Noise-resilient penalty operators based on statistical differentiation schemes

Classical approaches to penalized smoothing often rely on basis or kernel expansions, which constrain the estimator to a fixed span and may be restrictive for discretely observed data. We instead regularize a single noisy trajectory directly on its observation grid, using difference operators that remain genuine finite-difference approximations to derivatives while being statistically normalized and mutually decorrelated under a reference noise law. We extend this white-noise construction to a parametric family of covariance-adapted reference geometries, with the parameter estimated by generalized method of moments from an independent calibration sample. A first-order plug-in expansion shows that, within this moment family, efficient calibration minimizes the leading-order discrepancy between the plug-in and oracle covariance-adapted smoothers. Numerical experiments confirm the predicted rate at which this plug-in discrepancy vanishes and compare the resulting reconstruction against conventional discrete, basis, and kernel smoothers.