Empirical Evaluation of Numerical Mathematical Libraries

The accuracy of mathematical computations using floating-point numbers has a strong impact on the reliability of real-life systems. This presentation will provide a comprehensive analysis of the behavior of ten numerical mathematical libraries and evaluate them based on three criteria: consistency of results across libraries, rounding and accuracy, and execution time. The analysis aims to spot rounding and approximation errors and to identify groups of libraries with similar behavior. Finally, is presented an analysis of the grouping of the libraries from a behavioral perspective and discuss the results with respect to their historical and implementation wise relations.