We propose a theoretical framework based on asymptotic analysis to determine optimal sample transfer quantities in multi-source transfer learning, yielding an efficient algorithm (OTQMS) that enhances accuracy and data efficiency.