1 h 3 min
Seminar, Hosted by Toyota Technological Institute at Chicago (TTIC)
Quasi-Monte Carlo Integration and Discrepancy
University of Michigan
Recording
Abstract
A classical approach to numerically integrating a function f is the Monte Carlo (MC) method, which evaluates f at random points so that the estimation error scales as sigma(f)/n^(1/2) with n samples, where sigma(f) is the standard deviation of f. A different approach widely used in practice is quasi-Monte Carlo (QMC), which evaluates f at carefully chosen deterministic points so that the error scales roughly as 1/n. Both methods have distinctive advantages and shortcomings, and a key question has been to find a method that combines the advantages of both.
The talk introduces QMC methods and their connections to several areas of mathematics and to geometric discrepancy, then shows how recent developments in algorithmic discrepancy theory yield a method that combines the benefits of MC and QMC and improves on previous QMC approaches in several ways. The talk is self-contained and elementary, requiring no prior knowledge of discrepancy or integration, and is based on joint work with Haotian Jiang (University of Chicago).