On Theoretical Interpretations of Concept-Based In-Context Learning