Date:         Fri, 22 Aug 2003 09:19:34 +0100
Reply-To:     Roland <roland@RASHLEIGH-BERRY.FSNET.CO.UK>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Roland <roland@RASHLEIGH-BERRY.FSNET.CO.UK>
Organization: Universe Monitors
Subject:      Re: Tiny Roundoff Errors in 8.2 for Windows

You might find this page of interest:
http://www.datasavantconsulting.com/roland/floating.html

"Talbot Michael Katz" <topkatz@MSN.COM> wrote in message
news:200308220514.h7M5E6705418@listserv.cc.uga.edu...
> <<Sound of hand slapping forehead>>  Of course!  Thanks for the lucid
> explanation, Tim.
>
> On Fri, 22 Aug 2003 03:44:40 GMT, Tim Berryhill <tim@AARTWOLF.COM> wrote:
>
> >Dear TMK, It is a documented feature.
> >
> >Most numbers which can be expressed as precise base 10 decimals cannot be
> >expressed as precise base 16 heximals.  SAS numbers are floating point,
> >which means base 16 internally.  The amounts the numbers are adjusted
when
> >you round them are not perfect, so that after you add or subtract rounded
> >numbers you need to round again to keep them pretty.
> >
> >You might get more acceptable results by formatting your numbers 10.2 (or
> >one of the financial formats, as long as you specify 2 decimals) for
> >reporting, including formatting the tr_amount variable.  If your customer
> >accepts any rounding, they should accept the rounding implicit in the
10.2
> >format.
> >
> >My SAS machine is 30 miles away, but I suspect
> >DIFF=3-(0-3);
> >will also give you a tiny discrepancy because "twos complement" messed up
> >the last bit.
> >
> >Enjoy,
> >Tim Berryhill
> >
> >"Talbot Michael Katz" <topkatz@MSN.COM> wrote in message
> >news:200308220252.h7M2qR707234@listserv.cc.uga.edu...
> >> Hi, gang.
> >>
> >> I seem to be experiencing very small arithmetic roundoff errors in SAS
> 8.2
> >> for Windows.  Here is the code snippet :
> >>
> >>  run_bal = round(run_bal+tr_amount,0.01);
> >>  bal_dif = run_bal - prod_balance;
> >>  tr_amount = -bal_dif;
> >>  if tr_amount ne round(tr_amount,0.01) then do;
> >>   put "prec_num = " prec_num " tr_amount not round 27";
> >>  end ;
> >>
> >> Note that run_bal is rounded to 0.01 (i.e., dollars and cents) at the
> top.
> >> Prod_balance has also been rounded to 0.01, so bal_dif should be, too.
> >> But, somehow, tr_amount = -bal_dif is off by about 10**(-12).
> >>
> >> Has anyone else experienced something like this?  Is it a documented
> >> problem?
> >>
> >> --  TMK  --