danlavry wrote on Wed, 04 October 2006 20:03 
I too have an issue only with the section "what Nyquist did not say". Nyquist said that you need to sample FASTER THEN twice the highest frequency content, and that is a very important to understand.

I hate to be pedantic, but Nyquist said the following:
"Let x_c(t) be a bandlimited signal with
X_c(jΩ) = 0, for all Ω > Ω_N
then x_c is uniquely determined by its samples x[n] = x_c(nT), n=0, +1,+2,..., if
Ω_s = 2pi/T > 2Ω_N"
The above is verbatim from Oppenheim and Schafer's book, which is the standard text on DSP theory.
X_c(jΩ) = 0 means that there is
nothing outside the frequency band of interest.
Ω_s is the sampling frequency.
Ω_N is the
bandwidth of the signal.
the n=0,+1,+2... is
periodic sampling.
Assuming digital audio, Ω_s is the highest frequency component. There is a lot of work and information on the net (mainly related to digital communications) that makes use of bandlimited sampling. This is sometimes called subsampling or undersampling.
If the "X_c(jΩ) = 0" part is violated, then you will have aliasing. All real systems will have some degree of aliasing, even if it is minute.
Practically, clock jitter means that the "n=0,+1,+2..." part is violated. All real systems will have clock jitter, even if is is minute.
Nyquist also assumes infinite precision. Real systems use quantization.
What sets converters apart is how they handle the aliasing, jitter, and quantization issues.