In this paper, optimal training symbols are designed for training-based least-square channel estimation among various constant-envelope orthogonal frequency-division multiplexing (OFDM) systems to provide minimum achievable mean-squared error (MSE) performance under quasi-static multipath fading channels. In particular, the closed-form expressions of optimal training symbols are specifically given along with the minimum achievable MSE expressions. Simulation results show that the channel estimation technique using optimal training symbols outperforms the technique using other training symbols previously designed for OFDM.