The class is a simple one, no expression templates are used. The class is however a template, meaning that any base numeric type can be used with it. It has successfully tested with boost::multiprecision::mpfr.
## Scalar version
The scalar one allows very easily to produce higher order derivatives.
## Vector version
The vectorized one is harder to make work with higher order derivatives, but allows the simultaneous computation of the full gradient, in a single function call, making it more efficient than backward automatic differentiation.
It currently depends on Eigen for the vectorized part.
