noise shaping

Noise shaping is a clever sleight of hand that does not remove quantization noise but moves it out of the way. If you cannot avoid making a little rounding error on every sample, you can at least arrange for that error to pile up at frequencies you do not care about — high up, away from your signal — and then filter it off. It is like a tidy worker who cannot avoid making sawdust but always sweeps it into one corner where it is easy to clear, leaving the workbench clean.

The trick is a feedback loop that remembers its own mistakes. After each crude conversion, the converter computes how far off it was (the quantization error) and feeds that error back, through an integrator, to be added into the NEXT input. This forces the loop to keep correcting its running error, which mathematically pushes the noise spectrum to slope upward with frequency: tiny at low frequencies where your signal lives, large at high frequencies where nothing matters. A digital low-pass filter then discards the noisy high end, leaving a remarkably quiet low band. This is what lets a sigma-delta converter wring 20-plus real bits out of a humble 1-bit comparator.

Why this matters: noise shaping is the second half of the sigma-delta magic (oversampling is the first). Oversampling alone spreads the noise out evenly; noise shaping actively tilts it away from your signal, so you gain far more resolution per octave of oversampling than averaging alone could give. Higher-order loops (more integrators) shape more aggressively for even better in-band quiet. The honest catch: those higher-order loops can become unstable if pushed too hard, and the scheme only works because it is paired with oversampling — there must be empty high-frequency band to sweep the noise into.