Nonparametric Estimation of Trend in Directional Data

TL;DR

This paper introduces and analyzes nonparametric methods for estimating trends in directional data, such as paleomagnetic directions, respecting spherical geometry and using empirical process theory for risk assessment.

Contribution

It develops a class of spherical geometry-respecting trend estimators and provides a nonparametric error model with risk comparison techniques.

Findings

Proposes a richer class of directional trend estimators.

Establishes uniform laws of large numbers for risk estimation.

Provides criteria for selecting good trend estimators.

Abstract

Consider measured positions of the paleomagnetic north pole over time. Each measured position may be viewed as a direction, expressed as a unit vector in three dimensions and incorporating some error. In this sequence, the true directions are expected to be close to one another at nearby times. A simple trend estimator that respects the geometry of the sphere is to compute a running average over the time-ordered observed direction vectors, then normalize these average vectors to unit length. This paper treats a considerably richer class of competing directional trend estimators that respect spherical geometry. The analysis relies on a nonparametric error model for directional data in $R^q$ that imposes no symmetry or other shape restrictions on the error distributions. Good trend estimators are selected by comparing estimated risks of competing estimators under the error model. Uniform laws of large numbers, from empirical process theory, establish when these estimated risks are trustworthy surrogates for the corresponding unknown risks.