The Central Limit Theorem

Why does the average of many small, unrelated fluctuations so often look bell-shaped? Measurement errors, sums of independent bets, thermal noise, sample means in statistics — an astonishing range of aggregated randomness converges to the same universal curve, the normal distribution. This is not a coincidence of nature. It is a theorem about sums of independent random variables. It is one of the most consequential facts in all of mathematics. It justifies Gaussian approximations across statistics, physics, and engineering. It also explains why the normal distribution, rather than any other, deserves its central place in probability. The Central Limit Theorem (CLT) makes this precise. Given independent, identically distributed random variables with finite variance, the properly standardized sum converges in distribution to a standard normal random variable. This holds no matter what the original distribution looked like — Bernoulli, exponential, uniform, or wildly skewed. This universality is the heart of the theorem. The shape of the limit does not depend on the shape of the summands. It only depends on their having finite second moment, combined additively in large numbers. The big picture is this. The Weak Law of Large Numbers tells us that the sample mean converges (in probability) to the true mean. This is a statement about where the mass of the distribution concentrates. The CLT is a second-order refinement. It tells us the shape of the fluctuations around that mean, at the natural 1/n1/\sqrt{n} scale. Together these two theorems form the backbone of asymptotic statistics. The law of large numbers says convergence happens. The CLT says how it happens. It lets us compute confidence intervals, p-values, and error bars. This article gives the precise statement of the classical (Lindeberg–Lévy) CLT. It provides a complete proof using Characteristic Functions and Lévy's continuity theorem. It works through a derivation of the normal approximation to the binomial. It gives a second example involving exponential waiting times. It carefully discusses why each hypothesis — independence, identical distribution, finite variance — cannot be dropped. It closes by placing the CLT in its historical context. Finally, it points toward its quantitative refinements, the Concentration Inequalities that measure how fast this convergence actually occurs.

Key takeaways
  • The CLT says standardized sums of i.i.d. random variables with finite variance converge in distribution to N(0,1)N(0,1), regardless of the original distribution's shape.
  • It is a statement about convergence in distribution, the weakest of the standard Modes of Convergence of Random Variables, and no stronger conclusion holds in general.
  • The clean proof route is via Characteristic Functions: show the characteristic function of the standardized sum converges pointwise to e−t2/2e^{-t^2/2}, then invoke Lévy's continuity theorem.
  • Finite variance is essential — the theorem fails for heavy-tailed distributions like the Cauchy, whose stable limit is not Gaussian.
  • The rate of this convergence (how large nn must be) is governed by higher moments via the Berry–Esseen theorem, foreshadowing Concentration Inequalities.
  • Key idea: additive combination of many small independent effects, after centering and rescaling by n\sqrt{n}, always erases the details of the original distribution and leaves only the normal shape.

Where this lives today

The CLT underlies essentially all of classical inferential statistics: confidence intervals for a sample mean, the validity of zz-tests and (via a further Studentization step) tt-tests, and the justification for treating estimators as approximately normal in maximum likelihood theory. In statistical mechanics, it explains why macroscopic thermodynamic quantities — sums of enormous numbers of microscopic degrees of freedom — exhibit Gaussian fluctuations (e.g. the Maxwell–Boltzmann velocity distribution derivation via sums of independent momentum components). In signal processing and communications engineering, thermal (Johnson–Nyquist) noise is modeled as Gaussian precisely because it arises from the superposition of vast numbers of independent electron collisions. In finance, aggregated returns over many independent trading periods are frequently modeled as approximately normal (with well-known caveats about heavy tails, directly connected to the Cauchy-type counterexample above). In computer science, randomized algorithms and hashing analyses use CLT-style Gaussian approximations to bound the typical behavior of sums of independent random choices, and Monte Carlo methods rely on the CLT to construct error bars for simulation estimates. The theorem's quantitative refinement, controlling how quickly ZnZ_n's distribution approaches normal and how large deviations behave non-asymptotically, is the subject of Concentration Inequalities, which underlie modern high-dimensional statistics, learning theory, and randomized algorithm analysis.

Ask the tutor: How does the Berry–Esseen bound quantify the exact rate at which ZnZ_n's distribution approaches the normal, and how does that connect to modern concentration inequalities?

DefinitionConvergence in distribution and the standardized sum

Let X1,X2,…X_1, X_2, \dots be a sequence of real-valued random variables (not necessarily on the same probability space as each other in the abstract definition) with distribution functions F1,F2,…F_1, F_2, \dots. We say XnX_n converges in distribution to a random variable XX with distribution function FF, written

Xn→dX,X_n \xrightarrow{d} X,

if Fn(x)→F(x)F_n(x) \to F(x) at every point xx where FF is continuous. Now let X1,X2,…X_1, X_2, \dots be i.i.d. random variables with common finite mean μ=E[X1]\mu = E[X_1] and common finite variance σ2=Var(X1)∈(0,∞)\sigma^2 = \mathrm{Var}(X_1) \in (0,\infty). Define the partial sum Sn=X1+⋯+XnS_n = X_1 + \cdots + X_n and the standardized sum

Zn=Sn−nμσn=1n∑i=1nXi−μσ.Z_n = \frac{S_n - n\mu}{\sigma\sqrt{n}} = \frac{1}{\sqrt{n}}\sum_{i=1}^n \frac{X_i - \mu}{\sigma}.

The Central Limit Theorem asserts that Zn→dZZ_n \xrightarrow{d} Z, where Z∼N(0,1)Z \sim N(0,1), i.e.

lim⁡n→∞P(Zn≤z)=Φ(z)=∫−∞z12πe−u2/2 dufor every z∈R.\lim_{n\to\infty} P(Z_n \le z) = \Phi(z) = \int_{-\infty}^z \frac{1}{\sqrt{2\pi}} e^{-u^2/2}\, du \qquad \text{for every } z \in \mathbb{R}.

Intuition

Think of Xi−μX_i - \mu as a small independent "kick," positive or negative, with average size σ\sigma. Adding nn such kicks gives a sum whose typical size grows like n\sqrt{n} (this is exactly the additivity of variance for independent variables — Variance of Sums of Random Variables — since Var(Sn)=nσ2\mathrm{Var}(S_n) = n\sigma^2). Dividing by σn\sigma\sqrt{n} rescales the fluctuations to a fixed, non-degenerate scale. What's remarkable is that after this rescaling, any fine details of the individual kicks' distribution — skewness, discreteness, boundedness — wash out in the limit, leaving only the two numbers that survived rescaling: mean 00 and variance 11. A second way to see why the Gaussian in particular emerges: the normal distribution is the unique (up to location-scale) probability law that is a fixed point of the "convolve and rescale" operation — if Y1,Y2Y_1, Y_2 are independent standard normal, then (Y1+Y2)/2(Y_1+Y_2)/\sqrt{2} is again standard normal. The CLT says that every finite-variance distribution flows toward this fixed point under repeated convolution and rescaling, much as many dynamical systems flow toward an attracting fixed point regardless of initial condition.

Figure
Histograms of standardized sums of Bernoulli(0.3) trials converging to the standard normal density as nn grows
n = 1 (skewed) n = 10 n = 50 (≈ bell curve) 0 1 φ(z) = e^{-z²/2}/√(2π) Standardized sum Z_n = (S_n − nμ)/(σ√n); shape converges to N(0,1) regardless of the Bernoulli origin.
TheoremLindeberg–Lévy Central Limit Theorem

Let X1,X2,X3,…X_1, X_2, X_3, \dots be independent and identically distributed real-valued random variables with

E[X1]=μ∈R,Var(X1)=σ2∈(0,∞).E[X_1] = \mu \in \mathbb{R}, \qquad \mathrm{Var}(X_1) = \sigma^2 \in (0,\infty).

Let Sn=∑i=1nXiS_n = \sum_{i=1}^n X_i and define the standardized sum

Zn=Sn−nμσn.Z_n = \frac{S_n - n\mu}{\sigma\sqrt{n}}.

Then, as n→∞n \to \infty,

Zn→dZ,Z∼N(0,1),Z_n \xrightarrow{d} Z, \qquad Z \sim N(0,1),

that is, for every z∈Rz \in \mathbb{R},

lim⁡n→∞P(Zn≤z)=Φ(z).\lim_{n\to\infty} P(Z_n \le z) = \Phi(z).
LemmaLévy's Continuity Theorem — used, not proved here

Let YnY_n be a sequence of random variables with characteristic functions φYn(t)=E[eitYn]\varphi_{Y_n}(t) = E[e^{itY_n}]. If φYn(t)→φ(t)\varphi_{Y_n}(t) \to \varphi(t) for every t∈Rt \in \mathbb{R}, and φ\varphi is continuous at t=0t=0, then φ\varphi is the characteristic function of some random variable YY, and Yn→dYY_n \xrightarrow{d} Y. (This is a standard prerequisite result about <a class="wiki-link" href="/wiki/math-probstat-characteristic-functions" data-track="wiki_article.body.wiki_link">Characteristic Functions</a>: it is the analytic tool converting pointwise convergence of characteristic functions into convergence in distribution, via Fourier inversion and tightness arguments. We use it as a black box, exactly as one uses it to build the CLT proof; it does not itself depend on the CLT and so no circularity arises.)

Proof

We prove the theorem by computing the limit of the characteristic function of ZnZ_n and identifying it as the characteristic function of N(0,1)N(0,1), then invoking Lévy's Continuity Theorem above.

Step 1: Reduce to standardized i.i.d. variables. Let Yi=(Xi−μ)/σY_i = (X_i - \mu)/\sigma. Since X1,…,XnX_1, \dots, X_n are i.i.d., so are Y1,…,YnY_1, \dots, Y_n, and

E[Y1]=0,Var(Y1)=E[Y12]=1.E[Y_1] = 0, \qquad \mathrm{Var}(Y_1) = E[Y_1^2] = 1.

By construction,

Zn=1n∑i=1nYi.Z_n = \frac{1}{\sqrt{n}} \sum_{i=1}^n Y_i.

So it suffices to prove the theorem for mean-zero, variance-one i.i.d. summands; write Y=Y1Y = Y_1 and let φ(t)=E[eitY]\varphi(t) = E[e^{itY}] be its characteristic function.

Step 2: Local expansion of φ\varphi near t=0t=0. Since E[Y2]=1<∞E[Y^2] = 1 < \infty, the characteristic function φ\varphi is twice continuously differentiable (a standard fact from the theory of Characteristic Functions: existence of the kk-th moment guarantees φ∈Ck\varphi \in C^k with φ(k)(0)=ikE[Yk]\varphi^{(k)}(0) = i^k E[Y^k]), with

φ(0)=1,φ′(0)=iE[Y]=0,φ′′(0)=i2E[Y2]=−1.\varphi(0) = 1, \qquad \varphi'(0) = iE[Y] = 0, \qquad \varphi''(0) = i^2 E[Y^2] = -1.

By Taylor's theorem with Peano remainder (valid since φ′′\varphi'' exists and is continuous at 00),

φ(t)=φ(0)+φ′(0)t+12φ′′(0)t2+o(t2)=1−t22+o(t2)as t→0.\varphi(t) = \varphi(0) + \varphi'(0)t + \tfrac{1}{2}\varphi''(0)t^2 + o(t^2) = 1 - \frac{t^2}{2} + o(t^2) \qquad \text{as } t \to 0.

Step 3: Characteristic function of the standardized sum. Because Y1,…,YnY_1, \dots, Y_n are independent, the characteristic function of a sum of independent random variables is the product of their characteristic functions (a direct consequence of independence and the multiplicativity of expectation for products of independent functions of independent variables):

φ∑i=1nYi(t)=∏i=1nE ⁣[eitYi]=φ(t)n.\varphi_{\sum_{i=1}^n Y_i}(t) = \prod_{i=1}^n E\!\left[e^{itY_i}\right] = \varphi(t)^n.

Also, for a random variable WW and constant aa, φaW(t)=φW(at)\varphi_{aW}(t) = \varphi_W(at) directly from the definition E[eit(aW)]=E[ei(at)W]E[e^{it(aW)}] = E[e^{i(at)W}]. Applying this with a=1/na = 1/\sqrt n to W=∑iYiW = \sum_i Y_i:

φZn(t)=φ ⁣(tn)n.\varphi_{Z_n}(t) = \varphi\!\left(\frac{t}{\sqrt n}\right)^{n}.

Step 4: Take the limit as n→∞n \to \infty, for fixed tt. Fix t∈Rt \in \mathbb{R}. As n→∞n \to \infty, t/n→0t/\sqrt n \to 0, so by Step 2,

φ ⁣(tn)=1−12⋅t2n+o ⁣(1n),\varphi\!\left(\frac{t}{\sqrt n}\right) = 1 - \frac{1}{2}\cdot\frac{t^2}{n} + o\!\left(\frac{1}{n}\right),

where the o(1/n)o(1/n) term is a function of nn (depending on tt) that, multiplied by nn, tends to 00 as n→∞n \to \infty. Write this as

φ ⁣(tn)=1+cn(t)n,where cn(t):=−t22+n⋅o ⁣(1n)⟶−t22 as n→∞.\varphi\!\left(\frac{t}{\sqrt n}\right) = 1 + \frac{c_n(t)}{n}, \qquad \text{where } c_n(t) := -\frac{t^2}{2} + n\cdot o\!\left(\frac1n\right) \longrightarrow -\frac{t^2}{2} \text{ as } n\to\infty.

We now use the elementary complex-analytic lemma: if cn→c∈Cc_n \to c \in \mathbb{C} as n→∞n \to \infty, then (1+cnn)n→ec\left(1 + \frac{c_n}{n}\right)^n \to e^{c}. To justify this lemma: write (1+cnn)n=exp⁡(nlog⁡(1+cnn))\left(1+\frac{c_n}{n}\right)^n = \exp\left(n\log\left(1+\frac{c_n}{n}\right)\right) for nn large enough that ∣cn/n∣<1/2|c_n/n| < 1/2 (possible since cnc_n is bounded), using the principal branch of log⁡\log, which is continuous and satisfies log⁡(1+w)=w+O(w2)\log(1+w) = w + O(w^2) for ∣w∣≤1/2|w| \le 1/2. Then

nlog⁡(1+cnn)=n(cnn+O ⁣(cn2n2))=cn+O ⁣(cn2n)⟶c,n\log\left(1 + \frac{c_n}{n}\right) = n\left(\frac{c_n}{n} + O\!\left(\frac{c_n^2}{n^2}\right)\right) = c_n + O\!\left(\frac{c_n^2}{n}\right) \longrightarrow c,

since cn→cc_n \to c implies cn2/n→0c_n^2/n \to 0 (as cnc_n is bounded). Continuity of exp⁡\exp then gives (1+cnn)n→ec\left(1+\frac{c_n}{n}\right)^n \to e^c, proving the lemma. Applying this with cn=cn(t)→−t2/2c_n = c_n(t) \to -t^2/2:

φZn(t)=(1+cn(t)n)n⟶e−t2/2.\varphi_{Z_n}(t) = \left(1 + \frac{c_n(t)}{n}\right)^n \longrightarrow e^{-t^2/2}.

Step 5: Identify the limit and conclude. The function φ(t)=e−t2/2\varphi(t) = e^{-t^2/2} is precisely the characteristic function of the standard normal distribution N(0,1)N(0,1) (a standard computation via completing the square in the Gaussian integral, part of the basic theory of Characteristic Functions). It is continuous everywhere, in particular at t=0t=0. We have shown

φZn(t)⟶e−t2/2for every fixed t∈R.\varphi_{Z_n}(t) \longrightarrow e^{-t^2/2} \qquad \text{for every fixed } t \in \mathbb{R}.

By Lévy's Continuity Theorem, this pointwise convergence of characteristic functions to a function continuous at 00 implies

Zn→dZ,Z∼N(0,1).Z_n \xrightarrow{d} Z, \qquad Z \sim N(0,1).

This is exactly the claimed conclusion. ■\blacksquare

∎

The proof technique here is the method of characteristic functions: reduce a distributional limit statement to a pointwise limit of an analytic object (the characteristic function), where algebraic operations (products, rescaling) are transparent, then transfer back via a continuity theorem. This is the standard and cleanest route to the CLT; a direct argument with distribution functions or densities is far more delicate because convolution of nn arbitrary distributions has no simple closed form, whereas the characteristic function turns convolution into multiplication.

Worked examples

Example 1: De Moivre–Laplace as a special case — binomial counts of a biased coin. Let X1,X2,…X_1, X_2, \dots be i.i.d. Bernoulli(p)\mathrm{Bernoulli}(p) with p=0.3p = 0.3 (a coin strongly biased toward tails, chosen asymmetrically on purpose). Then

μ=E[X1]=p=0.3,σ2=Var(X1)=p(1−p)=0.3×0.7=0.21.\mu = E[X_1] = p = 0.3, \qquad \sigma^2 = \mathrm{Var}(X_1) = p(1-p) = 0.3 \times 0.7 = 0.21.

Let Sn=X1+⋯+XnS_n = X_1 + \cdots + X_n be the number of heads in nn tosses, so Sn∼Binomial(n,0.3)S_n \sim \mathrm{Binomial}(n,0.3). Take n=400n = 400. Then

nμ=400×0.3=120,σn=0.21×400=84≈9.1652.n\mu = 400 \times 0.3 = 120, \qquad \sigma\sqrt n = \sqrt{0.21 \times 400} = \sqrt{84} \approx 9.1652.

We estimate P(110≤S400≤135)P(110 \le S_{400} \le 135), i.e., the probability the number of heads lands in a specific asymmetric window around the mean. Standardizing (with continuity correction, since SnS_n is integer-valued and we are approximating a discrete distribution by a continuous one):

P(110≤S400≤135)=P(109.5≤S400≤135.5)≈P ⁣(109.5−1209.1652≤Z≤135.5−1209.1652).P(110 \le S_{400} \le 135) = P(109.5 \le S_{400} \le 135.5) \approx P\!\left(\frac{109.5 - 120}{9.1652} \le Z \le \frac{135.5-120}{9.1652}\right).

Compute the two standardized bounds:

109.5−1209.1652=−10.59.1652≈−1.1457,135.5−1209.1652=15.59.1652≈1.6912.\frac{109.5 - 120}{9.1652} = \frac{-10.5}{9.1652} \approx -1.1457, \qquad \frac{135.5-120}{9.1652} = \frac{15.5}{9.1652} \approx 1.6912.

So the CLT approximation gives

P(110≤S400≤135)≈Φ(1.6912)−Φ(−1.1457).P(110 \le S_{400} \le 135) \approx \Phi(1.6912) - \Phi(-1.1457).

Using standard normal table values Φ(1.69)≈0.9545\Phi(1.69) \approx 0.9545 and Φ(−1.15)≈1−Φ(1.15)≈1−0.8749=0.1251\Phi(-1.15) \approx 1 - \Phi(1.15) \approx 1 - 0.8749 = 0.1251:

P(110≤S400≤135)≈0.9545−0.1251=0.8294.P(110 \le S_{400} \le 135) \approx 0.9545 - 0.1251 = 0.8294.

This illustrates the theorem concretely: even though X1X_1 is a highly asymmetric 0/10/1 variable (mean 0.30.3, not 0.50.5), for n=400n=400 the sum's fluctuations around its mean are already well-approximated by a normal curve, letting us compute a probability about a discrete combinatorial quantity ((400k)0.3k0.7400−k\binom{400}{k}0.3^k 0.7^{400-k} summed over kk) using only a table of Φ\Phi.

Example 2: Sum of i.i.d. Exponential waiting times — a genuinely non-symmetric, skewed distribution. Let X1,X2,…X_1, X_2, \dots be i.i.d. Exponential(λ)\mathrm{Exponential}(\lambda) with rate λ=2\lambda = 2 (so mean waiting time 1/λ=0.51/\lambda = 0.5), modeling, e.g., inter-arrival times of customers in time units where the average gap is half a unit. The exponential density λe−λx1x>0\lambda e^{-\lambda x}\mathbf 1_{x>0} is strongly right-skewed — nothing like a bell curve — which makes it a good stress test of the CLT's universality claim. We have

μ=1λ=0.5,σ2=1λ2=0.25,σ=0.5.\mu = \frac{1}{\lambda} = 0.5, \qquad \sigma^2 = \frac{1}{\lambda^2} = 0.25, \qquad \sigma = 0.5.

Let SnS_n be the total waiting time for n=100n = 100 arrivals; note Sn∼Gamma(n,λ)S_n \sim \mathrm{Gamma}(n, \lambda) exactly (a fact independent of the CLT, from the exponential's memorylessness and additivity under convolution). We ask: what is P(S100>55)P(S_{100} > 55), the probability the 100th arrival is unusually late? Standardize:

nμ=100×0.5=50,σn=0.5100=5.n\mu = 100 \times 0.5 = 50, \qquad \sigma\sqrt n = 0.5\sqrt{100} = 5.
P(S100>55)≈P ⁣(Z>55−505)=P(Z>1)=1−Φ(1)≈1−0.8413=0.1587.P(S_{100} > 55) \approx P\!\left(Z > \frac{55 - 50}{5}\right) = P(Z > 1) = 1 - \Phi(1) \approx 1 - 0.8413 = 0.1587.

For comparison, the exact Gamma tail probability (computable via the Poisson-process identity P(S100>55)=P(N(55)≤99)P(S_{100} > 55) = P(N(55) \le 99) where N(55)∼Poisson(λ⋅55=110)N(55) \sim \mathrm{Poisson}(\lambda \cdot 55 = 110)) is approximately 0.17400.1740 by direct computation — reasonably close to the CLT's 0.15870.1587 given that n=100n=100 is only moderately large and the exponential distribution has substantial skewness (skewness coefficient 22, versus 00 for the normal). This gap is precisely the kind of higher-order correction quantified by the Berry–Esseen bound (of order 1/n1/\sqrt n here), which measures the rate at which the CLT approximation improves — a preview of the finer, non-asymptotic viewpoint of Concentration Inequalities.

CounterexampleWhy finite variance cannot be dropped: i.i.d. Cauchy summands

Drop the hypothesis σ2<∞\sigma^2 < \infty and keep everything else (i.i.d., "nice" continuous density). Let X1,X2,…X_1, X_2, \dots be i.i.d. standard Cauchy random variables, with density

f(x)=1π(1+x2),x∈R.f(x) = \frac{1}{\pi(1+x^2)}, \qquad x \in \mathbb{R}.

The Cauchy distribution has E[∣X1∣]=∞E[|X_1|] = \infty (the integral ∫∣x∣f(x) dx\int |x| f(x)\,dx diverges logarithmically), so neither the mean nor the variance exists; there is no μ,σ\mu, \sigma with which to even form the standardized sum ZnZ_n of the theorem. But we can still ask what happens to the unstandardized average. The Cauchy characteristic function is φ(t)=e−∣t∣\varphi(t) = e^{-|t|} (a classical fact, computed by contour integration). By independence,

φSn(t)=φ(t)n=e−n∣t∣,soφSn/n(t)=φSn ⁣(tn)=e−n∣tn∣=e−∣t∣.\varphi_{S_n}(t) = \varphi(t)^n = e^{-n|t|}, \qquad \text{so} \qquad \varphi_{S_n/n}(t) = \varphi_{S_n}\!\left(\frac{t}{n}\right) = e^{-n\left|\frac{t}{n}\right|} = e^{-|t|}.

Remarkably, Sn/nS_n/n has exactly the same distribution as a single X1X_1 for every nn — the sample mean of nn Cauchy variables is itself standard Cauchy, with no shrinking of spread at all as n→∞n\to\infty. This shows two failures at once. First, there is no n\sqrt n scaling that produces a non-degenerate normal limit: the natural CLT-style normalization Sn/nS_n/\sqrt n has characteristic function e−n∣t∣→0e^{-\sqrt n |t|} \to 0 for every t≠0t\neq 0, which is not continuous at t=0t=0 (it does not converge to a valid characteristic function at all — the limiting "distribution" would need infinite mass at ±∞\pm\infty). Second, the correctly-scaled quantity Sn/nS_n/n does converge in distribution — trivially, to Cauchy itself, not to any Gaussian. This is a genuine instance of a different stable limit law (Cauchy is 11-stable) replacing the Gaussian precisely because the finite-variance hypothesis fails. It demonstrates that "i.i.d. plus nice continuous density" is not enough; the theorem's finite-second-moment hypothesis is load-bearing, not decorative.

Scope, edge cases & common traps

SituationWhat happensWhy
I.i.d., finite varianceCLT holds exactly as statedcore theorem
I.i.d., infinite variance (e.g. Cauchy, Pareto with tail index ≤2\le 2)No Gaussian limit; possibly a different stable law, or no limit under n\sqrt n scaling at allTaylor expansion of φ\varphi to order t2t^2 fails
Independent but not identically distributed, uniformly bounded variancesCLT can still hold under the Lindeberg condition (a strictly more general theorem)different, more delicate proof; not covered by Lindeberg–Lévy as stated
Dependent (e.g. Markov chain) sumsCLT can still hold with a different asymptotic variance (sum of covariances), but the classical statement does not apply directlyindependence used essentially in Step 3 of the proof
Finite nn (no limit taken)ZnZ_n is not exactly normal; only the limit isCLT is an asymptotic, not exact, statement
Discrete summands (e.g. Binomial)Continuity correction improves finite-nn accuracyapproximating a discrete cdf by a continuous one

The mathematician behind it

Origins · 1 minPierre-Simon Laplace and how this was discovered
PL
Pierre-Simon Laplace
1749–1827 · Paris, Bureau des Longitudes and the École Militaire

Laplace was a central figure of the French Enlightenment's mathematical physics and probability, whose Théorie analytique des probabilités (1812) synthesized and vastly extended the theory of errors, generating functions, and limit theorems that would later be recognized as an early general form of the Central Limit Theorem. Working from problems in astronomy, geodesy, and the theory of observational error, Laplace sought to justify why averages of many independent measurement errors cluster in a bell-shaped pattern — the same question that motivates the CLT today. His techniques, based on generating functions and asymptotic expansions, anticipated the characteristic-function method used in the modern rigorous proof.

History
- 1733 — Abraham de Moivre proves the normal approximation to the binomial distribution, the first special case of the CLT.
- 1810 — Pierre-Simon Laplace extends de Moivre's result and gives an early general limit theorem for sums of independent errors, in the spirit of the CLT.
- 1901 — Aleksandr Lyapunov gives the first fully rigorous proof of a general CLT (allowing non-identical distributions), using characteristic functions and an explicit error bound (the Lyapunov condition).
- 1920–1922 — Jarl Waloddi Lindeberg and Paul Lévy independently establish the sharp, essentially minimal sufficient condition (the Lindeberg condition) for the CLT to hold for independent, not-necessarily-identical summands.
- 1941–1942 — Andrew C. Berry and Carl-Gustav Esseen prove a quantitative rate of convergence (the Berry–Esseen theorem), turning the CLT's qualitative limit into an explicit non-asymptotic bound.
Does the CLT require the underlying distribution to be continuous or symmetric?

No. The theorem as stated only requires the summands to be independent, identically distributed, and to have finite mean and finite positive variance. The distribution can be discrete (as in the Bernoulli example), highly skewed (as in the Exponential example), or have any shape whatsoever. The proof works entirely through the first two moments captured in the Taylor expansion of the characteristic function near t=0t=0; no other feature of the distribution — its symmetry, its support, its smoothness — enters the limiting characteristic function e−t2/2e^{-t^2/2}. This universality is exactly what makes the CLT powerful: only two numbers, μ\mu and σ2\sigma^2, survive the limiting process.

How large does n need to be for the normal approximation to be "good enough" in practice?

There is no universal threshold; it depends on how far the standardized third and fourth moments (skewness and kurtosis) of X1X_1 are from those of the normal distribution. A common rule of thumb is n≥30n \ge 30 for mildly skewed distributions, but the second worked example shows a heavily skewed exponential still has a visible gap between the CLT approximation and the exact value at n=100n=100. The quantitative version of this question is answered by the Berry–Esseen theorem, which bounds sup⁡z∣P(Zn≤z)−Φ(z)∣\sup_z |P(Z_n\le z)-\Phi(z)| by a constant times E∣X1−μ∣3/(σ3n)E|X_1-\mu|^3/(\sigma^3\sqrt n), giving an explicit, distribution-dependent rate of order n−1/2n^{-1/2}.

Is convergence in distribution the strongest conclusion one can hope for here?

Yes, in general. Convergence in distribution is inherently the right — and typically the only available — mode of convergence for the CLT, because ZnZ_n for different nn are built from overlapping but growing sums and do not settle down pointwise or in probability to a fixed random variable; indeed ZnZ_n and Zn+1Z_{n+1} are highly correlated but their difference does not vanish. One should not expect almost sure or in-probability convergence of ZnZ_n to a single limiting random variable; the CLT is fundamentally a statement about the limiting shape of a sequence of distributions, not about pathwise behavior of the sequence of random variables itself.

How does the CLT relate to the Weak Law of Large Numbers?

The Weak Law of Large Numbers states that Xˉn=Sn/n→μ\bar X_n = S_n/n \to \mu in probability, a first-order statement about where the sample mean concentrates as n→∞n \to \infty; it says nothing about the rate or shape of the remaining fluctuation. The CLT is the natural second-order companion: it shows that the fluctuation Xˉn−μ\bar X_n - \mu, magnified by the factor n/σ\sqrt n /\sigma, converges in distribution to a fixed, non-degenerate limit rather than to zero or to something distribution-dependent. In this sense the CLT sharpens the WLLN by revealing the exact scale (1/n1/\sqrt n) and exact shape (Gaussian) of the law's convergence, and the WLLN can in fact be recovered as an immediate corollary once one knows ZnZ_n converges in distribution to a finite random variable, since Sn/n−μ=σZn/n→0S_n/n - \mu = \sigma Z_n /\sqrt n \to 0 in probability by Slutsky's theorem.

Can the CLT be extended to random vectors or to sums where the summands are not identically distributed?

Yes, on both counts, though these are separate theorems beyond Lindeberg–Lévy as stated here. The multivariate CLT replaces the scalar variance σ2\sigma^2 by a covariance matrix and shows standardized sums of i.i.d. random vectors converge in distribution to a multivariate normal, proved by an entirely analogous characteristic-function argument using the Cramér–Wold device to reduce to the one-dimensional case. For independent but non-identically-distributed summands, the Lindeberg–Feller CLT gives a sharp sufficient (and, in a suitable sense, necessary) condition — the Lindeberg condition — controlling the contribution of each summand's tail, generalizing the classical result proven here.

Why does the proof use characteristic functions rather than moment generating functions?

Characteristic functions φX(t)=E[eitX]\varphi_X(t)=E[e^{itX}] always exist for every random variable, since ∣eitX∣=1|e^{itX}|=1 is bounded, whereas moment generating functions E[etX]E[e^{tX}] can fail to exist (be infinite) for heavy-tailed distributions such as the Cauchy or even for distributions with only some finite moments. Working with characteristic functions lets the proof apply to the widest possible class of finite-variance distributions without additional integrability assumptions. Characteristic functions also enjoy the same key algebraic property needed in the proof — turning sums of independent variables into products — while Lévy's continuity theorem gives a clean, general bridge back to convergence in distribution.

Before this

Where this leads