Why averaging shots reduces image noise

WHAT THIS TESTS The question probes understanding of signal versus random noise and the square-root law, a core concept in computational photography and astrophotography.

A GOOD ANSWER COVERS Principle: the true scene signal is the same in every shot, but read and shot noise are random and approximately independent across frames with zero mean. When you average N aligned frames, the signal stays constant while the random noise partially cancels: its standard deviation drops by a factor of the square root of N. So the signal-to-noise ratio improves proportionally to the square root of N, meaning four frames roughly halve noise and you need four times as many frames to halve it again. Limitations: the scene and camera must be static or frames precisely aligned, otherwise averaging blurs motion; only random, independent noise averages out, so correlated noise like fixed-pattern noise, hot pixels, or quantization persists and needs dark-frame subtraction; gains diminish because of the square-root relationship, so doubling quality costs four times the captures; and very long stacks raise storage and capture-time costs.

COMMON WRONG ANSWERS Saying noise decreases linearly with frame count, which overstates the benefit. Or claiming all noise types average away, ignoring fixed-pattern noise. Or forgetting that motion requires alignment first.

LIKELY FOLLOW-UPS Why square root and not linear. What is the difference between random and fixed-pattern noise. How does frame alignment or registration enable handheld stacking. How does dark-frame subtraction remove correlated noise.

ONE CONCRETE EXAMPLE Photographing a dim night sky, a single short exposure is grainy because read and shot noise dominate the faint signal. Stacking 16 aligned exposures averages out the random component, cutting its standard deviation by about a factor of four and revealing faint stars that were buried before. To get the next factor-of-two improvement you would need four times as many frames again, illustrating the diminishing square-root returns. Separately, you subtract a dark frame to remove hot pixels and fixed-pattern noise, since that correlated noise is identical every frame and therefore does not average away no matter how many shots you stack.