Averaging orthogonal projectors

Dimensionality is a major concern in analyzing large data sets. Some well known dimension reduction methods are for example principal component analysis (PCA), invariant coordinate selection (ICS), sliced inverse regression (SIR), sliced average variance estimate (SAVE), principal hessian directions (PHD) and inverse regression estimator (IRE). However, these methods are usually adequate of finding only certain types of structures or dependencies within the data. This calls the need to combine information coming from several different dimension reduction methods. We propose a generalization of the Crone and Crosby distance, a weighted distance that allows to combine subspaces of different dimensions. Some natural choices of weights are considered in detail. Based on the weighted distance metric we discuss the concept of averages of subspaces as well to combine various dimension reduction methods. The performance of the weighted distances and the combining approach is illustrated via simulations.