Regional earthquake populations are widely observed to have power-law magnitude-frequency statistics. Whether earthquake distributions on individual faults are similarly power-law is presently the subject of considerable research effort. Recent careful observations suggest that event statistics on rough faults are power-law whereas earthquake distributions on smooth faults are better described by a characteristic model in which there are more large earthquakes than would be expected by extrapolation from the distribution of small events. Here the relation between heterogeneity and the magnitude-frequency distribution is investigated in a cellular automaton that includes new nearest-neighbor stress transfer rules which produce realistic stress concentrations. The results agree with the suggestion that smooth faults have more characteristic earthquake distributions and support the view that an earthquake's size is controlled by rupture arrest. ¿ 1999 American Geophysical Union