Closed lindonroberts closed 4 years ago
Looking online, consensus seems to be that just using a full square matrix is better for symmetric matrices, since LAPACK assumes this anyway. So, just use a normal np.array instead of Hessian class.
Fixed in v1.2 (using np.ndarray for Hessian throughout)
Speed up trust region solve by implementing Hessian-vector multiplication in C (e.g. with Cython).