II · Probability & randomness

Concentration inequalities

What it is

Chernoff, Hoeffding, Chebyshev — bounds on how far an empirical sum strays from its expectation.

Where it lives

A/B test sample sizes, why "law of large numbers" actually converges, randomised algorithms' bounds.

The key insight

For n samples of a [0,1] variable: P(|mean − μ| > ε) ≤ 2 exp(−2nε²). The 2× factor comes from a two-sided bound.