Autozeroing

Autozeroing is a circuit's way of measuring its own imperfection first, then subtracting it out before measuring anything else. Every real amplifier and comparator carries an input-referred offset — a built-in disagreement between what the input should read as zero and what the device itself treats as zero — caused by device mismatch that no amount of careful layout fully eliminates. Autozeroing does not fix that mismatch; it measures the resulting error during a dedicated phase and cancels it during the phase that actually matters.

The technique works only because offset and signal occupy different places on the time axis. During the autozero phase the amplifier's input is deliberately shorted or grounded, so whatever appears at its output in that moment is pure offset, with no signal contribution at all — that value gets sampled onto a storage capacitor. During the signal phase, the real input is reconnected and the amplifier produces its usual output, offset and signal summed together exactly as always; subtracting the stored offset value recovers the signal alone. This only works cleanly because offset is a slowly-varying, nearly-DC quantity compared to the signal being measured — if the offset itself changed as fast as the signal, there would be nothing left to subtract against.

Autozeroing earns its real value by refreshing that correction every single cycle, which is what separates it from a one-time factory trim. A trimmed offset is measured once, in the factory, and assumed to hold; it cannot track offset drift from temperature, aging, or supply variation after the chip ships. Autozeroing instead re-measures and re-cancels the offset on every clock cycle of normal operation, so whatever offset exists right now — not whatever existed at test time — is what gets subtracted. The correction never goes stale because it is never more than one cycle old.

The same mechanism that cancels fixed offset also substantially suppresses low-frequency flicker noise, and that is not a coincidence. Flicker noise, the 1/f noise dominant in MOS devices at low frequency, behaves like a slowly, randomly wandering offset rather than a fixed one — from the autozeroing circuit's perspective, that is indistinguishable from offset itself, so the same sample-and-subtract operation that removes a fixed offset also removes most of the low-frequency noise power sitting in the same part of the spectrum. What autozeroing cannot touch is thermal noise, which is spread flat across all frequencies and has no concentration near DC for the sampling operation to isolate and remove — and the sampling process itself introduces its own wideband noise, since every sample-and-hold operation folds high-frequency noise back down into the band of interest.

That residual folded noise, plus the comparator and switch non-idealities already covered in charge injection and kickback, is the real cost side of the technique. Autozeroing requires switches to disconnect and reconnect the input, and those switches bring exactly the charge-injection and clock-feedthrough errors that corrupt precision sampled circuits generally; it also requires extra clock phases, extra settling time, and typically halves the amplifier's usable bandwidth, since one full clock cycle out of every two is spent measuring offset rather than signal. Autozeroing is therefore chosen specifically where DC accuracy and low-frequency noise matter more than raw bandwidth — precision instrumentation amplifiers, chopper-adjacent sensor front ends, and slow, high-resolution data converters — rather than applied universally.

Autozeroing trades bandwidth for accuracy by spending part of every cycle looking inward instead of outward — measuring its own error before trusting its own measurement. The same discipline that makes a circuit self-correcting against offset and flicker noise is also what caps how fast it can run, which is why the technique shows up precisely where precision, not speed, is the design's actual priority.