A general uncertainty relation between learning ability and generalization ability suited to overfitting was revealed in the modeling of BP neural network. Tests of numerical simulation were carried out to determine the field of values of overfitting parameter in the uncertainty relation, and the relations between the value assignment of parameters and the complexity of function category, which can be represented with the aid of complex correlation coefficient. An overfitting discriminant which improves the generalization ability of network was also presented for freeing of overfitting in the modeling of cases using BP neural network.