SVMs and the input distribution
If the input distribution has high density near the target boundary, the sample will likely contain points near the boundary, so that largemargin or smallmargin classifiers will be similar. If the input distribution has low density near the boundary, then the sample will have few nearboundary points, giving advantage to a largemargin classifier  but then also, the probability of drawing a nearmargin point during outofsample use is low, so E_out for lowmargin classifiers is not much affected.
Why does this not limit the advantage of largemargin classifiers in practice?
