Brian wrote:
> Does the filter (%+./) solve the big and
> fractional weights issue?
It definitely solves the fractional weights issue, by normalizing everything
to integers. I'll have to remember that, that's a neat tool.
However, it doesn't solve the big weights issue, as e.g. one of the weights
may be a large prime (so GCD is <: 1, meaning in the best case the numbers
get no larger). Actually, because +. is a fairly expensive operation, it
may have exacerbated the performance issue (as opposed to, e.g., Roger's
[0,1)-scaling mechanism which relies on the cheap + function).
Here's a demonstration of the performance issue:
wP =: (([?@#+/@]){(;@((#~)</.]# ;@(i.@#)))@])
0 ": (%"1<./)ts&> '100 wP 5 3 4 2 * ' ,L:0 ] 10 <@":@^ 1+i.5
1 1
3 7
35 56
190 841
2346 6847
BTW, without thinking about it too much, I believe your code is equivalent
to wPE =: (?@# +/) { I.@] , or with your scaling mechanism as a
preprocessor, wPES =: wPE (% +./) .
-Dan
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
