README.md
In barbehenna/coleReg: COmpound Loss Emporium with Regression

ColeReg: COmpound Loss Emporium with Regression

This project is an R "sub-package" containing nonparametric regression methods for the simultaneous estimation of normal means under squared error loss. ColeReg was used in the development of the methods accompanying A nonparametric regression approach to asymptotically optimal estimation of normal means (2022); however, these methods have been integrated into sdzhao/cole and are maintained here for reference only.

wtf1() implements the constrained and penalized least-squares estimator, $\hat{d}$ in the paper. To support this function we offer two different optimizers: MOSEK and CVXR, and three different methods to select the tuning parameter: high-probability upper bound, plug-in, and cross-validation (TV1.oracle.highprob(), TV1.oracle.plugin(), and cv.wtf1(), respectively)
wtf0() implements the monotone-only least-squares estimator given bandwidth $h$, $\tilde{d}_h$ in the paper. The bandwidth rate achieving asymptotically optimal risk is used by default. A fast, accurate approximation is used for the knots by default; however, optimal.knots(), gives the best knots for the risk estimate.
f.kde() implements Brown and Greenshtein's $f$-estimator (https://www.jstor.org/stable/30243684) with their suggested bandwidth as default.