Noise, averaging & SNR

Every real measurement arrives wearing static: thermal hiss in circuits, grain in photos, jitter in sensors. Noise is randomness riding on signal — which means the statistics you know is also the toolkit for fighting it.

Build the intuition

Noise is a distribution in motion

Model each measurement as truth plus a random draw: x[n] = s[n] + noise, the noise typically bell-shaped (it's a sum of many tiny disturbances — the central limit theorem at work in your electronics). Mean zero, spread σ: the noise floor. One glance at a sensor's σ tells you what size of effect it can hope to see.

x[n] = s[n] + \varepsilon[n], \quad \varepsilon \sim \mathcal{N}(0, \sigma^2)

Averaging: the √n discount

Average n repeated measurements: the signal part adds coherently (it's the same every time) while the noise partially cancels (random signs). Noise σ shrinks by √n — average 100 frames and the hiss drops tenfold. The same law that sized polls now cleans telescope images. Averaging is the moving-average filter from signals, viewed statistically: smoothing is noise reduction is low-pass filtering.

\sigma_{avg} = \frac{\sigma}{\sqrt{n}}

SNR: the one number that decides

Signal-to-noise ratio — signal power over noise power, usually in decibels — is detection's currency. Communication links budget it (every dB costs antenna size or transmit power); astronomers integrate for hours to buy it; audio gear advertises it. Below ~0 dB, signal drowns; engineering is largely the art of climbing the SNR ladder.

See it move

InteractiveThe law of large numbers

Flips40

After 40 flips: 23 heads = 57.5%. Single flips are pure chance — but the running average is drawn, inevitably, toward 50%. Randomness has long-run structure.

Averaging in its purest form: each flip is a noisy measurement of 0.5, and the running mean's wobble shrinks like 1/√n — watch the noise floor sink.

A worked example

Rescue a faint star

A star registers 2 units per pixel per frame; the camera's noise is σ = 10 — the star is invisible (SNR 0.2).
Stack 400 frames: signal stays 2; noise falls to 10/√400 = 0.5.
New SNR = 4 — the star stands clearly above the floor. Amateur astrophotographers do exactly this every clear night; the math is this lesson's one formula.

Out in the world

Voyager's whisper

Voyager 1 transmits ~20 watts from beyond the solar system; arriving signal power is around 10⁻¹⁶ watts — far beneath the receiver's own noise. NASA wins by narrowing bandwidth, integrating over time, and error-correcting codes: a 47-year masterclass in trading time and statistics for SNR.

Common confusion, cleared

“With clever processing you can always recover a signal from noise.”

Averaging needs repetition, filtering needs the signal and noise to occupy different frequencies. Information genuinely below the floor with no structure to exploit is gone — statistics sets the price, not the wish.

“Doubling the measurements doubles the quality.”

It buys √2 ≈ 1.4× — the square-root law is sublinear, which is why each additional dB of SNR gets progressively more expensive.

Recap

Measurements = signal + bell-shaped noise; σ is the floor.
Averaging n trials divides noise by √n — coherence beats randomness.
SNR is the detection currency; engineering largely means buying more of it.

Progress saves in this browser.