Geometrizing rates of convergence under local differential privacy constraints (Q2215754)

This is an article of theoretical inference. It is a work that complements the article [\textit{D. L. Donoho} and \textit{R. C. Liu}, Ann. Stat. 19, No. 2, 633--667 (1991; Zbl 0754.62028)] to characterize the differentially private minimax rate of convergence. The authors deepened the study by geometrizing rates of convergence under local differential privacy constraints restrictions with a depth of theory. They introduce data privacy protection and they present the concept of differential privacy as an alternative. In the articles cited is Donoho and Liu [loc. cit.]. They show preliminary information and ratings including locally differentially private minimax risk, testing affinities and minimax identities. They define and develop a general lower bound on private a-private minimax risk and show some results. Then they develop attainability of lower bounds, including to establish upper bounds on the private minimax, a general result of reachability, optimality of affine estimators. They also feature constructing rate optimal privatization mechanisms and estimators. The authors end with examples of concrete estimation problems and a study of the binary search estimator.