[Rcpp-devel] Ceres nonlinear least squares solver

[Douglas Bates]

Dirk forwarded a posting on Google's blog about the Ceres nonlinear
least squares solver in C++ to his Google+ followers and asked if
anyone was interested in using Rcpp to link to this C++ code.   As a
person with some experience in nonlinear least squares and also having
experience with Eigen (http://eigen.tuxfamily.org) which is used by
Ceres, I decided to take a look.

There are many nice features of Ceres including great flexibility and
both numeric and automatic differentiation.  However, like many
optimization codes it does not allow reverse communications, which
would make for much easier integration with languages like R.
Features like automatic differentiation require that the function for
evaluating residuals be written in pure C++ (i.e. evaluation of an R
expression would not be allowed) so that operators and functions can
be overloaded.   A person has already asked about this on the
ceres-solver Google group - in his case he was interested in defining
the residual function in python.

So it may be worthwhile doing the linkage but the real advantages of
the code will probably be lost. As a first cut it would probably be
best to use only the dense solvers, which are based on Eigen, and not
the sparse solvers based on SuiteSparse as used in the Matrix package.
 Eigen is a template library implemented as header files.  Suitesparse
is C code with a peculiar, idiosyncratic template mechanism based on
Makefiles and accessing the compiled code is not fun.