Interval estimation and Monte Carlo simulation of digital communication systems

The authors quantify the accuracy of bit error rate (BER) estimates produced by Monte Carlo simulations by carefully applying confidence interval estimation techniques. Due to numerical difficulties, some previous work in this area has assumed that the BER statistic possessed a Gaussian distribution. The authors demonstrate that in some important regions the estimate is decidedly non-Gaussian, and application of central limit theorem arguments can result in errors in excess of an order of magnitude. They investigate the accuracy of common approximations and the feasibility of exact calculation of confidence intervals, and present a novel polynomial class approximation. By combining this approximation with more conventional approaches, an algorithm is developed for estimating confidence intervals of BER estimates. The algorithm is nonrecursive and numerically stable, requires a trivial amount of compute time to evaluate, has a small margin of error, and can be used for all error rates less than 0.5