On dimensionality and sample size in statistical pattern classification