Variance-informed Rounding Uncertainty Analysis for Floating-point Statistical Models

Advancements in computer hardware have made it possible to utilize low- and mixed-precision arithmetic for enhanced computational efficiency. In practical predictive modeling, however, it is vital to quantify uncertainty due to rounding along other sources like measurement, sampling, and numerical discretization. Traditional deterministic rounding uncertainty analysis (DBEA) assumes that the rounding errors equal the unit roundoff $u$. However, despite providing strong guarantees, DBEA severely overestimates rounding uncertainty. This work presents a novel probabilistic rounding uncertainty analysis called VIBEA. By treating rounding errors as i.i.d. random variables and leveraging concentration inequalities, VIBEA provides high-confidence estimates for rounding uncertainty using higher-order rounding error statistics. The presented framework is valid for all problem sizes $n$, unlike DBEA, which necessitates $nu<1$. Further, it can account for the potential cancellation of rounding errors, resulting in rounding uncertainty estimates that grow slowly with $n$. We show that for $n>n_c(u)$, VIBEA produces tighter estimates for rounding uncertainty than DBEA. We also show that VIBEA improves existing probabilistic rounding uncertainty analysis techniques for $n\ge3$ by using higher-order rounding error statistics. We conduct numerical experiments on random vector dot products, a linear system solution, and a stochastic boundary value problem. We show that quantifying rounding uncertainty along with traditional sources (numerical discretization, sampling, parameters) enables a more efficient allocation of computational resources, thereby balancing computational efficiency with predictive accuracy. This study is a step towards a comprehensive mixed-precision approach that improves model reliability and enables budgeting of computational resources in predictive modeling and decision-making.