Take any simple distribution — a die, a coin, a lopsided spike — and add up n independent copies. The sum has a shape you can compute exactly by convolving the base with itself n times. Standardise it, and as n grows the shape forgets what it started as and becomes the bell. No sampling here: the pmf is exact, and every claim is runnable.
source A. M. Lyapunov (1857–1918), St. Petersburg — the Central Limit Theorem under the Lyapunov condition (1901), completing the line from Chebyshev and Markov (the method of moments). Room: THE NUMBER. Rendered, not quoted.
Pafnuty Chebyshev (1821–1894) gave probability the method of moments: if the moments of a normalised sum converge to the moments of the Gaussian, so does the distribution. His student Andrei Markov made the argument rigorous.
Aleksandr Lyapunov (1857–1918) closed the line in 1901 with a different tool — the characteristic function — and proved the theorem for independent, non-identical summands under a single moment condition. Same St. Petersburg lineage that runs through SONIA’s other THE NUMBER spheres: Chebyshev → Markov → Lyapunov.
The pmf of a sum of two independent variables is the convolution of their pmfs. So the pmf of a sum of n copies is the base pmf convolved with itself n times — exact, deterministic, no random draws.
pS = p ∗ p ∗ … ∗ p (n times)
Probability is conserved at every fold: the convolved pmf still sums to 1. The live witness confirms it for the current base and n.
The sum has mean nμ and variance nσ². Standardise it:
Z = (S − nμ) / √(nσ²)
By construction Z has mean 0 and variance 1 — measured straight off the exact convolved pmf, not assumed. The witness reads the variance of Z for the current n.
Choose a base pmf on a small integer grid and how many copies to add. The engine convolves it n times — no sampling — then standardises. The base shape is what the limit will forget.
The convergence ladder (excess kurtosis and the max gap to the matched Gaussian both fall toward 0 as n rises):
| n | Z var | excess kurtosis | |exc. kurt| | max|pmf−φ| |
|---|
Whatever base you feed, the standardised sum approaches the same standard normal. That universality — the base is forgotten — is the whole content of the theorem. It needs finite variance and the √n scaling; the TAMPER (7) shows what happens without them.
Lyapunov did not need the summands to be identical. He needed a moment ratio to vanish. With δ=1, the Lyapunov ratio for identical bounded copies is
Ln = ρ₃ / (√n · σ³)
where ρ₃=E|X−μ|³. As n grows, Ln→0 — the condition holds — and the CLT follows. The witness reads Ln for the current n and confirms it is falling.
“Averages are Gaussian for any distribution.” Cut. Only with finite variance. The Cauchy distribution has none: the average of n Cauchy draws is again Cauchy — it never narrows, never bells.
“Divide by n to standardise a sum.” Cut. That is the law of large numbers, not the CLT. Dividing by n sends the variance to 0 — a spike, not a bell (see 7).
“You must sample to see the bell.” Cut. Convolution gives the exact pmf. No Monte-Carlo, no noise — the limit is a theorem, not a histogram.
The red team’s move: standardise by n instead of √n — divide the centred sum by nσ rather than √n·σ. It looks like a normalisation, but the variance of the “Z” is now 1/n, which collapses to 0.
Break √n to n and the “standardised” sum does not approach the bell — it collapses to a spike at the mean (variance → 0). The NORMALISE witness (3) turns red and DATA OUT reads COLLAPSED. The CLT genuinely needs √n scaling and finite variance; the attack is real and it is caught.