AI Summary & Key Takeaways
AI InsightThe finite-sample performance of the proposed estimator is illustrated through simulation experiments and an application to spectrometric data.
📌 3 Key Takeaways
Abstract
We propose a family of kernel estimators k-nearest neighbor (kNN) for scalar-on-function regression based on a distance-exponent weighting scheme within the kNN neighborhood. Using the random radius kNN as a data-adaptive local scale, the method extends the classical functional kernel-kNN regression by introducing an additional parameter that controls the weight distribution among neighboring observations. Under standard assumptions on the functional design and a mild annulus-mass condition, pointwise consistency and explicit bias–variance bounds are established. Under local Hölder smoothness and polynomial small-ball behavior, the estimator achieves the usual convergence rate, while the exponent parameter affects only the leading constants. A leading bias expansion is derived, making the exponent’s influence explicit via weighted moments of the limiting distribution of normalized distances. In particular, the analysis identifies situations in which increasing the exponent can reduce the leading bias constant without altering the order of the variance. Practical selection of the tuning parameters is discussed through cross-validation over finite grids. The finite-sample performance of the proposed estimator is illustrated through simulation experiments and an application to spectrometric data.