In this thesis, we study derivative estimation via a pre-processing framework. We propose the pre-processed derivative estimator (PPDE) for estimating mixed partial derivatives in nonparametric regression. The key idea is to separate regression and de...
In this thesis, we study derivative estimation via a pre-processing framework. We propose the pre-processed derivative estimator (PPDE) for estimating mixed partial derivatives in nonparametric regression. The key idea is to separate regression and derivative estimation: an arbitrary regression estimator is first trained as a pre-processing, and derivatives are then recovered by applying one-dimensional local polynomial regression to slices of the pre-processed function and iterating these sliced operators to obtain mixed derivatives. We develop both L2- and L∞-PPDE, derive risk bounds that connect the error of the pre-processing to the error of the derivative estimator, and show that if the pre-processing attains minimax-optimal rates over Hölder classes, then PPDE achieves the minimax-optimal rates for mixed partial derivatives in both norms. Furthermore, when deep neural networks are used as pre-processings, PPDE inherits their structural adaptivity to low intrinsic dimensionality, enabling efficient estimation even when the data dimension is high. Simulation studies in low- and high-dimensional settings and for mixed derivatives compare PPDE with existing seminal works, and show that PPDE is competitive and remains numerically stable in higher dimensions.