does it represent a bias towards less variance over the different gradients one can sample at a given point?
Yup, exactly.
Yup, exactly.