Date:         Thu, 7 Jul 2005 07:10:30 -0700
Reply-To:     Jack Hamilton <jfh@STANFORDALUMNI.ORG>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Jack Hamilton <jfh@STANFORDALUMNI.ORG>
Subject:      Re: precision & size limits
Comments: To: Richard Ristow <wrristow@MINDSPRING.COM>
In-Reply-To:  <5.1.0.14.2.20050707012427.05ea5a60@pop.mindspring.com>
Content-Type: text/plain; charset="us-ascii"

Implementing infinite-precision arithmetic in the data step and PROC IML
would probably suffice - can't many other statistical procedures be emulated
in IML?



--
Jack Hamilton
Oakland, California




> -----Original Message-----
> From: SAS(r) Discussion [mailto:SAS-L@LISTSERV.UGA.EDU] On
> Behalf Of Richard Ristow
> Sent: Wednesday, July 06, 2005 10:37 pm
> To: SAS-L@LISTSERV.UGA.EDU
> Subject: Re: [SAS-L] precision & size limits
>
> At 01:10 AM 7/7/2005, Jack Hamilton wrote:
>
> >I've asked for an infinite (or 1 x 10**32000, which is
> pretty close to
> >infiite for most purposes) data type in SAS, similar what already
> >exists in REXX, but have always been told that it would be too
> >slow.  Too slow compared to what, I wonder - calculating numbers by
> >hand?
>
> I'd have called computation speed a non-starter. If
> arithmetic is, say,
> three orders of magnitude slower (probably a great over-estimate,
> though I'd expect at least one order of magnitude), well, the existing
> data representation is still there and still fast. If you do need huge
> dynamic range or huge precision, arithmetic three orders of magnitude
> slower is surely faster than "calculating numbers by hand".
>
> I would guess the proposal founders on three issues:
> A. It's very rare to need it.
> B. Statistical procedures almost always do arithmetic
> combining several
> variables. You'd need not just arithmetic for the new representation,
> but arithmetic to coerce standard form to expanded and then calculate.
> C. This sounds like a deal-breaker: Most statistical procedures have
> internal data structures, for things like SSCP matrices. Those are
> almost certainly built around the assumption that all numbers are 8
> bytes, and all of the same, standard, datatype. (The latter means that
> the structure doesn't need any indication of datatype.) Probably most
> statistical procedures would have to be heavily rewritten around a new
> structure that accepts multiple datatypes.
>
> There's also changing the representation of the PDV, but I expect that
> would be much simpler.
>