Optimization.jl
Optimization.jl copied to clipboard
Manual Hessian of Lagrangian function
For solvers that want the hessian of the Lagrangian (like Ipopt and IPNewton etc.) $\dfrac{d^2}{dx^2} f(x) + \lambda^T c(x)$ we currently need to supply a function for the hessian of $f$ and one that calculates the hessian of each element $c_i(x)$. If the constraints $c(x) \in R^{n_c}$ are computed by an array-valued function, it becomes rather cumbersome and inefficient to rearrange the computations into $n_c$ separate hessian computations with separate sparsity patterns etc.
Would it be possible to supply lag_hess
and lag_hess_prototype
instead of hess, cons_h
and corresponding sparsity patterns?