Nonparametric Estimation of Trend in Directional Data
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.