[Paper Review] Advances on nonparametric regression for functional variables

Frédéric Ferraty, André Mas+1 more|arXiv (Cornell University)|Mar 3, 2006

Statistical Methods and Inference7 references22 citations

TL;DR

This paper advances nonparametric functional regression by deriving exact asymptotic expansions for bias, variance, and mean squared error of the kernel estimator, including precise constants and asymptotic normality. It introduces a functional wild bootstrap for bandwidth selection and confidence band construction, validated on simulated and real data.

ABSTRACT

We consider the problem of predicting a real random variable from a functional explanatory variable. The problem is attacked by mean of nonparametric kernel approach which has been recently adapted to this functional context. We derive theoretical results by giving a deep asymptotic study of the behaviour of the estimate, including mean squared convergence (with rates and precise evaluation of the constant terms) as well as asymptotic distribution. Practical use of these results are relying on the ability to estimate these constants. Some perspectives in this direction are discussed including the presentation of a functional version of bootstrapping ideas.

Motivation & Objective

• Address the lack of exact constant computation in asymptotic theory for functional nonparametric regression estimators.
• Provide precise asymptotic expansions for bias, variance, and mean squared error with explicit constants.
• Establish asymptotic normality of the kernel estimator for functional data with exact variance terms.
• Develop a practical functional wild bootstrap procedure for bandwidth selection and confidence band construction.
• Bridge theoretical results with practical applications using simulated and real functional datasets.

Proposed method

• Adapt the Nadaraya-Watson kernel estimator to functional data using a kernel function $ K $ and bandwidth $ h $.
• Derive asymptotic expansions for the mean squared error by conditioning on the functional covariate $ \mathcal{X} $ and using functional density estimation.
• Use functional versions of Lebesgue-type integrals and norms to handle infinite-dimensional spaces.
• Apply Slutsky's theorem and the Central Limit Theorem to derive asymptotic normality of the estimator.
• Propose a functional wild bootstrap by resampling residuals with random signs, adapted to the functional context.
• Implement the wild bootstrap for automatic bandwidth selection and construction of pointwise confidence intervals.

Experimental results

Research questions

• RQ1 What are the exact leading terms in the asymptotic expansion of the bias and variance of the functional kernel regression estimator?
• RQ2 How can the exact constants in the mean squared error be estimated for practical inference?
• RQ3 What is the asymptotic distribution of the functional kernel estimator, and can it be used for constructing confidence bands?
• RQ4 Can a functional version of the wild bootstrap be effectively used for bandwidth selection in nonparametric functional regression?
• RQ5 How does the proposed functional wild bootstrap perform on real and simulated functional data compared to standard methods?

Key findings

• The paper derives exact asymptotic expansions for the bias, variance, and mean squared error of the functional kernel regression estimator, including precise constants.
• Asymptotic normality of the estimator is established with a limiting distribution that depends on the exact variance term $ \sigma_{\varepsilon}^{2}M_{2}/M_{1}^{2} $, enabling inference.
• The functional wild bootstrap procedure is proposed and validated on both simulated and real functional datasets for bandwidth selection.
• Theoretical results on the accuracy of the functional wild bootstrap remain open, suggesting a key direction for future research.
• The method enables automatic bandwidth selection by minimizing estimated mean squared error, and supports construction of confidence bands via the asymptotic distribution.
• Theoretical conditions ensure the consistency of the estimator, with convergence rates depending on the bandwidth $ h $ and the functional density at the point of interest.