Its nice to see that you've noticed that Perlin's "improved noise" paper forgot to normalize his gradients. I have seen the error migrate into various implementations meant to replace the original...