Dynamical Stability
We define stability as when an gradient descent equation is stable to perturbations at convergence points.
Recall convergence can [see page 7, happen] either when we reach a peak or minima in the weight/time spectrum. Observe that for a minima if we add some small constant it (moving it along the weight/time spectrum) we can continue traveling down the negative gradient and return back to the minima. For a maxima however the negative gradient will only push us further away from the maxima and towards a minima.
Minima are stable to perturbations. Maxima are unstable to perturbations.