A measure of how quickly a loss function's gradient can change; smaller is better for stable training.