imagine to sample a sinusoidal waveform with three samples per cycle: after
AD you see something like a saw-tooth waveform, that is a signal with the
fundamental frequency of the sampled signal plus additional harmonics. When
you play back to analog with a good DA the anti-aliasing filters after the
DA remove all those harmonics of the saw-tooth signal and leave to pass only
the sinusoid at the original frequency.
With oversampling systems, most of this work is done in the digital domain:
the original signal is oversampled, then filtered digitally (more accurately
than with analog filters), then converted back to analog with a high
sampling rate, say 8X or more than the original AD sampling, then again
filtered in the analog domain with a smooth filter.
in mathematical words, 3 points per cycle are enough to define one and only
one sinusoid.
of course real signals are not sinusoids, but composed by many sinusoids. in
any case we can assume that all sinusoids below Nyquist are reproduced well;
problems arise when there are components above Nyquist; these components
might be just our signal sweeping up, harmonics, noises, sounds from other
species, etc. We need to block them with an a-a filter to avoid they are
"reflected back" into the range we are interested in. In other words the
Nyquist frequency is like a mirror. We need to avoid signals hitting the
mirror.
Sharper the a-a filter, wider the bandwidth we can safely convert to
digital.
Gianni
|