A shim for implementing ERGM terms (and proposals?) in C++?

[krivit]

This has been discussed for a while and with various parties, and there is clearly some demand for it. Although a C++ ERGM engine already exists (https://github.com/fellstat/ernm/), there are reasons to try to offer the capability for writing extensions for ergm in C++:

1. There is an enormous amount of existing investment into ergm on both C and R side, and terms written within its API, which would have to be rewritten.
2. To replace ergm, ernm would need to be augmented for valued networks, and it is not clear whether its API could easily support operator terms (i.e., terms with other models as arguments).

This is to discuss desiderata and approaches to the problem.

[krivit]

In principle, if hypothetical class constructors CppNetwork(Network *nwp) and CppModelTerm(ModelTerm *mtp) could be made lightweight enough,

extern "C" C_CHANGESTAT(c_mychangestat) {
  CppNetwork nw(nwp);
  CppModelTerm mt(mtp);
  for(auto j: nw.neighbors(head)) mt.stat[0] += nw[tail, j];
  if(edgestate) mt.stat[0] *= -1;
}

could implement, in this case, the change stat for the number of triangles in the network. Here, extern "C" is needed so that R_FindSymbol() in the model initialisation code could find it in the library's namespace.

A simple macro could fold the boilerplate (extern "C" and the constructor calls) into itself, i.e.,

C_CPPSTAT(mychangestat, {
  for(auto j: nw.neighbors(head)) mt.stat[0] += nw[tail, j];
  if(edgestate) mt.stat[0] *= -1;
})

These constructors would be called every time a change stat is evaluated, so they should do the absolute minimum, and, in fact, ideally, they would get optimised away by the compiler.

[krivit]

statnet/ergm@cpp-shim has the beginnings of an implementation.