Through maximalizing the margin, we can obtain the optimal projection vector, and avoid the small sample size problem due to singularity of the within-class scatter.