Optimal Confidence Bands for Shape-restricted Regression in Multidimensions

TL;DR

This paper develops adaptive, shape-constrained confidence bands for multivariate regression functions in a white noise model, ensuring guaranteed coverage and optimal width, especially for low-complexity functions.

Contribution

It introduces a multiscale, kernel-based method for constructing confidence bands that adapt to smoothness and dimensionality in multivariate shape-restricted regression.

Findings

Confidence bands guarantee coverage for all sample sizes and functions.

Bands adapt to smoothness and intrinsic dimensionality of the target function.

Achieve near-parametric width for low-complexity functions.

Abstract

In this paper, we propose and study construction of confidence bands for shape-constrained regression functions when the predictor is multivariate. In particular, we consider the continuous multidimensional white noise model given by $d Y(\mathbf{t}) = n^{1/2} f(\mathbf{t}) \,d\mathbf{t} + d W(\mathbf{t})$, where $Y$ is the observed stochastic process on $[0,1]^d$ ($d\ge 1$), $W$ is the standard Brownian sheet on $[0,1]^d$, and $f$ is the unknown function of interest assumed to belong to a (shape-constrained) function class, e.g., coordinate-wise monotone functions or convex functions. The constructed confidence bands are based on local kernel averaging with bandwidth chosen automatically via a multivariate multiscale statistic. The confidence bands have guaranteed coverage for every $n$ and for every member of the underlying function class. Under monotonicity/convexity constraints on $f$, the proposed confidence bands automatically adapt (in terms of width) to the global and local (H\"{o}lder) smoothness and intrinsic dimensionality of the unknown $f$; the bands are also shown to be optimal in a certain sense. These bands have (almost) parametric ($n^{-1/2}$) widths when the underlying function has ``low-complexity'' (e.g., piecewise constant/affine).