Quote - For instance, changing the direction of the normal. You have the normal vector, a direction in 3D space, and to shift that you need two numbers, whether polar coordinates or local cartesian. And a greyscale only gives you one.

They don't change the normal directly. Instead the image you load for the bump acts as a height map, ie the renderer uses the 2D pixel field and calculates slopes between adjacent pixels. This gives you a vector that you can use to reproduce the same 'bump' in the set of normals of your surface. You control the amount of bump by controlling the magnitude of this vector.

As for gradient bump, that's just poser's way of doing normal maps. Normal maps are colored vector maps, with each pixel (RGB) storing the vector (xyz) of a normal. The renderer then just reads the normal from the file and pastes it on the surface. This way you get the same result as a bump but with far less calculation. The limitation being that since the vector is fixed, you cant change the amount of bumpiness easily.

People greyscale textures because people are lazy, even though the map you get this way is often wrong/inverted.