With 44,100 sample per second, Nyquist is 22,050. If you want flat clean response (passband) to 20,000, then the transition band of the filter is 20,000 to 22,050, which is quite narrow, only 2,050 wide or 0.14 octaves. A sharp/narrow filter like this is hard to implement in real time on the hardware available in a normal DAC chip, without passband side effects. So manufacturers "cheat" and extend the stop band above Nyquist to 24,100 Hz. This *doubles* the width of the transition band, making the filter easier to implement. But what about eliminating supersonic noise and their aliases, which is the very purpose of the filter? Well it happens that aliases reflect around Nyquist, so any aliases of high frequency noise (above Nyquist) are above 20 kHz, pushing them above the passband.

Put differently, in this situation the passband limit is 20k, transition band is 20k to 24.1k, Nyquist is 22,050 which happens to be smack-dab exact center of the filter transition band. Not by coincidence. Worst case scenario, the filter passes energy at 24,000 Hz which is by definition noise since it is above Nyquist. The alias of this frequency is its arithmetic mirror image across Nyquist, which is 22050 - (24000 - 22050) = 20,100. That's above the passband and inaudible. Conversely, the highest passband frequency (20,000) has an alias of 24,100, which the filter fully attenuates.

This is why we so often see the "sharp" filter in DACs that Amir reviews having a stop band above Nyquist at exactly 24,100 Hz. It's an engineering kludge to bend (break) the rules in a way that makes the filter easier to implement while avoiding passband noise or distortion. DAC chips usually don't do this at higher sampling rates. It's not necessary, since higher sampling rates have a much wider transition band, the DAC can fully attenuate by Nyquist with no engineering sleight-of-hand needed.