Nonparametric confidence bands and honesty

For a specified interval I and function class $\mathcal F$, a finite-sample honest band requires the infimum over f in that class of the probability of $L_n(x)\le f(x)\le U_n(x)$ for every x in I to be at least one minus alpha. Asymptotic honesty places a lower limit over n outside that infimum. In contrast, pointwise asymptotic validity for each fixed f permits the worst function to change with n and does not supply the uniform statement.

Simultaneity over x is another separate quantifier. Intervals that each have 95% coverage at a fixed location need not cover an entire curve with probability 95%. If errors at five locations were independent and each interval covered exactly with probability 0.95, the joint probability would be 0.95 to the fifth power, about 0.774. Dependence changes this calculation but does not turn marginal coverage into an automatic simultaneous guarantee. A continuum of points also needs control between any plotted grid nodes.

Decompose estimation error into the centered random fluctuation $\widehat f-\mathbb E_f\widehat f$ and the bias $\mathbb E_f\widehat f-f$. A calibrated simultaneous fluctuation bound alone covers the expected estimator, not necessarily the true function. If a deterministic bias bound B(x) holds uniformly on the announced class, adding it to the stochastic radius gives a valid triangle-inequality band. Estimating bias introduces its own error, and a bootstrap that reproduces centered fluctuations does not automatically remove that error.

Undersmoothing can make bias negligible relative to stochastic width under suitable uniform regularity conditions, at the cost of larger noise. Bias correction can instead change both the center and its variance. Neither method is a universal repair without assumptions. Choosing bandwidth from the same data adds selection dependence, so a fixed-bandwidth calibration needs a theorem before being reused unchanged. Exact finite-grid Gaussian calculations are useful teaching examples because both the stochastic event and the approximation remainder can be written down separately.

In standard nonparametric density and regression settings, full nested smoothness classes can prevent a band from being both uniformly honest over their union and as narrow as the smoother class would permit when it were known. This is an inference obstruction, not a claim that adaptive point estimation is impossible. Quantitative self-similarity or separation conditions can exclude signals whose unresolved fine-scale behavior is too hard to detect. Other choices accept a wider band or weaken uniformity, each changing the stated statistical promise.

Distribution estimation illustrates the importance of the target. The empirical cumulative distribution admits a distribution-free uniform band because it averages a structured family of bounded indicators. Recovering a density differentiates that function, an unstable operation without derivative control. Smooth perturbations can have small cumulative differences and large density differences at high frequency. Extra smoothness restrictions are therefore substantive information about the target, not cosmetic regularity assumptions that may be omitted when transferring a confidence statement.