At a point , the maximum growth is in the direction of the gradient (), and the maximum decrease is in the opposite direction (). This can be used for optimization (e.g. stochastic gradient descent).