From: Joel Young <jdy@cs.brown.edu>
I believe there is a numeric error in big_integer_0.3. I am working to isolate it to a single operation.
Ok. I have it isolated to a piece of test code: ///////////////////////////////////////////////////////////////// #include <iostream> #ifdef _CLN #define BOOST_BIG_INTEGER_USE_CLN #elif _GMP #define BOOST_BIG_INTEGER_USE_GMP #endif #include "boost/big_integer.hpp" #include "boost/rational.hpp" typedef boost::big_integer bi; typedef boost::rational<bi> data_t; // Main Program int main(int argc, char* argv[]) { data_t a = data_t(bi("3280993285"))/bi("281474976710656"); data_t b = data_t(bi("-2706319761"))/bi("140737488355328"); std::cout << (a - b) << '\n'; /* __correct__ values: 8693632807/281474976710656 __incorrect__ values: 4398665511/281474976710656 */ } ///////////////////////////////////////////////////////////////// If you compile the above code with cln or gmp support it gets the correct answer (checked with matlab and with double in C++) and if you compile it with plain old big_integer then it gets the wrong answer. Above was using gcc 3.3.2 on redhat 9 Joel