I have to disagree. A step response is an artificial signal that violates the Nyquist criteria.
Not as I understand the point to which you are responding.
A step in a digital audio file, reconstructed through the Shannon-Whittaker interpolation formula, produces a band-limited analogue audio step that "rings" [1] at the step. The corollary is that an analogue audio step when band-limited, sampled and quantized
can produce a step in the digital audio file as long as the sampling points correspond to the step-height crossings of the band-limited ringing (unlikely in practice but theoretically possible).
The same argument applies in the case of an impulse in the digital audio file and its both-ways correspondence with the band-limited sinc function in the analogue audio.
The only case I can see for a violation of the Nyquist criterion lies in the
strict requirement that sampling frequency/2 [Fs/2] > maximum signal frequency content [Fmax]. That is, any frequency content
exactly at Fs/2 gets aliased to Fs/2, so content at exactly Fs/2 is not allowed and would violate the Nyquist criterion.
However, I
think that for a finite length digital audio file with finite quantization I can always choose Fs/2 = Fmax + epsilon for a value of epsilon that is as small as I like but non-zero. So I think that any sequence of sample values in a
real-world audio file always reconstructs to meet the strict inequality of the Nyquist criterion.
Even if this is wrong (and I would be happy to be convinced it's wrong), aliasing from Fs/2 to Fs/2 is not a practical problem. So I always have difficulty understanding the practical significance of assertions I see at times that illegal sequences of samples are possible in a digital audio file.
[1] I don't like the word "ringing" in this context - it isn't ringing in the sense I normally use - it's just a resemblance.