Well, I believe in many cases the "fastest response" hardware is the best solution. Say, you want to count pulses from a 10.000 RPM, high-resolution encoder. Either you implement some CPU that can handle way more than 1MHZ pulses, or use dedicated/custom hardware. In that case, raw '51 is out of question. But if the only case where high frequency is possible is read error conditions, I highly doubt the problem would appear.


Say, our signal is like this:
_ = low, - = high, # - high-freq noise.

If, as you say, this happens only when the encoder comes to a complete stop, only four variants are possible:
low-noise-low,
low-noise-high,
low-noise-low,
high-noise-high,

low-noise-low:

Readouts on edges of A and B. Noise gets antialiased and rounded to "1". We get a common "turn-back" scenario. First, the encoder comes to complete stop, then changes direction.

A:-----______--------____###____------____---__
B:---_____------_______-------_____-----____---

We round the noise to "1".

A:-----______--------____---____------____---__
B:---_____------_______-------_____-----____---

We have normal turn-around.

High-noise-High:
The other turn-back scenario:

A: __--__---____----##----____---__
B: _--__--____----______----____--__

# = -

A: __--__---____----------____---__
B: _--__--____----______----____--__


Scenario High-noise-low. Our encoder stops, then continues running in the same direction as before.

A: __--__---____-------####______----___--__

B: _--__--____------_________------___---__-

Filtered:

A: __--__---____-----------______----___--__
B: _--__--____------_________------___---__-

Our edge readout is delayed by the length of the noise, but the "error" is the margin of the noise-generating area.

Second variant of this problem (with low to high transition) isn't much different. Our readout appears maybe a bit "ahead of schedule" but shouldn't cause problems.


A: __--__---____------____####-----___---___
B: _--__--____------____--------____---____

Applying antialias to 1

A: __--__---____------____---------___---___
B: _--__--____------____--------____---____

If we never approach frequency at which our smoothing function creates a continuous wave, do we risk missing anything?

Of course if we do approach, or reach the speed, like that...

A: ------____---------___---___--__--___---
B: ---_____-------______---___--__--___---_

into:

A: --------__-----------_-----_--------_---
B: -----___---------____-----_--------_----

then the whole idea will fail. Signal treated like noise will get filtered out.