i
i
i
i
i
i
i
i
9.5. Sampling Theory 231
representation of the filtered signal. Of course, we have lost the high frequencies,
but that’s better than having them get scrambled with the signal and turn into
artifacts.
Preventing Aliasing in Reconstruction
From the frequency domain perspective, the job of a reconstruction filter is to re-
move the alias spectra while preserving the base spectrum. In Figure 9.48, we can
see that the crudest reconstruction filter, the box, does attenuate the alias spec-
tra. Most important, it completely blocks the DC spike for all the alias spectra.
This is a characteristic of all reasonable reconstruction filters: they have zeroes
in frequency space at all multiples of the sample frequency. This turns out to be
equivalent to the ripple-free property in the space domain.
So a good reconstruction filter needs to be a good lowpass filter, with the
added requirement of completely blocking all multiples of the sample frequency.
The purpose of using a reconstruction filter different from the box filter is to more
completely eliminate the alias spectra, reducing the leakage of high-frequency ar-
tifacts into the reconstructed signal, while disturbing the base spectrum as little
as possible. Figure 9.52 illustrates the effects of different filters when used dur-
ing reconstruction. As we have seen, the box filter is quite “leaky” and results in
plenty of artifacts even if the sample rate is high enough. The tent filter, result-
ing in linear interpolation, attenuates high frequencies more, resulting in milder
artifacts, and the B-spline filter is very smooth, controlling the alias spectra very
effectively. It also smooths the base spectrum some—this is the tradeoff between
smoothing and aliasing that we saw earlier.
Preventing Aliasing in Resampling
When the operations of reconstruction and sampling are combined in resampling,
the same principles apply, but with one filter doing the work of both reconstruction
and sampling. Figure 9.53 illustrates how a resampling filter must remove the
alias spectra and leave the spectrum narrow enough to be sampled at the new
sample rate.
9.5.6 Ideal Filters vs. Useful Filters
Following the frequency domain analysis to its logical conclusion, a filter that is
exactly a box in the frequency domain is ideal for both sampling and reconstruc-
tion. Such a filter would prevent aliasing at both stages without diminishing the
frequencies below the Nyquist frequency at all.