I have now completed a compile time version of double which (as far as I have been able to test) the same results as the runtime equivalent. I have implemented mpl-support for all arithmetic operations plus all comparison operations. To test the accuracy, I have tested it against Cromwells sine-implementation (which by the way had a bug: In stead of 1-(angle*angle*/((2*i+2)*(2*i+3))*next_term it had (angle*angle*/((2*i+2)*(2*i+3))*next_term-1) and got the same result as a runtime-equivalent implementation of the same algorithm. There are four ways to construct a double: integral_to_double<mpl::int_<-355> > or integral_c_to_double<int,-355>::type string_c_to_double<3,'.',1,4,1,5,9,2,6,5,3,5,8,9,7,9,3,2,3,8,4,6,2,6,4,3,3,8,3,2,7,9> typedef BOOST_DOUBLE(3.141592653589793238462643383279502884197169399375105820974944592) pi; double_<mantissa<0x6487ed51,0x085a3000>,1,0> <-Equivalent of pi. The algorithm behind string_c_to_double is not very efficient, so if anyone has a better idea for an implementation, feel free to contribute :) BOOST_DOUBLE is a non-compliant macro that relies on compile time behaviour of floating operations. I have not yet added code to handle +-infinity,QNan,and SNan, but this is relatively easy. 0.0 is not represented correctly (has wrong exponent when converted to double) +0.0 does not equal -0.0 yet. On the upside, this implementation uses a 61 bit mantissa, so it is possible to utilize the full mantissa without rounding when calculating e.g. sine. This is done using boost::metamath::detail::plus_double_double<a,b>::type boost::metamath::detail::times_double_double<a,b>::type boost::metamath::detail::divides_double_double<a,b>::type The user should then deal with rounding before returning the final value: boost::metamath::detail::round<a>::type; or boost::metamath::detail::cutoff<a>::type; if no rounding is wanted. The source is available from the vault: http://boost-sandbox.sourceforge.net/vault/ <metamath.zip> Regards, Peder Holt