Modeling Snow Crystal Growth I:

Rigorous Results for Packard’s Digital Snowflakes

Janko Gravner and David Griffeath

CONTENTS
Digital snowflakes are solidifying cellular automata on the tri-
1. Introduction angular lattice with the property that a site having exactly one
2. Extreme Boundary Dynamics occupied neighbor always becomes occupied at the next time
3. Densities for Exactly Solvable Rules step. We demonstrate that each such rule fills the lattice with
4. Density of Hex 1 and Its Cousins an asymptotic density that is independent of the initial finite set.
5. Density of Hex 1456 and Hex 146 There are some cases in which this density can be computed
6. Proof of Theorem 1.1 exactly, and others in which it can only be approximated. We
7. Macroscopic Dynamics also characterize when the final occupied set comes within a
8. Thickness and Related Issues uniformly bounded distance of every lattice point. Other issues
addressed include macroscopic dynamics and exact solvability.
9. Exact Solvability
Acknowledgments
References

1. INTRODUCTION
Six-sided ice crystals that fall to earth in ideal winter con-
ditions, commonly known as snowflakes, have fascinated
scientists for centuries. They exhibit a seemingly end-
less variety of shape and structure, often dendritic and
strangely botanical, yet highly symmetric and mathemat-
ical in their designs. To this day, snowflake growth from
molecular scales, with its tension between disorder and
pattern formation, remains mysterious in many respects.
Study of snowflakes dates back at least to the sixteenth
century [Magnus 55], and includes early contributions
from such scientific giants as Kepler [Kepler 66], Hooke
[Hooke 03], and Descartes [Descartes 37]. With the ad-
vent of cameras came the first snow crystal album: more
than five thousand photos collected by W. Bentley be-
ginning in 1885 [Bentley and Humphreys 62]. Although
rather few people have ever seen such crystals with their
own eyes, Bentley’s images helped establish snowflake
designs, simplified and idealized, as universal icons for
wintertime. The most significant scientific advances of
the past century were due to Nakaya [Nakaya 54] in the
2000 AMS Subject Classification: Primary 37B15;
Secondary 68Q80, 11B05, 60K05 1930s, who classified natural crystals into dozens of types,
Keywords: Asymptotic density, cellular automaton, exact first grew synthetic crystals in the laboratory, and discov-
solvability, growth model, macroscopic dynamics, thickness ered an elaborate, still perplexing morphology diagram,

c A K Peters, Ltd.
1058-6458/2006 $ 0.50 per page
Experimental Mathematics 15:4, page 421
422 Experimental Mathematics, Vol. 15 (2006), No. 4

which predicts the predominant type of snowflake aris- turns to ice. Since real snowflake growth favors the tips of
ing at any given temperature and supersaturation level. the crystal, Packard proposed that exactly one occupied
(In particular, the familiar essentially two-dimensional (frozen) neighbor should cause solidification, but exactly
crystals arise only for certain parameter values; in other two should not. Thus, in one of his digital snowflakes a
conditions, columnar “needles” form.) All this and much site joins the crystal if and only if it has exactly one occu-
more is explained superbly in a recent popular account by pied neighbor, while in another it joins if the number of
Libbrecht and Rasmussen [Libbrecht and Rasmussen 03], occupied neighbors is odd. In the present paper we will
which also contains a great many state-of-the-art pho- refer to these rules as Hex 1 and Hex 135 , respectively.
tographs of breathtaking beauty. There is a companion Packard’s snowflake automata have been widely publi-
web site [Libbrecht 06]; see also [Libbrecht 05] for a cur- cized since the 1980s to illustrate how very simple math-
rent scholarly review. ematical algorithms can emulate complex natural phe-
Over the past century, geometric structures inspired nomena. A multicolor image of Hex 135 occupies nearly
by snow crystals have begun to adorn the world of all of page 189 in Wolfram’s 1984 article [Wolfram 84a],
mathematics. Most celebrated is the Koch snowflake and the same graphic is reproduced as the first color
[King 64], introduced by H. von Koch [von Koch 04] in plate of Steven Levy’s 1992 book [Levy 92]. More re-
1904. One of the earliest known fractals, a closed curve cently, the first 30 updates of Hex 1 are illustrated on
with Hausdorff dimension 2 log 2/log 3, is obtained in page 371 of [Wolfram 02]. The central tenet of [Wolfram
the limit by starting from an equilateral triangle (with 84a], already familiar from the established universality
• markers at its vertices) and repeatedly applying the of Conway’s game of life [Berlekamp et al. 04], was that
substitution scheme in the diagram below to each piece “Simulation by computer may be the only way to pre-
between markers: dict how certain complicated systems evolve.” Implicit
in this perspective is the inadequacy of mathematics to
analyze complexity. In the discussion of Hex 135 and the
corresponding caption, he writes,
More recent variations on Koch’s construction include
Snowflakes grown in a computer experiment
Gosper’s flowsnake [Gardner 76] and the pentaflake
from a single frozen cell according to this rule
[Dixon 91]. While none of these designs resembles a real
show intricate treelike patterns, which bear
snow crystal to any great extent, their blend of elemen-
a close resemblance to real snowflakes.. . . The
tary polygonal shapes with infinitely fine branching detail
only practical way to generate the pattern is by
evokes the same iconography as Bentley’s album.
computer simulation.
The building blocks for snowflakes are hexagonally ar-
ranged molecules of natural ice (Ih). Just how the elabo- Levy’s account reiterates the claimed verisimilitude:
rate designs emerge as water vapor freezes is still poorly
An elementary schoolchild could look at any
understood. Only very recently have a few rudimen-
of the gorgeous pictures of computer screens in
tary movies of synthetic crystal growth been produced
Packard’s collection and instantly identify it as
[Libbrecht and Rasmussen 03, p. 57]. The solidification
a snowflake.
process involves complex physical chemistry of diffusion-
limited aggregation and attachment kinetics. Theoretical So how do these digital snowflakes evolve? The left
research and mathematical modeling to date have mainly frame of Figure 1 shows a representative snapshot of Hex
focused on the evolution of dendrite tips. See [Gleick 134 after 218 updates starting from a single occupied
87, pp. 309–314] for a popular account of the challenges, cell. (The graphic in [Wolfram 84a] and [Levy 92] is
[Meakin 98] and [Pimpinelli and Villain 99] for recent quite similar.) Letting At denote the crystal at time t,
scholarly monographs, and [Adam 05] for a current re- started from A0 = {0}, it turns out that A2n occupies,
view. with a certain density, the hexagon of lattice cells within
In 1984, Packard [Packard 86] introduced a supremely 2n steps of the origin for each n, but that shapes with
simple cellular automaton (CA) model for crystal solid- apparently fractal boundary arise in the limit along in-
ification. On a honeycomb lattice of hexagonal cells, termediate subsequences of the form tn =
a2n  when a
start with a single “seed” cell of ice surrounded by va- is not a dyadic rational. For instance, the limit shape
por. At each subsequent discrete-time update, any va- for Hex 1 and Hex 135 along the a = 13 subsequence is
por cell neighboring the requisite number of frozen cells exactly the Koch-type snowflake starting from a regular
 Gravner and Griffeath: Modeling Snow Crystal Growth I: Rigorous Results for Packard’s Digital Snowflakes 423

hexagon (with • markers now in the middles of edges) • While computer visualization and empirical calcu-
and based on the substitution scheme shown in the dia- lation are indispensable tools, subtle properties of
gram below, applied to each nonstraight segment between these dynamics cannot possibly be gleaned from sim-
markers: ulations alone. Deductive reasoning plays a funda-
mental role in the analysis.

Turning to formalities, our basic setup features solid-

ification CA on the triangular lattice T (to reflect the
arrangement of water molecules in ice crystals).1 For
notational and computational convenience, we represent
T by Z2 , with the neighborhood N of (0, 0) consisting
Here, the top and the bottom choices apply to a con-
of itself and the six sites (±1, 0), (0, ±1), and ±(1, 1).
cave and a convex vertex, respectively. (A vertex is con-
The neighborhood of an arbitrary x is then x + N . This
vex (concave) if one makes a right (left) turn at it while
representation is handy because integers are much more
moving counterclockwise on the curve.)
familiar than Cayley graph representations of T. On the
In other words, the crystal oscillates between hexag-
other hand, some symmetries are not so easy to spot.
onal shapes and other familiar mathematical forms with
(Figures 1 and 8 are the only ones in which dynamics on
increasingly complex boundary. To illustrate the scien-
T are depicted.) The CA simulator MCell [Wójtowicz
tific insight offered by such simple CA rules, Wolfram
05] is ideally suited for empirical investigation of digital
makes the following intriguing prediction based on his
snowflakes, since its Weighted Life rule set supports this
time trace of Hex 1 up to time 30 [Wolfram 02]:
Z2 embedding. Indeed, MCell was a crucial resource dur-
For example, one expects that during the ing the early stages of our work, and any conscientious
growth of a particular snowflake there should reader of this paper will surely need to enlist its aid, or
be alternation between tree-like and faceted that of some similar program.
shapes, as new branches grow but then col- We denote by At ⊂ Z2 the set of occupied sites at
lide with each other. And if one looks at real time t. Often, sites in At are called 1’s, and sites in
snowflakes, there is every indication that this Act are called 0’s. The set At grows in discrete time
is exactly what happens. And in fact, in gen- t = 0, 1, 2, . . . . That is, At ⊂ At+1 ; such CA are called
eral the simple cellular automaton shown above solidifying. Whether x ∈ / At belongs to At+1 depends
seems remarkably successful at reproducing all only on the number of sites it sees in At , that is, on
sorts of obvious features of snowflake growth. |(x + N ) ∩ At |. Thus, the rule is given by a func-
tion π : {1, 2, 3, 4, 5, 6} → {0, 1} such that for x ∈/ At ,
We intend to address the level of realism of digital
π(|(x + N ) ∩ At |) = 1 iff x ∈ At+1 . As above, we spec-
snowflakes and other lattice models for ice crystal growth
ify π by listing all n for which π(n) = 1. Our canonical
in a sequel to this paper [Gravner and Griffeath 06].
choice of the initial set is A0 = {0}, although we will also
Here our goal is a rigorous study of the complete fam-
study the dynamics started from an arbitrary finite set,
ily of Packard rules. For instance, we will show that each
and use assorted infinite initial sets as props. Note that
crystal fills the lattice with a characteristic asymptotic
for any solidification dynamics and every A0 , the final set
density, independent of the finite initial seed. We also
A∞ exists as a sitewise limit of At , and A∞ = ∪t≥0 At .
identify a subclass of exactly solvable rules for which the
Our basic assumption on π, reflecting the fact that
density is a computable rational number (e.g., Hex 135
tips of a growing snow crystal are favored, is that π(1) =
and Hex 134 have densities 56 and 21 22 , respectively), and 1. We call solidifying CA with this property digital
a complementary class with inherently more computa-
snowflakes. It is not clear what else should be assumed
tional complexity of the limit state A∞ . Ultimately, our
about π, so we propose to catalog all possible behaviors
story here has two main morals:
of such automata.
• As CA dynamics go, Packard’s digital snowflakes are There are 32 candidates for digital snowflakes. Of
not very complex. The patterns they trace are not these, 16 are trivial: when π(2) = 1, the appropriate ver-
periodic, but nearly so, and in some cases exactly 1 We have chosen to conform to accepted mathematical terminol-
described by a finite recursion. Consequently, they ogy, although T is often called the hexagonal lattice in the popular
are quite amenable to mathematical methods. and scientific literature.
424 Experimental Mathematics, Vol. 15 (2006), No. 4

FIGURE 1. The occupied set of Hex 134 at time 218, started from {0}. Let τx = inf{t : x ∈ At } be the time x is occupied
and d the distance on the triangular lattice. The two frames depict periodically shaded contours of constant τx (left) and
constant σx = τx − d(x, 0) (right, where only the 0 contour {x : σx = 0} is black).

sion of the extreme boundary dynamics (see Section 2) is To summarize, Packard’s snowflakes enjoy three im-
very easy to analyze, and causes the dynamics to grow a portant properties (to be precisely described in Sec-
full hexagon from a singleton, while from any other finite tion 2):
seed they grow a full hexagon apart from a region at fi-
nite distance from the rays defined by the extreme points • Starting from a single occupied cell, the light cone
of the hexagon. (Many real snowflakes also grow as ex- CA forms an impenetrable web of occupied sites that
panding hexagons, but there is no mystery here.) From divides further solidification into independent finite
now on, we refer to only the remaining 16 rules, i.e., to domains with simple boundary conditions.
those with π(1) = 1 and π(2) = 0, as digital snowflakes. • Boundary effects within each domain are controlled.
Next, let us explain our basic approach and summarize
our asymptotic density results. • The light cone CA is additive, so the web from a
The right frame of Figure 1 highlights key structural general finite seed is representable as a superposition
features that underlie our analysis. Note first that any of webs from each of its individual cells.
digital snowflake advances at speed 1 along the axes of The first two properties ensure a recursive represen-
the lattice, since the closest unoccupied site in these di- tation of the dynamics, while the last is crucial for the
rections always has only one occupied neighbor. Within asymptotic density’s independence of A0 .
each of six wedges formed by the axes, the black cells are The delicacy of our results is conveyed effectively by
those that solidify at the edge of the light cone, i.e., at comparison to analogous solidification on Z2 with range-
the maximum speed of propagation allowed by a nearest- 1 Box neighborhood consisting of a central cell and its
neighbor rule. This process induces symmetric copies of eight nearest neighbors: (0, 0), (0, ±1), (±1, 0), (±1, ±1).
the space-time pattern of a one-dimensional CA. Because There are 128 such rules with π(1) = 1; see [Griffeath 06]
π(1) = 1 and π(2) = 0, this is the additive xor rule, ar- for a brief introduction and colorful graphics of Box 1,
guably the most familiar of all cellular automata. Conse- Box 157, Box 1357 , and Box 136 crystals. Although
quently, the black cells form discrete versions of a famous snowflake-like recursive carpets emerge in a great many
fractal known as the Sierpiński triangle. Of course, dig- cases, any and all of the three properties above may fail.
ital snowflakes continue to solidify after the edge of the For instance, we will see in Section 6 that the density of
light cone passes, as seen in the gray portions of the right Box 1 , provided it exists at all, can depend on the initial
frame of Figure 1. But the Sierpiński lattice effectively seed. Also, for the “odd” rule, Box 1357 , the light-cone
divides the crystal into independent finite regions with web “leaks,” and growth is apparently chaotic. Although
all 1 boundary conditions, within which subsequent dy- there are many fascinating problems connected with the
namics evolve. Box neighborhood, and exact computations are feasible
 Gravner and Griffeath: Modeling Snow Crystal Growth I: Rigorous Results for Packard’s Digital Snowflakes 425

in some cases, there is no hope of a complete analysis as sites on a side, one would still be led to the conclusion
in the present Hex setting. that A∞ = Z2 . In fact, the four dynamics are all distinct
Fix a set S ⊂ Z2 . Let µ
be 2 times the counting eventually and presumably have different densities less
measure on  · S. We say that S has asymptotic density than one.
ρ if µ
converges to ρ · λ as  → 0. Here λ is Lebesgue In contrast to density results, which are macroscopic
measure on R2 , and the convergence holds in the usual in nature, our next result addresses the most basic mi-
sense:

croscopic properties of final configurations. Call a set
f dµ
→ ρ · f dλ (1–1) S ⊂ Z2 thick if

for any f ∈ Cc (R2 ). sup{d(x, S) : x ∈ Z2 } < ∞.

Theorem 1.1. To each of the 16 digital snowflakes there Here, d is distance in any chosen norm, say
·
∞ . For
corresponds a ρ ∈ (0, 1], the asymptotic density of A∞ , snowflakes with density between 0 and 1, thickness of A∞
that is independent of the finite seed A0 . and Ac∞ is one rough notion of an almost-periodic final
state. In the following theorem, A0 is assumed to be an
We will index the densities by our notation for the re- arbitrary finite set.
spective rules, and give more information on their values
in the next theorem. Theorem 1.3. The eight exactly solvable rules have the
following properties:
Theorem 1.2. The densities are exactly computable in (1) The final set A∞ is always thick.
eight cases:
(2) Hex 13456 always has A∞ = Z2 . For the other rules
ρ13 = ρ135 = 5/6 ≈ 0.8333, with density 1, there exist initial conditions for which
ρ134 = ρ1345 = 21/22 ≈ 0.9545, A∞ contains infinitely many 0’s.
ρ136 = ρ1356 = ρ1346 = ρ13456 = 1.
(3) Ac∞ is always thick for rules with density less than 1,
In six other cases, one can estimate, within ±0.0008, and never thick for those with density 1.

ρ1 ≈ 0.6353, For the eight rules that are not exactly solvable, A∞ is
ρ14 , ρ145 ≈ 0.9689, never thick, and Hex 1 always has thick Ac∞ .
ρ15 ≈ 0.8026,
It is an intriguing open question whether Ac∞ is thick
ρ16 ≈ 0.7396, for the seven rules not covered by Theorem 1.3. We sus-
ρ156 ≈ 0.9378. pect that it is for all of them, but have no argument.
The rest of the paper is organized as follows. Prelim-
Finally,
inaries in Section 2 describe precisely various structural
ρ146 ∈ (0.995, 1), ρ1456 ∈ (0.9999994, 1). properties of the additive web that decomposes a digital
snowflake into independent regions of finite size. Slight
variations in how these regions solidify are identified case
Perhaps surprisingly, ρ14 > ρ134 , testimony to the fun- by case. Sections 3 and 4 then detail the first eight den-
damentally nonmonotone nature of these rules. sity calculations of Theorem 1.2 by deriving and solving
We refer to the first eight rules in Theorem 1.2 as recursions for |A2n −1 |. Section 3 handles exactly solv-
exactly solvable. In Section 9, we will develop a rigorous able cases, first the simplest rules: 13 , 135 , 136 , and
foundation for this terminology. 1356 , and then those obeying slightly more complicated
It is tempting to conjecture that ρ14 = ρ145 and ρ146 = dynamics: 134 , 1345 , 1346 , and 13456 . For all these
ρ1456 since the two dynamics of each pair are identical snowflakes, the limit of |A2n −1 |/(3 · 4n ) is evaluated ex-
starting from A0 = {0} on finite arrays up to 500 × 500 plicitly. Next, Sections 4 and 5 develop and analyze cor-
in size. This question remains open, but one should resist responding recursions for Hex 1 and the other seven rules
such empirical conclusions. For instance, as we shall see that are not exactly solvable. Now, due to certain messy
later, observing Hex 1456 from A0 = {0} on even the interactions, existence of the normalized limit of occupied
world’s most extensive graphics array, with millions of cells is established by a novel application of the renewal
426 Experimental Mathematics, Vol. 15 (2006), No. 4

theorem, but this density is implicit and can only be ap- One can also prove by induction that
proximated numerically. Section 4 handles six densities
that we are able to estimate within 0.0008. The cases T2n −1 ∩ {(x, y) : x = 2n − 1}
146 and 1456 in Section 5 require a different rescaling = {(2n − 1, y) : 0 ≤ y ≤ 2n − 1} (2–2)
argument for the upper bound, since their densities are
extremely close to, but less than, 1. and
In Section 6 we complete the proof of Theorem 1.1 by
T2n ∩ {(x, y) : x = 2n } = {(2n , 0), (2n , 2n )}, (2–3)
showing that A∞ has an asymptotic density ρ starting
from {0} in the formal sense of (1–1), that ρ agrees with since the two “buds” at time 2n create two versions of
the corresponding value obtained in Sections 3–6, and the dynamics that do not interact through time 2n+1 − 1.
that the same asymptotic density occurs when the initial The name of the dynamics stems from the fact that it
seed is an arbitrary finite set A0 . (A technicality for rules preserves exclusive union: TnA xor B = TnA xor TnB . This is
with π(3) = 0 is also handled at the end of this section.) immediate for n = 1, and then again follows by induction.
Then, Section 7 introduces two distinct rule-dependent It is helpful to consult Figures 2–4 while reading the
macroscopic dynamics Sa for the limit of 2−n Atn , where remainder of this section. The darker sites in those fig-
tn =
a · 2n . Theorem 7.1 thereby extends to ures form initial conditions (the reasons for which will be
general a the already-mentioned substitution-scheme explained later). For growth from a single seed at the
limit for Hex 1 and Hex 135 in the case a = 13 . A more origin, the lowest row (two rows) in the top three frames
sophisticated approach is required to handle both holes of Figures 2 and 3 (Figure 4) should be deleted, and then
formed by colliding branches of the snowflake and the the origin placed at the leftmost lowest site. Also, time
case of irrational a. Examples are given to illustrate should be diminished by 1 in Figures 2 and 3.
the exotic dependence on a of the Hausdorff dimen- The relevance of additive dynamics to digital
sion of the boundary of Sa . Finally, Sections 8 and 9 snowflakes becomes apparent when we note that if A0
address thickness (Theorem 1.3) and exact solvability, does not include any site to the right of the y-axis,
respectively. Theorems 1.1–1.3 yield a natural and pre-
cise division of digital snowflakes into two complexity An ∩ {(x, y) : x = n} = TnA0 ∩y-axis ∩ {(x, y) : x = n}
classes (Section 9). This distinction, based on the no- (2–4)
tion of automaticity of the final set, is potentially widely in any of our 16 dynamics. For a general solidifica-
applicable in the computational theory of cellular au- tion CA, the light cone of a set L0 is the set Ln =
tomata. L0 + N + N + · · · + N , where N is repeated n times,
2. EXTREME BOUNDARY DYNAMICS i.e., the set of points that can possibly be influenced by
L0 at time n. In particular, if L0 = A0 , which we as-
Our basic tool is the additive dynamics Tn , also referred sume from now on, An ⊂ Ln . The extreme boundary
to as xor , addition mod 2, or rule 90 (see, e.g., [Willson comprises sites y ∈ Ln with (x + N ) ∩ Lcn = ∅. Then
84]). Among several equivalent definitions, we choose the (2–4) means that our rules perform six copies of the addi-
following. We declare the neighborhood of 0 to consist of tive dynamics (appropriately mapped) at their extreme
(−1, 0) and (−1, −1), and the additive rule to be exactly boundaries, and create an “additive web” consisting of
one solidification dynamics with this two-point neighbor- black sites x with σx = 0 in the right frame of Figure
hood, i.e., according to this rule a site changes its state 1. We will call this the extreme boundary dynamics, and
to 1 iff exactly one of x + (−1, 0) and x + (−1, −1) is in the sites so created primary boundaries. Also note that
state 1. The initial set T0 will always be a subset of the properties (2–1) and(2–2) dictate impenetrable bound-
y-axis, our default choice being the singleton {0}. If we aries for our rules. For example, if A0 = {0}, then after
want to emphasize that T0 = A, we use the notation TnA . time 2n − 1 the dynamics create a triangle of 1’s given by
Many properties of this rule are well known and easy to (2–1) and (2–2), and after that time the dynamics inside
check. Nevertheless, we will explain them briefly as they and outside the triangle are independent. Even (2–1) it-
are used. self, since it extends all the way to the extreme boundary,
Observe first that with the canonical choice T0 = {0}, separates the dynamics into six independent wedges. For
many purposes, then, it will suffice to look at one of these
{(x, 0) : 0 ≤ x ≤ n} ∪ {(x, x) : 0 ≤ x ≤ n} (2–1)
wedges, typically the one in the first quadrant and below
⊂ Tn ⊂ {(x, y) : 0 ≤ y ≤ x ≤ n}. the line y = x.
 Gravner and Griffeath: Modeling Snow Crystal Growth I: Rigorous Results for Packard’s Digital Snowflakes 427

Digital snowflakes do more than add sites at their ex- will begin our study of the Hex 13 rule in the next sec-
treme boundaries. At time 2n , for example, the two buds tion.
in (2–3) do not merely spread into two new triangles out- We emphasize that some of the sites that we use as ini-
lined by the additive dynamics; they also grow into the tial conditions are not present initially; the hole “frames”
empty triangle between them. To be more precise, in are typical examples. Indeed, some of these initial sites
the eight rules with π(3) = 1, there is a pair of occupied may never be created, as in the Hex 1 case. However,
sites (2n , 1), (2n , 2n − 1) ∈ A2n +1 ; in the remaining eight they are very convenient for definitions and for symme-
cases, (2n + 1, 2), (2n + 1, 2n − 1) ∈ A2n +2 . In fact, it try considerations. In each case, it is straightforward to
is more convenient to interpret the initial buds as being verify that the dynamics behaves as if these sites were
at sites below and above these two; they are marked • present initially.
in the top left configurations of Figures 2 and 4. As we On the other hand, if π(3) = π(4) = 1, then the four-
explain below, these two buds generate their own sec- site (now marked • in Figure 3) becomes occupied at time
ondary extreme boundary dynamics (spreading into the 2n +2n−1 , and the resulting “live” vertex gives rise to an-
two smaller triangles outlined in the figures) until they other secondary dynamics in the hole. Three secondary
collide. boundary dynamics originating at the four-site and the
Assume first that π(3) = 1. Then the two buds gen- two one-buds all collide at time 2n + 2n−1 + 2n−2 − 1 to
erate exactly (appropriately rotated and deformed) addi- create a triangular hole (as in the middle top of Figure
tive dynamics at their extreme boundaries. This is be- 3). The remaining two secondary boundary dynamics
cause one of the two endpoints (sites generating the left- collide at the same time and create a smaller quadrilat-
most set in (2–1)) sees three occupied sites, two of which eral hole. This one, however, is of a type different from
are contributed by the boundary conditions (occupied that of the original, since only two of its vertices are live.
sites in primary boundaries), while the other endpoint is This mechanism is iterated, as illustrated by a larger hole
shared by the neighboring additive dynamics. Therefore, example in Figure 3. Our analysis of Hex 134 dynam-
the said two extra buds generate two copies of additive ics will therefore require three types of hole dynamics,
extreme boundary dynamics (in the smaller triangles), generated by different initial conditions.
which at the time 2n + 2n−1 − 1 generate an occupied The situation is again different when π(3) = 0. Now
secondary row and diagonal of length 2n−1 , separated by the two secondary buds are a little off center (by one
a 0 at (2n , 2n−1 ). This 0, let us call it the four-site, sees site, to be precise). They still generate additive bound-
four occupied sites. ary dynamics, but the final interaction inside the re-
If π(4) = 0, the four-site (marked by ◦ in the top mid- sulting hole generates two holes of different sizes, and
dle of Figure 2) will not get occupied immediately, and each successive generation of holes has one of a smaller
the vertex of the wedge between the row and the diagonal size. We give a more-precise description for the Hex 1
of 1’s is “dead.” The two secondary boundaries, together rule; others are similar. In this case, all points (2n , y),
with the primary ones {(x, 2n ), (x, x − 2n ) : 2n + 2n−1 ≤ 1 ≤ y ≤ 2n − 1, see at least two occupied sites to their
x ≤ 2n+1 −1}, form a “hole,” which is invaded by dynam- left at time 2n − 1 (hence thereafter) and thus will never
ics that emanate from two one-buds at (2n + 2n−1 , 2n−1 ) get occupied. The two secondary buds collide at time
and (2n + 2n−1 , 2n ), created at time 2n + 2n−1 (each 2n + 2n−1 , but the secondary row and diagonal (created
marked by • in the top middle of Figure 2). This hole is at time 2n + 2n−1 ) are now not separated. Neverthe-
in turn divided into two smaller holes (by secondary 1’s) less, the 0 at (2n + 2, 2n−1 + 1) (marked ◦ in Figure 4)
at time 2n + 2n−1 + 2n−2 − 1, etc. This hole-filling mech- sees four occupied points, so it never gets occupied, and
anism is illustrated in the bottom of Figure 2, where the the resulting vertex of the wedge is dead.2 The two sec-
two descendant holes are outlined in the first two frames. ondary boundaries, together with primary ones, create a
It is important to note that the parallelogram hole in hole. This hole starts being filled by buds that appear
the top row of this figure is equivalent, modulo boundary at time 2n + 2n−1 + 2 at (2n + 2n−1 + 1, 2n−1 + 1) and
corrections, to the small square hole with darker shaded (2n +2n−1 +1, 2n ). The fact that these buds are off-center
1 boundary conditions. (Match the marked first two oc- has two consequences. The first is minor and technical: it
cupied sites with the same marks in the parallelogram is necessary to start the analysis with a basic wedge and
hole.) This consequence of symmetries of T will be ex- holes that incorporate the buds in their initial conditions.
ploited throughout. Armed with these observations, we 2 In rules with π(4) = 1, such as Hex 14 , this 0 becomes a 1, but

then growth from the vertex stops.

428 Experimental Mathematics, Vol. 15 (2006), No. 4

FIGURE 2. Hex 13 : Basic wedge dynamics (top) at times 18, 24, and 32; final configuration of a hole of size 10; size-34-hole
dynamics at times 16, 24, and 32.

The second fact is crucial for the analysis of these rules. Assume first that the initial occupied set is the origin,
Namely, the two holes that result when the additive dy- A0 = {0}, and define
namics from two secondary buds collide are of unequal
|A2n −1 |
size, their sizes differing by exactly 2. This in turn cre- ρ13 = lim .
ates holes of a larger and larger variety of sizes (as shown
n→∞ 3 · 4n
in the two bottom frames of Figure 4, where the second- We will see shortly that the limit exists, and that
generation descendant holes are also outlined, since they
bn
evolve slightly out of phase and are thus difficult to iden- ρ13 = lim ,
n→∞ 4n
tify). That the interaction in smaller holes still creates
impenetrable boundaries and dead wedge vertices is guar- where bn are defined by certain wedge dynamics. Namely,
anteed by the following lemma. run the dynamics An inside the first quadrant with one
boundary condition on the axes. Let
Lemma 2.1. Assume that the initial configuration consists
of two points, A0 = {(0, 0), (2n − k, 2n − k), 0 < k ≤ Bn = {(x, y) : 1 ≤ x ≤ 2n , 1 ≤ y ≤ 2n }.
2n−1 }. For any digital snowflake, Then bn is the size of the final occupied set inside Bn ,
{(y, 2n − 1) : 0 ≤ y ≤ 2n−1 − k + 1} ⊂ A2n −1 . bn = |A∞ ∩ Bn | = |A2n ∩ Bn |.

The second equality, i.e., the fact that these dynamics do

Proof: This is a simple consequence of the speed of light. not occupy any site on Bn past time 2n , can be proved
Namely, consider the light cone of the point (2n − k, 2n − by induction along the lines of arguments given below.
k) at time t. Outside this light cone, the dynamics started Smaller wedge numbers are obtained with the same
from {(0, 0)} and the one started from A0 agree through initial condition as for bn but counted only inside
time t. Apply this observation at time t = 2n − 1.
Wn = {(x, y) : 1 ≤ x ≤ 2n , 1 ≤ y ≤ x}

and defined by
3. DENSITIES FOR EXACTLY SOLVABLE RULES wn = |A∞ ∩ Wn |.
We begin with the Hex 13 rule, since it represents the The dynamics inside this wedge (with given initial con-
simpler of two exactly solvable cases with nontrivial ditions) will be called the basic wedge dynamics. See the
density. top row of Figure 2. Note that bn = 2wn − 2n .
 Gravner and Griffeath: Modeling Snow Crystal Growth I: Rigorous Results for Packard’s Digital Snowflakes 429

By direct enumeration, b0 = 1, b1 = 4, b2 = 14, b3 = Thus, starting from A0 = {0}, by (3–1),

54, etc. Moreover, since A2n −1 started from A0 = {0} 5 n
contains, modulo boundary corrections, six copies of the |A2n −1 | = · 4 − 3 · 2n + 3,
2
dynamics on Wn , it follows that
and ρ13 = 56 .
|A2n −1 | = 3bn − 3 · 2n + 1. (3–1) Hex 135 , Hex 136 , and Hex 1356 all satisfy precisely
the same recursions (3–2) and (3–3). Also, it is easy to
Next, consider the dynamics on the {0, . . . , L} × check that for small holes and wedges, Hex 13 and Hex
{0, . . . , L} box with initial occupied sites consisting of 135 evolve identically. Thus, it follows by induction that
the frame {(x, y) : x ∈ {0, L} or y ∈ {0, L}}. We call Hex 13 and Hex 135 crystals agree exactly for all t when
L the size of such a hole. Holes of sizes 10 and 34 are A0 = {0}. In particular, ρ135 = ρ13 = 56 . Since Hex
featured in Figure 2. Let H(L) be the final occupation 136 and Hex 1356 both fill small holes and wedges, it
count minus the frame, also easily follows that these two rules solidify completely
from a singleton.
H(L) = |A∞ ∩ {(x, y) : 0 < x < L, 0 < y < L}|. We turn next to Hex 134 , another exactly solvable rule
with density less than 1. The domain counts bn and wn
We define are defined in exactly the same manner as before, giving
hn = H(2n + 2). b0 = 1, b1 = 4, b2 = 16, b3 = 62, b4 = 246, etc.
As mentioned in Section 2, the hole interactions (see
Thus h0 = 2, h1 = 6, h2 = 20, etc. Actually, it is slightly
Figure 3) are more complicated now. There are three dif-
more convenient to use hn = hn −2n+1 . By observing the
ferent types of hole dynamics, all run in the {0, . . . , L} ×
wedge dynamics from time 2n to time 2n+1 , we obtain
{0, . . . , L} box. We define the frame F = {(x, y) :
our first basic recursion:
x ∈ {0, L} or y ∈ {0, L}}. All occupation numbers are
wn+1 = 3wn + 2(wn−1 − 2n−1 ) + hn−1 . counted strictly inside this frame, and in addition ex-
clude any other initially occupied sites. We will always
This recursion is illustrated by the top row of Figure 2, use L = 2n + 2 and count occupation numbers at time 2n
where the five smaller wedges are outlined. Here one (which in every case differ by 1 from the final occupation
thinks of the sites in Z2 as centers of the squares, and numbers).
the outlined regions include their boundaries. The sites The principal hole dynamics use initial occupied sites
outside the outlined regions correspond to the interior of consisting of the “frame minus 3 corner sites,” F \{(0, L−
the pictured size-10 hole, from which the bottom row and 1), (0, L), (1, L)}, and defines the occupation numbers hn .
rightmost column are removed. Therefore Then h0 = 3, h1 = 7, h2 = 22, etc. Again, we set
hn = hn − 2n+1 .
bn+1 = 3bn + 2bn−1 + 2hn−1 . (3–2) The secondary hole dynamics start with two opposite
corners missing: F \ {(0, L − 1), (0, L), (1, L), (L − 1, 0),
A second basic recursion is obtained by observing the (L, 0), (L, 1)}. Call the resulting occupation numbers sn .
hole dynamics until time 2n−1 : Then s0 = 2, s1 = 6, s2 = 20, etc.
hn = 2bn−1 + 2hn−1 . The third and last hole dynamics start with a trian-
gular hole, i.e., the initially occupied set
This is illustrated in the bottom row of Figure 2, where
the interiors of the first descendant holes are outlined. (F \ {(0, L − 1), (0, L), (1, L)}) ∪ {(x, y) : y < x},
Equivalently,
and define tn . Then t0 = 2, t1 = 5, t2 = 13, etc.
Our system of recursions is given by (3–2), again il-
hn = 2bn−1 + 2hn−1 . (3–3) lustrated in the top frames of Figure 3, and
Express hn in terms of b’s using (3–2), and then plug hn = 2bn−1 + wn−1 + tn−1 + sn−1 ,
into (3–3) to obtain the second-order equation bn+2 =
sn = 2bn−1 + 2sn−1 ,
5bn+1 − 4bn , valid for n ≥ 1. Hence bn = α · 4n + β.
Computing the constants α, β from b1 and b2 , we get tn = 2wn−1 + hn−1 .

5 n 2 To understand these new equations, consult the bottom

bn = ·4 + . row of Figure 3. The first successor holes (outlined) are a
6 3
430 Experimental Mathematics, Vol. 15 (2006), No. 4

FIGURE 3. Hex 134 : Basic wedge dynamics (top) at times 24, 28, and 32; final configuration of a hole of size 10;
size-34-hole dynamics at times 16, 24, and 28.

triangular one and a square one of secondary type. The 4. DENSITY OF HEX 1 AND ITS COUSINS
former creates a hole of the principal type, while the lat-
In this section we will analyze, in this order, Hex 1, 14,
ter creates two holes of the secondary type, as seen in the
145, 15, 16 , and 156 . We still use the definition,
bottom middle frame. Therefore, we have
|A2n −1 |
ρ = lim , (4–1)
5 7 n→∞ 3 · 4n
hn = bn−1 + tn−1 + sn−1 − 2n ,
2 4 where A0 = {0}. However, for these rules we cannot give
sn = 2bn−1 + 2sn−1 , (3–4) the exact value of the density; instead, we will demon-
3 strate that the limit exists. It is, for now, convenient to
tn = bn−1 + hn−1 + 2n .
2 redefine
bn
ρ = lim n ;
n→∞ 4
At this point the problem could be solved by matrix ma-
we prove that (4–1) also holds at the end of Section 6.
nipulation, but it is easier to eliminate sn and tn using
The numbers bn are defined by appropriate wedge dy-
the first and third equations of (3–4), then eliminate hn
namics, defined slightly differently than before. Namely,
using (3–2). This yields the equation bn+3 − 5bn−2 +
these dynamics run inside the first quadrant with one
3bn−1 + 4bn = 0. It follows that bn is a linear combina-
√ boundary condition on the axes, and with the initial oc-
tion of 4n , φn , and (−φ−1 )n , where φ = (1 + 5)/2 is
cupied set A0 consisting of the single point (apart from
the golden ratio. Computing constants, the final result
the axes) (1, 2). Now set
is, for n ≥ 1,
Bn = {(x, y) : 1 ≤ x ≤ 2n , 1 ≤ y ≤ 2n + 2}
√  √ n
21 n 15 − 5 1+ 5 and let bn be the size of the final occupied set inside Bn ,
bn = ·4 + ·
22 55 2 i.e.,
√  √ n bn = |A∞ ∩ Bn |.
15 + 5 1− 5
+ · .
55 2 We also make use of the smaller wedge

Wn = {(x, y) : 1 ≤ x ≤ 2n , 2 ≤ y ≤ x + 1},
21
Again (3–1) holds, so ρ134 = 22 .
and define
Once more, it is easy to prove that from A0 = {0},
wn = |A∞ ∩ Wn |.
Hex 1345 generates exactly the same A∞ as Hex 134 ,
while Hex 1346 and Hex 13456 solidify completely. Note that bn = 2wn − 1 − (2n − 1) = 2wn − 2n .
 Gravner and Griffeath: Modeling Snow Crystal Growth I: Rigorous Results for Packard’s Digital Snowflakes 431

FIGURE 4. Hex 1 : Basic wedge dynamics (top) at times 18, 24, and 32; final configuration of hole of size 10; size-34-hole
dynamics at times 17, 26, and 32.

Appropriate hole dynamics on the {0, . . . , L} × Our second recursion is generated by hole dynamics
{0, . . . , L} box have initial occupied sites consisting of run until growth from the initial two buds collides:
the frame {(x, y): x ∈ {0, L} or y ∈ {0, L}} together
with (1, 2) and (L − 2, L − 1). As before, hkn = 2bn−1 + hkn−1 + hk+1
n−1 − ek , 2n−1 ≥ 4k + 8. (4–3)

H(L) = |A∞ ∩ {(x, y) : 0 < x < L, 0 < y < L}|, The restriction on n, obtained from Lemma 4.1 below,
is not optimal, but is one that works in all cases. (For
but now we need a much larger variety of hole counts: small k one can get away with a less-restrictive condition,
which varies from case to case and is useful for compu-
hkn = H(2n − 2k), k = −1, 0, 1, . . . .
tations.) The error terms ek , which keep this recursion
Here we view k as fixed and n large enough that this from closing, are the result of slightly “dirty” interac-
makes sense. We also abbreviate hn = h−1 n .
tion between the two growing buds (cf. the bottom of
Two basic recursions will be derived in a manner sim- Figure 4).
ilar to the Hex 13 analysis. The only difference between We bound ek using the following fact about additive
the six rules is the correction term in (4–2) below. Equa- dynamics. Starting from T0 = {0}, perform 2n − 1 steps
tion (4–3) is the same in all cases. to generate T2n −1 . Fix an , 0 < ≤ 2n , and let Z =
We now proceed with Hex 1 . Z(n, ) be the union of all connected components (in the
The first recursion is obtained by observing Wn from triangular lattice sense) of 0’s in T2n −1 ∩ {(x, y) : y ≥
time 2n to 2n+1 : 2n − }.

wn+1 = 3wn + 2(wn−1 − 2n−1 ) + hn−1 − 2(2n−1 − 1) Lemma 4.1. Every (x, y) ∈ Z has y ≥ 2n − 2 .
= 3wn + 2wn−1 + hn−1 − 2n+1 + 2.
Proof: Focus on the line {y = 2n − } and consider an
Note (as in the top left frame of Figure 4) that two rows interval of a 0’s flanked by 1’s at both ends. Find the first
and columns are now removed from the size-10 hole to 1 below, say at distance b, the leftmost 0 of this interval.
match the sites not covered by the five wedges. Therefore Paint this occupied site red . Then the column of b + 1
bn+1 = 3bn + 2bn−1 + 2hn−1 − 2n+1 + 4. (4–2) sites (b 0’s and the red site) must have to its immediate
left a column of b+1 1’s. The dynamics now ensures that
For example, b0 = 1, b1 = 4, b2 = 14, b3 = 50, b4 = a = b and that the red site is connected by an occupied
182, . . . ; and h0 = 0, h1 = 2, h2 = 8, h3 = 36, h4 = diagonal to the 1 at the right border of the initial interval
154, . . . . of 0’s.
432 Experimental Mathematics, Vol. 15 (2006), No. 4

This proves that the worst case is that in which T2n −1 ∩ Write
{y = 2n − } consists of − 1 0’s flanked by two 1’s, in n−1 k  
 k
which case b = − 1. η(n) = (2i + 1)2 = n2 2n − n2n + 2n − 1.
i
One can in principle compute ek for any k from quanti- k=0 i=0

ties ew h w
k and ek , which we now define. First, ek is the final Using (4–3) repeatedly, and the above bounds, one ob-
occupation count in the region at the tip of the growth tains
n
that gives bn , consisting of a (2k + 3) × (2k + 3) box to-
gether with two lattice triangles, a (2k + 3) × (2k + 3) and hn = 2k bn−k − en ,
k=1
a (2k + 1) × (2k + 1) one. Here is the region for k = 1,
labeled with x’s: where
−3η(n) ≤ en ≤ 4η(n). (4–4)
1 1 1 1 1 x x x x x
Therefore
x x x x x x
x x x x x x x n−1

x x x x x x x x bn+1 = 3bn + 2bn−1 + 2 2k bn−1−k − en−1 − 2n+1 + 4
x x x x x x x x x k=1
n

x x x x x 1 x x x 1
1 x x 1 = 3bn + 2k bn−k − en , (4–5)
k=1
1 x 1
Next, ehk is the final count of occupied sites in the where en = en−1 + 2n+1 − 4 for n ≥ 1 and e0 = −1.
interaction area in the middle of H(2n − 2k). This area The sequence bn /4n satisfies a renewal equation, and
consists of a (2k+3)×(2k+3) box and four triangles—the the renewal theorem [Feller 68, p. 330] gives
two above and their reflections across the main diagonal ∞ ∞
of the box. (The set of occupied sites is symmetric with bn 1 − 14 n=0 4−n en 5 1  −n 
ρ1 = lim n = 3 1 ∞ −k
= − 4 en .
n→∞ 4 6 6 n=1
respect to this reflection at all times.) 4 + 2 k=2 k2
In all small cases we use, we have checked that the last (4–6)
occupied site in the interaction area (or in fact anywhere Summability of 4−n en guarantees existence of the limit,
outside the two smaller holes that are filled recursively) but ρ1 could be 0 or 1 in principle.
gets added at time 2n−1 . If this is the case, a speed- To obtain nontrivial bounds on ρ1 , we enumerate h9n
of-light argument shows that ew n can be computed in the for n ≤ 11 directly, employ (4–2) and (4–3) to compute
(2k+3)×(2k+3) box with an additional row of 2k+3 sites bn exactly for n ≤ 23, and then use (4–5) to compute en
at the bottom, while the interaction area for computation for n ≤ 22. For n ≥ 23, we estimate en as follows.
of ewk adds to this an additional column of 2k + 3 sites on
The lack of symmetry between the upper and lower
the right. Then (4–3) holds under the restriction 2n−1 ≥ bounds in (4–4) can be removed by counting 0’s instead
2k + 6. As mentioned earlier, we use this property of of 1’s. That is, write bcn = 2n (2n + 2) − bn and hcn =
small cases for computations. (2n + 1)2 − hn . Then the analogues of (4–2) and (4–3)
In any case, by Lemma 4.1 above (and its proof), lead to
n

ek = 2ew h
k − ek .
bcn+1 = 3bcn + 2k bcn−k − en−1 − 4 · 2n − 6. (4–7)
Direct enumeration gives e−1 = 2, e0 = 10, e1 = 24, k=1
e2 = 44, e3 = 66, e4 = 92, e5 = 128, e6 = 170, e7 = 212,
e8 = 258. Although these en are different from the previous ones,
Observe that if we know hK we do not introduce new notation, because they satisfy
n , n ≤ N , and ek , k ≤ K−1,
then we can use (4–2) and (4–3) to compute bn up to the same bounds (4–4). Changing (4–7) into an equation
n ≤ N + K + 3. for bn , and using the result to obtain the upper bound
An explicit formula for ek is apparently too much to for en in (4–5), we get
hope for, but using the trivial facts that ehk and ew k are −3η(n−1)+2n+1 −4 ≤ en ≤ 3η(n−1)+2n·2n −3·2n −6.
nonnegative and bounded above by the number of sites
(4–8)
in their respective regions, we do have the bounds
The dominant term in both upper and lower bounds is
−3(2k + 3)2 < −ehk ≤ ek ≤ 2ew 2
k < 4(2k + 3) . 3n2 2n . This can in fact be improved to 52 n2 2n , since a
 Gravner and Griffeath: Modeling Snow Crystal Growth I: Rigorous Results for Packard’s Digital Snowflakes 433

better bound for ek is obtained by considering where the The estimates of en for large n are as follows (this time
two wedge dynamics and the hole dynamics must agree. we do not bother to symmetrize):
A little geometric argument demonstrates that this agree-
−3η(n − 1) + 6 · 2n + 4 ≤ en ≤ 4η(n − 1) + 6 · 2n + 4.
ment is achieved at least within a region between lines
of slope 2 and 12 through the center of the interaction For this case we compute bn up to n = 20, to get
region. We omit the details, since it is much easier to 4−20 b20 ≈ 0.968854 and
improve the bounds on ρ1 by computing more ek ’s for
small k than by trying to improve (4–8). 0.968618 < ρ14 < 0.969044.
Our computations yield 4−23 b23 ≈ 0.635280, and so, Next, we turn to Hex 16 . All definitions, as well as
using (4–6) and (4–8), the rigorous bounds recursions (4–2) and (4–3), remain exactly the same as
for Hex 1 . Note that the “6” part of the rule influences
0.635248 ≤ ρ1 ≤ 0.635312.
only sites that have no further influence elsewhere, so
Our second rule is Hex 14 . With minor changes our this rule has exactly the same interactions as Hex 1 . For
analysis also applies to Hex 145 , so we will omit that example, we can obtain A∞ for Hex 1 and then perform
case. a single Hex 16 step to obtain A∞ for Hex 16 .
The first basic recursion now reads, for n ≥ 3, The extra step does affect the computation of ek , mak-
ing in small cases the interaction areas the same as for
wn+1 = 3wn + 2wn−1 + hn−1 − 2 · 2n − 2, the Hex 14 rule. To repeat: for ehk we add two layers of
2k + 3 sites at the bottom and on the right of the cen-
and this time
tral (2k + 3) × (2k + 3) box and one layer of 2k + 5 sites
bn = 2wn + 2n .
at the top and on the left of that box, while for ewk the
This yields, with corrections computed separately for added layers are only at the bottom and on the left. The
low n, restriction in (4–3) is 2n−1 ≥ 2k + 8.
We get b0 = 1, b1 = 4, b2 = 16, b3 = 58, b4 = 212,
bn+1 = 3bn +2bn−1 +2hn−1 −6·2n −4−1{n=1} +2·1{n=2} , etc., and e−1 = 4, e0 = 15, e1 = 32, e2 = 56, e3 = 83,
(4–9) e4 = 115, and e5 = 157. By computing h6n up to n = 9,
for n ≥ 1. We have b0 = 3, b1 = 8, b2 = 24, b3 = 78, we obtain 2−18 b18 ≈ 0.739664 and
b4 = 280, . . . , and h0 = 4, h1 = 8, h2 = 25, h3 = 80, . . . .
The second recursion remains (4–3). 0.738902 < ρ16 < 0.740279.
What does change in this case is computational.
Next in line is Hex 156 . The first recursion now is
Namely, in all small cases we use for our estimates, the
similar to (4–9),
computation for bn ends at time 2n + 1, and at time
2n−1 + 1 in the interaction area. For computation of ehk bn+1 = 3bn +2bn−1 +2hn−1 −6·2n +10−6·1{n=1} , (4–10)
this forces us to add two layers of 2k + 3 sites at the bot-
tom and on the right of the central (2k + 3) × (2k + 3) while (4–3) is still the second recursion. In addition,
box and one layer of 2k + 5 sites at the top and on the bn = 2n (2n + 2) up to n = 3; after that, permanently
left of that box. For computation of ew empty triangles (of six sites) appear. Much later (see
k , the added layers
are of course only at the bottom and on the left, creating the discussion in Section 8), larger permanently empty
a (2k + 5) × (2k + 4) box. (The restriction in (4–3) then regions appear.
is 2n−1 ≥ 2k + 8.) In this way, we get e−1 = 2, e0 = 11, In this case, two extra steps are required (again, for the
e1 = 27, e2 = 49, e3 = 82, e4 = 125, and e5 = 170. small cases we have checked) to finalize the configuration
Thus in Bn . The last of these two steps, however, merely fills
n two sites next to the axes and thus does not affect any
bn+1 = 3bn + 2k bn−k − en , computations. The interaction areas are therefore the
k=1
same as for Hex 16 .
where en = en−1 + 6 · 2n + 4 + 1{n=1} − 2 · 1{n=2} for This time, we get b0 = 3, b1 = 8, b2 = 24, b3 = 80,
n ≥ 1, e0 = 1, and en satisfies the same bounds (4–4). b4 = 276, etc., and e−1 = 1, e0 = 9, e1 = 25, e2 = 55,
In this case the renewal theorem gives e3 = 84, e4 = 106, and e5 = 151. By computing h6n up
∞ ∞ to n = 9, we obtain 2−18 b18 ≈ 0.937935 and
bn 3 − 14 n=0 4−n en 11 1  −n 
ρ14 = lim n = 3 1 ∞ −k
= − 4 en .
n→∞ 4 6 6 n=1
4 + 2 k=2 k2 0.937183 < ρ16 < 0.938559.
434 Experimental Mathematics, Vol. 15 (2006), No. 4

Our last rule in this section is Hex 15 . In this case, As in (4–7), let bcn = 2n (2n + 2) − bn and hcn = (2n +
the first recursion is 1) − hn . Then
2

bcn+1 = 3bcn + 2bcn−1 + 2hcn−1

bn+1 = 3bn +2bn−1 +2hn−1 −6·2n +8−6·1{n=1} , (4–11)
and hcn satisfy the analogue of (4–3), with the same
and we get b0 = 3, b1 = 8, b2 = 22, b3 = 70, b4 = 238, bounds (4–4) on ek . Hence
etc., and e−1 = 0, e0 = 6, e1 = 18, e2 = 42, e3 = 66, n

e4 = 84, and e5 = 120. By computing h6n up to n = 9, bcn+1 = 3bcn + 2k bcn−k − en−1 , (5–2)
we obtain 2−18 b18 ≈ 0.802578 and k=1

where en = 0 for n ≤ 28. Furthermore, en satisfies the

0.801822 < ρ16 < 0.803199. bounds (4–4). The renewal theorem therefore guarantees
the existence of a density, and, by (4–4),
We note that again small cases require two steps to fi-
∞
nalize the configuration in Bn . Both steps fill two sites bcn 1  −n 
1 − ρ1456 = lim n = − 4 en
next to the axes and are thus not problematic. The new n→∞ 4 24 n=29
feature is different hole dynamics, requiring one extra ∞
1  −n 2 n 282
step (beyond 2n−1 ) to resolve. Therefore, the interaction < 4 ·n 2 < < 5.3 · 10−7 .
8 n=29 8 ln 2 · 228
areas are the same as for Hex 16 .
Unfortunately, a nontrivial upper bound seems very
difficult to produce by a brute-force approach along the
lines of Section 4. Instead, we devise a rescaling argu-
5. DENSITY OF HEX 1456 AND HEX 146
ment.
For concreteness, we will concentrate on Hex 1456 . The Let R be a time at which there is a permanent 0 in AR
same techniques applied to Hex 146 yield the bounds in (with the initial condition that produces bn ). Clearly we
Theorem 1.2. The definitions of bn and hn are the same can choose R to be a power of 2, and arbitrarily large. We
as in Section 4. The first recursion now is can also ensure that the 0 is at distance at least R/2 (in

·
∞ distance, say) away from the part of the boundary
bn+1 = 3bn + 2bn−1 + 2hn−1 − 6 · 2n − 2, (5–1) of AR strictly inside the first quadrant.
Now run the dynamics in time steps of size R, with
while the second is still (4–3). We note that if hn ≡
the proviso that whenever a hole is formed, the clock
(2n + 1)2 , then bn ≡ 2n (2n + 2). That is, if every hole
inside this hole is reset so that time 0 corresponds to the
fills completely, then An fills the lattice and ρ1456 = 1.
formation of the hole. The dynamics then uses different
At first, simulations suggest that this indeed happens.
clocks on different regions of the plane. The point is
However, after more systematic experimentation one
that every second time, starting with time 1, every added
discovers that H(82) and H(84) do not fill completely;
occupied site is part of a version of AR (its translation
a final triangle of 0’s consists of 25 sites. For the lower
and/or rotation). The other times are used for filling in,
bound, it is also important to observe that all the other
and holes cannot be guaranteed.
holes of sizes at most 130 do fill in. Therefore, all boxes
As we know, the buds inside the holes interact, but
of sizes 2n − 44, n ≥ 7, do not fill in. (In fact, they leave
the interaction will not reach the existing 0 until it is of
2(n − 7) + 1 unfilled triangles of size 25.) Since there are
size R/2, therefore up to (unrescaled) time at least 2R/4 .
infinitely many such boxes generated by the dynamics,
Up to that time, therefore, all the occupied sites will see
the number of unfilled 0’s is infinite. This, however, does
a permanent 0 within distance 4R.
not establish that ρ1456 < 1. In fact, we do not even have
Next, let dn be the site count of sites within distance
a nontrivial lower bound for the density yet, so we start
4R of a permanent 0, at time 2n +1. Since the interaction
with this easier task.
can destroy a 0 and therefore affect sites at distance 4R
A zero-creating hole will occur for the first time inside
from the interaction area, dn satisfy the same recursion
an H(2n + 2) when 2n−22 − 44 ≥ 84; hence n ≥ 28. Thus
as bn , but with error bounds multiplied by 81 · R2 . If
bn = 2n (2n + 2) for n ≤ 29. (This explains why a naive
dcn = 2n (2n + 2) − dn , then
simulation started with a single occupied site will never
n

produce a hole: one would need a system of a size more dcn+1 = 3dcn + 2k dcn−k − en−1 , (5–3)
than 230 ≈ 109 .) k=1
 Gravner and Griffeath: Modeling Snow Crystal Growth I: Rigorous Results for Packard’s Digital Snowflakes 435

where in this version

|en | < 103 R2 n2 2n .

Moreover, we already know that en = 0 for n ≤ R/4.

Therefore
dcn 
lim sup n
< 103 R2 n2 2−n ,
4
n≥R/2

which clearly can be made arbitrarily small for large

enough R.
It follows that a positive proportion of occupied sites
sees a permanent 0 within a fixed distance. Therefore
the final density of 1’s is necessarily strictly less than 1.

6. PROOF OF THEOREM 1.1

We start with the case A0 = {0}. What we need to
FIGURE 5. The division of W for m = 1 and  = 3 into two
prove, roughly, is that the proportion of A∞ -sites in any cases.
nice large subset of Z2 is about ρ. We accomplish this
by the division of space into “mesoscopic” units filled in
Fix an  > 0. It is clear that |4−(n−m−i) wn−m−i − 12 ρ|,
by the basic wedge dynamics (which gives wn ) and much
i = 1, . . . , , can be made simultaneously smaller than 
smaller problematic regions.
if n is chosen large enough. (Note that 12 · 4−(n−m−i)
To this end, let us return to the recursion for wn .
is exactly the area of the corresponding triangle in the
This recursion can be “unrolled.” That is, when not
division.)
both π(3) and π(4) are 1, wn can be written in terms
Pick a closed set S ⊂ W , with nice boundary. For
of wn−m−1 , . . . , wn−m− , and hkn−m− , for some m ≥ 0,
example, we may assume that S is convex and ∂S consists
≥ 2, and suitable k’s. This is the case we will focus on
of finitely many linear pieces. How many points from
for the remainder of the proof. When π(3) = π(4) = 1,
2−n (A∞ ∩ Wn ) does S contain? In the sequel, C is a
then hkn−m− is replaced by a suitable linear combination
“generic constant,” which is allowed to vary from one
of hn−m− , sn−m− , and tn−m− , and the proof is easily
appearance to another.
adapted. (The right side of Figure 5 illustrates the differ-
First, the triangles that intersect ∂S together contain
ence between resulting divisions.) The unrolled recursion
at most C · length(∂S) · 22n−m points. Second, all the
corresponds to a division of Wn into smaller triangles and
parallelograms in W have combined area at most C · 2− ,
parallelograms. If n is large, and Wn is normalized by
and so they contain at most C · 22n− points. A crude
1/2n , this partition approximates a partition of the unit
upper bound on the total length of the boundaries of the
triangle W = {(x, y) : 0 ≤ y ≤ x ≤ 1}. In this way, W
division sets is 4 times the number of division sets, which
is divided into covering triangles (corresponding to wi )
can be (equally crudely) bounded by the reciprocal of the
and parallelograms. We define these sets to be closed,
area of the smallest set in the division. Therefore, the
so there is some intersection along boundaries. Here, m
total boundary length is at most C · 4m+ , and so within
determines the largest size of a triangle used, which is
(microscopic) distance C of these boundaries there are
1/2m+1 times the size of W . Furthermore, gives ex-
at most C · 2n+2m+2 points. Finally, we know that the
actly the number of different triangle sizes used. The top
error terms (if any) contribute at most C · n2 · 2n points.
of Figure 5 pictures the division for m = 1 and = 3,
Therefore, S contains at least
with the remaining parallelograms shaded. The maxi-

mum diameter of sets in the division clearly halves with (ρ − ) · area(S) − C · length(∂S) · 2−m − C · 2−
each successive m. Moreover, the area covered by the
− C · 2−n+2m+2 − C · n2 · 2−n · 4n
parallelograms (uncovered by triangles) halves with each
successive . View m and as large, but much smaller points, for large enough n. An analogous upper bound
than n. also holds. By letting first n → ∞, then m, → ∞, and
436 Experimental Mathematics, Vol. 15 (2006), No. 4

finally  → 0, it follows that However, interaction between the secondary boundary

dynamics is slightly more complicated. When the bound-
1
|S ∩ 2−n (A∞ ∩ Wn )| → ρ · area(S). ary dynamics from the two secondary buds (cf. Section
4n
2) collide, they (together with the primary boundaries)
The proof of Theorem 1.1 for A0 = {0} is now con- create a hole that is less than or equal in size to that from
cluded in a straightforward manner. A0 = {0}. In fact, the hole’s side can be diminished by
For general finite A0 , Theorem 1.1 mostly follows from at most h − 1. By Lemma 4.1, the hole generates smaller
properties of additive dynamics, as we explain below. holes (all of which are smaller, and bounded perturba-
The proof is somewhat delicate, which is not a complete tions of the A0 = {0} case), provided that all the wedge
surprise, since sensitivity to perturbations in initial con- vertices that were dead before are still dead . The prob-
dition is widespread among generic cellular automaton lematic vertices are created in the middle of the holes
rules. In fact, there is a “box snowflake” for which the when secondary boundary dynamics collide. If π(3) and
density of A∞ has been proved to depend on the initial π(4) are not both 1, collisions always result in dead ver-
set. Namely, we have shown in [Gravner and Griffeath 98] tices, as is easy to check. We will deal with this case
that Box 1 solidification yields density 49 starting from first.
a singleton. Later, Dean Hickerson [Hickerson 99] engi- Pick a closed convex set S ⊂ W with piecewise linear
neered finite initial seeds with asymptotic densities 29 64 boundary and an  > 0. In the first step, perform m
61
and 128 . For instance, the latter is achieved by an inge- steps of the Sierpiński triangle construction. That is,
nious arrangement of 180 carefully placed occupied cells take out of W the central triangle congruent to W/2,
around the boundary of an 83 × 83 grid. We do not know then three triangles congruent to W/4, . . . , and finally
the highest density with which Box 1 solidification can 3m triangles congruent to W/2m . Discard the points in
fill the plane, nor whether any seed fills with density less W outside these triangles (see top left of Figure 6 for the
than 49 . m = 3 example). In the second step, mark by M the two
To return to hexagonal digital snowflakes, we will as- smaller triangles from each triangle obtained in the first
sume that A0 has sites on the y-axis, the lowest of which step, as in Figure 6. In the third step, mark four triangles
is at 0 and the highest at (0, h), but no sites to the right in the remaining quadrilateral hole, then four triangles in
of the y-axis. We will concentrate on proving the den- each of the remaining two holes, and so on, for a total
sity result for the part of A∞ between lines y = 0 and of m iterations and 4 · 2m triangles. The top of Figure 6
y = x+h, which, by symmetry and the extreme boundary illustrates these three steps for m = 3. Note that when
dynamics, is clearly enough. We start with the follow- m = m() is large enough, the area of the exceptional set,
ing fact about additive dynamics. Assume the seed is i.e., the set of points outside of marked triangles, is below
T0 = A0 ∩ y-axis. By additivity, when 2n − 1 > h, the /4. For a small enough δ = δ(, m), the area of We , the
rightmost column of T2n consists of two copies of T0 , sep- δ-neighborhood of the exceptional set, is below /2.
arated by 2n − 1 − h 0’s. The column immediately to the Now take a large n and consider 2−n A∞ ∩ W on one
left of these 0’s consists either of all 1’s (flanked by two of the marked triangles WM , off We . By Lemma 2.1, this
0’s) or else of all 0’s, flanked on the top by a 1 and on configuration exactly equals the configuration of sites in
the bottom by an additional 0, below which there is a 1. a basic wedge if n ≥ n0 (δ). Therefore, by taking n even
In the second case we can iterate and continue moving to larger if necessary,
the left until we encounter a column of 1’s (flanked by a
0 on both ends). Assume that this column consists of D |2−n A∞ ∩WM ∩Wec ∩S| ≥ ρ·area(WM ∩Wec ∩S)−6−m−1 ,
1’s. An important point is
by the already proved result for basic wedges. Since the
2n − h ≤ D ≤ 2n . number of marked triangles is less than 2 · 6m , it follows
that
Furthermore, if this column is created (by the boundary
|2−n A∞ ∩ S| ≥ ρ · area(S) − ,
dynamics) at time t, then at time t + D + 1 the boundary
dynamics creates a horizontal and a diagonal segment for n ≥ n0 (), as desired. The analogous upper bound is
that together with the above column seal off a triangle of proved in the same fashion.
0’s. Into this triangle (which is a bounded perturbation When π(3) = π(4) = 1, the middle vertices in the
of the one started from a singleton), the Hex dynamics holes are live as soon as D < 2n . Although h > 0 does
spread as when they fill basic wedges. not necessarily imply that D < 2n , there is for each h a
 Gravner and Griffeath: Modeling Snow Crystal Growth I: Rigorous Results for Packard’s Digital Snowflakes 437

There are two kinds of macroscopic dynamics for dig-

ital snowflakes, both denoted here by Sa , a ∈ [0, 1]. The
sets Sa will be a strictly increasing family of closed sub-
sets of the hexagon co(N ) ⊂ R2 . We start with the
simpler case, which we call simple hole dynamics (SHD).
This evolution will be associated with the 12 rules for
which π(3) and π(4) are not both 1.
Assume a < 1 and write the dyadic expansion a =
0.a1 a2 . . . with infinitely many 0’s. We set a = 0.a2 a3 . . .,
a = (a ) , etc., and use the following transformations:
reflection σ about the line y = x, rotation ρα through an
angle α, and deformation δ by the linear transformation
FIGURE 6. The three steps and marked triangles for m = 3.
(x, y) → (x − y, y). We will define only the portion of Sa
Regions outside them are shaded.
inside the triangle W = {(x, y) ∈ R2 : 0 ≤ x ≤ y ≤ 1};
other parts are then obtained by symmetry, i.e., by an
unique T0 that has D = 2n , namely the one generated appropriate use of the above transformations.
by the additive dynamics starting from {0} and run for We will first recursively define Sa ⊂ W , and a corre-
h time steps (and translated back so that its lowest site sponding hole dynamics Ha ⊂ W on dyadic rationals a
is at 0). that have only finitely many 1’s in their dyadic expan-
In this case, therefore, the division into marked trian- sion, starting with Sa = {(0, 0)} and Ha = ∅ for a = 0.
gles in step 3 has two possible forms, one for D = 2n The rules described in the next paragraph and Figure 7
and one for D < 2n . Both are indicated in the bottom are a natural extension of the recursions for wn and hn
of Figure 6. In either case, the proof is then a minor for the corresponding 12 rules. (We remark that when
modification of the above. no two successive ai ’s are 1, one can define the boundary
This concludes the proof of Theorem 1.1. To finish of Sa by a Koch-type substitution scheme, as outlined in
this section, we assume π(3) = 0 and prove (4–1), or Section 1 for a = 13 and in [Gravner and Griffeath 98]
equivalently, that for the dynamics on Wn , for Box 1 solidification; such algorithms are in general
precluded by collisions inside the holes.)
1
4−n |A2n −1 ∩ Wn | → ρ. If a1 = 0, then take Sa = 12 Sa
. If a1 = 1 and a2 = 0,
2
then Sa consists of five pieces:
In fact, the recursion for wn can be used to show that  
Sa = 12 W ∪ 12 , 0 + 12 Sa
∪ 12 , 12 + 12 Sa

A2n +Cn2 ∩ Wn = A∞ ∩ Wn , since the largest interaction  
area inside the holes is on the order of n2 . (It is possible ∪ 12 , 0 + 14 σSa

∪ 12 , 12 + 14 ρ−π/2 δSa

.
that in all cases, A2n +c ∩ Wn = A∞ ∩ Wn , for some If a1 = 1 and a2 = 1, then Sa consists of seven pieces:
small constant c, but we cannot prove this.) Then, in  
the division as in the case A0 = {0} above, only the Sa = 12 W ∪ 12 , 0 + 12 Sa
∪ 12 , 0 + 12 Sa

 
dynamics inside triangles that intersect the right edge ∪ 12 , 12 + 14 σW ∪ 12 , 12 + 14 ρ−π/2 δW
 
of W is not done by time 2n . These contain at most ∪ 12 , 14 + 14 Ha

∪ 1, 12 + 14 ρπ Ha

.
C · 22n−m points, and the proof is concluded as before.
If a1 = 0, then
(Growth inside some parallelograms might also persist,
but such sites are already incorporated into the error.) Ha = (1, 1) + 12 ρπ σSa
∪ (1, 0) + 12 ρπ/2 δSa
,
and if a1 = 1, then
7. MACROSCOPIC DYNAMICS Ha = (1, 1) + 12 ρπ σW ∪ (1, 0) + 12 ρπ/2 δW ∪ 12 Ha


In this section we assume A0 = {0} and study At for a ∪ 1, 12 + 12 ρπ Ha
.
large finite t instead of A∞ . As the radius of At increases
It is clear from the construction that both Sa and Ha
linearly in t, it is natural to ask whether these dynamics
are well defined and increasing on dyadic rationals. More-
have a hydrodynamic limit as space and time are scaled
over, we claim that |a − b| < 2−n implies
proportionally. Computer simulations certainly suggest √ √
so, and are validated by our analysis below. dH (Sa , Sb ) ≤ 2 · 2−n , dH (Ha , Hb ) ≤ 2 · 2−n , (7–1)
438 Experimental Mathematics, Vol. 15 (2006), No. 4

a1 = 0 a1a2 = 10 a1a2 = 11

Sa
Ha

Sa

Sa


Sa

Sa :

a1 = 0 a1 = 1
Sa
Ha

Ha :

FIGURE 7. SHD. Dark areas are fully covered by appro-

priately mapped W , while arrows indicate the placement of
suitably transformed sets.
FIGURE 8. Occupied set of Hex 13 , started at {0}, at times
a · 210  for a = 23 , 14
15
, and 62
63
in darker, lighter, and darker
where dH is the Hausdorff metric. To prove (7–1), we shade, respectively. This illustrates problems with surmising
proceed by induction on n. Note that we can assume fractal properties from pictures: the 23 case appears to have
the “most fractal” boundary, which in fact has dimension 1.
a < b and, by monotonicity, that ak = 0, for k ≥ n and
The fairly large fractal dimension in the other two cases is a
bn = 1. The assumption also means that ak = √ bk for consequence of growth that appears at the exposed corners
k < n. Since diameter(W ) = diameter(δW ) = 2, it is and is not visible at this scale.
easy to verify (7–1) for n = 0 and n = 1. For n ≥ 2,
we can use the induction hypothesis on a , a and b , b .
Note that all a sets are normalized by 12 and all a sets dimension dimH of the said boundary. The answer in
by 14 , which yields (7–1). many cases is yes, and we turn to this task now.
The uniform continuity (7–1) immediately implies As in [Gravner and Griffeath 98], we need to count the
that we can extend both Sa and Ha uniquely to a ∈ [0, 1], number of buds of size 2−n . Let us denote this number
so that (7–1) holds for a, b ∈ [0, 1], S1 = H1 = W , and by fn (a) for the triangle and gn (a) for the (half) hole
dynamics. Then
Sa = ∩b>a Sb = closure(∪b<a Sb ), Ha = ∩b>a Hb ⎧
⎪ 
= closure(∪b<a Hb ), ⎨fn−1 (a ) if a1 = 0,
fn (a) = 2fn−1 (a ) + 2fn−2 (a ) if a1 a2 = 10,
⎪
⎩
where b in all cases ranges over dyadic rationals. The fol- 2fn−1 (a ) + 2gn−2 (a ) if a1 a2 = 11,
lowing result identifies Sa as a subsequential limit of At .
and

Theorem 7.1. Assume that tn = a · 2n , for a ∈ [0, 1] and 2fn−1 (a ) if a1 = 0,
A0 = {0}. The 12 rules that do not have both π(3) = 1 gn (a) =
2gn−1 (a ) if a1 = 1.
and π(4) = 1 exhibit subsequential convergence to SHD:
When a is rational, these equations reduce to a finite lin-
2−n Atn → Sa , ear recursion, so that fn (a) ∼ cλn , for some c > 0 and
λ ∈ (0, 4]. If λ ≤ 2, then it follows immediately that
in the Hausdorff metric, as n → ∞, uniformly in a.3
dimH (∂Sa ) = 1. If λ > 2, then these are Mauldin–
Williams [Mauldin and Williams 88] fractals. Hence
Proof: For dyadic rationals a, the proof follows from ar-
the box-counting and Hausdorff dimensions agree, and
guments in Section 6. The case of general a ∈ [0, 1] is
dimH (∂Sa ) = log λ/ log 2.
then a consequence of monotonicity of At and St .
To illustrate, for N ≥ 2 let us take a with periodic
binary representation that repeats N − 1 initial 1’s and
The sets Sa are difficult to visualize for a’s that are not
then a single 0, so a = (2N − 2)/(2N − 1). Index the
dyadic rationals. In particular, it is clear that they have
shifted configurations with superscripts. We get
a very complicated boundary. It is therefore natural to
ask whether one can determine the fractal (Hausdorff) N −1
gn1 = 2gn−1
2 3
= 4gn−2 = · · · = 2N −2 gn−(N −2)
3 The limit also equals ρS in the weak sense, but Hausdorff
a
convergence seems more to the point. = 2N −1 fn−(N −1) .
 Gravner and Griffeath: Modeling Snow Crystal Growth I: Rigorous Results for Packard’s Digital Snowflakes 439

a1 = 0 a1a2 = 10 a1a2 = 11
Ha3

a1 = 0 or a1 a2 = 10, the recursion for Sa is the same as
Sa
Ha1

Sa

Sa

before, while if a1 a2 = 11, then
Sa

  
Sa : Sa = 12 W ∪ 12 , 0 + 12 Sa
∪ 12 , 0 + 12 Sa
∪ 12 , 0 + 14 W
 
∪ 12 , 12 + 14 ρ−π/2 δW ∪ 12 , 14 + 14 Ha1


a1 = 0 a1 = 1 
Sa
Ha2

∪ 1, 12 + 14 ρπ Ha3

.
If a1 = 0, then
Ha1 :
Ha1 = 12 Sa ∪ (1, 1) + 12 ρπ σSa
∪ (1, 0) + 12 ρπ/2 δSa
,
a1 = 0 a1 = 1
Ha2 = Ha3 = (1, 1) + 12 ρπ σSa
∪ (1, 0) + 12 ρπ/2 δSa
,
Sa
Ha1
while if a1 = 1,
Ha3

Ha2 : Ha1 = 12 W ∪ (1, 1) + 12 ρπ σW ∪ (1, 0) + 12 ρπ/2 δW
∪ 12 ρπ Ha2
,
a1 = 0 a1 = 1 Ha2 = (1, 1) + 12 ρπ σW ∪ (1, 0) + 12 ρπ/2 δW ∪ 12 ρπ Ha1

Sa
Ha3

∪ 1, 12 + 12 ρπ Ha3

Ha3 :
Ha3 = (1, 1) + 12 ρπ σW ∪ (1, 0) + 12 ρπ/2 δW ∪ 12 ρπ Ha3


∪ 1, 12 + 12 ρπ Ha3
.
FIGURE 9. DHD. Live vertices within the H’s are indicated
by arrows. These are extended to a ∈ [0, 1] in the same way as
before, yielding the following result.

Also, Theorem 7.2. Assume that tn = a · 2n , for a ∈ [0, 1], and

A0 = {0}. The four rules with π(3) = 1 and π(4) = 1
fnN −1 = fn−1
N
= fn−1 , exhibit convergence to DHD:
N −1
fnN −2 = 2fn−1 N
+ 2fn−2 = 4fn−2 ,
N −2 N −1
2−n Atn → Sa ,
fnN −3 = 2fn−1 + 2gn−2 = 12fn−3 .
in the Hausdorff metric as n → ∞, uniformly in a.
Recursively thereafter we get fnN −k = ck fn−k , using
It follows from Section 6 that for these four rules, most
N −(k−1) N −(k−2)
fnN −k = 2fn−1 + 2gn−2 = (2ck−1 + 2k−1 )fn−k . finite seeds give rise to macroscopic dynamics different
from DHD. The more general construction is quite anal-
Therefore c1 = 1, and for k ≥ 1, ck satisfy the recursion ogous, however, so we will not describe it in detail.
ck = 2ck−1 + 2k−1 . It follows that ck = k · 2k−1 , so Dimension computations are similar to those for SHD,
fN = N · 2N −1 fn−N , and λ = 2 · (N/2)1/N . Hence but more complicated. Now, there are fn (a) and gni (a),
i = 1, 2, 3, that satisfy
1 log2 N ⎧
dimH (∂Sa ) = 1 − + . 
N N ⎪
⎪fn−1 (a ) if a1 = 0,
⎪
⎨2f  
n−1 (a ) + 2f n−2 (a ) if a1 a2 = 10,
For example, dimH (∂S2/3 ) = 1 and dimH (∂S14/15 ) = fn (a) =
⎪2fn−1 (a ) + gn−2 (a )
⎪
 1 
dimH (∂S245/255 ) = 54 ; the dimension achieves its max- ⎪
⎩
imum (within this collection of examples) of approxi-
3
+ gn−2 (a ) if a1 a2 = 11,
mately 1.289 at a = 62 63 , and tends to 1 as N → ∞.
and

See Figure 8. 3fn−1 (a ) if a1 = 0,
Note that we do not know how to determine dimH (Sa ) gn1 (a) = 2
gn−1 (a ) if a1 = 1,
when a is irrational, nor the maximum possible value 
of dimH (Sa ). 2fn−1 (a ) if a1 = 0,
gn2 (a) = 1  3 
The diverse hole dynamics (DHD) is associated with gn−1 (a ) + gn−1 (a ) if a1 = 1,
the four Hex rules that have π(3) = π(4) = 1. Besides Sa , 
2fn−1 (a ) if a1 = 0,
there are now three hole dynamics Hai ⊂ W , i = 1, 2, 3, gn3 (a) =
all initialized at ∅ when a = 0 (cf. Figure 9). When
3
2gn−1 (a ) if a1 = 1.
440 Experimental Mathematics, Vol. 15 (2006), No. 4

FIGURE 10. Dynamics in H(84) at times 34, 49, 56, 58, 66. Arrows indicate which directions of boundary dynamics (from
nearby buds) lead to the residual hole.

For the same example a = (2N −2)/(2N −1) as in the SHD two buds: b1 at (2, 2n−1 +4) (moving eastward) and b2 at
case, a similar but more involved computation yields (2n−1 +2, 2n−1 +4) (moving diagonally toward the north-
west), and their two symmetrically located counterparts
dimH (∂Sa ) b1 and b2 . Now let us pause at time t2 = t1 + 2n−2 − 1
 
=
1
· log2
1 7
N · 2N + 2N −
11 5
− 1{N . (49 in the example). The two buds b2 and b2 would have
even}
N 3 9 9 9 created two diagonals

8. THICKNESS AND RELATED ISSUES (2n−1 + 2 − 2n−2 + 1, 2n−1 + 4)

to (2n−1 + 2, 2n−1 + 4 + 2n−2 − 1)
Let us begin the proof of Theorem 1.3 by considering the
eight digital snowflakes that are not exactly solvable. For and
the argument that A∞ is never thick, we address only the
hardest case, Hex 1456 ; the same method applies equally (L − (2n−1 + 4) − 2n−2 + 1, L − (2n−1 + 2))
well to the other seven rules. to (L − (2n−1 + 4), L − (2n−1 + 2) + 2n−2 − 1)
The key fact (initially suggested by simulations) is
that holes of sizes in the sequence below, obtained by if not for the interaction between them. By additivity,
doubling and subtracting 4, we need to eliminate the intersection between the two,
which consists of two sites, as is easy to check. At the
84, 164, 324, 644, . . . , same time, b1 creates a column
create holes, each of which is twice as large as the pre- (2n−2 + 2, 2n−1 + 4) to (2n−2 + 2, 2n−1 + 4 + 2n−2 − 1).
vious one. Although this should follow from a general
rescaling property of the dynamics (which we do not even Note that this does not interfere with the diagonal
know how to formulate), the property is a consequence created by b2 (on which the smallest x-coordinate is
of a finite chain of basic interactions, as illustrated by a 2n−2 + 3).
sequence of intermediate times for H(84) in Figure 10. Now it is easy to see that the gap of two sites created
Before we write out the details, note that the configu- by the interaction between b2 and b2 creates two buds
ration in the hole is symmetric with respect to the map that generate the same boundary dynamics (in the north-
(x, y) → (L − y, L − x), and write L = 2n + 2n−2 + 4, west direction) as would a single 1 at time t2 − 1 imme-
n ≥ 6. The first important configuration to consider diately southeast from the gap, at (2n−1 + 2, 2n−1 + 2n−2
occurs at time t1 = 2n−1 + 2 (t1 = 34 when n = 6). Con- +2). Call this seed b3 . Also consider the bud b4 (and its
centrate on the extreme boundary dynamics generated by symmetric counterpart b4 ) at (2n−2 + 3, 2n−1 + 2n−2 + 4)
 Gravner and Griffeath: Modeling Snow Crystal Growth I: Rigorous Results for Packard’s Digital Snowflakes 441

0 0 the smallest wedges and holes contain 1’s, and choose

0 02 01 a suitable R, an upper bound on the distance between
0 1 0x any 0 and the set of 1’s. An easy induction shows that
0 01 R provides the same bound for larger wedges and holes.
0 0 General initial sets are handled by arguments in Section
FIGURE 11. Local configuration at time t − 1. 6. (Note, however, that the maximal distance of a 0 from
(immediately northeast from the top of the column in the set of 1’s in A∞ may not be bounded independently
the previous paragraph), which creates extreme bound- of A0 .) Thickness of Ac∞ is equally easy to establish.
ary dynamics moving eastward. The next time to pause To finish the proof of Theorem 1.3, we need to demon-
is just before these two interact, at time t3 = t2 +2n−3 −1 strate point (2).
(which is 56 in our example). At this time, b3 creates a We start with Hex 13456 . First note that a finite seed
diagonal will generate six rays of 1’s via the extreme boundary
dynamics. Assume that there is a nonempty connected
(2n−1 + 2 − (2n−3 − 1), 2n−1 + 2n−2 + 2) (in the hexagonal lattice sense) set of 0’s in A∞ . If this
to (2n−1 + 2, 2n−1 + 2n−2 + 2 + 2n−3 − 1) set is infinite, then it must extend to infinity, say between
the eastern and northeastern ray of 1’s. From now on we
plus two new buds at its ends, one to the west of the refer to the sites within that portion of A∞ . Start at some
bottom point, one to the north of the top point. It fol- point between the rays, and move up to the topmost 0;
lows that the boundary dynamics from b4 is free to make call it 01 . Assume first that its northeast neighbor is
one more step free from interruption from b3 to create a 0; call it 02 . There cannot be any 0’s above 02 (or else
column the topmost 0 would see at least three 1’s). Then the
(2n−1 − 2n−3 + 2, 2n−1 + 2n−2 + 4) northeast neighbor of 02 must be a 0, call it 03 , again
with no 0’s above it.
to (2n−1 − 2n−3 + 2, 2n−1 + 2n−2 + 4 + 2n−3 − 3).
Continuing in this fashion, we create an infinitely ex-
This column would be higher if not for interference from tended diagonal of 0’s. But the boundary dynamics alone
b4 , which creates a horizontal column at the same time. precludes this, for otherwise they would never create a 1
These two, together with an additional site from b3 , seal directly above the diagonal, and then directly above that,
the hole at time t3 + 1. Three new sites are occupied etc., finally eliminating the ray of 1’s. Therefore the right
inside at the next time, but the remaining sites remain 0 neighbor of 01 must be a 0, again call it 02 , and the entire
forever (if n ≥ 6). column above 02 must be 1’s. By the same argument, the
Since the dynamics create a hole of every size in the right neighbor of 02 , named 03 , must be a 0 and the entire
sequence, infinitely many times and no matter what the column above it 1’s. This procedure eventually creates a
finite seed A0 (cf. Sections 4 and 6), it follows that there horizontal ray of 0’s, which is equally impossible by the
are arbitrarily large islands of 0’s in the final state A∞ . boundary dynamics. The only remaining possibility is
Therefore A∞ is not thick. that the connected set of 0’s is finite. But then simply
The argument that Ac∞ is always thick for Hex 1 is observe that the rightmost among its top 0’s must see
quite simple. Namely, we will show that every 1 in A∞ at least three 1’s. This contradiction demonstrates that
that is at ∞ -distance 3 or more from A0 must have a 0 Ac∞ = ∅, as claimed.
within ∞ -distance 2. Assuming that this is not the case, We conclude this section by showing that Hex 13456
let a 1 at x be a counterexample. Let t be the time at is the only case for which A∞ does not have infinitely
which x becomes 1. We label 0x the 0 at x at the time many 0’s regardless of A0 . It is easy to observe that the
t − 1. We will also assume, without loss of generality, other three candidates (i.e., rules with density 1) may
that the single 1 in x + N at this time is to the left of x. leave some 0’s: Hex 136 and Hex 1346 will not fill a
It follows that the two 0’s labeled 01 in Figure 11 must domino 00 in a sea of 1’s, while Hex 1356 will not fill a
also become 1 at time t (to avoid being a 0 with two triangle 000 in a sea of 1’s.
neighboring 1’s), and then so does 02 . This forces the To generate infinitely many 0’s, let the initial set A0
remaining 0’s in Figure 11 and yields a contradiction, consist of a column of L 1’s. Then the primary boundary
since the 1 cannot be isolated from other 1’s. encloses a triangle with this column as one side, at time
Let us turn next to the exactly solvable rules. To show L. This repeats every time the boundary dynamics re-
that A∞ is thick when A0 = {0}, we simply note that creates the initial set, hence infinitely many times. To
442 Experimental Mathematics, Vol. 15 (2006), No. 4

show that there is an A0 with infinitely many 0’s in A∞ , To our knowledge, the simplest nontrivial example
it is therefore enough to find an L for which the described of an exactly solvable CA is Diamond 1 solidification,
triangle does not completely fill in. which is a modification of Hex 1 using the neighborhood
By exhaustive computer search, the smallest such L’s {(0, 0), (0, ±1), (±1, 0)}. In this case (the “intricate, if
for rules Hex 136 , Hex 1346 , and Hex 1356 are, respec- very regular, pattern of growth” depicted on [Wolfram
tively, 5, 19, and 42. We do not know whether there are 02, p. 171]), it can be shown by induction that x ∈ / A∞
smaller seeds with this property, but it seems that very iff max{k : i1k = 1} = max{k : i2k = 1}. It is easy to
small random initial seeds (those that fit into a 5×5 box, construct a (two-state) finite automaton that checks this
say) are very likely to have A∞ = Z2 . condition, and the density ρ of A∞ evidently must satisfy
the equation ρ = 12 + ρ4 , so that ρ = 23 . It is also worth
noting that the first-quadrant portion of this A∞ is a
fixed point of the substitution system 1 → 10 11 , 0 → 01 .
10

9. EXACT SOLVABILITY Although, by Cobham’s theorem [Cobham 72, Theorem

The phrase “exact solvability” is popular in statistical 3], [Allouche and Shallit 03, Theorem 14.2.3], a substi-
physics [Baxter 82], but its exact meaning is difficult to tution representation of A∞ must exist in any exactly
pin down. By contrast, as we will now explain, in the dis- solvable case, we have no simple explicit construction for
crete world of deterministic CA this concept does have a our Hex examples.
natural rigorous framework based on computational com- The proof that Hex 13 is exactly solvable is quite a bit
plexity. We will restrict our formulations to the final set messier than the recursion for bn , since symmetry seems
A∞ , although generalization to the entire space-time evo- difficult to exploit. We now sketch the construction of the
lution is straightforward. (In fact, at the cost of an extra required finite automaton, as encoded in Figure 12. The
dimension, it is easy to represent the evolution of any number of states will be determined by a number of possi-
two-state CA as the final configuration of a solidification ble division squares. Except in three cases, these squares
rule.) Intuitively, we require that “for a given x ∈ Z2 , are further divided into two triangles by the northeast di-
it is computationally easy to decide whether x ∈ A∞ .” agonal. Each of these triangles is either a wedge, in which
This notion of course depends on an initial set A0 , which case it is equipped with an arrow at the vertex from which
will be assumed to be {0} unless otherwise specified. For- the wedge propagates in the dynamics, or a half-hole that
mally, we call a solidification CA exactly solvable (from is equipped with two arrows. The half-hole encodes dy-
A0 ) if there exists a finite automaton that upon encoun- namics filling in one-half of the diamond-shaped holes.
tering x as input, decides whether x ∈ A∞ . Representa- We need to distinguish wedges and half-holes with differ-
tion of x as input is given as (±i11 , ±i21 , i12 , i22 , . . . ), where ent orientations, so there are six of the former and four
i11 , i21 are the most significant binary digits of the first and of the latter. The first exceptional (undivided) division
second coordinates of x; i12 , i22 the next-most significant, square is a full hole, which has a single representative and
etc. (Some initial i1k ’s or i2k ’s may be 0, and the repre- encodes filling a square hole. The other two exceptions,
sentation is finite but of arbitrary length.) This means arbitrarily named B-square and C-square, have four rep-
that A∞ is automatic [Allouche and Shallit 03], or equiv- resentatives each and encode filling the “complex” part
alently a uniform tag system [Cobham 72]. More general of a wedge. (A B-square is marked by three arrows, while
representations of inputs are sometimes desirable, partic- a C-square is marked by a single arrow, as shown in Fig-
ularly if one wants to study CA on more general lattices, ure 12.) The number of possible division squares can be
but this one suffices for our present purposes. Note that reduced by observing that two half-holes cannot appear,
exact solvability puts a limit on the computational com- and neither can two wedges with arrows that originate
plexity of a CA: for example, it cannot be universal.4 at the corners of the southeast diagonal. This makes a
Also, by [Cobham 72, Theorem 6] such A∞ cannot have total of 29 possible division squares.
an irrational density, in contrast to the limit set in [Grif- Assume that both initial signs are +. Draw the ini-
feath and Hickerson 03] generated by a suitable initial tial division square, representing the initial state of the
seed in the game of life. automaton, which has two wedges, each with an arrow
at the lower left corner. The situation is reflected when
4 Note that since there are one-dimensional universal CA [Cook
both the initial signs are −, while in the case of a + and
05], there are two-dimensional universal solidification CA. A uni-
versal CA can generate any recursive language, hence one that is
a − the initial state is a C-square.
not recognizable by a finite automaton.
 Gravner and Griffeath: Modeling Snow Crystal Growth I: Rigorous Results for Packard’s Digital Snowflakes 443

00 W 01
B
W

W W B
10 11

W H
H W
B
W
W B FIGURE 13. Hole correction. First correction (on the larger
scale) is in the direction (0, 1), next in the direction (0, −1).

W
C One can also prove that Hex 134 is exactly solvable by
W
the same method. The algorithm is a little more involved,
W
since it necessitates several more types of full holes and
half-holes, but there is nothing conceptually new, so we
omit the details.
W
C
W
As it turns out, we have already established enough
C structure to establish the relative complexity of the eight
W F remaining digital snowflakes. Indeed, the following result
W
[Gravner and Hickerson 06], together with Theorems 1.1–
1.3, proves that none of them can be exactly solvable from
F W any initial set.
W
F
W Lemma 9.1. If an automatic set S ⊂ Z2 has a strictly
F
W positive asymptotic density, then it is thick.

W H In conclusion, we remark that complexity analysis of

H H cellular automata by means of finite automata was initi-
W ated by S. Wolfram in [Wolfram 84b]. The perspective
of Wolfram’s paper is in a sense opposite to ours, in that
it measures how the complexity of the set of all possi-
FIGURE 12. Division algorithm. Transition according to ble configurations increases over time. In contrast, we
the value of i1 , i2 is indicated in the top line, and pieces are consider a single configuration at the “end of time.”
labeled in the obvious fashion.
ACKNOWLEDGMENTS
We thank Ken Libbrecht for sharing with us his unmatched
expertise on real snowflakes, and Mirek Wójtowicz for de-
After reading (i11 , i21 ), the next state is one of the four veloping the MCell cellular automata explorer. We are also
grateful to Dean Hickerson, who pointed out that the rules
subsquares, two of the same type as the initial one, the
Hex 14 , Hex 145 , Hex 146 , and Hex 1456 , started from a
other two B-holes. This division is of course dictated by singleton, all generate different final sets.
how the dynamics fills the square. The next two bits The first author was partially supported by NSF grant
induce further division, and so forth. Figure 12 provides DMS–0204376 and the Republic of Slovenia’s Ministry of Sci-
a check that any of the 29 division squares gets divided ence program P1-285. The second author was partially sup-
into 4 division squares. ported by NSF grant DMS–0204018.
A minor final complication is that the holes are not
divided into two smaller holes by exact wedges, but by REFERENCES
their translation in a coordinate direction by 1. These
[Adam 05] J. A. Adam. “Flowers of Ice—Beauty, Symmetry,
corrections do not accumulate, since they go in opposite and Complexity,” a review of The Snowflake: Winter’s
directions from one scale to the next, as illustrated by Secret Beauty. Notices Amer. Math. Soc. 52:4 (2005),
Figure 13. 402–416.
444 Experimental Mathematics, Vol. 15 (2006), No. 4

[Allouche and Shallit 03] J.-P. Allouche and J. Shallit. Auto- [King 64] B. W. King. “Snowflake Curves.” Math. Teacher
matic Sequences: Theory, Applications, Generalizations. 57 (1964), 219–222.
Cambridge: Cambridge University Press, 2003.
[von Koch 04] H. von Koch. “Sur une courbe continue sans
[Baxter 82] R. J. Baxter. Exactly Solved Models in Statistical tangente, obtenue par une construction géométrique
Mechanics. New York: Academic Press, 1982. élémentaire.” Arkiv för Mathematik, Astronomi och
[Bentley and Humphreys 62] W. A. Bentley and W. J. Hum- Fysik 1 (1904), 681–702.
phreys. Snow Crystals. New York: McGraw-Hill, 1931 [Levy 92] S. Levy, Artificial Life: The Quest for a New Cre-
(reprinted New York: Dover, 1962). ation. Pantheon Books, 1992.
[Berlekamp et al. 04] E. R. Berlekamp, J. H. Conway, [Libbrecht 05] K. Libbrecht. “The Physics of Snow Crystals.”
R. K. Guy. Winning Ways for Your Mathematical Plays, Reports on Progress in Physics 68 (2005), 855-895.
second edition, vol. 4. Wellesley, MA: A K Peters 2004.
[Libbrecht 06] K. Libbrecht. “Snowflakes and Snow Crys-
[Cobham 72] A. Cobham. “Uniform Tag Sequences.” Math. tals.” Available online (http://www.its.caltech.edu/
Systems Theory 6 (1972), 164–192. \char126atomic/snowcrystals/), 2006.
[Cook 05] M. Cook. “Universality in Elementary Cellular Au-
[Libbrecht and Rasmussen 03] K. Libbrecht and P. Ras-
tomata. Complex Systems 15 (2005), 1–40.
mussen. The Snowflake: Winter’s Secret Beauty. Osce-
[Descartes 37] R. Descartes. Les Météores, 1637, edited by ola, WI: Voyageur Press, 2003.
Adam et Tannery. Paris, Vrin, t. IV, 1965.
[Magnus 55] O. Magnus. De Variis Figuris Nivium, 1555.
[Dixon 91] R. Dixon. Mathographics. New York: Dover, 1991.
[Mauldin and Williams 88] R. D. Mauldin, S. C. Williams.
[Feller 68] W. Feller. An Introduction to Probability Theory “Hausdorff Dimension in Graph Directed Constructions.
and Its Applications, vol. 1. New York: Wiley, 1968. Trans. Amer. Math. Soc. 309 (1988), 811–829.
[Gardner 76] M. Gardner. “Mathematical games.” Scientific [Meakin 98] P. Meakin. Fractals, Scaling and Growth Far
American 133 (December 1976), 124–128. from Equilibrium. Cambrige: Cambridge University
[Gleick 87] J. Gleick. Chaos: Making a New Science. New Press, 1998.
York: Penguin, 1987.
[Nakaya 54] U. Nakaya. Snow Crystals: Natural and Artifi-
[Gravner and Griffeath 98] J. Gravner and D. Griffeath. cial. Cambridge: Harvard University Press, 1954.
“Cellular Automaton Growth on Z2 : Theorems, Exam-
[Packard 86] N. H. Packard. “Lattice Models for Solidifica-
ples and Problems.” Advances in Applied Mathematics
tion and Aggregation.” Preprint, Institute for Advanced
21 (1998), 241–304.
Study, 1984 (reprinted in Theory and Application of Cel-
[Gravner and Griffeath 06] J. Gravner and D. Griffeath. lular Automata, edited by S. Wolfram, pp. 305–310. River
“Modeling Snow Crystal Growth II.” In preparation. Edge, NJ: World Scientific, 1986.
[Gravner and Hickerson 06] J. Gravner and D. Hickerson. [Pimpinelli and Villain 99] A. Pimpinelli and J. Villain.
“Asymptotic Density of an Automatic Sequence Is Uni- Physics of Crystal Growth. Cambridge: Cambridge Uni-
form.” In preparation. versity Press, 1999.
[Griffeath 06] D. Griffeath. “The Primordial Soup Kitchen.” [Willson 84] S. J. Willson. “Cellular Automata Can Generate
Available online (http://psoup.math.wisc.edu/extras/ Fractals.” Discrete Appl. Math. 8 (1984), 91–99.
1ormore/1ormore.html), 2006.
[Wójtowicz 05] M. Wójtowicz. “Mirek’s Cellebration: A 1D
[Griffeath and Hickerson 03] D. Griffeath and D. Hickerson.
and 2D Cellular Automata explorer.” Available online
“A Two-Dimensional Cellular Automaton with Irrational
(http://www.mirwoj.opus.chelm.pl/ca/), 2005.
Density.” In New Constructions in Cellular Automata,
edited by D. Griffeath and C. Moore. Oxford: Oxford [Wolfram 84a] S. Wolfram. “Computer Software in Science
Univ. Press, 2003. and Mathematics.” Scientific American 251 (September
1984), 188–203.
[Hickerson 99] D. Hickerson. Personal communication, 1999.
[Hooke 03] R. Hooke. Micrographia, 1665. New York: Dover, [Wolfram 84b] S. Wolfram. “Computation Theory of Cellular
2003. Automata.” Comm. Math. Phys. 96 (1984), 15–57.

[Kepler 66] J. Kepler. Strena Seu de Nive Sexangula, 1611; [Wolfram 02] S. Wolfram. A New Kind of Science. Wolfram
The Six-Cornered Snowflake, translated by Colin Hardie. Media, 2002.
Oxford: Clarendon Press, 1966.

Janko Gravner, Mathematics Department, University of California, Davis, CA 95616 ([email protected])

David Griffeath, Department of Mathematics, University of Wisconsin, Madison, WI 53706 (griff[email protected])

Received August 16, 2005; accepted in revised form March 26, 2006.