Alvaro Herrera <alvherre@commandprompt.com> writes:
> Jan Urbański wrote:
>> Oh, one important thing. You need to choose a bucket width for the LC
>> algorithm, that is decide after how many elements will you prune your
>> data structure. I chose to prune after every twenty tsvectors.
> Do you prune after X tsvectors regardless of the numbers of lexemes in
> them? I don't think that preserves the algorithm properties; if there's
> a bunch of very short tsvectors and then long tsvectors, the pruning
> would take place too early for the initial lexemes. I think you should
> count lexemes, not tsvectors.
Yeah. I haven't read the Lossy Counting paper in detail yet, but I
suspect that the mathematical proof of limited error doesn't work if the
pruning is done on a variable spacing. I don't see anything very wrong
with pruning intra-tsvector; the effects ought to average out, since the
point where you prune is going to move around with respect to the
tsvector boundaries.
regards, tom lane
