That is likely to be a very dummy question, but I'm stuck here... I have three groups (10 samples each) and 23 evaluated variables. I'm trying to pick up the subset of variables that best separate samples back in the 3 original groups. I tried pca (no separation at all) and Discrimminant/canonical analysis (since I have 10 samples per group I can only choose 9 variables. Some randomly picked variables perform well (6% error), but my question is about another way (other than randomly testing) to find the subset of variables which gives better separation of groups.