Interactive demo · Probability and Statistical Inference · Chapter 6

Consistency: does the estimate settle on the truth as n grows?

An estimator θ̂n is consistent when, for every tolerance ε > 0, the probability P(|θ̂n − θ| > ε) tends to 0 (Definition 6.7). Panel (a) follows the estimate as each simulated data set grows one observation at a time; panel (b) shows the exact value of that probability for every n, the fraction of the simulated paths outside the band, and the bounds that prove consistency: Chebyshev's MSE/ε² (Theorem 6.10) and, for a proportion, Hoeffding's 2e−2nε². One of the five estimators never settles.

Estimator

Tolerance ε

Simulated paths K

Largest sample size

Panel (b) vertical scale

Try these

Path j uses its own seeded generator (seed + 7919 j). Cauchy data are X = tan(π(u − 1/2)) with u uniform; the median of n values is the order statistic X(⌈n/2⌉). Paths are evaluated at n = 1, …, 20 and then at about 25 sizes per decade. At most 20 paths are drawn in panel (a); all K enter the fraction in panel (b).

In the lecture notes