RE: [boost] Re: Re: test_fp_comparisons and rounding errors

From: Rozental, Gennadiy (gennadiy.rozental_at_[hidden])
Date: 2003-07-11 13:35:00

> > This is what BOOST_CHECK_CLOSE is doing basically.
> >
> Yes, that's the point.
> What I proposed is that the user can combine the macro with
> the comparator (I changed the macro name), because the 
> comparison method is often application specific. I suggested 
> to have an absolute error comparator class as a default, but 
> not as a fixed method inside the macro. The problem with any 
> fixed scheme is that it can't be general enough, and the test 
> library is highly general.
> I'm not sure if any fixed comparison method should be 
> supplied at all. I would choose between my proposal, were 
> specific targets can supply their own comparators, or not 
> having fp comparison at all in the test library itself.

I believe Boost.Test need to provide at least some way to check that
calculated fP value is as expected. I couldn't use BOOST_EQUAL and do not
want to write predicate every time I need to validate floats. I won't be
generic enough to cover all the cases. But in majority of the cases it
should work. In all other cases you could use BOOST_CHECK_PREDICATE with
custom algorithm. It also will show mismatched values (at least try to).
 
> >For majority of simple cases (not very big, nor very small 
> values) they 
> >should behave similarly. For edge cases relative error comparison 
> >produces better results IMO
> 
> No. It is not a matter of value magnitude or edge/non-edge 
> cases. It is a matter of the number and type of operations 
> performed yielding the values to be compared. This is why the 
> best method is highly domain specific.

Tolerance is indeed highly domain specific. And I do not force you with how
to choose one. As for what is better fit in more cases absolute vs. relative
comparison, I believe it's second. Could you show me an example when
relative error comparison isn't correct while absolute is? While there many
example of reverse situation. Most of the references I have on FP values
comparison recommends to use relative errors either.
 
> Fernando Cacciola

Gennadiy.