On Sat, Aug 9, 2008 at 1:58 AM, Roger Hui <[EMAIL PROTECTED]> wrote:
>> 5*y should be statistically valid
>> for far larger samples than our psuedo-random number
>> generator would be good for.
>
> Can you please explain that? The default RNG used
> in ?y has a period of _1+2^19937 (4.32e6001).
> http://www.jsoftware.com/help/dictionary/d640.htm
Ok, it turns out I was woefully wrong here. The problem
is not very obvious for the case I was experimenting
with, but for smaller cases this become ridiculous.
coclass 'poissondistr'
create=: 3 :0
n=: y
possible=: (^-y) +/\@:* 1 */\@, y&%
thresh=: possible 1+i.y
)
rand=: 3 :0
r=. thresh I. t=. y [EMAIL PROTECTED] 0
while. #i=. I. r=#thresh do.
thresh=: possible 1+i.n=: n+1
r=. (thresh I. i{t) i} r
end.
)
cocurrent 'base'
d2=: 2 conew 'poisondistr'
($thresh__d2); 'rand__d2 2^10' ts~ 2^6
+--+----------------+
|13|6.36691e_5 40320|
+--+----------------+
(compare that 13 with 5*2)
--
Raul
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
