Interactive demo · Nonparametric Statistical Inference · Chapter 5

Confidence bands and the tube formula

For a linear smoother the band r̂(x) ± c σ̂ ‖ℓ(x)‖ fails somewhere on [a, b] when the Gaussian process W(x) = Σ Zi Ti(x), T = ℓ/‖ℓ‖, leaves [−c, c]. The tube formula gives P(max|W| > c) ≈ 2(1 − Φ(c)) + (κ0/π)e−c²/2, where κ0 = ∫‖T'(x)‖dx is the length of the curve traced by T on the unit sphere: the smaller h, the longer the curve and the wider the band. With non-constant variance, σ̂2(xi) enters inside the norm. The band is for r̄ = E r̂, the smoothed curve; the simulation on the four-bump function shows what that means.

Data

Bandwidth h (local linear, tricube)

Variance

Level α

Bandwidths searched (Bonferroni, Remark 5.106)

Try these

κ0 is computed from the weights on a fine grid (central differences, trapezoid rule) and c solves 2(1 − Φ(c)) + (κ0/π)e−c²/2 = α/m. Constant variance: σ̂2 from (5.86). Variance function: the log squared residuals smoothed by local linear regression with its own cross-validated bandwidth, exponentiated, with or without the correction −E log ε2 = 1.2704 for Normal errors. The simulation uses the true σ = 0.2.

In the lecture notes