Attenuators
Flatness, Accuracy and Monotonicity
An attenuator can be accurate and useless. These three columns describe three unrelated ways of being wrong.
Flatness, accuracy and monotonicity name three different defects, not one defect with three spellings. Flatness is the variation of attenuation across the frequency band at a fixed setting. Accuracy is the distance between the attenuation asked for and the attenuation delivered. Monotonicity is the property of a step attenuator whose attenuation never falls as the command word rises. A part can sit inside its accuracy limit and still break a servo loop, which is why the three columns are read separately.
The mechanism decides which defects are possible. An attenuator lowers a level in a controlled way and leaves the frequency alone; the variable kind takes a command, an analog voltage or a binary word. The function, what a variable attenuator does, has its own primer; the concern here is how a commanded part misses the ideal.
Analog voltage or binary word
A voltage variable attenuator answers an analog level: attenuation moves continuously with the control voltage, and the curve is linear in neither volts nor decibels unless the designer has shaped it. A digital step attenuator answers a word. Its attenuation comes from weighted sections, for example 0.5, 1, 2, 4, 8 and 16 dB, that the word switches in and out. The split of voltage variable against digital step decides which column applies: monotonicity only means something where there is a word. Two figures frame the range. Residual insertion loss, the attenuation still present at the minimum setting, cannot be zero, and the range above that floor is the easiest column to write and the least useful, because a part is almost never run at its maximum.
What is flatness sweeping?
Flatness holds the setting still and lets the frequency move. At any fixed code the delivered attenuation is not the same at every point of the band, and that spread is the flatness figure. The defect is quiet: a level set at one frequency silently becomes another level toward the band edges, so a mid band calibration does not transfer to the ends. The column carries its own trap. Flatness is defined at a fixed setting, so a figure that does not say which setting it was taken at answers a question nobody asked, and two sheets quoting flatness at different settings and bands are not comparable.
Accuracy is one distance
Accuracy compares two attenuations, the one requested and the one obtained, and it is read at a frequency. That is where the column misleads. A figure taken at a calibration frequency says nothing about the same figure elsewhere in the band, which is flatness territory; two sheets quoting accuracy under different settings and frequencies are not comparable. The distance also travels: decibel losses of cascaded elements add, so a stage delivered a fraction of a decibel away from its request shifts the whole chain after it by the same fraction. The arithmetic behind these columns is gathered in the parameter index.
How does an accurate part lose a loop?
Monotonicity fails at the carries of the word. Take the six sections above. The word 011111 switches on the five smaller sections and holds the 16 dB section off: a nominal 15.5 dB. The next word, 100000, switches all five off and the large one on, for a nominal 16 dB. Give every small section a tenth of a decibel more than its nameplate, and the large section a tenth less. The five together deliver 16.0 dB; the large one alone delivers 15.9. The word has advanced one step, and the attenuation has fallen by 0.1 dB, while no code sits more than half a decibel from the value printed beside it: an accuracy limit of that width passes the part, and a sheet quoting accuracy alone still passes. A servo loop asks that carry for more attenuation, receives less, asks again, and hunts. The loop never reads the accuracy column; it reads the steps.
The fourth defect rides along
Attenuation almost never travels alone. Each change of setting brings a change of phase with it, and none of the three columns above says a word about that phase. It stays invisible until a second channel appears, at which point two paths leveled separately stop tracking. Phase shift with attenuation is read beside the other three, not after them. The table holds, on four lines, what each term varies and what it breaks.
| Term | What varies | What breaks |
|---|---|---|
| Flatness | Attenuation across the band at a fixed setting | A level set at one frequency becomes another at the band edges |
| Accuracy | Delivered attenuation against requested attenuation | The error enters every cascaded stage, because losses add |
| Monotonicity | Attenuation from code to code as the word rises | A servo loop hunts at a carry, accuracy still in band |
| Phase shift | Phase alongside attenuation, setting against setting | Any second channel, once the paths stop tracking |
Checks before the columns are trusted
- Ask a flatness figure for its setting and its band; without both, it answers nothing.
- Treat residual insertion loss as the floor of the range, never as an error.
- Hold one code and sweep; that spread is the flatness of that code alone.
- Walk every code and record each attenuation; carries hide between named values.
- Calibrate the analyzer that day, SOLT or TRL, and torque couplings with the proper wrench.
Common mistakes
- Comparing range columns and stopping there; a part is almost never run at its maximum.
- Quoting one accuracy number as though it held across the band.
- Assuming a voltage variable attenuator responds linearly, in volts or in decibels.
- Testing only the codes named on the sheet and missing the carries.
- Forgetting the phase that rides with the attenuation until a second channel appears.
- Trusting yesterday's calibration; it drifts with temperature and time.
One sweep and one walk
Open the next sheet at the flatness line and ask the two questions it depends on: which setting, and which band. Then take the part to a bench qualified for the tenth of a decibel. Hold a code and sweep, so the flatness of that code becomes a curve, not a claim. Walk the word from its floor to full scale, one code at a time, attenuation recorded at each, eyes on the carries. Those two passes answer what the columns argue over: what the part delivers, at which frequency, under which word, and which way it moves when asked for more.