Re: Is this a numerical problem?

[George Barrick]

 

Re: Is this a numerical problem?

 

                                          2010.08.19.14:16:35 UT

 

Hey GNU-Octave folks,

 

      May I suggest that you cut the guy some

slack?  Lukas, that is.  You should not be

telling him to submit a bug report, because

the phenomenon that he is describing is not

a bug.  If you're telling him to submit

because his bug report will be ignored; OK.

 

      A lot of the people who use Octave do not

have a thorough and complete education about

the issues that come up with the representation

of rational numbers as binary floating-point.

The thing is, that we're using 64 bits to

describe a number, of which twelve are used

as exponent.  Just 56 bits make up the mantissa

(some IEEE formats use tricks to achieve 57).

 

      The question a user should always ask is:

 

   Can my number be represented exactly

     using 56 bits?

 

--or--

 

   Can the number be represented exactly in

     base 16 using 14 hexadecimal digits?

 

      If the answer is yes, then the floating

point representation will not have much effect

upon rounding, truncation, modulus calculations,

etc.  If the answer is no, then the user should

expect the unexpected.

 

      I'm just saying that, on a forum such as

this, we should give the guy a scientific answer.

We should not be flippant, and tell him to submit

a bug report that will be ignored.

 

George                 gbarrick_at_kent_dot_edu

                                gbarrick_at_walsh_dot_edu