Statistical testing of famous software

When using commercially available software, users assume they are working with reliable and accurate software. A model for statistical testing of software was used to investigate the computational accuracy of the elementary functions of three microcomputer software packages. Based on the characterization of the usage environment, test domains were constructed for the functions. Random samples of arguments were generated from the test domains and the values computed. The computed values were compared with "correct results" and the data based on the comparisons were statistically analyzed. The packages were tested for the performance of the functions over critical or problem sub-intervals, reliability of computations within packages, and the analysis of variance among packages. The results indicate that errors in elementary functions do occur, especially at the higher levels of precision, and there are statistically significant differences in the accuracy of the elementary functions across software packages.