Tessellations

Chaim Goodman-Strauss

Tessellations — the patterned repetition of small ceramic or stone

pieces — appear all over the world, in virtually every culture with a tra-
dition of permanent, formal building, over the last several thousand years,
perhaps reaching an apotheosis in the splendid and sumptuous traditions
of the Islamic world. The technique splendidly solves an important design
question: how to create a pleasing ornamental surface of any needed size,
as in Figure 1.
The craft of tile work, per se, requires formally constructed buildings
with permanent surfaces that can be tessellated, structures only possible in
more complex civilizations. However, human preoccupation with geomet-
ric ornament is far older still, and far more widespread. In virtually every
human society, textiles, wall ornament and pottery illuminate a fundamen-
tal human desire for the pleasing repetition of motifs: one can ask for no
better study than Dorothy Washburn and Donald Crowe’s Symmetries of
Cultures [20].
Truly archaic examples are difficult to identify, perhaps because mate-
rials such as fiber, wood or leather are preserved only in the most unusual
circumstances. But the remarkable ochres found in the Blombos caves,
carved some 77,000 years ago, show unmistakable evidence of geometrical
thought. Exploring pattern and symmetry are truly as human an endeavor
as can be, as central to the human experience as language and rhythm.
This essay is appearing in a volume on the ties of mathematics to the
broader culture, and exploring this theme in some depth would be natural,
perhaps expected. But I would like to take a somewhat different approach,
and discuss something of the contemporary and mathematical culture of
tessellations.
Perhaps because of the subject’s deep roots in the human experience,
the study of tessellations links from the arts through recreational mathemat-
ics all the way to one of the deepest threads in contemporary mathematics,
the theory of computation, the question of what can and what cannot be
computed. It is notable that many of the remarkable constructions in this

1
Figure 1: The late cosmatesque floors of San Giovanni in Laterano, com-
pleted in 1425.

essay are not due to professional academics, but to, well, there is no better
description than enthusiasts! Few areas of contemporary mathematics are
so amenable to important contributions from the non-professional— we’ve
included a number of open questions that perhaps you will answer for the
rest of us!
If you are interested in exploring further, you will find no more com-
prehensive source than Branko Grünbaum and Geoffrey Shepherd’s mas-
terpiece, Tilings and Patterns [7]. Doris Schattschneider’s M.C. Escher:
Visions of Symmetry [18] illuminates the tessellation work of the inimitable
artist, who anticipated many themes later explored by academic mathemati-
cians. The Symmetries of Things demonstrates contemporary topological
tools for examining these issues [2]. The many works of Martin Gardner,
and the volume of essays, Mathematical Recreations [9], in his honor, are
very useful to the young tessellator. Finally, most importantly, overshadow-
ing all of these, are the countless works by innumerable artists and artisans
through time and space, from whom this mathematical inquiry draws inspi-
ration.
In this essay, we won’t try to survey every possible interesting corner—
there’s just too much!— but set our to explore an ancient question, some-

2
what formalized here, to be sure: What can fit together, and how? and
follow a particular thread that leads us to one of the deeper issues in con-
temporary mathematics, that some problems are undecidable, can be proven
to be unanswerable by mechanical means. cross ref elsewhere in
La Matamatica

1 What can tessellate?

Let us start with the non-convex, equilateral pentagon, discovered by Livio
Zucci (Figure 2), the unique equilateral pentagon for which the vertex angles
satisfy 2A+E = B +C +2D = B +C +E = 2π, describing some of the ways
copies of the tile can fit together around a vertex (for example, two copies
of the A vertex can fit with an E vertex, as shown). These relationships
imply several others, and copies of the tile can form a remarkable variety
of configurations and patterns. You really must have a set of these tiles
to play with in order to appreciate their rich behavior — I suggest buying
several copies of this volume, saving one, and cutting out sets of tiles from
all the rest.1

D
B C A
A
E
A E

Figure 2: Livio’s pentagon.

It’s quite a subtle problem: just what tessellations can this tile form–
and what can’t it? In Figure 3 are two patterns that can each be extended to
1 Alternatively, materials can be downloaded from the author’s website
http://mathfactor.uark.edu/downloads/tiles.pdf

Figure 3: Two “isohedral” tessellations admitted by Livio’s pentagon.

3
Figure 4: Three “anisohedral” tessellations by Livio’s pentagon. Can the
configuration at bottom right be completed into a tessellation of the entire
Euclidean plane? More importantly, how might one tell?

isohedral tessellations — tessellations with a symmetry that acts transitively

on the tiles themselves. In more familiar language, in such a tessellation,
each tile has the same relationship to the whole as every other tile does.
In Figure 4 we begin to see the rich behavior of Livio’s pentagon. It is
clear that three of the patterns in the figure can be extended to anisohedral
tessellations of the entire plane, the tiles lying in more than one aspect with
relation to the whole, the symmetry not acting transitively on the tiles. If
we continue the pattern shown at upper left across the infinite Euclidean
plane, we will have a periodic tessellation — that is, one formed of regularly
repeating units (or “fundamental domains”). Another way to put this is
that if you were to turn your back for a moment, a friend could slide the
tessellation some distance, yet when you look again you would be unable to
tell that anything had changed. On the other hand, the tessellation is still

4
Figure 5: Each triangle and each quadrilateral admits at least some tessel-
lation.

anisohedral — the tiles lie in a few different aspects relative to the whole
tessellation, indicated by the way they are colored.
Though it exhibits a definite pattern, the tessellation with concentric
rings of tiles, shown at upper right, is non-periodic: Because the rings are
all centered on a particular point, if your friend were to shift the tessellation
in any way, you would certainly notice. Is the figure at lower left a portion
of a periodic or of a non-periodic tessellation? No finite figure can really
tell us — this pattern can be completed either way, but if this tessellation is
to be periodic, the units must be fairly large, larger than the image shown
here.
But what about the configuration — or tessellation of a disk-like region
— at the lower lower right of the figure? Can it be extended to a tessellation
of the entire plane? Most important of all, what criteria could we apply to
work this out?

2 What tiles admit tessellations?

The question before us, the primary question underlying the study of tes-
sellations, is how do local constraints inform global structure? As Livio’s
example richly shows, the question is remarkably subtle.
Given a tile, we can ask the most basic question of all: Does the tile
admit a tessellation of the plane, that is, does there exist a tessellation of
the plane formed of copies of the tile, with no gaps or overlaps? More to
the point: how can we tell? As we shall learn, there are surprising and deep
lacuna in our understanding.

5
 But to begin with, as suggested by Figure 5, it is not difficult to see
that every triangle admits a tiling of the plane, as does every quadrilateral,
convex or non-convex, though it is considerably more subtle to ask about
all tilings admitted by a given triangle or quadrilateral!
The Euclidean Tiling Theorem, which we state and prove at the end of
this essay, provides some powerful constraints on the number of neighbors
a tile can have in a tessellation of the Euclidean plane. In particular, in a
tessellation by copies of a single tile, the Euclidean Tiling Theorem tells us
that it is impossible for every copy to have seven or more neighbors.
Now if the tile is convex, then every copy will have at least as many
neighbors as it has sides, and so we can go further and be sure that if a
convex tile admits a tessellation then it has six or fewer sides.
We’ve already seen that every quadrilateral and every triangle admits a
tiling, so the remaining questions here are: Which convex pentagons admit
tessellations? Which convex hexagons admit tessellations?2
If we add one more constraint, that the tessellations be isohedral, as
those in Figure 3 are, these questions are quite easily answered: It is not
particularly difficult to enumerate all isohedral tessellations, as in [7], or
with more topological methods in [2], checking as we go to see which of
these are by pentagons or hexagons, whether convex or not. Doing so,
we find there are exactly five families of pentagons and three families of
hexagons that admit isohedral tessellations.
The topic is instantly richer when we consider anisohedral tiles, tiles
that admit tessellations, but only anisohedral ones. In 1918, K. Reinhardt
showed there can be no convex anisohedral hexagon, but starting in the late
1960’s, a number of convex anisohedral pentagons began to be discovered, by
Richard Kershner, then later by Richard James III, Marjorie Rice and finally
Rolf Stein. A few of these are illustrated in Figure 6 and an interactive
Mathematica notebook can be obtained at [15]. But to this day, the question
remains quite open: What are the convex anisohedral pentagons?
We can go further and ask for non-convex anisohedral hexagons and
pentagons. Though there has been much investigation and many examples
are known, no fully general picture has emerged.
Indeed, it can be remarkably difficult to determine whether a given
tile does in fact admit a tessellation of the plane. This clearly depends on
2 As with so much of the topic of tessellations, Grünbaum and Shephard’s Tilings and

Patterns [7] remains the definitive account. We refer the reader to their Chapter 9 for
a more detailed account of the history of these questions and for further bibliographic
references, many of which appear in a wonderful collection of articles in honor of Martin
Gardner [9]

6
Figure 6: Four periodic tessellations by anisohedral pentagonal tiles, dis-
covered (clockwise from upper left) by Kersher, Rice, James and Stein, with
isohedral numbers two, three, three and two respectively — that is, for ex-
ample, in any tessellation by the pentagon at upper left the tiles are in a
minimum of two aspects with respect to the whole. Still today, there is very
little understanding of which convex pentagons admit tessellations, much
less how to address the Monotiling Problem on page [[cross reference]].

7
Figure 7: Voderberg’s tile, here with n = 10: ten (!) copies are surrounded
by two others. Careful inspection reveals the tile really is in just one piece.

the ways copies of the tile can fit together, but even within a small region,
this problem is full of surprises. For example, in 1936, H. Voderberg gave
a remarkable construction for any given n, producing a tile for which two
copies can surround n others! His tiles don’t admit tessellations for n > 2,
but there is no reason, at present, to rule out the possibility of tiles that
have this property, and do tile the plane. Analyzing, then, what can and
cannot occur, begins to seem quite subtle. Many other interesting examples
can be found throughout [7]. We pose:

The Monotiling Problem: Does a given tile admit a tessellation of the

plane?

For any of the tiles we have seen so far, this is a relatively easy problem
to settle, even if the tiles have strange properties. With just a little exper-
imentation (or a glance at Figure 4) we can decide that Livio’s Pentagon
does admit a tessellation, or that the Voderberg tile in Figure 7 does not.
But what about the remarkable tiles, discovered by Casey Mann and
Joseph Myers, shown in Figure 8? Which of these does or does not admit
a tessellation?
It is not too difficult to prove, at least, that the tile at left does not
admit a tessellation: it can be viewed as a cluster of hexagons, with a
total of seven edges bulging inwards and four bulging outwards. In any
configuration of N tiles, there are at least 3N unmatched inward-bulging
edges, which must lie on the boundary of the configuration. If the tile
were to admit a tessellation, we could cover disks of arbitrary radius R; the
number of tiles in such a configuration, and hence the number of unmatched

8
Figure 8: The first three tiles are among several discovered by Casey Mann
in 2000 and 2007; the two on right are among several discovered by Joseph
Myers in 2003. Which of these admits a tessellation of the plane?

inward-bulging edges, would be approximately proportional to R2 , but the

number of the edges on the boundary would be approximately proportional
to only the circumference of the disk and hence to R. For small radii, this
may be possible, but asymptotically, at some point, we have too many tiles
and not enough room for excess unmatched edges. The tile cannot admit a
tessellation.
We can similarly prove the third tile does not admit a tessellation, but
this will not work for the second: it seems we must enumerate all possible
configurations, covering disks of larger and larger radii. Eventually, we run
out of configurations, and will have proven that this tile cannot admit a
tessellation.
The Heesch number measures the complexity of such a tile, as the
largest “combinatorial radius” of disks it can cover — that is the number
of concentric rings that copies of the tile can form. In Figure 9 we see the
first tile has Heesch number 2. The current world record, incidentally, is
held by the tile at the center of Figure 8, with Heesch number 5 [11]. But
still today, no one knows the answer to the question: For all n, is there a
tile with Heesch number at least n?
The two tiles at right do admit tessellations, though I suspect that this
is humanly impossible to discover! (I’d very much like to know if it is!) But
at least there is a mechanical means available: If a tile admits a periodic
tessellation, we can find this out using another brute force enumeration of
all possible configurations, this time checking as we go whether or not we
have one that can be used as a fundamental domain. The isohedral number,
defined in Figure 6 above, measures the complexity of this.
Myers has found an amazing variety of strange examples [14]— the two
at right in Figure 8 are just the most notable of a great many exotic tiles.

9
Figure 9: This tile does not admit any tessellation of the plane, and has
Heesch number 3 — a fourth ring can be started, as at lower left of the
configuration, but not completed [12].

Remarkably, the tile at far right has isohedral number 9; the tile just to
the left holds the current world-record, with isohedral number 10. But still
today, this remains an open question: For all n, does there exists a tile with
isohedral number at least n?
We might settle the Monotiling Problem through the brute force enu-
meration of configurations: we can always find out if a tile does not admit
a tessellation, and we can always find out if a tile does admit a periodic
tessellation. But it is unknown whether or not there is a procedure to settle
this in general, whether or not it is possible to discover whether or not an
arbitrary tile does in fact admit a tessellation. Could it be that this problem
is undecidable?
This could only be so if there exist tiles that do admit tessellations of
the plane, but do not admit any periodic tessellations, that is, if there exist
aperiodic tiles. This is far more subtle than merely admitting non-periodic
tessellations: as noted in Figure 11, even the humble 2×1 rectangle manages
that simple task! For a tile to be aperiodic, it somehow would force non-
periodicity, somehow wrecking symmetry at all scales! This is such an
astounding property that one might plausibly suppose no such tile could
exist, and that the Tiling Problem is in fact decidable.

10
Figure 10: The simplest tessellation by this tile, with isohedral number 10,
the current world-record. The shaded region, found through the brute force
enumerations of all possible configurations [14], is a fundamental domain of
the smallest possible size for this tile.

Figure 11: The 2 × 1 rectangle does admit non-periodic tessellations — for

example, one may tessellate the plane with 2×2 squares split horizontally or
vertically at random — but admits periodic ones too, and so is not aperiodic.

11
 The examples we’ve seen should chasten us: at the very least, the
simple-sounding problem “How does a tile behave” is astoundingly subtle,
and despite much attention, this question has not been conclusively settled!
Today, no one can say whether or not the Monotiling Problem is decidable.

3 Aperiodicity and the Tiling Problem

More generally, we can ask the

Tiling Problem or Domino Problem: Does a given set of tiles admit a

tessellation of the plane — does there exist a tessellation formed by copies
of tiles in the set?

This problem arose as an aside in 1961, as the logician Hao Wang

worked on one of the remaining open cases of Hilbert’s Entscheidungsprob-
lem (“Is a given first order logical formula satisfiable?”) [19]. Wang noted
that the closely related “Completion Problem” is undecidable and cannot
be decided in general by any mechanical procedure whatsoever!

The Completion Problem: Does a given a finite collection of tiles admit

a tessellation of the plane containing a given “seed” configuration?

That is, can we “complete” the seed configuration to form a tessellation of

the entire plane using copies of some or all of the given tiles?
The proof is simple, and rests on encoding Alan Turing’s “Halting Prob-
lem” for machines inside of Wang’s Completion Problem for tiles: For any cross reference to Turing
Machine
given Turing machine, Wang produces a set of tiles and seed configuration,
in such a way that the seed configuration could be extended to a tessellation
by copies of the tiles if and only if the machine never halts.
Of course the Halting Problem is a touchstone of undecidability: in
1935, through a simple and elegant construction, Turing proved there can
be no procedure to decide whether a given Turing machine will halt or not
[[cited elsewhere]]; consequently, there can be no procedure to tell whether
one of Wang’s corresponding sets of tiles, with its seed configuration, can
complete a tessellation of the plane.
Wang’s construction is easy enough to illustrate by an example: Con-

12
 A
0 0 0 0 0 0 t=0
B
0 A0 0 0 0 0
0 0 0 0 0 0 t=1
φ(A0)
A 0 0 B0 0 0 0
0 0 1 0 0 0 t=2 φ(B0)
B 0 A0 1 0 0 0
0 0 1 0 0 0 t=3 φ(A0)

C 0 0 B1 0 0 0
0 0 0 0 0 0 t=4 φ(B1)
0 0 0 C0 0 0
B φ(C0)
0 0 0 1 0 0 t=5
0 0 0 1 B0 0

Figure 12: Wang emulates the run of a Turing machine as a tiling problem.

sider the Turing machine specified by

φ A B C
0 0RB 1LA 1RB
1 1RB 0RC 0LH
This machine will work on an infinite tape; at each step, each cell is
marked 0 or 1 and the machine will be in a particular state A, B or C, reading
one particular cell. The transition function φ determines the action of the
machine, depending on its state and the marking it is reading; for example,
if the machine is in state A reading 0, as in the upper left of the table, the
machine will leave a 0 in that spot on the tape, move right one cell, and go
into state B. If it is in state B reading a 0, it leaves 1 on the tape, moves one
cell to the left, and goes into state A. There is one special “halt” state H —
if the machine enters this state, it can do no more, and the process halts.
Beginning in state A, on a tape marked with all 0’s, we can illustrate
the first few steps of the run of the machine, at left in Figure 12.
The essential point is that this illustration itself satisfies completely
local rules: it is composed of pieces that must fit together in a certain
manner. We can encode this as a tessellation, shown at right above.
With only a little care, we then have a set of tiles, shown in Figure 13,
that can emulate the machine. It is possible to cover the plane with copies
of these tiles, so that labels on adjacent edges match,3 as shown above.
3 It is easy enough, if one prefers, to use unmarked tiles that simply are required to

13
 1 0

A0 0 1 0

A0 0 1 0 1 A1
φ(A0) φ(A0) φ(A0) φ(A1) φ(A1) φ(A1)

0 B0 B1 A0 A1 0
0 1 B0 B1 0 1
φ(B0) φ(B0) φ(B0) φ(B1) φ(B1) φ(B1)
A0 A1 1 0 C0 C1
C0 0 1 C1
φ(C0) φ(C0) φ(C0)
1 B0 B1 H

Figure 13: The tiles Wang uses in his Completion Problem, to emulate our
sample machine.

Must they emulate this machine? In any tessellation containing the

initial “seed tile”, at upper left in Figure 13, there must be, inductively
row by row, a faithful representation of the run of the machine; this can be
completed into a tessellation of the entire plane if and only if the machine
never halts — note that the tile at bottom right in Figure 13 corresponds
to the machine entering the halt state, and no tile can fit beneath it. As
the Halting Problem is undecidable, so too is the Completion Problem.
(On the other hand, note that it’s easy enough to tile in other ways,
if we don’t place the seed tile. For example, we could just cover the plane
with copies of the filler tile at upper right in Figure 13.)
This example highlights the deep connection between undecidability
and computational universality: The celebrated Church-Turing thesis in ef-
fect asserts that anything we might mean by computation can be realized
by a Turing machine and thus by anything that can emulate a Turing ma-
chine. Wang’s Completion Problem is undecidable precisely because it has
this property, precisely because completing a tessellation from a seed tile
is “computationally universal” and can emulate any computation (albeit
wildly inefficiently!).
As a further aside, Wang posed the Tiling Problem (which is trivial for
the sets constructed above since we may cover the plane with just copies of
the blank filler tile). Just as we saw with tessellations by copies of a single
fit together: we may convert the labels into geometric jigsaw-like bumps and notches:

14
 Figure 14: The Robinson tiles are aperiodic.

tile, Wang noted that if the Domino Problem were in fact undecidable, there
must exist sets of tiles that do admit tessellations, but none of which are
periodic — if every set of tiles either does not admit a tessellation or admits
a periodic tessellation, then we have an algorithm for answering the Domino
Problem: enumerate configurations, covering larger and larger disks, until
we run out of possibilities (No, the tiles do not admit a tessellation), or until
we discover a fundamental domain (Yes, the tiles admit a tessellation).
Wang reasonably conjectured no such aperiodic set of tiles could exist
— after all, somehow, just by local rules, symmetry would have to broken
at all scales — but within a few years Robert Berger, and then Raphael
Robinson gave subtle proofs that the Domino Problem is undecidable, along
the way producing aperiodic sets of tiles [1, 17].
Berger’s proof that his set is aperiodic rests on their forcing the emer-
gence of a particular kind of non-periodic, hierarchical structure; a few years
later Robinson gave a greatly simplified construction which we momentarily
describe. Both authors used these forced hierarchical structure as a kind of
scaffolding in their demonstrations of the undecidability of the Tiling Prob-
lem, as we will soon sketch in Section 4. The construction has been widely
imitated and generalized, producing many other aperiodic sets of tiles, as
we discuss in Section 3.2.

3.1 Robinson’s aperiodic set of tiles

Robinson’s six tiles are shown in Figure 14. Let us show that:

Theorem The Robinson tiles are an aperiodic set.

Proof We have two tasks: we must show that the Robinson tiles can tile
the plane, but also that they cannot tile periodically. We first notice that
the last tile in Figure 14 is “cornered” and the other five are “cornerless.”
The first and last “cross” tiles have a “bump” pointing out of each edge;
the other four “passing” tiles have one “bump” and three “nicks”. The two

15
Figure 15: Robinson’s proof rests on showing that tiles form blocks that
themselves fit together as the tiles do, on a larger scale, forming still larger
blocks, ad infinitum.

cross tiles have an “elbow”, shown as pointing up and to the right.

In any configuration of these tiles, at every corner, three cornerless tiles
must meet one cornered tile, as shown at top left in Figure 15.
What tiles can be filled in the positions adjacent to the elbow of a
cornered cross tile, as indicated in the top middle of Figure 15? We can
only have two passing tiles, as shown at top right, and thus a cornerless cross
adjacent to those. We quickly see that every cornered cross can only belong
to the 3 × 3 configuration shown at bottom left in Figure 15 — though we
do have a choice for the orientation of the central elbow.
In turn, at the ends of this larger central elbow, only two passing tiles
can fit, as shown at bottom right in Figure 15, and hence a chain of passing
tiles, forcing the placement of a cornerless cross, at the upper right of the
figure above. In just the same way that the original tiles could only form
3 × 3 blocks, these 3 × 3 blocks can only fit together to form a 7 × 7 block,
these 7 × 7 blocks can only fit together 15 × 15 blocks, and these into 31 × 31
blocks, and so on. In Figure 16 we can see a nested sequence of such blocks.
Do the Robinson tiles admit a tessellation? Let us explicitly construct
a tessellation: fix a crossed corner tile. We may place a 3 × 3 block to the
southwest; this in turn may be in the northwest of a 7 × 7 block, which
will be in the northeast of a 15 × 15 block, which will in its turn be in the
southeast of a 31 × 31 block, and so on. Continuing in this way, we have a

16
Figure 16: Constructing one of uncountably many tessellations admitted
by the Robinson tiles, as a countable sequence of ever more momentous
choices.

well-defined tessellation of the plane: every point in the plane will lie within
a specified well-defined tile.
In fact, there are uncountably many distinct tessellations by the Robin-
son tiles! We had an infinite sequence of choices in the construction: at each
stage, the next largest block might have lain in any of four positions rel-
ative to what we had already placed. There are two subtleties: not every
sequence actually will produce a configuration that covers the entire plane
(for example, if we always placed the next block to the southwest). But
still uncountably many do. Different sequences of choices might have given
congruent tessellations, but as any given tessellation would have only been
congruent to countably many others, the Robinson tiles still admit uncount-
ably many non-congruent tessellations of the plane.
Can the Robinson tiles admit a periodic tessellation? As shown in
Figure 17, a cross tile in the center of an (2n − 1) × (2n − 1) block lies at
the vertex of a marked 2n−1 by 2n−1 square; squares of the same size do
not overlap.
Suppose there were a periodic tessellation admitted by the Robinson
tiles. Then translation by some vector v would leave the tessellation invari-
ant. (That is, if you were to turn your back while I shift by v, when you

17
Figure 17: A heirarchy of squares: the Robinson tiles do not admit periodic
tessellations. Note the complementary systems of overpasses and branches.

18
again take a look you would be unable to tell that anything had happened.)
But this is impossible: there is some size square S so large that v would
be unable to take S to another square of the same size, and so shifting by
v could not leave the tessellation invariant.

3.2 Aperiodic hierarchical tessellations

Berger’s great discovery, that hierarchical structure can be forced through
local rules, led to many notable and beautiful examples. The Robinson
tiles we just saw are quite typical in many respects: from some hierarchical
structure we wish to form, in that case, a hierarchy of larger and larger
squares, as in Figure 17, we then find a way of marking tiles that forces
that structure to emerge. Frankly, in all the simple examples, there is a
touch of magic to this!
Once we somehow have actually found such tiles, the proof that they
actually do what we wish, that they actually force such a hierarchical struc-
ture, tends to follow familiar lines: we show they must form larger blocks,
which themselves behave just as the original tiles do, and so must themselves
form blocks that are larger still, which... ad infinitum.
But what hierarchical structures are there in the first place, for us to
try to enforce? Before we describe a few more examples, let us pause for a
moment to discuss substitution tessellations

3.3 Rep-tiles and substitutions

An old puzzle asks how an farmer can divide his L-shaped piece of land
into four plots for his for jealous sons; each plot must be just the same
size and shape as the others. The answer is shown in Figure 18: a large L
can be subdivided into four small L’s. A tile that can be subdivided into
smaller copies of itself is often called a “rep-tile”, which is kind of a pun in
English (it is made of replicas of itself). Rep-tiles readily can be used to
form hierarchical structures; just as in Figure 16, we can assemble our tiles
into larger copies, which we then assemble into larger copies still, and so
on. With care we can assemble a tessellation of the entire plane, one that
automatically comes with a hierarchical structure.
A few more rep-tiles are shown in Figure 19, with portions of a hier-
archical tiling. More generally, we might have two or more different kinds
of tile, subdivided into copies of themselves in some more complex manner.
For example, the two “golden” triangles at left in Figure 20 can each be
subdivided into copies of themselves, as can the “silver” triangles at right.

19
Figure 18: A rep-tile, the “chair”, and a corresponding hierarchical tessel-
lation.

Figure 19: A variety of rep-tiles and corresponding tessellations, clockwise

from top left: the sphinx, Conway’s pinwheel, the dimer and the half hex.

Figure 20: Two triangle substitutions, at left “gold” and at right “silver”,
with corresponding tessellations.

20
Figure 21: A configuration of Penrose rhombs; note the sinuous ribbons, the
ever irregular arrangement of black dots, and the appearance of the rhombs
as cubes in perspective — all with nested five-fold symmetries.

Many examples of these kinds of substitutions on tiles are known —

Ludwig Danzer in particular has found many infinite families with very
strange properties [4]— but the fact is, no one really understands what is
possible, or what is not, or why. There is room for many more discoveries!

3.4 Aperiodic tiles

Now the rep-tiles we’ve just seen are not themselves aperiodic. The un-
marked tiles can form a great variety of tessellations, some of which are
periodic and some of which, like the hierarchical tilings we’ve illustrated,
are non-periodic. The tiles do not force this structure. Which hierarchical
structures can be forced?
Once Berger pointed the way, a number of notable, elegant examples
were found. We’ve just seen Robinson’s, after which Roger Penrose soon fol-
lowed, forcing hierarchical structure based on subdivisions of the pentagon.
His rhombs (Figure 21) and his kites and darts [5] are probably the most
famous of all aperiodic sets of tiles. In the late 1970’s Robert Amman gave
several remarkable examples, finally published in Tilings and Patterns [7]—
still today, his tiles are among the very simplest known. Ludwig Danzer,

21
Joshua Socolar, Charles Radin, myself and a few others have given addi-
tional sets of this general kind. But all together, remarkably few sets were
known, and all seemed like art, or magic.
In 1988, Sharir Mozes gave the first general construction, for systems
based on special subdivisions of the rectangle. For any subdivision of a
particular kind, he was able to produce new tiles that faithfully forced the
corresponding hierarchical structure to emerge.
At this point, may I modestly report that the complete solution—
quite indebted to all these examples— appeared in [6]? Given virtually any
substitution — and there are certainly infinite familes of examples — we
can produce a set of tiles to force it. But still, this construction is so general
that in practice it produces large, unwieldy sets of tiles. There is still room
for novel, simple, magical constructions, and from time to time a new one
is found.
In particular, towards the end of Section 2, we glossed over a particular
famous question: Does there exist an “einstein”, an aperiodic monotile, that
is, a single tile that admits only non-periodic tessellations?
To some extent, we’ve found that this becomes almost a semantic, or
perhaps sociological question— what do we mean, really? Are we allowed to
mark our tiles? Can we allow more complex matching rules than just fitting
the tiles together? What if we allow the tiles to overlap? Who decides, and
how, what the rules should be?
Generalizing very slightly, Penrose gave a reasonable example [16] and
just recently Joan Taylor and Socolar gave another one, and David Fletcher
yet another. But if we require unmarked tiles, simply fitting together, as
shown throughout Section 2, then we do not know the answer, whether or
not there is an “einstein”.
More critically, what other kinds of aperiodic sets of tiles are possible?
All of the examples we have seen force hierarchical structure, but by no
means is that the only possibility. Space limitations do not allow discussion
here, but there is a literature on “quasiperiodic” tiles, sets of tiles which force
a completely different kind of non-periodic structure[3, 10], and in 1996,
Jarkko Kari gave a completely different means of forming aperiodic tiles [8].
But that’s it. These three methods are just a shadow of what is possible.
Precisely because of the undecidability of the Tiling Problem, discussed in
a moment, we know that we can never fully answer this question: What are
other means of constructing aperiodic sets of tiles? Mysteries abound.

22
Figure 22: Berger’s basic idea for the proof the undecidability of the Tiling
Problem: emulate Wang’s proof, on larger and larger domains. (Compare
to Figure 12.)

4 The Undecidability of the Tiling Problem

And this will always be so. As we mentioned in Section 3, Berger proved
that we can never find a general solution to the Tiling Problem, we can
never fully understand how to determine whether or not any given set of
tiles admits a tiling.
Just as Robinson simplified Berger’s initial construction of an aperi-
odic set of tiles, he also streamlined Berger’s proof that the Tiling Problem
is undecidable, and we can at least get the idea here. Recall that Wang’s
Completion problem was proven undecidable by emulating the action of a
Turing Machine, but required the placement of an initial “seed” configura-
tion. Wang himself could not see how to force this placement and still carry
out the emulation. Berger’s idea was instead to begin by constructing a set
of tiles that can only admit a non-periodic, hierarchical structure. Within
this structure, we hang increasingly large domains in which we can emulate
finite, but longer and longer runs of a given Turing machine. All of these
domains can be tiled, fully, if and only if the machine itself does not halt.
In Figure 22, we sketch the basic idea. Once the corners of each domain
are placed, as our “seeds”, the remaining tiles will be forced, emulating the
behavior of our given Turing machine for some limited period of time. We
will be able to complete domains of arbitrary size — and thus a tessellation
of the plane — if and only if the Turing machine does not halt. As this is
undecidable, so too will be the Tiling Problem.

23
 a
a
a
a
a

Figure 23: Our domains will be nested, and as at left larger domains will
have to be attenuated in order to “flow” around smaller ones. We allow this
by introducing new tiles. For example, if an edge marked a in our original
simulation were to be attenuated to allow room for a smaller domain, we
introduce a tile marked as shown at right.

But now, how can we ensure these arbitrarily large domains are them-
selves present? In Figure 24 we illustrate Robinson’s solution: let us ignore
every other level of squares in the hierarchy forced by the Robinson tiles; a
sequence of nested squares remain — these form our domains.
We need one final trick: smaller domains lie within larger ones, getting
in the way of our emulation. All we must do is to flow around them, at-
tenuating our simulation, as sketched in Figure 23. Where two tiles would
once have met, there is now a string of what we might call “edge tiles” —
tiles which emulate the markings on our original edges.
Returning to Figure 24, we see the first three levels of nested domains,
attenuated to flow around those that are smaller, leaving 3 × 3, then 5 × 5,
then 9 × 9 cells for the runs of a machine, using squares of the 2nd, 4th and
6th levels of Robinson’s hierarchical structure. In general, from the (2n)th
level, we obtain a domain of 2n + 1 cells width and height.
With the idea in hand, the rest is fairly technical but relatively straight-
forward. We overlay the Robinson tiles with additional markings to encode
this additional structure, as well as with markings similar to those in Fig-
ure 13: We now can reliably emulate finite runs of the machine. For a given
machine, the corresponding set of tiles admits a tiling if and only if the
machine runs for arbitrarily long time — that is, if and only if the machine
does not ever halt.
We know so little, and there is so much to explore. This topic is rich
with questions that need little formal training — but much imagination!—
to explore, and I hope many readers take up the challenge!
I’d like to conclude with a final image, of a single shape of tile that
can form remarkable tessellations with scaled copies of itself. These tiles,

24
Figure 24: Here is the heart of Robinson’s proof that the Tiling Problem
is undecidable. Onto his aperiodic set of tiles, we paint a variety of new
markings that encode domains and the Turing machines that are emulated
within them. The domains will be nested as shown: 3 × 3 domains within
the 4 × 4 blocks of the second level of the original hierarchy, then 5 × 5
domains within the 16 × 16 blocks of the fourth level, 9 × 9 domains within
64 × 64 blocks, etc. We can complete a tessellation by the tiles we thus
produce if and only if we can emulate arbitrarily long runs of the machine
— that is, if and only if the machine does not halt.

25
when drawn to infinite detail, are all equivalent up to scaling, and admit
myriad tessellations that cover almost the entire plane — the structure of
these missing points in such tilings is a rich and fascinating open area to
explore!

A The Euclidean Tiling Theorem

In Section 2 we promised a statement of the powerful Euclidean Tiling The-
orem; it’s well worth proving, too, since the techniques are so elementary,
yet broad and useful. The key lemma applies, in fact, to tilings of any
surface!
Before we begin, we note there’s no limit on how many sides a polygon
can have and still admit a tiling of the plane, if we allow adjacent pairs of
tiles to meet along as many successive edges as we please:

But there is a strict constraint on the number of neighbors a tile might

have, or the number of tiles that can meet at a vertex, at least on average,
at least if we require all our tiles to be about the same size. For example,
it’s perfectly possible to tessellate with tiles meeting three-to-a-vertex, each
having seven neighbors, but these heptagons will have to be increasingly
distorted in order to fit together. On the other hand, if all of our heptagons
are about the same shape and size, or even are exactly the same shape and
size as in all the examples we’ve seen so far, then this is impossible.
In essence, we will define a way to measure a kind of combinatorial
imbalance— do our tiles have too many neighbors, or too few? The theorem
will state that this imbalance must average out to nearly zero over very large
configurations.
We make this precise: Consider a tessellation by tiles that are “reason-
ably round”: the tiles are topological disks and if a configuration roughly
covers a large disk, then the number of tiles within it is roughly proportional
to the area of the disk and the number of edges on its boundary is roughly
proportional to the circumference of the disk. More precisely, a tessellation
is “reasonably round” if there are constants c1 and c2 so that in any disk of
radius R, there are at least c1 R2 tiles and the boundary of the configuration
of these tiles has no more than c2 R edges.
This can be guaranteed in many ways, and a complete discussion can be

26
Figure 25: What can copies this tile form? What structure is there in the
holes in tessellations by this tile? What other self-similar tiles can you find?

27
found in Chapter 3 of Grünbaum and Shephard’s masterpiece Tilings and
Patterns [7]. For example, we can simply require that all of our tiles can
inscribe, and in turn be inscribed in a pair of particularly sized disks, and
that they all have a bounded number of neighbors. But we can loosen even
these restrictions somewhat and still maintain the key property of being
“reasonably round.”
In a Euclidean tessellation, there can be tiles with a great many neigh-
bors, or a tremendous number of tiles meeting at a given vertex, but these
have to be balanced out by tiles with just a few neighbors, or a small num-
ber of tiles meeting at some other vertex. This balancing condition is very
precise, and fairly easy to state. In order to do so, let us contrive a way to
measure how far from “balanced” a given tile is.
Suppose a given tile meets n others in the tiling (if, somehow, a tile
meets another more than once, we just count with multiplicity), and that
reading around, at the first vertex of the tile, q1 tiles meet, then q2 , etc.,
until at the last vertex, qn tiles meet. Let’s call 2π/q1 , 2π/q2 , . . . “valence
angles” of the polygon; then the “imbalance” of the tile is Σ − (n − 2)π,
where Σ is the sum of all the valence angles. This looks strange, but this
has a simple interpretation. Remember that the sum of the vertex angles
of any n-sided polygon in the Euclidean plane is exactly (n − 2)π. This
imbalance measures how far we are from being able to straighten out all the
valence angles into vertex angles.4
4 We may regard this imbalance as a kind of combinatorial curvature which exactly
measures the total curvature the tiles would have to have if were to modify them so
the valence angles were their actual vertex angles. For example, we might modify the
tessellation by octagons and squares, so each vertex angle really is 120◦ . An octagon
with 120◦ vertex angles must have negative curvature, totalling −2π/3 over its surface.
Similarly, a square with 120◦ vertex angles must have positive curvature — one way to
construct such a square is to take one face of the spherical cube, with total curvature
one sixth the total curvature of the sphere, or 4π ÷ 6 = 2π/3. In essence all that follows
here can be summed up in these terms: The total curvature of tessellation by tiles with
curvature must, in the large, approach zero if we are to be able to flatten it out into a
Euclidean tessellation without too much distortion.
But how can we use K as a stand in for curvature? The celebrated Gauss-Bonet
theorem relates the total curvature of a disk-like region to the net amount of turning
we experience as we traverse its boundary. If the region has zero-curvature, this net
turning will be exactly 2π — try it on any flat disk-like region in the plane! But around
a positively curved region, this turning will be less than 2π and if the total curvature
is negative, the turning will be greater than 2π. The term “holonomy” is shorthand
for this difference: it is just 2π minus the amount we turnR as we traverse the boundary
of a region. It is remarkable that the total curvature, S κ (that is, the total of all
the Gaussian curvature κ across S), of a disk-like region S is exactly the same as the
holonomy around its boundary.
Now what happens as we walk around a polygonal disk with n straight sides, but

28
 For example, in the tessellation by octagons and squares, each vertex
has valence three, and all the valence angles are 2π/3. Each octagon has
an imbalance of 8 × 2π/3 − (8 − 2)π = −2/3π and each square has an
imbalance of 4 × 2π/3 − (4 − 2)π = +2/3π. In the configuration pictured
here the imbalances don’t quite cancel out — there are more squares than
octagons — but in a very large region, there will be about the same number
of octagons as squares, at least relative to the total number of tiles, and the
net imbalance, relative to the area of the configuration, will approach zero.
This next theorem provides strong topological restrictions on Euclidean
tessellations:

The Euclidean Tessellation Theorem: For any Euclidean tessellation

by reasonably round tiles, considering configurations C(r) of tiles meeting
larger and larger disks of radius r, the average imbalance of the tiles in C(r)
must approach zero.

The following lemma contains the essential truth of the matter: let us
define the imbalance of a configuration in the same way we did for individual
tiles. A configuration is itself just a large polygon subdivided into tiles.
Ignoring the tiles themselves, suppose the configuration has n vertices, with
valences q1 , . . . qn ; then define the imbalance to be Σ − (n − 2)π, where
Σ is the sum of the valence angles 2π/q1 , . . . , 2π/qn . We’ll pause for this
deceptively powerful lemma — which holds even for tessellations in other
spaces such as the sphere or hyperbolic plane — and then use it in the proof
of our theorem:

The Tiling Lemma: The imbalance of any configuration is exactly equal

to the sum of the imbalances of the tiles within it. Consequently, if a config-
uration has imbalance K and n vertices on its boundary; then |K −2π| < πn
possibly having some curvature across its interior? If the interior angle at a given vertex
is α, then we turn π − α at that vertex. If Σ is the sum of all these interior angles, the
total turning is πn − Σ and our holonomy, and hence total curvature, is 2π − (πn − Σ) =
Σ − (n − 2)π — exactly the same as K if our “valence angles” really were our actual
interior angles.

29
 One route is to prove it inductively, chopping our configuration into
smaller ones for which we presume the lemma holds. We’ll prove the lemma,
though, using Euler’s celebrated theorem for planar maps, giving us a strin-
gent restriction on the numbers of vertices (v), edges (e) and tiles (f ) in
any bounded configuration:

v−e+f =1

This does require us to continue to make two assumptions we’ve already

been using: that our configuration and tiles are topological disks.
Try it! For example in the configuration of octagons and squares we
saw earlier, there are ninety-one vertices, one hundred twenty-five edges and
thirty-five tiles, and indeed, 91 − 125 + 35 = 1! Euler’s Theorem generalizes
readily and plays a large role in shaping a wide range of mathematical
phenomena, such as vector fields or the topology of surfaces. Here we can
get control of the imbalance of configurations, and thus exactly how many
neighbors, on the whole, can a tile have in any tessellation.
All we really have to do is make a careful count of all the valence angles
in all the tiles, applying Euler’s Theorem at just the right moment. With
that in mind, let’s fix a configuration C, and let v, e and f be the number
of vertices, edges and tiles in C. Let Σ be the sum of the valence angles of
C and n be the number of vertices on the boundary of C.
For a tile t ∈ C, let Kt be the tile’s imbalance, nt be the the number of
edges it has, and Σt be the sum of its valence angles. Then summing this
imbalance over all the tiles in C, we have



Kt = Σt − (nt − 2)π
t∈C t∈C

Now t∈C Σt is the sum of all the valence angles in the entire tiling.
Each of the v − n vertices in the interior of C contributes exactly
 2π to
this sum, and all the rest contribute a total of Σ. Therefore t∈C Σt =
2π(v − n) +Σ.
Next, t∈C nt counts the number of edges of each tile — this counts
each of the e − n edges in the interior of C twice (since each such edge
 two tiles) and each of the n edges on
meets the boundary of C just once.
So t∈C nt = 2(e − n) + n = 2e − n and so t∈C nt π = 2πe − πn.
Lastly t∈C 2π = 2πf , giving us a grand total of


Kt = 2π(v−n)+Σ−(2πe−πn)+2πf = Σ−πn+2π(v+f −e) = Σ−(n−2)π
t∈C

30
The sum of the imbalances of the tiles is indeed the imbalance of the con-
figuration!
Now we restate this, obtaining our bound on the imbalance K: Each
valence angle is more than 0 but less than 2π and consequently the sum Σ
of all the valence angles is more than 0 but less than 2πn. Remembering
that K = Σ − (n − 2)π we have 0 − (n − 2)π < K < 2πn − (n − 2)π or

−πn < K − 2π < πn

In other words, the imbalance of a configuration is very strictly bounded by

the length of its boundary.

It is very nice that this observation holds no matter what space we are
tiling: it is just as true on the sphere or in the hyperbolic plane! What is
different in the Euclidean plane is that the area of a disk grows much faster
than the radius of a disk. Our Euclidean Tessellation Theorem follows
immediately from the lemma:
If we take configurations C(r) of reasonably round tiles meeting disks
of larger and larger radius r, by the definition of reasonably round there
will be constants c1 and c2 so that the number of tiles Nr in C(r) will be
at least c1 r2 and the number of vertices on the boundary of C(r) will be
less than some c2 r. By our lemma, the imbalance Kr of C(r) will satisfy
−(c2 rπ − 2π) < K < c2 rπ + 2π and |K| < c2 rπ − 2π; thus |Kr |/Nr <
(c2 rπ − 2π)/(c1 r2 ), which tends to zero as r increases.

To sum this all up, we have shown that in a Euclidean tessellation
by “reasonably round” tiles, on average, over larger and larger regions,
the average imbalance must be closer and closer to zero. This has many
immediate corollaries. For example:

Corollary: No tile admits a Euclidean tessellation in such a way that

every copy of the tile has seven or more neighbors.

Any tessellation by copies of a single tile is reasonably round automati-

cally — all the tiles are exactly congruent. If every tile has seven neighbors,
then every tile is negatively imbalanced, and so the average cannot approach
zero. The number of sides does matter if a tile is convex: in a tessellation by
convex tiles, every tile will have at least as many neighbors as it has sides.
We thus know, for example:

Corollary: No convex n-gon, n ≥ 7, admits a Euclidean tessellation.

31
References
[1] R. Berger, The undecidability of the Domino Problem, Memoirs Am.
Math. Soc. 66 (1966).
[2] J.H. Conway, H. Burgiel and C. Goodman-Strauss The Symmetries of
Things, A.K. Peters (2008).
[3] N.G. de Bruijn, Algebraic theory of Penrose’s non-periodic tilings, Ned-
erl. Akad. Wentensch Proc. Ser. A 84 (1981), 39-66.
[4] D. Frettlöh and E. Harriss Tilings Encyclopedia http://tilings.math.uni-
bielefeld,de/tilings/index
[5] M. Gardner, Extraordinary nonperiodic tiling that enriches the theory of
tilings, Scientific American 236 (1977), 110-121.
[6] C. Goodman-Strauss, Matching rules and substitution tilings, Annals of
Math. 147 (1998), 181-223.
[7] B. Grünbaum and G.C. Shepherd, Tilings and patterns, W.H. Freeman
and Co. (1987).
[8] J. Kari, A small aperiodic set of Wang tiles, Discrete Mathematics 160
(1996), 259-264.
[9] Mathematical Recreations: A Collection in Honor of Martin Gardner,
D. Klarner (ed.), Dover Publications (1998).
[10] T.T.Q. Le, Local rules for quasiperiodic tilings, Proc. of NATO ASI
Series C 489, “The Mathematics of Long Range Aperiodic order”, (ed.
R.V. Moody), Kluwer (1997), pp. 331-366.
[11] C. Mann Heesch’s tiling problem, Amer Math Monthly 111 (2004),
509-517.
[12] Personal communication.
[13] S. Mozes, Tilings, substitution systems and dynamical systems gener-
ated by them, J. D’Analyse Math. 53 (1989), 139-186.
[14] J. Myers Polyomino, polyhex and polyiamond tiling,
http://www.srcf.ucam.org/∼jsm28/tiling/
[15] E. Pegg, Pentagon Tilings, a Wolfram Demonstrations Project,
http://demonstrations.wolfram.com/PentagonTilings/

32
[16] R. Penrose Remarks on Tiling: details of a (1 +  + 2 )-aperiodic set,
The mathematics long range aperiodic order, NATO Adv. Sci. Inst. Ser.
C. Math. Phys. Sci. 489 (1997), 467-497.

[17] R. Robinson, Undecidability and nonperiodicity of tilings in the plane,

Inv. Math. 12 (1971), 177-209.

[18] D. Schattschneider M.C. Escher: Visions of Symmetry second edition,

Harry N. Abrams.

[19] H. Wang Proving theorems by pattern recognition- II, Bell System Tech-
nical Journal 40 (1961), 1-42.

[20] D.K. Washburn and D.W. Crowe, Symmetries of Culture: Theory and
Practice of Plane Pattern Analysis, University of Washington Press
(1987).

33