Correlated Double Sampling

Correlated double sampling takes two measurements close together in time — a reference or reset level, then the actual signal level — and subtracts one from the other, canceling out whatever error happens to be present in both samples identically. It is the third major technique this series covers for fighting offset and fixed-pattern error, distinct from autozeroing's amplifier-level measure-and-subtract and chopper stabilization's frequency-domain relocation, and it is the dominant technique used in image sensor pixel readout and many precision data-converter front ends.

The word "correlated" is doing real work in the name, and it is the entire condition the technique depends on. A reset sample captures whatever baseline error — amplifier offset, pixel-level fixed pattern variation, a reference voltage that drifted slightly — exists at that specific moment; the signal sample, taken shortly afterward with no intervening reset event, carries the identical baseline error plus the actual signal on top of it. Subtracting the two removes exactly the baseline error, leaving only the signal, but this cancellation only works because both samples share the same physical instance of that error — correlated double sampling is named for the requirement, not just the result.

This is the same fixed-pattern-cancellation principle autozeroing uses at the amplifier level, applied instead at the sensor or pixel level where the dominant error source is different. Autozeroing shorts an amplifier's input, measures its own output offset, and subtracts that stored value during the signal phase — a technique aimed specifically at one amplifier's internal mismatch. Correlated double sampling more commonly targets variation across an entire array of nominally identical elements, most visibly in a CMOS or CCD image sensor, where every pixel's own reset level differs slightly from its neighbors due to threshold-voltage mismatch across the array; reading each pixel's own reset level and subtracting it from that same pixel's signal level removes that pixel's specific offset without needing to characterize or calibrate the array in advance.

The technique's limitation is exactly as sharp as its name implies, and it is a frequent source of confusion for anyone who assumes double sampling removes noise in general. If a fresh reset event occurs between the two samples — a new reset pulse that re-establishes the baseline rather than simply reading an already-settled one — the thermal kT/C noise captured at that reset is its own independent random draw, uncorrelated with whatever kT/C noise accompanied the previous reset. Subtracting two independent noisy quantities does not cancel their noise; it combines their noise powers, meaning a naive double-sampling scheme that resets between every pair of samples can actually leave a design with more total noise than a single, uncorrected sample would have shown, not less.

This is fully consistent with — not an exception to — the kT/C noise floor this series has already established as untouchable by anything except a larger sampling capacitance. Correlated double sampling cancels systematic, repeatable error: fixed pattern offset, slow drift, low-frequency flicker noise that barely changes between the two closely-spaced samples. It does nothing for the thermodynamic noise floor set by a fresh reset event's own random thermal draw, because that noise is not repeatable or shared between samples in the way a fixed offset is — the same physical reason kT/C noise resists every other circuit-level trick this series has covered applies here as well.

Where this technique is actually deployed reflects exactly which error it is good at removing. Image sensor pixel readout chains use correlated double sampling specifically to remove pixel-to-pixel fixed pattern noise from reset-level mismatch across a sensor array containing millions of nominally identical elements, a problem no single amplifier-level autozeroing scheme could address since the error varies pixel by pixel rather than living inside one shared amplifier. Pipeline and sigma-delta analog-to-digital converter stages use related double-sampling schemes to cancel their own comparator and amplifier offsets stage by stage, closer in spirit to autozeroing but applied at each conversion stage rather than once per amplifier.

Correlated double sampling completes the trio of precision-analog offset-fighting techniques this series has covered, and the differences between the three matter more than the similarities. All three exploit the same underlying idea — take a measurement of error, then remove exactly that error from the measurement that matters — but they do it at different points in a system, against different sources of error, with different limitations on what they can and cannot touch. An amplifier's own mismatch, a signal's frequency-domain placement relative to a noise corner, and an array element's own reset-level variation are three different problems wearing the same solution shape, which is exactly why a real precision system often needs more than one of these techniques working together rather than picking a single favorite.