It seems that I have a rounding issue with SAS PROC FORMAT

It seems that I have a rounding issue with SAS PROC FORMAT. The code is as below:

proc format;
value testf 
    -1.85E-13--1.85E-13 = 'negative'
    0-0 = 'zero';
run;

data test;
    input number best32.;
    cards;
0
;
run;

data test2;
    set test;
    format varlabel $50.;
    varlabel = put(number, testf.);
run;

The code is easy: first, create a format, with two options, a negative number really closed to zero and zero itself. Second, create dataset test, with a single observation whose number variable equ开发者_如何转开发als to 0. Third, create another dataset, with applied format. I expect to see test2 with number = 0 and varlabel = 'zero' but actually I see test2 with number = 0 and varlabel = 'negative'. Anyone knows why and how to fix this problem? Much appreciated for any suggestion/help.

This is a floating point precision problem. When testing floating point values for equality SAS employs a "fuzz" value of 1e-12 (by default). The expression a = b evaluates to true if abs(a-b) < 1e-12. In your case, 1.85e-13 is sufficiently close to zero that SAS will give it the format 'negative' (why it always assigns this and never 'zero' I don't know).

One way of handling this is to reduce the fuzz value SAS associates with your format:

proc format; 
  value testf (fuzz=1E-13) -1.85E-13--1.85E-13 = 'negative' 0-0 = 'zero'; 
run;

This format assigns the expected varlabel 'zero'.

If this sort of thing is a practical problem for you, you might want to think carefully about potential numeric precision issues in your code, as they're capable of generating some nasty and difficult-to-spot bugs in the wrong circumstances.

This link gives some information on numeric representation issues in SAS which may prove helpful.