Interesting. I used a similar trick to compute a lower bound on the levenshtein distance. I was able to skip the real levenshtein distance calculation, if the lower bound already exceeded a threshold ( I used it for clustering ).