In my opinion, it's because the footprint unit length (length of the square) shrinks linearly but the mipmap shrinks in a log scale. Smaller footprint lengths will be squeezed close to the level 0 mipmap.
Stents-
Near the horizon the L value approaches 0 so the lowest resolution mipmap will be chosen. This lowest mip will also be chosen if the L value is close enough to 0, resulting in a range of values that correspond to the lowest mip. In this case the lowest mip is just a solid color. Trilinear filtering only helps so much here, since we’re blending a solid color with a very blurry image, the result looks like just a solid color.
In my opinion, it's because the footprint unit length (length of the square) shrinks linearly but the mipmap shrinks in a log scale. Smaller footprint lengths will be squeezed close to the level 0 mipmap.
Near the horizon the L value approaches 0 so the lowest resolution mipmap will be chosen. This lowest mip will also be chosen if the L value is close enough to 0, resulting in a range of values that correspond to the lowest mip. In this case the lowest mip is just a solid color. Trilinear filtering only helps so much here, since we’re blending a solid color with a very blurry image, the result looks like just a solid color.