Central limit theorem

The formula

The classical (Lindeberg–Lévy) statement. Take X₁, X₂, …, Xₙ independent, identically distributed, with mean μ and finite variance σ²:

(X̄_n − μ) / (σ/√n)  →  N(0, 1)   as n → ∞

Reading: center the sample mean, scale it by its own standard deviation, and the result marches toward the standard normal — regardless of the shape you started with.

The working consequence engineers actually use every day:

σ_x̄ = σ / √n

Reading: the standard deviation of an average of n independent readings is the single-reading scatter divided by √n. Four times the samples, half the noise. Always √n, never n.

And the stack-up form — a sum of independent contributions:

σ_total = √( σ₁² + σ₂² + … + σ_k² )

Reading: independent errors add in quadrature, and the CLT is what lets you call the combined result normal even when the individual terms are anything but.

How fast the normal shape arrives is bounded by the Berry–Esseen theorem:

|F_n(x) − Φ(x)| ≤ C·ρ / (σ³·√n)      (C < 0.5)

Reading: the worst-case gap between your averaged distribution and the true normal shrinks like 1/√n, scaled by how skewed (ρ is the third absolute moment) the raw distribution is ^[8].

Where you meet it

The averaging button on the bench. Set the scope or DMM to average 16 acquisitions and the random noise drops by √16 = 4. That instrument feature is the CLT in firmware — and it's also why the improvement stalls: getting the next factor of 4 costs you 256 averages, and no amount of averaging touches a bias.
Tolerance stack-ups at the design review. Five parts, each toleranced ±0.010 in. Worst case says the stack is ±0.050. Treat each as uniform across its tolerance band (σ = 0.010/√3 ≈ 0.0058), stack them RSS, and the assembly sits at σ ≈ 0.0129, so ±3σ ≈ ±0.039 — and the stack distribution really is close to normal even though every input was flat. That's the argument behind every statistical-tolerancing note on a drawing, and the CLT is what makes it legitimate.
Control charts on the production floor. An X̄ chart plots subgroup means against ±3σ limits and assumes those means are normal. Individual parts off a grinder are frequently skewed; means of five are far closer to Gaussian. The CLT is the unstated license printed on every Shewhart chart in the building.
Uncertainty budgets. A GUM-style budget combines a dozen Type B terms — rectangular here, triangular there — RSSes them, then quotes a coverage factor k = 2 for "approximately 95%." That 95% is a normal-distribution number. It applies to the combined result only because the combination of many comparable terms goes normal.

How it works

The mechanism is convolution. The distribution of a sum is the convolution of the individual distributions, and convolution smooths. Add two uniform variables and you get a triangle. Add a third and the corners soften. By twelve you're close enough to Gaussian that early simulation software generated normal deviates by summing twelve U(0,1) draws and subtracting 6 — the variance works out to exactly 12 · (1/12) = 1, so the result approximates N(0,1) with no scaling at all. Each convolution pass erases more memory of the starting shape; the bell curve is what's left when the memory is gone.

What the theorem requires, in engineering terms:

Independence, or close to it. Correlated readings don't buy you √n. Average 100 samples of a signal riding on 60 Hz pickup inside one power-line cycle and you've averaged 100 copies of nearly the same error. The averaging button reduces what's random between acquisitions and faithfully preserves what isn't.
Finite variance. Heavy-tailed processes — the occasional cable-bump glitch, a flaky connector that produces rare huge outliers — converge slowly or not at all. The pathological case is the Cauchy distribution, where the mean of a million samples is exactly as scattered as one sample.
No dominant term. The formal version is Lindeberg's condition ^[6]; the bench version is that one contributor can't own the stack. If a single fixture error is 10× everything else, the total is shaped like that error, not like a bell. RSS the budget, then look at what fraction the biggest term carries.

Convergence speed depends on the skew of what you're averaging, and this is where the folklore "n = 30 and you're normal" earns its skepticism. Skewness of a mean of n iid samples falls as skew/√n: average 30 exponential holding times and the result still carries skewness 2/√30 ≈ 0.37 — visibly lopsided. Symmetric inputs converge fast; skewed inputs converge like a government program.

The mistake people actually make: the CLT normalizes the middle of the distribution of the mean. It says nothing about individual readings, and its tails converge last. A Cpk calculation or a 1-in-10,000 exceedance claim lives three-plus sigma out, exactly where the normal approximation is weakest. Averages go normal; extremes don't — maxima and worst-cases follow extreme-value statistics, a different theorem entirely. If the review board asks about the tail, "CLT" is not the answer.

History

The first version belongs to Abraham de Moivre, a French Protestant refugee doing mathematics in London. In a Latin pamphlet dated November 12, 1733, he showed that the binomial distribution — coin-flip counts — is approximated by what we now call the normal curve, the first appearance of the normal probability integral; the result was folded into later editions of his Doctrine of Chances ^[1]^[2]^[4]^[9].

Pierre-Simon Laplace turned a coin-flip trick into a general theorem. In a 1810 memoir he proved that suitably rescaled sums of independent, identically distributed variables tend to the normal distribution, inventing characteristic-function methods along the way, and built the machinery into his Théorie analytique des probabilités of 1812 ^[3]^[4]^[7]. For most of the 1800s the result sat there — widely used, not rigorously nailed down.

Rigor arrived from Russia: Aleksandr Lyapunov, better known in Huntsville for the stability theory that runs through every controls course, proved the theorem in 1901 under an explicit moment condition, using characteristic functions in essentially the modern way ^[4]^[5]. Jarl Waldemar Lindeberg of Helsinki found the near-sharpest condition in 1922 — the no-dominant-term requirement above ^[4]^[6]. The name came in between: George Pólya's 1920 paper title used "zentraler Grenzwertsatz" — central limit theorem — "central" because he judged it the center of probability theory, not because of the centering in the formula ^[1]^[4].

Related tools

/tools/snr-enob — averaging gain of √n in SNR terms is this theorem with dB units on
/tools/adc-resolution — quantization error is the classic uniform distribution that averages toward normal
/tools/rms-peak — converting the σ the CLT gives you into the peaks the hardware sees
/tools/crest-factor — Gaussian-noise peak-to-RMS assumptions come straight from CLT-shaped noise
/tools/thermal-noise-floor — Johnson noise is Gaussian because it's the sum of countless independent electron events
/tools/bearing-life-l10 — the counterexample: failure life is Weibull, a tail problem the CLT does not cover

Sources

Written by HE in our own words from the cited sources — engineering judgment included, your stamp still required. All entries →