Directional Nonlinear Principal Components: a measure transportation approach

Abstract

One of the earliest techniques in multivariate analysis, Principal Component Analysis (PCA) remains among the most popular ones in the area. PCA, however, is a linear and bidirectional technique (principal directions $\pm\mathbf{u}$ are defined up to a sign), which is fine under Gaussian or elliptical assumptions but constitutes a major limitation outside that restrictive context. This has prompted the development of numerous nonlinear approaches. We show here how recent results in measure transportation allow for a nonlinear and strictly directional (disentangling $\mathbf{u}$ and $-\mathbf{u}$) concept preserving the intuitive interpretation of the classical linear concept by maximizing expected squared arc lengths along data-driven principal curves rather than principal directions.