When do gradient optimisations converge to saddle points