I agree with earlier comments that the experiments performed were not sufficient, were somewhat confusing, furthermore did not entirely support the theoretical results derived in the paper. As Dave noted earlier only about 50% of the experiments actually conformed to the theoretical results.

I was not entirely sure about how general the results were. I think generative and discriminative classifiers chosen in the paper make assumptions about the distribution of samples that are not made by other classifiers in general, hence the generality of the proofs is questionable.