betaFeedback

WorksMathematics

Deep Nonlinear Sufficient Dimension Reduction for Discretely Observed Functional Data

Abstract

We develop a modular framework for deep nonlinear Sufficient Dimension Reduction (SDR) of discretely observed functional predictors with Euclidean or metric-space-valued responses. A presmoothing encoder converts the functional input into finite-dimensional coordinates, and a neural network learns a sufficient representation from them. The framework separates the choice of encoder from that of the SDR criterion. We study kernel embedding and discrete functional principal component analysis, each combined with either a generalized martingale difference divergence network or a Fréchet cumulative covariance network. These combinations share objectives of the martingale difference divergence form. {Under input smoothness and general Hölder regularity of the target operator, we establish approximation and generalization bounds that jointly account for functional truncation, network approximation, statistical estimation, numerical integration, measurement noise, and encoder estimation. The complete theoretical analysis rigorously demonstrates how errors in the encoded coordinates propagate through the network to the deep nonlinear SDR estimator, and quantifies the final estimation error in relation to observation frequency and model complexity.} Simulations and an application to satellite-derived vegetation trajectories and crop-composition responses illustrate the statistical and computational advantages over RKHS-based nonlinear functional SDR methods.

Paper details

vlnc:W_e3c7ew5j
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC-BY-NC-SA-4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…