Improvements to and limitations of Latin hypercube sampling

Two techniques are proposed to improve the performance of Latin hypercube sampling. To improve the statistics for each variable, it is proposed that realizations for each variable be acquired by finding the probabilistic means of equiprobable disjunct intervals in the variableʹs domain, instead of using the cumulative distribution function directly, as is currently done. To reduce error in correlations between variables (correlated or uncorrelated), it is proposed to perform a single-switchoptimized method on the realizations for one variable at a time, instead of using matrix manipulation, as is the current custom. Limitations in hypercube sampling will be discussed, and numerical results involving a simple Poisson process will be offered.