It should be possible to only evaluate the objective to check the trustregion. It should also be possible to only update on success if quasi-newton. For Newton's method this might not be a good idea because the hessian can often be evaluated with the gradient and objective cheaply. Hope for many accepts. We should accept and return QNVars, not individual things.
It should be possible to only evaluate the objective to check the trustregion. It should also be possible to only update on success if quasi-newton. For Newton's method this might not be a good idea because the hessian can often be evaluated with the gradient and objective cheaply. Hope for many accepts. We should accept and return QNVars, not individual things.