Non parametric estimation of convex bodies and convex polytopes

Overview & Synopsis

In this thesis, we are interested in statistical inference on convex bodies in the Euclidean space ℝ^d. Two models are investigated. The first one consists of the observation of n independent random points, with common uniform distribution on an unknown convex body. The second one is a regression model, with additive subgaussian noise, where the regression function is the indicator function of an unknown convex body. In the first model, our goal is to estimate the unknown support of the common uniform density of the observed points. In the second model, we aim either to estimate the support of the regression function, or to detect whether this support is nonempty, i.e., the regression function is nonzero. In both models, we investigate the cases when the unknown set is a convex polytope, and when we know the number of vertices. If this number is not known, we propose an adaptive method which allows us to obtain a statistical procedure performing asymptotically as well as in the case of perfect knowledge of that number. In addition, this procedure allows misspecification, i.e., provides an estimator of the unknown set, which is optimal in a minimax sense, even if the unknown set is not polytopal, in the contrary to what may have been thought. We prove a universal deviation inequality for the volume of the convex hull of the observations in the first model. We show that this inequality allows one to derive tight bounds on the moments of the missing volume of this convex hull, as well as on the moments of the number of its vertices. In the one-dimensional case, in the second model, we compute the asymptotic minimal size of the unknown set so that it can be detected by some statistical procedure, and we propose a decision rule which allows consistent testing of whether of that set is empty.