Hyunwoo Lee: Accelerating Exact Real Computation Using Dynamic Dependence Analysis

Abstract: In many computational tasks involving continuous data, floating-point arithmetic is often insufficient to guarantee reliable results due to rounding error and the resulting numerical instability. To address this limitation, languages and runtimes that support real numbers directly have been proposed and the most common approach is reiteration-based Exact Real Computation (ERC). However, the reiteration induced computation overhead makes ERC impractical for computations that require high precision. In this talk, we introduce a dynamic dependence analysis technique for ERC, which is designed to reduce redundant computation. By preserving ERC’s exactness guarantees while lowering computation overhead, we aim to make exact real computation viable for problem classes that were previously difficult to solve in practice.