> The Nyquist theorem never stated anything about accuracy, only that to
> reproduce any given frequency, it must be at least half the sample rate.
> If you sampled a 10khz sine wave at 20khz, it would become a square wave
> - certainly NOT accurate. Most filtration in CD players would wind up
> rounding the edges off anyway, but it's still not accurate - as compared
> to the source.
At the risk of spiraling way off-topic, I feel compelled to address this
very common misunderstanding. The Nyquist Theorem provides the
mathematical underpinning for ∗exact∗ transformation of a continuous
representation of audio into a discrete representation. If you sample a
10kHz sine wave at 20.01kHz, you get a 10kHz sine wave coming back out.
There are no "edges" to round off, because the digital-to-analog
reconstruction is not done by connecting the dots.
Of course, this is all in theory. Paul's caution about the difference
between theory and practice is quite correct. Since perfect filters don't
exist, one has to sample at noticeably more than 2x the highest frequency
to be represented. The question is whether 44.1kHz is enough, or whether
you need 96k or even 192k, and if higher sample rates are necessary, what
type of filter is optimal.
It's worth noting that there's more than just representation of audio to
be considered. Any non-linear processing (such as compression or VA
synthesis) is going to tend to produce aliasing, and higher sampling rates
can greatly reduce the impact of that aliasing on audible frequencies.