[boost] library for differentiation (interest research)

Subject: [boost] library for differentiation (interest research)
From: DE (satan66613_at_[hidden])
Date: 2010-01-19 13:14:53

hi all
recently i have interested in evaluating of derivatives of complicated
functions and discovred that there is a finely developed technique
which is called algorithmic differentiation (visit www.autodiff.org if
you are interested)

the resulting code using an imagined lib may look like this

  struct
  {
    template<typename type>
    type operator()(type x)  //defining function
    { return sin(x); }
  } foo;
  //...
  double
    x = 42.,
    f = foo(x),  //evaluating function value
    df = deriv(foo, x);  //evaluating derivative value

notice that the function' code is untouched
the very amazing thing is that the function may be arbitrarily complex
it can even be a matrix inversion or eigenvalues evaluation so we get
e.g. derivatives of eigenvalues with respect to system parameters
gradients, jacobians and higher order derivatives are also possible

there are numerous libraries implementing this technique and the
questions are:
 - does it fit into boost collection of libraries?
 - is it worth building a distinct boost (well-designed) library or
 adopt a specific lib or to use other (non-boost) libs?
and the like

personally i am going to investigate the so called tangent mode (or
direct mode which is very simple) and probably implement it if i don't
find any suitable implementation for my own purposes

--
Pavel