Interactive demo · Nonparametric Statistical Inference · Chapter 9

Intersecting confidence intervals

At a point x, compute the kernel estimate r̂j(x) for bandwidths h1 < h2 < …; its standard deviation sj falls as h grows while its bias grows. Put an interval Dj = r̂j ± Γsj around each estimate and walk up the bandwidths for as long as all the intervals so far still have a common point. The last such bandwidth j+ is used at x. While the bias is small the intervals all contain r(x); once the bias dominates they stop overlapping. The rule needs no knowledge of the smoothness, and the bandwidth it picks changes with x.

Data (Doppler, n = 1024)

Rule

Interval half-width Γ (in standard deviations)

Noise level used

Point x for the middle panel

Try these

Nadaraya–Watson estimator with a Gaussian kernel at the design points xi = i/1024, 50 bandwidths log-spaced from 1/1024 to 1; sj(x) = σ‖ℓ(x, hj)‖. The book's intervals are r̂j ± 2ρj with ρj = κsj, so Γ = 2κ; its theory asks for κ > √(2 log n). The oracle index j* = max{j : biasj ≤ ρj} uses the true function. Lepski's rule here: the largest hj with |r̂j − r̂k| ≤ Γsk for every smaller bandwidth hk. The estimated noise level is the first-difference estimator of Section 5.6.

In the lecture notes