On-line cursive handwriting recognition is both a pattern recognition and search problem. Its computing complexity is very high. Some new methods are proposed to overcome these problems. First, some prior knowledge, such as middle of primitive and delayed stroke, is used to adjust the reference line. Then, based on the special primitive attribute of handwriting cursive, a fast valid decoding algorithm, that is, the primitive level building beam viterbi (PLBBV), is presented. Finally, a combining classifiers algorithm that uses all the classifiers' measurement to make decision is represented. The tests on the Unipen training set and the laboratory set reveal the feasibility of the proposed method.