Numerical Recipes: Member functions of a class as arguments - function-pointers

Question

My question concerns Numerical Recipes. I'm trying to use the conjugate gradient solver “frprmn.cpp” to minimize the negative log-likelihood function of a problem that depends on data and a bunch of other parameters. In order to have all pertinent parameters and data available to evaluate the log-likelihood, I have written a class, and both the log-likelihood function and the gradient function are member functions of that class. When I call frprmn with these functions as arguments, I get the error message
argument of type 'DP (kalman_yield_only::)(Vec_I_DP&) {aka double (kalman_yield_only::)(const NRVec&)}' does not match 'DP (*)(Vec_I_DP&){aka double (*)(const NRVec&)}'
The objective log-likelihood function and the gradient function are defined as below in a file called kalman_yield_only.cpp, which also defines the class of the name kalman_yield_only.
DP kalman_yield_only::llh(Vec_I_DP& theta)
{
...code...;
DP L_num=...;
return L_num;
}
The gradient function is:
void kalman_yield_only::llh_grad(Vec_I_DP& theta, Vec_O_DP& grad)
{
...code...;
}
The same file also has a member function that calls the solver:
void kalman_yield_only::optimizer(Vec_IO_DP& theta)
{
const double ftol = 1e-6;
double fret;
int iter;
NR::frprmn(theta,ftol,iter,fret,kalman_yield_only::llh,kalman_yield_only::llh_grad);
}
The whole thing is then called in main by
int main(int arg, char* pszArgs[])
{
...code for data and parameters...;
kalman_yield_only ks(...data and parameters...);
...code for theta...;
ks.optimizer(theta);
...code for doing stuff with theta...;
return 0;
}
In kalman_yield_only.h, I have the lines
class kalman_yield_only
{
public:
kalman_yield_only(Mat_IO_DP& , Mat_IO_DP& , Vec_IO_DP& , Vec_IO_DP& , Mat_IO_DP& , Mat_IO_DP& , DP& , DP& , DP& , DP& , DP&, Vec_IO_DP& );
// All sorts of data and parameter object declarations...
// Member functions
DP llh(Vec_I_DP &);
void llh_grad(Vec_I_DP & , Vec_O_DP & );
void optimizer(Vec_IO_DP& );
};
The corresponding constructor in kalman_yield_only.cpp reads:
kalman_yield_only::kalman_yield_only(Mat_IO_DP& Y_in, Mat_IO_DP& Z_in, Vec_IO_DP& vH_in, Vec_IO_DP& c_in, Mat_IO_DP& mT_in, Mat_IO_DP& Q_in, DP& maxZ_in, DP& maxT_in, DP& maxQ_in, DP& maxc_in, DP& tol_in, Vec_IO_DP& Maturities_in )
{
… code...;
}
I think I tried all permutations of adding and deleting the NR:: and kalman_yield_only:: prefixes in the function call, to no avail.
When I compile the example file “xfrprmn.cpp” from the numerical recipes using the same compiler and make-file infrastructure, it compiles and runs without problems. I cannot see much difference to my own code, except that my objective and gradient routines are member functions of a class, so I wonder if this is what trips me up. Any help is thoroughly appreciated.

Answer 1

I found a workaround whereby I adapt the Numerical Recipes routines so as to take as arguments pointers to my class object instead of pointers to objective and gradient functions. Inside the routines, I changed the code such that they call the member functions of the class object they were handed. Would still be interested if there's a more elegant solution.

Answer 2

argument of type 'DP (kalman_yield_only::)(Vec_I_DP&) {aka double
(kalman_yield_only::)(const NRVec&)}' does not match 'DP
(*)(Vec_I_DP&){aka double (*)(const NRVec&)}'
This means you're trying to use a method (which needs a hidden this pointer) where a function is expected.
Numerical Recipes in C++ is not really C++, as it expects you to use global/static variables!!!!
I think I would edit Numerical recipes so that the function you're passing is a template argument
-- means at the place of usage it is not defined, other than it needs a vector argument and returns a double:
template<typename FUNCTION>
frprmn(..., const FUNCTION &_rF, ....)
{
...
}
Writing frprmn like that still lets you pass a normal function pointer.
Then define the function as the operator() of your class containing all the information:
struct YourClass
{
double operator()(const NRVec&) const;
};