Statistics and Computational Methods Seminar Series (Spring 2026)

Speaker: Lasse Vuursteen (Duke University)

Title: The Price of Differential Privacy in Adaptive Estimation and Uncertainty Quantification

Abstract: Privacy has become a central concern in modern statistical practice, and differential privacy offers a mathematically rigorous framework for guaranteeing it. Imposing a differential privacy constraint on a statistical inference procedure can, however, come at a cost - one that is especially pronounced in high- and infinite-dimensional models, where the amount of information per parameter is already limited. In some settings, privacy is essentially free, and optimal inference is possible with no asymptotic penalty. In others, the constraint severely degrades the quality of inference, or makes the task altogether impossible. Quantifying this cost is therefore essential: it tells us whether a given task remains feasible under privacy and, when it does, how the required sample size scales.
In this talk, I will discuss estimation and uncertainty quantification in high- and infinite-dimensional settings, highlighting the sharp differences between the private and non-private regimes and the ideas that explain them.

Link Teams: https://teams.microsoft.com/meet/321143349190968?p=8dhYpCisiA8HL2PBpU