Assistant Professor of Statistical Science
My research develops theoretical foundations for statistical inference under modern constraints, focusing on problems where classical methods face fundamental limitations. I work at the intersection of statistics, information theory, and machine learning, aiming to characterize performance limits and develop optimal methods that achieve these limits.
Privacy-Preserving Statistical Inference
I study the fundamental trade-offs between data privacy and statistical utility in differential privacy frameworks. My work establishes tight minimax bounds for inference problems including nonparametric regression, hypothesis testing, and density estimation under privacy constraints. I have developed a federated privacy framework that generalizes both local and central differential privacy models, revealing precisely when privacy can be maintained without sacrificing statistical accuracy and when fundamental costs must be paid. Current work extends this to adaptive estimation problems and explores how privacy requirements reshape classical experimental design trade-offs.
Communication-Constrained and Distributed Learning
I develop optimal methods for statistical inference when data is distributed across multiple locations with limited communication bandwidth. My research establishes minimax rates and optimal testing procedures for federated hypothesis testing, showing that consistent inference remains possible even with severe bandwidth constraints. This work extends to meta-analysis settings, where I quantify how combining summary statistics from multiple studies affects statistical power. These results provide crucial guidance for network infrastructure design and experimental planning in bandwidth-limited environments.
High-Dimensional Bayesian Inference
I investigate theoretical properties of Bayesian methods for high-dimensional data analysis, particularly focusing on spike-and-slab priors for model selection. My work establishes Bernstein-von Mises theorems for semiparametric inference in high-dimensional regression, characterizing when Bayesian procedures achieve optimal frequentist properties and providing guidance for prior design.