For convex objectives with Lipschitz-continuous gradients, SGD with a decaying learning rate satisfying