We aim to improve the accuracy of handwritten Chinese character recognition using two advanced techniques: discriminative feature extraction (DFE) and discriminative learning quadratic discriminant function (DLQDF). Both methods are based on the minimum classification error (MCE) training method of Juang et al. [7], and we propose to accelerate the training process on large category set using hierarchical classification. Our experimental results on two large databases show that while the DFE improves the accuracy significantly, the DLQDF improves only slightly. Compared to the modified quadratic discriminant function (MQDF) with Fisher discriminant analysis, the error rates on two test sets were reduced by factors of 29.9% and 20.7%, respectively.