Feed any distribution — you need not know its shape, only its mean μ and variance σ² — and the fraction of mass more than kσ from the centre can never exceed 1/k². One inequality, no assumptions. And the same P. L. Chebyshev built the polynomials that oscillate flattest of all, and proved a prime always sits in (n, 2n]. Every claim here is runnable.
source Pafnuty Lvovich Chebyshev (1821–1894), St. Petersburg — Chebyshev’s inequality, Chebyshev polynomials Tn, and Bertrand’s postulate. Room: THE NUMBER. Rendered, not quoted.
Pafnuty Lvovich Chebyshev (1821–1894) founded the St. Petersburg school of mathematics. His inequality gives probability an unconditional grip on any distribution with a finite variance.
His students carried the line forward: Andrei Markov (the Markov chains & the Markov property) and Aleksandr Lyapunov (stability, the central limit theorem) — the same lineage that runs through SONIA’s other THE NUMBER spheres. Chebyshev’s inequality is the seed of the weak law of large numbers.
For any random variable X with mean μ and variance σ² > 0, and any k > 0:
P( |X − μ| ≥ kσ ) ≤ 1/k²
No shape assumed — not normal, not symmetric, not anything. k=2 → ≤ 0.25; k=3 → ≤ 0.1111. The bound only says something for k > 1 (at k=1 it reads ≤1, no information). It is tight: a two-point distribution hits it exactly.
Chebyshev polynomials are defined by Tn(cos t) = cos(n·t) and grow from a two-term recurrence:
Tn+1(x) = 2x·Tn(x) − Tn−1(x)
T₀=1, T₁=x, T₂=2x²−1, T₃=4x³−3x. The minimax property: scaled to monic (Tn/2n−1) it has the smallest sup-norm on [−1,1] of any monic degree-n polynomial — exactly 21−n. It equioscillates between ±1.
Pick a distribution and draw a sample from a fixed-seed generator. You do not tell the engine the shape — it estimates μ and σ from the draws, then measures how much mass strays past kσ.
For each k, the measured tail fraction beyond kσ sits under the guarantee. Blue bar = empirical; the red mark is the bound:
| k | empirical P(|X−μ|≥kσ) | bound | tail vs bound |
|---|
The guarantee is unconditional: whatever you feed, the fraction past kσ obeys 1/k². The bounded and bell distributions leave it slack; the extremal spike presses it to equality — that is why 1/k² cannot be improved.
Chebyshev proved Bertrand’s postulate: for every n ≥ 1 there is a prime in (n, 2n]. The witness checks all n from 1 to 50 live.
“You must know the distribution to bound the tail.” Cut. Chebyshev needs only the mean and the variance. Shape is irrelevant; the bound holds for all of them at once.
“1/k² is a loose over-estimate you can always beat.” Cut. It is tight — the extremal two-point distribution attains it exactly. No universal bound is smaller.
“The bound tells you something at any k.” Cut. For k ≤ 1 it reads ≥1 and says nothing; it only bites for k > 1.
The red team’s move: relabel the bound as 1/k instead of 1/k². It is still an over-estimate, so the sample never crosses it — but it is no longer Chebyshev’s guarantee. The tightness witness (2) tests it against the extremal distribution and catches the mislabel.
Break 1/k² to 1/k and the extremal distribution — whose tail is exactly 1/k² — no longer meets the stated bound with equality. The witness recomputes and turns red. The attack is real and it is caught.