Default Interval rounding policies incomplete

[Steven Robbins <smr@…>]

Report originates at ​http://bugs.debian.org/440178

The rounding policy requirements (​http://boost.org/libs/numeric/interval/doc/rounding.htm) list e.g. tan_down(), but none of the implementations in rounded_arith.hpp implement it.

The result is that this code fails to compile:

#include <boost/numeric/interval.hpp>

int main( int ac, char* av[] )
{
    boost::numeric::interval<double> I(0.1, 0.2);
    I = tan(I);

    return 0;
}

[Steven Robbins <smr@…>]

This bug is still present in 1.35.0.

[Ben Galehouse <bgalehouse@…>]

According to the documentation at http://www.boost.org/doc/libs/1_35_0/libs/numeric/interval/doc/rounding.htm in the "transcendental function" section, the standard library routines for tan, etc, do not typically satisfy the needed rounding properties and therefore the templates which implement them are disabled by default.

I have used tan with the interval library. If you pass it a policy based on rounded_transc_std, it works fine. e.g. I have written a rounded_control specialization for the mpfr_class type from the gmpfrxx interface to mpfr. (​http://math.berkeley.edu/~wilken/code/gmpfrxx/) With it I can declare the specialization

template<>
struct rounded_math<mpfr_class> 
  : save_state_nothing<rounded_transc_std<mpfr_class> >
{};    

and then code like

 j =  boost::numeric::interval<mpfr_class> (0.1,0.2);
  j = boost::numeric::tan(j);

  std::cout << "[" << j.lower() << "," << j.upper() << "]" << "\n";

compiles, executes, and gives plausible looking results. Therefore, I think this is working as designed.