Confidence v.s. Critique: A Decomposition of Self-Correction Capability for LLMs