Note that LDA likely does not give the optimal separating hyperplane when minimizing out-of-sample error. That honor likely belongs to SVMs, and in practice tweaking the kernel and identifying relevant non-linear relationships generally become more important to finding a reasonable classification boundary.