Interactive demo · Probability and Statistical Inference · Chapter 5

How good is the normal approximation? The exact Berry–Esseen error

For gamma, Bernoulli and Poisson observations the sum Sn = X1 + … + Xn has a known law (gamma, binomial, Poisson), so the error of the central limit theorem can be computed exactly, with no simulation: the Kolmogorov distance dn = supz |P(Zn ≤ z) − Φ(z)| for the standardised sum Zn = (Sn − nμ)/(σ√n). The Berry–Esseen inequality (5.4) bounds it by Cρ/(σ3√n), with ρ = E|X − μ|3. The top panel shows the error as a function of z and the one-term Edgeworth expansion that predicts it; the bottom panel shows dn against n, with the bound for the book's constant C = 33/4 and for C = 0.4748.

Law of one observation

Parameter

Sample size n

Try these

Gamma(k, 1): mean k, variance k, skewness 2/√k; k = 1 is the exponential and k = 1/2 is χ21/2. Bernoulli(p) and Poisson(λ) are lattice laws: Zn jumps, and the jumps alone keep dn of order 1/√n. Edgeworth: P(Zn ≤ z) ≈ Φ(z) + φ(z)[γ(1 − z2)/6 + S(y)/σ]/√n, where the sawtooth S(y) = ⌊y⌋ − y + 1/2 at y = nμ + zσ√n appears only for lattice laws. Distances: exact at the atoms for lattice laws; a 1601-point grid refined by golden-section search for gamma laws.

In the lecture notes