Workshop: 10th International Workshop on Software Correctness for HPC Applications (Correctness '26)
Authors: Mohit Tekriwal and Ignacio Laguna (Lawrence Livermore National Laboratory) and Yichen Tao and Jean-Baptiste Jeannin (University of Michigan, Ann Arbor)
Abstract: Traditional deterministic error analysis provide worst case error bounds and prior work on probabilistic error analysis do not model structural floating-point properties, leading to looser bounds for cancellation heavy cases. In this work, we use hybrid error analysis to provide tighter error bounds for long summation and dot product numerical kernels. We extend the FPChecker tool to track global and local rounding error to compute key statistics related to structural floating-point properties, such as the Sterbenz lemma, which is used to compute modified Bernstein's tail bound. We then compare our derived error bound with Higham and Mary's probabilistic error bound and show that our bound faithfully models the structural floating-point behavior. Finally, we show the applicability of the derived bounds on a physical example of computing the magnitude of Fourier modes for density wave function used in compressible-flow simulations.
Back to 10th International Workshop on Software Correctness for HPC Applications (Correctness '26) Archive Listing
Back to Full Workshop Archive Listing