On consistency and sparsity for sliced inverse regression in high dimensions (Q1750280)

Sliced Inverse Regression is a technique for dimension reduction of a dataset \(X\) with dimensions \((n \times p)\) based on a response variable \(y\). In this paper the authors prove that, under mild conditions, the SIR is consistent if and only if \(\lim p/n\) is \(0\). When this condition is not satisfied, a new procedure called DT-SIR (Diagonal Thresholding SIR) is introduced. The DT-SIR is proved to be consistent under some additional assumptions.