Mail Archives: djgpp-workers/1998/12/10/10:36:47
Subj: Re: RNG random()
To: eliz AT is DOT elta DOT co DOT il (Eli Zaretskii)
CC: djgpp-workers AT delorie DOT com
Dear Eli Zaretskii,
On 12-10-98 at 06:07:14 EST you wrote:
>
>
> On Wed, 9 Dec 1998 Kbwms AT aol DOT com wrote:
>
>
> After reading this, I'd say that the sentence about ``unfavorable
> properties'' of the lagged-Fibonacci generators which therefore
> ``should be avoided'' is way too strong. Especially since the paper
> then shows that almost every RNG on Earth has similarly bad properties
> under some circumstances.
>
First, you ought to take issue with L'Ecuyer on whether his statement
``is way too strong.'' Secondly, the paper confines its attention to
multiple-recursive random number generators (MRGs). That hardly qual-
ifies as an indictment of "almost every RNG on Earth." It is true that
LCGs have poor structural qualities, too, but those problems are not
addressed there. L'Ecuyer and others have produced RNGs that do not
suffer from the unfavorable properties of lagged-Fibonacci generators.
When an investigator is faced with a choice of RNGs and is aware of
the shortcomings of some of the choices, the poor choices will be
avoided (or should be).
>
> However, this certainly does NOT mean that we should dismiss `random',
> `rand48', or any other widely-used RNG, just because they are all bad
> anyway. On the contrary, we should IMHO try to provide several
> different RNGs, so that whoever needs to switch them, could do that.
> We also should not stop putting small improvements into the existing
> RNGs: even if these improvements do not (and can not) make their
> worst-case behavior better, they make them better in other aspects.
>
In the case of `random,' I can demonstrate it fails a simple test. You
are not suggesting that such evidence should be ignored, are you?
I suppose the best approach is to report the problem and let others
take their chances. Making the generator at issue ``somewhat'' better
is certainly no match making it 100% fixed.
In the next few weeks I will be submitting to DJGPP three RNGs of
L'Ecuyer's with periods ranging from 2^191 to 2^377 that are free
of the problems discussed by L'Ecuyer.
K.B. Williams
- Raw text -