Let us consider a simple example of commercial - even cheap-and-dirty: An asynchronous serial-bit receiver, sampling exactly once per bit-time, usually avoids the transition interval, with some small error possibility: A quick-and-dirty engineering solution might simply adjust its time-base to a new (even random-arbitrary) offset: when it catches packet errors on its time-synchronization ... of course, that packet is lost already.
Advanced technique {commercial, military} might sample exactly twice per bit-time - evenly spaced - and, detecting on both, correct an erred packet - choose the un-erred packet - and switch to that as the preferred time-synchronization.
But this allows for the samples to be taken from near quarter-phase of the bit-times: when the bit-waveforms are most irrectangular - recall the sinc function (sin x/x) as the temporal transform of the bit in the rectangular low-pass-filtered frequency domain (a TEMPEST emission control ideal) - "the Fourier transform of a top-hat in one domain is a sombrero in the other" - and the detector must continually switch between erred packets ... and without adaquate parity-check, that's impossible.
3 or more odd samples per bit, repositions the valued sample to the central third or narrower: the first and last samples are nearest the transition uncertainty. [Even numbers of samples tend not to center the valued sample - and, 4 samples per bit still has that quarter-phase trouble, albeit, a bit-transition can be detected instead of the bit directly, by so valuing both 2nd and 3rd samples: leaving the 1st and 3rd least certain]