Tricky Perlin Noise Question. How does the Grad Function Use Normalized Vectors?

Question

I\'m currently in the process of understanding improved Perlin noise. I understand the theory entirely however, I am confused by an aspect of its common implementation, such