Under regularity conditions √n(θ̂n − θ) is approximately N(0, 1/I(θ)), and no well-behaved estimator does better: 1/I(θ) is the Cramér–Rao bound. Each preset pits the MLE (blue) against a natural competitor (pink) on R simulated samples of size n. The top panel shows the two sampling distributions of √n(estimate − truth) with the Normal limits (gold: variance 1/I; dashed pink: the competitor's limit); the bottom panel shows n times the mean squared error against n on log axes. The MLE's curve settles on the bound, the competitor's on a higher level, and the ratio of the two levels is the asymptotic relative efficiency. The last preset breaks the rules: for the Uniform(0, θ) endpoint there is no Fisher information and the MLE converges at rate 1/n, faster than any "efficient" estimator.
Normal: N(0, 1), mean against median. Cauchy: location 0, scale 1; the MLE is computed by Newton's method from the median, with a weighted-mean (EM) step whenever Newton's step would lower the likelihood. Gamma: shape α = 2 and scale β = 1 both unknown, the shape estimated by maximum likelihood (Newton on the profile equation log α − ψ(α) = log X̄ − mean log X) or by moments, X̄2/σ̂2. Exercise 9.6: N(0.5, 1) data, ψ = P(X > 0) = Φ(0.5), estimated by Φ(X̄) or by the fraction of positive observations. Uniform(0, 1): the maximum against 2X̄. Each sample size uses its own stream of the seeded generator, so the histogram is the point of the lower curve at the chosen n.
code/ch09_parametric_inference.py (demo_estimators, cauchy_mle_demo, gamma_mle_demo, eff_limits).