Suppose that a random variable X of interest is observed perturbed by independent additive noise Y. This paper concerns the “the least favorable perturbation” , which maximizes the prediction error in the class of Y with . We find a characterization of the answer to this question, and show by example that it can be surprisingly complicated. However, in the special case where X is infinitely divisible, the solution is complete and simple. We also explore the conjecture that noisier Y makes prediction worse.
We dedicate this work to our colleague, mentor, and friend, Professor Larry Shepp (1936–2013)
We thank Professor Dan Crisan (Imperial College London) and Dongzhou Huang (Rice University) for helpful discussions. We also wish to thank an anonymous referee whose helpful and detailed comments have greatly improved the quality of this manuscript. The first-named author gratefully acknowledges the support of ARO-YIP-71636-MA, NSF DMS-1811936, ONR N00014-18-1-2192, and ONR N00014-21-1-2672.
"The least favorable noise." Electron. Commun. Probab. 27 1 - 11, 2022. https://doi.org/10.1214/22-ECP467