signal processing

Interactive direct-coupled filter design

This technique makes it easy to exploit several degrees of freedom, while also assuring exact filter response over any bandwidth and wide ranges of positive elements.

passband width, within desired bounds. Filter configurations can be By Thomas R. Cuthbert changed simply by clicking on worksheet buttons that activate preprogrammed macro procedures. The resulting filter configuration is andpass filter design has been immediately displayed by the incorpobased on readily available lowpass rated drawings, ladder network proand embedded totypes for more graphs and other than 50 years [1]. aids are easily Direct-coupled filadded for enters are probably the hanced visibility most important type of design relationof bandpass filter, ships. Another consisting of a casbutton enables cade of resonators choice of either coupled by reactive singly or doubly components. terminated filter The LC resonators transfer funcare usually parallel tions. (nodal) or series This article (mesh) resonators, describes a comeach nominally pletely general tuned to the band transition from center frequency. lowpass protoDesign of microwave types to directcavity resonators is coupled filter commonly related networks. It also to the lumped LC illustrates the resonator model. possible combiAssumption of a narnations of courow passband width plings between has been essential in resonators and relating direct-coudiscusses the pled bandpass filters common layout of to lowpass prototype Figure 1. Evolution from (a) lowpass to (b) classical bandpass to (c) direct-coupled network. the nine worknetworks, and the sheets. The general procedure for stray capacitance or to limit resintrinsic passband distortion has either using Excel workbook file UCHOOSE onator voltages. been ignored or accounted for by furis presented. This article describes nine workther approximations [2]. sheets in one Excel file that simplify A new and useful method has direct-coupled filter design using two recently been described for designing Lowpass and bandpass networks to seven resonators in 210 nodal conall-pole direct-coupled filters that are A typical fourth-degree (N=4) lowfigurations. Another 210 equivalent distortion-free, no matter how wide pass prototype network is shown in mesh (series) resonator filters are the passband. Furthermore, wideFigure 1(a). The load resistance is easily obtained as dual networks deranging choices of positive element unity and element numbering prorived from the normalized nodal devalues are determined by arbitrary ceeds from the load toward the source. signs [4: page 34]. Modern spreadparallel resistances that do not affect Its passband extends from 0 to 1 sheets include point-and-click the filter response [3]. This method of rad/s. The response is said to be allconstrained optimizers that can obchoices has now been generalized to pole when c 3=0; otherwise a trap is tain reasonable objectives by autoinclude elliptic-function filters [4]. formed by g3 and c3 to produce a stopmatically adjusting the parallel resisFor N resonators, at least N2 band null (infinite loss). The classical tances and other parameters, such as degrees of freedom are available to bandpass topology in Figure 1(b) is achieve various objectives, such as minimizing element spread while constraining, eliminating, or standardizing elements. The N2 parallel resistance values are easily adjusted to obtain positive elements and can be upper-bounded to minimize effects of

24

www.rfdesign.com

September 2000

derived by replacing all lowpass C s with parallel LCs and all lowpass Ls with series LCs, resulting in alternating nodal and mesh resonators, all tuned to 0=1 rad/s. The bandpass frequency scaling factor is: QBW =

the AC branch in Figure 1(b). If c 3 =0 the latter branch is like any other series branch. Otherwise, the AC branch In Figure 1(b) produces two null responses at geometrically symmetric stopband frequencies:

0 , 0 = b! a b a

(1)

- = 2 + 1 , +
1 (2QBW gk!ck )

1 , -

Branch A-C in Figure 1(b) can be transformed into several equivalent topologies that are also sel0 dom practical [5: pages QBW = , 0 = b! a 5245]. Except for lowb a impedance systems, series (2) resonant branches are often upset by parasitic In the kth parallel LC, C= capacitance to ground, Qk and L =1/Qk so that their especially for narrow passanti-resonance is at 0 =1 band widths (large QBW). rad/s. Conversely, in the kth Figure 2. An exact bandpass replication of a lowpass elliptic filter response. series LC, L= Q k and C The equivalent directcoupled bandpass filter in =1/Q k so that their resoFigure 1(c) is obtained by replacing all nance is at 0=1 rad/s. These singly power in those resonant branches. series branches in Figure 1(b) by subloaded Q values are the meaningful Particularly note the transformation networks that do not disturb the filter ratios of reactive to real (transmitted) of the A'C' branch in Figure 1(a) to

where the geometrically symmetric passband edge frequencies are b and a. The bandpass prototype elements are then defined by:

(3)

Figure 3. A universal exact replacement for a series elliptic-function filter branch.

26

www.rfdesign.com

September 2000

response but can change nodal impedance levels. The impractical topology between nodes AC in Figure 1(b) is altered to the equivalent, but practical, topology in Figure 1(c) by adding a parallel LC resonator at node B, coupled by two parallel LC traps in series. The traps are anti-resonant at the two stopband frequencies in (3); note that as c k0, and -0 and +.

Thus, in Figure 1(c), ck=0 corresponds to Lbc and Cab=0. That reduces the branches between nodes AB and BC to the same form as those between nodes DE and EF, respectively. (Node C is redundantly labeled D for later convenience.) Comparison of parallel branches BG and EG, where G is the ground terminal, will be made in the next section. The networks in

Figures 1(b) and 1(c) both produce the response shown in Figure 2. Bandpass geometric frequency symmetry is observed when the abscissa is scaled logarithmically.

Exact series branch transformation

The replacement of each series branch in Figure 1(b) is accomplished by the universal subnetwork shown in Figure 3. Resonator B and the two traps have been mentioned. Trap AB is assumed anti-resonant at + and trap BC is assumed anti-resonant at - ; their topological names will be LC and CL, respectively. Pairs {CT, LT} and {CV, LV} are combined (absorbed) into parallel resonators A and C in Figure 1(c). The equations in Figure 3 that define the elements shown are in terms of the loaded Q values from (2) and the parallel resistance parameters denoted in Figure 1(c) (parallel resistances are the reciprocal conductances at 0=1 rad/s conventionally associated with inverter transformations [4:50]). In Figure 3, equation QB=LLPQBW refers to LLP=g3 in Figure 1(a), and the lower stopband null defined by (3) appears throughout. Three inductances are shown as inverses to facilitate parallel combination with others, as well as for element elimination (L-1=0). There is no end resonator at node F in Figure 1(c) to absorb the {CV,LV} pair in Figure 3, so two constraints are required to eliminate them. Inspection of the equation for CV in Figure 3 shows that C V=0 when -=0 and R b=Q B2R c. Then LV-1=0, and additionally Cb=0, the latter explaining why there is no capacitor from node E to ground terminal G in Figure 1(c). These two constraints are necessary at one or both ends of the filter when the lowpass prototype, Figure 1(a), has a series branch at one or both ends. These occurrences produce external end couplings, like branch EF in Figure 1(c). The universal subnetwork in Figure 3 can be turned end-for-end by interchanging R a and R c in the included equations. The resulting equations are similar, so programming the branch sequences is all that is required.

Wideband directcoupled topologies

It has been reported that there are nine possible wideband nodal topologies using N=2 to N=7 resonators

28

www.rfdesign.com

September 2000

is the set {LC, CL, L, C}. Examination of the N=4 topology in Figures 1 and 4 shows that external (end) top coupling results from a series lowpass prototype branch connected at that end. That is denoted in Figure 4 by the lightly shaded pair of convex and concave dashed arcs that involve three branches. As in branches EF and EG in Figure 1(c), the external (end) top-coupling element and the adjacent parallel branch must be L and C or C and L, respectively (not C and C, for instance, and not LC). Internal top couplings in Figure 4 are denoted by single concave arcs that involve only two branches. Those pairs may be {LC, CL} or {CL, LC}, or {L, C} or {C, L}.

Worksheet layout
The features of a standard format for each of the nine Excel worksheets are described in this section; the N=4 case shown in Figure 5 is typical. The relevant lowpass (LP) prototype network and the direct-coupled bandpass (BP) network it yields are shown. All eight possible coupling sets are listed and sorted in an order favoring capacitive end couplings. A coupling set is selected by clicking on a numbered yellow button, and those coupling symbols automatically appear in the BP topology. Set 4 would be selected in Figure 5 to correspond to the filter in Figure 1(c), but set 3 is shown for the following Example 1. Clicking on the yellow choose source type button in Figure 5 invokes a dialog box for selection of either a lossy source (finite RS to define a doubly terminated filter) or a lossless source (RS eliminated to define a singly terminated filter). When the doubly terminated case is chosen, input dialog boxes solicit the Final R S (unnormalized) source resistance and prototype source resistance gN+1 (g5 in this case). These two values then appear in cells B8 and B14, respectively, and input

Figure 4. The nine possible nodal direct-coupled filter topologies.

N=2 2

N=3A N=3B N=4 N=5A N=5B N=6 N=7A N=7B 4 4 8 16 16 32 64 64

LC trap appears inductive and a CL trap appears capacitive). Table 1 shows the number of coupling combinations for the topologies in Figure 4. For example, the coupling combination for N=4 resonators in Figure 1(c)

Table 1. The number of coupling combinations for the topologies in Figure 4.

[3][4]. These are shown in Figure 4, where the /\ symbol indicates an L, C, LC, or CL coupling branch. Figure 3 shows that there must always be an equal number of inductive and capacitive top couplings (at 0=1 rad/s, an

30

www.rfdesign.com

September 2000

resistance R 4 is computed and displayed in cell B19. When the singly terminated case is chosen, other input dialog boxes solicit the input resistance corresponding to the 0 dB passband reference [4:16]. That value is normalized to the Final R L and displayed as input resistance R4 in cell B19, cell B14 is blanked, and cell B8 displays eliminated as shown in Figure 5. Input resistance R4 is the normalized impedance when 0=1 rad/s (normalized passband center frequency). The user must enter the input data, which include %BW, f0 (MHz), final (unnormalized) load resistance, and the g and c values. Those cells in column B are shaded yellow. Parallel resistance values at 0=1 rad/s are shown in cells B16 through B19; R1 cannot be varied because it is constrained by the end coupling, and R4 is constrained by the source. Initial values of the two free parallel resistances must be established, but the user must enter the log10 of these resistances in cells C 17 and C 18 (also shaded light yellow). By dealing with

variables in dB (except for a constant), the variables are usually in the range 0.5 to 1; however, variables going negative can never produce negative parallel resistances (R2=10^C17, for instance). That artifice and normalization enhance scaling for optimization. The resultant data consists of design parameters, normalized L 1 and C values, the summed squares of any negative normalized values (shown in red), unnormalized L and C values, the max/min ratios of Cs and Ls and the worst of the two, the estimated midband loss (L 0 in dB) f o r r e s o n a t o r u n l o a d e d Q u= 1 0 0 [6:341], and the relative node-voltage phases at the passband center frequency (for tuning). If c3 were not zero, - for the traps adjoining node 3 would be displayed according to (3). The user can obtain immediate results for wide-ranging choices, especially when the point-and-click constrained solver optimizer under the tools menu controls the parallel resistances.

Example 1: Singly-terminated all-pole filter

This and the following example illustrate results obtainable with UCHOOSE. This example is the N=4 case shown in Figure 5; topology set 3 is an all-pole filter, and the singly-terminated 0.2 dB Chebyshev gi values have been published [6:109]. A 60% passband width is specified with geometric center at 100 MHz, and resulting normalized and unnormalized bandpass edge frequencies are shown in cells D5D8. The button covering cells A7B7 was used to select a lossless source and specify an unnormalized input resistance of 200 . The unnormalized 50 load then requires a normalized input resistance R4=4 at 0=1 rad/s as automatically entered in cell B19. It is important to note that this example will produce the response in Figure 6 no matter what values are entered into variables cells C17 and C18, and that changing the %BW in cell B4 only scales the response versus frequency. Thus the user is free to choose these two or three variables in

32

www.rfdesign.com

September 2000

Figure 5. The N=4 spreadsheet for the singly-terminated filter in example 1.

Figure 6. The N=4 singly terminated all-pole response for example 1.

pursuit of various objectives. The first step is to obtain log variables in cells C17 and C18 that produce all positive element values; such values are guaranteed to exist [4:72]. When variables cells C17 and C18 are set to 0.0, (R2=R3=1.0), all elements happen to be positive, and cell B40 shows a sum of squares value of zero. With this lucky guess for parallel resistances (but see Example 2), the worst ratio in cell D39 was 16.1532 for the inductors. The second step is to start solver (in the tools menu) with all elements positive, with the objective of minimizing target cell D39 by changing cells C17 and C18. The options in solver should be set for automatic scaling, quadratic estimates, central differences and Newton search. Starting from 0.0 in cells C17 and C18, the uncon-

strained worst element ratio was reduced from 16.1532 to 6.8207 with log R 2 =0.0001 and log R 3=0.3986 (there are multiple solutions), and all elements remain positive. If some elements had become negative, solver constraint B40=0 could be added to maintain positive elements (seldom required). Nonlinear optimization results sometimes vary, depending on starting values of the variables; a little experimenting is good practice. Also, optimization in this unconstrained step is underdetermined. Many nonunique solutions will be obtained, depending on starting values. The third step is to add N-2 = 2 constraints and continue optimization from any solution to the preceding unconstrained optimization. Figure 5 shows the unique results of requiring L2=L3= (by applying constraints B24=0 and B26=0 in solver). Figure 5 shows that the inductor ratio is only 6.7182, but L 34 is large and C 23 and C 3 are small (unnormalized reactance in ohms at f0 is available by dividing any value in cells B21B36 into the final RL; any values less than 0.1 will require more than 500 reactance in this case). Alternative constraints L12=L34 and L2=L4 also produce a solution, but trying to force L2=L4= will not. Any solution produces the input resistance behavior shown in Figure 6. The constant current source delivers power into the lossless network proportional to that resistance. The dB loss with respect to the peak resistance (209.4257 corresponding to 0.2 dB ripple above the 200 at 100 MHz is shown on the secondary ordinate in Figure 6.

Example 2. Doubly terminated elliptic filter

This example shows how lowpass prototype gi values are obtained given unnormalized stopband attenuation requirements, how to obtain starting variable values for positive elements, and how to predict effects of uniform resonator loss on a lossless filter. A five-resonator, doubly terminated elliptic filter centered at 100 MHz is required to provide about 50 dB attenuation above 162 MHz, and each res-

34

www.rfdesign.com

September 2000

Figure 7. The N=5A spreadsheet for the doubly terminated elliptic filter in example 2.

onator will have the same finite unloaded Q. In this case, a design passband ripple of about 0.01 dB is suitable. Tables of elliptic filter values for 5% ripple are available [5:210211]; the inverse of (3) maps any normalized bandpass frequency, , to its corresponding lowpass frequency, : 1 = QBW (4) The normalized stopband frequency in this case is =162/100=1.62, corresponding to =2.005 rad/s. The elliptic filter tables show that at S=2.0, the stopband attenuation shelf is Amin=49.22 dB. The corresponding gi and ci element values are shown in Figure 7 (odd N elliptic-filter data are normalized to 1 load resistance; even N data must be renormalized to 1 ). Topology set 8 is selected, although three other sets have four traps. The first step is to obtain three values in log variables cells C19C21 that produce all positive element values; manually trying all four 0.5, 0, 0.5, or 1.0 or any combination produces several negative elements. Instead of manually searching for variable val-

Figure 8. The N=5A lossless and lossy response for the elliptic filter in example 2.

36

www.rfdesign.com

September 2000

ues for all positive elements, solver can be started from say three 0 variable values in variables cells C19-C21 with the objective of minimizing target cell B45 (a standard optimization strategy for obtaining constraint feasibility). The nonunique result was R 2=0, R 3=0.1625, and R 4=0.3000 so that the sum of the squared negatives in cells B24B41 is zero (in cell B45). The second step is to start solver with all elements positive, with the unconstrained objective of minimizing target cell D44 by varying changing cells C19-C21. Starting from the values in step one, the worst element ratio was reduced from 7.6224 to 5.3259, and all elements remained positive. The third step is to add N2=3 constraints and continue optimization from the solution to the preceding unconstrained optimization. Figure 7 shows the unique results obtained by requiring L12=L34, L23=L45, and L4 (by applying B25=B29, B27=B31, and B30=0, respectively in solver); the inductor ratio is only 8.3001 after constrained minimization. As soon as the bandwidth and prototype gi element data were entered, the estimated midband reflection and dissipative loss, L0 in dB, was shown in cell D16 for uniform resonator Q u =100. Fortunately, L0 is inversely proportional to resonator Qu, so the midband loss for Q u =50 is 0.43382=0.8676 dB. Figure 8 shows the insertion loss for this design for Q u , Q u =100, and Qu=50; the midband loss values closely agree with the above estimates. The only dissipative effect other than some symmetrical rounding of the passband is a decrease in the null depths. A relatively simple calculation [7:1010(5.10)] shows that the critical Qu, which just erases the nulls adjacent to the passband, is Qu=18.3, causing a midband loss of about 2.37 dB.

Conclusions
A new and useful transformation of a significant two-terminal LC network branch (dipole) has been described and applied to formation of direct-coupled filters. This exact transformation does not use the minimum number of inductors and capacitors (noncanonic), and thus allows arbitrary nodal impedances (parallel resistances) and converts series LC resonators to parallel resonators. The given dipole is the result of the standard lowpass to bandpass reactance transformation

for both all-pole and elliptic filters, which produces both series and parallel resonators. Transformation of all series branches to parallel form produces practical direct-coupled bandpass filters having only parallel resonators and undistorted response over any passband width. Those filters are easily converted to dual direct-coupled filters having only series resonators. The transformation results in L and C coupling pairs, including LC and CL null-producing traps, which can be placed in arbitrary order. For N resonators, 2N7, there are 210 combinations of parallel resonator directcoupled filters which have been programmed on nine worksheets in one Excel file. Upon selection of a listed coupling set, source, and readily available lowpass prototype element values, a normalized direct-coupled filter is defined for any values of the arbitrary N2 parallel resistances. A feasible region in parallel-resistance space for positive element values is guaranteed and easily reached by a special feature that uses the pointand-click spreadsheet optimizer. Then the worst element spread can be minimized and, subsequently, the N2 degrees of freedom can be used to eliminate elements and/or constrain them in constructive ways. Interactive use of these spreadsheets enables quick design of practical filters for any specified midband frequency and bandwidth. Constraints meeting the users needs are possible because of the wide-ranging set of positive element values available without distorting the filter response. The midband insertion loss is accurately predicted for moderate values of resonator unloaded Q. These lossless filters degrade uniformly with increasing dissipation due to the guaranteed geometric symmetry over all frequencies. Further information on the inexpensive, documented worksheet file is available from the author [8].

[3] Cuthbert, T. R. (1997). Wide-band Filters with Wide-Ranging Choices. RF Design. July:48-61. [4] Cuthbert, T. R. (1999). Broadband Direct-Coupled And Matching RF Networks. Greenwood, AR, published by author. [5] Zverev, A. I. (1967). Handbook of Filter Synthesis. NY: John Wiley. [6] Matthaei, G. L., L. Young, and E. M. T. Jones (1964). Microwave Filters, Impedance Matching Networks, and Coupling Structures. NY: McGraw-Hill. Also Boston: Artech House (1980). [7] Amstutz, P. (1978). Elliptic Approximation and Elliptic Filter Design on Small Computers. IEEE Trans. Circuits and Systems. Dec.:1001-1011. [8] Cuthbert, T. R. (2000). UCHOOSE Users Guide. Greenwood, AR, published by author.

About the author

Thomas R. Cuthbert Ph.D., P.E. teaches and consults at industrial sites. He was director of advanced technology at Rockwell and manager of microwave technology at Texas Instruments. He studied at M.I.T., Georgia Tech, and SMU, and has authored three books. He can be reached by fax at: 501.996.5618. email: [email protected] or on the Web: www.trcpepcom.

References
[1] Lawson, A. W and R. M. Fano (1948). The Design of Microwave Filters, in Microwave Transmission Circuits, M.I.T. Rad. Lab. Series, V9, pp.661-706. NY: McGraw-Hill. [2] Cohn, S. B. (1957). Direct-coupled Resonator Filters. Proc. IRE, V45. Feb.:187-196.

38

www.rfdesign.com

September 2000