On the reconstruction of convex sets from random normal measurements (Q2349857): Difference between revisions

The authors are solving the following problem, coming from applied mathematics. Assume that we can measure the normal vectors \(x_i\) at \(N\) random boundary points of a convex body \(K\) in \(R^d\) of surface area 1. How to approximately reconstruct \(K\) from these measurements, and how far the reconstructed body is likely to be from the original one? The authors suggest the following natural reconstruction procedure (see Lemma 6): let \(m\) be the average of the \(x_i\), then, by the Minkowski theorem, there exists a unique polytope with the facets \(F_i\), such that \(x_i-m\) is the exterior normal vector of \(F_i\) of length \(Vol (F_i)\). Normalizing this polytope to the surface area 1, we obtain an asymptotically good approximation of the original body \(K\) in the following sense (Theorem 1): The probability that the Hausdorff distance between the properly shifted approximation and \(K\) is greater than \(L\) is smaller than \(\exp( C N L^{d^2/2 + 3d/2} )\), where \(C\) is a negative constant depending on \(K\). The proof is based on introducing the so called convex-dual distance between surface area measures. This distance is weaker that the bounded Lipschitz one, but stronger than the Hausdorff distance between convex bodies. Some minor remarks on the exposition: there should be no \(K\) in formula (19); Proposition 2 should claim that \(\tilde \nu\) depends only on \(\nu\), otherwise the statement is tautological by taking \(\tilde \nu = \mu_K\).