Title:Double Data Piling for Heterogeneous Covariance Models

Abstract:In this work, we characterize two data piling phenomenon for a high-dimensional binary classification problem with heterogeneous covariance models. The data piling refers to the phenomenon where projections of the training data onto a direction vector have exactly two distinct values, one for each class. This first data piling phenomenon occurs for any data when the dimension $p$ is larger than the sample size $n$. We show that the second data piling phenomenon, which refers to a data piling of independent test data, can occur in an asymptotic context where $p$ grows while $n$ is fixed. We further show that a second maximal data piling direction, which gives an asymptotic maximal distance between the two piles of independent test data, can be obtained by projecting the first maximal data piling direction onto the nullspace of the common leading eigenspace. This observation provides a theoretical explanation for the phenomenon where the optimal ridge parameter can be negative in the context of high-dimensional linear classification. Based on the second data piling phenomenon, we propose various linear classification rules which ensure perfect classification of high-dimension low-sample-size data under generalized heterogeneous spiked covariance models.