Add up enough independent random nudges — measurement errors in a lab, the heights of a crowd, the daily ups and downs of a stock, the votes in a poll — and no matter what the individual nudges look like, their (properly scaled) total draws the same bell-shaped curve. This uncanny universality is the Central Limit Theorem, the reason the normal distribution turns up everywhere in science. It is easy to state and, with the right tool, astonishingly clean to prove. That tool is the characteristic function.

For a real random variable X with law \mu_X, the characteristic function is the expectation of a rotating unit phasor:

Read the middle expression as an expectation, the right one as a Lebesgue integral against the distribution of X. In fact \varphi_X is nothing but the Fourier transform of the measure \mu_X (with the probabilist's sign and normalisation convention). Every fact you know about Fourier transforms — that they turn convolution into multiplication, that they are invertible, that they encode smoothness and decay — is about to earn its keep in probability.

The properties that make it useful

A handful of properties follow straight from the definition and the linearity of the integral. Each is small; together they are everything we need.

• Normalisation: \varphi_X(0) = \mathbb{E}[e^{0}] = \mathbb{E}[1] = 1.
• Bounded: |\varphi_X(t)| \le 1 for all t, with equality at t = 0.
• Hermitian: \varphi_X(-t) = \overline{\varphi_X(t)}; if X is symmetric about 0 then \varphi_X is real-valued.
• Uniformly continuous: \varphi_X is uniformly continuous on \mathbb{R} (dominated convergence, with the constant dominator 1) — no matter how rough the underlying distribution.
• Affine maps: for constants a, b, \varphi_{aX + b}(t) = \mathbb{E}\!\left[e^{it(aX + b)}\right] = e^{itb}\,\varphi_X(at).

And now the property that does all the heavy lifting — the reason we changed coordinates into Fourier space at all:

By induction, for independent X_1, \dots, X_n we get \varphi_{X_1 + \cdots + X_n}(t) = \prod_{k=1}^{n} \varphi_{X_k}(t), and if they are identically distributed with common characteristic function \varphi, this collapses to a single power:

Uniqueness, inversion, and moments

A change of coordinates is only useful if you can change back. Two facts guarantee it. First, the uniqueness / inversion theorem: the characteristic function determines the law completely — if \varphi_X = \varphi_Y as functions, then X and Y have the same distribution. There is even an explicit inversion formula recovering the distribution function from \varphi (a Fourier inversion in disguise). So passing to \varphi loses nothing: it is a faithful re-encoding of the whole distribution.

Second, moments live in the derivatives at zero. Differentiating under the integral sign (legal whenever the relevant moment is finite) pulls down a factor of iX each time:

In particular \varphi_X'(0) = i\,\mathbb{E}[X] and \varphi_X''(0) = i^{2}\,\mathbb{E}[X^{2}] = -\mathbb{E}[X^{2}]. Feeding these into a Taylor expansion about t = 0 gives the local shape we will need. If X has mean \mu and variance \sigma^{2} (so \mathbb{E}[X^{2}] = \sigma^{2} + \mu^{2}), then as t \to 0,

Read the coefficients: the constant term is 1 (normalisation), the linear term carries the mean, the quadratic term carries the second moment. The entire behaviour near the origin is governed by just the first two moments — which is precisely why, in the limit theorem to come, only the mean and variance survive.

Lévy's continuity theorem — the engine

Everything so far lets us compute with characteristic functions. The last ingredient lets us take limits, and it is the crank that turns algebra into a theorem about distributions.

This is extraordinary leverage. Convergence in distribution is a statement about cumulative distribution functions matching at every continuity point — awkward to check directly. Lévy says you may instead check pointwise convergence of a single function, \varphi_{X_n} \to \psi, plus a mild continuity condition at the origin (which rules out probability mass escaping to infinity). Prove that one function converges, and the whole distribution converges. That is the engine; here is what it drives.

The Central Limit Theorem, proved in four lines

Let X_1, X_2, \dots be independent and identically distributed, with finite mean \mu = \mathbb{E}[X_i] and finite variance 0 < \sigma^{2} < \infty. Form the sum S_n = X_1 + \cdots + X_n and standardise it — subtract its mean n\mu and divide by its standard deviation \sigma\sqrt{n}:

Now the proof. Recentre by writing Y_i = X_i - \mu, so \mathbb{E}[Y_i] = 0 and \mathbb{E}[Y_i^{2}] = \sigma^{2}, with common characteristic function \varphi_Y. Since Z_n = \sum_i Y_i / (\sigma\sqrt{n}) is a scaled sum of independent copies, the affine rule and the product rule combine into a single power:

Expand \varphi_Y near 0 using the moment expansion above. Because Y has mean 0 and variance \sigma^{2}, the linear term vanishes and \varphi_Y(s) = 1 - \tfrac{1}{2}\sigma^{2}s^{2} + o(s^{2}). Substitute s = t/(\sigma\sqrt{n}), so s^{2} = t^{2}/(\sigma^{2} n) and the \sigma^{2} cancels beautifully:

That last step is the classic limit (1 + c/n)^{n} \to e^{c} with c = -t^{2}/2. And e^{-t^{2}/2} is exactly the characteristic function of the standard normal N(0,1) (a Gaussian is its own Fourier transform). The limit function is continuous at 0, so Lévy's continuity theorem converts this pointwise convergence into convergence in distribution:

Notice what the argument used and what it ignored. It used only the mean and variance — the first two Taylor coefficients — and it never touched the shape of the individual X_i. Bernoulli, uniform, exponential, dice, anything with a finite variance: the third-and-higher moments are swept into the o(1/n) and forgotten. That forgetting is the universality of the bell curve. The chart below is this proof, drawn: it plots \varphi_{Z_n}(t) for a sum of uniforms and watches it flatten onto e^{-t^{2}/2} as you raise n.

At n = 1 the solid curve is a single uniform's characteristic function — a \operatorname{sinc} that dips below the axis and oscillates, visibly not a bell. Push n up and each raise-to-the-power irons the oscillations flat, pinning the curve onto the dashed Gaussian. By the time the two are indistinguishable, so are the distributions.