Right now, the scalar * matrix definition promotes the matrix and scalar to autodiff variables. It should just go through and do element-by-element multiplications with a promoted return type. There can be a single definition in math, which would just template out the return type to be promoted. Here's what it current looks like (there was already a relevant FIXME note in the code):
template<typename T1,typename T2,int R2,int C2>
inline Eigen::Matrix<var,R2,C2> multiply(const T1& c,
const Eigen::Matrix<T2, R2, C2>& m) {
// FIXME: pull out to eliminate overpromotion of one side
// move to matrix.hpp w. promotion?
return to_var(m) * to_var(c);
}
There may be other cases like this one, but I think a single implementation can replace the first three definitions of multiply() currently in stan/agrad/rev/matrix/multiply.hpp.
Right now, the scalar * matrix definition promotes the matrix and scalar to autodiff variables. It should just go through and do element-by-element multiplications with a promoted return type. There can be a single definition in math, which would just template out the return type to be promoted. Here's what it current looks like (there was already a relevant FIXME note in the code):
There may be other cases like this one, but I think a single implementation can replace the first three definitions of
multiply()currently instan/agrad/rev/matrix/multiply.hpp.