Published Nov 21, 2026 · Last updated Nov 21, 2026 · 5 min · IndieRF

Select-at-Test Equalizer Resistors: 6 States, 77.6% Yield

One bridged-T gain-slope equalizer, 2 to 18 GHz, with six resistor pairs picked at test: 388 of 500 units land within 0.5 dB, and the misses cluster between states.

Six resistor pairs on one constant-resistance bridged-T cancel any amplifier gain slope from 1 to 6 dB across 2 to 18 GHz, and 388 of 500 simulated units (77.6%) land within a 0.5 dB peak-to-peak budget when the tester solders in the best pair. The 112 misses cluster between states: at nominal parts, from 5.35 to 5.92 dB of unit tilt, no pair is within budget.

The sheet opens with all six states, E96 values, the unit-tilt spread set to uniform from 1 to 6 dB and the default part tolerances (1% resistors, 5% L and C). Switch states on the Test states panel and the S21, residual and return loss plots follow.

One board, a few stocked resistors

Select-at-test splits the design in two. The capacitor and inductor set where the equalizer acts, and they are the same on every board. The resistor pair sets how deep it acts, and the technician picks it after measuring the unit's slope. On this example every board carries C1 = 0.184 pF and L2 = 0.461 nH; only R1 and R2 change.

That only works if changing the resistors leaves the match alone. A bridged-T with two R0 arms, a bridge arm Z1 and a shunt arm Z2 is R0 at both ports at every frequency when

Z1Z2=R02,S21=R0R0+Z1Z_1 Z_2 = R_0^2, \qquad S_{21} = \frac{R_0}{R_0 + Z_1}

Here Z1 is R1 ∥ C1 and Z2 is R2 in series with L2, so R2 = R0²/R1 and L2 = R0²·C1 hold the product for any R1. Each pair moves the loss shelf, and the input stays at 50 Ω in all six states. The E96 rounding of R2 and the 49.9 Ω arms is the only thing that breaks the identity, and it leaves the worst return loss of any state at 59.4 dB. Nobody will measure that through an SMA launch.

StateR1 (Ω)R2 (Ω)Tilt reached (dB)Ripple (dB p-p)Loss (dB)
1 dB33.2750.9950.1453.42 to 4.41
2 dB48.751.12.0250.2303.85 to 5.88
3 dB63.439.23.0570.2703.99 to 7.05
4 dB76.832.43.9600.2404.02 to 7.98
5 dB90.927.44.8450.2584.00 to 8.84
6 dB11022.65.9360.4623.94 to 9.87

The least loss in the band stays at about 4 dB in every state, the cap the wizard was given. That loss buys the flatness: one shelf can only follow a tilt that is linear in f while it still has loss left at 18 GHz, and the 6 dB state runs out first, with 0.46 dB of ripple.

The step between states comes from the budget

A unit whose tilt falls halfway between two states is off by half the step, as a straight line across the band, on top of the state's own ripple. So the step can be at most

step≤2×(budget−ripple)\text{step} \le 2 \times (\text{budget} - \text{ripple})

A 0.5 dB budget with 0.2 dB of ripple allows 0.6 dB steps. On this example the 4 dB state's 0.240 dB of ripple allows 0.52 dB, and the 6 dB state's 0.462 dB allows 0.076 dB. A 1 dB step is wider than either, so the rule predicts gaps.

The rule is a worst case. The straight-line error and the ripple partly cancel, so the real gaps are much narrower than the rule says: between the 4 and 5 dB states only 4.47 to 4.57 dB is uncovered. Use the rule to pick a starting step, then let the coverage chart say where the misses are.

What the SAT yield and the coverage chart say

The coverage chart plots each state's error against the unit's tilt at nominal parts, the best state in blue and the budget dashed. On this example it marks five stretches where no state is within 0.5 dB: 1.50 to 1.62, 2.55 to 2.70, 3.57 to 3.60, 4.47 to 4.57 and 5.35 to 5.92 dB. The last one is the big one, because the 6 dB state has the worst ripple.

The SAT yield then adds the part tolerances. Each of 500 boards draws its C1, L2 and arms once, draws each state's resistors on their own, draws a unit tilt from 1 to 6 dB, and picks the state with the least peak-to-peak error. 388 pass, 77.6%, with a Wilson interval of 73.7% to 81.0%. The picks per state, 57, 103, 96, 82, 103 and 59, are the stocking list: the end states get half as many units as the middle ones, because each covers only one side.

Another 0.1 dB of budget buys most of the 22% back without touching the board: at 0.6 dB, 472 of 500 units pass (94.4%), and only 5.45 to 5.72 dB stays uncovered. Keeping 0.5 dB means a finer step, and halving it to 0.5 dB doubles the stock to eleven resistor pairs and eleven rows on the test procedure. Which one is cheaper depends on who owns the system budget, so ask the systems engineer before redesigning the board.

Cable loss is easier than gain slope

A cable's loss rises like √f, which is the shape a shelf makes on its own. The cable example flattens 5 m of Huber+Suhner SUCOFLEX 104, 1.69 dB at 2 GHz to 5.50 dB at 18 GHz, to 0.033 dB of ripple with two shelves on E96 parts, inside the printed table of the sheet (200 MHz to 26.5 GHz). From the Coax Loss calculator, "Equalize this cable" opens the same wizard on whatever cable and length you have.

References

  • O. J. Zobel, "Distortion correction in electrical circuits with constant resistance recurrent networks," Bell System Technical Journal, vol. 7, 1928: the constant-resistance bridged-T and Z1·Z2 = R0².
  • H. W. Bode, "Variable equalizers," Bell System Technical Journal, vol. 17, 1938.
  • The note's numbers are the schematic engine's on the linked example, checked against a nodal solve of the same parts to 1e-6 dB, and the fit against the scipy fits in the IndieRF research notes for the equalizer plan.

Questions? Contact

Discussion

No comments yet. Start the thread with a measurement, a correction, or a worked example.