Preprint on Arxiv: https://arxiv.org/abs/1909.03540

In this project we developed methods for constructing confidence intervals and hypothesis tests for the linear component of high-dimensional single-index models. Single-index models generalize linear models by allowing the reponse \(y\) to depend on a non-linear function of a linear combination of the predictors: \[ y = g(\langle x, \beta \rangle ) \] This can avoid curse of dimensionality because we are not dealing with a fully non-parametric model \(y = g(x)\) which suffers from curse of dimensionality. Even for a high-dimensional \(\beta\), we still estimate \(\beta\) if it is sparse and the design is elliptically symmetric.

The code for experiments in the paper is available on github: https://github.com/ehamid/sim_debiasing

Using Hermite expansions to estimate link function Histogram comparing efficient and plain de-biased estimates
More efficient estimation using Hermite polynomials

Updated: