In general? I don't know. Nonlinear filters can behave very differently, so it's hard to guess anything about them in generalurosh wrote:mystran wrote:My question was actually something like this: is there any way to perform bandlimiting of nonlinear filter with out going down the oversampling road?
I don't know about any revolutionary "trick" to avoid the generation of harmonics in the square(), abs() or max() rectification stage. IMHO you can't avoid them.
But the branch related harmonics/aliasing can be avoided/reduced with the methods mentioned below:
- One thing always works: Band-limit the input signal. But this may reduce the precision of the detector and effectiveness of your application.
- For A/R detectors, you can smooth the attack/release transition by moving the branch out of the actual detection signal and use two filters instead, where one controls the "cutoff" freq of the other to create a smooth A/R transition.
- It's often possible to simply avoid the A/R branch completely and use a one stage detector instead (which means the filter stage becomes fully linear).
- Non-causal/Look-ahead techniques can also improve the situation for specific type of signals and can most of all keep the phase information intact(moving averages, FIR filters, Hilbert detector, median, windowed max, ect). Further on, the step response can be controlled in a wider range.
But it isn't easy to avoid or smooth branch "corners" in the latter context.
