WorksMathematics
Renewable Estimation for Generalized Nonparametric Regression with Dynamic Spline Bases and Likelihood Pullback
Abstract
Streaming datasets require statistical models to be updated as new observations arrive, often without repeated access to the full history. In generalized nonparametric regression, a fixed spline space may limit approximation accuracy when the initial representation is insufficient. However, when the basis is expanded, the historical summaries represented in the previous spline space must be re-expressed in the enlarged spline space, which requires a suitable mechanism. We propose a renewable penalized-likelihood algorithm for canonical exponential-family nonparametric regression that combines nested cubic-spline refinement with a Gram-metric pullback of the retained historical likelihood quadratic under the streaming setting. The algorithm incorporates incoming-batch likelihood contributions and updates historical score and curvature summaries at the actual fitted center. The Gram-metric pullback preserves the retained historical quadratic over the embedded previous space. Under explicit regularity and algorithmic conditions, we establish L2-consistency at rate O_p(N^{-2/5}) for a fixed target in H^2[0,1]. We further establish that the energy-norm discrepancy between the renewable estimator and a matched offline estimator is of smaller order, where the offline method is fitted based on the full data and using the same realized terminal spline space and penalty. Simulations demonstrate the benefits of dynamic bases when the initial spline representation is restrictive, with additional gains from historical summary pullback. Additional experiments assess computational cost and storage of the proposed method relative to the offline method. Applications to real datasets also illustrate the proposed method.