Printable PDF
Department of Mathematics,
University of California San Diego


Ilse Ipsen

North Carolina State University

BayesCG: A probabilistic numeric linear solver


We present the probabilistic numeric solver BayesCG, for solving linear systems with real symmetric positive definite coefficient matrices. BayesCG is an uncertainty aware extension of the conjugate gradient (CG) method that performs solution-based inference with Gaussian distributions to capture the uncertainty in the solution due to early termination. Under a structure exploiting 'Krylov' prior, BayesCG produces the same iterates as CG. The Krylov posterior covariances have low rank, and are maintained in factored form to preserve symmetry and positive semi-definiteness. This allows efficient generation of accurate samples to probe uncertainty in subsequent computations.

January 20, 2022

11:30 AM 
(the passcode is the first prime number > 100)