The previous couple blog posts touched on a special case of sphere packing.
We looked at the proportion of volume contained near the corners of a hypercube. If you take the set of points within a distance 1/2 of a corner of a hypercube, you could rearrange these points to form a full ball centered one corner of the hypercube. Saying that not much volume is located near the corners is equivalent to saying that the sphere packing that centers spheres at points with integer coordinates is not very dense.
We also looked at centering balls inside hypercubes. This is the same sphere packing as above, just shifting every coordinate by 1/2. So saying that a ball in a box doesn’t take up much volume in high dimensions is another way of saying that the integer lattice sphere packing is not very dense.
How much better can we pack spheres? In 24 dimensions, balls centered inside hypercubes would have density equal to the volume of a ball of radius 1/2, or (π/2)12 / 12!. The most dense packing in 24 dimensions, the Leech lattice sphere packing, has a density of π12 / 12!, i.e. it is 212 = 4096 times more efficient.
The densest sphere packings have only been proven in dimensions 1, 2, 3, 8, and 24. (The densest regular (lattice) packings are known for dimensions up to 8, but it is conceivable that there exist irregular packings that are more efficient than the most efficient lattice packing.) Dimension 24 is special in numerous ways, and it appears that 24 is a local maximum as far as optimal sphere packing density. How does sphere packing based on a integer lattice compare to the best packing in other high dimensions?
Although optimal packings are not known in high dimensions, upper and lower bounds on packing density are known. If Δ is the optimal sphere packing density in dimension n, then we have the following upper and lower bounds for large n:
The following plot shows how the integer lattice packing density (solid line) compares to the upper and lower bounds (dashed lines).
The upper and lower bounds come from Sphere Packings, Lattices, and Groups, published in 1998. Perhaps tighter bounds have been found since then.