Fitting peaks with very low statistics