Fri, 25-Sep-2020 / 1:15pm / Zoom: https://stanford.zoom.us/meeting/register/tJwkf-uvqjoqHNIWxY4HHon4K107QMo22PVR
Symmetric properties of distributions arise in multiple settings, where for each of them separate estimators have been developed. Recently, Orlitsky et al. showed that a single estimator, called the profile maximum likelihood (PML), achieves the optimal sample complexity universally for many properties and any accuracy parameter larger than $n^{-1⁄4}$, where $n$ is the sample size. They also raised the question whether this low-accuracy range is an artifact of the analysis or a fundamental limitation of the PML, which remained open after several subsequent work.
In this talk, we provide a complete answer to this question and characterize the tight performance of PML in the high-accuracy regime. On the positive side, we show that the PML remains sample-optimal for any accuracy parameter larger than $n^{-1⁄3}$ using a novel chaining property of the PML distributions. In particular, the PML distribution itself is an optimal estimator of the sorted hidden distribution. On the negative side, we show that the PML as well as any adaptive approach cannot be universally sample-optimal when the accuracy parameter is below $n^{-1⁄3}$, and characterize the exact penalty for adaptation via a matching adaptation lower bound.
Based on joint work with Kirankumar Shiragur. The full papers are available online at https://arxiv.org/abs/2004.03166 and https://arxiv.org/abs/2008.11964.
Yanjun Han is a 6th-year PhD candidate in Electrical Engineering at Stanford University, advised by Tsachy Weissman. His research interests include high-dimensional and nonparametric statistics, information theory, online learning, applied probability, and their applications.
Packard Building
Electrical Engineering Department
350 Jane Stanford Way
Stanford, CA 94305-9515
