Hello, I've uploaded to a subversion repository on SourceForge my Lazy Types library. This library provides template classes for dual numbers (used for automatic differentiation) and complex numbers. Expression templates are used in order to evaluate only those components which are actually needed. All of the standard-library math functions (and arithmetic operators) are supported. http://lazytypes.sourceforge.net/ http://sourceforge.net/projects/lazytypes The dual-number formalism is especially useful because it enables the compiler to generate analytic expressions for the derivative of "any" function (where "any" is restricted to those functions which compute using operators and functions the library knows about, which include all of the standard ones). These are *not* finite-difference approximations, but exact answers (to machine precision). Functions which use conditional statements, loops, etc. are fine, although the derivative might not have a meaning if evaluated at a point of discontinuity. This code has been released under the BSL 1.0. Although some Boost headers are used (enable_if, type_traits, etc.), Boost.Proto is not used. Is there any interest in having these libraries in Boost: 1. As-is (plus more documentation, tests, namespace change, etc.) 2. Only if rewritten to use Boost.Proto (which I am willing to do, so long as the performance of the current implementation can be matched). The project Wiki page (to which http://lazytypes.sourceforge.net/ redirects) has some documentation and usage examples. Here is a quick example of taking the derivative (and the second derivative) of a multivariate function: #include <lazy_types.hpp> namespace lt = lazy_types; ... template <typename X> X f(X &v1, X &v2) { using namespace lt; return pow(v1, 3) + pow<3>(v1) + pow<2>(v2 + v1) + v2*v1/5.0 + 3.0; } // A pow<> template is provided which works with the expression-template types. ... lt::dual<double, 2> fv1(2.0, 0), fv2(4.0, 1); lt::dual<double, 2> fv = f(fv1, fv2); std::cout << fv << " " << fv.eps(0) << " " << fv.eps(1) << std::endl; // This computes the derivative of the function f with respect to both of its arguments evaluated at the point where the function itself has been evaluated. // An example of computing the second derivative: template <typename X> X df(X &v1, X &v2, size_t ei) { lt::dual<X, 2> d1(v1, 0); lt::dual<X, 2> d2(v2, 1); return f(d1, d2).eps(ei); } lt::dual<double, 2> df0 = df(fv1, fv2, 0); lt::dual<double, 2> df1 = df(fv1, fv2, 1); std::cout << df0.eps(0) << " " << df0.eps(1) << std::endl; std::cout << df1.eps(0) << " " << df1.eps(1) << std::endl; If there is interest in adding these types (or some rewritten version of them) to Boost, please let me know, and I'll work on getting the library ready for review. Sincerely, Hal Finkel