Deep Learning Gradient Check

When it comes to test how correct the gradient in deep neural nets, how can we check? I have found good solution though, couldn't understand why it requires us to calculate the estimate gradient twice.