Abstract:
|
Directional statistics are concerned with observations that are not unrestricted p-dimensional vectors as in classical multivariate statistical analysis but directions, axes or rotations. Directional data can be found in various fields, including astronomy, medicine and environmetrics to cite a few. In the present paper, we propose concepts of ranks, signs, and quantiles based on measure transportation ideas for data taking values on unit (hyper)spheres. These ranks and signs allow for distribution-free inference for directional data without the usual assumption of rotational symmetry, while the corresponding quantile contours define a data-driven system of parallels. We show the consistency of empirical versions of the quantiles and demonstrate their usefulness on simulated and real data.
|