Our colloquium series offers a diverse platform for research scholars, faculty, students, and industry experts to share and exchange ideas, fostering discussion and networking across mathematics, statistics and data science.
Our own Professor Alexander Katsevich will speak at this week's colloquium on "High-dimensional Laplace asymptotics up to the concentration threshold."
Abstract: High-dimensional Laplace-type integrals are a basic tool for approximating difficult quantities across mathematics, statistics, physics, chemistry, and engineering. They arise in posterior normalization and expectation formulas in Bayesian statistics, in partition functions and free energies in statistical physics, in saddle-point and fluctuation calculations in field-theoretic models, and in molecular simulation and theoretical chemistry. A major obstacle, however, is that in many modern problems both the large parameter lambda and the dimension d grow, and the classical rigorous theory no longer applies in these regimes.
In this talk I will describe recent results that substantially extend the reach of rigorous Laplace asymptotics. Until now, explicit high-dimensional expansions with remainder bounds were essentially limited to the Gaussian-approximation regime, which requires d^2/lambda -> 0. Our results go far beyond that barrier and remain valid essentially all the way up to the natural concentration threshold d/lambda -> 0. In other words, they identify the true high-dimensional range in which one can still obtain explicit asymptotic formulas with rigorous error control. This closes a major gap between formal practice and rigorous theory, and represents a significant advance on a century-old problem of extending Laplace asymptotics to high dimension.
The results have several important consequences. They provide explicit higher-order formulas for logarithms of Laplace integrals, yielding rigorous versions of corrections that in physics are often interpreted as fluctuation or loop terms. They also give constructive approximations to the underlying concentrating probability measures, leading to practical approximations for posterior expectations, marginal likelihoods, and approximate sampling schemes. As a result, they are directly relevant to uncertainty quantification and Bayesian inverse problems, while also placing a broad class of calculations in statistical physics, chemistry, and related areas on a firmer mathematical foundation.
This is joint work with Anya Katsevich, Department of Statistical Science, Duke University.
Speaker Bio: Alexander Katsevich is a Professor of Mathematics at the University of Central Florida whose research focuses on inverse problems, microlocal analysis, tomography, and medical imaging. He earned his Ph.D. in Mathematics from Kansas State University in 1994 and held a postdoctoral fellowship at Los Alamos National Laboratory before joining UCF in 1996. He is the recipient of the 2016 Marcus Wallenberg Prize.
Read More