Applied Math Seminar- PINS: Proximal Iterations with Sparse Newton and Sinkhorn for Optimal Transport

Who: Di Wu (PhD student, University of Maryland, College Park)

What: Applied math seminar

When: Friday 27, 2026, 2:00-3:00pm

Where: Phillips 730

 

Title: PINS: Proximal Iterations with Sparse Newton and Sinkhorn for Optimal Transport

 

Abstract: Optimal transport (OT) is a core problem in optimization and machine learning, where we often want both high accuracy and good runtime. Entropic regularization and the Sinkhorn algorithm improve scaling, but they can suffer from numerical instability and slow convergence when the regularization parameter is small. In this talk, I will present Proximal Iterations with Sparse Newton and Sinkhorn (PINS), a method that targets highly accurate solutions for large-scale OT. The method uses proximal updates with a sparse Newton step, together with Sinkhorn structure, to speed up convergence while keeping computations sparse. We provide theory that supports global convergence and explains how sparsity reduces overall cost. Experiments show that PINS reaches accuracy close to unregularized OT solutions, converges faster than related methods, and is less sensitive to the regularization parameter.