no, the error bounds hold for any error. But it helps if you have a skewed distribution and capture most of the active items.