Three-Point Approximations for Continuous Random Variables

This paper compares a number of approximations used to estimate means and variances of continuous random variables and/or to serve as substitutes for the probability distributions of such variables, with particular emphasis on three-point approximations. Numerical results from estimating means and variances of a set of beta distributions indicate surprisingly large differences in accuracy among approximations in current use, with some of the most popular ones such as the PERT and triangular-density-function approximations faring poorly. A simple new three-point approximation, which is a straightforward extension of earlier work by Pearson and Tukey, outperforms the others significantly in these tests, and also performs well in related multivariate tests involving the Dirichlet family of distributions. It offers an attractive alternative to currently used approximations in a variety of applications.