In this letter, we propose a nonlinearity-tolerant channel estimation technique for training-aided single-carrier polarization-division-multiplexed (PDM) coherent optical systems. Simply by compensating the effects of fiber nonlinearity on the received training sequence, linear channel estimation and equalization can be improved; this approach is very efficient, which requires a negligible computational effort since training only occupies a very small portion of the total data. The concept is verified using simulations for both QPSK and 64-quadrature amplitude modulation (QAM) systems and experimental measurements for a 120-Gb/s PDM 64-QAM system with 800-km erbium-doped fiber amplifier (EDFA)-only transmission. The experimental results show that the system performance is improved by 0.4 dB at the optimal launched power and 1.6 dB in the highly nonlinear region, compared with conventional training-aided channel estimation approaches.