Fréchet regression for random objects with Euclidean predictors (Q132099)

The paper deals with estimation in models of observation where data are non-Euclidean and specifically do not lie in a vector space. For random elements on a metric space, \textit{M. Fréchet} [Ann. Inst. Henri Poincaré 10, 215--310 (1947; Zbl 0035.20802)] introduced generalizations of mean and variance, which are referred to as Fréchet mean and Fréchet variance. In the paper under review, the global and local Fréchet regression models are developed which extend the notion of Frechet mean to that of conditional Fréchet mean. The obtained asymptotic pointwise rates of convergence are optimal in the sense that for Euclidean objects, they correspond to the known optimal rates, and under the same regularity conditions as satisfied for Euclidean objects, the rates remain the same for random elements on metric spaces. This is demonstrated in the Wasserstein space of distributions as one of many example spaces. For random objects on a Hilbert space, such as regression models with vector predictors and functional data responses, a limit distribution is found. \par For models that involve responses that lie in a Riemannian manifold (including responses consisting of probability distributions and correlation matrices), it is reported that the proposed approach works well in simulations. Both global and local Fréchet regression are demonstrated for demographic and brain imaging data. Local Fréchet regression is also illustrated with response data which lie on the sphere.