Even if the learning rate is very large, every iteration of gradient descent will improve the loss (מועד ב 2022)