The Filtered Nearest Neighbor Method for Generating Low-Discrepancy Sequences

We introduce the filtered nearest neighbor method for generating low-discrepancy random-number sequences. Simulations show that these sequences have lower discrepancy than either pseudo random or quasi random numbers when used to generate small samples (N ≤ 300) in high dimensions (p ≥ 24). They are therefore useful for approximating high-dimensional expectation integrals when function evaluation is expensive.