Abstract
Linear regression is often deemed inherently interpretable; however,challenges arise for high-dimensional data. We focus on further understandinghow linear regression approximates nonlinear responses from high-dimensionalfunctional data, motivated by predicting cycle life for lithium-ion batteries.We develop a linearization method to derive feature coefficients, which wecompare with the closest regression coefficients of the path of regressionsolutions. We showcase the methods on battery data case studies where a singlenonlinear compressing feature, $g\colon \mathbb{R}^p \to \mathbb{R}$, is usedto construct a synthetic response, $\mathbf{y} \in \mathbb{R}$. This unifyingview of linear regression and compressing features for high-dimensionalfunctional data helps to understand (1) how regression coefficients are shapedin the highly regularized domain and how they relate to linearized featurecoefficients and (2) how the shape of regression coefficients changes as afunction of regularization to approximate nonlinear responses by exploitinglocal structures.