Uncoupled isotonic regression via minimum Wasserstein deconvolution

• Main Authors: Rigollet, Philippe (Author), Weed, Jonathan (Author)
• Other Authors: Massachusetts Institute of Technology. Department of Mathematics (Contributor)
• Format: Article
• Language: English
• Published: Oxford University Press (OUP), 2020-08-21T13:00:11Z.
• Subjects: Article
• Online Access: Get fulltext

Isotonic regression is a standard problem in shape-constrainedestimation where the goal is to estimate an unknown nondecreasingregression functionffrom independent pairs (xi,yi) whereE[yi] =f(xi),i= 1,...n. While this problem is well understood both statis-tically and computationally, much less is known about its uncoupledcounterpart where one is given only the unordered sets{x1,...,xn}and{y1,...,yn}. In this work, we leverage tools from optimal trans-port theory to derive minimax rates under weak moments conditionsonyiand to give an efficient algorithm achieving optimal rates. Bothupper and lower bounds employ moment-matching arguments that arealso pertinent to learning mixtures of distributions and deconvolution.
National Science Foundation (U.S.) (Grants DMS-1712596, DMS-TRIPODS-1740751)
United States. Office of Naval Research (Grant 00014-17-1-2147)
Chan Zuckerberg Initiative Donor-Advised Fund (DAF) (2018-182642)
National Science Foundation (U.S.). Graduate Research Fellowship (DGE-1122374)