Well, this guy has it nailed. He cites Flajolet and Martin, which was (I
thought) as good as you could get with only a reasonable amount of memory per
statistic. Unfortunately, their hash table is a one-shot deal; there's no way
to maintain it once the table changes. His incremental update doesn't degrade
as the table changes. If there isn't the same wrangle of patent as with the
ARC algorithm, and if the existing stats collector process can stand the extra
traffic, then this one is a winner.
Many thanks to the person who posted this reference in the first place; so
sorry I canned your posting and can't recall your name.
Now, if we can come up with something better than the ARC algorithm ...