We first performed a set of simulations to determine how well the proposed methods were at classification. We generated p = 1000 dimensional vectors for two populations. We considered the following sample size combinations (n0, n1) = (15, 15), (20, 10), (50, 50), and (70, 30), where nk is the number of samples in the group with g = k (k = 0, 1). All the genes were assumed to be independent with a normal distribution and variance 1. We assumed a model in which a fraction π of the genes was differentially expressed between the two classes, π = 0.05 and π = 0.5 were considered. We examined two scenarios. For the first scenario, there was a big change in differential expression in the differentially expressed genes, a shift of 5 units in the mean. In the second scenario, the fold change was only a 1.5 unit difference in mean. For each simulation setting, 100 datasets were generated, and the classification error rates were estimated using three-fold cross-validation. No optimization was performed; we set λ = 10. The results
