Because this algorithm doesn't take surfaces facing away from lights
into account, the forward differencing calculation will produce ``lights''
on the surface even when no light is falling on the surface. Use the
result of
to scale the shift so that the bump
effect tapers off slowly as the surface becomes more oblique to the light
direction. Empirically, adding a small bias (.3 in the authors'
experiments) to the dot product (and clamping the result) is more visibly
pleasing because the bumps appear to taper off after the surface
has started facing away from the light, as would actually happen for a
displaced surface.