Using a "pure L1" metric for tf analysis