[18855] in s-news-athena
Re: [S] Does the new version of S-Plus fix num acc. and RNG problems?
daemon@ATHENA.MIT.EDU (Tim Hesterberg)
Wed Aug 4 14:27:29 1999
Date: Wed, 4 Aug 1999 11:20:49 -0700 (PDT)
Message-Id: <199908041820.LAA01245@quake.statsci.com>
From: Tim Hesterberg <timh@statsci.com>
To: s-news@wubios.wustl.edu
Reply-To: timh@statsci.com (Tim Hesterberg)
H. D. Vinod wrote:
>Does the new version of S-PLUS solve some of the serious
>numerical accuracy problems pointed out in McCullough (1999)?
> ...
We urge everyone to read the article,
"Assessing the Reliability of Statistical Software: Part II", 1999
B. D. McCullough, The American Statistician 53(2), 149-159.
In general S-PLUS did very well, better than the competitors.
There were some areas where S-PLUS was criticized, some of which
have been discussed on S-news. Our comments are available at:
www.statsci.com/Hesterberg/reliability
A quick summary of the comments:
(1) The period of the S-PLUS random number generator (RNG) is about
4.6*10^18, not 2.1*10^9. The RNG is documented in Venables and Ripley
(1997) page 167.
(2) The RNG fails a number of DIEHARD tests for randomness, including
some which the article indicated it passed. The primary reason for
these failures is that the RNG produces 31 random bits (the value of
2^32*runif(n) is always even), while the tests assume 32 random bits.
This has an insignificant effect on normal uses of the RNG. For users
who need 32 random bits, a patch is available that pastes together 16
bits from each of two consecutive numbers; the result passes all
DIEHARD tests.
(3) A patch is available for "pchisq" and "pgamma", so that
pchisq(x, y) and pgamma(x, y) will be computed more accurately for
x > 10^11 and y ~= x.
(4) Minor corrections to the article.
We wish to thank McCullough, Brian Ripley, and Douglas Bates for
providing corrections and other information that are included in
the comments.
Tim Hesterberg
========================================================
| Tim Hesterberg Research Scientist |
| timh@statsci.com www.statsci.com/Hesterberg |
| (206)283-8802x319 MathSoft, Inc. |
| (206)283-0347 (fax) 1700 Westlake Ave. N, Suite 500 |
| Seattle, WA 98109-3044, U.S.A. |
========================================================
-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news
