The normal distribution — the familiar symmetric "bell curve" — describes an extraordinarily wide range of natural and social phenomena, from human heights to measurement errors to exam scores. Its ubiquity is explained by the Central Limit Theorem: whenever a quantity results from adding up many small, independent random effects, the result tends toward a normal distribution, regardless of how those individual effects are themselves distributed.
Measure the heights of a thousand adult men. Most cluster around an average height, with fewer and fewer people as you move toward unusually short or unusually tall. Plot this on a graph and you get a symmetric, bell-shaped curve — the normal distribution. This same characteristic shape appears again and again: exam scores, measurement errors, blood pressure readings, IQ scores, and countless other natural and social phenomena.
A normal distribution is completely determined by just two numbers: the mean (μ) — where the centre of the bell sits — and the standard deviation (σ) — how wide or narrow the bell is. A small σ gives a tall, narrow bell (most values close to the mean); a large σ gives a short, wide bell (values more spread out).
For any normal distribution: about 68% of values fall within 1 standard deviation of the mean, about 95% fall within 2 standard deviations, and about 99.7% fall within 3 standard deviations. This simple rule of thumb is used constantly across science, medicine, and quality control to quickly judge whether an observed value is unusually extreme.
This formula, despite its intimidating appearance, is built from familiar pieces: e (Euler's number), π, and the mean μ and standard deviation σ. The exponent −(x−μ)²/2σ² is what creates the characteristic symmetric, bell-shaped decay away from the mean.
The special case with μ=0 and σ=1 is called the standard normal distribution. Any normal random variable X can be converted ("standardised") to this form via Z = (X−μ)/σ, allowing a single set of reference tables (or software functions) to answer probability questions for any normal distribution whatsoever, regardless of its particular mean and standard deviation.
If X₁, X₂, ..., Xₙ are independent, identically distributed random variables with finite mean μ and finite variance σ², then as n grows large, the distribution of their sum (or average) approaches a normal distribution — regardless of the original distribution of each individual Xᵢ. This remarkable universality is why the normal distribution appears so often in practice: many real-world quantities genuinely are sums of numerous small, roughly independent contributing factors.
Abraham de Moivre (1733) first derived the normal distribution as an approximation to the binomial distribution for large numbers of trials. Pierre-Simon Laplace substantially extended and generalised this work. Carl Friedrich Gauss (1809) independently derived it in the context of astronomical measurement errors, which is why the normal distribution is also commonly called the Gaussian distribution.
Despite its ubiquity, many real-world phenomena are decidedly not normally distributed — income distributions, city population sizes, and word frequencies in language, for example, typically follow highly skewed, "heavy-tailed" distributions (like the log-normal or power-law distributions) rather than the symmetric normal curve. A common statistical error is to assume normality without actually checking it.
The standard proof uses the moment generating function (or characteristic function) of the sum of the Xᵢ. As n increases, the characteristic function of the standardised sum converges pointwise to exactly the characteristic function of the standard normal distribution — and a theorem in analysis (Lévy's continuity theorem) guarantees that this pointwise convergence of characteristic functions implies convergence in distribution of the corresponding random variables themselves. The Lindeberg-Lévy version (the "classical" CLT) requires identically distributed variables; more general versions (Lyapunov's CLT, 1901; the general Lindeberg condition, 1922) relax this to merely independent variables satisfying a technical uniform smallness condition.
The normal distribution generalises naturally to multiple dimensions: the multivariate normal distribution is characterised by a mean vector and a covariance matrix (see the Matrix entry), which fully captures how each pair of variables is correlated with every other. This multivariate generalisation is fundamental to modern statistics, machine learning (particularly Gaussian processes and many probabilistic graphical models), and quantitative finance (particularly for modelling correlated asset returns in portfolio theory).
The specific mathematical form of the normal density — involving e raised to a negative quadratic exponent — can itself be independently derived (following an approach originally due to Herschel and Maxwell, in the separate context of statistical mechanics) from just two natural, physically-motivated assumptions: that the probability distribution of errors in multiple independent dimensions should not depend on the choice of coordinate orientation (rotational symmetry), and that errors in different, independent directions should behave independently of each other. Remarkably, these two comparatively modest and intuitive assumptions alone force the specific Gaussian exponential form — a beautiful example of how strong mathematical structure can emerge from surprisingly weak, natural starting assumptions.
The apparent mathematical elegance and ubiquity of the normal distribution has sometimes led to its inappropriate over-application in statistics and, historically, occasionally in the misapplication of statistical methods to social and biological contexts (including, controversially, in the early 20th-century eugenics movement, which frequently misapplied normal-distribution-based reasoning to human traits, populations, and social outcomes in scientifically and ethically deeply flawed and now thoroughly discredited ways). Modern statistical practice strongly emphasises explicitly checking distributional assumptions (via specific normality tests and diagnostic visual plots) before applying normal-distribution-based statistical methods, rather than simply assuming normality by unexamined default.