The Central Limit Theorem states that the sum (or average) of many independent random quantities tends toward a normal distribution — regardless of how each individual quantity is itself distributed. This remarkable universality is why the bell curve appears so pervasively throughout statistics, science, and nature, and the theorem underlies the vast majority of statistical inference methods used in practice.
Roll a single die: the result is completely unpredictable, and equally likely to be any of 1 through 6 — a flat, uniform distribution, nothing like a bell curve at all. But now roll ten dice and add up the total. Do this experiment many times, and plot all your totals. You'll get a beautiful, symmetric bell-curve shape — even though each individual die roll wasn't bell-shaped at all.
This theorem explains why so many real-world measurements — human heights, exam scores, manufacturing tolerances, measurement errors — tend to follow roughly bell-curve patterns: they're often the net result of adding together many small, roughly independent contributing factors (genetics, environment, and nutrition all affecting height, for example). It's also the essential mathematical foundation underlying most everyday statistical methods: opinion polls, clinical drug trials, quality-control testing, and much more all rely on this theorem to justify their standard analytical methods and to correctly compute margins of error.
Francis Galton's ingenious 19th-century physical device, the "Galton board" (or "bean machine"), drops balls through a triangular grid of pegs, each peg randomly deflecting the ball left or right. Even though each individual deflection is a simple, unpredictable 50-50 random event, the balls collectively pile up at the bottom in a strikingly clear bell-curve pattern — providing a simple, direct, and visually compelling physical demonstration of the Central Limit Theorem in action.
Let X₁, X₂, ..., Xₙ be independent, identically distributed random variables with finite mean μ and finite variance σ². Define the standardised sum:
Then as n→∞, Zₙ converges in distribution to the standard normal distribution N(0,1) — regardless of the original distribution of the individual Xᵢ.
There's no single universal answer — it depends on how far the original distribution is from normal to begin with. As a common statistical rule of thumb, n≥30 is often considered sufficient for reasonably well-behaved original distributions, but distributions that are highly skewed or have very heavy tails may require considerably larger sample sizes before the normal approximation becomes genuinely accurate and reliable.
The Central Limit Theorem is the essential mathematical justification for why sample means from opinion polls and scientific studies can be treated as approximately normally distributed, which in turn is precisely what allows researchers to construct valid confidence intervals and correctly compute the margins of error you routinely see quoted alongside professionally-conducted opinion poll results (e.g., "52% support, ±3% margin of error").
Statistical process control methods in manufacturing rely directly on the Central Limit Theorem: even if individual manufacturing defects or measurement errors don't follow a normal distribution themselves, the average measurement across a properly sized sample of manufactured products will approximately do so — enabling manufacturers to reliably set statistically meaningful and well-calibrated quality-control thresholds and tolerance limits.
A common and important misconception is that the Central Limit Theorem implies most naturally-occurring quantities are normally distributed. It actually says something more specific and precise: it's the sums or averages of many independent random quantities that tend toward normality — not necessarily every individual random quantity itself. Individual measurements can still follow decidedly non-normal distributions even while their sums or averages closely follow the CLT's normal approximation.
Abraham de Moivre (1733) proved an early special case for the binomial distribution — showing that as the number of coin flips grows large, the binomial distribution's shape approaches the normal curve. Pierre-Simon Laplace substantially generalised this result over subsequent decades. The specific term "central limit theorem" itself was coined by George Pólya (1920), reflecting the theorem's genuinely central, foundational importance within probability theory more broadly, rather than referring to any literal notion of a mathematical "centre" or "limit point."
The classical Lindeberg-Lévy version requires the Xᵢ to be identically distributed. Aleksandr Lyapunov (1901) proved an important generalisation requiring only independence (not identical distribution), provided a specific technical condition on higher moments (the "Lyapunov condition") is satisfied. Jarl Waldemar Lindeberg (1922) further weakened and generalised the required conditions (the "Lindeberg condition"), producing what is now considered the most general classical, broadly applicable version of the theorem for independent (but not necessarily identically distributed) random variables.
The Central Limit Theorem itself only guarantees convergence to the normal distribution as n→∞, without specifying how quickly or accurately this approximation becomes valid for any specific, finite sample size n. The Berry-Esseen theorem (proved independently by Andrew Berry, 1941, and Carl-Gustav Esseen, 1942) provides an explicit, quantitative bound on precisely how far the true distribution of the standardised sum can be from the normal distribution, for any given finite n — a result of considerable practical importance for correctly judging exactly when the normal approximation can be safely and reliably relied upon in real applications with limited sample sizes.
The Central Limit Theorem's key assumptions — finite mean and finite variance — genuinely can fail for certain probability distributions. The Cauchy distribution, notably, has neither a well-defined finite mean nor variance, and sums of independent Cauchy-distributed random variables do not converge to a normal distribution at all (remarkably, they remain distributed as a scaled Cauchy distribution, regardless of how many are summed together) — providing an instructive and mathematically important cautionary counter-example demonstrating that the CLT's stated conditions are not merely a technical formality, but are genuinely necessary for the theorem's conclusion to hold.
The Central Limit Theorem generalises to random vectors (yielding the multivariate normal distribution, see the Normal Distribution entry) and even further to certain classes of random functions and stochastic processes (yielding, for instance, Donsker's theorem, sometimes called the "functional central limit theorem," 1951) — extending this same fundamental underlying mathematical principle of emergent, universal statistical regularity from simple sums of numbers to considerably more complex, high-dimensional, and even infinite-dimensional probabilistic and statistical settings.