Towards Better Statistical Understanding of Watermarking LLMs