oversampling
Oversampling means deliberately sampling much faster than the Nyquist theorem strictly requires — not because you need the extra speed, but because spreading the work over more samples buys you cleaner data. The everyday picture: to measure something fuzzy more precisely, take many readings and average them. The random scatter partly cancels, and the average is steadier and finer than any single reading. Oversampling is exactly that, applied to a converter.
The mechanism rests on a fixed fact: a converter's quantization noise is a fixed total amount of power, and sampling faster spreads that same power thinly across a wider frequency band. Your real signal only occupies a small slice of that band, so when you digitally filter away everything outside the slice, you also throw away most of the noise but keep all the signal. The payoff is precise: oversampling by a factor of 4 cuts the in-band noise enough to gain one extra bit of resolution; in general each factor-of-4 in sample rate adds one bit. To gain 2 bits you oversample 16 times; 3 bits needs 64 times.
Why this matters: oversampling lets a cheap, modest converter punch above its weight, recovering resolution in software that the hardware alone does not have. It also relaxes the analog anti-alias filter dramatically, since the Nyquist edge is now far away, so a gentle filter suffices. The honest catch is the cost in speed and data: each extra bit costs 4 times the sample rate and processing, so there are diminishing returns. And it only helps if there is some noise or dither to average — perfectly noiseless oversampling of a steady DC gives the same code every time and gains nothing. Sigma-delta converters combine oversampling with noise shaping to push the idea much further.
A 12-bit ADC sampled at 16 times the needed rate, then digitally averaged and decimated, gains log4(16) = 2 effective bits, behaving like a 14-bit converter in the signal band. Going to 64 times oversampling would gain 3 bits.
Sample fast, filter narrow — most of the noise falls outside your band.
Oversampling only converts to extra bits if there is enough noise (or added dither) to randomize the quantization error across samples. Averaging a perfectly quiet, unchanging input just returns the same quantized code and buys nothing.