Sample and map from one convex: generation using conjugate temporal methods

The method of the generated model library is to divide the ideal model into two blocks: Define first how to spread the noise (e. u such that ρ = ∇u♯e^–u. While this seems to bind the sample effectively (from the log-concave distribution e^–u) and action (pushing the particles with ∇u), we look at simple examples (eg Gaussians or 1D distributions) This choice is suitable for practical tasks. We study another phenomenon, where ρ is fixed as ∇w *♯e^–Wthere w * Is the convex conjugate of a convex potential W. We call this method converging time intervals, and we show very accurate results in these examples. Because ∇w * Monge map between the log-concave distribution e^–W and ρ, we depend on the smoothness of the cheeks to lift the algorithm to recover W from ρ samples, and parameters W as a convex aural network. We also deal with the general case of a sample where the personality of ρ is known only until it is normal, and propose an algorithm to learn W In this situation.

• † Crest-Ensae, Institut Polytechnique de Paris