The Central Limit Theorem: Why Averages Turn Into Bell Curves

The formal statement

The Central Limit Theorem (CLT) is arguably the single most-used result in statistics. It says that the mean of a large sample drawn from any distribution with finite variance is approximately normally distributed — regardless of whether the underlying population is uniform, exponential, bimodal, or anything else non-normal. Formally: let X₁, X₂, …, Xₙ be independent, identically distributed random variables with mean μ = E[Xᵢ] and variance σ² = Var(Xᵢ) < ∞. Define the standardised sample mean:

X̄ₙ = (X₁ + X₂ + ... + Xₙ) / n
Zₙ  = (X̄ₙ − μ) / (σ/√n)

Then:  Zₙ → N(0, 1)  in distribution as n → ∞

Equivalently, X̄ₙ is itself approximately N(μ, σ²/n) for large n. That single line justifies the 95% confidence interval x̄ ± 1.96·σ/√n, the z-test, the t-test and most of the machinery of classical inferential statistics — none of it requires the original population to be normal, only the sample mean drawn from it.

Why it's true: characteristic functions

The cleanest proof uses characteristic functions — essentially the Fourier transform of a probability distribution, φ_X(t) = E[e^(itX)]. Standardise each variable as Yᵢ = (Xᵢ − μ)/σ, so E[Yᵢ] = 0 and Var(Yᵢ) = 1, and write Zₙ = (Y₁ + ... + Yₙ)/√n. Because the Yᵢ are independent, the characteristic function of the sum is the product of the individual ones:

φ_Zn(t) = φ_Y(t/√n)ⁿ

Taylor-expand near t=0, using E[Y]=0 and E[Y²]=1:
log φ_Y(t) = −t²/2 + O(t³)

log φ_Zn(t) = n · log φ_Y(t/√n) = −t²/2 + O(1/√n)  →  −t²/2  as n → ∞

So φ_Zn(t) → e^(−t²/2), the characteristic function of N(0,1).

By Lévy's continuity theorem, convergence of characteristic functions to that of N(0,1) implies convergence in distribution to N(0,1) — that's the whole proof, and notice it never assumed the shape of the original distribution beyond a finite variance.

How fast does it converge?

The CLT guarantees convergence eventually, but says nothing about speed on its own — that's what the Berry-Esseen theorem supplies. It bounds the worst-case gap between the true CDF of Zₙ and the standard normal CDF Φ:

sup_x |P(Zₙ ≤ x) − Φ(x)|  ≤  C · ρ / (σ³ · √n)

ρ = E[|X − μ|³]     (third absolute moment)
C ≤ 0.4748          (best known constant, Shevtsova 2011)

For a Bernoulli(p) variable near p = 0.5 this works out to a maximum error near 1/(2√n) — about 5% at n = 100 and 1.6% at n = 1000. The practical upshot: n ≥ 30 is a fine rule of thumb for symmetric, well-behaved distributions, but a heavily skewed one such as Exponential often needs n ≥ 100 or more before the normal approximation is trustworthy.

The Galton board made it physical

Francis Galton built a physical demonstration of the theorem around 1876: the Galton board, or quincunx, drops a ball through a triangular array of pins, deflecting left or right with equal probability at each row. A ball's final horizontal position is the sum of n independent ±1 steps — a Binomial(n, 0.5) variable — and by the CLT, Binomial(n,p) → N(np, np(1−p)) as n grows. The height of each landing bin is proportional to a row of Pascal's triangle, and with enough rows the histogram of balls visibly settles into a bell curve. The same logic generalises far beyond coin flips: any quantity that arises as the sum of many small, independent contributions — measurement error, height, blood pressure — tends to be approximately normal for exactly this reason.

Where it breaks

The theorem has real preconditions, and violating them matters. Heavy-tailed distributions with infinite variance — a Pareto distribution with tail index at or below 2, or the Cauchy distribution — fall outside the CLT entirely; their sums converge instead to a Lévy-stable distribution, which is why financial returns and internet traffic with power-law tails don't obligingly average out to normal. The Cauchy distribution is the extreme case: it has no defined mean or variance, and the sample mean of n Cauchy variables is itself Cauchy-distributed for every n — averaging never helps at all. Dependent observations, such as correlated time series, also break the classical i.i.d. assumption, though a functional version of the theorem (Donsker's theorem, 1951) recovers a limit — the partial-sum process converges to Brownian motion, the foundation underneath stochastic differential equations and the Black-Scholes model.

Frequently asked questions

Try it live

Draw repeated samples from a Uniform, Exponential, Bimodal or Poisson distribution in Central Limit Theorem and watch the histogram of sample means converge to a bell curve regardless of the source shape. Everything runs client-side in your browser.