quantization noise

Every time a converter rounds a true voltage to the nearest level, it throws away the leftover fraction. Gather up all those tiny discarded scraps across a whole waveform and they look and behave like a faint hiss riding on the signal. That hiss is quantization noise — not interference from outside, but the built-in cost of rounding, the audible (or measurable) ghost of the steps in the ruler.

Because the rounding error on any sample is somewhere between minus half an LSB and plus half an LSB, fairly evenly, its size works out to a fixed fraction of an LSB (about 0.29 LSB in RMS terms, the equivalent heating value, not the average). From this comes the famous rule for the best-possible signal-to-noise ratio of an ideal converter: SNR is about 6.02 times n plus 1.76 dB, where n is the number of bits. Each extra bit halves the step, halves the noise, and buys you about 6 dB more dynamic range. A 12-bit converter tops out near 74 dB; a 16-bit near 98 dB.

Why this matters: quantization noise is the theoretical best a converter could ever do — a floor you cannot get below by improving the analog parts, only by adding bits. Real converters always do somewhat worse because thermal noise, distortion, and reference wobble pile on top. One clever escape exists though: if you sample much faster than you need (oversampling) and filter, you spread this fixed noise power over a wider band and keep only the slice in your signal band, recovering effective bits. Noise shaping in sigma-delta converters pushes the trick further.