A1525
Title: Statistical deconvolution via entropic iterative proportional fitting
Authors: Antoine Jacquet - Sciences Po (France) [presenting]
Alfred Galichon - New York University (United States)
Guillaume Pouliot - Rice University (United States)
Abstract: A linear programming approach to deconvolution is proposed for recovering the latent distribution of $X$ from observations of $Y = X + \varepsilon$, when the noise distribution is known or can be simulated. From an entropically regularized formulation, primal and dual iterative proportional fitting algorithms with closed-form updates are derived, in the spirit of Sinkhorns algorithm for optimal transport. Numerical experiments suggest that the proposed approach is tractable and compares favorably with existing procedures.
