WorksMathematics
Finite-Smoothing Efficiency and Inference in Single-Index Models
Abstract
When the error is independent of the predictors, its distribution can carry information about a single-index direction beyond the conditional mean. We develop Single-Index Kernel Density Regression (SI-KDR): at each candidate direction, it refits the link and evaluates a leave-one-out kernel log-likelihood of the residuals. Under Gaussian design, the population score vanishes at the true line at every positive smoothing tuple. At a fixed positive-curvature tuple, the leading covariance in the sample expansion is the conditional-mean efficiency bound times , where is the information gain from independence and is the variance inflation from smoothing. We derive by following each observation through link refitting, the dual role of its residual as evaluation point and kernel center, and training standardization. Case-weight covariance estimates guide tuning for point estimation. Tuning separately at each candidate gives an exact conditional test. When predictors outnumber observations, independent screening that retains the support preserves the covariance comparison and the exact test, and becomes the squared-noncentrality ratio of a retained-coordinate test relative to the conditional-mean oracle. In four non-Gaussian error distribution designs at , covariance-trace tuning reduced direction risk by 56% to 74% relative to reference bandwidths, and score-inversion coverage was 93.8% to 96.6%.