Sphere packing

Henry Cohn

IAP Math Lecture Series

January 16, 2015
The sphere packing problem
How densely can we pack identical spheres into space?
Not allowed to overlap (but can be tangent).
Density = fraction of space filled by the spheres.
Why should we care?

The densest packing is pretty obvious.

It’s not difficult to stack cannonballs or oranges.

It’s profoundly difficult to prove (Hales 1998, 2014).

But why should anyone but mathematicians care?

One answer is that it’s a toy model for:

Granular materials.
Packing more complicated shapes into containers.
Sphere packing is a first step towards these more complex
problems.
Varying the dimension
What if we didn’t work in three-dimensional space?

The two-dimensional analogue is packing circles in the plane.

Still tricky to prove, but not nearly as difficult (Thue 1892).

What about one dimension? What’s a one-dimensional sphere?

Spheres in different dimensions
Sphere centered at x with radius r means the points at distance r from x.
r

Ordinary sphere in three dimensions, circle in two dimensions.

Just two points in one dimension:
r
x

The inside of a one-dimensional sphere is an interval.

r
x
One-dimensional sphere packing is boring:

(density = 1)

Two-dimensional sphere packing is prettier and more interesting:

(density ≈ 0.91)

Three dimensions strains human ability to prove:

(density ≈ 0.74)

What about four dimensions? Is that just crazy?

Some history
Thomas Harriot (1560–1621)
Mathematical assistant to Sir Walter Raleigh.
A Brief and True Report of the New Found Land of Virginia (1588)

First to study the sphere packing problem.

Claude Shannon (1916–2001)
Developed information theory.
A Mathematical Theory of Communication (1948)

Practical importance of sphere packing in higher dimensions!

We’ll return to this later.
Distances and volumes in Rn

The distance between (a, b, c, d) and (w , x, y , z) is

q
(a − w )2 + (b − x)2 + (c − y )2 + (d − z)2 .

Just like two or three dimensions, but with an extra coordinate.

(n-dimensional Pythagorean theorem)

4d volume of right-angled a × b × c × d box =

product abcd of lengths in each dimension.

Just like area of a rectangle or volume of a 3d box.

Higher dimensions work analogously. Just use all the coordinates.

Why should we measure things this way? We don’t have to.

We could measure distances and volumes however we like, to get

different geometries that are useful for different purposes.

But we’re going to focus on Euclidean geometry today.

Applications

Anything you can measure using n numbers is a point in n

dimensions.

Your height, weight, and age form a point in R3 .

Twenty measurements in an experiment yield a point in R20 .

One pixel in an image is described by a point in R3 (red, green, and

blue components). A one-megapixel image is a point in R3000000 .

Some climate models have ten billion variables, so their states are
points in R10000000000 .

High dimensions are everywhere! Low dimensions are anomalous.

All data can be described by numbers, so any large collection of
data is high dimensional.
Big data

Classical statistics originated in an information-poor environment.

Goal: extract as much information as possible from limited data.

Put up with a high ratio of computation and thought to data.
Only valuable data is collected, to answer specific questions.

Nowadays massive data sets are common. We often collect vast

quantities of data without knowing exactly what we are looking for.

How can we find a needle in a high-dimensional haystack?

Curse of dimensionality
Volume scales like the nth power in n dimensions:
rescaling by a factor of r multiplies volume by r n .
Exponential growth as the dimension varies!
Volume of 2 × 2 × · · · × 2 cube in Rn is 2n .
210 = 1024
2100 = 1267650600228229401496703205376

21000 = 107150860718626732094842504906000181056140481170
553360744375038837035105112493612249319837881569
585812759467291755314682518714528569231404359845
775746985748039345677748242309854210746050623711
418779541821530464749835819412673987675591655439
460770629145711964776865421676604298316526243868
37205668069376
Why is exponential volume growth a curse?

Exponentials grow unfathomably quickly, so there’s a ridiculous

amount of space in high dimensions.

Naively searching a high-dimensional space is incredibly slow.

Nobody can search a hundred-dimensional space by brute force,
let alone a billion-dimensional space.

To get anywhere in practice, we need better algorithms.

But that’s another story for another time. . .
How can we visualize four dimensions?
Perspective picture of a hypercube.
Shadow cast in a lower dimension (“projection”).

Fundamentally the same as a perspective picture.

Projections can get complicated.
Cross sections
Cubes have square cross sections.
Hypercubes have cubic cross sections.

Cubes have other cross sections too:

So do hypercubes:
Hinton cubes
Charles Howard Hinton (1853–1907)
A New Era of Thought, Swan Sonnenschein & Co., London, 1888

(Hinton introduced the term “tesseract” and invented the automatic pitching machine.)
Alicia Boole Stott (1860–1940)
On certain series of sections of the regular four-dimensional
hypersolids, Verhandelingen der Koninklijke Akademie van
Wetenschappen te Amsterdam 7 (1900), no. 3, 1–21.
Information theory

Shannon’s application of high-dimensional sphere packings.

Represent signals by points s in Rn .

E.g., radio with coordinates = amplitudes at different frequencies.

In most applications, n will be large. Often hundreds or thousands.

Send stream of signals over this channel.

Noise in the communication channel
The key difficulty in communication is noise.

Send signal s in Rn ; receive r on other end.

Noise means generally r 6= s.

The channel has some noise level ε, and r is almost always within
distance ε of s.

Imagine an error sphere of radius ε about each sent signal, showing

how it could be received.
ε

r
s
How can we communicate without error?
Agree ahead of time on a restricted vocabulary of signals.

If s1 and s2 get too close, received signals could get confused:

r
s1 s2

Did r come from s1 or s2 ? Therefore, keep all signals in S at least

2ε apart, so the error spheres don’t overlap:

2ε

s1 s2
This is sphere packing!

The error spheres should form a sphere packing. This is called an

error-correcting code.

For rapid communication, want as large a vocabulary as possible.

I.e., to use space efficiently, want to maximize the packing density.

Rapid, error-free communication requires a dense sphere packing.

Real-world channels correspond to high dimensions.

Of course some channels require more elaborate noise models, but

sphere packing is the most fundamental case.
What is known?

Each dimension seems to behave differently.

Good constructions are known for low dimensions.

No idea what the best high-dimensional packings look like (they

may even be disordered).

Upper/lower density bounds in general.

Bounds are very far apart:

For n = 36, differ by a multiplicative factor of 58.
This factor grows exponentially as n → ∞.
Packing in high dimensions

On a scale from one to infinity, a million is small, but we know

almost nothing about sphere packing in a million dimensions.

Simple lower bound: can achieve density at least 2−n in Rn .

How to get density at least 2−n
Consider any packing in Rn with spheres of radius r ,
such that no further spheres can be added without overlap.

No point in Rn can be 2r units away from all sphere centers.

I.e., radius 2r spheres cover space completely.

uncovered point could be center of new sphere

Doubling the radius multiplies the volume by 2n .

Thus, the radius r packing has density at least 2−n (since the
radius 2r packing covers all of space). Q.E.D.

This is very nearly all we know!

Nonconstructive proof

That proof showed that the best packing in Rn must have density
at least 2−n .

But it didn’t describe where to put the spheres in an actual

packing. It’s a “nonconstructive proof.”

In fact, nobody has any idea how to build such a good packing.
Our constructions are all much worse.

For example, using a rectangular grid gives density

π n/2
(n/2)!2n

(this is not obvious), and the factorial ruins everything.

Packing in high dimensions (large n)
We just showed:
density at least 2−n .

Hermann Minkowski (1905):

at least 2 · 2−n .

. . . [many further research papers]. . .

Keith Ball (1992):

at least 2n · 2−n .

Stephanie Vance (2011):

at least e6 n · 2−n when n is divisible by 4.

Akshay Venkatesh (2012):

−γ
at least e 2 n log log n · 2−n for certain sparse sequence of dimensions.
First superlinear improvement on 2−n !

For comparison, the best upper bound known is 2−0.599n , due to Grigorii
Kabatiansky and Vladimir Levenshtein (1978).
 3
log(density) vs. dimension
2

0
Upper bound by Cohn & Elkies
-1 Best packing known
Lower bound by Ball
-2

-3

-4

-5

-6
4 8 12 16 20 24 28 32
-7

-8

-9

-10

-11

-12

-13

-14
The most remarkable packings

Certain dimensions have amazing packings.

R8 : E8 root lattice
R24 : Leech lattice [named after John Leech (1926–1992)]

Extremely symmetrical and dense packings of spheres.

Must be optimal, but this has not been proved.

Connected with many areas in mathematics, physics

(such as string theory, modular forms, finite simple groups).

What makes these cases work out so well?

How could we pack spheres?
Lattice: integer span of n linearly independent vectors. I.e., for
basis v1 , . . . , vn , center spheres at

{a1 v1 + a2 v2 + · · · + an vn | a1 , . . . , an ∈ Z}.

v2

v2-v1 v1

(0,0)

Packing radius = half shortest non-zero vector length.

Periodic packing

Union of finitely many translates of a lattice.

Spheres are not restricted to just the corners of a fundamental

domain.

No reason to believe densest packing must be periodic, but

periodic packings come arbitrarily close to the maximum density.

By contrast, lattices probably do not.

Quadratically many parameters don’t give enough flexibility
to fill all the gaps in an exponential amount of space.
The best sphere packings currently known are not always lattice
packings, but many good packings are.
Simplest lattice: Zn , lousy packing.

Better: for n ≥ 3, “checkerboard” packing

Dn = {(x1 , . . . , xn ) ∈ Zn | x1 + · · · + xn is even}.

D3 , D4 , D5 are best known packings in those dimensions, and

provably best lattice packings.
What about n ≥ 6?

Holes in Dn grow greater and greater.

A hole is a local maximum for distance from nearest lattice point.
What about n ≥ 6?

Holes in Dn grow greater and greater.

A hole is a local maximum for distance from nearest lattice point.
Two classes of holes in Dn (for n ≥ 3):
I (1, 0, . . . , 0) at distance 1 from lattice.
I (1/2, 1/2, . . . , 1/2) at distance
s 
1 2
 2 r
1 n
+ ··· + = .
2 2 4
How big are the spheres in a Dn packing?

Packing radius for Dn : nearest points are

(0, 0, . . . , 0) and(1, 1, 0, . . . , 0)
√ √
at distance 2, so use spheres of radius 2/2.
Wonderful properties of dimension 8

p √
When n = 8, radius n/4 of deep hole equals distance 2
between lattice points.

We can slip another copy of D8 into the holes! This doubles the
packing density.

Result called E8 lattice.

E6 , E7 are certain cross sections of E8 .

These are densest lattice packings for n = 6, 7, 8. Almost certainly

densest packings.

Leech lattice (n = 24) is similar in spirit, but more complicated.

Why is the sphere packing problem difficult?

Many local maxima for density.

Lots of space.

Complicated geometrical configurations.

No good way to rule out implausible configurations rigorously.

High dimensions are weird.

Visualization, take two

There are many ways to try to visualize four dimensions.

But they become incomprehensible as the dimension increases.

Is there any way we can conceive of a million dimensions?

What could it look like?

Let’s start by thinking about cubes, because they are easy to pack.
Even the cubes go crazy
Let’s look at the cube {(x1 , . . . , xn ) | |xi | ≤ 1 for all i}.
2n vertices (±1, . . . , ±1).
√ √
Distance from center to vertex is 12 + · · · + 12 = n.
Imagine n = 106 . How can this be?
Concentration of volume
Almost all volume in high-dimensional solids is concentrated near
the boundary.

Why? Imagine a sphere. Fraction of volume less than 90% of way

to boundary is 0.9n . This tends rapidly to zero as n → ∞.
For example, 0.91000 ≈ 1.75 · 10−46 .

r 0.9r

We can replace 0.9 with 0.99, 0.999, or whatever we want.

[Aside for those who know calculus:
As n → ∞ with c fixed, fraction of volume less than 1 − c/n of
the way to the boundary is (1 − c/n)n → e −c .]

All low-dimensional intuition must be revised.

Boundaries, where all the interaction takes place, become

increasingly important in high dimensions.

Relevant for packing and for analysis of algorithms.

Also the second law of thermodynamics (increase of entropy)!

Hard computational problems
Hard to tell how dense a lattice packing is!

density = (sphere vol.) · (# spheres/unit vol.)

2nd factor = 1/(vol. of fundamental domain)

Absolute value of the determinant of the basis.

Hard part is figuring out which radius spheres will fit. How long is
the shortest nonzero vector in the lattice?

NP-complete (Ajtai, 1997).

In other words, many search problems can be reduced to it. No

proof is known that it cannot be solved efficiently, but that is
almost certainly the case.

There are good algorithms that give pretty short vectors, at least
in low dimensions. These vectors are short enough for some
applications.
Practical applications

Cryptosystems based on the difficulty of finding short lattice

vectors, or ones near a given point.

Many direct applications in computational mathematics, and

numerical exploration.
Goldreich-Goldwasser-Halevi cryptosystem (1997)

Public key: high dimensional lattice

Private key: secret nearly orthogonal basis, that makes it easy to

find the nearest lattice point to any given point

Encode messages as lattice points.

To encrypt, add a small random perturbation. Can decrypt using

the private key, but not without it.

This system has weaknesses (Nguyen 1999), but there are other,
stronger lattice-based systems.
Recognizing algebraic numbers:

The number

α = −7.82646099323767402929927644895

is a 30-digit approximation to a root of a 5th-degree equation.

Which equation?

Of course, infinitely many answers, but we want the simplest.

Trivial example: recognizing that

0.1345345345345345345345345345345

is an approximation to 1/10 + 345/9990.

Let C = 1020 (based on precision of α), and look at lattice
generated by

v0 = (1, 0, 0, 0, 0, 0, C ),
v1 = (0, 1, 0, 0, 0, 0, C α),
v2 = (0, 0, 1, 0, 0, 0, C α2 ),
v3 = (0, 0, 0, 1, 0, 0, C α3 ),
v4 = (0, 0, 0, 0, 1, 0, C α4 ),
v5 = (0, 0, 0, 0, 0, 1, C α5 ).

Lattice vector (a0 , . . . , a5 ∈ Z):

5
!!
X
a0 v0 + · · · + a5 v5 = a0 , a1 , a2 , a3 , a4 , a5 , C ai α i
i=0

ai αi ,
P
For this to be small, we want small ai ’s and really small i
since C is huge.
We can find a short vector using a computer:

(71, −5, 12, −19, 13, 2, 0.000004135 . . .).

That tells us that

71 − 5α + 12α2 − 19α3 + 13α4 + 2α5 ≈ 0.

(More precisely, it is about 0.000004135/C ≈ 4 · 10−26 .)

In fact, this is the equation I used to generate α.
How do we get good sphere packings?

There are several potential UROP problems:

Computer searches in low dimensions.

Nonconstructive proofs in high dimensions.

Computer searches

Yoav Kallus, Veit Elser, and Simon Gravel, Method for dense
packing discovery, Phys. Rev. E 82, 056707 (2010)

Étienne Marcotte and Salvatore Torquato, Efficient linear

programming algorithm to generate the densest lattice sphere
packings, Phys. Rev. E 87, 063303 (2013)

Yoav Kallus, Statistical mechanics of the lattice sphere packing

problem, Phys. Rev. E 87, 063307 (2013)

These papers recover the densest lattices known up to 20

dimensions.
Can we push the calculations further, to unknown territory?
What about periodic packings?
The Siegel mean value theorem

What does it mean to average over all lattices?

It’s not quite obvious, but there is a canonical probability measure

on lattices with fixed determinant (i.e., fundamental cell volume).
Key concept: SLn (R)-invariance.

Recall that SLn (R) is the group of all n × n real matrices with
determinant 1. They act on Rn by left multiplication and preserve
volume.

What does the average pair correlation function look like? It

measures the average number of neighbors at each distance.

Siegel mean value theorem: it is exactly the same as for a Poisson

distribution (uniformly scattered points).
More precisely: given nice f : Rn → R, the average of
X
f (x)
x∈Λ\{0}

over all lattices Λ of determinant 1 equals

Z
f (x) dx.
Rn

Why is this true? There is enough symmetry to rule out any other
possible answer.
R
Specifically, by linearity the answer must be f dµ for some
measure µ on Rn \ {0} that is invariant under SLn (R). This is only
one such measure, up to scaling, and some simple consistency
checks determine the constant of proportionality.
Meta principle: averaging over all possible structures is the same
as having no structure at all.

This is certainly not always true. It generally depends on having a

big enough symmetry group.

Breaks down a little for higher correlation functions, but get

Poisson statistics modulo obvious restrictions (e.g., three points in
a row for three-point correlations).
Why do we care? Sphere packing bounds.
Siegel mean value theorem ⇒ density lower bound of 2 · 2−n .

Let B be a ball of volume 2 centered at the origin.

For a random lattice of determinant 1, the expected number of

nonzero lattice points in B is vol(B) = 2.

These lattice points come in pairs, and some lattices have a lot of
them. Since the average is 2, some lattices must have none.

No lattice points in B means lattice gives a packing of B/2.

Get packing with one copy of B/2 per unit volume, density

vol(B)
= 2 · 2−n .
2n
Vance’s idea

The extra factor of 2 came because lattice vectors occur in pairs of

the same length.

What if we impose additional symmetry?

Intuition: the average number of neighbors remains the same, but

now they occur in bigger clumps, so the chances of no nearby
neighbors go up.

Vance (2011): quaternion algebras

Venkatesh (2012): cyclotomic fields

Is this the best one can do? Only certain symmetry groups work
here: we need a big centralizer to get enough invariance for the
Siegel mean value theorem proof, and only division algebras will do.
Cyclotomic fields are the best division algebras for this purpose.

Other sorts of groups will distort the pair correlation function away
from Poisson statistics. But is that good or bad?

Is there anything similar one can do besides imposing symmetry?

The results so far cannot be the end of the story.

In summary

High dimensions are deeply weird. Even four dimensions can be

puzzling, and a million dimensions are outlandish.

We really don’t know what the best sphere packing in a million

dimensions looks like. Our estimate of its density might be off by
an exponential factor. We don’t even know whether it is ordered
(like a crystal) or disordered (like a lump of dirt).

But these problems really matter. Every time you use a cell phone,
you can thank Shannon and his relationship between information
theory and high-dimensional sphere packing.
For more information

Papers are available from:

http://research.microsoft.com/~cohn

Specifically, see Order and disorder in energy minimization:

http://arXiv.org/abs/1003.3053