Title: Stratification and Antithetic Mechanisms for Subsampling

Presented by Dingyi Wang, Chinese Academy of Sciences

DATE: Wednesday, March 25, 2026, 3:30 PM, AUST 434

Meeting Link: WebEx link

Coffee will be available at 3:00 PM in the Noether Lounge (AUST 326)

Abstract: Massive datasets present computational challenges for statistical estimation, making subsampling a critical tool for balancing accuracy and cost. In this talk, we present two frameworks designed to improve estimation efficiency over existing probability-based methods. First, we present Maximum-Variance-Reduction Stratification (MVRS), a mechanism that partitions the data to reduce the variance of the subsample estimator. MVRS incurs only linear additional computational costs and can be seamlessly combined with existing subsampling designs to further boost efficiency. Second, we introduce an antithetic subsampling framework that departs from standard independent sampling by intentionally inducing negative dependence among observations. This designed negative correlation reduces the variance of the resulting estimator. We establish consistency and asymptotic normality for both estimators and demonstrate their superior accuracy via extensive experiments.