On large deviations