Speaker: 杜航，MIT

Title: Optimal spectral algorithms in high-dimensional statistical inference

Language: Chinese

Time & Venue: 2026.01.13 11:00-12:00 南楼613

Abstract: Spectral algorithms receive a lot of attention in the study of high-dimensional statistical inference, due to their efficiency and ease of implementation. However, it is not always clear a priori which matrix has the most informative spectrum for a specific inference problem. In this talk, I will introduce ideas from statistical physics that guide the choice of optimal spectral algorithms, in the sense that they succeed all the way down to the information-theoretic threshold. I will illustrate these ideas on three multi-model learning problems, focusing on both the heuristic derivation and rigorous mathematical analysis. This talk is based on joint work (in progress) with Henry Hu and Saba Lepsveridze (MIT).