Explain why multimodal distributions in training samples should be avoided when using the maximum likelihood classifier