We present an adaptive approach to the recognition of handwritten mathematical symbols, in which a recognition weight is associated with each training sample. The weight is computed from the distance to a test character in the space of coefficients of functional approximation of symbols. To determine the average size of the training set to achieve certain classification accuracy, we model the error drop as a function of the number of training samples in a class and compute the average parameters of the model with respect to all classes in the collection. The size is maintained by removing a training sample with the minimal average weight after each addition of a recognized symbol to the repository. Experiments show that the method allows rapid adaptation of a default training dataset to the handwriting of an author with efficient use of the storage space.