voL
37, pp. 1583-1596, 1931.
(21 Cauer. W. "Vierpole," Elek. Nach. Tech., vol. 6, pp. 272-282, 1929.
13) Schoeffler, i. D. "The synthesis of minimum sensitivity networks,"
l.E.E.E. Trans. CT., vol. C T I 1 , pp. 271-276, 1964.
W Calahan. D . A. "Computer generation of equivalent networks,"
l.E.E.E. Int. Con. Rec. Pt.l, pp. 330-337, 1964.
15) Cheetham, B. M . G . "A new theory of continuously equivalent
circuits," l.E.E.E. Trans. CAS, vol. CAS21, pp. 17-20, 1974.
1*1 Nathan, A. "Matrix analysis of networks having infinite gain opera-
tional amplifiers," Prcc. IRE, vol. 49, pp. 1577-1578, 1961.
[71 Morse. A. S. "The use of operational amplifiers in active network
theory." M . S. Thesis, University of Arizona, 1964.
[81 Fiacco, A. V., and McCormick, G . P. Nonlinear Programming:
Sequential Unconstrained Minimization Techniques, i. Wiley & Sons,
1968.
CI Kuester, 5. L., and Mize, J . H . Optimization Techniques with Fortran,
McGraw-Hill, 1973.
0.5(11
PAPER 21
iensitiwBty a™ partitBoni miisiruiBsatiOui
traosfomatras In aotove ©qyowaleont
[networks
L . G . G r a n t and J . I . S e w e l l
Abstract: Following the extension of equivalence transformations to a general class of electrical networks
containing active devices such as operational amplifiers, the theory is developed to accommodate the mini-
misation of the sensitivity of the network response function to component variations. Theorems for passive-
component sensitivity invariance and for updating routines applicable to the optimisation step are proved.
Previous limitations imposed upon the optimisation routine by equality readability constraints are com-
pletely eliminated by a partitioned approach. This leads to a new form of transformation. The partitioned
transformation can be implemented in both discrete and continuous forms, and applied to the optimisation
of component values and sensitivities in active networks.
1 Introduction
is termed the semi-relative sensitivity. Noting that
Equivalent-network theory is well established in the passive-
1,2
det IY ] R U = N(s)
filter field, and the application to sensitivity studies and (4)
component value optimisation is well documented. For " det [Y ]„ R £>(*)
3
active networks it has been shown recently that
we define as follows the entry sensitivities Eft o f the
(1) numerator and the denominator to the entry y kt of Y : R
det [ K ] „ , (6)
generalised Howitt transformation matrices. The solution o f H W
EM
constraints that must be satisfied to maintain realisability.
In particular, the equality constraints can prove to be quite
=I I N D
(7)
Assuming that the active elements are ideal operational < , = 6 ,(5*p + &
fc KQ + 6 K P 8 K Q ) - ( 5 K P 6 L Q + & & ) KQ LP (8)
amplifiers, whose effects upon the network can be rep- dy | A' if e is a conductan.
M P q p q
Theorem I
Paper T 1 3 8 E , first received 28th October 1977 and in revised form
6th January 1978
Dr. Grant and Dr. Sewell are with the Department of Electronic
Engineering, University of Hull, Hull HU6 7 R X , England N' N
Now
Putting
det [ Y \ „ R det [ n 5 - § ] R jy
r
S = YQ E N D
, IS " = I [5], (12)
= det n det [ K ] , det § R 7 p
3
for a scaJed equivalence transformation. Form.F such that
T T N D
= » I S " = s.d.e. o f [ Y Q ] E
det y = det [ y ] , , fl
all
conductances
r r JVI T
= det [T|^5]
= s.d.e.of§ y V( r'£" (§V G n
T N D T I
= s.d.e. of § K j i T (§ r
=* det [KAlo.pa = det [ t L v g ] M
r r
-1
where
function; the value o f sum squared sensitivity function at 2m — l ) / 2 pairs o f entries. Now consider Y re-ordered so r
s = / ' l is7-7016. The circuit o f Fig. 2 results f r o m sensitivity as to place these symmetrical entries in an (n — 2m — 1) x
minimisation and has a comparative sensitivity value o f (n — 2m — 1) symmetrical submatrix in the top right hand
6-8393. This solution is a local as opposed to a global mini- corner. That is, f o r m
mum. The increase in number o f elements is apparent, but
this is not unexpected as the limiting solution to these
problems is a fully connected network with equal element
sensitivity values. Y* = (14)
The results shown here were obtained using one-sided
transformations only, because of the problems reported in
f = n - 2m - 1
Reference 3 and the difficulties arising f r o m the nonlinear
form o f the equality constraints generated by the quadratic where Y, is an / x / s y m m e t r i c matrix.
transform. The following well-known matrix properties are used:
(i) the interchange o f columns /' and / o f a matrix A is
obtained by forming the product A j<; and r -,
( i i ) the interchange o f rows /' and j o f a matrix A is
obtained by forming the product 7} *jA
where the topological transformation matrix 7} ,y is the unit
matrix w i t h its / t h and / t h rows interchanged.
Let the unloaded nodes (those not connected to an
06666 amplifier, and not the input node / ) be p , , p ,. . . ,pallet 2
Y -+ Y* take the f o r m
r
i i M
~ Tf-*f ••• r 2 -+2 ' r , -»1 *r Cf*n-m 'Cf.i+n-m-l •• •
3 Partitioned transformation
=
T , •>\T ->l
r Tj • • • T ^fY*T ^nn.2
rf c • • • T * -.
Cf n m (16)
The equality constraints, which are necessary to keep sym-
metric the passive part o f the admittance matrix during
In practical terms, this amounts to simply re-ordering the
optimisation, have prevented, up to now, the formulation
nodes.
of a continuous quadratic transformation f o r active RC
5
The following results emerge and are proved where
networks like that proposed for passive networks. Essen-
necessary.
tially, continuous techniques depend on the transformation
being small and the ability to make linear approximations
to quadratic constraints. I f the constraints are inequalities, Theorem 3
good linear approximations may be made; however, it is
det [ 7 - . J p l 7 = ( - -5 p a )(«„6 / p +S 6 )
LP IA
almost impossible to make efficient linear approximations
to quadratic equality constraints. The elimination o f - « « 0 - « P , X 1 -«/P)(1 - 8 i ( I ) + 8 ,6 . p (j
Tv = (13)
det [Y ],H det [Y ].,
r D(s)
Proof
where Y is formed by removing the /'th row o f Y . Inspec
r R
See Reference 3.
b = 1
where /. / are the unique solutions o f the equations. The
Proof signs are ignored, as inverse transforms are taken later. Then
'i : <i," 0 0
1\ ...det [ T C R N . )4 * 0M (21)
A = Y' =
0 : det [T ti -m +2 1 Iff det [ 7 " , , c m + 3 ] 3 T . ..
0: 7 •••det [ T ^ . n ] ^ * 0 (22)
0 : A Y
0 : Eqns. 21 and 22 are solved easily by a computer
_
algorithm using the result of theorem 3.
then We observe from eqns. 19 and 20 that the problem o f
applying transformations to Y to retain invariant [det Y ] J
r r
6
f "So . o "m + 1
n' : °
Theorem 7 £ _ (25)
s —
I f / I is ( / - l)x ( / - 1), Y is ( / - 1) x / and B, C . . .Z are m 1i • To 5.
/ x /, then
m+l f
det [A YBC... Z]j = £ £ . . . I det / I det [Y], b
Considering eqn. 14, eqn. 24 becomes
b e 2
n —m
x det [B] bc . . . det [ Z ] , z
f
/Voo/
\n — m—\
By extension of theorems 5 and 6, writing Y"
n.v 5o'+ ;n§r':n 'F,V +
1 n 1
V
+ 1oJ'o§o + 1 o > 2 § i - n o ^ c ; ' (26)
A = ^ , -.1 7*r, -»2 • • • 1y->f"
= X £ . . . X d e t / l d e t [K*]. det [T .„ , U . .. a Ci 7
T
where 1^ = x]'Y %'. Putting r | ' = § ' gives
s
• det [r „- ] er m z;
(17) 7
and
n' = n'nn'". (28)
a transformation
transformation vwhich retains the symmetry o f the f x f
det [Y h r top-right-hand
p-right-hand submatrix.
subr Now
T 1 D
then the transformation K*' = t i K ' f is a scaled equivalence E N D
' = (/, + m + n xr ^ (^ . m l
r
+5 v)-'
transformation with scaling factor co/0, and the necessary
2
symmetry properties o f Y are conserved under the trans- = ilf+m - r\ x T
+ riV* . . . )
formation.
Once we have Y*' in symbolic f o r m , inverse transform- x E N D
( l , . m . l - Z T
x + T
S ?x>+...)
ations are applied to produce K '; thus r
N D N D N D 2
= E - ( r f E +E f ) x + 0(x ) (41)
~ ^r, +1 - - • Trf-*f Y T Ci ^ M + 2 • . . T f +- C n m (33)
The sensitivity minimisation techniques o f Section 1 can
The matrix Y' is then rebuilt f r o m YR by inserting suitable
now be implemented using linear programming methods;
symbolic rows and columns corresponding to the con-
this avoids matrix inversion each time the semirelative
strained nodes, to form the symmetric nodal admittance
sensitivities are updated.
matrix of the passive part o f an equivalent active network.
The symmetry o f this matrix has been ensured entirely by
the f o r m o f the transformation and not by equality con-
straints imposed on the optimiser in a previous implemen- 5 Conclusion
3
tation. A method for the minimisation o f the sensitivity o f active
networks using equivalent-network transformations has
4 The continuous partitioned transformation been produced, and appropriate sensitivity updating
routines derived. A new transformation is given which
The partitioned transformation may be implemented in a eliminates equality constraints and enhances the application
continuous f o r m . Let o f the quadratic transformation, both in the minimisation
o f element spread and in the sensitivity o f active networks.
(34)
It also enables a continuous f o r m o f the transformation to
where I , I f * i are identity matrices, q , § are o f the
f t m t m
be used. The choice between discrete and continuous forms
form described in theorem 8 and x is a small real scalar. is somewhat subjective, as some assessment of the effect o f
Then i j , g clearly obey the conditions o f theorem 8 and the truncation steps in the continuous form must be
weighed against the avoidance o f matrix inversion present
Y*' = i\Y*% (35) in the discrete implementation. The equivalent networks
produced by the two forms can differ quite considerably.
is a generalised scaled equivalence transformation. Substitut-
Both methods will produce new networks containing more
ing eqn. 34 into eqn. 35 gives
elements than the original, an effect observed particularly
Y"(x) = (7, + ^)V*(/ +ix, when the optimisation o f total sum squared sensitivity is
+ m n f + m t l
attempted.
2
= Y* + (T,K* + Y*S)x + (r,Y^)x (36)
Y" = Y* + (x]Y* + Y*$)x (37) The efforts of D.G. Spridgeon in writing and running some
of the programs used in this work are particularly appreci-
The consequence o f eqn. 37 is that, for a small x, a trans- ated. The work was supported by the Science Research
formation may be applied producing an equivalent network Council.
1 HOWITT, N.: 'Group theory and the electric circuit', Phvs. Rev.. 4 F I A C C O , A.V., and McCORMICK, G.P.: 'Nonlinear program-
1931, 37, pp. 1583-1596 ming: sequential unconstrained minimisation techniques' {Wiley,
2 S C H O E F F L E R . J.D.: 'The synthesis of minimum sensitivity net- 1968)
works'. IEEE Trans., 1964. CT-11, pp. 271 - 2 7 6 5 C H E E T H A M , B.M.G. 'A new theory of continuously equivalent
3 G R A N T , L . G . , and S E W E L L , J.I.: 'A theory of equivalent active circuits', IEEE Trans., 1974, CAS-21, pp. 17-20
networks', ibid., 1976, CAS-23. pp. 350-354
PAPER 22
Some further sensitivity theorems in
active equivalent - network theory
J. Dunning-Davies, Faith D. Faulkner and J.I. Sewell
Abstract: The relevance of various sensitivity measures in the study of active equivalent networks is inves-
tigated. It is known that the summed differential sensitivity of the passive components in an active RC
network is invariant under the transformation; it is now shown that the similar sums of large-change sensitivities
and the large-change multiparameter sensitivity are both invariant for component changes within specified
bounds. Even more practical sensitivity measures derive from integral definitions over a specified frequency
interval. It is demonstrated that the sums of integral sensitivities are noninvariant under transformation
but the integral of the multiparameter sensitivity function again remains invariant. These results are of
significance in any attempted optimisation of the sensitivities of active-filter structures.
1 Introduction ykl of Y R as
1
The theory of equivalent networks '" h-s been extended dN
N
determinants. In particular, the voltage transfer function interest between nodes p, q and sensitivity calculations
T , which is a function of the complex frequency s and
v
assume the bilinear properties of network functions. 5
Two networks N, N', consisting of active devices (in the Now large-change sensitivity is defined to be
same positions in both networks) together with passive
3
elements, are scaled equivalent with respect to some
voltage transfer function T iff, T' =HT \ H&J%. De- A e
v V V
* T Ae
fining the entry sensitivities and of the numerator
and denominator of T (eqn. 1) with respect to the entry
v From eqn. 2
„ V V P N D ^M.
(4)
dD/be
1 + Ae I YSE = 1 ND
[S] P P
diagonal P
as A e - » 0 , S ^ -.^"(differential sensitivity). Now and for a scaled equivalent network
diagonal
= I h i c s n c V r ' ^ a V ]
and for changes in e of up to 10%, eqn. 3 can be expanded diagonal
to give a bounded large-change sensitivity function
= 1 i Yjn (n r E ' (i r
T T T i s D T
e b
diagonal
- L L kl , T T
= I I I [ E U [ s U K J i J .
a
P b
2
EH W = 1 i i i s u m a p u u
(5) a b p
=
I 11^1 "b &ob (8 = the Kronecker de'ta)
a b
The computation of such a sensitivity measure during
each cycle of optimisation would obviously be expensive. = I [S]aa = I YSE ND
(8)
° diagonal
However it is possible to update eqn. 5 without further
evaluation of the complete function. Direct application
4
of the updating theorem to eqn. 5 gives the current
Let y= YQE ND
(A Y ) G
T
— then
sensitivity function of some element e', during a generalised
scaled equivalence transformation, in terms of the original
entry sensitivities as
I YSE °(AY )T N
G \ = £ M <
diagonal U q
, T
sli =e zzzz[iv );i[E N D
un rt T
Now
h I a b
- w I I I I rf I Yg'E"°X*Y ) G
be' diagonal \U
D
-i i rF i ftTi -i On = I [ r » r o H T
[ 0 j V £ N D
( { V ]
diagonal
[V ika pit 1 ab I? ] bl
T
WAr )£] [(TjV c ^ t t V l
= I T
i YjE (AY f^r ND
G
I YZE* diagonal ^U J \D
all diagonal
conductances
(10)
diagonal
I ND
Y£E (AY ) E G
T D
D diagonal
Henu
I Slv (ii)
all
diagonal
conductances
Y l E m }
{ A Y a ) T
E D
( A Y c ) T
E The same proof will hold when considering capacitances.
(7)
194 ELECTRONIC CIRCUITS AND SYSTEMS, NOVEMBER 1978, Vol. 2. No. 6
Tlieorem 2: The bounded large-change multiparameter sensi- Using the same methods of proof as for eqns. 8 - 1 0 . it
tivity of all the passive components of one type in an RC follows that
network with embedded operational amplifiers is invariant
S^; ~ S% (17)
under a generalised scaled Howitt transformation.
Hence theorem 2. and of course this also implies the in-
froof: In reality T = T (s, <?,, v v e„) hence variance of the differential multiparameter sensitivity of the
same networks.
A'(c, + Ac, ,
AT. =
D(e, + Ac, . .e„ + i f „ ) £>(e,
| isV a.v
| Dn\ . . . . c „ ) [ . V ( c , . . . . c„l + Ac, (e, nl + ''n)
I Be, A<<„ — If,.
dD ZD
+ Ac, — (e, , . . . c„) + . . . + A e n — (e, . . . . c „ ) ]
dD 3£>
flic,,. ..e )\D(e
n c ) + Ac, — (c,
n
•«n) + -<•„)]
dc. 3c„
(12)
A C
+ Z ;Z
i k l u o e
i Hence
£
3>'h!
Z Ae, Z Z ^
3e,
sn + (3a -i/3e)
n (3flo/3e)
5
(3a /3e) n " '(bajbe)
i - z - , z z - ^ M ^ ^ ! ? 3A//3g = 3a /3e n
(15) N ~ a„
s" +-
which is a definition for bounded large-change multi-
parameter sensitivity.
Using an induction proof similar to that given in Appendix Let A/ have n distinct roots s, = a, + jp t = 1, t . « and
7, it may be shown that 3fl,/3e = a/. Tlien
A/
Sit, • Z
diagonal
Vi*Y
r ) E-- a
N TD ND 1
Z
diagonal
<.AY ) E
C
T ND
A/ "
III l + ^ + (19)
S — S i
where
'n-1
+ 1 X AY
(- G)'E N
\ diagonal U }
ELECTRONIC CIRCUITS AND SYSTEMS, NOVEMBER 1978. Vol. 2, No. 6 195
1 1 4
r"i [ T r + / ( e - " ) ] r
where
I S i A U , - u , da><\eAeW\
e \U\u~\V\ £ |f*(s,)l
Jo
W -1 _1
a„ b„ bm
sinh [ ^ - ) ~ |G(s )|sinh
r i^J^
B
\F{*t)\
\sJ \ => <\eU\ s w + )eV\Z 2
ffi K +(co-(3,) ]'
+ sinh" IH'I I I
lC(*r)l
+ ieW| I 2 2
r, * r ,
[7? + ( e - t ) ] "
r J
\G(s ,)l\G{s )\
r r2
sinh
and for the case of a lowpass filter the integral of sensitivity V(e , - e
r r j ) 2
+(7,, -7 r ! ) J
ui=0
A similar expression may be obtained for the third term.
(22)
Direct application of theorem 3 establishes the non-
Theorem 3: The sum of the integrals of the differential
invariance of the sum of the integrals of the moduli of
modulus sensitivities of all the passive components of one
expressions of the type given in eqn. 26. Hence the
type in an RC network with embedded operational ampli-
corollary.
fiers is noninvariant under a generalised scaled Howitt
transformation. Theorem 4: The integral of the modulus of the bounded
large-change multiparameter sensitivity of the passive com-
Proof : Since the summations concerned are finite
ponents of an RC network with embedded operational
amplifiers is invariant under a generalised scaled Howitt
t f lSl"tdw = [ liS^lrfw (23) transformation.
Under a generalised scaled Howitt transformation it has al- Proof: With reference to eqns. 15-17 the bounded large-
ready been established that change multiparameter sensitivity function is invariant
under transformation, and hence the integral of the mod-
I = I (Sl )', v
i.e. invariance. ulus is also.
However
E% Qki
(27)
The results demonstrate again the care required in speci- D be
fying sensitivity measures. The sums of the integrals of the 1 T N D D
£ Y E (AY G ?E (28)
moduli of differential and large-change sensitivities are U diagonal
noninvariant and can be grouped with the other non-
invariant measures such as the sum of sensitivity squares or where A Y is the matrix of entry increments.
a
and the integral of the modulus of this function exhibit a node /. Let f=n — 2m, the number of nodes not
complete invariance, when the network is subject to a fairly connected to an amplifier. Also e = e - and e Ae = 0, tj ;i /; fc/
general class of equivalence transformations, and this con- V(i Eqn. 27 gives:
firms an underlying practical feature. j f f m m
ii J ,
all conductances
= —
n L
£>
Z
,
L,
a = 1 b = 0 + I = l t = 1
± 0 4261
S S L = I W ^ A V f E " = P n m
v n > m
D diagonal '
06666 e r r A e i r E ^ E f f
D
+ 1 e fAe .{E™+E?
a af r Era)
(=1
5 Acknowledgment
m
The work reported in this paper was supported by the U K +l e f [ l Ae r i l (E^-E^)(Ef. -E? f l f .)
Science Research Council.
• t r f lt
'•/ o - ; o
) s l 3
a=\ u = I b= a+l
(•/TiVD j _ r . V f l _ r-.VD
£
_ ,-.VD,
\ oo ' £ bb o6 £ 6a )
l-inallv. for a network with n + 1 nodes, in + 1 amplifiers.
c c L £
i aa ^ bb ab ba I ./"=/-!
f m /' r" m + 1 m * I
+ 1 1 e Ae (F™-E^iFg-EZo.)
a0l a0l
V V V V 1 if
aa — - aa*- aa u
t
a - l 1=1 «=lb=«tl ( - I s= 1
e
+ ab Al> b {E^ a + t 'J'f — C ~ ab
+ i i ^ . ^ ( f - s ' - ^ x f s - £ • { , „ )
f r D j _ I'D _ r-D _ pD
m m
£ f •f fe « A„ ,r-wn /r.VD*
+ I Ie,,o. ™ . K - So.) "0,Ae ^J>^E^ Fg-E )
a0l l){ a0l
+ e^Ae^E™ - E%S)(ES, ~ E? ) ia
=
+ I. m
+ e Ae (-E™ )(-E? )
ll0s ll0s s l0s
T ND T D
I &. = - k l Y E (A Y) E
,
(£"£ + f £ b - f ? 6 - £ g , ) all U diagonal
conductances
The same process can be used to identify the other terms in
eqn. 6.
PAPER 32
E7.1
and applies equally t o passive and a c t i v e networks. columns 1, ...,p, and t h a t Y i s obtained from N by
Using ii-domain f o r m u l a t i o n o f the nodal charge d e l e t i o n o f the zero rows and columns p+1, .,.,n,
equations, i t i s demonstrated both i n theory and n+p+1, ...,2n.
by example t h a t a wide class o f d i s t i n c t but equi- The submatrix R o f T includes the c o n s t r a i n t s due
v a l e n t networks can be obtained using two sided to the a c t i v e devices {assumed here t o be VCVS's),
scal;.r transformations o f a reformulated MNA the vector I represents the c u r r e n t v a r i a b l e s i n t -
m a t r i x (RMNA). For s i m p l i c i t y o f p r e s e n t a t i o n , roduced as a r e s u l t o f the MNA method, and S con-
only c i r c u i t s w i t h two non-overlapping clock pha- t a i n s e n t r i e s due t o the r e s u l t i n g c u r r e n t balance
ses w i l l be considered; the corresponding time equations. Note t h a t i n p u t i s a t node 1, and t h a t
s l o t s w i l l be r e f e r r e d t o as odd and even respect- (1) represents the equations f o r even e x c i t a t i o n .
ively. Since Vy has been normalised t o 1, the t r a n s f e r
e
e
f u n c t i o n s H ( 3 ) , H ( z ) may be equated t o V
e e e o 2
0
2. EVALUATION OF TRANSFER FUNCTIONS USING MODI- and V 2 r e s p e c t i v e l y , which are r e a d i l y evaluated
FIED NODAL ANALYSIS
!l 0
A v a r i e t y o f e q u i v a l e n t presentations o f the nodal JO .
equations o f SC networks are i n c u r r e n t use. The I .
I
p r e s e n t a t i o n described i n t h i B s e c t i o n combines •!• '
i •
nodal a n a l y s i s i n the s-domain [sj w i t h the modi-
f i e d nodal approach (MNA) [ 9 ] , t o produce a m a t r i x
d e s c r i p t i o n which always includes the complete - i f:~?j
o
admittance m a t r i x o f the passive p a r t o f the n e t -
I. •
work i n symmetric form. D e t a i l s o f the m a t r i x i,
f o r m u l a t i o n are shown i n ( 1 ) ; since the e v a l u a t i o n lo 0
of s o l u t i o n s i s not i n v o l v e d , arguments f a v o u r i n g
a f u l l clock cycle d e f i n i t i o n o f a, as used i n 1
n
[ 1 0 ] , do not apply. The scheme i s based on t h e
elemental switch time f o r the d e f i n i t i o n o f a,and
has the advantage o f r e t a i n i n g the t o t a l system
d e s c r i p t i o n w i t h i n the m a t r i x .
697
E7.1
and i f s , t e D
necessary t o replace the u n i t sector 2 p + l Y 2 o + 2 , 3 r U b u
second c i r c u i t i s the same as (reap, a scalar mul- where { J , k, . . r } = U,...n} and the sum i s taken
tiple oft t h a t o f the f i r s t . over a l l permutations o f the second s u b s c r i p t s .
Prom (3) every non-zero element i n t h i s sum has
3. REFORMULATION OF THE MNA the form
occur i n the c o r r e c t p o s i t i o n s . I f i n a d d i t i o n , i = i
the s u b n a t r i x o f G c o n s i s t i n g o f elements i n rows It f o l l o w s t h a t d e t S = 1; s i m i l a r l y det S = 1 e 0
1
and i f m e I ,
-1 t 1
-3 P, Y P, . (S
e>rm = ( e'rm s v r e (1
'-- n
'
_ _ _ J - _ ? ° _ 2 _ J pS, Hence f o r each I , ?> Jn I n
P, Y_P„ |
2 oe 1 2 oo 2 (^e^e'tm = 1 I Cs^)
e ' t q q r ' e'rm c s
I
=
! I O w ^qr< e>rm s
!2) q=l r = l
or, simply, GTH = T
I f two d i s t i n c t capacitor matrices C an3 C w i t h 1 S *-CS . The lerma now f o l l o w s from the c o n s t r u c t -
0 0
698
1 2 0
E7.1
Consider the s e t o f equations presented i n matrix a t i o n o f the RMNA shown i n (7). Only q u i t e simple
form i n ( 6 ) . The presence o f the four submatrices adjustments are r e q u i r e d t o o b t a i n the analogous
I _ p e f f e c t i v e l y removes rows and columns p + l . . . ,
n f conditions f o r any o r a l l o f the other t r a n s f e r
n,n+p+l, ..2n i n the e v a l u a t i o n o f the determinant f u n c t i o n s . The approach i s a n a t u r a l extension o f
of any submatrix o f T c o n t a i n i n g these four ident- e a r l i e r methods [ l ] , [ 4 ] , and r e a l i s a b i l i t y con-
i t y blocks. Therefore by Cramer's Rule and Lemma s t r a i n t s on C do not occur as conditions i n the
2- V i = v , $1°= Vj° V 1 e { l , . . . , p } , where fv ,
e
i
e e
theorem but must be separately considered.
V,°, i = 1, ... p) are s o l u t i o n s o f the o r i g i n a l
set o f equations ( 1 ) . Theorem: Let the non-singular matrix T* be the
RMNA of a SC networks with n nodes, invut and out-
put being at nodes 1 and 2 respectively. Let G
and H be scalar block diagonal matrices as in (7)
with the following properties:
0 0 .v.. ( i ) G is non-singular
1 t
( i i ) the first row of H i s l " U j
t
( i i i ) the second row of H is k u 2
( i v ) the ( 2 n + l ) f i row of G is t u + i 2 n
vector Xo o f
9
Comparing (1) w i t h ( 6 ) , i t i s seen t h a t a number T Xo " "2n+l •••< >
of new unknowns have been introduced i n t o the sec- Now by c o n d i t i o n (v)
ond s e t o f equations, i n the v e c t o r s J-y J2' 3 - f
J ! _ 1 G u
2n+l = 2n+l u
e
s t r u c t u r e o f T ensures t h a t ( X ) ^ = V^ = 1 (see Q
699
1 2 1
E7.1
k > 0 c q k C 1 2 }
' c + c + c + c < •••
L L
l 2 ^3 ^4
i s a r e a l i s a b l e capacitor m a t r i x . Moreover, i f P
i s derived from Q using (5) , and A. and B from the
r e l a t i o n s AVP = V, P^TB = W { c f ( 7 ) ) , then GTi: i s
a transformation o f the RMNA T o f the f i l t e r satis-
f y i n g the conditions o f the theorem, where G and
H are i n ( 7 ) . Therefore, each SC network whose ~j"c,» t a n u f
capacitor matrix c' s a t i s f i e s (10), ( 1 1 ) , ( 1 2 ) above
and whose switching and a m p l i f i e r arrangements FIG. I : SECOND ORDER A C T I V E SC F I L T E R
are as i n Fig 1 i s equivalent t o the f i l t e r i n
Fig 1. Two examples are shown i n Figs 2 and 3; =j= H IP* P f
note t h a t the o r i g i n a l network corresponds t o
k = q = 1. Analysis o f these three equivalent
networks using the computer program SCNAP 2 shows
good agreement-of the t r a n s f e r f u n c t i o n s .
Note t h a t , since k and q can take a continuous
range o f values, the same i s t r u e o f the capacit-
ance values o f the networks i n the equivalence
class. Moreover, t h i s class o f equivalent net-
works i s only a small subset o f the class o f a l l
networks which are equivalent t o the f i l t e r o f
Fig. 1 and can be obtained using the conditions o f FIG. 2 : EQUIVALENT F I L T E R WITH k=1, q«0.99
the theorem. Hence o p t i m i s a t i o n w i t h respect t o
such parameters as component sum and spread, o r z - s q - f t 6 > p»
s e n s i t i v i t y would seem t o be one possible a p p l i c -
a t i o n o f the theory.
A general t r a n s f o r m a t i o n and o p t i m i s a t i o n program ,\ 0'6klia P P
6. CONCLUSIONS
41ff04ffr
f J_ 3 0 p r
700
1 2 2
PAPER 34
Proc. IEEE I n t e r n a t i o n a l Symposium on C i r c u i t s and Systems, Philadelphia, May 1987.
•Department o f E l e c t r o n i c Engineering, U n i v e r s i t y o f H u l l , H u l l ,
U n i t e d Kingdom, " D e p a r t m e n t o f E l e c t r o n i c s and E l e c t r i c a l E n g i n e e r i n g ,
U n i v e r s i t y o f Glasgow, Glasgow, U n i t e d Kingdom.
ABSTRACT TX = w (1)
Provided a s i n g l e network r e a l i s a t i o n o f a d e s i r - and o f t h e transformed system
ed t r a n s f e r f u n c t i o n i s a v a i l a b l e , an i n f i n i t e 1
GTHX = W (2)
c l a s s o f d i s t i n c t b u t e q u i v a l e n t networks can be
generated u s i n g s p e c i a l congruence t r a n s f o r m a t i o n s are i d e n t i c a l . R e s t r i c t i o n s t o t h e e n t r i e s o f G
of the capacitor matrix. Continuity properties o f and H ensure n o t o n l y t h e i n v a r i a n c e o f c e r t a i n
these t r a n s f o r m a t i o n s enable improvements t o n e t - s o l u t i o n s , b u t a l s o t h a t T" = GTH has t h e s t r u c t u r e
work parameters, such as c a p a c i t o r sum and spread, o f an RMNA m a t r i x . The r e s u l t i n g networks may be
dynamic range o r s e n s i t i v i t y , t o be achieved w h i l e exact o r scaled e q u i v a l e n t s w i t h r e s p e c t t o any o r
d e s i r e d t r a n s f e r f u n c t i o n s are preserved. To a l l o f t h e v o l t a g e t r a n s f e r f u n c t i o n s , as d e s i r e d
implement t h e method, a program SCNAPT, i n c o r p o r a - [1], [2].
t i n g a c o n s t r a i n e d o p t i m i s a t i o n s u b r o u t i n e , has
Equivalence t r a n s f o r m a t i o n s can a l s c be express-
been developed; as i l l u s t r a t i o n , t h e a p p l i c a t i o n t o
ed as congruence t r a n s f o r m a t i o n s
a range o f two-phase s t r a y s i n s e n s i t i v e switched t
73C
liii) P
phase networks, f o r example, o n l y those c a t e g o r i e s
'Vl2= o " °
l 2
of connections l i s t e d i n Table I can ever c o n t r i b u t e .
(iv) R K = K [P 1 = K
e 22' o P
o( oi 2 2
One method o f overcoming t h e problem o f low rank i n
(v) itoe 1st column of R is U the c a p a c i t o r m a t r i x , t h e r e f o r e , i s t o i n c o r p o r a t e
-1 a d d i t i o n a l f u n c t i o n a l l y redundant, o r 'phantom',
(vi) t/ie 2nd row of P is
„ .
k. k>0,
t e —2 c a p a c i t o r elements i n t o t h e i n i t i a l network date i n
ivii) C = Q CQ is a realisable capacitor mat-
ru, o r d e r t o i n c r e a s e t h e rank o f t h e i n i t i a l c a p a c i t o r
where U denotes the i t / i u«it vector. m a t r i x C. A f t e r a p p l y i n g an e q u i v a l e n c e t r a n s f o r m -
I f H (z) and H '(2) respectively denoze the even a t i o n ( 3 ) , any f u n c t i o n a l l y redundant c a p a c i t o r s i n
input/e^en output Voltage transfer function of N the network w i t h c a p a c i t o r m a t r i x C' are removed t o
1
and o/ ifte network with capacitor matrix C and t/je y i e l d t h e f i n a l e q u i v a l e n t network [ 3 ] .
same s u i t e m n g arjd amplifier arrangements as N, then
(B) = k H (z) Capacitor Plato I Capacitor P l a t e 2
ee Typo
connected switching coonocted (witching
F u r t h e r d e t a i l s and n o t a t i o n may be found i n [ 2 ] . to phase to phase
_ 1 I i/p of network Q v i r t u a l ground e
Since P = S "' Q S , P = ^ Q S and t h e e n t r i e s
E
of S ,^ ,K and K i r e ?ixed°real numbers determined ap-anp / p
II v i r t u a l ground a 0 a
by t n e f n i f i a l ne?work N, t h e c o n d i t i o n s o f t h e t h e o -
rem c o n s t i t u t e , i n e f f e c t , a s e t o f c o n s t r a i n t s on III i/p o f network e v i r t u a l ground o
the e n t r i e s o f Q. Where t h e i n i t i a l network i s
s t r a y s f r e e , some s l i g h t adjustments a r e n o r m a l l y rv v i r t u a l ground. • op-asp o/p a
r e q u i r e d t o ensure t h a t a s a t i s f a c t o r y c l a s s o f
e q u i v a l e n t networks i s generated; t h i s aspect w i l l V op asp o/p o v i r t u a l ground o
now be c o n s i d e r e d .
VI v i r t u a l ground o op-cap o/p a
731
f u r t h e r extended by using the f a c t t h a t the i n p u t s W i t h i n each o f the e q u i v a l e n t t o p o l o g i e s , the
o f o p e r a t i o n a l a m p l i f i e r s are v i r t u a l grounds t o c a p a c i t o r elements n o r m a l l y take a continuous range
weaken c o n d i t i o n ( v i ) . of values; o c c a s i o n a l l y , however, the values a r e
unique up t o a s c a l a r m u l t i p l e i n minimum element
FORMULATING THE OPTIMISATION PROBLEM topologies. A s e q u e n t i a l augmented L a g i a n g i a n
technique i s used i n the o p t i m i s a t i o n s u b r o u t i n e i n
The mathematical c o m p l e x i t y o f the o p t i m i s a t i o n SCNAPT, the m i n i m i s a t i o n subproblems being solved
problem means t h a t a c o m p u t a t i o n a l approach i s by a quasi-Newton method; the o b j e c t i v e f u n c t i o n s
e s s e n t i a l i n a l l b u t the s i m p l e s t cases. A F o r t r a n should t h e r e f o r e be c o n t i n u o u s , and p r e f e r a b l y
77 program SCNAPT has t h e r e f o r e been developed c o n t i n u o u s l y d i f f e r e n t i a b l e , over a s u i t a b l e con-
which f i r s t determines e q u i v a l e n c e and r e a l i s a b i l i t y s t r a i n e d space. Many network parameters o f i n t e r e s t ,
c o n d i t i o n s from the i n i t i a l network data, and then f o r example, n o r m a l i s e d c a p a c i t o r sum ( t o t a l sum v
o p t i m i s e s a chosen o b j e c t i v e f u n c t i o n s u b j e c t t o s m a l l e s t c a p a c i t o r ) o r spread ( l a r g e s t c a p a c i t o r f
these c o n s t r a i n t s . I t i s necessary t o s p e c i f y a t smallest capacitor) give r i s e t o o b j e c t i v e functions
t h e o u t s e t whether the i n i t i a l network i s i n s e n s i - which are not continuous on a l l o f the c o n s t r a i n e d
t i v e t o s t r a y s , so t h a t a p p r o p r i a t e c o n s t r a i n t s space, b u t are continuous on r e g i o n s c o r r e s p o n d i n g
and o b j e c t i v e f u n c t i o n s are u t i l i s e d i n the program. to a f i x e d topology. Since r e g i o n s c o r r e s p o n d i n g
I t i s e v i d e n t from t h e theorem and from t h e t o f i x e d t o p o l o g i e s are u s u a l l y simply connected,
d i s c u s s i o n i n the p r e v i o u s s e c t i o n t h a t e q u i v a l e n c e o p t i m i s a t i o n o f such parameters can be c a r r i e d out
and r e a l i s a b i l i t y c o n d i t i o n s can be expressed i n f o r each o f t h e e q u i v a l e n t t o p o l o g i e s i n t u r n i n
terms o f t h e e n t r i e s o f Q; the e q u i v a l e n c e c o n d i - o r d e r t o o b t a i n the g l o b a l optimum.
t i o n s are m o s t l y l i n e a r e q u a l i t y c o n s t r a i n t s and
Suppose a network N i s g i v e n , and t h a t , proceed-
the r e a l i s a b i l i t y conditions quadratic i n e q u a l i t y
constraints. Since t h e o n l y v a r i a t i o n i n network i n g as above, e q u i v a l e n t networks N ,N. w i t h
components under the e q u i v a l e n c e t r a n s f o r m a t i o n s i s d i s t i n c t t o p o l o g i e s T j , . . . , T r e s p e c t i v e l y have
t o c a p a c i t o r elements, e v e r y o b j e c t i v e f u n c t i o n can been o b t a i n e d whose components are scaled so t h a t no
be expressed as a f u n c t i o n o f the e n t r i e s o f C , c a p a c i t o r v a l u e i s l e s s than 1. For each i = 1 ,. . . ,
and hence a l s o , by ( 3 ) , o f the e n t r i e s o f Q; how- *' ^ C j (a) C^(q,
ever , where t h e i n i t i a l network i s i n s e n s i t i v e t o
denote the c a p a c i t o r elements p r e s e n t i n T., where
s t r a y s , o n l y those e n t r i e s o f C should be i n v o l v e d
q i s a v e c t o r i n the space o f reduced v a r i a b l e s . To
which correspond t o one o r more o f the types o f
c a p a c i t o r c o n n e c t i o n l i s t e d i n Table I . o b t a i n an e q u i v a l e n t network w i t h t o p o l o g y T. and
m i n i m a l spread S^ffJ^' u s i n g SCNAPT, i t i s r e q u i r e d
F o r m u l a t i n g the, problem i n terms o f the e n t r i e s t 0
minimise { F (<i) = max{Cj (gj ,... . c ^ (g_) ) )
i
o f Q r e s u l t s i n n v a r i a b l e s f o r a network w i t h n
nodes; t h i s would be e x t r a v a g a n t on computer time s u b j e c t t o e q u i v a l e n c e and r e a l i s a b i l i t y c o n s t r a i n t s ,
f o r even q u i t e small networks, and f o r l a r g e r n e t - and, a d d i t i o n a l l y ,
works c o u l d j e o p a r d i s e the o p t i m i s a t i o n process C
i<i> ^(gj i1
completely. The l i n e a r e q u a l i t y c o n s t r a i n t s , once
a l l o t h e r p o s s i b l e c a p a c i t o r c o n n e c t i o n s being con-
t h e s e have been d e t e r m i n e d , are t h e r e f o r e used w i t h -
s t r a i n e d t o zero. Since the v a l u e o f t h e s m a l l e s t
i n t h e program t o reduce t h e number o f v a r i a b l e s ;
c a p a c i t o r i s f o r c e d down t o 1 by the m i n i m i s a t i o n
t h e r e d u c t i o n achieved i s n o r m a l l y around 75%. I f
process, , min „ , .
t h e r e are s t i l l too many v a r i a b l e s f o r o p t i m i s a t i o n
t o proceed e f f i c i e n t l y , t h e user has the o p t i o n o f Vai' - a Va'
imposing a d d i t i o n a l c o n s t r a i n t s which w i l l reduce s u b j e c t t o t h e above c o n s t r a i n t s , a n d the g l o b a l m i n i -
b o t h the number o f v a r i a b l e s and t h e c o n s t r a i n e d mum i s
S m i n S (
space over which the o b j e c t i v e f u n c t i o n i s m i n i m i - min = i l i '
sed. A s i m i l a r approach i s adopted t o minimise n o r m a l i s e d
c a p a c i t o r sum; note t h a t t h i s t e c h n i q u e a v o i d s t h e
Each p o i n t x i n the c o n s t r a i n e d space corresponds problem o f i d e n t i f y i n g t h e s m a l l e s t c a p a c i t o r s .
t o a s p e c i f i c network N(x) which i s e q u i v a l e n t t o Sometimes t h e number o f e q u i v a l e n t t o p o l o g i e s i s so
t h e i n i t i a l network and has the same s w i t c h i n g and large t h a t i t i s only p r a c t i c a b l e t o consider a
a m p l i f i e r arrangements. I f x i s an i n t e r i o r p o i n t , sample o f the p o s s i b i l i t i e s .
then N(x) can be expected t o have t h e most g e n e r a l
e q u i v a l e n t t o p o l o g y , whereas, i f x i s a boundary A number o f o b j e c t i v e f u n c t i o n s have a l r e a d y been
p o i n t , N(x) w i l l have fewer c a p a c i t o r elements, i n c o r p o r a t e d i n t o SCNAPT, b o t h f o r the purpsoe o f
s i n c e a t l e a s t one o f t h e r e a l i s a b i l i t y c o n s t r a i n t s i d e n t i f y i n g the e q u i v a l e n t t o p o l o g i e s , and i n o r d e r
i s active. The most g e n e r a l e q u i v a l e n t t o p o l o g y t o improve s p e c i f i c performance f a c t o r s such as
i n c o r p o r a t e s a l l the c a p a c i t o r c o n n e c t i o n s which c a p a c i t o r sum and spread. There i s no d i f f i c u l t y
occur i n any o f the o t h e r e q u i v a l e n t t o p o l o g i e s ; in principle i n incorporating further objective
because o f t h i s p r o p e r t y , an e q u i v a l e n t network N f u n c t i o n s , f o r example, weighted o b j e c t i v e f u n c t i o n s
G which r e f l e c t those parameters c o n s i d e r e d t o be most
w i t h the most g e n e r a l t o p o l o g y can be generated i m p o r t a n t , p r o v i d e d they can be s i m p l y e v a l u a t e d
from the i n i t i a l network by u s i n g i n d i v i d u a l capa- w i t h i n the program, and are c o n t i n u o u s on i d e n t i f i -
c i t o r elements (which are c o n t i n u o u s on the whole able s i m p l y connected r e g i o n s o f the c o n s t r a i n e d
c o n s t r a i n e d space) as o b j e c t i v e f u n c t i o n s and then space. A d d i t i o n a l range c o n s t r a i n t s may be used t o
maximising w i t h SCNAPT. The r e m a i n i n g e q u i v a l e n t p r e v e n t improvements t c some parameters from r e s u l t -
t o p o l o g i e s can be o b t a i n e d by successive m i n i m i s a - ing i n unacceptable d e t e r i o r a t i o n i n o t h e r s .
t i o n s o f c a p a c i t o r elements p r e s e n t i n N .
732
1 2 8
p a r a m e t e r s w h i c h a r e p o s s i b l e i n a g i v e n c a s e depend c
' 4 C '=a(c-b)C
b o t h on t h e t o p o l o g i c a l s t r u c t u r e and on t h e compon- 1
7 2
I n some o f t h e e q u i v a l e n t t o p o l o g i e s , s w i t c h e s
a r e e f f e c t i v e l y l o s t b e c a u s e t h e y become r e d u n d a n t ; c;ic ),
3 " i c,'
t h i s o c c u r s i n the f i r s t order s e c t i o n i f = C5 =
0, s i n c e no c a p a c i t o r s a r e c o n n e c t e d t o node 4 , and
i n t h e l o w p a s s f i l t e r w h e r e many s w i t c h e s a r e s h a r -
ed i n t h e e q u i v a l e n t n e t w o r k s i n F i g . 3 b . This pro- C ( O
c e s s c a n be r e v e r s e d by t h e u s e o f r e d u n d a n t s w i t c h
pairs [3].
CONCLUSION c
• ^» h c C
3 •<V<5"" 3 ' i l - < , c l C i C
4 '<li* 2 • Y3 < M,hC
4' Y4- 7< 1C
ACKNOWLEDGEMENT
7
REFERENCES 1 i j ^ i iil^H™
1. D . J . G i l b e r t and J . I . S e w e l l , "On t h e g e n e r a t i o n
of equivalent switched c a p a c i t o r networks", Proc.
ECCTD'85, p p . 6 9 7 - 7 0 0 , P r a g u e , 1985.
2. D . J . G i l b e r t and J . I . S e w e l l , M u l t i p h a s e e q u i v a -
- - - - - I I - - - 7 -
-- — I J
lence theory for switched c a p a c i t o r networks",
P r o c . 29th Mid-West S y m p . C i r c . S y s t . , L i n c o l n , Cj =kaC[,C +C^=abC ,C}=abc ,C^=adC ,CJ»bcC ,C£scdC
2 2 3 4 5 6
Nebraska,1986.
C e c C c C
733
12 7
PAPER 35
1 2 8
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS. VOL. CAS-34. NO. 12. DECEMBER 1987 1611
I E
or stuck-at-zero.
V. AREA A N D T I M E CONSIDERATIONS
This testability enhancement technique is employed in imple- Manuscript received May 6, 1987; revised August 12, 1987 L. Ping was
menting a testable serial-parallel multiplier. A prototype of multi- supported by the Royal Society and British Telecom.
The authors are with the University of Glasgow, Glasgow G12 8QQ, United
plier with 4-bit multiplicand and 3-bit multiplier has been imple- Kingdom.
mented using 4-/i m CMOS (NORA) technology, Fig. 4, the total I E E E Log Number 8717615.
INTRODUCTION
c + r + G y ( 2 ) 13
( i T ^ r ^ r =
Equation (2) is equivalent to
SH rt>i r>i rO-i
1 1
/ 4*-' 2z" \ l + z' —t
AV={ jT TG}V + TJ (3a) ii
7
-JUL
with
A = C+T + G. (3b)
12 20 21
is
Fig 2. (a) L U D type signal-flow-graph simulation of the circuit in Fig. t(a).
A~D„ +B (5) (b) L U D SCF simulation of the circuit in Fig. 1(a).
13 0
I E E E TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-34. NO. 12, DECEMBER 1 9 8 7 1613
TABLE 1
D E S I G N D A T A FOR THE S C F I L T E R EXAMPLES OF F I G . 1 A N D F I G .
F i f t h o r d e r Lowpass E l l i p t i c Case
p a s s b a n d edge 1. s t o p b a n d edge 1.555723827
passband ripple 0.0436 dB srophand r i p p l e 41.9 dB
G i - GL - 1 LI 0 CI 0 867058 L2 1.223976 C2 0.L37493
C3 1.S1194B L4 0.940651 C4 0 406594 C5 0.669168 L5 0
F i f t h o r d e r Lowpass E l l i p t i c Case
p a s s b a n d edge 3.325 kH2 s t o p b a n d edge 4.8 kHz
CI 1.125332 C2 1.125332 C3 3.664828 C4 1.128354
C5 5.732167 C6 7.104417 C7 1.000000 C8 1.938541
C9 1.000000 C10 1.000000 Cll 7.464650 C12 3.953S23
C13 4.867762 C14 7.053555 C15 1.295663 C16 1.654369
C17 1.351268 CIS 1.000000 C19 1.905143 C20 1.000000
C21 1.000000 C22 1.595858 C23 2.100672 C24 1.430766
number o f c a p a c i t o r s 24 number o f s w i t c h e s 26
number o f op amps 6 total capacitance 62.4922
capacitance spread 7,4646 clock frequency 32 kHz
s i x t h o r d e r Bandpass E l l i p t i c Case
upper p a s s b a n d edge 1 kHz l o w e r p a s s b a n d edge 0.9 kHz
upper s t o p b a n d edge 1.08 kHz l o w e r 6 t o p b a n d edge 0.829 kHz
CI 1.000000 C2 1.000000 C3 23.35666 C4 13.72274
C5 2.353740 C6 2.807107 C7 2.180592 C8 1.030186
C9 3.930910 C10 1.000000 Cll 7.101668 C12 4.910849
C13 4.977363 C14 5.842078 C15 6.643089 C16 5.464835
C17 1.000000 C18 1.000000 C19 12.62665 C20 10.11870
C21 1.000000 C22 1.375713 C23 1.432187 C24 1.000000
number o f c a p a c i t o r s 24 number o f s w i t c h e s 26
number o f op amps 6 total capacitance 116.8751
capacitance spread 23.3567 clock frequency 8 kHz
20 Examples
CONCLUSION
PAPER 37
Li P i n g , B S , M S
Prof. J.I. S e w e l l , B S c , P h D , C E n g , F I E E
0
C 6
9
.n T L
l T 3 C
t C
5 T 7 C
119,
V V
1 1 ~ > 1 1 3 I 1
SC„*g i n
i i ; i j I i
SC,-9„
su sc
su t"SC 5 ' 5L 6 t'SC .g ; L
1 1 1 ]
1 1 X, I I X ? 1 1 X 3 1
b
1 1 1 1 I 1
in in in in in in
SL 2
sc 2 2 "SC 3 3
S L
6 "5C >g
w L
1 1 X, 1 1 X; 1 I X, 1
b
\06kn l.OOkn I.00W1 1.23W) lOOkn l.65kn
h i
72nF 85nF
2.0nF
A
80.1 nF 66.5nF
Cj = 6.370F L = 4.789H 4
c , = 6.370F L = 4.330H b
c , = 3.585F
b Leap-frog type S F G simulation of the passive circuit Fig. 2 7th-order elliptic lowpass ladder
c Corresponding aclive-RC realisation (normalised) a Circuit diagram
I S l s
9i. = 9t=
C . - 3.450F c = 0.1717F2 L,=4.137H
1
to make the matrix inversion (sC + G)~ i n eqn. 3 some- Cj = 5.60IF C = 0.80I6F L, = 3.828H
4
l T
s~ D A L L
0
3 The LU decomposition simulation method
'D 1
-I/.. I + _0_
(8) Perform the bilinear transform on eqn. 1 (let sampling
period T = 2 s for simplicity)
Then the so-called L U D - t y p e S F G and corresponding
1 +
active-/?C implementation can be obtained according to v + G\v = j (9)
eqn. 8 (see Fig. 3). This circuit configuration w i t h altered 1 + .
component values can also realise an 8th order bandpass
T o produce a system w i t h o u t delay-free loops, the f o l l o w -
filter. N o feed-through loop exists i n the circuit. O n l y i n
ing manipulations are required. Eqn. 9 is equivalent to
the lowpass case is one extra op-amp required when com-
pared w i t h the leapfrog approach. I n other cases the + z
number of amplifiers can be made canonical. A + r +- J (10a)
(1
with
A = c+ r+ G (10ft)
F o r simplicity rewrite eqn. 10a as
I06kfl 35 9 n F
^ I J - ^ -(<D4r + 2C)TX"] (1+4/
0
A (14)
42.8nF 43.8nF 4 B . 0 n F 2.54nF 4 9 8 n F
It II I The corresponding S F G is shown jn Fig. 5, and is suit-
Z16nF U 35nF
able f o r SC realisation [ 1 0 ] . The realisation of input
f u n c t i o n (1 + z) can be accomplished by m u l t i p l y i n g by
1
"Out z " giving (1 + z " ' ) . but i n t r o d u c i n g a delay of one
period.
F o r digital realisation T can be decomposed as i n the
Fig. 3 LUD type simulation of circuit in Fig. 2a
leapfrog method, r = A D Al, i n this way the number
L L
a Diagram
b Corresponding aclive-RC realisation Iscaled for maximum dynamic and denor-
of multiplications is reduced. The corresponding stan-
malised. 3.4 kH2) dard L L ( L U D - l e a p f r o g ) can be realised directly by the
circuit shown i n Fig. 6a.
N o w it is essential to take into consideration some
F o r a prototype w i t h more loops than nodes, it is pos-
factors related to the operation speed. Let
sible to start w i t h a l o o p description of the network
instead o f eqn. 1. I n more general cases, h y b r i d descrip- m = order of the filter (15a)
tions of the p r o t o t y p e can be used.
n = number of nodes of the prototype (15b)
4 T h e L U D a n d leapfrog ( L L )approach for Ts = sampling period (15c)
discrete systems
T„ = time for one multiplication operation (\5d)
It is now logical to examine the application of this theory
to discrete implementations. T o a f f o r d some means of T* = time for one a d d i t i o n operation (15e)
comparison w i t h existing approaches. Fig. 4 shows cas- F o r an ordinary odd-order lowpass RLC ladder,
caded biquad and wave digital realisations of a standard m = In - 1. Examining eqns. 15 and F i g . 6a carefully, it
can be seen that when the transfer factors {c,} of the hori- along horizontal branches must be performed serially, for
zontal branches i n Fig. 6a are general numbers, the oper- the upper line f r o m left to right and for the lower line
f =2 816 x 10""
;) A ^ J l ^ J ) A^nLj^A u^x__
o/p
ation speed is limited by f r o m right to left. Since 7" > 7^, the limit is dominated
m
A = SAS
S fls = SA 4D AlS
L L G =
S S2GS
l
V = S' V
s J = SJ
S (17)
Fig. 5 LUD-type SFC for the elliptic filter A new system is obtained w i t h a transfer function differ-
j, = 0.2056 6, = -0.9668 r , = -0.0850 ing f r o m that o f eqn. 11 only by a constant.
a, = 0.1420 b, = -0.9668 c, = -0.1509
aj = 0.1412 fi, = -2.011 c, = -0.1201 (A, + V<t>B + <VG,)V = 4>(1 + z)J,
s S (18)
u, = 0.2070 6. = - 1.045
fc, = -2.138 J, = -2 Scaling is carried o u t so that A, w i l l decompose i n t o
ft, = -1.093 tf ; = -2
b- = - 1.093 A =UJD U,
S S (19)
B. = SA 4D AlS
L L (20) can be expressed in (n — 1) x (n — II and (n - II x n
matrices, respectively,
In the lowpass case there are (n - 1) inductors in the (iii) Type M3 (Fig. 6d)
prototype (Fig. 2a). D Lis (« — 1) x (n — 1) and A is L
A V A v A
A y A V A v A
o/p
f_.= 0 5 7 8 7
A v A A v A A v A A v AA
f = 0.585
l
u 2 > 3 4
0.765
(1 + z~ ) and f o l l o w the procedure eqns. 9-14. The rele-
vant matrices are The type M l realisation shown in Fig. 6b follows imme-
diately.
4.86
4.4 Comments
0.085 1 7.04
Since the transfer coefficients of the horizontal branches
0.151 7.08 in Figs. 6b-d are all 2*, and assuming shift and add has
0.121 1 4.83 the same cost as addition, then for M l and M 2 the oper-
a t i o n speed is limited by
1 0.085
0.1<! T . > ^ T . + 3Tm
(27)
(24a)
I 0.121
1 and for M 3 the limit is given by
m + 5
0.968 I 1 T + 2T
a m
(28)
l I 1.06 1 1
4r = A l l these new systems demonstrate a high degree of
I I 1.09 1 1
parallelism [ 1 ] , and types M l and M 2 will show low
sensitivity; thus a c o m b i n a t i o n of the features of the
(24b) biquad cascade and the wave filters is possible. As these
new structures only use some simple matrix operations
(see eqn. 14), they are particularly attractive i n the case
where an array processor is available.
2G = (24c) It is i m p o r t a n t to distinguish between the number of
multiplications and the number of m u l t i p l i c a t i o n coeffi-
cients to be stored i n the new structure. T r a d i t i o n a l l y , in
biquad and wave realisations these t w o processes are
The standard L L realisation can now be d r a w n (Fig. 6a). inextricably linked. In some of these new methods
T o produce a type M I realisation, continue by choos- (standard L L , types M l and M 2 ) , because of their sym-
ing the scaling matrix S ( f r o m eqn. 17) metrical structures, t w o identical multiplications {c,} can
be undertaken serially by one multiplier. F o r hardware
S = diag [0.603, 0.444, 0.735, 0.765] (25) implementation, therefore, the number of multipliers
required is nearly cononical f o r these circuits. The adders
Then f o r eqns. 17-20 can be shared i n a similar way.
The methods introduced i n this Section can be applied
directly to bandpass but not to highpass and bandstop
design. The difficulty is that f o r these latter cases the
A = SAS = U.D.UT transfer f u n c t i o n is not of zero value at z = — 1
(corresponding to s = oo i n continuous domain).
2~ J
1 However, the input function of eqn. 14, ( I
z
+ " ' ) l = - i = 0 - This implies that the transfer f u n c t i o n
2
1
reference. F o r the bandstop case some different tech- By comparing the power of (jQr in the denominator
niques are adopted which are beyond the scope of this and numerator it can be seen that in order to retain the
paper. output i„ nonzero, the coefficient of (jni"" i n det | Y\
must be zero. This leads to the f o l l o w i n g :
5 C o m p a r i s o n of t h e various approaches
Remark: T h e system given i n eqn. 30 has nonzero
In the above Sections four types o f digital filters simulat- response at Q = 0 only i f T is singular. The response is
ing passive ladders have been proposed (standard L L given by
and types M 1 - M 3 ) . N o w they are compared w i t h each
other as well as w i t h cascade biquad and wave realis-
ations. n *i
J- ti at Q = 0 (34)
5.1 Sensitivity estimates 4?, i det ( T ) , , + gm det (D„„
W h e n the coefficient matrices i n eqns. 11 and 18 are rea- Otherwise, i f r is nonsingular, then v = 0 at fi = 0. n
lised by digital circuits, they will deviate f r o m their ideal Using this result f o r a lowpass filter, it would seem
values owing to finite wordlength. This also happens mandatory to ensure that T is singular. However, f o r
when they are realised by SC circuits, owing to element L U D it is easily seen that deviation i n the entries o f r
deviations. I t is observed that the sensitivities o f all four may cause it to become nonsingular; this is also true for
types o f circuits differ most significantly at u> = 0. This type M 3 f r o m the relation that r = S' (BJ4]S~K Only l
7m+ 5
Number of additions 7m 6m-2 3m + 1 3m + 1 3m + 1
2
Number of multiplier coefficients to be 3m - 1 3m + 1 I'm + 1) 3m + 5 §(m + l ) 3m-5
stored (scale factor f , not included)
c 2 2 2 2
3vn - 1 3m + 4
Number of delays m m + 1 m + 1 m + 1
2 2
Lower bound on T, (T and T are
M M
m +5 m + 1 m + 5. m + 5 m + 7 .
multiplication and addition times, r. + 3 r „ T. + 37\, T. + 2TM
respectively)
m = order of function
frequency, kHz
5 -0.4
5 -60
Fig. 7 Ideal response of elliptic filler
a Overall ideal response (wave digital with 16 bitsl
b Ideal passband response of elliptic filter
PAPER 38
1 4 4
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS VOL ."'6. SO " . .111.1 1 S»Sy 1041
Meut Vilue 101 . 70 DV sio E n The statistical estimations reported in Table 11 indicate that an
8 x 8 neural net controller which has the above characteristics is
DUtribntion 54
expected to compute configuration matrices within a period of
120 ns, and with at least 98 percent of average throughput
T A B L E II efficiency.
PERFORMANCE E S T I M A T I O N OF T H E 8 X 8 N E U R A L N E T CONTROLLER
V. CONCLUSIONS
Avoraga rank Avenge Throng hps I Analog Computation
Efficiency
This letter has demonstrated the efficiency of neural networks
S 110a*
to arbitrate the packets at the input of an 8 X 8 crossbar switch
6 U0u with nearly optimal throughput performance. The proposed ar-
4 1009 120u chitecture computes configuration matrices in a lapse of time of
2 100O UOna
approximately 100 ns, with an average throughput efficiency of
at least 98 percent.
The number of neural interconnections of this architecture is
totally optimal configuration matrices, the analog computation of the order of A/ (instead of N for a fully connected neural
3 4
which results from the set of synaptic weights of (4) leads to very net), N being the dimension of the crossbar. This reduction in
satisfactory throughput performance. the number of neural interconnections will facilitate hardware
implementation of the proposed neural net arbitrator with VLSI
IV. REAL-TIME SIMULATION AND OPTIMIZATION OF AN
technology, and perhaps of larger size controllers, e.g., 16x16.
8 x 8 NEURAL NET CONTROLLER
As for the 8x8, the resistance of the larger neural nets can be
Given an arbitrary input request matrix, the problem of maxi- optimized through computer simulations of the real-time evolu-
mizing the throughput of the crossbar reduces to finding numeri- tion of their neurons for various samplings of the input request
cal values for the set of synaptic weights defined in (4), and the matrices.
characteristics of the amplifiers of an N x N neural net which Finally, in view of the simplicity of this architecture, i.e.,
would compute optimal configuration matrices in the least synaptic weights having the same sign a_.d absence of exterjal
amount of time. It is clearly impossible to simulate the analog input currents, an implementation based on opto-electronic or
computation of a configuration matrix for each of the 2" N X N photonic technologies would probably be more suitable for the
binary matrices. For N = 8, there are already of the order of 10" larger size and faster switch controllers.
possible input request matrices. In order to estimate the perfor-
mance of the net, we have generated separate subsets of 500 ACKNOWLEDGMENT
input request matrices by choosing their matrix elements to be The authors are indebted to Dr. S. M. Walters for suggesting
" 1 " with a probability x and "0" with a probability 1 - x. X the problem and for stimulating discussions.
being a positive number between 0 and 1. Within each subset, the
analog computation of a configuration matrix associated to each REFERENCES
input request matrix has been real-time simulated by integrating 11] T. P. Troudet and S. M. Wallers, "Hopfield neural network architecture
for switch control," lo be published.
on a VAX 8650 the set of coupled differential equations (1). The [2] J. J. Hopficld, "Neural networks and physical systems with emergent
selected electronic characteristics of a VLSI implementation of collective computational abilities," in Proc. Natl. Acad. Sci. USA, vol.
the neural cells are reported in Table I for the 2-jim CMOS 79, pp. 2554-2558. 1982.
(31 , "Neurons with graded response have collecuve computational
technology. properties tike those of two-state neurons," in Proc. Nail. Acad. Sci.
In Table I , G„, u^ , p and y denote the average gain, offset,
r USA, vol. 81, pp. 3088-3092, 1984.
input resistance and propagation delay of the amplifiers respec- [4) J. J. Hopfield and D. W. Tank, "Neural" computation of decisions in
optimization problems," Biol. Cybern , vol. 52. pp. 141-152, 1985.
tively, and Q is the capacitance per connection. Such a distribu-
tion of the input resistances of the neurons is chosen in order to
break the symmetry of the neuron time-constants so as to en-
hance the performance of the neural dynamics. The TWINTOR in Bandstop Switched-Capacitor
For N = 8, the resistances between neurons of the same row
Ladder Filter Realization
and the same column have been randomly chosen distributed
around 4 kfl within 1 percent. The initial voltages of the "free" L I P I N G A N D J. I. S E W E L L
neurons have been randomly distributed in the interval [-2.5
mV, + 2.5 mV] around the ground potential. The stability of the Abstract — A new design for strays-insensitive bandstop switcbed-eapaci-
analog computation with respect to noise has been tested by tor ( S Q ladder filter structures is introduced. A two channel scheme
simulating, at the input of each inverter, a white noise gaussianly obviates the need for term cancellation in realizing bandstop-type opera-
distributed around 0 V with a standard deviation of 10 jiV. With tors and is less demanding on opamp settling time.
this set of parameters and initial conditions, the time-evolution of
the 8 x 8 neural net controller has been simulated for a period of Manuscript received November 9, 1988. This letter was recommended by
120 ns. The results of the simulation are reported in Table I I for Associate Editor T. R. Viswanathan.
subsets of input request matrices having an average rank of 2, 4, ing.TheUniversity
authors are with the Department of Electronics and Electrical Engineer-
of Glasgow, Glasgow G12 8QQ, U.K.
6, and 8 respectively. (The rank of an input request matrix being I E E E Log Number 8926456.
with
! = / i d * UI
(c) l
a= a(n
Fig. 1. (a) A Ihird-order normalized low-pass ladder, (b) Low-pass to band-
stop transformation, (c) The signal flowgraph of leapfrog-type simulation. ^-(/i-'-^/U-'+ji).
Topologically, decompose T into
I. INTRODUCTION
T= A D Al
L L (5)
Switched-capacitor (SC) filter structures based on passive lad-
der simulations have attracted much attention because of their where A is an incidence matrix of the inductors in the ladder,
L
low sensitivity properties. However, an instability problem exists D is a diagonal matrix of reciprocal inductance values. With this
L
in the design of bandstop SC ladders by stray-insensitive LDI (4) can be rewritten in the form
integrators [1]. A second-order building block technique has been
AV=Q(A fV + 2GV) + ( - l + Q)J (6a)
proposed in [2] to overcome this difficulty. L
l x T
Following a matrix leapfrog method [3], [4] a passive low-pass when n even (7a)
reference RLC ladder, Fig. 1(a), is described by the nodal admit-
tance matrix equation Q [ / ( » ) - / ( » - 2 ) ] =-C x°(n) + C x'(n-l) 1 2
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS. VOL. 36. NO. 7. JULY 1 9 8 9 1043
lor y = * x . C , i l , C j = P . C j = 1
for y»(jlx.C,=p.Cj=1.Cj=l
T 2T 3T a 5T 6T
(b)
even
channel
~ ° channel
channel
-0-
5- >odd
channel
(c)
Fig. 2. (a) A TWINTOR circuil. (b) The clock waveform, (c) Two channel
equivalent connectors of TWINTOR's.
2
Notice that the denominator (1 - z~ ) is exactly realized without a ladder structure, some simplifications are possible by separat-
term cancellation. ing signals into two channels, Fig. 2(c). The first equivalence in
It can be seen from Fig. 2(b) that now the clock period is 27* Fig. 2(c) is obvious. For the second equivalence, notice that a
compared to T in a conventional LDI integrator SC circuit. This sampling signal of an even (odd) channel opamp output in a odd
means that the operation speed for the whole circuil, determined (even) period is actually the signal held from the previous period,
by sampling frequency, can be doubled without requiring an l
therefore a delay factor, z~ , is realized. A number of switches
increase in opamp speed. are saved by this two channel technique.
By selecting suitable capacitance values 4> and * can be easily An overall sixth-order bichannel bandstop SC ladder is shown
implemented. When TWINTORS are connected together to form in Fig. 3 with the low-pass RLC ladder of Fig. (la) as reference
14 7
TABLE I
D E S I G N D A T A FOR THE S I X T H - O R D E R S C BANDSTOP F I L T E R
Specifications fot t h e B a n d s t o p SC F i l t e r
Component V a l u e s for t h e B a n d s t o p SC F i l t e r
nuober of c a p a c i t o r s 40 rtunber o f s w i t c h e s 30
number o f o p araps 6 total capacitance 439.51
capaci tance spread 3 8 . 29
REFERENCES
Manuscript received November 2, 1988. This work was supported in pan by the
University of California MICRO Program and TRW, Inc. under Grant 87-066. This
ACKNOWLEDGMENT teller was recommended by Associate Editor T. R. Viswanalhan.
The author is with the Integrated Circuits and Systems Laboratory, Electrical
The authors are grateful for the Valuable help from R. K.. Engineering Department, University of California, Los Angeles, CA 90024.
Henderson. IEEE Log Number 8926455.
PAPER 39
jln°perf@riinasi©e filter m
mmlm systems
L. Ping, BS, MS
Prof. J.I. Sewell, BSc, PhD, CEng, FIEE
network design based on matrix principles. Attention is decrease. This means that | H(P) | must have zero deriv-
given to the properties of the system descriptions of the ative with respect to any parameter p, at jm , and conse-
m
circuits. It is shown that matrix symmetry is a crucial quently the sensitivity is also zero, i.e.
factor to ensure optimal performance of the systems. Two
concepts considered by many other authors, boundedness Pt S\H(P)\
= 0 at s = jwm (2)
I H(P) | d Pi
Paper 6942G (E10), first received 7th November 1988 and in revised
form 10th July 1989
and it may be reasonably expected that over the whole
The authors are with the Department of Electronics and Electrical
passband the sensitivity will remain small, a reassuring
Engineering, University of Glasgow, Glasgow G12 8QQ. United argument used by many other authors for ladders as well
Kingdom as various simulation methods [23-25].
IEE PROCEEDINGS, Vol. 136, Pt. G. No. 6. DECEMBER 1989 327
:L 5 0
x A Ax 1 7 X*(s C + s G + r)X = 0
m m (14)
Mil = max — f (7) (X* denotes the transposed conjugate of X) or
1*0 x x
blocks [19]. In Section 4 it will be shown that higher should be added to ensure that no roots lie on the ima-
order networks, based on a symmetric matrix decomposi- ginary axis. This can be checked by evaluating
tion approach, can also be designed to meet this condi- det | Y(jw) |. In most cases the system in eqn. 11 is
tion. designed to realise a transfer function H(s) which has no
328 I EE PROCEEDINGS, Vol. 136, Pi. G, No. 6, DECEMBER 1989
15 1
poles on the imaginary axis. If the system is properly From eqns. 27 and 28
designed without redundancy so that the order of the
V*G V=iJ*R J
system is equal to the order of H[s\ or in other words it aM in
is observable from the output, then it will have no poles -{J* - 2y*GjRjJ (29)
on the imaginary axis either, as in this case H{s) and the
system have the same set of poles. If G,„ is non-negative, then from eqn. 24 / f is also non- in
The non-negative property of the symmetric matrices negative, which means that (7* - 2 Y'GJRjJ - 2G K| in
C, T and G can be easily tested. For instance, decompose is a non-negative number. Theorem 1 follows from eqn.
Cinto symmetric L U form [28] 29 immediately.
(37)
= Re { y*J] — y*G y in (28) Yin + 011 + z,-.
/££ PROCEEDINGS, Vol. 136, Pi. G, No. 6, DECEMBER 1989 329
15 2
So p is just the reflection function and the upper bound tributioti to Y of a branch admittance jq between nodes k
network theory.
In the proof of the boundedness, no condition has
been imposed on C and T except that they must be sym-
metric and non-negative. Accordingly, zero-sensitivity (43)
with respect to symmetric deviation can be achieved at
the frequency points where the transfer function attains
its upperbound.
So
4 Sensitivity formulas dY,'dq =jM (441
k ah
The above result provides only an estimation of sensi- ft is easily seen that
tivity for symmetric deviations. More general sensitivity
T 2 1
formulas are now derived. To simplify the problem, only y M y=v
ab k and VM y=\v \ab k (45)
single-input/output system will be considered.
where v is the voltage across jq .
k k Then eqns. 41a and b
Suppose a single-input/output system meets the condi-
are reduced to
tions of theorem 1 and eqn. 30. Let the system in eqn. 1 la
be excited by another arbitrary input K instead of J and dvjdq t = (2g„„! v„ \)~'(-fpv 2
k + |v | ) k
2
(46a)
let the response be U. The system can be written as
and
(38)
d\v„\ldq k = (2g„ \ \y> n K Im [>n (46b)
then the output of the new system is related to the old
system by Eqn. 46b is zero at the frequency points where p = 0 or
equivalently \v„ \ attains maximum bound. This is just the
u„ = (2g„„vj-'CpY T
- Y*)K (39)
well known zero-sensitivity property for doubly terminat-
C indicates conjugate). The derivation of eqn. 39 is given ed ladders. An alternative derivation of eqn. 46 based on
in Appendix 10.1. a topological approach is given elsewhere [31].
2 T
'-<fc-Ir (486) - c + ^-r
T 2
+G 154)
2
^. T From eqns. 51 and 54 it can be seen that the following
J? c + r relationships hold when J = [ 0 ] :
r 2
T T r T
x [n\A Ax[n\ = .v (ii).r(n) r (n)Gv{n) (55«)
which can be seen as a generalised form of the standard
state-space equation (eqn. 4), by introducing a second- T
If G is non-negative, then v [n\Gv{n) is a non-negative
order term. number, and therefore
As bilinear transformation will keep the stability pro- T T r
perty and map the imaginary axis in the s domain to the x {n)A Ax(n) < jr (n)x(n) 155b)
unit circle in the z domain, it is easy to show: no matter what the value of x(n). The theorem follows
from eqns. 7 and 9. Incidentally, from eqn. 54 it can be
Remark: The system in eqn. 47 has no poles outside the seen that matrix A is orthogonal if G — [ 0 ] .
i a T
unit circle z = e if C, T and G are all symmetric non- Conditions for the continuous time-domain systems
negative and has the same boundedness and sensitivity have no direct practical applications; however, for com-
properties as indicated by theorems 1 and 2, except that pleteness a derivation is given in 10.2 Appendix.
Jt r
eqn. 47 is evaluated on the unit circle z = e ° .
A = 2' £ ?
Y - c ^ ir + G\ [I I ] + or the right-LUD form
+ c r
T +
_-L j\T 2 UsC + G)V= -L U+J r (57a)
(53a) l J
{ U=s- L V v (57b)
Proof: First, only if C and T are both non-netative can Crii.i-i) + 'el,-. ,))*'.• = + Vii.i)f (
T
1 / 1 dvjdc,, = (2y„v,r\-f>V + V*)jm dCdc V
1
= (2<J„„[~„r ' M - p V i V j + Vfi-j) 159)
V
1 .1 Similar formulas can be derived in the same way for
other elements.
1 (..I For the lowpass case, the right-LUD system (eqn. 57)
J: O I results in identical circuit structures to those derived
c It 1 from a leapfrog approach. The left-LUD simulation (eqn.
-oW
56) has a significant advantage for elliptic transfer func-
tions in that there are no capacitor-coupled op amp
loops in the circuit implementation. The right-LUD
simulation may produce such loops, and high-frequency
1 oscillations may exist in these loops, resulting in undesir-
D I.I
able noise problems.
-
c1 2 c
r ^ i
8
n^ 10
/
\P + 4
-1
;
z~ l
-T r + 2
\
rG V
II
r, c, c c«
II c,„
II
n
7 c
=
1 + z''
T T 7 ^ Y
< > 60
4 c, c .4c. c / 4 c 11 L
' W =
- ( j ~ ~ i \ T + 2 G
) V + ( X + Z ) J 62fl
< >
I I II
7 L y
l = yr^ w 62b
( >
or the right-LUD form (with L L\ = T/2 V) r
out
PV= -~^— (L U
l r + GV)+ ' _' J + Z
(63a)
1- z 1
1— z
2
c 3
V = ^—nL\V (63b)
12 c A A 13 1— z
V
1
8 10
tive elements in the prototype design is discussed. When
the prototype relation (eqn. 11) is produced by nodal-
Fig. 2
1 11 a SFG representation of a continuous domain equation
b Corresponding active-RC realisation
c Left-LUD type active-RC realisation
d Right-LUD type active-RC realisation
•0.5
- 0 9 j-
/'-'21"22 \ F
/- 32 33 / P
\ cn j
"D i
I — s 2g *
i
1 1 1 1
h 1 P
1 1-—' 33
11 L7 P
22
-2.1
Fig. 4 Simulated response of the circuil in Fig. J C
Cl6 c 6
c 7 C13 6.4.1 Matrix method: Eqn. 52 has already given one
II II 11 II
II H II II realisation of the system (eqn. 47). Direct implementation
\
r ,L° ]
'o
I' I of the multiplication of X and U by A and B can require
a high number of multiplications and additions, as
usually A and B are full matrices. If A and B are decom-
Fig. 3 posed according to eqn. 53, the multiplications of A and
a Terminated all-pole LC ladder prototype B are then carried out by a sequence of multiplications
Component values Tor normalised ladder by factors. Efficiency is achieved by taking into consider-
C, =0.93106 C , - -0.073694 L , = 1.4326
C = 1.9453 C, = -0.071313 L. = 1.4804
ation matrix sparsity. Numerical methods, such as L U
3
! |
dependent factors are —1/(1 - z " ) and z " ' / ( • — z " ) , as
shown in eqn. 60 which can be seen as a stable and exact
LDI-type realisation. The system expression of eqn. 49
involves matrices P, C and T, whereas that of eqn. 60
involves matrices P, T and G. Usually, matrix C is more
complicated than G. Therefore, implementation based on
eqn. 60 may result in simpler structures. However, no
output vector simple method for oscillation suppression has been found
for standard LDI-type [11-12] digital structures nor for
the exact structures based on eqn. 60. Therefore, struc-
tures based on eqn. 49 are suitable for digital design, if
both sensitivity and noise due to number-truncation-
induced oscillations are considered. Realisations based
V X
on eqn. 60 are interesting for other discrete systems, in
particular for SC filter design, due to the fact that the two
L D I factors coincide with the transfer functions of a pair
of stray-insensitive integrators, resulting in efficient fabri-
cations [6].
7 Conclusions
capacitor network design', IEEE Trans., 1987. CAS-34, (12), pp. X = (73)
1611-1614
21 P I N G , L , H E N D E R S O N , R.K., and S E W E L L . J.I.: Matrix with
methods for swiiched-capacitor filter design'. Proceedings of I E E E
ISCAS. Espoo. Finland, June 1988, pp. 1021-1024 C = (74a)
22 P I N G . L., and S E W E L L , J.I.: 'Filter realisation by passive circuit
simulations', Proc. IEE C. 1988, 135,(4). pp. 167-176 and
23 S E D R A . A S, and B R A C K E T ! ' , P.O : Filter theory and design T
T
C = ADA C C C (76«)
•>•->
T = A D Al
v r (766) - A l - Y = : l
W c l
(80a)
where D and D care diagonal branch-admittance r
yc = T
A V
C (77a) 2
x'(n)x(n) = - O >»' «(")
c c
V T = A\V (lib)
PAPER 40
16 0
400 M H z SWITCHING RATE GaAs behaviour and stability in all switching states. The filter archi-
SWITCHED CAPACITOR FILTER tecture is shown in Fig. 1. where capacitors C,,. C and C f c
0 2361 FD
nput 39 0 IpF 0 lpf(F3s
Introduction: Gallium arsenide (GaAs) technology has an
n — m
important role in the realisation of high speed analogue
1 01?668pF "T " J _ 0V
sampled data systems. Reference 2 describes the design and C7C:06p
optimisation of a second order switched capacitor (SC) LpF j. -
0V OV
bandpass filter for implementation using GaAs technology
and with a maximum design switching frequency of 250 MHz. Fig. 1 Second order SC filter for implementation
Since Reference 2 was published, the filter chip described has
been fabricated and very encouraging experimental results The circuit was layed out using M A G I C with a specially
obtained. The objective of this paper is to report these results. written technology file for the Anadigics 0 5/rm GaAs process
and a layout plot is shown in Fig. 2 The size of the chip,
Switched capacitor filter design: The second order switched which was fabricated by Anadigics, is 3 3 x 2 6 mm and the
capacitor bandpass filter was designed to realise a Q-factor of power consumption is 440 mW.
16, midband gain of unity and midband frequency of l/25th of
the switching frequency for a maximum switching frequency of Testing and experimental results: The chip was bonded within
2
250 M H z . The circuit used the single-stage, double-cascode, a dual-in-line package, which was mounted in a dual-in-line
double-level-shifting operational amplifier design of Reference socket on a circuit board. The input signal level for all tests
3, which had a simulated gain of 60 dB, typical gain- was + 7 d B m , o r 0 5 V R M S .
bandwidth product of 3-5 GHz and a minimum settling time The filter was found to operate very accurately up to and
of 630 ps. The switch control circuit used is that of Reference beyond its design maximum switching frequency of 250MHz.
4, which features low power consumption and low signal The amplitude/frequency response was measured for switch-
dependence of clock feedthrough. The switch capacitor circuit ing frequencies of 300 MHz, 350MHz and 400MHz. The
was optimised for high frequency operation using the tech- responses for 300 M H z and 400 M H z are shown in Figs. 3a
nique of References 2 and 5, whereby amplifier load capac- and b, respectively. Expanded passband plots for the above
itances are introduced to maintain acceptable settling switching frequencies are shown in Figs. 4a and b. Peak gains,
0 1696J2 I
Fig. 2 Layout plot for GaAs switched capacitor filter
Total size = 3 3 x 2 6 mm
References
1 ALLEN, p. E., and BREEVOORT, C M . : 'An analogue circuit perspective
of GaAs technology'. Proceedings of 1987 I E E E ISCAS, Phila-
delphia, USA, May 1987
frequency, MHz 2 H A I G I I , D . G „ TOUMAZOU, C , HARROLD, S. J., SEWELL, J. 1., and
STEPTOE, K.: 'Design and optimisation of a GaAs switched capa-
citor filter'. Proceedings I E E E 1989 International Symposium on
Circuits and Systems, Portland, Oregon, USA, June 1989
3 TOUMAZOU, c , and HAIGH, D. G.: 'Design of a high-gain, single-
stage operational amplifier for GaAs switched-capacitor filters',
Electron. Lett.. 1987, 23,(14), pp. 752-754
4 HARROLO, s. J.: 'Switch driver circuit suitable for high order
switched capacitor filters implemented in GaAs', Electron. Lett.,
1988, 24, (15), pp. 982-984
5 H A I G H , -D. G., TOUMAZOU, c , and BETTS, A . K . : Chap. 10 in H A I G H ,
frequency, MHz
D. G., and EVERARD, J. K. A. (Eds.): 'GaAs technology and its impact
on circuits and systems' (Peter Peregrinus, 1989)
6 HARROLD, s. j . , VANCE, l. A. w., and H A I G H , D . G.: 'Second order
Fig. 4 Measured expanded passband response
switched-capacitor bandpass filter implemented in GaAs', Elec-
a 300 MHz switching frequency
tron. Lett., 1985, 21, (11), pp. 494-496
b 400 MHz switching frequency 7 HAIGH, D. G., and TOUMAZOU, C : 'Fast settling high gain GaAs
operational amplifiers for switched capacitor applications', Elec-
Tsble 1 M E A S U R E D F I L T E R P E R F O R M A N C E D A T A tron. Lett., 1989, 25, (1), pp. 734-736
8 LARSON, L . E., CHOU, c. s., and DELANEY, M . J. : 'An ultra high speed
-3dB frequencies F 0 Q-factor GaAs M E S F E T operational amplifier', IEEE J., 1989, SC-24, (6),
Switching Peak pp. 1523-1528
frequency gain
considerable improvement over the previous results presented Indexmg terms: Codes, Boolean functions
in Reference 6. The Q-factor accuracy is reasonable for a
Some recent work concerning the strict avalanche criterion
300 M H z switching frequency, but the Q-factor is reduced for
for a Boolean function has been motivated by the claim that
the higher switching frequencies. The noise at the output of
a certain cryptographically useful property will be true of any
the filter was measured for a measuring bandwidth of 10kHz function satisfying the criterion. In the letter it is observed
The peak noise is at — 75 dB relative to the maximum signal that not only is this claim untrue, but that possession of the
level. Observation of the filter output signal at the resonant property in question is in fact precluded by satisfaction of the
frequency indicated a DC offset voltage of less than 50 mV strict avalanche criterion.
and reasonably low levels of clock feedthrough. Future testing
on an improved jig could improve these results.
1
Definitions and motivation: Webster and Tavares introduced
Conclusions: A significant increase in the maximum switching the strict avalanche criterion (SAC) for a Boolean function to
frequency for switched capacitor filters has been demon- combine the concepts of completeness and the avalanche
PAPER 41
1 G 3
Li P i n g , M S , P h D
P r o f . J . I . S e w e l l , B S c , P h D , C E n g , FIEE
required. The summing amplifier and several resistors are summation stage reactive *.wc c<y'
3
an extra cost. ,
V,„ J,'W. W; W,
summation
passive ladder stage staae reactive two port
Fig. 2 Actiue-RLC allpass circuit
.• = « " ' L , . y , = * ~ ' C ,
i
is
passive ladder stage
sC + -r + G)y = j (6)
Fig. 3 SFGs for allpass realisation
a LUD type
where J = [ - ghv , 0 , . . . , 0 ] and C, T and G are admit-
in
Case A: 5-domain
tance matrices f o r m e d by the contributions of capacitors, » = * = s " ' . a > = » . i i = j = l . . i = 2,/I=0
inductors and resistors, respectively. V is defined as V = Case B: ;-domain
* = 1/(1 - zl"',* = ;-'/(! - ; '|. i = b = d = g = i. I = 2.0 = I
[ft . ] , where alternate signs are i n t r o - Case C: 2-domain
duced to insure that all entries in eqn. 6 are positive. <J> = 1/(1 =-'"'/(! -z'),a = 3 =d - g - l,b= I +z.fi = 0
The matrix decomposition method f o r active RC b Leapfrog type
Case A: 5-domajn
network design described i n References 9-11 can be l
®^ti/=s~ .a = b = d = g= ],tt = 2
readily applied here. Eqn. 6 can be written in the L U D Case B: z-domain
1
4> - 1/(1 - z " ) , ill - 2" 7(1 - r"'),h= I + :"'.u =* = d = g = 1
f o r m , (because all of the capacitors in F i g . 2 are con-
nected to earth, C is simply diagonal and no real decom-
position of C is necessary)
( C + s~ G)Vl
= s~ (- A W
l
L + J) (8a)
l T
tV = s DA y L L (86)
B o t h eqns. 7 and 8 are linearised w i t h respect to s~' so
that they can be realised directly w i t h active-RC circuits.
For a sixth order circuit, the signal flow graphs (SFGs) of
Figs. 3a and b (case A) and the simulation circuits of
Figs. 4a and b (incorporating the summation stages) can
be obtained. Inversion i n the output stage is incorporated
in the simulation part. Notice the different termination
stages in L U D and leapfrog realisations.
The summing amplifier employed in the output stage
in Fig. 4 need not be realised explicitly i n delay equalised
filter systems. Provided the allpass filter is succeeded by
an amplitude filter stage, the virtual earth of the input
Qa=1 Qd=
integrator of the amplitude stage can be directly con-
nected to P,. to realise the summation f u n c t i o n . I n a fully
differential implementation, separate inverters are not
required and thus realisations w i t h a canonical number
of op amps are possible.
most efficient t o use the so-called b i l i n e a r - L D l method, where 1 = 2. jJ = 1. This can be again represented by an
as it is both exact in frequency response and efficient in SFG, Fig. 3rt (case B), including the output stage. Notice
terms of implementation cost. Such a structure could be that the corresponding digital implementation is canon-
derived by the technique of Reference 11, which places ical w i t h respect to both multiplier coefficients and
real zeros in the ladder prototype to introduce the cancel- delays. The SC realisation is straightforward, but this
lation of capacitors after bilinear transformation. method results in a large capacitance spread. There is
However, a more straightforward derivation is presented another realisation of eqn. 13, by choosing 2 = 1 + z
here, which uses a continued fraction expansion f o r z- and P = 0, Fig. 3a (case C) results. Note that % and b can
d o m a i n transfer functions [ 6 - 8 , 12]. be exchanged w i t h o u t affecting the overall transfer func-
tion. The corresponding SC circuit can be obtained by
3.1 LUD method for SC and digital ladder design replacing the branches in the S F G by SC elements, see
For an alipass function i n z-domain Fig. 6a. The singie z~' of eqn. 10 is realised by a
rearrangement of switching in the sample-and-hold and
1 + . other i n p u t / o u t p u t circuitry. The sampled input f r o m an
zP(z)±z"P(z->) even phase is transferred to the output summing a m p l i -
I + fier d u r i n g the subsequent odd phase. The unit delay is
P(z) + z"P(z-
realised when the output is sampled in the even phase of
1 + z the next clock period. This scheme usually results i n a
= 1 - (9)
1 + Y(z) lower capacitance spread than f o r case B.
{— sign for n — 2m, + sign f o r n = 2m 4- 1) L U D type SC circuits will always require an even
number of op amps, which is canonical f o r even order
T o avoid the noncausal term z in eqn. 9, the transfer cases, but not f o r odd cases.
f u n c t i o n is modified to
+ z
l
z~ HJ z)
< = z' 1
(10) 3.2 Leapfrog method for SC and digital ladder design
1 + Y(z)
Use of 13 to derive directly the leapfrog type circuit
which introduces only a single extra delay. Define causes difficulty in realisation o f the termination terms.
Instead, it is easy to verify the equivalence between eqn.
1
(D = T = (11) 13 and the f o l l o w i n g system:
1 -z-'
V 'C, +
<J>~'L, +
(12)
4- 'C„ +
0>-'L„
where
C = C - G (16)
1 1
y=0x x " — y = yx The right m a t r i x decomposition structure can be derived
Fig. 5 Digital realisations of<t> and 4* [ 9 - 1 1 ] by rewriting eqn. 15. as
For n odd, eqn. 12 w i l l terminate w i t h a 4*" ' C term. m + 1
\C + ®G)V = <D(- (17a)
'J)
Positive values of {C,} and { L , } are guaranteed [ 1 2 ] . By T
W=*VD A V (17b)
analogy w i t h eqn. 2, it can be seen that a ladder simula- L L
t i o n is appropriate. The passive ladder part in Fig. 2 can The S F G of F i g . 3b (case B) can be used t o represent eqn.
again be used t o realise eqn. 12 by means of a 'passive 17 and an illustration of SC replacement is given i n Fig.
network", w i t h admittance y ^ f ' C , and impedance 6b, which is comparable to an independently reported
1
z, = (J>~ L . A l t h o u g h physically unrealisable, it can be
j
1
realisation [ 7 ] . Notice that the z " factor in eqn. 17a
used as prototype f o r SC and digital simulations. A nodal cancels the noncausal factor of 1 + z i n eqn. 10.
description can be set up f o r the ladder section of Fig. 2 Leapfrog SC simulation can always realise a circuit
in terms o f f and Q>: w i t h a canonical number o f op amps, provided the a m p l i -
tude and allpass stages are considered together. However,
-C + ®r + G)V = (\ +z)J (13) for narrow band SC design, it will result i n a larger total
capacitance than the L U D method.
L U D SC and digital circuits can be obtained [ 9 - 1 1 ] by F r o m Fig. 3b (case B), i t is seen that there is a delay-
rewriting eqn. 13: free loop at the termination stage: v -*—g t
and requires only a canonical number of multipliers for racies in the values of {C \. { L , [ and all unity-valued ele-
{
digital implementation. The number of additions ments of Figs. 4 and 6 w o u l d affect the sensitivity.
summing stage
"C
S/H stage
summing stage
" C
5/H stage
Fig. 6 SC allpassfilters
It will be proved that the amplitude response of the
a LUD lype b Leapfrog type
structures introduced i n this paper, unlike their biquad
counterparts, are completely insensitive w i t h respect to
deviation of most o f the element values and are bounded
\_ 1
for a few terminating elements. The transfer functions o f
C,l-z-' Figs. 2 , 4 and 6 have the f o r m
bav,/J l
bo- — = H a = +| a (18)
I + dvJJ,
Fig. 7 Network transformation to eliminate delay-free hop
For Figs. 2 and 4, a=b=d=g= 1, and a = 2 , and, f o r
1
Fig. 6, b = b, + b z~ and a = fc, =b =a
2 = d = g= 1. 2
required is also relatively small (roughly 2n). The digital
leapfrog realisation requires one more delay than the
Remark : For the circuits in Figs. 2 , 4, and 6, i f a, b (or fc,,
equivalent canonical L U D realisation (case B).
b ) and a are fixed, then \H \ = 1, regardless of all other
2 a
d c\H„
I (9k K ) (201 -2 < • (24r!
Ji all forward
\H„\ cd 1 +1 y < y a j | | 2
paths
Similar formulas can be derived for the circuits of Fig. 6.
with
4b.
F o r the z-domain circuit o f Fig. 6a, note that all A and
{ A } are, again, real. Referring to Fig. 3a (case C),
k
1 + z
1 + •/,/!>,
1 z 2 Ji/fi 2\ 1
1 +• 1 + 1 +
1 +Z 1 +z 1 + z
(23) Fig. 8 Sixth order bandpass SC elliptic LUD filter
2 2 JJvi 2\ 1
1 +• 1 + 1 +
1 + z 1 + z + z
tic L U D - t y p e SC circuit, see Table 1. A l l of the circuits
As (1 — z)/(l + z) = j tan (OJT/2) is imaginary for z = have been scaled f o r m a x i m u m dynamic range. The P„
e i a T
, one can see that \H \ = 1. A similar result can be A
point o f either circuit of Fig. 6 can be directly connected
proved f o r the circuit in Fig. 6b. to the input of the circuit o f Fig. 8. The amplitude and
The sensitivity formulas for the remaining elements in delay responses are shown in Figs. 9a and b.
the circuits of Fig. 4 can be derived (since Y(ju>) = J\/v l
The f o l l o w i n g f o r m u l a is used to measure the overall
is a pure imaginary number) as system sensitivity:
1
_a_d\H A
= +•
IWI € 1 (24a) (26)
da l+\Y(joj)\ \H \ CI dc,
The system delay sensitivity can be denned in the same parameters are listed in Table 3. As in Reference 3, t o p o l -
way. F o r comparison, t w o cascade biquad SC circuits are ogy 1 has quite a small spread, but very high sensitivity,
designed f o r the delay equalisation stage, using b i q u a d whereas topology 2 has an i m p r o v e d sensitivity at the
topologies 1 and 2 o f Reference 3. The resulting design cost of high spread, see F i g . 10. Other biquad topologies
C. 1.000 C ,
t 1.023 C b2 1.023 C , 2.333 C„ 2.386
C, 9.903 C 2 2.380 C 3 3.273 C, 1.000 C 1.000
5 C„ 1.011
C, 24.22 C B 2.478 C, 8.660 C , 1.000
0 C , , 1.000 C, 1 000
2
Total capacitance 109 units 138 units 102 units 311 units
Capacitance spread 29 units 29 units 26 units 62 units
Number of op amps 6 6 6 6
Number of switches 27 28 32 32
Number of capacitors 21 21 24 24
S/H and summation stages excluded
frequency, Hz x 1 0
combined
0 3
response
12
20
1.0
40
equalisation stage
60 08
78 84 90 10 2
01
0 1
-80
0i
^ 06
10
100
04
1 9
1 20
passband
28
140
8.1 90 108
frequency, kHz frequency, kHz
16 7 Acknowledgements
/
The authors wish to thank R.K. Henderson f o r his
invaluable help and suggestions. H e l p f u l discussions w i t h
8 B. N o w r o u z i a n are also gratefully acknowledged.
leapfrog LLUD
UD
/ /
N /' /' ••
I.'' i - ^' i ^ -,. ~- / I ~r- 8 References
72 81 9.0 9.9 10 8
frequency. kHz
1 S E D R A , A.S., and B R A C K E T T , P.Q.: 'Filler theory and design:
active and passive'. Matrix, Portland O R , USA, 1978
2 G H A U S I , M.S., and L A K E R , K.R.: 'Modern filler design: active
R C and switched-capacitor' (Prentice-Hall, Englewood Cliffs. New
Jersey, USA, 1981)
3 L A K E R , R.K., G A N E S A N , A., and F L E I S C H E R , P.E.: Design
LUD
and implementaion of cascaded switched-capacitor delay equalizers',
15 IEEE Trans.. 1985, CAS-32, (7), pp. 700-711
biquad f c 4 N O W R O U Z I A N , B., and L E E , L.S.: 'Minimal multiplier realisation
i of bilinear-LDI digital allpass networks', IEE Proc. G, Circuits,
Devices and Systems, 1989, pp. 114-117
5 SV1HURA, M X , and N O W R O U Z I A N , B.: A new approach to the
12
design of bilinear-LDI switched-capacitor filters having low pass-
leapfrog ^ / / band sensitivities'. Proceedings of I E E E Pacific Rim Conference on
Communications, Computers and Signal Processing, Victoria,
3i
\ /
7
I
'
t t
t tl
t
Canada, June 1989, pp. 193-197
6 N O W R O U Z I A N , B., and B R U T O N , L.T.: 'Novel approach lo ihe
exact design of digital L D I allpass networks'. Electron Leu., 1989,
25, (221, pp. 1482-1483
btquad
topology 2 7 N O W R O U Z I A N , B.: 'A new synthesis technique for the exact
design of switched-capacilor L D I allpass networks'. Proceedings of
PAPER 42
1 7 1
R.K. H e n d e r s o n , B S c , P h D
Li P i n g , M S , P h D
Prof. J . I . S e w e l l , B S c , P h D , C E n g , F I E E
misation techniques have therefore been studied [18, 19]. the values and the derivatives of p( v) where
A new algorithm is proposed which permits direct w
application of Rernez-type a p p r o x i m a t i o n methods to the N = 2 + Lijij + 1)
c (4)
problem [ 2 0 ] , By observing the similarity between the i = i
group delay f u n c t i o n and a filter amplitude function, the The aim of the curve-fitting problem is to find the lowest
techniques and theorems for amplitude approximation order a p p r o x i m a t i n g polynomial which fits the specifi-
are still valid. A stable, accurate algorithm is then devel- cations in eqns. 2 and 3.
oped for arbitrary group delay correction.
p(x)
A transmission f u n c t i o n is designed by w o r k i n g w i t h a
magnitude squared f u n c t i o n . It has the f o l l o w i n g pro-
perty:
to f o r m a p o l y n o m i a l of order N, where
N + 1= N { - M (5)
The remaining M specifications must be met by adjusting
the M positions of the touch points.
The relation between the behaviour and the order o f a
polynomial is a complicated issue. T o decide the
m i n i m u m order is a difficult problem and some theoreti-
cal discussion can be f o u n d in Reference 2 1 ; however, i n
most cases the order determined by eqn. 5 is satisfactory.
A N t h order p o l y n o m i a l can always be interpolated by
N 4- 1 constraints. Osculatory N e w t o n interpolation [ 2 2 ]
can be used to interpolate a number of derivatives w i t h
certain computational advantages over other inter-
Fig. 1 Solution of minimum curve-fitting problem by polynomial pix) polation methods [ 2 3 ] .
curves, u(.x) = m(x) + 6 and f(.x) = m(.x) — <5, can be seen
2.3 Bilateral method
as boundary functions and p(.x) is sought to fit between
Assume that, f o r the specifications
them. A t a series of points, the so-called touch points, p(x)
(i) all fi-, are odd
will touch u(x) and l(x) alternately, which implies that
(ii) the touch points are assigned alternately to u(x)
p(.x) w i l l have the same zero- and first-order derivative
and l(x), i.e. {x„\i = 1, 3 M„} and {x \i = 2, 4,
values of u(.x) or /(.x). I n a general sense, u(x) and /(.x) can ti
Interpolate p(x) such that Theorem I: Define a A'th order p o l y n o m i a l q(x) subject
1
to the f o l l o w i n g A' + 1 interpolation conditions
p "!*,,-) = u"%x„) r = 0, I , • H, ~ 1
<7(x ) = e(x ) mi mi \= 1. 2, (llfl)
= 1, 3, 5, M, (da)
lr|
q (.\„.) = 0 r = 0, 1 2, ; = 1,2, M (116)
p'"(x„.) = /<"(x„.) 0, I . . . . , f t - 1
:
and
2, 4, 6, . . . W, (661
and (x - x )q(x) H .
<V.x) 1, 2, M (Hi)
MiHx)
p(fl) = A n(b) -. (fir)
then the N e w t o n system (eqn. 9) can be solved for the
Thus, exactly N — M specifications are met by inter-
c touch point increments { A x , . j by
polation. It now remains to adjust {.\„} to make
{p^'\x„)} satisfy the other M specifications. Ax„ = lim <5,-(x) (12)
Definitions
Proof of theorem I: Consider the solution of the N e w t o n
upper error f u n c t i o n e„(x) = p(x) — u(x) system for touch point movements Ax„ by interpolation
of q(x). T w o remarks are necessary f o r the proof.
lower error f u n c t i o n ej(x) = l(x) — p(x)
mid-function m(x) = (l(x) + u(x))/2 Remark 1: Suppose u(x) or l(x) (and so e(x)) are differen-
tiable up t o /i,th order at all the touch points. The func-
search f u n c t i o n s(x) = max [e„(x), efcx)] tion gJ(x), formed by differentiating the interpolated
p o l y n o m i a l p(x) w i t h respect to a touch p o i n t x „ , is
combined error -e (x)u if p(x) > m(x)
e(x) = therefore itself an N t h order p o l y n o m i a l of x. I t can be
function e^x) if p(x) ^ m(x) calculated by interpolation subject to the f o l l o w i n g con-
F r o m assumption (i) above, { f i , — 1} are restricted to be straints:
even, so, i n general, the touch points are now points of 1
?!"(x,;) = 0 r = 0, 1 , .
inflection ( F i g . 2) and s(x) w i l l change sign in the neigh-
b o u r h o o d o f each touch point {x„}. I f the p o l y n o m i a l is ; = 1,2,. , M
manipulated such that p(x) does not cross u(x) or l(x) at
(13a)
these points, p(x) must possess an extra order of tangency
to u(x) or l(x), having then up to the ftth order tangency fl!*- >(xj 1
= e<*>(x .) 11 r = 0, l , . . . , / v (136)
at {x„} required f o r p(x) to be a solution. A t this stage,
max [s(x)] = 0 i n the neighbourhood of {*„•}. Notice that Proof: Eqn. 13a is evident as the fij interpolated deriv-
f r o m assumption (ii) there must be at least one m i n i m u m atives of p(x,j) are fixed w i t h respect to another touch
of e(x), denoted as x ,,,on [ x „ _ , , x , J.Therefore, i f
m i + point x„, i ^ / The p r o o f of eqn. 136 follows. Suppose
that one of the touch points on u(x), x„, changes to x' = ti
dp"'(x)
0 x
Af( mM)
Sx
Ax, M j ti
e(x ) ml
e{xm
lim P r(x„.) - p<"(x„.) (16)
(9) d-0
[^'''(x,, + h)
|r 2
„ ( « - >>f
which updates the variable vector by
A.x„
1
3 /"-"(.*„•) {*.,-}«-{**.} /=1.2,...,Af (27)
This indicates that the interpolation ordinates are simply
<?""-"(*„•)
I, 2. (19) exchanged with the locations of the extrema and reinter-
e?"'{x ) u polated (Fig. 2). It may be expected that ordinates
separated by an excessively large ripple will be brought
Proof: A single row of the Jacobian system (eqn. 9) is together and those separated by an insufficiently large
ripple will be moved apart. When the {x„} are close to
X i/iU^.) A.v„ = e(x ) mJ j - \ , 2 , M (20) the solution, the {x \ are also close to {.x,,}, and the
mi
{*„• [x, + ( x
t ml - x )IHi}
n i'=l,2, M (28)
= gf\x t j ) Ax, ;
( 2 2
>
This can be seen as a generalisation of the Remez method
Eqn. 19 follows. of eqn. 27 in which, instead of moving the abscissa all the
From eqns. life and 6a, x„ is a (/ij — 2)th order zero of way to the extremum, it is moved by a fraction of the
q{x) and (/i, — l)th order zero of e{x). They can be distance dependent on the order of the touch point.
expanded by a Taylor expansion at x„ as
1
2.6 Unilateral method
q(x) = A (x - X / "
q + 0[(x - x„H (23)
In most filter applications, emphasis is given to one of the
e{x) = A (x-x r e li + OUx-x, r*^i
(24) bounding functions. F o r example, in the passband region
of a filter, u(x) is most important as it determines the
From eqns. 19, 23 and 24,
points of maximum transmission. All the high-order
touch points (with / i > 1) could be assigned to u(x) for
(
p(x)
be calculated from the limiting values of the increment
polynomial q{x) and error function e(x) in the proximity
of the touch point, x„ + h. The distance h must be chosen
suitably according to word length and order of touch
6
point. A suggested rule is x 1 0 ~ / N for double preci-
sion arithmetic.
The computational cost of the whole procedure is very
3
small as it involves only repeated interpolation. The 0(n )
step of solving the Newton system has been reduced to
2
an 0 ( n ) interpolation. Each evaluation of the inter-
polated polynomial costs 0(n) multiplications.
A.x„ = <M.x ) mi
2.7 Computer algorithm
The approximation methods proposed in Section 2.6 can
= (Xml - x
M x
m i ) be implemented on a computer by the following algo-
rithm:
Step I Read the boundary functions u(x) and /(x) as
(26) piecewise linear functions on range [a, t>]. Read the
number of touch points and specified orders / j , .
i = 0, 2, 4, n; choose x , to maximise m
2.8.3 Cluster method: In most cases, the boundary func-
{p\x) - u(.v), x _ , sS x $ x
m i I i + ,} tions are only given by values and the derivatives are not
available. Although the derivatives can be calculated by
i = 1, 3. 5, . . . , n — 1; choose x , to maximise m numerical differentiation, this is notoriously inaccurate
for high orders. The polynomial obtained by a Newton
{((x) - p(x), x _ , ^ x s S . v , , . * , }
m i
interpolation may become totally unreliable in the neigh-
Step 5: Compute improved touch point estimates by bourhood of a high-order touch point. A better condi-
one of the methods in Section 2.3. The extended Remez tioned method is to interpolate the polynomial at a
method indicates that cluster of points with first-order tangency to the bound-
ary function. A /i,th-order touch point with ft odd t
X„ = X„ + ((x - X„)/fli)
mi
requires (/i, + l)/2 first-order touch points distributed in
Step 6: Compute convergence estimate for klh iter- the neighbourhood of x „ . In practice, it is found that a
6
ation as spacing of 10" (with normalised passband width of 1)
can be chosen. The error caused by this approximate
method can be controlled and made much smaller than
Z «<*-> the allowed ripple [the separation of u(x) and /(x)].
E
* = ~; —
2.8.4 Damping: The term damping is used for the
X|u(.x„,)-/(x .)| m l
process whereby the step sizes determined by Newton's
method may be reduced to avoid divergence. A form of
Terminate if e < tolerance or e„ > e
t k+ , (divergent) damping can be useful where Newton's method is used to
predict adjustments A x „ , and the touch points would
2.8 Software considerations cross one another or move entirely outside the approx-
2.8.1 Interpolation : The interpolation of the polynomial imating region [a, ft]. In these cases (usually far from
at Step 3 can be done by oscillatory Newton inter- solution), it is found useful to limit the movement of the
polation. This is an extended form of the well known touch points to half the distance in the direction of their
Newton interpolation, whereby a number of derivatives closest neighbour. In this way, no touch point may cross
can be matched to the specified function. or escape the region [a, ft], and yet the direction required
It is particularly important in a filter problem to take to reduce the extrema is observed.
care with the accuracy of construction and representation
of the polynomials within finite word length arithmetic. 2.8.5 Convergency, accuracy and storage: Computa-
Cancellation errors are particularly severe as the touch 2
tion costs are 0{n ) for passband approximation. Stop-
points of a filter function become closely spaced near a band approximation requires solution of a matrix system
band edge. A typical calculation would be 3
with 0 ( n ) efficiency. Convergence is quadratic near solu-
tion, a property of algorithms based on Newton's
/[.x,o.-X ] = / ( X
' o )
'{ - ( X l l )
(29) method. Divergence occurs only in those cases where the
M
The zeros of T(x) are contained in the numerator poly- This corresponds to a special case of the unilateral
nomial. In a filter transfer function they are most effec- method where all touch points are tangential to the lower
tively assigned to the imaginary axis of the s-plane (real boundary which is zero. It remains to compute the
x-plane locations) and placed in the stopband region for attenuation margin 0 (Fig. 4). Two methods can be
maximum attenuation. By making this restriction, the p(x)
*ts0
and compute initial convergence estimate e, . p Fig. 4 Scheme for approximation of an arbitrary lowpass /titer specifi-
Step 4: Solve stopband approximation problem on cation by a rational function
I X i a . *,*,-] using N(x) such that
applied: the heuristic method [26] permits the approx-
u(x) = PV(X)ID(X) imation methods of this section to be used. However, a
variant of the method of Temes and Smith [1] has
l(x) = 0Ux)/D(x) proved to be more stable and more easily extended to
multiband approximations. A Jacobian matrix of the
and compute initial convergence estimate e,. peak positions with respect to the touch points must be
Step 5: Terminate if £_ and £, < tolerance or set up and solved.
k > maxiter.
S^lx^o) • 3r(x„ ) 0
Ax
Owing to the special properties of the numerator and dx.,o
M 0
L P _
Vix^)
3.1.1 Passband design: Any of the methods of Sections
2.3-4 are suitable for the design of the passband function msn -*isnr
D(x). It is found that the bilateral method has very good
global convergence. The unilateral method is then useful Ax, s0
1
to ensure that the function does not exceed maximum
transmission (i.e. 0 < T(.x) < 1) for passive filter realis- (34)
ation. The touch points are all fixed to the upper bound- Ax,„
1
ary. P _
0 96
derivatives
Arbitrary rational function (IIR) filter responses are
Re and Im shown in Figs. 6-7. An llth-order function is fitted
. 3y t . touching the passband boundaries and meeting the stop-
band boundary to within some constant dB error (Fig. 6).
by chain matrix calculations. A bandpass approximation with asymmetric stopband
In the case of certain classical approximations, where and sagging passband is shown in Fig. 7. High-order
the points of maximum or minimum transmission touch points can be placed in the passband to trade off
(pO'cu) = 0 or p(j<o) = 1) are known, the explicit calcu- between group delay, stopband attenuation and passive
lation of p(jw) is not necessary for Orchard's algorithm. sensitivity properties. A series of such passbands is shown
However, in general, Orchard's method requires the for- in Fig. 8, including inverse Chebyshev and elliptic forms
2
mation of p\jo)) by Hurwitz factorisation of \p\ as in as special cases. The effects on stopband attenuation and
classical synthesis, which is an ill-conditioned procedure group delay is shown in Table 1. As high-order touch
[1]. A n extension of Orchard's method is described points are introduced into the response, there is a signifi-
1
below which works with more general forms of \p\ but cant reduction in the peaks of the group delay, which
eliminates any root-finding requirement. follows from the progressive smoothing of the amplitude
2
The value of |/>| and its deriatives at the touch points characteristic. There is also an accompanying improve-
{*„} can be chosen as the objective function for the ment in sensitivity behaviour of the resulting passive filter
Newton scheme. The derivatives of \p\ with respect to 2
realisation, as would be expected from Orchard's cri-
the element values {y } are required for the construction
k
terion [29], since the higher order touch points ensure
of the Jacobian matrix and are given by (let *„• = tuj-) adherence to maximum transmission over a wider inter-
val. This property is inherited by active simulations,
j^(d'\p(u, )\ i
where sensitivity considerations are of major concern.
2
Re (37) There is some attendant cost, with a corresponding loss
dx' dco'
in stopband attenuation. Trade-off between these factors
must be considered in design. It is obviously not possible
for r = 0 , . . . , //, and i = 1, 2 , . . . , M.
to dictate the location, sequence and order of the touch
Notice that here p (the conjugate of p) and dp/Sy are k
points in a general manner, but it is recommended that
obtained from the approximate network with guessed
they are deployed in regions where the filter character-
component values, which can be generated by Orchard's
istic is most critical.
algorithm and then the remaining part of eqn. 37 can be
calculated by a numerical differentiation. Here it is also Practical examples of the use of amplitude-shaping
found to be efficient to use the 'cluster' method men- capabilities are given in Figs. 9-10. In Fig. 9 a length 69
1
tioned in Section 2.8.3. The objective function \p(w,i)\ differentiator is designed with a high-order touch point at
and derivatives are obtainable from a unilateral passband 0 H z for increased D C linearity [4]. A 14th-order wide-
approximation of Section 2.6. This provides a direct link band S C filter with 200 k H z clock frequency shows a
between approximation and ladder design procedures, 1 dB passband droop due to sine (x) weighting The orig-
bypassing the traditional Hurwitz factorisation step. inal 0.1 dB flat equiripple passband behaviour can be
c -60
-80
-100
-120
0.0 0.4 0 8 12 16
u
frequency, Hz x 10
0 0
0.4 06
-1.2
0.00
-005 8. -18
-010
-2.4
-0.1 5
-3 0
-0.25 Fig. 7
a lOth-order asymmetric bandpass filter with sagging passband
b Passband detail
-0.3 0 L
0.0 0.5 1.0 1.5 2.0
f r e q u e n c y H z x 10
0 6
B
Fig. 6 8.-0.8
a I llh-order lowpass filler with arbitrary passband and stopband specifications
b Passband detail 12 15 18 3 12 15 18 < 12 15 18 2
frequency.HzxlO frequency.HzxlO frequency.HzxlO
2 5
nonidealities in a lOth-order S C l e f t - L U D ladder filter
ifl
00 , ^
[30]. In fact, the optimisation process involves a few iter-
ations of circuit analysis and redesign to inverse weighted
specifications. However, it is significantly more efficient
-1.2
T a b l e 1: C o m p a r i s o n of g r o u p delay a n d stopband attenu- -1.5 12 15 18 3
a t i o n o f f i l t e r s in F i g . 8 12 15 18 3 frequency, Hzx10
frequency.HzxlO
Filter Name G r o u p delay Stopband 3 6
number variation, ms rejection. d B -0.0B 0.0.
-03
1 inverse 0.7 22 -06
Chebyshev -09
2 2-10-2 1.35 43 -1.2
3 4-6-4 1.4 47 -1 5
4 2-2-6-2-2 1.9 53 12 15 18 3 12 15 18 3
70
'e so
X
CD correc.ee
~°. 50
c
b
30
20
sincCx) distorted
10
predistorted
00 02 04 06
o frequency. H z
-1 0
c o m p a r i s o n of linearity at D C
frequency, Hz xlO^
9.4 9 8 10 2
optimised
pre-distorted distorted
frequency, Hz
f r e q u e n c y , Hz x 10
F i g . 11 Optimisation of distortion in left-LUD SC filler due to op-amp
and switch nonidealities
-20
4 Group delay approximation by allpass functions
-40
The allpass function in the continuous time s-domain is
m -60 X{-s)
x> T(S): (38)
c" X(s)
o
dX(-s)
-100 2 Re <X(s)
ds
T(W) = • (39)
X(s)X(-s) DM
-120
4.1 New Remez-type algorithm Using Muller's quadratic interpolation method, this typi-
Consider now the problem of fitting the delay function to cally only requires two to three iterations per root [31].
lower and upper boundaries L(a>) and V(<.o) in a minimax Accuracy is preserved in the algorithm by avoiding
sense over a frequency interval w to io (Fig. 12). It is io ki representation of polynomials in coefficient form. Instead,
required that Newton interpolated form is used at step 3 and factored
form at step 4. Both forms are well conditioned on the
N(w) approximation region, allowing high-order functions and
Lieu) + C ^ $ Uivj) + C (40)
X{j(o)X(-ja)) narrow band allpass functions to be designed.
No theoretical proof has been obtained of con-
vergence. However, experience has shown that con-
vergence is good and that five or six cycles will generally
u(oj) suffice. The mechanism of the algorithm is dependent on
the similarity between the group delay function t(co) and
the denominator magnitude function D(w) in eqn. 39. The
numerator is observed to be a smooth function over the
approximating region, which warps the delay function of
I 00 the denominator. Further theoretical investigation is
being undertaken.
Digital allpass functions can be obtained by bilinear
transformation. The delay specifications must be pre-
2
warped by a factor of cos (coT/2) ,
Group delay equalisation can be performed by com-
bining Um) and V(w) with the additive inverse of the
group delay function of the amplitude filter. In this case,
the total group delay of the allpass and amplitude filter
stages will meet the desired specifications.
2 2
\af + (OJ + ft,) af + (cu - ft,) J
06
The numerator function can be calculated from eqns. 43
and 44
OA
N(u>) = D M x T(O>) (44)
interpolation abscissae can be arbitrarily spaced on the it draws on the same numerical methods of interpolation
approximation region (normally equidistantly). and factorisation.
0 1 2n
60
corrected response
0 08
I/I
n
<" 0.06
o
004 20
_j i i— i
90 180 270 360 100 200 300 400 500 600
]
f r e q u e n c y . H z x 10
f r e q u e n c y . Hz xlO
Fig. 14 28lh-order stepped allpass group delay approximation
h i g h - o r d e r touch point
equaliser r e s p o n s e
corrected response
oto
008
" 0.06
004
20 2.5 30 353 40
002 b f r e q u e n c y , Hz x 10
Fig. 16
a Equalisation of lOth-order elliptic filter delay response by 28th-order allpass
function
90 180 270 360
b Passband detail
5
f r e q u e n c y . H z xlO
PAPER 43
Canon sea i design of integrated ladder filters
x
P a p e r 7 7 6 0 G ( E 1 0 ) . received 1 I t h J u n e 1990 CV
r = s' W (4b)
Prof. Sewell is with the D e p a r t m e n t of E l e c t r o n i c s a n d E l e c t r i c a l E n g i n -
where W is the vector of intermediate variables.
eering. U n i v e r s i t y of G l a s g o w , G l a s g o w G 1 2 8 Q Q . U n i t e d K i n g d o m
R . K . H e n d e r s o n is with C S E M , C a s e P o s l a l e 41. C H - 2 0 0 7 N e u c h a t e l .
2.2 Right matrix decomposition
Switzerland
r can also be factorised as
L i P i n g is with the D e p a r t m e n t of E l e c t r i c a l and E l e c t r o n i c s E n g i n e e r -
ing. U n i v e r s i t y of M e l b o u r n e , P a r k v i l l e . V i c t o r i a 3052, A u s t r a l i a r = r,r r (5)
Remark I: The nodal description of a doubly-terminated where A(s) is the determinant of Y and A l n is the appro-
ladder synthesised from the minimum node two-port sec- priate cofactor. It can be shown that
tions in Fig. 1 has the following properties: n -1
n o .
(9)
A(s)
N(s)
(10)
D(s)
A(i) = | C | s " + a„
o—
II
h and the numerator becomes
ii - t
N(s) = s"]~I [ s c ( + l i , - 5~')',• + , . , ] (12fc)
Here c l+ and
l i(for all i) cannot both be zero,
+
Fig. 1 Minimum node two-port sections a (2n — l)th-order polynomial with only odd terms. If
N(s) = s"-' n W i . i - * " ' y i + i.J H36) 4 Canonical ladder simulation by modified
prototype
i= 1
and it is easily shown that N(s) must be an even poly- The numerators of the transfer functions considered in
nomial.
this paper are restricted to be polynomials with purely
Theorem 1 establishes some necessary conditions for a even or odd terms. This means their zeros are restricted
transfer function to have a canonical realisation. It to lie on or have quadrantal symmetry about the imagin-
appears that these conditions are also sufficient for reali- ary axis. Such a constraint is valid for most filter design
sability provided that the transfer function is stable. problems. With this constraint and from theorem 1 the
problem concerned in this paper has been greatly simpli-
It is seen from theorem 1 that the constraint on the
fied : whether a given transfer function can be realised by
parity of the numerator is related to the singularity of the
a canonical standard ladder depends solely on the
matrices C and T. The singularities, however, cannot be
numerator parity.
arbitrarily chosen.
The most common example of a transfer function with
the wrong parity is an even-order 'pure' elliptic function
Theorem 2: A doubly-terminated ladder has a nonzero with an even-order numerator. It is well known that this
response at co = oo only if C is singular and has a kind of function cannot be realised by a doubly-
nonzero response at co = 0 only if T is singular. terminated ladder [13], Traditionally the solution to this
problem is to modify the numerator by moving a pair of
Proof: Let s = j<o. From eqn. 12 it can be seen that when finite zeros to infinity [7], which incurs the penalty of a
CO -> OO,
loss of stopband attenuation (Fig. 2).
1
A(s)-|r|s" + „_ a l X "- (14)
xlO
and from eqn. 9 the numerator is at most to the power of
1
s"" . Therefore if | C | is not zero then eqn. 9 must be
zero. Similar reasoning can be used at co -> 0.
T a b l e 1: U p p e r b o u n d s o n s y s t e m o r d e r for v a r i o u s classes 8
of ladder filter w i t h n n o d e s
Classes Constraint U p p e r b o u n d of
s y s t e m order 100 200 300 400 x l 0 2
f r e q u e n c y , Hz
Lowpass W(O) # 0 T singular 2n -1
Bandpass 2n Fig. 2 Modified and 'pure' eighth-order bandpass elliptic functions
Highpass H ( ° o ) / 0 C singular 2/7 - 1
Bandstop H(0) / 0 W(co)#0 In - 2
both C a n d V singular
A new scheme is introduced here to realise the transfer
function exactly without any sacrifice of filtering quality.
Let H(s) be a transfer function with all its zeros on the
It is easily seen that canonical designs can be achieved imaginary axis or at infinity. If the constraints of
only by bandpass, odd-order lowpass and odd-order theorem 1 are not met, or more precisely if N{s) has the
highpass. F o r other cases, constraints given in theorem 2 wrong parity, some simple manipulation of the transfer
(sC+s' rl
+ G)V = {s + tufs^V (17)
This system can now be expressed in realisable form by
matrix methods [5, 6],
l
cy
r = s~ w-c,-\-j) (186)
Fig. 4 Canonic eighth-order active RC ladder filters
a L e f l - U L D active RC circuit
4.1.2 Right matrix decomposition form: Let r = r,T .
r b Right-ULD active R C circuit
Then the system can be written as
Fig. 3 Eighth-order prototype ladder designed from partitioned trans- Ay=-<t>(B,H' + Gy)-((of+\)(-J) (23a)
fer function
x
W = V(B y + 4w?B[ ' X - y >
r (23b)
4.2 Canonical discrete ladder simulations The prototype in Fig. 3 is simulated by the two canonic
The design of a discrete ladder simulation starts from the eighth-order left-ULD and r i g h t - U L D S C circuits shown
modified prototype system eqn. 17, which after bilinear in Fig. 5. The sensitivity behaviour of the new structures
transformation becomes must be examined, as they are no longer strictly ladder
simulations and would seem to depart from Orchard's
low-sensitivity criterion [1], From the following exam-
r + G)V ples, and many others studied by computer simulation,
T 1 + :"' 2 1- z'
the sensitivity for the new structures has been confirmed
to be much better than their biquad counterparts, and
J + ~ cjfJ (20)
T 1+ . 2 1 very close to traditional ladder simulations.
T4^ ^
c 2 8
left - U L D
9 0 9 5 10 0 11 0
1 <-l | C 5 I C 9 I C13 4 C, 7 I C; ? | L 2 5 |
t r e q u e n c y . Hz x!0 3
5 C i r c u i t design examples
-0 2
The circuits of Fig. 5 are generated by the P A N D D A
filter compiler [14] to realise the 'pure' eighth-order ellip-
tic bandpass function of Fig. 2. The design data are given
in Table 2. A passband sensitivity comparison is shown
in Fig. 6 with a corresponding cascade biquad desi n. A 6
ULD Biquad
- 1 0
Total c a p a c i t a n c e 188.03 193.91
C a p a c i t a n c e spread 57 38 41.54
Average capacitor 5.88 6.46
N u m b e r of c a p a c i t o r s 32 30 200 WO 600 800
N u m b e r of s w i t c h e s 34 36 f r e q u e n c y , Hz KlO'
N u m b e r of o p a m p s 8 8
Fig. 7 Modified and 'pure'eighth-order lowpass elliptic functions
C a p a c i t o r v a l u e s of l e f t - U L D S C ladder filter in Fig 5a
1.36 ations
2 3 5
c = 2.16
c „ = 3.21
c = 2 05
c =
c =
2 5 2.57
1.00 ULD Biquad
J 6 3B 2 3 3 0
c „ = 1.00 C 3 2 = 1.00
Total c a p a c i t a n c e 189.43 107.83
Clock f r e q u e n c y = 200 k H z ; P a s s b a n d ripple < 1 d B ; S t o p b a n d Capacitance spread 75.16 14.08
attenuation > 50dB; L o w e r p a s s b a n d e d g e = 9 k H z ; Upper pass- Average capacitor 5.92 3.59
b a n d e d g e = 11 k H z N u m b e r of c a p a c i t o r s 32 30
N u m b e r of s w i t c h e s 34 36
The passive ladder of Fig. 3 can also be used to realise N u m b e r of o p a m p s 8 8
the partitioned transfer function of an eighth-order
C a p a c i t o r v a l u e s of r i g h t - U L D S C ladder filter in F i g . 5 6
lowpass function in Fig. 7. The circuits of Fig. 5 are once
again employed. T h e design data for a r i g h t - U L D S C c, =13.15 c 2 = 7516 c 3
= 2.21 c. = 7 96 c = 13.99 5
realisation are given in Table 3. A sensitivity comparison 1.04 c 7 = 1.00 = 6.15 c, = 2.19 C.o = 11.61
is shown in Fig. 8. The l e f t - U L D has a serious peak c,, = 1.00 c, a = 6.33 = 2.30 c , = 1.27 c„ = 1.04
4
c, =
2 1.00 c„ = 1.66 c 2 3 = 6.01 c . = 1.45 c = 2.49
2 2 5
design, and would not normally be considered. The right- c = 3.11 c = 1.00 c = 4.27 c = 2.61 c „ = 1.00
3 6 2 7 2B 2 s 3
U L D design has very low sensitivity, well below the c, = 1.00
cascade biquad realisation. In this case, the ladder has a
3
c 31 = 1.00
Clock frequency = 2 0 0 kHz; Passband edge = 10 kHz; Passband
larger total capacitance. ripple < 0.1 d B , S t o p b a n d a t t e n u a t i o n > 7 0 d B .
left - U L D
0
to 12
+ s
J (24)
nghl -ULD The output is now i„, which differs from [;„+, by only
a constant g . The rank of matrix T will be n — 1 since
L
made in order to provide an efficient matrix nodal row equations can be arranged into the following form,
description. F o r arbitrary prototypes, more general providing a canonical solution:
hybrid matrix descriptions can be adopted to minimise
the size of the matrix systems and their resulting simu-
lated circuits. The drawback of the hybrid method is that £.-„_, + c _ m 2 0
there is no unified rule. The exact design method depends
on the individual prototype structures and the selection
of the internal variables, which can be seen from the fol-
lowing examples. + s , L l,
m 0
An even-order lowpass circuit shown in Fig. 9a has
n + 1 nodes but the filter order is 2n. The rank of T is n
as there are n inductors in the circuit, making the total
number of opamps required 2n + 1; so even a leapfrog 0 I J (25)
design cannot directly provide a canonical circuit. -1
L2
C |
J9,n T
1-2
era
m- 3 m-2
Vn-l Lm n.
/-YV\
l
C 3 m-3 m-2
J c. m-U L m- m-
\
8
9
UL
A n objection to the use of active ladder simulation filters 1 S E D R A . A.S.. and BRACKET"!'. P.O.: Filler theory and design:
is the difficulty of guaranteeing a canonic filter circuit. By active and passive' (Pitman. London. 19791
2 G R E G O R I A N . R., M A R T I N . K.W.. and T E M E S . G.C.: 'Swiiched-
standard methods, a passive ladder realisation may not
capacitor circuit design'. Proc. IEEE. 1983. 71. pp. 941-965
exist or the simulation of the ladder w o u l d yield non- 3 L E E , M.S.. and C H A N G . C . : 'Switched-caparitor filters using the
canonic circuits. This paper has presented a set of condi- LDI and bilinear transformations'. IEEE Trans.. 1983. CAS-30. | I 2 | .
tions on the transfer function and prototype structure pp. 873-887
whereby a canonic simulation is obtainable. For the A M A R T I N . K.. and S E D R A . A S . : Exact design of switched-
capacitor bandpass filters using coupled biquad structures'. IEEE
exceptions, a new design method and new canonic ladder Trans.. 1980. CAS-27. |6|. pp. 469-474
structures are proposed. These circuits are applicable to :' LI P I N G : 'Theory and methodology for integrated ladder tiller
all active technologies and are stray insensitive in design". PhD Thesis. University of Glasgow. 1990
switched-capacitor implementation. Despite departing 6 LI P I N G . H E N D E R S O N . R.K.. and S E W E L L . J.l : Matrix
f r o m the strict conditions for low-sensitivity passive methods for switched capacitor filter design'. Proceedings I E E E
ISCAS. Helsinki. June 1988. pp. 1044-1048
ladder simulation, the circuits are shown to have very
7 D A N I E L S . R.W.: 'Approximation methods for electronic filler
good sensitivity properties. A l t h o u g h competitive w i t h design' (McGraw-Hill, New York, 1974)
biquad cascade realisations for certain types of transfer 8 L I P I N G , and S E W E L L , J.I.: 'Filter realisation by passive ladder
function, there is some cost in component spread for simulation', IEE Proc. G, Circuits. Devices & Svst.. 1988, 135, (4), pp.
others. The design is formulated i n a very regular manner 167-176
9 L I PING, H E N D E R S O N . R.K , and S E W E L L , J.I.: A method-
in terms of m a t r i x equations, m a k i n g i t highly suited to ology for integrated ladder filter design', submitted for publication
computer implementation. The problem of o b t a i n i n g a 10 LI P I N G , and S E W E L L , J.I.: 'The L U D approach to switched-
canonic simulation of an arbitrary ladder sturcture is capacitor filter design', IEEE Trans.. 1987. CAS>-34, (12), pp. 1611-
demonstrated t o be one of properly choosing the system 1614
variables t o ensure a sparse matrix description. Some 11 L I P I N G , and S E W E L L , J.I.: 'High performance circuit structures
and symmetric matrix systems', I EE Proc. G. Circuits. Devices and
example techniques are illustrated for symmetric Syst.. 1989,136, (6), pp. 327-336
bandpass elliptic ladder filters. 12 Pearl, M.: 'Matrix theory and finite mathematics' (McGraw-Hill.
New York, 1973)
13 H E R R E R O . J . L . . and W I L L O N E R . G . : 'Synthesis of fillers'
8 Acknowledgment (Prentice-Hall, Englewood Cliffs. NJ. 1966)
14 H E N D E R S O N . R.K., L I P I N G , and S E W E L L . J.l.: A design
The authors w o u l d like to acknowledge the support of program for digital and analogue fillers: PANDDA'. Proceedings
the U K Science and Engineering Research Council. E C C T D '89, Brighton, September 1989, pp. 289-293
PAPER 44
19 2
Abstract—This paper is concerned with the design uf high silicon nitride, was used, and novel switching circuits with
frequency switched-capacitor filters implemented using GaAs the introduction of diodes to reduce the signal depen-
M E S F E T technology. The design techniques developed are illus-
dence of clock feedthrough were explored. Excellent filter
trated by means of a 2nd-order switched-capacitor bandpass
filter designed to operate at a switching frequency of 500 MHz. results for switching frequencies up to 100 M H z were
The design incorporates a recently proposed operational ampli- obtained.
fier architecture with a simulated gain of 60 dB and a recently T h i s paper is concerned with the design, optimization,
proposed switch driver circuit Amplifiers are characterized in and testing of a G a A s 2nd-order bandpass filter using a
terms of settling time performance and a general procedure is
presented for optimization of the switched-capacitor circuit to recently proposed high-gain operational amplifier [8],
majdmize switching frequency. Measurements on a fabricated which can have a low frequency gain of over 60 d B , and a
chip confirm the design of the filter and its components. recently proposed switch control circuit [9]. T h e amplifier
is characterized in terms of its minimum settling time and
the S C circuit is optimized for maximum possible clock
I. INTRODUCTION frequency. T h e design procedure is evaluated by measure-
0O98-4O94/9l/080O-O825$01.0O ©1991 I E E E
19 3
GND GND MI R
MO-/-6V
M 0 9
Q-° 'P
-8V
(b)
Fig. 3. Switch control circuits for GaAs technology, (a) Basic architec-
ture, (bl Circuit adopted (Gatewidths: Ml—>0 fim: M2—IS um: M3.
M4—24 m ) .M
2O0r 70
— gan.dB
--phase, deg
2 0 L
' ° i k 10k KX* 1M 10M10CM16 DG100G
frequency, Hz
TABLE I
K E Y PARAMETERS FOR J F E T M O D E L O F G A A S MESFET
VTtJ -1 V
BETA 0.067E-3 mAV" 2
constant T is given by
T = [Ci + CL + C i C L / C f ] / g m . (2)
5.3 CL e f f
(6)
f.-rinO/d) (1) Kg
where T is a time constant and d is the settling value T h e t e r m CL e f f / x is referred to as the amplifier " l o a d -
tolerance. M a t s u i et al. [15] have shown that the time ing f a c t o r . "
828 I E E K TRANSAC1"10NS O N i. I I U '. I ! S A N D S Y S T E M S . V O L . Vi. N O . s. A I X . L ' S T 1W1
V. A M P L I F I E R C I R C U I T AND CHARACTERIZATION Cf
G a A s M E S F E T ' s have a number of disadvantages for
the design of high gain operational amplifiers, including
the lack of a suitable P-channel device, the general lack
OA
of enhancement mode devices (depletion mode devices WOT
are easier to manufacture), and a low value f o r the 1„ AnpUfor
ifder Its!
open-circuit voltage gain (transconductance/output con- N Luriwit pulse
ductance) of the device, typically only 20. The require-
ment f o r high b a n d w i d t h and voltage gains of the order of
60 d B has led to the development of a range o f circuit Fig. 7. Test setup for amplifier characterization.
versus loading factor, w h i c h forms the starting point f o r states may t h e n be determined f r o m (11) for these re-
o p t i m i z a t i o n o f the SC f i l t e r circuit. maining states. The expressions obtained f o r x and the
A C L , are given in Table I I I . I t can be seen that i f / • = f k
Switching states may be divided into critical and non- x or A C L = A C L w i l l become indeterminate. Thus it is
y t
critical switching states. C r i t i c a l switching states are those important to choose carefully the two switching states f o r
in w h i c h the a m p l i f i e r is receiving charge at its input or which A C L is the same.
conveying charge to a subsequent integrator. It is desir- T h e addition o f a m p l i f i e r load capacitance w i l l not
able that i n a critical switching state, the operating point increase m i n i m u m settling time because the a m p l i f i e r
on the a m p l i f i e r characterization curve should be close to gatewidth factors are adjusted to remain at prescribed
that f o r m i n i m u m settling t i m e . Other switching states are operating points on a m p l i f i e r characterization curves. T h i s
denoted noncritical and usually include the guard I N T w i l l , however, increase chip area and power consumption
state. Effective load capacitance CL eff, as defined in (4), and therefore, the values o f added load capacitance must
is frequently close to zero i n noncritical switching states be carefully optimized. The application o f the expressions
since Ci and CL are usually very small i n practice. I t can in Table I I I to optimize an SC circuit w i l l be illustrated i n
be seen f r o m the characterization curves i n Fig. 8 that this Section V I I I .
leads to possible unstable oscillation. W e now present a
technique f o r avoiding such instability and m a i n t a i n i n g
control of a m p l i f i e r settling behavior by specifying the VII. E F F E C T OF AMPLIFIER PARASITIC
a m p l i f i e r operating point in all switching states. This CAPACITANCES
control is achieved by the i n t r o d u c t i o n of grounded capac- A m p l i f i e r i n p u t and o u t p u t capacitances can each be
itance at the a m p l i f i e r o u t p u t terminals. considered as the sum o f two contributions. One, w h i c h is
For each switching state, denoted ;', let the values o f primarily due to interconnections, is independent o f am-
C L e f f and Ci / Cf before a d d i t i o n o f load capacitance be p l i f i e r gatewidth factor and can be included i n the capaci-
denoted CZ, e f f , and a,, respectively, and let the value o f tances Ci and C L o f the embedding circuit in F i g . 4, and
added load capacitance be A C L / . F r o m (4), the aug- hence allowed f o r i n the circuit o p t i m i z a t i o n . T h e other
mented effective load capacitance f o r the ;'th switching c o n t r i b u t i o n derives f r o m M E S F E T model capacitances,
state may be w r i t t e n which are proportional to gatewidth factor x. A m p l i f i e r
input capacitance is related to M E S F E T gate-source ca-
CL eff, = CL e f f ; + ACL,(1 + a,). (9)
pacitance, w h i c h is typically much larger than the gate-
D e n o t i n g the loading factor required i n this switching drain capacitance, which determines a m p l i f i e r o u t p u t ca-
state as / , and the a m p l i f i e r gate w i d t h factor x, we may pacitance. W e therefore now extend the above circuit
write o p t i m i z a t i o n procedure to allow f o r a m p l i f i e r i n p u t capac-
CL eff) = x f . (10) itance, which scales w i t h the gatewidth factor o f the
t
T A B L E IV OUTPUT
EXPRESSIONS FOR C I R C U I T OPTIMIZATION INCLUDING A M P L I F I E R
INPUT C A P A C I T A N C E
2
Ax + Bx + C = 0
••••here A = f,b k - f b,
k
B = Ml + a ) - + a,)+ b,CLeff
k - b CL k k eff,
C - CL e f f ( l + a,)-CL
t eff/l + a) k
x f j -CL eff,
ACL, = ACL, =
1 + a, + xb.
xf.-CLeff,
ACL, = INPUT S9
1 + a, + xfc,
• — B
O7O206P OT7M8P
where b denotes k c / C f . F r o m (10) and (12), we have
t t
ov ov
j/ f = C L e f f + A C L , ' 1 + <!,. +**>,•)
f (13)
Fig. 9. 2nd-order S C filter for implementation.
which may be w r i t t e n f o r each switching state. Constrain-
ing, as before, the added load capacitances to be equal i n
which the a m p l i f i e r behaves like an ideal transconduc-
two switching states, j and k, we again solve f o r A C L , =
tance, but the computer simulations to be given i n Section
A C L and x. I n this case, x is given by the solution to a
t
I X c o n f i r m that reasonably accurate prediction o f settling
quadratic equation, which is given i n Table I V . Having
time is obtained using the transconductance approxima-
obtained the value o f x, the added load capacitance in
tion f o r these operating points. T h e settling curves f o r the
the and k switching states A C L , = A C L is given by the t
three operating points to be used are shown i n Fig. 10(b),
expression given i n Table I V . Finally, (13) f o r the remain-
illustrating the d i f f e r e n t degrees o f damping. T h e settling
ing switching states gives the added load capacitances
times f o r a 0.5% settling value tolerance are given in
A C L , f o r these states according to the expression i n
Table V . W e now apply the optimization method o f
Table I V . 1
Section V I .
V I I I . O P T I M I Z A T I O N O F A N SC FILTER
A. General Considerations
B. Amplifier I
W e now describe the o p t i m i z a t i o n o f the G a A s 2nd-
The values o f Ci, CL, and Cf (as d e f i n e d i n F i g . 4) f o r
order bandpass f i l t e r in Fig. 1. T h e circuit is shown again
amplifier 1 in the three switching states are given i n Table
in Fig. 9 w i t h some modifications; capacitor C I 1 has been
V I . The values o f C L e f f (as defined i n (4)) and a = Ci /Cf
replaced by an equivalent capacitor "T" network i n order
are also given. I n the case o f a m p l i f i e r 1, the " E " phase is
to reduce capacitor spread; we have added a b u f f e r am-
the critical phase, and its loading factor f is to corre- e
p l i f i e r f o r driving o f f - c h i p loads; capacitors Co, Cb, and
spond to the o p t i m u m operating point on the characteri-
Cc w i l l be discussed later. T h e circuit uses the high-gain
zation curve. W e assume that the operating points i n the
p u s h - p u l l a m p l i f i e r o f Fig. 5. T h e switch c o n t r o l circuit to
noncritical " O " and " I N T ' switching states are identical
be used is that shown i n Fig. 3(b).
and we denote the corresponding loading factor f . T h i s a
Since the a m p l i f i e r in Fig. 5 has the property that the makes all parameters i n the " O " and " I N T " switching
quiescent input voltage is equal to the negative power states identical (see T a b l e V I ) and therefore, they may be
supply voltage, the ground connections f o r switches SI treated f o r optimization as a single switching state. A p p l i -
and S3 i n the f i l t e r o f Fig. 9 have to be connected to the cation of the equations i n Table I I I ( w i t h j = e and k = o)
negative power supply. The settling time characterization to the data i n Table V I yields added load capacitor value
curve f o r the a m p l i f i e r has already been shown i n Fig. 8 and a m p l i f i e r gatewidth factor given by
and it is shown again in F i g . 10(a) w i t h the a d d i t i o n o f
chosen operating points (note that the horizontal scale is
/.CL eff.
now loading factor as mentioned in Section V ) . Since ACL, = ACL„ (14)
points to the l e f t of the m i n i m u m settling time point / , - / „ ( ! + «.)
correspond to underdamped behavior it was considered
safer, i n view of process tolerances, to select the o p t i m u m CL eff.
operating point f o r the critical switch phases to the right (15)
/ . - / . ( ! + ««)'
of the m i n i m u m corresponding to a loading factor o f
0.012 pF/nm, indicated " O p t " i n F i g . 10(a). For noncriti-
cal switching states, we shall adopt operating points corre- Table V I I shows values o f A C L and x f o r a range o f " O "
sponding to loading factors o f 0.004 and 0.006 p F per j i m ,
indicated " P I " and " P 2 " in F i g . 10(a). Points " P I " and 'The optimization technique (including scaled amplifier input capaci-
tances) had not been developed at the time the integrated filter was
" P 2 " are strictly outside the range o f loading factor f o r
designed.
H A I G H el ai: D E S I G N O P T I M I Z A T I O N 831
Ve sefUrg
ve settling
T A B L E VI
S C F I L T E R L O A D I N G FOR A M P L I F I E R 1 IN F I G . 9 TABLE VIII
SC F I L T E R LOADING FOR A M P L I F I E R 2 IN F I G . 9
Phase Ci (pF) C/(pF) CL (pF) O - e f f (pF) a( = Ci/C[)
Phase a (pF) Cf (pF) CL (pF) CZeff(pF) a( = C//C7)
E 0.2545 0.7131 0.2545 0.5998 0.3569
O 0 0.7021 0 0 0 E 0 0.7021 0.1767 0.1767 0
Int 0 0.7021 0 0 0 O 0.1767 0.7021 0 0.1767 0.2517
Int 0 0.7021 0 0 0
(16)
switching states are given in Table V I I I . In this case, /,-/.•.(!+ 0
phases " E " and " O " are both critical phases, and the
CL e f f „
required loading factor should be the same and corre- (17)
spond to the o p t i m u m operating point (we denote the
19 9
T A B L E IX
S I M U L A T E D AND D E S I G N F I L T E R S E T T L I N G T I M E S (ps)
E-phase E-phase
Amp Sign Int-phase O-phase Isolated Coupled
m 0
I
I
Fig. 12. Macromodels for S C filter, (a) Amplifier macromodel. (b) Switch macromodel.
Fig. 13. Simulated amplitude response of integrated filter with ideal curve, (a) 250-MHz switching frequency, (b) 500-MHz
switching frequency.
H Q * ] ® 0
Fig. 14. Layout plot for GaAs S C filter, total size 3.3 mm x 2.6 mm.
allowed for in the above design procedure. F o r the results an appropriate clock signal to the digital input, the buffer
denoted "Isolated" in Table I X , the interaction was bro- acts as a sample and hold circuit where the sampling
ken by replacing the other amplifier by an ideal voltage instant may be freely chosen. Alternatively, with a dc
source. T h e results denoted "Coupled" include the inter- digital input, the circuit acts as a nonsampling buffer.
action. It can be seen that the interaction between ampli- S P I C E simulation of the circuit in switched and non-
fiers increases settling times by about 3 0 % . F o r the " O , " switched modes showed that in both cases there is negligi-
"Int," and noncoupled " E " cases, there is reasonable ble performance deterioration over the required fre-
agreement between design settling times and those ob- quency band.
tained by simulating the filter.
CvtnloBI
z
12
» 12
(a)
OanldBI
-12
20 22
b)
Fig. 16. Measured expanded passband response, (a) 250 MHz switching frequency, (b) 500 MHz switching frequency.
-U -
_.
o a
9
m
FmiHlcy IHHl)
analog applications. Comparison o f the computed and are also applicable to other technologies, including
measured response curves indicates a need f o r more CMOS.
realistic computer simulation o f the integrated system,
perhaps along the lines proposed in [24]. I n spite o f the
results achieved, the operational a m p l i f i e r used i n the ACKNOWLEDGMENT
design is significantly less fast than more recent designs T h e authors g r a t e f u l l y acknowledge the invaluable help
[26] and advanced designs i m p l e m e n t e d using state-of- o f numerous colleagues, i n particular, Heloisa Facanha o f
the-art technology [27]. Thus f u t u r e o p p o r t u n i t i e s are University College L o n d o n , w h o wrote the GaAs layout
opened up f o r SC systems operating w i t h gigahertz clock technology f i i e ; B r i a n W i s h a r t o f Glasgow University, w h o
rates. I t is thought that the techniques f o r a m p l i f i e r carried out S C N A P simulations; Steve Newett and L i z a
characterization and o p t i m i z a t i o n o f an SC circuit f o r W o n g o f S E R C R u t h e r f o r d A p p l e t o n Laboratory, w h o
m a x i m u m switching frequency that have been presented carried out additional design rule checking; Steve H a r t ,
2 0 3
I E E E T R A N S A C T I O N S O N C I R C U I T S A N D S Y S T E M S . V O L . JR. N O . K. A U G U S T l<NI
David Goldsmith, and Shiva Carver of Interactive Solu- [24] D. G . Haigh and C . Toumazou. "On computer simulation of
integrated switched capacitor circuits." in Proc. 1989 European
tions, L t d . , w h o helped w i t h M I N N I E ; and Graham
Conf. Circuit Theory and Design. Sept. 1989.
Johnson o f Imperial College, w h o provided C A D support. [25] J . I. Sewell. Li Ping, and R. K. Henderson. "Matrix methods for
switched capacitor filter design." in Proc. 1988 IEEE Int. Symp.
Circuits mid Systems. June 1988.
REFERENCES
[26] D. G . Haigh and C . Toumazou. "Fast settling high gain GaAs
operational amplifiers for switched capacitor applications," Elec-
[1] S. Masuda et al., "CMOS sampled differential push-pult cascode
tronic Letters, vol. 25. pp. 734-736, May 1989.
operational amplifier," in Proc. 1984 IEEE ISCAS, pp. 1211-1214.
[27] L . E . Larson el al., "A 10 G H z operational amplifier in GaAs
May 1988.
M E S F E T technology," Digest 1989 IEEE Int. Solid State Circuits
(2] D. G . Haigh and J . T . Taylor, "High frequency switched capacitor
Conf.. pp. 72-73. Feb. 1989.
filters for C M O S technology," in Proc 1988 IEEE ISCAS. June
1988.
[3l P. E . Allen and C . M. Breevoort, "An analogue circuit perspective
of GaAs technology," in Proc. 1987 IEEE ISCAS, May 1987.
|4) S. J . Harrold, I. A. W. Vance, and D. G . Haigh, "Second-order
switched capacitor bandpass filter implemented in GaAs." Elec-
tronics Letters, vol. 21, pp. 494-496, May 1985. ,
David G . Haigh (M'69-M'70-SM 90) received
[5] GaAs Integrated Circuits, J . Mun, E d . Oxford: BSP Books, 1988.
the B.Sc. degree in electrical engineering from
[6] L . E . Larson, K. W. Martin, and G . C . Temes, "GaAs switched Bristol University, Bristol, England, in 1968, and
capacitor circuits for high speed signal processing," IEEE J. Solid the Ph.D. degree from the University of London
State Circuits, vol. SC-22, pp. 971-981, Dec. 1987. in 1976.
{7} W. S. Lee, "Sensitivity of GaAs integrated circuit building blocks to From 1968 until 1972 he worked in the Cir-
the effect of backgating," Electronics Letters, vol. 23, pp. 587-589, cuit Theory Group at the G E C ( U K ) Hirst Re-
May 1987. search Centre on the design of microelectronic
[8] C . Toumazou and D. G Haigh. "Design of a high-gain, single-stage active filters. In 1972, he moved to Imperial
operational amplifier for G a A s switched-capacitor filters." Elec- College (London), where he became a research
tronics Letters, vol. 23, 2nd pp. 752-754 July 1987. fellow in 1980, becoming responsible for a group
[9] S. J . Harrold, "Switch driver circuit suitable for high order switched working on integrated high-precision filters. He left Imperial College in
capacitor filters implemented in GaAs," Electronics Letters, vol. 24. 1986, becoming an independent consultant, principally to British Tele-
pp. 982-984, July 1988. com. In 1987, he joined the staff of the Electronic and Electrical
[10] D. G . Haigh, C . Toumazou, S. J. Harrold. J . 1. Sewell. and Engineering Department of University College London, where he is
K. Steptoe, "Design and optimisation of a GaAs switched capacitor specifically interested in the field of analog integrated circuit design with
filter," in Proc. IEEE 1989 Int. Symp. Circuits and Systems, June applications to signal processing and particularly in high-frequency GaAs
1989. circuits.
(11] C . Toumazou, D. G . Haigh, S. J . Harrold, K. Sleptoe, J . 1. Sewell, Dr. Haigh is secretary of the Circuits and Systems Chapter of the U K
and R. Bayruns, "Design and testing of a GaAs switched capacitor and Republic of Ireland Section of the I E E E . He is also editor of
filter," in Proc. IEEE 1990 Int. Symp. Circuits and Systems, May Proceedings of the Institute of Electrical Engineers, Part G , covering
1990. circuits, systems, and devices, and he is editor of the I E E Books Series
112] D. G . Haigh, C . Toumazou, and A. K. Bens, GaAs Technology and on Circuits and Systems. He was a co-recipient (with R. Jeffers) of the
its Impact on Circuits and Systems, D. G . Haigh and J . K. A. 1973 J . J . Thomson Premium of the 1 E E and also a co-recipient (with W.
Everard, Eds. London: Peter Peregrinus, 1989, ch. 10. Saraga and R. G . Barker) of the 1979 Darlington Award of the I E E E
|13] R. Gregorian and G . C . Temes, Analogue MOS Integrated Circuits. Circuits and Systems Society.
New York: Wiley, 1986.
(14] H. C . Yang, J . R. Ireland, and D. J . Allstot, "Improved operational
amplifier compensation techniques for high frequency switched
capacitor circuits," Proc. 30th Midwest Symp. Circuits and Systems,
pp. 952-955, Aug. 1988.
Chris Toumazou (M'86) received the B.Sc. de-
[15] K. Matsui et at., "2-micron C M O S switched capacitor circuits for
gree in engineering and the Ph.D. degree in
analog video L S I , " in Proc. 1982 IEEE ISCAS pp. 241-244, May
electrical engineering from Oxford Polytechnic,
1982. Oxford, England in 1983 and 1986, respectively.
[16] C . Toumazou and D. G . Haigh. "Design of GaAs operational
From 1986 to 1987 he was a postdoctoral
amplifiers for sampled data applications," IEEE Trans. Circuits
research fellow in the Information Engineering
Syst., vol. 37, pp. 922-935, July 1990. Section of the Department of Electrical Engi-
[17] , "Design and applications of GaAs current mirror circuits," neering, Imperial College, London, England,
Proc. Inst. Elec. Eng.. Pan G . vol. 137. pp. 101-108, Apr. 1990. where his research was high-frequency precision
(18] R. Spence el al., "Minnie—The way ahead for analogue C A D . " analog filter design in GaAs technology. In 1987
Silicon Design, p. 6, Mar. 1986. he was appointed lecturer in the Department of
[19] C. Toumazou and D. G . Haigh, "Level-shifting differential to Electrical Engineering at Imperial College, where both his lecturing and
single-ended convenor circuits for GaAs M E S F E T implementa- researcii activity are in the area of analog integrated circuit design. His
tion." Electronics Letters, vol. 23, pp. 1053-1055, Sept. 1987. research interests include current-mode analog signal processing, high-
[20] "Some designs and a characterisation method for GaAs frequency analog integrated circuit design in bipolar and GaAs technol-
operational amplifiers for switched capacitor applications." Elec- ogy, and artificial intelligence applied to analog circuit design.
tronics Letters, vol. 24, pp. 1170-1172. Sept. 1988. Dr. Toumazou is European Laison Officer for the Analog Signal
[21] C . A. Makris and C . Toumazou. "Operational amplifier modelling Processing Committee of the I E E E Circuits and Systems Society and is a
for high-speed sampled-data applications. Proc. Inst. Elec. Eng.. Member of the Circuits and Systems Chapter of the U K and Republic of
Part G , vol. 137, pp. 333-339. Oct 1990. Ireland Section of the I E E E . He is also a member of the I E E Profes-
[22] A. K. Betts, D. G . Haigh. K. Steptoe, and J . T. Taylor. "Design sional Group E10 Committee on Circuits and Systems. He is co-editor of
techniques for switched capacitor filters using GaAs MSI technol- Analog IC Design: The Current-mode Approach, and has co-organized a
ogy," in Proc. IEEE 1990 Int. Svmp. Circuits and Systems. May number of special sessions and a workshop on the topic of current-mode
1990. analog design at various I E E E Symposia. He has also contributed
[23] A. K. Betts, J . T. Taylor, and D. G . Haigh. "An improved charac- several chapters to various books in the field of analog I C design. He
terisation technique for amplifiers used in high-speed switched has authored or co-authored some 80 technical publications in the area
capacitor circuits." in Proc. 1989 European Conf. Circuit Theory of analog circuit design and currently runs a short course to industry in
and Design. Sept. 1989. this area.
2 0 4
H A I G H ei nl : D E S I G N OPTIMIZATION
Stephen Harrold received the B.Sc degree in From 1%6 to I96S he was a senior research
applied physics and electronics from Durham associate at the University of Newcastle upon
University. England, in 1978. and ihe Ph.D. Tyne. From 1%8 to 1985 he held faculiy posi-
degree in engineering science from Exeier Uni- lions at ihe University of Hull, England. In 1985
versity. England, in 1982. 4 J'JB^^Km n e w a s
appointed as Professor of Electronic
From 1982 (o 1985 he was a senior research d "r^^^^M Systems in ihe University of Glasgow. Scotland.
engineer with S T C Technology Lid. Harlow. M ^B^B^E a n
5'
d n n c D e c a
Dean of Engineeringm e
England. From 1985 lo 1987 he was a project IB JJ™*^^C, N' research inieresis are concerned with the
officer in the Department of Electrical Engi- ^•T development of C A D design tools for front-end
neering & Eleclronics at UM1ST. Manchester. M ^pkV signal processing including Si and GaAs imple-
England. Since that date he has been a Lecturer mentations, integrated fillers, and signal con-
in the Department of Electrical Engineering & Electronics. UMIST. His verters.
current research interests include high-speed IC design, with particular Dr. Sewell has served on the U K R 1 Chapter Committee of the C A S
emphasis on III-V technologies, and automated testing of analog and Society since 1975. in ihe positions of Vice Chairman (1978-1980) and
mixed-mode IC's. Chairman (1980-83). He was also a member of Ihe U K R I Section
Committee from 1980 to 1983.
Dr. Harrold has published work and filed patents in various topic
areas, and he received the I E R E Marconi Premium in 1987.
PAPER 45
2 06
Abstract—The design procedures for integrated ladder niter Some constraints will first be stated to define a class of
design by operational simulation are regularized in terms of matrix equations that are directly realizable. These equa-
matrix operations. This provides a systematic framework for the
derivation of standard circuits, such as leapfrog and coupled- tions are linear in s~' so that op-amp circuits can be used
biquad, as well as several novel structures. Each circuit is now to perform additions, multiplications, and integrations.
seen to belong to a certain family, dependent on the type of Techniques are then introduced to render the matrix
matrix factorization employed. A detailed comparison of the description of the ladder prototype realizable, by decom-
properties of the different structures is undertaken for a range posing it into several linear subequations. A large family
of filter specifications, demonstrating that ladder filters offer of circuit structures is revealed depending on the choice
good solutions to several hitherto difficult filtering problems.
The methods are applicable to all active filter technologies of matrix decomposition, including the existing leapfrog,
including active- RC and switched capacitor (SO. coupled biquad, and L U D ones as specific cases. A l l
circuits are insensitive to parasitic capacitance in SC
implementation.
I. INTRODUCTION
LDI-transformed SC ladders are discussed initially [3]
Manuscript received July 15, 1990; revised March 27, 1991. This work II. D I R E C T L Y R E A L I Z A B L E M A T R I X SYSTEMS
was supported in part by the Science and Engineering Research Coun-
cil, U K . This paper was recommended by Associate Editor T. T. Vu. The basic building blocks for active-/?C and SC net-
L. Ping was with the Department of Electronics and Electrical Engi- works are shown in Fig. 1, realizing the elementary trans-
neering. University of Glasgow. He is now with the Department of fer functions of feedthrough, inverting, and noninverting
Electrical and Electronics Engineering, University of Melbourne, Mel-
bourne, Parkville, Victoria 3052, Australia. integration.
R. K. Henderson was with the Department of Electronics and Electri-
cal Engineering, University of Glasgow. He is now with Centre Suisse
d'Electronique et de Microtechnique SA, Neuchatel. Switzerland.
Directly Realizable Matrix Systems
J. I. Sewell is with the Department of Electronics and Electrical
Engineering, University of Glasgow, G I 2 8QQ. Scotland. U K .
Matrix methods are known to be an efficient means of
I E E E Log Number 9100871. representing large interconnected networks. The inverse
O098-4094/91/080O-0853$0I.OO C1991 I E E E
854 I E E E TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. 38, NO. 8. AUGUST 1991
(a)
and nonsingular with all the diagonal elements
nonzero.
With the above assumptions a directly realizable matrix
equation will have the following form:
°\ 1°
(2)
X1° k
where [X ) are vectors of variables, [J ] are input vectors,
k k
C„„D„ + C.
(1) (sC + s-'T + G)V=J. (3)
1- Z"
To ensure that all the entries in C, I \ and G are positive,
This equation can be directly represented by the SFG in
we introduce alternating signs in V, i.e., let V -
Fig. 2(a). Let the variables {u,} and {w,} be the voltages of
[ r „ - L ' , i : - v , • ].
op-amp outputs, then the SFG can be replaced by the SC 2 3 > 4
A = (2/T)C,
B = (T/2)T, Fig. 3. 6th-order bandpass ladder prototype.
(6b)
Multiply (5a) through by z ~ 1 / 2
to get D. Right Matrix Decomposition
1 B can also be factorized as
(6c)
B = B,B f (8a)
A = (2/T)C,
B = (r/2)r, The following pair of equations to (6c):
i / 2
f=z- J. (6d)
Since the transfer functions from J and J' to the output AV= - < D [ B , H ' + GV + ( - / ) ] (8b)
differ only by a delay of a half period, we will not
distinguish between them in the following discussion. (8c)
As system (6) can be made identical to (3) by replacing
\jr = <j> = j - ' there is virtually no difference between
f From the rules given in Section I I , (7) and (8) can be
active-/?C and L D I transformed ladder design, apart from realized by SC circuits, provided that the relevant matri-
replacing a pair of LDI integrators by a pair of inverting ces are obtained by certain decompositions. The one-to-
and noninverting continuous domain integrators. For this one correspondence between the circuit elements and the
reason, only L D I SC ladder design will be detailed, as- matrix entries indicates that the efficiency of the SC
suming that all the techniques can be used directly for implementation in terms of numbers of capacitors is re-
active-/?C design. lated to the sparsity of the system matrices. Consequently,
a good simulation of a prototype by matrix methods will
B. System Linearization by Matrix Decompositions attempt to maintain the sparsity property of (3).
\A,W=(-<t>B-G)V-{-J) (7b) the current vector of the inductance branches. This con-
firms that topological decomposition of B, derived as
\A V = VW
r (7c) shown at the bottom of Table I , yields the same structures
where W is the vector of intermediate variables. as those by a conventional leapfrog method. In general, if
2 09
856 I E E E TRANSACTIONS ON CIRCUITS AND SYSTEMS. VOL 38. NO. 8. AUGUST 1991
y-4_Y Y-\i—\' I h M
S—II £—I'
A
41 li-
ft)
Fig. 4. Various S C niter realizations, (a) Left-LUD type S C circuit, (b) Left-direct ( I A ) type S C circuit, (c) Right-LUD
type SC circuit, (d) Right-direct (BI) type SC circuit.
topological decomposition is applied to the left-hand ma- IV. O T H E R SYSTEM LINEARIZATION APPROACHES
trix A, it cannot be assured that the resulting A, and A r
A. Inverse Matrix Approaches
will be square. Consequently, the resulting system (6) may
violate Assumption (iii) in Section I I . I f A, and A , r If a left-direct decomposition is used, i.e., A, = I, A-r
C, Jo C, J° C,, J« C,4 J-
—II—\—II—T—ii—k-Hi—i
o Vo \o <jfV \o
"V.
(d)
Fig. 4. (Continued)
l
the right-hand side with coefficient ¥ = <P = s~ require requirements of preserving matrix sparsity and maintain-
resistors. The identity matrix / on the right-hand side has ing a canonical number of variables, derivations for such
the fewest nonzero entries and hence requires a minimum structures are not straightforward and are explored in this
number of resistors. Minimum capacitor or resistor prop- section. Rearrange (3) as
erties are useful for certain fabrication technologies such
as continuous-time transconductance-C [13]. ^ A + <t>B + zD + D y 1 n = (l + z)J' (1
la)
The disadvantage of this method is that even when C
or T is tridiagonal, C - 1 l
or T~ is usually a full matrix, A = 2 / TC + T/2T - G, + G„
representing a fully interconnected system. This will cost B = 1TT
extra elements, the number increasing with the square of D = 2G
l l
C. ^0 C, T T
I f—II f—U
\°
X
V!"
J. " ).
1 I 1
Fig. 5. U L - L U type S C circuit.
C, 1 + Z -'
UL as c = L U.
bs b (13g)
C, 1-z-'
Also let
f=-(l+D A-')J ls (13h)
Then (11a) can be linearized as r. I
l
(14a)
-{2+D V; )j is
(14b) M 1
—
l
\L ,w -vv.,w.-u„u; j.
b b
PING el at.: A M E T H O D O L O G Y
T T T T
W' MM' HI
r r
-ih
T
1 1
v" —M—i h si
(a)
T T T T T T
II f—lh
3BT 5
c
» v T
M l r-ir—+MI ]„ it )1 I I )
1 1 1 1 1 1 '
(b)
Fig. 7. Exact bilinear-LDI S C fillers, (a) Left-ULD type S C circuit, (b) Right-ULD type S C circuit.
C) Cj Ct Ci Ci CM CI; Cu
II I
II
3
1
10 13
Lee and Chang [4] from a topological basis for SC cir- Alternatively, the following systems can be used for left
cuits. Application of the methods of Section I I to (17) and right decompositions:
results in a similar range of circuit structures.
The input of (17a) has a (1 + z) or (1 + z ) multiplier. _ 1 A,W = -(<t>B + D)V-2(-J) (18a)
The realization of the noncausal factor ( 1 + z ) can be AV = V W - A f \ - J )
r (18b)
- 1 _ 1
accomplished by multiplying by z giving (1 + z ) , and
introducing a delay of one period. Several realizations are lAV= -<t>(B,W+ DV)- ] (19a)
possible. l
1
\W =VB V-2Br J-
r (19b)
Direct realizations of (l + z ' ) for left-decomposition
l
designs and (1+ z " ' ) / ( l - z~ ) for right-decomposition It can be verified that (18) and (19) are equivalent to the
designs can be achieved by using some special circuit original system (16), respectively. They can be directly
arrangements [15], Fig. 6. simulated by SC circuits with L D I type integrators. Notice
860 I E E E TRANSACTIONS ON CIRCUITS AND SYSTEMS. VOL. 38. NO. 8. AUGUST 1991
5- 6 7 0 12 3 IS 7 22
e
la
43 47 1 53 55 00 57 106 59 12
44 65 71 46 77 SO 61 52 69 54 56 97 5fl 03 60 62
95 09
o e o e o e o e o e o i
r- r r r r f~ f"~ t B—U-
\ Ja 'o >o 'o ?Q *e 'o ?e ro e
<J <F
kP k j kP
24 23 26 25 28 27
k j k? kF kh
30 29 32 31 34 33 36 35 38 37 40 39 4
out
Fig. 9. L e f l - L U D S C ladder simulation with broken loops.
if J has only one nonzero input, i.e., J = [y,,0, • • • 0], it is Frequency IkHzl
l 5 8 11 It 17
preferable to have the first column of A f and Bf' with
as few nonzero entries as possible. This number is mini-
l 1
mized to one when Aj~ and B," are upper triangular
matrices, which occurs when U L or IA or IB decomposi- -30 -
tions are selected. Fig. 7 shows examples of exact left-and
right-ULD circuits.
Both the above methods require some extra compo- -50 -
nents that may introduce a significant layout overhead for
low-order filters. They may be simply dispensed with by
replacing the 0 + z ~ ' ) factor in the numerator function 2 -70 -
by 1 or z ~ \ resulting in the same circuit structures as a
LDI ones. However, this incurs a penalty warping func-
- 1 l/2
tion of (1 + z ) or z" cos"'(<u7y2), and a zero at half 90
the sampling frequency is lost. The distortion introduced
in the passband by cos(<o7"/2) can be corrected by pre-
warping the original prototype ladder. This can also be
conveniently combined with sin(^r)/jr correction resulting
in a x / t a n ( * ) function, which can then be superimposed
on the frequency response specifications [16]. If the sam-
pling frequency is very high compared with the center
frequency, as is often the case in practice, x / t a n ( j : ) = 1
110
-130 -
Si (a)
TABLE II
D E S I G N D A T A FOR A 2 0 T H - O R D E R S C L A D D E R FILTER
C,5 0.51578
lower passband edge 0.9075 upper passband edge 1.1065
lower stopband edge 0.877 upper stopband edge 1.132
passband ripple < 0.03 dB
lower stopband atten. >95dB upper stopband atten. >85 dB
Component Values for the SC Ladder
NODE1 NODE2 VALUE NODE1 NODE2 VALUE NODE1 NODE2 VALUE NODE1 NODE2 VALUE
4 3 24.38 4 5 10.37 6 5 9.095 6 7 40.92
8 7 39.09 10 9 47.51 10 11 6.576 12 11 25.31
12 13 4.401 14 13 5.049 14 15 4.493 16 15 43.78
18 17 46.54 18 19 7.795 20 19 14.39 20 21 4.113
22 21 2.712 24 23 16.32 26 23 18.05 26 25 28.15
28 25 18.62 28 27 30.77 30 29 28.71 32 29 9.335
32 31 22.63 34 31 9.969 34 33 34.52 36 33 3.243
36 35 31.09 38 37 32.58 40 37 15.40 40 39 28.45
2 39 6.541 2 41 30.91 43 44 1.000 45 46 1.000
47 48 1.000 49 50 1.000 51 52 1.000 53 54 1.000
55 56 1.000 57 58 1.000 59 60 1.000 61 62 1.000
63 64 3.764 65 64 4.341 63 68 1.670 65 68 3.558
71 68 1.000 65 74 6.805 71 74 13.62 77 74 1.000
71 80 1.000 77 80 10.02 83 80 2.349 77 86 1.000
83 86 5.139 89 86 2.533 83 92 1.000 89 92 1.852
95 96 10.65 97 96 1.000 95 100 1.000 97 100 10.70
103 100 4.128 97 106 1.463 103 106 4.049 109 106 1.000
103 112 1.243 109 112 1.000 24 3 2.993 2 21 1.383
115 3 1.000
total capacitance 814 units capacitance spread 47.5 units
capacitance unit lpF clock frequency 800 kHz
number of capacitors 73 number of switches 81
number of op-arnps 20
lower passband edge 10.1 kHz upper passband edge 12.45 kHz
lower stopband edge 9.18 kHz upper stopband edge 13.00 kHz
the transfer function. The response error thus caused can VI. COMPARISONS AND APPLICATIONS
be eliminated in the approximation procedure, by replac- Two notable novel categories of ladder simulations
ing poles at infinity by ones at - 2 f on the real axis. The
s presented in Table I are left-LUD and right-LUD (and
negative capacitance required can then be incorporated in their duals by U L decompositions) circuits. Both have
the synthesis of the passive ladder prototype [16]. The some characteristic features.
low-sensitivity properties are not influenced by the intro- In the left-LUD method the below-diagonal and
duction of negative elements [17]. above-diagonal elements of A are separated to matrices
T
For high-order S C filters, these inductor and negative L and L , respectively, effectively removing all capaci-
a a
capacitor pairs can be used to cancel the off-diagonal tor-coupled op-amp loops. They also demonstrate excel-
2 15
lent properties regarding component spread and dynamic and d((o) over the passband
T
range for bandpass design. However, since Z.„ and L a
must be square, the dimension of the intermediate vari- 5=- s(w)do> (23a)
able vector W is n (the number of nodes), and so alto- width of passband Vssband
gether W and V contain 2n variables. Sometimes this is 1
more than necessary. For instance, the left-LUD simula- J
width of passband •'passbandd(u>)dw. (23b)
tion of Fig. 4(a) could also be used to realize a fifth-order
low-pass function. Alternatively, right-LUD methods Normally the chip area required for fabrication of an
would use only 2n - 1 variables, a canonical number for SC filter is measured by
an odd order prototype. T=
c E c f (23c)
For right matrix decompositions, B and B can be all capacitors
t r
made rectangular and the dimensions of V and W are not but to reflect the influence of capacitance spread the
necessarily equal. Consequently, the dimension of W can following index will also be used
be smaller than that of V. In the case of a low-pass filter
simulated by L U D methods, it is mandatory that T be C-f E c/]' . / 2
(23d)
singular [17] so that a zero column in L will appear after
r all capacitors
LU decomposition. It can be then deleted, making L an r An overall performance index of an SC filter can be
n x ( n - l ) matrix. Thus one variable is saved in form- defined by
ing W.
CS
There are also sensitivity problems in Iowpass design /»« — (23e)
for all the methods mentioned above except right-LUD.
Deviation in the entries of B, caused by the inaccuracy of For these indices, it is desirable to have lower S, T , C, e
the element values associated with these entries, may and P (the lower limit is 0). The maximum op-amp output
cause B to become nonsingular, introducing a zero at will always be assumed to have been scaled to 1, so that
ft) = 0. Extra zeros introduced at the origin can be viewed D will always be a positive number less than 1. It is
either as an advantage or disadvantage; for instance, desirable to have D close to 1. which means that all the
low-frequency noise suppression can be facilitated by op-amps have equal output swing in the passband.
these zeros. The right-LUD method does not have this For a bandpass filter, the relative bandwidth is defined
problem as it involves multiplication of matrices with only by
(« - 1 ) rows or columns. The resulting matrices can never +
RBW=(a> -a>-)/a, , a> = (a <o~ ) ' + / 2
(24)
m m
have a full rank n. +
where a i and a>~ are the upper and lower band-edge
The right-LUD method, however, does have the draw-
frequencies, respectively. It is known that RBW has a
back of possessing capacitor-coupled op-amp loops. Un-
great influence on the system performance. For sixth-order
desirable large component spread and poor dynamic range
elliptic designs, let the passband ripple be fixed as 0.1 dB,
are also observed for certain bandstop designs.
stopband attenuation 50 dB and f , / f ratio 25 ( 2 i r / = m m
These arguments indicate that right-LUD is a good <o ,f, is the sampling frequency). Computations of S, T ,
m c
candidate for low-pass design. It is more complicated to D, and P are displayed against relative bandwidth (Fig.
reach any conclusion about bandpass designs as the per- 11). The same analysis has been performed for higher
formances the circuits vary dramatically according to the order elliptic type designs and similar trends are ob-
relative bandwidth. This can be seen from some compara- served.
tive studies.
B. Narrow Bandpass Filter Design
A. Comparison of Bandpass SC Ladder Designs
From Fig. 11 it can be seen that the left-decomposition
The following indices are used as global measures of
designs have very good total capacitance over the
system sensitivity and dynamic range, respectively:
narrow-band range, but the biquad method is better
1/2 around R B W = 1 . Regarding the sensitivity index S, all
ux
:,- d\H(u>)\ the ladder designs are much better than the biquad
*(«)- E (22a)
\H(io)\ Be, method over the whole range as expected [1]. The plots
l/M for cascade biquads are occasionally discontinuous; this is
< * ( « ) - { n i * u / « » ) i } (22b) due to the fact that E-type and F-type biquads are se-
lected according to Q-factor required, and discontinuity
where (c,) and (W ) are the sets of capacitances and
m
of internal nodal voltages may take place when the design
op-amp output voltages, respectively, and M is the num- is switched from E-type to F-type or vice versa affecting
ber of op-amps. the dynamic range index. Another reason is that the
As the passband behavior is of most interest to filter pairing of biquadratic sections is carried out to achieve
designers, define two indices for system sensitivity and minimum total capacitance, which does not take into
dynamic range, which are the average measures of s(a>) consideration voltage levels.
PING el at: A METHODOLOGY
863
LEFT-DIRECTIIAI
n LEFT-LUD
UL-LU
B I QUAD
1 RIGHT-LUD
1 S
LU-UL
ie i—i—i—i i i 111
io i o
Relative bandwidth
a)
B I QUAD
LEFT-DIRECT!IA)
10
UL-LU
LU-UL •
LEFT-LUD
I—' I I I I I i — i — r r-t- h l J
1 0 1
Relative bandwidth
<b
Fig. I I . Performance comparison of 6th-order bandpass filter realizat ions, (a) Dynamic range, (b) Sensitivity, (c) Total
capacitance, (d) Performance index.
I E E E TRANSACTIONS ON CIRCUITS AND SYSTEMS. VOL. 38, NO. 8. AUGUST 1991
/
/
RIGHT-LUD
/
/
LEFT-LUD /
/
LU-UL
1 0
.'/
J.
UL-LU
FX
B I QUAD
• a
LEFT-DIRECT(IA)
r 1
Relative bacdwduj
(c)
10 3 B I QUAD
LU-UL
LEFT-DIRECT(IA)
\ UL-LU
10
RIGHT-LUD
LEFT-LUD
1
I I I I I—I 1 I I 1 I
Relative bandwidth
(d)
Fig. I I . (Continued)
2 18
PING €t al: A M E T H O D O L O G Y
IQUAD
RIGHT-OiRECTtBll
Me ptuts Input
RIGHT-LUD
LeFT-DIRECT(IA)
U)
BIQUAD
RIGHT-DIRECT!BI)
RIGHT-LUD
LEFT-LUC
LEFT—DIRECT!IA>
(b)
Fig. 12. 6th-order asymmetric bandpass filler realization, (a) Overall response, (b) Passband response, (c) Sensitivity, (d)
Dynamic range.
T A B L E III TABLE IV
D E S I G N D A T A FOR A W I D E - B A N D PROTOTYPE L A D D E R O F F I G . 3 D E S I G N D A T A F O R W I D E - B A N D S C L A D D E R SIMULATIONS
C, 0.3465 c 2
0.1580 1-2 0.2484 Left- Left- Right- Right-
C, 0.6119 c. 0.0596 0.3006 LUD Direct (LA) LUD Direct (BI)
C 5 0.4836 Ls 4.0362
G i n 1.0000 0.8802 c, 1.000 1.000 3.147 3.147
Sampling frequency 100 kHz c 2
9.174 9.174 9.392 9.392
Lower passband edge 300 Hz c, 1.000 1.000 1.000 1.000
Upper passband edge 3400 Hz c. 3.620 3.620 9.821 9.821
Passband ripple < 0.2 dB 1.615 1.615 3.387 1.615
Lower stopband edge 10 Hz c 6
1.266 1.266 1.266 1.266
Lower stopband ait. >30dB c. 4.445 3.435 3.435 3.435
Upper slopb'and edge 5000 Hz c a
1.375 1.375 4.044 4.044
Upper stopband att. >45 dB c, 3.736 3.736 2.391 5.433
21.697 25.451 22.474 22.474
c„ 4.632 5.433 1.690 3.736
21.169 22.474 11.201 25.451
c„ 1.573 1.573 2.935 1.000
1.000 1.000 1.000 1.573
A comparison of the overall performance indices indi- c 1 5
1.612 i.oob 2.074 1.473
cates that the left-LUD method is the best candidate for 1.000 1.000 1.000 1.000
narrow-band design RBW < 1 and the biquad method is C,7 1.000 1.000 1.000 2.137
c 8.326 9.325 9.953 7.070
best for bandpass design with RBW around 1. C,
1 8
From Fig. 11 it is seen that both ladders and biquads total 111.199 117.373 150.109 123.014
are far from ideal for a very important area of filtering spread 21.697 25.451 51.232 25.451
866 I E E E TRANSACTIONS ON CIRCUITS AND SYSTEMS. VOL. 38. NO. 8. AUGUST W l
UL-LU
EIQUAD
RIGHT-LUD
1
LU-UL /
LEFT-DIRECT!IAI /
LEFT-LUD
Relative bandwidth
RIGHT-DIRECT BI
LEFT-DIRECT 1AI
a
UL-LU
! 10
LEFT-LUD
o
B I QUAD
\ RIGHT-LUD
I
8
applications, that is, for the voice band application with narrow and very wide passbands; the leapfrog method is
passband from 300 Hz to 3400 Hz whose R B W > 3. It is the best choice for sharp transition low-pass filter design;
found that for this type of wide-band design the most and cascade biquads are the best choice for moderately
significant factor causing the deterioration of all designs selective low-pass and bandpass filter design. Some spe-
is the location of lower band finite zeros, which approach cial design techniques can be applied to produce efficient
zero as the RBW increases. Realization of these zeros circuits in other technologies such as continuous-time
requires large capacitance spread. The deterioration pro- transconductance-capacitor filters.
cess can be reduced if the lower band finite zeros are
replaced at the origin. Since the zeros in the lower stop- APPENDIX
band are very close to origin anyway, this shifting will
When A and B are tridiagonal, L and U are also a b
have only a slightly deleterious effect on the characteris- tridiagonal as well as triangular. Separate the diagonal
tic. The performance of a 6th-order design of this t>pe is and off-diagonal parts of the matrices
shown in Fig. 12. The design data of four SC circuits are
given in Tables I I I and Table IV. The left-LUD has the La = L ad + L ao (25a)
lowest capacitance spread and best dynamic range. V „ ' V + U (25b)
M b o
alternative ladder circuits with cascade biquads (Fig. 13). nonzero entries only on the first lower off diagonal, and
This result echoes the conclusions for the previous spe- V has nonzero entries only on the first upper off-diago-
bo
nal,
cific example—that left-LUD design is still the best choice
for this family of filtering applications, exceeding all oth- 0
ers over the whole bandwidth range. 0
I' .1., —
D. Low-Pass Design
Fig. 14 shows a sweep of performance index P against
transition bandwidth ratio (passband edge/stopband
edge) for various 5th-order elliptic low-pass filters. The
passband ripple is fixed at 0.1 dB; the passband edge is U„ = b (26)
1 kHz and the sampling frequency is 200 kHz. Note that
only cascade biquad and right-LUD circuits provide one- 0
op-amp-per-pole realizations; the others require six op-
where * stands for the nonzero entries.
amps. Overall, the right-LUD is the best choice followed
by biquad. The low sensitivity and good dynamic range of In (13) assume that U and L are also tridiagonal as bs
the right-LUD are its main advantages despite the lower and triangular matrices. Separate U and L as as bs
software package [8] has been developed employing this Since in (28a-c) the number of constraints is less than the
scheme together with many other sophisticated approxi- number of variables, we can assign
mation, ladder synthesis, and optimization algorithms.
L„st = l (29)
V I I . CONCLUSIONS
which guarantees the readability of system (14). From
(28) and (29) we have
A methodology has been developed for the design of
acWe-RC and SC ladder simulation filters. A wide range ^ d - ( L a d - L a o V j U b o y x(U -U LjL. )
bd bo a (30)
of circuits can be derived by adopting different matrix
and remaining variables can be solved from (28):
factorizations, notably L U and U L decompositions, in-
cluding both existing and novel structures. A detailed Lbso ~ Uasa-LaoUbd (31a)
comparison of various SC circuit structures has been
(31b)
undertaken and some notable conclusions are: the left-
LUD method is the best choice for filters with very The matrices in (30) and (31) may be singular and the
868 1F.F.F TRANSACTIONS ON CIRCUITS AND SYSTEMS. VOL. 38, NO. 8. AUGUST 1991
normal inverses do not exist. In these circumstances, [15] T . - H . Hsu and G. C. Temes. "Improved input stage for bilinear
switched-capacitor ladder filters." IEEE Trans. Circuits Syst., vol.
Moore-Penrose's generalized inverse can be used [18]. As CAS-30, pp. 758-760. Oct. 1983.
the matrices in (30) and (31) are all diagonal, the proce- [16] R. K. Henderson, "Computer-aided design of switched-capacilor
dure to obtain their Moore-Penrose inverse is very sim- filters," Ph.D. dissertation. Univ. Glasgow, 1989.
[17) Li Ping, "Theory and methodology of integrated ladder filter de-
ple. The Moore-Penrose inverse of a diagonal matrix sign," Ph.D. dissertation, Univ. Glasgow, 1990.
D = d i a g [ d , - • -,d ] is also a diagonal matrix given by
n nn [18] C. R. Rao and S. K. Mitra, Generalized Inverse of Matrices and Its
M = d i a g [ m , - • • ,m ] with
u nn
Application. New York: Wiley, 1971.
PAPER 47
IEEE JOURNAL O F S O L I D S T A T E C I R C U I T S . V O L : • . NO 1. S E P T E M B E R 199:
Abstract—Conventional S C realizations of wide-band filters ments of the amplifiers and switches compared to other
demand large capacitance spread and exhibit serious sensitivity solutions [ 7 ] . The strategy is to retain E-type damping
problems. The U I . - L U structure is a ladder type simulation
which demonstrates superior sensitivity performance and (low capacitance spread) simultaneously with right-hand
maintains low capacitance spread. The design illustrates how matrix decomposition (low sensitivity), which is not read-
the facilities of a modern filter compiler can be utilized to solve ily available with existing approaches. A modern filter
some quite difficult practical problems in a real application. compiler | 9 ) has been used and exercised considerably to
Results from fabricated devices confirm the predicted proper- provide an optimum solution.
ties.
I I . W I D E - B A N D F I L T E R DESIGN
I . INTRODUCTION
An audio receiving filter for mobile telephony appli-
Template
Frequency (kHz)
Cascade - biquad
Coupled - biquad
Leapfrog
3 -
a i i
21
Fr«qu«ncy (Mi)
in Fig. 2. The sensitivity measure being used is It can be seen that both l e f t - L U D and coupled-biquad
circuits suffer from a low-frequency sensitivity peak,
Mu'l = i s whereas the leapfrog and cascade-biquad circuits exhibit
H(u) | dc poorer sensitivity performance at the higher band edge.
I E E E J O U R N A L OF S O L I D - S T A T E CIRCUITS. V O L . SO ') SEPTEMBER I'M:
TABLE I
:
COMPARISON or CAPACIFANCT. COSTS F OH FII.IT:K R F ATION
UL-LU 746.4 1 IM
LUD 703,3 1 12.5
Couplcd-E-Biquad 756.4 1 12.5
Leapfrog 1 714.7 957.8
Biquad 728.8 122.5
"6
"j rYYY\-_-3~
J G
,„tw U , ?>-,
The new U L - L U structure (detailed in the next section) troid to help with matching and to reduce offsets. The
maintains a low sensitivity over the whole passband. T a - area-to-perimeter ratio of unit to nonunit capacitors is kept
ble I gives a comparison of capacitance cost and high- constant so that parasitic peripheral capacitance effects are
lights the penalty of choosing a leapfrog realization. The cancelled. The size of the filter is 3027 jim x 8 9 4 nm.
conclusion of these studies is that the U L - L U structure which is relatively small considering the complexity of
offers the best solution to this filtering problem. F i g . 3 the response. Passband details of the measured frequency
shows the structure of a passive prototype synthesized response can be seen in F i g . 5. The response meets the
from this transfer function. Note that the position of a template very well and over a number of devices there
series capacitor between the first and second nodes is re- was hardly any deviation. T h i s illustrates the low sensi-
quired to ensure E-type terminations [7] in the S C circuit tivity of the filter to process variations and hence the ro-
realization shown in F i g . 4. The simpler structure of the bust nature of the design. For commercial production this
filter towards the output is due to the cancellation of is an important factor in ensuring an increased yield and
feedthrough capacitors by the specially positioned real reliability of the devices. T h e noise floor at 1 k H z was
axis zeros (capacitors C and C in F i g . 3 therefore have
h s - 6 0 dB and was approximately level at this value over
negative values). ihe entire passband: no harmonic level could be detected
The circuit uses a clock frequency of 128 kHz and has above the noise floor.
been fabricated onto silicon using a 3-^m single-metal,
double-poly. 5-V process. The amplifiers used are single-
stage folded cascode without compensation capacitors, III. DERIVATION OF THE U L - L U STRUCTURE
stability being ensured by the capacitors in the filter: the The derivation utilizes a U L - L U decomposition to-
input devices are laid out as cross-coupled common cen- gether with a bilinear-LDl ladder design | 2 ] . Starting from
2 26
9
/
Simulated
i: x
Measured
Frequency (KHX)
Fig. 5. Passband detail of tenth-order niter
the nodal equation of a passive prototype ladder damping at the input of the S C realization and this is con-
sistent with E-type damping already implicit at the output.
(sC + s r + G)V = J, _ l
(1)
The virtue of E-type damping is to reduce capacitance
applying a bilinear transformation, and separating out a spread in the termination sections of the S C realization.
pair of L D I integration operators * = 1 / ( 1 - ; " ' ) and Notice now that the output is w „: as U is the upper tri- h h
]
• = z ~ / ( \ — z ~ ' ) . equation ( 1 ) becomes angular. w differs from the output by only a constant. hn
W = V-'L
a V. W = U V, and D = DU ~
a h h Again the correction, resulting in tan (x) / x prewarping.
s h
]
particular design illustrates the potential of modern filter pass filters using coupled biquad structures." IEEE Trans. Circuits
Syst.. vol. 27. no. 6. pp. 469-474. June 1980.
compilers and their application to difficult practical design
[5] Li Ping and J. 1. Sewell. "The L U D approach to swiiched-capaciior
problems. filter design." IEEE Trans. Circuits Svst.. vol. 34. no. 12. pp. 161 I -
1614, Dec. 1987.
[6) Li Ping and J. 1. Sewell. "Filter realisation by passive ladder simu-
REFERENCES lation." Proc. Inst. Elec. Enu.. Pt. G. vol. 135. no. 4. pp. 167-176.
Aug. 1988.
(1) H. J. Orchard. "Induciorless fillers." Electron. Leu., vol. 2. pp. 224- |7] P. Korhonen and J. Korhonen "An analogue CMOS circuit for mobile
225,June 1966. radio base station signal processing." in Proc. ESSCIRC '89. pp. 196-
(2] Li Ping, R. K. Henderson, and J. i. Sewell, "A meihodology for in- 199.
tegrated ladder filler design." IEEE Trans. Circuits Svst.. vol. 38. no [8] P. E . Fleischer and K. R. Laker. " A family of active switched capac-
8, pp. 853-868, Aug. 1991. itor building blocks," Bell Svst. Tech. J., vol. 58. no. 12. pp. 2235-
[3] M. S. Lee and C . Chang, "Swiiched-capaciior fillers using ihe LDI 2269. Dec. 1979.
and bilinear transforms." IEEE Trans. Circuits Svst.. vol. 28. no. 4, [9] R. K. Henderson, L i Ping, and J . I. Sewell. "A design program for
pp. 265-270. Apr. 1981 digital and analogue filters: P A N D D A . " in Proc. ECCTD (Brighton.
[4] K. Martin and A. S. Sedra. "Exact design of switched-capacitor band- U . K . ) . Sept. 1989, pp. 289-293.
2 2 8
PAPER 48
2 29
R.K. HENDERSON
Centre Suisse d 'Eleclronique el de Microtechnique, Neuchatel, Switzerland
L I PING
Department of Electrical and Electronic Engineering, University of Melbourne, Parkville, Victoria 3052, Australia
J.I. S E W E L L
Department of Electronics and Electrical Engineering, University of Glasgow, Glasgow, G12 8QQ Scotland
Abstract. A review of the progress in automated design of analog integrated filters is presented. Such tools are
ahead of other analog circuit automation in terms of the acceptance by designers and practical applicability. A
survey of the present-day commercial and academic systems is made and the range of facilities available is com-
pared. The problems faced in the design of this type of software are typical of the problems of analog design systems
in general; lack of openness for introduction of new design knowledge, difficulties of dealing simultaneously with
expert and novice users, poor integration in design environments, and user-interface problems. The structure of
a typical system is studied and the computer methods used within are discussed with regard to such issues as speed,
flexibility, and ease-of-use. Some future directions for analog filter compilers are proposed.
57
2 30
design methods. This need has been made particu- between levels is accomplished by a synthesis step
larity evident by the continual emergence of new which converts a design f r o m a behavioral to a struc-
technologies, recently switched-current and tural description. For example, filter synthesis com-
MOSFET-C circuits [ 3 - 5 ] . mences with the description of the design in terms of
a frequency response and terminates with a netlist of
Several filter compilers are now offered commer-
parameterized building blocks, e.g., op amps, transcon-
cially and many more reside in academia. Filter design
ductors, capacitors, resistors, and switches. Several syn-
aids are also increasingly found as extensions to digital
thesis steps follow before we finally arrive at silicon;
signal processing packages, standard circuit analyzers,
for example an analog ceLl generator w i l l translate from
and mathematics and systems simulation languages.
building block to sized device schematic and a layout
This article w i l l look at some of the issues faced in the
tool f r o m sized device to layout geometry. A typical
design of this kind of software as a small illustration
organization of such a system is shown in figure 1 and
of the problems facing the current generation o f analog
the "state-of-the-art" is summarized in table 1 which
CAD tools. One particularity significant issue, in view
compares some of the most reputed systems (this list
of the rapid technological developments over the past
is by no means exhaustive).
30 years, is the extent to which "technology indepen-
dent" filter design can be achieved. This means the Some pertinent comments follow:
degree to which the shared design methodology of filter 1. There is a very dominant forward path in the com-
in various technologies can be exploited to provide pilation process. Most compilers adopt the principle
reusable programs (a major goal of present-day com- of making fast synthesis modules and placing the de-
puter science). Good design software must not only be signer in control of a "weak" optimization (manual
flexible enough to adapt easily to the changing possi-
bilities offered by the technology but must also be able System Specriicaion
2. Background
Layout Synmesu
from a high level design description into layout [7], Layout - Design Tool iDau
58
Analog Integrated Filter Compilation 219
Continuous-time/
Switched-capacitor
Passive RLC
approximation
Commercially
Equalisation
synthesis
Extended
synthesis
available
Cascade
Classical
biquad
Ac.tive-RC
Layout
Special
Allpass
Ladder
Name features/
comments
S/F1LSYN
filtorX
•• m • •• • • 9 m m m m
•• •
m m m
The original passive filter synthesis program. New extended
to various filter implementations and design methods |8],
A compendium of programs written by graduate students at
• • •
VITOLD synthesis. Sensitivity and THDoptjmoation.(13]
Commercial package from Mentor Graphics [14],
•• • •
AutoFilter
High order touch point approximations.VVide variety of
#
•
PANDDA m m m m filter structures including new topologies, Non-ideal optimtsation.fl7]
• m • •• •• •
IMSYS ANACAD product. Good practical design options .[15].
First OTA-C filter synthesiser (16).
••
FIESTA
• • •
SCSYN and ladder structures. Noise and capacitance optimisation.(18J
• •
e
Dedicated SC filter layout synthesis tool. Amplifier and
•• •
CAST/ALEX
• ••
switch sizing.|20]
Japanese contribution to filter automation. [21]
•• • •• •• • •• • ••
SCULPTOR
Promising PC synthesis tool in popular environment of
PSpice SPICE analysis tools.
Belt Northern Research SC synthesis toolplus University of Waterloo
SCDS analysis software market ad by Cadence [22]
iteration round the tools, viewing circuit simulation decomposition is rarely automated leaving this step to
results and altering specifications). The alternative be performed by rule-of-thumb and system designer's
"strong" optimization approach consists of a core cir- expertise. This is a prospective area for future auto-
cuit simulator, a multivariable optimization package, mation. The aim would be to help system designers to
and a graphical user's interface for viewing simulation set-up realistic specifications and to obtain quickly a
results. In this case, an ideal design is taken as a start- feel for the trade-offs involved between blocks and the
ing point and circuit element values are manipulated possibilities offered by the technology.
by the optimizer to improve the circuit response i n the 3. A t each o f the filter description levels there is a
presence of nonidealities such as switch resistance or potential to save the state of the design in some data
amplifier bandwidth. However the "weak" approach format. There is very little standardization o f these for-
is more efficient since there are generally many fewer mats (except at netlist or layout level) leading to a d i f f i -
filter specifications than circuit element values, but it culty in porting information between systems. For ex-
relies on designer expertise to interpret simulation ample, a standard frequency domain specification
results and modify specifications accordingly. The format or transfer function description format would
"strong" approach trades speed for generality and ease- allow much more shareable and extendable use of ap-
of-use tending toward the "push-button" end of design proximation software. A t present however, advanced
automation. Both (and a range of possibilities i n be- approximation techniques become an inaccessible part
tween) are necessary for a complete system. of a single system and there is very poor reuse of even
2. Very few existing compilers offer the complete standard software. Standard simulator formats for ana-
cycle as depicted in figure 1. In particular, the system log and switched-capacitor circuits are emerging by
59
2 32
default as SPICE and SWITCAP. Layout descriptions allow the digital tool to claim " f u l l mixed analog/digital
have already the GDSII and E D I F formats available. synthesis capability" while protecting the analog tool's
In certain cases this is the domain of analog behavioral
modeling languages such as the projected analog exten- interests.
sion to V H D L .
4 . Compilers are normally devoted to only one 3. Computer Methods f o r Analog Filter Compilation
design strategy and technology. Commercial compilers
in particular restrict themselves to the mainstream The previous section has taken an external view of the
design flow of cascade biquad designs in switched- existing filter compilers and the trends in their develop-
capacitor technology derived from classical approxima- ment. The present section w i l l take an internal view
tions (Butterworth, Chebyshev, etc.). This limits the and w i l l examine some of the computer techniques
designers ability to combine and compare technologies being used inside such systems.
(e.g., active-/?C and SC) and the benefits of different
topologies (e.g., biquad and ladder). Not enough work
has been put into an open framework of description for 3.1. Approximation
the design methods and the networks.
5. The importance of having more powerful approx- In filter approximation a realizable transfer function
imation software is emerging [23]. As the design steps must be computed to meet specifications of amplitude
close to the fabrication technology are optimized the and delay in time or frequency domains. Obtaining a
"room for improvement" shifts toward the higher levels good approximation is a struggle between conflicting
of design (see also point 2) such as approximation. For demands of filter selectivity, group delay variation, time
example, traditional approximations based on classical domain response, and transfer function order. Approx-
functions (Butterworth, etc.) yield fruitions w i t h e r imation software can either help or hinder this process
passbands which are suitable for frequency division depending on its flexibility, speed and ability to satisfy
multiplexing applications. However, filters with shaped both advanced and beginner users.
frequency response can also compensate for signal dis- Table 2 shows a survey of the most successful algo-
tortions f r o m other parts of a communications system rithms for filter approximation. From a software per-
(typically antialiasing and transmission line losses). spective, a major remaining issue is simply the degree
Compared to a standard equalizer/filter solution the in- of accessibility to the user. Entering the specs and view-
corporation of an equalizing capability in the filter ing the responses of classical filters such as Butterworth
results in smaller circuits and better overall perfor- or elliptic is a fairly simple matter of entering fixed
mance. Approximation for such designs is increasingly parameters and viewing the approximated solution. The
being recognized as a necessary utility in a compiler order of the response can be determined from passband,
and several packages offer some facility to optimize a stopband ripple, and frequency edge specifications (in
frequency response to an arbitrarily shaped template fact any one o f these can be left open, to be determined
[17, 19, 22]. automatically f r o m the specs on the other three). There
Instrumentation and data communications require is usually very little exploration of responses to be done
approximations based on phase and time domain re- (often the specification w i l l even define the order and
quirements. There is very little software available for
such tasks and mostly designs are either "handcrafted" Table 2. Comparison of different filter approximation algorithms.
or cast in terms of problem to be solved by an optimiza- Uu' Speed Flexibility
Approximation
Mat hod
Props/lie*
tion package. NOVICE FAST LESS CLASSICAL Butterworti Chebyahey etc Rat past [24]
FLEXIBLE and stopband frequency domain
amplitude specs.
6. The majority of compilers originate from univer-
'Pole ptaoer* aloonftm family,
sities. As with all analog C A D software it covers a very REM6Z ataws M i n t e d lroc*>artcy dorrsun
pastband and sopbend ipecs
[24]
narrow and highly specialized group of users. The ratio Equrippia group delay i poofcabora
ALLPASS GROUP
Remvi-type algonffim* appWaWe [25|
of difficulty of development of the software to poten- DELAY
EQUALISATION
Lost »ftoa*il ten joint ampkrude/
group delay opemMOon.
make use of the schematic capture, layout, database, MORE SIMULATED Slewresi but avoids local mi rum a
EXPERT SLOW FLEXIBLE ANNEALING problem* and a very fleiib'e [19]
and user-interface standards). Closer integration would
60
2 33
H h
3.2. Prototype Synthesis
a(x) = C z-7C,<l-z->
2
the "push-button" user a default ladder structure can RULE-OF-THUfcffl Cascade biquad of cunt Choose i
fixed sequence o< polo* and nros
POLE ZEBO FWRINC acaxtfog to pre-dofinoQ' rule
normally be proposed using inbuilt knowledge of the
allowable synthesis steps. For the expert who wants to Cascade bulged c r c u a Te»t al
COMBINATORIAL
improve on the silicon area, or who needs a special POLE-ZERO
PAIRING
structure of prototype the ladder synthesizer can pro-
pose a prioritized choice of structures. ITERATIVE Design a fixed iftvour* ol ladder
LADDER by maidwg n> trmrittei (unction
DESGt: itBratr>e<y with deeired one
Cascade biquad design requires a choice of a pole
sequence and corresponding zero sequence ("pole-zero t
Syntheu* a taddtt by decofflpeting
ftt r a n i tor hincson using polynomial
PASSIVE menpuiaBon. Accuracy probfem*
pairing") of which there are a factorial number. It is EXPERT
LADDER
SYNTHESIS
Oten neecta expert but very
general
61
2 34
all possible combinations. However the assessment 1. The netlist can be generated directly from the matrix
needs to be based on some performance measure of the form by representing each ,v variable by an active
filter, e.g., area and sensitivity. A compiler can have
a built-in performance measure or can allow a user to device (op amp, transconductor. etc.) output and
supply his own based on access to analysis results. The linking the appropriate block as represented by an
expert should also be permitted to enter a user-defined a„ f r o m output ito input /'. The y vector represents
pairing while the "push-button" user can be offered
an input signal (normally only one), and they should
default pairings based on reasonable rule-of-thumb
choices, e.g. increasing Q factor for lowest noise be linked to the corresponding .v input via the blocks
transmission. indicated by the b terms.
2. A library of the standard cells and their block trans-
fer functions can be set up and modified according
to changes in the technology or development of new
cell topologies, e.g., low offset integrator structures.
3.3. Network Compilation 3. Being based on a matrix representation the scheme
is ideal for computer implementation and storage.
Network compilation is the stage at which a filter proto- A library of matrix manipulation routines can be
type is decomposed into a netlist of ideal linear net- made available to allow quick coding of new designs.
work blocks (capacitors, transconductance amplifiers, 4. The matrix can be analyzed directly by substituting
switches, op amps, etc.). There is very little standard- the numerical values of the block transfer functions
ization here, and the algorithms used are often rewritten and solving by conventional L U decomposition
for each different structure of filter. Yet a certain degree methods. This provides a quick first check before
of technology independence can be maintained at this an external circuit simulator is called and can supply
level so that the designs do not need to be recoded when estimates of sensitivity and dynamic range. Note that
the implementation of the building blocks change. The since the matrix rank is normally dependent on the
main task at this stage is the construction of a linear- order of the filter rather than the number of compo-
ized flow graph of the circuit in terms of the building nents in the eventual circuit that the analysis is very
blocks. The flow graph structure itself stays essen- efficient. The matrix is built up f r o m the transfer
tially constant between implementations. The basic functions of the building blocks rather than those
integrator or resonator building blocks do change in of the individual components inside them (as it
implementation (switched-capacitor, switched-current, would in a standard M N A scheme). This avoids
transconductor-D, active-RC). reanalyzing each occurrence of a building block in
A convenient and portable representation of a flow the circuit to determine its transfer function.
graph is a matrix system. Several authors make use of 5. The matrix allows scaling for minimum area and
such a system but no convention has emerged [30-32]. maximum dynamic range to be performed with ease.
One possible form is a pseudostate space description Simple row and column multiplications are required.
62
Analog Integrated Filter Compilation 223
F i l t e r Class
[l.290OOOft»05
u n d Definition
eouiripple
Ij.OOOQOOeHK -3.0Oe*0i - -
Fig. 3. Screen shot of X F I L T showing template specification of lower passband and approximation of eighth order filter.
3. The order of the transfer function (particularly for layout tools. These have been the slowest parts of the
ladders) design cycle to automate because they deal with d i f f i -
cult nonlinear design problems and are near the fast-
It is difficult to make general rules but usually some
changing demands of the technology. Nevertheless, sev-
helpful defaults can be proposed, e.g., leapfrog struc-
eral systems have been presented dedicated to switched-
tures are good for low-pass odd order designs, coupled-
capacitor filters [33]. These programs interpret the
biquad, and L U D are best for bandpass, whereas cas-
loading demands and required charging times from the
cade biquads excel at bandstop designs. Based on exper-
netlist as specifications on amplifier and switch designs.
imental running of a compiler over a range o f specs
They then dimension the devices in order to minimize
a rule base can be established to make a reasonable sug-
the power and area of the circuit either by an optimizer
gestion of a structure. I t is as important to preserve
or by using rearranged device equations. This is nor-
knowledge o f successful as well as unsuccessful trial
mally the least accessible part of the process. Analog
designs to avoid repreating the same mistakes. Other-
block generators are making efforts to become more
wise this involves the "weak" optimization of a designer
open to the designer but are still evolving [34-35].
comparing possibilities by repreatedly rurLning the com-
Often the solution is just to use a handcrafted standard
piler. This is not necessarily a bad thing in terms of
amplifier cell and to accept the overspecification of the
designer "psychology" as it keeps him or her as a
design.
valuable part of the design process.
Analog routers capable of taking into account special
requirements o f sensitive and noisy nets, power sup-
3.4. Module Synthesis and Layout plies and variable analog transistor sizes and styles have
been developed [ 3 6 ] . The inherent regularity of the
The interface between tho jlock-level network descrip- filter topologies makes this a more amenable task than
tion and silicon is filled by analog cell generators and for more general classes of analog network. Several
63
2 36
dedicated layout strategies for switched-capacitor net- form of the transfer function are specified band by band
works have been reported [18-21]. Normally a fixed (in this case the passband). Passbands can be assigned
topology is adopted (as in random logic layout), row of forms anywhere between equiripple and maximally flat
op amps, row of capacitors, row of switches. The routing and stopbands can have user-defined zero distributions.
problems are greatly simplified and dedicated algorithms Of particular note is the 20 dB/decade slope in the pass-
can be used. Sensitive nets such as the connections to band for preemphasis. Three zeros have been placed
the virtual ground of the op amp are known in advance at the origin to reduce capacitance spread over an
and crosstalk can be avoided. Capacitors are designed elliptic-style zero distribution. Two notches are placed
in units with constant area-perimeter ratios for high- in the upper stopband but are not seen because of the
accuracy matching. The user can often control the rela- frequency range of the plot. Figure 4 shows the designer
tive positions of switches, op amps, and capacitors. evaluating different realizations of the filter. Depending
on the designer's level of expertise so-called automatic,
interactive or expert modes of design offer progressively
4. Automated Filter Synthesis Example more possibility for user intervention and customiza-
tion of the circuit design. For example, capacitance
In this section an example of a complete synthesis of spread can be reduced by different pole-zero pairing
a nonstandard filter w i l l be illustrated f r o m the X F I L T algorithms. I n automatic mode, all possible combina-
compiler. Figure 3 shows a screenshot of a filter fre- tions are tried, in interactive mode the designer chooses
quency response being defined to the filter compiler. f r o m among certain simple preprogrammed pairing
The response characteristics are defined in stopband rules aimed at making certain specific trade-offs while
and passband by a piecewise template of lower and in expert mode he chooses the sequence by hand. The
upper bounds on amplitude and delay. The order and designer w i l l normally also try a number of different
.20e*02
F i l t e r lyptft Design l.25**0i 2.50e-03
F l l t a r Design node type—•
Blqued Dealer. Node autonotic
filouad Polo P e i r i n * Rule none 5tert Frequency J.000OOOe-01
5.0O»00e'03
of Points
1.200O0Oe*O2
O.OOOOOOe'OI Set feiis Fernet.
Md Tenplete Points Add Point Finish
nod Group fenpleto Points Erase Point Finish
F i l t e r Oesivn Software
If any problem found, pleese contest
64
Analog Integrated Filter Compilation 225
circuit structures, comparing component area, power, has been established. More work is now necessary on
dynamic range, noise, and sensitivity before arriving unifying the framework and offering better access to
at a final choice. A selection of ladder simulation and the design facilities. Integration with digital tool envi-
cascade biquad structures are available. For this filter- ronments and more standardization o f data formats is
ing problem, an eighth order cascade biquad circuit to be hoped for in this direction. Greater gains in terms
provides the best cornprornise between area and sensi- of design efficiency should now become available at a
tivity considerations. Normally the final structure w i l l system level. For example, by allowing the designer to
have been selected after a number of iterations between more easily observe trade-offs between different blocks
approximation and circuit realization stages. The meas- in his filter system design and offering more guidance
ured results o f the filter (figures 5 and 6) show excellent in the setup of reasonable specifications. This means
agreement with the original template and ideal using the compiler in fast first-cut design mode and by
response. Finally a filter layout is shown in figure 7. storing informaton about already explored design space.
The op amps and capacitors are standard cells taken Filter compilers should lead the way toward more inte-
f r o m a library. grated analog system design exploration tools.
5. Conclusions Acknowledgments
As regards the basic algorithms for filter design the field The assistance of Wolfson Microelectronics L t d . and
is now fairly mature. Moreover, the reliability with L u Yue at the University of Glasgow in producing
which filter compilers can produce good quality designs screen shots o f X F I L T is gratefully acknowledged.
7.2--
-7.2
.894 4 3 . 2
•eauency (KHz)
Fig. 5. Ideal passband response (solid) amd measured passband response (dotted) of eighth order S C filter.
65
238
I5
I3
Fig. 6. Overall ideal response (solid) and measured reponse (dotted) of eighth order SC fiiici
66
239
1
A n a l o g Integrated Filter C o m p i l a t i o n 22
7. N.K. Dutt and D.D. Gajski. "Design synthesis and silicon com- 28. H.J. Orchard. "Filter design using iterated analysis." IEEE Trans.
pilation," IEEE Design Test Computers, Vol. 7. No. 6. pp. 8-23. Circuits Systems, Vol. CAS-33. No. 11, pp. 1089-1096. 1985.
1990. 29. A.S. Sedra and P.O. Bracket!, Filter Tlieory and Design: Active
8. G. Szemirmai, S/FILSYN User Manual. DGS Associates. Santa and Passive. Pitman: London. 1979.
Clara. CA. 30. Li Ping. R.K. Henderson, and J.I. Sewell. "A methodology for
9. D.G. Nairn and A.S. Sedra, "Auto-SC, an automated switched- integrated ladder filter design," IEEE Trans. Circuits Systems.
capacitor filter silicon compiler," IEEE Circuits Dev. Mag.. Vol. vol. CAS-38, No. 8. pp. 853-868. 1991.
4, pp. 5-8, 1988. 31. A. Muralt and G.S. Moschytz, "Performance optimization of
10. G.V. Eaton, D.G. Nairn, W . M . Snelgrove. and A S Sedra. elliptic SC ladder filters using signal flow graph transformations."
"SICOMP: a silicon compiler for switched-capacitor filters," in in Proc. IEEE ISCAS, pp. 2205-2208, 1990.
Proc. IEEE ISCAS. pp. 321-324, Philadelphia, 1987. 32. W.M. Snelgrove and A.S. Sedra, "Synthesis and analysis of state-
1 1 C . Ouslis, M . Snelgrove, and A.S. Sedra, "A filter designer's space active filters using intermediate transfer functions." IEEE
filter design aid: filtorX, in Proc. IEEE ISCAS. Singapore, pp. Trans. Circuits Systems, Vol. CAS-33, pp. 287-301, 1986.
376-379, 1991. 33. T. Kaneko, Y. Akazawa, and M . Natagani. "Switched capacitor
12. W. M . Snelgrove and A.S. Sedra, FILTOR2: A Computer Filter and active-RC filter layout using a parameterizable generator,"
Design Package, Matrix Publishers: Illinois, 1979. in Proc. IEEE ISCAS. Singapore, pp. 2012-2015, 1991.
13. Y. Therasse, L. Reynders, R. Lannoo, andB. Dupont, "A switched 34. M G R . Degrauwe, B.L.A.G. Gofrart, C. Meixenberger, M . L . A .
capacitor filter compiler," VLSI Systems Design, pp. 85-88, 1987. Pierre, J.B. Litsios. J. Rijmenants, O.J.A.P. Nys, E. Dijkstra,
14. Mentor Graphics, Autofilter Users Guide, 1990. B. Joss, M.K.C.M. Meyvaen, T.R. Schwartz, and M.D. Pardoen,
15. J. Assael, P. Senn, and M.S. Tawfik, "A switched-capacitor filter "Towards an analog system design environment," IEEE J. Solid-
silicon compiler," IEEE J. Solid-State Circuits, Vol. SC-23, pp. State Circuits, Vol. SC-24, No. 3, pp. 659-671, 1989.
166-174, 1988. 35. R. Harjani, R.A. Rutenbar, and L.R. Carley, "OASYS: A frame-
16. M R. Kobe, E. Sanchez-Sinencio, and J. Ramirez-Rodrigo, work for analog circuit synthesis," IEEE Trans. Computer-Aided
"OTA-C filter silicon compiler" (to appear). Design, Vol. CAD-8, No. 12, pp. 1247-1266, 1989.
17. R.K. Henderson, Li Ping, and J.I. Sewell, "A design program 36. J.M. Cohn, D.J. Garrod, R.A. Rutenbar. and L.R. Carley,
for digital and analogue filters: PANDDA," in Proc. ECCTD, "KOAN/ANAGRAM II: new tools for device level analog place-
Brighton, U K , pp. 289-293, 1989. ment and routing," IEEE J. Solid-State Circuits, Vol. SC-26. No.
18. R.P. Sigg, A. Kaelin, A. Muralt, W.C. Black, Jr., and G.S. 26, pp. 330-342, 1991.
Moschytz, "Automated SC filter generation with a CMOS gate-
array-type filter chip," in Proc. IEEE ISCAS, Helsinki, pp.
1725-1728, 1988.
19. K. Theqvist, H . Alarotu, P. Ylisirnio, and H . Kaskelma, "Adap-
tive simulated annealing used in the synthesis of switched-
capacitor filters," in Pmc. IEEE ISCAS, Helsinki, pp. 1729-1731.
1988.
20. M . Negahban and D. Gajski, "Silicon compilation of switched-
capacitor networks," in Proc. European Design Automation Conf.,
Glasgow, pp. 164-168, 1990.
21. A. Barlow, K. Takasuka, Y. Nambu, T. Adachi, and J. Konno,
"An integrated switched capacitor filter design system," in Proc.
IEEE Custom Integrated Circuits Conference, Boston, 1989.
22. Bell Northern Research, SWCAP Tools Reference Manual, 1990.
23. M . Ismail and J. Franca, ed., Introduction to Analog Design Auto-
mation, Kluwer: Boston, Chap. 6, 1990.
24. R. W. Daniels. Approximation Methods for Electronic Filter
Design, McGraw-Hill: New York, 1974. Robert K . Henderson received the B.Sc. (Hons.) degree in elec-
25. R.K. Henderson, L i Ping, and J.I. Sewell, "Extended Remez tronics and electrical engineering with computing science and the
algorithms for filter amplitude and group delay approximation," Ph.D. degree in electrical engineering in 1986 and 1990 respective-
IEE Proc., Part G. Vol. 138, No. 3, pp. 289-300, 1991. ly, both from the University of Glasgow, Scotland.
26. G. Cortellazo and M R. Lightner, "Simultaneous design in both From 1989-1990 he was engaged at the University of Glasgow
magnitude and delay of IIR and FIR filters based on multiple as a research assistant. He is currently a research engineer at the
criterion optimisation," IEEE Trans. Acoustics Speech Signal Centre Suisse d'Electronique et de Microtechnique S. A . , Neuchatel,
Process., Vol. ASSP-32. No. 5, pp. 949-967, 1984. Switzerland. During his doctoral work he co-authored a compiler
27. C. Xuexiang, E. Sanchez-Sinencio, and R.L. Geiger, "Pole-zero for analogue filters called PANDDA. His research interests are in
pairing strategy for area and sensitivity reduction in SC filters." the areas of analog signal processing, numerical algorithms and
in Proc. IEEE ISCAS, pp. 690-611, 1986. computer-aided circuit and system design.
67
2
L i Ping received the B.S. degree from Nanjing Institute of Posts and John I . Sewell received the B.Sc. (Hons.) degree in Electrical Engi-
Telecommunications, China, in 1981, M.S. degree from Shanghai neering from the University of Durham in 1963 and the Ph.D. degree
Jiao long University, China, in 1984, and Ph.D. degree from Glasgow from the University of Newcastle upon Tyne in 1966.
University, U.K., in 1990. From 1966 to 1968 he was a Senior Research Associate at the
He was a research associate at Shanghai Institute of Railway Tech- University of Newcastle upon Tyne. During the period 1968 to 1985
nology from 1984 to 1986 and a Royal Society-British Telecom he held faculty positions at the University of Hull. In 1985 he was
Research Fellow at Glasgow University in 1987. From 1988 to 1990 appointed as Professor of Electronic Systems in the University of
he was a consultant and later an engineer with Wolfson Microelec- Glasgow and in 1990 he became Dean of Engineering. His research
tronics, Edinburgh, U.K. Since 1990 he has been a lecturer at the interests are concerned with the development of CAD design tools
Department of Electrical and Electronic Engineering, the University for front-end signal processing including Si and GaAs implementa-
of Melbourne, Australia. His research interests are in analog inte- tions, integrated filters and signal converters.
grated circuits and telecommunications. Dr. Sewell has served on numerous IEEE committees in the UK
and also on the IEE Professional Group on Circuit Theory. He has
been involved with organizing committees for a number of Confer-
ences, Colloquia and Workshops under the auspices of IEEE and
IEE. He is Fellow IEE and Senior Member IEEE.
68
2 41
PAPER 5 2
2 42
capacitor section.
C 0 2032
C 0.1184
(v) Finally, the filter is scaled in frequency by the 1 2
appropriate choice of transconductor and capacitor unit I Ml iH-
values. V:„/R
Ti2 "joe
1 758 ?1 319
.1 763 ?1.430 TlZ
For a given passive prototype ladder, various forms of shows a sixth-order elliptic bandpass prototype w i t h
the second-order matrix equation (eqn. 1) can be con- 0.1 d B passband ripple and 50 d B stopband attenuation.
structed, depending on the choice of variables used to In the ^-representation this is described by the matrices
f o r m the vector V. We use the terms ^-representation,
/-representation and ^/-representation to refer to the use 1 0 o\
of nodal voltages, branch currents and mixed variables, 6 =±l0 0 0 ,
respectively, i n V. The choice of representation is * V
0 0 1/
governed mainly by t w o factors. First, the order of the
matrices (which equals the number of variables i n V)
should be kept to a m i n i m u m so that the resulting active
circuit is canonical, i.e. has one integrator per pole of the c =
desired transfer function. Secondly, a representation
should be chosen which leads to the matrices C, C and T
and
being as sparse as possible, as the sparsity of these
matrices is reflected in the complexity of interconnect in
the resulting active circuit. +
Li L,
Each R L C ladder has a ' m i n i m u m inductance' and a
' m i n i m u m capacitance' version. Identical matrices are 1 I 1
r = + +
obtained if the K-representation is used f o r the former
and the /-representation f o r the latter. However, the
L 2
rT r
correct representation must be used f o r a particular 0 '-2 L
J L
4
prototype to ensure canonicity. Fig. 1 shows the -1
m i n i m u m inductance version of a fifth-order elliptic Another example is the asymmetric Chebyshev bandpass
^4
0 627
1 59^ 2(C-2C,
C 0116
2 L,8 508 2(C-2C,)
20.80)00481
20.80 ) 0 0461
1
Fig. 4 Another sixth-order elliptic bandpass RLC ladder
• , o\
0 - l , Fig. 5 First-order section using conventional transconductor
(Ci + c 4
2C
I-
and
,2Cj '2C
I L,t +
L 4
. 2C:
0 (4a-e)
2
CR2 2C
-1
0
300/3 300/3
100/5
L
II 1 600/5 i | "l
<
1
f 1 1
L100/5
100/5 1L 100/5
AGNDO-i
agorg-g
4.2 Left-inverse decomposition type 1 (LID1) serious problem for the sixth-order elliptic ladder of Fig.
The auxiliary variables are defined by 11, but for higher-order filters care should be taken with
the choice of prototype. Essentially this means placing
gX = sCV (9)
capacitors such that long chains of directly connected
where g is again a scaling factor w i t h the dimensions of inductors are avoided. Some passive prototypes also exist
conductance, and eqn. 9 is substituted into eqn. 1, giving whose F m a t r i x may not be inverted at all, due to the
determinant being zero. The only examples of such filters
gX = J - GV - s Ty (10) k n o w n to the authors are bandstop. A question yet to be
The L I D 1 design equations are obtained by rearranging investigated is to what extent an inverse matrix filter pre-
eqn. 9 and m u l t i p l y i n g eqn. 10 by the inverse of T : serves the low-passband sensitivity properties of its
passive ladder prototype.
g-'CV-- r'-x in)
l 4.3 Left-inverse decomposition type 2 (LID2)
gr~ x= -s- (12)
The auxiliary variables are defined by
F r o m eqns. 11 and 12, the main features of L I D 1 can be
co„x = y s (13)
deduced. First, both integrated terms (i.e. those contain-
ing the factor s~') are vectors, so only a single value of where io is a scaling factor w i t h the dimensions of angu-
0
transconductance is required in the active ladder. Sec- lar frequency, and eqn. 13 is substituted into eqn. I ,
ondly, both nor.integrated terms on the right-hand side giving
of eqn. 12 are generally asymmetric, so they must be real-
ised using unidirectional capacitivc coupling paths. In
m,cx = j - GV - s 'ry (14)
other words, the dependence of V u p o n X is not the same Rearranging eqn. 13 and m u l t i p l y i n g eqn. 14 by r 1
as the dependence of X upon V in these terms, so each gives the L I D 2 design equations
branch must be realised by a separate capacitor, rather
than symmetric branches being realised by a single capa- cu„r 'cx= r ~ ' J - r x
GV - s~'y (15)
citor connected between the nodes concerned. A single and
capacitor can be used for symmetric nonintegrated
branches arising f r o m off-diagonal terms on the left-hand y= t (16)
sides of eqns. 11 and 12, as long as first-order sections As i n the other left-inverse decomposition, only one value
w i t h high-impedance outputs are used. of transconductance is required, together w i t h unidirec-
Fig. 11 shows the L I D 1 realisation of the sixth-order tional nonintegrated paths. The distinguishing feature of
elliptic prototype of Fig. 2. This was obtained by substi- L I D 2 is the term f ' C , which gives the o p p o r t u n i t y to
t u t i n g the K-representation matrices, (eqn. 3a-e), i n t o obtain a relatively sparse transconductor ladder is a
eqns. 11 and 12 and translating each row of each equa- prototype can be used for which T = C. As shown in
tion i n t o a first-order section of the type shown i n Fig. 7. Section 2, this c o n d i t i o n is satisfied for a symmetric
A general feature of inverse m a t r i x decompositions is bandpass filter derived by the transformation o f each
that the number of components and the density of inter- component o f a lowpass ladder individually. Fig. 12
connect may be high in the transconductor ladder i f the shows the sixth-order bandpass elliptic transconductor
sparsity of T (or C) is lost upon inversion. This is not a ladder obtained by substituting eqns. 4<j-e into eqns. 15
2 1 3 2
0.225 0.225 0 294 0.294
_L
.g Z-Z- g Z-Z g Z-Z'
1* I* 1*
'1 u
ns— TT 6
0 707 0 515 0 777
Hki_
i* r
z" z"-g z' z~ -g Z* Z- -g
I T T 1 ±X 111T1111
0 045 0045 0 003 0 335 0060 0013 0 013 0 011 0 335 0 266 0003 0003 0 045 0 080 0 266
1 3 5 6 1 3 4 6 1 3 4 5
' , n O
4HH
0.636 0 917 0.836
HK
lljo0 116
0 045 | 0 045 0 042 0.003 1 0 003 0 003 0 042 0 045
\rf=>
F i g . 12 Elliptic bandpass ladder obtained by LI decomposition type 2
0.225 0.244
and 16 and implementing each row of each design equa-
tion w i t h the first-order section of Fig. 7.
g
4.4 Right-inverse decomposition (RID)
For the R I D , X is defined by
i
gX=s~ rV (17) "0.707 "0 515 "0.777
This is substituted into eqn. 1 to give r-ov ,o u
IN' IN'
V -
^
N V^ ^ /FCGQ
\JajfWy
400pF
39 00pF
39.00pF F C G D
FCGQD
2 80pF
65T0pF
V
2' "OUT
6510pF
34 35pF
I OUT" V,
FCGQD
FCGQD 5 b
7.60pF 34.35pF
T39.40pF
V 'OUT"
39.40pF
OUT OUT
300/3 300/3
900 u A 900 u A
900/5 p 900/5^
2 3
frequency, MHz
Fig. 16 Measured amplitude response of elliptic lowpass filler F i g . 17 Schematic of double-input folded cascode transconductor
V
I N " 'oiff
g/W
V I N - I O U I
n7l.9pF
63.25pF
H
birr* V*
b FCGQ
V* loin
32.1pF
89A0pF
FCGO
89 40pF
FCGQ10
7.95PF 7 95pF
75.25pF
OUT V
FCGQ
i|| |53.55PF
OUT" OUT'
F i g . 20 Fully differential elliptic bandpass filler [bias tines omitted)
CT 1 b [CT, 0
V ' O U T
V" 'OUT
1 0
69.35pF
7 55DF 7 55pF
H i ii—
67.15pF
OUT" V
FCGQ
FCGQ 'i ; rc R h
77.05pF
__66.8pF 5 b
7 55 pF 7 55pF
V" I
5 0
66.8pF
77.05pF
77.05pF
OUT OUT
F i g . 23 Fully differential Chebyshev bandpass filler tbias lines omitted)
\\
fir
20
40 60
60
-80
200 300 400 500 600 700 800
80 F i g . 28 Measured elliptic bandpass filler response for three values of
reference clock frequency
0 05 10 1.5 20
a 865 kHz
f r e q u e n c y , MHz [3 965 kHz
Fig. 24 Measured amplitude response of Chebyshev bandpass filter 0 1.065 MHz
6 Conclusions
Table 4 : C o m p a r i s o n o f d i f f e r e n t r e p r e s e n t a t i o n s o f RLC
prototype
Table 6: C o m p a r i s o n o f d i f f e r e n t decompositions of
second-order matrix equation
100 /££ Proc.-Circuits Devices Syst., Vol. 141, No. 2. April 1994
2 5 4
PAPER 54
2 5 5
711
Eroc. I.E.E.E. Interratuxal S^tposiurt an Circuits and Sy^tsrs, VbL.5, Icrrkn, Nfery 1994.
During the past decade several S C biquads have been No.C Total C C Spread
reported and used in practical applications. Most of the
TypeE 22 315.90 78.49
biquads proposed have been analysed individually and the
TypeF 22 706.13 191.48
investigations have been based on the characteristics of the
L G F Type 21 329.02 95.09
biquad itself. However the various filters proposed are
SSGI 27 351.27 94.78
intended for different applications, frequency ranges,
SSGII 27 351.27 94.78
sensitivity requirements, dynamic range etc., therefore it is
G-T 21 329.02 95.09
difficult to judge the suitability of the biquads. By using
M-S 21 328.95 95.16
X F I L T filter compiler! 1], a systematic comparison for the
Nagarai 27 213.24 47.11
realisation of S C filters using most popular S C biquads is
given in this paper based on filter system realisation. Table 2. Bandpass filter design results
c) Narrow bandpass filter. The filter response is given in
Although the software has the facility to design a SC system
with combined biquads. we restrict cur comparison here to
systems only composed of one kind of SC biquad, in order No.C Total C C Spread
to simplify the question and make the guidelines of TypeE 20 1150.40 333.81
selecting S C biquads more practical for the S C filter TypeF 20 1112.97 332.04
designer who does not have an SC filter compiler or only a L G F Type 20 1126.48 337.18
compiler with limited biquad structures. SSGI 25 1235.07 168.60
SSGII 25 1152.21 168.59
2. BIQUAD C O N F I G U R A T I O N S G-T 20 1126.48 337.18
M-S 20 1126.40 337.20
The biquads that we adopt here are: Nagaraj 27 5795.65 1050JS1
Type-E:Fleischer and Laker's E-type biquad in [2]. Table 3. Narrow band bandpass filter design results
Type-F:Fleischer and Laker's F-type biquad in [2].
FGL Type: Modified Fleischer and Laker's biquads[3]. d) Highpass Filter: The higtipm» filter frequency response
G-T Type: Gregorian and Temes's biquad presented in [4].
M-S Type: Martin and Sedra's biquad in [5]. No.C Total C 1 C Spread
SSGI Type and SSGII Type: Sanchez-Sinencio, Silva- TypeE 18 538.77 156.35
Martinez and Geiger's type-I and type-II biquads in [6]. TypeF 18 318.89 62.88
Nagaraj type: Nagaraj's biquad proposed in [7]. L G F Type 18 484.69 137.61
SSGI 23 366.23 66.53
3. F I L T E R R E A L I S A T I O N C O M P A R I S O N SSGII 23 366.23 66.53
G-T 17 522.23 123.55
a) Lowpass Filter : A 6th-order elliptic filter response is M-S 17 458.29 135.93
shown in Fig. 1(a). The Table 1 shows the realisation Nasaraj 21 158.32 27.58
statistics of different biquad cascades. Table 4.4th-order highpass filter realisation
2 5 6
712
e) Bancistop Filter: The bandstop filter response is shown SSGI and SSGII biquads have very good dynamic range
in Fig. 1(e) and the design results are given in Table 5. performance and they are best candidates for narrow band
filters in total capacitance and capacitance spread. They also
No. C Total C C Spread give best capacitance spreads in lowpass filter realisations.
Type E 29 220.97 29.24 However, SSGI and SSGII structures usually have slightly
TypeF 29 1309.69 361.34 high sensitivity to component change.
L G F Type 28 219.72 33.18 G-T biquad has lowest sensitivity and small non-ideal
SSGI 36 259.41 25.45 effect in the filter realisation.
SSGH 36 259.41 25.45 Generally, type-E, L G F , M-S, and G-T structures have
G-T 28 555.12 107.12 similar total capacitance and capacitance spread, except in
M-S 28 213.04 22.88 bandstop, where G-T biquad needs large total capacitance
Nasaraj 36 142.55 22.66 and capacitance spread. They also have lower sensitivity
Table 5. 8th-order Elliptic bandstop filter realisation and good dynamic range, and small non-ideal effect.
Type-F biquad is good in total capacitance and capacitance
4. S E N S I T I V I T Y AND DYNAMIC R A N G E spread in highpass case, but poor in wide bandpass, and
COMPARISON bandstop cases. It has fairly good sensitivity performance
and dynamic range, and small non-ideal effect
The following indices are used as global measures of system
sensitivity S(co) and dynamic range D(C0), respectively ACKNOWLEDGEMENT
REFERENCE
D(co) = -^£201og|H (cD)| M
Fig.1(a) Lowpass Filter Response Fig. 1(b) Wide Bandpass Filter Response
JO
40
x
a. -40
-to
200 joboo 00
1000 2000 3000 4000 5000 1000 7 000
Fraoeact (H«)
Frequency ( H t )
Fig. 1 (c) Narrow B andpass Filter Response Fig. 1 (d) Highpass Filter Response
70
ao
GO
00
0000
Prcqococy ( H O
SOBoo ooco
3
• JO
40
•00 -
60
7 200 400 800
Frequency ( H E )
aoo i ooo
Fig. 1(e) Bands top Filter Response Fig.2(a) Comparison of Lowpass Filter Sensitivities
100
20
* I0D mm * Baou
•t-O
2 OOO MO (OOO 1O0O 0000
<H»> ooo 3 WO
XI
Fig.2(b)Comparison of Wide Bandpass Filter Sensitivities Fig.2(c) Comparison of Narrow Bandpass Filter Sensitivities
too
000
V MO
600
M*fl«'*l
200
Fig.2(d) Comparison of Highpass Filter Sensitivities Fig.2(e) Comparison of Bandstop Filter Sensitivities
2O0
50
100 mot 4
Oj> ft M O >
£v -00
TVB*-l,U-t, ft LO
1
50
Jo IE J J two 80 3oo
714
Fig.3(a) Dynamic Range Index of Lowpass Filter Fig.3(b) Dynamic Range Index of Wide Bandpass Filter
T f p o - 6 4 Ty p o - F
SSGI4SSG
H .30
ISOO JOOG
1000 2000 3000 4000
Frequency (He)
Fig.3(c) Dynamic Range Index of Narrow Bandpass Fig.3(d) Dynamic Range Index of Highpass Filter
QT,a-8,aiOF
5
s
SSOI&SSQI ^ ^ ^ ^ aaoi.8saii.aa-T
Typo-EfiTypo-F
H 15-J
aa •
i 0
iio " *io " tto tfo 1060 u6o
1 0 7 0 0 1B0OO tSDOO 2 0 0 0 0 2 0 1 0 0 2 0 2 0 0 20300 Fw^xoej (Hi)
Frequency (Ha)
Fig.3(e) Dynamic Range Index of Bands top Filter Fig.4(a) Non-ideal Response of I owpass Filter
Fig.4(b)NcHi-ideal Response of Wide Bandpass Filter Fig.4(c) Non-ideal Response of Narrow Bandpass Filter
004111 T*OB-B.Q-S.a.T,at.OP
i0 I •003960
Fig.4(d) Non-ideal Response of Highpass Filter Fig.4(e) Non-ideal Response of Bands top Filter
2 5 9
Publication Page
[1] Table for the voltage transfer functions of single amplifier double ladder 261
feedback systems
[3] Table for the voltage transfer functions of single amplifier double ladder parallel
[15] Correction to Symbolic analysis for computer-aided circuit design - The 311
interpolative approach
[19] Network partition methods for use in symbolic analysis by interpolation 316
[25] Inclusion of amplifier finite gain and bandwidth in analysis of switched- 337
capacitor filters
Page
[28] Compact matrix scheme for use in computer analysis of switched capacitor 348
networks
[31] Symbolic analysis of ideal and non-ideal switched capacitor networks 355
[33] Efficient computer techniques for the exact analysis of all non-ideal effects of 360
switched-capacitor networks in the time domain
[49] Efficient sensitivity analysis for large non-ideal switched capacitor networks 372
[53] Efficient noise analysis methods for large non-ideal SC and SI circuits 376
2 6 1
PAPER 1
2 6 2
lL..:tion. Expressions are also given for point B is positioned between the Yi
Finally, it might be noted from Table 1 the product terms arising from the table. admittance and the Z impedance. The
3
that, if stabilisation and improvement in
It is well known that the burden of tedious circuit in Fig. 1 lacks symmetry, in that
sensitivity performance is to be effected by
calculations required to find the voltage the upper and lower ladders contain
use of Darlington-compounded transistors
transfer functions of networks having different numbers and arrangements of
(as in Fig. 3c). it suffices to do this only
certain regular forms may be much cased elements.
for transistor T . in each circuit, whether
or eliminated by the use of tabular forms When writing the suffixes to the ele-
i.n.i.c. or e.n.i.c., because the factor 1 — u,
in which the elements are set out in a ments, the following procedure is sug-
appears in four cases, whereas the factor
systematic array o l rows and columns. gested. Begin with the feedback elements
I — a, docs not appear at all, these being
1
One such array for the analysis of a single- in the upper ladder and denote these as
the significant factors.
ladder feedback network has been given Y>. Y . ) ' „ , where n is the order of the
•Breadboard' models of each e.n.i.c. by Aggarwal.'
4 ;
(Fig. 4) have been built and tested in the network. The Z elements should then be
In practice, it is often necessary to given the suffixes 1, 3, (2/i — 1), and
laboratory, and found to convert satis-
determine the voltage transfer ratio for finally the elententsconnectcd to earth in the
factorily. They have not yet been tested
exhaustively for stability and drift under lower ladder should be numbered as pre-
variable operating conditions and en-
vironment.
B. R. MYERS 21st April 1965
Department of Electrical Engineering
University of Notre Dame
Notre Dame, Ind., USA
(2n-5) 2n-3 I2n-1)
References
1 LARKY, A. 1.: 'Negative impedance con-
verters'. IRE Trans., 1957, CT-4, p. 124
2 HAKIM, s. s.: "Some new negative-impedance
convenors', Electronics Letters, ]y6i, 1, p. 9 l2n-4K2n-3) (2n-2)(2n-l)
elements subsequently omitted. (ft) Sets of four-component products ill The procedure is carried out until the
two Zs (mil in(/ }'\ last set for a particular M i l n e of n is
Derivation j)f table A l l Z . and Z . are multiplied by the sum reached, this having n factors in Zs and
(
It is possible to derive a recurrence of all the > between k and / and by the ii factors in Vs. The last term has only
formula to give the voltage transfer func- sum of the elements in the Y, row be- one factor.
tion front which the pattern for the table tween / ~ 1 and 2/i; i.e.
can be observed.
For a complete network with the first / - 1
series clement denoted as Z, and the last (5)
feedback element as >',, let the transfer z z,k
()', -i.)
(-/TI
ratio — EJE, be denoted as GO',/)
[/odd,/even,'/ < / ; if / > / , C ( / , / ) - 0], where k and / are odd. Concluding remarks
Initially assume a simple feedback system
A general pattern has been derived for
with one series input element and one (c) Sels of terms in three Zs and three Ys the terms which appear in the voltage
feedback element. Then, extending the
A l l Z Z,Z„, are multiplied by (i) the
k
transfer functions of single-amplifier
network towards the input three elements
sum of all Vs with suffixes between k and multiple-loop feedback systems. The Z >'
at a time (i.e. not disturbing the virtual
/, ( i i ) the sum of all Vs with suffixes table presented has been found exceed-
earth), the expression for the current at
between / and ///. and (iii) the sum of ingly convenient and labour-saving when
node B (Fig. 1) is given by the following
V,s, starting at m — 1, to 2//. Here applied to feedback networks intended for
recurrence relation:
k < 1 < m, and all k, I and in are odd, use as RC active filters.
Z r ' [ G ( 1,2/;) - G(3, lii)] beginning with 1, 3, 5, respectively; i.e.
2n- 1 2/1-3 . -5
+ Z - ' [ C ( 3 , In) - G(5,2//)] . (1) (r, + y, )
z z,z„
3
( l r l l ( ^ * — Il)
-5 l=i k - \ t </-/+!
(6)
in which, for example, G(5. 2»i) is the same
as G ( l , 2 / i — 4) with all the suffixes
increased by 4. Continuing the above (d) Sets of terms in four Zs and four Ys The authors have pleasure in acknow-
expression, it becomes Multiply Z^Z,Z Z m t by (ii the sum of ledging the interest of Prof. R. L . Russell.
Professor of Electrical Engineering in the
G ( l , 2//) = Z,(Y. - Y, - . . . + r , „ ) - M l - Z^Y, -- Y,,)] G(3, 2n) University of Newcastle upon Tyne. It
is also a pleasure to acknowledge the sug-
-r [ Z , ( K , -J- Y )] is G(S, In) -'r t Z , ( } ' 6 -f >' )] G(7, 2//) - f
6; ...
gestion by an IEE referee of the particular
+ lAO'<*,-» + Y ,„. .. ,]G(2n
L M3 L - 1,2//) (2) recurrence relations in eqns. 1 and 2. The
financial support of the Department of
The transfer function can now be written down by inspection.
Scientific and Industrial Research is also
G ( l , 2) = Z,Y, acknowledged.
G ( l , 4) = Z , ( y . - >',) -r [1 - Z I Y , + Y )] A G(3, A) A. G. J. HOLT 13th April 1965
1. I . S E W b L L
= z , ( y , -r r ) -r [ i -r z , ( r
t 2 - y )]z y,
2 3 3
Reference
+ 11 + z ( y s 4 J- y ) ] + z y,} - z,(y, +
u 5 y 4 5 )z y, . . .
5 (3)
1 AGGARWAL, G. K.'. 'ZY table for analysing a
ladder network', Froc. IEE, 1964, 111, p. 758
The following pattern emerges for higher all ys with suffixes between k and /»
values o f n: (ii) the sum of all ys with suffixes be-
There is a set of double product terms tween / and m, (iii) the sum of all Ks with
of the type ZY and a set of four compo- suffixes between /// and p, and (iv) the sum
nent products with two factors in Z and of y,s, starting with p — I , up t o ' 2 n .
two factors in Y etc. The last term has /< Here k < I < m < p ; all are odd.
factors in Z and n in Y. In Table 1 the 2/i-t 2/1-3 2/1-5 2/1-7
Zs and ys are so arranged that the voltage z,z,z z m p 'z (y, + y„
ratio - ( £ , / £ , ) can very easily be ob- p= 7 m= 5 1=1 k
j-k+l
tained. The forward transfer ratio £ / £ i a m- 1
P-l
can, o f course, be written down when s (y.+ y, {Y + r y r l r + 1 1 ) (7)
- ( £ , / £ , ) is known. -m+l
Table I
Use of the Table
ZY TABLE
The procedure for obtaining terms from-
the Table is as follows:
1
1 I 1
1 1 2 3
4
PAPER 3
320 I K L T . THAN'••ACTION'S O S C 1 K C 1 1 T THEORY SKl'TKMUEK
Final]}-, erii li of the follcwinp. procedures may be used I•> [icnerato ui the amplifier are infinite, its output impedance is zero, and a
from other structures that also exhibit the phenomenon (2). virtual f-tirt 11 p.'int exisu at the point A.
The eViucnl numbeiinp; is earned out as shown in Y\^. 1, where the
Procedure J: Split, any node of 5 in any way and insert a resistor
last subscript of each clement denotes the double ladder to which
and an inductor in parallel combination between the two halves
it belongs.
of the node. Repeat this procedure as many times as desired on the
resulting structure. DERIVATION OF T A B L E
Procedure 2: Replace any resistor (inductor) of S by two resistors A recurrence formula for this system may be derived to give the
(inductors) in scries. Connect an inductor (resistor) between the voltage transfer function, and the application of this to various
node joining the two resistors (inductors) and any other node of the orders systems leads to the pattern from which the table may be
structure. Repeat the procedure as many limes as desired. derived.
Using a notation similar to that previously employed, let the
That Procedure 1 produces structures which exhibit phenomenon first series element be denoted as Z , where a is the respective 10
1
(2) follows frurn Theorem 7 of Lee . That Procedure 2 also produces double ladder suffix, and. the last feedback element as Y/i; the
such structures follows from the observations that latter sullix is always unity, since the system is arranged to have the
1) Inequalities analogous to (7a) and (Sa) can be written for final feedback component in ladder 1. The transfer ratio —Ei/E t
the new structure. is denoted as 0'(i, / ) [i odd, / even, »'</; if »>/, G(i, / ) = 0]. From
2) The following selection for the column matrices B and C, inspection of the system it can be seen that for both parallel branches
the network increases symmetrically three elements per node per
branch until the node (2n — 1) in each ladder is reached. The final
section can be regarded as the common adding circuit well known in
analog computation.
Considering a node such as B, but treating the elements for a
B and C = general node in the symmetrical pattern, i.e., node 2r + l i the
equation for continuity of current is
+ (C2rtl — e i r ^ Z j V t l J (1)
and subsequent division of (7a) by (Sb), lead;- to a lower dividing by «*, and rearranging gives
bound on Isnun/smiaj similar to (9), the only dillerence being
that h, li, and f, (r r j , and r ) are replaced by sums of
Jf t
G(2r + 1, 2n), - G(2r+ 1,2/1), Y + 7 ,. ,
2r 2 2rt
1066 C O R R E S P O N D E N C E 327
iawt
W<.;(U-J)
LZZ3
T A B L E I
ZY TABLE
n 1 2 3 4
Series
Element Z.i Zn Z„ Zu Zn Zrhi - i),
1
Upper
Ladder Y„ Ya Kii Ya Yu
1
Lower
Ladder Y,H Ym Ym Ym Yin Y(U _ » (Jn - 1)1
1
Serin
Element Z» Z„ Zn Zi> Zn Zrta - n ,
2
Upper
Ladder Y„ Yn Ya Yu Y„
2
Lower
Ladder Y,u I'm K«i Ym Ym Y(tM - t) <t» - 1)1
2
U S E OF T A B L E Seta of Terms in Four Z'a and Three K's: AU Z,i Z.i Z«i are mul-
Following a procedure similar to that outlined in Holt and tiplied by
Sewell [1], the terms may be obtained from the table as follows. 1) all Z *
2) the sum of all W s with suffixes between q and v
Numerator 3) the sum of all Yi's with suffixes between v and w
Sets of Terms in Tim Z's and One Y: All Z<, and Z,i are mul- 4) the sum of Y/i's starting with w + 1 to In plus the sum of
l2
:
tiplied Ijy the. sir.n of the V/i's and Y s beginning with >' ui up
(>
Y/t's storting at p + I up to 2n; p,q,v,w are odd and q < v < w.
to Yi„ and + ,)•. up to }' „, respectively, w here i, j nre odd.
2 This is repeated with branch subscripts interchanged, and there
Seti of Tirms in Three Z's and Two Y's: All Z„ Z,, are multiplied is a finai set of terms in this group given as follows.
by AU Z \ Z,\ are multiplied by
r
2) the sum of all )Ys with suffixes lying between q and ii 2) the sum of all JYs with suffixes between p and q
3) the sum of Y.n'a starting at v + 1 to In plus the sum of 3) the sum of all Ki's with suffixes between v and u>
Y/i's starting at p + 1 up to 2«; 11,7,11 are odd and 17 < v. 4) the sum of Y;i's starting with q + 1 to 2n plus the sum of
Y/t's starting with w + 1 to 2n; p,q,v,w are odd and p < q,
This is repeated again with the branch subscripts interchanged. v < w.
2 6
This procedure is continued until tlie hist term having 2;i factors in F r o m the above it can be seen that for each increase in rank of the
Z and 2n — 1 in ) ' is reached. network another Z multiplier appears. T h e above expressions are
continued until the last term is reached; in general there are
Denominator
mn — (m — 1) terms in the numerator.
Sits of Terms in Zs: This is just the sum of all Z,, and Z,., where
i is odd. Denominator
Sets of Terms in Tuo Z's and One Y: All Z , Z,, are multiplied p 'Verms in Z'.
by '.he sum of all the Y's lying between suffixes p and q; this is also
repeated for branch two; p,q are odd and p < q.
Sets of Terms in Three Z's and Tuo Y's: All Z,i Z,\ Z are (12)
A
k, - 1
multiplied by * j odd
(L - 1 • • - M - 1 )
1) the sum of all }"s with suffixes between p and q
2) the sum of all IVs with suUixes between q and o. Terms in Y:
This is repeated for branch two; p,q,v are odd and p < q < v.
This procedure is continued until the last term which has n factors
in Z and 2 iu Y is reached.
£
B-L
£ £ k.z. (3 2 , . . )
0-3 4
OI-1 0 odd p odd
(/-.•••M-2)
G E N E R A L SYSTEMS
11 > 0 i • " > A „ _ ,
I t is now possible to consider the general case which is of order n
and rank m. The recurrence relations are derived in a similar manner,
the boundary conditions now being
21-p+l
.,)}]
1) G(l,2n)„ = G(l,2n) for all m (6)
2) (13)
g(2n - l,2n).
2) = Y, (7) Terms in £K»:
r
EE £
2«-L 0-2 i-2 /»-3 \
i.e., assume that branch 1 is the highest order ladder.
T, - I
t , odd
II
A-1 / ft-L
E
2J-JU +I
Y ,A
2
J
(9)
CONCLUSION
SEE
fc-I p-i
P CHJJ
*T-I
k i odd
^ P M ri ^I.fl )l S C I B ~J~ i^2Z ( 2 / + J
general terms using a digital computer, since a program would
involve only the arithmetic of the subscripts, and reduces error by
limiting the operations involved to multiplication and summation,
avoiding any redundant procedures.
2NI
2N, m
(10)
£ r + m
£ E n ACKNOWLEDGMENT
E K Y „ T + E
2(-*,+I
Y T L . +
E
E
at^-ID+I
E
rf-I
r
»
II K K F R K N C t S
[1) A. G . .7. ITfilt flntl J . I . S"\vell, "Table fur Uie voltapc transfer {unctions of sinnle
amplifier uf.nhl'.-lii'litt'r fn-iihai-k m-slc-ins." h:irrir ,r l.tlt.. vol. 1. p. 7f). tuiiS.
when b=c K =0 lllt
\2) (I. K \--'-:iru:il. " / . ) ' Uiil.: I..r ;II.»UMMS ix l:i.|.lrr n r l u u r t . " /'."C. //•-'/;.
1 on 111. p. 7JS, i'Mi.
2 6 8
PAPER 4
2 6 9
I.E.E. C o n f e r e n c e P u b l i c a t i o n rio. 51 , pp. 515-526
1
^ r. — ^ ->rp^- ~ 7 T ' " n n T - rv; ~<-_- ^ T" ^ -<
1 „ Introduction
I t i s w e l l kr.ov;n i 1 , 2j t h a t t h e g r e a t e s t s o u r c e o f e r r o r i r . a n a l y s i s
by t h e n o d a l a d m i t t a n c e m a t r i x a p p r o a c h o c c u r s i n t h e e v a l u a t i o n o f t h e
n e c e s s a r y c o f a c t o r s and d e t e r m i n a n t s . I n p a r t i c u l a r , i f numbers of l a r g e
m a g n i t u d e a r e m u l t i p l i e d and r o u n d - o f f o c c u r s i n t h e c o m p u t e r ; t h e n when
t h e s e numbers, e s p e c i a l l y i f they a r e n e a r l y e q u a l , a r e s u b s e q u e n t l y sub-
t r a c t e d from e a c h o t h e r , an e r r o r o f g r o s s m a g n i t u d e c a n r e s u l t . One
b r u t e f o r c e method o f r e d u c i n g t h i s e r r o r i s t o r e s o r t t o m u l t i p l e l e n g t h
a r i t h m e t i c on t h e c o m p u t e r , t h i s u n f o r t u n a t e l y c a n i n v o l v e c o m p l e x mach-
i n e c o d e p r o g r a m m i n g and a l a r g e m a c h i n e t o e x e c u t e t h e r e s u l t i n g p r o g -
ram.
1 .1 Con.iacture
Any c h a r a c t e r i s t i c f u n c t i o n ( b e i t a t r a n s f e r r a t i o , d r i v i n g p o i n t
o r t r a n s f e r i m m i t t a n c e ) d e s c r i b i n g a l i n e a r p a s s i v e s y s t e m w i t h no r . h .
p l a n e z e r o s o r a n i d e a l f e e d b a c k s y s t e m * i n w h i c h t h e r e i s no p o s i t i v e
f e e d b a c k w i l l c o n s i s t of a r a t i o of p o l y n o m i a l s , i n the f r e q u e n c y vari-
able, with p o s i t i v e c o e f f i c i e n t s only. As t h e s e c o e f f i c i e n t s a r e p o s i t i v e
any s u b t r a c t i v e s t e p i n the d e r i v a t i o n of t h e s e i s b o t h r e d u n d a n t and
e r r o r i n t r o d u c i n g , not o n l y i s the s t e p of s u b t r a c t i o n redundant but i t
i m p l i e s r e d u n d a n c y of p r e v i o u s m u l t i p l i c a t i o n and a d d i t i o n s t e p s . Hence,
f o r t h e s y s t e m s u r d e r c o n s i d e r a t i o n , i t s h o u l d be p o s s i b l e t o d e r i v e a n a l -
y s i s techniques which e l i m i n a t e these redundant s t e p s , thus improving both
a c c u r a c y and s p e e d o f c o m p u t a t i o n .
/1.2
* An i d e a l f e e d b a c k s y s t e m i s one i n w h i c h t h e a c t i v e d e v i c e s c a n be
represented as a simple c o n s t r a i n t , One e x a m p l e o f t h i s i s t h e i d e a l
v o l t a g e o p e r a t i o n a l a m p l i f i e r , w h i c h c a n be r e p r e s e n t e d a s a v o l t a g e
constraint.
J . x . S e w e i i i s a t the Department o f E l e c t r o n i c E n g i n e e r i n g , U n i v e r s i t y
of H u l l .
C. I-.'i.jhtingale i s a t t h e D e p a r t m e n t o f E l e c t r i c a l E n g i n e e r i n g , U n i v e r s i t y
of N e w c a s t l e upon T y n e .
2.
';. 2 J.-:.-lir'er.f'-cior. o f t h e C s r - q c t - y e
? o r p a s s i v e l a d d e r s , t e c h n i q u e s v.'hich u t i l i s e t h e sfsove p r i r . c i t l e
r a v e been i n e x i s t e r . e e f o r some c o n s i d e r a b l e t i r . e \y sr.d v a r i a n t s en t h e
method a r e v : e l l knov.r. [c\ . Hov;ever, f o r many a c t i v e retv.-orks, l i t t l e
v.-orr: a l o n g t h e s e l i n e s has beer, done, p r o b a b l y t h e . t a i n r e a s o n be i n s cue
t o added c o m p l e x i t y v.'hich t h e a c t i v e d e v i c e i n t r o d u c e s . Some work has
been c o m p l e t e d a s s u m i n g t h e d e v i c e s a r e i d e a l j_'3,6j and y i e l d e n c o u r a g i n g
results.
U n f o r t u n a t e l y i f t h e a c t i v e d e v i c e c a n n o t be assumed i d e a l and be
r e p r e s e n t e d by some s i m p l e c o n s t r a i n t , t h e s e i d e a s c a n n o t be used and
some g e n e r a l method o f a n a l y s i s has been used. Thus a l l t h e d i s a d v a n t -
ages o f a g e n e r a l method r e t u r n . I f , on t h e o t h e r hand, t h e s y s t e m can
be d i v i d e d i n t o two s u b - s y s t e m s , one o f w h i c h r e p r e s e n t s t h e s y s t e m when
a l l d e v i c e s a r e assumed i d e a l , and t h e o t h e r r e p r e s e n t i n g t h e c o r r e c t i o n
r e q u i r e d t o produce t h e a c t u a l s y s t e m , t h e n a p o s s i b l e s o l u t i o n i s
obtained.
Suppose t h e system i s r e p r e s e n t e d b y
Y Y Y (1)
= I © c
where Y i s t h e a c t u a l system m a t r i x
Y j i s t h e i d e a l system m a t r i x
Y i s the correction matrix.
c
The r e q u i r e d r e s p o n s e , c o u l d i n g e n e r a l be f o u n d f r o m e v a l u a t i n g
a p p r o p r i a t e c o f a c t o r s o f Y. The r e s p o n s e when t h e d e v i c e s a r e assumed
i d e a l can be d e t e r m i n e d f r o m Y j by one r f t h e t e c h n i q u e s r e f e r r e d t o
already. The a c t u a l r e s p o n s e can be f o u n d b y d e t e r m i n i n g t h e e f f e c t
w h i c h Y p r o d u c e s on t h e s y s t e m and t h e n c o m b i n i n g t h i s w i t h t h e r e s -
c
ponse due t o Y j .
I t i s t h e r e f o r e p o s s i b l e t o observe t h e s i g n i f i c a n c e o f t h e device
n o n - i d e a l i t i e s and p a r a s i t i c s , and a t t h e same t i m e u t i l i s e a h i g h - s p e e d ,
h i g h l y a c c u r a t e p r o c e d u r e f o r t h e main t a s k o f a n a l y s i s .
3ecause o f t h e way i n w h i c h t h e e f f e c t o f Y i s e v a l u a t e d i t i s n o t
c
I n the m a j o r i t y o f p r a c t i c a l s i t u a t i o n s i t i s o f g r e a t advantage i f
Y c becomes t h e n u l l m a t r i x i . e .
Y = Yj
Thus i t i s p o s s i b l e t o i n c l u d e t h e m i n i m i s a t i o n o f Yc o r some r e l -
a t e d f u n c t i o n , i n the general m i n i m i s a t i o n procedure r e q u i r e d i n the
synthesis.
A s e c o n d a r y f e a t u r e o f t h e method, i s concerned w i t h t h e a n a l y s i s o
s t r u c t u r e s , w h i c h d e p a r t f r o m n o r m a l t o p o l o g i e s , w i t h o u t r e c o u r s e t o gen-
e r a l .T.-Vthodo. ?o; i n r . t x n c e , i n t h e case o f b r i d g i n g e l e m e n t s i n p a s s i v e
l a d d e r , ; n n i y ; i i n i t v:c be r.occc"-.ry t o r e s o r t t o come goner.nl t c c h n i q u
ana f o r f e i t t h e b e n e f i t s o f c o n t i n u a n t o r cuu.ulant methods. V.'ith t h i s
t e c . r . i c u e t h e a r i d r i n g 5lc-r.5r.t3 ccv.id be a s 3 c c i ' . t e d v;it'r. t h e c o r r e c t ! ; . ' ,
m a t r i x .v.*.; tr.e r e m a i n d e r o f t h e c i r c u i t a n a l y s e d by r e c u r r e n c e methods.
I t i s a l s o p o s s i b l e 1 0 a n a l y s e r . u l t i t s r - i r . a i a m p l i f i e r 3 t r u e t u r e s by
t r . i s t e c h n i q u e , t h e p o s i t i v e f e e d b a c k b e i n g a s s u r e d as a c o r r e c t i o n tor::..
I n t h e case 0 : a d e v i c e w h i c h can be r e d u c e d t o s o r e s i r . p l e c o n s t -
r a i n t i t i s p o s s i b l e t o e l i m i n a t e a colur.r. o r row f r o m the a c r . i t t a n c e
m a t r i x , f o r one v a r i a b l e i s a d i r e c t f u n c t i o n o f a n o t h e r . The t e r r s
produced i n t h e e l i m i n a t i o n o f t h e i n i t i a l rov; c r column can he s e p a r a t e d
i n t o Y as c o r r e c t i o n e l e m e n t s .
c These a r e termed c o r r e c t i o n e l e m e n t s
s i n c e i f i d e a l o p e r a t i o n a l a m p l i f i e r s were b e i n g used t h e y w o u l d be z e r o .
U n f o r t u n a t e l y i t i s n o t a l w a y s p o s s i b l e t o reduce a d e v i c e t o s i m p l e
current or voltage constraints without considerable d i f f i c u l t y ( i . e . i t
i s n e c e s s a r y t o produce an e q u i v a l e n t c i r c u i t ) . Even i n t h e case o f a
t r a n s i s t o r , v e r y complex e q u i v a l e n t c i r c u i t s a r e proposed w h i c h r e q u i r e
a c o m p u t e r p r o g r a m o f f a i r c o m p l e x i t y t o d e t e r m i n e them.
As i t i s p o s s i b l e t o produce a Y m a t r i x f o r any l i n e a r d e v i c e w h i c h
i s n o n - i d e a l , e i t h e r by measurement o r by m a t h e m a t i c a l means v i a t h e
m a n u f a c t u r e r s d a t a , i t w i l l be assumed t h a t t h i s r e p r e s e n t a t i o n i s s u f f i c -
ient. The terms o f such an a d m i t t a n c e m a t r i x w i l l i n g e n e r a l be r a t i o n a l
functions. I t i s t h u s n e c e s s a r y t o assume t h a t a d m i t t a n c e m a t r i c e s t o be
c o n s i d e r e d have r a t i o n a l f u n c t i o n s as t h e i r e l e m e n t s . Obviously the
p a s s i v e e l e m e n t s w i l l n e v e r c o n t r i b u t e any t e r m l a r g e r t h a n a 2nd o r d e r
p o l y n o m i a l o v e r a 1 s t o r d e r one.
1
i.e. Y = T Y T 0
T I (2)
. C
: Y
O J
The method i s n o t r e s t r i c t e d t o a m p l i f i e r s , b u t may be used f o r
s y s t e m s c o n t a i n i n g X.I.Cs., g y r a t o r s o r any l i n e a r n e t w o r k d e v i c e . How-
e v e r , as most d e v i c e s a r e f a b r i c a t e d f r o m o p e r a t i o n a l a m p l i f i e r s i t s
m a j o r use w i l l be c o n c e r n e d w i t h such a c t i v e e l e m e n t s .
3. The Development o f t h e C o r r e c t i o n ? u n c t i o n X a t r i x
I t i s now n e c e s s a r y t o d e t e r m i n e t h e e f f e c t o f t h e t e r m s o f t h e
c o r r e c t i o n m a t r i x Y on t h e system, r e s p o n s e f u n c t i o n , o r c h a r a c t e r i s t i c
c
/3.1
27 2
_
l - i i - _ V - ^ * c i ^ -"
s
t h i s gives : = f_ + f (:.,
The d e t e r m i n a t i o n o f I T ar.d f„ i s e a s i l y d e m o n s t r a t e d by c o n s i d e r i n g
t h e method o f e v a l u a t i n g e q u a t i o n 3, w h i c h i s a g e n e r a l s t a t e m e n t o f a
s p e c i f i c p r o b l e m f o r w h i c h the s o l u t i o n i s known [ 6 j . A s i n p i e p r o o f o f
the g e n e r a l technique f o l l o w s .
Vhooy..-.
D e :
I'iore e x p l i c i t l y l e t ( x ? ) , (jg^Op (z^)j1^he " i r s t , second and
n
t h i r d colurr.nn o f any o f tho n m a t r i c e s X', ...X , whore p nay o r may not
equal c and/or r .
Then t h e t h e o r e m states,
n
=
l ? f , l £lr*'| <5>
( a l l p o s s i b l e p, q, r f o r p, q, r = 1, ...n)
The r e s u l t i s e a s i l y seen u s i n g t h e t e n s o r n o t a t i o n . L e t 2 i j k be
t h e f u n c t i o n o f i j k such t h a t f o r i p j = k £ikj = - 1 a c c o r d i n g t o
w h e t h e r i , k , _ j i s an even o r odd p e r m u t a t i o n o f 1,2,3, ( O t h e r w i s e £ijk = o ) .
Then by a w e l l xnown f o r m u l a
jxj = eijk x .. x .
2 x 3 k
Thus ! £ x r'
oecom.es
x
li x c
2 j ' 3k tit
e x a m i n a t i o n o f t h e e q u a t i o n ( 5 ; f o r r , s , t = 1 , ...n shows ' i t t o be equiv-
a l e n t t o t h e r i g h t hand s i d e o f ( 5 / which p r o v e s t h e theorem.
2 7 3
where X i s a d e t e r m i n a n t f o r m e d f r o m terms o f t h e t y p e
y y y y
i j ?i,?j - ?i,j i,?j.
?or a matrix the formula i s
n-1,n-1
,(7)
n-2
^ n-2
4 <n-
u
-n-1,n-1 pj-.-j L Aa i n _i j n _ i ^
^ n-2
i s an n-1 x n-1 m a t r i x f o r m e d by t a k i n g a l l t h e p o s s i b l e c o m b i n -
a t i o n s o f rov;s o r columns f r o m 3, b a c k - s u b s t i t u t i n g t h e s e i n A and d e v e l -
o p i n g new terms o f t h e f o r m
A3 / ^ab,s v/here p = 1
^_ ^
s=
n-2 ..—c
t o r m u i a e remain, ur.cr.s.
. • _ - -
1
.--3 w i _ _ oo a p p r e c i : »ou .re:.. \.-/ _„ i s n e c e s s a r y t o c o r s i c o r c. t o :
- —o
I n most n e t w o r k s i t i s h i g h l y
n e c t e d t o an a c t i v e d e v i c e , t h u s i t i s most p r o b a b l e t h a t t h e r e w i l l be a
number o f rows and columns o f Y w h i c h o n l y c o n t a i n z e r o . As t h e p r e s e n t
c
p o l i c y i s t o back s u b s t i t u t e columns, i t o u g h t t o be p o s s i b l e t o r e d u c e
the number o f c o m b i n a t i o n s c o n s i d e r e d by e l i m i n a t i n g t h e z e r o columns
f r o m Y , t h i s a l s o r e d u c e s s t o r a g e space. The o r i g i n a l column n u m b e r i n g
c
has o f c o u r s e t o be r e t a i n e d i n a c o u n t e r .
3ecause t h e number o f c o m b i n a t i o n s i s r e d u c e d a d i f f e r e n t d i v i s o r i s
r e q u i r e d t h a n t h a t i n e q u a t i o n ( a ) . O b v i o u s l y no t e r m i n a z e r o column
can be used as a p i v o t . I f nc i s t h e number o f n o n - z e r o columns, back
s u b s t i t u t i o n o f a reduced Y m a t r i x w i l l y i e l d
nc
, nc
~r combinations. Thus t h e number o f excess c o m b i n a t i o n s
r=0
produced v/hen z e r o columns a r e b a c k - s u b s t i t u t e d i s
n nc
n <w nc
Thus i f o n l y columns w h i c h a r e n o t z e r o a r e b a c k - s u b s t i t u t e d t h e n t h e
m a t r i x ( b ) becomes
T ; U + ; L
^ n - l ,n-1 ~ ^ P1,?J V.PJ £, A3
^•1 II
c l.U
;c n
r=0
r.-2
2 r-
?
r=0
v cr u Z C -.'J c _ ^. — ev o ,— 1. "0 G C r.'J. Z c Z ~ L/ j-. C . Z~* C ^ c " "J \ l ^ 1. C '* Z "' "
+
; o r . s i c e r : = A, + A 2 ~ ••••
k = 1
m m-1
X / 1, ^ LVi * ^ A. j + V . 3 ... W
d d 3 ^
j = i +1 i = 1
VJ's a r e •.•/eights c a l c u l a t e d as shown l a t e r .
y
hj_ i s any m a t r i x X ± r e d u c e d by one o r d e r .
A-j_j i s any m a t r i x X-j_ w i t h a p i v o t column o r row f r o m X j s u b s t i t u t e d
and r e d u c e d by one o r d e r .
3^ i s any m a t r i x r e d u c e d by one o r d e r .
3 j i i s any m a t r i x °X^ w i t h a p i v o t column o r row f r o m 'A± s u b s t i t u t e d
and r e d u c e d by one o r d e r .
The v a l i d i t y o f t h i s r u l e may be e s t a b l i s h e d i n t h e f o l l o w i n g way,
t a k e t h e case o f m = 2, i f t h e e l e m e n t s o f t h e Y' m a t r i x a r e examined i t
i s f o u n d t h a t t h e y can be w r i t t e n down as t h e s e r i e s :
°.
Y'ii jj = (x, x 2 t i,j>- zjti.pj, x^pi.j^r
r = 0
c
+
(x^pi.pj; x 1 i.i,j)- x ii,pjjx 1 1 (pj, jj)r
r = 0
q
•- 7 ^ (x 2 i?i,?:jx 1 \i, j)- x 2 (i,pjj x, ? i , j P r
4
r = 0
c
(>: (,?i.?ij
2 x 2 L i, j ; - x 2 (i,pj; x tpi.^r
r = 0
n-2
n-2
r = 0
I f t h e terms a r e c o l l e c t e d , t h e e x p r e s s i o n can t h e n be s i m p l y w r i t t e n
:
i n t e r m s o f t h e .. and 3 m a t r i c e s above. ? o r m > 2 t h e p r o c e d u r e i s r e p -
eated over a l l p a i r s o f m a t r i c e s .
T 5 x 7 3 1
. •: T Vl2 "3 21 c 2 )
I f t h e zero hack s u b s t i t u t i o n i s o m i t t e d , one s e t o f terms o f t h e
t y p e A-. w i l l be m i s s i n g ; now i n g e n e r a l t h e r e a r e
n - 1
of these terms.
ience:
W. = 1 -
"> n-1
t c
r=0 r
£ S
Once t h e m a t r i x [A + 3 j _ - , -| h
n been s y n t h e s i s e d , f u r t h e r r e d u c -
t i o n i s a c h i e v e d by s t a n d a r d p i v o t a l c o n d e n s a t i o n on t h e s i n g l e m a t r i x .
4. Stor?:;"G R e q u i r e m e n t s and Pro-rani L i m i t a t i o n s
j'or t h e two main methods c o n s i d e r e d i t i s n e c e s c a r y to determine the
s t o r a g e r e q u i r e m e n t s f o r each method.
4.1 Total Back-Substitution
In a l l c a l c u l a t i o n s i t i s assumed t h a t :
Y - li x ); x ( n p + 2 + nd). Array Y = network m a t r i x
(np = max. o r d e r o fn u m e r a t o r Y)
(nd = max. o r d e r o fd e n o m i n a t o r Y)
2 - (N - r.a) x (K - na) x (npc + 2 + n d c ) . A r r a y C = c o r r e c t i o n
matrix. (c)
npc = max. order of numerator C
ncc = max. order of denominator C
Y1 - (X - na) x (X - na) x ( n p n + 2 + n d n ) . A r r a y Y1 = ideal system
matrix. (lo)
npn = max. o r d e r o f n u m e r a t o r Y1
ndn = max. o r d e r o f d e n o m i n a t o r Y1
The w o r k i n g a r r a y Y".< has t o c o n t a i n t h e terms p r o d u c e d by substitu-
tion,
Y'.T - (X - na) x (X - na) x (npw + 2 + ndw)
where now and new a r e computed as follows:
npw.. = 2(npw.. ^ -t- ndw_. _^ )
ndw = 4 r.dw_.._ (
+ [ n d W N _ _ 2 ] [2TNC
na (N -na) - 3])
..-ua
i = 1,2, (N - r . £ - 3)
4.2 Xiniral 3sc>-3vb3tituticr.
The two a r r a y s G, Y1 a r e as i n e q u a t i o n s ( s ) , (lO).
Y1 - w o r k i n g a r r a y f o r t n e 1 node r e d u c t i o n .
Yi - (X-na) x (X-na) x ( 2 b p o + n d c ] + 4 ndc + 2)
YV," - (X-na) x (X-na) x (npw + 2 4- ndw)
where npw and ndw a r e computed as follows:
npa^ = max. J r.pa. 1 + 4ncc; 2 ( n p c + ndc) + nda^ ,
nda = nda^_, + 4 ndc
i = 1,2,3
r.pa = 2 ( n u n -r ndn/' nda = 4 ndn
o o
npvK = 2(npv.' ._, T ndw^ ^ )
ndw,. = 4 ndw._,
i = 1,2...(X - na - 3)
where now = noa„, ndw = nda_,
' o - y o 3
YC - 4 ( 2 -r n?i._ , + nt>w , + ndi . + ndw,.
v ,) v
* A-na-4 ' X-na-3 X-na-4 iM-na-3
4.3 C o r — r i s e n of Storage R"uirem%rts
Total Back-Substitution Xinimal 3ack-Substitution
X = 4, np = 1, npn = 1
na = 1, r.d = 0, ndn = 0 1 ,1 SO 1 ,1 82
.." = 5, np = 1 , npn = 1
na = 1, nd = 0, ndn = 0 S,"'0o 6,536
X = 5, up = 1 , npn = 1
na = 1, nd = 3, ndn = 0 61,500 33,666
X::tra s t o r a g e w i l l be n e c e s s a r y i n b o t h methods, f o r w o r k i n g areas
and polynomial handling.
I t can be seen t h a t as t h e number o f nodes i n c r e a s e t h e demand or.
s t o r e g e space becomes s e v e r e . This i s p a r t i c u l a r l y evident i n the t o t a l
b a c k - s u b s t i t u t i 0 n m e t h o d.
5. X- rrlr: -yr.:. Correction Tunction Xatrix
5.1 C o r r e c t i o n 7-jr.cticr. X a t r i "
I n most c i r c u m s t a n c e s i t i s u s u a l l y a d v i s a b l e t o d e r i v e c o r r e c t i o n
2 7 8
6. Analysis of he I d e a l System
The method employed here depends l a r g e l y upon t h e system u n d e r i n v e s -
t i g a t i o n , t h e r e q u i r e m e n t s f o r such a method have been c i t e d e a r l i e r .
A system w h i c h has been o f c o n s i d e r a b l e i n t e r e s t r e c e n t l y i s t h e
m u l t i - l o o p f e e d b a c k one j_7j , t h e ZY methods j,5,o] a r e a p p l i c a b l e t o t h i s
class of networks. A program A L A u t i l i s i n g t h e s e t e c h n i q u e s has been
w r i t t e n i n A l g o l and i s s u i t a b l e f o r i n c l u s i o n i n a c o r r e c t i o n f u n c t i o n
analysis.
The d e v e l o p m e n t o f o t h e r a n a l y s i s p r o c e d u r e s a p p l i c a b l e i n a c o r r e c t -
i o n f u n c t i o n a p p r o a c h i s t h e s u b j e c t o f f u r t h e r work.
7. Use o f o r o r r a m
As t h e m a j o r l i m i t a t i o n i s du- t o s t o r a g e space a v a i l a b l e on t h e
c o m p u t e r , a b l o c k s t r u c t u r e i s employed t o m i n i m i s e t h e s t o r a g e demanded
a t any one t i m e by t h e program. Sven so, a KD?9 w i t h 16K s t o r e w i l l o n l y
p e r m i t t h i s c o m p u t a t i o n on c i r c u i t s w i t h a maximum o f 5 nodes. I t may be
p o s s i b l e t o i m p r o v e t h e s t o r a g e demands by e m p l o y i n g m a g n e t i c tape s t o r a g e
and s e g m e n t a t i o n , t h i s u n f o r t u n a t e l y w i l l t e n d t o s l o w down t h e c o m p u t a t -
ion.
Of t h e two programs used, i t i s o b v i o u s t h a t t h e one u s i n g t h e A i n i -
mal B a c k - S u b s t i t u t i o n Method i s s u p e r i o r b o t h i n t i m e t a k e n f o r e x e c u t i o n
and f o r s t o r a g e r e q u i r e m e n t s . A l l examples have t h e r e f o r e been computed
w i t h t h i s program.
An example o f a 2nd o r d e r m u l t i - l o o p S u t t e r w o r t h c i r c u i t ( P i g . 1 } v.a.s
examined whan t h o a c t i v e d e v i c e ( i d e a l l y an i n f i n i t e g a i n v o l t a g e a m p l i f -
i e r ) d e p a r t s from i d e a l . The i d e a l a n a l y s i s u s i n g program A L A took 4 sees,
o f c o m p u t i n g t i m e and t h e c o r r e c t i o n f u n c t i o n a p p r o x . 2 mins. u s i n g t h e
.-.Igol c o m p i l e r . The g r a p h s , F i g u r e 2, show t h e i d e a l r e s p o n s e and two
c o r r e c t e d responses f o r d i f f e r e n t a m p l i f i e r s .
6. Conclusions
A g e n e r a l technique f o r t a k i n g account o f n o n - i d e a l i t i e s o f a c t i v e
d e v i c e s i n n e t w o r k a n a l y s i s has been p r e s e n t e d . T h i s e n a b l e s t h e use o f
f a s t and a c c u r a t e r e c u r s i v e p r o c e d u r e s i n t h e e v a l u a t i o n o f an i d e a l r e s -
ponse and c o r r e c t i o n terms a r e s u p p l i e d s u b s e q u e n t l y . This technique pro-
v i d e s a u s e f u l a n a l y s i s method f o r i t e r a t i v e s y n t h e s i s where speed i s
vital. F o r t h e c o r r e c t i o n terms, w h i c h a r e e s s e n t i a l l y n u m e r i c a l and
r e q u i r e a I o n g e r t i m e f o r e v a l u a t i o n , i t may be p o s s i b l e t o compute t h e s e ,
27 9
r.imi:
•.'•:_£••.•: c: re
C e r t a i n l y t h e e v a l u a t i o n o f t i e i d e a l response f u n c t i o n i s v e r y a c c u r a t e
as no r e d u n d a n t a l g e b r a i c s t e p s a r e p e r f o r m e d . There w i l l bo r o u n d - o f f
e r r o r due t o m u l t i p l i c a t i o n b u t t h i s i s n o t a c c e n t u a t e d by s u b s e c u e n t
subtractions. I t i s never p o s s i b l e t o c o m p l e t e l y e l i m i n a t e c o m p u t a t i o n a l
errors. The e v a l u a t i o n o f t h e c o r r e c t i o n t e r m s i s s u b j e c t t o t h e c o n v e n -
t i o n a l s o u r c e s o f e r r o r , b u t as t h e s e terms a r e u s u a l l y o f s m a l l m a g n i t u d e
and m i n i m i s a t i o n o f them i s u s u a l l y i n t e n d e d , any e r r o r s h e r e a r e o f s e c -
ondary importance.
CM
ft
V3
<5 O i
•5G
R
Vi
19 Ul
H
ii
Ul
I
6
VI
to
HI
V)
1
Ul «5
5 s;
»
A 7
- j
0» o o
eg
9 —
<3 =VWv
2 °^ f*
UJ
ST Nttf-9 i
2 81
PAPER 5
R e p r i n t e d f r o m The Radio and Electronic Engineer. V o l . 3 7 . N o . 4, A p r i l 1 9 6 9 .
U.D.C. 621.372.63
I o
1 2 3 4
5 1 0 0
>'n -yi2
2 - } ' 2 \ 0 0
t Department of Electronic Engineering, The University of Fig. 2. Admittance matrix for the 5-terminal active device
Hull. shown in Fig. 1.
J. I. S E W E L L and F . W . STEPHENSON
for practical and operational cases, the parameter K malization is suitably performed. The voltage opera-
is a constant which tends to infinity. tional amplifier is therefore a practical proposition
Included in these tables are the matrices for the and in fact is the most common type at the present
common amplifier configurations, which are obtained time.
by eliminating certain nodes. The realization of extremely high input admittances
An interesting point to note is that the single is not particularly difficult with a common-base
operational amplifiers are equivalent when included circuit or with suitable feedback. But zero output
in a network. This can be shown by simple trans- admittances present a somewhat more difficult
formation.' problem, although circuits approaching this (output
resistances of the order of 50 MQ pr even 1000 MC2)
The hybrid amplifiers are given in Tables 5 to 7. 2
are not unknown. This makes the current operational
If there is 1 : 1 conversion from voltage to current, amplifier quite feasible, although, perhaps, rather
then A = C. more complex than the above.
3.3. A Comparison of the Different Types of Amplifiers The transimpedance operational type is probably
the most suitable of all, i.e. zero input impedance
In the physical realization of the various opera-
and zero output impedance, both of which can be
tional types it is obvious that some will be easier to
realized fairly accurately and simply.
construct than others. With modern techniques using
f.e.t.s it is possible to obtain extremely high input The amplifier which presents the most practical
resistances ( 1 0 11
ft) and, for a limited frequency problems in construction is the transadmittance one,
range, v , - » 0 . The realization of very low output
t
which requires high input resistance and high output
impedances ( < 100 fl) is common practice and appli- resistance. Although realization of this type is not at
cation of suitable feedback can produce very low all beyond present techniques.
values indeed. Thus y -* oo, at least in comparison
22 With regard to usage, if the forward gain parameter
with surrounding network elements, provided denor- (voltage, current, impedance, admittance) of an
1 2 3 4 1 2 3 . 4
1 0 0 0 0 0 0 0 0
General 2 0 0 0 0 0 0 0 0
3 - ^ 3 3 Ayi3 J\33 0 K 2
-K 2
K 0
4 Ay iir -Ay'44 0 v
44 K2
-K 2
0 K
1 0 0 X 0 0 0 X 0
Difference 2 0 0 X 0 0 0 X 0
3 X X X X X X X X
4 Ay^ - Ay xlr
X K 2
-K 2
X K
1 0 X 0 0 0 X 0 0
Paraphase 2 X X X X X X X X
3 -<<>'3 3 X .)'3 3 0 -K 2
X K 0
4 Ay** X 0 >44 K 2
X 0 K
1 0 X X 0 0 X X 0
Single in/out 2 X X X X X X X X
3 X X X X X X X X
4 A)u 4 X X v 4
4
K 2
X X K
1 2 3 4 1 2 3 4
1 >'n 0 0 0 K 0 0 0
General 2 0 0 0 0 K 0 0
3 - B y u ^ 2 2 0 0 -K 2
K2
0 0
4 Byll -S.V2 2 0 0 K 2
- K 2
0 0
1 yw 0 X 0 K 0 X 0
Difference 2 0 J'22 X 0 0 K X 0
3 X X X X X X X X
4 By ii X 0 K 2
- K 2
X 0
1 X 0 0 K X 0 0
Paraphase 2 X X X X X X X X
3 X 0 0 - K 1
X 0 0
4 By u X 0 0 K 2
X 0 0
1 X X 0 K X X 0
Single in/out 2 X X X X X X X X
3 X X X X X X X X
4 By u X X 0 K 2
X X 0
1 2 3 4 1 2 3 4
1 0 0 0 0 0 0 0 0
General 2 0 0 0 0 0 0 0 0
3 -c C 0 0 -K K 0 0
4 c -C 0 0 K -K 0 0
1 0 0 X 0 0 0 X 0
Difference 2 0 0 X 0 0 0 X 0
3 X X X X X X X X
4 c -c X 0 K -K X 0
•
1 0 X 0 0 0 X 0 0
2 X X X X X X X X
Paraphase
3 - C X 0 0 -K X 0 0
4 C X 0 0 K X 0 0
1 0 X X 0 0 X X 0
Single in/out 2 X X X X X X X X
3 X X X X X X X X
4 C X X 0 K X X 0
J. I. SEWELL and F . W . S T E P H E N S O N
1 2 3 4 1 2 3 4
I 3'n 0 0 0 K 0 0 0
General 2 0 3*22 0 0 0 K 0 0
3 ~ AVu3*33 -Dy 23'33 2 3'3 3 0 - K 3
K 3
K 0
4 A v 11 3*44 "" AV2 2 3'44 0 3*44 K 3
- K 3
0 K
1 3*11 0 X 0 K 0 X 0
Difference 2 0 3*2 2 X 0 0 K X 0
3 X X X X X X X X
4 AV 11 3*44 "- £ 3 * 2 2 3*44 X 3*44 K 3
' -A: 3
X K
1 3*n X 0 0 K X 0 0
Paraphase X X X X X X X X
3 -Dyi 13*33 X 3*33 0 -A: 3
X K 0
4 AVll3*44 X 0 3*44 A: 3
X 0 K
1 3*11 X X 0 K X X 0
Single in/out 2 X X X X X X X X
3 X X X X X X X X
4 Dy n >>44 X X 3*44 K 3
X X K
1 2 3 4 1 2 3 4
1 3*11 0 0 0 K 0 0 0
Voltage output 2 0 0 0 0 0 0 0 0
3 - AVu3*33 ^3'3 3 3'3 3 0 - K 3
K 2
K 0
4 Dy u 3*44 - ^3*44 0 3*44 K 3
- K 2
0 K
1 0 0 0 0 0 0 0 0
Voltage output 2 0 3*2 2 0 0 0 K 0 0
3 -^3'33 ^3*2 2^33 3*3 3 0 - K 2
K 3
K 0
4 ^3*44 "-Dy y .
22 iA 0 3*44 K 2
- K 3
0 K
1 3'ii 0 0 0 K 0 0 0
Current output 2 0 0 0 0 0 0 0 0
3 -By u c 0 0 - K 2
K 0 0
4 By,, -c 0 0 K2
-K 0 0
1 0 0 0 0 0 0 0 0
Current output 2 0 3'2 2 0 0 0 K 0 0
3 -c By 22 0 0 -K K 2
0 0
4 c -By 22
0 0 K - K 2
0 0
4 ^ 4 4 "-Dy y .
22 44 0 ^44 K 1
-K3
0 K
1 .Vu 0 0 0 K 0 0 0
Complete 2 0 0 0 0 0 0 0 0
hybrid 3 -£.)'3 3 C 0 0 -K2
K 0 0
4 ^llJ'4* — Ay 44 0 J'44 K> -K2
0 K
-
1 0 0 0 0 0 0 0 0
Complete 2 0 3'2 2 0 0 0 K 0 0
hybrid 3 - ^ 3 3 Dy y 33 >'3 3 0 -K2
K3
K 0
22
2
4 c -By 2 2 0 0 K -K 0 0
1 >'ll 0 0 0 K 0 0 0
Complete 2 0 0 0 0 0 0 0 0
hybrid 3 -Dy^y^ Ay 3 J'3 3 0 -K3
K 2
K 0
3
4 By -C 0 0 K2
-K 0 0
J. I. S E W E L L and F . W . STEPHENSON
Y} Y,Y 2
e s y*
4 (>U4 + Y ) ( 2Y) y i l + Y + Y )+ Y (Ay„-
l 2 2 2 yu + Y: + Y 2
yu + Yt + Y,
(1)
This expression reduces to the more familiar one: j iy**+Y ) 2
Y^Ay^-YJ YMy - Y)
Y>1* = _ (2) + yn + Y 4A 2
e> Y y^ + Yt + Y, i + Y 2
2
between input and output voltages. From Table 10, , yn^y^A- Y)2 , Y (y„A-Y ) 2 2
and
Y, = Y 2 = y 3 = y 4 = y
then
e A = YjY + y^A)
s Y { y A Y )
* [(2Y+Y ){2Y 5 + ytt+ " ~
2Y
Y(y A+Y)]
4i
(4)
Now, if y Ai and A -> oo, then
e 2Y ±=
•(5)
Y5
Fig. 5. Non-inverting amplifier (II).
2
y, ', „ f., . y >,3
- Y 1 y 4 +
i • ,
\ yu+yi2 + Y + Y 2 3
y\ 2
y n + yi2 + y + v 2 3
' -(y**A + Y ) 4
fy22+yi2+yi+y4+y5 Yy 3 i2
y* +
y\2
,vii +yn + Y + y J 2 3 Ivii + y 12 + y> + Y 3
3'n+.vi +y +y i 2 2 J
f>-22+yi2+y.+y4+y5"
L }'u +yu+ Y +Y
2 }J
Table 11
Admittance matrix for the voltage follower (Fig. 6)
yn -yiz
J. I. S E W E L L and F . W. STEPHENSON
Si s„ 4
2
y ! - y Y t 2
- Y Y„
T
- ^ Y,
1
v
-> 11 1 1
' .' n • X >i j'n + Z >/
ij
-Y,Y 2 ! ~Y Y 2 N
\ Y
yu+lY, y-n+lY,
y„ -y y„
2 Y„Y f
0-44+y/)
y^j>44-y/)
+ y/(^44-y/)
>-u + E Y t >'n + Ey.-
3 4
y> 2 y,y 2
S,o-
s y. 0
3 y
Yt + Yz + y^ 33
Y +Y + y
l 2 il
J
n-I0-
y +>44
Y (y^A -Y )
l lA 2
2
^ 4 4 ^ , 4 - y ) n o-
Y + Y +y
t 2 n
2
y. + y +>-u 2
5. Acknowledgments
The authors have pleasure in acknowledging finan-
cial support from the Ministry of Technology.
6. References
Fig. 8. Differential output network. 1. Sewell, J . I . , 'The Application of Analogue Computing
Techniques to the Design of Active Filters', Ph.D. Thesis,
University of Newcastle-upon-Tyne, 1966.
If Al3 1, then 2. Baxandall, P. J . and Shallow, E . W., 'Constant-current-
source with unusually high internal resistance and good
e3 Y x temperature stability', Electronics Letters, 2, No. 9, pp.
.(14)
351-2, September 1966.
3. Kerwin, W. J . , Huelsman, L . P. and Newcomb, R. W.,
If y = pC and yi = l/R, then the network can
2
'State-variable Synthesis for Insensitive Integrated Circuit
be used as a dual integrator giving positive and Transfer Functions', Stanford University Technical Report
negative integrals of the input signal. No. 6560-10, p. 4.
I
<N
—.
ca
ca
5S
+
<N
C/5
00
II ^
0) en
ca
U3
9
4>
ca
Hi
9.
ca o (J
CJ
E S
•a
ca
w UJ
-^/VvV—I
I—-W\A/—'
ca 1>
(/i •AWv-|l
UJ
CA
to
3 O
in
« 3 i/i
UJ
in UJ in
CA
<u
ks- A s - UJ
UJ
•=> ca
1 in
U.
K Co IV)
o, CS
April 1969
255
2 91
PAPER 9
I N T . .1. C O N T R O L , 1970. V O L . I 1. NO. 3, 479-488
1. Introduction
I n recent years there has been an increasing interest in the d e r i v a t i o n of
algorithmic methods o f analysis f o r various networks. This has been brought
about b y the demand for faster and more accurate analysis techniques f o r use
on computers. General analysis methods are either not sufficiently accurate
(Skwirzynski 1965) or occupy so much computer store to make them i m p r a c t i c -
able. These difficulties are accentuated when iterative synthesis is considered,
and a t h i r d factor, the time occupied by the analysis step, becomes a m a j o r
concern.
Many o p t i m i z a t i o n procedures used i n automatic synthesis require the
c o m p u t a t i o n o f the p a r t i a l derivatives o f the error f u n c t i o n each cycle o f itera-
:
t i o n . W i t h m a n y methods this is not p a r t i c u l a r l y eas\ to accomplish, and i f
numerical d i f f e r e n t i a t i o n is being used inaccuracies are i n e v i t a b l y present.
Analysis met hods which w i l l meet the abovedemands are q u i t e wellestablished
f o r passive networks ( B a r t l e t t 1930, Herrero and W i l l o n e r 1966). B u t active
networks have not received as much a t t e n t i o n , although some progress has been
made in the case of m u l t i -loop feedback networks ( H o l t and Sewell 1966). These
techniques do satisfy the accuracy, time and storage demands cited above b u t the
c o m p u t a t i o n o f p a r t i a l derivatives, t h o u g h possible, is not s t r a i g h t f o r w a r d and
requires e x t r a p r o g r a m m i n g o f the computer.
The accumulant method enables the accurate and speedy analysis o f a m u l t i -
loop structure and permits an analytic assembly o f the derivatives.
Z x 1 0 0 0 1
-1 ]'„ 1 0 e 2
1 Z 3 1 0 0 h
o - i y 4 i o
— 1
* 2.Y-4 1 ^ 2.V-4 c
2.\'-4
0 - 1 Z t v _, 0 '2.V-3
o o - i y ._ .
iA 2
_ c
2.\ —
1
voltage gain G 21 =
. • j ri (^1 ^2.V-2)
i n p u t impedance Z — j^ TF :! (2)
n
(^ 2 •'2.V-2)
1
f o r w a r d transfer impedance Z 1 2 =
2.V-2^
Fig. 1
I
e
0 Z
! e
2 Zj %
I
The terms i n brackets are called accumulants, i n this case t h e y are referred
to as single accumulants. They are similar to cumulants (Herrero and W i l l o n e r
1966) b u t because the f o r m of the H m a t r i x is slightly modified, d i f f e r e n t pro-
perties result.
Accumulant theory and applications 481
r Z l 1 0 0
- l r,+ r 2 3 i 0 Y 2
o -1 z 3 1 0
o o -1 Y i + Y i s
Y A
~ 1 -^2.V-4l -^2.V-4
+ 1 2.V-4, 2.Y-3
0 -1 Z . _,
2 v 0
0 0 —
1 ^2A'=2 _
Fig. 2
Y
'2N-4
— [
«1 1
0 0
—1 (/,, 1 0 . a 2
0 -1 «3 1 0 0
0 -1 «4 I 0 . « 4
where a, is an
a d m i t t a n c e or
impedance.
1 a
n-2 1 a
n-2
0 -1 0
0 0 - 1 a„
482 ./. / . Sett-ell
H = H + H, X
«1 1 0 0 0 0 0
-1 1 0 n 0 0
0 - 1 1 0 0 0 0
0 -1 °4 0 0 0
H =
-1 a
„-2 1 0 "a
0 - ] • 1
0 1 a„ 0
final o n e .
1 0 0
-1 1 0 a2
0 -1 «3 1 0 0
0 0 -1 «4 1 0 o 4
\H\ = \H \
1 +
1 a _ (1 2 I a n-2
0 - l aH - l - 1
0 -- 1 0
W h e n n is even
1^1=1^1-1,
n odd:
1^1 = 1^1
H is o f the correct f o r m f o r evaluation by the Euler method. The case when
1
I « , + &, 1 0
0 - 1 1 0 0
0 0 - 1 a, + b x 1
-1 a„_, 0
771
+ 2 2 b b (a
OJ 2j 1 a ^ j ) (a 2 i + 1 fl,^) (a,, - +1
+ .. .
+ (b b . 2 t . • b _ )(a a .
n 2 1 3 . •(t ^a _ a ).
n H l n
n-2
m•
(wis even).
Proof
(flj a ) © {b
n 1 b __ ) results f r o m the sum o f t w o matrices:
n 2
A B
a 1 1 0 0 - 0 0 0
-1 a, 1 0 a 2 0 b 2 0 0
0-1 flg 1 0 0 0 0 0 0
0-1 cr 4 1 0 . 0 6 4 0 0
+
484 J. I. Sen-ell
«1
1 0 0
-1 a i 1 Co
0 - 1 «3
1
0 -1 a
t -I 0 0
0 -1 1 «.•
0 0 ai+l
0
- 1
a 1
,l-l
+ K +l «r-l)(Or+2 «<»)}
»l Hi — 1
+ 2 2 &2.MK a
r)(«r+l «2i-l) + (ai «r-l)(«r+2 «2,-l)}
J = i +1 i = l(r+l)/2]
X a
{(«2i+l 2,-l)(«2j-l «»)}
m (r/2)-l
b b
+ 2, Z, 2i 2i K «!(-l){(«B+l -«r)K+l %-l)
J = [(r+l)/2] < =1
x
+ («2f+l «r-l)(«>+2 Oy-i)} {(ffly+i "*)}
lr/2]-l [r/2J-2
b
+ 2 2 2 i h j ( « i «2i-i)(«2i-fi
j=i+l i=l
X (
+i «>)(«>+! « „ ) + («2> + l «r-l)(«r+2 <n)}
+ .. .
,((l-2)/2 \,»/2 ,
+ n K n « 2 y - ! « „ •
V i =l ' \i = l
Accumuhnl theory and applications 485
Proof
This follows s i m p l y by a p p l y i n g (4.1) and the rule f o r s p l i t t i n g eumulants
(Herrero and W i l l o n e r 1966).
= K o„ + (a 1 a„_ ) 2
m
+ 2 K (O,. «2i-l){K-l + («2i+l «„- )} 2
£=1
m »! - 1
+ 2 S&a&fyK a
u-i)lfl*+i a
2j-i)
;'=i + l i=l
x a
{(«iy+i + (% i + »- )}2
+...
, <n-2)/2 \/i/2 \
b
+^ II 2ijlJJ^V-l J«n-
Proof
This follows b y a p p l y i n g the rule f o r expanding cumulants by the final t e r m
to the expansion o f (4.1).
= a r _ l ) ( f l r + 1 a , , )
da
m
a
+ 2 M«i «>_i)(«y +1 «2i_i)(oa+i J
! = ((r+l)/2]
[r/21-1
b a a a
+ 1 2i( l 2 i - l ) («2/+l «r-l)( r+l 0„)
1 = 1
»i »( —1
a a a
+ 1 I M 2 ; ( « l «Y-l)(«r-l 2i-l)(«2i+l i j - l ) ( V+l «n)
j = i + l !=[<r-M)/2|
HI |r/2]-l
f
+ 1 I M ;( 'i 2 «2i-i)(»2t+i a _i)(a
r r + i « ;-l)
2
a
x («2,+i n)
(r/2]-l [r/2!-2
a a
+ 1 2 M 2 > ( « 1 «2/-l)(«2i-M «2,-l)(«2>+l -«r-l)( r-rl n)
;'=i-fl 1= 1
+...
,(H-2)/2 \ / i i / 2 \ a I 1 71/2 \
+ n y n « y - - k , v + i i ,
\ ,=i ' v=i « r I o,i j=i )
S ; ' is the Kronecker delta.
486 ./. 1. Sewell
Proof
Apply rule 4.3 to 4.2 gives :
(«1 « r - «r-M « „ ) © (bi b b_
r! r i >> _. )
tl 2
a
= {(«! 0 r + ("l ,-z)} K - l «,.) + («1 «r-l)(«r-2- ""J
/ft
fl
+ I M[(«l Or-lK+K «r-l)]( r-l "2i —1'
i' = [<r-l)/2]
X
+ («1 Cf - )(rt'r-2
r 1 «2i-l)} {(«2i+l «,,)}
(r/2!-l
+ 2 «2i-l){[(«2,-l « r - l ) « r + («2f+l « r - 2 > ] («r-M ««)
1 =1
+ («2.^1 «r-l)(«r-2 <*»)}
m [r/21-1
b a
+ 2 2 Z i h j ( l «2i-l){[(«2i-l «r-2)]
;' = [ ( r + l ) / 2 | £= 1
x (a r + 1 a. ._!) + (a,,..!
2; a -i)(a ^_
r r a _ )}
2j x
x(«2 i J + a J
[ r / 2 1 - l [r/21-2
b
+ I 2 2ib (a 2J 1 a i-i)K-i
2 «2>-i){[K;+i «y-iK-
j=i+l i=l
a f a
+ («2>+l r- )] ( V l
2 «J + ( 2; +l «r-l)(«r-2 «»)}
m ;n — 1
+ 2 2 6 6
2i 2j{[(«l a -l)«r+( -l
r
a
Or-a) ](«> + ! a _i)
2i
+ ( <l!-2)/2
n 6
\ /n/2
2i)(n« ;-i)«,r 2
\
and 1 because 2 i > n —2. Otherwise the split accumulants are evaluated as
indicated earlier.
r/2-1
+ 2 2i(«l 6
«2i-l)(«2/+l «r-l)(«r+l «»)
i=l
m-1
+ 2 6
2>(«1 °r-l)(«r+l fl
2;-l) K+l «»)
J=[(r+l)/2]
+ .. .
Accumulant theory and applications 4S7
Proof
Consider the expansion of (4.1), d i f f e r e n t i a t i n g w . r . t . b leads directly to the r
above expression.
C = aa
2 1 2 + 1,
C = a {a a
3 3 1 2 + ]) + a , 1
ill/2)
C n = aC_
n n l + C _ -
n 2 I8 ". 2s
8=1
0 0 - 1 Z 3 1 0 *3
e 0 - 1 Y,+ Y Y*
o ^ ffi ffi e
\
2.V-4 1 T
3 2.Y-4
e
e
0 ^ //2.V-4 2.V-4
+ Y //2.V -4
0 0 - 1
0 0 1 I 9V-9
488 Accumulant theory and applications
Fig. 3
2N - !
A'-l
(_l)2.v-i+ ^ Y j m & x Z»_ )®{Y
x m
Y,//<2i-2>/
i=i
'2.Y-2 X ( - l)2-V-2i + l
G 21 = -
:
\ Z , Y 2 X ) ©( Y m 3" /<2.v-4)")
/
J 1
i =l U //2 //(2i-2)) '
(T ,
9 =
( 3 "^(2.Y-2X2.V-l))\
—i\v)e! ;;
^//(2-Y-4)) 1
7. Conclusions
Computer programmes have been w r i t t e n using the algebra o f accumulants,
and have been f o u n d t o be fast a n d accurate. T h e m a i n advantage o f accumu-
lants is t h a t p a r t i a l derivatives can be computed w i t h ease, thus presenting an
ideal analysis m e t h o d f o r inclusion i n i t e r a t i v e synthesis. T h e derivatives are
also o f value i n sensitivity w o r k .
The s p l i t t i n g technique enables a large structure t o be dissociated into a
number of small analysis problems and thus alleviating t h e demand f o r large
storage on a machine a t a n y p a r t i c u l a r instant.
The m e t h o d is also suitable f o r literal analysis b y computer.
ACKNOWLEDGMENTS
The author wishes t o acknowledge t h e financial support o f the M i n i s t r y of
Technology a n d D r . A . G. J . H o l t of the U n i v e r s i t y o f Newcastle upon T y n e f o r
advice and encouragement.
REFERENCES
BARTLETT, A. C , 1930. The Theory of Electrical Artificial Lines and Filters (Chapman
& Hall).
HEBRERO, J. L . , and W I L L O X E R , G., 19(36, Synthesis of Filters (Prentice-Hall).
H O L T , A. G. J., and SEWELL, J. I . , 1965, Proc. Instn elect. Engrs, 112, 2227; 1966,
Trans. Circuit Theory, 13, 326.
SEWELL, J. I . , and N I G H T I N G A L E , C , 1969, I.E.E. Conference Publication, No. 51,
p. 515.
S K W I R Z Y N S K I , J. K . , 1965, Marconi Rev., 28.
30 2
PAPER 13
3 0 3
J. I . SEWELL
Department of Electronic Engineering. University of Hull.
Hull, HUB 7 R X
1. Introduction
I t has been demonstrated recently (Rollett 1968} how versatile active
multi-port circulators can be, when used in conjunction with resistors and
capacitors to produce active filter networks. T h i s communication attempts
to establish a general interconnection system for multi-port circulators.
2. General configuration
Consider the following system
Gc,
i=i . . . n.
Gc n overall circulator admittance matrix. I t is skew-symmetric and of
order m . 6'c = 0. n is the number of parallel systems.
n 0
r
S ( j ', are similarly defined but operate on the Gcc matrix.
(j
= n , a x :
• V . ; - i = *V ("'l.y-i
C r — V / _ r
«V =5 , ' = .- 1 1 °11 —
°21. —
=,v=°
~ 0 I - 1 .. - r " 0 - 1 1 ... 1"
- 1 0 1 .. I 1 0 -1 ... -1
9'tCll
1 - 1 0 .. . - I - 1 1 0 ... 1 .
1 - 1 1 .. 0 - 1 1 - 1 ... 0_
little significance. F o r gCCn >g CCiil a circulator of positive sequence and reduced
is u n b a l a n c e d but still reciprocal between the ports, the matrix remains skew-
symmetric but the entries are not equal. A second circulator m a y be used in
the above connection to ensure identical conductances at. and between, all the
ports.
2.2. Two circulators in parallel, one of negative sequence, and ] -port overlap
T h e significance of this connection is demonstrated simply using third-
order circulators :
Active circulator si/stf>iis 079
= S : r
•V ' * i i ' = °- -V "• S.J=\. .So ' = 0.
L S. =\2l
- 1
I
' v . = './••• i (I
= a n t
I f 9cc ~(J^n
n * capacitors are connected across ports 2. 3 as indicated
in the figure, the admittance matrix becomes
0 - 9c 0
~9r 0 f'r
9c 0 sC 2 -<Jr
0 9c 0
a a
•c, sc 2
O
• f w r ^ — o
I 4
3. Conclusions
There are many possible configurations which remain to be investigated,
the few shown here demonstrate interesting effects which result from inter-
connected circulators. One higher order combination examined has the
parameters :
H = 2 , H , = 1. /»j = -t. )(., = 1. / » = 3, K. 2 n = I. oc = u + \, a 2 1 = + l.
K u = {\, 1, - l , , v = .s' ' = ,sv = s ' = o, s ' = o,
n 2 2 1 v = i -
REFERENCES
PAPERS 1 4 + 1 5
'38 I E E E T R A N S A C T I O N S O N C I R C U I T T H E O R Y . V O L . C T - 2 0 . NO. 6. N O V E M B E R 1973
Abstract—Symbolic analysis using direct matrix and interpolation- active element is one which may have any degree of frequency
formula methods are discussed. The techniques accommodate passive dependence and can itself only be described by a symbolic
networks and active networks containing general frequency-dependent
rational-function matrix. An entry of the nodal admittance
devices, themselves represented by rational-function matrices. Subse-
quently, this leads to the development of partitioned analysis. The matrix Y will have the form
suppression of numeric common factors and the determination of the
order of complexity of the network are synonymous problems, and
.v !•/(*) =
simple algorithmic solutions emerge. Simple rules for the choice of <*//(*)'
sample points leading to accurate analysis are given. Implementation
yields a very fast computer-analysis method capable of accuracies com- Suppression of the internal nodes will yield a terminal matrix
r
INTRODUCTION
'dm
T H E I N C R E A S E D application of the computer in design-
ing electrical networks has emphasized the need for
efficient analysis programs which are particularly useful in
where r signifies a reduced matrix. The problem then is to de-
termine the coefficients of these polynomials.
iterative synthesis. In this area of work, symbolic analysis, in A fundamental theorem of algebra states that a polynomial
which the network functions are obtained as functions of s, of mth degree is uniquely defined by the value at m + 1 values
the complex frequency variable, with either literal or numeric of its argument. Thus if a network matrix determinant or co-
coefficients, becomes quite attractive. factor is evaluated at m + 1 frequencies, the resultant de-
terminant values completely define the corresponding mth-
The traditional approach to symbolic network analysis using
degree network-function polynomial. If a sample value k
topological techniques has the inevitable time-consuming tree- p
proach. This technique involves the determination of the co- [K][A]=[A] (1)
efficients of the describing polynomials by subjecting the net-
work to random excitations v/hich lead to either the inversion where [A] is a vector of functional values, [A] is a vector of
of a purely numeric matrix or the evaluation of an interpola- the required coefficients, and [K] is the (m + 1) X (m + 1)
tion formula. The computer implementation provides a very Vandermonde matrix of sample points.
fast analysis. By exploiting the standard form of the Vandermonde
matrix, a solution for the coefficient vector may be obtained
In the analysis of real active networks, device modeling
directly from
presents a complex problem. However, provided the device
can be regarded as linear, a matrix can always be made to fit Ml - [ A T ] - ' [ A ] . (2)
the device characteristics, although in general this matrix will
contain symbolic rational-function entries. An algorithm by Traub [8] has been used for the inversion
of [K] and it provides a fast and, for samples within certain
FORMULATION O F SYNTHETIC ANALYSIS bounds, accurate solution. The numerators and denominators
of the rational functions of the reduced matrix are assembled
The general network to be considered contains resistors,
by repeating this process.
capacitors, inductors, and general active devices. A general
The main advantage which accrues from such a strategy is
that once [K] has been inverted, it may be used for a variety
Manuscript received February 21, 1973; revised May 15, 1973.
J. K. Fidlcr is with the Department of Electrical Engineering Science, of analyses, providing the sample frequencies chosen remain
University of Essex, Colchester, England.
the same. A second point is that the accuracy of the inverse
J. 1. Scwcll is with the Department of Electronic Engineering, Uni-
versity of Hull, Hull, Yorks., England. matrix can be improved subsequently at will. By using an
F I D L E R AND S E W E L L : A N A L Y S I S FOR C O M P U T E R - A I D E D DESIGN 739
iterative technique [9] for a small number of cycles (not more (m - I )/2 (s + O,)
than 6), improvements of 3-4 orders of magnitude in the error F(s) =
r=0
£
n
of the inverse, for matrices of up to order 30, have been (=0
A ) + A / _ Km-0/21
{ - h i m - «(»«
L (7)
1=0
the points where the network function is undergoing change, coefficients, respectively, and [ A T ^ ] and [ K ] are Vander-
that is, in active areas of the j-plane. monde matrices of samples.
1
"/hen m is odd, [Af* '] = [K^\, and when m is even, this
In the general case, the computation of (2) or (3) for com-
identity can be assured by increasing m by unity. Hence,
plex samples may seem somewhat daunting; however, if the
samples are chosen in as many complex-conjugate pairs as [AM] = [*<«>]-'[ZW]
possible, the property of network functions that A ( s ) = A(s) m 2
[A ] = [*•<'>]-'[Z< >].
may be recognized, the bar denoting complex conjugate. Thus
the number of determinant evaluations will be reduced, re- The Vandermonde matrix is approximately half the size
quiring \(m + 1) for m odd and ({ml2) + 1) for m even. ((m + l)/2, m odd; (ml2) + l , m even) of thai required in (2),
Taking m as odd, (3) becomes and again, only one inversion is required. If imaginary samples
740 I E E E TRANSACTIONS ON C I R C U I T T H E O R Y , NOVEMBER 1973
PROGRAM OPERATION
then difficulties will arise due to the buildup of numeric com-
mon factors. It is obvious that these common factors occur Programs utilizing both approaches have been written and
for the same reasons as the algebraic ones in the inversion of used on an I C L 1905 E and a PDP-10 in both batch process
matrices in symbolic analysis [ 4 ] , and an adaptation of the and conversational modes. Considerable experience has been
factor suppression techniques used in that work is appropri- gained in the analysis of both active and passive networks
ate here. using these techniques. Polynomials, with wide-ranging co-
A new term at any stage in a pivotal condensation process efficients, of order up to 39 are currently being generated with
is given by coefficients accurate to at least 4 significant figures. To com-
pute the complete 2 X 2 admittance matrix in rational form
n n
, _ ij kkdikdkj ~ n, n jd,jd k k kk d k k for a fourth-order active multiloop circuit takes 3 s and a
y
'' d d d d n seventh-order passive elliptic filter (8 = 6 0 ° ) takes 7 s with a
t j k k l k k l k k
One measure of derivative sensitivity for a network function F with where D' is the derivative of D with respect to e.
respect to a network element e is given by the following: Equation (3) may be used to provide simple methods fur the evalua-
e&F tion of the change A f in the network function F due to a finite change
(1) Ae I.; the network element e.
~F~Ze'
In [ 2 ] , a method of computing the change in a response voltage due
First-derivative sensitivity is usually useful as a guide only in cases to a finite change in a network element was described. The network
where very small changes in element values are made. For larger was current driven, and therefore the function of interest was transfer
changes, we may define the finite-change sensitivity function impedance. For a single input current flowing from the reference node
into node k, the voltage at node m is given by
• MEASURED theoretical values obtained from (3), (4), and (7). A l l standing
I I PREOICTED wave modes were observed for ;V = 1, 2, 4, a n d 6, i.e., for
N + 1 a prime number, but not for TV = 3, 5, 7, a n d 8.
F , - 3 4 34 KHZ The ladder oscillator presented here differs from other
(36.43) multimode oscillators [8], [9] in that its mode of oscillation c a n
be ascertained from the relative phase o f the node voltages
without measuring the frequency. T h e oscillator has also been
• IN P H A S E
- OUT OF P H A S E realized as an active RC circuit by replacing the inductors with
gyrators a n d capacitors [10].
F -25.97 KHZ
s
(25.74) REFERENCES
1
I n the above paper, o n p. 739, (5) should be replaced with
(m- l)/2
( I m [&(k,)] R e [P ]r - I m [P ] r Re [A(M])fr - o,) + ( R e [ A ( M ] R e [P ] r + Im [A«r )] Im f [P ])a>A
r
r=0
2 2
w (Re
r [P ]r + Im [/>,]) <"-J>/2 j 1
. [I (* 2
- l",* + 0>, ) J
(5)
1=0 J
l*r
where
i=0
2 2
and in (6) o n the same page the minus sign between (s - cu, ) should be plus, such that the last bracket o f that expression reads
2 2
(i + rVVU-V - co ). r
PAPER 16
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-21, NO. 3, MAY 1974 345
Abstract—A method Is given for Improving the accuracy of the orders may be correct, gross coefficient errors are not
coefficients of network polynomials of high order produced by symbolic uncommon when the system order exceeds fifteen.
analysis on a computer, By partitioning the polynomials into sets of
One method which has been successfully applied to the
coefficients the order of the individual matrices involved is reduced and
so is the error when these are inverted. The problem is reduced to the analysis of high-order systems is to partition the network
Inversion of only two Vandermonde matrices. Considerable increases in [ 6 ] , and the results show improvement in accuracy and
accuracy arc demonstrated. speed of analysis.
The technique to be described now involves polynomial
INTRODUCTION
partitioning in an attempt to increase accuracy and speed of
A T T H E present time there is considerable interest in symbolic analysis on a computer.
symbolic analysis which generally implies the
POLYNOMIAL INTERPOLATION A N D PARTITIONS
analysis of networks by computer to yield the characteriza-
tion of the network functions as rational functions of the Following the interpolative approach [4] and considering
frequency variable s with numeric coefficients. Techniques any required polynomial
evolved so far fall into the groups of topological, eigenvalue J
f(s) = a 0 + as
{ + o i 2 + •• •+ a^T
[1], matrix inversion [2], and interpolative methods
[3], [ 4 ] . It is also possible to generate polynomials with subjecting the network to m + 1 samples of the variable
literal coefficients and an efficient method for semiliteral (complex) will enable the coefficients a; to be determined
analysis has recently been announced [5]. from
As the complexity of the network increases, so do the F = aA (1)
orders of the resultant polynomials, and the problem of
where A is the column vector of coefficients, F the column
coefficient accuracy arises. In fact, although the polynomial
vector of sampled responses, and a the [m + 1) x (w + 1)
Vandermonde matrix of samples. A may be found by
Manuscript received May 18, 1973. direct inversion or interpolation formulas. In direct
The authors arc with the Department of Electronic Engineering,
University of Hull, Hull H U 6 7 R X , England. inversion, error in the determination of A comes from two
346 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, MAY 1974
As,) 1
o r * ,
A
T 0
1 5
I+ 1 (3)
(2)
S S
v . .
.1 l +6 - l Uh-l
T o solve for the coefficients a general matrix inversion
The sample matrix is not square since all the coefficients can
routine would be required. However, if D = b I, t being q q
only be determinea with all the samples. However, on
the unit matrix and b some complex constant, then (3) q
partitioning the coefficients into sets containing h each, (2)
becomes
becomes
F = £C4 (4)
m
where Z — diag (a ,a ,-• ',a -i) and C is a modified 0 i p
i+ 1
This yields a threefold choice in the method for determining
b Vq; the radius can be changed between coefficicrt sets,
q
* R A+'l 3
-.2A- 1 with <p constant, the initial set angle can be changed with
l + A- I 5
i +A- 1 1+A- s
i + A-l_ q
k ;
Therefore a general entry of C is D, = p,*V *«""/, and it
i 1
is possible to compute A from A = C~ Z~ F using
- 1
Vandermonde matrix inversion only, as I = diag
>
(o- ~ ,a0 • • ,a~l j) and a, is Vandermonde too.
l
S (p-I)A J ( -1)*+1
P s (p-l)*+2 A still further improvement can be achieved. If the
JP- 1)*+ I
J
_((.-l)A+2 change in b is simple, then as the samples of a are q 0
i+ 1 *l + 1 1
J-,*!, • V A - I those of o, can be k s ,k s ,- • • fi^^ or q 0 q l t
2
J
.(p-l)A J
(p-l)A-n ( p - l)A+2 a, = c K where K = diag (\,k ,k ,-•
0 q • ,*$"'), k, being a
q q q
L i + A-l i + A-1 4
i + A- I 1 - 1 1
complex quantity. T h u s c , " = A , " ^ " a n d £ " ' can be
a
(p-l)A obtained by inverting only one Vandcrmonde matrix.
3
(|>-1)A+1 This method has removed some degree of arbitrariness
concerning the sample locations in the complex plane.
Certain further constraints govern the choice of circular
sample locations. Time savings can be made in computing
which can be rewritten as
the sampled responses svhen conjugate samples are used,
F , = a,A 0 + D q \ A l + •••+ C r ' ^ r i and with p = constant V, this property can only be
q
FRIDAS AND SEWELL: SYMBOLIC ANALYSIS OF NETWORKS 347
TABLE I
COMPARATIVE COMPUTER PROGRAM REQUIREMENTS
10
30-N0DE TIME (SECS.) 39 TO
/
/ LINEAR
llO-HODE TIME (SECS.) 72 1 13
10
10
/
PAPER 19
Network partition methods for use in symbolic
analysis by interpolation
C . P H R Y D A S and J . I. S E W E L L
(Department of Electronic Engineering, University of Hull, Hull, HV6 7RX, England)
Symbolic analysis of networks of high order and complexity is not a straightforward process. The problems encountered
are usually enormous demands for computer storage and time together with inaccurate results. A number of methods
are presented which can drastically reduce the computer requirements and in some cases keep the computational error
to a minimum. The principle of the methods is to partition the networks into a number of parts, and techniques of
division and recombination ure developed.
In recent years a variety of techniques for the symbolic it would therefore seem necessary to overcome the effects
analysis of networks have been developed. These generally which impede true partitioning in symbolic work.
fall into the categories, topological, eigenvalue , matrix in- 1
One solution to the elimination of c o m m o n factors on
10
2 3 4
version , state space and interpolative . T h e analysis has partition uses an artificial loading t e c h n i q u e . T h e methods
been extended to produce polynomials with completely outlined now obviate the need for any artificial elements.
s s
literal , semi-litera) and numerical coefficients in both pas-
6
sive and active networks. More recently , requirements in
tolerance, statistical, and stability analysis have demanded NETWORK PARTITION
the production of the whole inverse matrix. T h i s is a more
difficult problem to tackle generally, and the difficulties In Figure 1 let N represenr a network with a nodal admit-
encountered with the symbolic analysis of large networks tance matrix Y ; rhe sets 5, /, E containing all internal and
are sufficient to justify restricted attention to their solu- external nodes respectively, hence S = I U E. T h e nodes
tion alone although partition methods are of value in the are numbered in such a way so that if X, £ E, and /ly £ / ,
above analysis areas. A n y form of symbolic analysis will then A, < uy V i, j. Dividing N into d parts in an arbitrary
require more computer time than straightforward numeri- way, Nj, i = 1,2, • • •, d, then the s e t s S , , /,', £ , may be de-
cal methods, hence speed is a vital concern. fined as for N. Clearly, 5,- C S, / , C / and £,• C / , V i .
As the order of the network is increased, inevitable prob- L e t C;y be a set defining the c o m m o n nodes between
lems with the computer requirements for speed and stor- N j , Nji thus Cjj = E , n Ej. Define n, u, € as the total num-
age are encountered. T h e computation error apparent in ber of elements in S, I, E respectively, and similarly u,,
the results also gives considerable cause for concern. Some e, for S{, /,, £;.
techniques have been developed which do improve the per- A straightforward analysis of N can be achieved as indi-
formance for higher order networks and these include par- cated previously". A n improved condensation formula for
1
titioned polynomial interpolation and use of the F a s t evaluating the sampled responses is:
8
Fourier T r a n s f o r m . However, with large increase in net-
work complexity, these difficulties still increase in severity
and sparse matrix methods and network partition techni- d
1 0
"I; = ( ' U j ' ^ k ^ k j ~ °4k"ik"kjV"k*lJ<+\
q u e s provide ways of tackling the problem.
T h e principle and advantages of partitioning large net- n n n .
works is not new and has been amply explored in frequency
d
and time domain analysis since the early days of K r o n . N o l <r a (n n<& <»
only can large networks be considered, but, especially in
the case of electronic circuits such as integrated amplifiers
in which many sections are duplicated, the symbolic solu- where d^ is the denominator of rhe gh entry of the nodal
tion of these circuiis can be stored and called as required. admittance matrix,
Many networks are ill-conditioned, and by partitioning and
suitable normalization, various difficult sections with dras- k = ii, n - 1, H - 2, • • • , e + 1 and i, / = 1, 2, • • • , k — 1
tic component difference from the norm can be accommo-
dated. T h e application to symbolic analysis is not as T h e terms n'ijld,j and 'hj/djj are the new and old numera-
straightforward as initially anticipated' because of the tor and denominator, entries of the admittance matrix. T h e
4
interruption of common factor suppression mechanisms . pivotal condensation formula needs to be applied w — e + 1
However, the improvements due to partitioning can times and the reduction procedure to be repeated for every
amount to 90% increases in speed for some networks and sample point (a value of the complex frequency). Sparse
1. 2.
il
/ Active network,N
Next, elimination of node n — 1 produces
"B
Nooes u "JL. >.!= 1. (4)
d'.'fi
' ij
Clearly, for k node eliminations, [n, n — 1, u — I, • •
n — k -t- 1} C Jg, equation (4) would be of the form
FIGURE 1. General network N
8
(«»-)
'n-k + \,n-k+\ ; j = I.2.- - k
S
<J
(4>
matrix methods and optimal ordering of pivots can be used (5)
at this stage to improve the accuracy.
When analysing a necwork by partition the network may When all the internal nodes of S' are eliminated, the inter- B
be divided into an arbitrary number of parts, but for sim- mediate state of the semi-reduced system has the form:
plicity a two part division, Figure 2, is considered. Let N , A
B
PB = ( » t l * i , _ * i ) - n +
PB-
(b) y f j = 0 V/'and i € (S A - C A B ) with i « / The next stage of analysis may be executed by eliminat-
ing the internal nodes I , which is the set I = {» — ii , A A B
c m a v t a e a n v
( ) yfj, yfj ^ values for {/, ; } C C A B n — tig — 1, • - , n — fig — fi + 1} . On eliminating the A
In cases (a) and (b) the entries of Y are formed entirely given by:
from Ng and N\ respectively, whereas for case (c) both
subnetworks may contribute. ,r" **B
l „ k
\ A B B
ln"- \
Set yjj = njj/djj, then equation (2) becomes (7)
p b + i n z U ) A
U) \ t * i " ^
•j
i , 2, , « - Ms "
d-ij 1/ d?ij '
On eliminating all the nodes in I , we have A
B B B
n - = rAd + n d^ d- = d*d - (3)
IJ i] i] IJ tj' IJ 'J IJ
l B = {», n - 1, n — 2, • • • n — )1 + l } B
B B B Nodes,t
n- • = {n*d
"IJ y
'ij t] u + "d*n
t ] )nt ] ' n d-d
n m n j • - (d*d
,
' i j )dt j ' ' i n 'nm n nj'
, K L
Nodes
B
d'
"ij = (d*d
v
ij -)d-
i]' m d nj-
But,
* \ Nodes Nodes
\
Thus,
FIGURE 2. Two part division networks
k J
(in
1=1 i=**i
where
P = J | p,-and * = 1, 2, • • • , d - 1.
K
d
where
(12)
PA ~ ^ ' n - M ^ + l . n - ^ + l '
Y
Hl = L
v Yi +piY2
'/= ( w i 2 ) ' where is the admittance matrix of the kth interconnected
system, H% is the reduced matrix and p£ the last pivotal
i=3
numerator.
Selecting A'2 as the next subnetwork, analysis will yield When k = d - 1.
d
Y
<} = ( Wl2 Y
,) P +Pl 2
2
Y
Elimination of the interconnecting nodes of Nj would
r
Define the admittance matrix of the kih intercon- THE NUMERICAL IXTERCOXXECTIOX METHOD
nected svstem bv
General!)', the analysis of relatively simple networks bv the
G fc = G * - l + P k - i Y,,,A
k = 1, 2,
1 .d. interpolative approach is efficient and quite accurate. But
for very high orders and ill-conditioned systems the accuracv
0 , P = 1,
n
would tend to degrade after the solutions are interconnected.
Instead, numerical interconnection can be used when the
r
where P^ is the last pivotal numerator of the reduced sys- terminal matrix Y will be given numerically for every fre-
tem G t . quency point. The results would be more accurate bur
If k - 1, then G] = Y | . The first stage of analysis is to generally more effort is required and only the frequencv
eliminate all internal nodes of A/j producing the matrix domain response can be determined.
G^ and the polynomial p\. Next, with k = 2, G2 = +
P 1 Y 2 . Elimination of nodes in I2 and the interconnecting
nodes C j 2 gives G and pi- The procedure continues until
2
r r
k = d, when the required solution is given by Y = G j. THE SYMBOLIC INTERCONNECTION METHOD
The network partition method relying on artificial load-
10
i n g can be derived from the theory of simultaneous coup- It is possible to directly express the symbolic terminal
ling. Let Yfc be the contribution of the kth subnetwork to matrix of a two part interconnected system in terms o f the
that part of the admittance matrix G j which is affected by two individual solutions and hence substitution will produce
both the previous combined system and the new part. Thus; the general partitioned analysis.
For simplicity, consider two subnetworks N , A/ with A fl
G*-Gl_ I + P*_,(Y -Y k t + Y )
t only one interconnecting node. Y ^ , Y are their terminal f l
(13)
where 0
VUPB AIPB
= G,„ F* - l . and Y • Y, - Y; 0
t
AIPA A\PA (14)
N d + d D d
"11 = P B n u A i ^ A A x A v u = n ? i
B d = d d
Independent coupling "12= - « 1 2 1 2 . 12 U \ 2
d d
< = A A l n 4 l " n 4 A 2 » l l V p A >
;=i
d
°ii = * A A A x
where
B
and similarly for D j.
B
d The terms D ^ j , N ,Df 1 1 may be obtained f r o m the
s
individual analyses of and . V in the same way as njy,
P U P d f j , nfj, d f j are obtained.
i=i It is possible to continue the analysis in an iterative man
ner by considering n-, d- as the solution of a single part
the last dummy pivotal numerator of the above system. and selecting a new part. However, the terms N J J , D J J and
r
On elimination of all the interconnecting nodes, Y = Yf. p for the analysed interconnected system need to be com-
2 3 4 3 2 2
T c = 44(t - l ) , * + ( 7 r - 13r + 9 8 r - 29r - 204)<i
and 4 3
+ ( 6 r + 26r + 75r - 57r + 159W - (39» + 40r 2 3 2
+ n d
P = p «22''llf l P,\ 1\ 22 - 14r),
The major error at this stage is polynomial division and
as this is not encountered in the formation of Yjj, computa- Symbolic interconnection:
tion error is expected to be kept to a minimum. In fact, the
2 2 4 2
accuracy of the interconnection polynomials is only slightly T D = ( 9 r - 8r - \ )d + (13r + 169r - U6r)d -
inferior to that of the fundamental ones.
(18r + 2)
An attempt to generalize symbolic interconnection to
cater for a number of interconnection nodes has encoun-
The degree of approximation in the estimated computer
tered problems with the enormous number of polynomial
time is displayed in Figure 4. The actual times are taken
combinations required.
from computations on an ICL 1905E, and the difference
between actual and estimated is accounted for as routine
COMPARISON OF THE NETWORK machine calling times. Figure 5 shows the relative merits of
PA R TITION METHODS the various partition methods and the following conclusions
The number of arithmetic operations required to be exe- may be drawn:
cuted in each method of analysis is a good indicator of the (i) Tearing the network into relatively large subnetworks
relative efficiency. Since the straightforward symbolic may not produce the best results.
analysis is the basis of all partition methods, it is also con- (ii) The analysis speed may fall with the increase in the
sidered as the basis for estimating the computing time. number of parts.
If T is the estimated computing time, then
(iii) There exists an optimum number of parts for which
T = (effort to load the numerical nodal admittance the computation time is a minimum.
matrix for all samples) (iv) It is possible to analyse large networks within practical
+ (effort to compute all the polynomial responses) limits.
+ (effort to invert the Vandemonde matrix) Conclusions (ii) and (iii) may not be immediately appa-
+ (effort to compute the polynomial coefficients) rent for the symbolic interconnection technique but they
Assuming an n-node, e-terminal network having N con- c
can be verified by observing the continuously decreasing
ductances, N resistors, NQ capacitors, N inductors, with
K L
separation of the curves for equal increases in the number of
R as the total number of coefficients of the polynomials parts. The curves will cross each other for very large net-
representing the active devices, then from the Appendix works. This phenomenon is due to the relatively low effort
required for interconnection.
T = (m + 1X12R + 32Q + 1 UN R + 12SN L + 84/V +c
In Figure 6 the existence of an optimum number of
parts is displayed. Comparing the network partition methods
68N ) C
+ 11 m(m + 1)
2 2
+ 26e (m -r l ) (15) 0
Actual
where m is the maximum order and Q is the total number
of entries of all matrices entered as data. 10
To facilitate comparison a very simple RC ladder network
has been selected. Equation (15) then reduces to Estimated
4 2 10
T= 1 3 » + 178n - \62>i,
4 I
complex conjugate and passivity properties have been used.
Now assume the network is divided into d identical sub-
networks each of r nodes;then for
1 1 1 1
0 II '
Independent coupling: 10 15 20 25 30 35 40
Number of nodes
4 } 2 2
T A = 13(r - l)d + 13(3r - 2)d + (13 3r - 1 lOr + I6)d FIGURE 4. Actual and estimated computing times for
4
+ ( 1 3 r + 178r + 80r - l\i)d 2
+ 110, straightforward analysis
Number of ports
Symbolic interconnection
Independent coupling
Double coupling
Simultoneous coupling
it may be seen that the symbolic interconnection has the minimum. Typical error performance of symbolic inter-
greatest speed, while the independent coupling method connection is shown in Figure 7. The high accuracy property
takes the second place f o r a relatively small number of of this method is a very attractive feature and it is hoped
parts. The performance of independent coupling rapidly that a similar characteristic will be repeated for the general
deteriorates for a large number of parts. multiple node symbolic interconnection case.
The first three network partition techniques have similar The comparative performances of the partition methods
7
accuracy, especially when polynomial partitioning is used. when applied to a conventional high input-impedance
However, the symbolic interconnection method shows very amplifier circuit, Figure 8, are given in Table 1; a two part
considerable increases in accuracy. As no interpolation is division is used. The times quoted are those experienced
required for the computation of polynomials of the inter- in computing the overall 2 X 2 admittance matrix in sym-
mediate interconnected systems, the error is reduced to a bolic form. Because of the relatively low order of the net-
work, the improvements due to partition are less dramatic
than encountered with higher order circuits. This is parti-
cularly true of the symbolic interconnection case whose
performance easily exceeds the other methods f o r higher
order networks.
Symbolic interconnection More impressive results have been obtained in the analy-
sis of larger active networks, for instance, partitioned analy-
sis (double coupling) of the SN 72741 amplifier improved
the analysis time by 70% to 4 sees. For medium sized net-
works up to 100 nodes the partitioned methods work quite
S 10 efficiently. A n active network with 144 nodes proved about
__Double coupling the present limit, the restrictions on network complexity
come mainly from accuracy problems in the resultant poly-
Simultaneous coupling
nomial coefficients. Very large polynomials generally have
Independent coupling a wide range of coefficient magnitudes and the smaller ones
are easily submerged by the consequences of a finite word
length. Further errors occur in the evaluation of the
sampled tesponses in very high order systems, although
numerical methods for controlling these are available.
2 6 10 14 18 22 26 30
Number of parts
Simple inversion of the Vandermonde matrix would often
lead to serious errors due to ill-conditioning. A number of
FIGURE 6. Speed comparison of the partition methods for methods are known for overcoming this; two alternatives'' *
a 61 node RC-ladder have been used in this work.
y
TABLE 1. Com oarative performance parameters o[ pa 'i-
lion methods in attiplifier analysis
Mogmtude
Figure of
Merit
Computer Straight-
Require- forward/
Method Parameter ments Partition
O -I5v
2N4250 T ^Hk
2N4250
200 ^
500 FM ~ i HI 2N4250
2N4B69A
43kfl
lOOkfl 3.3kfl
The pivotal condensation formula is used in reducing the = (m + 1)(112<V + \2SN R L + 84N + 68N C C + 32Q + 12K)
numerical nodal admittance matrix to its terminal form. At additions,
an intermediate stage, say, when reducing the kth order
matrix to one of order (k — 1) the formula is applied effort for Vandemonde matrix inversion using a Fast Fourier '
2
(k — l ) times. Transform version
If T £ is the effort required to reduce a matrix from
m
( 1)2
rcond = ( " l ) +
2 *" = 2de\m + l) 2
additions.
1
68N )
= - (m + 1)[«(» - l ) ( 2 n - 1) - e(e - l)(2e - 1)] G
6
+ 25(m + l ) [ n ( n - l ) ( 2 n - 1) - e(e - \)(2e - 1)1
For passive netwotks the nodal admittance matrix is
symmetric; using this properly gives + 1 lm(m + l ) 2 2
+ 26e (m 2
+ l ) equivalent additions.
P A P E R 20
3 2G
1. Introduction
T h e sensitivities of n e t w o r k responses t o element v a r i a t i o n s are v e r y
i m p o r t a n t i n c i r c u i t design a n d therefore e f f i c i e n t means f o r t h e i r c o m p u t a -
t i o n have been sought. A p o p u l a r m e t h o d has made use of t h e a d j o i n t
n e t w o r k concept ; i t is q u i t e f a s t a n d applicable i n b o t h t i m e a n d f r e q u e n c y
domains ( D i r e c t o r a n d R o h r e r 1969 a, b ) . M a t r i x i n v e r s i o n techniques are
a n i m p o r t a n t a l t e r n a t i v e (Sud 1975) b u t generally r e q u i r e r e p e t i t i o n a t each
f r e q u e n c y p o i n t . L a r g e change sensitivities ( G o d d a r d et al. 1971) are also
v a l u a b l e a n d have f o u n d a p p l i c a t i o n i n s t a t i s t i c a l design ( B u t l e r 1971).
D i f f e r e n t i a l sensitivities i n s y m b o l i c f o r m ( C u t t e r i d g e a n d D i M a m b r o
1970, 1971) t e n d t o be more general a n d e f f i c i e n t t h a n the single f r e q u e n c y
p o i n t techniques. T h e use of i n t e r p o l a t i o n is e v i d e n t i n these techniques.
Since i m p r o v e m e n t s t o the accuracy of the i n t e r p o l a t i v e methods have been
made ( F r i d a s and Sewell 1974, Singal and V l a c h 1974) t h e y h o l d greater
appeal. S t i l l f u r t h e r increases i n accuracy are possible using an i t e r a t i v e
c o r r e c t i o n procedure (Fridas a n d Sewell 1975).
P a r t i t i o n methods have been e m p l o y e d t o good e f f e c t (Fridas a n d Sewell
1976) i n t h e analysis of h i g h order a c t i v e n e t w o r k s . General i m p r o v e m e n t
i n speed a n d accuracy were experienced. I n t h i s w o r k the n e t w o r k p a r t i -
t i o n i n g methods are developed t o produce f i r s t - a n d second-order sensitivities.
2. Differential sensitivities
r
L e t ,A be a linear a c t i v e n e t w o r k w i t h n nodes, of w h i c h e are t e r m i n a l
nodes, and h a v i n g a nodal a d m i t t a n c e m a t r i x Y, F i g . 1. I n t h e usual n o t a -
t i o n G, T, C represent the conductance, inverse i n d u c t a n c e a n d capacitance
components of Y. A is an active device m a t r i x whose entries are r a t i o n a l
functions in s :
2
Y =G+ .4 + - + sC = - [V + s(G + A) + s C\
s s
J
\Y \
D D
7
where Y and Y are a p p r o p r i a t e co-factor matrices.
{ j D A s A ,-- = f ( y ) , where
r s
y is a general e n t r y of Y, t h e n
rs
2
^ = I I ^ ^ . V i , j = 1, 2 ... e and k = \ , 2 ... V (1)
-" t i m
cx . k r = 1 ,fi ty „
r cx .
k
T h e second-order d e r i v a t i v e s f o l l o w :
2 2
r V
c;r,.r.T,
I" „ I", .I"| , "I| cy rcy V
,. = 1 = p = rx m
ru
cx k
rv
cx (
2
" " ?V r ?/
+ y y i ^ _ L 4 ^ ! vA- = i , 2 ... V - \
,= i »= i r(/ , cx .cx, r k
\/l = k+ 1, A- + 2 . . . K (3)
c*D
A s i m i l a r expression results f o r . I n c o m p u t i n g e n t r y sensitivities i t
is o b v i o u s t h a t a passive component m a y c o n t r i b u t e t o o n l y one or f o u r e n t r y
terms.
Symbolic network sensitivities 27
^ = c Ji
Hfr,
and
?D
— = c„"
i} 1
where C , C' are a d j o i n t matrices, whose entries are d e t e r m i n e d in general b y
the f o l l o w i n g procedure :
2
«,/> = { V ' - ' » n , , " - ' - » « ' - ' > } / n < ' - V , , A I . j = l . 2 - ^ i A + +
) , r 1 ( r 1 , , r I ( r , , ( , 2 , l 2
6,/' = {6 / y - »»a- - -» f t - «*-y - }/" '- *- i.^i. + J = . - »
f o r & = ra, n — l , . . . e + l ; i = 1, 2 . . . & — 1 and r=n—k+ 1.
1
I n i t i a l l y n,./°» = n , ; 6 / ° » = 8, ( K r o n e c k e r delta) ; a*- > „
7 f 7 +li = 1. When
r = n — e these operations are h a l t e d and n = n (n-e) i j
{ r )
i j
rf <D = n . -e-l>
n t
N o w let
,/. = •„...<»-«>, j = 1, 2 ... n
f
„..'(0) = n ..(«-o and &,./«»» = 6 / " - > f
l , r I ) , r 1 ( r 1 , r f
« y<'>={» y«- V-' - -««-' - V - }/ *- i
I 1 +
< 1 (r l,
6 '''» =
lV {6 /''-iS.,'fr->'-n .' '- V - }/^ i
/ (7 +
1, 2 ... e
c P 9 = b
k »j"< j = e+l,e +2 ...n
<V" = °> i = 1, 2 ... e : jV<7
C "9 = 6 "
Cjj " = 0, 1 1
i= I , 2 ... e a n d M ^ ; j = 1, 2 . . . n
p, (o
c>/ ={^•/\ ;'-'^•;'^/'}K•^l. :• -l' ^ *,j=«+1, e + 2 . . .
) ( f *
6 <0)
S i " ' = K " n - V „ / } K i . ' . i = e+l,e n
+ 2...n + e+1
C//" = 0, i = e + 1, e + 2 ... w ; j = 1, 2 . . . e ; j ^ q
1
Finally
Cjj' = c - = 0, 1
0 i=l,2.../i: J = 1, 2 ... e
c
/ / = V> i, j = e+\, e+ 2 ... n
r J rf
The p o l y n o m i a l s of } ' , C' and C' are therefore d e t e r m i n e d b y i n t e r p o l a t i o n
at an a p p r o p r i a t e n u m b e r of samples. Hence a classical s e n s i t i v i t y measure
N N8
the matrices A A as
V B
Y A = - ^ A A and Y B = ^ - A B
PA PB
A t y p i c a l n u m e r a t o r e n t r y of Y , w i l l be
n I = n a
V uu^Ph +n t B j B p A (6)
= LL-f ^ (7)
?.T, P> CX k P CX k
(
'
2 8 T
c N^ _ _L_ c A / t t fdN i } oP dN if cP
1
dx cx k t P dx cx, k P \ cxk cx, cx t dx . K
2
N u cP cP N u cP
i
P cx k oxi P dx dx k {
where t = n — e . I l
dx dx,
k r =
£ 1 s =
£
1 p = i £< 9 = 1
- / 9 ^ /
ydnjdnp/J\ \ dx
f d n) \/ \ fdxd n
k
r
t
p /
a n d t h e d e r i v a t i v e s of Z) are p r o d u c e d s i m i l a r l y .
I J D J D
Now if Z and Z
r s are t h e general entries of a d j ( Y / ) a n d a d j { Y j )
r s
ax, p £ V A d x
k ) p
fo k
3x . y \A 3x,
x X S x x
P r =1 4 =i \ ^rJ )\^ k^ l) P \ k dl dx x dx k
2
N u / d P \ / d P \ N tl dP
,13
- ' ' ' - ' ' ^ ( W t e J - ' - F S ^ >
A g a i n , s i m i l a r expressions o b t a i n f o r the d e r i v a t i v e s of D.
30 C. Phrydas and J. I. SeweH
T h e f i r s t - a n d second-order d e r i v a t i v e s as w e l l as t h e n e t w o r k f u n c t i o n s
can be c o m p u t e d f r o m t h e analyses of s u b n e t w o r k s . The network f u n c t i o n
n u m e r a t o r s , last p i v o t a l n u m e r a t o r and t h e i r d e r i v a t i v e s are a l l t h a t are
r e q u i r e d . T h e process m a y be repeated f o r an a r b i t r a r y n u m b e r of p a r t s .
4. Results
T h e p a r t i t i o n e d m e t h o d f o r d e r i v i n g f i r s t - a n d second-order n e t w o r k
sensitivities has been a p p l i e d t o b o t h passive a n d a c t i v e n e t w o r k s . Mainly
f o r comparison purposes a 30-node R C ladder was chosen a n d t h e sensitivities
of the 2-port a d m i t t a n c e f u n c t i o n s w i t h respect t o a n u m b e r of t h ° elements
computed. U s i n g a s t r a i g h t f o r w a r d technique w i t h o u t p a r t i t i o n a n d com-
p u t i n g f i r s t - and second-order s e n s i t i v i t y p o l y n o m i a l s of t h e above f u n c t i o n s
w i t h respect t o six elements r e q u i r e d 105 sec a n d 16-4 K words of store on
PARTITION
LINE
O «15V
rVVV-
75kf?
l/Ps
39kO
30pF
® 250
0/P
75kfl
50fi
®
kQ 50k0 IkO 5k0
0-15V
Appendix
Let u = f ( x , x, 1 2 x ) where x = x ( t , t , . . . , t ) a n d t . are independent
n i j 1 2 m k
variables.
A s s u m i n g the c o n t i n u i t y of d e r i v a t i v e s i n v o l v e d gives ( S o k o l n i k o f f 1939)
I L - V d f d x
< i A n
( A l )
d t k r f h ^ d t b
E q u a t i o n (9) f o l l o w s d i r e c t l y f r o m ( A 1).
D i f f e r e n t i a t i n g eqn. (A 1) w i t h respect to t,, gives
dt ot,
k dt, dXf dt k
\ d X f d t , dt k c.r,- ct dt, k
2
? / \ CXj " cf cx
+
r.r, \ 3 t J ct k ,?i ex, dt ct. k
2
C X
y IL ™l\ f f i 4. y f L -1 :
since i n n e t w o r k analysis
ex,
^ = 0, V>,J
and
2 2
8 UI 8 U|
= 0
REFERENCES
PAPERS 24 + 30
3 3 5
/ = YV. /f y\i y\ n
y\% yf
The nomenclature of 11 ] is retained, / ' 7 ° being the vectors of even
e
and odd phase currents; V , V° are the even and odd nodal voltage
K y%\ y'n'n AS A&
ot
e
vectors; S , S" the even and odd switching matrices; 1° the even /'? ylf v
yw y f f
y1S y?
and odd network current matrices; C the capacitance nodal admittance
matrix; and z" = e~1 n
where r is half the clock period. The Y matrix
>°n. y°n\ AS y°n°n. -Vn -
is always symmetric for a passive network, and the main diagonal terms where node 1 is input and node n is output. In practical networks quite
are positive, but in contrast to the ordinary definite admittance matrix a number of the ^-parameters will be zero. Cleaily the input admit-
the signs of all off diagonal terms are not necessarily negative. Some tance during even phase switching is
rows and columns are zero so, by definition, these and the correspond-
ing variables can be eliminated; however, it is wise to examine the inclu-
sion of active devices before executing this step.
-y\l
When active devices are present in the network, assemble the nodal
matrix for the passive part Y as P indicated. Now an ordinary ampli- Similarly the voltage transfer functions, for even sampling input and
fier will be operational during the even and odd switching phases and output
therefore will contribute equally to even and odd equations hence
- m
yf ~ AS,
[/f"H~°~i vl [v°] for even sampling input and odd sampling output
yjL -A$_
y f ' ....
The Y matrix, for instance, may be any of those quoted for various
A provided, of course, these functions exist practically. The number of
types of amplifier [4) or indeed the admittance matrix of any active characteristic functions available in an S C network is obviously poten-
device. Addition of the active and passive matrices will yield the com- tially greater than in a conventional active network.
plete nodal matrix When ideal operational amplifiers are the active elements in the net-
work it is possible to utilize conventional constraint theory [5] in the
analysis. Consider an ideal voltage amplifier with voltage gain -m and
As expected, Y is not symmetric for an active network. connected between nodes i , / , hence
Some simple rules for the assembly of the matrices now emerge.
e -V,
When writing down the switching matrices S , S" all port voltages must ; -m Vj or V, =
be retained as independent variables and all voltages at amplifier ter-
minals must remain as independent variables. When constructing the
current matrices I', 1° the currents at either the network ports or the in an S C network the constraint becomes Vf = -V°lm, Vf = -Vflm
amplifier terminals must not become zero. hence the column i(odd) is divided by m and subtracted from K,(o<Jd),
column i(odd) is then discarded and similarly with columns i(even) and
After the assembly of the Y matrix, inspection will show the corre-
I {even). The rows involving i f and / / are redundant, since these ex-
sponding rows and columns that are zero, these are now eliminated press the current outputs from a voltage source, and are, therefore,
to produce a reduced matrix Y R. It is then possible to determine, by discarded. The matrix Y once more is square, but cofactor techniques
C
cofactor methods, the appropriate transfer functions from YR. For have now to be used in evaluating the required characteristic functions
as there is no longer a direct relationship between the rows and columns
general analysis it is probably more informative to apply pivotal con-
of the matrix and the network nodes. Of course, when m — ~ the pro-
Manuscript received October 9. 1979. cess merely amounts to striking out columns i(odd), i(even) and rows
The author is with the Department of Electronic Engineering, Uni-
versity of Hull. Hull HU6 1 R X , England. j{odd) J (even).
85
P R O C E E D I N G S O F T H E I E E E , V O L . 68, NO. 2. F E B R U A R Y 1980
When a difference amplifier is present then this may be characterized With even input sampling V , = 0, / i = 0 so remove column and row
by V = m(Vi - V,). Hence K^/m = Vf • Vf and V%lm = K," - Vf and
k
1 {odd) and from the remaining matrix
when m - ~ , = f / , f ° = V° thusi ccolumn i(even) is added to col-
umn /(even) and the former discarded; column i(odd) is added to col-
umn j(odd) and the former discarded. Again, the current outputs from
the voltage source are arbitrary, hence rows k(even) and k(pdd) are
discarded. At the present time, the application of different amplifiers -
in SC networks in not considered especially practical for capacitor " Alt
grounding reasons since no virtual earth exists at the input. 2
Crd-z" )
Consider the circuit of the lossy differentiator shown. Take the pas- !
sive network alone, then following the rules given: " C 3 -z- (C 3 - Ci)
which is the transfer function of a lossy differentiator.
"1 0 0 0" "1 0 0 0"
For odd input sampling V\ = 0, l \ - 0, so remove column and row
0 1 0 0 0 1 1 0 Heven) and from the remaining matrix
s° = 1° =
0 1 0 0 0 0 0 0
.0 0 0 1. .0 0 0 1.
0 1 0 0 0 1 0 0
e
s = 1' = _ -C (i-z-'\C -C ))
l 3 2
0 0 0 1 0 0 0 0 2
C (l - ; - | C
3 3 -c \)
2
0 0 0 1_ .0 0 1 1.
- £l
" c, -c, 0 0"
c 3
/f -c, c, + c 3 0 -c 3
r'c, -z-'(C, +C ) 3 0 «?
/f 0 0 0 0 o 0 0
'f 0 0 c 3
0 -z-'(C 2 -C ) 3 0
/? -z-'(C, + C ) 0 3 -z"'(C 2 - c )
3
-Ci C, + C 2 +C 3 0 -c 3
*?
/? 0 0 0 0 0 0 0
0 0 -i-'c, 0 -C 3 0 /4.
Eliminate zero columns and rows 3(even), 3(o{j d'), to give YR. Apply , REFERENCES
the amplifier constraints, hence remove columns 2(even), 2(odd) and 11J C . F . Kurth and G . S. Moschytz, "Nodal analysis of switched-
rows 4 (even), A(odd) to give matrix Yc capacitor networks," IEEE Tram. Circuits Syst., vol. CAS-26, pp.
93-105, 1979.
[2) , "Two-port analysis of switched-capacitor networks using
four-port equivalent circuits in the z-domain," IEEE Trans. Cir-
cuits Syst., vol. CAS-26, pp. 166-180, 1979.
"if C, 0 -z-'C, 0 >f" 131 K. R. Laker, "Equivalent circuits for the analysis and synthesis of
switched capacitor networks," Bell Syst. Tech. J., vol. 58, no. 3,
n -C, -C 3 z-'C, z"'C 3
pp. 729-769, 1979.
[ 4 | J . I . Sewell and F . W. Stephenson, "Matrix tables for the general-
n -z-'C, 0 C, 0 If ized 5-terminal amplifier," Radio Electron. Eng., vol. 37, no. 4,
pp. 247-255, 1969.
z-'C, -z-'(C, - C j ) -C, C
[S 1 A. Nathan, "Matrix analysts of networks having infinite gain opera-
A - 3. tional amplifiers," Proc. IRE, vol. 49, p. 1577, 1961.
86
3 3 7
PAPER 25
5
well be replaced by Ihe double d.f.t processes. However, this If y 2 is assumed to be a real constant, normalised to unity,
2
is not a matter for consideration, because the main purpose of then in a switched network the frequency-dependent current
this letter is to show how to apply the sampling theorem to term during any lime interval n is given by
s.a.w. filter analysis. T h e discussion can be applied easily to
other time-consuming s a w. filter analysis, such as equivalent
circuit analysis or bulk spurious response analysis, as long as
t ,„ and
m are determinable beforehand.
References
where T is the sampling period. This becomes
1 P E N U N U R I , D.: A numerical technique for SAW diffraction simula-
tion', IEEE Trans., 1978, MTT-26, pp 288-294
2 SAVAGE, E. B.: 'Fast computation of s a w. filter responses including
diffraction', Electron. Leu., 1979, 15, pp. 538-539
3 SAVAGE, E. B„ and MATTHAEI, G. L.: 'Compensation for diffraction in 1
. -' 1 6
SAW filters'. Proceedings of the I E E E ultrasonics symposium. New
Orleans, 1979 (to be published)
4 PAPOULis, A . : 'Signal analysis' (McGraw-Hill, New York, 1977),
chap. 5, pp. 139-145
IV*""'" I tb
e °""<'V(x)dT di
p-lh./S.H I I»./»il.j,'f ) J,
O013-5194/80/12O46O-O3S1.50/O Tb, J f I d r
"° f -(6o/t,M»-0
f K ( t ) d T + I ( t ) dl
V
Tb
° K ( r ) dx f m d t
Tb 0
a n c
Let a / f c T = a, 60/^1 = 0.
0 0 ' since an s.c. network is a
INCLUSION O F AMPLIFIER FINITE G A I N discrete system
A N D B A N D W I D T H I N A N A L Y S I S OF
S W I T C H E D - C A P A C I T O R FILTERS
/ „ = - « £ e-"--"V(.) + a £ V(i)
1=0 « «= 0
Indexing terms: Amplifiers, Switched-capacitor networks
+ «"£' e-""-<-"V{i)-0L "X V(i)
The exact mathematical solution of the amplifier finite gain i =p i=P
and bandwidth problem in switched-capacitor filter networks
is presented. The theory is incorporated into a computer Applying the convolution theorem of the ^-transformation'
analysis program and typical results for a filter circuit are yields
included.
or in closed form
25
or / = [?, + Y )V.
A
2
, M ' ~-
y
(1 - z- e- "")2 2
+ ^IzlUDTJl
(1 - z ' e - )
VIA' 2 l e T [
/ ' , P and V', V are the phase-1 and phase-2 currents and
voltages; V is the passive part matrix assembled in an accepted
p
2 2 1
manner. Because of frequency-dependent amplifier par- Now / = / ' + / , and V = V + V , so splitting the current
ameters, the active matrix Y will not have the simple form as
A equation into direct-phase and delayed-phase components
indicated previously ' gives
F o r a typical operational amplifier, the gain can be ex-
pressed as m = - a / ( f e + bi s), a single pole in the s-domain.
0 0
-, az-'(l - 2
z-'e- » )V'{z)T
_ Z L ! ) ^ ^
A simple voltage amplifier has a Y matrix given byA 2 2 T 2 2>T
(I - z e- » ) (1 - z~ e~ )
0 0
1 7 2 2 r
(1 - z - ^ - ' ) (1 - 7 - e - " )
40 \ m
1000 It is then glued onto the fibre end.
This process is complicated, and it is quite difficult to hold
50 the microsphere lens. We report here a simple, accurate and
\ production-oriented technique for lens mounting. A typical
6a \ microsphere lens attached to a fibre end is shown in F i g . 1. I n
70 \ K300 the mounting procedure, the end of the fibre is first properly
m
WOO 1.001s cleaved and cleaned, and adhesive is applied to it by dipping.
-m Wis lueiil
References
1 SEWELL, J. i.: 'Analysis of active switched-capacitor networks', Proc. Fig. 1 Photograph of mounted microsphere lens
IEEE, 1980, 2, pp. 292-293
2 K U R T H , c F . and MOSCHYTZ, G : 'Nodal analysis of switched-
capacitor networks', IEEE Trans.. 1979, CAS-26, pp. 93-105 With the aid of micromanipulators and a low-power micro-
3 JURY, E. I . : 'Theory and application of the z-transform' (Wiley, scope, the fibre end is roughly centred over a microsphere lens
1964) and lowered so that the epoxy contacts the lens. T h e epoxy
4 ALLSTOT, D. j . , BRODERSON, R. w., and GRAY, p. R.: 'MOS switched then holds on to the lens by surface tension so that it is picked
capacitor ladder filters', IEEE J. Solid-State Circ, 1978, SC-13, pp. up as the fibre moves upwards. D u e to the combination of
806-814 adhesive surface tension and gravity, the lens moves until it is
on, or very close to, the fibre axis. T h u s the technique can be
0013-5194/80H 20462-02$! .50/0 described as self-centring.
10
09
The light power which can be launched into an optical fibre -18 -12 -6 0 6 12 18
from a diode laser by simple butt coupling is small, being lens offset,urn [j^]
limited by the fibre acceptance angle. Several coupling Fig, 2 Calculated dependence of normalised coupling efficiency on lens
techniques have been reported which increase the coupling offset. Curves for offset parallel and perpendicular to laser junction plane
efficiency.' 6
O f these, the microsphere lens glued onto the are shown
6
fibre endface provides one of the highest coupling efficiencies. parallel
The high efficiency arises in part from the ability to select perpendicular
-lfT
Spacing d 1 - z
r
«(1 - «
T
)V'(z) az-'e '^l -f-' )K'(z)
P= J r
l - z " V »
conductor
4 5 6 7 8
?
spacing d. p m ( « l 0 )
Fig. 3
a Calibration rig
h Frequency change against spacing
Reference
1 DAVIES. R , BOOTH. P- and N F W T O S . B : An improved T R A P A T T
oscillator'. IEEE Trans.. 1980. M T T - M , p 171
n
" Technique reported b> the '-ssey C o some years ago on their
F E T M I C amplifiers
PAPER 26
Digital active network analysis
F.G.A. Coupe, B . S c , Ph.D., and J.I. Sewell, B . S c , Ph.D., C.Eng., M.I.E.E., Sen. Mem.I.E.E.E.
Abstract: An analysis technique is presented and developed which enables analogue active networks and
digital filters to be combined into a new type of neiwork. ihe digital active network. The poiemial
effects of limit cycle noise aic extensively considered and for a practical example both the results of
computer analysis and actual measurements aje given.
1
T (13)
1
•nz) 0- 0
8 1
0
•s i (17)
i
This result is important in the analysis which follows.
0 0-
3 Digital admittance matrix
A digital admittance matrix may be formed by intercon- where the general element^,- may be defined:
necting digital amplifiers such that each digital amplifier
simulates an element of an admittance matrix. As each K, = (product o f untransformable elements in ith row)
digital amplifier would normally contain an a.d. and d.a.
converter, a considerable simplification can be made Now, by definition,
IEE PROC. Vol. 127. Pi. CI. So. 4. A L'UL'ST 1 980 182
Each sampled term may now be Z-transformed to give the Hence .-!.\>(rl is superseded by a voltage-transfer vector
digital admittance matrix: Q (:)
N where
¥(•) l
-[Z\Y-i}]- [Z\Y-dY{_s)\] (21)
<h<:) "</(-)
(28)
those elements o f Y(s) derived from digital amplifiers Hence a digital active network constructed solely f r o m
which include amplitude quantisation. Thus the noise digital amplifiers cannot show limit cycle noise providing
output current l ( s ) will be N
that a.d. convertors are identical and aligned.
The limit cycle oscillation analysis has considered only
the effect of amplitude quantisation producing amplified
Ms) = Y (s)V%(s)
N
(23)
quantisation noise. The actual amplitude depends on the
By taking the Z-transform absolute size o f AN and thus, for a given dynamic range,
a large number of quantisation levels is preferable. The
l (:) = Y (z)V (i)
N N N (24) presence o f numerical round-off in the other stages o f the
digital amplifier will also affect the limit-cycle
Now l ( z ) will excite the whole digital active network
N
6 , 7 8 , 9
criterion. ' However, coefficient accuracy will only
by way o f the digital impedance matrix: 10
be a problem in digital active networks i f the coefficients
V (z) = Z(z)Y (z)V (z)
0 N s (25)
hypothetical;
and thus a voltage transfer matrix A {z) N may be defined: 1
samplers
A (z)
N = Z(z)Y (z) N (26)
5 Element resolution
Now by definition {g + s C i ) and sC z cannot be trans-
In a digital amplifier the presence of an amplitude quantiser formed and thus Y may be written as
v
for the element >>,-,-. The fractional error in y j will be t Thus the inverse is
Aj'rj _ A/ AK
(36) 1
•^7 0
1 + sC,
However, by definition A F < £ l.s.b. and thus / will not (41)
change, making A / = 0. Hence 1
sC2
AV
(37) Further
{1 -exp(-sT )} s
e x - s K r
Z^ T +, s^C , ^)
sfe P ( ' »)
{1 -exp(-sT )} s
- «2 7- exp ( - sA. r )
2 s
1
2
s C 2 J (42)
0
(38) C, \z-oc
-a 1 (43)
M YUM i 1
0
Hence no digital amplifier
transadmittance may be C 2 z - 1
1
measured to an accuracy greater than —• where
M
6 Example: digital gyrator a = exp(-£?/C,) (44)
A 2-port capacitively loaded digital gyrator forms a suitable Eqn. 42 similarly gives
example to illustrate the analysis o f a digital active
network, and i t is shown i n Fig. 5. The basic equations
describing the gyrator may be written down:
/ , ( s ) = fe + s C , ) K , (*) + / , ( * ) V *(s)
2 _ I
( 3 9 )
A ( s ) = -Mn)V*^) + sC V (s)
2 2
C2 ' z - \
where
(45)
(' - exp (-sT )
s
= g1
fx(')
r exp (-sK,T ) s
Thus Y(z) may now be obtained
(1 - a ) z -
• (sr>
? / \—<"f,(s)
V digital y
Y{z)
amplifiers
> > + K
: + 1
(z-l) Parameter Parameter meaning Value
(I -a)
c, ** sampling frequency 26-6 kHz
Z(z) = T, sampling period 18-75(is
(z - a ) c, port-1 shunt capacitance 9-6 uF
C, port-2 shunt capacitance 10uF
. c,cV C 2 9, .9. transconductances 10mS
R. port-1 shunt resistance 5 0 , 1 0 0 . 2 0 0 , 300
M(2)
400, 500 P.
If, , K fractional delay 1
(47)
where 7 C o m p u t e r analysis
K K
(1 -a)gig2 The denominator {M(z)} o f the digital impedance matrix in
M{z) = z >+ '(z l)(z-a)+T, (48)
gC 2
eqn. 48 was analysed to find the pole positions, and the
modulus o f the roots nearest the unit circle are listed in
The digital gyrator as defined is a mixture o f analogue Table 2. Gearly these roots will lie within the unit circle
components and digital amplifiers, and thus it is possible and hence the digital gyrator will be stable.
that limit-cycle oscillations may be present w i t h certain sets
Table 2 also shows the resonant frequency and the
of component values. Thus the voltage-transfer vector
magnitude o f z and z , at that frequency. Figs. 6 and 7
n 2
(see Section 4) must be derived.
show the frequency responses for the magnitudes o f z n
Thus the voltage-transfer vector Q (z) N may be written A digital machine was constructed (Fig. 5) which consisted
900 r
* z * > * ' ( - l ) + * * , l ( ,
~ a ) z computed
2
gCi R = 5O0n
s
(50
QN(.Z)
~g7 , Z\SiT^
- (z — a)z
K i + 1
R»=500n
c,c 2
M(z)
R«=tOnn R.=400rt
In order to determine whether limit-cycle oscillations can
be present theoretically, the two elements o f the above
Ft. = 300n R. =300fl
vector must be evaluated for a range o f frequencies, for
values o f z, and for various values o f the components
R = 200A R = 20011
g,g\ ,g2,C and C .
x 2
R.=wn R = 50n
elements z and z o f the digital impedance matrix were
M 2 ]
s
Shunt resistance, P.
10 Acknowledgment
of two cross-coupled identical digital amplifiers. Each The work reported in this paper was supported by the UK
digital amplifier contained a 5-bit voltage-sensing a.d. Science Research Council.
convertor, or 4-quadrant digital multiplier and a 10-bit
current generating d.a. convertor. The relevant parameters 11 References
and component values are shown in Table 1.
1 SHINNERS. S.M.: 'Control system design' (Wiley. 1964)
The digital amplifiers were aligned by setting the 2 T O U , J.T.: 'Digital and sampled-data control systems' (McGraw-
input to the 4-quadrant multiplier to maximum and Hill, 1959)
adjusting the current generating d.a. convertor refer, nee 3 J U R Y , K.I.: 'Sampled data control systems'(Wiley, 1958)
voltage until a transconductance o f lOmS was obtained. 4 B L A C K M A N , R.B.: 'Data smoothing and prediction' (Addison-
Wesley, 1965)
The a.d. convertor was multiplexed between the two
5 B E T T S , J.A.: 'Signal processing, modulation and noise' (English
amplifiers but because of circuit complexities two Universities Press, 1970)
separate d.a. convertors were used. 6 OPPENHEIM, A.V., and W E R N S T E I N , C.J.: 'Effects of finite
The frequency responses for the magnitude o f the register length in digital filtering and the fast l-'ourier transform',
Proc.IEEE. 1972, 60, pp. 9 5 7 - 9 7 6
digital-input impedance z and forward-transfer
n
7 A V E N H A U S E . E . : 'On Ihe design of digital filters with
impedance z were measured and are plotted in Figs. 6
2 i
coefficients of limited word length', IEEE Trans., 1972, AU-20,
and 7 as a broken line. The measured frequency responses pp. 2 0 6 - 2 1 2
have taken into account the limit cycle oscillations present 8 J A C K S O N , L.B.: 'Roundoff-noise analysis for fixed point digital
when « „ = 200, 300, 400 and 500 fi. These osculations filters realised in cascade or parallel form', ibid., 1970, AU-18,
pp. 107-122
were found as theory predicted.
9 S A N D B E R G , I.W., and K A I S E R , J I".: 'A bound on limit-cycles
in fixed point implementations of digital filters', ibid., 1972,
9 Conclusions AU-20, pp. 110-112
10 WHITING, P A.: Tinite word length effects in digital processes'.
The main advantage of digital active networks are that they Electron, 1974, pp. 4 6 - 5 1
PAPER 28
3 4 9
References cess. The nodes in the reduced mairi.x are renumbered lo gen-
1 SCHMIDT,L.-P.. and I T O H . T . : 'Spectral domain analysis of dominant erate a minimal sel A/' for lime interval i. Repeating this for
and higher order modes in fin-line'. IEEE Trans'. 1980. MTT-28, all the lime intervals introduces a global minimal set of nodes
(9) A f and a minimal sel of equations. For a network . I with \
s
2 M E N Z E I . . w., C A L L S E N , H . . A D E L S E C K , li.. and S I C K I N G , K.; 'An imegra- physical nodes, the total number of relevant nodes in a time
led Ka-band receiver front-end'. 11th E u M C . Amsierdam. 1981. 1 :
interval is m = .\ - n' . where n[ is the number of swiiches
s
symp. dig., pp. 361-364
closed during interval i. then Af = {1. 2. 3 ... m'), and the
3 M L T N E L , H . , and C A L L S E N . H . : 'Fin-line PIN-diode attenuators and
lotal number of rows or columns in )' is m = , m'. For
swiiches for 94 G H z range'. Electron, ten., 1982, 18, pp. 541-542
sensitivity and other studies it is essential not to lose the orig-
inal node identities, so an index vector is used to record all
0013-5194/82/190839-02$!. 50/0 movements of rows and columns. The complete index vector is
a concatenation of the interval index vectors and simply maps
the definite capacitance matrix t ' i n t o all the Y''.
However, it is apparent that Y will still be a very sparse
matrix. Sparse techniques could be used, but these can be
avoided by the following method. Note that all the on-
diagonal partitions of Y are frequency independent and that
C O M P A C T M A T R I X S C H E M E FOR USE IN all off-diagonal portions have a multiplier and are fre-
C O M P U T E R A N A L Y S I S OF S W I T C H E D - quency dependent. Commence with interval p and remove, by
CAPACITOR N E T W O R K S pivotal condensation, all nodes apart from the input and
output nodes and any others to be retained for sensitivity
calculations and the like. This yields a matrix of the form
Indexing terms: Circuit theory and design, Switched-capacitor
networks. Computer analysis 0 0 ... 0 -'Y;-'-' -r'Y!'
Y" :
- 1 y i 1 Y" 0 0 0 0
The matrices that normally occur in the analysis of multi- 0 z-'Y" 0 0
phase switched-capacitor networks are quite large and very
>"' 0
sparse. A technique is presented which reduces the matrix 6 0 0 - . - - ' > ' ' - ' •'-' Y ' ' 0
order and increases the fill. Storage requirements are mini- 0 0 0 ... 0 - -'Y;-'-'
:
Y"
mised and the arithmetic handling of large arrays is reduced
without recourse to sparse matrix procedures. where Y'l indicates a reduced matrix. This process is repeated
down to interval 2 and yields
For the analysis of multiphase switched-capacitor (SC) net-
works a number of basic matrix representations are possible. K-ICr" -ii"".--
-r'r;' 1-IL--
The modified nodal approach ( M N A ) does have advantages
deriving from the structure of the matrix, but will generally
require manipulation to avoid singularities and sparse tech-
niques to improve the numerical routines. Here a definite f is much more compact than Y, of lower order and greater
1 3
nodal admittance m e t h o d ' is favoured since it allows simple density. In fact Y need never be created, the submatrices of f
reduction in order and increased fill techniques to be im- can be generated sequentially when needed and the space
plemented. Network parameters can be computed in a reused for other submatrices in the successive condensation
straightforward manner by condensation and cofactor tech- steps. For transfer function calculations only array storage of
niques. m 2 1 2
[ m + 2(p — l ) ] -t- [ m " ] is required, compared with m for 1
i
IC] + lY ]V ({n+l)T)
A
'l =
The definite matrix Y contains all the conductance terms due
A
or / = YV.
The C' are of lower order than C in general and can be
derived from it by contraction representing the switching pro- v
1 KURTH, c. F., and MOSCHYTZ, G. S.: 'Nodal analysis of switched If the junction is perfect, if the power meters are perfectly
capacitor networks', IEEE Trans.. 1979, CAS-26, pp. 93-105 matched, and if the directional coupler D C is perfect the
2 HOKENEK, E., and MOSCHYTZ, G. s.: Analysis of general switched- centres of the circles will be at — 2 and at 2 exp [ + j(n/3)] in
capacitor networks using indefinite admittance matrix'. IEE Proc. the complex reflection-coefficient plane. A change in the refer-
G, Electron. Ore. A Syst., 1980, 127, (1), pp. 21-33 ence planes, maintaining symmetry, will simply rotate the
3 SEWELL, j. i.: 'Analysis of active switched-capacitor networks', three centres around the origin without changing their relative
Proc. IEEE, 1980, 68, pp. 292-293 positions.
4 LAU, j . , and SEWELL, J. l.: 'Inclusion of amplifier finite gain and
We ask how far from these ideal positions the centres may
bandwidth in analysis of switched-capacitor filters', Electron. Lett.,
1980, 16, pp. 462^*63 move if the five-port whilst remaining symmetrical, reciprocal
5 HERBST, D., FETTWEJS, A., HOEFTLINGER, B., KLEINE, U.. N 1 E N T I E D T , W., and loss-free, is not perfectly matched, again assuming a per-
PANDEL, H., and SCHWEER, R.: An integrated seventh-order unit fect directional coupler and perfectly matched power meters.
element filter in VIS-SC technique', IEEE J. Solid-State Circuits, We assume, therefore, that its scattering matrix can be written
1981, 16, pp. 140-146 in the form
6 NOSSEK. J A., and TEMES, G. B : 'Switched-capacitor filter design
using bilinear element modelling', / £ £ £ Trans., 1980, CAS-27, pp.
y a p p
481^*91
a y a fi
[S] = P a y a (1)
0013-5194/82/190840-02$!. 50/0 P p a y
a p P a
1
a = a0 + da
Introduction: Hansson and Riblet have demonstrated that a
five-port symmetrical star junction can be used as a key com- P = Po + W (4)
ponent in a six-port reflectometer of the kind described by
2
Engen in an excellent review article. T h e basic properties of and treat Sa, dp and y { = | >• | exp (j<t>)} as small quantities.
such junctions have been known for many years, and were With suitably redefined reference planes, and neglecting
PAPER 29
3 52
666 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS. VOL. CAS-31. NO. 7, JULY 1984
F(z)=t z> a i
;-o the result of the error mechanism and is really not acceptable.
Table 1 shows the peak magnitude error for a number of filters,
can be determined by interpolating with n + 1 samples { z , } and
from which an ominous trend is clearly visible.
n + 1 correspondingly sampled responses F(z,). The process cor-
> Consider the two planes w and z, Fig. 3. These are related by
responds to solving A=*o~ F where a is the Vandermonde
1 Z - 1 = M>T where r is the radius of a new sampling circle
matrix of samples { z , } . Now o " can be formed using a Traub
centered on {1,0}. Performing an FFT with a normal set of
technique [3]. Alternatively any standard interpolation method
samples from the unit circle i n w and a corresponding set of
(Langrange, etc.) could be used for the process. The samples { z , )
sampled responses derived from the circle in z will yield
can be chosen anywhere in the complex plane, with some prefer-
ence for a circular array, which generally gives best accuracy. It n
transpires [4] that i f a circle of unity radius and centre {0,0} is H w
) =
° T, a w
'i '-
1
used, then o " F is exactly equivalent to performing an FFT. This i-0
has many computational attractions, as much effort has been Replacing w by (z - l ) / r gives
invested i n providing extremely fast and accurate FFT routines.
However, two problems arise with the application of the FFT n * ) - ia;(z-i)« -i. r
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-31, NO. 7, JULY 1984 667
T A B L E II
ERRORS FOR BANDPASS FILTER WITH VARIOUS F F T
INTERPOLATION SCHEMES
' OB >
TABLE I • -\
P
1
AT. "
P j 1
( a „ a , ,- • •, a _ , } , - D. = + 1 2 /
J2 k/
the i th set of samples Z . Let 0, = C^,, hence z = C~'e " '
f = i
. j2«k/i
r e wA r - 0 , 1 , 2 , - - S o C = r/. Also 6 = r , / r
h e r e f ( 0
l
then o, = o B , where 0 = d i a g { l , b bf,- • - ,b'~ ).
( Hence: it
2
Co Co ^0
2
C, c, c r
2
r
Bo'
1
>-t
each On" 'F, involves the transformed FFT as outlined. The inver-
sion of the [ C ] matrix presents few problems since its order is
equal to the number of partitions which is usually quite small,
and as it is a Vandermonde matrix a Traub technique can be
1
used. The calculation of B~ is trivial.
The two final columns of Table I I show the further improve-
ments in accuracy with partitioned polynomial interpolation, a
scheme with 2 partitions in the numerator and 4 in the denomina-
Fig. 3. w and z planes with sampling circles. tor providing the best results.
668 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, VOL. CAS-31, NO. 7, JULY 1984
The partitioned polynomial approach also provides a solution suffers by a very serious problem of slow convergence [3]. The
to a second problem with the FFT, that is overinterpolation. As problem is overcome by the algorithm proposed in [4], which
the number of samples presented to an FFT routine must satisfy determines D by an extensive use of the adaptive predictor.
p
2", it will generally be necessary to round up to the nearest Further improvements are, however, possible by using the con-
integer satisfying this condition. This means having to evaluate strained version of the predictor [5], as will be shown in the
the system response more times than necessary. Worst cases present letter. The resulting algorithm is faster than that pro-
occur for orders of 2" (requiring 2" + 1 sample points) which posed in [4] and allows a reduced computational cost.
+ 1
must be rounded up to 2" , giving 2" - 1 over evaluations and In the following we will use the notations:
almost doubling the computer time for system evaluations. By 1) D is the tap vector of the predictor filter of order N, i.e.,
N
integer i, then partitioning the system requires fewer function be adaptively computed by the method proposed in [6].
evaluations than the basic FFT. I n some instances quite consider- 2) P(z) is the polynomial transfer function of the filter, i.e.,
able savings can be made.
The application of a transformed sample plane and polynomial
P(0= I d z-
k
(2)
partitioning to interpolative analysis of SC filters demonstrates *-o
improvements in accuracy and the number of function evalua-
tions on even moderate order networks. I t is expected that in The roots of P(z) give an estimate of the frequencies of the
,2 /
analysis of high-order SC networks these routines should produce sinusoids of the process. These roots take on the form e ~- * ,
significant improvements in accuracy and computer time. This is t = 1,2,-•-, A/, since we will use N = 2M and P(z) is con-
a subject of further work. strained to have its roots on the unit circle.
3) D ( z ) is the tap vector of the constrained predictor of
B H
Abstract —A new algorithm is proposed for determining the F I R filter •1,2,- ,M-l
for estimating the frequencies of a process consisting of sinusoids in white
noise. The algorithm is based on an extensive use of the constrained M-i M-i
predictor and is characterized by a very fast convergence. VL = f o — - o 1 o _ - _ o l (5)
The use of finite impulse response (FIR) filters for the spectral M M
analysis of a process is well known. When the process to be
where D , i =1,2,- • •, M is obtained by (3) with H *= i . Such a
BJ
Rewrite as
[. . .]
ABSTRACT
- ?i
V I - [» » k - l ] -k =
-
L
[ V - B
V i ]
(3)
SYMBOLIC ANALYSIS OF I D E A L S C NETWORKS
d = B N adj B
One method f o r e f f e c t i n g t h e s y m b o l i c a n a l y s i s
of SC networks ( i d e a l and n o n - i d e a l ) employs a top- w h e r e n i s d e t e r m i n e d by r a n k B_. T h i s g i v e s a max-
o l o g i c a l f o r m u l a t i o n [ l ] . H e r e a c o m p a c t MNA f o r m - imum v a l u e f o r n, i n p r a c t i c e t h e d e g r e e i s u s u a l l y
u l a t i o n h a s been u t i l i s e d , w i t h s w i t c h i n g i n t r o - l e s s a n d z e r o c o e f f i c i e n t s may b e s u p p r e s s e d .
d u c e d v i a c o n t r a c t i o n a n d a c o m p a c t i o n scheme ^2^
for multiphase networks. Totally arbitrary multi- Then
p h a s e s w i t c h i n g a r r a n g e m e n t s c a n be accommodated.
I n s t a n d a r d MNA form £3] - l - i - l J N z ' L
: H ( z ) = [A
3 Bj i = 0 -1 =
1=0 -i
d.z D <z)
i=o
(4)
w h e r e D ( z ) i s t h e d e n o m i n a t o r common t o a l l t e r m s
and m <^ n.
I f IAI = 0 a n d | B | = 0 t h e n d e f i n e a new v a r i -
a b l e 2= z + e w h e r e e = c o n s t a n t , o f s u c h a v a l u e
to ensure t h a t t h e m a t r i x p e n c i l £ B+A} i s r e g u l a r . E
node v o l t a g e s v ( t ) a n d b r a n c h c u r r e n t s i ( t ) in
t i i a e s l o t A, a r e : Ttie f r e q u e n c y domain r e s p o n s e follows ijassdlately from substitution
k
(t)~ Ln e q u a t i o n
V c 0 v(t) j(t)
r
y («)
+ =
i(t) o 0 0 e(t) - In)
(5) where * k + 1
u
- e~^ H*l
p r o c e s s and i f c a r e i s n o t t a k e n , c o n s i d e r a b l e con-
c e r n can a r i s e a s t o t h e v a l i d i t y o f t h e P com- k
puted .
is)
-1166-
3 58
- —
-2 . WP_ (z)
1
"DP(2)
X NP , ( a ) HP . ( a ) -P KM (s) KM (9)
-P
_ _
1 K
K
-k = — . I - i
(C)G.
ke
At i=l
(14)
w h e r e A t i s t h e t i m e s t e p a n d a ^ , K_ a r e e a s i l y
«• FDSAXi SJTTTCBSB
p r e d e t e r m i n e d |[lo] . T h e o r d e r o f a p p r o x i m a t i o n N
i s a l s o determined as i n , and may r e q u i r e H=l0-
For t h e g e n e r a l i s e d s t a t e space d e s c r i p t i o n i t can
be shown t h a t
N " I -1
i/ A C
At e ke
At i=l
. . (15)
T h e MNA e q u a t i o n s f o r t h e n o n - i d e a l SC n e t -
work p r o d u c e a s t i f f s y s t e m , l a r g e l y due t o t h e
values of switch G a n d G „,.. I t h a s b e e n f o u n d
on off
that equation (15) normally g i v e s the b e s t answers
and t h e r e f o r e t h e g e n e r a l i s e d s t a t e s p a c e MNA
f o r m u l a t i o n h a s been adopted. Incidentally i fa
NAM d e s c r i p t i o n i s u s e d t h e s i z e o f t h e m a t r i x i s
s m a l l e r a n d s l i g h t l y b e t t e r c o n d i t i o n e d , t h i s may
h a v e r e l e v a n c e i n t h e a n a l y s i s o f l a r g e SC n e t -
works . 10 11 12 13 M 15
3
H2 X19
A program SCNAP 3 w r i t t e n i n F o r t r a n 77 h a s
P i g . 1. F i f t h Order I l l l p t l c Lowpa«« SC PI1 t a r
been u s e d t o a n a l y s e a v a r i e t y o f SC f i l t e r n e t -
works b o t h f o r t h e i d e a l a n d n o n - i d e a l s i t u a t i o n s .
The r e s u l t s compare v e r y f a v o u r a b l y w i t h t h o s e
p r o d u c e d by f r e q u e n c y domain p r o g r a m s . F i g 1 shows REFERENCES
t y p i c a l r e s u l t s f o r a 5 t h o r d e r SC e l l i p t i c l o w -
pass f i l t e r w i t h both i d e a l and n o n - i d e a l s w i t c h e s . T a n a k a , M. and M o r i , S . : " T o p o l o g i c a l formula-
tions f o r the c o e f f i c i e n t matrices of s t a t e
The i n i t i a l s t e p t o p r o d u c e t h e v a r i o u s p o l y - equations f o r switched c a p a c i t o r networks",
n o m i a I s i s o b v i o u s l y t h e most t i m e c o n s u m i n g p a r t I . E . E . E . T r a n s CAS, V o l CAS-29, No. 2, pp 1 0 6 -
of t h e a n a l y s i s . The a d v a n t a g e s f o l l o w a f t e r t h i s , 115, 1982.
s i n c e many p o i n t s o f f r e q u e n c y a n a l y s i s c a n be
computed by s i m p l e p o l y n o m i a l e v a l u a t i o n . This J o h n s o n , D.G., S e w e l l , J . I . a n d M e a k i n , A.D.:
has s i g n i f i c a n t i m p l i c a t i o n s i n n o i s e a n a l y s i s . " I n t e r p o l a t i v e a n a l y s i s of switched capacitor
O t h e r d i r e c t a p p l i c a t i o n s a r e i n o p t i m i s a t i o n and n e t w o r k s " , I . E . E . C o l l o q u i u m on D e s i g n S o f t -
p a r t i t i o n e d a n a l y s i s o f l a r g e SC f i l t e r n e t w o r k s . w a r e , C o l l o q u i u m D i g e s t , pp 6/1 - 6/7, 1 9 8 4 .
-1 1 6 7 -
3 59
4. K a d e r A., " I n v e r s i o n o f p o l y n o m i a l n e t w o r k
matrices with p a r t i c u l a r r e f e r e n c e t o sen-
s i t i v i t y a n a l y s i s " P r o c . I . E . E . , P a r t G, No.4
pp 1 7 0 - 1 7 2 , 1 9 8 1 .
ACKNOWLEDGEMENT
T h i s work was s u p p o r t e d by t h e S c i e n c e a n d
Engineering Research Council.
-1168-
3 60
PAPER 33
3 6 1
L. B. W o l o v i t i arid J . I . S e w e l l
Department o f E l e c t r o n i c s and E l e c t r i c a l E n g i n e e r i n g
U n i v e r s i t y o f G l a s g o w , G l a s g o w , G12 8QQ, U.K.
373
i n t e g r a t i o n t e c h n i q u e Q2] i s u s e d t o d i s c r e t i z e
e q u a t i o n ( 2 ) . T h i s t e c h n i q u e h a s t h e added b e n e f i t s
that i t d i r e c t l y handles equation (2) in i t s i m p l i c i t
form, t o g e t h e r w i t h any a l g e b r a i c e q u a t i o n s . By
-2 + I D
a c c u m u l a t i n g c h a r g e s i n t h e v e c t o r o j x ) and t h e n
I a p p l y i n g t h e n u m e r i c a l i n t e g r a t i o n to t h i s v e c t o r ,
I the r e s u l t i n g method i s more a c c u r a t e and e f f i c i e n t
t h a n i n d i v i d u a l l y i n t e g r a t i n g t h e MOSFET t e r m i n a l
charges.
The n u m e r i c a l i n t e g r a t i o n t e c h n i q u e used i s a
b a c k w a r d d i f f e r e n c e a p p r o x i m a t i o n to the d e r i v a t i v e
( p r e d i c t o r ) , f o l l o w e d by a c o r r e c t o r b a s e d on G e a r ' s
The method o f e q u a t i o n formulation has a dramatic a p p r o a c h [12j . The o r d e r 1 method i s u s e d a s a
e f f e c t on t h e s i z e and s p a r s i t y o f the r e s u l t i n g s t a r t e r and t h e r e a f t e r t h e o r d e r 1 and o r d e r 2
m a t r i c e s . To a l l o w the i n c l u s i o n o f v o l t a g e s o u r c e s f o r m u l a s ar3 u s e d i n a v a r i a b l e - o r d e r , v a r i a b l e
(dependent and i n d e p e n d e n t ) , t a b l e a u type f o r m u l a - t i m e - s t e p m a n n e r . The o r d e r i s l i m i t e d to 2 t o a l l o w
t i o n methods are d e s i r a b l e f o r t h e i r g e n e r a l i t y and t h e i n c l u s i o n o f PWL n o n l i n e a r i t i e s [ l 3 ] . Because
s p a r s i t y p r o p e r t i e s . A f e a t u r e o f these methods i s o f t h e l a r g e number o f b r e a k p o i n t s i n the c l o c k
the need f o r a p i v o t i n g a l g o r i t h m t o ensure t h a t a s i g n a l e x c i t a t i o n common t o SC n e t w o r k s and t h e
non-zero d i a g o n a l i s o b t a i n e d , which then a l t e r s the c o n s e q u e n t r e q u i r e m e n t s f o r r e s e t t i n g t o low o r d e r ,
s t r u c t u r e o f the m a t r i c e s [4] . A comparison the h i g h e r o r d e r methods a r e n o t i n v o l v e d . The d i s -
between the f o r m u l a t i o n methods d i s c u s s e d i n £l] c r e t i z a t i o n o f e q u a t i o n (2) l e a d s t o a s e t o f n o n -
as w e l l as the mixed n o d a l t a b l e a u (MNT) [i] on the l i n e a r equations
b a s i s o f o p e r a t i o n counts f o r t h e LU decomposition
of the r e s u l t i n g m a t r i c e s , found the MNA f o r m u l a t i o n (3)
[8] t o be the b e s t approach. By t r a n s f e r r i n g as much which must be s o l v e d a t each t i m e - p o i n t t . A n
of the m a t r i x p r o c e s s i n g t o a p r e - p r o c e s s i n g s t a g e , m o d i f i e d Newton-Raphson a l g o r i t h m i s used t o s o l v e
t!.e time-domain response i s o b t a i n e d v e r y e f f i c i e n t - e q u a t i o n ( 3 ) , which leads t o the i t e r a t i o n
ly. Taking i n t o account the b l o c k - m a t r i x s t r u c t u r e
of e q u a t i o n (1), the s p a r s i t y o f b l o c k s A i and B i , ,, k. . k+1
J(x )Ax
and t h e form o f t h e MNA e q u a t i o n s used t o assemble n —
e q u a t i o n (1), a h i g h l y e f f i c i e n t b l o c k compaction
k+1 k k+1
a l g o r i t h m t y p i c a l l y reduces the system o f e q u a t i o n s x x + Ax (4)
—n —n
by 50-90%. T h i s reduced m a t r i x i s t h e n r e - o r d e r e d
to achieve a m i n i m a l o p e r a t i o n count and LU de- k k
where J(Xn) i s the Jacobian e v a l u a t e d a t x„. Each
composed u s i n g a sparse v a r i a n t o f Gaussian e l i m i n a -
t i o n . To o b t a i n the time-domain response these LU i t e r a t i o n r e q u i r e s the s o l u t i o n of a s e t o f l i n e a r
f a c t o r s are used r e p e a t e d l y i n a f o r w a r d e l i m i n a t i o n e q u a t i o n s . F o r t u n a t e l y the nonzero s t r u c t u r e o f
and back s u b s t i t u t i o n , t h e r e f o r e t o o b t a i n the Jacobian does n o t change from i t e r a t i o n t o
maximum e f f i c i e n c y a code g e n e r a t i o n scheme [9~\ i s i t e r a t i o n . I t i s sparse and n o r m a l l y d i a g o n a l l y
used t o generate code f o r t h e s o l u t i o n process. dominant. T h e r e f o r e sparse m a t r i x techniques[14]
can be used t o g r e a t e f f e c t . The i t e r a t i o n (4) i s
t e r m i n a t e d when the i t e r a t i o n s s a t i s f y the converg-
NONIDEAL TIME-DOMAIN ANALYSIS ence c r i t e r i a . The u s u a l c r i t e r i a i s based on the
node v o l t a g e s and n o n l i n e a r c u r r e n t s meeting
To ensure charge c o n s e r v a t i o n , i t i s necessary s p e c i f i e d e r r o r t o l e r a n c e s . These checks are n o t
to s e l e c t charges as the s t a t e v a r i a b l e s f o r t h e s u f f i c i e n t t o ensure convergence and may l e a d t o
MOSFET capacitances Clo] , l e a d i n g t o the s e t o f f i r s t charge n o n c o n s e r v a t i o n [ l o ] , T h e r e f o r e the t e r m i n a l
order nonlinear algebraic d i f f e r e n t i a l equations nodes o f dynamic and n o n l i n e a r devices are checked
t o s a t i s f y the convergence c r i t e r i a
d a j x ) + Gx + f 2 ( x) w(t) te[0,T] (2)
1
dt — IV^ — I < E + c
1
max { i v * * ! . Iv"!}
where <l(x) i s the charge f u n c t i o n , <i_(x) = C x + f ^ ( x ) The t e r m i n a l charges o f c a p a c i t o r s and MOSFETS are
checked t o s a t i s f y
18 the n o n l i n e a r charge f u n c t i o n
k + 1 k
^2 i s the n o n l i n e a r c u r r e n t f u n c t i o n lQ - Q| < e q + c r max((|Q. l, k+1
lQ l|k
c i r c u i t b l o c k s h a v i n g a l a r g e number o f feedback
where e , e and are the a b s o l u t e v o l t a g e , charge
loops w i t h w i d e l y d i f f e r i n g c i r c u i t t i m e - c o n s t a n t s . v q
r e l a t i v e tolerance.
s i m u l a t i o n techniques QlJ are n o t s u i t e d t o s o l v i n g
these networks. To overcome the problem of t h e
' s t i f f n e s s ' o f the e q u a t i o n s , an i m p l i c i t n u m e r i c a l The o v e r a l l a l g o r i t h m f o r s o l v i n g e q u a t i o n (2) i s
g i v e n below.
374
3 63
375
3 6 4
17)
S p r i n g e r - V e r l a g , New Y o r k , 1 9 7 8 .
B.Dembart a n d A.M.Erisman, " H y b r i d s p a r s e -
S 5 11 M M M I
TIME (SECS)
m a t r i x methods", I E E E T r a n s . CT, V o l . CT-20,
No.6, 1 9 7 3 , p p . 6 4 1 - 6 4 9 . Rg.2 Time response of 5th order Nossek filter
18) J . A . N o s s e k a n d G.C.Temes, " S w i t c h e d - C a p a c i t o r with 500Hz sinusoidal excitation
f i l t e r design using b i l i n e a r element modeling",
I E E E T r a n s . CAS, V o l . CAS-27, No.6, 1980, p p .
488-491.
19) S . J . H a r r o l d , I.A.W.Vance, J.Mun a n d D.G.Haigh,
"A GaAs s w i t c h e d - c a p a c i t o r b a n d p a s s f i l t e r I C " ,
P r o c . 1985 I E E E GaAs I C Symposium, M o n t e r e y ,
November 1985, 4pp.
j v
?
I
nc (sees)
Fig.3a Clock waveforms for bandpass filter Rg.3b Impulse response for bandpass filter
376
3 65
PAPER 36
General analysis of large HI near switched capacitor
oet^fv® rlcs
Indexing terms: Networks, Circuit theory and design. Matrix algebra. Algorithms
spectral analysis is possible. Extended sparse matrix (nT, (n + 1)T] a periodically switched linear network can
methods [14, 15] i n c o n j u n c t i o n w i t h interpretive code be represented by the system differential equation
generation have been employed where appropriate, and a
C v„ (t) + G v„. (t) = w . (i) k •• N (7)
wide range of matrix and numerical c o n d i t i o n i n g tech- k A k k n k
such that the clock signals (and therefore the network) time d o m a i n a square wave can be exactly represented by
does not vary in I . Here k denotes the feth time-slot. H k a simple p o l y n o m i a l . Sinusoidal approximations w i l l be
As shown i n Fig. 1, these time-slots are not necessarily of required f o r frequency response considerations, and
equal duration. because of the small time increments involved (owing to
In.2 In.3 In.N the switching frequency) the p o l y n o m i a l approximations
for sinusoidal signals are very accurate; typically orders
9
ffN = T m = 3 to 9 ensure errors o f < 1 0 ~ . Substituting expr. 8
in eqn. 7 gives the new system differential equation
C b Jt)
k m + G v Jt) k K = X < f ' t
where
T —«- 1
__ . «-l
P (0 = i ? - { [ s C
t
1
k + G ]-'} l
(11)
Fig. 1 Definitions for a N-sloi switched linear network
t e /
„ ( t ) = J"(') -.»
v _ (nT + ff _ ) are the initial conditions.
' (0 otherwise k l t t
v (nT
k + a„) = v(nT + o )~ k (4) t (nT
k + a ) = P v ^{nT
k k k + + f (13)
1=0
That is, u„, (t) are functions o f time, and v {nT + o ) are
t k k
tv (nT
k + a )z-=V (z)
k k fc=l,...,N-l = (15)
»=o
The coefficients a" k are obtained by defining
(5)
n=0
h ( l 6 )
- = S
and therefore m
/7.t = w . . i ( « i J (17)
X AnT Vf + a )z~" K = X M"T + T)z'"
n= 0 n= 0 and then applying eqn. 53, w h i c h gives
= z ^ z ) - V^O)) = zV^z) (6)
< * = f •/„/"* (18)
since K^O) = 0. 1= 0
the extended state transition matrix (eqn. 11), and apply- E=PyP . S 1 PiPx (25)
ing an I M N approximant o f even order M [ 2 7 ] , then k = 1, . N - 1
Pk* l
eqn. 54 gives Ek = (26)
I k = N
P (z )
k k * I Re l2KJi \zJv k k C k + GJ " '] (19) Matrices E and E are frequency independent and are k
(expr. 55) to eqn. 12 gives tion in the excitation T-W (z) as a preprocessing step, k
giving
Mil 2K/r i! t r
«,.»(rj* £ R e (20)
(27)
i=0 1= 0
Guidelines for selecting the order of p o l y n o m i a l approx-
imation m are given i n the Appendix and the I M N where
approximant order M and constants of a p p r o x i m a t i o n
K and Z , elsewhere [ 2 7 ] . The I M N approximant is
t
P>. k = Ek i.k- B
(28)
reputed t o be stable and accurate over a wide range of Eqn. 24 then reduces to
systems applications and this has certainly proved true i n
the SC network applications outlined here w i t h M ^ 10.
< z / - E ) * V z ) - £ Z »"*(*) (29)
4 ^-domain analysis
which can be very efficiently solved for ^ z ) using
I n the general analysis of switched networks it is o b v i - methods discussed below. The solutions f o r V (z) f o r k
ously essential to have an efficient frequency analysis. I t k = 1, N — 1 are then obtained by block
w o u l d be possible t o perform a Fourier transform o n a back-substitution,
time d o m a i n solution, but however efficiently this might
be implemented i t is hardly likely to provide an accept-
able solution. Previous attempts [ 6 - 1 0 ] at nonideal y (z) = P V _ (z)
k k k l + -LW (z) k k = 2, N - 1 (30)
analysis become very time-consuming at this point of
transformation i n t o the Z - d o m a i n , p r i m a r i l y because of Using the direct Gauss elimination method to solve eqn.
3
the need for A C analysis i n each clock phase at each fre- 29 requires 0 ( v ) flops. This cost is excessive f o r large
quency point. Fortunately, w i t h the present ' o r m u l a t i o n , networks, especially when m a n y frequency points are
2
a fairly straightforward solution is at hand. evaluated. I t can be reduced to 0 ( v ) flops by trans-
T o o b t a i n the Z - d o m a i n series of final states V (z) for f o r m i n g the system to upper Hessenberg f o r m , which
k
(zl-H)y = b (33)
5 Solving the discrete system
where
A substantial amount of frequency independent pre-
processing can be performed i n solving eqn. 23. The H= T ' P ' E P T
approach used is based on the method developed i n Ref- l ,
y=T- p- V N (z) (34)
erence 16 f o r ideal SC networks. First, all the P are fre- k
eqn. 35 as and
o) = w + nw,
b= xiffju) (36)
0
i=0 1= 0 vectors W k
y by the transformation matrix, w h i c h requires a further (ii) B u i l d R H S of eqn. 33 using eqns. 36 and 37
2
v flops. The total solution process then requires approx- (iii) Solve eqn. 33 f o r V ^ e i a o T
) using the Hessen-
2
imately 5 v flops, which is a dramatic improvement over berg method
the direct approach. (iv) Calculate y (e °" )k
i lT
using eqn. 30
(c) Spectral analysis
(i) F o r selected n calculate weights D „ using eqns. k
8 Results
z = e J m o T
(39)
The theory developed above was implemented i n the
A p p l y i n g Poisson's f o r m u l a to l.W (z) k gives
program Q U I C K S C N A P using 16-digit double-precision
m m / m arithmetic throughout. I n the preprocessing section
T
ZrV {e* )
k = £ I Ev„ m a x i m u m use is made of sparse techniques, utilising new
n-0 Li-0 \l = 0
interpretive code and o p t i m a l ordering algorithms [ 1 5 ] .
, - jntaT F u l l matrix techniques are used i n the frequency response
x W [nT k + o + lh - r » ) ^ « ,
k k calculations. A dynamic storage allocation scheme is used
throughout. Numerical accuracy has been checked by
comparison w i t h measured results and w i t h a selection of
B
= i I \t i.K ly m<o
ti - no,) computed results f r o m other established programs, when
' n= 0 L( = 0 1= 0 agreement was better than five significant figures. A rep-
(40) resentative sample of large SC networks is references i n
x exp {J{u> - nw$p k + lh - t»))J k
Table 1.
400
Table 5: Run statistics for SWITCAP
350
Example Number of number of Pre-Pr., Time/pt, Storage SCNAPNIF
300
nodes slots - ic sec words
250
1 28 2 6.14 0.207 11852
2 41 4 — — — in S W I T C A P (ideal
«! 2 0 0
3 30 6 — — —
4 69 2 25.9 1.621 34188 150 OUICKSCNAP
5 86 36 — — —
<~> 1 0 0
6 77 5 — — —
50
give a fair comparison against currently available soft- The frequency analysis algorithms are highly suited to
ware, indicating improvements i n speed by a factor of vectorisation, which could provide further dramatic
200. T h e preprocessing times also show a speedup factor improvements when implemented o n a suitable pro-
of 10, directly attributable to the enhanced sparse tech- cessor.
niques utilised. N o t e that Example 5 is beyond the capi-
bility of p r o g r a m S C N A P N I F . Tables 4 and 5 show 10 Acknowledgment
performance figures produced by programs specifically
designed f o r the analysis o f ideal SC networks, and it is The authors gratefully acknowledge the financial support
interesting to observe that although Q U I C K S C N A P of the U K Science and Engineering Research Council.
undertakes a complete nonideal analysis, comparison
shows that its performance lies between S C N A P I F and 11 References
S W I T C A P , being approximately twice as fast as the
latter. While S W I T C A P can approximate a nonideal 1 L I O U , M.L., K U O , Y . L . , and L E E , C . F . : 'A tutorial on computer
analysis, the r u n times are i n terms of hours. S W I T C A P aided analysis of switched capacitor circuits', Proc. IEEE, 1983, 71,
pp. 987-1005
requires that the network is not disjointed i n any clock 2 K U N I E D A , H : 'Effects of finite gain-bandwidth products of op
phase, and this explains the limited selection of successful amps on switched capacitor networks — Approached via equivalent
examples in Table 5; although a straightforward f i x is representations'. Proc. I E E E Int. Symp. on Circuits & Systems,
possible, this requires m o d i f i c a t i o n of the circuit by the Rome, May 1982, pp. 464-467
user. 3 B H A T T A C H A R Y Y A , B.B., and R A U T , R.: 'Analysis of switched
capacitor networks containing operational amplifiers with finite D C
Similar conclusions are apparent f r o m the graphs gain and gain-bandwidth product values', IEE Proc. G, Electron.
shown in Figs. 2 and 3, which display the relative per- Circuits & Syst., 1983,130, (4), pp. 114-123
4 C U T L A N D , N J , L A U , C , and S E W E L L , J.I.: 'General computer
formances of the programs f o r two-phase SC networks.
analysis of switched capacitor networks including nonideal ampli-
fiers'. Proc. I E E E Int. Symp. on Circuits & Systems, Rome, May
9 Conclusions 1982, pp. 17-20
5 F A N G , S C . , T S I V I D I S , Y.P., and W I N G , O.: ' S W I T C A P , a
switched capacitor network analysis program', IEEE Circuits &
A new analysis scheme has been demonstrated to main-
Syst. Mag., 1983, 5, pp. 4-10 and pp. 41-46
tain accuracy, while p r o v i d i n g orders o f magnitude 6 L I O U , M.L.: 'Exact analysis of linear circuits containing period-
improvement i n speed o f analysis. The computational ically operated switches with applications', IEEE Trans., 1972,
cost o f the analysis increases linearly w i t h the number of CT-19, pp. 146-154
time-slots and quadratically w i t h the number of nodes. 7 S T R O M , T., and S I G N E L L , S.: 'Analysis of periodically switched
linear networks', / £ £ £ Trans., 1977, CAS-24, pp. 531-540
This affords the user a facility to breach the existing
8 W E H R H A H N , E . : 'Exact analysis of periodically switched linear
vN = 500 boundary, while containing time and storage networks'. Proc. I E E E Int. Symp. on Circuits & Systems, Chicago,
requirements w i t h i n bounds typical of generally available April 1981, pp. 884-887
Cx(t) + Gx(t) = nit) x(0) = X o (46) The inverse Laplace transforms (eqns. 50 and 51) are
c o m p u t e d using a numerical q u a d r a t u r e a p p r o x i m a t i o n
where x(t) is the u n k n o w n response vector, w{t) is the
of the Laplace t a n s f o r m inversion integral [ 2 6 , 2 7 ] . T h e
excitation vector, C and G are constant real matrices.
constants f o r the a p p r o x i m a t i o n (K a n d z,), are tabulated l
The S I L T m e t h o d provides an attractive s o l u t i o n tech-
complex constants and f o r an I M N a p p r o x i m a n t o f even
nique f o r eqn. 46. The excitation vector is a p p r o x i m a t e d
order M these occur i n M/2 complex conjugate pairs,
by the vector o f mth order polynomials,
w h i c h allows the c o m p u t a t i o n to be halved. Hence the
extended state transition m a t r i x a p p r o x i m a t i o n is
P (t)=
m £ot„t* (47)
t=o M/2
/fAr) * £ Re [ 2 K / A t [ z / A t C + G ] " ' ]
i i (54)
Substituting eqn. 47 i n t o eqn. 46 and t a k i n g the Laplace
t=i
transform,
and the excitation response m a t r i x a p p r o x i m a t i o n
X(s) 1
= [ s C + G\ - \ c x a + £ ^1 (48)
PAPER 49
3 73
Abstract - A new sensitivity analysis method in the interpretable code generation, polynomial a p p r o x i m a t i o n
frequency domain is presented f o r non-ideal switched o f the excitation, descretization o f the whole system and
capacitor networks that is more general and efficient than Hessenburg techniques are f u l l y exploited.
the previously published works. There are no restrictions
on the number of clock phases, circuit configuration and II. GENERAL SENSITIVITY ANALYSIS METHOD FOR ARBITRARY
input signal type (which can be either continuous or UNEAR NETWORKS
sample-and-hotd). By performing a substantial amount of
frequency independent pre-processing and using adjoint For an arbitrary linear system, T X - W , where T i s
network techniques, the computational costs are reduced the system matrix, X is the unknown response vector and
dramatically and a highly efficient method results. The W denotes the excitation vector. I f the output o f interest
method has been implemented in a SC network analysis <t> ( a linear c o m b i n a t i o n o f the components o f X ) is
program called SCNAP4. Numerical results of the T
* - d X , where d is a constant vector and d T
denotes the
sensitivity analysis of a fifth order elliptic low pass filter transpose, then standard sensitivity theory f o l l o w s :
are given. T h e a d j o i n t vector X , 1
is d e f i n e d by X] - -d'T" o r
T
T X , - - d a n d the sensitivity o f <P w i t h respect to t h e
I. INTRODUCTION (MI* ffT d\V
vector o f parameter changes h is x] — X - x] —
dh ah ah
The sensitivity analysis o f ideal SC networks is w e l l The computational procedure f o r the adjoint approach
established and forms part o f the designers standard tool- can be summarized:
k i t . The present trend to extend the application o f SC step 1: Solve f o r X ;
networks to h i g h frequency front-ends w i t h multi-rate step 2: Calculate X , ;
clocks raises the serious question o f sensitivity o f the step 3: Formulate dT/ah and a w / a h ;
c i r c u i t structures w i t h respect to non-ideal parameters step 4: Calculate a « / a h .
such as switch resistance and finite a m p l i f i e r parameters -
gain/bandwidth, input and output impedances; sensitivity T h e method is clear and straightforward. The advantage
o f parasitic capacitances is also required. o f this procedure is that the vectors X and X , are o n l y
A number of programs can undertake various calculated once, irrespective of the number of
frequency and time d o m a i n analysis o f non-ideal SC parameters. The transpose solution X , can be e f f i c i e n t l y
networks [ 1 - 4 ] . A range o f techniques have been utilized obtained using similar steps to those used i n solving f o r
to increase speed, improve accuracy and cope w i t h large X , w h i c h does not require another network analysis.
V,(z) - EW (z) 2
independent and can be pre-calculated. Hence the
(1) computation of 3T/3h and 3W/3h follow in a
-P N zl z£W (z)
N
s t r a i g h t f o r w a r d manner and the general procedure f o r
£"' denotes the inverse Laplace transformation; p is (b) T h e method is also very e f f i c i e n t . A f t e r p r e -
k
defined as the extended state transition matrix (EST); B processing of all frequency independent equation
; k
denotes the i d e n t i t y m a t r i x . I n the c o m p u t a t i o n o f (c) The results o f the method are reliable and accurate.
frequency response, p , B are frequency independent, T h e o n l y a p p r o x i m a t i o n employed i n this method is the
k i k
=0 (9)
T h i s paper presents a new non-ideal sensitivity analysis
and rewrite as f o l l o w s
scheme f o r SC networks. The method is both general and
1
-M; (10) e f f i c i e n t . N o restrictions are imposed regarding circuit
c o n f i g u r a t i o n , number o f c l o c k phases or e x c i t a t i o n
Hence
ap, signal. The implementation of the method is
(11) s t r a i g h t f o r w a r d . N u m e r i c a l results c o n f i r m the v a l i d i t y
3h
1406
3 75
ACKNOWLEDGMENT
REFERENCES
switched capacitor networks including all parasitics". Proc. IEEE ISCAS, Fig. 2a. Sensitivity with respect lo C i with nonideal opamps (GB=159kHz)
C4
Hi V
s
i.
7i % C.VJ 7I s
Ca>+—it- ' g, v 3 4«U S &i
:
j§HI-@
5i
s s ,
%,s y,4 'yi,4 %,s
i I ®—*-\ t-
® x y
1407
376
PAPER 53
3 77
565
Eroc. I . E . E . E . IriterjHticnal Syrrxsiur, cn Circuits and SystHTS, VbL.5, Lcrrfcn, May 1994.
noise analysis. In addition, numerical inversion of the transition matrix and Hessenberg technique for solving the
Laplace transformation, the Hessenberg technique, adjoint system as well as the pre-processing of a large
extensive sparse matrix routines and interpretable code amount of frequency independent material reduce
generation have been used to improve the efficiency of computation costs dramatically and hence a highly efficient
noise analysis. The method has been implemented in method results. By using spectral analysis technique the
SCNAP4 and numerical results are given in the last section. fold-back effects can be accurately evaluated. In the
following sections, these aspects will be discussed in detail.
I INTRODUCTION
II ADJOINT NETWORK TECHNIQUES
Since the early years of last decade, noise analysis for SC
circuits has been discussed frequently [1-7]. However most The task of noise analysis can be divided into two different
of the presented techniques are only adequate for small aspects.
circuits. General and efficient methods were seldom
mentioned. Recently, switched-current circuits have gained (a) reducing the overall number of system solutions;
much attention. Naturally, their noise characteristics are of
main concern to designers. Generally, two types of noise are (b) efficient methods for each individual system solution.
considered in SC circuits, white noise from switches and 1/f
noise from operational amplifiers. For switched-current The second aspect has already been considered when
circuits, since the MOS transistors operate in saturation SCNAP4 was being developed. A solid platform for noise
regions, both thermal and flicker noise exist analysis has been built successfully. In the following, we
will focus on the first issue.
Unlike the traditional analogue case (switch free), the
folding of wideband white noisefromhighfrequencybands For general linear switched networks, if the system contains
to baseband, caused by the sampling nature of switched- m noise sources, the problem can be described by
circuits. leads to excessive computational costs.
Traditionally, in order to calculate the whole noise spectral
density, the folding transfer functions from each noise I -p,1"Vj(z)- " EWj(z) "
source to output have to be evaluated. For example, a circuit p 2 I Vj(z) = SW (z)
2
566
H (o)) = - X i W
i i (4) The noise analysis method presented above has been
implemented in the program SCNAP4. Two types of noise
Assuming that there is no correlation between the input source (voltage and current) are available, they can be either
noise sources, the total noise power spectral density at the flicker noise or white noise. The foldover wideband white
output is then calculated by superposition noise from high frequency bands to baseband can be
calculated.
1 n=0i=l
bandpass filter is selected. The op-amp noise components
are given to be 8QnV/VHz with comer frequency 1kHz.
where S[(£2) is the ith input noise power spectral density The output noise spectrum density is illustrated in Figure 2.
and p is the number of bands to be considered. Here the nth A close agreement between simulation and results from [S]
band is defined as the frequency interval from (n - l)cu /2 can be noted. The fold back effects are also shown in Fig. 2.
s
to no) /2 where <o is the sample frequency. Several For switched-current circuits, a third-order elliptic ladder is
n 8
considered. The noise bandwith is about 160MHz and the
comments can be made at this stage.
comer frequency is set to lKHz. If the reference noise is
lnA/vlfz , then the circuit response and noise behaviour
1) It is evident that after solving the adjoint system (eq. (3»,
Hj(o)) can be obtained directly by merely one subtraction are illustrated in Figure 3 and Figure 4 respectively. The
since each W; contains at most two nonzero entries. Instead CPU time for analysing these circuits are given in Table 1.
of m different network analyses, only one adjoint system It is evident that even for circuits with moderate size,
solution is needed for all m noise sources, the original SCNAP4 is able to give noise analysis results in fairly short
system solution can be saved. Therefore computational time.
efficiency is achieved.
V CONCLUSIONS
2) The adjoint system is solved by using the Hessenberg
approach which has already been proved very effective in This paper presents a general and efficient noise analysis
frequency domain analysis of switched linear networks. method applicable to both SC and SI circuits. Standard
circuit non-idealities are included and complete folding
3) The folding effects from high frequencies to a specific effects for broadband noise are incorporated. An adjoint
band are evaluated up to a user given noise bandwidth. This description of the non-ideal network is employed and only
is based on the fact that the noise is normally bandlirnited the solution of this is required. Numerical results compare
by op-amp frequency response and time constant effects well with practical results measured from integrated
from switch resistances and circuit capacitances. circuits. The speed of analysis on a relatively model work
station indicates that noise analysis of large switched
379
567
systems is feasible and the application to some degree o f [9] V . Z a k i a n , "Properties o f I ^ , and J ^ , approximants and
noise optimisation is a realistic possibility. applications to numerical i n v e r s i o n o f Laplace
transforms and initial value problems". J. Maths and
ACKNOWLEDGEMENT A p p l i c . vol.50, pp.191-222,1975
[10] K . S i n g h a l and J . V l a c h , " C o m p u t a t i o n o f the t i m e
Z.Q. Shang gratefully acknowledges the f i n a n c i a l support o f d o m a i n response by n u m e r i c a l i n v e r s i o n o f the
the B r i t i s h C o u n c i l . Laplace transformation", J. F r a n k l i n Inst., vol.299, pp.
109-126, Feb. 1975
REFERENCES [11] J.B. Hughes. "Analogue techniques for large scale
integrated circuits," Ph.D Thesis, University of
[1] J.H.Fischer."Noise sources and calculation techniques Southampton, March 1992
for switched capacitor filters," IEEE J. of Solid-state [12] G. Wegmann. "Design and analysis techniques for
circuits, vol. SC-17, No.4, Aug. 1982, pp.742-752 dynamic current mirrors," Thesis No. 890, Ecole
[2] C.A.Gobet and A. Knob,"Noise analysis of switched Polythechnique Federate de Lausanne, 1990
capacitor networks", IEEE Trans, on CAS, vol. CAS- [13] S.J. Daubert and D. Vallancourt," Operation and
30. No. 1. Jan. 1983. pp.37-43 analysis of current copier circuits." Proc. EEE Pt. G.
[3] JJlabaey, J.VandewaUe and H.De Man,"A general and vol. 137. pp.109-115. April 1990
efficient noise analysis technique for switched capacitor [14] J. A. Barby. "Switched-current filters models for
filters", Prcc. of IEEE ISCAS. Newport Beach. May. frequency analysis in the continuous-time domain,"
1983. pp.570-573 Prcc. of IEEE ISCAS. Chicago, May.. 1993. pp. 1427-
[4] S. W. L i , F. W. Stephenson and J. Velazquez- 1430
Ramos,"Noise analysis of switched capacitor filters,"
Proc. of IEEE ISCAS. Montreal. May. 1984. pp. 1312-
5
1315 Vf2
[5] CK.Pun. J.I.Sewell and A.G.Hall,"Noise analysis for lth2
switched capacitor networks in symbolic form," Proc.
of 29th Midwest symposium on circuits and systems. lth3
Louisvill, Aug. 1985. pp.807-810 M2
[6] J.Gotte and C.Gobet, "Exact noise analysis of SC circuits
and an approximate computer implementation." I K K K
Trans, on CAS. vol. CAS-36, No.4. pp.508-521. Apr. vfl
1989 lth 1
[7] L. Toth and K. Suyama, "Exact noise analysis of 'ideal'
-€H
Ml
SC networks". Proc. of IEEE ISCAS, New Orleans,
May.. 1990. pp. 1585-1588
[8] Z.Q.Shang and JXSeweUVTifficient sensitivity analysis Figure 1. Memory cell with noise sources
for large non-ideal switched capacitor networks," Prcc.
of IEEE ISCAS, Chicago. May. 1993. pp. 1405-1407
-90
-100
1 1 0
-
i -
u
-5 -120 J band=50 ~' _
•a
aa
-
-130 V baseband r
^ *=^«=<=» - =
-140
0.0e+O 1.0e-r4 2.0e+4
Frequency(Hz)
Figure 2. Noise behaviour of sixth-order bandpass SC filter
with folding back effects
20
-40
-60
80
3 8
10 10 10 10 10 10 10
Frequency(Hz)
Figure 3. Frequency response of 3rd order elliptic ladder SI circuit
16
A
14
12 band=8
S 10
E 8
baseband
0
8
10 10 10 10 10 10 10
Frequency (Hz)
Publication Page
PAPER 17
728 ! E F . E TRANSACTIONS ON COMMUNICATION?, V O L . COM-23. NO. 7, J U L Y 1975
Abstract—A design for an adaptive three-port electronic circulator encountered in telephone systems and also in the various
is presented. The circuit has particular use in telephone work and
filtering requirements [ 5 J .
this model can provide isolation when the terminating load varies
I t has been demonstrated [ 1 ] that line matching can be
over the range 200 f l - 1 kS). The whole circuit is realizable in micro-
electronic form. achieved in a most accurate way. However this refers only
to the static case and will not apply to dynamic matching
where there can be changes in line impedance with chang-
INTRODUCTION ing length, quality of line, and terminations. Some form of
ADAPTIVE CIRCULATORS
T h e great virtue of the audio circulator is that the design
is simple [ 3 ] , it is easily extendable to an >i-port version, T h e principles are most easily observed by considering a
and a large variety of combinations are possible [ 4 ] . T h u s three-port circulator, which is the lowest order of practical
the circulator can be used for all the isolation problems use in telephone work. I n a typical handset application
[ 2 ] , the microphone is connected to port-1, the subscribers
line to port-2, and the earphone is fed with a signal propor-
Paper approved by the Associate E d i t o r for W i r e C o m m u n i c a t i o n
of the I E E E Communications Society f o r p u b l i c a t i o n w i t h o u t oral tional to the. difference of the signals at these two ports.
presentation. M a n u s c r i p t received October 1, 1974; revised Januarv T h e circulator provides the required isolation on transmit
21,1975.
R . P. Lodge is w i t h M a r c o n i C o m m u n i c a t i o n Systems, Chelms- and receive. I n the transmit mode, the difference signal
ford, England. will produce the side-tone and, with ideal matching, uni-
.1. I . Sewell is w i t h the D e p a r t m e n t of Electronic Engineering,
U n i v e r s i t y of H u l l , H u l l , HUO 7 R X , England. form speech level? are ensured. If the circulator impedance
LODGE AND S E W E L L ' . A D A P T I V E E L E C T R O N I C C I R C U L A T O R 729
B F W 10
P O R T -1 PORT-3
PORT-
A M P L I F I ER
RECTIFIER
I
B I A S
AMPUFI E .
SUMMER
RECTIFIER
2
at port-2 perfectly matches the line, then in the t-insmit rection feedback loop should be as high as possible.
mode there is theoretically zero voltage at port-3, and in However, this is restricted by the possibility of oscillation,
the receive mode no signal will appear at port-1. Any mis- although this may be removed by some form of damping.
match with the line will produce a finite rejection port When an input is applied to the second port, simulating
voltage and it might be assumed that this could be used as an incoming line signal, the feedback mechanism is differ-
a feedback signal to control a voltage variable resistor ent. Assuming that there is no input signal at port-1, there
( W R ) defining the circulation impedance at port-2. is no signal at the inverting input of O A 2. T h e input
However a phase-sensitive rectifier would be required to voltage divides between the F E T and the load resistance,
derive a correction signal, which will vary the impedance thus the output of the summing amplifier goes negative to
in the appropriate direction. a great extent. With a high loop gain and small damping
Fig. 1 shows a better alternative circuit for an adaptive factor, the F E T channel becomes cut off and all the signal
circulator. T h e circulator is balanced when ports-1 and -3 is passed to O A 4. There will now be isolation at port-1 and
are terminated in Ro and when the channel resistance of an earphone fed with the difference of the port-1 and -2
the F E T is equal to the load at port-2. If, under these signals will receive the full incoming one. However this
conditions, the input voltage atport-1 is V and there are no leads to two problems: the adaptive feature is no longer
other input signals, the output from operational amplifier operative as the line feeds the earphone circuit direct; and
( O A ) 1 is also V. T h e port-2 voltage is then V/2. T h e dif- the circulator suffers from voice switching. A fairly con-
ferential amplifier formed by O A 2 thus has two equal in- venient method of overcoming these two difficulties is to
puts and, therefore, no signal will appear at port-3. For un- provide feedback of the signal at port-3 to the inverting
balanced condit ions, when the F E T channel resistance and input of O A 1, isolation in the path is required to prevent
the load are different, there is an output from O A 2 and any feedforward. I n the transmit mode, there is no signal
thus the port-3 voltage will be finite. With the gain of at port-3 and operation is as previously described. On re-
amplifier 1 double the gain of amplifier 2. a feedback signal ceiving an incoming signal at port-2, the extra feedback
to the F E T gate is derived by comparing the actual provides an input for amplifier 2 which will ensure that the
port-2 Voltage with the desired value. I t is usually neces- F E T will not now be cut off. Voice switching is eliminated
sary to provide some bias, in series with the correction and the adaptive property retained.
signal, to ensure that the F E T channel resistance is
centered on the mean load resistance. As the r j , charac- PRACTICAL CIRCUIT
teristic is not entirely linear, it is usually necessary to
A circuit employing the common 741 amplifier as the
apply local feedback to obtain a wider linear range.
basic amplifier unit was constructed. T h e rectifiers were
For good third-port rejection, the gain round the cor-
simply a single diode with a smoothing capacitor; some
730 I E E E TRANSACTIONS O S COMMUNICATIONS, J U L T 1975
1.4
P O R T - I I N P U T 10
1.2
E 1.0
R.= s o o n '
/
inn
R.= 200 n
/ L
">0 I K 10 K SO K
F R E Q U E N C Y Hi
buffering with a voltage follower may prove necessary there are wider rejection bands for the greater load devia-
before the summer, again a conventional amplifier circuit. tions; this is due to better controllability of the larger
The smoothing capacitors also furnish the required damp- error signals present in these circumstances. Nevertheless
ing and 0.22 provides a satisfactory response. the circuit will provide third-port rejections of between
The r d , ( N ) value of the F E T selected ( B F W 10) was in
0
30-40 d B for variations in port-2 terminations over the
the region of 200 fi. T h e initial matching range was chosen range 200-1000 ft, for a 5-kHz band. T h e isolation charac-
as 200-1000 fi. W h e n the load resistance has some value teristics of the circulator when fed with a signal at port-2
R , the difference voltage generated is
L
and subjected to load variations at port-2 are shown in
Fig. 3. T h e rejection is greater than 44 d B over a 20-kHz
V, V i a R L
bandwidth, there being negligible effect due to load change
V = °
2 RL + rj,(0N)'
from 600 fi until approximately 1 kfi is reached. T h i s can
be expected since the third stage, which is purely resistive
where rd, ti)' is the actual drain-source resistance apparent and accurately adjusted, is responsible for producing this
(0
in the circuit and includes any alteration induced by local degree of isolation.
feedback; and V is the input voltage at port-1. For the
i n I t is therefore possible to conclude that adaptive
signal levels ensuring linear operation of this particular electronic circulators art: quite suitable for telephone usage.
circuit, an input voltage of about 30 m V resulted in a Isolation is possible over a good bandwidth, and with more
maximum value of V as 5 m V . T h e gate-source voltage
d careful design, considering the time constants involved in
required for a channel resistance of 1000 fJ for the B F W 10 rectification, F E T nonlincarity, and feedback loop gain,
is —2.1 V , thus a differential gain of about 400 is required. together with better amplifiers, a superior performance
T h i s is realized by the amplifiers preceding the rectifiers, could be ensured. One disadvantage of the present circuit
which have adjustable gains to compensate for differing is the need for low input voltages, as is common with most
voltage drops across the diodes. A n added feature is that V V R applications of F E T ' s , since the signal inevitably
these amplifiers also prevent loading of the F E T and the produces some modulation of the channel resistance. T h i s
line. can be overcome by employing more elaborate gate-drain
Fig. 2 shows some typical response curves, when the feedback circuits and using F E T ' s especially designed for
circuit was fed with 30 m V at port-1 and the rejection W R work.
voltage at port-3 is measured. T h e bias on the F E T gate The voice-switching problem of the simple circuit is
is set to match a 600-ft load at port-2. I t is apparent that overcome by adding an extra feedback path. The circulator
within a 5-kHz bandwidth, the rejection is best for loads can be used to match complex impedances, but obviously
in the region about 600 Q. A point of interest is to note that the variable correcting circuits will be more sophisticated.
L O D G E AND S E W E L L : A D A P T I V E E L E C T R O N I C C I R C U L A T O R 731
I 1
IK 10 K
F R E Q U E N C Y Hi
The operating conditions considered here would only R . P . Lodge was bora in Huddersfield,
England, on June 21, 1952. I n 1974 he re-
correspond to unbalanced line working in a telephone
ceived the B . S c . degree in electronic engineer-
system. B u t with duplication of the circulators [ 2 ] normal ing from the University of Hull, Hull, E n g -
operation is easily accomplished. A n added advantage is land.
He recently joined Marconi Communica-
that the whole circulator circuit can easily be realized in
tion Systems, Chelmsford, England.
microelectronic form.
ACKNOWLEDGMENT
PAPER 23
3 88
A n Improved Adaptive E l e c t r o n i c C i r c u l a t o r for Consider F i g . 1 and note that the controlling signals are de-
Telephone Applications veloped from a separate pilot lone and that the impedance
measuring and adjustment bridge can have both pilot, transmit
H. G A Z I O G L U , D. A. HOMER, A N D J. 1. SEWELL, MEMBER, IEEE and receive voice signals present in one arm simultaneously.
The circulator section automatically removes the pilot signal
from the isolation path and the voice frequencies are filtered
Abstract-A previous attempt at realizing an adaptive circulator for
out in the c o n t r o l path. T h e line input impedance is simulated
telephone applications suffered from limitations due to voice switching
by an RC circuit, F i g . 2 ; a three element parallel/series circuit
problems, resistive part matching only, and some difficulties with adap-
is considered initially, but results are s h o w n for a five element
tive operation for broadband signals. The use of pilot tones enables the
circuit to be free from voice switching problems and to accomplish a network for more general application. T h e three element net-
close impedance match with that of a telephone line. Analysis of the work will suffice for short lines, but the more c o m p l e x net-
circuit in an adaptive hybrid configuration produces the theoretical work is necessary for the longer subscriber lines. T h e possible
bounds on performance. Experimental results display the performance combinations of length and diameter of cables on subscriber
characteristics of the adaptive circuits and various other hybrid circuits lines is almost infinite. F o r British Post O f f i c e lines, the three
on a selection of subscriber lines with a range of terminations. networks given in F i g . 2 represent o p t i m u m fits to the fre-
quency responses of average overall lengths taken in three
groups and t y p i c a l c o m b i n a t i o n s of C u cable of differing d i a m -
INTRODUCTION eters, a line termination of 6 0 0 R was used. E x p e r i m e n t a t i o n
There is an increasing interest in the inclusion of a variety will show that these n e t w o r k s provide a suitable median from
of electronic features in the telephone subscriber's loop. Apart w h i c h the adaptive feature can work.
from dialing facilities and subscriber carrier systems, studies With the three element circuit it is obvious that a 3-term
concerned with circuits providing a m p l i f i c a t i o n [ 1 , 2 1 , active controller could be used. Analysis of such a network d e m o n -
hybrids [ 2 , 3 , 4 ] and adaptive h y b r i d s [ 5 , b, 7 ] are being strates that the high frequency response is a direct function of
actively pursued. T h e previous attempts at realizing adaptive the series resistance, whereas the low frequency response is
hybrids had a number of limitations, some providing the adap- dependent upon both resistances. T h e break points of the
tive feature only at certain frequencies [51 or over a limited simulated line characteristic can be directly controlled by vary-
frequency band [ 7 ] . T h e adaptive c i r c u l a t o r [ 6 ] provided ing the capacitance value. Hence a c o n t r o l loop e m p l o y i n g a
only resistive part matching and although some precautions pilot tone well above the voice band, say 10 k H z can be used
against voice switching were taken, these were not entirely to adjust the high frequency response and another using a pilot
satisfactory; the adaptive operation of the circuit w i t h broad- tone below the voice b a n d , say 12 H z , can adjust the low fre-
band signals also presented problems. quency response. A further pilot tone just outside the b a n d ,
The possibilities of adaptively m a t c h i n g line impedance say 5 k H z , can be used to c o n t r o l the capacitance value. H o w -
changes in the subscriber e q u i p m e n t , at the interface with line ever, investigation shows that in practice the high frequency
amplifiers, or at the 4-wire to 2-wire transition points are as- asymptote is of little significance, because the major effect is
suming increasing importance. T h e line characteristic is vulner- out of band and a n y w a y this varies little from line to line and
able to the influence of change in terminations and cable para- with various terminations. A n y influence on the upper break
meters affected by p h y s i c a l means due to either m a n or nature. frequency can be covered by capacitance control T h e major
The m i s m a t c h against static balancing impairs sidetone rejec- change c o m e s at the low frequency e n d , manifesting itself as
tion and reflection characteristics resulting in problems of no quite substantial changes in loop resistance, lower break fre-
small significance. quency and slope of the impedance characteristic. F o r sim-
A further requirement of future developments is to provide plicity of analysis a single term controller, operating from a 12
isolation and matching characteristics by circuits suitable for Hz pilot will be investigated in detail. A dual pilot tone c i r c u i t ,
mass p r o d u c t i o n by integrated circuit means. with adaptive parallel R and C elements is of more general
application for broadband signals, however extension of the
THE ADAPTIVE CIRCULATOR theory is straightforward.
The basic 3-port circulator is well k n o w n ( 8 ) and the re- With reference to F i g . 1, for the purposes of analysis, it is
placement of one section by an adaptive one has been reported not necessary to consider the pilot tone signal but only the
[ 6 ] . A t t e n t i o n here will be c o n f i n e d to a different adaptive components of the transmitted signal E and the received signal
section w h i c h may be used in c o n j u n c t i o n with standard sec- V. Assuming ideal operational amplifiers, the output signal of
tions to form a fully adaptive circulator. the circulator section is given by:
4Zj V
Paper approved by the Editor tor Communication Electronics o f the
-+ 2E
IEEE Communications Society for publication without oral presenta- (1)
Z n + Z , Zn + Z,
tion. Manuscript received November 25, 1977; revised March 10, 1979.
This work was supported by the Science Research Council.
The authors are w i t h the Department o f Electronic Engineering. Now E corresponds to the port-1 voltage, V to the port-2
University o f Hull, Hull HU6 7 R X , England. voltage, the port-3 voltage will be E D = EA ' / 2 , hence:
0 0 9 0 - 6 7 7 8 / 7 9 / 0 8 0 0 - 1 2 1 8 $ 0 0 . 7 5 © 1979 IEEE
389
I E E E T R A N S A C T I O N S O N C O M M U N I C A T I O N S , V O L . C O M - 2 7 , N O . 8, A U G U S T 1979 1219
PILOT
TONE
FROM FIRST
SECTION OF TO THIRD
5ECTION OF
CIRCULATOR
CIRCULATOR
IAI
FILTER RECTIF E R
[Bl
FILTER RECTIF ER
0-29 M F 0 21MF
HI— HI— (X + I ) 2
+ Y 2
(5)
270R
- v W -
- A V -
«MR E q u a t i o n ( 5 ) represents a circle of center ( - 1 , 0 ) a n d radius
(b) 2/K. Similarly for some constant value of I/AR = R the ex-
pression for the magnitude of equation ( 4 ) yields
0-28MF
-Hh- R 2
I 1 2
2
2R
I70R + Y = (6)
-vw- R 2
- 1 _ R 2
- 1_
36R 2
which represents a circle in the XY plane with center l(R +
(c) 2 2
\)I(R - 1), 0 ] and radius 2R/(R - 1). C o n s t a n t transfer
Tig. 2 Optimum balance networks for (a) 1 km line (mixed diameter ratio and rejection ratio circles c a n be plotted as s h o w n in F i g .
Cu) (b) 4 k m line (mixed diameter Cu) (c) 7 k m line (mixed diameter Cu)
3. A s can be seen infinite rejection is obtained at the p o i n t 1 +
/ 0 and this corresponds to the c o n d i t i o n ZQ = Z j . T h e rejec-
tion rapidly decreases as the ratio changes f r o m u n i t y , al-
2Z, V ^0
Z
l though the transfer gain is relatively insensitive to the change
Ea = + £ (2)
z + z.
a
and remains close to 0 d B .
One of the major roles of this circulator is as an adaptive
hybrid in the telephone set, and most of the experimentation
has been carried out with that area o f operation in m i n d . If
which consists of a transmitted component E o T and a c o m - the circulator is to be used in the mode already indicated [ 6 ] ,
ponent to be rejected E o R Set a = Z IZ 0 X then the transfer with the .aouthpiece connected to port-1 and the earpiece fed
and rejection ratios become respectively with a signal from across ports-1 and -2 giving a d e f i n e d side-
tone injection, then the transmission properties o f the adaptive
2 section are of little consequence as the third section will pro-
A T = (3) vide complete isolation a n y w a y . A n o t h e r c o n n e c t i o n of the
V a + 1
circulator is with the mouthpiece again across p o r t - 1 , the line
across port-2 and the earpiece across port-3. N o w a n i n c o m i n g
£
oR a —1 signal is transmitted through the adaptive section and both
(4)
E a + f rejection and transmission characteristics of the section are of
1220 I E E E T R A N S A C T I O N S O N C O M M U N I C A T I O N S , V O L . C O M - 2 7 . N O . 8. A U G U S T 1979
0 01
tY
K--0-21dB
K.-0-27dB
003
002
001
H, U d B R.S-SdB
001
0 02
003
0 01
significance. A performance figure can be defined as: program was incorporated into others w h i c h simulate the be-
havior of the single and dual pilot tone adaptive circulators.
The results from these programs indicate the possibility of
approaching the above theoretical limits.
5 *
8
I 4 s
M / V
s
S LyW
HI
8
L HI
*
60
T
v W
HI
1 HI
Lj
HI
8
1222 I E E E T R A N S A C T I O N S O N C O M M U N I C A T I O N S , V O L . C O M - 2 7 . N O . 8, A U G U S T 1979
50
40
600R
S
CD
30
100R
100R
U
10 K 600R
Id 2K
v OK
20
10
inevitably produces a variable resistance in parallel with a tice, operating conditions can o c c u r that cause line termina-
variable capacitance and this resistance value decreases with tions to appear largely reactive, this results in some of the re-
negative bias. j e c t i o n curves demonstrating rapidly changing characteristics.
Practical measurements were taken using an input signal of H o w e v e r , the results obtained for the range of resistance ter-
100 m V and results for the rejection between ports-1 and -3 in minations used, quite adequately encompass the worst per-
the dual tone adaptive circulator circuit are c o m p a r e d w i t h the formances experienced with both capacitive and inductive
performances of some other possible h y b r i d c i r c u i t s in Figs. terminations. It is therefore reasonable to use these responses
5-7. T h e graphs in F i g . 5 show measured responses of the dual as a good indicator of the average performance with practical
tone adaptive c i r c u l a t o r and of the conventional passive h y b r i d terminations. T h e best overall performance achieved f r o m a
of the B P O 7 0 0 type telephone set, under t y p i c a l short line dual tone circuit was an average in excess o f 3 4 d B rejection
conditions for a range of terminations. O v e r the normal tele- over the whole voice band for a line impedance change of 3 : 1 .
phone band and average terminations improvements in rejec- T h e characteristics on the whole deteriorate at higher frequen-
tion are observed. F i g . 6 demonstrates the improvement due to cies, but since the out of band behavior of the high frequency
the addition of the adaptive feature to the circulator when asymptote has been ignored, this trend is to be expected. E x -
working on a m e d i u m length line. T h e results for the non- periments with the upper pilot tone at 10 k H z indicate a slight
adaptive circulator with the passive RC m a t c h i n g circuit are improvement in performance figures and the circuit also has
computed. There is a moderate improvement for the 6 0 0 R the advantage of relaxing the requirements o f the high-pass
termination, to be expected since the n o m i n a l c o m p o n e n t s of filter for this tone, a feature o f some p r a c t i c a l importance as a
the passive matching network are derived for a median 4 k m 3rd order elliptic filter will suffice.
line of mixed diameter C u sections with a 6 0 0 R termination.
T h e circuit has been completed by the addition of oscil-
Quite marked improvements are apparent as the terminations
lators to provide the pilot tones; these oscillators were of the
are varied, displaying the activity of the adaptive feature. F o r
Wien-bridge type. It is an advantage if circuits for inclusion in
the longer lines the line impedance is m u c h higher and the
telephone sets do not require an independent power source.
F E T ' s are increasingly pressed into a non-linear V V R operating
E x p e r i m e n t s in deriving ± 1 5 V supplies from D C power de-
mode when the c o n t r o l is not as linear and fine. T h e measured
livered d o w n the line f r o m a central supply indicate that s u c h
responses displayed in F i g . 7 also snow the poorer p e r f o r m -
circuitry could be powered from a local exchange. Some
ance of the single, low f r e q u e n c y , pilot tone model. In prac-
studies on the reduction o f power c o n s u m p t i o n have indicated
3 93
r
40
100R
30
600R 10K
(0
2K
u
LJ
20
100R
10 2K
Passive RC m a t c h i n g circuit
10K
30
100 R
600 R
20 -
' /
UJ \
I0OR '
2K 600R
10
2K
I0K I0K
that circuit simplifications are possible. In actual fact, the computer simulation. At low system blocking the traffic carried is
circulator loop can be broken and the third stage eliminated determined solely by the server limitations while at higher blocking the
without any reduction in the performance factor. traffic carried is determined by channel limitations.
The overall results obtained indicate that the realization of
an adaptive hybrid for telephone applications with integrated Performance characteristics have been investigated for large
circuit technology is quite feasible. T h e theoretical conditions scale mobile radio systems using dynamic channel assign-
for very high rejection ratios, of the order of 40 d B , are quite ment. 1 - 9
These systems utilize small radio coverage cells and
strict. F o r applications requiring such responses under wide reuse channels within a metropolitan area. In d y n a m i c c h a n n e l
variations of line c o n d i t i o n s , it will be necessary to employ assignment systems, channels are assigned to serve calls on the
higher grade operational amplifiers in some positions in the basis of instantaneous demand in the system subject only to a
circuit. Detailed analysis of the c o n t r o l system reveals a stand- prescribed channel reuse constraint, i.e., any c h a n n e l m a y be
ing error responsible for about 3 d B loss in rejection. A rede- assigned to any cell in the system provided that the c h a n n e l
sign should be possible here. A more Linear and wider range of is not being used w i t h i n a specified n u m b e r of cells f r o m that
variation of resistance and capacitance c o m p o n e n t s may be cell at the time that the c h a n n e l assignment is made. T h e speci-
afforded by techniques e m p l o y i n g multipliers, w i t h very little fied n u m b e r of cells required between cells that may use the
increase in area on an integrated c i r c u i t . same c h a n n e l is called the reuse interval. Earlier s t u d i e s 1 - 4
Publication Page
[46] A method for the evaluation of multirate sigma delta systems 396
[50] Design of high order sigma-delta modulators with minimum weighted noise 401
[51] Issues in the design of low oversampling ratio single bit sigma-delta modulators 406
[55] The application of redundant number systems to digital sigma-delta modulators 412
3 96
PAPER 46
j y<
A METHOD FOR T H E E V A L U A T I O N OF M U L T I R A T E
SIGMA-DELTA SYSTEMS
Abstract - Simulation techniques for sigma-delta based estimation, least squares function fitting) are all incorporated as
systems are discussed. An extension to the next state simulation network elements and treated in the same way as simple elements
method for block diagrams is presented which allows efficient such as gains, delays and quantisers. Most element parameters
handling of multirate systems. This is achieved by the additional (such as the gain of elements, or the number of levels in a
definition of clock dependent inputs for each element to define quantiser) are taken from inputs, which are usually replaced by a
how the element behaves in a multirate situation. constant, but which can be connected to a variable source to
allow parameter sweeps.
I. INTRODUCTION
In the block simulator B L O S I M [4], blocks are interconnected by
Relatively little has appeared in the literature concerning the means of first-in-first-out (FIFO) buffers. This makes sample
simulation of sigma-delta converters. It is widely realised that rate changes very easy to implement. However there is
simulation using traditional analogue simulation tools is not considerable overhead in the buffer management routines which
feasible due to the prohibitive C P U time required. Most limits the application of this technique to the simulation of sigma-
researchers have managed by using purpose written computer delta systems. Subsequent research has been directed in the field
programs which model the difference equations representing the of program compilers for multiple processor D S P [5]. The
specific sigma delta modulator topology being investigated. The sample rates in a sigma-delta modulator system are well defined
program described here allows the user to describe the structure before the simulation commences and are constant (ie. they are
of complete sigma-delta modulation systems which may contain synchronous). This fact has been used to devise a block diagram
several sigma-delta ADCs, D A C s , sigma-delta based multipliers simulation method which retains many of the advantages of
along with the associated filters in a form similar to a S P I C E B L O S I M (generality, ease of use, hierarchical definition of
deck, but at a signal flow graph level [1]. blocks, block libraries, consistency checks etc.) but handles
synchronous multirate circuits more efficiently.
The circuit is described at a system level and consists of a
network of gains, delays, quantisers, filters etc. The algorithm The program executes in the four stages shown in Fig 1. The
then reduces the network description to a list of function calls network description file is first read and comments stripped out
which can be executed repeatedly to simulate the action of the and macro definitions expanded. The resulting network of
circuit. This is almost as fast as a purpose written program. elements is then converted to an unordered list of operations.
Alternatively, the algorithm could be made to produce a C Each operation is represented by a data structure containing the
program file for compilation by the computer system's C function to be executed, the input and output nodes, dependent
compiler. The resultant program would be almost and clock dependent nodes, values of any parameters and any
indistinguishable from a well written, purpose written program, other information required Lists of pointers to these operations
but would be much quicker to create and easier to modify. are then created by a scheduler which ensures that the operations
Sigma-delta systems are usually multirate and this is handled are executed in a valid order, and that only nodes which need to
efficiently. be recalculated are actually recalculated in multirate systems.
Running the simulation then simply involves executing the
Many non-idealities can be included, such as finite integrator gain operations in the specified order.
and saturation, linear and nonlinear settling (slew limiting),
comparator hysteresis, dc offsets, quantiser level spacing etc. can D. R E P R E S E N T A T I O N O F M U L T I R A T E S Y S T E M S
all be modelled [2]. The method is compatible with (can include)
table based techniques [3]. Three node types are defined: an ordinary node value (eg. the
input and outputs of gains, delays and quantisers), a signal node
To be most useful, a simulation tool for sigma-delta modulation or list of values (eg. for complete signals, F F T records, F F T
systems must allow many parameters to be swept through a range window functions etc.) and a file node type (eg. for reading in
of values, eg.: D C input level, integrator offsets, amplitude and input signals, printing out results, communicating with other
frequency of input signal (to determine the variation in SNR), processes such as externally defined elements or graph plotting
number of quantisation levels or the word length of the digital programs). Evaluating a node involves executing all the
filters, gain values (to determine sensitivity due to capacitor operations with outputs connected to that node.
mismatch, integrator gain and bandwidth). There is also a
variety of analyses which may be performed on various signals in Each element as entered by the user, is represented in the
the system (eg. the raw output and the decimated output). A computer by a number of operations. A single operation may
flexible approach has been adopted whereby all the signal have many inputs (the dependent nodes) and a single output, all
sources, and all the analysis features ( F F T , power spectral of which may be simple nodal values, complete signals, or files
1328
0-7803-0593-0/92 $3.001992 I E E E
(Fig. 2). The operation also has a number of ciocA-dependem each simulation run, and during a warm up period (to allow filter
nodal inputs, which determine how the operation will behave in transients to die down) and during the actual simulation period.
multirate systems. Elements with several outputs are realised by The system can be simulated many times in a single program run.
several operations.
The user does not need to understand how the network of
Three rules govern the behaviour of all the operations in a elements entered is transformed into a network of operations and
multirate system: the user never comes into contact with the clock dependent
1. A node will only be calculated if any operation contributing to that inputs. The modular approach adopted allows new element
node is known to be active. definitions to be added easily without altering the core of the
2. An operation will only be active if at least one of its clock program.
dependent inputs is connected to a node known to have changed.
3. To allow forratechanges, it is possible to force a node to be either Some Common Elements Expressed as Operations
active or not by applying a forced clock to the node. This rule
overrides the two above. Gain elements including nonlinear gains such as saturation,
overflow, absolute value, quantisers, piecewise-linear non-
The rule for operation ordering is based on the standard linearities, dB/linear conversion etc. are each represented as a
precedence rule: single operation (Fig. 3). The left hand clock dependent input
4a An operation cannot be evaluated until all it's inputs (the dependent ensures that the output is recalculated whenever the input
nodes) are known. changes. The right hand clock dependent input ensures that the
4b A node cannot be recalculated until all the active operations directly operation is included in any recalculation of the output node, even
contributing to that node satisfy Rule 4a if the input has not changed. Input elements are similar to gain
elements but with no dependent node. The output of a signal
Scheduling Algorithm source usually has a forced clock applied to provide a starting
condition for the algorithm.
The scheduling must be done in two passes. The first pass
determines which nodes change and therefore must be
recalculated, and the second pass then places only the operations The delay element is represented by two operations (Fig. 4).
which are active into a computable order. This is because the output must be available for calculation before
the input has been calculated. The definition of a delay element
Pass One effectively "opens up" any loops (loops must contain a delaying
A l l nodes are unknown. element) and changes the topology from having loops to a tree
A p p l y r u l e 3. T h i s w i l l make some nodes structure, without any loops. Referring to Fig. 4, 'delay 1' is
known ( e i t h e r t o have changed o r n o t ) dependent on 'delay 2' and so 'delay 2' is always executed first
and p r o v i d e s a s t a r t i n g p o i n t f o r t h e The link between the two operations is not realised using a node
next stage. ensuring that no loops are formed. Note that all this is hidden
R e p e a t e d l y a p p l y r u l e s 1 and 2 t o d e t e r m i n e from the user who simply enters an instruction line 'delay xl x2'
w h i c h o t h e r nodes a r e known t o be to implement a delay from node x l to node x2.
active.
Any nodes w h i c h cannot be d e t e r m i n e d a r e Sample-rate changes are implemented by a modified linear gain
deemed i n a c t i v e . element For both the mathematical resample (Fig. 5), and the
sample & hold element (Fig. 6), the new sample rate is forced
Now all nodes are known to be either active or not. upon the output node (making use of Rule 3, above). A
• Pass Two decimating FIR filter (Fig. 7) is efficiently realised using a
cascade of a circular buffer operation followed by an FIR
All nodes w h i c h a r e n o t a c t i v e are known operation which takes a vector dot product of the vector of
(because t h e y have n o t changed) . A l l previous inputs with the filter coefficients. The rate change is
o t h e r s a r e n o t y e t known. implemented between the two operations so that the vector
Repeatedly t r y t o e v a l u a t e every node. If a product is taken only when necessary. This scheme also works
node can be e v a l u a t e d (Rule 4) t h e n add for non-decimating FIR filters. Interpolating FIR filters are more
to the l i s t of p o i n t e r s t o operations efficiently realised by a poly phase structure [1].
a l l t h e o p e r a t i o n s s a t i s f y i n g Rule 2.
A c l e a r node o p e r a t i o n must be e n t e r e d
a u t o m a t i c a l l y ahead o f t h e s e . This More complex elements can have several outputs. The additional
node i s now known. outputs come from additional operations dependent on the output
I f any nodes can n o t be d e t e r m i n e d , t h e n t h e of the main operation.
network contains e i t h e r a delay f r e e
l o o p or has no i n p u t - e r r o r . Significant memory can be saved when implementing certain
operations which act on complete signals, eg FFTs, by
The above scheduling algorithm must be invoked once for each performing the operation in place, where the output data is stored
distinct set of clock states. in the same memory locations as the input data. The destruction
of the input data can be delayed by assigning the outputs of all
The definition of an operation ensures that the network operation operations dependent on the input to the FFT (or other) element
is well defined in multirate systems, even when signals at as dependent nodes to the FFT element This gives the illusion
different rates are combined. Rate changes can have either a of separate input and output nodes even although storage is not
mathematical resampling action (ie. output is zero between input assigned to both. Elements which can use this technique include
samples), or have a sample and hold type response, depending the FFT, PSD (for conversion of a FFT record to a squared
on the connection of the clock dependent inputs. Each operation magnitude spectrum), log_sig (for conversion to a dB scale) and
has six function pointers associated with it (the functions can be various signal shortening operations to reduce the number of data
empty). These define the run time operation during the very first points for graph plotting.
and very last clock cycle, immediately prior and immediately after
1329
HI. THE PROGRAM IN USE [3] G.T Brauns, "Table-Based Modeling of Deha-Sigma
Modulators using ZSIM", IEEE Trans. Comp. Aided. Des., vol
Fig. 8 shows a simple sigma delta modulator in flow graph form 9, no. 2, p 142-150, Feb 1990.
suitable for simulation. The nodes must be labelled and entered [4] D. Messerschmitt, "A Tool for Structured Functional
as a net list An input signal must also be included which can be Simulation", IEEE Sel Areas in Comm. vol SAC-2, no. 1, pp
built up from a variety of ac and dc sources, or taken from a file. 137-147,Jan 1984.
The output must also be sent to file (which can be either a [5] J. Bier et al, "Gabriel: A Design Environment for DSP".
terminal, disk file, or a UNIX pipe to an external program). IEEE Mirco, pp 28-45, Oct. 1990.
Normally some form of analysis would be done on the output [6] B. Boser, K.-P. Karmann, H. Martin, and B. Wooley,
before it is sent to a file. The 'sweep' element is similar to a dc "Simulating and testing oversampled analog-to-digital
source except that its value changes from one simulation to the conveners," EEEE Trans. Computer-Aided Des., vol CAD-7, pp
next (but within the same program run). This may be used to 668-674, June 1988.
modulate the input signal (either the amplitude or frequency) or to [7] Kuniharu Uchimura et al, "Oversampling A-to-D and D-to-
adjust the value of some of the gains, or to adjust the word- A Converters with Multistage Noise Shaping Modulators," IEEE
lengths in digital systems. This provides great flexibility and is Trans. ASSP, vol. 36, no. 12, pp. 1899-1905, Dec 1988.
very simple to use. [8] K. C.-H. Chao, S. Nadeem, W.L. Lee, and C.G Sodini,
"A Higher Order Topology for Interpolative Modulators for
Normally, built in, single operation integrators would be used for Oversampling A/D Converters," IEEE Trans. Circuits &
the integrators in Fig. 8. These can include non-ideal effects [2]. Systems, vol. 37, no. 3, pp. 309-318, March 1990.
By quantising the input and modelling numerical overflow [9] T.C. Leslie and B. Singh, "An improved Sigma-Delta
propeTly. digital systems can be simulated in the same manner. Architecture", IEEE ISCAS 1990, p372-375
Both analogue and digital elements may exist in the same system. [10] J.C. Candy, "A Use of Double Integration in Sigma Delta
Modulation", IEEE Trans. Comm., vol. COM-33, pp. 248-258,
Determination of SNR. TOP and IMP. March 1985
The SNR, THD and IMD can be estimated by performing a least FIGURES
squares fit of a number of reference signals (eg. DC, the input
signal and its harmonics) onto the output of the modulator [6]. pre- convert re-order run
This can be done in either in the time or the frequency domain. processor to - o operations circuit
The least squares element has several inputs which are signals
(the reference signals) and a number of outputs representing the operations
signal power, the noise power, the proportion of each reference
signal in the signal being analysed and the power of each of
reference signal. These can be then be combined to calculate the
SNR, THD, IMD as required. Fig. 9. shows the analysis part of convert to ops | run
a network. The user constructs the analysis network in the same
way as for Fig. 8. Each blockrepresentsa single element, which library ot element types
may be composed of several operations. The SDM could be that
of Fig. 8. Its output is stored to form a signal. The multirate
logic automatically ensures that the least squares estimation, ratio, Fig. 1 The program runs in four stages and makes use a large
dB conversion, rectangular to polar and print operations are only library of element types.
executed only once, after each simulation ran. The great
flexibility afforded by the above technique arises by not splitting
the simulation and analysis parts of the program. dependent ( ~ generic
inputs ""S - output
operation
IV. CONCLUDING REMARKS
A method for efficient simulation of multirate sigma-delta
systems has been presented. The algorithm has been encoded
using the ' C language, under the UNIX operating system. The
program has proved to be very useful and powerful. The
program has been used to evaluate many different topologies and clock dependent
has replicated the results of the purpose written programs used to inputs
analyse a third order cascade [7], a fourth order single loop ADC
[8] and a converter with a multibit quantiser but with single bit
feedback [9]. Fig. 2 Each operation has a single output, but can be dependent
on several input nodes. The clock dependent inputs
determine the exact behaviour of operations in multirate
ACKNOWLEDGMENTS networks.
Support from Wolfson Microelectronics, Edinburgh, UK and
the Science and Engineering Research Council is gratefully
acknowledged.
gam
REFERENCES linear/
nonlinear
[1] R.E. Crochiere and L.R.Rabiner, "Multirate Digital Signal
Processing", Pretice-Hall, Englewood Cliffs NJ, 1983
[2] C M . Wolff and L.R. Carley, "Simulation of Delta-Sigma
Modulators Using Behavioural Models," IEEE ISCAS 1990,
p376 Fig 3 Representation of a simple gain.
1330
400
resample
output
input
delay 1 delay 2
sample
Ahold input circular FIR output
buffer [vector dot
sroduct)
\ i
Clock
Fig. 6 The sample and hold operation is executed at the rate of F i 7 Decimating FIR filter. The input is stored at the input
the output rate ^ ^ cajcuia^d the lower rate.
o u t p u t i s o n l y a t
Integrator 1 Integrator 2
quantiser
Fig. 8 A second order sigma-delta modulator [10]. For simulation, the nodes must be labelled and the system entered
as a net List.
open
tsignal power
file
oise power
.SNR
SDM store
Least ratio ratio print
Squares todB to file
Estimation
ac Element
source JB-> ref.1
inear rectangular
to polar K i
print
to file
ref.2 magnitude j
phase
sweep signal nodes file node
ref.3
1331
4 01
PAPER 50
Proc. I.E.E.E. I n t e r n a t i o n a l Symposium on C i r c u i t s and Systems, Chicago, 1993.
Abstract - This paper discusses the design of sigma- this ensures that the sigma-delta loop is not delay free [4). There is no
delta modulators specifically for digital audio. constraint on where in the unit circle the poles may lie; the constraint is
Approximation techniques for the signal transfer function that the multipling constant must be unity while the high frequency
and the noise transfer function are described which give gain must be around +3 dB. A convenient way to satisfy these
an audible noise improvement over classical functions. requirements is to use BuUerworth or Inverse Chebyshev high pass
The noise transfer function shapes the quantisation noise filter pole positions and to adjust the cutoff frequency to move the
to be least audible to the ear, while the signal transfer poles towards or away from the unit circle to adjust the gain to the
function is given preemphasis to match the dynamic range desired value. A simple computerised iteration process can be used to
of the modulator to the dynamic range of typical audio achieve this.
sources. These functions are then transformed into
snitched capacitor networks for the loop filter in an ADC. B. NTF numerator design
These give a combined improvement of 24 dB in weighted
SNR for the example modulator, a seventh order, 32 times The zeros of the N T F can be chosen almost independently from
oversampling one bit modulator. the poles because, for reasonable oversampling rates, moving the
zeros through the signal band has very little effect on the response in
the noise band. Rather than using classical approximations (either
Butterworth with all zeros at D C , or inverse Chebyshev with an
I. INTRODUCTION equiripple noise response), the method adopted here involves placing
the zeros so as to minimise the audible effect of the shaped
Sigma-delta modulation has recently become the preferred method quantisation noise. Fig. 1 shows the typical response of the ear to
for conversion between analogue and digital domains for audio low level sounds. This curve, referred to as the F-curve, was derived
signals. The main reason for this is the inherently high linearity which
can be achieved with standard C M O S 1C processes without the need
for trimming. T h i s paper discusses the design of the loop filter
transfer functions for higher order modulators which are optimised for
the particular case of audio.
Audio Weighting Curves
Three areas of audio sigma-delta modulator design are described
20
here. Firstly, the determination of stable noise transfer functions
( N T F ) designed using arbitrary approximation methods so as to
minimise the audible effect of quantisation noise. The possibilities of
introducing preemphasis into the signal transfer function ( S T F ) are
then introduced and a direct method for the design of analogue loop
modmed
filters is given. The paper concludes with an example.
curve
20
To avoid confusion with standard filler terminology, the terms
signal band and noise band are used here. The signal band refers to
the audio region (typically 0 to 20kHz) and is the stop band of the
-40
N T F and is contained in the pass band of the S T F . The noise band
refers to all other frequencies.
Y ^ f u i r ' ( l - z ')
n+i 2
= -an.i X 2 1 + (b-2)z> + z- (7b)
(4)
where F(z) is the F-curve and Dec(z) is the Decimation filter response.
The magnitude of Dec(z) can often be assumed to be 1/STF over the
I 3 > output
C. STF Design
NTF : . N _ y . D _ Nq D N
n l f « f - mf
_ >
Dmf Q" D + N Q D N„„ (8)
and:
• I =Mq (10) The loop must then be simulated in the time domain and capacitors
scaled to limit the maximum internal swings. It is possible to reduce
where I is the column vector of unknown capacitances, UQ is the capacitance spread by reducing some of the internal signal swings.
column vector of desired loop filter numerator coefficients and the While in the case of ordinary S C filters this is at the expense of
columns of A contain the coefficients of the transfer functions from dynamic range, the action of the noise shaping serves to greatly
each feedback input to the output which are now known as the values reduce the effect of thermal noise from the integrators and switches
of the b coefficients have already been determined. The transfer towards the comparator. Only the noise from the first op-amp and the
function coefficients for A are easily determined from the switches around it (and to a lesser extent, the second) are critical to the
representation of the loop filter given in Fig 4. The seventh order loop performance of the complete modulator.
signal input
1 1 1 1 1 1
1 1 1 1
D z OJz) tl(z) .output
1
to com
parator
t 1 t 1
feedback
from
comparator
4
IV. EXAMPLE
The example is a very high order (7th). very low oversampling Noise Transfer
ratio (32) modulator. This was chosen to investigate the potential Chebyshev Functions
performance of very low oversampling one bit modulators. The hit
rale al the output of this modulator is only a faclor of two over
ordinary 16 bit pulse code modulation (PCM).
The simulations were all performed using a mullirate discrete time O 5000 10000
simulator [5].
Froquoncy (Hi)
V. CONCLUSIONS
Rg. 6 Weighted noise transfer functions.
It has been shown that significant audio improvement can be
achieved in sigma-delta modulator designs by shaping the N T F and
S T F with non-classical functions. The direct transformation of these
functions into S C loop fillers has also been illustrated. The methods Minimal Weighted
should also be applicable to other application areas where a non-flat Noise Designs
noise spectrum and/or non-flat overloading characteristics are • pra-amphaai!
appropriate.
no-preomph
REFERENCES
PAPER 51
P r o c e e d i n g s o f IEE C o l l o q u i u m on O v e r s a m p l i n g Techniques and S i g m a - D e l t a M o d u i a t
London, March 1994.
Wolfson Microelectronics Ltd, Lutton Court, Bernard Terrace, Edinburgh EH8 9NX
Abstract The linear delta modulator is shown to be a special case of the more general
sigma-delta modulator, in which the signal and noise transfer functions are equal. Thus
the extensive knowledge of delta-modulator systems may be applied to sigma-delta
modulators and the more recent research interest in sigma-delta modulators applied to
further refine delta modulator systems. High order sigma-delta systems may be
designed which retain the single pole -6dB/octave overloading characteristics, but with
the advantages of higher order noise shaping. Techniques for adaptive step size, both
instantaneous and syllabic may be adopted to give companding operation. By damping
the loop integrators slightly, the cutoff phenomenon may be used to give a noise gating
effect.
These techniques have been used for a design study for the integration of a switched
capacitor speech coder/decoder with a bit rate of 32kbits/s to 64kbits/s. The resulting
system gives a significant performance improvement with similar complexity compared
with a linear pulse code (PCM) approach using the same bit rate. In this application,
there is no requirement for digital decimators and interpolators to process the digital
data.
3/1
408
the system idling and so encode the input. The general sigma-delta loop is shown in
Figure 1. For a delta modulator N s i g ~ D , while for a basic sigma-delta modulator
=
NsiG N q . The signal transfer function (STF) is defined as the transfer function
from the signal input to the output and the noise transfer function (NTF) is defined as
the transfer function between the quantiser and the output (assuming that the quantiser
has been replaced by a linear gain).
Figure 2 shows how the quantisation noise in a third order delta-modulator can be
rejected at the frequencies where the ear is most sensitive. This modulator gives an
SNR of 33dB over a 4kHz bandwidth (eight times oversampling). The delta modulated
signal is reconstructed using an integrator and a lowpass filter to remove most of the
quantisation noise.
The loop filter is implemented on silicon using standard switched capacitor filter
methods.
Companding Step size adaptation, as used in delta-modulator systems [3], can also
be used for sigma-delta modulation. This allows the oversampling ratio to be
significantly reduced while maintaining a single bit output per sample (Figure 3).
Syllabic companding extracts lowfrequencyenvelope information from the encoded bit
stream using a coincidence detector, which looks for strings of three or four ones or
zeros in the bit stream, together with a filter of time constant of the order of 10ms. The
coincidence detector has zero output, except when the last three or four samples are all
the same. This indicates that the modulator is potentially overloaded and that the step
size should be increased. The output of the coincidence detector is low pass filtered
with a cutoff frequency which allows the signal level variations to pass through. The
filtered coincidence level may be used directly to set the step size. Different
companding ratios may be implemented by changing the gain or shape of the mapping
between the filtered coincidence signal and the step size.
For syllabic companding it is necessary to encode and extract the signal level from the
bit stream in order to determine the current step size. This is done by considering the
effective operating point on an un-companded system. Figure 4 shows the limited
dynamic range of an un-companded second order sigma-delta modulator and Figure 5
shows the corresponding filtered coincidence levels. The 3-bit algorithm gives a better
indication c f the operating point over the amplitude region with best SNR performance.
The coincidence level is used to convey the signal envelope information. Figure 6
shows the coincidence level and corresponding step size for a companded system, and
Figure 7 shows how the SNR vs input amplitude curve of Figure 4 has been
"stretched" out over a wider dynamic range.
The long time constant required by the syllabic filter is better implemented using simple
digital methods rather than with switched capacitors. This gives more flexibilty in the
forming of the current step size from the coincidence filter output. The filter can be
efficiently implemented using a digital rate multiplier or with bit serial techniques. For
speech, the attack time constant is made smaller than the decay time constant in order to
provide a better match to the envelope of syllables. The variable step size required in
the analogue feedback path can be implemented by using a range of feedback capacitors
and also by switching the capacitor to the reference several times within a single sample
period. The non-linear function for mapping coincidence rate to step size may be
3/2
409
The use of a coincidence detector is less effective for the third order system. To operate
the higher order system in its stable region requires that the input is smaller (maximum
peak -6dB relative to the feedback level) and consequently long trains of ones or zeros
cannot occur. Syllabic companding can be implemented however, by extracting the
magnitude of the reconstructed signal and using the magnitude to adapt the step
size [5], Companding causes the modulator to be used near its optimum operating
point for wider ranges of signal levels. For high order systems this can assist system
stability - any tendency to go unstable will be counteracted by an increase in step size.
An alternate form of companding again looks for patterns in the recent history of the bit
stream, but adapts the step size immediately relative to its previous size.
Bit E r r o r s Since each of the bits in the bitstream carries equal weight to all the
others, analysis of the effect of bit errors is simplified. A single bit error results in an
impulse of twice the current step size. A single bit error will have negligible effect on
the step size. Because of the oversampling, most of the energy in the impulse is
removed by the subsequent filter. The effect of a bit error in a non-oversampled PCM
system can be far greater, resulting in a full scale impulse with little subsequent
filtering.
Conclusions A variety of low bit rate, low complexity, coders have been studied
with a view to their cost and practicality for integration. Simulations have been
performed using real speech signal sources to compare performance. Recent work has
allowed sigma-delta modulators to be optimised for speech applications. A number of
techniques, originally devised for delta modulation, have also been employed.
References
[1] D. Hossack and J.I. Sewell, "Design of High Order Audio Sigma-Delta
Modulators with Minimum Weighted Noise", Proc. IEEE ISCAS, Chicago,
ppl80-183,May 1993
[2] M.A. Gerzon and P.G. Craven. "Optimal Noise Shaping and Dither of Digital
Signals," 87th AES Convention, preprint 2882 (J-l), New York, Oct 1989
[3] Chong Kwan Un, Hwang Soo Lee and Joo Seok Song, "Hybrid Companding
Delta Modulation", IEEE Trans on Communications Vol 29, No. 9,
pp 1337-1344, September 1981
[4] J.C. Candy and G.C. Temes, "Oversampling Methods for A/D and D/A
Conversion", in "Oversampling Delta-Sigma Data Converters", J.C. Candy
and G.C. Temes (Eds), IEEE Press, Piscataway, NJ, ppl-25, 1992
3/3
4 10
signal
input clock fs
clocked
comparator
feedback
input ^ output
clock fs
clocked
signal input I comparator
single bit output
Loop
Filter J
coincidence syllabic
detector filter
variable
level
DfiC
instantaneous combine
compand logic
3/4
EC
10 -
20
Input Signal Level (dB)
Fig.4. SNR vs input level for second order system.
100
80
60
40
3-bit
20
4-bit
i.s
relative
step size
coincidenence level
30
20
CO
Effective Operating
Point
Derived Signal Level
20 i
3/5
4 12
PAPER 55
4 13
481
E r a : . I . E . E . E . I r ^ e r r e t i c r a l Synprjsiun an C i r a i i t s and Systers, \ f e l . 2 , I m ± n , May 1994.
in oversampled digital to analogue converters [ I). Since the 1 all represent the decimal value -1.25. The symbol
modulator operates on oversampled data, the circuitry used is used to indicate a weight of - I . The advantage of the
must be fast. Conventional pipelining techniques cannot be redundant number system is that additions can be performed
applied to increase throughput due to the recursive nature of without long carry propagation chains. Figure I shows a
the loop. This paper discusses a technique which overcomes SBNR adder. Each adder cell consists of three subcells A,
this problem. B and C which perform an addition of digits with various
ranges of values. The final result is in the same SBNR
A sigma-delta modulator consists o f a loop filter and a format as the input and the carry/borrows can only propagate
comparator in a feedback loop. I f the comparator is a maximum of two digits. Thus the addition time is
considered as an additive noise source, then the transfer independent of the word length. A detailed discussion for
functions from the input to the output, and from the noise implementing these cells is given in [2].
source to the output can be determined. These are termed the
signal transfer function (STF) and noise transfer function Ideally, in a properly designed sigma-delta modulator, the
(NTF) respectively. Proper design of the loop filter ensures additions will never overflow, although this is difficult to
the modulator is stable. The signal transfer function is guarantee. Furthermore, the left most cells may produce
designed to pass the signals of interest while the noise
.1 ... i.. it
transfer function is designed to reshape the noise spectrum ii
appearing at the output so that there is little noise in the
frequency band o f interest. A digital implementation of a
sigma-delta modulator must perform all the calculations
required to obtain a single output sample before work can
start on the next sample. This dependency of the current
output on the immediately previous one limits the
application of pipelining techniques to speeding up the loop
operation. The delay between an input and its corresponding
output is termed the latency. For high speed operation it is
6
important to minimise the latency in feedback loops.
Recede
Recent work has overcome this latency problem for the case
v.! 1.1 1. 1.,
of high speed infinite impulse response (MR) filters [2].
This has been achieved by adopting a redundant number Figure I . Fully parallel adder using SBNR arithmetic
4 14
482
The SBNR adder of Figure I may be used directly to Figure 3. Second order resonator used for loop filter
construct a high speed sigma-delta modulator. However feedback coefficients (fl and f2) are determined by the
there is a further feature of redundant arithmetic which may solution of a set of linear equations involving the desired
be exploited. When conventional two's complement zeros of the open loop filter obtained by algebraic
arithmetic is used in a ystolic type structure, it is the least manipulation of the NTF [4]. Since the input to the gains
significant bits which are performed first since the more fl and f2 can only take on the values +1 and - I , no
significant bits depend on the least significant bits through multiplier circuit is required. The values of the resonator
cany propagation. This is not the case with redundant coefficients (b) are directly related to the open loop poles
arithmetic which allows the order of execution to be reversed which are at the same locations on the z-plane as the NTF
giving systolic type structures which operate most zeros. This is the only gain for which a full SBNR number
significant digit (msd) first. This is in line with the needs to be multiplied.
requirements for a sigma-delta modulator where it is only the
top few digits of the filter output which are required. Data The single delay around the two integrator loop places
skewed msd first has the property that gains of less than one stringent demands on the latency of the integrators. For
may be performed with negative latency. This is illustrated baseband sigma-delta modulators an additional delay can be
in Figure 2. Parallel data enters at the top and is skewed tolerated. The poles move off the unit circle giving the
such that the digits become available in groups of three. NTF zero at a similar frequency as before, but not infinitely
However after the right shift by three digits (a divide by deep. The depth of the NTF notches is of little consequence
eight operation), the digit representing the output is to the output noise level.
available before the digit of equal significance has entered at
the top. Thus the divide by eight operation can be It can be shown that the magnitude of the coefficient b in
considered to have a latency of - 1 . the feedback loop is small while the feedback coefficients
tend to be much larger. If the feedback coefficient is
SIGMA-DELTA TOPOLOGY sufficiently truncated the addition of the two words can be
achieved by merging the two words rather than by an
The sigma-delta modulator topology used in this work is arithmetic circuit.
based on a loop filter constructed from the cascade of
integrators and resonators [3]. Figure 3 shows a single SBNR ADDER
resonator. The gains si and s2 are for scaling only and are
initially assumed to be unity. The values of the comparator For the example sigma-delta modulator, it was decided that
the data would be skewed by one pipeline delay every third
\ digit, as was illustrated in Figure 2. This allows parallel
it It
computation of the top three digits which are required in
S.
O ll it parallel by the comparator. Three numbers need to be added
for each integrator. Rather than cascading two dual input
'I adders of the type in Figure I, a two stage design was
derived using the carry save concept of conventional two's
complement arithmetic. This is shown in Figure 4. Three
SAewod Data' SBNR numbers are added and the effect of transfer
propagation limited to the next three more significant digits.
\ This adder produces its output after a one clock cycle delay
from the pipeline cuts indicated.
SBNR COMPARATOR
Figure 2. Parallel to skewed data conversion followed by The sigma-delta modulator requires a comparator. In
divide by eight operation conventional two's complement arithmetic, the output of the
4 15
483
Figure 4. Three input pipelined adder operating on three at a time skewed data
comparator is simply the sign bit of the input. However coefficients (with sign applied by the comparator output) to
this is the result of a long carry propagation chain. In start entering the adder pipelines as early as possible. The
SBNR there is no specific sign bit. The sign of the overall single pipeline delay on each of the three inputs of the two
number is the sign of the left most non-zero digit. Logic is adders represent the pipeline delay internal to the adder.
required to extract this from a parallel SBNR word.
Fortunately the comparator performance in sigma-delta Note that whilst it is convenient to consider the resonator in
modulators is non-critical because any errors are reshaped Figure 6 as being multiplexed, this is not truly the case. At
with the full noise shaping function of the loop. In practice the same time as one adder is computing the most
only the top three or four digits need to be examined as the significant digit of the output, other parts of the same adder
additional truncation noise is less significant than the are computing other digits of the other resonators and even
quantisation error. different digits of the previous values of the same word. The
multiplexer for adjusting the scaling gain (between 1/32, 1/4
In the circuit presented, the comparator has one full pipeline and 112) and the multiplexer for feeding the output of one
delay in which to come to its decision and the speed of the resonator back into the input for the next must both take the
comparator circuitry is not a limiting factor. skew into account. These multiplexers are actually taking
different digits from each of their inputs to form the output.
SBNR S I G M A D E L T A MODULATOR
The number in brackets beside some of the gain blocks
A sixth order modulator consisting of a cascade of three indicates negative effective latency. These numbers need be
resonators was designed (Figure 5). The signal transfer taken into account when counting pipeline delays around
function was designed with signal band gain of -6dB to loops. Note that there is the correct multiple of three
ensure that a full amplitude input signal (+1 - I ) did not pipeline delays around all the resonator loops. However
cause the modulator to become unstable. The noise transfer there are only five delays around the loop which links the
function was designed with Butterworth poles so that the output of the resonator back to its input. This ensures that
signal transfer function could have a low pass Butterworth the data representing the output of the first resonator arrives
response and be implemented with a single feed-in at the right time to be interpreted as input to the second and
coefficient into the first integrator. similarly for data leaving the output of the second resonator
and entering the third. The output of the final resonator is
Each integrator output in the cascaded resonator topology is not feed back to the input - during this time the system
dependent only on the previous output of the previous input is sampled and multiplexed into the first integrator.
integrator. This fact is used to calculate the data in the three
resonators by using a single pipelined resonator with three The scaling factors were chosen to the nearest power of two
pipeline delays corresponding to a single z"' delay. The in order to roughly equalise the signal amplitudes at each of
coefficients to the single resonator are multiplexed and the the integrator outputs. The resonator coefficients were also
state information for the three resonators contained in the quantised to powers of two. Because the resonator
pipelines (Figure 6). The multiplexers are arranged such that coefficients are small (1/128) two extra pipeline delays
the function of the third resonator is performed first, become available which allows closing the resonator loop
followed by the second and then the first. This is to allow with only three pipeline delays (corresponding to a single z-
the comparator to come to a decision and for the feedback I delay) and so giving NTF zeros which lie exactly on the
4 16
484
Comparator
CH/3
Figure 6. Sixth order sigma-delta modulator using single pipelined resonator section
unit circle. Additional adders could use the three additional with an undistcrted single tone. The signal to noise ratio
delays available to place the NTF zero frequencies more from this simulation is 85dB with the input 20dB smaller
accurately. It is likely that a low latency SBNR multiplier than full scale. The modulator was designed for an
could be used to give programmable NTF zero frequencies oversampling ratio of 50.
with arbitrary precision.
This paper has concentrated on using SBNR arithmetic. The
Since the two's complement number system is a subset of fact that it was desirable to calculate three digits in parallel
SBNR there is no hardware overhead in converting input in suggests that a higher radix redundant number system may
two's complement form to the SBNR form required by the be more suitable and more efficient in hardware.
first adder. The word length of all SBNR data was 5 integer
digits and 20 fractional digits and this could be increased CONCLUSIONS
without decreasing the throughput rate.
Features of SBNR arithmetic have allowed the design of a
The sixth order system of Figure 6 has been simulated at the sixth order modulator (Figure 6) which produces a new
gate level using a commercial digital simulator. The output sample every third clock cycle, with the clock rate limited by
spectrum (Figure 7) shows the expected sixth order shaping a single proprogation delay through the adder of Figure 4.
Many other design decisions could have been made, ranging
from fully parallel designs where the order and word length
could both increase without affecting throughput, to more
heavily pipelined systems using a single adder.
REFERENCES
[lj J.C Candy, G.C. Temes (Eds), "Oversampling Delia-Sigma Data
Converters," IEEE Press, 1992
[2] R.F. Woods, J.V. McCanny. "Design of a High Performance IIR
Digital Filler Chip," IEE Proceedings Pan E, vol. 139, no. 3, May
1992, ppl95-202
[3] R.W. Adams, P.F. Ferguson and A. Gansesan, "Design of Single
Bit Noise-Shaping Loops with High Order Loop Filters", AES
89th Convention, Los Angeles, Sept. 1990, Preprint #2974.
[4] D M. Hossack, J.I. Sewell, "Computer-Aided Design of High Order
.0001 .001 .01 .1 Sigma-Delta ADCs," Proc. IEE Colloquium on Advanced A-D and
Frequency (sample rate = 1.0) D-A Conversion Techniques and Applications, pp 2/1-2/4,
London, May 1993
Figure 7. FFT of output of sixth order system
4 17
Publication Page
PAPER 11
A course in computer-aided
electronic circuit design for
undergraduates
A. G . Martin and J . I. Sewell
The problem of including a course which demands the successful circuit in a real design situation. Past experience
use of a computer by undergraduates «i engineering has shown that the point of unreality is reached when,
curricula has been of interest for some time. It is of course having produced a suitable design, the student wishes to
essential that the student engineer acquire a respect for breadboard his ideas for final proof. Particular points of
computational techniques and especially that he develop friction occur with designs marginally near the specification
a sense of how best to set a computer to work. when neither student nor tutor is sure of the performance.
With computer-aided instruction, as with any means of The policy of requiring designs to be proved on paper
learning, the object must be to create an initial exercise because of the impossibility of breadboarding many
which the student thinks he can solve. Once guided on to modern designs is rarely convincing. Computer checking
the right path the student should uncover a succession of of designs not only eliminates the breadboard problem
further goals, providing stimulation to learn for himself. because it is quicker but also acquaints the embryonic
Open-ended problems of this sort are often met by the designer with a modern technique. I f additionally, as in
designer and though design teaching is often omitted from this case, the student is meeting computer aid for the first
engineering courses it can provide a natural slot for time, an immediate effective demonstration of its versatility
computer familiarization. is achieved.
A well-constructed computer-aided design course A simplified representation of modern circuit-design
introduced at an early opportunity can demonstrate the philosophy is given by the design Row chart in Figure 1.
power of computational techniques without the student Some details of certain real-life situations have been
first stumbling through the constructed problems of omitted and the chart is terminated prematurely since no
introductory programming. As well as stimulation, some details of computer aids for circuit production are in-
disincentives must be made apparent, by pointing out that cluded. However, such material can be found elsewhere and
inspired paper work produces quicker but fewer results anyway it is not particularly significant for a first design
and by careful user monitoring to prevent wastage. A n course. I n the time allowed for the design course it is not
appropriate moment in an open-ended problem can be possible to complete the whole scheme, and therefore
chosen for introducing time limits so that cut and try has certain steps are omitted or left for the brighter student.
to be replaced by planned optimization. Hence in the computer-analysis step only a.c. steady-state
analysis was undertaken by all students. A limited number
Design projects of students did arrive at the stage of undertaking a simple
tolerance analysis of the circuit on the machine. However,
Electronic design work has been a feature of the under- this process could not amount to more than investigating
graduate course at the University of Hull for some time. worst-case effects on circuit response.
Projects are introduced just before the halfway stage in the
Convincing the student that he can solve the problems
course, when about 40h of circuit theory and electronics
is a matter of posing suitable problems. The student is
course-work have been given. The design laboratory is
invited to choose one from the following three:
intended to act as a focus to encourage the student to
consolidate the analytical techniques he has just learned. 1. Transistor pre-amplifier. Design an a.c. pre-amplifier
The motivation for learning in such a course is the which has an input impedance greater than 1 MQ and
prospect of producing with some independent effort a a voltage gain of 20 + 1. The maximum output
I2O0
JI
IOOUF 20kfi 20k!? 2Ok0
OOI32uF
IOOUF
O OI3 2uF
OOI32iiF
8Ok0
1
33k0
From the flow diagram of Figure 1, it can be seen that
the course has concentrated on a simplified path through
the chart. Obviously any course is limited mainly by time
available, and hence only the design of linear circuits was
clone. The computer was used only for a.c. analysis and
PAPER 27
4 2 3
Int. J. Elect. Enging Educ, V o l . 18, pp. 29-34. Manchester U.P.. 1981. Printed in Great Britain
J. I. SEWELL
1 INTRODUCTION
29
4 2 4
30
sake of stability within departments during the transition period, which may
last a number of years, some concrete plans have to be made and adhered to.
Since the degree courses in electronic engineering at the University of Hull
had traditionally consisted of four years full-time study, it was somewhat
logical to consider the pursuance of the proposals for 'enhanced' degrees where
the basic academic unit was four years in length.
2 BASIC E L E M E N T S O F E N H A N C E M E N T
2
Following the Merriman Report, the I E E . issued a guide outlining the
contents of degree courses in electrical and electronic engineering, which would
be the normal route to qualification as a chartered engineer in the future. The
salient points are listed and the ensuing course proposals are an attempt to
implement these in the light of local circumstances and resources.
(i) The complete course will be about five years in duration.
(ii) The enhanced course will contain 20-30% more academic material than
existing undergraduate courses.
(iii) The industrial involvement will be of about 18 months' duration.
(iv) There will be close links between the educational establishment and
employers of electronic engineers. A proper integration of practical and
theoretical studies is sought, with no rigid separation of industrial and
academic aspects of engineering.
(v) The course should contain topics such as industrial organisation, com-
munication and man-management skills.
(vi) The major technical components of the course will be supported by
suitable practical and project work, and contain elements of the design
approach.
(vii) The entry grades of A-level (General Certificate of Education, Advanced
Level) candidates should total nine points or over, equivalent require-
ments being demanded for other qualifications.
Superimposed upon these requirements are extra ones felt to be necessary for
the complete operation of the courses in the university environment.
(a) The course should offer some degree of choice of specialisation at the
highest level possible.
(b) There should be adequate provision for those students who find the
enhanced course studies too demanding, and need a safety net of a Pass
Degree course of non-enhanced standard.
(c) It is not at all clear that overseas students will wish to complete the full
enhancement requirements, and provision for a non-enhanced alternative
is necessary.
^ON-
ENHANCED
3
ASS
lOOSSi
3.Sc. B. S c . Ennanced
Pass Honou r Degree
Non- and (l.Eng.
Enhanced Pass or
Dipl.Eng?)
by a complete year during the fifth year. The fourth year of academic study
offers three options to the student, who can specialise in communications,
automatic control or computer engineering. O n completion of the fourth year,
the B.Sc. degree is awarded with the usual gradings of Honours and Pass
categories. Overseas students and those not wishing to proceed to full enhance-
ment may terminate at this point. T o complete the enhanced course, students
must submit a dissertation on their industrial training and experience at the
end of the fifth year, when the fully enhanced degree will be awarded
(Dipl.Eng., M.Eng. or whatever).
The complete details of the academic content of each year are shown in Figs.
2 and 3. It is perhaps easiest to consider each year as consisting of a block of
basic subjects, which begin with the foundations of electronic and mathemati-
cal studies and develop in natural progression. Hence, in the first year this
comprises mathematics, combinational logic and integrated logic circuits,
computer appreciation, electronic circuits, circuit theory, electromagnetics,
system dynamics, mechanics and engineering drawing. Core subjects in the
FIRS 1
! SECOND YEAS THIRD YEAR
i
Ka \ - e ~ a : i I i ec ; r o . " <
' c S v s i et"s and Digital E l e c t ror. i cs
;
Co-: i o-ia ' -cc S c a-- f. i c r o Co~D'j:e- Sy?:e~ ~
; 1
1 " »
' e c r a i ec Loc c C i r c ~ ;'. . L ! c c i "O ". i c Dev <
' ;es Tc 1 e c o n n u - i c a i i o r . s
CO-OJ'.C'- Arorc: a: . o- 1 r o j j c tor-. Te 1 e - Svste-s T^or-.
i i L-C- r r ' c C i r c . ' : -: co^-tj" 'Cat i cnt A j ' . o r a ; ic Co-.: r c l
C i .- - L> : * e: r . Ne • i / o ^ - Ana ! y i i s Aef i e ' s . waveguides, 1ra^s-
r r , :
t i c : ' 'D"-ac.r-:-; i c s t t e ; : o ~ a p e : i s~ ^'S^io" Lines-, C'Diica Wave-
S, : : o - I- \ ~ a ~ ' c h D i c i i a ' i 1 e: t ro- > guides
Na-nt-a; : :
p ; ;
ract ca i Circuit t - e s i c~ S t g n a i = A".a 1 v? s ;
" a : "ie™.at i c s Ges i •- : c 1
nee ' i n
c j
r
' c: ; ca ' r
C o n u - • ca : ;
C Q~D j t e r P"" Og a "V t -• c
r r> r r
E- J ' ->ee i r. _ a' ; E » ' o - : - :
St"':: Siair ~(:C^c'.OC\
1
_a " a: o • . *c ' - LSDC-'ot Dr\ WO "-.
second year are electronic systems and feedback, electronic devices, introduc-
tory telecommunications, network analysis, electromagnetism, digital elec-
tronic systems, mathematics, and solid state technology. In the third year these
are digital electronics, microcomputer systems, telecommunications, systems
theory, automatic control, aerials, waveguides, transmission lines, optical
waveguides, mathematics and signals analysis. The fourth year presents three
options of electronic specialisation with a few subjects common to two or more
options.
Added to the basic core are other essential and interlinking features which
progress from year to year. Fig. 4 shows the practical and project content. The
proportion of project work is increased steadily year by year until in the fourth
year each student undertakes a major project investigation which may be
directly related to an industrial problem. Fig. 5 displays the influence of the
computer in electronic engineering education. It is necessary now, more than
3
ever , to ensure the correct balance of these studies. All graduates should feel
competent in using both general machines for problem solving and the more
4 CONCLUSIONS
There is a sense in which engineering degree courses have never been easy.
They have always covered a wide range of introductory subjects and therefore
the workload imposed on students is usually heavier than in some other
disciplines. Fortunately these aspects have not proved completely unacceptable
in the past, since the motivation provided by vocational studies tends to
prevail. It is essential however, when attempting to implement the proposals for
enhancement, to maintain a correct balance of subjects in such a manner that
the student can recognise that his specialised technical knowledge is expanding
at a rate which exceeds his growing appreciation of general engineering
ambience. The problem is nowhere more critical than in electronic engineering
where there is the constant pressure of new technology. But what value is this
1
FIG. 6 Industrial lecturers' input.
5 REFERENCES
[ 1] The Institution of Electrical Engineers, 'Qualifying as a chartered electrical engineer'. The
Merrimun Report (1978).
[2] The Institution of Electrical Engineers. Engineering degrees and the chartered electrical
engineer' (1979).
[3] Martin, A. G . . Midgley, D. and Sewell. J. I. 'Computer-aided engineering education'. Proc. of
Conference on the Teaching of Electronic Engineering in Degree Courses, University of Hull,
pp. 17. 1-17.2 (1973).
ABSTRACTS-ENGLISH, F R E N C H , G E R M A N , SPANISH
Publication Page
[57] Accurate semi-symbolic analysis of large non-ideal switched linear networks 468
[58] A systematic approach for ladder based switched-current filter design 473
PAPER 56
431
E F F I C I E N T DESIGN OF LADDER-BASED A C T I V E F I L T E R S
AND E Q U A L I S E R S
1
Lu Yue, BSc, MSc,
2
N.P.J.Greer, MA, MSc, PhD,
1
J.I.Sewell, BSc, PhD, C Eng^FIEE,
1
4 32
ABSTRACT
1. INTRODUCTION
The oldest and probably best known method of active filter design is a cascade approach.
This commences by factorising a prescribed high-order transfer function into a product of
second-order terms, which are then realised by an appropriate choice of active biquadratic
sections. This method offers many attractions; principally it is straightforward to
implement and various practical circuit requirements are almost automatically satisfied,
such as dynamic range scaling which can be ensured by simple voltage or current scaling
in each section, and in SC realisations the capacitance spread is generally quite modest.
Consequently, many of the filter compilers available[l] adopt this approach as standard.
However for high order monolithic filters, the cascading of biquadratic sections leads to an
unacceptably high sensitivity of the response to component parameter variations. It is well
known that appropriately designed LC ladders have very low sensitivities to component
tolerance and hence active filters simulating the internal workings of doubly terminated
RLC ladder prototypes are widely used in high precision integrated filters. Passive ladder
prototypes for standard lowpass approximations are readily available from filter tables and
component values of all other types of frequency response (bandpass, bandstop and
highpass) can be obtained by applying standard frequency transformation and scaling
2
4 33
Progress from a prototype ladder to an active realisation can follow "direct" simulation,
when unwanted elements such as inductance or resistance are replaced by simulated
elements, or "signal flow graph" simulation which involves the representation of selected
currents and voltages (or combinations) of the prototype by proportionate voltages,
currents or charges in the active circuit. A powerful matrix based approach[2] has been
developed recently and applied to the design of SC, active-RC, and digital ladder filters
and equalisers. In the case of SC filters this has led to configurations which are far from
intuitively obvious but which offer improved performance with respect to parameters such
as settling time and capacitance spread. The matrix approach has also been used to develop
novel filter structures, ideally suited for transconductor-C realisation[5].
This paper presents various solutions to the problem of efficient active realisation of
ladder-based designs. One difficulty concerning canonic realisation of certain filter
functions has been recognised previously[6]; here, a general solution and a wide range of
its implementations are examined. The application of advanced approximation
techniques[7] also permits minimal realisations and improved performances in
filter/equaliser design. The filter compiler incorporating these techniques is demonstrated
and results from a fabricated video filter and group delay equaliser are presented.
A passive ladder can be represented by the standard nodal admittance matrix equation
J = (G+sC+s-lr)V (1)
Where V is a vector representing the nodal voltages and J is a vector representing the input
current sources. G,C, and T are admittance matrices formed by the contributions of
resistors, capacitors and inductors respectively. Equation (1) represents a set of equations
of second order in the Laplacian variable s. It is well known that a set of linear first order
algebraic equations can represent a signal flow graph and be realised by active building
blocks[8]. The matrix method[2] enables direct decomposition of the second order matrix
3
434
to give two inter-related first order equations, and these first order systems are then directly
implemented using the active building blocks.
Either the C or the T matrix can be factorised, leading to left and right decompositions
respectively.
C=C,C r (2)
If r is factorised as
r = r, r r (4)
then the RC Right Decomposition method is obtained and the following pair of equations is
equivalent to (1)
1
(C + s- G)V + s-^/W = s-!j (5a)
-s-lr V + W = 0
r (5b)
For the above decompositions, in order to maintain the sparsity of the matrix, no matrix
inverse is involved. If inverse matrix is used in the design process, some special circuits
can be obtained.
4
435
If the auxiliary variables are defined by W = sCV and (3b) is multiplied by C~l then the
design equations become
_1
Choosing W = s rV and multiplying (5b) by T"l the system equations become
1 1 1
(C + s- G)V + s- W = s- J (7a)
-s^v + r^Wr^o (7b)
The equations (3,5,6,7) are equivalent to (1). Each set can be realised by active-RC
1
circuits. In these equations the matrix multiplied by s" is called integrated matrix and it will
be observed that in the active-RC circuit realisation, the entries in the integrated matrices
are realised by resistors and the entries in non-integrated matrices are identified with
capacitors. It can be shown that equation (6) yields a minimum capacitor realisation while
(7) produces a minimum resistor realisation[2].
The design of a switched-capacitor filter starts from the prototype system equation (1),
which after bilinear transformation becomes
- 1
V =J (8)
Tl +z 2 1-z
Rearranging gives
- A + \(/B + D V = J(l + z- ) 1
(10)
5
436
where
The bilinear transformation has the advantage of both stability and exactness.
Unfortunately, bilinear integrators are sensitive to the stray capacitance and are not
practically useful. Instead a modified SC ladder-based structure utilising LDI integrators[2]
can be formed.
Again, the standard numerical techniques for matrix decomposition of A or B yield a range
of circuit implementations whose suitability in various applications have been exarnined[2].
For the sake of completeness, the two inverse decompositions are shown, though no
economical SC circuits are known to result from their application.
1 1
(A + <DD)V + <DW = J(l+z- )/(l-z- ) (15a)
- y V + B-1\V = Q (15b)
6
Generally the SC Left Decompositions demonstrate excellent properties regarding
component spread and dynamic range for bandpass designs. However, these structures
sometimes need more opamps in circuit realisation. They are not suitable for lowpass
filter design because of a peak in sensitivity at zero frequency.
The SC Right Decomposition approaches can demand fewer opamps in circuit realisation
and they demonstrate good sensitivity performance for lowpass designs. Undesirably
large component spread and poor dynamic range are unfortunately observed for certain
bandpass designs.
TC Topological Decomposition
T
T=ADA (16)
W=(sg)-1DATV (17)
where g is a scaling factor with the dimension of transconductance, often set as g=l/aR
where the R is the ladder filter terminal resistor and a takes an optimum value close to the
fractional bandwidth of the filter.
7
4 38
s -l 2
g A W + ( C+s-lgG)V = s-!gj
g (18a)
1 1 T
gD" W + s- A V = 0 (18b)
If the prototype ladder is equally terminated, the filter can be realised with one value of
transconductance. In a transconductor-capacitor realisation, the entries in non-integrated
matrices are realised by capacitors while the components in integrated matrices are
implemented by transconductances. All non-integrated matrices in the Topological
Decomposition are symmetrical, so realisation can utilise both conventional and low
impedance input transconductor structures. The auxiliary voltages W are directly
proportional to the currents in the inductors of prototype ladder. The topological
decomposition will generally lead to equivalent leapfrog topologies in active-RC and SC
and transconductance-capacitor designs, if conventional transconductors are used. In the
bandpass case, this decomposition cannot be guaranteed to generate stable active circuits
and alternative decompositions are required[5].
w=(s-irv)/ g (19)
1 1
gr- w- - v =o s (20a)
s -l 2
g w -lgG)V = s-igj
+ ( g C + s (20b)
In (20b), the integrated vector V is premultiplied by the matrix G. Only when the
prototype ladder has equal resistance at both terminations, would the transconductor-
capacitor filter require one transconductance value in the complete realisation. Because all
the non-integrated matrices in TC Right Inverse Decomposition are symmetrical, it can be
realised both using conventional transconductor and low impedance input transconductor
building blocks.
The Left Inverse Decomposition results from decomposing the C matrix into two matrices
as (2). Defining the vector of auxiliary variables by
8
4 39
The general TC Left Inverse Decomposition design equations are obtained by substituting
(21) into (1) and rearranging:
Only low impedance input transconductors can be used in the realisation of Eq.(22a,b),
because some of the non-integrated matrices are asymmetric. For C/ = I and C = C, the r
Since r~l and C are generally full and tridiagonal respectively, a relatively large number of
capacitors is required in the circuit realisation. However since all non-integrated matrices
_
( r l G and C) are symmetric, they can be implemented using conventional
transconductors.
1 l
g 2r- CW+(s- g+gr-lG)V = g T - l j (24a)
-s-!gW+V = 0 (24b)
This requires fewer capacitors than Left-IC because the only coupling capacitors are those
of the product matrix F'^C, these coupling paths must be implemented using low
impedance inputs since P^C is generally asymmetric. A singularly good application [5]
of this decomposition is for prototypes where C = T, when r ~ l C = I which is very
sparse and symmetric, and produces a highly efficient realisation with conventional
transconductors.
The important feature of equations (22-24) is that all of the integrated vectors are multiplied
by a single constant g, this allows the equations to be implemented as a transconductor-
capacitor circuit with a single value of transconductance.
9
440
The equalisation of group delay is becoming increasingly important with high frequency
applications such as video communication systems. In analogue systems this equalisation
is often provided by one or more allpass biquadratic stages. However, where the equaliser
required is of order greater than two, sensitivity considerations would indicate that it is
preferable to use a ladder derived circuit. A ladder-based technique has been developed for
digital implementation[10] and extended to SC realisations[ll]. It is essential to
incorporate the advantages of this approach into the design of transconductor-C equalisers
for direct application in high frequency environments.
P(-s)
H(s) = k
POO (25)
Define
E(s)
if n is even
0(s)
Y(s) = .
O(s)
E
if n is odd
(s) (27)
H(s) = l l ^ l l = "
1+Y(s) 1+Y(s) ( 2 8 )
Since P(s) is Hurwitz, Y(s) can be expanded as a continued fraction, and equation (28) can
then be realised as the combination of a singly terminated RLC ladder and an active
summing stage. The singly terminated passive ladder network part can be simulated by a
transconductor-C circuit using the above matrix decomposition method and the overall
transconductor-C group delay equaliser structure is shown in Fig.2.
10
441
Previous work has shown that the problem of finding a canonical ladder-based active filter
can be solved by finding a canonical ladder prototype[6]. The conditions required for a
transfer function to be realisable by a canonical doubly-terminated ladder are that
numerator of the transfer function of a canonical even-order doubly-terminated ladder is an
odd polynomial, and that the numerator of the transfer function of an odd-order doubly-
terminated ladder is an odd polynomial if ICI is non-singular or an even polynomial if III is
non-singular. For most filter design problems the numerator of the transfer functions are
polynomials with purely even or odd terms. So the numerator parity alone determines
where a given transfer function can be realised by a canonical ladder. The solution is to
augment the transfer function, unrealisable by canonical ladder, to produce a transfer
11
function which is realisable by canonical ladder, and when the ladder is simulated by an
active circuit the original transfer function behaviour is then restored by a change in input
circuitry. If H(s) is a transfer function with all its zeros on the imaginary axis or at infinity,
then parity manipulation can be effected by s, 1/s, sAs^+coi^). The parity of the modified
transfer function H'(s) facilitates realisation by a canonical prototype ladder. A system
realising the original transfer function H(s) can be obtained by multiplying the input vector
J by the inverse of the modifying function.
1 1
(sC+s- r + G)V=s" J (29a)
1
(sC+s" r + G)V=sJ (29b)
1 2 1
(sC+s- r + G)V=(s+o)i s" )J (29c)
In an active-RC circuit realisation, for RC Left Decomposition form, the system (29a-c)
can be decomposed in the following ways.
1 1
C/W + (s" r + G)V = s- J (30a)
-s-lW + C V = 0 r (30b)
1
C/W + (s- r + G)V = 0 (30c)
-s-lW +C V = C / - l j
r (30d)
1 2 1
C/W+(s- r+G)V=o>i s- J (30e)
-s-iW+C^Q-ij (30f)
Now canonical active-RC networks can be obtained directly from equations (30). These
circuits use n opamps, where n is the order of original transfer function H(s). The
1
coefficients of s" on the right-hand-side of the equations are implemented by resistors.
The different parity arrangements and decompositions available lead to different circuit
realisation efficiency. Eq.(30a,b) produce the same topological structure as the original
one, the only difference is that the input capacitance is changed to resistance. Analysis of
Eq.(30c,d) and (30e,f) shows that canonical realisation generally leads to the introduction
of extra passive components. However, if UL or IA decompositions are used, the number
of input branch elements can be minimised. For those two decompositions, only the input
node is changed in case Eq.(30c,d) or one additional resistor is added in case (30e,f). If
other decompositions are used, they can introduce up to N (number of nodes in prototype
ladder) extra input capacitor or resistor branches.
12
443
1
(C + s- G)V + s ^ W ^ (31a)
1 1 1
-s- r v + w = - s - r / - j
r (3ib)
1 1
(C + s- G)V + s" r/W = J (31c)
-s-ir V + W = 0
r (31d)
1 1
(C + s- G)V + s" r/W = J (31e)
1
- - i r v + w = -s- rricoi2j
s r (3if>
Eq.(31c,d) demonstrates a very efficient canonical form with input branches changed from
resistance to capacitance. The equations of (31a,b) and (31e,f) can only be efficiently
realised by UL or IA decompositions, since r/~lj produces a vector with only one non-
zero entry, which can be produced by one input branch only.
1
W + (s"lr+G)V = s" J (32a)
1 1
-s- C" W+V = 0 (32b)
1
W + (s" r + G)V=0 (32c)
1 1
-s- C" W + V = C - ! j (32d)
1 2 1
W + (s- r + G)V = u)i s- J (32e)
1 1 1
-s- C" W + V = C- J (32f)
Equations (32a,b) can be very efficiently realised with the input branch changing from a
capacitor to a resistor. Realisation of equations (32c, d) would introduce N extra
capacitors, but without a resistive input branch. The system equations (32e,f) introduce N
extra capacitors plus the original resistor input branch. In the case of (32c, d) and (32e,f),
the characteristic minimum capacitor realisation of RC Left Inverse Decomposition would be
lost.
13
444
1 1
(C + s' G)V + s- W = J (33c)
1
-s- V + r-lW = 0 (33d)
1 1
(C + s" G)V + s- W = J (33f)
1 1 1 2
-s- V + r-lW = - s - r - o ) i J (33f)
SC Left Decomposition canonical realisations for systems (29a), (29b) and (29c) become
The circuit realisation of Eq.(34a,b) and (34e,f) retains the basic topology with only the
introduction of 2 more switches and 1 more capacitor. The circuit realisation of Eq.(34c,d)
14
4 4 5
1
is very efficient, since J has only one non-zero input and if A f is an upper triangular
matrix, which occurs when UL or IA decompositions are selected, only one input branch is
required.
SC Right Decomposition canonical realisation for systems (29a), (29b) and (29c) are
2
(A + <DD)V + <DB/W = (To)i /2+2/T)J (35e)
l 2
-¥B V + W = -2WE[- a>i 3
r (35f)
Eqns.(35a,b) and (35e,f) will yield identical topologies, having 2N more switches than a
realisation based on Eq.(13a,b). However, Eqns.(35c,d) lead to a very efficient circuit
realisation with N fewer capacitor input branches than demanded by Eqns.(13a,b; e,f)-
2
W + (<DB + D)V = 2T<D(Di J (36e)
1 1 2
-»FA- W + V = A- (T(Di /2+2/T)J (36f)
The realisation of Eq.(36c,d) introduces N more capacitors, and Eq.(36a,b) and (36e,f)
have same topologies which introduce 2(N -1) switches and 2JV capacitors.
15
44R
1
(A + 4>D)V + W = 2B" JAT (37c)
/ 1
-4 V + B " W = § (37d)
2
(A + 0>D)V + <DW = (Tcoi /2+2/T)J (37e)
/ 1 1 2
-4 V + B - W = - 2 W B - 0 ) i J (37f)
Eq.(37c,d) can produce an efficient circuit realisation and Eq.(37a,b) and (37e,f) have
same topologies which have 2(N -1) more switches and N more capacitors.
The U L and IA decompositions are recommended for all cases of canonic SC filter
design, since they minimise the number of the input branches significantly and generally
produce efficient circuit configurations.
A typical 8th-order bandpass filter which satisfies the specification of a speech processing
channel is designed by canonical techniques using the Left-UL decomposition. The
simulated circuit response is shown in Fig.4 and closely agrees with the approximation.
The design statistics for different circuit designs is given in Table I, and the sensitivity
comparison of a ladder-based design and a cascade based design is given in Fig.5. The
lower curve is the sensitivity characteristics of Left-UL realisation and the upper curve is
the sensitivity characteristics of Sedra's[12] biquad cascade realisation. The index of
multiparameter sensitivity is defined as
2-, 1/2
aiH(co)|
S(co) = 8 . 6 8 6 ^ /100.0
|H(CO)| a Ci
16
3.3) Canonical Ladder Simulation by Transconductor-Capacitor Circuits
1
CV = -s- [AgW + GV] (38a)
1 T 1
W = (sg)" [DA V-JA" ] (38b)
The validity of these equations is dependent upon the matrix A being square and having an
inverse, which is not usually the case. Even when A is square-invertable, the
implementation of s ' ^ A ' l j in Eq.(38b) and s'^A^coi^J in Eq.(38f) would introduce a
extra N (dimension of A matrix) transconductors with different transconductance values,
and this destroys the main advantage of a ladder-based transconductor-capacitor filter.
However, Eq.(38c,d) will always yield an efficient circuit structure no matter what form
A matrix assumes.
CV = J - s - ^ G V + gW) (39c)
1
r - l W = (sg)" V (39d)
1
C V = J - s " ( G V + gW) (39e)
1 2 1
r - l w = (sg)- (V-coi r" J) (39f)
17
4 4 8
-
The terms (sg)~lr~lj in Eq.(39b) and (sg)~l(0i2r lj in Eq.(39f) can only be realised
by N extra transconductors with different transconductance values. However for
Eq.(39c,d), an efficient circuit realisation can be obtained.
Only the canonical realisation for Eq.(23) is given here, the other variations on the TC Left
Inverse Decomposition method give exactly the same results. The canonical equations can
be written as
1
C V = s" gW (40a)
g 2 p - i w = s-igr-ij - gr-iGv - s-^v (40b)
1
CV = s" gW+J (40c)
1 1
g 2p-1W = - r - G V - s" gV g (40d)
CV = s' gW + J 1
(40e)
2
g 2 -l r - l g r - l ( 0 i J - gr-iGV - s-!gV
W = s (40f)
- -
The terms s ' l g r J in Eq.(40b) and s"lgr l(Oi2j in Eq.(40f) can only be realised by N
extra transconductors with different transconductance values. Only Eq.(40c,d) can be
realised by canonical form.
H ( S ) 2 2
^ 2 2
" ( s + p f ) ( s + p ^ ( s + p i ) ( s + pj)
Because the transfer function has even order denominator with even order numerator, a
canonical prototype ladder cannot be realised. Actually, the ladder derived from the
transfer function is shown in Fig.6, which has 5 nodes, so it requires fifth order matrices,
which in turn would lead to a tenth order transconductor-capacitor filter as shown in Fig.7.
If a canonical design approach is applied, the transfer function is changed to the form
^
H ( S ) 2 2 2 2 2
" (s + p )(s + pi)(s + p5)(s +p5)
18
4 4 9
and a modified ladder is obtained as in Fig.8. This is now a four node network, and using
any of the design equations (38c,d), (39c,d) or (40c,d), a canonical realisation can be
implemented. Fig.9 shows the canonical realisation circuit structure of Eq(39c,d).
Compared to the non-canonical realisation in Fig.7, two transconductors can be
eliminated.
The canonical filter design method above is called Transfer Function Modification
Approach, in which we focus on finding a canonical ladder prototype, and then using
matrix design method to obtain a canonical transconductor-capacitor filter. Because the
limitation of equal transconductance in transconductor-capacitor filter design, a large
number of designs cannot be realised in canonical form. To overcome the difficulty, an
alternative canonical design approach called Mixed Variable Representation Approach is
developed. Instead of finding a canonical ladder prototype, we try to find the best variable
representation in order to get a canonical transconductor-capacitor design from a non-
canonical ladder prototype. Generally a non-canonical ladder prototype contains more
nodes than a canonical ladder and this number also exceeds the number of poles being
realised. Hence when a voltage variable representation is used the size of matrix is
increased and this leads to more transconductors in a transconductor-capacitor filter
realisation. It is also noticed that the extra nodes are usually introduced by series L C or LR
branches. In a Mixed Variable Representation Approach, both voltage and current
variables are selected in a non-canonical prototype ladder to form vector V, and a compact
matrix form is obtained. Then the standard matrix design method is used to achieve a
canonical transconductor-capacitor implementation. The application of this approach and
the selection of the variables are very much dependent upon the structure of the non-
canonical ladder. To demonstrate the application of the approach, we redesign the 8th-order
Butterworth filter using the Mixed Variable Representation Approach. Examining the
circuit in Fig.6, we notice that there are five nodes in the prototype ladder, so using a nodal
voltage representation would requirefifthorder matrices in matrix design method, which in
turn would lead to a tenth order transconductor filter. Instead, we can construct the vector
V from two nodal voltages (Vi and V3), and the currents through the two L C series
branches (I2 and I4). We use voltage (VI2 and V14) proportional to the currents and the
terminating resistance, in order to preserve dimensional consistency. It is not necessary to
simulate the output voltage (V5) of the passive ladder explicitly, because all of I4 passes
through the termination resistor and V14 can therefore be treated as the output voltage. For
the mixed voltage-and-current representation of the eighth order Butterworth bandpass
ladder the matrices are:
19
4 5 0
fV. \ I 0
n
R
in.
f v
0 °]
1 I 0 -1 0
J= 0 ,v = ,G = (43a,b,c)
v 3
R 0 1 0 -1
0 v 0 -1 1,
0
(\_
0 0
0 0 0 A U
-1
0
-L
2
2
0 0 o 2
0
c = R ,r = R C 2
(43d,e)
0 0 c 0
3
o 0 0
0 0 0
V R 2
J -1
0 0 2
R C
47
The field of filter automation is now fairly mature, however it is still a challenging area
with the continuing development of VLSI technology. Most filter compilers developed
today are normally devoted to only one design strategy and technology and are usually
restricted to the design of cascade biquad or provide very limited ladder-based structures
( often leapfrog ). These restrictions limit the designer's ability to combine and compare
techniques and examine the benefits of different topologies. For these reasons, a new
active filter and equaliser compiler called XFILT[14] has been developed to implement
passive-RLC, active-RC, switched-capacitor and transconductor-capacitor filters and
equalisers and incorporates a wide variety of filter structures proposed for ladder-based
realisations together with more conventional biquad configurations.
The X F I L T system structure is shown in Fig. 12. A graphical interface utilises standard
Xll[15] and is provided both as a user friendly interface to the designer and as a system
manager of all the software in XFILT. It offers a menu-driven interface for checking and
facilitating the entry of design parameters. The specifications of the filter can be read in
from menus or existing files. The menu-driven interface can prompt the user for the input
needed and reject incorrect input. Graphical display is another characteristic of the
20
interface. The graphs of the approximation function and different circuit responses are
given. A special feature of the XFILT graphical interface is that it provides a graphical
editor, utilising this the user can specify/modify arbitrary, or classical magnitude and group
delay templates with a mouse.
The filter circuit design falls into two structural categories: cascade biquad and passive
ladder simulation. Ladder simulation requires a passive ladder prototype which can either
be synthesised internally or read from an external file. There are three ladder prototype
synthesis modes: expert for experienced designers whereby complete control is exercised
over the classes and sequence of pole/zero removals; interactive in which only the sequence
of removals is selected and automatic when no manual intervention is necessary. All
ladder-based filter structures are derived by matrix decomposition methods given in
Section 2. In the cascade biquad mode, a variety of designs are available, including type-
E, type-F and all-pass biquads. The structure of the software enables the addition of any
required biquadratic sections, active-RC, SC and transconductor-C, in an easy manner.
Analysis facilities for frequency response, group delay, and amplitude sensitivity are
provided. The embedded general simulator is SCNAP4[16], though SPICE and
SWITCAP can also be called.
21
then used to modify the template. The software automatically produces a new
approximation and circuit realisation. The main attraction of this approach is the low order
optimisation space, which always corresponds to the order of the original filter. The
process is quick and converges for realistic deviations due to practical circuit non-idealities.
For a lowpass video filter with specifications of passband edge frequency 1MHz, stopband
edge frequency 1.887MHz, passband ripple 0.28dB and stopband attenuation 50.5dB, a
5th-order elliptic ladder prototype will satisfy the requirements. Using the above matrix
decomposition method, a fully differential transconductor-C circuit realisation is given in
Fig.13, where all tranixonductors have the same value, and Fig.14 gives the SPICE
simulation result of the circuit. The transconductance spread is 1 and the capacitance spread
is 23.54. If a cascade of biquadratic stages is used, it is not possible to have equal valued
transconductances, but the alternative equal valued capacitance realisation demands a
6
transconductance spread of 7.3xl0 ! Fig. 15 shows the sensitivity comparison between a
ladder-based realisation and a cascade biquad configuration, and the obvious advantage of
ladder-based structure is clearly apparent.
The filter group delay variation within the passband is l.l|0,s, whereas the original
specifications require a maximum variation of 0.4|is. An initial 12th-order equaliser
design, which reduces the group delay variation to 0.13 |is with an equal ripple behaviour,
was considered. However with the help of XFILT compiler, the maximum group delay
variation can still be satisfied by utilising a 6th-order equaliser, though the equal ripple
behaviour no longer applies. The comparative sensitivity analysis for the ladder-based
equaliser and a typical cascaded biquad equaliser is shown in Fig. 16; the ladder-based
structure is again significantly better. A fully differential realisation of the 6th-order
equaliser is given in Fig. 17. The filter and equaliser were fabricated on a 1 micron CMOS
process, having double polysilicon and double metal and using a 5V power supply.
Fig. 18 shows the overall amplitude response of the filter and equaliser, the measured
results are typical ones from the range obtained . It will be noted from the circuit
response, that there is a loss of attenuation in the stopband including the disappearance of
the second notch, though the level of stopband attenuation remains within specification.
This retention of performance is due to deliberately allowing an initial filter approximation
with a very generous stopband attenuation to absorb process variations and inaccuracies,
and deterioration due to noise, though the noise level in the passband is typically
-128dBm. The passband response of the fabricated circuit shows some variation in ripple
level and corner frequency. Detailed investigation of the equaliser amplitude response
22
shows an arbitrary rippling behaviour over the passband of up to l.ldB, whereas the
computed amplitude response of the all-pass equaliser shows a completely flat
characteristic over a wide frequency range. These variations in equaliser response can be
directly attributed to the accuracy problems in realising the very wide spread in capacitance
2120:1 (transconductance spread is 1). It was also noted that the dynamic range of the
equaliser was considerably less than the filter. It is now possible to utilise various signal
scaling techniques at different stages of transconductance-C ladder-based designs and it
would be sensible to invoke these in any re-design. The group delay responses. Fig. 19,
show very effective equalisation within the passband, the initial computed group delay
variation of the filter is 1.02|as and the equalised value reduces to 0.395|J.s, the actual
response of the combined circuit shows a variation of only 0.1|is which easily satisfies the
initial specification.
Use of X F I L T has enabled further improvements to the equaliser. Simply increasing the
number of iterations in the group delay approximation stage improves the equal ripple
nature of the response at no extra cost in circuit terms. A more significant improvement in
silicon area can be achieved by utilising the arbitrary amplitude approximation facility for
the filter. The use of a 4th order touch point near the band edge reduces the group delay
variation of the filter to 0.72|is and thus lowers the equalisation requirements
considerably. Fig.20(a) gives the passband response of the modified filter and Fig.20(b)
shows the comparison of modified filter group delay and original filter group delay. A 4th-
order equaliser can now be used. There is a cost of about 4dB loss of attenuation in the
stopband of the filter.
6. CONCLUSION
Efficient ladder-based active filters and equalisers for active-RC, SC, and transconductor-
capacitor are presented. The canonical realisations of these filters are studied. Several
examples are given to demonstrate that the proposed canonical design method can be used
to considerable advantage. The XFILT software development has been based on the ladder
filter design methods given here, together with many cascade designs. A practical video
filter and equaliser IC implementation example is presented and the test results are given.
23
4 5 4
7. A C K N O W L E D G E M E N T S
Technical support in the implementation and testing of the video filter and equaliser by
Wolfson Microelectronics is gratefully acknowledged. This work was supported by the
Science and Engineering Research Council and the Department of Trade and Industry.
REFERENCES
24
[14] Lu Yue, R.K.Henderson, and J.I.Sewell,"XFILT: an X-window based modern filter
and equaliser design system," 11th European Conference on Circuit Theory and Design,
Davos, 1993, pp.305-310
[15].D.A.Young, X window system programming and applications with Xt, Prentice
Hall, 1989
[16].Z.Q.Shang and J.LSewell,"Efficient sensitivity analysis for large non-ideal switched-
capacitor networks," I E E E ISCAS, Chicago, May 1993, pp.1405-1407
25
° (a) (b)
Fig. 1(a) First order section using a conventional transconductor
Fig. 1(b) First order section using a low impedance input transconductor
tZ7
HI
/-e e -e \
— [
Vout
ZD
Fig.2 Transconductor-C ladder based group delay equaliser structure
20
o
20
40 Circuit Response
3J
60
80
100
120
1000 2000 3000 4000 5000
Frequency (Hz)
20
40
60 Circuit Response
80
100
Approximation
120
500 1000
i
1500 2000
Frequency (Hz)
8.00
ca
4.00
L a d d e r Based Realisation
on 2.00
0.00
680
Frequency (Hz)
tn tn t n \5 V5 J
\out
VI
2 V2 r v7l
Wl W2 W3 \V4 r w5
i
in
S L_f} LLff LM J_f}
m m ILTTTO rrrtxo. crrrm rrrin
rr j
5f f[[| l n]|| pi
C4
Q c 2 R
R
1
z g
Vout
VI V2 I V3 I V4 T
I Wl I W2 I W3 W4
X T
VI I V2 1 V3 I V4
Wl JH W2 W3 x- W4
I
I
-g /.+1
Vin
BQ
Frequency ( M H z )
SPECIFICATIONS
APPROXIMATION
< T
PROTOTYPE
DESIGN
BIQUAD F I L T E R
Ed DESIGN
LADDER FILTER
DESIGN
C/3
s
ON
F I L T E R N E T L I S T OUTPUT
SIMULATION
F R E Q U E N C Y RESPONSE, SENSITTVITY, NOISE
O
OPTIMISATION
as
V2
1
V3P
rc.o
03
SUA
dflA
I
o-o
CO
a*
3 = t?0 e
VHP in
AC*
1
CO?
<L>
d6*
00
c a z z
B 5 « y s
20
20
40
\
r
3
60
a.
80
-100-
-120-1 i 1 1 r—*==i • 1 - i •
0.0e+0 1.0e+6 2.0e+6 3.0e+6 4.0e+6 5.0e+6 6.0e+6
Frequency (Hz)
Fig. 14 SPICE simulation of transconductor-C video filter response
50
40
"P. 30
Cascade Based
20
Ladder Based
Z
1/3
10
C\J
CD
<V
00
£8
1>
oo
in
0)
09
ID
3S
oo
a!
CD
-a
a>
CO
(%/flP) A'liAiiisuas
4 6 3
L
VHP
VCP
VCt.
JL
053 «
°JZD
CCMD
Y2
V7
VBP
3
VCN
9
8*
G2M
fOA
u
1CD
CCMO
V?
va
VHP
VCN
ACNO
C'OJ
I
0N9V
i •rf3A
CM
£A
*** 60
ONSV IA
N3A
TP
iA
1
c
CA
0N3
« ! H 8
Hi
m m in
s? s
CO
J 10
CO
cu GO
0> 8"
0) CD
CD
41
CO s 00
cu
co
cu
CD
CVi
cci
6S
cu
CO
CD
00
CD
CM CO 00
CM o
(HP) a p m n d i u v
4 65
CD
+
cu
CM
4> CD
V)
CD
V)
4>
V)
VJ 00
(A
00
v>
1)
CD
eg a? CO
CD
01 co 1)
N
4>
CU
V)
60
CD
CU
V)
0)
V)
ON
CM
I
CO CM
0.0
-0.5
-1.0
E
<
-15 H
-2.0 T
0.09+0 4.0e+5 8.0e+5 1.29+6
Frequency (Hz)
Fig,20(a) Video filter passband response using a 4th order touch point
2.00e-6
173
1.00e-6 H
a
Modified Filter Delay
3
O
1=
0.00e+0
0.00e+0 1.00e+6 2.00e+6 3.00e+6
Frequency (Hz)
Fig. 1(b) First order section using a low impedance input transconductor
Fig.20(a) Video filter passband response using a 4th order touch point
26
468
PAPER 57
Accepted f o r l.E.E.E. International Symposium on Circuits and Systems, Seattle, May 1995
In this paper, an accurate semi-symbolic analysis The main contribution of this paper is the generation of
method based on polynomial interpolation techniques modified polynomials in a single variable z by
for large non-ideal switched linear networks (SC and interpolation, as required for analysis of large non-
SI) is presented. For the computation of frequency ideal switched linear networks. Ii is demonstrated that
responses, the semi-symbolic method demonstrates high accuracy can be manintained for large networks
significant improvements in speed over direct and the application to the computation of frequency
numerical methods and this increases with circuit size. responses is very attractive. The extension to noise
It is shown that semi-symbolic techniques can be used computation in switched linear networks is considered.
in the computation of noise for circuits containing a
limited number of clock waveform? but the
2. P O L Y N O M I A L I N T E R P O L A T I O N M E T H O D
performance rapidly deteriorates with network size.
Examples demonstrate the comparative speed and A large non-ideal switched linear network can be
accuracy performance over a range of circiuts. described by a system of equations the in z domain [21]
1. I N T R O D U C T I O N
1 p •v,(=)- IW,(z) "
- l "
The symbolic analysis of linear analogue circuits has
V (z)
2 ZW (z)
2
attracted the attention of many researchers over a P
2 1
problem, the escalation of computer time and memory Ck is the capacitance matrix of the kth time-slot and p k
requirements with increase in circuit size. Two typical is the extended state transition matrix. The r.h. side
solutions proposed are expression approximation! 16] contains polynomial approximations to the excitation
and hierarchial decomposition[17],some improvements vectors and appropriate sampling weighting factors, the
have resulted. However, despite all the effort, these vector V contains node voltages and branch currents.
methods still experience great difficulty with medium Since the pk malices and components of the r.h. side
to large networks. Whilst attention has been focussed vector can only be assembled by numerical techniques,
on continuous-time analogue circuits, some work has closed form symbolic solutions for the system
ben carried out on the symbolic analysis of ideal SC responses are impossible to obtain. However, a
networks! 18] and only exploratory investigations have polynomial interpolation scheme provides a natural
been reported into the semi-symbolic analysis of non- solution, this can be summarised as follows:
ideal SC networks[19] and noise performance [20].
for(i=0; i<P; i++) ( /* P is polynomial degree */
The only realistic scheme for large networks of any solve T(Zi)V = W ;
type is a semi-symbolic one, when polynomials in s or
calculate D(Zi) = detT(Zi);
z or both, with purely numeric coefficients, are
N(z ) = D ( z ) V :
i i
generated. It is well known that generation of
polynomials by interpolation is a very efficient 1
technique. However, the generation of polynomials in DFT (FFT) to generate the numerator polynomial:
mixed variables (s,z) is really the basic requirement for DFT(FFT)to generate the denominator polynomial:
the analysis of non-ideal switched networks (SC or SI).
For a system like (1). the denominator of the transfer I P. H,,(z)
function can be determined bv (4)
-P : H (:)
J : LB:
D(.- .) = det(z .I- E) where E = P P _
J ( P,
1 M w r
Then
For continuous-time systems, the above procedures are
straightforward, in the switched network case some
care has to be taken. This can demonstrated in a H 2 1
H , 2 J J<*>:
simple 2-phase clock system when: The error resulting from evaluation of H t interpolated
ja>o { polynomials is given in F i g . l . graph (b) and shows a
I -P, H,(z)
(2) great improvement in accuracy. Table 1 shows this
zl H,(2) technique applied to a wide range of SC circuits and
B~<-
several deductions can be made. Potentially all
solving the matrix equation (2) gives (M is the number of time slots) need to be interpolated,
P,B f J(0O2 though some are zero in the majority of circuits and
2
The numerator and denominator polynomials are then efficiency gains decrease because of the increased
generated by interpolation using a DFT(FFT). When number of transfer functions. The polynomial
the frequency response is evaluated from the resultant interpolation methods can preserve acuracy even in the
polynomials and compared with directly computed analysis of quite large circuits.
results, the error is quite considerable, even for small
circuits. For a small SC treble tone control filter, the 3. SYMBOLIC M E T H O D FOR NOISE
error reponse for one of the transfer functions is shown EVALUATION
in Fig. 1, graph (a). The reason for this is that
3 -s • r 12 By utilising transpose techniques, all transfer functions
from each noise source to the output can be obtained by
solving an adjoint system. The problem can be
formulated as
<t> = d ' T ' W ,;
fractional powers of z are intoduced by the time-slot where D are sin(x)/x factors.
k
weighting factors and if the polynomials are re-defined After two adjoint system solutions at one sampling
to account for this, very high orders result and low frequency, noise transfer functions can be interpolated
accuracy follows. There is a simple solution, again from:
consider a 2-phase system and separately excite each X
l _ X
l l X
I 2 ' U
'
time-slot:
V 2y X
*-21
D-
I - P i " "H„(z)-
(3) The total noise power spectral density at the output is
.H,,(z) 0 calculated by superposition
P m 10-
S (tt>) = ' S; (co — n c o )
T
I I <D, s
,=0 i =o'
where Sj(£2) is the ith input noi^e source power
spectral density and p is the number of bands to be
considered. Two circuits were tested and the run time
statistics are illustrated in Table I I . These show that for
switched linear networks, the symbolic method for
noise evaluation is less efficient than its numerical
counterpart and an examination of the relative
theoretical computational costs w i l l provide an
explanation for this.
4. T H E O R E T I C A L COSTS OF EXPRESSION
PxMxM
E V A L U A T I O N AND M A T R I X SOLUTION
2
Fig.2b P M dependence of the
Two basic definitions are utilised. A flop (floating expression evaluation
point operation) is defined to be the time required for a
particular computer system to execute the C code theoretical factors when comparing the computational
A+=C*D efficiency of numerical and symbolic approaches.
Similarly, a trop (trigonometric function operation) is Generally, a > P which means expression evaluation is
the time required for execute another C code faster than numerical matrix solution. For instance
A+= B*cos(OD) with the treble tone control filter, these two factors can
In SCNAP4, the last time-slot matrix solution takes be calculated as follows
2
2
about 5 N flops, where N is the matrix dimension. The a = ( 2 x 2 + 3 ) x 8 =448
2
solutions of all other slots require another 2(M - 1)N 2
P = 9 x ( 2 x 4 x 2 ) = 288
flops, this makes a total of (2M + 3 ) N flops per 2 The ratio a : P = 0.09:0.06 which are the values in the
2 tf example entry in Table I . Since P increases
frequency point. Fig 2a shows clearly the (2M + 3 ) N
quadratically with respect to the number of time-slots,
dependence of matrix solution approach. For symbolic
2 expression evaluation for multirate circuits is expected
expression evaluation, the network has M transfer
2 to be less efficient than for 2-phase circuits.
functions to be calculated, it requires about 2 P M
2
trops where P is the polynomial degree. The P M
The total cost situation for noise analysis by expression
dependence of the expression evaluation is shown in evaluation becomes more significant. I f a circuit has
Fig. 2b m noise sources, the total cost is m/3. Now, in large
non-ideal switched linear networks where there are
10' many switches and other active devices, m is likely to
be very large, hence the application of symbolic
1 . techniques to noise analysis is not very feasible. This
10
trend can be seen in Table I I , where limited noise
evaluation at only 100 frequency points compares
10 1
poorly with full SCNAP4 noise analysis [22].
5. CONCLUSIONS
10"
An application of a semi-symbolic analysis method to
large non-ideal switched-linear networks is presented.
10"
z 3 4 3
Accuracy and speed are compared with an equivalent
10 10 10 10 10° numerical simulator. Noise calculation utilising
(2M+3)xNxN symbolic expression evaluation approach is proposed,
2
Fig. 2a (2M + 3 ) N dependence of but it is shown to be of limited efficiency.
matrix solution approach ACKNOWLEDGEMENT
2 2
Let a = (2M + 3)N , P = Y(2PM ),where y is a
Financial support by the British Council and the State
computer system related constant. On a SUN-Sparc
Education Commission of PR China is gratefully
ELC station, one trop takes about nine times longer
acknowledged.
than one flop, therefore y = 9. a and (5 are two useful
REFERENCES numerical and symbolic parameters by using hybrid
decomposition method". Proc. I E E E ISCAS. Singapore,
[I] G. Gielen, H. Walscharts. W. Sansen. "ISAAC: a Jun. 1991, pp. 794-797
symbolic simulator for analog integrated circuit/'. IEEE [13] G. Gielen, P. Wanibacq. W. Sansen. " Symbolic
JSSC, vol. SC-24, no. 6. pp. 1587-1597. Dec. 1989 approximation strategies and the symbolic analysis of
(2] S. Seda, M. Degrauwe, W. Fichtner, "A symbolic analysis large and nonlinear circuits", Proc. I E E E ISCAS.
tool for analog circuit design automation", Proc. 1EF.F. Singapore, Jim. 1991. pp. 806-809
ICCAD. 1988. pp.488491 [I4[ S. -M. Chang. J. F. Mackay and G. M. Wierzbr. "Matrix
|3] A. Libertore, S. Manetti, "SAPEC - a personal computer reduction and numerical approximation during
program for the symbolic analysis of electronic circuits", computation techniques for symbolic analysis". Proc.
Proc. I E E E ISCAS. Helsinki, May 1988, pp. 897-900 I E E E ISCAS, San Diego. May 1992, pp. 1153-1156
[4] A. Konczykowska, M. Bon, "Automated design software [15] G. Gielen, P. Wambacq, W. M. Sansen," Symbolic
for switched-capacitor IC's with symbolic simulator analysis Methods and applications for analog circuits: A
S C Y M B A L " , Proc. DAC, Anaheim, 1988, pp.362-368 tutorial overview", Proc. I E E E , vol. 82, no. 2, pp. 287-
[5] S. -D. Shieu, S. -P. Chan, "Topological formulation of 304, Feb. 1994
symbolic network functions and sensitivity analysis of [16] F. V. Fernandez, P. Wambacq, G . Gielen, A. R.-
active networks", I E E E Trans. CAS, vol. CAS-21, no 1 Vasquez, W. M. Sansen, "Symbolic analysis of large
pp. 39-45, Jan. 1974 analog integrated circuits by approximation during
[6] G . E . Alderson, P. M. Lin, "Computer generation of expression generation", Proc. I E E E ISCAS, London.
symbolic network functions - new theory and Jun. 1994, pp. 1.25-1.28
implementation", I E E E Trans. CT, vol. CT-20, no.l, pp. [17] S. Jou, M. Perng, C . Su and C K. Wang, "Hierarchical
48-56, Jan. 1973 techniques for symbolic analysis of large electronic
[7] P. Sannuti, N. N. Puri, "Symbolic network analysis - an circuits", Proc. I E E E ISCAS. London, Jun. 1994. pp.
algebraic formulation", I E E E Trans. CAS, vol. CAS-27. 1.21- 1.24
no. 8, pp. 679-687, Aug. 1980 [18] D.G.Johnson and J.I.Sewell,"Improved z plane
[8J J . K . Fidler, J . I. Sewell, "Symbolic analysis for polynomial interpolative analysis of switched capacitor
computer-aided circuits design - the interpolative networks'" I E E E Trans.CAS. vol.CAS-31, no.7. pp.666-
approach", I E E E Trans. C T , vol. CT-20. no. 6, pp. 728- 668, 1984.
741, Nov. 1973 [19] C.K.Pun and J.I.Sewell, "Symbolic analysis of ideal and
[9] K. Singhal, J. Vlach, "Symbolic analysis of analog and non-ideal switched capacitor networks". Proc.IEEE
digital circuit", I E E E Trans. CAS, vol. CAS-24. no. 11, ISCAS, Kyoto, June 1985, pp. 1165-1168
pp. 598-609, Nov. 1977 [20] C.K.Pun, A.G.HalI and J.I.Sewell, "Noise analysis of
110) M. M. Hassoun, P. M. Lin, " A new network approach to switched capacitor netorks in symbolic form", Proc. 29th
symbolic simulation of large-scale networks", Proc. Midwest Symp., Lincoln.Na., August 1986, pp. 807-810
I E E E ISCAS, Portland, May 1989, pp. 806-809 [21] Z. Q. Shang, J. I. Sewell, "Efficient sensitivity analysis
[ I I ] M. M. Hassoun, " Hierarchical symbolic analysis of for large non-ideal switched capacitor networks", Pioc.
large-scale systems using a Mason's signal flow graph I E E E ISCAS, Chicago, May, 1993, pp. 1405-1407
model", Proc. I E E E ISCAS, Singapore, Jun.1991, pp. [22] Z. Q. Shang, J. I. Sewell, "Efficient noise analysis
802-805 methods for large non-ideal SC and SI circuits" Proc.
[12] T. Matsumoto, N. Nakayama, K. Tsuji, "Symbolic I E E E ISCAS, London, Jun., 1994, pp. 5.565-5.568
analysis for large networks and systems including both
circuit circuit No.of polyn. sample DFT expr. cal. SCNAP4 peak err
name size slots degree (sec)^ (sec) (100 pts) (100 pts) (dB)
spft 112 24 20 241.14 16.42 23.61 89.06 le-8
lplO 60 16 15 24.26 4.46 7.25 17.99 le-9
nosl 1 48 4 15 1.17 0.33 0.59 3.47 le-5
b 8
P 43 2 11 0.20 0.06 0.15 1.70 le-9
bp6 36 2 10 0.15 0.04 0.06 1.18 le-6
nos5 24 4 8 0.27 0.12 0.31 0.88 le-11
bp2 15 2 5 0.05 0.03 0.05 0.23 le-10
tf 8 2 4 0.03 0.01 0.06 0.09 le-14
int 5 2 2 <0.01 0.01 0.04 0.07 le-14
PAPER 58
474
Left-UL Decomposition
If A is decomposed as A = U L . the system design
equations are
(a)Typical circuit (b) Block symbol
UW = (i|)B+D)I-(l+z-').l (10a)
Fig.l SI Filter First Order Building Block
LI=-<)>W (10b)
In practice, any enhanced type memory cell can be
employed to form a first order building block, such as Left-IA Decomposition
cascode, regulated cascode, folded cascode, or class AB For the case A = I A , the design equations are
U
(a)
Ml 1
H (b)
1
decomposition of a typical 6th-order matrix system is Fig.3 Left Decomposition Input Factor SI Realisation
shown in Fig.2.
2.2. R I G H T M A T R I X DECOMPOSITION
APPROACHES
- *A MX = J (23)
can be utilised.
Fig.5 6th-order Right Decomposition SI Structure In the dynamic range scaling phase, a frequency sweep
analysis of the Filter passband is performed, to determine
Four right decompositi ns follow: the maximum signal levels of the internal currents.
Scaling can be done by multiplying column j of M by the
Righl-LU Decomposition maximum value of the variable X j for all variables in X.
If B is decomposed as B=LU, the design equations are This is equivalent to creating a new variable X j / X j a x -
m
The input current J1 in Fig.5 is design results are shown in Table I, where the Left-UL
l + z~ is seen to be the most efficient realisation. By trying the
J1 = J, (20) all design methods, we can achieve, about 39.04%
1-z"
This can be realised by any bilinear SI integrator. Fig.5 reduction in g spread and 38.73% reduction in g
m m
is (he realisation structure based on the basic building sum. The simulated Left-UL circuit response is shown
block in Fig. 1. in Fig.6 and Fig.7 shows the circuit schematic with the
biasing circuits omitted. To compare the sensitivity
performance, the total multiparameter sensitivity
simulation over all gms has been carried out. The
sensitivity simulation results of a comparative biquad
realisation, having a maximum g spread of 63.78, are
m
5. CONCLUSION
0.OO«*O 5.00«*5 | ;o**6 I 50..6 2.00«*6
Frcqucocf (Hi)
A new approach to realise exact ladder based SI filters
has been presented. The bilinear transformation is used
in the design procedures. Eight different SI ladder based Fig.6 10th-order bandpass SI filler response
structures can be obtained from one prototype ladder.
Therefore it provides SI filter designers with a choice of
circuit realisations based on different requirements such
as power, area, maximum transistor aspect ratios,
sensitivity or noise performance. Techniques to improve ES T
_, ? s ea I a lis a
ACKNOWLEDGEMENTS 1
PAPER 5 9
Accepted for l.E.E.E. International Symposium on Circuits and Systems, Seattle, May 1995
M U L T I R A T E S C AND SI F I L T E R S Y S T E M D E S I G N
BY X F I L T
Lu Yue and J.I.Sewell
— •- Department of ElecffomcsancTElectric^'Engineefirig
University of Glasgow
Glasgow G12 8LT, U.K
Tel: +44-141-3304253, Fax: +44-141-330 4907
EMail: sewell@elec.glasgow.ac.uk,luyue@elec.glasgow.ac.uk
ABSTRACT
following the anti-aliasing filter becomes a decimaior implements a sampling rate increase from F to L F . s S
d / l _1 2 N
(1) cascade, a new design methodology is proposed
' l + a,z +a z~ +...+a z~ 2 N
consisting of two FIR filters with the higher sampling
frequency and one IIR filter with the lower sampling
where the unit delay period corresponds to the high frequency whose transfer function is
sampling period 1/MF in decimator case or 1/LFs in S
the interpolator case. A well known modification of the HF(Z) = HDIow(z)Hsc(z)Hllow(z) (13)
original z-transfer function (1) leads to [6].
Correction of sin(x)/x effects is generally required in
N(M-I)
multirate designs and this is readily facilitated by
1
(bo + ^ z " + b z - + . . . + b z - ) 2
2
K
K c
X i "' z
XFILT within one section .
:
H d / i (z)
l + d,z- 1 M
+ d,z- 2 M
+ ... d z' + v
= K
"' 4. M U L T i R A T E F I L T E R DESIGN
(2) EXAMPLE
Hd/i(z) can be decomposed into two parts. Decompose
4.1 SI BANDPASS FILTER DESIGN
Hd/i(z) as:
Hd/i(z) = Hi(z)H2(z) (3) A simple example will illustrate the improvements that
where can be obtained in SI filter design. A video frequency
N(M-n bandpass filter with a passband from 4.25MHz to
z K
H , ( z ) = ( b + b , z ' ' + b z ~ + . . . + b z ~ ) J^z
0 2 K 6.8MHz, stopbands from 0.1 Hz to 3.75MHz and from
i=0 7.25MHz to 60MHz. passband ripple 2dB and stopband
(4) attenuation 30dB, can be realised by an 8th-order
and elliptic filter using a cascade of four SI biquads with a
single sampling frequency of 72MHz. Two critical
H (z) = 2
parameters in the circuit realisation of SI filters are gm
1 M 2 M N M
Ud z- 1 + d z- 2 + ... d z- + N spread (deriving from the spread of transistor aspect
(5) ratios) and g sum. A small g spread will maintain
m m
H((z) is of FIR transfer function form with sampling lead to a lower power consumption. The design results
frequency M F and can be realised by a FIR decimator
C are given in Table 1. To reduce the spread of transistor
structure. H2(z) is of the IIR transfer function form aspect ratio and gm sum, a multirate multistage
with sampling frequency F and can be realised by an s
structure is adopted as shown in Fig.7, where the
HR decimator. lowpass filter has a 72MHz sampling frequency and the
highpass filter has a sampling frequency of 36MHz.
The transfer function for the whole sampled data system The lowpass filter also has an anti-aliasing function up
is to 67MHz. The lowpass filter is realised by cascaded
H(z) = H D ( Z ) H S D W H I ( Z ) (9) biquads and has a g sum of 547.81 units and a g
m m
Z r
HSD( ) f ° middle filter transfer function and H](z) for of 19.79. The overall multirate SI system has a gm
the interpolator transfer function. For the decimator and sum of 985.23 and a g spread of 26.19. The
m
interpolator, the transfer function can also be expressed multirate design technique has reduced the g sum by m
4 81
Input Decimator
SC Filter
Interpolator
L=4
(b)
M=4
Fig.10 Frequency response of baseband filter
fs:320kHz fs:80kHl fs:.U0kHz
4.3 SINGLE-PATH FREQUENCY-TRANSLATED
Fig.9 Baseband bandpass filter system SWTTCHED-CAPACTTOR BANDPASS FILTER SYSTEM
Table II shows the comparison of three different design The single-path frequency-translated (SPFT) approach is
results. Design 1 is a traditional design using a single a paricularly effective solution to the very narrow band
filler problem. The basic design performance 5.CONCLUSION
parameters for a typical 80Hz bandwidth filter at 20
kHz are quoted from [1] and shown in Table 3. where
Design 1 is a single rate realisation, and Design 2 a In this paper, some strategies and algorithms for
multirate implementation. The parameters of the multirate SC and SJ_ filter design by XF1LT have been
scheme produced by XFTLT are given as Design J."and presented. Several design examples demonstrate the
the overall system is shown in Fig.l 1. utility of the multirate X F I L T in producing improved
solutions to difficult filter problems.
[npul DtdDBtcr lotopcietor
Flltr M=12 Pltir
SC Filer
ft:U&Hi 16UU'
ACKNOWLEDGEMENTS
Fig.l 1 SPFT Bandpass Filter System
The first and last filters are simple 2nd order bandpass Financial support by Science and Engineering Research
sections, the decimator and interpolator both have Council, UK and Department of Trade and Industry. UK
bandpass characteristics and the centre filter is ladder is gratefully acknowledged.
derived. Only 4 clock waveforms are used and the
whole structure is strays-free.
REFERENCES
0
k
SO 1985
[5] Lu Yue, R.K.Henderson, J.I.Sewell, " X F I L T : An
X-window Based Modern Filter and Equaliser Design
s -100 System", European Conference on Circuit Theory and
ISO
0 50000 100000
Frequency (Hit
i
t50000
Design, Davos, Switzerland, Aug. 1993, pp.305-310.
[6] R.Crochiere and L.Rabiner,"Multirate digital signal
processing", Prentice-Hall, Inc., Englewood Cliffs,
New Jersey, 1983
(b)
Fig. 12 SPFT system response