Gradient descent for quadratic functions diverges when the learning rate exceeds "2 divided by the maximum curvature"

In gradient descent, the loss can start to increase even if the learning rate is raised only slightly. For a quadratic ...