1981 Book TheMathematicalGardner PDF

THE
MATHEMATICAL
GARDNER
edited by
David A. Klarner
PRINDLE. WEBER & SCHMIDT WADSWORTH INTERNATIONAL

Boston. Massachusetts 02116 Belmont. California 94002
© Copyright 1981 by Wadsworth International
Softcover reprint of the hardcover 1st edition 1981
10 Davis Drive, Belmont, California
All rights reserved. No part of this book may be reproduced or transmitted in

any form or by any means, electronic or mechanical, including photocopying,
recording, or any information storage and retrieval system, without permis-
sion, in writing, from the publisher.
Wadsworth International is a division of Wadsworth, Inc.
Library of Congress Cataloging in Publication Data
Main entry under title:
The Mathematical Gardner
Bibliography p
1. Mathematical recreations - Addresses, essays,

lectures. 2. Gardner, Martin, 1914
I. Klamer, David A.
QA95.M3676 793.7'4 80-22046
ISBN-13: 978-1-4684-6688-1 e-ISBN-13: 978-1-4684-6686-7

DOl: 10.1007/978-1-4684-6686-7
Cover image "Bees in Clover" by Marjorie Rice. Used by permission ofthe artist.
Cover and text design by Susan G. Graham in collaboration with the staff of Prindle,
Weber & Schmidt. Composed in Baskerville by H. Charlesworth & Co., Ltd. Dust
jacket and color insert printed by Bradford & Bigelow. Text and covers printed and
bound by The Alpine Press.
Dedication and
Introduction
-~-
T he articles in this book are dedicated to Martin Gardner, the world's

greatest expositor and popularizer of mathematics. While our papers are
confined to this single subject, Gardner's interests and accomplishments
have a wide range of subjects. Hence, we have entitled the book the
Mathematical Gardner, and would like to see other volumes such as the
Magical, the Literary, the Philosophical, or the Scientific Gardner accompany it.
Of course, our title is also an appropriate pun, for Martin Gardner's
relationship to the mathematical community is similar to a gardener's
relationship to a beautiful flower garden. The contributors to this volume
comprise only a small part of a large body of mathematicians whose work
has been nurtured by its exposition in "Mathematical Games"; Martin's
column which appears every month in Scientific American. More than just a
mathematical journalist, Martin connects his readers by passing along
problems and information and stimulating creative activity. Thus, he is a
force behind the scenes as well as a public figure.
Two people were particularly helpful in putting this book together.
Ronald Graham and Donald Knuth not only made contributions them-
selves but also helped in the quiet solicitation of manuscripts from the other
authors. The project was carried out secretly, although Dr. I. J. Matrix
seems to have gotten a hold of the manuscript and he is coming out with a
pirated edition. Because of the clandestine nature of the project, the
contributors were selected on the basis of personal contact rather than as a
consequence of a published appeal for papers. Many who might have
wished to contribute have my apologies now. A published call for papers
not only would have spoiled the surprise, but it would have resulted in an
overwhelming avalanche of material.
1Il
DEDICATION AND INTRODUCTION
In the nuts and bolts of making the book there are others to thank.
Foremost among these is Dean Hoffman, Many of his ideas were
incorporated in the papers he edited. He also encouraged me to press
on and carry out the project in spite of the time it was taking from my
research work. I would also like to thank my wife for all of her help and
patience. Kara Lynn graciously posed as Dr. Matrix's beautiful Eurasian
daughter Eva for the centerfold, but this had to be left out at the
publisher's insistence. There is a lot of correspondence involved in
making a book of this sort, and my secretary Elizabeth Newton was
invaluable in this regard. Finally, I would like to thank Susan Graham
and the staff at Prindle, Weber and Schmidt for their sincere interest
and help in making this book. In particular, Theron Shreve deserves
considerable credit.
David A. Klamer
IV
Table of Contents
-~-
GAMES
Anyone for Twopins? 2

Richard K. Guy
Pretzel Solitaire as a Pastime for the Lonely Mathematician 16

N. G. de Bruijn
Some Remarks about a Hex Problem 25

Claude Berge
A Kriegspiel Endgame 28
Jim Boyce
Mental Poker 37
AdiShamar
Ronald L. Rivest
Leonard M. Adleman
Cheap, Middling, or Dear 44

V. Chvatal
A Random Hopscotch Problem, or How to Make Johnny Read More 51

David Berengut
v
TABLE OF CONTENTS
GEOMETRY
Wreaths of Tangent Circles 62

Solomon W. Golomb
Bicycle Tubes Inside Out 75

Herbert Taylor
Flexing Surfaces 79
Robert Connelly
Planting Trees 90
Stephan Burr
Slicing It Thin 100

Howard Eves
How Did Pappus Do It? 112

Leon Bankoff
Two.DIMENSIONAL TILING
Fault-Free Tilings of Rectangles 120

R.L.Graham
Disections Into Equilateral Triangles 127

W.TTutte
In Praise of Amateurs 140

Doris Schattschneider
VI
T ABLE OF CONTENTS
Some Problems on Plane Tiling 167

Branko Griinbaum
G. C. Shephard
Angels and Devils 197

H. S. M. Coxeter
THREE-DIMENSIONAL TILING
Packing Problems and Inequalities 212

D.G. Hoffman
Can Cubes Avoid Meeting Face to Face? 226

Raphael M. Robinson
Packing Handed Pentacubes 234

C.]. Bouwkamp
My Life Among the Polyominoes 243

David A. Klamer
FUN AND PROBLEMS
Disappearances 264
Donald E. Knuth
Noneuclidean Harmony 265

Scott Kim
Vll
T ABLE OF CONTENTS
Magic Cuboctahedrons 270

Charles W. Trigg
Games, Graphs, and Galleries 274

Ross Honsberger
Probing the Rotating Table 285

William T Laaser
Lyle Ramshaw
N UMBERS AND CODING THEORY
Supernatural Numbers 310

Donald E. Knuth
The Graph Theorists Who Count-and What They Count 326

Ronald Read
Error-Correcting Codes and Cryptography 346

N.J. A. Sloane
Vlll
THE
MATHEMATICAL
GARDNER
Games
-~-
Anyone for T~opins?
-~-
Richard K. Guy
UNIVERSITY OF CALGARY
The bowling game of Twopins (pronounced "Tuppins") is played by

two people, with columns of pins lined up as in Figure I. Each column
contains one or two pins. The columns are spaced so that the bowler may
knock out anyone column, or any two neighboring columns. Ifa column of
two is hit, both pins fall. After a shot, the pins are not reset before the
opponent takes his turn. The game ends when the last pin falls and the
person knocking it down is declared the winner. For a proper shot, at least
two pins must fall; you are not allowed to remove a single column when it
contains only one pin, and the game may end with some isolated single pins
still standing. In Figure I, for example, you may remove column d only,
but not columns b, c, e or h, unless you remove an adjacent column at the
same time. After removal of d, the opponent cannot remove columns c and e
because these are not neighboring.
Twopins is considered an impartial game because in any position, the
available options are the same for each of the two players. In contrast, chess
is a partisan game because, in any position, Black has a different set of
available options from White. The theory of impartial games in which the
last player is declared the winner, is not as widely known as it deserves to be.
It was discovered independently by Sprague [21] and Grundy [12] and by
various people since. They found that every position in any impartial game
has a nim-value; that is, the position is equivalent to a nim-heap, or a heap of
ANYONE FOR TWOPINS?
beans in the game ofNim [4,2,15]. There is a simple rule for finding the
nim-value of a position:
Take the mex of the nim-values of the options.
The mex (minimum excluded value) ofa set of nonnegative integers is the
least nonnegative integer not in the set. For example, mex {5, 3,0, 7, I} = 2
and mex 0 = 0, therefore, the nim-value for the Endgame (when there are
no options and the game is finished) is zero.
The importance of the nim-value, or Sprague-Grundy function,
derives from the fact that all (positions in) impartial games form an
additive Abelian group. Indeed, so do all last-player-winning games,
including the partisan ones, but the Sprague-Grundy theory applies only to
the subgroup of impartial games.
The sum (or disjunctive combination) of two or more positions, (not
necessarily in the same game) is played as follows:
The player whose turn it is to move chooses one of the component

games and makes a legal move in that component.
The compound game ends when each component has ended, and the last
player is again the winner. It is easy to see that this kind of addition is
associative and commutative.
The identity of the group is, of course, the Endgame, and the negative
of any position is the same position with the opposing player to move. (In
impartial games each position is its own negative.) Most people have come
across examples of the T weedledum and T weedledee principle, in which a
symmetry strategy, mimicking your opponent's moves, enables you to win
FIGURE 1
Ready for a shot at Twopins.
MATHEMATICAL GARDNER
a game. This additive group is not only mathematically pretty, but is also
important practically, since many games break up into sums of separate
games in the normal course of play. A typical move in Twopins, for
example, breaks a row into two shorter rows, and therefore, the next move
must be made in one of the two new rows.
The main result of the Sprague-Grundy theory for impartial games
with the last player winning is summarized in the theorem:
The nim-value of the sum of two games is the nim-sum of their

nim-values.
To find the nim-sum of two nonnegative integers, add them in binary

without carrying. This is the operation used by Bouton [4] in his original
analysis of Nim (see also [2,15]). Indeed, now that we have the Sprague-
Grundy theory, Nim is seen to be the archetype of all impartial games: a
typical Nim position is the disjunctive sum of games of Nim, each played
with one heap.
dr = Game played with rows of beans Game played with heaps of beans
There is no legal move in which r beans may be taken.

r beans may be taken if they A heap of exactly r beans may be
comprise a whole row. removed completely.
r beans may be taken from either r beans may be taken from a
end of a longer row. (larger) heap, leaving a non-
empty heap.
° °
r beans may be taken in either of the last two circumstances,
(leaving or I rows). (leaving or I heaps).
r consecutive beans may be taken r beans may be taken from a heap
strictly from within a longer of r + 2 or more, leaving the
row, leaving 2 nonempty remainder as two non-empty
rows. heaps.
rconsecutive beans may be taken
°
from a row, if this leaves just
or 2 rows.
A heap of r beans may be taken,
or r beans may be taken from a
heap of ~ r + 2, leaving the
rest as two non empty heaps.
r consecutive beans may be taken r beans may be taken from a
from a longer row, leaving 1 or la~er heap, with the rest left
2 rows. as I or 2 heaps.
r beans may be taken in any of these circumstances,
(leaving 0, 1 or 2 rows). (leaving 0, 1 or 2 heaps).
TABLE1
Meaning of the octal code digit dr.
ANYONE FOR TWOPINS?
The game of Twopins was discovered by Elwyn Berlekamp in the

course of his ingenious analysis [3, Chapter 16] of the well-known paper-
and-pencil game, Dots-and-Boxes, or Dots-and-Squares [10]. It contains,
as special cases, the games of Kayles [8,19,9] and Dawson's Kayles [6,7],
which we'll soon describe and whose analyses are already known. In fact,
Guy and Smith [14] investigated a large class of "take and break" games,
played with rows or heaps of beans. These may be called octal games because
the rules can be described by a code name in the scale of eight:
do -d 1 d 2 d 3 ...
where do = 0 or 4 (split a row or heap into two nonempty rows or heaps

without removing any beans) and, 0::::; d, ::::; 7 for r ~ 1; the meaning of the
digits is given in Table 1. For example, the code name for the game called
Kayles by Dudeney [8] and Rip Van Winkle's Game by Loyd [19,9] is 0·77.
It is the special case ofTwopins where every column contains two pins, so
that the rules can be concisely stated as: take 1 or 2 adjacent columns.
Analysis of octal games was first prompted by a problem proposed by
T. R. Dawson, the fairy chess expert [6,7]. We call it Dawson's Chess. It is
played on a chessboard with 3 ranks and n files (Figure 2). White and Black
pawns occupy the first and third ranks, respectively, and the game is
"losing chess' in that the capture is obligatory and the last player loses.
Those who know how pawns move and capture will soon see that pairs of
pawns become blocked on a file after any pawns in the neighboring files
have been swapped. So Dawson's Chess may be played with a row of
beans, with the option to take any bean, provided that its immediate
neighbors, if any, are removed at the same time. You can check that in
octal code, this is the game 0 ·13 7.)
The game as Dawson originally proposed it is in mise're form; that is,
the last player loses. The analysis of misere games is inordinately more
complicated than the normal form, where the last player wins. Because
misere Nim involves only a small change of strategy near the end, people
have often been deceived into thinking that strategies for other impartial
games can be similarly modified. For the vast majority of them this is not
FIGURE 2
Ready for a game of Dawson's chess.
true. (See Grundy and Smith [13] or Conway [5, chapter 12], who give an
analysis of the first few positions of the misere forms of Dawson's Chess (in
the form 0·4 and ofKayles [po 145]. More extensive analyses will be found
in Chapter 16 of [2].)
It is not difficult to show [14] that the games 0·137, 0·07 and 0·4 are
closely related. We call the form 0·07 Dawson's Kayles. It may be played
with a row of beans; a move is defined as the taking of two adjacent beans.
Thus, it is the special case of Twopins in which each column contains a
single pin. The nim-values for Kayles and Dawson's Kayles, played with a
row of n beans, were found [14] to be periodic, apart from some irregular
values for small values of n, with periods 12 and 34, respectively.
It is unrealistic to ask for a complete analysis of Twopins, since its
positions are too various. How many essentially different positions are there
with n columns? Because there are just two kinds of columns, the simple
answer is 2n positions, but we do not need to investigate all of these because
Berlekamp has already pointed out various equivalences between positions,
which you may easily verify:
1 O*iJk··· = *ijk ... = OOijk ... ,

2 ... ijhO*lmn ... = ... ijh + *lmn ... ,
3 ... ijhOO*lmn ... = ... ijh**lmn ... ,
where 0 represents a column with one pin (on its own it can be removed by
neither player and is the Endgame); * represents a column with two pins
(which may be removed by either player, therefore it is equivalent to a nim-
heap of 1); the game star, [5, p. 72] and letters represent columns of either
sort; and the sum sign on the right of equation 2 is the disjunctive sum we've
already described.
We need to analyze then, only those Twopins positions which have a
star at either end, as in equation 1, and in which O's (columns of 1 pin) do
not occur except in blocks of at least three, as in equations 2 and 3. Binary
sequences of this kind were enumerated by Austin and Guy [1], who had 0
and 1 in place of our * and O. The relevant number, tno of such Twopins
positions is a~3~ 2 in their notation and the difference of 2 in rank is due to
the 2 stars at either end of the row. The quantity tn satisfies the recurrence
tn = 2tn- 1 - tn- 2 + tn- 4 ·
In fact
where in is the Fibonacci number
4 j, = )s{(I+/S)" - C- v'5)"} 2
ANYONE FOR TWOPINS?
However, it is unnecessary to analyze positions which are mere reflections

of those already analyzed, so we next ask for the number of symmetrical
positions, Sn'
The center ofa symmetrical position is of one of the 4 types, A,B,C,D,
shown on the left of Figure 3, if n is odd, where? denotes either 0 or *. If n
is even, replace the central symbol by a pair of equal ones. The central
symbol (if n is odd) may be replaced by
(a) * * *, (b) 0 * 0, or (c) 000
r'
to yield symmetrical positions with two more columns, except that (a) may
* ... A
A ... * * * ... * **•
... * 000 * ... B
B ... * 0 0 0*··· ... * 0 000 0*··· C
C ... * 00 o 00* ... ... * 00 O?O 00* ... D
(""'OOOO?O 000··· D
D ···000 ? 000···
···000***000··· A
FIGURE 3
The four types of center for a symmetrical position.
not be used in cases Band C, and (b) may not be used in A or B. If n is even,
replaces the central pair of symbols by
(a) * * * *, (b) 0 * * 0, or (c) 000 O.
Let An denote the number of symmetrical n-column positions of type A,
etc., so that
An = An- 2 + D n- 2,
Bn = A n - 2,
Cn = Bn - 2 ,
Dn = Cn - 2 + D n- 2,
and insert a coefficient 2 to allow for the ambiguity in D,
sn = An + Bn + Cn + 2Dn
= (An-2 + B n- 2 + Cn - 2 + 2D n- 2) + (An-2 + Cn - 2 + Dn- 2)
= (An-2 + B n- 2 + Cn - 2 + 2D n- 2) +
(An-4 + B n - 4 + Cn - 4 + 2D n - 4 )
Thus, Sn satisfies the recurrence

5
and has value
Sn = i[(n+ 1)/2J
where L J is the floor function (greatest integer not greater than) and i is
the Fibonacci number 4.
Therefore the number, un' of unsymmetrical Twopins positions, not
counting reflections as distinct, is
I I I I . nn
un ="2 (tn - Sn) = 4in -"2 i[(n+ 1)/2J + 2j3 sm 3'
and the total number, not counting reflections, is

1 1 1 1 . nn
Vn = - (tn + Sn) = - in + - J;(n+ 1)/2J + r;; sm - .
2 4 2 2y 3 3
The more general case
was discussed in [1], in which 0 occurs in blocks oflength at least k. Here we

extend the analysis to obtain the corresponding sequences S~k), U~k) and v~k)
for general k. The formulae are generally true for k ~ 1, but for k = I no
restriction is implied (apart from the requirement of * at each end) and it is
easy to see that for n ~ 2 (and k = 1),
tn = 2n - 2 , S
n
= 2[(n-1)/2 J, un = 2n - 3 _ 2[(n-3)/2J
,
vn = 2n - 3 + 2[(n-3)/2 J
From now on, we will omit the superscripts (k).
First, we use the fact [1] that
6
so that
tm - tm - 1 = tm - 1 - tm - 2+ tm - k - 1
= tm- 2 - tm- 3 + tm- k - 1 + tm- k - 2
= tk+l - tk + t m - k - 1 + t m - k - 2 + ... + t2 + tl
and since tl = t2 = ... = tk = tk + 1 = 1, we have
m-k-l
7 tm - tm - 1 = L
i= 1
ti'
ANYONE FOR TWOPINS?
a convenient algorithm for calculating {tm }. We may also sum this formula
to obtain
m-k-l
8 tm = I + L (m - k - i)t i •
i= 1
Formulas 7 and 8 were not given in [1].

We next establish formula 5 in the more general form
9 sn=sn-2+sn-k-l.
CAsE A k = 21 - I odd, n = 2m - I or 2m.
~--- 2m-1 -----+ ~--- 2m -----+

* ... ... * ... ... * * ... * * ... *
~--m----+
*···*00000*···* *···*000000*···*
~ m-l--+ ~k=21-I--+ ~ m-l --+ ~k+ 1=21 +
*···*0000000*···* *···*00000000*···*
~ m-l-I ~- 21+ I ---+ ~ m-l-I --+ ~-21+ 2---+
*000······ ···000* * 000 000 *

FIGURE 4
Symmetrical Twopins positions with O's in blocks of ~ k.
From Figure 4 we see that the number of symmetrical positions is

sn = s2m-l = s2m = tm + t m - I + t m - 1- 1 + ... + tl
sn-2 = s2m-3 = S2m-2 = t m - 1 + t m - 1- 1 + t m - I - 2 + ... + tl
Sn - sn-2 = tm - t m - 1 + t m - I
m-21
= tm - I + L ti•
i= 1
by 7, so that sn - sn-2 = s2(m-I)-1 = s2(m-l) = sn-k-l' as required.

CASE B k = 21 (even), is similar to Case A, but we have to treat
n = 2m and n = 2m - I separately:
S2m = tm + t m - I + t m - 1- 1 + ... + t 1 ,
S2m - 1 = tm + tm - 1- 1 + tm -1- 2 + ... + t 1,
S 2m - 2 = tm - 1 + tm -1- 1 + tm -1- 2 + ... + t l'

S2m-3 = t m - 1 + t m - I - 2 + t m - I - 3 + ... + t 1 ,
S 2m - S 2m - 2 = tm - tm - 1 + tm -I'
S2m-1 - S2m- 3 = tm - t m - 1 + tm - I - 1 ,
and we obtain 9 in either case by the use of 7, as before.

The generating functions for tn and Sn are
T( ~ (k) i z( 1 - z) S() ~ (k) i

z,k) = i~O ti Z = (1 _ Z)2 _ zk+l' z,k = i~O Si Z
z( 1 + z)
= 1- z2 _/+1·
Formulae 7 and 9, together with
1 1
un = 2" (tn - Sn), Vn = 2" (tn + sn)
enable us to calculate the values in Table 2, where dots indicate that the
sequences are constant for earlier positive values of n.
Of these sequences, the only ones to appear in Sloane's Handbook
[20] are
the powers of two , P)
n = 2 - 2,
n
the Fibonacci numbers, S~3), and

sequence #102, s~2).
This last appeared in [11] as an example of a sum, having taken over the
generalized diagonals
3x+ ~ =n- 1
of entries in Pascal's triangle. It is also given in [16,17,18] with factor-
izations and a discussion of divisibility properties. For example,
s~2) is even just if n = 7m - 3, 7m - 2 or 7m.
The highest power of2 which divides #~ _3 is the "ruler function"; the
highest power of 2 in 2m.
3 divides s~2) just if n = 13m - 3, 13m - 2, 13m or 13m + 6.
We have however, wandered away from the game ofTwopins. What is
the best hit to make in Figure I? Berlekamp's equivalence equation I tells us
that column h can be ignored, and equation 2 that we can remove e without
affecting the position. Equivalence equation 3 then enables us to put band c
together and the position is
*** + **.
ANYONE FOR TWOPINS?
n ... 3 4 5 6 8 9 10 II 12 13 14 15 16 17 18 19
t• ... 1 2 4 12 21 37 65 114 200 351 616 1081 1897 3329 5842 10252
5,. ... 1 223 4 5 7 9 12 16 21 28 37 49 65 86 114
k=2
u• ... 0 0 2 4 8 15 28 51 92 165 294 522 924 1632 2878 5069
v• ... 1 235 8 13 22 37 63 108 186 322 559 973 1697 2964 5183
20 21 22 23 24 25 26 27 28 29 30
----------------------------------------------------------------
179913157255405972291706252994265254569221111618192 2839729 4983377
151 200 265 351 465 616 816 1081 1432 1897 2513
8920 15686 27570 48439 85080 149405 262320 460515 808380 1418916 2490432
9071 15886 27835 48790 85545 150021 263136 461596 809812 1420813 2492945
... 4 5 6 8 9 10 11 12 13 14 15 16 17 18 19 20 21
... 1 2 4 II 17 27 44 72 117 189 305 493 798 1292 2091 3383 5473
... 1 2 2 3 3 5 588 13 13 21 21 34 34 55 55 89
k =3 ... 0 0 2 4 6 II 18 32 52 88 142 236 382 629 1018 1664 2692
... 1 2 3 5 II 16 26 40 65 101 163 257 416 663 1073 1719 2781
22 23 24 25 26 27 28 29 30 31 32
8855 14328 23184 37513 60697 98209 158905 257114 416020 673135 1089155
89 144 144 233 233 377 377 610 610 987 987
4383 7092 11520 18640 30232 48916 79264 128252 207705 336074 544084
4472 7236 11664 18873 30465 49293 79641 128862 208315 337061 545071
... 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
... 1 2 4 II 16 23 34 52 81 126 194 296 450 685 1046 1601 2452 3753
... 1 2 2 3 3 4 5 6 8 9 12 14 18 22 27 34 41 52 63
k=4
... 0 0 2 4 6 9 14 22 36 57 90 139 214 329 506 780 1200 1845
... 1 2 3 5 10 14 20 30 45 69 104 157 236 356 540 821 1252 1908
24 25 26 27 28 29 30 31 32 33 34 35
5739 8771 13404 20489 31327 47904 73252 112004 171245 261813 400285 612009
79 97 120 149 183 228 280 348 429 531 657 811
2830 4337 6642 10170 15572 23838 36486 55828 85408 130641 199814 305599
2909 4434 6762 10319 15755 24066 36766 56176 85837 131172 200471 306410
TABLE 2 (continued on next 2 pages)

Values ofr~k), s~k), u~k), v~k) for k = 2,3, "',9.
... 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
... 1 2 4 7 [[ 16 22 30 42 61 91 137 205 303 443 644 936 1365 1999

... 1 2 2 3 3 4 4 6 699 13 13 19 19 28 28 41 41
k = 5 ... 0 0 1 2 4 6 9 12 18 26 41 62 96 142 212 308 454 662 979
... 1 2 3 5 7 10 13 18 24 35 50 75 109 161 231 336 482 703 1020
25 26 27 28 29 30 31 32 33 34 35 36 37
2936 4316 6340 9300 13625 19949 29209 42785 62701 91917 134758 197548 289547
60 60 88 88 129 129 189 189 277 277 406 406 595
1438 2128 3126 4606 6748 9910 14510 21298 31212 45820 67176 98571 134476
1498 2188 3214 4694 6877 10039 14699 21487 31489 46097 67582 98977 145071
... 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
... 1 2 4 7 [[ 16 22 29 38 51 71 102 149 218 316 452 639 897 1257

... 1 2 2 3 4 4 5 6 7 9 10 13 14 18 20 25 29 35
k=6
... 0 0 1 2 4 6 9 12 16 22 31 46 68 102 149 216 307 434 611
... 1 2 3 5 7 10 13 17 22 29 40 56 81 116 167 236 332 463 646
26 27 28 29 30 31 32 33 34 35 36 37 38
1766 2493 3536 5031 7165 10196 14484 20538 29085 41168 58282 82561 [[7036
42 49 60 69 85 98 120 140 169 200 238 285 336
862 1222 1738 2481 3540 5049 7182 10199 14458 20484 29022 41138 58350
904 1271 1798 2550 3625 5147 7302 10339 14627 20684 29260 41423 58686
... 8 9 10 II 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
... 1 2 4 7 11 16 22 29 37 47 61 82 114 162 232 331 467 650 894

... 1 2 2 3 344557710 10 14 14 19 19 26 26
k = 7 ... 0 0 1 2 4 6 9 12 16 20 27 36 52 74 109 156 224 312 434
... 1 2 3 5 10 13 17 21 27 34 46 62 88 123 175 243 338 460
27 28 29 30 31 32 33 34 35 36 37 38 39
1220 1660 2262 3096 4261 5893 8175 11351 15747 21803 30121 41535 57210
36 36 50 50 69 69 95 95 131 131 181 181 250
592 812 1106 1523 2096 2912 4040 5628 7808 10836 14970 20677 28480
628 848 1156 1573 2165 2981 4135 5723 7939 10967 15151 20858 287110
(continued)
ANYONE FOR TWOPINS?
... 9 10 II 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
... 1 2 4 II 16 22 29 37 46 57 72 94 127 176 247 347 484 667

... 1 2 2 3 3 4 4 5 5 6 8 10 II 14 15 19 20 25
k=8
... 0 o 2 4 6 9 12 16 20 25 32 42 58 81 116 164 232 321
... 1 2 3 5 10 13 17 21 26 32 40 52 69 95 131 183 252 346
28 29 30 31 32 33 34 35 36 37 38 39 40
907 1219 1625 2158 2867 3823 5126 6913 9367 12728 17308 23513 31876
27 33 37 44 51 59 70 79 95 106 128 143 172
440 593 794 1057 1408 1882 2528 3417 4636 6311 8590 11685 15852
467 626 831 1101 1459 1941 2598 3496 4731 6417 8718 11828 16024
... 10 II 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28
... I 2 4 II 16 22 29 37 46 56 68 84 107 141 191 263 364 502

k=9
... I 2 2 3 4 4 5 5 6 688 II II 15 15 20 20
... 0 o 2 4 6 9 12 16 20 25 30 38 48 65 88 124 172 241
... I 2 5 7 10 13 17 21 26 31 38 46 59 76 103 139 192 261
29 30 31 32 33 34 35 36 37 38 39 40 41
686 926 1234 1626 2125 2765 3596 4690 6148 81<\B 10754 14326 19132
26 26 34 34 45 45 60 60 80 80 106 106 140
330 450 600 796 1040 1360 1768 2315 3034 4014 5324 7110 9496
356 476 634 830 1085 1405 1828 2375 3114 4094 5430 7216 9636
TABLE 2
(continued)
Even without knowing the nim-values, you can see that the (only) good
moves are to take out column d or column a.
Figure 5 shows a Twopins-wheel which enables us to read off the nim-
value of any Twopins position of eight or fewer columns of pins, provided
that we know the nim-values for a row of n pins in Kayles or Dawson's
Kayles (for Dawson's Chess, slide the nim-values one place to the left);
n 0 2 3 4 5 6 7 8 9 10 11 12
Kayles 0 2 3 1 4 3 2 1 4 2 6 4
Dawson's Kayles 0 0* 1 1 2 0 3 1 1 0 3 3 2
Suppose for example, you want the nim-value of
***000**.
FIGURE 5
A Twopins-wheel for finding the nim-values of small Twopins
positions.
Find this arrangement in the outer ring (running from 12 o'clock to 3
o'clock), spiral in from the first and last stars, and meet in a cell containing
the value 4. Thus is the nim-value.
What is the best move in the Dawson's Chess game in Figure 2? Our
advice is to allow your opponent the privilege of the first move. It is a P-
position (previous-player-winning) and has nim-value O.
* Note that a single pin must remain standing in Dawson's Kayles.

ANYONE FOR TWOPINS?
References
1 Austin, Richard, and Guy, Richard. 1978. Binary sequences without
isolated ones. Fibonacci Quart. 16.
2 Ball, W. W. Rouse, and Coxeter, H. S. M. 1974. Mathematical Recreations
and Essays. 12th ed. Toronto: Univ. of Toronto Press. pp. 36-39.
3 Berlekamp, E. R.; Conway,]. H.; and Guy, R. K. 1980. Winning Ways.
New York: Harcourt, Brace Jovanovich.
4 Bouton, Charles L. 1901-2. Nim, a game with a complete mathemat-
ical theory. Ann. Math. Princeton (2) 3: 35-39.
5 Conway,]. H. 1976. On Numbers and Games. New York: Academic Press.
6 Dawson, T. R. 1934. Problem 1603. Fairy Chess Review. p. 94.
7 _ _ . 1935. Caissa's Wild Roses. Fairy Chess Review. p. 13.
8 Dudeney, H. E. 1958. Canterbury Puzzles. N.Y.: Dover. pp. 118-119,
220.
9 Gardner, Martin. 1960. More Mathematical Puzzles of Sam Loyd. N.Y.:
Dover. pp. 5, 122.
10 _ _ . 1974. Mathematical Games: Cram, cross cram and quad-
raphage: new games having elusive winning strategies, Sci Amer. 230 2:
106.
11 Green, Thomas M. 1968. Recurrent sequences and Pascal's triangle.
Math. Mag. 41: 13-21.
12 Grundy, P. M. 1964. Mathematics and games. Eureka. 27: 9-11.
13 Grundy, P. M., and Smith, C. A. B. 1956. Disjunctive games with the
last player losing. Proc. Cambridge Philos. Soc. 52: 527-533; M.R. 18: 546.
14 Guy, Richard K., and Smith, Cedric A. B. 1956. The G-values for
various games. Proc. Cambridge Philos. Soc. 52: 514-526; M.R. 18: 546.
15 Hardy, G. H., and Wright, E. M. 1960. An Introduction to the Theory of
Numbers. 4th ed. Oxford: Oxford Univ. Press. pp. 117-120.
16 ]arden, Dov. 1966. Recurring Sequences. Riveon Lematematika 2nd ed.
86-91.
17 _ _ . 1946-47. Third order recurring sequences. Riveon Lematematika
1: 74; 1952-53 6: 41-42.
18 _ _ and Katz, A. 1947-48. Table of binary linear third order
recurring sequences. Riveon Lematematika 2: pp. 54-55.
19 Loyd, Sam. 1914. Cyclopedia of Tricks and Puzzles. New York: Dover.
p.232.
20 Sloane, N. J. A. 1973. A Handbook of Integer Sequences. New York:
Academic Press.
21 Sprague, R. P. 1935-36. Ober mathematische Kampfspiele. Ttfhoku
Math. J. 41: 438-444; Zbl. 13: 290.
Pretzel Solitaire as a
PastilDe for the Lonely
MathelDatician
-~-
N. G. de Bruijn
EINDHOVEN UNIVERSITY OF TECHNOLOGY
In general, playing solitaire card games is a rather sad form of amusing

oneself. Starting from a random position of cards on the table, arranged in
some predefined format, we have to try to find a sequence of legitimate
moves that turn the position into some predescribed goal. The moves are
usually irreversible, and therefore dead positions can result where no further
move is possible. If we are fortunate enough to reach the goal, we get some
feeling of satisfaction. If we run into a dead end, however, we usually feel
sad. The reason for distress is not that luck was against us, but that we
wonder whether the failure had been unavoidable. We are sad because we
feel as though we may have spoiled the game. More often than not, we
have forgotten the original position, and now it is impossible to retrace our
steps. Failing to solve a solitaire problem is particularly humiliating ifit is a
game with complete information; that is, all of the cards are facing up.
1. It is not in a mathematician's nature to be proud of having been lucky

enough to solve a problem more or less by accident. We want to have some
method or other by which we can find a solution if there is any. And if there
is no solution, we want to be able to prove that there is no solution.
~16&.
PRETZEL SOLITAIRE
What the mathematician aims for is to be able to study a given

position of the cards, and to produce without touching a card, either a
solution or a proof that no solution exists. It is clear that this can only be
done with solitaire games that are not too difficult. The game we suggest
for this purpose of study has two integer parameters, which the player may
fix according to his taste and ability, in order to get something that is
neither too difficult nor too simple. We shall call it the k x n pretzel. The
game is old, certainly for the case k = 4, n = 13 (see section 5 for a
description of how the 4 x 13 is usually played), and it is difficult to trace
its origin. The name is new, and was chosen because a pretzel is a kind of
biscuit in the shape of a knot.
2. For the k x n pretzel, we take a deck of cards consisting of k suits of n

cards each. To make things easier, let's say k = 4, and we'll call the suits
S,H,D,C (for spades, hearts, diamonds and clubs). The cards in each suit
will be numbered from 1 to n (the 1 is usually called "ace"), and our
notation will be SI, ... ,Sn, ... ,Cl, ... ,Cn.
After shuffling the deck, we put the cards face up on the table in a
rectangle with 4 rows and n columns. Next we take away the aces and put
them in a new column in front of the rectangle, in the order S 1, HI, D 1, C 1.
We get a k x (n + 1) rectangle with 4 empty spaces which we call holes;
each hole is the size of a card. We still call them holes even when 4 of them
are in the last column. Let us imagine a k x (n + 1) frame around the
cards in order to distinguish between the holes and the infinite outside
world. In the course of the game we are never allowed to put cards outside
the frame. As an example of a position in the 4 x 4 pretzel, we display
1
+1 +4 .3 .2
,,"I +2 +3 ,,"3 .4
Now we start playing. A move consists of putting a card in a hole

provided that the next lower card (in the same suit) is lying directly to the
left ofthe hole. So in the position 1 anyone of the cards S2,H3,H2,D2 may
be used for a first move. With the first move we have filled one hole and
obtained a new one. This new hole allows us to make our second move.
We want to reach this final position;
~17&.
2
+1 +2 +3 +4
41 42 43 44
We shall call this position the goal.

A position is called solvable if a sequence of moves exists that leads to
the goal (the goal itself is also called solvable). If there is no such sequence,
we will term it unsolvable. A position is called dead ifit is different from the
goal and does not allow a single move. Dead positions are unsolvable, of
course.
The rules of the game were explained here for the case k = 4, n = 4,
but for all other cases we have exactly the same rules.
3. The position 1 is solvable; for example, by the following sequence of

moves (each move is indicated by the name of the moving card)
3 S2,S4,C4,C3,D4,D2,D3,C2,H4,H2,S3,S4,D4,C3,C4.
Let us now first change the notation. The columns in 1 do not play
any role in our game, so we may just as well print the position in a single
line, with separation marks:
4. The 4 x 2 and the 4 x 3 pretzel are very dull games, but the 4 x 4
pretzel can be quite interesting. The 4 x 5, 4 x 6 are usually still playable
(in the sense explained in section 2).
For larger values of n, the author does not have enough experience to
say whether an average player will be able to settle the majority of cases
without touching the cards, and without using pencil and paper and/or
computer.
5. Let p(k,n) be the probability that a random k x n pretzel position is

solvable. It seems to be difficult to determine p(k,n) exactly, unless n :::;; 2. If
n = 2, the only unsolvable positions are those with a ryclic blockade. In a
cyclic blockade we have a cyclic arrangement of a subset of the suits, for
instance, H -+ D -+ S -+ H, such that (Dl H2), (Sl D2) and (HI S2) are
pairs of neighbors. If k = 4 and n = 2, there are 40 such cases with a 3-cycle,
168 with one 2-cycle, and 6 with 2-cycles-2l4 unsolvable positions
altogether. The total number of positions is 1680.
It can be estimated that p( 4,4) is about 0·45. The author has played
several hundred 4 x 4 pretzels at one time or another. In all cases he was
PRETZEL SOLITAIRE
able to decide whether or not the position was solvable (although he often
deviated from the no-touching rule by allowing "safe moves" -see section
14 below). Moreover, a computer search dealt with 2473 random
positions, of which 1123 were solved and the remaining 1350 were
established to be unsolvable.
The author has less experience with n = 5 and n =6. A very rough
guess would be thatp(4,5) is of the order ofO·l.
The 4 x 13 pretzel still seems to have a reasonable number of solvable
cases. Usually it is not played with the expectation of reaching the goal but
rather to get as far as possible. The player tries to get a long sequence
Sl,S2, ... , Sp in the first row, and similarly Hl, ... ,Hq, Dl, ... ,Dr,
C 1, ... , Cs. When the board is dead, he takes away all the cards that do not
belong to these sequences, shuffles the deck, and fills the rectangle again,
leaving a hole directly to the right ofSp,Hq,Dr,Cs. Now the game starts
over. After this second round, a third round may follow, and the player is
considered to be successful ifhe reaches the goal position in the third round.
Yet it happens now and then that the goal position is reached in the
first round. Let us say that it happens in approximately one percent of all
cases. This only gives a lower estimate for p( 4,13), since nobody knows how
many solvable positions have been spoiled. From this kind of experience
one may guess that p( 4,n) is not exponentially small if n tends to infinity.
6. In section 5 we mentioned that the 4 x 4 pretzel game was run on a

computer. The program was based on studying the graph of the game. The
points of that graph are all possible positions, and there is an oriented edge
from P to Qifwe can get from position P to position Qby a single move. If
P is a position, then by S(P) we denote the set of all points that can be
reached by a finite sequence of moves if we start at P. The question is
whether the goal position Po belongs to S(P).
The set S(P) is, for any given P, easily built up in a computer's
memory by trying to find new points obtainable from old points in a single
move. The execution is stopped if no further points of S(P) are found
(whence P is unsolvable) and also if Po is reached. In that case P is solvable,
and then we need not know the full S(P).
If P is unsolvable, then the number of elements in S(P) is rarely more
than 60; however, the largest number ever found was 380. With solvable P,
the number of points inspected before Po was reached, was of the order of
150, and the largest number ever found was 802. The size of S(P) is not
very interesting, however. If it is very big or very small, it is usually quite
easy to calculate P by mere thought.
In the pretzel there can be many different sequences of moves leading
from one position to another, because of the possibility of interchanging
moves or sequences of moves that do not interfere with each other. This
means that it would be inappropriate to run a computer program based on
the backtracking technique applied to sequences of moves.
7. With the 4 x 4 pretzel and arbitrary solvable starting position P, it is

usually not too difficult to find a solution mentally although one may
occasionally hit on tricky cases. But if the starting position is unsolvable,
how do we show it? Determining the full S(A) (of Section 6) is usually
hopeless for a non-computerized brain. But there are many other ways to
prove unsolvability.
One possibility is cyclic blockade. In its simplest form, it was
mentioned in Section 5. Often it is more complicated, like the following
example. We fix our attention on a certain card cl ' Then we show that, if it
ever moves, a card C2 must have moved before C1 moves for the first time.
Then we note that a card c3 cannot move before c2 moves for the first time.
So in any solution there is a moment where c2 moves for the first time, and
C1 and c3 have not moved yet. Next we find card C4 that has to be moved
before c3 moves for the first time, etc. In the course of this argument we
have a growing set of cards that have not moved yet at the moment we
have in mind, and that strengthens further argumentation. If sooner or
later we run into a loop, observing that C1 must have moved before all the
others, we have proved that the original position was unsolvable.
The mathematician who plays the game will no doubt discover a bag
of additional tricks which augment his proving ability.
Such tricks are not always easy to put into words. Often they are a
mixture of forward and backward analysis. Backward analysis is what we
just explained when discussing cyclic blockade. Forward analysis involves
seeing what happens to all possible first moves in the starting position: if
from P we can get in a single move to either P 1 ,P2 ,P3 or P4 , and if
P 1 ,P2 ,P3,P4 are all unsolvable, then P is unsolvable. We have to be careful
not to mix forward and backward arguments on a large scale, since one
can easily get to an impossibility proof based on wishful thinking, which is
worse than having no proof at all. Usualiy such a proof can be safeguarded
if we replace it by something based on the truncation theorem of section 8.
Failure to establish a solution may help to construct a proof of
unsolvability, and failure to construct such a proof may help to find a
solution. In short; it is in the spirit of the working mathematician!
8. IfPis an integer, 1 ~ P ~ n, then Es (P) is the set of spades ~ p (that is,

{Sl,S2, ... ,Sp})' Similarly we define EH, ED, Ec and we put
F(p,q,r,s) = Es(P) UEH(q) uED(r) uEc(s).
+-
Such an F(p,q,r,s) will be called a trunk; if k 4 the definition is similar.
If P is a position in the k x n pretzel and if F is a trunk, then we shall
call "P truncated by F" the position we get by taking away from P every card
not belonging to F. This is no longer a position in the sense of Section 2
(unless F = F(n,n,n,n)) yet we can play the game on truncated positions,
keeping the definition ofa move the same as in Section 2. We stipulate that
whatever Fis, the truncation of the k x n pretzel still has to be played in the
PRETZEL SOLITAIRE
original (k + I) x n frame (therefore the k x n pretzel is not the same game

as the k x (n + 1) pretzel truncated by F(n,n,n,n)). We take Po truncated
by F as the goal in the truncated pretzel, where Po is the goal in the original
pretzel. Also, a truncated position is called solvable if there is a sequence of
moves leading to the new goal.
Let P 1 and P2 be positions (or truncated positions) such that a single
move leads from P1 to P2. Let F be a trunk, and let Pt be Pi truncated by F.
If the moving card belongs to F then the same move can be used to
transform Pt into P!. Ifit does not belong to Fwe have Pt = P!. Applying
this to a sequence of moves, with a fixed F, we get a theorem that can serve
as a necessary criteria for solvability:
THEOREM If P is solvable, then every truncation of P IS

solvable.
9. The simplest applications of the truncation theorem of Section 8 are

those where some truncation of a position is dead. For example,
is unsolvable since truncation by F( 1,3,3, 1) leads to

Sl--D2H2*H ID3---*D IH3---*Cl----*
which does not admit a single move.
In more complicated cases we show that the truncated position is
unsolvable but not dead. Or we show that every move possible from our
starting position leads to a position for which we can select an unsolvable
truncation, etc.
10. A move is called safe unless it turns a solvable posItIon into an

unsolvable one. In an unsolvable position all moves are safe, but in many
solvable positions we can make unsafe moves.
Quite often we can prove a move to be safe. Consider a position P
where a card c lies at a place p. Moving c from p to a place q is certainly safe
if we have both 1 and 2:
1 The place directly to the right ofp is either outside the frame,
that place is occupied by the next higher card in the same
suit, or there is no card higher in that suit.
2 By no sequence of moves leading from P to the goal Po, can
place q ever be occupied by a card other than c.
The purpose of condition 1 is to guarantee that there is no further use

for card c at place q; condition 2 guarantees that c cannot do any harm at
place q.
As to condition 2 we note that it is often easy to say that certain cards

can never get to certain places. For example, if we have a row SlH3C2-
D4S3D2D3 then we can say, by inspecting the places on the left, that the
only possible candidates for the place now occupied by D2 are
S6,H7,C5,D6 and S4.
11. Another source of safe moves is the following. If in a sequence of

moves the only moving cards are diamonds, and if at the end of the
sequence all diamonds are in their proper places (that is, the places they
have in the goal position), then the sequence consists of safe moves. This
statement remains true if both occurrences of the word "diamond" are
replaced by "diamonds and spades," or by "diamonds, spades and clubs."
These statements are easily deduced from the truncation theorem.
12. If in a position P there is only one possible move, then that move is
obviously safe.
Now assume that we are in a position P with just two possible moves,
with cards C1 and C2' and assume that C2 is neither the next higher nor the
next lower to C1 in the same suit. (This guarantees that the moves do not
interfere: after moving C1 the old move for C2 is still possible, and vice
versa). Moreover we assume that Cl transforms P into a position in which C2
is the only available move. Then we can say that C2 is a safe move for P.
This simple idea for proving safety can be generalized to more than
two noninterfering moves and to noninterfering sequences of moves. The
lonely mathematician will no doubt discover many arguments of this kind.
We finally mention that in any position, a move can be proved to be
safe by showing that all other moves are bad (that is, lead to unsolvable
posi tions ) .
13. In Section 1 we said that the game should be analyzed without

touching a card. This may sometimes be frustrating, and therefore it is not
unreasonable to modify the rules by allowing the player to carry out moves
as long as he can prove that they are safe.
A further mnemonic aid the player might allow himself is connected
with the kind of blockade analysis mentioned in Section 7. Analyzing what
happens when a certain card moves for the first time, the player might put
a penny on top of each card that he knows he's been unable to move until
that moment.
14. Thus far our pretzel game satisfied these conditions:
1 Every suit has the same number of cards.

2 In the frame (see (1)), all rows have the same length.
3 The number of holes equals the number of suits.
PRETZEL SOLITAIRE
We can ignore the above three statements and still have a game for which
almost everything holds true that was discussed in the previous sections. As
an example we present
with solution
D2,S5,S4,H3,D4,S2,D3,S3,D4,S4,S5.
15. We present a number of exercises, taken from a collection of random
positions. The exercises marked t are quite difficult: the author's bag of
tricks does not seem to contain nice arguments. Cases 1-9 are 4 x 4's, 10-
16 are 4 x 5's. The solutions are listed below.
1 SlD2-D3S3*HI-C2H2D4*Dl-S4C3H4*ClS2H3
-C4*
2 Sl-H2H3S2*HIH4S4-D2*DI-C3S3C2*CID4D3
C4-*
3 SlC3-C4H4*HI-D4H2C2*DID3S4S2H3*CI-D2
-S3*
4 SlS2D4C3D3*HIC4-H3H4*DlS4C2H2-*Cl-D2
-S3*
5 SlC4C2D3-*HIH2H3H4-*DlS4S3C3D4*ClD2
-S2-*
6 SlC4H4-H2*HID3D4C3S4*DlS3-H3-*ClS2
-C2D2*
7t SlD3C2-H2*HlS3D4D2-*DlH4C3H3C4*Cl
-S4S2-*
8 SlC2C3S2-*HI--D2H4*DID3H3S4H2*CIC4S3
-D4*
9 SlD4S4S2H4*HIC4D3H2C2*DI-C3-H3*ClS3
-D2-*
10 SlD4C4D3-C5*HI-D5-H5H2*DIC2S3H4H3S4*
C 1S2S5-C3D2*
11 SlS5D2H2-D4*HIH4S4S3-D5*DIH5C5C2D3
-*ClS2-C3H3C4*
12 SlC4D5D3S5C2*HI-D4-C3S2*DIC5H5H3-H4*
CID2H2S4-S3*
13 Sl-C4H4C5D4*HID3C2S2H2H5*DID5-D2
- H3*C 1C3S3S5-S4*
14t SlS5S2D5H3D4*HI-C2D3S3H4*Dl
-H2C4H5C3*CIC5-S4D2-*
15 SlC4-D2C5S4*H1S5D5H5D4D3*D1H2S2H4-C3*
C1S3H3-C2-*
16 SlC4H2D4-D2*H1C3C5D3S5H5*Dl-H4S4-C2*
Cl-D5H3S3S2*
Solutions
1 Solvable: D3,D4,H3,S3,S4,D2,D3,H2,S2,S3,C2,C3,D4,
S4,C4,H3,H4.
2 Unsolvable. Truncation by F(l, 1,3,3), then dead in two
moves.
3 Unsolvable. Truncate by F(4,1,3,4). Then D3 is a safe
move. Next D2,C2,C3,S2,S3 is a sequence of safe moves
leading to a dead position.
4 Unsolvable. Truncate to F(4,1,4,1), then dead in one
move.
5 Solvable: D4,C4,D3,C3,S4,S3,D2,D3,C2,C3,S2,S3,S4,
C4,D4.
6 Unsolvable: H3 and D4 are blocking each other.
7 Unsolvable.
8 Unsolvable. Truncate by F(2, 1, 1,4).
9 Solvable: C4,H2,D4,S2,S4,S3,S4,D2,C2,C3,C4,D3,D4,
H3,H4.
10 Unsolvable. Move H2 is safe, the following 14 moves are
unique and lead to a dead board.
11 Solvable: D4,H3,H4,H2,C4,C5,D3,C3,S3,S4,H3,H4,
H5,D4,C4,D2,D3,S5,S2,S3,S4,C2,D4,S5,C3,C4,C5,
D5.
12 Unsolvable. After truncation by F(4, 1, 1,4) a dead board
results in 2 moves.
13 Unsolvable. Move S2 is safe because move D3 leads to an
unsolvable position (truncation by F(l, 1,5, 1)). For the
same reason C3,C2,D4 are safe. The following 13 moves
are unique (apart from interchanging the 10th and 11th)
and lead to a dead board.
14 Unsolvable.
15 Solvable: H4,C3,S3,S4,S5,C5,C2,H2,D3,D5,H3,C3,
H5,H4,D4,H5,C4,C5,S2,S3,D2,D3,D4,S4,S5,D5.
16 Solvable: D2,C2,S5,D4,H3,H4,H5,D3,D5,C3,C4,H2,
S2,S3,C5,H3,H4,S4,S5,D4,D5,H5.
Some Remarks about
a Hex Problem
-~-
Claude Berge
UNIVERSITY OF PARIS
h e readers of Gardner's rubric (see Scientific American: July 1957;

October 1957;July 1975; December 1975) know the depth and the beauty
of the game of Hex, launched almost simultaneously by Piet Hein in
Denmark and by John Nash in America. This game is especially appealing
to mathematicians. Two players play alternately by putting a peg in an
empty hole of the lozenge board shown in Figure 1. The first player uses
black pegs, and wins if he manages to construct a chain of black pegs
connecting East to West; his opponent tries to construct a chain of white
pegs connecting North to South. The best size for a board seems to be
14 x 14, and because a proof by contradiction shows that the first player
has a winning strategy, his opponent can require him to put the first peg in
a restricted area of the board. Of course, all kinds of handicaps can be used
if the two players are not of equal strength, and the aesthetic interest of the
game is nearly unchanged.
I will pose a Hex problem that I wish to present in dedication to
Martin Gardner. Desirable attributes of a "problem" may be, for example:
the first move is unique; or the first move is paradoxical, or the underlying
idea is unexpected, etc. (We recall Sam Loyd, the greatest chess problemist
of all time, who claimed that his main goal was to compose a problem
where the first move is the opposite of what 999 players out of 1000 would
propose.)
NORTH
BCDEFGH
WEST EAST
ABCDEFGHI JKLMN
SOUTH
FIGURE 1
Empty lozenge board.
Here, to avoid unnecessary complications and to permit a few

intermediate moves, we shall not adopt the criterion of uniqueness of the
key move, and we shall not require that the diagram have the same
number of black pegs and white pegs, because a feasible position may come
from a play with handicaps. The problem to be posed is not meant to be
difficult by the complexity or the number of variations, because White can
easily complete a connecting chain between North and South playing
either J9 or M8: since Black cannot play simultaneously on these two holes,
it appears that he is about to lose the game. The problem is for Black to play
and win!
A simple analysis of the position shows that the black pegs can be split
into two separate groups which are connected to the West side, but not to
the East side; in fact, these two groups are completely surrounded by two
disjoint walls of white pegs. The first is:
All, BlO, Bll-CIO, Cll, Dll-DIO, EIO, Ell-FlO, Fll, G ll-G 10,
HIO, Hll-IlO, Ill,J11-JlO, KIO, Kll-LIO, LIl, LI2-M11, M12,
Nl2, N13, N14.
and the second is:
A8, B8, C8, C9-D8, D9, E8-E9, F8, F9-G8, G9, H9-H8, 18, 19-J8,
J9, K9-K8, L8, L7, L6-M6, M5, N4, N3, N2, Nl.
SOME REMARKS ABOUT A HEX PROBLEM
N(white)
r;~B~;C~D~~E~F~G~~H~~I~J~K~.L~M~N~
'" 14
E(black)
S(white)
FIGURE 2
Black to play and win.
As you can see, none of the critical holes (like e9, D8, etc .... ) belong both
to the first wall and to the second one. Therefore no combination can be
expected to break through "at least one" of the white walls. So who would
guess that only the black peg L9 can initiate a new path to connect West
and East? After the intermediate black moves M6, MIl and LIO, the black
lines of pegs will be: K9, J 10, II 0, I9, H9, G I 0, FlO, F9, E9, etc ....
It would be nice to solve some Hex problem by using nontrivial
theorems about combinatorial properties of sets (the sets considered are
groups of critical holes). I t is not possible to forget that a famous chess
problem of Sam Loyd (the "comet"), involving parity, is easy to solve for a
mathematician aware of the Konig theorem about bipartite graphs; also in
chess, the theory of conjugate squares of Marcel Duchamp and Alberstadt is
a beautiful application of the algebraic theory of graph isomorphism (the
two graphs are defined by the moves of the kings).
The use of a mathematical tool may be unexpected and therefore
adds some new interest to a game; but Hex exists as a most enjoyable game
in its own right, for mathematician and layman alike.
A Kriegspiel Endgame
-~-
Jim Boyce
STANFORD UNIVERSITY
Kriegspiel is one of the most interesting variants of chess: each player

tries to mate his opponent, using ordinary chessmen and following the
ordinary rules, but neither player knows where the other player's pieces
are. Instead, both players have a concealed board on which they can keep
track of their own position and guess at the locations of the opponent's
pieces. There is also a third participant in the game, namely the kriegspiel
referee; he has a third set of chessmen, with which he keeps the actual
position. When it is White's move, White suggests a possible move to the
referee. Ifit is legal in the actual game position, it becomes White's official
move; otherwise White must try additional moves until one is legal. Then it
is Black's turn, and the game continues in this fashion. Any legal move that
places the opposing king in check is announced to both players. There are
other rules (which do not concern us) that involve captures and pawn
moves; further details can be found in [1].
This article analyzes the ending king and rook vs. king in kriegspiel.
In normal chess, this is well known to be an elementary mate [2], but the
problem is by no means simple under the kriegspiel ground rules. In fact,
experienced chess players have been known to spend hours on this problem
without solving it. Therefore the reader is encouraged to try his own hand
at the task before looking at the solution below.
A KRIEGSPIEL ENDGAME
FIGURE 1 FIGURE 2
Standard algebraic notation. A sample kriegspiel position.
Movement and Notation

In this article, White has a king and a rook; Black has only his king. Moves
are written in a version of algebraic notation, in which the files are lettered
a-h starting at White's left and the ranks are numbered 1-8 starting at
White's end of the board (see Figure 1). All figures show the board as
White sees it: the known locations of his pieces are marked, and each
possible location for the Black king is marked.
The rules and conventions will be clear after we consider a simple
example or two. Suppose that White tries to move his king to d7 in the
position of Figure 2. If the referee says that such a move is illegal, White
sees the position in Figure 3. Suppose White now tries to play his rook to e2
and the referee announces check. That means that the Black king must be
on e8. Black makes a legal move from e8 and it is White's move in the
position of Figure 4. White moves his king to'c6, a move he knew was legal
before he tried it. After Black moves, his king can no longer be on d8; if it
was there, it had to move away. He can however, still have a king on f8. A
FIGURE 3 FIGURE 4
Position after K -d 7 is illegal. Position after R-e2 gives check.
FIGURE 5 FIGURE 6
Position after K-c6. Position after R-e8 gives mate.
king on f8 would also have had to move, but a king on f7 could move to f8.
The position after Black's next move appears in Figure 5. If the Black king
is on c8, White can mate by moving his rook to e8. Figure 6 shows that
result. Of course, if White makes that move, a king on f7 or f8 could
capture the rook and assure a draw.
The sequence of moves from Figure 2 to Figure 6 would appear as
follows in algebraic notation: 1. K--d7, R-e2+, 2. K-c6, 3. R-e8#.
Note, first of all, that there is no number before R-e2+ . That is because K-
d7 was illegal and White was still looking for his first move. The symbol +
denotes a check and the symbol # denotes checkmate. A draw (stalemate
or capture of the rook) is denoted by the symbol =, which would appear in
the same place as a + or a #.
King and Rook vs. King

We will see that White can almost always force a win in this endgame, even
under kriegspiel conditions. Black can draw only in certain starting
positions where White cannot unite his king and rook before Black captures
the rook. Even if Black might be able to capture the rook (given what
White knows about the position), in general, he will not. For if White does
put his rook next to the Black king, Black must guess where to move to
make the capture. However, this article considers only those positions
where White can force mate against any (even the most clai"rvoyant)
defense.
White's plan consists of several stages. First, he must make sure that
his rook is safe from capture. Next he plays to a position where all of the
possible squares for the Black king are in a rectangle where one corner of
that rectangle is a corner of the board and the opposite corner is at the
rook. White wants his king in that rectangle also to keep the Black king
away from his rook. He then forces the Black king back until it can occupy
only those squares on a single edge. It is then a fairly simple task to mate
the Black king.
~30&
This article demonstrates a simple (albeit slow) way to force

checkmate by considering these types of positions:
I Both kings are in the same quadrant of the board as seen from the
rook.
2 The Black king is restricted to one or two quadrants of the board.
3 The White rook is (can be) safe from capture.
Then it examines the positions of Type I in more detail:
la Black king confined to one rank (or file).
Ib Black king confined to two ranks.
Ie Black king confined to three ranks.
Id Black king confined to four ranks.
Ie Black king confined to more than four ranks.
I Both kings in the same quadrant. In Figure 7, both kings are above
and to left of the rook. White wants to confine the Black king to smaller and
smaller rectangles of the board. With this in mind, White tries to move his
king to e4 and his rook to f3. (He would be even happier to move his rook
to the fourth rank.) The game might continue. 1. K-e4, R-f3 (or ifK-e4 is
legal, continue as in A, below), 2. K-e4, K-e3, K-f5 (or B), 3. K-e4,
R-f4. This particular sequence of moves reduces the smaller dimension of
the quadrant and results in a position like that after White's first move. If
K-e4 was legal on move 2 or 3, then White reduced the larger dimension
and obtained a position similar to Figure 7. There are other possible
responses (by the referee) to some of White's attempts. If 1. K-e4 is legal,
we have line A: 2. R-f3 or 2. R-f3 +, 3. R-f4. The second case results
in a position where the smaller dimension of the relevant quadrants is
decreased, but the Black king can be in either of two quadrants. The other
FIGURE 7 FIGURE 8
Both kings in one quadrant (as White to move.
seen by rook); White to move.
~31&
possibility is that 2. . .. K-e3 is legal. Line B continues 3. K-e4, R-f4.

The first move is familiar. The second reduces the smaller dimension, but
results in a position in which the Black king is restricted to one quadrant,
while the White king is not in that quadrant. A similar sequence of moves
can be played from any analogous position (except if the smaller dimension
is one). Subsequently it will be shown that the White king can end up in
the same quadrant with the Black king in all cases without increasing the
smaller dimension. Therefore, White will be able to restrict the king to the
edge of the board.
2 Black king in one or two (adjacent) quadrants. Figure 8 shows a
position where the Black king might be in either of two quadrants. White
wants to reach a position similar to Figure 7; for example, White king on
b4, rook on a3, Black king somewhere in ranks 4-8. The first step is to move
the rook (and the king if necessary) so that one of the quadrants that might
contain the Black king has a side oflength two. From Figure 8, play could
proceed 1. K-b2, 2. R--c3. Then White plays his king into that quad-
rant, 3. K-b3, 4. K-b4. If the black king interferes, White simply
retreats his rook. Then both kings will be in the same quadrant, one with a
dimension of two. Finally, White moves his rook to the edge, 5. R-a3.
White can ignore any checks that occur during that sequence of moves.
White can use the same plan if the Black king is restricted to one quadrant.
Sometimes, White doesn't need to move his king and rook over to the
second and third lines from the edge. In Figure 9 White plays his king
around to c6 and moves his rook to c5 to get a position that arises after the
first move in Section I. 1. R--d4 would also result in a position with both
kings in the same quadrant, but the smaller dimension of the quadrant
would be larger. The game in Figure 9 could continue 1. K--d4, 2. K-
c5, 3. K--c6, 4. R-b5.
3 The rook is (or can be) safe. If the king and rook are united in the
center of the board and the Black king is not restricted as in the previous
FIGURE 9 FIGURE 10
White to move. Problem 1. How many moves
save the rook?
FIGURE 11 FIGURE 12
White to move. White to move.
sections, then White wants to limit the Black king in some way. One
position which resembles that in Section I has the rook on a corner square;
for example, hI, and his king nearby, say, g2. So White moves his pieces
toward the corner. If the Black king interferes, then White knows that its
position is restricted and proceeds as in Section I or 2. Otherwise, White's
pieces get to the corner and White can follow the line in Section I.
If the rook is subject to capture, White's first task is to protect the
rook. This is usually straightforward. Sometimes, there are several ways to
defend the rook. Figure 10 is a position with an undefended rook: can the
reader discover all of the ways to save it? (The answer appears at the close
of this article.)
la Black king confined to one rank. The previous analysis does not apply
to positions with both kings in the same quadrant if the quadrant is a single
line on the edge of the board. Figure 11 shows such a position . White can
mate quickly by forcing the king back into the corner and mating him
with the rook. But White must be a little careful to avoid stalemate. The
game could end 1. K-d8, R-g7, 2. K-d8, 3. K-c7, (not 3. K-c8=),
4. R-g6, 5. R-a6#. When the Black king prevents a king move, the rook
makes a move and the Black king must then retreat.
Ib Black king confined to two ranks. White can improve on the line
given in I if the black king is already restricted to two lines. I t is
unnecessary to advance the rook along with the king when chasing
the Black king into the corner. From Figure 12, play could continue
1. K-e7, 2. K-d7, K-d6, 3. K-d7, 4. K-c7, K-c6, K-d8,
5. K-c7, 6. K-b6, K-c8, 7. R-a6#. If 5. K-c7 is illegal White can
play 5 .... R-h7. Two moves later, 7. R-h7 = would be a seriol)s mistake,
but one easily avoided. (The line in section I shows that White can shrink
the area available to the Black king. One point not mentioned then is that
this can stalemate Black. It is easy to determine that, as in the last example,
White will have a different move which checkmates. That, of course, is
what White should play.)
FIGURE 13 FIGURE 14
White to move. Problem 2. How quickly can
White force mate?
Ie Black king confined to three ranks. There are a few ways to improve
the line in Sections I and 2 when the Black king is limited to three ranks or
files. Figure 13 shows a position after White has played R-fS +. The Black
king is in one of two quadrants. White can avoid the bother of going over to
the edge of the board with his pieces. He plays l. R-f6, 2. K-e7. IfK-e7
is legal, the black king is in the quadrant on the right; ifit is illegal, the king
is on the left.
Figure 14 shows another way White can save time in some positions.
On 1. R--dS+ Black may escape to the other side of the board and last
until White's twelfth move. How can White do better? (See the answer
below.)
Id Black king confined to four ranks. Figure 15 is similar to Figure 13.
White wants to play R-f4. If he plays it now and it is check, Black's king
will be in one of two quadrants, and White is too far from the edge for the
idea in Figure 13 to work. White can still avoid the time and effort
involved in the line of Section 2 by playing 1. K-e6. Ifit is illegal, he plays
FIGURE 15 FIGURE 16
White to move. White to move.
FIGURE 17 FIGURE 18
Problem 3. White to move . Problem 4. White to move.
... R-g5 and has reduced the smaller dimension. If it is legal he plays
2. R-f4. If that results in a check, he is in the line that follows Figure 13.
Ie Black king is confined to more than four ranks. When the Black king
is limited to a large section of the board, White wants to shrink the region
as quickly as he can. Figure 16 shows the Black king limited to the largest
area possible. If the region is large enough, White can try to shrink the
smaller dimension (instead of the larger one, as in Section I) of the quad-
rant the Black king occupies. From Figure 16, play could continue I. R-gI,
2. K-e3, 3. R-fl. If either rook move is a check, White next moves his
rook two squares to the left to produce a position in which both kings are in
the same quadrant and the smaller side is at most four squares. Figure 17 is
more difficult. Again, the goal is to limit the Black king to at most, four
lines, without playing king and rook to the edge of the board as in 2.
Figure 18 arises from Figure 10 if 1. K-f2 is illegal. How does White
play to restrict the Black king to a region of no more than four lines? (Note
that a 3 x 4 rectangle contains all possible squares for the Black king, but
the edges are not defined by the edges of the board and the rook.)
This article has shown that White, with a king and a rook, can
checkmate Black, who has only his king, in a game of kriegspiel. Sections
la and Ib show that White can checkmate ifhe forces the Black king to the
edge of the board. Section 3 suggests that White can easily reach a position
with all possible positions of the Black king in a rectangle bounded by a
rank and file controlled by his rook and two edges of the board. Sections I
and 2 show that White can proceed from such a position to one where the
rectangle is smaller, where we say that rectangle A is smaller than rectangle
B if the smaller dimension of A is less than the smaller dimension of B or if
the smaller dimensions are equal and the larger dimension of A is less than
that ofB. The arguments suffice to show that White can mate from a wide
variety of positions, and on a rectangular board of arbitrarily large size.
By carefully studying the strategy detailed here, it can be shown that
White can force mate from the starting position in Figure 10 in at most 39
moves. Therefore there is no need to worry about games that are drawn
because of the "50 moves without a capture" rule.
Acknowledgem.ents
This article is the result of one of the assignments in the class CS204 taught
at Stanford University in the fall of 1978. I would like to thank the
professor of the course, Donald E. Knuth, who suggested the problem of
the kriegspiel ending after he had learned of it from some friends in
Germany; and the T.A. of the course, Chris Van Wyk, who discussed this
problem at length and helped prepare the article; and the rest of the CS204
students who made several valuable suggestions.
Solutions
I Three moves save the rook, R-fl, K-fl, and K-f2. After each of the last
two, Black cannot take the rook because then White's move would have
been illegal. If the move is illegal, White can simply move the rook away
as he pleases.
2 White mates with I. K-c7, R-e6, (IfK-c7 is legal, White wins with a
similar but shorter line.) 2. K-d7, K-d6, 3. K-e7 and mates in nine
more moves as in lb.
3 White limits Black to a "small" rectangle with I. R-e4, R-c4. If the first
move is a check, White replies 2. R-e2. If the second move is a check,
White continues 3. R-e4.
4 White first moves the rook to safety, 1. R-h5, 2. R-a5. White then
moves his king near the rook 3. K-d2, 4. K-c3. If the second king
move is illegal, he moves his rook to safety again R-g5, and keeps trying
to move his king near to it with
5 K-e3, K-c3, 6. R-a5. If these two king moves are illegal, then the
Black king must be on d4. White plays a tempo move, 6. R-h5, to make
the king move and then tries again. This time, he must succeed. If3. K-
d2 is illegal, White plays R-a4, 4. R-h4 to produce a position similar
to what arises if it is legal.
Bibliography
I Compayne, Charles. 1976. Kriegspiel. Games and Puzzles 50: 12-15.
2 Fine, Reuben. 1941. Basic Chess Endings. David McKay.
Mental Poker
-~-
Adi ShalJlir, Ronald L. Rivest

and Leonard M. AdlelJlan
MASSACHUSETIS INSTITUTE OF TECHNOLOGY
ABSTRACT
Can two potentially dishonest players playa fair game of

poker without using any cards-for example, over the phone?
This paper provides the following answers:
1 No. (Rigorous mathematical proof supplied.)
2 Yes. (Correct and complete protocol given.)
Once there were two "mental chess" experts

who had become tired of their pastime. "Let's
play 'Mental Poker,' for variety" suggested
one. "Sure" said the other. "Just let me deal!"
Our anecdote suggests the following question (proposed by Robert

W. Floyd): "Is it possible to playa fair game of 'Mental Poker'?" We will
give a complete (but paradoxical) answer to this question. First we will
prove that the problem is intrinsically insoluble, and then describe a fair
method of playing "Mental Poker".
What does it lIlean to

play Mental Poker?
The game of Mental Poker is played just like ordinary poker (see Hoyle [2])
except that there are no cards; all communications between the players
must be accomplished using messages. Perhaps it will make the ground
rules clearer if we imagine two players, Bob and Alice, who want to play
poker over the telephone. Since it is impossible to send playing cards over a
phone line, the entire game (including the deal) must be realized using
only spoken (or digitally transmitted) messages between the two players.
We assume that neither player is beyond cheating. "Having an ace up
one's sleeve" might be easy if the aces don't really exist! A fair method of
playing Mental Poker should preclude any sort of cheating.
A fair game must begin with a "fair deal". To accomplish this, the
players exchange a sequence of messages according to some agreed-upon
procedure. (The procedure may require each player to use dice or other
randomizing devices to compute his hand or the messages he transmits.)
Each player must know which cards are in his hand, but must have no
information about which cards are in the other player's hand. The dealing
method should ensure that the hands are disjoint, and that all possible
hands are equally likely for each player.
During the game the players may want to draw new cards from the
"remaining deck", or to reveal certain cards in their hand to the opposing
player. They must be able to do so without compromising the security of
the cards remaining in their hand.
At the end of the game, each player must be able to check that the
game was played fairly and that the other player has not cheated. If one
player claimed that he was dealt four aces, the other player must now be
able to confirm this.
FIGURE I
Bob encrypts the cards and sends them to Alice in scrambled order.
MENTAL POKER
rTiIll~
~~
Alice chooses one for Bob, and encrypts another for herself, and sends
them both to Bob .
Bob decrypts both cards, and returns Alice's encrypted card to her.
The above set of requirements makes a "fair game" of Mental Poker

look rather difficult to achieve. To make things easier, we'll assume that
both players own computers. This enables the use of complicated protocols
(say, involving encryption). We do not assume that either player will trust
the other's computer. (The players could program their computers to
cheat!)
We suggest that you might find it an interesting challenge to attempt
to find on your own a method for playing Mental Poker, before reading
further.
FIGURE 2
Alice decrypts her card, and then they compare cards to see who has
won.
~39~
Swnmary of Results
We will present two solutions to the problem of playing Mental Poker:
1 A rigorous proof that it is theoretically impossible to deal the

cards in a way which simultaneously ensures that the two
hands are disjoint and that neither player has any knowledge
of the other player's hand (other than that the opponent's
hand is disjoint from his).
2 An elegant protocol for dealing the cards that permits one to
playa fair game of Mental Poker as desired.
The blatant contradiction between our two results is not due to any
tricks or faults in either result. In fact, we will leave to the reader the
enjoyable task of puzzling out the differences in the underlying assumptions
that account for our seemingly contradictory results.
The Impossibility Proof

For the sake of simplicity, we consider the minimal non-trivial case of
dealing two different cards (one to each player) from a deck of three cards
{X, Y, Z}. The impossibility prooffor this case can be easily generalized to
any combination of cards and hand sizes.
If a legal protocol for this case exists, then after exchanging a finite
number of messages, Alice and Bob each know their card but not their
opponent's card. These messages must coordinate the two players' choices
of cards to prevent them from getting the same card.
Suppose that for a particular deal
the messages exchanged are M i , ... , Mnl
the card Alice actually gets is X, and
the card Bob actually gets is Y.
We define SA to be the set of cards that Alice could have gotten in any deals
where exactly the same messages are exchanged. (Since each player may
want to make some random choices in order to get a card which is
unpredictable to the other player, different deals could arise with the same
sequence of messages having been exchanged.) Obviously, the card X is in
SA·
If SA were to contain just the card X, then the deal would violate our
requirement that Bob should have no information about Alice's card.
Clearly the sequence of messages uniquely determines Alice's card in this
case, so in an information-theoretic sense he has (total) information about
her card. Furthermore, in any physically-realizable (and terminating)
protocol for the deal, Alice has only a finite number of random com-
~40&
MENTAL POKER
putations possible, so that Bob can actually determine Alice's card by

examining those which are consistent with the given message sequence.
On the other hand if SA contains all three cards, then Bob cannot get
any card-regardless of which card he gets, the message sequence is
consistent with the possibility that Alice's card is the same. Consequently,
SA must contain exactly two cards.
The set SB of cards Bob can get without altering his external behavior
is similarly defined, and it must also contain exactly two cards. However,
the total number of cards in the deck is three, so that SA and SB cannot be
disjoint. (In our example, Z belongs to both sets.) Thus it could happen
that both Bob and Alice get the card Z in the case that the message
sequence is M l , ... , Mn' Thus the protocol cannot guarantee that Bob and
Alice will choose distinct cards. We conclude that a fair deal is impossible.
A Protocol for the Deal

The following solution meets all the requirements for the problem. First of
all, Bob and Alice agree on a pair of encryption and decryption functions E
and D which have the following properties:
1 EK(X) is the encrypted version of a message X under key K,

2 DK(EK(X) ) = X for all messages X and keys K,
3 EK(EJ(X) ) = EJ(EK(X) ) for all messages X and keys J and K,
4 Given X and EK(X) it is computationally impossible for a
cryptanalyst to derive K, for all X and K,
5 Given any messages X and r, it is computationally impossible to
find keys J and K such that EJ(X) = EK(Y)'
Property 3, the commutativity of encryption, is somewhat unusual but

not impossible to achieve. Properties 4 and 5, (especially 4), essentially state
that E is cryptographically strong or unbreakable.
As an example of a function with the above properties, consider
EK(M) == MK (mod n)
where n is a large number (prime or composite with a given factorization)
which is known to both Bob and Alice, and where
gcd[K, cjJ(n)] = 1.
[cjJ(n)] is Euler's totient function, which can be easily computed from the
prime factorization of n.)
The corresponding decoding function is
DK(C) == C L (mod n),
where
L' K == l[mod cf>(n)].
Since
EK[EJ(M)] == EJ(EK(M)] == M JK (mod n),
E satisfies property 3. (For more details on the cryptographic strength and
importance of this function see [1,3,4].) We describe this particular
encryption function here only to demonstrate that the kind of encryption
functions we desire apparently exist; we will not make use of any particular
properties this function has other than 1 ... 5.
Once Bob and Alice have agreed on the functions E and D (in our
example this means agreeing on p), they choose secret encryption keys B
and A respectively. These keys remain secret until the end of the game,
when they are revealed to verify that no cheating has occurred.
Bob now takes the fifty-two messages:
"TWO OF CLUBS",
"THREE OF CLUBS",
"ACE OF SPADES"
and encrypts each one (whose bit string is considered as a number) using
his key B. (That is, he computes EB ("TWO OF CLUBS"), etc.) He then
shuffles (randomly rearranges) the encrypted deck and transmits it all to
Alice.
Alice selects five cards (messages) at random and sends them back to
Bob; these messages Bob decodes to find out what his hand is. Alice has no
way of knowing anything about Bob's hand since the encryption key B is
known only to Bob.
Now Alice selects five other messages, encrypts them with her key A,
and sends them to Bob. Each of these five messages is now doubly
encrypted as EA(EB(M)), or equivalently EB(EA(M)), for each M. Bob
decrypts these messages obtaining EA(M) for these five messages and sends
them back to Alice. Alice can decrypt them using her key A to obtain her
hand. Since Bob does not know A, he has no knowledge of Alice's hand.
Michael Rabin suggested a nice physical analogy for the above
process. We can view encryption as equivalent to placing a padlock on a
box containing the card. Bob initially locks all the cards in individual
undistinguishable boxes with padlocks all of which have key B. Alice selects
five boxes to return to him for his hand, and then sends him back five more
boxes to which she has also added her own padlock with key A to the clasp
ring. Bob removes his padlock from all ten boxes and returns to Alice those
still locked with her padlock, for her hand. Notice the implicit use of
commutativity in the order in which the padlocks are locked and unlocked.
MENTAL POKER
Should either player desire additional cards during the game, the
above procedure can be repeated for each card.
At the end of the game both players reveal their secret keys. Now
either player can check that the other was "actually dealt" the cards he
claimed to have during play. By property 5 neither player can cheat by
revealing a key other than the one actually used (one which would give
him a better hand).
The above procedure can also be easily extended to handle more than
two players. (Details left to the reader.) Another obvious generalization is
to use commutative encryption functions in secret communications systems
to send arbitrary messages (rather than just card names) over a com-
munications channel which is being eavesdropped.
Conclusions
Initially we proved that the card-dealing problem is insoluble, and then we
presented a working solution to the problem. We leave it to you, the
reader, the puzzle of reconciling these results. (Hint: each player would in
fact be able to determine the other player's hand from the available
information, if it were not for the enormous computational difficulty of
doing so by "breaking" the code.)
Acknowledgements
We would like to thank Robert W. Floyd, Michael Rabin, and Albert
Meyer for their motivation and valuable suggestions.
References
I Diffie, Whitfield and Hellman, Martin E. 1976. New Directions in
Cryptography. IEEE Trans. lrifo. Theory IT-22: 644--654.
2 Morehead, A. H., Frey, R. L. and Mott-Smith, G. 1947. The New
Complete Hoyle. New York: Garden City Books.
3 Pohlig, Stephen C. and Hellm<l:.n, Martin E. 1978. An Improved
Algorithm for Computing Logarithms over GF(p) and its Cryptographic
Significance. IEEE Trans. Info. Theory IT-24: 106-110.
4 Rivest, Ronald L., Shamir, Adi and Adleman, Leonard M. 1978. A
Method for Obtaining Digital Signatures and Public-Key Crypto-
systems. CACM 21: 120-126.
This research was supported by NSF grants MCS78-05849 and MCS78-04343;

and by ONR grant NOOO 14-76-C-0366.
Cheap, Middling
or Dear
-~-
Vasek Chvatal
MCGILL UNIVERSITY
O n page 128 of a book entitled The Games and Diversions of Argyleshire,

written by R. C. Mac Lagan and published in 1901 by David Nutt in
London, one finds the following description of an old Scottish game.
CHEAP, MIDDLING, OR DEAR
This also is played by two. The letters C, M, D, representing

respectively the words from which the game is named, are written on a
slate, with some interval between them. Under C the figures 1,2,3 are
placed, under M 4, 5, 6, and under D 7, 8, 9, thus:
C M D
1, 2, 3 4,5,6 7,8,9
Player A, who is to play first, marks one of the figures from any of the
groups, concealing it from player B, whom he challenges to guess to
which group it belongs, saying "My father bought a horse at a fair". B
asks, "Cheap, middling, or dear?" A answers him, naming the group
from which he has selected his figure. Thus if his figure were 5, the
answer would be "middling". B then guesses one of the three numbers,
and if he hits upon 5, that is a gain to him of 5, but if he says 4 or 6,
then the 5 is scored to A. In any case the 5 is blotted out. B then leads,
each playing in turn, till all the figures have been expunged. The total
marks credited to each are then ascertained, and he who has the
highest number is the winner.
Weare going to describe the results of an analysis of Cheap, Middling or

Dear. For the benefit of readers not acquainted with game theory, we shall
first sketch a few basic facts concerning matrix games.
Let us begin with a game of matching coins, somewhat related to
Cheap, Middling or Dear but much simpler. Here, each of the two players
hides a nickel, a dime or a quarter. If the two coins match, then they go to
the first player; otherwise they go to the second player. Clearly, neither
player has a safe winning strategy: given bad enough luck, one may lose in
every single play of the game. And yet we claim that the second player is
better off at least in a statistical sense: if he plays correctly then he can
expect steady winnings over long periods of time. Naturally, he must make
every choice in some random fashion, so that the opponent will be unable
to predict its outcome. At the same time, however, the probabilities of the
outcomes must be fixed: in a long run, the second player ought to playa
nickel 7/34 of the time, a dime 12/34 of the time and a quarter the
remaining 15/34 of the time. Assuming that he does follow these in-
structions, we now consider a long sequence of plays. Each nickel chosen by
the first player is met by the second player's nickel 7/34 of the time, by a
dime 12/34 of the time and by a quarter 15/34 of the time: no matter how
the first player behaves, he cannot obliterate the element of chance
introduced into the game by the second player's random choices. Within
this group of plays, the second player loses a nickel 7/34 of the time and
wins a nickel 22/34 of the time. Similarly, in those plays in which the first
player chooses a dime, the second player loses a dime 12/34 of the time and
wins a dime 22/34 of the time. In those plays in which the first player
chooses a quarter, the first player loses a quarter 15/34 of the time and wins
a quarter 19/34 of the time. Within each of these three groups, the second
player's average winnings come to 100/34 cents per play. Of course, these
winnings are not guaranteed: they are only expected in the same sense as an
unbiased coin is expected, but not guaranteed, to show heads fifty percent
of the time. In this sense, the first player can protect himself from heavier
losses than 100/34 cents per game by playing a nickel 10/17 of the time, a
dime 5/17 of the time and a quarter the remaining 2/17 of the time. No
matter how the second player behaves now, every time he plays a nickel,
dime, or quarter, his greatest expected wins are respectively
10 5 2 50 10 5 2 50
- -'5 + -'10 + -'25 = - -'5 --'10+ -'25 = -
17 17 17 17' 17 17 17 17'
10 5 2 50
-'5+ -'10 --'25 = - .
17 17 17 17
This game may be represented by the matrix
[ -25
-1~ ~~ _~~l.
-25 2;J
The first player chooses one of the three rows (corresponding to nickel,
dime and quarter, respectively) and the second player, unaware of the first
player's choice, chooses one of the three columns; the corresponding entry
in the matrix specifies the amount won by the first player. Generally, each
matrix specifies a game in which the first player chooses a row, the second
player chooses a column, and the corresponding entry a ij indicates the
amount won by the first player. (Cheap, Middling or Dear may be viewed
as this kind of a game, with each row of the matrix corresponding to a set of
unambiguous instructions for player A to be followed throughout the nine
rounds, and with each column corresponding to a similar set of instructions
for B. The number of rows and columns in this matrix is enormous.) Again,
the only sensible strategy for each player is to make his choices at random,
so that they cannot be predicted by the opponent. Suppose that the first
player chooses each row i = 1,2, ... ,m with a relative frequency Xi and that
the second player chooses each column j = 1,2, ... ,n with a relative
frequency Yj. The resulting average winnings of the first player come to
m II
L L
i= 1 j= 1
a ij xiYj
per play. In matrix notation, this quantity may be recorded as xAy. The
row vector X with components xl,x2, ... ,xm and the column vectory with
componentsYltY2, ... tYlI share a characteristic feature: all the components
are nonnegative and their sum equals one. Such vectors are called stochastic.
In the late nineteen twenties, John von Neumann (Zur Theorie der
Gesselschaftsspiele. 1928. Math. Ann. 100: 195-320) proved the celebrated
minimax theorem asserting that
max min xAy = min max xAy

x y y x
for an arbitrary matrix A, with the maxima taken over all stochastic
vectors x and the minima taken over all stochastic vectors y. To put it
differently, for every matrix A there are stochastic vectors x* andy* such
that
xAy* ~ x* Ay* ~ x* Ay
for all stochastic vectors x and y. Now x* and y* ma~ be thought of as
optimal strategies in the game represented by A: the first player can expect
to win at least
min x* Ay = x* Ay*
y
per play whereas the second player can expect not to lose more than
max xAy* = x* Ay*
x
per play. The quantity x* Ay* is referred to as the value of the game. (The
value of the coin matching game is -50/17.)
After these preliminaries, we return to Cheap, Middling or Dear. To
break the suspense, let us reveal that the game is biased in favour ofB: its
value is about -3.8.(That is, if both players follow their optimal strategies
then B may expect to score about 24.4 whereas the corresponding figure for
A is only about 20.6.) Each conceivable position in the game is specified by
the set S of numbers left on the slate. If the size of S is odd, then A is about
to choose a number; if the size of S is even, then B is about to choose a
number. If the player about to choose a number follows his optimal
strategy then he may expect to score at least v points during the rest of the
game, no matter what strategy his opponent adopts. Similarly, if the player
about to guess a number follows his optimal strategy, then he may expect to
score at least w points during the rest of the game, no matter what strategy
his opponent adopts. By the minimax theorem, the sum v + w equals the
total of the number in S. We shall refer to the difference v - w as the value
of S. We are about to describe an easy way of computing the value of every
S, the optimal ways of choosing a number from S and the optimal ways of
guessing a number in a specified group in S. Our claim may be verified by
induction on the size of S; we omit the tedious proofs.
Each round begins when one of the players names the group from
which he has selected his number. The optimal strategy is:
1 larger groups are always preferable to smaller ones

2 among the three-number groups, Dear is preferable to
Middling and Middling is preferable to Cheap,
3 all the two-number groups are equally attractive,
4 among the one-number groups, Cheap is preferable to
Middling and Middling is preferable to Dear.
Naturally, the selection of the number itself has to be randomized. The

strategy is:
5 from a group containing two numbers x and y,

choose x with probability y/(x + y),
choosey with probability x/(x + y),
6 from a group containing three numbers x,y and z,
choose x with probability yz/ (xy + xz + yZ).,
choose y with probability xz/ (xy + xz + yz),
choose Z with probability xy/(xy + xz + yZ).
These rules provide a complete description of optimal strategies for that

player who is about to choose a number. In fact, almost all of the optimal
choices have the above form; the only exception is the fact that
7 in positions with one Cheap, two Middling and two Dear,

one cannot go wrong by playing Cheap.
Now we turn to evaluating the various positions. First, let us assume

that two of the three groups have been erased completely. In that case, the
value of {x} is -x, the value of {x,y} is (x 2 + y2)j(x + y), and the value of
{x,y,z} is
1 (~z _ ~(x2 + y2) _ xz(x 2 + Z2) _ yz(/ + Z2)).
xy + xz + yz x +y x+ z y+z
In explicit terms, the last formula says that
613
the value of {1,2,3} is - -330 ~ -1 .85757575 ,
1 f { } . 64913.
the va ue 0 4,5,6 IS - 18315 = -3.544253344,
. 2129473.
the value of {7,8,9} IS - 389640 = -5.465232009.
Next, let us note that the rules 1-6 described above, combined with
the twenty-one values presented so far, point out a simple way of
evaluating several different positions. For example, we shall consider S
= {1,2,3,4,5, 7,9}. Since S has an odd number of elements, A is about to
choose a number. An optimal play may take the following course:
A chooses Cheap, B chooses Cheap, A chooses Middling,
B chooses Dear, A chooses Cheap, B chooses Middling,
A chooses Dear.
The two players might as well agree to proceed in the following order:
A chooses Cheap, B chooses Cheap, A chooses Cheap,
A chooses Middling, B chooses Middling,
B chooses Dear, A chooses Dear.
Thus we conclude that the value of S is
value of {1,2,3} + value of {4,5} - value of {7,9}
613 41 130 21491
= --+ - - - = - - - ~ -5.4270202.
330 9 16 3960
In order to provide a similar way of evaluating each of the 512
positions, we shall require a few additional building blocks. Again, let S be
a position with two of the three groups completely erased. By the out-of-
phase game on S = {x,y}, we shall mean the game with A choosing in both
rounds; by an out-of-phase game on S = {x,y,Z}, we shall mean the game
with A choosing in the first round only and guessing in the remaining two
rounds. The out-of-phase value of {x,y} is - (x 2 + y2)/(x + y) and the out-
of-phase value of {x,y,z} is
xy(x 2 + y2)
+ xz(x + Z2) + YZ(y2Z2)).
1 ( 2
xyz+
xy + xz + yz x +y x+ Z y+Z
In explicit terms, the last formula says that
the out-of-phase value of {1,2,3} is ~~~ ~ 2.948484848,

. 124313 .
t h e out-o f-p h ase va1ue 0 f { 4,5,6 } IS = 6.787496587,
18315
4185793
the out-of-phase value of P,8,9} is 389640 ~ 10.74271892
Now we know all we need for a fast evaluation of an arbitrary position. For
example, we shall consider S = {1,2, 3,5,6}. An optimal play may take the
following course:
A chooses Cheap, B chooses Cheap, A chooses Middling,
B chooses Cheap, A chooses Middling,
The two players might as well agree to proceed as follows:
A chooses Cheap, B chooses Cheap, B chooses Cheap,
A chooses Middling, A chooses Middling.
Thus, we conclude that the value of S is
out-of-phase value of {1,2,3} + out-of-phase value of {5,6}
= -
G~~) e~) = - ~~~ ~ -2.5969696.
Finally, we turn to the optimal guessing strategies. Suppose that one
of the players is choosing a number from a three-number group {xty,z}. Let
a,b,c denote the values of S - {x}, S - {y}, S - {z}, respectively. Clearly,
the optimal guessing strategy for the other player is identical with the
second player's optimal strategy in the game specified by the matrix
-x-a x-a
[ y-b x-a]
-y-b y-b .
z-c z-c -z-c
This player ought to
guess x with probability ( + 1 + ) [y(x+c-a)+z(x+b-a)],

2 xy xz yx
guess y with probability ( + I +

2 xy x~ y~)
[x(y + e - b) + ~(Y + a - b)),
guess ~ with probability 2(xy + x~

I +
y~)
[x(~ + b - e) + y(~ + a - e)).
Similarly, if one ofthe players chooses a number from a two-number group

and if a,b stand for the values of S - {x}, S - {Y}, respectively, then the
other player's optimal guessing strategy amounts to the second player's
optimal strategy in the game specified by the matrix
[ -x-a x-a].
y-b -y-b
This player ought to
guess x with probability 2(x ~ y) (x + y - a + b),
I
guessy with probability (+ (x +Y + a - b).
2x y)
Technical Report SOCS-79.3

A Random Hopscotch
Problem or Ho", to Make
Johnny Read More
-~-
David Berengut
STATE UNIVERSITY OF NEW YORK AT BINGHAMTON
Many of our most popular games combine the element of randomness-

produced either by the roll of a die, the twirl of a spinner, or the deal of a
shuffled deck of cards-with apia ying surface or board consisting of spaces
arranged in a sequence, either in a closed loop (as in Monopoly), or with
separate start and finish points (as in cribbage or backgammon). The
movement of pieces in these games might be whimsically described as
random hopscotch. As a mathematical statistician, I have more than a passing
interest in the probabilistic aspects of such games.
This article originated from a problem that David Klamer mentioned
to me. A teacher of young children had devised a game of random
hopscotch with the aim of stimulating her students' interest in acquiring
reading skills. The board consisted of a sequence of spaces with separate
starting and finishing points. Various spaces contained instructions to
perform various reading tasks, while the remaining ones were blank. Each
student in tum took a run through the board, making moves according to
the throw of a modified die whose faces consisted of two 1's, two 2's, and
two 3's (so that only moves of 1, 2, or 3 spaces, each equally likely, were
possible). When the student landed on a space containing an instruction,
he was required to perform the appropriate reading task.
By introducing the element of chance-as well as the challenge to

control the outcomes of the tosses-the teacher was very successful at
holding the attention of her students. The question posed by the teacher
was the following: In which spaces should she place the instructions so as to
maximize the expected number of tasks a student would have to perform?
Implicit in this question, of course, is the assumption that the number of
spaces on the board and the number of tasks are both fixed.
Frolll Words to NUlllbers

The key to solving many mathematical problems lies in simply formulating
the problem the right way; to a large degree, this is the case here. Since it is
often easier to solve mathematical problems in a general way rather than
for specific cases, I choose to let the number of spaces on the board and the
number of tasks both be arbitrary, and denote these quantities by nand t
respectively.
To solve any problem, we must first know exactly what the problem is
saying. What, precisely, is meant by the expression expected number rif tasks to
be performed? The expected value of a random variable is a standard notion
in the theory of probability-in brief, it is the probability-weighted
average of the possible values of the random variable. If T represents the
number of tasks performed in a hypothetical run through the board, T is a
random variable which can take on anyone of the integer values from 0 to t
inclusively, each with a certain probability (as yet undetermined). Letting
P(j) denote the probability that T takes the value j, where j is any integer
between 0 and t, the expected value of T, written E( T), is given by the
formula:
E(T) = OoP(O) + loP(l) + ... + (t -1) op(t -1) + top(t).
It is convenient to number the spaces on the board sequentially from
the starting point so that the first space after the starting point is numbered
I, the next space 2, and so on, with the final space on the board being
numbered n. Let i1 , i2 , ••• , it represent the numbers of the spaces, in
ascending order, which contain tasks. For any given play of the game then,
the value of T is simply the number of spaces among the set {i l' i2 , ... , it} on
which the player lands. A convenient mathematical device here is the use
of indicator variables, which can only take on the values 1 or 0, according to
whether or not a particular event occurs. In this case, let 11 be the indicator
variable for the event that the player lands on space i1; that is, 11 equals I if
the player does land on space i1 during the course of the game, and it
equals 0 ifnot. In a similar fashion, let 12 be the indicator variable for the
event that the player lands on space i 2 , and so on, up to It. Then, clearly, 11
+ 12 + ... + It simply counts the number of times during the game that
the player lands on a space containing a task. We will call this number, T.
Calculating the expected value of T is thus equivalent to calculating
the expected value of 11 + 12 + ... + It. At this point, we make use of a
A RANDOM HOPSCOTCH PROBLEM
fundamental property of expectations: the expected value of a sum is the

sum of the expected values. (Thus, for example, the expected number of
rainy days in a year for a given location equals the expected number of
rainy days in January plus the expected number of rainy days in February
plus, etc.) In symbols, then, E(T) equals E(It) +E(12 ) +···+E(lt ).
How do we calculate E(11 ), for example? From the definition of
expectation, it follows that E(11 ) is given by O·P(O) + I·P(I), or simply
P( I). P( I) is the probability that 11 takes the value I; the probability that
the player will land on space i 1 • Similarly, E(12 ) is simply the probability
that the player will land on space i 2 , and so on. If we let Pi denote the
probability that the player will land on space i, where i ranges from I to n,
then we can write E( T) = Pil + Pi2 + ... + Pi,.
The problem, then, is to evaluate the numbers PI, P2, ... ,Pn. Once we
have these numbers, we are in a position to solve the teacher's problem:
choosing the spaces ii' i2 ... , it to place the tasks so as to maximize E(7).
Obviously, these are the t spaces whose corresponding p-values are the t
largest among all the values PI' P2' ... , Pn" In effect, the teacher ought to
choose those t spaces which have the greatest chance of being landed on!
Determining which spaces these are, however, is a nontrivial prob-
lem. To illustrate the method of solution, I want to consider first a slightly
simpler version of the game.
A Simpler Problem
Suppose the rules of the game are altered slightly, so that each move along
the board can be of only one or two spaces, determined by the flip of a coin.
In this case, can we calculate the probabilities of landing on each of the
spaces, PI' P2' ... , Pn?
Well, PI is easy enough-it's just the probability of moving one space
on the first move, namely t. Now P2 is the probability oflanding on space 2.
This can happen in either of two distinct ways: by moving two spaces on
the first move, or by moving one space on each of the first two moves. The
first way has probability t of happening; the probability of the second way
is just the probability of getting 2 tails in 2 consecutive coin tosses. Since the
coin tosses are independent of each other, the multiplicative principle tells
us that this probability is simply the product of the probability of a tail on
the first toss and the probability of a tail on the second toss, namely t·t or
t. Since the two ways of arriving at space 2 are exhaustive and distinct, the
overall probability of landing on space 2, P2' is the sum of their
t t
probabilities + or -i.
The same type of argument, which requires listing all the possible
paths, could be used to calculate P3 and all the remaining p's; obviously,
the task will become increasingly tedious as one moves further up the
board. Fortunately, this difficulty can be finessed by using a more
ingenious argument to obtain the p's in a recursive manner.
The Recursive Approach

Suppose we want to calculate Pi' the probability that the player will land
on space i, where i exceeds 2. Any move landing on space i must have
originated from either space i - 2 or i-I. Thus, there are two distinct
ways of landing on space i that can be described:
1 the player eventually lands on space i - 2 and then moves

2 spaces on the subsequent move (thus passing over space
i-I); or
2 the player eventually lands on space i - I and on the
subsequent move advances 1 space (see Figure 1).
Because successive moves are independent, the probability of landing

on space i via path 1 is the probability of eventually landing on space i - 2
(namely Pi-2), multiplied by the probability of moving 2 spaces on the
next move, namely 1, which gives 1Pi-2; similarly the probability of
arriving via path 2 is 1Pi-l. Thus the overall probability of landing on
space i (PJ equals tPi-l + tPi-2. Therefore, each term in the sequence PI'
h, ... , Pn beyond the second is simply the average of the two preceding
terms!
t,
Knowing PI = h = i enables us to conclude that P3 = 1(1 + i) or
i. In this manner, we can recursively calculate Ps (= ~~), P4 (= U), etc. Of
course, if we were interested in calculating the value of P20, say, by this
method, we would first have to calculate the previous 19 p-values. This
would certainly be a tedious computation by hand. Can we obtain an
explicit formula for Pi' for any i? The answer is an emphaticyes. Indeed, the
t
formula for Pi turns out to be disarmingly simple: Pi = + t( - 1)i (which
··dentaIIy te IIs us that P20 equaIs 23:145:728
mCI 097 152) .
; - 1
; - 2
I i-I I
FIGURE 1
The two distinct ways of arriving at the /1h space.
Id.. ,'-d,
-~-
=. -
d i-I-
--4
i-2 i-I i+1
FIGURE 2
Relations between successive p-values.
Obtaining this formula from the recursive formula is a fairly

straightforward matter, and follows from the simple observation that the
average of two numbers lies midway between the numbers. Thus, Pi lies
midway between Pi-2 and Pi-I. The difference between Pi and Pi-1 is
therefore in magnitude, half the difference between Pi-1 and Pi- 2' and it is
easy to see that the differences are opposite in sign. Letting di denote the
difference Pi - Pi - 1, for any i, means that di = - tdi _ 1. This is
illustrated in Figure 2.
By repeated application of the relation di = - !di _ l' we get di =
-!di - 1 = -t( -tdi - 2) = ... = (-t)i- 2d2· But d2 equalsPz - PI' which
. 4
IS 11Th
3 - "2 or 4. us di equa1s 41( -"21)i-2 , or ( -"2l)i . F·ma II y, rna k·mg use 0 f t h e
fact that Pi can be written as the sum (Pi - Pi-d + (Pi-1 - Pi-2) + ...
+ (P2 - PI) + Pt, or equivalently di + ~i-1 + .... + d2 + PI' we get by
direct substitution that Pi equals (-t)'+ (-t)'-l+ ... + (-t)2+!.
Using the well-known formula for the sum of a finite-length geometric
series yields the answer Pi = t+ t( _!)i. Table I lists the first 12p-values.
An interesting observation at this point is to note that the sequence of
p-values oscjllates in a regular fashion around the value t, with the
oscillation decreasing in magnitude the farther one proceeds along the
sequence. Moreover, the subsequence with odd indices Pt, PJ, P5, ...
increases monotonically toward the limiting value of t, while the com-
plementary subsequence P2, P4' P6' ... decreases monotonically to the same
limiting value. Clearly, then, the largest p-value is Pz, the second largest is
P4' and so on, while the smallest p-value is PI' the second smallest is PJ, etc.
We are finally in a position to answer the teacher's question (at least
for the modified, simpler version of the game) To maximize the expected
~55~
i Pi
1 .5000
2 .7500
3 .6250
4 .6875
5 .6563
6 .6719
7 .6641
8 .6680
9 .6660
10 .6670
11 .6665
12 .6667
TABLE 1
Values of Pi for i up to 12. Simpler version of game.
number of tasks performed, the tasks should be placed on the even-

numbered spaces, beginning with space 2 and proceeding sequentially
down the board until either (a) all the tasks are placed or (b) the end of
the board is reached; in case (b), any remaining tasks are placed on the
unoccupied (that is, odd-numbered) spaces, beginning at the end of the
board and working backwards.
What is it Worth
to be SDlart?
Being a mathematician of somewhat applied bent, I wondered how much
is gained by placing the tasks optimally on the board. To make the
question more precise, what is the difference in expected number of tasks
performed between the optimal placement and the worst possible place-
ment, which is obtained by simply reversing the roles of odd and even in
the optimal scheme? The answer involves only straightforward algebra,
and turns out to depend on whether or not n (the number of spaces on the
board) is at least twice as large as t (the number of tasks); if this is the case,
the answer is (1 - 4 -/)/3, or (1 - 4t - n )/3. This number can never exceed
t, no matter how large n or t may be. In most cases, however, it will be
quite close to t. For example, if there are 5 tasks in all, and the board
contains 10 or more spaces, the expected number oftasks performed under
the optimal placement is 3.4443, as compared with a value of 3.1113 under
the worst placement; the difference is 0.3330 (341/1024 to be exact).
Although thi~ number may seem an insignificant difference, it becomes
increasingly important the more often the game is played. For example, if
each student in a class of 30 plays the game once, the difference between
the optimal and the worst placement of the tasks translates into approxi-
mately 10 extra reading tasks being performed.
The Original Problem

Having solved the simpler problem, let us try to apply the same methods to
solving the teacher's original problem, where moves of one, two, or three
spaces are equally likely. Certainly an analogous argument works for
producing a recursive relation; in this case, the relation becomes Pi = (Pi-l
+ Pi-2 + Pi-3)/3, for values of i exceeding 3. In order to get the ball
rolling, we need to calculate Pl' Pz, and h· Now Pl is obviously t. Space
2 can be landed on either by an initial move of two spaces or by two
consecutive moves of one space each, hence P2 equals t + (t)2, or~. There
are four ways oflanding on space 3: (I) an initial move of three spaces; (2) a
move of two spaces followed by a move of one space; (3) a
move of one space followed by a move of two spaces; (4) three consecutive
moves of one space each. Thus P3 equals t + (t)2 + (t)2 + (t)3 = ~~. The
recursive relation, in conjunction with these three values, enables us to
calculate in principle the value of Pi for any i. Table 2 gives the first twelve
p-values.
i Pi
I .3333
2 .4444
3 .5926
4 .4568
5 .4979
6 .5158
7 .4902
8 .5013
9 .5024
10 .4980
II .5006
12 .5003
TABLE 2
Values of Pi for i up to 12. Teacher's version of game .
.657&
Unlike the case of the simpler game, however, there is no easy way of
deriving on explicit formula for Pi from the recursive relation. * Neverthe-
less, some interesting observations can be made. Since the recursive
relation states that any p-value beyond the third is simply the average of
the preceding three values, it follows that it must exceed the smallest of
those three values but be smaller than the largest of them. By repeated
application of this argument, one is able to conclude that for any three
consecutive values in the sequence {Pl' P2' ... }, the largest of the three
values exceeds all subsequent terms in the sequence, while the smallest of
the three values is in turn exceeded by all subsequent terms in the
sequence. An immediate consequence of this remark is the fact that the
largest ofPl' h, h, namely P3' is the largest of all the p-values, whereas the
smallest of the three, namely Pl' is the smallest of all the p-values.
Is there a regular pattern to the way in which the p-values are
ordered, as there was in the simpler model? You might be tempted to
speculate that the largest p-values, in descending order, would be given by
the subsequence P3' P6' P9' .... Alas, this is not the case, as an examination
of Table 2 reveals: the pattern first breaks down with P12 ( =0.5003) which
is not the fourth-largest, but only the sixth-largest p-value, being exceeded
by both Ps (=0.5013) and Pll (=0.5006). Similarly, a listing of the
smallest p-values in ascending order reveals no regular pattern: Pi> P2' P4'
P7, P5, P10, etc.
Table 1 does suggest that the sequence of p-values converges to in t
the limit; in fact, this can be proved, although the proof requires some
rather sophisticated mathematics. An easy consequence of this fact,
however, is that no three consecutive p-values can lie on the same side of
t; for if there did exist three such p-values, then by earlier remarks, all
subsequent p-values must lie between the largest and smallest of the three;
t
hence they must be further from than the one of the three which is closest
to t. This contradicts the convergence of the p-values to .1.
The p-values are forever oscillating about the value t with no more
t.
than two consecutive p-values on the same side of There is no apparent
pattern, however, to the oscillation. Table 3 gives the pattern of oscillation
for the first 25 p-values.
Suppose we apply what we've learned to the specific example
considered earlier. If there are five tasks to be placed on the board, and the
board contains at least eleven spaces, then the optimal placement is on
spaces 3, 6, 8, 9, and 11; this gives a value of 2.6127 for the expected
number of tasks performed. By contrast, the worst possible placement
would be on spaces I, 2, 4, 5, and 7, giving a corresponding value of
2.2226. The difference between the best and the worst placement is 0.3901
* The more advanced reader may be interested to know that Pi can be expressed in
the form (1/2) + [(-I)i/4] [(1 + ~)i+ (1- ~)i].
i I 2 3 4 5 6 7 8 9 10 II 12
Pi vs. 0.5 - - + - - + - + + - + +
i 13 14 15 16 17 18 19 20 21 22 23 24 25
Pi vs. 0.5 - + + - + - - + - - + - -
TABLE 3
Position offirst 25 p-values relative to 0.5. += above, - = below.
expected tasks; it is interesting that this difference is larger than in the case
of the simpler game.
SolDe Final Words

This problem illustrates how certain aspects of a solution may be
generalized for all cases, whereas other aspects may be specific to the
particular case considered. The argument leading to the recursive relation
for the p/s is valid for all versions of the game, but the pattern of the
resulting solution depends very strongly on how many possible moves can
be made. Of course, the two most common versions of random
hopscotch-based on moves of 1 to 6 spaces or of2 to 12 spaces-have not
been considered here, but the reader should now be in a position to analyze
these versions with the aid of a calculator.
After working his way through this article, the reader can have the
fun of generalizing the results. After all, all work and no play makes Johnny
a dull boy.
Geometry
-~-
Wreaths of
Tangent Circles
-~-
SololDon W. GololDb
UNIVERSITY OF SOUTHERN CALIFORNIA
I t is well known that a circle can be exactly surrounded by six other circles
the same size as the original circle (Figure 1). Generally, n identical circles
of radius s can be used to exactly surround one circle of radius r, for all
n 2:: 3 where, by elementary trigonometry we have (see Figure 2)
. n s
Sln- = --,
n r + s
from which r = s (csc n/n - 1). Ifwe fix r = 1, then as n increases, s decreases,
as shown in Table 1. .
FIGURE 1
A circle exactly surrounded by six identical circles.
WREATHS OF TANGENT CIRCLES
II
a =-
n
FIGURE 2
The case n = 7 of n identical circles surrounding a given circle.
We next consider the case where the inner circle and the n
surrounding circles may all have different radii (Figure 3). This problem is
encountered when one has circular coins in a variety of sizes, and wishes to
surround one of them exactly with several of the others. First, we need a
precise definition of the notion that n circles exactly surround a given
circle:
DEFINITION A circle C* is said to be exactly surrounded by the n

circles C1 ,C2 , ••. ,Cn if each circle Cj is (ex-
teriorly) tangent to C*, to Cj - 1 , and to Ci+ l'
where the subscripts are taken modulo n.
Given the radii r l' r2' ... , rn of the surrounding circles, is there a simple
formula for the radius r of the surrounded circle C*?
The first surprise is that for n > 3, the size of the surrounded circle
depends not only on the magnitudes of the radii of the circles surrounding
it, but on the sequential order in which these circles occur! In general, with
n given surrounding circles, there may be as many as t(n - I)! different
sizes for the exactly-surrounded circle in the middle, depending on the
n s
3 6.46410 = 3 + 2~
4 2.41421 = I + :;2
5 1.42592
6 1.00000
7 .76642
8 .61991
9 .51980
10 .44721
TABLE I
The radius s of each of n circles exactly surrounding a unit circle.
FIGURE 3
A circle exactly surrounded by seven unequal circles.
permuted order of the surrounding circles. (In fact, if the n radii of the
surrounding circles are algebraically independent real numbers, there will
indeed be t(n - I)! different values for the radius of the exactly-
surrounded circle.) The importance of the order of surrounding circles is
exemplified in Figure 4.
In Figure 4, the circles C6 and C;; are the same size. Yet C6 placed
between C4 and Cs, accomplishes something in surrounding C*, while C;;,
between C1 and C2 , does not. IfC;; were enlarged a tiny bit, it would force
C1 and C2 to separate, but would still not accomplish as much as it would
between two circles more nearly its own size!
Roughly speaking, the radius of the exactly-surrounded circle is
maximized if the surrounding circles are arranged so that each circle is as
close as possible to others of the same size. Conversely, this radius is
minimized if adjacent circles are as disparate as possible in size. If it were
merely a matter of stringing the n circles out along a line, the maximizing
and minimizing strategies would be clear (Figure 5). (The straight line can
be thought of as a central circle C* of infinite radius, which of course
FIGURE 4
Circles C6 and C;' are the same size, but play very different roles in
surrounding C·.
FIGURE 5
Maximizing and minimizing strategies for placing n unequal circles
along a line.
cannot be "surrounded" by a finite number of finite circles.) We number

the circles C1,C2,C3, ... ,Cn so that their corresponding radii satisfY'1 ~'2
~ '3 ~ ... ~'n·
NOTE The notion of "minimizing" is not well-defined if Cn is smaller
than the circle simultaneously tangent to C1,C2, and C*. For present
purposes, we will not consider the minimization problem for this extreme
case.
As to the problem of placing the n surrounding circles so as to
maximize the radius of the surrounded circle, the empirical algorithm
illustrated in Figure 6 is suggested, continuing from C1 with C2,C4, C6, ... in
one direction, and with C3, C5, C7 , ••• in the other. An empirical algorithm
for minimizing the radius of C* is illustrated in Figure 7, where, starting
with C1, we continue in one direction with Cn,C2,Cn-2,C4,Cn-4,C6' ... '
and in the other direction with Cn-1,C3,Cn-3,C5, Cn- 5, ... . For neither of
these algorithms has the optimality been proved or disproved, for all
possible choices of radii of the circles C1,C2,C3, ... ,Cn.
When n = 3, there is, except for rigid Euclidean motions (rotation,
reflection, and translation), only one way to arrange the three circles
C1 ,C2 , and C3 to be mutually tangent, and this uniquely specifies the
radius of the exactly surrounded circle C* (see Figure 8). IfC1,C2 , and C3
have radii a,b, and c, respectively, then the radius, of C* can be expressed
as
FIGURE 6
Heuristic algorithm for maximizing the radius of C*.
FIGURE 7
Heuristic algorithm for minimizing the radius of C*.
abc
r =---------------r===========
ab+ bc+ ca+ 2Jabc(a+ b+ c)
When n = 4, if the circles C1 ,C2 ,C3 , and C4 all have different radii,
then there are t( 4 - I)! = 3 essentially different ways in which they can
be arranged so as to surround a central circle C*, as shown in Figure 9.
We will consider the special case when only two distinct radii occur
among the four surrounding circles. Suppose that C 1 and C2 have radius a,
while C3 and C4 have radius b. The only two distinguishable cases that
occur are shown in Figure 10.
In Case I, by Pythagoras' Theorem we have (a + b)2 = (a + r)2
+ (b + r)2, from which 2ab = 2ar + 2br + 2r2, or r2 + (a + b)r - ab
= O. Thus, by the quadratic formula,
Ja 2 +6ab+b 2 - (a+b)
r= ~-------------'-----'-
2
Integer values of r occur for many integer choices of a and b. Thus r(3,2)
= 1, r(1O,3) = 2, r(12,5) = 3, etc. (They are related to Pythagorean
Triplets by the rule that if r(a,b) = r, then (b + r, a + r, a + b) is the
Pythagorean Triplet, which is exemplified in Figure 10, Case 1.
FIGURE 8
Three surrounding circles uniquely specify C*.
FIGURE 9
The three essentially different sequential orderings where n = 4.
Conversely, if (A,B,C) is a Pythagorean Triplet, with A2 + B2 = C2, then
we have a diophantine solution of Case I with a = (A - B + C) /2, b =
(-A + B + C)/2, r = (A + B - C)/2.) It is interesting that a coin of
radius 1 can be exactly surrounded by two coins of radius 2 and two coins
of radius 3, provided that they are arranged as in Case I. If on the other
hand, they are arranged as in Case II, they may seem to fit perfectly, but we
shall see that they do not.
In Case II, let h = hl + h2 . By Pythagoras' Theorem, we see that
hi = (a + R)2 - a2 = 2aR + R2, h~ = (b + R)2 - b2 = 2bR + R2, and
(if we observe that the dotted line has length h), h2 = (a + b)2 -
(a - b)2 = 4ab. Thus,
4ab = h2 = (h 1 + h2)2 = hi + h~ + 2hlh2 = 2(aR + bR + R2 + h1h2),
from which
CASE I CASE II
FIGURE 10
The two distinguishable cases of exactly surrounding a central circle
with two circles of radius a and two circles of radius h.
~67~
and
4a b2 2
4ab + 2(a + b)R + R2 = R2 + (a + b) 2 + R2 - 4ab + 2(a + b)R
4ab(a + b)
R
R2(a 2 - 6ab + b2) - 4ab(a + b)R + 4a 2b2 = O.
By the quadratic formula,
4ab(a + b) ± jI6a 2b2{(a + b)2 - (a 2 - 6ab + b2)}
R = 2(a2 _ 6ab + b2)
and with the requirement R > 0,
2ab{2fob - (a + b)} 2ab
R = 8ab - (a + b)2 = (a + b) + 2fob·
We see that for integral a and b, R is rational if and only if 2ab is a per-
fect square. In particular, R(3,2) = 12/(5+ 4J3) = (12/23) (4J3 - 5)
= 1.006019 -, an increase of about 0.6% over r(3,2) = 1. This difference
is too small to have caused much concern "in practice." However, as the
ratio alb increases, so too does the ratio R/r. Thus, while r(IO,3) = 2, we
have R(10,3) = 60/(13 + 4Jl5) = 60/71 (4Jl5 - 13) = 2.10586-;
an increase in the radius R of C* of some 5% over r.
When the circle C* is exactly surrounded by one circle of radius band
n circles of radius a, the order of the surrounding circles does not affect the
radius r of C*. When n = 2, this is a special case of four mutually
(externally) tangent circles, where the circle C* of radius r is exactly
surrounded by three circles, of respective radii a,b,c. As previously
mentioned, and as derived in [I], the formula for this more general case is
abc S3
r - . - ------=--==
- ab + bc+ ca + 2jabc(a + b + c) - S2 + 2jS1S3
where Sl = a + b + c, S2 = ab + be + ca, S3 = abc, are the three ele-
mentary symmetric functions of a,b,c.
Our special case, when a = c, is more elementary, and satisfies
4br(b + r)
a= (b _ r)2 , b > r.
~68~
FIGURE 11
A circle exactly surrounded by two circles of radius a and one of radius
b.
Figure 11 shows us that:

(a+ r)2 =a 2 + x 2, (a+ b)2 =a 2 + (x+ r+ b)2.
Thus,
x+r+b=J2ab+b 2
with
x= J2ar+r 2.
Hence,
r+b=J2ab+b 2 -J2ar+r 2,
r2 + 2rb + b2 = (2ab + b2) + (2ar + r2) - 2 J (2ab + b2) (2ar + r2),
(ab + ar - rb)2 = (2ab + b2) (2ar + r2),
a2b2 + a2r2 = 2a 2br + 4abr 2 + 4arb 2,
(b - r)2 = (l/a) (4br) (b+ r),
a = 4br(b + r)/(b - r)2.
The apparent symmetry between band r is not realized geometrically, since
clearly b > r. If a and b are given, we find from the quadratic formula that
(a + 2b) - 2Jb 2 + 2ab

r = b -'-------'---"----
a - 4b
FIGURE 12
A circle of radius I exactly surrounded by a straight line and two
circles of radius 4.
The case a = 4b is a "removable singularity," for In this case r = tb.

However, in the related equation
(a + 2r) + 2Jr2 + 2ar
b= r a - 4r '
the case a = 4r is a genuine singularity, corresponding to h = 00.

Specifically, with r = 1 and a = 4, we see from the 3-4-5 right triangle in
Figure 12 that the three mutually tangent circles are also tangent to a line
(that is, a circle of radius b = 00).
There are many cases of integer triples (r,h,a), including (1,2,24)
(1,3,12), (5,7,420), (6,14,105), etc. These may be obtained from such
formulas as:
FIGURE 13
A circle of radius b exactly surrounded by three circles of radius a and
one of radius b.
.870&'
FIGURE 14
A circle of radius b exactly surrounded by four circles of radius a and
one of radius b.
1 = n, b = n + I, a = 4n(n + 1)(2n + I)
r
2 r = n, b = n + 2, a = 2n(n + I)(n + 2)
3 r = 2n, b = 2n + 8, a = n(n + 2)(n + 4), etc.
Let us also require the circle C*, surrounded by one circle of radius b
and n circles of radius a, to have radius r = b. For each n ~ 3, this
configuration uniquely determines a ratio Q.n = a/b. We will discuss the
values of Q.n for n = 3, 4, 5, and 6 in some detail.
In Figure 13, we see the case n = 3. Here we have (2a)2 = x 2 +
(a + 2b)2 and (a + b)2 = x 2 + b2. Thus, (2a)2 - (a + b)2 = (a + 2b)2
- b2, from which a2 - 3ab - 2b 2 = O. By the quadratic formula, Q.3 = a/ b
= (3 + fo)/2 = 3.5615528 ....
When n = 5, of course Q.5 = a/ b = I, since this is reduced to the
configuration in Figure 1. However, when n = 4, the situation is somewhat
more complex. From Figure 14, we observe each of the following:
(a + b)2 = b2 + x 2,
(a + b)2= a2 + y2,
(2a)2 = (y + b)2 + (x - a)2.
Thus x = a2+ 2ab, y = Jb 2+ 2ab, and a2 - 2ab - b2 = bJb 2+ 2ab

-a a2+ 2ab, from which
3a 2 - ab - b2 = J(a 2 + 2ab)(b 2 + 2ab) ,

FIGURE 15
When a circle of radius b is exactly surrounded by n circles of radius a
and one of radius b, it is evident that 1t=8+ (n-l)</>, where
8 =cos- 1 [b/(a+ b)] and </> =sin- 1 [a/(a+ b)].
and
9a 4 - 8a 3 b - IOa 2 b2 + b4 = O.
Thus Q4 = alb is a root of 9x 4 - 8x 3 - lOx 2 + 1 = 0, which has the
numerical value Q4 = 1.5684897 .... (In principle, all quartic equations
have solutions which can be expressed in terms of radicals, but the explicit
solution in radicals for this case is too complicated to be worth including
here.)
The case n = 6 (see Figure 15) was first posed as a problem in an
unpublished letter from Gary A. Ford to Martin Gardner in 1973. Ford
commented that the problem was inspired by the arrangement of coins
(specifically, dimes and quarters), and that the best he and his colleagues at
the University of Maryland had been able to do was to express Q6 as a root
of a tenth degree polynomial. (Ford subsequently published this problem
in MIT's Technology Review [2].) Can we do as well or perhaps better using
the methods already illustrated? More generally, is it possible to express Qn
either algebraically or trigonometrically, as a function of n, for all n ~ 3?
We will see that a general trigonometric expression exists, from which a
polynomial equation can always be obtained. In particular, we will express
Q6 as a root of a polynomial of degree eight.
In Figure 15, we see that the straight angle in the central circle of
radius b is the sum of an angle fJ and 5 copies of an angle q> where fJ
= cos- 1 [bl(a+ b)]and q> = sin- 1 [al(a+ b)]. For general n, the result is
1t = cos - 1 C~ a b) ) + (n - 1) sin - 1 C~
a b) )-
If we set al(a + b) = a and b/(a + b) = P, then a + P= 1, and the ratio Qn

= alb = alP, with
1
For practical computation, this formula is sufficient to allow oe, and
therefore Qn' to be determined to any degree of accuracy. However, it is
also possible, for each n, to replace the equation 1 by a polynomial
equation having oe as a root. Moreover, it is also possible to obtain a
polynomial equation having Qn = oe/(1- oe) as a root, for if J(x) = 0 has
x = oe as a root, then it is readily verified, by direct substitution, that g(x)
=J[x/(1 + x)] = 0 has oe/(I- O() as a root.
The method is to rewrite 1 as (n -1) sin- 1oe = 11: - cos- 1(1 - oe),
and then take the cosine of both sides, to obtain
cos [( n - 1) sin - 1oe] = cos [11: - cos - 1(1 - oe)] = 0( - l.
Let sin -10( = z. Then it is well known that cos(n - l)z is a polynomial of
degree n - 1 in cos z, and cos Z = cos(sin -10() = J
1 - 0(2. Thus, in the
worst case, oe may be the root ofa polynomial of degree 2(n - 1). In fact,
for n odd, Qn satisfies a polynomial equation of degree :::;'n - 1 for all
n ~ 3; and for n even, Qn satisfies a polynomial equation of degree
:::;, 2(n-2), for all n~4.
We will illustrate the actual computations for the (already solved)
cases n = 3 and n = 4.
When n = 3, cos 2z = 0( -1, 2cos 2z -1 = oe -1, 2cos 2z = oe, 2(1- oe 2)
= oe, 2oe 2 + oe - 2 = 0, and 0( satisfiesJ(x) = 2x2 + X - 2 = O. Q3 then satis-
fiesg(x) =J[x/(I+x)] =0, and (1 +X)2g (X) =2x 2 +x(l+x) -2(I+x)2
= x2 - 3x - 2 = O. Then Q3 is the root (3 + ji7)/2 = 3.56155 of
g(x) = x 2 - 3x - 2 = o.
Similarly, when n=4, cos 3z=oe-l, 4cos 3 z-3 cos z=oe-l,
Jl - 0(2 {4(1 - 0(2) - 3} = 0( - 1, (1- 0(2)(1 - 40(2)2 = (oe - 1)2,
(1 + O() (16oe - 8oe + 1) = 1 - oe,
4 2 160(5 + 16oe4 - 8oe 3 - 8oe 2 + 20e = 0
and since oe = 0 is not a possible solution, oe is a root ofJ(x) = 8x 4 + 8x 3
- <!x 2 - 4x + 1 = O. Hence C2t = oe/ (1 - O() satisfies g(x) = J[x/ (1 + x)]
= 0, so that (1 + X)4g(X) = 8x + 8x 3(1 + x) - 4x 2(1 + x)2 - 4x(1 + x)3
+ (1 + x)4 = 9x4 - 8x 3 - 10x 2+ 1 = 0 has ~ = l.56849 ... as a root.
In Table 2, we give the polynomials for Qn for 3 :::;, n :::;, 9, with the
corresponding values of Qn. Many patterns in the coefficients of these
polynomials are readily apparent. The even and the odd values of n clearly
correspond to separate populations of polynomials.
Finally, we mention the elegant result that when three circles, of
respective radii a, b, and c, are mutually externally tangent to one another
and to a common tangent line, with a ~ b ~ c, then [3]
III
-
Ja+vfb
-- -
- Jc.
(This generalizes the situation depicted in Figure 12.)
n degree polynomial for Q. Q.

3 2 x 2 - 3x - 2. 3.56155
4 4 9x 4 - 8x 3 - I Ox 2 + I. 1.56849
5 4 x4 - IIx 3 + x 2 + 7x + 2. 1.00000
6 8 25x 8 -188x 7 + 236x 6 + 436x 4 - 2x 4 -180x 3 0.73403
- 68x 2 - 4x + I.
7 6 x 6 - 31x 5 + 40x 4 + 42x 3 -7x 2 - Ilx - 2. 0.58027
8 12 49x 12 - 956x 11 + 5090x 10 - 3036x 9 - 11121x 8 0.48015
+ 1800x 7 + 10140x 6 + 4200x 5 - 86.';x4 - 972x 3
- 222x2 - 12x + I.
9 8 x 8 - 55x 7 + 259x 6 + 77 x 5 - 215x 4 - 10 I x 3 0.40977
-17x 2 + 15x+ 2.
TABLE 2
Polynomials and values for Q., 3 ~ n:S;; 9.
NOTES TO TABLE 2
1 It appears that for n = 4k + 1, the polynomial of degree
n - 1 which we obtain always has x = 1 as a root. For n = 5,
this corresponds to Qs = 1. However, for n = 9, 13, 17, ... ,
the number Qn satisfies a polynomial equation of degree
Sn-2.
2 For n > 4, none of these polynomials have been proved
irreducible. Note that the polynomial for n = 5 is the
product of x - I and the irreducible cubic x 3 - lOx 2 - 9x
- 2. For n = 9, the factorization is
(x - 1) (x 7 - 54x 6 + 205x s + 282x4 + 67x 3 - 34x 2 - 17x - 2).
References
I Beecroft, Philip. 1842. Properties of circles in mutual contact. Lady's and
Gentleman's Diary, pp. 91-96.
2 Ford, Gary A. 1974. Technology Review, problem June 5, vol. 76: 57-8.
(See also problem NS 13, vol. 81, November 1978, p. 84.)
3 Trigg, C. W. 1940. Problem E432, American Math. Monthly, 47: 487.
4 _ _ . 1941. Solution to Problem E432 American Math. Monthly 48: 267-
68.
Bicycle Tubes
Inside Out
-~-
Herbert Taylor
UNIVERSITY OF SOUTHERN CALIFORNIA
The old rubber sheet geometry discussed surfaces which could be bent,
stretched, or twisted, while they were kept smooth and whole. One popular
topological pastime is to try to visualize what a bicycle tube would look like
turned inside out. As far as I know, these curiosities have no serious
implications for mathematics, but they can be used to cultivate flexibility
in visual thinking.
Let us start by moving the surface of Figure IA to that of Figure IB as
an example. The reader is asked to visualize, or draw, a sequence of
FIGURE IA FIGURE IB
pictures moving Figure IA to Figure IB, without cutting, and without

letting one part of the surface touch another part. Figure I C is a possible
sequence.
@
/
FIGURE IC
The next exercise involves cutting a temporary hole in the surface.

Instead of merely turning a bicycle tube inside out, how about turning a
more complicated surface inside out? It will soon be apparent that the
mildly complicated Figure 2A could become very complicated, so, to
simplify things we paint the inside black, and the outside grey. We are
going to cut a small hole in the surface temporarily and put a rim on the
hole to keep track of it.
Rim
FIGURE 2A
Figure 2B shows a sequence for turning the surface inside out;

shrinking the complicated part, and passing it through the hole.
After the hole is closed up, the black will cover the whole outside,
whereas it formerly covered the whole inside. An advantage of the method
of turning surfaces inside out, just pictured, is evident in Figure 3. This
method makes it just as easy to see what happens to a sphere with many
handles when it is turned inside out as to see what happens to the bicycle
tube.
BICYCLE TUBES INSIDE OUT
2A
2
FIGURE 2B
3A 38
3D
3C
FIGURE 3
FIGURE 4
The last sequence will aim for a simple picture of what Figure 4 will
look like, after one of the two linked surfaces is turned inside out.
These pictorial ideas are not new. They occurred to me 25 years ago,
and perhaps to several people before that. The twister of Figure 5 was
posed recently by Dennis L. Johnson, who is well-versed in the theory of
knots. Now the reader is invited to finish up with a little light exercise,
moving from Figure 5A to Figure 5B in the same fashion as IA --+ IB was
done.
FIGURE 5A FIGURE 5B
Flexing Surfaces
-~-
Robert Connelly
CORNELL UNIVERSITY
Suppose a closed polyhedral surface is built from flat pieces of stiff

cardboard taped together along their edges. Will the surface flex? That is;
will it change its shape continuously without ripping the tape or bending
the cardboard? As an example, let us consider the octahedron shown in
Figure 1. If one builds this out of cardboard it turns out to be very rigid
and does not flex. However, if the top is slightly smaller than the bottom,
the top will pop down as shown in Figure 2. To do this, one must bend the
cardboard. Figure 1 will not continuously move into Figure 2 without
distortion.
FIGURE I FIGURE 2
~79~
In 1813 Cauchy, the well-known french mathematician, proved that

any convex polyhedral surface is rigid. (It is understood that the flat natural
faces are the pieces of cardboard and are held rigid). So it seems natural to
conjecture that all surfaces, convex or not, are rigid. Unfortunately, this
"rigidity conjecture" is false. There is an embedded polyhedral surface,
without self-intersections, that flexes. In what follows I will describe some
of the examples that I have found, and subsequent modifications by others
which refute this conjecture.
The Construction
To understand why the forthcoming surfaces flex, we describe some of the
flexible octahedra of R. Bricard, a French engineer who discovered them
in 1897. These surfaces do have self-intersections so we regard them as a
collection of incompressible, in extendible rods connected by flexible
rubber nodes at their endpoints. To build these octahedral frameworks we
start with a skew quadrilateral aba'b' as in Figure 3, with opposite sides of
the same length. It turns out that there is then a line L in 3-space such that
the quadrilateral is symmetric about L. That is, if the quadrilateral is
rotated 180 0 about L, it is rotated into itself. We think of aba'b' as the
a a'
FIGURE 3
equator of the octahedral framework. Choose a point c not on L; the line of

symmetry. Join c to each of the nodes a,b,a',b' by a rod. It is not hard to
check that this framework flexes as it is. So flex it and join it at each instant
to the congruent framework c'(a'b'ab) obtained from the first by rotation by
180 0 about L. Note that c is rotated into c'. The union is one of the flexible
octahedra of Bricard and is easy to build. See Figure 4 for the completed
framework. Note that if all the triangles are "filled in" the resulting surface
will have many self-intersections. Our goal is to reduce and simplify these
self-intersections as much as possible.
FLEXING SURFACES
c'
FIGURE 4
In Figure 5 we see another version where all the rods start out lying
flat in a plane. The line of symmetry L is perpendicular to this plane and as
the framework flexes, the vertices do not remain in a plane, but this is a
very convenient position for starting.
Another slight variation on this framework is to start with the points a
and a' in a horizontal plane H as in Figure 5. Then choose points b,b' at a
height I> > 0 above Hand e,e' at height b > I> above H so that all the points
project orthogonally onto the picture of Figure 5. Line L is again
perpendicular to H, the octahedral framework is still flexible, and the
boundaries of the triangles ab'e and a'be' link; that is, they cannot be pulled
apart without breaking.
To construct the embedded flexible surface, we start with the surface
that is used for Figure 5. Instead of filling in all of the triangles with flat
planar pieces, we change the surface somewhat, still keeping the rods of the
old framework as edges in our surface. We regard the octahedral surface as
being made of two pieces~a bottom and a top. Let us say that the bottom
b a
FIGURE 5
~81&
FIGURE 6
FIGURE 7
FIGURE 8
FIGURE 9
FIGURE 10
FLEXING SURFACES
is exemplified in Figure 6. We push down on each of the triangular faces to

get a new surface that looks like Figure 7, in which each triangle is replaced
with an upsidedown bottomless tetrahedron-or, a pit.
Similarly we replace the top surface (Figure 8) with the surface in
Figure 9, where each triangle is replaced by a bottomless pyramid. Just as
the surfaces of Figures 6 and 8 flex, so do the surfaces of Figures 7 and 9,
with the extra vertices; the apexes of the pyramids move rigidly with
respect to their bases.
We next glue the surfaces of Figures 7 and 9 together along their
common boundary to get the surface of Figure 10. This surface is flexible,
just like the surface of Figure 5, but unfortunately, it has a couple of self-
interactions-s and s'. Figure II shows the parts of the surface of Figure 10
that intersect: points sand s' correspond to the crossing points of Figure 5.
In order to get rid of sand s' we build what I call a crinkle, which is
again, based on the Bricard flexible octahedra. Choose a planar quadri-
lateral defg with opposite sides equal, de = jg, ef = gd as in Figure 12, with
b'
FIGURE 11
I
I
I
g I
I
I
I
I
I
e
,
I
d I
I
I
L'I f
I
I
I
V
FIGURE 12
~83~
FI RE 13
Glue
Mountain fold
Valley fold
FIGURE 14
fl RE 15
FIGURE 16
FLEXING SURFACES
the segment de intersectingfg. Choose a point h directly over the center of

the circle through difg and h' -the same distance under the center. Thus
hd = he = !if = hg = h'd = h'e = hJ = h'g. Then the frameworks h(difg)
and h'(difg) flex in conjunction. (The quadrilateral difg actually remains
coplanar.) The union of the triangular faces, hif, hfg, hgd, h'if, hJg, h'gd is
the crinkle; an octahedron with two triangular faces removed (see
Figure 13), with boundary hdh'e. The distance from d to e remains fixed
during the flex. Figure 14 shows how to build a crinkle.
To construct the final embedded flexible surface, take the surface of
Figure 10 and cut out, as shown in Figure 15, one small quadrilateral hole
around each of the self-intersection points. Then insert a crinkle of the
appropriate size into each of the holes. If the crinkle is positioned properly,
there will be no self-intersections in the resulting surface (Figure 16). Since
de remains at a fixed position in the crinkle, the surface with the crinkle and
the two holes removed flex in conjunction. Thus, the whole crinkled surface
flexes. I t looks something like Figure 17.
This surface was one of the first flexible surfaces I found, and I paid
no attention to the simplicity of construction or how few vertices would be
needed. Subsequently N. H. Kuiper and Pierre Deligne modified my
construction to get a surface with 11 vertices and 18 faces. They started
with the framework described in the text following Figure 5. Instead of
adding four pits to the bottom surface, they added only one (as in
Figure 18). The other three triangles they kept flat. For the upper surface
FIGURE 17
FIGURE 18
FIGURE 19
FIGURE 20
FLEXING SURFACES
FIGURE 21
they only added two mountains (as in Figure 19-shown with two views).
When these new upper and lower surfaces were glued together along their
common boundary, the line segment c'b intersected the sides of the two
mountains above ca. Due to the slight raising of c,c',b,b' this is the only
place where the surface intersected itself. They then removed ca and the
inside of the two triangles with ca as an edge and placed a carefully
proportioned crinkle, as in Figure 13, in the hole that was created. Points
d and e in Figure 13 fit into c and a respectively, and the apexes of the two
mountains are hand h'. Figure 20 shows two views of the upper and lower
surfaces glued together with ca removed from the upper surface. Figure 21
shows two views of the final flexible surface with the crinkle added.
To top this, Klaus Steffen found a flexible surface with only 9 vertices.
He started with two identical crinkles like those in Figure 14. They were
joined with two other triangles as in Figure 22, where the figure has a
symmetry about a vertical line so that the corresponding lengths are equal.
The result was a flexible surface, something like Figure 23.
,
I
,
I I
\ /
/
/
~----------------------------~/
_ _ _ Valley fold
_ ._ ._ Mountain fold
FIGURE 22
FIGURE 23
FLEXING SURFACES
SOllle Conjectures
An interesting property of the previous examples is that as they flex, the
volume enclosed by these surfaces remains constant. I do not see however,
how to prove that the volume is constant for every possible flexible surface.
CONJECTURE I If a triangulated polyhedral surface flexes, its

volume remains constant during the flex.
Even more surprising things seem to be true. Let P and r be two 3-

dimensional polyhedra in 3-space. We say P is equivalent to r by
dissection, and we write P ,. . ., r, if we can dissect P into a finite number of
polyhedral pieces, P 1 ,P2 , •.. ,Pk , and then reassemble them to get r. (So P
= PI U .. , U Pkl Pi n P j c (boundary P;) n (boundary P j ), for i =f=. j, r
= Pi u ... U Pic, Pi n Pj c (boundary Pi) n (boundary Pj), for i =f=. j, and Pi
is congruent to Pi for i = I, ... ,k.) M. Dehn's solution to the problem posed
by Hilbert resulted in the finding that the regular cube and tetrahedron of
the same volume are not equivalent by dissection. (See page 35 of Martin
Gardner's Second Book if Mathematical Puzzles and Diversions.) However,
suppose that Pt is the 3-dimensional solid enclosed by one of the above
flexible surfaces at time t. A result of]. P. Sydler implies that Po""'" Pt for all t
in the flexing interval. A good discussion of Hilbert's third problem and this
(non-trivial) result ofSydler can be found in the recently translated book of
Boltianskii (Boltianskii, V. 1978. Hilbert's Third Problem. New York: John
Wiley and Sons.) Still the general question remains.
CONJECTURE 2 If Pt is the polyhedral solid enclosed at time

t by any flexing polyhedral surface, then
Po""'" Pp for all t.
Even to see the specific dissections for the surfaces described above
would be interesting.
This material is based upon work partially supported by the National Science
Foundation under Grant No. MCS-7902521.
Planting Trees
-~-
Stefan Burr
CITY UNIVERSITY OF NEW YORK.
Your aid I want, nine trees to plant

In rows just half a score;
And let there be in each row three.
Solve this: I ask no more.
In 1821, John Jackson published this mathematical conundrum in a book

of problems called Rational Amusement for Winter Evenings [4]. These days,
verse is not as popular, and a modern-day puzzle poser might even dispense
with the trees, saying: Arrange nine points on a plane so that there occur
ten rows of three points. When a mathematician encounters such a
problem, he feels a natural urge to generalize it and then wants to make it
more precise. This leads to the following version: Given a positive integer
p, how can p points (p 2 3) be arranged on a plane, no four in a straight
line, so that the number of straight lines with three points on them is
maximized? We will call this maximal number oflines l(p).
The formidable mathematicianJ.J. Sylvester pursued the elusive 1(P)
in the nineteenth century, and it has attracted sporadic attention since,
from amateurs as well as professionals. In the past, amateurs have often
made valuable contributions to mathematical problems like this. It is a
shame that the amateur mathematician seems to be a dying breed today.
This is partly due to the inaccessibility of a lot of mathematics. There are
however many areas, especially those related to combinatorics, where an
amateur can work. Unfortunately, the public doesn't hear much about
these accessible questions. Generally, they hear about the glamorous
problems like Fermat's Last Theorem, where even a professional has little
~90~
PLANTING TREES
•
• • •
• • •
p =3 1 =1 p =4 1 =1 p =5 1 =2
(a) (b) (e)
p
A =6
(d)
1 =4 p=7
(e)
1 =6 p =8
(0
1 =7
p=91=10 p = 10 1 = 12 p = 11 1 = 16
(g) (h) (i)
FIGURE 1
chance of making major progress. One of the charms of combinatorial

geometry then, (of which this orchard problem is an example) is the fact
that amateurs can often make substantial contributions to it.
Figure 1 illustrates some orchards corresponding to p = 3,4, ... ,11.
All of these are optimal solutions, achieving 1(P) lines. Only two other
values of 1(P) are known exactly-we will get to these shortly. But for now,
note that four points are no better than three, and that sometimes one
orchard is contained in another, such as for p = 10 and 11.
A natural question to ask is whether the above arrangements are
unique solutions. The answer is no; Figure 2 shows a different arrangement
FIGURE 2
..691~
for p = 8, derived from the one for p = 7; of course the additional point
can go anywhere on the new line.
However, there is another way in which these arrangements are not
unique: they can be projectively transformed. We won't attempt to explain
completely what a projective transformation (also called a projection) is,
but one type can be literally visualized. Tilt the page and look at the
diagrams on the slant: the distances and angles change with the change in
perspective, but straight lines still appear straight, so that the transformed
arrangement still has the properties we want.
One problem arises from doing this: As with railroad tracks, parallel
lines can become non-parallel, changing the nature of an arrangement,
and vice versa. However, mathematicians were perhaps the inventors of
the now-fashionable idea of turning a problem into an opportunity, and
long ago they did so for projective transformations. They took the ordinary
plane and added an imaginary line at irifinity consisting of points at irifinity. A
set of parallel lines were considered to intersect at some point on the line at
infinity. Such a point was considered the same if one went in the
diametrically opposite direction, so that a set of parallel lines intersected in
just one point. In fact it was now true that any two lines met in exactly one
point. This created the so-called Euclidean projective plane and projective
geometry, which has proved itself to be a fruitful source of interesting
mathematics.
In our case, putting points at infinity is very useful for simplifying and
making more symmetrical some of the more complicated diagrams. It was
also useful for finding some of them in the first place, and for proving some
of the results we will refer to. As an example of putting points at infinity, let
us put the points of the top line in the arrangement for p = 9 at infinity;
this leads to Figure 3.
The arrows labeled a, b, and c indicate the directions of the three
points at infinity in the arrangement; any of the arrows could just as well
have pointed in the opposite direction. Of course, the line at infinity is to be
considered one line of the arrangement.
Figure 4 shows the only other orchards that are known to be optimal,
namely p = 12 and p = 16. Note that for p = 16 the orchard contains an
optimal 7-point orchard. Each arrangement contains three points at
.~ ~ .
FIGURE 3
PLANTING TREES
p = 12 I = 19 p = 16 I = 37
(4a) (4b)
FIGURE 4A FIGURE 4B
infinity and the line at infinity. These diagrams can be projected in such a
way that the points and line at infinity in each become real points, but the
symmetry would be lost, and they would be hard to draw on a small piece
of paper.
What about other values of p? Table 1 tells what has been learned
about 1(P) for p = 3,4, ... ,25, giving the best-known lower and upper
bounds. The twelve cases in which 1(P) is known exactly are indicated with
an asterisk, and for them the upper bound is dispensed with. Table 1, and
in fact almost all of the results here, are taken from a paper by myself, B.
Griinbaum and N. J. A. Sloane, entitled "The Orchard Problem" [1],
although many of those were taken, in turn, from previous work.
From p = 20 on, all the bounds in Table 1 are the result of two
general theorems, given below.
THEOREM 1 1(P) ~ LP(P - 3)/6J + 1, where LxJ denotes the

largest integer ~x.
THEOREM 2 Ifp ~ 4, then 1(P) ~ L(P(P -1)/2 - f3p/7l)/3J,

r
where x1denotes the smallest integer ~x. (Of
course, 1(3) = 1.)
We will not prove Theorem 1, but we will give some indication how it
comes out of the theory of cubic curves, that is, curves satisfying an algebraic
equation of degree 3. Figure 5 shows the symmetric cubic curve defined by
the equation (x - 1) ((x + 2) 2 - 3y2) = 8, together with twelve points on it,
including three points at infinity. These twelve points are arranged in the
same way in the twelve-point orchard in Figure 4a.
p Lower bound Upper bound

for 1(P) for 1(P)
3 1*
4 1*
5 2*
6 4*
7 6*
8 7*
9 10*
10 12*
11 16*
12 19*
13 22 24
14 26 27
15 31 32
16 37*
17 40 42
18 46 48
19 52 54
20 57 60
21 64 67
22 70 73
23 77 81
24 85 88
25 92 96
TABLE 1
Why does the use of a cubic curve produce good arrangements of

points? The secret lies in the fact that certain cubic curves can be given a
parametric representation, based on the so-called Weierstrass ellipticfunctions.
This representation gives every point on the curve a real number of at least 0
and less than 360, in such a way that three points are in a line if, and only if,
the three corresponding numbers add up to a multiple of 360. (We could
make the magic number anything we liked, but 360 is convenient, and lets
us think of the numbers as being something like angles.) Figure 6 shows the
curve of Figure 5, but with most lines removed and the numbers of the
points given by the parametric representation inserted. It is easy to check
that each straight line in Figure 5 satisfies the above criterion (including
the line at infinity).
It is not hard to work out, for any p, how to choose p numbers of at
least 0 and less than 360 in such a way that three of them add up to a
PLANTING TREES
I
I
I t·O
~I
~
300
330
ISO
/.120
270 I
------~ -----
180
1~
(x - 1)(x + 2)2 - 3Y1 = 8 I
I
FIGURE 5 FIGURE 6
multiple of 360 as often as possible. Such a choice leads to an orchard, and

therefore to a lower bound on l(p). The resulting lower bound is that of
Theorem 1. In 1868, Sylvester [6] proved a lower bound that is the same as
in Theorem 1 except when p is a multiple of 3. When this is the case,
Theorem 1 is better by one. (It is always very satisfying to improve on the
work of someone who is as brilliant as Sylvester-even by a bit!)
Theorem 1 accounts for every lower bound in Table 1, except for p
= 7, 11, 16, and 19. In each of these cases, it turns ou t to be possible to use an
arrangement of p - 1 points on a specially-chosen cubic curve, and then to
add one point not on the curve to get a better orchard than that associated
with Theorem 1. As an example, Figure 7 shows the construction of the 16-
point orchard of Figure 4b. Note that the extra point 0 is not O-since it is
not on the cubic, it has no number. To make the figure easier to see, only the
lines through 0 have been given.
Now let's turn to upper bounds for 1(P), including Theorem 2. Here
the approach is quite different. A useful tool is what we will call the graph of
an orchard. Draw the points of an orchard, and then connect two points
with a line segment (called an edge) if they are not in line with a third
FIGURE 7
. \-l .
FIGURE 8
point. For example, Figure 8 shows the graphs corresponding to the 7-

point and 8-point orchards of Figure I and the alternative 8-point orchard
of Figure 2. The fundamental difference between the two 8-point orchards
shows up very clearly in their graphs; for instance, one contains a triangle
and the other doesn't.
Suppose that a point in an orchard has k lines of three going through
it. Then, in the corresponding graph, the point is on P - 1 - 2k edges, since
the k lines eliminate 2k points that would have otherwise been adjacent to
the point in the graph. (We call the number of such edges the degree of the
point.) Because of this fact, it must be that if P is even, every point in the
graph must have odd degree, and if p is odd, every point in the graph has
even degree (including zero), since p and p - 1 - 2k have opposite parity.
Furthermore, consider the number of edges in the graph of an
orchard. If an orchard has no three points in line, then every pair of points
gives an edge in the resulting graph, and it is not hard to see that such a
graph (called a complete graph) has p (p - 1) /2 edges. However, any time
three points are on a line in the orchard, three edges are eliminated from
the graph.
Therefore, if the orchard has I lines, the corresponding graph has
p(P - 1) /2 - 31 edges. Hence, if e is the number of edges, we get
e= p(P - 1) - 31
2
so
Since e :2: 0,
I ~ lP(P; 1) J
But if p is even, the degree of every point of the graph is odd, which means
the degree is at least one. For this to happen, it must be true that e :2: p/2, so
I ~ lP(P ~ 2) Jwhen p is even.
~96~
PLANTING TREES
With a little fiddling, the above inequalities for 1 can be combined

into the following:
This result is not as strong as Theorem 2. For that theorem, we need a

theorem of Kelly and Moser [5]. This says that, in any arrangement of p
points such that not all p are on one line, there must be at least r3p/71 pairs
of points which are not on a line with any third point. For an orchard with
at least four points, this means that in the corresponding graph, e 2 [3p/7],
so that by the previous paragraph,
which is essentially Theorem 2.

It will be noticed that this proof of Theorem 2 doesn't really use the
clever construction of a graph from an orchard very heavily. However, the
idea of the graph of an orchard is very useful in dealing with special cases.
In fact, in the only cases in which a better upper bound is known than that
given by Theorem 2, orchard-graphs have proved useful. These cases are
p = 8, la, 12, and 14; each time the upper bound (given in Table 1) is one
better than that given by Theorem 2. We will indicate a prooffor p = 8-
the others are similar but rather long.
To prove that 1(8) ~ 7 (which is the same as saying 1(8) = 7, since a
7-line orchard exists) we must show that no 8-point, 8-line orchard exists.
Assume, to the contrary, that such an orchard did exist. Consider its graph;
the number of edges e is given by
p(P-I) 8'7
e= - 31 = - - 3'8 = 4.
2 2
Furthermore, because 8 is even, every point in the graph has odd degree.
The only possible 4-edge graph satisfying this requirement must consist of
four separated edges, as shown in Figure 9. (In Figure 9 the graph is drawn
abstractly, showing the connections while ignoring the actual positions of
the points.)
IIII FIGURE 9
1 2 3
4 5 6
A-
• 7 8 9
FIGURE 10
Consider the points A and B in the graph. Turning to the correspond-

ing orchard, we can use a projective transformation to put them on the line
at infinity. By studying the graph, we see that the other six points in the
orchard must fall on three lines emanating from A, and also on three lines
emanating from B. The situation is shown somewhat schematically in
Figure 10. (The only way in which Figure 10 is schematic is the even
spacing of the lines.)
The six points (other than A and B) in the orchard must fall among
the nine intersection points shown in Figure 10. Moreover, we have
accounted for only six of the lines in the (hypothetical) orchard; the other
two lines must use the nine intersection points. The only way to add two
new lines is if they go through the points 1,5,9 and 3,5,7 -but this
accounts for only seven points, namely A, B, I, 3, 5, 7, and 9. (Further-
more, the lines A5 and B5 still have only two points on them.) Therefore,
the sought-for orchard is impossible, and 1(8) is indeed 7.
Although more could be said about this problem, and even more
about related problems, this discussion should convey the flavor of the
subject. We will close with some comments on what more might be done.
Obviously, the best thing of all would be to determine 1(p) exactly for all p.
This seems difficult, but may be possible. In [1], we conjectured that
1(P) = LP(P - 3)/6J + 1
for all p =1= 7, 11, 16, 19; in other words, that Theorem 1 tells almost the
whole story.
One could hope at least, to narrow the gap between Theorems 1 and
2. By ignoring the brackets in these theorems and subtracting, we see that
the gap is approximately 4/21 p - I for each p. The gap therefore grows
fairly slowly, as Table 1 confirms.
In any case, it would be very interesting to determine 1(P) or to
narrow the gap for some small p. Clearly the place to start is p = 13, 14, or
(possibly) 15. In view of the above conjecture, it would be best to try to
lower the upper bounds in Table 1. Besides, attacking the upper bounds
should not involve knowing a lot about cubic curves or the like, whereas
.,a98~
PLANTING TREES
attacking the lower bounds might. Such an attack probably would require
the pursuit of many lines of argument in great detail. One would start
out with the approach used in the case p = 8 given here, but many alter-
native subcases would have to be examined separately. Perhaps an artful
computer program would be useful in this project.
It has been the purpose of this paper to shed some light on an
interesting corner of combinatorial mathematics. I hope also to inspire
interested amateurs to try their hands at this or similar problems. (For
reading on some other questions in the field, see Branko Griinbaum's fine
book, Arrangements and Spreads [3].) It seems particularly appropriate to
include such problems in this book because the person who has done the
most to make mathematics popularly accessible today is Martin Gardner.
He has, in fact, discussed some aspects of this tree-planting problem in his
column [2]. Surely, planting trees is a fruitful activity for a mathematical
gard (e) ner.
References
1 Burr, S. A.; Griinbaum, B.; and Sloane, N. J. A. 1974. The Orchard
Problem. Geometriae Dedicata 2: 397-424.
2 Gardner, M. 1976. Mathematical Games. Scientific American, 102-109.
3 Griinbaum, B. 1972. Arrangements and Spreads. Providence, R.I.: Amer.
Math. Soc.
4 Jackson, J. 1821. Rational Amusement for Winter Evenings. London:
Longman, Hurst, Rees, Orme, and Brown.
5 Kelley, L. M., and Moser, W. O.J. 1958. On the Number of Ordinary
Lines Determined by n Points, Ganad. ]. Math. 10: 210-219.
6 Sylvester, J. J. 1886. Problem 2572. Math Questionsfrom the Educational
Times 45: 127-128.
~99~
Slicing it Thin
-~-
Howard Eves
UNIVERSITY OF MAINE
Over six hundred years ago, the Blessed John Colombini of Siena
founded a new religious order, originally devoted to nursing and burying
the victims of the rampant bubonic plague that swept away more than a
third of the population of Europe. The group, known as the Jesu'ats (in no
way related to the Jesuits-which at the time had not yet been founded),
was officially approved by Pope Urban V in 1367. As time went on, the
order declined, and in 1606 a partially successful attempt was made to
revive the group. But certain abuses, apparently involving the manu-
facture and sale of distilled liquors in a manner not sanctioned by Canon
Law, crept in. This, along with a difficulty in maintaining a reasonable
membership quota, led to the order's abolishment by Pope Clement IX in
1668. The order had existed for just over three hundred years.
I t was in 1613, shortly following the attempt to revive the J esuat
order, that a young Italian boy of fifteen, by the name of Bonaventura
Cavalieri, was confirmed as a new member. Cavalieri spent the rest of his
life in the ministry of the J esuat order. Because of this dedication, and
because of the subsequent dissolution of the order and the close similarity
between the title Jesuat and the much more familiar title Jesuit, many
major encyclopedias, biographical dictionaries, histories, and source books
of today erroneously state that Cavalieri was a Jesuit instead of a Jesuat.
We have here an excellent example of a perpetuated error hidden in our
written histories. Errors of this sort are not uncommon, some having
endured over long periods of time.
SLICING IT THIN
Bonaventura Cavalieri was born in Milan, Italy, in 1598. As a young

man he studied under Galileo. Later, in 1619, he received an appointment
as a professor of mathematics at the University of Bologna, serving in this
capacity until his death in 1647, at the relatively young age offorty-nine.
He was one of the most influential mathematicians of his time amd he
wrote a number of works on geometry, trigonometry, astronomy, astrol-
ogy, and optics. Among the first to recognize the great value of Napier's
invention oflogarithms, he played a leading role in their early introduction
into Italy. His most important contribution to mathematics, however, is
the treatise Geometria indivisibilibus, published in its initial version in 1635.
This work is devoted to the precalculus method of indivisibles, a method that,
like so many mathematical achievements of more modern times, can be
traced back to the early Greeks~in this case to Democritus (ca. 410 B.C.)
and Archimedes (ca. 287-212 B.c.)~though the direct stimulus probably
lay in attempts at integration that had been made by Johann Kepler
(1571-1630). In any event, the publication of Cavalieri's Geometria
indivisibilus in 1635 marks an important milepost in the history of
mathematics and in the history of the calculus in particular.
Cavalieri's great treatise is voluble and unclear, and it is difficult to
learn from it just what Cavalieri meant by an "indivisible." It seems that
an indivisible of a given planar piece is a chord of that piece, and the
planar piece can be regarded as made up of an infinite parallel set of such
indivisibles, or atomic parts. Similarly, it appears that an indivisible of a
given solid is a planar section of that solid, and the solid can be regarded as
made up of an infinite parallel set of these kinds of indivisibles, or atomic
parts. Cavalieri argued that if we slide each member of a parallel set of
indivisibles of some given planar piece along its own axis so that the end
points of the indivisibles still trace a continuous boundary, then the new
planar piece will have the same area as the original planar piece, since the
two planar pieces are made up of the same indivisibles. Similarly, if we
slide each member of a parallel set of indivisibles of some given solid along
its own plane, so that the boundaries of the indivisibles still form a
continuous surface, then the new solid will have the same volume as the
original solid, since the two solids are made up of the same indivisibles.
This last result can be strikingly illustrated by taking a vertical stack of
cards and then pushing the sides of the stack into curved surfaces; the
volume of the disarranged stack is the same as the volume of the original
vertical stack.
The above remarks, slightly generalized, yield the so-called Cavalieri
principles:
1 If two planar pieces are included between a pair of parallel

lines, and if the lengths of the two sections intercepted by them
on each line parallel to the including lines are always in a given
ratio, then the areas of the two planar pieces are also in this ratio.
2 If two solids are included between a pair of parallel planes,

and if the areas of the two sections intercepted by them on each
plane parallel to the including planes are always in a given ratio,
then the volumes of the two solids are also in this ratio.
Cavalieri's hazy conception of indivisibles as atomic parts making up

a figure spurred considerable discussion, and serious criticisms were leveled
by some students of the subject, particularly by the Swiss goldsmith and
mathematician Paul Guldin (1577-1642). Cavalieri accordingly recast his
treatment in the hope of meeting these objections, but his new attempt was
scarcely any more successful than his first. The French geometer and
physicist, Gilles Persone de Roberval (1602-1675), claimed to have
invented the method before Cavalieri, but this matter of precedence is
difficult to settle, for Roberval was consistently tardy in disclosing his
discoveries and so became embroiled in several arguments of priority.
Roberval's tardiness has been explained by the fact that for forty
years, beginning in 1634, he held a professorial chair at the College
Royale. This chair automatically became vacant every three years, to be
filled by open competition in mathematical contests in which the questions
were set by the outgoing incumbent. Naturally, to perpetuate his position,
Roberval saved his discoveries to be used for questions in the competition.
In any case, Roberval ably employed the method of indivisibles to the
discovery of a number of areas, volumes, and centroids. The method, or
some procedure very like it, was also effectively used by Evangelista
Torricelli (1608-1647), Blaise Pascal (1623-1662), Pierre de Fermat
(1601?-1665), Gregoire Saint-Vincent (1584-1667), Isaac Barrow
(1630-1677), and others.
Cavalieri's two principles constitute valuable tools in the computation of
areas and volumes, and they can easily be made rigorous with the use of
modern integral calculus. Accepting the principles as intuitively apparent,
many problems in mensuration can be solved that normally require the
more advanced technique of the calculus. Indeed, many authors of begin-
ning textbooks on solid geometry have, on pedagogical grounds, advocated
the assumption and then the consistent use of Cavalieri's second principle,
as this greatly simplifies the subject for the student. For example, in
deriving the familiar formula for the volume of a tetrahedron (V = Bh/3),
the sticky part is first to show that any two tetrahedra having equivalent
bases and equal altitudes on those bases have equal volumes. The inherent
difficulty here is reflected in all treatments of solid geometry from Euclid's
Elements on. With Cavalieri's second principle, however, the difficulty
simply melts away.
We now turn to some illustrations of the use of Cavalieri's principles.
Some of the examples constitute teasing puzzles for those who enjoy
challenging mathematical diversions. We first formulate a convenient
definition: Two planar pieces that can be placed so that they intercept
sections of equal length on each member of a family of parallel lines, or
~102&
SLICING IT THIN
two solids that can be placed so that they intercept sections of equal area on
each member of a family of parallel planes, are said to be Cavalieri congruent.
It follows from Cavalieri's principles that two figures that are Cavalieri
congruent have equal areas (in the one case) or equal volumes (in the other
case).
Dlustrations
1 Let us find the area of an ellipse of semiaxes a and b. Consider the
ellipse
x2 y2
-2 + -2 = I, a > b,
a b
and the circle
x2 + y2 = a2,
plotted on the same rectangular coordinate frame of reference, as shown in
Figure l. Solving each of the above equations for y, we find, respectively,
y = (~) (a 2 _ X2)1/2, Y = (a 2 _ X2)1/2.
It follows that corresponding ordinates of the ellipse and the circle are in
the ratio b/a. The corresponding vertical chords of the ellipse and the circle
are also in this ratio, and thus, by Cavalieri's first principle, so are the areas
of the ellipse and the circle. We conclude that
area of ellipse = (~) (area of circle)
This is essentially the procedure Kepler employed in finding the area of an

ellipse of semiaxes a and b.
x'+r=a'
FIGURE I
~103~
FIGURE 2
2 Now let us find the familiar formula for the volume of a sphere of
radius r. In Figure 2 we have a hemisphere of radius r on the left, and on
the right, a circular cylinder of radius r and altitude r with a cone removed
whose base is the upper base of the cylinder and whose vertex is the center
of the lower base of the cylinder. The hemisphere and the gouged-out
cylinder are resting on a common horizontal plane. We now cut both solids
by a plane parallel to the base plane and at a distance h above it. This
plane cuts the one solid in a circular section and the other in an annular, or
ring-shaped, section. By elementary geometry one can easily show that
each of the two sections has an area equal to n(r2 - h2). It follows, by
Cavalieri's second principle, that the two solids have equal volumes.
Therefore the volume of the sphere is given by
v = 2 (volume of cylinder-volume of cone)

= 2 (nr 3-"31 nr 3) = "34 nr3.
The whole trick here, of course, was to find a "comparison solid" (in this
case, a gouged-out cylinder) that is Cavalieri congruent to a hemisphere.
3 As a second illustration of the planar case of Cavalieri's principles,
consider the planar piece, bounded by a straight line and two curved arcs,
pictured in the left of Figure 3, wherein the two distances marked mare
equal to one another. One readily obtains, for a comparison area, the
FIGURE 3
.8104~
SLICING IT THIN
FIGURE 4
planar piece made up of a semicircle and an isosceles right triangle

pictured in the right of Figure 3. It follows that the required area is
4 For another illustration of the solid case of Cavalieri's principles, let

us find the volume of the spherical ring obtained by removing from a solid
sphere of radius r a cylindrical boring of radius a coaxal with the polar axis
of the sphere (pictured in the left of Figure 4). Consider a sphere of
diameter equal to the altitude of the spherical ring and placed with its
center on the same horizontal plane that contains the center of the
spherical ring (pictured in the right of Figure 4). Now cut the two solids by
a horizontal plane at distance h from the centers of the two solids. In the
spherical ring we get an annular section of area
n(r2 - h2) - na 2 = n(r2 - a2 _ h2).
In the sphere we get a circular section of area
n(k 2 - h2) = n(r2 _ a2 - h2).
We conclude, by Cavalieri's second principle, that the volume V of the
spherical ring is the same as the volume of a sphere of radius k. That is
4
V = -nk
3
3
.
It is interesting that all spherical rings of the same altitude have the same volume,
irrespective of the radii of the rings.
5 In Problem E 465, The American Mathematical Monthly, March 1941,
one is asked to find the area enclosed by the curve
b2y2 = (b + X)2 (a 2 - x2),
where b;;::: a > 0. The real graph of the curve (see Figure 5) consists of a
noncircular ring passing through (0, ±a) and (±a,O), along with an
~105&
(-b,O) (-a,O) (a,O) x
FIGURE 5
isolated point at (-b,O). The curve is symmetrical in the x-axis. Consider

the circle x 2 + y2 = a2. We shall show that the required area is equal to
the area of this circle. Because of the symmetry of the figure in the x-axis, all
that we have to do to accomplish our aim is to show that the crescent-
shaped area between the curve and the circle in the second quadrant is
equal to the crescent-shaped area in the first quadrant. This we accomplish
by showing that the intercepted ordinates in the two crescents at equal
distances from the origin are equal to one another. To this end, letYl andY2
be the corresponding positive ordinates of the given curve and the circle for
the same x. Then we have
Yl - Y2 = b
I (b + x) J2
a - x 2 - J2 XJ2
a - x2 = b a - x 2.
We see that, except for sign'Yl - Y2 has the same value for + x and -x. It
now follows that the two crescent-shaped areas are Cavalieri congruent,
and the area enclosed by the given curve is equal to na 2 •
Puzzles
Designing a proper comparison solid to assist in finding a required volume
by Cavalieri's second principle can sometimes constitute a very enjoyable
puzzle. Perhaps there is a reader who might like to try his hand at some
puzzles of this sort. Here are a few; hints and suggestions appear at the end
of the paper.
6 Find, by Cavalieri's second principle, the volume of a torus, or anchor
ring, formed by revolving a circle of radius r about a line in the plane of the
circle at distance c ~ r from the center of the circle.
7 It is not difficult to show that there is no polygon to which a given
circle is Cavalieri congruent (because equally spaced chords between two
sides of a polygon change length uniformly, whereas equally spaced chords
~106~
SLICING IT THIN
FIGURE 6
in a circle do not). One might also think that there is no polyhedron to

which a given sphere is Cavalieri congruent. Show that this is not so by
obtaining a tetrahedron that may serve as a comparison solid for a sphere.
8 An oblique plane through the center of the base of a right circular
cylinder cuts off from the cylinder a cylindrical wedge, called a horif (see
Figure 6). Find, by Cavalieri's second principle, the volume of a hoof in
terms of the radius r of the associated cylinder and the altitude h of the hoof.
9 A generalized prismoid is any solid having two parallel base planes and
having the areas of its sections parallel to the base planes given by a
quadratic function of their distances from one base.
a Show that the volumes of a prism, a wedge (a right triangular

prism turned so as to rest on one of its lateral faces as abase),
and a pyramid are given by the prismoidal formula
h(U+ 4M+ L)
V= 6 '
where h is the altitude, and U, L, and M are the areas of the

upper and lower bases and midsection, respectively.
b Show, by Cavalieri's second principle, that the volume of a
generalized prismoid is given by the prismoidal formula.
e Show that (I) a sphere, (2) an ellipsoid, (3) a horif (see
Puzzle 8 above) , and (4) a Steinmetz solid (the solid
common to two right circular cylinders of equal radii and
having their axes intersecting perpendicularly) are all ex-
amples of a generalized prismoid, and thus find expressions
for their volumes.
* * * * * *
We conclude with a warnmg and with a surprIsmg property of

Cavalieri congruence. First, the warning.
In elementary textbooks the Cavalieri principles are accepted as
intuitively evident. Now consider two solids included between a pair of
parallel planes and having the perimeters of the two sections cut by them
on any plane parallel to the including planes always equal in length. One
may feel, on intuitive grounds, as in the case of the Cavalieri principles,
that the lateral areas of the two solids must then be equal. One would be
led to this conclusion by thinking of the lateral surfaces of the two solids as
each being made up of a number of atomic elements in the form ofloops of
very thin string (forming perimeters of section of the solids). Since
corresponding loops are of the same length and there is a one-to-one
relation between the loops on one solid with those on the other, it would
seem that the lateral areas of the two solids must be equal. To show that
r
this is not necessarily so, consider two square prisms P and of equal bases
r
and altitudes, where P is a right prism and is an oblique prism having a
pair of opposite lateral faces perpendicular to the base (see Figure 7). The
lateral area of the oblique prism is greater than the lateral area of the right
prism. This points out the danger of relying, in mathematics, upon one's
intuition.
N ow let us consider the surprising property of Cavalieri congruence.
We first formulate a definition. Two triangles ABC and A'B'C' lying in the
same plane are said to be affine reflections of one another if AA', BB', CC' are
parallel with their midpoints on some line m (see Figure 8). The two
following related theorems can be established.
'fiIEOREM Any two coplanar triangles of equal area can be

carried one onto the other by not more than three
affine reflections.
FIGURE 7
SLICING IT THIN
C t'-------f--;---+----~ C'
FIGURE 8
THEOREM Any two triangles of equal area are Cavalieri

congruent.
The above two surprising theorems have only recently been discov-
ered. The analogue in three-space of the second theorem is not true. That
is; two tetrahedra of equal volume are not necessarily Cavalieri congruent.
We shall not undertake here the task of establishing these results.
Hints and Suggestions

for the Puzzles
6 Place the torus on a plane p perpendicular to the
axis of the torus.
Take for a comparison solid a right circular cylinder of radius rand
altitude 2nc, and place it lengthwise on the plane p. Cut the torus and the
cylinder by a plane parallel to p. The section A in the torus is an annular
region of outer and inner radii a and b, say, and the section A' in the
cylinder is a rectangle of length 2nc and width w. Now
A = na 2 - nb 2 = n(a 2 - b2 ) = n(a + b)(a - b) = 2nc(a - b)

and
A' = 2ncw = 2nc(a - b).
Since A = A', it follows that the volume of the torus is equal to the volume
of the cylinder. That is
~109&
7 Let AB and CD be two line segments in space such that (I) AB = CD

= 2ry0i, (2) AB and CD are each perpendicular to the line joining their
midpoints, and each are distance 2r apart, (3) AB is perpendicular to CD.
The tetrahedron ABCD may serve as the comparison polyhedron.
8 Divide the hoof into two equal parts by a plane p through the axis of
the cylinder and let A be the area of the resulting cross-section of the hoof.
Construct a right prism having as its base a square of area A (the base lying
in the plane p) and having an altitude equal to the radius r of the cylinder.
Cut from this prism a pyramid whose base is the base of the prism not
lying in p, and whose vertex is a point in the other base of the prism. This
gouged-out prism may serve as a comparison solid for one of the halves of
the hoof.
Or one may choose for a comparison solid for half the hoof, a solid
whose base is a right triangle oflegs rand h, whose altitude is r, and whose
upper base is a line parallel and equal to the hypotenuse of the lower base.
The volume of the hoof is 2hr 2 /3.
9b Any section, which is a quadratic function of the distance from one
base, is equal to the algebraic sum of a constant section area of a prism, a
section area (proportional to the distance from the base) of a wedge, and a
section area (proportional to the square of the distance from the base) of a
pyramid. Thus the generalized prismoid is equal to the algebraic sum of the
volumes of a prism, a wedge, and a pyramid. Now apply part (a).
9c The section of the ellipsoid
x2 y2 Z2
-
a2
+b-2 +c- -
2 -
formed by the plane at distance z from the xy-plane is the ellipse

x2 y2 Z2
2+2=
a b
having semiaxes
The area of this ellipse is
showing that the ellipsoid is a generalized prismoid. We find
V = 4nabc.
3
~llO~
SLICING IT THIN
For the Steinmetz solid we have V = 16r3 /3. There is a story that says
the electrical genius, Charles Proteus Steinmetz (1865-1923), was once
presented with the problem of finding the volume of the solid now named
after him. He astounded everyone by immediately giving the answer,
without any use of paper and pencil. Asked how he arrived at the result so
quickly, he declined to tell. It is believed that he recognized the solid as an
example of a generalized prismoid, and then used the prismoidal formula.
Adapted from one of the lectures in the author's lecture sequence entitled Great
Moments in Mathematics.
.alII&.
Ho~ Did Pappus Do It?
-~-
Leon Bankoff'
MATHEMATICIAN AT LARGE
One of the chief joys of mathematics is problem-solving and the greatest

thrill of problem-solving is the attainment of an elegant result-preferably
by elegant means. Pappus was a master of the elegant theorem, as
evidenced by the countless "theorems of Pappus" found in the mathe-
matical literature, and it comes as no surprise that he was successful in
finding an ingenious way of proving the truth of what he called "an
ancient theorem". This is a proposition described in Book IV of his
Mathematical Collection and has to do with a sequence of consecutively
tangent circles inscribed in the Arbelos, or the Shoemaker's Knife, a
geometrical figure first treated by Archimedes in his Book of Lemmas. We
may safely presume that this Circle Theorem was probably known
empirically in times gone by, but had never before been proved.
If an arbitrary point C is chosen anywhere on a line segment AB and if
semicircles are described on the same side of AC, CB and AB as diameters,
the space within the curvilinear triangle bounded by the three semicircular
arcs is called an Arbelos. One of the three properties of the Shoemaker's
Knife considered by Archimedes pertained to the calculation of the
diameter of an inscribed circle touching each of the three given arcs.
Pappus started with this inscribed circle and extended his arena of
operations by following it with a series of successively tangent circles, each
touching two of the arcs of the Arbelos, as shown in Figure 1. According to
this "ancient theorem" the distance from the center of the first inscribed
circle to the line AB is equal to twice its radius; the distance from the center
~112~
How DID PAPPUS Do IT?
FIGURE 1
of the second inscribed circle to AB is four times its radius-and for the n-th
circle, 2n times its radius.
In the related configuration shown in Figure 2, the arc CB is omitted
and the first inscribed circle touches arcs AC and AB and the line segment
CB. Again, a chain of successively tangent circles is bounded by the arcs of
the mixtilinear triangle (a triangle which may have some curved sides) and
now the distance from the center of the n-th circle of the chain to the
baseline AB is equal to 2n - I times the corresponding radius. For
example, the distance from the center of the seventh circle of the chain to
the line AB is thirteen times its radius.
At first glance the proof of these properties appears to be quite simple.
We twentieth century mathematicians, smug in our possession of sophisti-
cated tools, look upon a problem of this sort as rather routine. All we have
to do is to apply principles of Inversion to the basic configuration and let
the inverted figure stare back at us, stripping the veil of mystery from the
original figure and revealing the naked truth of the Pappus Circle
Theorem.
For the benefit of those readers not well-acquainted with Inversion, a
brief review of the procedure may be in order. Erect a perpendicular to AB
at C, cutting the circumference of arc AB in D. Then with A as center and
with AD as radius, describe an arc representing a portion of the circle of
FIGURE 2
.a113~
A c B
FIGURE 3
inversion. Since AD2 = AG' AB, the arc AB becomes the extended line GD
perpendicular to AB at G; the arc AG inverts to the line tangent to the arc
AB at B. Since the circle of inversion cuts the arc GB orthogonally, the
semicircle on GB is self-inverse and remains unchanged. As for the Pappus
chain of circles, they are transformed into the chain of equal tangent circles
bounded by the two vertical parallel lines, just as their original counter-
parts were bounded by the arcs AB and AG. In Figure 3, it is immediately
apparent that the distance of the center of the n-th inverted circle from the
baseline is 2n times its radius. A similar inversion applied to Figure 2 yields
the result Yn = 2n - I, where Yn is the height of the center of the n-th circle
above the baseline.
Now let us return to the topical question: "How did Pappus do it?"
After all, he had to work without the assistance of Inversion (that was to
come fifteen centuries later). There was nothing in Euclid, Archimedes or
Apollonius to serve as a clue for the attainment of his astonishing result. So
Pappus prepared for his attack on the main problem by first proving
several pertinent introductory propositions or lemmas.
We commend Pappus for his remarkable achievement and we are
grateful to him for bequeathing us his complicated, yet rigorous proof. We
appreciate the difficulties he encountered by having to forge his own tools
for the solutions of his problems. But if we apply present day standards of
evaluation and criticism, we are compelled to conclude that the proof
offered by Pappus is not one we would want to foist upon a trained
mathematician, let alone a bright, eager high school student. This becomes
evident if we pursue our natural curiosity about the Pappus proof to the
point of actually locating it and wading through its intricacies. But where
are we to find it? Although sixteen centuries have gone by since the days of
Pappus, no-one has ever taken the trouble to do an English translation of
the Collection in its entirety. However, an abbreviated version of a small
~114~
FIGURE 4
part of the works of Pappus was undertaken by Sir Thomas Heath and
among the topics covered are the Circle Theorem and its associated
lemmas. Heath's scholarly treatment was published in his History of Greek
Mathematics by the Oxford University Press in 1921. However, in his con-
densed Manual of Greek Mathematics (Oxford, 1931) he merely mentions
the theorem but deftly bypasses any explanation of the method of solution.
Other works on Greek mathematics, such as those by James Gow and Ivor
Thomas, assiduously avoid tackling the lemmas and the main proof.
My personal library boasts a copy of the Collection in Latin,
translated from the Greek by Commandino in 1659. To do justice to
Pappus's proof one would have to plow through eleven legal-sized pages of
densely packed small print to try to ferret out what Pappus had to say. On
the other hand, there does exist a two-volume edition in French, faithfully
translated and meticulously edited by Paul Ver Eecke and published in
Paris in 1933 by Desclee de Brouwer et Cie. Nineteen pages of this edition
are devoted to the description of how Pappus did it. In his footnotes, Ver
Eecke frequently cites a German edition by Frederic Hultsch, published in
Berlin in 1876 and consisting of three volumes.
Having travelled thus far, the reader is entitled to at least a summary
of the approach undertaken by Pappus. The following outline will consist
of brief comments on the four diagrams designated here as Figures 4 to 7
inclusive. In Figure 4, illustrating the first lemma, the self-explanatory
diagram serves to show that KE· EL = EB2, a relation to be used later on.
The modern equivalent of this result would be the establishment of Land K
as anti-homologous points with respect to E.
In Figure 5 the circle FGH centered at A is any circle touching the
B K DM L c
FIGURE 5
B D M
FIGURE 6
arcs of the semicircles erected on the diameters BD and BC, with D lying on
Be. Starting with the diameter HF parallel to BC, the rest of the figure is
constructed as shown. Then by setting up ratios BCIBG and BFIBL in the
similar triangles BGC, BKH and BLF, BED, Pappus arrives at the ratio
2BMIKL = (BC + BD)/(BC - BD). Since KL = 2r, this relation can be
written as BMlr = (BC + BD) I(BC - BD).
To arrive at the main part of his proof, Pappus now establishes a few
more essential lemmas. From the similar triangles BKH and FLC he
obtains BK·LC = AM2. And, since BCIBD = BLIBK, he develops two
more relations, BL·CD = BC·2r and BK·CD = BD·2r.
We now go to Figure 6, in which circles (A) and (P) are any two
tangent circles placed as shown. The projections of A and P upon BC are
denoted by M and N respectively. Pappus has already established that the
ratios BM/AS and BNIPO are constant and are equal to (BC+ BD)/(BC
- BD). Using the other established lemmas and setting up appropriate
relationships in similar triangles, he finds that FH = FB. Finally, with a
bit more maneuvering, Pappus succeeds in showing that (AM + d)/d
= PN/d', where d and d' are respectively the diameters of the circles (A)
and (P).
Only now can Pappus attack the main theorem. Referring to
Figure 7, we use the lemmas ingeniously contrived by Pappus to find that
B K DM L
FIGURE 7
.8116~
BK·LC = KL2 and that BK·LC = AM2, whence KL = AM, or PI = dl ,

where PI is the distance from A to the line BC. Let P2 and d2 denote the
corresponding elements of the second circle. Then the last lemma gives us
(PI + dl)/d l = Pz/d2 , so that Pz = 2d2 · Continuing in the same manner,
we obtain the same result as that shown by inversion, namely that the
distance of the center of any circle of the Pappus Chain to the baseline BC is
equal to 2n times its radius. A similar argument applied to the case shown
in Figure 2 results in the conclusion that the required distances are 2n - I
times the corresponding radii.
The foregoing is a condensed version of Heath's synopsis of how
Pappus did it; that is, how Pappus developed a framework on which to
build his famous Circle Theorem. If we plod through the Pappus proof
with painstaking patience we are convinced that Pappus was a genius who
knew his way around geometrical configurations.
Like Euclid and Archimedes, Pappus proved his theorems by pure
geometry with occasional ventures into a primitive sort of algebra, one
involving nothing more complicated than the equating of ratios. We have
seen how Pappus might have proved his famous Circle Theorem ifhe had
only had access to the principles oflnversion. Let us imagine how he might
have done it with the help of our more flexible algebra.
Adopting a fresh notation for Figure 8, let D and E denote the centers
of any two tangent circles (radii x andy respectively), bounded by the arcs
AB and AC (centered at 0 and P), and let F, H denote the projections of D
and Eon AB. Let AO = R, AP = r, DF = mx and EH = ny.
From the relation F0 2 - FP2 = D0 2 - Dp2, we obtain
or
(R + r - 2AF) (R - r) = (R + r) (R - r - 2x),
x (R - r) .. Y (R ~ r)
thus,
AF
=
(R + r)' and sImIlarly AH = (R + r)'
A HF P 0 c
FIGURE 8
~117~
Then
(R + r)
HF = AF - AH = (x - y) (R _ r)·
Now
DE2 = HF2 + (DF - EH)2. So
(X + Y )2 = (R + r)2(x - y)2 + (mx-ny.

)2
1
(R - r)2
Now express the area of triangle DOP first by the Heronian Formula and
then by the traditional "half the base times the altitude". We find that
mx(R - r)
-JRrx(R - x - r) = 2 .
Similarly, for triangle EHO, JRry(R - y - r) = ny(R - r)/2.

Solving for x andy, we obtain
4Rr(R - r)
2 x ---~--~--~~
- 4Rr + m2(R - r)2
and
4Rr(R - r)
3 y - -------,;---------;;-
- 4Rr + n2(R - r)2
Equations 2 and 3 yield the relation
x n2(R - r)2 + 4Rr
4
y
= m2(R - r)2 + 4Rr
Now equation 1 can be converted to the form
(~) [4Rr + m2(R - r)2] + ~) [4Rr + n2(R - r)2]
= 2[2R 2 + 2r2 + mn(R - r)2],

which together with equation 4, yields
4(R - r)2 = (R - r)2(m 2 + n2 - 2mn).
Finally, m - n = ±2, and since n > m, we obtain n - m = 2.
Note that D could be chosen as the center of the initial circle or of any
other circle of the Pappus chain and that the ratio EH/y and the
corresponding ratios for successively tangent circles ad infinitum follow the
pattern n - m = 2, that is, the successive ratios belong to an arithmetic
progression with a common difference of 2.
Whether solved by geometry, algebra or inversion, the Pappus Circle
Theorem continues, after many centuries, to arouse our mathematical
imagination and to intrigue us with its elegance.
8118~
Numbers and
Coding Theory
-~-
Fault-free Tilings
of Rectangles
-~-
R. L. Graham
BELL LABORATORIES
Imagine that we have an unlimited supply of rectangular tiles of size 2 by

I and we wish to tile the floor of a rectangular room of size p-by-q. Of
course, we must cover all pq square units of floor area. Furthermore, two
tiles are never allowed to overlap. As an example, we show in Figure I a
tiling of a 5-by-6 rectangular floor.
-.--
FIGURE I
A tiling of 5-by-6.
I t is easy to see that if such a tiling is to be possible then pq must be

even, since the area of each tile is 2. On the other hand, ifpq is even, then at
least one of p and q must be even, say p = 2r. In this case we can place the
tiles as shown in Figure 2 to construct the desired tiling.
~120~
FAULT-FREE TILINGS OF RECTANGLES
2 ---- r
p = 2r
FIGURE 2
A tiling of 2r-by-q.
Fault-free Tilings
Ifwe examine the tiling shown in Figure I more carefully, we notice that it
contains a "fault-line", that is, a straight line completely cutting through
the rectangle which doesn't cut through any tile. Let us call a tiling which
has no such fault-linesfault-free. Ifwe think ofa tiling as a cross-section ofa
wall built of bricks then it is clear why we might like to avoid fault-lines. In
Figure 3 we show a fault-free tiling of a 5-by-6 rectangle.
A curious phenomenon occurs however, in trying to construct a fault-
free tiling of a 6-by-6 rectangle. (The reader is encouraged to find one
before proceeding further.) The same difficulty occurs for a 4-by-6
rectangle. In fact, there are no fault-free tilings for either of these cases!
This leads to a question* that no mathematician can resist asking:
Exactly which p-by-q rectangles have fault-free tilings?
J
I
5
I
I
6
FIGURE 3
A fault-free tiling of 5-by-6.
Answering the Question

To begin with, a rectangle cannot have a fault-free tiling ifit has no
tiling at all, that is, ifits area is an odd number. Stating this another way, a
necessary condition for a p-by-q rectangle to have a fault-free tiling is that
pq is divisible by 2.
~ 121 f:>
q q
(a) (b)
FIGURE 4
A fault-free tiling of 2-by-q?
As we have already seen, however, this is not enough. Suppose for

example we try finding a fault-free tiling of a 2-by-q rectangle (where q
;;::: 2). There are just two ways to fill in the left-hand end of the rectangle,
which we show in Figure 4. In either case however, (because q ;;::: 2) we
must form a (vertical) fault-line. We therefore conclude that such
rectangles have no fault-free tilings, even though their area is even.
In a similar spirit, consider what happens in the attempt to avoid
fault-lines when tiling a 3-by-q rectangle. There are basically just two ways
of tiling the end of the rectangle (see Figure 5).
The start shown in (a) is never any good since if q = 2 there is a
horizontal fault-line and if q > 2 there is a vertical fault-line. What
happens in case (b)? We show the various possibilities in Figure 6.
As before, case (a) creates fault-lines and must be discarded. The
remaining possibility is to place Tile I as shown in (b). However, this forces
use to place the additional Tiles 2 and 3 as shown in (c) (nothing else will
fit into the gaps created in (b)). But notice that the "profile" of the pattern
q q
(a) (b)
FIGURE 5
Trying to tile 3-by-q.
q q q
(a) (b) (c)
FIGURE 6
Still trying to tile 3-by-q.
~ 122&.
FAULT-FREE TILINGS OF RECTANGLES
6 L
FIGURE 7
Potential fault-lines in 6-by-6.
in Figure 6 (c) is exactly the same as that in Figure 5 (b). Thus, we have
only delayed the problem of filling the indentation created in Figure 5 (b).
Eventually we must face the issue of completing the tiling (q is finite, after
all) and this can only be done by finally filling this annoying indentation
with a vertical tile as in Figure 6 (a). However, as soon as this happens we
have formed a fault-line!
So, we conclude that no 3-by-q rectangle has a fault-free tiling.
In a similar way (but with a few more cases) it follows that no 4-by-q
rectangle has a fault-free tiling as well. (Naturally, from the symmetry of
the situation this means that no p-by-4 rectangle has a fault-free covering
either.) Of course, one must resist the temptation to generalize at this point
by trying to show that the same holds for 5-by-q rectangles, etc. After all,
we do have a fault-free tiling of a 5-by-6 rectangle.
What we have shown at this point is that another necessary condition
for a p-by-q rectangle to have a fault-free tiling is that both p and q must be
at least 5.
This still leaves the nonexistence of a fault-free tiling of the 6-by-6
square unexplained. This gap can be filled by the following stunning
argument of S. W. Golomb and R. I. Jewett.
Suppose for the moment that we have managed to find a hypothetical
fault-free tiling of the 6-by-6 square. In the square there are 5 vertical and
5 horizontal fault-lines which, we assume, must all be broken by tiles (see
Figure 7). Notice that each tile breaks exactly one potential line.
Furthermore (and this is the crucial observation), ifany fault-line (say L in
Figure 7) is broken by just a single tile, then the remaining regions on either
side of it must have an odd area, since they consist of6-by-t rectangles with
a single unit square removed. However, such regions are impossible to tile
by 2-by-1 tiles. Each of the 10 potential fault-lines must be broken by at
least two tiles. Since no tile can break more than one fault-line, then at least
20 tiles will be nect;ssary for the tiling. But the area of the 6-by-6 square is
only 36 while the area of the 20 tiles is 40! Thus, we have reached a
contradiction. No such tiling of a 6-by-6 square can exist.
~123&.
p+4
000
q + 4
FIGURE 8
Extending fault-free tilings.
We can summarize what we know up to this point as follows: Necessary

conditions for the existence of a fault-free tiling of a p-by-q rectangle are:
1 pq is divisible by 2;
2 P ~ 5, q ~ 5;
3 (p,q) =1= (6,6).
Surprisingly, it turns out that these conditions are also sufficient. That
is, if p and q satisfy 1, 2 and 3 then the p-by-q rectangle will always have a
fault-free tiling. One way this can be proved is by starting with small fault-
free tilings, such as 5-by-8, 6-by-8 and our earlier 5-by-6, and building up
larger fault-free tilings from these. For example, if we have a fault-free
tiling of a p-by-q rectangle then in Figure 8 we show how to form a fault-
free tiling of a (p + 4) by (q + 4) rectangle. A similar construction can be
used to form a fault-free tiling of a (p + 2r + 2) by (q + s + 2) rectangle
from the original p-by-q for any positive integers rand s.
Other Tiles
It is only natural to wonder about the possibilities offault-free tilings when
other-sized tiles are used; for example, 3-by-l or 7-by-5. At first, an exact
characterization of just which p-by-q rectangles can be appropriately tiled
appears to be a hopelessly difficult problem. However, it turns out that
there is a surprisingly beautiful answer to this question. Suppose we are to
use tiles of size a-by-b where we can assume by changing our units if
necessary, that a and b have no common factor greater than 1. (We can
also assume that we don't have a = b = I since otherwise no fault-free tilings
are possible, except for the trivial tiling consisting of a single tile!) As in the
.8124~
FAULT-FREE DLINGS OF RECTANGLES
previous result, there are two basic necessary conditions: one dealing with a
divisibility condition and one dealing with a size condition.
Regarding divisibility, the area pq of the rectangle must be divisible
by the area ab of the tile if any tiling, fault-free or not, is to be possible.
Other requirements are necessary as well. By coloring the tile and the
rectangle with ab colors in an appropriate cyclic manner, it can be argued
that it is necessary that
I' Both a and b each divide at least one of p and q. (This is

stronger than just requiring that ab divides pq).
Regarding size, there should be an analog in the general case to the

earlier condition 2 which required p, q 25. The corresponding condition is
unexpected:
2' Each ofp and q must be able to be expressed as a sum xa + yb

with positive integers x andy in at least two ways.
Basically, this guarantees that there is enough freedom in placing tiles

across the sides of the rectangle; that is, we don't always have to place the
same numbers of tiles horizontally and vertically. For the case a = 2, b = 1,
2' reduces to the earlier condition 2, since 2, 3 and 4 do not have two
representations as x'2+y'1, x,y>O, while any integer 25 does; for
example, 5 = 1 . 2 + 3' 1 = 2' 2 + I' I, etc.
Curiously enough, the impossibility of a fault-free tiling of a 6 x 6
square with 2 by I tiles remains as the unique anomalous exception. In all
other cases it is possible by construction techniques similar to those
mentioned before to produce the required fault-free tilings whenever p and
q satisfy the necessary conditions. The general result is summarized in the
following statement.
THEOREM A fault-free tiling of a p-by-q rectangle with a-by-b

tiles exists (where we assume pq > ab and (a, b)
= I) if and only if
I' Each of a and b divides p or q;
2' Each of p and q can be expressed as xa + yb,
x,y > 0, in at least two ways;
3' For {a,b} = {1,2}, (p,q) =f= (6,6).
The actual proof of this result is not difficult and is left to the energetic
reader. We remark that m can be expressed as xa + yb in at least two ways
if and only if m - ab can be expressed as xa + yb in at least one way (assum-
~125~
ing (a,b) = 1). In general, such integers do not form an interval as they do
for a'= 2, b = 1. For example, by using 3-by-2 tiles, it is possible to find
fault-free tilings for II-by-18 and 14-by-15 rectangles but it is not possible
for a 12-by-12 rectangle (in particular, 11 = 3-3 + 2-1 = 3-1 + 2-4 but
12 = 3-2 + 2-3 is all).
What Next?
It is typical of this business that one answer leads to n more questions. For
example, how many fault-free tilings does a rectangle have? What if we can
use two sizes of tiles instead ofjust one? What about these same questions in
3 (or more) dimensions? What if we require each fault-line to be broken by
at least 2 tiles? At least n? We have reached the frontier of our current
knowledge on this topic. We encourage the interested reader to explore this
fascinating byway of geometry and discover for himself the gems which
must surely lie waiting to be discovered.
* Another more general question is: How many different fault-free tilings does a
p by q rectangle have? We don't confront this question here.
~126&.
Dissections into
Equilateral Triangles
-~-
w. T. Tutte
UNIVERSITY OF WATERLOO
I am delighted to have the opportunity to contribute to this Collection

honouring Martin Gardner. I once wrote a paper for his column in Scientific
American, about dissections of rectangles into squares [7]. Perhaps another
article on dissections would be appropriate here.
The afore-mentioned paper was concerned with "squared rectan-
gles"; that is, dissections of rectangles into squares. A dissection of a
rectangle or square into unequal squares is a perfect rectangle or square
respectively. The paper was based on earlier work of Brooks, Smith, Stone
and Tutte published in the Duke Mathematical Journal in 1940 [2]. This
earlier paper gives an example of a perfect square and describes a method
(exploiting symmetrical subgraphs) for constructing others. There is also a
note on possible generalizations. One of these is the study of dissections of
equilateral triangles into other equilateral triangles. There are two later
papers on this generalization by myself, ([6] and [8]), and one by Brooks,
Smith, Stone and Tutte [3].
I have no cause to complain about the reception of papers on squared
rectangles by mathematicians and other scientists. They are all pleased
that there should be a connection between KirchofPs Laws of electrical
flow and the dissection problems of pure geometry. But there seems to be
..a 127~
no corresponding interest in the triangular problem. Yet, in my opinion the

theory of triangulated triangles is a simple and elegant generalization of
those theories of squared rectangles. In the remainder of this paper I will
try to justify the preceding statement, in the hope of awakening the interest
that I feel these dissections deserve.
When we are discussing dissections into equilateral triangles it does
not really matter whether the original figure is a triangle or a paral-
lelogram (with angles of 60° and 120°). A dissection of a triangle can be
changed into that of a parallelogram by deleting one constituent triangle
(at a vertex) and adding another. A dissection of a parallelogram can be
changed into that of a triangle by adding new constituent triangles on two
adjacent sides.
I t is the dissected parallelogram that is the natural analogue of the
squared rectangle. Suppose we start with a squared rectangle R and shear it
so that it becomes a dissection of a parallelogram into rhombuses. Each
rhombus can be dissected along an appropriate diagonal to form two
congruent equilateral triangles. The squared rectangle R is thus transform-
ed into a triangulated parallelogram P. Because of this observation we
can regard the theory of squared rectangles as a special case of the theory of
triangulated parallelograms.
Suppose that R is a perfect rectangle. Does it make sense to say that P
is perfect also? Given any triangulated triangle or parallelogram we can fix
one of its sides as "horizontal". Then any constituent triangle T has one
horizontal side. Let us say that T is positively or negatively oriented
according to its position above or below its horizontal side. Let us define
the "size" of T as plus or minus the length of its side, depending on whether
T is positively or negatively oriented. Then we say that a triangulation is
"perfect" if no two of its constituent triangles have the same size. With this
definition we can say that the perfect rectangle R shears into a perfect
parallelogram P. For the two congruent triangles of P, arising from any
square of R, have opposite orientations and therefore have different sizes.
Figure 1 shows a perfect parallelogram that cannot be obtained by
shearing a perfect rectangle. In [6] I have satisfied myself that this
parallelogram has the least number of constituent triangles, namely 13,
consistent with perfection. For the sake of comparisons with squared
rectangles, I like to say that its order is 13/2. For each square of a rectangle
corresponds when sheared to two triangles. The number entered in each
triangle of Figure 1 is its size.
At this stage in papers on squared rectangles, there appears a diagram
showing the relation between a squared rectangle and its electrical
network. Figure 1 fulfills the same purpose in the theory of triangulated
parallelograms. But now the term "electrical" must be used in a general-
ized sense. It refers to a form of "electricity" not yet found in physics, (as
far as I know). But we think of it as flowing in directed graphs and obeying
analogues of KirchhofPs Laws.
~128~
DISSECTIONS INTO EQUILATERAL TRIANGLES
y
FIGURE 1
The parallelogram P is on the left and the corresponding electrical

network is on the right. Each vertex of the network N corresponds to a
maximal horizontal segment, made up of sides of constituent triangles in P
and is shown on the same level. The network is directed. Each of its darts,
or directed edges, corresponds to a constituent triangle of P (the direction
of the dart is from the apex to the base of the triangle). The size of each
triangle is written against the corresponding dart of N and is called the
current in that dart. The vertices of N corresponding to the upper and lower
horizontal sides of P are called the positive and negative poles of N
respectively.
We observe that N is a balanced digraph. (Digraph is short for directed
graph.) Balanced means that at each vertex the number of incoming darts is
equal to the number of outgoing ones. It is not difficult to see that N is
planar (for any P), and that it can be drawn in the plane so that the order
of the darts around each vertex is the order of the corresponding triangles
around the horizontal segment of P. In such a drawing, incoming and
outgoing darts must alternate around each vertex. We describe a plane
figure with this property as an alternating map.
At each vertex v of N, the sum of the currents in the darts directed to v
is called the current leaving N at v. Its negative is the current entering Nat v.
Thus, the current entering the positive pole is equal to the current leaving
the negative pole, and each of these numbers measures the length of a
horizontal side of P. But the current entering or leaving N at any non-polar
vertex is zero, and the current leaving is the sum of the sizes of the triangles
of P having their bases in the corresponding horizontal segment. This rule
of currents at non-polar vertices is our analogue of KirchofPs First Law.
Let us define the potential of a horizontal segment of P as its distance
above the lower horizontal side rif P, measured parallel to a slanting side rif P.
~129&.
Transferring this number to N we call it the potential of the corresponding

vertex. Now the vertices of N incident with a dart D are called the head and
tail of D, with the direction of the dart being from tail to head. We can
assert the usual form ofKirchhoffs Second Law, as follows. The total fall of
potential around any circuit C of N, regardless of the directions of the darts,
is zero. We note that the current in a dart D is equal to the fall of potential
from tail to head. We can thus apply the Second Law by saying that the
sum of the currents in any circuit C is zero, provided that we agree that
currents in darts going against the chosen direction of the circuit are to be
replaced by their negatives.
The modified Kirchhoff Laws can be used here like the conventional
ones in the theory of squared rectangles. Having drawn an alternating map
M, with at least four darts at each vertex, we can choose positive and
negative poles, incident with a common face. We can solve the equations
for a set of currents. We can then reverse the construction of Figure I to get
a corresponding triangulated parallelogram. With luck this will be perfect.
The complete theoretical justification of this procedure may be difficult,
but the principle is clear.
It is now possible to deduce properties of triangulations from those of
alternating maps. Thus, we can infer from the Euler Polyhedron Formula
that the alternating map corresponding to a triangulated parallelogram
must have either a 2-sided face or a non-polar vertex incident with only
four darts. In either case two of the constituent triangles of the paral-
lelogram must be congruent. This result leads to the assertion in [2],
couched in a different terminology, that there is no perfect equilateral tnangle.
Consider a squared rectangle R with its ordinary electrical network G.
Let it be sheared to form a triangulated parallelogram P, and let the
corresponding N be derived. We find that N can be obtained from G by
replacing each edge with two oppositely directed darts. In the correspond-
ing alternating map these two darts bound a digon. The modified
Kirchhoff Laws for N are easily seen to be equivalent to the ordinary ones
for G, on the assumption of unit resistances. The squared rectangle, then, is
still a special case in the triangular theory.
Precedents set by the studies of squared rectangles would lead us to
expect that someone, at this stage, would have constructed extensive
catalogues of triangulated parallelograms. Alas, it is not so.
A detailed discussion of the Kirchhoff equations for an undirected
graph G would be out of place here. Let us note that the theory can be
based on the Kirchhoff matrix K( G) of the graph concerned. We suppose the
vertices ofG to be enumerated as Vl,V2,''''V n • Then K(G) has n rows and n
columns. Its /h diagonal entry is the number of edges joining vj to other
vertices. The non-diagonal entry in the zth row and/h column is minus the
number of edges joining Vi to Vj' Thus, K(G) is a symmetrical matrix and its
elements sum to zero in each row and column.
et130~
The Matrix-Tree Theorem asserts that if we strike out from K(G) the
lh row and column then the determinant of the resulting matrix K/G) is
the number T( G) of spanning trees of G. Let us agree that the resistance of
each edge is to be unity, and let us choose two vertices as positive and
negative poles. It is convenient to take T(G) as the magnitude of the current
entering G at the positive pole and leaving at the negative, for then the
currents in the various edges are integers. These currents constitute the full
flow in G with respect to the chosen poles. The potential differences in this
flow can be expressed as the determinants of certain submatrices of K(G),
appropriately multiplied by + I or - 1. In particular, the potential
difference between the poles is the determinant of the submatrix obtained
from K(G) by striking out the two rows and two columns corresponding to
the poles.
All this generalizes very nicely to a theory of unsymmetrical electri-
city in directed graphs. Let G be any directed graph, not necessarily
balanced. We define K(G) much as before. Thel h diagonal entry is the
number of darts directed to vj from other vertices. The non-diagonal entry
in the Z"lh row andlh column is minus the number of darts directed from Vi to
Vj. We note that this K( G) is not necessarily symmetrical. Its elements sum
to zero in the rows but not necessarily in the columns.
We can define Kj(G) as before, and its determinant is found to be a
number of spanning trees ofG. But not all the spanning trees ofG are to be
counted -only those in which each edge is directed away from Vi. We call
these structures the out-arborescences ofG on Vj. A spanning tree ofG in which
each edge is directed towards Vj is an in-arborescence of G on Vj. (In the
undirected case, the value of det Kj(G) is independent of j, but in the
directed case this is not necessarily true.)
Let us choose a positive pole vp and a negative pole vq, and use the
modified Kirchhoff Laws to calculate a corresponding distribution of
currents. The current entering at vp is not necessarily equal to the current
FIGURE 2
~131~
leaving at v . However, we can find a full flow in which current det Kq(G)
enters at v qand current det Kp( G) leaves vq. Determinants of appropriate
submatric:s of K( G) fix the currents and potential differences in this full
flow. These currents and potential differences are found to be integers.
The rule for the potential difference between the poles is the same as
before.
As an example, let us take the graph of Figure 2, with VI as positive
pole and v6 as negative. The currents of the corresponding full flow are
indicated.
We easily verify the matrix K(G) to be as follows.
2 0 -1 -1 0 0
-1 1 0 0 0 0
-1 -1 3 0 -1 0
0 0 0 1 0 -1
0 -1 0 0 1 0
0 0 -1 0 -1 2
Let us write Tj(G) for the number of out-arborescences of G on Vj'

Either by inspection of Figure 2 or by evaluating det Kl (G) and det K6 (G)
we find that Tl (G) = 6 and T6 (G) = 3. Hence the flow depicted in
Figure 2 is full, for a current of3 enters at VI and a current of61eaves at v6'
The fact that the currents in the darts from VI to V2 and from V2 to V5 are
zero, is an immediate consequence of the modified First Law.
What happens if we reverse the direction of every edge, to get a new
digraph G'? As far as the non-diagonal elements are concerned we replace
K(G) by its transpose. But diagonal elements may change. IfG is the graph
of Figure 2 then K(Gj is the following matrix.
2 -1 -1 0 0 0
0 2 -1 0 -1 0
-1 0 2 0 0 -1
-1 0 0 1 0 0
0 0 -1 0 2 -1
0 0 0 -1 0 1
The digraph G'is shown in Figure 3, again with a full flow.

We have Tl (Gj = 8 and T6 (G') = 9. There is an evident 1 - 1
correspondence between the out-arborescences of G' on Vj and the in-
arborescences of G on Vj'
When we restrict our theory to balanced digraphs we avoid many
complications. For if N is a balanced digraph, the elements of K(N) sum to
zero in the columns as well as in the rows. Hence we can show, by
elementary determinant theory, that det Kj(N) is independent of}. The
number of out-arborescences of N on a given vertex V is the same for each v.
.8132 ~
DISSECTIONS INTO EQ.UILATERAL TRIANGLES
FIGURE 3
We therefore refer to this number simply as the tree-number T(N) of N. As

an example we can take the balanced digraph of Figure 1. If the vertices
are numbered from above to below in the diagram, then
K(.N)
[
= -~
2
-1
-1
-1
-1
0
3
0
-1
-1
-1
3
-1
-1
0
-1
3
J]
-1
2
It follows that in the full flow, the fall of potential between the poles is
3 -1 -1
-1 3 o = 20.
-1 -1 3
We deduce that the flow shown in Figure 1 is full. Accordingly T(N),

being equal to the current entering at the positive pole, is 19.
Given a balanced digraph N we can reverse all its darts to obtain a
digraph N'. Clearly N' is balanced. Moreover if N defines an alternating
map in the plane, then so does N'. But now K(N') is simply the transpose of
K(N). Hence T(N') = T(N). Hence, on each vertex of N the numbers of
in-arborescences and out-arborescences are equal.
Let two vertices be chosen to be positive and negative poles in both N
and N'. Consider the two full flows. The current entering at the positive
pole is the same in each, by equality of tree-numbers. Moreover the fall of
potential between the two poles is the same in each. For it is given in N by
the determinant of a submatrix symmetrically related to the main
diagonal, and in N' by the determinant of the transpose of this matrix.
If N corresponds to a triangulated parallelogram, then so does N'.
The parallelograms, P and P'respectively, have the same size and shape, as
a result of the preceding paragraph. Unfortunately this effect is not well-
.a133~
illustrated by the digraph of Figure 1. Reversing its darts merely gives an

isomorphic digraph; moreover each pole is invariant under the isomor-
phism. If N corresponds to a squared rectangle R, then P and P' correspond
to two ways of slicing R. In shearings, we can move the upper horizontal
side either to the left or to the right with respect to the lower one.
On page 478 of [6] we are shown two perfect triangulated paral-
lelograms which are the P and P' corresponding to an alternating map N.
Each parallelogram has horizontal side 3441 and slanting side 2999. Each
is dissected into 36 constituent triangles, which are all different sizes. Each
has two congruent constituent triangles of side 129. But apart from this,
there is no repetition of triangle-size between the two dissections.
The leaky or asymmetrical electricity relevant to the theory of
triangulated parallelograms is discussed in more detail in [3], [6] and [9]. In
the most general theory, the conductance (reciprocal of resistance) of a
dart is not restricted to 1, but is an indeterminate over the integers.
A squared rectangle R has sides in two perpendicular directions, and
either of these directions can be taken as horizontal. The two choices lead
to two electrical networks G and G1 . These, however, are simply related. If
each is completed by the adjunction of a new edge joining the poles then
the two completed networks, or c-nets, are plane duals. Since the two
corresponding systems of electrical equations give rise to the same pattern
of squares, it seems reasonable to expect the determinants det K;(G) and
det Kj(G 1 ) to be equal. This is known to be the case; it can be shown that
dual-plane-connected graphs have equal tree-numbers.
Is there anything corresponding to a pair of dual c-nets in the theory
of triangular dissections? There is indeed, but to exhibit it we must pass
from triangulated parallelograms to triangulated triangles. As an example
we can adjoin two new constituent triangles to the triangulated paral-
lelogram P of Figure 1 to obtain the triangulated triangle of Figure 4.
The corresponding modification of the electrical network of Figure 1
is obvious. To represent the new triangle of size 20 we need a new dart on
FIGURE 4
~134~
the left side, directed from the old positive pole X to the old negative pole
r. For the triangle of size 19 we need a new vertex Z and a new dart from Z
to X. We can say that Z represents the apex of the triangulated triangle,
even though this apex is not strictly a horizontal segment.
The currents in the new electrical network obey the modified
r
KirchhotTLaws with Z as positive pole and as negative pole. The number
of out-arborescences from Z can be shown to be 39; the side-length of the
dissected triangle. The number of out-arborescences from r (or from any
vertex but Z) is obviously zero. This corresponds to the fact that the
current entering the network at Z is zero.
It is, in some ways, convenient to make Z coincide with r in the
electrical diagram, so that the diagram exhibits a new alternating map M.
The dart originally incident with Z can be called the polar dart, and it can
be distinguished in the diagram by a cross-bar. Figure 5 shows an electrical
diagram of this kind corresponding to the triangulation of Figure 4. The
vertices correspond to the maximal horizontal segments of the
triangulation.
Figures 4 and 5 exhibit one example of a general procedure. Given
any alternating map M we can mark in it a polar dart D, and we may then
expect to derive a corresponding triangulated triangle. To form the
corresponding electrical network we detach D from its tail-vertex rand
give it a new tail Z incident with no other dart. After this, Z is taken as the
r
positive pole, and is taken as the negative pole. The calculated currents
are interpreted as triangle-sizes. (Sometimes we find a zero current, and
then complications arise.)
Returning to Figure 4 we observe that the sides of its constituent
triangles make up three families of parallel segments, and that anyone of
these families could be chosen as horizontal. We might, for example,
consider maximal segments parallel to the left slanting side of the dissected
triangle, as shown in Figure 4, and make these segments correspond to the
vertices of another electrical network. This is shown, with a marked polar
20
FIGURE 5
11
FIGURE 6
edge, in Figure 6. The remaining family of maximal segments, correspond-

ing to the right slanting side, gives rise to the electrical network of Figure 7.
In Figures 5, 6 and 7 we find three different electrical networks, all
carrying the same set of currents. Such triads of alternating maps can be
regarded as analogues of the pairs of dual maps in the theory of squared
rectangles. Our analogue of duality is called triality in [6] and trinity in [8].
The three maps making up a triad are said in [8] to be trine to one another.
Given any alternating map we can define its trine maps directly,
without constructing a related triangulated triangle. It is shown in [8] how
a triad M 1 , M 2 , M3 of alternating maps can be derived from a bicubic
map M, and how the triad associated with any triangulated triangle has its
corresponding M.
20
FIGURE 7
~136~
In a bicubic map, the vertices can be distinguished as black and

white, each edge joining a white vertex to a black one. Moreover, the faces
can be 3-coloured (say in red, green and blue) so that each edge separates
faces of two different colours. Apart from permutations of the three colours,
the 3-colouring is unique. We can form an alternating map Ml of the
corresponding triad as follows. We give it one vertex in each red face of M.
We form its darts from the edges of M, separating green and blue faces.
Each of these is directed from its black end to its white, and each is
extended at each end through the adjacent red face of M to the vertex of
Ml contained in that face. The other members of the triad are defined
similarly, each with a permutation of the three face-colours. (In [8] the
definition of a triad is given, not directly in terms of the bicubic map M but
in terms of its dual map. This dual is an Eulerian triangulation of the
sphere.)
Suppose M 1 is derived from an undirected planar map N by
replacing each edge by two oppositely directed darts. It is then found that
one other member of the triad (say M 2 ) is similarly derived from the dual
map of N. The third member of the triad is an oriented form of the medial
map of N. This observation justifies our assertion that trinity is a true
generalization of duality.
Since dual plane maps have equal tree-numbers we may expect trine
alternating maps to have equal tree-numbers too, and this has been proven
to be the case. (Proofs are given in [6] and [8]. A simpler proof was
discovered recently by K. A. Berman [1].)
The common tree-number of three trine alternating maps M l , M2
and M 3 can be regarded as a property of the associated bicubic map M. In
[3] it is interpreted directly in terms of the structure of M, as a count of
matchings, not trees.
It seems appropriate to conclude the exposition here. Any reader
interested in further study can proceed to the papers on our list of
references. However, a warning should be given regarding one minor
point. In a triangular dissection it may happen that six constituent
triangles meet at a single point. Such a point is called a cross. It corresponds
to a cross in the theory of squared rectangles; that is, to a point at which
four constituent squares meet. In neither theory is it usual to encounter a
cross in an interesting example unless it has been forced to occur by some
symmetry in the construction. But when crosses do occur, it is necessary to
introduce new conventions so that we may still have pairs of dual c-nets or
triads of trine-alternating maps. In short we subdivide some of the maximal
segments at the crosses. It is the resulting subsegments that correspond to
the vertices of the associated electrical networks. Each cross, in the
triangular case, is the intersection of three maximal segments. The rule is
that exactly two of them (the choice is arbitrary) are to be subdivided
there.
~137~
Addendultl
Writing the above revived my interest in triangulated parallelograms, and
I set out to calculate one or two more. I calculated two, and they seem to
be sufficiently interesting to record here. They are perfect triangulations-
each of the order llt. Each has horizontal side 401 and slanting side 264.
There is no case of a constituent triangle in one being congruent to a
constituent triangle in the other. This is the only case known to me of two
perfect triangulations of the same parallelogram with no constituent size
repeated from one dissection to the other.
The electrical networks of the two parallelograms are related by a
reversal of darts. These networks, with their currents, are shown in Figures
8 and 9. I have verified that the flows shown are full. In each case X is the
FIGURE 8
FIGURE 9
~138~
r
positive pole and is the negative pole. A current of 40 1 enters at X and
leaves at r, and the fall of potential from X to r is 264. The reader should
by now have no difficulty in deriving the actual triangulations.
References
I Berman, K. A. 1978. Spanning trees, arborescences and 4-valent graphs.
Thesis. Waterloo.
2 Brooks, R. L.; Smith, C. A. B.; Stone, A. H. and Tutte, W. T. 1940. The
dissection of rectangles into squares. Duke Math., 7: 312-340.
3 _ _ . 1975. Leaky electricity and triangulated triangles. Philips Res.
Reports 30: 205-219.
4 Duijvestijn, A. J. W. 1978. Simple perfect squared square of lowest
order. J. Combinatorial Theory B 25: 240-243.
5 Sprague, R. 1939. Beispiel einer Zerlegung des Quadrats in lauter
verschiedene Quadrate. Math. Zeitschrift 45: 607.
6 Tutte, W. T. 1948. The dissection of equilateral triangles into equilat-
eral triangles. Proc. Cambridge Phil. Soc., 44: 463-482.
7 _ _ . 1961. Squaring the square. In The 2nd Scientific American Book oj
Mathematical Puzzles and Diversions, ed. Martin Gardner. New York:
Simon and Schuster.
8 _ _ . 1973. Duality and trinity. Colloquia Mathematica Societatis Janos
Bolyai. 10: l459~ 14 72.
9 _ _ . 1976. The rotor effect with generalized electrical flows. Ars
Combinatoria I: 3-31.
~139~
In Praise of ADlateurs
-~-
Doris Schattschneider
MORAVIAN COLLEGE
One of the most appealing aspects of Martin Gardner's column

"Mathematical Games" is its presentation of mathematical problems
designed to intrigue amateurs and encourage their personal efforts at
solution. By his own insistence, Gardner is an amateur mathematician and
gives no special deference to formal mathematical education-his column
pays tribute to the efforts of the mathematical "great" and the mathemati-
cally unknown, names often appearing side by side with no titles to
distinguish one from the other. Amateurs are his most avid followers and
enjoy the challenge of matching their wits against others in solving
problems. Amazingly, their lack offormal mathematical education is often
an advantage rather than a hindrance and their ingenious solutions to
problems sometimes top the efforts of the professionals.
A striking case in point occurred as a result of Gardner's July 1975
column "On Tessellating the Plane with Convex Polygon Tiles".
Challenged by the column, Richard James decided to try his own hand at
solution and his obvious approach (altering the familiar) produced a
solution that was overlooked in a formal mathematical scheme. The
subsequent report of] ames's discovery aroused intense curiosity in Marjorie
Rice, providing her with staying power for a thorough and methodical
search (carried out mostly at her kitchen counter), which ultimately
yielded a wealth of new results. The invitation to write this article gives me
the chance to relate some details of the events that one Gardner column set
~140~
IN PRAISE OF AMATEURS
in motion-events that still continue to ripple. It is a tribute to the

thousands of amateurs who have made Gardner's column such a success.
Tiling problems have been a favorite subject of Gardner's over the
years. More than a dozen columns in the last twenty years have been
devoted in large measure to this subject. How to fit pieces together snugly
to fill a desired space seems to be one of our earliest childhood pastimes.
Even as adults we continue to be engaged by these problems, either for
pleasure or out of necessity, with tiles, bricks and the like. One of the most
basic tiling questions to ask is "What shape tile will fill the plane with its
replicas without gaps or overlaps?" Many obvious shapes come to mind-
no mathematician is needed to supply a lengthy list of examples. Some of
the beautifully shaped tiles found in ancient mosaics around the world
attest to the imagination of decorative artists in solving the problem. The
most general answer to our question is not known. In order to get partial
answers, conditions on the tiles are specified and then these specialized
questions become problems whose solution is sought. What if the tiles are
composed of stuck-together squares (polyominoes)? Or equilateral tri-
angles (polyiamonds)? Or regular hexagons (polyhexes)?
Gardner's July 1975 article discussed the problem in which the tiles
were convex polygons. "What convex polygons will tile? Explicitly
describe conditions on a convex polygon to insure that it tiles." Here,
Gardner had chosen a topic which had been worked on by many
mathematicians over a 50-year period and it had been announced that the
problem was completely solved. It is easily discovered that any triangle or
any quadrilateral can tile, but that convex polygons offive or more sides do
not always tile. For instance, regular pentagons do not tile but any
pentagon having a pair of parallel sides will tile. Regular hexagons do tile
but many other hexagons do not. Convex polygons having seven or more
sides cannot tile. This last statement declared by Gardner as "not hard to
show", is best described as a mathematical "folk theorem". Thus, everyone
quotes it, but no one can seem to cite a complete and accurate proof.
Fortunately, a recent article by Ivan Niven, in the December 1978
American Mathematical Monthly, fills this gap in the mathematical literature
and presents a thorough as well as convincing proof.
Gardner based his article on a 1968 paper by R. B. Kershner, which
surveyed the answers to the question, "What convex polygons will tile?"
The complete answer for hexagons (3 types) and partial answer for
pentagons (5 types) had been found by K. Reinhardt in 1918, and
extensive investigations by Kershner had yielded 3 more types of pentagons
which tile. Kershner felt the list was now complete (Figure I). His fasci-
nation with the problem was described in a letter quoted by Gardner: "For
reasons that I would have difficulty explaining, I have been intrigued by
this problem for some thirty-five years. Every five or ten years I have made
some kind of attempt to solve the problem. Some two years ago I finally
discovered a method of classifying the possibilities for pentagons in a more
[j
TYPE.
. ".
G
TYP£I
•
',. , c
TYP£7
.•
~ . o 0
C
b • ,
D
TYPE'
o TYP£I
°
ABC C
•
~
0L/
•
b
1. A + B + C = 360°. s. A = 60°, C = 120°,

2. A + B + D = 360°, and a = b, c = d.
and a = d. 6. A + B + D = 360°. A = 2e,
3. A = C = D = 120°, and a = b = e, c = d.
anda=b,d=c+e 7. 2B + e = 2D + A = 360°,
4. A = e = 90°, a = b = c = d.
and
and a = b, c = d. 8. 2A + B = 2D + e = 360°,
and a = b = c = d.
FIGURE I
The eight types of convex pentagons which tile, reported by Martin
Gardner in July, 1975.
convenient way than Reinhardt's .... The result of this investigation was
the discovery that there were just three additional types of pentagon ... that
can pave the plane. These pavings are totally surprising. The discovery
of their existence is a source of considerable gratification." Kershner's
fascination with the problem would prove to be contagious. Gardner's
exposition removed the subject from dusty mathematical journals (where it
had been unquestioned for many years) and placed it in the hands of a wide
readership, including many amateur puzzle enthusiasts. It is here that our
story begins.
4!3142&'
When RichardJames III saw Gardner's article, he read only the first
part; he decided to test his puzzle-solving skills before reading the
remainder of the article. He wrote (in a letter kindly supplied by H. S. M.
Coxeter), "Before reading Mr. Gardner's description of Mr. Kershner's
eight tessellating pentagons, I set about to try it for myself. The first thing
that came to mind was taking ... octagons (with squares filling the holes)
and adapting them so that pentagons replace the squares (thus moving the
octagons out of a lattice into "parallel" strips). The octagons neatly split
into four pentagons. The description was A = B = E = 90°, C = D
= 135°, and a = b = 2c = 2e. The tessellation was interesting but the
pentagon was dull. Rotating the cross in the octagon (and making other
adjustments as needed) produced [the pentagon and tiling shown in
Figure 2]." James sent his discovery to Martin Gardner with the inquiry:
"Do you agree that Kershner missed this one?" The exciting news was
immediately communicated by Gardner to Kershner and to a few other
mathematicians. Kershner's good-humored response and the James tessel-
lation were reported to readers in the December 1975 "Mathematical
Games" column. An amateur had, with one new example, shown that the
list of tessellating pentagons was not complete. Were there still others to be
discovered?
Perhaps an aside is in order here; a glimpse at the "mathematical
grapevine" maintained by Gardner (which I would like to denote MG 2 ,
for Martin Gardner's mathematical grapevine-shown in Figure 3). MG 2 is
kept humming year round by a steady flow of correspondence, telephone
calls, and personal conversations. When researching or writing a column,
Gardner contacts experts with knowledge and/or the latest information on
the subject; he seeks their comments on the accuracy of his manuscript
before submitting it for publication. Conversely, when Gardner receives
correspondence on any problem, before filing it away for future reference,
he sends copies to those he knows are actively interested. In this way, the
latest news is made available to those interested; its accuracy is checked and
comments are returned to Gardner; and most importantly, those working
on a common problem are put in touch with each other by Gardner. H. S.
M. Coxeter was one of those who received a copy ofJames's discovery from
Gardner; it was from Coxeter that I learned of the discovery. I couldn't
resist trying my own hand at the problem-it wasn't long before I had a
general description of a class of tessellating pentagons which generalized
the single example James had sent to Gardner. Following Kershner's
method of description, the general pentagon of James's type is one which
satisfies the equations: A = 90°, E= l80° -B, D = 90 0 +B/2, C= 180°
- B/2, a = b = c + e. After my communication of this to Gardner and
Coxeter, I found myself the recipient of copies of Gardner's correspondence
on "the pentagon problem" and soon was in direct communication with
others actively working on the problem. Ifmy bulging files on this problem
can be taken as an indicator of the correspondence generated by one
~143~
FIGURE 2
HOW RICHARD JAMES DISCOVERED A NEW PENTAGON THAT TILES
The familiar tiling by octagons and squares (with underlying square
grid in dotted outline) is shifted to see if pentagons might replace the
squares. A successful new tiling is produced. Further alteration of the
octagons and pentagons produces an example of a previously un-
discovered family of pentagons which tile.
FIGURE 3
MG 2
Gardner column, then his files surely must fill several rooms!
When the December 1975 issue of Scientific American was delivered,
another avid Gardner fan in California turned immediately to the
"Mathematical Games" column. Marjorie Rice, a San Diego housewife
and mother of five, was usually the first one in the household to read her
son's magazine. She had been intrigued by the July article on tiling by
pentagons and had "thought how wonderful it must have been [for
Kershner] to discover the new types of pentagon tiles." Now, reading of
James's newly discovered pentagon tile, her interest was strongly aroused
and she set out to see ifshe might find still other new pentagons which tile.
"I thought 1 would like to understand these fascinating patterns better and
see if 1 could find still another type. It was like a delightful new puzzle to
me and 1 considered how 1 could best go about this." Her search began
quite differently from that ofJames and soon became a full-scale assault on
the problem, extending over a period of two years.
Marjorie Rice had no formal education in mathematics beyond a
General Mathematics course required for graduation from high school in
1939. Thus, as she faced the challenge of finding new pentagonal tiles she
not only worked out her own method of attack, but also invented her own
notation as well. Her first step was to catalogue all of the information
available in the two Gardner columns on tiling pentagons (Figure 4). By
doing this, she hoped to discover any common relationships satisfied by the
pentagons and their tilings. "To begin, 1 needed to visualize the 9 types to
.8145~
:0 q)® ® W
"
8 ~ J\ A
E C B • B fO" £
B t;
.
... ~ a l:ltl .( ~ q 4 1:1 .(
.;J A b 8 ..3e e( j) f"e /) ~ q J)
/aO
~
~.
~
~ ~;f1:;;~
h <\ b <I b "
E. t! .6~ JJ .,~B £
... <!
"A h 8 <!. 4 'b C <>( 'J) <! '~'4o/ ~
(C+D)
I.A+B~C 3100 0
:l.A~P>t1> 3 '0° <x.d.
3. A = C '.D : 1.10 0 a:h d.~,.e.
-Yo A =C:y,,' .... ,t, (!xd
..1. II =6()o (!:1.:20 0 arh e:4.

6. A f f3~.lJ =..7.0· 1'\:.:Ie .,:/e t!.=e(
7-.:3/3 -Fe! :.:2D ,.. ...... ~~O· a',6 (! =ot'
~ aA 'B = .:I D ~ (! % 03(,0 0 11.,6: (! rd
9. f/ = 9()o e ~ZJ:"7'Oo J./J~E=.::2(!~B :3{;O' Q:D.e_e
FIGURE 4
Marjorie Rice's codification of information on the tiling pentagons of
types 1 through 8 and James's discovery.
see how they differed from one another. I listed the formulas [the equations
on sides and angles] on a 3 x 5 card and drew 10 pentagons on another card.
I drew lines in color within the pentagons to show the information in the
formulas, red for 360° combinations of 3 angles, blue for edges of the same
length, black for 360° combinations of 4 angles, greeen for other infor-
mation." "Now I could see that in [types] 7 and 8 each vertex had been
touched twice by a red or black line (the hooks counting for two times) and
that if I had drawn a line on [types] I or 2 between angles totaling 180°,
each vertex would be touched once by a line. Continuing to [type] 3, I used
the symbol "" to indicate 3 identical angles would come together, this
for 3 of the vertices. Three straight lines were needed then between [angles]
Band E. Every corner was touched three times. I saw that for every pattern
each vertex of the pentagon must be touched by a line or symbol the same
number of times."
This observation that each vertex of a tiling pentagon must be used
the same number of times in a tiling was the key to Marjorie's investigation
which followed. Her symbolic notation of information was further refined
to a form which suppressed all but essential information (Figure 5). "Here
I used [symbolic] pentagons in a form easier to draw." Lines connected
corners of the pentagon which would come together at a vertex of the tiling
by the pentagon. "This notation is the key to all my further work. By
labeling the corners (it didn't matter where I started), I could then develop
a sort of signature of letter combinations for every diagram and develop
these with little sketches."
~146~
J?p!',,-h,'" -
,iI "",,, ..
¥.- 3'OC' I( J.5tJp1fl ~,:zI,Fhr~"t
'I =3 ~o' 'f 3 S,m" - /#/;11",,,,1
.3 • 3'0 II' .3 .s#hY«' "~o"}
0
:3 ~ 3"" Y
.3 .slime -",(felt'"r
o
.3 : 3';0' A .3 """1r'''~t1r
r
3 • IIPOo. J -S,u"e-IdtfYp,,,,,!
.3 .. ,PO", .3 .set"'/!' (' ,;,,0)
.:J: 16'0°/ ~,.I/~re;lr
~·I'O· /I~ .st:lnrt!!.('"O)
!t!l]) A~EC
BSt!
"t . ~'f""'N/~ PS('d t(J ,;,~e"Tt'
61111f6S ,I' ;tsI~Nf'cQ/ /~ng'"
8Z1£
111\£ Il;D nd'~~ssa-~~ rei .,,,~ ~/I~,"
c,. t!I'" t'pr.n~'" jist!. r",
##1",,6,1.
/"It'";,,,e. re,r,ee5 c'f ~a;nf!.
5/2, /7t'(,us-try -I, Tht' ,P4l/en,
__ ;--t:;~.:>'e.rl'e5 e?""'It"'!
4' >'Alr,! "",11.4" ;1eCt!S5et'Y
f,,,
,
t"M/>,,#ern.(Or oTIr""
relartO>I,.JtII" s of eel9<!'j)
FIGURE 5
The pictorial notation developed by Marjorie Rice (here recording
information on the 9 types of pentagons known to tile) which was the
key to all her work.
In her own way, Marjorie had discovered a way to manage an

enormous amount of information that would emerge in a thorough
combinatorial search of possible combinations of angles (and sides) which
yield tiling pentagons. Mathematicians use symbols for objectivity, con-
ciseness and clarity-and good notation must be simultaneously suggestive
and definitive as well. Marjorie's pictorial notation looks like hieroglyphics
yet it records the possible combinations of angles with a simplicity not
possible with more conventional mathematical notation. The notation
eliminates completely the need to worry about repetition of cases caused by
assigning different letters to the angles of a pentagon.
Beginning with two copies of a single pentagon stuck together,
Marjorie considered how further copies of the pentagon could be added to
these to create a tiling of the plane. The information on how corners of the
pentagon met at a vertex of the tiling and how sides touched were all kept
track of by using her symbolic notation. If it became clear that a certain
combination was impossible (a tiling would not result), this case was
eliminated; when a tiling seemed possible she sketched an actual example
of such a pentagon and its tiling.
In order to make calculations quickly and test new pentagons to see
whether or not their angles might add up so as to create vertices of a new
pentagonal tiling, Marjorie arbitrarily divided 360° into units of 18°,
marking the divisions on a small protractor. Then using these units, an
angle of 36° was represented as 2, an angle of 108° as 6, and so on. "When
constructing a pentagon for trial I usually began with the 2 angles that
equal 180° numbered as 4 (72°) and 6 (108°) and adjusted them later on as
.a l47~
= 2D + C
2E + B = 360°
a=b=c=d
E e D
FIGURE 6
The first discovery (February 1976) ofa new type of tiling pentagon by
Marjorie Rice. The range of shapes that the pentagon can assume is
shown along with a tiling by one representative of this type. A tiling by
a representative of this type whose sides are in golden ratio is the
underlying grid for an Escher-like design of bees in clover, designed by
Mrs. Rice. (See Figure 16A.)
~148~
required. This was the busy Christmas [1975] season which took much of
my time but I got back to the problem whenever I could and began
drawing little diagrams on my kitchen counter when no one was there,
covering them up quickly if someone came by, for I didn't wish to have to
explain what I was doing to anyone. Soon I realized that many interesting
patterns were possible but did not pursue them further, for I was searching
for a new type and a few weeks later, I found it."
In mid-February, 1976, Marjorie sent her discovery to Gardner with
a sketch of the range of shapes which the pentagon could assume and tilings
by two different representatives of this new type of pentagon tile (Figure 6).
She wrote, "Here is a pentagonal tile that I believe really is different from
any you had listed though similar to type 7 and 8. One of the enclosed
examples in which the two sizes ofline are in golden proportion makes a very
pleasing arrangement, I think." Again, Gardner dispatched the discovery
to several interested parties via MG 2 including Kershner and myself. It
was verified to be indeed a new addition to the list of tiling pentagons;
Kershner wrote to Marjorie to ask how she discovered it, and acknowled-
ged that he had erroneously eliminated this as a possible type in his search.
As happens with a great deal of reader correspondence, the discovery was
not reported in Gardner's column but filed away for future reference.
(Other items had also been received by Gardner in response to the James
tessellation-at least one quilt and a beautifully woven rug were inspired
by the design. See color plate Y.)
When I examined the material that Marjorie had sent to Gardner
and compared it with Kershner's types 7 and 8, it appeared that these three
types of tiling pentagons (I named hers type 9) might all be examples in a
still larger class of tiling pentagons. I made the following conjecture and
sent it to Gardner: "Any pentagon having four equal sides and containing
four different angles P, Q., R, S, such that 2P + Q. = 360 0 and 2R + S = 360 0 ,
tiles the plane." In less than two weeks I received a letter from Marjorie in
which she showed that she had considered the conjecture and proved it
false. "As the symbols below show, there are only 8 possibilities-[8 ways
in which the corners of such a pentagon can come together so that only the
equations 2P + Q. = 360 0 and 2R +S = 360° are used]. Four of them 1, 5, 6,
and 8 can tile the plane, the other four seem to be impossible for the
reasons illustrated. 6 seems only to work when two adjacent angles equal
180° thus making it a type 1. It does however give two interesting ways of
assembling a special type kind of type 1 as shown on the enclosed sheet."
(Figure 7).
This was my first correspondence from Marjorie and my first
encounter with her notation and method of checking possibilities by
construction. It was so far from the conventional ways that mathematicians
use that I puzzled over the diagrams trying to figure out what she was
saying and how these proved anything. Her "reasons" that some pentagons
considered would not tile were little sketches, not algebraic or geometric
~ 149&'
~
" . ",
:
~"
O
I
U
A+",fO·
.,
I I,
6,' ..
III "A I I I '
71 '. : I: ::
I ..... I ~', II I D
,,.:.' .1t.f I,' I I 'til
~ I I or I~"
'
". ""1
.~~ 1~
",.,~ .. ~"".'
.lJ2· 11':"
FIGURE 7
The pictorial "proof" that Schatlschneider's conjecture was false.
Case 6, having 4 equal sides and 2 adjacent angles whose sum is 180°,
assembles into blocks in 2 distinct ways (note the change in the tiling
across the dashed line).
~150~
arguments that mathematicians require for proof. Probably because it

seemed so obvious to her, she had not enclosed any explanation of her
pictorial notation. In her notation, the heavy chicken tracks ~
represent the equations 2P+ Q = 360°, 2R +S = 360° which would bring
together these corners at a vertex of the tiling, and the lighter curves
connecting corners of a pentagon represent the remaining equation on how
corners must come together at a vertex of the tiling. Recall that the sum of
the interior angles of a pentagon is 540° so that this fact, together with the
two equations involving P, Q, R, S imply that Q + S + 2 T = 360°, where
T is the 5th angle of the pentagon. She assumed no further equations were
necessarily satisfied by angles of a pentagon and so a tiling by that
pentagon could use only the angle relationships represented symbolically.
Thus her cases 2 and 7 were eliminated because (as her arrow indicates)
another equation on angles would have to be added if the pentagons were
to tile in this way. Her cases 3 and 4 are impossible to construct-her
sketches tell this pictorially. It was not hard for me to verify algebraically
that no pentagon could satisfy the relationships on angles shown in these
cases. Marjorie had constructed several examples of her case 6 and always
2 adjacent angles seemed to add up to 180°. But this observation did not
constitute a mathematical proof. In trying to prove her observation, I
found that the assumed angle relationships did not force this fact. It was
Kershner who later supplied an elegant proof, showing the usefulness of his
generalized laws of sines and cosines. This provided a very striking
illustration of an amateur's intuition and observation using elementary
tools leading to a correct conclusion, but the necessity of more sophisticated,
mathematical means and a trained mathematical mind to provide irrefut-
able proof
The four cases which remained did indeed tile and were already
known-her cases 1 and 8 were Kershner's types 8 and 7 respectively. Her
case 5 was her new discovery (type 9) and case 6 was type 1. Thus the
conjecture was completely disposed of.
Now that Marjorie had established direct contact with Kershner and
with me, MG 2 was no longer an intermediary; however, Gardner was kept
informed of the continuing work on the problem. No doubt Marjorie was
encouraged by the praise for her work received in correspondence, but it
was the problem itself which continued to entice her. Although she was
busy with family events she kept returning to the puzzle, drawing tilings
and considering and reconsidering possibilities in whatever snatches of
time she could find. The problem was like a partially finished jigsaw puzzle
laid out in a spare room-worked on intently for a while until a small
satisfaction is achieved, then abandoned. It is not forgotten, however, and
lures you back again and again to tempt you to add a few more pieces and
see a little more of the pieced-together scene.
I asked Marjorie to write me all she had done on the problem and
keep me informed of any new results since I had been asked to write an
~151&
pI
AI
A6f - CD AC:O- 6E
FIGURE 8
Symbolic representation of pentagons in which each angle is "used"
once in a vertex of the tiling.
article on the pentagonal tiling problem for Mathematics Magazine. In

March, 1976, I received her codified analysis of how she had considered
ways in which a pentagon could tile. The diagram showed groups of
pentagons considered-each group corresponded to a set of angle re-
lationships satisfied by a pentagon. These angle relationships (and only
these) were to be reflected in an associated tiling of the plane by the
pentagon. How the angles came together then forced any conditions on the
sides of a tiling pentagon. The first group, called "p 1" (for pentagon-1),
considered pentagons in which each angle was "used" once in a vertex of
the tiling. Thus three different angles "came together" to sum to 360° and
the two remaining angles "came together" to sum to 180°. Only 2
pentagons, types 1 and 2, were in this category (Figure 8). The next 12
groups (listed 1 through 12 under a heading ''p2'') considered pentagons in
which each angle was used twice if different vertices of the tiling were listed.
Group 12 in this listing is the collection of pentagons which Marjorie had
written to me about earlier, in response to my conjecture. On this codified
listing (Figure 9), Marjorie explained "There are 3 tests, the first is whether
the group itself is possible (5 and 6 are not). Then sketches are made of
combinations indicated by each member of a group to see if the proper
angles will come together. If they combine successfully, it will be obvious
which lines [sides of pentagons] must be equal lengths. The last test is
whether it can be translated into specific angles-if it can, it will tile
successfully." Marjorie had written "no" next to each symbolic pentagon
in the list which would not tile in the way specified; next to those that would
tile, she put the type number from Kershner's list. In addition, for each
pentagon that tiled she produced an illustrative tiling showing that it could
tile in the way specified (Figures 9a, 9b). In her listing, types 1, 2, 4, 6,7,8,9
had occurred and she had 26 different tilings. Several of these tilings were
new. It is not surprising that types 3, 5, and James's tile (which I call type
10) did not occur on this list because if all of the angle relationships which
are satisfied at the vertices of tilings by these types are given, then each
angle of a pentagon tile is used three times.
A few weeks after receiving this information, Marjorie sent an even
more extensive list. "Regarding the 2-pentagon patterns I sent you
earlier- I have found on rechecking that I had missed quite a few-have
gone over them much more carefully and here is the revised list with
~152~
examples." Her revised list showed the same 12 groups but many more
cases considered. Now she had found 35 pentagons with angle relationships
that led to plane tilings. Some angle combinations yielded two or more
distinct tilings; thus there were a total of 45 sketched tilings accompanying
this listing. Although no new types of tiles were discovered, the range of
tilings was greatly expanded. Her letter indicated that she was already at
work on the "p3" groups of pentagons-those whose angle relationships
would use each angle of a pentagon three times. "The majority are quickly
@.~ ,,®~ I)~ -?~

I AAII "0
_I
",sz·lfac.
C/J-zjl
"' •• ·04 ....
f:1-CI}
..
....
....... -e6~(". A •• · . , .
... c-e..
.1 AAlv ... ~ ~~ ~~"

• I .(~ •• ·.teo..
• C-II
i'MII .. !Q\
.., .....,
.. ~1>··Ie.
~.,. .. ~/J
FIGURE 9
The 12 groups of pentagons considered which might form "2-
pentagon" patterns. Each angle of such a pentagon is "used" twice in
the list of angle relationships occuring at the vertices of the tiling.
Sketches of tilings are shown in Figures 9A and 9B for each successful
combination of angles. This listing and collection of tilings is the second
one made by Marjorie Rice.
~153&.
FIGURE 9A
seen to be impossible, so it doesn't look like such a formidable job to go

through those remaining .... Among them are types 3 and 5 and Mr.James's
new one [type 10)'''
In October, 1976, I received another bulging envelope from
Marjorie. She had made a new listing of all of the pentagon tilings she had
discovered thus far-58 in all. She had reorganized her listing; this time
she had arranged the pentagons (and associated tilings) into 12 classes,
each class corresponding to which sides of the pentagon must be equal. Six
pages of illustrative tilings demonstrated the thoroughness of her work.
e4154~
FIGURE 9B
Everyone of the ten types of tiling pentagons occurred, and there were
many new tilings. She closed her letter "this is as far as I can go with my
limited knowledge, so I am through looking. Perhaps there are still others I
have not come across."
In mid-November, I sent Marjorie a preprint of a paper by Branko
Griinbaum and Geoffrey Shephard which showed the 24 tile-transitive
tilings by pentagons of types I through 5 and also the first draft of my
article on the pentagon problem for publication in Mathematics Magazine.
This was an expanded version of a talk I had given at a Conference on
Recreational Mathematics at Miami University in Oxford, Ohio. John H.
Conway had been at that conference and showed great interest in the
problem and the contributions made by James and Rice. He admitted that
he had once set out to find all tiling pentagons but had abandoned the
.8155~
problem when it became too time-consuming. At the end of my article I

raised some natural questions: "Is the list of pentagons which can tile in an
edge-to-edge manner complete?" "Can we find the complete list of all
equilateral pentagons which tile?"
Marjorie couldn't ignore the questions. I received another letter in
December, 1976. "Had thought to spend no more time on pentagons but
they weren't so easy to lay aside." This time she had tackled the questions
raised in my article. In reply to the last question, she had sketched all of her
tilings by equilateral pentagons and the angle relationships forced by the
tilings. She had also tackled the first question. In the article, I had
explained "block-transitive" tiling by pentagons, noting that all of the
recent discoveries-Kershner's, James's, and hers were pentagons which
could not tile isohedrally, but for which a block of two or three stuck-
together pentagons was necessary to generate the tiling. This notion was
new to Marjorie and she reexamined all of her tilings which began with
two pentagons stuck together and noticed that "most of them consist of 4
tiles forming 2 hexagons which tile in one of 6 ways." Focusing in on one
possible dissection of such a block into four congruent pentagons had led
her to discover several new edge-to-edge tilings. Two weeks later
(December 27, 1976-the Christmas season seemed to be her most
creative time!), came the exciting news: "I have been working with this
idea further and have some new patterns, and to my surprise and delight-
2 new types, closely related." Indeed, she had discovered types 11 and 12,
and the accompanying ti1ings were very striking (Figure lO).
The discovery of the new types had been an unexpected bonus in her
methodical analysis of 2-b10ck transitive designs. She had found that the
double hexagons could be dissected into four congruent pentagons in nine
distinct ways (Figure 11) and these dissections together with the variety of
C)'
A f'o,
e E e ..t.
..t.
(!.
P <'- D
,]) ='()~ ,]) :,0"

./3~E:18(Jo ,sto~-~/IO"
A 1'", toE "Jr.,," A"'A fE =3';""

f. +-f! 1'8 ".3r.(J0 I!. ~t! ~B "3,,,0
.a. .. ,6 &t 1't! =6
..tl.~..e"c 'b .A ,....e =b
FIGURE lO
~156~
IN PRAISE O f AM
ATEURS
FI G U RE 10 (con
Ti lin g pentagons t'd)
of types 11 an d 12
discovered by M
D ec em be r, 1976. ar jo rie Rice in
s
FIGURE 11
Consideration of how "double hexagon" blocks (represented here
symbolically) could be dissected into four congruent pentagons and
the ways in which such double hexagon blocks can tile, produced a
wealth of new tilings by pentagons .
.8158~
tilings by their blocks, led to over 50 different tilings by pentagons (the

tilings were "2-block transitive"). Through the spring she continued to
pursue the idea and found several dissections leading to 3-block pentagon
tilings and even some 4-block pentagon tilings.
By now the original MG 2 circulation of information on the pentagon
problem had grown and information was traveling to three continents.
Another group of amateurs-II th year school children in New South
Wales, Australia, had spent a week investigating the problem of discover-
ing convex equilateral pentagons which tile and the ways in which they
tile. Led by their teachers, George Szekeres and Michael Hirschhorn, they
had made good progress on the problem. One particular equilateral
pentagon was capable of a great variety of tilings (this was the special tile
"case 6" of Marjorie Rice's first correspondence with me). Hirschhorn had
discovered many unusual tessellations using this pentagon, including two
beautiful central tessellations with just six-fold rotational symmetry
(Figure 12).
In the summer of 1977 I provided Marjorie with the article "The 81
Isohedral Tilings of the Plane" by Branko Griinbaum and G. C. Shephard.
The paper contained careful sketches illustrating the 81 distinct types of
tilings and I felt that Marjorie might find among these some whose tiles
could be dissected into congruent pentagons, thus creating new pentagonal
tilings. Throughout the fall, she continued her previous work and also
utilized the Griinbaum-Shephard article in a far more sophisticated
manner than I had anticipated. "The isohedral tilings by Griinbaum and
Shephard were of much interest. I have copied them into a four page
version I could more easily use." She had, in fact, represented each of the
FIGURE 12
Michael Hirschhorn's central tessellation by an equilateral pentagon,
fashioned into an engraved silver pendant.
~159~
IH25 m26 m27 m28
...
FIGURE 13
(U pper half) Several tilings by shaped tiles as they appear in "The
81 Types ofIsohedral Tilings in the Plane", by Branko Griinbaum and
G. C. Shephard (Mathematical Proceedings of the Cambridge Philo-
sophical Society. September, 1977. pp. 190-191). (Lower half) The
same tilings as redrawn in symbolic form by Marjorie Rice.
~160&.
Griinbaum-Shephard tilings of shaped tiles by a symbolic marked tiling

showing only the topological network of the tiling and the action of the
symmetry group on the tiles (Figure 13). Using these, she had reanalyzed
all of her previous tilings and discovered several new ones.
It was Christmas season, 1977, when she sent me a fat envelope
containing her voluminous work. Again, there was a Christmas surprise.
"] ust a couple of weeks ago, this new type of pentagon turned up (and 1
thought there would be no more). This one (Figure 14) like type 4 has two
opposite 90 0 angles but the requirements for the lengths of the sides are
different." Her illustrative tiling by this new type 13 showed an interesting
pattern of interlocked bow ties, made up of 4 of the pentagons. 1 had just
received the galley proofs of my article on the pentagon problem and so was
able to insert this latest discovery before publication. (The article, in the
] anuary 1978 issue of Mathematics Magazine, contains a fuller account of the
mathematical details of pentagons which tile.) Still Marjorie's work on the
problem did not end -she was determined to try to see if she could prove
that all pentagons which could tile had been found. Although her attempt
at proof was not complete, her thorough combinatorial check ofa1l2-block
and 3-block patterns reduces the remaining task considerably. Perhaps by
the time this story is published the question as to whether there are still
other pentagons which tile will be answered. Hirschhorn is confident that
all equilateral convex pentagons which tile have been found (these are
described in the Mathematics Magazine article); his argument utilizes a
computer in a proof by elimination of possible angle relationships.
What makes a person pursue a problem so steadfastly as Marjorie?
She was not trained to do this, nor paid to do it, but obviously gained
personal satisfaction in her patient and persistent search. No doubt her
personal history is like that of many amateurs who are inspired by
Gardner's writing.
She was born in 1923 in St. Petersburg, Florida, a first child. At age 5,
she began school in Garden Valley School, a one-room country school with
grades 1 through 8, having a total of about two dozen pupils. "My mother
wished me to have a good start and had taught me well at home so 1 was
placed in the second grade." She was a shy child, loved to read and "could
easily become absorbed in a book or in my daydreams and forget the world
around me." She did well in school; "arithmetic was easy and 1 liked to
discover the reasons behind the methods we used." "I was interested in the
colors, patterns and design of nature and dreamed of becoming an
artist .... " Her later years at the school "were enriched by two very fine
teachers, Miss Keasey and Miss Timmons ... [who] helped make up for the
deficiencies of a small country school."
"When 1 was in the 6th or 7th grade our teacher pointed out to us one
day the Golden Section in the proportions of a picture frame. This
immediately caught my imagination and though it was just a passing
incident, 1 never forgot it. I've continued reading on a wide variety of
.6161 ~
B = E = 90·
2A + D= 360·
2C + D = 360·
a =e
a+e=d
FIGURE 14
Tiling pentagon oftype 13 discovered by Marjorie Rice in December,
1977.
subjects over the years and have been especially interested in architecture
and the ideas of architects and planners such as Buckminster Fuller. I've
come across the Golden Section again in my reading and considered its use
in painting and design. A book that was especially helpful and inspiring to
me in this regard was The Geometry of Art and Life by Matila Ghyka."
Marjorie's interest in art continued-she became especially interested in
textile design and the works ofM. C. Escher. As she pursued the problem of
pentagons and their tHings, she produced some beautiful geometric designs
and imaginative Escher-like patterns (Figures 15, 16).
While in high school the family moved to Pine Castle, near Orlando,
~162~
3 - y'1 y'1- 1
FIGURE 15
A beautifully symmetric tiling by Marjorie Rice which radiates
outwards from the center and can be continued to fill the plane. Two
pentagons, each having two 90° angles and three 120° angles, create the
design.
Florida. At Orlando senior high Marjorie studied shorthand and typing to

prepare for future employment and did poorly at both. She regretted that
she could not take mathematics beyond the required general course. After
graduating from high school at 16 she was employed first in the office of a
laundry, then in a small printing firm until her marriage to Gilbert Rice in
1945. "During those years I frequented the public library and learned
much about science, psychology and other subjects I had missed out on in
school. I also started a correspondence course in commercial art .... " After
the Rices' first son was born they moved to San Diego, California. "These
were very busy years for us both" -an understatement to be sure. The
Rices raised five children and Gilbert started his own trade typesetting
shop. Marjorie was drawn back into mathematics by her children. "When
my oldest son, David, was in junior high, ... the 'new math' was just
beginning to be used .... I wanted to study his lessons and keep up with him
~163~
FIGURE 16A
Underlying grid for Bees in Clover (see Plate I.)
FIGURE 16
Three Escher-like plane-filling designs by Marjorie Rice are based on
the geometric grids of some of her unusual tilings by pentagons.
(M. C. Escher used a well-known tiling by pentagons as the underlying
grid for some of his plane-filling designs.)
as he learned and he encouraged me to do this-but these were busy days.

I soon fell behind and gave up the idea. My interest in his lessons continued
however and I could often find solutions to his problems by unorthodox
means, since I did not know the correct procedures. He shared with me the
mathematical games he learned i:n class, such as Hex and three dimen-
sional tic-tac-toe .... "
~164~
FIGURE 16B
Underlying grid for Fish (see Plate II.)
"J enjoy puzzles of all kinds, crosswords, jigsaw puzzles, mathemati-

cal puzzles and games, and have purchased books of mathematical puzzles
over the years. Those of a geometric nature are a special delight. Thus
when my son's Scientific American arrives I first turn to Martin Gardner's
"Mathematical Games" section." Her absorbing fascination with such
puzzles and keen perception of shapes, proportions and designs is tellingly
revealed in her account ofa recent trip. "In November, 1974, my husband
and I started on ajourney which would take us around the world .... I had
~165~
FIGURE 16c
Underlying grid for Hibiscus (see Plate IlL)
much interest and curiosity concerning designs and proportions that were
different and unfamiliar. I looked for such things wherever we traveled,
taking notes of proportions of houses, doors, windows, division of space, the
designs of grilles over windows .... I especially enjoyed the delightful and
colorful textile designs often in big bold patterns that were often worn in
Ghana and Nigeria." "We kept our luggage to a minimum on this trip ...
but I did slip in a small paperback book, Work This One Out, 105 puzzling
brain-teasers by L. H. Longley-Cook. Thinking on these problems helped
pass the time quickly when we had long periods of waiting .... "
The mind and spirit are the forte of all such amateurs-the intense
spirit of inquiry and the keen perception of all they encounter. No formal
education provides these gifts. Mere lack of a mathematical degree
separates these "amateurs" from the "professionals". Yet their dauntless
curiosity and ingenious methods make them true mathematicians. Martin
Gardner has awakened many such mathematicians.
~166~
PLATE I
Bees in Clover by Marjorie Rice.
PLATE II
Fish by Marjorie Rice.
PLATE III
Hibiscus by Marjorie Rice.
PLATE IV
Bees and Wasps by M. C. Escher
PLATE V
A hand woven rug (created by Australian A. G. Bomford of Can-
berra, Australia and reproduced here with his permission) captures
Richard James's design. The rug is 2.03 by 1.31 meters, 46,986
stitches.
SODle ProbleDls on
Plane Tilings
-~-
Branko GriinbaulIl
UNIVERSITY OF WASHINGTON
G. C. Shephard
UNIVERSITY OF EAST ANGLIA
Although the art of tiling is as old as human history, the science of tiling
seems to have been curiously neglected until recent times.
The advanced state of the art in the Middle Ages is demonstrated by
the remarkable tilings on many mosques and other Saracenic buildings-
the fourteenth-century Spanish palace known as the Alhambra is one of
the prime examples (Figures 1 and 2). There is little doubt that it was a
visit to the Alhambra that helped to inspire the Dutch artist M. C. Escher
in some of his well-known creations (Figure 3).
The first treatment oftilings from a mathematical point of view seems
to be that of the German astronomer Johannes Kepler (1571-1630)
(Figure 4) in his book Harmonice Mundi published in 1619 (Figure 5).
Kepler's geometrical discoveries were so overshadowed by his contri-
butions to physical astronomy that, incredible as it may seem, they were
mostly forgotten for nearly three hundred years. Apart from Kepler's book,
very little work on tilings seems to have been done before the end of the
nineteenth century; so the science of tilings, by which we mean the active
investigation of their mathematical properties, is barely a hundred years
~167~
FIGURE 1
A view of part of the Alhambra palace at Granada in Spain. This dates
from the fourteenth century and, like many other Saracenic buildings,
is remarkable for the wealth of tilings used as decoration. A few of the
tilings from Alhambra are shown in Figure 2.
FIGURE 2
Eight tilings from the Alhambra. These sketches were made by the
Dutch artist M. C. Escher when he visited the Alhambra in 1936. The
subject of tilings had a profound effect on the artist's later work.
~168~
SOME PROBLEMS ON PLANE DLINGS
FIGURE 3
One of Escher's less well known tilings. In Escher's work there are
many examples of tilings in which the tiles resemble animals or
inanimate objects.
FIGURE 4
The famous astronomer Johannes Kepler, whose pioneering research
of tilings was forgotten for almost three centuries.
e.t 169~
FIGURE 5
Kepler's book Harmonice Mundi, which appeared in 1619, contains the
first mathematical treatment of tilings ever published. The plates from
this book, reproduced above, show some of the tilings by regular and
star polygons which Kepler investigated.
old. During this century a large amount of material has been published,
much of it by crystallographers, engineers and other non-professional
mathematicians. The subject still abounds with open problems; we shall
draw attention to some of them in the following pages.
Recently, with an increased awareness of the educational merits of
geometry, the investigation of tilings has come to be regarded as an
appropriate mathematical activity in schools and colleges. It is interesting
to note that periodicals intended for mathematics teachers seem to be
devoting more and more space to this topic. This resurgence of interest can
be partly attributed to Martin Gardner's beautiful articles in the Scientific
~170~
SOME PROBLEMS ON PLANE TILINGS
American. By reporting new discoveries, he has repeatedly stimulated the

interest of his readers in the subject. In the following pages we shall have
occasion to mention several of these articles and we hope that the gratitude
which the mathematical community should feel towards Martin Gardner,
the expositor par excellence, will be made apparent.
1 What is a tiling? From a mathematical point of view it is a family ff
= {T1 , T 2 , ••. ,} of closed sets ~ (the tiles) which cover the plane without
gaps or overlaps of non-zero area. As it stands, this definition is much too
general, and throughout the article we shall restrict attention to tilings for
which the tiles T i , i = 1,2, ... , are topological disks; that is, each is
obtained from a circular disk by a continuous deformation. In particular,
each tile is connected and simply-connected. Thus tiles which are
fragmented into two or more disjoint parts, or have holes in them, are not
permitted. Some of the commonest tilings are those in which the tiles are of
a limited number of different shapes. This situation is conveniently
described in the following way. Let 9' = {P 1 ,P2 , ••• ,Pk } be a (finite) family
of closed sets such that every tile T; in ff is congruent to one of the Pi
(i = 1, ... , k). Then 9' is called a set of pro to tiles for ff and we say that 9'
admits the tiling ff. If 9' contains exactly k distinct sets (by which we mean
that no two are congruent) and all of these are used in ff, then ff will be
called k-hedral.
When k = 1 we call the tiling monohedral. Such tilings are very
familiar. Examples are provided by the regular tilings (Figure 6) which
have been known from time immemorial. The uniform tilings (Figure 7),
known to Kepler (see Figure 5) and probably discovered by him, 2-hedral
(dihedral) or 3-hedral (trihedral). Many other k-hedral tilings with small
FIGURE 6
The three regular tilings. These familiar tilings have been known since
time immemorial.
.a171&.
FIGURE 7
The eight uniform tilings that are not regular. In each, the tiles are
regular polygons and there exist symmetries of the tiling which map
any chosen vertex onto any other vertex. One of the tilings (denoted
by (3 4 .6)) occurs in two mirror-image forms as shown. It is believed
that as a group, these tilings were discovered by Kepler in the early
seventeenth century.
FIGURE 8
Are these proto tiles of monohedral tilings? I t is by no means easy to
answer this question.
values of k are known, such as those shown in Kepler's drawings K, T, X

and Aa of Figure 5.
It might be thought that monohedral tilings, in which all the tiles are
the same shape, are mathematically trivial and therefore, hardly worth
investigating. But this is far from true. A few of the great variety of
possibilities are shown in Figures 21, 23, 28 and 29. To convince the reader
of the difficulties, we show a few prototiles in Figure 8. For some of these it
is not known with certainty whether or not they admit monohedral tilings,
and in every case it is by no means easy to settle the question. Problems of
this type were discussed by Martin Gardner in his article in Scientific
American (August 1975). The basic question is the following:
.e:A 173 ~
PROBLEM I Is there any well-defined procedure (or algo-

rithm), not employing trial-and-error, for testing whether or not
a given prototile P admits a monohedral tiling of the plane?
This is a very difficult problem, and to discuss all its ramifications
would take us deep into the subject of mathematical logic. Although this is
not appropriate here, in Section 5 we will briefly note the connection
between this problem and that of finding aperiodic tiles. It is worth
mentioning that one of the few conditions known for testing whether a
given prototile admits a tiling or not is the so-called Conway criterion (see
Figure 9). This is of surprisingly wide application, but it is only a sufficient
condition for a tiling to exist and is not necessary. Therefore, it does not
provide a solution to Problem I.
The depth of the problem as well as the extent of our ignorance is
shown by the fact that we even do not know all the shapes of polygons that
(a)
b
(b)
c
d
FIGURE 9
Is it possible to tile the plane with the tile shown in (a)? This question is
answered in the affirmative by the Conway criterion which says that a
prototile admits a monohedral tiling if it is possible to divide its
boundary into six parts (as indicated by the black circles in (b)) in such
a way that parts a and d are translates of each other, while the other
four parts b, c, e and f, are curves each of which has a center of
symmetry (indicated by an open circle. In the tiling four of the six
adjacents of the given tile are obtained by a 180 0 rotation of the given
tile about the open circles, and the other two are translates of the given
tile. The criterion applies to any tile-not just to polygons like the tile
shown here.
~174~
2rr/3
x
a
rr-x-y
a
y
x
FIGURE 10
The three types of hexagon that admit monohedral tilings of the plane;
in fact each will also tile isohedrally. These types were discovered by
K. Reinhardt in 1918. Relations between angles (marked x,y, z, t) and
between sides (marked a, b, c) are indicated.
admit monohedral tilings. Every triangle and every quadrangle admit such
tilings, as do three families of hexagons (see Figure 10). (A family is
defined as a set of polygons satisfying certain stated relationships between
its sides and angles.) But the number of families of pentagons is still
uncertain. At the present time, thirteen families of pentagons that admit
tilings are known (Figure 11), but whether or not this list is complete is still
not settled. Here again Martin Gardner made a contribution to the subject
by publishing, in Scientific American (July 1975), what was thought at that
time to be a complete list of pentagons-namely, those discovered by K.
Reinhardt in 1918 and by R. B. Kershner in 1968. Several readers pointed
out that the list was incomplete, and the new tiles that turned up are
included in Figure 11. This leads us to our second question.
PROBLEM 2 Do there exist any families of convex pentagons,
other than those shown in Figure 11, that admit monohedra1
tilings of the plane?
Recently D. C. Hunt and M. D. Hirschhorn claim to have proved
that a list of equilateral pentagons that tile the plane, compiled by Doris
Schattschneider in 1978, is complete. However, their proof has not yet
been published.
2 Even if a prototile admits a monohedral tiling of the plane, there is no

a priori method of deciding in how many different ways it will tile. Let us
say that a tile is monomorphic if it admits a unique tiling. The commonest
example of a monomorphic tile is the regular hexagon-the only tiling
possible is the regular tiling shown in Figure 6. On the other hand, a square
is not monomorphic. In fact, by "sliding" rows of squares relative to one
another, it is evident that a square admits an uncountable infinity of
different tilings. In Figure 12 we show some less familiar examples of mono-
morphic tiles, and in considering these an important question of definition
e,t175&.
a
1112 b a II-X a+b
I X
b T(a + b)
a
L.._ _ _ _ _ ~ II-X 11/2
1112
2n13
a X
211/3
II-X
Y
a a
a 211/3
X
211-2x
211 - 2y
x
II-X 2a+b 211-2y

a
a
1112 I 211-2x
1112 11- TX
II -
C)~
~xa
2a a
X~---../
11/2 I
11- TX
FIGURE 11
The thirteen types of pentagon that are known at present to admit
monohedral tilings of the plane; the first five types also tile isohedrally.
The list is taken from Schattschneider's paper (Mathematics Magazine,
51 (1978), pp. 29-44) in which further details and examples of the
tilings can be found.
~176~
(e)
(0
(g)
(b) (i)
FIGURE 12
Nine examples of monomorphic prototiles.
~177~
FIGURE 13
The prototiles (e) and (f) of Figure 12 admit the tilings shown above.
Since the two possible tilings with each are mirror images of one
another, the tiles are regarded as monomorphic.
arises. The tiles (e) and (f) of Figure 12 each appear to admit two ti1ings
(Figure 13), but in each case these are mirror-images of each other. Should
we count these as different? The most appropriate answer is a matter of
opinion, but for various reasons we prefer not to distinguish between mirror
images and so we assert that the tiles (e) and (f) are monomorphic.
Do there exist tiles which are dimorphic, that is, tiles which admit
precisely two monohedral tilings? The answer is that such tiles exist and
examples are shown in Figure 14. A trimorphic prototile, and the three
corresponding tilings are shown in Figure 15; this example is unique in the
sense that all other known trimorphic prototiles differ trivially from this
one. The discovery of dimorphic and trimorphic prototiles is not easy, so
we propose the following.
PROBLEM 3 Find additional, essentially different, examples of
dimorphic and trimorphic prototiles.
PROBLEM 4 Do r-morphic prototiles exist for any finite value
of r ~ 4?
The following IS a more technical question of considerable theoretical
interest.
~178~
FIGURE 14
Two dimorphic proto tiles and the corresponding tilings .
.6179~
PROBLEM 5 Does there exist a tile which only admits a

countable infinity of monohedral tilings?
Thus Problems 4 and 5 are asking whether the large gap in values
between r = 3 and an uncountable infinity, can be bridged.
If we try to extend these ideas to k-hedral tilings (those in which k
shapes of tiles are used) then new considerations enter. Let us consider k-
hedral r-morphic prototiles (k ~ 2, r ~ 1). Their existence was established
by H. Harborth in 1977. He pointed out that a suitable rhomb and a tile
which is a union ofrhombs (see Figure 16) solved the problem for k = 2,
and hence for k ~ 2. (This was done by cutting either of the tiles into an
appropriate number of pieces in such a way that these pieces can only be
fitted together again by reconstituting the original tiles.) However, the
construction of Figure 16 does not solve the problem completely; it does so
only if we insist that in all the tilings, copies of every prototile must be used.
Without this condition the example fails because a rhomb on its own
admits an uncountable infinity of tilings. We therefore pose the following
question.
FIGURE 15
A trimorphic proto tile and the corresponding tilings.
~180~
(a) (b)
(c) (d)
FIGURE 16
Harborth's construction for a pair of prototiles which admit exactly r
tilings (assuming that both kinds of tile occur in each tiling). In this
diagram, r = 4. One tile is a rhomb with angles 2n/p and (p - 2)n/p
where p = 6r - 7, and the other is obtained by fusing together 2r - 2
rhombs as shown.
PROBLEM 6 For every value of k ~ 2 and r ~ I does there

exist a set g of k prototiles such that exactly r distinct tilings are
admitted by g even if copies of all the prototiles in g need not
be used?
An example that solves Problem 6 in the case k = r = 2 is shown in
Figure 17.
3 It is well known that in 1900, D. Hilbert posed a series of problems
that have had a major effect on the development of mathematics. One of
these-the eighteenth problem-was concerned with tilings. The follow-
ing is a brief account of that problem.
To state the problem we need to use the word isohedral. This is defined
as follows. Every tiling ff has a group S(ff) of symmetries, that is, rigid
motions of the plane that map ff onto itself. (See, for example, Figure 18,
where several different kinds of symmetry are indicated.) An informal way
of thinking of a symmetry (not without reservations as to its mathematical
validity) was described by Fourrey in 1907. He suggested tracing the tiling
FIGURE 17
Two dihedral tilings using the same two pentagonal prototiles. These
are the only tilings possible, so the pair of pentagons forms a dimorphic
set. Notice that here neither prototile on its own admits a tiling.
FIGURE 18
A tiling with, in the left part of the figure, its symmetries (except
translations) marked. Triangles represent centers of 3-fold rotational
symmetry, solid lines represent lines of reflection, and dashed lines
represent glide-reflections. Any vector connecting two of the solid
triangles is a translational symmetry.
~182~
FIGURE 19
Two monohedral tilings using the same prototile. The right tiling is
isohedral because, given any two tiles, there exists a symmetry of the
tiling which maps one onto the other. The left tiling is not isohedral
since, for example there is no symmetry which maps the tile A onto the
tile B.
onto a transparent sheet and then every symmetry of the tiling corresponds
to a motion of the sheet (including the possibility of turning it over) such
that the tracing again fits exactly over the tiling. A tiling :Y is isohedral if,
given any two tiles ~, ~ of :Y, there exists a symmetry of:Y that maps ~
onto ~. Every isohedral tiling must be monohedral-the distinction
between the two concepts is illustrated in Figure 19.
Hilbert's eighteenth problem can be stated in the following way.
Does there exist a prototile which admits a monohedral tiling but no
isohedral tiling? In fact the problem was originally posed in a three-
dimensional context, and there is reason to suppose that in the plane case
Hilbert believed the answer to be trivially in the negative. But he was
wrong! In 1935 H. Heesch found a counterexample (see Figure 20), that is,
a tile which admits an infinity oftilings, but none of these tilings is isohedral.
Since then, other examples have been found; some of the convex pentagons
of Figure 11 have this property.
At first it may appear that the difference between monohedral and
isohedral tilings is slight-but this is far from the truth. For example, if we
restrict attention to isohedral tilings, then Problems 1, 2, 4 and 6 can be
solved. In particular, in connection with Problem 2, H. Heesch and
..0183~
(a)
(b) (c)
FIGURE 20
Heesch's tile, shown here, was the first to be discovered that admits a
monohedral tiling but no isohedral tiling. Many other examples are
now known-for example several of the pentagons in Figure 11 have
this property. In (a) we show how the tile is built up from squares and
half-squares, and in (b) and (c) we show examples of the (infinitely
many) possible tilings using this prototile.
O. Kienzle gave in 1963 a complete list of polygons that admit isohedral

tilings. This is possible because all isohedral tilings can be described. It can
be shown that there are 81 types of which 4-7 can be realized using
polygons. Some examples are shown in Figure 21. The actual method of
classification into types is too technical to describe here; it has only recently
been satisfactorily resolved in a general context. (This fact probably
explains why many of the older papers dealing with classification problems
contain many obscurities and errors.)
4 One of the most remarkable theorems concerning tilings is the
Extension Theorem. A special case of this is known as Wang's Theorem
(named for its discoverer), but the general case has only recently been
proved and no proof has yet been published. Suppose that f/ is a given
(finite) set of prototiles, each one a topological disk, and suppose that,
however large R is, we are able to tile over a circular disk DR of radius R
using the tiles of f/. Then the Extension Theorem asserts that f/ admits a
tiling of the plane. The phrase tile over needs clarification. It means that
it is possible to construct a set (or patch) of tiles in such a way that their
.8l84-&.
overlaps have zero area, yet they completely cover a region which includes
DR' For this theorem to be true, it is essential that the prototiles be bounded
and finite in number. For example, in Figure 22 we show a single prototile
copies of which will tile over arbitrarily large disks, but which does not
admit a tiling of the plane. In this case, of course, the theorem does
not apply since the prototile is unbounded.
Some of the consequences of this theorem are surprising. For
example, it implies that if we can tile over a quarter-plane using the tiles of
fI', then we can tile the whole plane. At first sight it may appear obvious
that if one can construct larger and larger patches of tiles, without limit,
then eventually one will arrive at a tiling of the plane. But this becomes
much less clear when one realizes that as R increases, it may.be necessary to
2 3
4 5 6
7 8 9
10 11 12
FIGURE 21
Twelve examples of isohedral tilings.
~185~
--
/ \
\
"- ........ - ./
FIGURE 22
Here each tile is a semi-infinite strip ending in a semicircle. Such tiles
will clearly tile over arbitrarily large circular disks, but will not admit
a tiling of the plane. This example shows that for the truth of the
Extension Theorem it is necessary to assume that the tiles are bounded.
continually rearrange the tiles and it may happen that none of the patches
used to cover the disks can appear as part of the final tiling.
A related problem is known as Heesch's problem. It is known that
there exist prototiles P which do not admit a tiling of the plane, yet P can
be completely surrounded by copies of itself (see Figure 23). By completely
surround P we mean that we can construct a ring R of tiles around P (leaving
no gaps) in such a way that each point of P is at a distance greater than
some fixed positive number from every point ofthe plane that is outside R.
By the phrase P is completely surrounded twice we mean that P is surrounded by a
ring R, and then R is surrounded by a second ring R'.
PROBLEM 7 Does there exist a proto tile P which does not
admit a tiling of the plane yet can be completely surrounded
twice by copies of P?
More generally it is clear what we mean when we say that a tile P can
be surrounded r times, and if r is the largest such integer it is called the
Heesch number for the tile P. .
PROBLEM 8 Do any tiles exist with Heesch number r = 3,4;

5, ... ?
5 A tiling ff is called periodic if there exist two non-parallel translations

which are symmetries of ff. Examples of periodic tilings appear in Figures
13, 15 and 18; in each case one can think of the tiling as being made up ofa
small patch of tiles repeated in a lattice arrangement. A set of prototiles 9'
is called aperiodic if it admits a tiling of the plane, but no such tiling is
periodic. Until recently no aperiodic sets were known; the first example
was discovered by R. Berger in 1966, and other examples have been found
.8186&.
SOME PROBLEMS ON PLANE TrLINGS
(a)
(c)
FIGURE 23
Heesch's tile shown in (a) has the remarkable property that it can be
completely surrounded by copies of itself, yet admits no tiling of the
plane. Even so, extensive patches can be built up, such as that in (c).
The shape of the tile is indicated in (a) -it consists of the union of a
square, an equilateral triangle, and a half equilateral triangle as shown
by the dashed lines.
by Robinson, Penrose, Ammann and others. Figure 24 shows a tiling using

the first set of aperiodic tiles discovered by R. Penrose.
For a delightful account of Penrose's aperiodic tiles known as kites
and darts, the reader should consult Martin Gardner's article in the
Scientific American of January 1977. These tiles have many remarkable and
unexpected properties, not all of which have been adequately explained.
Since 1977, several new sets of aperiodic tiles have been discovered by
Robert Ammann, to whom we are grateful for permission to reproduce the
tiles shown in Figure 25.
The subject of aperiodic tilings is much too big to deal with here. The
history of the subject and its connection with mathematical logic is a
fascinating chapter in the history oftilings, a chapter which is still not fully
written. Here we can do no more than mention these few facts. However,
no article attempting to survey the properties of tilings would be complete
without mentioning the main outstanding problem in this area.
PROBLEM 9 Does there exist a single prototile P which forms
an aperiodic set?
In other words, P may admit many monohedral tilings of the plane,
but no such tiling is periodic. I t can be shown that this problem is, sur-
prisingly, related to Problem 1, or at least H. Wang has shown a connection
between aperiodicity and the tiling problem, that is, the determination of an
~187~
(a) (b)
iJ$O Ob
oo© 0, O·~ 0 o 0 I I 0
00
_
0
2' 0 i
(c)
FIGURE 24
The first set of aperiodic tiles to be discovered by Roger Penrose. The
prototiles are shown in (a), and in (b) we indicate how the projections
and indentations can be replaced by numbers which give an equiva-
lent "matching condition": 0, 1, 2 must fit against 0, T, 2" respectively.
In (c) we show an example of the tiling.
~188&.
o \ L -_ _----'
FIGURE 25
Two sets of aperiodic tiles recently discovered by R. Ammann. Each
set of three tiles contains a small "key tile" which fits into indentations
in the larger tiles, and so restricts the ways in which the latter can abut
against each other.
algorithm for deciding whether a given set f/ of prototiles admits a tiling.

Wang's results only apply to a very special kind of "square tile with colored
edges", and we do not know to what extent similar considerations apply to
tiles of general shapes. However, it seems likely that an affirmative solution
to problem 9 would imply a negative answer to Problem I.
6 We conclude with some curiosities. The first of these is the enclosure
problem. In 1934 K. Reinhardt asked whether it is possible for two tiles-
each congruent to a proto tile P-to enclose another copy of P. In 1936 this
problem was solved by H. Voderberg, who produced the example shown in
Figure 26, which admits a periodic tiling. This prototile has the strange
property that not only do some tiles enclose a third, but other pairs of tiles
enclose two copies of P. We can say that Voderberg's tile has the 2-
enclosure property-more generally, a tile P is said to have the r-enclosure
property if two copies of P enclose a region equal to the union of r non-
overlapping copies of P. It is remarkable that however large r is chosen,
tiles with the r-enclosure property exist. A construction for such tiles is
shown in Figure 27. However, these tiles do not solve the following
(probably not very difficult) question.
~189~
(a)
(b)
(c)
FIGURE 26
Voderberg's tile. This has the remarkable property that two copies of
tile can enclose a third (b) or even two other copies (c). In (a) a
periodic tiling is shown using the Voderberg tile as prototile .
.a190~
PROBLEM 10 Does there exist a tile P with the r-enclosure

property (r ~ 3) and which admits a tiling of the plane?
In his paper Voderberg remarks, almost casually, that his tile admits
a tiling with "spiral form" (see Figure 28). The appearance of this is
sufficiently striking for it to have attracted considerable attention. In
January 1977, Martin Gardner published a diagram ofVoderberg's spiral,
and also Michael Goldberg's explanation of how it is constructed. Once the
method is explained, it is obvious that the spiral appearance has nothing to
do with the enclosure property; moreover, it is easy to find many different
tiles that admit spiral tilings.
However, all spirals that can be constructed by Goldberg's method
necessarily have an even number of arms, that is, rows of tiles spiralling
""B,.--_ _---, F
D ------
"
"".....--
E
"" "",
--~----
""" A
EL-----~
FIGURE 27
The construction of a tile with the r-enclosure property (in the figure, r
= 8). First construct a zigzag line ABCD with right angles at Band C
and the points ABCD on four equidistant parallel lines, shown dashed.
DE is an arc centered at A and FA is an arc centered at D. EE' is one
fourth of the arc ED. Denote AFBCE by S and let S' be the result of
rotating S about A until E coincides with E'. Then the tile bounded by
S, S' and EE' has the 8-enclosure property as shown in the right-hand
diagram. For values of r other than 8 the same construction is carried
out with E' chosen so that the arc length DE is [!(r + I)] times that of
EE'.
~191~
FIGURE 28
Voderberg's spiral tiling. This uses the same prototile as the tiling in
Figure 26.
outwards from the center. Recently spirals with an odd number of arms
have been found and examples appear in Figure 29. The tile used here
has been called versatile since it admits many other unusual tilings (see
Figure 30). However, there are still several open problems. For example, in
Figure 29 we notice that about half the copies of the prototile are
reflections of the others. Is this necessarily so?
PROBLEM 11 Does there exist a monohedral spiral tiling with

an odd number of arms, which uses only direct copies (no
reflections) of the proto tile?
Perhaps we should mention that the subject of spiral tilings, while
aesthetically attractive, has a big mathematical disadvantage-so far we
have not been able to say exactly what a spiral tiling is! Is it a genuine
mathematical concept, or is it merely psychological? We conclude with the
following problem, which is probably more suitable for general discussion
than for a mathematical investigation:
PROBLEM 12 Give a precise definition of a spiral tiling.
.8192&.
FIGURE 29
Monohedral spiral tilings with three and with one arms. The prototile
is known as a "versatile".
.a193~
FIGURE 30
An ornamental spiral tiling using the same proto tile as the spirals of
Figure 29.
References and Further Reading

In addition to the articles by Martin Gardner quoted in the text, the
following books and papers will be of interest. They are listed by the
sections of the paper to which they are relevant.
Introduction. Several collections of Escher's works have been pub-

lished; the most extensive is The World of M. C. Escher (Abrams, New York,
1971). A very interesting account of Escher and his tilings is given in B.
Ernst's book The Magic Mirror of M. C. Escher (Random House, New York,
1976). A discussion of the ti1ings to be found in the Alhambra appears in E.
Miiller, "Gruppentheoretische Ornamente aus der Alhambra in Grenada"
(ETH dissertation, Zurich, 1944). Kepler's book Harmonice Mundi, orig-
inally published in Linz in 1619, has been reprinted in Kepler's Complete
Works, edited by M. Caspar (Gesammelte Werke, Band VI, Beck, Munchen,
1940 and by Culture et Civilisation, Bruxelles, 1968). These texts are in
Latin; a German translation, Weltharmonik, by M. Caspar has also been
published (Oldenbourg, Munchen, 1967) .
..0194~
1 The three types of hexagons that admit tilings of the plane were
determined by K. Reinhardt in his thesis "Dber die Zerlegung der Ebene
in Polygone", Frankfurt University, 1918 (Noske, Leipzig, 1918).
Kershner's paper is "On paving the plane", American Mathematical Monthly
75(1968), 839-844; our list of pentagons is taken from D.
Schattschneider, "Tiling the plane with congruent pentagons",
Mathematics Magazine 51 (1978), 29-44. The announcement by M. D.
Hirschhorn and D. C. Hunt that the list of equilateral pentagons IS
complete also appears in the Mathematics Magazine 51 (1978), p. 312.
2 The subject of k-morphic tilings is considered in the authors' "Patch-

determined tilings", Mathematical Gazette 61 (1977),31-38, and Harborth's
example in "Prescribed numbers of tiles and tilings", Mathematical Gazette
61 (1977),296-299.
3 Hilbert's famous problems were printed (English translation) in

"Mathematical Problems", Bulletin rif the American Mathematical Sociery
8 (1902), 437-479 and reprinted in "Mathematical Developments Arising
from Hilbert Problems", Proc. Sympos. Pure Math., Vol. 28, (American
Math. Soc., Providence, R.I., 1976). The book by Heesch and Kienzle
is Fliichenschluss (Springer-Verlag, Berlin-Gottingen-Heidelberg, 1963). A
recent treatment of isohedral tilings is the authors' "The eighty-one
types of isohedra1 tilings in the plane", Math. Proc. Cambridge Philos. Soc.
82(1977),177-196.
4 A proof of the Extension Theorem will appear in the authors' book

mentioned below. Heesch's problem appears in his book Reguliires
Parkettierungsproblem (Westdeutscher Verlag, Ko1n-Op1aden, 1968).
5 Accounts of aperiodic tiles appear in R. M. Robinson's article

"U ndecidability and non-periodicity of tilings of the plane", Inventiones
Math. 12 (1971), 177-209, and R. Penrose's papers "The role of aesthetics
in pure and applied mathematical research", Bull. Inst. Math. Appl.
10(1974),266-271, and "Pentaplexity", Eureka 39(1978), 16-22. The
most up-to-date account is Martin Gardner's article cited in the text, and a
more detailed exposition will appear in the author's forthcoming book.
The connection between aperiodicity and the tiling problem is discussed in
Robinson's paper quoted above, and also in H. Wang, "Proving theorems
by pattern recognition II", Bell System Techn. Journal 40 (1961), 1-42.
~195~
6 Voderberg's papers are "Zur Zerlegung der Umgebung eines eben en

Bereiches in kongruente", J.-Ber. Deutsch. Math.- Verein. 46 (1936),
229-231, and "Zur Zerlegung der Ebene in kongruente Bereiche in Form
einer Spirale", ibid. 47(1937), 159-160. Goldberg's explanation of the
structure of spiral tilings appears in "Central tessellations", Scripta Math.
21 (1955),253-260. A short article "Spiral tilings and versa tiles" by the
authors appeared in Mathematics Teaching 88(1979),50-51.
The above are only a few references to the wealth of published
material on tilings in the plane. Further information and problems will
appear in the authors' book Tilings and Patterns, to be published by W. H.
Freeman and Company, San Francisco.
~196&.
Angels and Devils
-~-
H. S. M. Coxeter
UNIVERSITY OF TORONTO
About forty years ago, Abraham Sinkov and I wrote twin papers on the
subject of groups determined by the periods of two generators S, T and of
their commutator S- 1 T- 1 ST [Coxeter 1936; Sinkov 1936], never dream-
ing that twenty years later M. C. Escher would be using such groups
(unconsciously) as symmetry groups for a carved ball and four other works
of art [Escher 1971, Figures 112, 115, 226, 235, 244, 247; MacGillavry
1976, p. 18]. Those works have been reproduced here by kind permission of
the Escher Foundation, Haags Gemeentemuseum, The Hague.
Patterns in the Euclidean Plane

On ordinary wallpaper, a motifis repeated by translations in two directions
in accordance with the symmetry group called pl. Theoretically, a more
interesting pattern could be obtained by other symmetry operations such as
a half-turn which turns the motif upside down (so that b would yield q), or a
reflection which changes a left hand to a right hand (or b to d or p). A half-
turn is a rotation of period 2, but we might also use a rotation of period 3 or
4 or 6. By combining these "isometries" in every possible way, E. S.
Fedorov proved in 1891 that there are just seventeen planar symmetry
groups involving translations in two directions. Eleven of the seventeen
were unconsciously discovered long ago by the Moors in their decoration of
the Alhambra. Some of these eleven, and an additional five, were used by
~197~
the Bakuba and Benin tribes in Africa (south of the Sahara) in their
pottery, weaving, and basket-making [Crowe 1971; 1975]. The remaining
one, called p31m, occurs in a Chinese pattern [Fejes T6th 1964, p.40,
Plate 11.1].
Escher's patterns are more interesting, because their motifs are
mainly animals, ingeniously fitted together so as to cover the plane without
leaving any gaps. For instance, in Figure 1, every point of the plane
belongs either to an angel or to a devil or to the curve that separates one
from the other. The pattern, regarded as continuing so as to fill and cover
the whole plane, is symmetrical by quarter-turns (rotations through n/2)
about each point where wing-tips of four angels meet wing-tips of four
devils. I t is symmetrical also by reflections in certain vertical and
horizontal lines, and by any product of such rotations and reflections. More
economically, the whole symmetry group is generated by one rotation of
period 4 (say S) and one reflection T. For instance, if T is the reflection in
FIGURE I
Escher's Sketch for "Angels and Devils" .
.8198~
ANGELS AND DEVILS
one of the vertical mirrors, and S is a quarter-turn whose center is as near

as possible to that mirror, then S will transform T into the reflection
T1 = S-1 TS in one of the horizontal mirrors. The powers S' of the rotation
S (k = 0, 1,2,3) will transform T into reflections S-k TS k in all the sides of
a square.
The rotations Sk form the cyclic group (£:4 of order 4, generated by S.
The two reflections T and T1 generate the dihedral group Il2 of order 4, in
which the product T1 T generates a subgroup (£:2 (because the half-turn
T1 T has period 2).
The relations
are abstract definitions or presentations for the groups (£:4 and Il 2, re-
spectively, in the sense that every relation satisfied by the generator S, or by
the two generators T1 and T, is an algebraic consequence of these simple
relations. Writing S-1 TS for T 1, we deduce the presentation
~ = T2 = (S-1 TST) 2 = 1
for the infinite group generated by the rotation S and the reflection T.
This infinite group p4g [Coxeter and Moser 1972, p. 47] is the special
case [4 +, 4] of the group
[/+, 2p],
which has the presentation
S' = T2 = (S-1 TSTV = 1 (l~2,P~1).
Here S is a rotation of period I (that is, through an angle 2nl I) and T is a

reflection whose mirror is in such a position that the product T1 T, where
T1 = S- 1 TS, has period p. In other words, the powers of S transform the
mirror for T into the sides of a regular I-gon, and two adjacent sides (such
as the mirrors for T1 and T) form an angle nip. It follows that such a group
[1+, 2p] occurs whenever the plane can be tessellated with regular I-gons, 2p
round each vertex. Such a tessellation is denoted by {I, 2p} [Coxeter and
Moser 1972, p. 102]. For instance, {4, 4} is the ordinary tiling of squares
(the "squared paper" pattern). The only other possibility in the Euclidean
plane is {3, 6}, indicated by the heavy lines in Figure 2.
The crystallographer Caroline MacGillavry [1976, Plate 8], im-
pressed by Escher's rediscovery of most of Fedorov's symmetry groups,
pointed out that Escher, like the Africans, had missed the group
p31m ~ [3 +, 6]
(misnamed p3ml in [Coxeter 1969, p. 413] and [Coxeter and Moser 1972,
pp. 49, 136]). At her request, Escher filled this gap by creating a new
picture of red bees and yellow-green wasps (Plate IV). Here the edges of
~199&.
FIGURE 2
A fragment of the dual tesselations {3, 6} (dark) and {6, 3} (light).
the underlying tesselation {3, 6} are the obvious lines of symmetry of the
insects. These edges are perpendicularly bisected by the edges (appearing
as light lines in Figure 2) of the dual tessellation {6, 3}, whose vertices are
the points where the "elbows" of three bees meet those ofthree wasps. The
generator S of[3 + , 6] is the rotation through angle 2n/3 about such a point,
and we easily verify the relations
S3 = T2 = (TiT) 3 = I,
where T = S-l TS.
We might stretch the meaning of the symbol [1+, 2p] to include
[2 +, ex)] and [ ex) + , 2]
as the fifth and sixth of the seven frieze groups [Coxeter 1969, p. 48]. The
former, generated by a half-turn and a reflection, is the symmetry group of
the sine curve and of the frieze
... VI\VI\V ... ;
the latter, generated by a translation and a reflection (the translation being

along the direction of the mirror), is the symmetry group of the frieze
... DDDDD ....
.a200~
ANGELS AND DEVILS
FIGURE 3
Escher's Sphere with Fish.
Carved Balls
A more interesting extension of the Euclidean tessellations arises when we
replace the ordinary plane by a non-Euclidean plane-either the sphere or
the hyperbolic plane.
The surface of a sphere may be regarded as a plane whose lines are
the great circles. This idea occurred to Abu'l Wafa (940-998) [see Woepke
1855, pp. 352-357]. Rotation about a diameter of the sphere is regarded as
rotation about either of the points where this diameter penetrates the
surface. A regular tetrahedron {3, 3}, inscribed in the sphere, is symmetri-
cal by rotations of period 3 about its vertices. These rotations generate the
tetrahedral group of order 12, which is denoted by 214 because it is the
alternating group of degree 4: the group of even permutations of the 4
vertices. Four more tetrahedra may be inscribed in the same sphere so that
the whole set of 20 vertices belong to a regular dodecahedron {5, 3}
[Coxeter and Moser 1972, p. 35]. Adjoining to 214 a rotation of period 5,
we obtain the icosahedral group of order 60, which is denoted by 2ls because
it is the alternating group of degree 5: the group of even permutations of
the 5 tetrahedra.
.8201&
FIGURE 4
Polyhedron with Flowers.
Escher used m: 4 as the symmetry group for one of his carved globes
(Figure 3), and m:s for another (Figure 4). In the latter, the twelve flowers
are chiral, like periwinkles, and the tips of the petals are at the vertices of
the dodecahedron. In fact, this globe is evidently based on the above-
mentioned compound polyhedron (Figure 4a), which is denoted by
{5, 3} [5{3, 3}] {3, 5}
because its twenty vertices belong to a dodecahedron {5, 3} while its
twenty faces lie in the same planes as the faces of the dual icosahedron
{3,5}.
..0202~
ANGELS AND DEVILS
FIGURE 4A
J. F. Petrie's drawing of the five tetrahedra.
More directly relevant to our discussion of groups [1+, 2p] is his other
carved globe (Figure 5), because it is symmetrical by a rotation S, of
period 3, about a point where wing-tips of three angels meet wing-tips of
three devils. It is also symmetrical by a reflection T whose mirror is one of
three mutually perpendicular planes. Thus Sand T generate the group
[3 +, 4], and the underlying spherical tesselation {3, 4} is the octahedron
whose faces are the spherical triangles cut out on the sphere by these three
planes of symmetry. [3 +, 4] is sometimes called the pyritohedral group
because it is also the symmetry group of a crystal of iron pyrites-a not-
quite-regular dodecahedron. Abstractly, this group of order 24 is the direct
product
~4 X (£2
[Coxeter and Moser 1972, pp. 3, 39]. In fact, the defining relations
are satisfied by the permutations
S = (a be), T = (a b) (c d) (e J) ,
~203~
FIGURE 5
Sphere with Angels and Devils.
which generate the direct product of the 2(4 generated by

(abc) = S, (ab) (cd) = STSTS
and the group of order 2 generated by
(ef) = (ST)3.
For the sake of completeness, we may mention also the trivial groups
[2+, 2p]~ l)2P and [/+, 2] ~ (£, x (£2
(of orders 4p and 2/), which resemble [2 +, 00] and [00 +, 2], only now the
friezes are wrapped round a cylinder. In other words, [2+, 2p] is the
symmetry group of a p-gonal antiprism [Coxeter 1969, p. 149].
~204~
ANGELS AND DEVILS
Circle Limits
In 1958 I sent Escher a reprint which contained a drawing like Figure 6. In
reply, he thanked me and said that "[the illustration gave him] quite a
shock. Since a long time I am interested in patterns with 'motives' getting
smaller and smaller till they reach the limit of infinite smallness. The
question is relatively simple if the limit is a point in the center of a pattern.
Also a line-limit is not new to me, but I was never able to make a pattern in
which each 'blot' is getting smaller gradually from a center towards the
outside circle-limit, as shows your figure. I tried to find out how this figure
was geometrically constructed, but I succeeded only in finding the centers
and radii of the largest inner-circles. If you could give me a simple ex-
planation how to construct the following circles, whose centers approach
gradually from the outside till they reach the limit, I should be immensely
pleased and very thankful to you! Are there other systems besides this one
to reach a circle-limit?"
In response to Escher's letter, I told him that {4, 6} and {6, 4} are two
of infinitely many regular tessellations {p, q] consisting of congruent regular
p-gons fitting together, q at each vertex. If p and q are too big for the
tessellation to exist on a sphere or in the Euclidean plane, it requires a
hyperbolic plane, in which the regular p-gon has a smaller vertex angle. In
FIGURE 6
A fragment of the dual tessellations {4, 6} (dark) and {6, 4} (light).
e.t205~
one of Poincare's models, the "straight lines" of the hyperbolic plane

appear as arcs of circles orthogonal to a boundary circle 0 drawn in the
Euclidean plane. Angles are represented faithfully but distances are
distorted, with the points of 0 itself being infinitely far away. Although the
regular quadrangles formed by the dark lines in Figure 6 get "smaller
gradually from the center towards the outside circle-limit", we enter the
spirit of hyperbolic geometry when we stretch our imagination in order to
pretend that these quadrangles are all regular and congruent.
Escher's sketch-books show that he diligently pursued these ideas
before completing his Circle Limit IV (Figure 7), whose symmetry group (in
the sense of hyperbolic geometry) is [4 +, 6]:
~ = T2 = (S-l TST)3 = 1.
The generator S is (as in Figure 1) a quarter-turn about one of the points
where wing-tips offour angels meet wing-tips offour devils-that is, about
FIGURE 7
Escher's Circle Limit IV.
~206~
ANGELS AND DEVILS
a vertex of {6, 4} (drawn with light lines in Figure 6). The other generator
7 is the reflection in one of the dark lines in Figure 6 (each containing
infinitely many edges of {4, 6}). If the mirror for 7 is one of the straight
lines through the center, S transforms it into one of the dark arcs, and the
hyperbolic reflection 71 = S-l 7S appears in the Poincare model as the
inversion in the circle that carries that arc. Figure 7 differs from Figure 1 in
that the rotation 717 (about the tips of the angels' feet) is of period 3
instead of 2.
Escher used the analogous group [3 +, 8] for his Circle Limit II
(Figure 8), which is just as interesting mathematically although Bruno
Ernst [1976, p. 109] dismissed it with a little joke. The underlying
tessellation {3, 8} (Figure 9) has for vertices the centers of all the crosses.
The points where the three colors come together, being the centers of the
triangular faces of {3, 8}, are the vertices of the dual tessellation {8, 3}
[Coxeter 1979, p. 23, Figure 5]. On the other hand, the centers of the
FIGURE 8
Escher's Circle Limit II.
et207~
FIGURE 9
A fragment of the dual tessellations {3, 8} (dark) and {8, 3} (light).
crosses of one color are the vertices of a tessellation of quadrangles, {4, 8}.
(In Figure 9 the vertices of {3, 8} are black, white or dotted accordingly as
the corresponding crosses are black, white or grey.) In this sense, Escher
may be said to have anticipat~d my discovery of the regular compound
tessellation
{3, 8} [3{4, 8}]2{8, 3}
[Coxeter 1964, pp. 156-157] which consists of 3 superposed tessellations
{4, 8} whose vertices belong to a single {3, 8} while their faces have the
same centers as the faces of the dual {8, 3}, each used twice.
Conclusion
For any two integers 1 and p (l> 2, p> 1) there is a group [/+, 2p]
generated by a rotation S of period I and a reflection T of period 2, whose
commutator S-1 TST is a rotation of period p. The kind of plane involved
is spherical, Euclidean or hyperbolic depending on whether the number
(l- 2) (p - 1) is less than 2, equal to 2, or greater than 2. The finite group
[3 +, 4] (of order 24) is the symmetry group of Escher's Sphere with Angels and
Devils. He used also the two Euclidean groups [4 + , 4] and [3 + , 6], and it is
~208~
ANGELS AND DEVILS
interesting that, among the infinitely many hyperbolic groups [1+, 2p] with
(l - 2) (p - I) > 2,
he instinctively chose the two simplest: [4 +, 6] and [3 +, 8].
References
1 Coxeter, H. S. M. 1936. The groups determined by the relations S'
= T" = (S-l T- 1 STJ P = 1 Duke Math. Journal. 2: 61-73.
2 _ _ . 1964. Regular compound tessellations of the hyperbolic plane.
Proc. Royal Soc. London A 278: 147-167.
3 _ _ . 1969. Introduction to Geometry. 2nd ed. New York: Wiley.
4 _ _ . 1979. The non-Euclidean symmetry ifEscher's picture 'Circle Limit III'.
Leonardo 12: 19-25,32.
5 _ _ and Moser, W. O. J. 1972. Generators and Relations for Discrete
Groups 3rd ed. Berlin: Springer.
6 Crowe, D. W. 1971. The geometry of African art I. J. Geom. 1: 169-
182.
7 _ _ . 1975. The geometry of African art II. Historia Math. 2: 253-271.
8 Ernst, Bruno. 1976. The Magic Mirror of M. C. Escher. New York:
Random House.
9 Escher, M. C. 1971. The World if M. C. Escher. New York: Abrams.
10 Fejes T6th, L. 1964. Regular Figures. New York: Pergamon.
II MacGillavry, Caroline. 1976. Fantasy and Symmetry- The Periodic
Drawings if M. C. Escher. New York: Abrams.
12 Sinkov, Abraham. 1936. The Groups Determined by the Relations S'
= T" = (S-l T- 1 S7Y = 1. Duke Math. Journal 2: 74-83.
13 Woepke. F. 1855. Recherches sur l'histoire des sciences mathematiques
chez les orientaux, d'apres des traites inedits arabes et persans. J.
Asiatique 5: 309-359.
~209~
Three-Dimensional
Tiling
Packing Problellls
and Inequalities
-~-
D. G. Hoffntan
AUBURN UNIVERSITY
Four rectangular farms, each 7 miles by 8 miles, are to be carved out of a

square county, 15 miles on a side. How is this to be done? (Try it yourself
before peeking at the answer in Figure la on page 213.
Notice that the county has area 15 2 = 225 square miles, and each
farm has an area of 56 square miles. Hence there will be 225 - 4(56) = 1
square mile left over.
Here is a more general puzzle. Suppose that x and yare positive
numbers. Will four rectangle farms, each x miles by y miles, fit into a square
county, x + y miles on a side? Now the area of this county is (x + y)2 square
miles, and each farm has area ~ square miles. So unless
1
we don't have a chance of solving this puzzle. In other words, if the puzzle
has a solution, then the combined area of the four farms cannot be greater
than the area of the county, so that inequality 1 must hold.
This more general puzzle does have a solution, and I'm sure you've
found it if you solved the first puzzle, where x = 7 andy = 8. Figure 1 below
gives the answer to the first puzzle in (a), and the answers to the general
puzzle in (b), (c), and (d), depending on the relative sizes of x andy. The
arithmetic mean, or average of two numbers x andy is half their sum:
~212~
PACKING PROBLEMS AND INEQUALITIES
8 y
x
7
(a) (b)
x = 7 x<y
Y= 8
x y
x
x
(e) (d)
x=y x>y
FIGURE 1
1
2 AM. = - (x + y)
2
There is another "average" of x andy that is sometimes useful when x
andy are positive. It is called the geometric mean of x andy, and it is the
square root of their product:
3 C.M.=~
For example, if x = y, then AM. = t(x + x) = x, and C.M. = fo = x,
so these two means coincide if x = y. However, with x = 4,y = 16, we have
AM. = 10, while C.M. = 8. It turns out that the geometric mean is never
bigger than the arithmetic mean; that is,
~213~
Here is an easy proof, using our four farms puzzle. Figure I proves
that the puzzle does have a solution, therefore inequality 1 must be true. If
we take the square root of both sides of 1, we get 2 j ; ::; x +y, and
dividing both sides by 2 gives precisely inequality 4, concluding the proof.
Puzzles in which a number of objects are to be fit into a container are
called packing problems. Such puzzles are well known to Gardner readers,
recreational mathematicians, and smugglers.
The moral of the above example is that any solution to a packing
problem yields a proof of an inequality. If all the objects fit into the
container, then their combined area (or volume) must be no greater than
the area (or volume) of the container. Perhaps the reverse is true! That is,
can we start with an inequality known to be true, and make an amusing
packing problem out of it? Here is an example.
The arithmetic mean of three numbers, x,y and z, is their average, or
one third their sum:
The geometric mean of three positive numbers x,y and z is the cube root of
their product:
6 G.M.=~.
The inequality 4, for two numbers, is also true for three numbers:
To make this into a packing problem, multiply both sides by 3, and then
cube both sides. The result is
8
We notice that xyz is the volume of a brick with dimensions x by y by z,
while (x + y + z) 3 is the volume of a perfect cube, x + y + z on a side. So
here is our packing problem:
Will 27 identical bricks, each with dimensions x byy by z, fit into

a cubical box, x + Y + z on a side?
Of course, if this packing problem has a solution, that solution would
prove that 8, (and therefore 7) is a true inequality.
We mentioned above that 7 is in fact true for any positive numbers x,
y and z. This fact in itself does not guarantee that the bricks will fit in, but
only that their combined volume is no bigger than the volume of the box.
~214~
(The volume of my fishing pole is less than the volume of my briefcase, but
the fishing pole won't fit in my briefcase!)
At this point you might want to obtain a set of 27 bricks and try this
puzzle. Let me advise you on what dimensions you should choose for x,y
and z. There is one basic problem in choosing x,y and z. Many values of x,
y and z will yield puzzles that are far too easy! For example, x = l,y = 1,
and z = 100 is ridiculous; 27 sticks, each 1" by 1" by 100" will fi t in to a 102"
by 102" by 102" box with enough room left over for a tuba. Another
extreme is x =Y = z; here the 27 bricks are cubes that fit snugly into the box.
To be certain of avoiding all such trivialities, make sure that the three
dimensions are different, and that the sum of the two larger ones is less than
three times the smallest.
In other words;
9 0< x <y<z
10 y+ z< 3x.
If you make a set of bricks violating either 9 or 10, there might (and
probably will) be solutions that are easy to find.
The only other considerations in choosing the dimensions x,y and z,
besides 9 and 10, depend only on what materials you have at hand, and the
convenience of construction. No values of x, y and z satisfying 9 and 10
produce a puzzle easier or harder than any other such values, with one
possible exception: a puzzle withy = t (x + z)might be just a bit harder!
Here is a good example: x = 4,y = 5, z = 6. So, will 27 4 by 5 by 6 bricks
fit into a 15 by 15 by 15 box?
Let me assure you that this puzzle does have a solution. In fact, I
learned from both J. H. Conway and William Cutler, that there are
exactly 21 solutions, not counting rotations and reflections of the box,
provided of course that 9 and 10 hold.
A solution seems hard to find. I tried for hours to solve this puzzle
merely using pencil and paper. Finally, I gave up and called my friend
David Klamer, who has a power saw. He says that as he sanded the blocks
one by one, he stacked them into the solution shown below without
changing one block. Thus, he earned the distinction of being the first
person to solve my puzzle. (See the photos on the following page.) I've seen
solution times ranging from about 20 minutes to a few days. If you have
built a set, and you want to put your blocks back into the box (as you
should, if you are done playing with them), I enclose a solution in Figure 2
below, with x = 4, Y = 5, z = 6. The actual dimensions don't matter of
course, you can use Figure 2 for any x, y and z satisfying 9 and 10, just as
Figure I b works for any x <yo
The 27 bricks are arranged in three layers of nine bricks each, and
Figure 2 is a cross section through each of the three layers. Let us examine
this puzzle more closely. We are assuming that x,y and z are numbers
satisfying 9 and 10.
1
The first to build a set of blocks, and to find a packing of them was
D. A. Klarner. Shown above is the first set built, each block having
dimensions 7 by 8 by 10, being packed by a chip off the old block, Carl
Klarner. The photos were taken by Kara Lynn Klarner.
h216~
3
~217~
Top Layer Middle Layer
Bottom Layer
FIGURE 2
A perfect cube, x + y + z on a side, can be cut into 64 perfect cubes,

each i (x + y + z) on a side by nine cuts (see Figure 3a). These nine cuts
determine nine planes, which we shall call the special planes (see
Figure 3b). These nine planes fall into three sets of three mutually parallel
planes. Within each set, the middle plane is i (x + y + z) from the other
two, and these other two are each i (x + y + z) from the outside of the box.
Two special planes that are not parallel meet in a line. Altogether, there are
27 such lines, and we call them the special lines (see Figure 3c). Three
special planes, of which no two are parallel, meet at a point. Altogether,
there are 27 such points, we call them the special points (see Figure 3d).
Suppose now that this cube has been packed with 27 x by Y by z bricks.
Consider now one of the three sets of three mutually parallel planes.
They divide the cube into four slabs, each with thickness i (x + Y + z). Can
anyone of the 27 bricks in our packing lie entirely within one of these four
slabs, miraculously untouched by any of the three knife cuts?
e.t218 ~
/ / / /
/ / / /
/ / / /
FIGURE 3A FIGURE 3B
The cube cut. The nine special planes.
.------------"
II I
11-----------"A
/ I I II
I I / I I I
I I I
/ I / I
I
/ I I / I / / / / I
f--+-.!.---·---'!'-/ I
I' • • I
(- --i- - - f-( I
I
I I I
I• I I
I • I I I I
I • • I I
:. . .:
I • • • I I I
I
I I I
~-=---!"--"!-.j.--~ .~- rt· ...J
/' I I • • I /
I
I I
I I
I / I I/
I
I I /
/ I I I I I
ilI_____________ y1/ L!.I ____________ J!1/
FIGURE 3c FIGURE 3D
The 27 Gpecial lines. The 27 special points.
Since y + Z < 3x, we have x + y + Z < 4x, or t (x + Y + z) < x.

Thus the smallest dimension of our brick, x, is bigger than the width
i (x + y + z) of each slab. Unlike the magician's lovely assistant, no brick
has escaped all three of the parallel knife cuts.
To summarize what we have just proved: each brick in the packing is
cut by at least one of the three special planes in any of the three parallel
classes of special planes. As a consequence, the special point in which these
three special planes meet must be inside the brick. We have shown that
each of the 27 bricks in the packing has at least one special point inside it.
But there are only 27 special points altogether. Hence each brick has
exactly one special point inside it, and each special point is on the inside of
exactly one brick. (We have also proved, by the way, that there is no hope
of getting 28 or more bricks in the box, as there are not enough special points
to go around.)
Each special line has three special points on it, so much of it lies on the
inside of the three bricks containing these three special points. Can there be
gaps, as in Figure 4?
~2l9~
FIGURE 4
Gaps on a special line.
We will show that there can be no gaps. For each special line, let us
measure how much of its total length of x + y + z is on the inside of bricks
in the packing. Let us denote the sum of these 27 numbers, one for each of
the 27 special lines, by the symbol t. Each of the 27 bricks has a special
point inside it, and this special point has three special lines through it, in
three perpendicular dimensions. Hence each brick must gobble up at least
x +y + z units worth of special lines. Hence 27 (x +y + z) ~ t. On the
other hand, each of the 27 special lines is of length x +y + z, so t ~ 27
(x + y + z). The only alternative is that t = 27 (x + y + z), and that each
special line is completely within the three bricks on its three special points.
In particular, there can be no gaps, as in Figure 4. Hence each special line
looks like Figure 5, where a + b + c = x + y + z, and each of a, band c is
one of x,y or z. Since x <y < z, there are only two ways this could happen.
Either a, band care x,y and Z in some order, or a = b = c = y. Further, this
second case cannot happen unless y +y +y = x +y + z, or y = t (x + z).
We can prove that the second case, a = b = c =y, cannot happen at
all. Each of the 27 special lines is divided into three segments by the three
bricks containing this line. Altogether, there are 27·3 = 81 such segments.
Moreover, these 81 segments must consist of exactly 27 x's, 27 y's, and 27
z's, since each of the 27 bricks cuts off three segments-one of each length.
The paragraph above points out that each special line has either one or
three segments oflengthy. But there are only 27 segments oflengthy to go
around, so each special line has exactly one segment of length y. Hence
a = b = c = y cannot happen.
FIGURE 5
~ 220f$.
FIGURE 6
We have shown that each of the 27 special lines is entirely inside three
bricks of the packing, which divide it into three segments, of three different
lengths, x,y and z. This is quite a useful fact to know. For example, if you
attempt to solve the puzzle by putting the first nine bricks in a layer at the
bottom of the box as in Figure 6, you will be unable to complete the puzzle.
Do you see why?
Another useful fact to know is that no special plane can have a corner
like Figure 7 below, since whichever brick goes in the center of this plane
must abut both A and B, and yet C is in the way. Knowing these facts
certainly helps to solve the puzzle, but there are many more difficulties and
false starts to be encountered.
Notice that each of the nine special planes contains nine special
points, and hence cuts across nine bricks. Thus, this special plane has been
packed with nine rectangles, each either x-by:y, x-by-z or y-by-z, It is easy
to prove then, the following fact: these nine rectangles comprise exactly
three x-by:y's, exactly three x-by-z's and exactly three y-by-z's. For three
examples, see Figure 2 above. We have proved, by the way, that
A
C
FIGURE 7
~221~
11 3(xy + xz + yz) ~ (X + y + Z)2,

since the left hand side is the total area of the nine rectangles, and the right
hand side is the area of the square they pack.
If conditions 9 and 10 hold, there are exactly 78 ways, not counting
rotations and reflections, in which a square, x + ) + z on a side, can be
packed with three x by) rectangles, three x by Z rectangles, and three) by Z
rectangles.
The inequalities 4 and 7 are special cases of what is known as the
arithmetic mean, geometric mean inequality. This states that if Xl, X2, ... , Xn are
positive numbers, then
12 .y'-X-l'-X-2-'-"-'-'x-n ~ ~(Xl + X2 + ... + xn)·
n
The left hand side is called the geometric mean of the numbers, while the
right hand side is the familiar arithmetic mean.
To make 12 into a puzzle, multiply both sides by n, then raise both
sides to the nth power. The result is
13 nn(xl·x2·····Xn) ~ (Xl + X2 + ... + X3)n.
Now Xl' X2 •..•• Xn is the n-dimensional volume of an n-dimensional brick
with dimensions Xl'X2," ',Xn; and (Xl + X2 + ... + xn)n is the n-dimensional
volume of a perfect n-dimensional cube; each side of the cube has length
Xl + x 2 + ... + Xn. So will nn such bricks fit into the n-dimensional cube?
Let me assure those readers unaccustomed to thinking in four or more
dimensions, that such thought, contrary to rumor, does not cause,
significant brain damage if practiced moderately. Here is some expla-
nation, if the above paragraph made you shudder.
Just as a rectangle has two dimensions (a length and a width) and a
shoe box has three, an n-dimensional brick or box has n dimensions, Xl' X2,
.. " Xn and each dimension Xi is a positive number. Let us denote such a
brick, or box by the vector (Xl' X2, "', xn)' Thus (3,5) denotes a 3-by-5
rectangle, (4,6,2,6) denotes a 4-dimensional 4-by-6-by-2-by-6 "box", and
both (2,3,9) and (2,9,3) denote 2-by-3-by-9 ordinary 3-dimensional
boxes. (In one dimension, n = 1, (Xl) denotes a line segment Xl units long).
A measure of how much "n-dimensional stuff" there is in the box (Xl'
X2, "', xn) is the number Xl' X2····· Xm the product of its dimensions. It is
called the hypervolume of the box. Thus for n = 1, hyper-volume is length,
for n = 2, hyper-volume is area, and for n = 3, hyper-volume is usual
volume.
Here is a question which, as far as I know, has not been resolved:
14 For which positive integers n is it true that for any positive
numbers Xl' X2, "', Xm nn bricks, each with dimensions (Xl, X2,
"', xn), will fit into a perfect n-dimensional cube, Xl + X2 + .. ,
+ Xn on a side?
The problem is trivial for n = I, and Figures I and 2 show how to do
n = 2 and 3 for an) (Xl' X2) or (Xl' X2, X3)' How about n = 4 or more?
~222~
Let me remind the reader that the inequality 13 does not assure us
that the bricks will fit, it only fails to discourage us from trying. (If 13 did
not hold, we would be certain that the bricks would not fit into the box).
Let us call those dimensions n, for which a packing exists, the good
dimensions. Thus we know that 1, 2 and 3 are good. J. Selfridge has told
me that R. Robinson has proved the following theorem:
15 If m and n are good dimensions, so is mn.
In particular, since 2 and 3 are good, so are 4, 6, 8, 12 and 181398528
= 2 10 • 3 11 . Thus 44 = 64 four-dimensional bricks, each with dimensions
(Xl' X2' X3' X4)' will fit into a perfect 4-dimensional cube, Xl + x2 + X3 + X4
on a side.
One might suspect that all positive dimensions n are good. If this is so,
it need only be proved for all prime numbers n, since any positive integer is
a product of prime numbers, and solutions can be "multiplied together" by
15. The smallest dimension in doubt is thus 5.
I hope some ambitious readers will attempt to resolve this question:
Will 55 = 3125 5-dimensional bricks, each xl-by-x2-by-x3-by-x4-by-x5' fit
into a perfect 5-dimensional cube, Xl + X2 + x3 + X4 + X5 on a side?
There is only one little stumbling block in understanding Robinson's
proof of 15, and that is the fact 16 below. Once 16 is understood you will
see that it is easy to prove 15.
Before we state this fact, let us give an example. We showed in
Figure 1 that four (7,8) rectangles pack a (15,15) square. Hence four
(7,8,100) bricks will pack a box (15,15,100). Similarly, four 5-dimensional
bricks, each (7,8,100,14,47) will pack a 5-dimensional (15,15,100,14,47)
box.
16 Suppose that k n-dimensional bricks, with dimensions
(Xl,l' x l ,2' " ' , x l ,II)' (X2 ,l' x2,2' "', X2,1I) , "', (xk.1> xk.2' "', xk,lI)
fit into a box of dimensions (Yl,h, ""YII)' Then the k (n+t)-
dimensional bricks with dimensions
(X l ,l,X l ,2' "', Xl ,lI' <:1' <:2' " ' , <:/)'
(X 2 ,l' X2 ,2' " ' , X 2 ,1I' <:1' <:2' "', <:/)' "',
(Xk,l' Xk,2' "', Xk,lI' <:1' <:2' " ' , <:/)'
will pack a box of dimensions
(Yl'Y2' ""YII' <:1> <:2' "', <:/)'
We are now ready to prove Robinson's theorem (15). We know then
that m and n are good dimensions, and we have to show that mn is also a
good dimension. In other words, we must find a way to pack (mn)mll (mn)-
dimensional bricks, each with dimensions
(Xl,l' x l ,2' "', xl ,lI' X2 ,l' x2,2' "', x2,1I' X3 ,l' "', xm - 1 ,1I' Xm,l' xm,2' "', Xm,lI) ,
into a perfect (mn)-dimenslOnal box, with each side having length t, where t
is the sum of the mn-dimensions X 1,1' ... , x m ,lI' The proof will be easier to
.8223 ~
follow if we don't string out the mn dimensions of a brick in one line, but
write them in matrix form instead:
Xl,l' xl,2, ••• , Xl,lI)

x2,1, x2,2' ••• , x2,11
(
.
... ..
., .
Xm,l, Xm,2, ••• , xm,1I
To begin, divide the (mn)mll bricks into groups of mm bricks each. Thus
there will be (mn)mlljmm = mm(lI-l)n mll such groups. Let Sl = Xl,l + X 2 ,1 + ...
+ Xm,l' Since m is a good dimension, mm bricks, each with dimensions (Xl,l,
x2,1' ... , Xm,l), will pack a box of dimensions (S1> Sl' ... , Sl)' Using 16, we
see that the mm bricks in each of the mm(lI-l)n mll groups can pack a box of
dimensions
Sl' Xl,2, ... , Xl,lI)

Sl' X2,2, ... , x2,11
(
.
. . .
.. ,
...
Sl' Xm,2, ••. , xm,1I
making mm(lI-l)nmll such boxes. Divide these new boxes into groups of mm
boxes each. There will be mm(lI-l)n mll j m m = mm(lI- 2)n mll groups. The mm boxes
in each group can be packed into a box of dimensions
Sl' S2, Xl,3, ... , Xl,")

Sl' S2, X2,3, ... , X2,1I
··· . . ...
(
.
...
Sl' S2, Xm,3' ... , Xm,1I
where S2 = Xl,2 + X2,2 + ... + x m,2' (Here we have used the fact that mm
bricks, each of dimensions (Xl,2, x2,2' ... , Xm,2), will fit into a box of
dimensions (S2' S2, ... , S2); and then we used fact 16 again.)
Continue following this process n times, once per column. At each
stage, the boxes produced at the previous stage are divided into groups of
mm boxes each, and the boxes in each group are assembled into a new box.
The dimensions of these new boxes, (one per group), are obtained by
replacing every entry of the column in question by the sum of the column
entries. The result is this: the (mn)mll original bricks have all been packed
into mm(II-II)n mll = n mll boxes, each box with dimensions
Sl' S2' ... , SII)

s1> s2,,,,,slI
( .. .
. ,
..
"
.
Sl,S2"",SII
where Sj = Xl,j + X2,j + ... + Xm,j for i = 1, 2, ... , n.
~224 f:>
Now we do the same thing, this time working with the rows instead
of the columns. Divide the nmll boxes into groups of nil boxes each. There
will be nmnjnn = n(m-1)n such groups. Since n is a good dimension, nil bricks,
each with dimensions (Sl' S2' ... , SII)' will fit into a box with dimensions (t,
t, ... , t), where t = S 1 + S2 + ... + SII' (Note that t is the sum of all the mn
original brick dimensions x 1,1' Xl, 2, ... , X m , III as defined above, just after 16.
So t is the length of each side of the mn-dimensional cube we are trying to
pack.) Thus, using 16, we see that the nil boxes in each group will fit into a
new box with dimensions
t' t, ... , t )
Sl,S2,,,,,SII
( ..
. . .
.. ,
..
Sl' S2, ... , Sn
and there will be n(m-1)n such boxes. Repeat this for each of the m rows,
with the result that everything can be packed into n(m-m)1I = I box of
dimensions
t' t, ... , t)
( t,..t, ... , t.
..
.. ..
t, t, ... , t
which is precisely our goal. Thus all (mn)mn original bricks fit in, and mn is a
good dimension. This completes the proof
We have seen several examples in this chapter of a nice connection
between inequalities and packing problems: sometimes known inequalities
yield interesting packing problems, and conversely, sometimes an in-
equality can be proved by solving a packing problem.
The proof above illustrates a further connection. In some of the
standard proofs of the arithmetic mean-geometric mean inequality, one
step is to prove that ifit is true for m numbers, and ifit is true for n numbers,
then it is true for mn numbers. This proof can be stated so that each step of
the proof corresponds to a step in Robinson's proof above! So the moral is
this: sometimes a known proof of an inequality can be used to find a
solution to a packing problem.
Mathematics is rich with inequalities and surely many of these can
yield interesting packing problems. May I recommend to the reader the
excellent book Geometric Inequalities by N. D. Kazarinoff. (New
Mathematical Library, Vol. 4, Random House). This (and in fact all the
books in New Mathematical Library series) can be readily understood by
anyone with a reasonable background in high school mathematics, and an
open mind.
When next accosted by an inequality, use whatever force is necessary
to change it to a packing problem -you might be pleased by the result!
.8225~
Can Cubes Avoid
Meeting Face to Face?
-~-
Raphael M. Robinson
UNIVERSITY OF CALIFORNIA
In this paper, we shall study the tiling of space by congruent cubes. By

saying that the cubes form a tiling of space, we mean that they cover space
without overlapping. We will assume that the edges of the cubes are
parallel to the coordinate axes. There is no loss of generality in using unit
cubes, and we shall usually do so.
We will start by looking at the (easier) two-dimensional case. Here
we are concerned with tilings of the plane by congruent squares with sides
parallel to the coordinate axes. The simplest tiling is the standard tiling
shown in Figure 1. In this tiling, each square meets edge to edge with four
FIGURE 1
Standard tiling of the plane.
~226~
CAN CUBES AVOID MEETING FACE TO FACE?
FIGURE 2
Tiling of the plane with displaced rows.
other squares. We say that two squares meet edge to edge only if they share
a complete edge with each other. We can also tile a plane by sliding
different rows of the standard tiling by different amounts, as in Figure 2. In
this way, we can form tilings in which each square meets edge to edge with
only two other squares. Instead of displacing rows, we could displace
columns, as in Figure 3. It is easy to see that in any tiling of the plane by
congruent squares, the squares must lie either in rows or in columns, or
perhaps in both. Thus, in every such tiling, each square must meet at least
two other squares edge to edge. (A lattice point in the plane or in space is a
point (x,y) or (x,y,z) with integer coordinates.) Ifwe place a unit square
with its center at each lattice point in the plane, then we will obtain a
standard tiling of the plane.
We now turn to the three-dimensional case. If we place a unit cube
with its center at each lattice point in space, then we will obtain a standard
tiling of space. In this tiling, each cube meets face to face with six other
cubes. (We say that two cubes meet face to face only if they share a complete
face with each other.) This tiling may be decomposed into layers parallel to
the xy-plane. Modified tilings can be formed by sliding different layers by
arbitrary amounts in the x- andy-directions. Within each layer, we may
either slide the rows in the x-direction or the columns in they-direction by
FIGURE 3
Tiling of the plane with displaced columns.
various amounts, just as in the plane case. In some layers we may slide
rows, and in others we may slide columns. In this way, we can obtain
tilings in which each cube meets only two other cubes face to face.
A similar construction can be carried out, starting with layers parallel
to any coordinate plane. But the cubes do not necessarily fall into layers; in
fact, we shall give several examples of tilings which are not layered. In
describing these, it is convenient to use the word column to denote a stack of
cubes lined up in the direction of anyone of the coordinate axes.
The simplest example of an unlayered tiling can be obtained by
starting with the standard tiling and choosing one column in each
coordinate direction in such a way that the three columns have no cubes in
common, and then displace each column along itself.
An extension of this idea will produce a tiling in which a certain cube
avoids meeting any other cube face to face. To do this, start with the
standard tiling, and select a cube which is to avoid meeting other cubes
face to face. The six adjacent cubes may be incorporated into six
columns-two in each coordinate direction-which may be slid inde-
pendently without interfering with each other or the central cube. For
example, the cubes adjacent to the central cube in the x-direction may be
slid in the y-direction; the cubes adjacent to the central cube in the y-
direction may be slid in the z-direction; and the cubes adjacent to the
central cube in the z-direction may be slid in the x-direction. In each case,
an entire column is moved in the indicated direction. The central cube
then meets no other cube face to face.
In a similar fashion, we can construct a tiling in which infinitely
many cubes avoid meeting other cubes face to face. The tiling described
below was discovered by Hans Jansen in 1909. Start with the standard
tiling obtained by placing a unit cube with its center at each lattice point
(x,y,z). Then shift the cubes with centers (x,y,z) by t unit
in the x-direction ify is even and z is odd,
in the y-direction if z is even and x is odd,
in the z-direction if x is even and J is odd.
The forms of the cross sections z' = c, where c is an even or odd integer, are
shown in Figure 4. These planes pass through the centers of most cubes.
However, in the case of cubes which were displaced in the z-direction, the
cross section passes through faces of the cubes. This is indicated in Figure 4
by cross-hatching. Arrows indicate the direction of the motion of the other
cubes. The cubes with x,y, z all even or all odd remain fixed, and therefore,
are unmarked. They do not meet any other cubes face to face. Obviously,
the cubes which have avoided meeting any other cubes face to face
constitute one-fourth of all the cubes. On the other hand, given any tiling
of space by congruent cubes, it is easy to see that if a certain face of a given
cube is not shared by another cube, then there must be two cubes touching
~228~
z even
FIGURE 4
Cross sections of the Jansen tiling of space.
this face which meet each other face to face. Thus, while some cubes can
avoid meeting other cubes face to face, not all cubes can do so.
We now consider multiple tilings of the plane by congruent squares or
multiple tilings of space by congruent cubes. We assume that each point of
the plane or space is covered a finite number of times, and that each point
not on the boundary of any square or cube is covered the same number of
times. If this number is k, then we speak of a k-fold tiling. The tilings
considered previously are I-fold tilings, or simple tilings. Multiple tilings of
the plane are not very interesting. It is not hard to see that any k-fold tiling
of the plane can be formed by superposing k simple tilings. Hence every
square will meet at least two other squares edge to edge.
In space, something new happens. Not all multiple tilings can be
formed by superposing simple tilings. Furthermore, it is possible for all
cubes to avoid meeting other cubes face to face. This was proved in 1974,
when I discovered a 25-fold tiling of space by congruent cubes of which no
two meet face to face. In describing this tiling, it will be convenient to use
cubes whose edge is 5 instead of using unit cubes. This change of scale
obviously does not affect the problem. The cubes used will all have centers
at lattice points. We will use exactly those centers (x,y, z) satisfying one of
the following four conditions:
x ==y == z (mod 2), x +Y + z == 0 (mod 5);
x + 1 == y == z (mod 2), x + y + z == 1 (mod 5);
x == y + 1 == z (mod 2), x + y + z == 2 (mod 5);
x==y==z+ I (mod 2), x + y + z == 3 (mod 5).
Here we have used the congruence notation a == b (mod m) to mean that a
and b differ by a multiple of m. Thus the first line above states that the
coordinates x, y, Z are either all even or all odd, and that their sum is a
multiple of 5. In general, the first condition of each line depends only on
whether the three coordinates are even or odd, while the second condition
.a229~
gives the remainder when x + y + z is divided by 5. The first condition will

be called the parity condition. All of the points satisfying one of the parity
conditions will be said to form a parity class. We shall show that all of the
above cubes together form a 25-fold tiling of space, and that no two of them
meet face to face.
To show that we have a 25-fold tiling, it is necessary to check that
every lattice point is covered 25 times. Now the centers of cubes with edge 5
containing a lattice point (x',y',z') are exactly the lattice points (x,y,z) in a
cube with edge 5 whose center is (x',y',zl We start by considering the case
(x',y',z') = (0,0,0). Thus, we are given the cube with edge 5 whose center
is at the origin. The lattice points in this cube are the 125 points (x,y,z)
with each coordinate assuming the values -2, -1,0, 1,2. These form a
5 x 5 x 5 array. We look first at the parity class defined by x == y == Z
(mod 2). This consists of two parts: there are 27 points in a 3 x 3 x 3 array
with x,y,z all even, and 8 points in a 2 x 2 x 2 array with x,y,z all odd.
We now form the sections x + y + z = c, which are perpendicular to
a diagonal of the cube. The sections of the 3 x 3 x 3 array will contain
1,3,6,7,6,3,1 points, as shown in Figure 5. The sections of the 2 x 2 x 2
array will contain 1,3,3, 1 points. Considering all 35 points together, we see
that they will lie in equally spaced planes containing
1,0,3,1,6,3,7,3,6,1,3,0,1
points. Combining those in which x + y + z (mod 5) have the same value,
we obtain the totals shown below.
103 1
6 3 7 3 6
130 1
7 7 7 7 7
Thus, there are exactly 7 lattice points in the given cube with x == y == Z
(mod 2) that have any prescribed value for x + y + z (mod 5).
All of the lattice points in the given cube form the 5 x 5 x 5 array
mentioned above. Again we consider the sections x + y + z = c. Three of
the sections are shown in Figure 6. The numbers of points in all of the
sections are listed below; they are combined according to the value of x + y
+ z (mod 5).
• • • • • • • • •
• • • • • • • • • • •
•
• • • • • •
FIGURE 5
Oblique sections of a 3 x 3 x 3 array .
..a230~
• • •
• • • • •
• • • • • • •
• • • • • • •
• • • • • • • • •
• • • • • • • • •
• • • • • • • • • • • •
FIGURE 6
Oblique sections of a 5 x 5 x 5 array.
1 3 6 lO
15 18 19 18 15
lO 6 3 1
25 25 25 25 25
Thus each value occurs 25 times. This could also be seen from the fact that
x andy may be assigned arbitrary values from the list -2,-1,0,1,2, and
then z is a uniquely determined value, when the value of x + y + z (mod 5)
is given. Each value of x + y + z (mod 5) occurs 7 times for points with x
== y == z (mod 2), and therefore 18 times for other points. Now the other
three parity classes are permuted by permuting x,y,z cyclically; in other
words, the lattice points of the other three parity classes are interchanged
by a rotation of 1200 about the line x = y = z. Thus, in each of these parity
classes, any prescribed value of x + Y + z (mod 5) must occur exactly 6
times in the given cube.
Now consider the cube of edge 5 with center at any lattice point
(x',y',z'). Ifwe move this point to the origin by a translation, then a lattice
point (x,y,z) lying in the given cube will go into a lattice point (x-x',y-y',z-
z') lying in the cube of edge 5 centered at the origin. If the point (x',y',z') is
given, then the parity class of this point is determined by the parity class of
(x,y,z), and the reverse is also true. In particular, we will have
x - x' ==y - y' == z - z' (mod 2)
just when (x,y,z) lies in the same parity class as (x',y',z'). From the results
proven above, we see that there are 7 such points (x,y,z) in the given cube
with a prescribed value for x + y + z (mod 5). But in anyone of the other
three parity classes, there will be only 6 points in the given cube with a
prescribed value for x + y + z (mod 5).
We thus see that the cubes whose centers satisfy anyone of the four
conditions in the definition of the tiling will cover all lattice points which
satisfy the same parity condition 7 times, and other lattice points 6 times.
Thus the four sets of cubes together will cover all lattice points 25 times, and
therefore, furnish a 25-fold tiling. All of this is independent of the values
assigned to x +y + z (mod 5) in the four cases .
.8231&'
We now check that no two cubes meet face to face. If two cubes did so
meet, then the center ofthe second cube would be obtained from the center
of the first cube by changing one coordinate by 5 units. This would change
the parity class, but not the value of x + y + z (mod 5)-which would be
impossible since different values of x + y + z (mod 5) were used for the
four parity classes. Thus we have a 25-fold tiling in which no two cubes
meet face to face.
Notice that in this tiling, formed with cubes of edge 5, the entire tiling
is unchanged by a translation of 10 units along any coordinate axis, or by
the translation through a vector with components (5,5,5). It can be shown
that these properties are consequences of the fact that no two cubes meet
face to face.
It is difficult to make any drawing which gives an idea of this tiling.
The best we can do is to take cross sections, and indicate the centers of the
cubes used. In Figure 7, we show two of the cross sections, those in the
planes z = 0 and z = 5. The portion drawn indicates the sections of the
unit cubes with centers at lattice points (x,y,z), where x andy vary from 0
to 9. The dots indicate the centers of cubes of edge 5 used in the 25-fold
tiling. Each of the two diagrams is repeated both horizontally and
vertically. Other cross sections have a similar appearance.
In order to have a multiple tiling of space, it is necessary that 5 x 5
squares drawn in the same position in both sections contain the same
number of dots. In order that no two cubes meet face to face, it is necessary
that the dots always lie in different positions in the two sections, and that no
two dots are 5 units apart either horizontally or vertically in either section.
The same 25-fold tiling described above was published recently in [1,
section 12] with a somewhat different proof. (The paper [1] was devoted
mainly to the corresponding problem in higher dimensions, but with the
additional condition that the centers of the cubes form a lattice.) For some
other related results and historical remarks, see Stein [2].
It is an open question whether there is a k-fo1d tiling of space by
congruent cubes with k < 25 for which no two cubes meet face to face .
• •
• •
• •
•
• •
• •
.
• • • •
• •
• • ..• • • •
. • .. .. •
••
•
• • • •
z'" 0 z '" 5
FIGURE 7
Cross sections of the 25-fold tiling of space.
~232~
References
1 Robinson, R. M. 1979. Multiple tilings of n-dimensional space by unit
cubes. Mathematische Zeitschrift 166: 225-264.
2 Stein, S. K. 1974. Algebraic tiling. American Mathematical Monthly 81:
445-462.
POSTSCRIPT Since this article was written, Basil Gordon has

shown that it is possible to take k = 2. Indeed, there is a 2-fold
tiling of space by cubes of edge 2, each composed of eight unit
cubes of the standard tiling, in which no two cubes meet face to
face.
~233~
Packing Handed
Pentacubes
-~-
c. J. BouwkaDlp
TECHNISCHE HOGESCHOOL EINDHOVEN
T o refresh the reader's memory, there are 29 distinct three-dimensional

pentacubes; a pentacube being a solid built from five unit cubes such that
neighbor cubes have a full face in common and any component cube has at
least one neighbor. Of them, the 12 planar pentacubes, also called solid
pentominoes, are well known in the field of recreational mathematics.
Together they have a volume count of60 units, and boxes of various sizes
such as (2 x 3 x 10), (2 x 5 x 6), (3 X 4 X 5) can be packed with them
[1,2,3]. Among the non-planar pentacubes there are 5 that have at least one
plane of symmetry; each of them is its own mirror image. The remaining 12
pentacubes consist of 6 pairs, each pair containing a pentacube and its
mirror image. The pentacubes of a pair are in the same relation to each
other as one's left and right hands; I shall, therefore, call them a pair of
handed pentacubes. The complete set of handed (or chiral) pentacubes is
depicted in Figure 1, together with their identifying labels 1 through 12.
Again, their total volume count is 60 units and, as with the planar
pentacubes, it is natural to ask whether they can pack a 3 X 4 x 5 box, for
example.
1 was confronted with this question in] uly 1973 through a letter from
Dr. G. V. Baddeley of the School of Chemistry, University of New South
Wales, Australia. His principal interest is in stereochemistry and he wrote
"to ask if [I] have a solution to the 3-4-5 block formed from the chiral
pentacubes and if so, whether ... the solution is unique".
~234~
PACKING HANDED PENTACUBES
FIGURE I
The handed pentacubes and their labels.
Parenthetically, ifthe reader is lucky enough to have access to a set of

handed pentacubes, I challenge him to find a solution on his own before
reading further. Although the handed pentacubes are much more difficult to
handle than the planar ones, he should not give up too soon, in view of the
overwhelming number of distinct solutions that are now known to exist.
A very clever, yet simple way to get many solutions at the same time
was dicovered by J. M. M. Verbakel of Philips Research Laboratories at
Eindhoven. He partitioned the set of handed pentacubes into four sets of
three each, so as to be able to form four simultaneous blocks, here called
pianos. One ofVerbakel's constructions is shown in Figure 2, where the two
pianos at the left are the mirror images (or duals) of the two at the right. It
is obvious that the four pianos can be assembled into two simultaneous 2
x 3 x 5 blocks (and in three different combinations) and hence into a 3
X 4 X 5 block, which settles the existence question raised by Baddeley.
Moreover, the solution is not unique at all. As a by-product ofVerbakel's
construction it is clear that 2 X 3 X 10 and 2 X 5 X 6 blocks can be formed
from the handed pentacubes as well, again in more than one way (as in the
case of the planar pentacubes).
By computer I found all possible partitions of the set of handed
pentacubes in four simultaneous pianos. There are 8 different partitions
modulo rotation and translation (or 6 modulo reflection, rotation, and
~235~
FIGURE 2
Four simultaneous pianos built from the handed pentacubes.
translation). They are listed in Table l. The first four of them are
attributed to Verbakel and these are all self-dual. Partitions 5 and 6 are
each other's dual, and so are partitions 7 and 8. (The reader is invited to
realize the remaining seven partitions of Table 1 in solid blocks (indeed a
neat exercise in manipulating handed pentacubes). He should observe that
there are two different pianos (5,6,9) and two different pianos (5,6,10).)
For the time being I shall only consider compound blocks 3 x 4 x 5;
that is, blocks that are built from rectangular blocks of smaller size. (By the
way, it is easy to see that a compound 3 x 4 x 5 block must necessarily be
built from two 2 x 3 x 5 sub-blocks.) By computer, I determined the
components of all compound 3 x 4 x 5 blocks. The results· are listed in
I (I, 6, 9) (2,5,10) (3,8,11) (4,7,12)
2 (1, 6, 9) (2,5,10) (3,8,12) (4,7,11)
3 (I, 7, 9) (2,8,10) (3,5,11) (4,6,12)
4 (I, 7, 9) (2,8,10) (3,6,12) (4,5,11)
5 (I, 6, 9) (2,8,10) (3,5,11) (4,7,12)
6 (I, 7, 9) (2,5,10) (3,8,11) (4,6,12)
7 (I, 8, II) (2,9,12) (3,4, 7) (5,6,10)
8 (1,10,11) (2,7,12) (3,4,8) (5,6,9)
TABLE 1
The eight partitions of the set of handed penta cubes each generating
four simultaneous pianos .
..a236~
PACKING HANDED PENT ACUBES
Number rif Solutions for Total (S =

Type Pentacubes in first block first block second block self-dual)
I (1,2,5, 6, 9,10) I 3 3S
2 (1,2,5, 7, 9,10) I I I
3 (1,2,7, 8, 9,10) 3 4 12 S
4 (1,2,7,8,11,12) I I IS
5 (1,2,7,10, II, 12) I 2 2
6 ( 1,3,4, 5, 9, II) 3 I 3
7 (1,3,4, 6, 7,10) I I I
8 (1,3,4, 6, 8, 9) I 2 2
9 (1,3,4, 7, 8, II) I 2 2
10 ( 1,3,5, 6, 8,10) 1 1 1
II (1,3,5, 6, 9,11 ) 4 I 4
12 (1,3,5, 7, 9, II) 2 2 1+ 2 S
13 (1,3,5, 8,10,12) I I 1S
14 (1,3,6, 7, 9,12) I I IS
15 (1,3,6, 7,10,12) 2 2 1+2S
Hi (1,3,6, 8, 9, II) 1 I IS
17 (1,3,6, 8, 9,12) 2 2 1+2S
18 (1,3,6, 8,11,12) 2 I 2
19 (1,4,5, 7, 9, II) I I IS
20 (1,4,6, 7, 9,10) I I I
21 (1,4,6, 7, 9, II) I I IS
22 (1,4,6, 7, 9,12) 3 3 3+3S
23 (1,5,6, 8, 9,10) I I I
24 (1,5,6,8,10,11) 2 I 2
Total
28+ 30 S
TABLE 2
The 24 types of compound 3 x 4 x 5 blocks
Table 2. I distinguish 24 types according to the six different pentacubes

that pack the first block (which contains pentacube I). Also indicated is
the number of different 2 x 3 x 5 blocks that can be formed from these
pentacubes, as well as the number of solutions for the second block with the
complementary pentacubes not indicated. The total number of distinct
(modulo reflection and rotation-and, of course, modulo translation-
implicit in the sequel) combinations of each type is indicated in the last
column, where the letter S stands for self-dual. These numbers add up to
28 + 30S, which means that there are 2 x 28 + 30 = 86 different (modulo
rotation) combinations of two simultaneous 2 x 3 x 5 blocks. Some, but
not all blocks involved are built from pianos. By way of an example,
Figure 3 shows type-3 blocks with three solutions for the first block and four
solutions for the second block. As usual, each block is represented by two
rectangular cross-sections, one for the bottom (left) and one for the top
~237~
1 8 1 1 8
f-- ~
8 9 10 8 9 10
'--
9 10 9 10
-7
lJ7 f---
2
2 7
- f---
2
2 7
7 9 7
1----1
9
10 2
10 8
8 2
6 4
W 12 11 4 11
12
4 3 3
S
4 3
N3
11 S 11 12 ~ 6
Is I
12
5 11 5 S 11 5
6 6
FIGURE 3
The basic 2 x 3 x 5 blocks of type 3 for the construction of compound
3 x 4 x 5 blocks.
(right) layer. It is clear that only three of the seven blocks are built from
pianos. It is remarkable that all seven blocks are symmetric; that is, each of
them has a center of symmetry.
The total number of different (modulo reflection and rotation)
compound 3 x 4 x 5 blocks turns out to be 372. Of them, 28 have a plane of
symmetry and as many have a center of symmetry. The set of 28 with a
plane of symmetry is complete, because a 3 x 4 x 5 block with a plane of
symmetry is necessarily compound. For the actual construction of com-
pound blocks with symmetry, I refer the reader to Figure 4 showing a 2
e.t238~
3 S 3 1
~
S
-
12 1 8
-
-.J
~
10 8 12
10
FIGURE 4
This block and its dual can be assembled into a 3 x 4 x 5 block with
either center or plane of symmetry each in two different ways.
x 3 X 5 block built from the pentacubes (1,3,5,8,10,12). The correspond-

ing dual block can be built from the complementary pentacubes
(2,4,6,7,9,11). The two blocks can be joined (in two ways) to form a 3 X 4
X 5 block with a plane of symmetry. They can also be positioned so that
the resulting block has a center of symmetry-a gain-in two ways. Many
other examples of 3 X 4 X 5 blocks with either type of symmetry can be
found with piano partitioning as in Figure 2.
I now turn to 3 X 4 X 5 blocks that are simple; that is, not compound.
An easy way to prove their existence is to refer to Figure 5 showing the
construction of simple blocks with the help of four simultaneous pianos.
Since the pianos in Figure 5 can be permuted in several ways and since
there are eight different partitions into pianos (see Table 1), it is evident
that simple blocks are numerous as well.
Although I often tried to find a "random" solution by hand I was
never successful-it was kind of unfair to challenge the energetic reader
above to get a solution of his own. Yet, I felt sure that the few hundred
solutions known so far represented only a small fraction of the total set of
solutions. I was right!
To program the problem for the computer to obtain one or a few
solutions is one thing, to get all of the solutions by computer is quite
another. To eliminate rotations and reflections is not easy in this case, and
solutions showing symmetry should be so indicated in the output. Both by
FIGURE 5
Four simultaneous pianos assembled into a 3 x 4 x 5 block that is not
compound.
~239~
6 7 lJ 9 6 7 ..!.l9 r1. 5
9 11
11 314 12 12 314
5 10 8 10 8
f2 ~ f2
6 7 4 4 1 6
7 10 10
11 3 4 3 11 8
2 5 8
5 5 2 9 2
1011 415
llJ 8
7
112 9 9
613 21 3 2
FIGURE 6
Five simple and two compound 3 x 4 x 5 blocks each with a center of
symmetry. With pentacubes I and 2 fixed as indicated the set is
complete.
the general program as well as by a special program I found that the total
number of different 3 x 4 x 5 blocks with central symmetry is 742. This
number includes the 28 compound blocks mentioned before. Note that
their contribution to the total is less than four per cent.
Figure 6 shows all possible solutions with pentacubes 1 and 2 fixed as
indicated in the lower right corner. A solution is represented by the three
cross-sections of (from left to right) bottom, middle and top layers. Five of
the blocks are simple and two are compound (type 15).
~240~
3 1 3 8 12 12
r-- r-
8 6 9 6 5 10
9 10 7
11 5 7 4 2
~ I---
11 2 4
FIGURE 7
Block with center of symmetry that can be decomposed into two
"steps". The two steps can be reassembled into a 3 x 4 x 5 block in
various ways.
A beautiful block among those with center of symmetry is shown in

Figure 7. It is composed of two "steps" which can be assembled into
different blocks in various ways. There are many more blocks of this
structure.
The last example of a block with center of symmetry is given in
Figure 8. It is peculiar in more than one way. As the reader will observe,
pentacubes (7,8,9,10) are not indicated in the 3-layer diagram of the
block. As a matter of fact, there are two possibilities to fill in the missing
labels in connection with the congruence of blocks built from pentacubes
(7,8) and (9,10) in the lower part of Figure 8. It so happens that the two
resulting blocks are each other's mirror image (modulo rotation). Blocks of
this structure are very scarce; only ten of them exist.
2 3 12 2 3 I 12 5
I
5 12
6 I -
11
11 6 4
14 1 11 I
FIGURE 8
Example of a block that is turned into its mirror image by translation
of four of the twelve pentacubes. For details, see text.
~241~
It is about time to close this paper and to disclose the total number of
solutions of the 3 X 4 X 5 block formed from the twelve handed pen-
tacubes. Believe it or not, the total number is 29162 modulo reflection and
rotation, and 57554 modulo rotation. They were obtained in computer
slack time during a period of about two years.
References
I Bouwkamp, C. J. Catalogue oj solutions of the rectangular 3 x 4 x 5 solid
pentomino problem. 1967. The Netherlands: Technische Hogeschool
Eindhoven, Department of Mathematics, Eindhoven.
2 - - . Packing a rectangular box with the twelve solid pentominoes.
1969. ]. Combinatorial Theory 7: 278-280.
3 _ _ . Catalogue of solutions of the rectangular 2 x 5 x 6 solid pen-
tomino problem. 1978. Kon. Ned. Akad. Wetensch., Proc., ser A 81: 177-186.
~242~
My Life AlDong The
PolyolDinoes
-~-
David A. Klarner
STATE UNIVERSITY OF NEW YORK.
M y life as a mathematician was probably launched by my father. He

had a scientific mind, and he was a very constructive man with the attitude
that he could do anything. I also had an excellent high school teacher,
Nemo Debely, who nurtured my interest in science and mathematics; in
particular, he encouraged me to read Martin Gardner's columns in The
Scientific American about the time they began to appear in the late fifties.
Inspired by my father's example, I had the attitude toward mathematics
that I could do it. When Deb gave me a classical problem in geometry,
number theory, combinatorics, or whatever, I set to work thinking I would
solve it. Of course, I never managed to solve Fermat's Last Theorem, the
four color conjecture (now a theorem), or to prove that there is an infinite
set of prime pairs, but I learned a lot by trying. Mixed into this naive
period was my first contact with Martin Gardner.
Imagine my thrill as a high school student when Martin Gardner
replied to my letters and put me in contact with other people who had
interests similar to mine. (For example, it was through Martin that I first
made contact with Solomon Golomb who later incorporated many of my
results and problems in his book Polyominoes.)
During my induction into the mathematical world, I had a series of
wonderful teachers: Nemo Debely in high school,James E. Householder in
~243~
college, Leo Moser in graduate school, and Dick de Bruijn in my post

doctoral period. Martin Gardner belongs in this list too, but his role differs
from the others. For twenty years Martin has been an abiding inspiration,
and he helped me share my mathematical hobbies with many people.
In this article I have attempted to present some of my mathematical
findings dealing with polyominoes-a subject that seems too technical
for a popular account. I intend, nevertheless, to present things so that non-
specialists will understand them. The first few sections deal with enu-
meration problems. The technical papers covering this subject comprise
scores of printed pages intersperced with ugly formulas; I have left out the
proofs and oversimplified things throughout. The last few sections give a
very brief account of some exact packing (tiling) results. While this subject
is easier to understand, and probably more attractive to the average
reader, it has been covered pretty well in Martin's columns and books
already. Thus, I will not try to outdo the master Gardner.
Growing Polyominoes
Polyominoes dwell amid the integer points of the Cartesian plane. They
are composed of cells, that is, unit squares that have their vertices at integer
points. An n-omino is a union of n cells whose interior is connected.
Solomon Golomb describes an n-omino as a set of n cells of a large
chessboard where a rook path through the cells of the n-omino connect
any two cells of it. Figure I indicates a unit cell, three pentominoes (5-
• • • • • • • • • • • • •
• •
• •
• •
• •
• • • •
•
•
•
o• • • • • • • • • • •
•
•
•
FIGURE I
A cell, two translates and a rotation of the F pentomino, and a
heptomino with a hole in it.
~244&.
My LIFE AMONG THE POL YOMINOES
FIGURE 2
The 19 translation types of tetrominoes.
ominoes), and a heptomino (7-omino). The heptomino shown in Figure 1

underscores the fact that n-ominoes may not be simply connected; that is,
they may have holes in them.
Another point underscored by Figure I is that for each n there are an
infinite number of n-ominoes even though there are only a finite number of
shapes involved. Various types of n-ominoes are defined on this infinite set
by notions of equivalence given in terms of certain groups of isometries of
the plane. First of all, translation types are defined by saying two n-ominoes
are equivalent if some translation of the plane maps one onto the other. For
example, translation types of the tetrominoes (4-ominoes) are shown in
Figure 2. Rotation types arise when two n-ominoes are equivalent if there is a
sequence of rotations of the plane which maps one onto the other. Since a
translation can be accomplished by composing two appropriate rotations,
translational equivalence is a refinement of rotational equivalence. In
terms of Figure 3, this means that the representatives of rotational types
may be selected from the set of representatives of translation types. An
intuitive notion of a rotation type of n-omino is that it is a one-sided puzzle
piece; that is, it can't be turned over but it can be moved around freely
otherwise. Finally, two n-ominoes may be considered equivalent if they are
FIGURE 3
The 7 rotation types of tetrominoes.
~245~
FIGURE 4
The 5 reflection types of tetrominoes.
congruent. Since every isometry is a composition of reflections, two n-

ominoes are congruent only when there is a sequence of reflections which
maps one onto the other. Congruent n-ominoes form reflection types. The 5
reflection types of tetrominoes are shown in Figure 4.
Computers have been used to generate complete sets of n-ominoes for
small values ofn (say, n::;; 24). But the number ofn-ominoes increases so
rapidly with n that there would not be enough storage space in our solar
system to list all the n-ominoes, even for fairly modest values of n (say
n ~ 30). Ronald Rivest has suggested an ingenious algorithm to generate
the rooted translation type of n-ominoes. To root a translation type of n-
omino, one of the n cells is distinguished from the others by endowing it
with an umbilicus. There are n times as many rooted translation types of n-
ominoes as there are translation types. The basic idea of Rivest's algorithm
is to define a growth process which creates each rooted translation type of
•
12
2 11 2 6 11 2
5
D
4
3 10
1
5
9
I I
4
3
8
7 10
17
5
4 8 15
14
12 18
12 18
11 2 16
11 2
10 5 20
10 5 3 7 14 20
17 4 8 15 21
17 4 8 15 22
16
16 23
FIGURE 5
Stages of Rivest's growth process for a 9-omino.
~246&
n-omino just once. To grow a rooted translation type n-omino, begin with
the umbilicus which is labelled 1. Then, start with the cell above the
umbilicus and label the cells surrounding it 2,3,4,5 in a clockwise fashion.
Cells adjacent to cells of a polyomino will be called border cells. The growth
process is a sequence of stages. Each stage consists of a polyomino which has
its border cells labelled with the smallest positive integers. A larger
polyomino is grown from this one by adding labelled border cells which
belong to the n-omino being grown. New unlabelled border cells usually
appear and these are labelled with the smallest positive integers not
already used as labels. These label assignments are made in a clockwise
fashion beginning above and nearest to the umbilicus. It is conceivable
that one must wind around the umbilicus more than once to assign all the
border cells their labels. Figure 5 shows the stages of growth involved in
creating a particular rooted 9-omino.
Rivest's growth scheme defines a family tree for rooted translation-
type n-ominoes. The set of numbers p(A) assigned to the cells ofa rooted n-
omino A is its name. For example, the rooted 9-omino shown in Figure 3 has
{1,3,5,6,7,9,13,14,21} as its name. A rooted polyomino B is said to be
descended from a rooted polyomino A if A's name is part of B's name; that is,
p(A) 5; p(B). Part of the rooted translation type polyomino family tree is
shown in Figure 6.
FIGURE 6
The family tree for rooted translation type polyominoes.
~247&.
If n is greater than 1, then 3n - 1 is the largest integer used in the

name of any rooted n-omino. This is obvious when you reason as follows:
the largest integer used to label a border cell of a rooted monomino (1-
omino) is 5. Addition of a cell to a rooted polyomino increases its border
cells by 3 (at the most). Hence, the largest integer used to label the border
cells ofa rooted n-ominois, at most, 3n + 2. When n > 1, a rooted n-omino
is grown by adding a border cell to some rooted (n - 1)-omino. The
integer used to label this border cell is, at most, 3(n - 1) + 2 = 3n - 1.
Thus, since a rooted n-omino is grown by adding border cells with larger
and larger labels, the largest integer used to label a cell is 3n - 1. This
shows that every rooted translation type n-omino has a name which is an n-
elemen t subset of the set {I, 2, ... , 3n - I}. Of course, 1 is in every name, so
rooted translation type n-ominoes correspond one-to-one with some of the
(n - I)-element subsets of the (3n - 2)-element set {2,3, ... , 3n.;.... I}. The
number of these subsets is (3n - 2) = (3n - 2)!/(2n - I)!(n - I)!. Ift(n)
n- 1
denotes the number oftranslation type n-ominoes, then nt(n) is the number
of rooted translation type n-ominoes. The foregoing argument shows that
1 t(n) ~ ~ (3n -
n n- 1
2).
But it is easy to verify that
2
1
(n+ 1)
(3n +
n
1)/1 ;;
(3n -
n-1
2) 3(9n 2 - 1)
=2(2n 2 + 3n+ 1)
27
<-
4'
For n = 1,2, ... , so a simple induction implies
27)n-l 27
3 t(n) < ( 4 or [t(n)p/n <-
4·
Methods for obtaining exponential lower bounds for t(n) will be described
in the next section. The best published lower bound for t(n) first appeared
in my Ph.D. thesis. It was shown there that
4 (3.72)n < t(n) or 3.72 < [t(n)p/n
holds for all sufficiently large n.
These bounds show that the number of translation type n-ominoes is
growing exponentially. Note that if r(n) denotes the number of rotation
type n-ominoes, and c(n) denotes the number of congruence type n-
ominoes, then
5 t(n)/8 ~ c(n) ~ r(n) ~ t(n).
Thus, the number of n-ominoes of each type grows at the same exponential
rate. This means that even the most efficient algorithm designed to
generate all translation type n-ominoes could not do its job for very large n
.8248~
My LIFE AMONG THE POLYOMINOES
(say, n = 30) because there wouldn't be space enough in all the world to
store the output.
Counting Polyolllinoes
The set of translation type n-ominoes is too large to list, even for moderate
values of n. This does not rule out the possibility that one might be able to
compute the number of elements in this set. Recall that t(n), r(n), and c(n)
denote the number of translation, rotation, and congruence type n-ominoes
respectively. Since t(n), r(n), and c(n) are bounded above exponentially, by
8" say, it follows that each of these numbers consists of not more than n
decimal digits. Sometimes there are good algorithms to compute terms of
sequences which grow exponentially. For example, the decimal repre-
sentation of2" can be computed using about 210g2 n multiplications. Also,
there is an algorithm which only requires on the order of log2 n basic
operations to compute the nth term of the Fibonacci sequence 1, 1,2,3,5,8,
13,21, ... (The nth term of the Fibonacci sequence is formedby summing the
two terms immediately preceding it.) Again, the Fibonacci sequence is
growing exponentially, so its nth term has about n decimal digits. Is there a
good algorithm to compute t(n)? To put this even more vaguely, is there a
formula for t (n )?
In a sense, there is a formula for t(n). For example, Rivest's growth
process could be used to form the set of all rooted translation type n-
ominoes, the elements of this set could be counted, the total divided by n,
and the result would be t(n). Unfortunately, this algorithm requires about
nt(n) steps, and nt(n) is exponentially large. As far as I know, all known
algorithms for computing r(n), t(n), or c(n) have this problem; that is, the
computation times for the algorithms are bounded below by elementary
functions of r(n), t(n), or c(n) instead of being bounded above by some
polynomial in n. It is conceivable that there is no good algorithm to
compute these numbers, but the recently-born theory of computational
complexity seems far from the level of development required to settle this
problem.
Surprisingly, even the very elegant algorithm discovered by R. C.
Read runs in exponential time. Read's algorithm computes tk(n), the
number of translation type n-ominoes located in a strip k cells wide. This
algorithm can be adapted to compute t(n), r(n), and c(n), and Read was
able to find these numbers by hand computation for about a dozen values
of n. Figure 7 shows translation type n-ominoes located in a strip 2 cells
wide for n=I,2,3,4. Here is a table of values oft2(n):
n 1 2 3 4 5 6 7 8 9 10 11 12
t2(n) 2 3 6 11 20 37 68 125 230 423 778 1431
~249&.
FIGURE 7
Translation type n-ominoes located in a strip 2 cells wide.
Perhaps after studying the table of values of t2(n), the reader will
notice that from the fourth term onward, t2(n) is equal to the sum of the
three terms immediately preceding it; that is,
l' tn(n + 3) = t2(n + 2) + t2(n + 1) + t2(n)
for n = 1,2,3, .... This formula indicates that the sequence [t 2(1), t2(2),
t2(3), ... ] satisfies a third order difference equation. Read proved the more
general result that for each k the sequence [tk(1), tk(2), ... ] satisfies a
difference equation whose order depends on k. Implicit in Read's proof is
an algorithm to compute this difference equation. He also formulated his
computations in terms ofpowers ofa s~uare matrix M k. To compute tk(n) ,
we must compute M k, M k , ... , M;:-k+ . Unfortunately, the size of Mk is of
order 3k , and it is necessary to compute Ml with k andj both nearly half of n.
This means Read's algorithm runs in exponential time. I suspect a computer
implementation of Read's algorithm would make it possible to compute
t(n), r(n), and c(n) for many new values of n.
Other subsets of translation type n-ominoes have been considered. For
example, Murray Eden, one of the pioneers in this subject, considered b(n),
the number of translation type n-ominoes which have each of its rows a
connected bar of cells. Such n-ominoes look like board piles (Figure 8).
Here is a table of values of b(n):
n 1 2 3 4 5 6 7 8 9 10
b(n) 1 2 6 19 61 196 529 1517 4666 14827
Eden was able to show that (3.14t < b(n) for all sufficiently large n. Leo
Moser suggested improving Eden's bound to me when I came to the
University of Alberta to start my graduate work. I was able to show that
2' b(n + 4) = 5b(n + 3) - 7 b(n + 2) - 4 b(n + 1)
for n = 1,2, .... From this it follows that (3.20t < b(n) for all sufficiently
large n. About ten years later, Donald Knuth suggested that board piles
~250~
FIGURE 8
A typical board pile.
were row-convex n-ominoes, and the convex n-ominoes should be defined as

board piles that are both row and column convex. Ronald Rivest and I
found a way to compute d(n) (the number of convex translation type n-
ominoes) and we proved that (d(n))l/n tends to 2.309138 .... Ed Bender
derived an asymptotic formula for d(n) based on our formulas.
Estilnating the Number of n-ominoes

In previous sections, it was indicated that t(n) (as well as r(n) and c(n))-
the number of translation type n-ominoes, grows exponentially with n.
Furthermore, it was shown all known algorithms for computing t(n) run in
exponential time. It may be a consequence of old age and growing
pessimism, but I think it is possible that there is no good formula for t(n). In
fact, the number t(n), for n (say, about 30) may be "unknowable". What
does this mean? The claim that t(30) is equal to a certain decimal number
(with less than 30 digits) must be accompanied with a proof. Suppose there
is no short proof? That is, it may be that all proofs of this fact are so long
that they cannot be generated even by the largest and fastest computer in
less than a thousand years of continuous work-a bleak prospect.
When faced with a difficult problem, it is often best to try a simpler
(yet still related) problem. In my Ph.D. thesis, I showed that (t(n)) l/n tends
to a limit fJ as n tends to infinity. Furthermore, the sequence (t(I), [t(2)p/2,
[t(3)p/3, ... ) increases to fJ, so that fJ can be approximated from below by
computing [t(n)p/n for growing values of n. Unfortunately, this involves
computing t(n) and there is no known good algorithm to do this
computation. Also, this sequence of lower bounds seems to be increasing
very slowly to fJ, so that a very large value of n would be required to
improve the lower bound 3.72 < fJ which I established by a completely
different method.
Rivest and I devised a method for computing a decreasing sequence
of upper bounds for e. The computer cost for this decreasing sequence of
upper bounds grows exponentially, and we were not able to show that the
sequence actually converges to e. e
Our best upper bound is < 4.65.
Recently, Pat Woodworth and I refined this scheme, and it seems likely
that the new method gives rise to a decreasing sequence of upper bounds
which does convelge to e. So far we have not implemented our method
on a computer.
Thus, the number fJ is also elusive; not one digit of the decimal
expansion of e is known! It seems possible that there is no good
et 251 fs.
computation for 0, and that apart from the first two or so, the decimal
digits of 0 are unknowable. (Sometimes in talks on this subject, I have
facetiously discussed 0 as "Klamer's constant". None of the digits are
known and they may be unknowable! If 0 = 4 could be proved, will school
children be taught to count, "one, two, three, Klarner's constant, five,
... "?)
A table of the known values of c(n) (the number of congruence type n-
ominoes) concludes this section. These numbers (and a wealth of other
information concerning polyominoes) were computed by D. H.
Redelmeier. There is an outstanding conjecture that the sequence of ratios
c(2)/c(1), c(3)/c(2), ... increases; that is,
c(n) c(n + 1)
--- < for n = 2,3, ....
c(n-l) c(n)
If this conjecture is true, then this sequence increases to its limit which is
equal to the Klamer constant O. Furthermore, each of these ratios would be
a lower bound for 0; in particular, c(24)/c(23) < 0, that is, 3.8977 ... < 0
which beats the current lower bound 3.72 ... < O. Of course, this improve-
ment depends on the truth of the conjecture.
n ern) n ern) n ern) n ern)
I I 7 108 13 238591 19 742624232

2 1 8 369 14 901971 20 2870671950
3 2 9 1285 15 3426576 21 11123060678
4 5 10 4655 16 13079255 22 43191857688
5 12 11 17073 17 50107909 23 168047007728
6 35 12 £3600 18 192622052 24 654999700403
TABLE I
The number of different n-ominoes.
Polyotnino Jigsaw Puzzles

Martin Gardner's 1958 article on polyominoes introduced me to the
subject when I was a student in high school. Featured in the article were
puzzles and proofs invented by Solomon Golomb. In particular, it was
pointed out that the twelve reflection type pentominoes (simply called
pentominoes from now on) can be used to form various rectangles with
area 60. The shapes are (3 X 20), (4 xIS), (5 X 12), and (6 X 10). Years,
later, C.]. Bouwkamp used a computer to make complete lists of solutions of,
all box-filling problems using the twelve solid pentominoes. Besides the 2-
dimensional boxes already listed, Bouwkamp also filled boxes with shapes
~252~
FIGURE 9
Two ways to fill a 3 x 20 rectangle with pentominoes. The shaded area
can be rotated 180 0 about its center to obtain the second solution.
(2 x 3 X 10), (2 X 5 X 6), and (3 X 4 X 5). Back in 1958, Martin's readers

did not have the benefit of Bouwkamp's lists or his ingenious computer
programs. In fact, I sought long and hard to find the two ways offilling a 3
x 20 rectangle with pentominoes, and was one of many enthusiasts who
sent his solutions to Martin. The two solutions (which are unique) are
indicated in Figure 9.
In preparing this article, it occurred to me that no one had considered
problems dealing with the rotation type pentominoes, that is, the eighteen
"one-sided" pentominoes shown in Figure 10. But it turns out that Golomb
mentions this set in his book Polyominoes by asking that these pieces be
used to tile a 9 x 10 rectangle. Probably the other problems discussed in
this section occurred to Golomb also. If the reader plans to make a set of
eighteen one-side pentominoes, use cardboard-preferably, black on
one side and white on the other. Then when playing with the pieces, they
must be kept one side up, say, black.
Since there are eighteen rotation type pentominoes, they might be
used to fill rectangles with area 90. It turns out that four shapes are
possible, namely, (3 x 30), (5 x 18), (6 x 15), and (9 x 10). The problems of
~~OudJ
[1JJ~cU[b~
BLbDCJ<J1) FIGURE 10
The eighteen rotation type pentominoes.
~253~
lis I' J l:lSfT? ~ N yl6 ~ 21

Is QF5ll1~
FIGURE 11
Rectangles tiled with one-sided pentominoes.
filling these rectangles become easier as one progresses from the slender to
the fat rectangles. For example, the 3 x 30 rectangle took me about two
hours to solve, while the 9 x 10 only required about two minutes. This
varying order of difficulty probably reflects the fact that there are few
solutions to the 3 x 30 problem, but there are thousands of solutions to the
9 x 10 problem. No doubt, Chris Bouwkamp will be the first person to
quantify this statement! Solutions to these problems are shown in Figure 11.
lt may seem difficult to tile a 3 x 30 rectangle with one-sided
pentominoes. However, once one tiling has been found, it is often possible to
transform this solution into new ones via a rotation or an interchange. To
illustrate a rotation, consider the block of pieces numbered 6 through 13 in
the tiling shown in Figure 11. This block can be rotated 180 0 about its
center to obtain the first solution shown in Figure 12. Again, by using the
tiling shown in Figure 11, two blocks consisting of pieces 6 through 11 and
15 through 17 can be interchanged to reach the second solution shown in
Figure 12. There are five different tilings obtainable by transforming the
.a254~
Interchange
FIGURE 12
Tilings of the 3 x 30 rectangle with one-sided pentominoes derived
from the tiling shown in Figure 9.
tiling shown in Figure 11; two of these involve rotations, and the other
three involve interchanges. Of course, these transformations can be applied
to the tilings shown in Figure 12 also, and nine new tilings result. This
process can be continued to find many new tilings. Perhaps the reader will
enjoy seeing how many tilings can be derived from these by applying a
sequence of rotations and interchanges. Of course, any tiling can be
reflected to reach a new solution, but such reflections really shouldn't be
considered distinct even though in a technical sense they are.
Early in my life among the polyominoes I tried to find elegantly
defined sets of jigsaw pieces, and then turned my attention to seeing what
could be done with them. This sort of enterprise has just been illustrated
with the one-sided pentominoes. The elegance of this set of pieces is, in
part, the simplicity of its definition. Over the years I have asked if even
simpler sets of puzzle pieces might give rise to complex problems. My own
inventions involved using several copies of one particular polyomino or
polycube. (A polycube is the three-dimensional version of a polyomino.)
~255~
The problem has always been to characterize the rectangular boxes which
can be exactly packed with copies of one piece. An example of this is given
in the next section.
The N-pentacube
An N-pentacube is shown in Figure 13-the way it is positioned indicates
where its name comes from. Also, it is clear that the N-pentacube is really
just the N-pentomino with its five unit squares replaced with unit cubes. It
is fairly easy to prove that the N-pentomino cannot tile (exactly pack) a
rectangle. In fact, it is impossible to tile one end of a rectangle with N-
pentominoes. Hence, the N-pentacube cannot tile a box having one of its
sides equal to one unit. Can any rectangular boxes be tiled with N-
pentacubes? It turns out the answer is yes, and I will begin by characteriz-
ing the boxes which can be tiled having one side equal to two units.
If an a x b x c box has been exactly packed with pentacubes, then the
volume of each pentacube (five) must divide the volume abc of the box.
This means that one of the integers a, b or c is a multiple of 5. Consider the
special case a = 2, then either b or c is a multiple of 5 (say c). It is easy to
show that the N-pentacube cannot tile the end of a 2 x b x c box for b
= 1,2,3 even when the third side c is a multiple of 5. However, it is possible
to tile boxes with dimensions (2 x 4 x 5), (2 x 5 x 5), (2 x 6 x 5), and 2 x 7
x 5 as Figure 14 shows. These boxes can be stacked to form larger boxes
which can also be tiled. It is fairly easy to show that every integer b greater
than 3 can be expressed in the form b = 4w + 5x + 6y + 7z with w,x,y,z non-
negative integers. Thus, the 2 x 5 ends ofw (2 x 4 x 5) boxes, x (2 x 5 x 5)
boxes,y (2 x 6 x 5) boxes, and z (2 x 7 x 5) boxes, can be matched together
to form a 2 x b x 5 box tiled with N-pentacubes. Then k of these boxes can
be matched together along their 2 x b sides to form a 2 x b x 5k box. The
conclusion is that the set S2 of all 2 x b x c boxes which can be tiled with
N-pentacubes consists of all those boxes having b greater than 3 and c a
multiple of 5. Furthermore, the four bricks shown in Figure 14 serve as
a basis for S2; that is, every 2 x b x c box which can De tiled with the
N-pentacube can be tiled with the four bricks shown in Figure 14.
The set S2' consisting of all 2-layer boxes which can be tiled with N-
pentacubes, contains boxes which can be stacked to form boxes with an
FIGURE 13
The N-pentacube.
~256~
FIGURE 14
Four prime bricks for the N-pentacube. Cells marked with a small
circle in the top layer are attached to cells marked with a small cross in
the bottom layer.
.a257~
xlo
x lo
xl r
10 xl
xlo
W 10 xlo
FIGURE 15
A 3 x 5 x 8 box tiled with N-pentacubes.
FIGURE 16
A 3 x 4 x IS box tiled with N-pentacubes. If the shaded pieces are
removed and the remaining ends pushed together, a tiling of a 3 x 4
x 10 box results.
~258~
even number of layers. This means the four boxes shown in Figure 14 are a
basis for boxes having dimensions 2a x b x 5c with a,c = 1,2,3, ... and b
= 4,5,6, .... None of these boxes have just 3 layers. Is it possible to tile 3-
layer boxes with N-pentacubes? So far, I have discovered tilings for a 3 x 5
x 8 box (shown in Figure 15), and tilings for 3 x 4 x 5k boxes for k
= 2,3, ... (shown in Figure 16). All 3 x 4 x 5k boxes with k greater than 1
can be tiled with 3 x 4 x 10 and 3 x 4 x 15 boxes. I t has been shown that the
3 x 4 x 5 box cannot be tiled with N-pentacubes. It is not known which
boxes with dimensions 3 x 5 x n can be tiled with N-pentacubes, but there is
an algorithm to settle this question.
~
~
gJrl
~
FIGURE 17
A cube of side 5 tiled with N-pentacubes.
~259~
Every 4-1ayer box which can be tiled with N-pentacubes can be tiled
with 2-layer or 3-layer boxes already tiled themselves. What about 5-layer
boxes? The key problem here is to tile a cube with side 5. I discovered the
tiling shown in Figure 17 after considerable effort; in fact, I nearly fainted
when the last piece fell into place!
There is enough of a variety of boxes tiled with N-pentacubes that it
can be shown that every a x b x c box can be tiled provided abc is a
multiple of 5 and that a,b,c are all sufficiently large. As the discussion in
the next section indicates, we should expect something like this. The boxes
that can be tiled but cannot be cut into smaller boxes (that can be tiled) are
called primes. The set of primes is finite, but unknown in the case of the N-
pentacube.
Tiling Boxes with Bricks

One of the many benefits of my two-year stay in the Netherlands working
with de Bruijn, was meeting Frits Gobel, another dweller among the
polyominoes. At the time, Frits had a notebook which stored information,
problems, and conjectures concerning this subject. He recalled a problem
set by Golomb which I had solved in the American Mathematical
Monthly. In a nutshell the result is this: Every rectangle that can be tiled
with L-tetrominoes can be tiled with 2 x 4 and 3 x 8 rectangles.
Furthermore, the 2 x 4 and 3 x 8 rectangles can themselves be tiled with
L-tetrominoes. Another result of this sort was established by David
Walkup; namely, every rectangle that can be tiled with T-tetrominoes can
be tiled with a square of side 4. Also, the square of side 4 can be tiled with
T-tetrominoes. (Tilings of 2 x 4 and 3 x 8 rectangles with L-tetrominoes
together with a tiling of the 4 X 4 square with the T-tetromino are shown
in Figure 18.) Based on these and similar findings, Frits conjectured that
the set of rectangles R which can be tiled with a given polyomino has a
finite basis B; that is, there is a finite subset B of R, having the property that
every rectangle in R can be tiled with the rectangles in B. Frits also
conjectured that if the two rectangles A and B have areas with no common
prime divisor, then all rectangles with both sides large enough can be tiled
with A and B. He wanted to use this result to prove the first conjecture, but
it turns out that the second conjecture is false.
FIGURE 18
Prime boxes for the L-tetromino and the T-tetromino .
.a260~
Stimulated by Gobel's conjectures, I found a necessary and sufficient

condition for an a X b rectangle to be tiled with p x p and q x q squares.
Either the rectangle can be tiled withjust one size of square, or it can be cut
into two smaller rectangles each of which can be tiled with just one size of
square. This means one of the following holds true: 1) Both a and bare
divisible by p. 2) Both a and b are divisible by q. 3) Either a or b (say a)
is divisible by both p and q while b = px + qy with x,y, non-negative
integers. This means that if p = 2, q = 3, none of the squares 5 x 5,
25 x 25, 125 x 125, ... , 5k X 5\ ... can be tiled with 2 x 2 and 3 x 3 squares.
Of course, this destroys Gobels' second conjecture, and one approach to
proving the first conjecture. The first conjecture is true however; in fact, I
was able to prove a much stronger result.
Suppose R is an infinite set of rectangles with integer sides. The
rectangles are specialized in that they are located in the plane and
oriented. That is, an a X b rectangle has its side oflength a parallel to they-
axis and the other side parallel to the x-axis. Thus, if a =1= b, then a x band
b x a rectangles are different. Rectangles relocated by translations are
considered equivalent. Next, the tiling is also specialized by being
restricted to a divide and conquer approach; that is, one is only allowed to
tile by cutting a rectangle into two smaller rectangles and then to tile these
rectangles by the same method. For example, my theorem about tiling
rectangles with two sizes of sq uares asserts that the rectangle can be tiled if
and only if it can be tiled by the divide and conquer method. Also, divide
and conquer tiling does not permit tilings of the sort shown in Figure 19.
My theorem deals with tiling oriented rectangles using the divide and
conquer method. It asserts that every infinite set R of oriented rectangles
contains a finite subset B such that every element of R can be tiled with
elements of B using the divide and conquer method. Furthermore, an
analogous result holds in k-dimensional space.
There is a very nice theorem about tiling boxes with bricks discovered
by Dick de Bruijn. He defines a harmonic brick to be one with dimensions
a x ab x abc with a,b,c integers. (A similar definition is made for higher
dimensional bricks.) The result is that a box can be tiled with these bricks if
I
I
FIGURE 19
A tiling which cannot be accomplished by the divide and conquer
method.
.a261~
and only if the box can be tiled with all the bricks parallel; that is, the box
has dimensions ap X abq X abcT. This means a harmonic brick tiles a box if
and only if the box can be tiled by the divide and conquer method.
Another theorem which I proved asserts that copies of a rectangle tile a
larger rectangle if and only if the rectangle can be tiled by divide and
conquer. This made me suspect that, in general, a brick tiles a box in k-
dimensions if and only if the tiling can be done by divide and conquer.
David Singmaster came up with a counter-example. A 1 X 3 X 4- brick tiles
a 5 X 5 X 12 box, but there is no way to tile the box by divide and conquer!
It follows from my solution to G6bels' first conjecture that all but a finite
set of boxes which can be tiled with a given brick can be tiled by divide and
conquer.
Happy Birthday Martin!
~262&'
Fun and Problems
-~-
...........•.............. ~I ·~o~de~ h~~ ~agi~i~~s· ~~k~ ·the·ir· ~abbits· dis·appe~~; :
Enchanted words like "hocus pocus" can:not interfere
with laws of science:and facts of mathematics that are clear .
. . .... .. . . .. . . . .. . . .. . ... . .. .. . .. .
The prestidigitators, making use of devious schemes,.
(although they never tell you how)·transport things as in dreams:
At times· suspended, banished, null and void - or so it seems.
~
~.
: There must be something secret, yes, a trick that will:involve fIl
- when done with sleight of hand -·a force that's able to dissolve . ~
. , .............. " ...................' ................................ .
~
:I:
~ trl
~
~ N.B.: When the right-hand portions of this eight-line poem are interchanged,
~ ~
Ol
...,. tD o
a seven-line poem results. Which line disappears? -D. E. Knuth >-
t"'
rp ~
... ~
:;.:l
t:I
~ Z
trl
:;.:l
.
Enchanted words like "hocus pocus" can· transport things as in dreams:
with laws of science:suspended, banished, null and void - or so it seems.
~
=
The prestidigitators, making use of devious schemes, : involve
tD
(although they never tell you how):a force that's able to dissolve.
fIl
At times:i ~~nd~r· ho~· ~~gicia~~ ~·ake· th~i~ ~~bbit~ di~~pp~a~;·:
There must be something secret, yes, a trick that will:not interfere
- when done with sleight of hand -:and facts of mathematics that are clear.
Noneuclidean
Harlllony
-~-
Scott Killl
STANFORD UNIVERSITY
Euclid
One of the outstanding achievements of the ancient Greeks was the
formation of a deductive system of tonal harmony constructed on the
principal sounds of everyday experience. This system was set forth in the 13
books of Euclid's Elements if" Harmony, otherwise known as the "Harmonic
Series". Starting from just 5 axioms:
1 Given two notes, there is an interval that joins them. (Axiom

of Arpeggiation)
2 An interval can be prolonged indefinitely. (Axiom of the
Fermata)
3 A progression can be constructed when its key center and
one of its chords are given. (Axiom of the Sequence)
4 All tritones are equal. (The Devil's Axiom)
5 If a melodic line accompanying two melodic lines a fifth
apart makes the interior intervals on the same side less than
two tritones, the two melodic lines, if extended indefinitely,
resolve on that side where the intervals are less than the two
tritones. (The Fifth postulate)
~265~
Euclid was able to derive such theorems as:
The sum of the intervals of a triad is equal to two tritones.

He also investigated Apollonian progressions, and obtained the result
that every point inside a circle of fifths has an invertible counterpoint.
Euclid, however, was clearly not satisfied with the statement of his
fifth postulate. It lacked the concise elegance of the previous four. Indeed,
Euclid delayed the entry of the fifth postulate in his "Elements" as long as
possible, while employing the first four postulates to their full potential. At
one point he considered appending yet another rule, to be called the
"augmented fifth postulate", but rejected it as merely a "minor sixth."
During the following centuries, many people tried their hands at
vindicating Euclid of every flaw. Perhaps the fifth postulate was re-
dundant. Perhaps all pieces composed with the aid of the fifth postulate
could be composed with only the first four. Some attempted to show this by
deceptive cadence-the method of "resolution theorem-proving."
The 17th century Italian music theoretician, Gerolamo Saccheri,
composed whole suites of pieces contradicting the fifth postulate. He
expected to find a cross-relation; none appeared. Saccheri's work con-
stitutes one of the earliest examples of Noneuclidean composition. His
training, however, prevented him from realizing the importance of his
discovery. Unable to accept the new sonorities, he declared his com-
positions to be "repugnant to the nature of the melodic line", thus denying
his proper place in history. The fifth postulate would remain unresolved for
many years.
The Greek Ideal

The failure of Saccheri, like that of his predecessors, was due to the
dominant influence of Greek thought over musical theory. The Greeks
conceived of a universe governed by divine ratios. The same ratios ruled
the orbits of the planets, the harmonies of geometry, and the diagrams of
music. Many examples of simple ratios were to be found. The Pythagorean
theorem related the pitch of a hypotenuse to the lengths of the sides. If a
length was doubled, the pitch was halved. Pythagorean tuning systems
were based on the ratio 3/2, called the Golden Interval which was considered
to be the interval most pleasing to the ear.
Ultimately, the Greeks related all ratios to the overtone series, a
theme echoed for many centuries in the quotation "God created the
overtone series; all else is the work of man". From the overtone series they
derived the different tuning systems, or modes. Modal logicians believed
that tuning had a strong effect on a person's temperament. Ratio and
reality were inseparable; music, then was a revelation of the existing order
of the universe.
~266~
NONEUCLIDEAN HARMONY
Since there was only one music, there was no need to make it explicit.
To a modern musician, Euclid's common notations, such as "the score is
greater than the parts", sound curious. They allude to further notations
which are equally obscure and fail utterly as self-contained definitions.
Similarly, while composers following Euclid recognized the importance of
proper voice-leading, they felt obliged to notate only those steps which
might cause confusion.
This tendency towards abbreviation lead directly to a number of
confusing notational practices, including mathematicaficta, figured bass, and
ornamentation. The performer was required to fill in the missing steps in
the appropriate manner. The great mathematician J. S. Bach was
considered old-fashioned for his praotice of writing out all ornamentation.
The most subtle omission involved instrumentation. Composers frequently
relied on the specific properties of their instruments. If a performer is to
reconstruct the sound totally from scratch, much more information is
needed.
Greek theory was not without dissenters. Critics raised fundamental
questions of interpretation, some of which have been decided only in
modern performance practice. Zeno of Alea questioned the interpretation
of dotted notes. He argued that a whole note is impossible, since one must
dot a half note infinitely many times. The critic Epimenides made the
notorious self-criticizing statement "all critics are liars". Epimenides'
paradox in more modern form, "this song is unsingable," ultimately led to
Goders "Universe Symphony", which is either uncriticizable or incom-
plete. The question of the universality of Euclid's harmony, however,
remained unasked.
Noneuclid
The dissonance was finally resolved some 2000 years later by Noneuclid
(1874-1951). Near the end of July 1921, Noneuclid told a pupil that
"At one o'clock minus 12 minutes today, I made a discovery that will
ensure the supremacy of German geometry for the next 100 years." He
called the elements of his new system "at-one's" after the first two words of
his famous quotation, and his system came to be known "at-onal"
geometry.
Basically, 12-to-one geometry dispensed entirely with the notion of
parallel intervals, which had been the cause of much worry, and declared
that no intervals would be forbidden. Noneuclid took as his model the face
of a clock. Melodic lines were to be identified with the hands. Since the
hands always met at the center, there could be no parallel intervals. In
Noneuclidean harmony, each atone receives a number from 12 to one,
based on the numbers of a clock. Atone row, or series, is then constructed on
this scale from 12 to one. Series may be added, subtracted, summed, or
~267~
differentiated. The last operation (atone differentiation) is of particular

interest to psychoacousticians.
Noneuclidean harmony upset the Greek ideal of simple ratio.
Atonality requires the complete symmetry of equal-temperament, a tuning
system in which intervals cannot be expressed in simple whole number
ratios by flattening out the irregularities of the modes. Noneuclidean
harmony severed its ties to temperament, and thus to any direct connection
with reality.
For many years, the exact significance of the new harmonic system
eluded the critics. Certainly one could invent a new set of axioms, and
proceed to compose pieces in that style. But would the results be
meaningful? Would the structure be audible? Tonality and atonality
seemed totally contradictory. If one embraced the new system, what would
be the logical status of Euclidean harmony? Would the efforts of Euclidean
composers have been wasted?
Noneuclidean harmony forced musicians to broaden their attitudes
towards the relation between music and the world. Noneuclid had shown
that there existed harmonies in direct opposition to Euclidean harmony
but nevertheless maintained internal consistency. This meant that har-
mony must be treated as an abstract game, quite independent of its
descriptive power. Composers have now begun to shift their attention from
writing pieces in existing styles to the iI).vention of new styles themselves.
While all harmonies may be equally valid in a compositional sense,
the question remains as to which system accurately describes acoustical
reality. The final irony was revealed by recent work in psychoacoustics,
which has discovered that the ear does not hear precise whole number
ratios. Perceptual space is warped by many other factors, such as the
volume of a sound, and the rate at which it is moving. So the Euclidean
model is not correct even as a description of real sounds. In fact Euclidean
harmony is just one point on a continuum of possibilities which includes all
the varieties of atonality. For everday purposes Euclidean harmony
suffices, since the intervals of equal-temperament are close approximations
to those of the overtone series. The difference is only apparent in extreme
cases, especially when modulations occur at a very fast rate. No wonder,
then, that Noneuclidean harmonies were not realized earlier.
Recent Developments
With atonal geometry as their model, many other pathological com-
positional styles began to appear. Some people attempted to write pieces
which could be played in more than one way, creating compositions of two
or more functions. Others advocated the use of chance to express their
compositions' indeterminate form. This trend met with considerable
resistance. The acoustician Albert Einstein, who developed the revolu-
~268~
NONEUCLIDEAN HARMONY
tionary theory of relative pitch, was quoted as saying "God does not play
dice" .
In the 1870's, Georg Cantor began investigating the theory of
supersonic pitches. Previously, composers had assumed that all pieces of
unbounded speed were equally inaudible. Cantor showed otherwise, by
using notions of 2 against 3 correspondence and uncountable rhythms. His
research lead eventually to the ear-filling melodies, such as Piano's curve,
sung by such birds as the orioles and the cardinals. While these sounds
cover all frequencies, they are nondifferentiable. Cantor was also able to
show that the violin has more notes than a piano, by the "glissando"
method.
In 1976, the duo pianists Haken and Appel realized the long-
unperformed "Map Quartet", which suggests that every national anthem
can be sung with only four voices. Their realization used a computer
program, which systematically enumerated the enormous number of
permutations drawn from what was basically a very simple musical idea.
In so doing, Haken and Appel revived basic questions concerning the
nature of beauty. If no one can hear the structure of a composition can it
claim to be beautiful? Who is the composer of a piece of computer music?
The answers to these questions have yet to be found.
For further information on Noneuclid and discussions of truth in
harmony, the reader is referred to "Gadel, Escher, Bach: an Eternal Golden
Braid" by Douglas Hofstadter (Basic Books, 1979), which metaphorically
weaves together the works of the music theoretician Kurt Gadel, the artist
Maurits Cornelis Escher, and the mathematician Johann Sebastian Bach.
~269~
Magic
Cuboctahedrons
-~-
Charles W. Trigg
Los ANGELES CITY COLLEGE
The surface of a cuboctahedron consists of eight equilateral triangles

and six squares, with triangles and squares alternating around each of the
twelve vertices, as indicated in the orthogonal projection of Figure 1.
The squares form a vertex-connected network with two squares at
each vertex. They lie by twos in parallel planes. Between each parallel pair
is a medial plane containing four vertices of a medial square. Thus, the
vertices of the cuboctahedron lie on 9 squares, with three squares on each
vertex.
FIGURE 1
Orthogonal projection of cuboctahedron.
~270~
MAGIC CUBOCTAHEDRONS
.-.:::::_ _ _::--_---::::,., n
k~-----~~
FIGURE 2
A Schlegel diagram of the cuboctahedron.
When twelve elements having a sum 1: are distributed on the vertices

of a cuboctahedron in such a fashion that the perimeter sum of each surface
square is e, the cuboctahedron is said to be magic. Then, summing the
perimeters, 6e = 21:, so the magic constant e = 1:/3. It follows, from
consideration of three parallel squares, that the perimeter sum of each
medial square is 1: - 2(1:/3) or 1:/3, also. In the Schlegel diagram of
Figure 2, the quartets on the vertices of the medial squares are e,j,g,h;
a,c,p,m; and b,d,k,n.
The eight surface triangles lie by twos in parallel planes. When the
perimeter sums of the three squares surrounding triangles ajb and hpk are
equated, and the equation is simplified, we have a + f + b = h + P + k;
that is, on a magic cuboctahedron, the perimeter sums of opposite triangles
are equal.
When the elements on the vertices are the first 12 positive integers,
then 1: = 78 and e = 26. As a first step in identifying the magic
cuboctahedrons, the 33 partitions of 26 into four distinct integers $ 12 are
listed below. To achieve compactness, the integers la, II, and 12 are
represented by X, r, and Z, respectively.
12rz 14xr 239Z 258r 3472 349X 456r
13XZ 159r 248Z 267r 356Z 358X 457X
149Z 168r 257Z 259X 348r 367X 4589
158Z 169X 23Xr 268X 357r 3689 4679
167Z 178X 249r 2789 5678
With 1 assigned to fixed position a, two quartets having only 1 digit in
common are needed for distribution in the positions a,b,c,d and a,e,m,j
Eleven quartet pairs meet this requirement, namely: 12rz, I 69X; 12rz,
178X; 13X.(', 159 r; 13XZ, 168 r; 149.(', 168 r; 149.(', 178X; 14Xr, 158.(';
14Xr, 167.('; 158.(', 169X; 159r, 167.('; and 159r, 178X. With 1 fixed in
position, the distribution of each quartet on the vertices of its quadrilateral
can be completed in 3! ways. Thus there are 6 2 (11) or 396 initial
distributions to consider.
~271&'
3~ ______________~
FIGURE 3
Complementary magic cuboctahedrons.
One distribution of the first quartet pair is shown in the diagram on

the left in Figure 3. A quartet, 457 X, containing X and no other integers in
common with the other quartets, was selected for distribution on the outer
square. Then, a quartet containing 9 and 2 and one of 4, 5, and 7 was
sought for placement on the Jngb quadrilateral. The quartet 2789 was
r
chosen. A quartet containing 8 and and one of 4 and 5 was needed next.
When 348r was placed on the quadrilateral cgph, the digit 5 was left for
placement on the vertex k to complete the magic cuboctahedron.
In this manner, 40 magic cuboctahedrons have been identified. They
exist as 20 reflective twins. One twin is converted into the other by
reflection in the plane of vertices m, a, c,p. Only one of each twin pair has
been recorded in Table 1. Distributions that can be rotated to coincide are
not considered to be distinct.
Two integers will be said to be complementary if their sum is 13. Two
magic cuboctahedrons will be said to be complementary if their cor-
responding elements are complementary. The two cuboctahedrons in
Figure 3 are complementary.
In Table 1, the complementary cuboctahedron pairs are: A, B; C, D;
E, F; G, H; and I, ]. In each of the pairs, the a, b, c, d groups are
permutations of the same quartet of integers. The same is true of their e,f,
g, hand k, m, n, p groups, except that in the A, B and I, ] pairs, the quartets
are interchanged.
The ten cuboctahedrons, K through U, are self-complementary.
Within the pairs N, P; R, S; and T, U, the a,b,c,d and k,m,n,p distributions
are identical. Within the pairs N, P and R, S, the e,j,g,h distributions are
reversed. In the set R, S, T, U, the a,b,c,d groups are permutations of the
same integers, as the e,f,g,h and the k,m,n,p groups are also.
Since the corresponding triangles are opposite, the perimeter sums of
the triangles bordering the outer square in Figure 2 are the same as the
perimeter sums of the triangles bordering the inner square. Proceeding
MAGIC CUBOCTAHEDRONS
Penta- Locations on Schlegel Diagram Triangle Sums

gram a b c d e f g h k m n p fmn gnp hkp ekm
A I 2 r Z 6 9 8 3 5 X 7 4 26 19 12 21
B I Z r 2 X 7 4 5 9 8 3 6 18 13 20 27
C I 3 X Z 5 r 8 2 7 9 4 6 24 18 15 21
D I Z X 3 r 8 2 5 7 6 4 9 18 15 21 24
E I 9 4 Z 7 8 6 5 2 X 3 r 21 20 18 19
F I Z 4 9 8 6 5 7 2 r 3 X 20 18 19 21
G I 5 8 Z 4 X 9 3 7 r 2 6 23 17 16 22
H I Z 8 5 X 9 3 4 7 6 2 r 17 16 22 23
I I 7 6 Z 4 r 3 8 2 X 5 9 26 17 19 16
J I Z 6 7 5 9 2 X 4 r 3 8 23 13 22 20
K I 2 Z r 6 9 7 4 5 X 8 3 27 18 12 21
L I r Z 2 X 8 3 5 9 7 4 6 19 13 20 26
M I 9 Z 4 8 X 5 3 r 7 2 6 19 13 20 26
N I r 4 X 6 7 5 8 2 Z 3 9 22 17 19 20
P I r 4 X 8 5 7 6 2 Z 3 9 20 19 17 22
Q I 8 Z 5 6 9 7 4 r X 2 3 21 12 18 27
R I 9 r 5 7 6 3 X 4 Z 8 2 26 13 16 23
S I 9 r 5 X 3 6 7 4 Z 8 2 23 16 13 26
T I r 5 9 7 X 3 6 4 8 2 Z 20 17 22 19
U I r 5 9 X 7 6 3 4 8 2 Z 17 20 19 22
TABLE 1
Magic cuboctahedrons, constant = 26.
clockwise from the topmost triangle, the perimeter sums of the bordering
triangles are recorded in Table l. In every case the four sums are distinct.
The quartet of sums is the same in the pairs C, D; E, F; G, H; K, Q; L, M;
and R, S, as well as in the group X, P, T, U. In C, D, E, and F, the four
sums are in arithmetic progression. Alternate sums total 39 in C, D, E, F, G,
H, K, L, M, and Q In X, P, R, S, T, and U, there are consecutive sum pairs
that total 39.
Cuboctahedrons L, M, R, and S are supermagic with 9 squares and 2
triangles having perimeter sums of 26.
There is no distribution of the first 12 integers such that the perimeter
sums of the eight triangles are the same, since 2(78)/8 or 39/2 is not an
integer.
~273~
GaDles, Graphs,
and Galleries
-~-
Ross Honsberger
The Bulging Railway

Let's begin by giving Martin a little test (due to Murray Klamkin) of his
mathematical intuition. Imagine a flat stretch of straight railway track AB,
5000 feet long, which is immovably fastened down at each end (Figure 1).
In the heat of the summer the track expands 2 feet, causing it to buckle.
Assuming it bends symmetrically, how high do you think the bulge rises
above the ground at the middle? An inch? 4- feet? A tenth of an inch?
Since the total length of the track is now 5002 feet, each half would be
2501 feet. While it is clear that the buckle would. take the form of a gentle
curve, we can get a rough idea of the situation by assuming it to be
straight. On this basis we obtain an estimate of the height x by applying the
theorem of Pythagoras:
x = )2501 2 - 2500 2
= )(2501 - 2500) (2501 + 2500)
= )5001,
making it approximately 70 feet! This comes as a big surprise to many
people. The actual value, taking into account the curve in the track, is
closer to 67 feet. How close did you come, Martin?
~274-~
GAMES, GRAPHS, AND GALLERIES
A ---2500-----1 B
FIGURE 1
Aritlunetic Mastermind
A year or so ago a mathematical game called Mastermind swept across the
country. One player would secretly arrange 4 colored pegs in a row and the
second player, by a combination of guessing and reasoning, would try to
determine the colors and order of the pegs in as few steps as possible. Let us
playa similar arithmetic game, also due to Murray Klamkin.
We begin by having you choose 5 positive integers less than 100,
repetitions allowed:
It is up to me to determine these numbers as quickly as I can. I am

permitted to enlist your help through the following kind of dialogue. I
present to you a set of 5 numbers (Xt>X Z ,X3,X4,XS) and you respond, not as
the Code-maker does in Mastermind, by telling me how many of my x's are
correct and how many are in the correct places, but by telling me only the
single number SI obtained by multiplying the corresponding numbers in
our selections and adding:
SI = aIx l + azx z + a3x 3 + a4x4 + asxs·
Armed with this value Sv I then propose another slate (YI,Yz'Y3'Y4'YS). If!
am correct the game is over; otherwise you respond with the number Sz
given by
Sz = alYI + avz + a3Y3 + a4Y4 + asYs·
This procedure continues until I get them right.
It is clear that after 5 steps I am in possession of 5 equations in the 5
unknowns al,a Z ,a3,a4,aS and can determine what they are. Thus the game
never needs to go beyond 5 steps.
Of course, the idea is for me to determine your selections in as few
steps as possible. Would you believe, Martin, that I can always do it in 4
steps? Would you believe 3 steps? 2 steps? The fact is, it can be done injust
one step!
Since I know that each of your numbers is less than 100, I choose my
first slate to be
(l08, 10 6 , 104 , lO z, 1).
This leads you to compute
SI = 100000000a s + 1000000a4 + 10000a3 + 100a z + aI'
~275~
which displays your selections as 2-digit numbers from left to right. For
example, if
(a1,a2,a3,a4,aS) = (17,68,5,42,8),
then Sl = 1768054208, and it's all over.
The First n Positive Integers

I t is well known, and easy to establish, that
1 + 2 + ... + n = n(n + 1).

2
The following novel approach to this result is due to Donald Snow of
Brigham Young University.
Let
and
Then
S2(n + m) = 12 + 22 + ... + (n + m)2
= (12+ 22+ ... + n2) + (n+ 1)2+ (n+ 2)2+ ...
+ (n+ m)2
= S2(n) + (n 2 + 2n + 12) + (n 2 + 4n + 22) + ...
+ (n 2 + 2mn + m2)
=S2(n) + mn 2 + 2n(1+ 2+ ... + m) + (12+ 22+ ...
+ m2 )
= S2(n) + mn 2 + 2nS 1(m) + S2(m),
that is,
S2(n + m) = S2(n) + S2(m) + 2nS 1(m) + mn 2.
Interchanging nand m, we have
S2(m + n) = S2(m) + S2{n) + 2mS1 (n) + nm 2.
Then the equality S2(n + m) = S2(m + n) immediately simplifies to
2nS 1(m) + mn 2 = 2mS 1(n) + nm 2.
Since Sl (1) = 1, the substitution m = 1 gives
2n + n2 = 2S 1(n) + n,
.8276&.
n2 + n
and out pops our formula Sl (n) = -2-.
I never cease to be amazed at the remarkable relation

13 + 23 + 33+ ... + n3 = (1+ 2+ 3+ ... + n)2.
When he was visiting from Australia, Roger Eggleton pointed out the
following geometric demonstration of this fact. Since the area of k squares
of side k is k' k 2 = k 3 , let us arrange in the plane
I square of side 1, 2 squares of side 2, 3 squares of side 3, ... , n squares
of side n,
having a total area of 13 + 23 + 33 + ... + n3 • Let us do our best to build
up a layered square by adding the squares of a given size around the outside
of the part that has been constructed (Figure 2). This succeeds pretty well,
except for the "even" squares, 2,4,6, ... , which both overlap ( ~ ) and
leave uncovered gaps ( 0 ).
A moment's reflection shows that each overlapping square cor-
responds to a blank square of the same size, yielding the conclusion that the
total area is that of a square of side (1 + 2 + ... + n).
Herb Shank's Millunan's Route

In the middle of 1974, my colleague Herb Shank made an unusual
discovery about graphs. I would like to describe his result without
considering its proof To avoid getting involved with the explanation of
several technical terms in graph theory, let us confine ourselves to the
familiar setting of geometry. In so doing, the theorem will not be presented
in its full generality. In order to state the whole story, permit me to address
eJ/277~
FIGURE 3
the following single sentence to those readers who are acquainted with
simple graph theory:
our story begins with any planar, Eulerian graph which has an
odd number of spanning trees.
Suppose a configuration F is constructed in the plane by connecting
polygons together at single vertices. Any number of polygons may be used
in any arrangement you please. One polygon may even occur inside
another. We need to observe only three things:
1 every polygon is to have an odd number of sides,

2 no two edges are permitted to cross,
3 no polygon is to be connected into the figure at more than one
vertex (this avoids producing a "ring" of polygons).
Now, let the regions in F be colored, alternating red and a nice lime
green so that regions which border along the same edge have different
colors. The above conditions guarantee that such a coloring is always
possible. In order to keep our example reasonably simple, we illustrate
with two triangles and a pentagon (Figure 3). We use R for red and L for
lime.
"\
\
I
JG
"" '-.
-------
FIGURE 4
,/' /
/
~278~
/-_-7'\
/ ./ I
H( (-:----~
\ G
\
"-
'-
G: solid lines
H: dotted lines
FIGURE 5
Next, we embellish the figure by inserting any additional edges we

like between pairs of vertices of F, subject only to the condition that no two
edges cross either themselves or edges which are already in F (this allows
even for the introduction of multiple edges, that is, edges which connect
vertices that are already joined by an edge). Let the resulting figure be
called G (Figure 4).
From G we derive an independent figure H, called the dual graph, as
follows:
1 place a vertex in ea€h region of G, including the infinite

outside region;
2 join the vertices of two neighbouring regions of G by drawing
an edge from the one to the other so that it crosses through the
interior of an edge in their common boundary; insert such an
edge for each edge in their common boundary (Figure 5).
Finally, we assign the label Rand L to each vertex of H according to

the color of its region-red or lime (Figure 6).
FIGURE 6
8279&'
Now we have a labelled graph H, which, no matter what choices

might have been exercised in its construction, always possesses the
following remarkable property:
Beginning at any vertex V as a starting point, and proceeding
along any edge of that vertex, one will traverse each edge
exactly twice, once in each direction, and end up at V again, by
always making a hard right turn at each vertex labelled R and a
hard left turn at each vertex labelled L.
Such a tour can be charted nicely by placing a numbered arrow in the
appropriate direction on each edge as it is traversed. Thus, for example, we
might obtain the following-Figure 7:
FIGURE 7
Who Stole The Apples?

I imagine that many of us have spent time during our holidays working on
some of those brain teasers that call for "each man's occupation" or "who is
married to whom?". At the age one is often introduced to such a pastime, a
straightforward trial-and-error process of elimination is likely to be the
only approach at one's disposal. However, certain problems of this sort
have very ingenious solutions. For example, consider the following
problem:
Out of 6 boys, exactly two were known to have been stealing apples.
But who? Harry said "Charlie and George". James said "Donald and
Tom". Donald said "Tom and Charlie". George said "Harry and
Charlie". Charlie said "Donald and James". Tom couldn't be found.
Four of the boys interrogated had named one miscreant correctly, and
one incorrectly. The fifth had lied outright. Who stole the apples?
The problem is not without its appealing twists-Tom couldn't be found;
one of them lied outright. In my Ingenuity in Mathematics I gave a rather
~280~
involved formal approach to the problem. Recently my good friend and

colleague Scott Vanstone showed me the following brilliant solution which
translates the whole thing into a simple problem concerning a graph.
Making use of an observation of John Annulis (University of Arkansas at
Monticello), the solution is virtually transparent.
Let a graph (Figure 8) be constructed with a vertex for each boy (C for
Charlie, G for George and so on), and with an edge for each pair of boys
named in the statements (edge CG reflects Harry's statement, etc.). Thus
every edge but one has exactly one end at a thief, and the remaining edge,
due to the outright liar, has neither end at a thief. That is to say, the two
vertices which represent the thieves, have between them a total of 4
endpoints (one from each of four edges and none from the fifth edge). We
should note also that the thieves themselves, are not joined by an edge. We
see immediately that the culprits can only be Charlie and James (C and J).
... . : : - - - - - - - H
C
G
J
FIGURE 8
Chvatal's Art Gallery Theorem

Chapter 11 in my book Mathematical Gems II is entitled "Chvatal's Art
Gallery Theorem". It concerns a certain problem of guarding the
paintings in an art gallery. The way the rooms in museums and galleries
snake around with all kinds of alcoves and corners, it is not an easy job to
keep an eye on every bit of wall space. The question is to determine the
minimum number of guards that are necessary to survey the entire
building. The guards are to remain at fixed posts, but they are able to turn
around on the spot. The walls of the gallery are assumed to be straight.
Chvatal showed that if the gallery has n walls (that is, if the floor-plan is an
n-gon as in Figure 9), then, whatever its zig-zag shape, the minimum
number of guards needed is never more than [n/3], the integer part of n/3.
His proof is given in Mathematical Gems II. Now I would like to present a new
proof of this theorem. Chvatal's analysis is certainly first-class, but the
following proof by Steve Fisk of Bowdoin College is simply inspired.
First of all, the gallery is divided into triangles by drawing non-
intersecting diagonals in its interior (Figure 10). Next, the vertices of the
gallery are colored so that two vertices that are joined by an edge in the
figure are given different colors. It is an easy application of mathematical
.a281~
FIGURE 9
induction to prove that 3 colors suffice for this task. For n = 3, the entire con-
figuration consists merely of a single triangle, and 3 colors obviously are
sufficient. Let us suppose that 3 colors suffice for a triangulated n-sided
gallery (n ~ 3) and that we have at hand a gallery G with (n + 1) sides.
Clearly there is no shortage of diagonals (AB) which cut from G a
single triangle (ABC). By the induction hypothesis, the triangulated n-sided
configuration G', obtained from G by cutting off !!ABC, can be colored
with 3 colors so that adjacent vertices have different colors. Since only 2 of
the 3 colors are used at A and B, the third color can be used at C to extend
this into a successful3-coloring of the triangulated gallery G itself. Thus, by
induction, 3 colors suffice in all cases.
Suppose, therefore, that the vertices of our gallery are properly 3-
colored with the colors a, b, and c. Since there are only n vertices altogether,
all three of the colors cannot each be used more than n/3 times. The color
which is used least often, (say b), must occur only m times, where m ~ n/3.
Since m is an integer, this is the same as saying m ~ [n/3].
Since no two vertices of any triangle in the figure can be the same
color, each triangle must have a vertex of each color a, b, and c. Clearly a
watchman at the vertex colored b can survey the whole triangle from that
FIGURE 10
.8282fS.
vertex, and m watchmen at the m vertices colored b can observe all the
triangles, that is, the entire gallery.
The Track
Consider now the following engaging problem which, as far as I can
determine, is due to the brilliant young Hungarian mathematician Laszlo
Lovasz.
Around a track Tare n arbitrarily spaced depots Xl,X2,""X n' each
containing a quantity of gasoline (Figure 11). The total amount of gas is
exactly enough to take a car around T once. Prove that, no matter how the
gas is distributed over the depots, there is some depot at which an empty car
can gas up, proceed around T picking up the gas at the other depots as they
are encountered, and get all the way around the track.
x. x.
FIGURE 11
The following solution was contributed by Dean Hoffinan (Auburn

University), who learned of it from Lovasz himself.
Suppose a practice run is made, starting from any point on the track,
with a car that has lots of extra gas. Since the amount of gas required to
make the run is exactly the quantity picked up along the way, the car will
finish the lap with the same amount of gas it began with. Now, during the
run we keep an eye on the gas gauge, and we note that when we arrived at
depot Xi' the reading was lowest for the whole trip, say d gallons. During
the run, then, we were always carrying around an extra d gallons of gas
that was never needed. Clearly the same would have happened if this d
gallons of gas had been left at home. Without this unnecessary gas, we
would have come into depot x;just as the tank ran dry, a condition that we
achieve by starting the real trip at depot Xi' Beginning at Xi' the gas gauge
will never be pressed to go below zero because it never went below d gallons
during the practice run.
.8283&'
Lattice Cubes
The points (x,y,z) in 3-dimensional space which have all three coordinates
integers are called lattice points. They occur at the vertices of the unit cubes
into which space is divided by the planes x = a, y = h, z = c, where
a,b,c are integers. Obviously there is any number of cubes, having their (8)
vertices at lattice points, which sit in space with their faces parallel to the
coordinate planes. The length of the edge of such a cube is clearly an
integer.
There are also lots of cubes, having their vertices at lattice points,
which do not occur in "standard" position but sit obliquely (relative to the
grain of the lattice). Few of us, however, possess the intuition not to be
surprised at the all-inclusive result:
the length s of the edge of a cube C, having its vertices at lattice points,
is always an integer.
Since it doesn't matter which point of the lattice we choose to take as
the origin 0 of coordinates, let's use one of the vertices of C. Suppose the
three edges from 0 go to the vertices V1 (Xl'Yl,Zl), V2 (X2'Y2,Z2),
V 3{X3,Y3,Z3). Having edge s, C has volume V = s3. A well known formula
from freshman mathematics gives the volume of a cube in terms of the
edges at a vertex. For the appropriate choice of sign, we have
Xl Yl Zl
V =± x2 Y2 Z2
X3 Y3 Z3
Because all of the entries here are integers, the volume V must also be an
integer, that is,
S3 is an integer.
Now the length of the edge OV1 is
s= 1OV11 = Jxi + yi + zi = Jan integer.

Therefore we see s2 is also an integer. Consequently, s itself must be
rational:
S3
S = 2 is rational.
s
But, as we just observed, s = Jan integer. And if the square root of an
integer is not irrational, then it is not just rational, but integral. Thus s is an
integer.
This result can be generalized in a straightforward manner to cubes
in n-dimensional space, where n is odd.
~284~
Probing the Rotating
Table
-~-
William T. Laaser
STANFORD UNIVERSITY
Lyle Ramshaw
STANFORD UNIVERSITY
ABSTRACT
Consider the following generalization of the Problem of the

Rotating Table: the shape of the table is a regular n-gon; at each
of the n corners of the table is a well containing a glass that is
either upright or inverted; and the player has k hands. For each
k and n, the question then arises whether or not there exists a
procedure that is guaranteed to ring the bell after a finite
number of moves. In this paper, we will see that such a
procedure exists if and only if the parameters k and n satisfy the
inequality k ~ (I - lip )n, where p is the largest prime factor of
n.
~285~
Section 1
Introduction
In his column in the February, 1979 issue of Scientific American, Martin
Gardner presented the Problem of the Rotating Table as follows:
I open with a delightful new combinatorial problem of unknown
origin. It was passed on to me by Robert Tappay of Toronto, who
believes it comes from the U.S.S.R.
Imagine a square table that rotates about its center. At each
corner is a deep well, and at the bottom of each well is a drinking glass
that is either upright or inverted. You cannot see into the wells, but
you can reach into them and feel whether a glass is turned up or down.
A move is defined as follows: Spin the table, and when it stops,
reach each hand into a different well. You may adjust the orientation
of the glasses any way you like, that is, you may leave them as they are
or turn one glass or both.
Now, spin the table again and repeat the same procedure for your
second move. When the table stops spinning, there is no way to
distinguish its corners, and so you have only two choices: you may
reach into any diagonal pair of wells or into any adjacent pair. The
object is to get all four glasses turned in the same way, either all up or
all down. When this task is accomplished, a bell rings.
At the start the glasses in the four wells are turned up or down at
random. If they all happen to be turned in the same direction at this
point, the bell will ring at once and the task will have been
accomplished before any moves were made. Therefore it should be
assumed that at the start the glasses are not all turned the same way.
Is there a procedure guaranteed to make the bell ring in a finite
number of moves? Many people, after thinking briefly about this
problem, conclude that there is no such procedure. It is a question of
probability, they reason. With bad luck one might continue to make
moves indefinitely. That is not the case, however. After no more than n
correct moves one can be certain of ringing the bell. What is the
minimum value of n, and what procedure is sure to make the bell ring
in n or fewer moves?
Consider a table with only two corners and hence only two wells.
In this case one move obviously suffices to make the bell ring. If there
are three wells (at the corners of a triangular table), the following two
moves suffice.
I Reach into any pair of wells. Ifboth glasses are turned the same
way, invert both of them, and the bell will ring. If they are turned in
different directions, invert the glass that is facing down. If the bell does
not ring:
2 Spin the table and reach into any pair of wells. If both glasses
are turned up, invert both, and the bell will ring. If they are turned in
different directions, invert the glass turned down, and the bell will
ring.
8286&.
PROBING THE ROTA TING TABLE
Although the problem can be solved in a finite number of moves

when there are four wells and four glasses, it turns out that if there are
five or more glasses (at the corners of tables with five or more sides),
there are no procedures guaranteed to complete the task in n moves.
Next month I shall give one solution for the problem with four glasses
and discuss some generalizations of the problem developed by Ronald
L. Graham and Persi Diaconis.
There are several fine points about the rules which deserve clarifi-
cation, concerning the types of procedures that the player can legally
follow. First, the player must announce exactly which wells she is planning
to probe before she actually probes either of them. It is not legal for her to
probe one well, and then to decide which other well to probe on the basis of
what she finds in the first one. Secondly, once she has probed the wells and
turned the glasses in some way, the bell will not ring to indicate success
until after she has withdrawn her hands from the wells. That is, it is not
legal for the player to tryout several orientations of the glasses with her
hands still in the wells, hoping to ring the bell on at least one of them.
The solution for the problem with four wells and four glasses given in
Martin Gardner's column of March, 1979 makes the bell ring in no more
than five moves. If the reader is not already familiar with the Problem of
the Rotating Table, she is encouraged to reconstruct this solution.
The March, 1979 column also mentions two generalizations of the
Rotating Table Problem, suggested by Ronald L. Graham and Persi
Diaconis. They first suggested allowing the player more than two hands.
For each k and n, we can imagine a k-handed player and a table in the
shape of a regular n-gon, with n corners, wells, and glasses; the problem
then is to discover a finite procedure guaranteed to ring the bell, or to
prove that no such procedure exists. Graham and Diaconis also suggested
generalizing the problem by replacing the glasses with objects that have
more than two positions. And still other interesting generalizations are
possible; for example, Scott Kim considered replacing the square table by a
table with a richer group of symmetries, such as a cube.
This paper is devoted to the first of these generalizations. In
particular, we defineJ(n) for n ~ 2 to be the smallest integer k such that a
finite procedure guaranteed to ring the bell exists when a k-handed player
faces an n-cornered polygonal table, whose glasses have precisely two
states. Our goal is to determine the values of the function f
The rules that specify the Rotating Table Problem are sufficiently
subtle that it is hard to be sure that they have been completely specified.
Recall that we have already had to pause to clarify several fine points.
Thus, it behooves us to state the problem more formally, in the hope that a
more formal description will also be more precise. We will rephrase the
problem as an asymmetric two-person game between the Player and the
Table, involving the manipulation of circular strings.
~287~
Whenever we talk about polygons in this paper, we will be referring to

regular polygons that are free to rotate about their centers, but that cannot
be turned upside down. With this understanding, we define a necklace rif
length n to be a labelling of the vertices of an n-gon. A necklace oflength n is
like a string of length n, except that all circular shifts of the string
correspond to the same necklace. We will specify particular necklaces
either by drawing a picture, or by writing a string in angle brackets, where
the string represents the labels of the necklace in clockwise order from an
arbitrary starting point. For example, the necklace oflength six in Figure 1
could be written (010011), or (100110), or in four other ways; but
< 10 1100) represents a different necklace.
o
o
FIGURE 1
A necklace of length six.
At any point in time, the current state of the glasses in their wells
corresponds to a necklace oflength n over the alphabet {u, d}, where we will
use u and d for "up" and "down" to denote upright and inverted glasses
respectively. The first move of the Game of the Rotating Table is a move
by the Table; in particular, the Table determines the initial state of the
glasses by choosing a necklace over {u, d}. Call this state S 1.
Now it is time for the Player's first move, in which she must announce
the circular arrangement of her probing hands around the table. We will
call this arrangement a probe pattern. Formally, a probe pattern is
represented as a necklace over the alphabet {h,g}, where h for "hand"
denotes a position that the Player desires to probe, and g for "gap" denotes
an unprobed position. Naturally, the Player must choose a probe pattern
that has no more than k occurrences of the symbol h, since the Player has
only k hands. Call this first probe pattern Pl.
Next, it is the Table's move again. The Table gets to choose one of the
n possible ways in which the necklaces Sl and Pi can be superimposed.
This choice models the spinning of the Rotating Table. After the Table has
made this choice, the Player is told for each h in the probe pattern Pi
.8288~
PROBING THE ROTATING TABLE
whether the corresponding symbol of the state Sl is a u or a d. The Player

may then order that any or all of the u's in probed positions be changed to
d's, or vice versa. The result of these alterations defines the next state of the
table, a necklace S2 over the alphabet {u,d}. We will call a state necklace
monotone if it consists entirely of u's or entirely of d's. If the state S2 is
monotone, a bell now rings and the Player has won the game in one probe.
At this point, the game enters a loop. Suppose that the non-monotone
state Sj has just been determined. The Player then decides upon a probe
pattern Pj; the Table determines how Sj and Pj are to correspond; and
the Player is informed of the contents of Sj in those positions corresponding
to h's in Pj. The Player then decides upon new contents for these positions,
and her decisions determine the next state Sj + 1. If Sj + 1 is monotone, the
bell rings and the Player is deemed to have won in i probes. The Table's
goal is to prevent the Player from winning; we will say that the Table wins
if and only if the game continues forever. Recall that we have definedJ(n)
to be the smallest integer k such that k hands allow the Player to win; a
winning strategy for the Player exists if k 2.J(n), while a winning strategy
for the Table exists if k <J(n).
The remainder of this paper is devoted to the proof of the following
result:
THEOREM 1 For any integer n 2. 2, let J(n) denote the

smallest integer with the property that J(n)
hands are enough to enable the Player to win
the Game of the Rotating Table. Let p denote
the largest prime factor of n. Then, the value of
J(n) is given by the formula
Previous work established some special cases of this result. As

announced in Martin Gardner's column of March, 1979, Ronald L.
Graham and Persi Diaconis showed that J(n) = n - 1 if n is a prime
number, whileJ(n) :::;; n - 2 ifn is composite. For all a 2. I and b 2. 2, Scott
Kim showed the lower bound
where fxl denotes x rounded up to the nearest integer. And James Boyce
first conjectured the formula for J(n) given in Theorem 1.
By the way, Theorem I has the curious corollary that, except for
n = 2, the minimum number of hands needed is always an even number.
.8289~
Section 2
The Lower Bound
In this section, we will demonstrate that the formula given in Theorem I
constitutes a lower bound on the true value ofj(n). In particular, we will
prove the following Lemma.
LEMMA 1 Let n ~ 2 be an integer, and let p be its largest

prime factor. Then, if the Player is restricted to
fewer than (l-lfp)n hands, there is a winning
strategy for the Table in the Game of the Rotating
Table.
PROOF We begin by noting that, if n factors in the form n

= 1m, an n-gon can be viewed as I distinct and interleaved m-gons.
Because 6 = 2·3, for example, we can view a hexagon as a Star of
David, as shown in Figure 2.
Let p be the largest prime factor of the integer n ~ 2, and determine I

by the relation n = Ip. In fact, we could take p to be any prime factor of n in
this argument, but the largest prime factor gives us the strongest result. As
noted above, the states and probe patterns of the Rotating Table Game
can be viewed as composed of I interleaved p-gons. Our goal is to construct
a winning strategy for the Table, assuming that the Player has fewer than
(1 - IJp)n hands. The basic idea is to guarantee that, for all i, at least one
of the p-gons that constitute the state Sj will be nonmonotone, that is, will
contain both up and down glasses.
This condition is easy to establish initially, since the Table may
choose the state S1 arbitrarily. But we must show that the Table can
maintain this condition as the Game progresses. Suppose that the state Sj
contains a nonmonotone p-gon; choose one, and call it Sr. Let Pj denote
the probe pattern which the Player decides to employ on the ith move.
Since the Player is restricted to fewer than (I - Ifp)n = 1(P - 1) hands, at
least one of the I different p-gons of P j must contain at least two gaps g.
FIGURE 2
Viewing a hexagon as two interleaved triangles .
.a290~
Choose such a p-gon, and call it Pi. Also choose two gaps which appear in
Pi*' and suppose that they appear at distance} around Pi, where we
measure distance so that adjacent vertices are at distance 1. Now, consider
touring Si by walking around it in steps of} vertices each. Since p is prime,
Sr
this tour will visit every vertex of before returning to its starting place,
no matter what the value of}. During this tour, we will certainly come
across both u's and d's, since Sf was chosen to be nonmonotone. Hence, we
will come across some d immediately after a u. This implies that Si contains
a u and a d which are at distance j.
We are now ready to advise the Table how to proceed. It should
chose to align the necklaces Si and Pi so that the p-gons Sf and Pf are
superimposed, and furthermore, so that the u and d at distance} in Sf
correspond to the two g's at distance} in Pf. Since the Player cannot alter
the states of glasses which correspond to gaps, the resulting state Si+ 1 will
contain at least one nonmonotone p-gon. By using this technique
repeatedly, the Table can guarantee for all i that the state Si will contain a
nonmonotone p-gon. This constitutes a winning strategy for the Table.
Section 3
Perfect Powers
The lower bound result in Lemma I was the easy part. To complete the
proof of Theorem I, we are left with the more exacting job of constructing
winning strategies for the Player that use the minimal number of hands.
We begin this task with an insight that will allow us to use previously
constructed strategies for smaller tables as subroutines when constructing
strategies for larger tables.
We need a definition. Suppose that the size n of the table factors in the
form n = lm. Recall that we can then view an n-gon as l distinct,
interleaved m-gons. Suppose that some necklace T of length n has the
property that all of the m-gons of which it is composed are monotone. Note
that the necklace T must then consist of m copies of some string of length l,
>
strung together end to end; that is, T must have the form (X'" for some
string X. We will distinguish such necklaces by calling them perfect mth
powers. For example, an even-length necklace is called a perfect square
when the labels at the two ends of every diagonal are the same, implying
that the necklace has the form (XX> for some string X. Saying that a
necklace oflength n is a perfect nth power is another way of saying that it is
monotone.
Suppose that we are constructing a strategy for the Player to use
against a size n Table where n factors non trivially as n = lm, and suppose
that we get to a point where the current state S of the table is a perfect mth
power. We can finish the strategy in an easy way as long as we have at least
mf(l) hands. The idea is to mimic an optimal Player strategy for a size l
~291&.
table, with everything replicated m times; one might call this process
raising a strategy to a power. Take the first probe pattern of the smaller
strategy, and build a probe pattern for the larger strategy by stringing
together m copies. Since the current state S of the table is a perfect mth
power, each of our Player's m groups of f(l) hands will feel the same
sequence of ups and downs. We then instruct our Player to perform with
each group of hands whatever adjustments of the table state the smaller
strategy recommends. No matter what those adjustments are, the next state
of the larger table will also be a perfect mth power, and thus, we will be
able to continue mimicking the smaller strategy. This argument has
demonstrated the truth of the following Lemma.
LEMMA 2 Suppose that the Rotating Table Game is being

played with a table of size n, and that the current
state S of the table is a perfect mth power for some
integer m ;::: 2 that properly divides n. Then, there
exists a strategy that the Player may now begin to
employ that is guaranteed to ring the bell in a
finite number of probes, and that uses only
mf(n/m) hands.
Before we can apply Lemma 2, we first need to force the table into a
state that is a perfect power. The next Lemma claims that there exist
strategies that handle this part of the job without using too many hands.
LEMMA 3 If P denotes the largest prime factor of an integer

n ;::: 2, then there exists a strategy for the Player in
the Rotating Table Game of size n with the
following properties: it uses at most (1 - l/p)n
hands, and, in a finite number of moves, it either
rings the bell or forces the table into some state
that is a perfect pth power.
We will postpone the proof of Lemma 3 until later. Instead, let us now
show that Lemmas 1, 2, and 3 together allow us to complete the proof of
Theorem 1.
PROOF OF THEOREM 1 Theorem 1 claims that f(n) =

(1 - 1/p )n, where p is the largest prime factor of n. From
Lemma 1, we know thatf(n) ;::: (1 - l/p)n. It remains to show
thatJ(n) :::;; (1 - l/p)n. Our proof will be by induction on n;
thus, we assume that the Theorem holds true for all tables
smaller than n in size.
Our task is to demonstrate the existence of a winning strategy for the

Player that uses only (1 - l/p)n hands. We begin by using the strategy
.8292~
from Lemma 3; this either rings the bell, or forces the table into a state S
that is a perfect pth power. If n is prime, then p equals n, and any state that
is a perfect pth power is also monotone. In this case, therefore, the strategy
from Lemma 3 always rings the bell, and we are done.
On the other hand, if n is not prime, then p is a proper divisor of n, and
hence Lemma 2 applies. The strategy from Lemma 2 will take the Player
from the state S to a state that rings the bell. We only need to check that the
Player will have enough hands to implement the Lemma 2 strategy. This
boils down to checking that the inequality
1 (1 - ~)n "2:.pf(nlp)
holds. To see this, we will invoke our inductive hypothesis for the table size
nip. Let q be the largest prime factor of nip, and hence the second largest
prime factor of n. From the inductive hypothesis, we can conclude that
Since q ~ p, Inequality 2 implies Inequality 1, and thus our Player will

have enough hands to use the Lemma 2 strategy.
The rest of this paper is devoted to the proof of Lemma 3. It turns out
that this proof splits into two somewhat different cases. If the largest prime
factor of n is at least 3, then a relatively simple strategy called the up-down
strategy will work; this is discussed in Section 4. In Section 5, we will turn
to the trickier up-flip strategy that is needed when n is a power of 2.
Section 4
The Up-Down Strategy
In this Section, we will give the proof of Lemma 3 for all table sizes n other
than powers of 2. To develop some intuition for the general case, we will
start by handling the particular case of n = 12. The largest prime factor of
12 is 3. Therefore, Lemma 3 for the case n = 12 makes the following claim:
there exists a strategy that an 8-handed Player could employ against a 12-
cornered Table that either rings the bell or forces the table into some state
that is a perfect cube.
When n = 12, the table is a dodecagon. To understand the strategy
that we are about to construct, that dodecagon should be viewed as
composed of two disjoint, interleaved hexagons, and each of the hexagons
should be viewed as two disjoint, interleaved triangles. Figure 3a shows a
dodecagon with its vertices numbered like those of an ordinary, 12-hour
clock face; in somewhat different language, Figure 3a could be described as
a picture of the clock necklace
c = «(1)(2)(3)(4)(5)(6)(7)(8)(9)(10)(11)(12).
~293~
12
12:
'1'1 I" I'I '7 II II

'I" I
9 3
2 610 I" I 48 12 II
6
(a) (b)
FIGURE 3
(a) The clock necklace, and (b) its structural diagram.
In Figure 3b, the clock necklace C has been redrawn to reflect the
decomposition of the dodecagon into interleaved hexagons and triangles.
Each rectangle in Figure 3b represents the regular subpolygon of the clock
face whose vertices it encloses; at the upper left corner of each rectangle is a
label that gives the number of enclosed vertices. We will call such pictures
structural diagrams. We can draw a similar structural diagram for any
necklace oflength 12, in particular, for the states and probe patterns that
arise as we construct our strategy.
Our strategy will use three different probe patterns, called PI' P 2' and
P 3 . Figure 4 depicts the structural diagrams that these probe patterns
should have. Pattern Pluses six hands to probe precisely one of the two
hexagons; pattern P 2 uses six hands also, but probes precisely one of the
triangles from each of the two hexagons; and P 3 uses eight hands to probe
precisely two of the vertices of each of the four triangles. Furthermore,
these insights about what vertices each pattern will probe are valid
regardless of how the table rotates. We don't know which of the two
hexagons PI will probe, but it will definitely probe one of them;' the same is
true for P 2 and P 3 • It is this invariance under rotation that makes
structural diagrams useful.
No matter what structural diagram we draw, there will always be at
least one necklace that has that structural diagram, but there may be more
than one. In the cases of PI and P 2 there is exactly one; we must choose PI
= (ghghghghghgh) and P2 = (gghhgghhgghh) if they are to have the
structural diagrams in Figure 4. But we have some choice in the case of P 3 •
The two different necklaces (gggghhhhhhhh) and (ghhghhghhghh) both have
the structural diagram that we require for P3 ; and in fact, there are six
other necklaces that we won't bother to mention that also have that
structural diagram. It turns out that we can choose the P/s to be any probe
.8294~
PROBING THE ROT A TING TABLE
12: 12:
"I 'I g g g 1 3: I g g g
II '1'1 g g g 1 3: I h h h
II
'1"1 h h h 1 3: I h h h
II '1'1 g g g 1 3:1 h h h
II
P, P,
12:
'1'1 g h h
1'81
'1'1 g h h 1 3: I g h h
II
P,
FIGURE 4
Structural diagrams of the P;'s.
patterns that have the correct structural diagrams; precisely which ones we
choose doesn't matter.
Before exploring what these probe patterns can do, we pause to
extend our notation for states. Recall that the exact state of the table is
represented by a necklace of length n over the alphabet {u, d}. We can also
use necklaces to represent partial information about the state of the table,
by allowing the symbol e to appear as well: the symbol e for "either" will
denote a position that contains either an up glass or a down glass. For
example, the initial state of the table Sl has the formula Sl = (en), since
the Table gets to choose this initial state arbitrarily.
For the first three probes of our strategy, we instruct the Player to
probe with the patterns PI' P 2 , and P 3 , in that order, turning up every
glass that she feels, regardless of its initial state. The effects of these probes
can be seen by examining Figure 5, which depicts the structural diagrams
of the states Sl through S4. The probe with pattern PI forces one of the
hexagons to be monotone up in state S2. The six hands in pattern P2 probe
one triangle from each hexagon, and hence the second probe forces a third
of the four triangles to be monotone up in state S3. Pattern P 3 probes
precisely two of the three vertices of every triangle, and hence the third
probe turns up two of the three remaining e's. Thus, after the first three
probes, the state S4 of the table will be S4 = (eu l l ).
It might happen that, after one of these three probes, the remaining
e's all happen to represent glasses that are actually up. If so, the bell will
~295~
II
12: 12:
6: 1 3: I e e e I 3:1 e e e
II
6:1 3: I e e e II 3: e e e
6: 1 3: I e e e I 3: I e e e
II 6:1 3: I u u u I 3:1 u u u
II
8. 82
12: 12:
6:1 3:1 e e e II 3: u u u
II 6:1 3: I e u u I 3: I u u u
II
6:1 3:1 u u u II 3: u u u
II 6:1 3: I u u u I 3:1 u u U
II
83 84
FIGURE 5
Structural diagrams of Sl through S4'
ring, and the Player will have won. We might as well ignore this case,
therefore, and concentrate instead on the case where the final remaining e
represents a down glass, so that the state S4 is actually S4 = (dull).
We will refer to the portion of the strategy that we have discussed so far
as theftrst pass. The net effect of the three probes in the first pass is to force the
table into a state that is only one position away from being monotone up. We
will refer to this lone down glass as the exceptional glass, and we will also use
the term exceptional to refer to the triangle that contains it.
Where are we heading, anyway? We want to force the table into a
state that is a perfect cube. We can do that in, at most, three more probes
which we will call the probes of the second pass. These probes will also use
the patterns Pi' P 2 , and P 3 , in that order. First, let us consider probing S4
with Pi; recall that Pi probes one of the two hexagons. If the table spins so
that the Player feels the exceptional glass, the Player can merely turn that
glass up, and the bell will ring. Hence, we might as well assume that the
Player's hands probe the hexagon that is currently monotone up. We
instruct the Player to turn this entire hexagon down; the structural
diagram of the resulting state Ss is shown in Figure 6.
On to the second probe of the second pass. Recall that P2 probes one
triangle from each hexagon. If one of the triangles that the Player feels is
non monotone, it must be the exceptional triangle. In this case, we instruct
the Player to turn the exceptional glass up, which makes every triangle in
the resulting state monotone; this is equivalent to saying that the resulting
state is a perfect cube. Since that is our goal, we can tell the Player to omit
~296€>
12: 12:
'1'1 d u u 1 3:1 u u u
II '1'1 d u u 1 3: 1 d d d
II
'1'1 d d d 1 3:1 d d d
II 'I 3:1 d d d 1 3:1 d d d
II
s, S.
FIGURE 6
Structural diagrams of Ss and S6.
the rest of the second pass. On the other hand, it might happen that the
Player feels only monotone triangles. One of them will be monotone down,
and the other monotone up; we instruct the Player to leave the down
triangle alone, and to turn the up triangle down. This results in the state S6
whose structural diagram is shown in Figure 6.
We have now come to the third and final probe of the ~econd pass, a
probe using pattern P 3. Recall that P3 probes two of the vertices of each of
the four triangles. Currently, the exceptional triangle is in state (duu),
while all of the nonexceptional triangles are monotone down. Note that
the Player cannot get confused about which triangle is which. If the two
vertices that she feels from a triangle are both down, then that triangle
must be monotone down; we instruct the Player to leave those triangles
alone. The Player also has two hands in the exceptional triangle. Those
two hands feel either both of the up glasses, or one up glass and one down
glass. In the former case, we turn both up glasses down; in the latter, we
turn the down glass up. The net effect is to guarantee that the exceptional
triangle becomes monotone in the next state S7. But all of the other
triangles in S7 will also be monotone, in fact, monotone down. Thus, the
Player has forced the table into a state that is a perfect cube, as demanded
by Lemma 3.
We will call the strategy that we have just constructed the up-down
strategy, since the first pass turns every glass but the exceptional one up,
while the second pass turns every triangle but the exceptional one down.
Of course, we have to prove Lemma 3 for all n, not just for n = 12. It turns
out that we can handle all values of n other than powers of 2 by the up-
down strategy. The only confusing thing is the notation. On the other
hand, the up-down strategy fails for powers of 2, so we postpone our
discussion of those cases until Section 5.
PROOF OF LEMMA 3 WHEN n IS NOT A POWER OF 2 Let n

denote the size of the table, and let p denote the largest prime
factor of n; since n is not a power of 2, we have p 2 3. We are to
construct a strategy for the Player with the following properties:
~297~
it either rings the bell or forces the table into a state S that is a
perfect pth power, and it uses no more than (1 - l/p)n hands.
We begin by setting up some notation. Suppose that n has) prime

factors, of which P is the largest. If we put the prime factors of n into
increasing order, we can write
n = PIP2 ... Pi for }:2: 1, Pl 5, P2 5, ... 5, Pi = p.
From this factorization of n, we define some ancillary notations, Ii and ri;
the value of Ii is the product of the first up through but not including the ith
prime factor of n, while ri is the product of the ith through and including
the}th. More formally, for 1 5, i 5,}+ 1, we define the l's and r's by
1 $k<i i$k$i
Note that, for all i, we have l;Ti = n.

In order to understand the general case of the up-down strategy, one
has to view the n-gon in a fairly complicated way. If we start at the
outside, we have the n-gon, or equivalently, the rrgon, composed of Pl
interleaved r 2 -gons; each of the r 2 -gons is composed of P2 interleaved rr
gons; and in general, each r;-gon is composed of Pi interleaved ri + l-gons.
Finally, each rrgon is composed of Pi vertices; that is to say, ri equals Pi'
The Prgons are at the bottom of the hierarchy, and will play the role that
the triangles played in the n = 12 case.
The previous paragraph demonstrates the meaning of the r/s: they
give the sizes of the various polygons into which we are breaking up the n-
gon. The l/s play the complimentary role of counting how many r;-gons
there are. In particular, for each i, the n-gon contains precisely Ii distinct r;-
gons.
We can now get a reasonable grip on what a structural diagram looks
like in the general case. The rectangles are nested} levels deep, and, at each
level of the nesting, the Ii distinct r;-gons are each broken up into Pi distinct
ri + l-gons. Figure 7 is an attempt to show, in the general case, what the ith
level of nesting of a structural diagram looks like.
The up-down strategy in the general case will use} different probe
patterns, which we will call P l ,P2 , ... ,Pi . As in the case ofn = 12, it won't
matter which probe patterns the P/s are, as long as they have the correct
structural diagrams. We can describe the structural diagram that Pi is
required to have in terms of the ith level of nesting, shown in Figure 7. In
particular, the pattern Pi should probe all but one of the r i + rgons that
constitute each r;-gon. We will abbreviate this condition by saying that
such a pattern resolves the ith level if nesting.
We now choose the patterns Pi arbitrarily, subject to the condition
that Pi must resolve the ith level of nesting. The next thing to check is that
~298~
'j 'j 'j
'j+ I:II...__ -J 'j + 1 : LI_ _---I 'j + I: 1...1_ _---1
'j+I:I. . __. . . . Pj
'j :1
+ 1 L _ _---I
Pj
'j + 1:1...._ _--'
Pj
'j+ 1:1...1_ _........ 'j + I: 1

'j + I: 1
....__--'
~-------------------------~v~----------------------~
lj
FIGURE 7
The ith level of nesting of a structural diagram in the general case.
our Player will have enough hands to use the probe patterns Pi that we
have chosen. To see this, note that any pattern that resolves the ith level of
nesting uses exactly
hands. Since our Player is allowed (I - I /p)n hands where p = pi ~ Pi' she
will have enough hands to use the patterns Pi'
We now advise our Player to execute the first pass, that is, to probe
with each of the patterns from PI to Pi in order, turning up every glass that
she feels. The pattern PI probes the entire table, an r I -gon, except for a
single r2-gon; then, pattern P2 probes all of each r2-gon except for one of
the constituent T3 -gons. This process continues, with pattern Pi further
restricting the portion of the table that can still contain down glasses to a
single ri+ I-gon. Therefore, after probing with Pi' the table can include, at
most, one down glass. We might as well assume that there is exactly one
down glass, or else our Player has already won. We will use the term
exceptional to refer to thi.; down glass, and to the prgon that contains it. The
first pass has effected the "up" portion of the up-down strategy: it has
turned all but the exceptional glass up.
I t follows from our choice of the patterns Pi that Pi probes all of the
vertices of any rj + 1-gon of which it probes any of the vertices. This implies
that P j must in fact probe either all of or none of each rk-gon for any k > i.
This fact helps to explain our choice of terminology; in order to examine
only part of an rk-gon; that is, to resolve the kth level of nesting, one needs
to use a pattern P j where i ~ k. In particular, we note that all of the probe
patterns Pi except for the last one, Pi' must probe either all of or none of each
rrgon-that is to say-each prgon. We could rephrase this observation by
saying that the patterns Pi for i <j are perfect Pith powers.
~299~
On to the second pass. We advise our Player to run through the

patterns P 1 ,P2 , ••• ,Pi again, in order, with somewhat more complex
instructions. Think of the table as composed of interleaved Prgons. When
we start the second pass, all of the Prgons are monotone up except for the
exceptional one, which includes a single down glass. And, until the last
probe of the pass, we will be probing all of any Prgon that we probe any
part of. We instruct our Player as follows, for all but the last probe. If you
feel the exceptional Prgon, turn the exceptional glass from down to up.
This puts the table into a state that is a perfect Pith power, so you have
satisfied the demands of Lemma 3; omit the rest of the second pass. If you
feel a Pi-gon that is monotone up, it is nonexceptional, and it hasn't yet
been reversed; turn all of the glasses in such a Prgon down. If you feel aPr
gon that is monotone down, it has already been reversed; leave it alone.
We can show that only one Prgon can escape being probed during
the first j - I probes of the second pass. The argument is essentially the
same as the argument that we used to analyze the first pass. The probe with
PI will hit all of the Prgons except for those lying in a particular r2 -gon; the
probe with P 2 will hit all of the rest except for those lying in a particular r 3-
gon; and so on. By the aftermath of the probe with Pi-I' all of the Prgons
will have been probed except for those lying in a particular rrgon, and ri
= Pi. Furthermore, we can assume that the exceptional Prgon escapes
being probed; if our Player was lucky enough to probe the exceptional Pr
gon, she was instructed to correct the exception and quit. Therefore, the
first j - I probes in the second pass probe all of the non-exceptional Pr
gons, and turn them all down. By doing so, they accomplish the "down"
phase of the up-down strategy.
Going into the last probe of the second pass, then, the table contains
one exceptional Prgon which has Pi - I up glasses and one down glass,
while all of the other Prgons are monotone down. For this last probe, we
instruct our Player to use pattern Pi' which probes Pi - I of the vertices of
each Prgon. And here we reach a subtle point. So far, we have not needed
to use the assumption that Pi = P ~ 3, but now that assumption becomes
crucial. Note that, because Pi ~ 3, the Player will be feeling at least two
glasses from each Prgon, and hence she must feel at least one of the up
glasses of the exceptional Prgon. This allows her to figure out which of her
groups of Pi - I hands is probing the exceptional Prgon. We have to
outlaw the case ofPi = 2 at this point precisely because this argument fails.
IfPi = 2, the exceptionalPrgon is a diagonal in the state (du); if the Player
probes just one end of this diagonal, she might probe the down end, and
then she would not be able to distinguish that down glass from all the other
down glasses.
But for now, we are assuming that Pi ~ 3, so the Player will be able to
locate the exceptional Prgon. We instruct our Player to leave all of the
non-exceptional Prgons alone. For the exceptional Prgon, there are two
cases. Either she will feel the lone down glass, in which case she should turn
~300~
it up; or she will feel all Pj - I of the up glasses, in which case she should
turn them all down. In either case, the exceptional Prgon will become
monotone in the next state, and hence the demands of Lemma 3 will be
satisfied. This completes the proof of Lemma 3 for every integer n other
than powers of 2.
Section 5
The Up-Flip Strategy
Our remaining task is to construct the strategy implied by Lemma 3 when
the size of the table is a power of 2. As we noted above, the up-down
strategy comes to grief in this case, because the Player might not be able to
locate the exceptional diagonal during the last probe of the second pass. In
this Section, we will construct a strategy for the power of 2 case called the
up-flip strategy, which is similar to the up-down strategy but involves a
trickier second pass. Flipping a glass means turning it up if it is currently
down, or down- if it is currently up. During the second pass of the up-flip
strategy, we will instruct our Player on each probe to flip every non-
exceptional diagonal that she feels rather than to turn them all down. The
repeated flips put the table into states that have lots of up and down glasses
arranged in interesting ways, and it turns out that a careful Player can take
advantage of the resulting structure. To prepare for the subtleties of the
up-flip strategy, we need to study some facts about periods in circular
strings.
For any integer k and necklace T, we will say that Tis k-periodic, or
that k is a period of T, if and only if each symbol in T is the same as the
symbol k steps to its right around T. Putting it another way, a necklace is k-
periodic when it looks exactly the same after being rotated k positions to the
right. Every necklace is O-periodic; if a necklace is I-periodic, then it must
also be ( -I)-periodic; and, if a necklace is both I-periodic and m-periodic,
it must also be (l + m)-periodic. In fancier language, the set of all periods
of a necklace T forms a subgroup of the integers uilder addition.
Since a necklace of length n is always n-periodic, every necklace has
some positive period. We will refer to the smallest positive period of a
necklace as its basic period. We then have the following result. -
LEMMA 4 The periods of any necklace are exactly the

multiples of its basic period.
PROOF Certainly any multiple of the basic period will be a

period. To see the other implication, let k be the basic period of
a necklace, let m be some other period, and let q denote the
greatest common divisor of k and m. Recall that q can be written
in the form q = ak + bm for some integers a and h; hence, q must
~301~
also be a period of the necklace. Since k, the basic period, is the

smallest positive period of the necklace, we must have q = k.
And since q divides m, we conclude that m must be a multiple of
the basic period. Readers with algebraic background will
recognize this as the proof that the integers form a principal
ideal domain.
The notion of a period is pretty standard; in fact, it is closely

connected to the idea of perfect powers. Note that, if a necklace oflength n
has basic period k, then k divides n, and the necklace is a perfect (n/k)-th
power. But we need another concept as well, which is more exotic. This
other notion only makes sense for necklaces over an alphabet with exactly
two symbols; in our case, these two symbols will be u and d. We will call
each of these symbols the opposite or complement of the other. Then, we will
say that a necklace is k-alternating if and only if each symbol in the necklace
is the opposite of the symbol k steps to its right. That is, rotating the
necklace k steps to the right gives the same result as complementing every
symbol and not rotating at all.
If a necklace is k-alternating, it must also be 2k-periodic, because
complementing something twice is the same as not complementing it. It
can happen, however, that the necklace also has some periods smaller than
2k. For example, consider the necklace <010 10 I) of length six. This
necklace is 3-alternating and 6-periodic; but it happens to be 2-periodic as
well. Our next Lemma states that these small periods cannot arise if k is a
power of 2. This result probably seems irrelevant at the moment; be
patient.
LEMMA 5 If a necklace is 2i -alternating, then its basic period

is precisely 2i + 1; putting it another way, a 2i_
alternating necklace has no non-trivial periods.
PROOF Certainly i+ 1 will be a period; thus the basic period

of the necklace must divide 2i + 1 , so it must be a power of2. But
if the basic period were 2' for some l ~ i. then 2i would also have
to be a period of the necklace. No necklace can be both 2i_
periodic and 2i -alternating. Therefore, the basic period must be
.+ 1
precisely 21 .
With Lemma 5 out of the way, we are ready to start constructing the
up-flip strategy.
PROOF OF LEMMA 3 WHEN n IS A POWER OF 2 Suppose that

our Player is facing a table of size n = 2i. We are to construct a
strategy that uses no more than n/2 hands, and that either rings
the bell or forces the table into a state that is a perfect square. If
~302~
n = 2, there is a trivial one probe strategy that suffices. The case

of n = 4 is not that trivial, but there is a four probe strategy that
does the job: just take the first four moves of the solution to the
original Rotating Table Problem given in Martin Gardner's
column of March, 1979. Therefore, we can restrict our con-
sideration here to those cases where n ;;::: 8, or equivalently,
whereJ;;::: 3. It is somewhat surprising that the case ofn = 4 has
to be treated specially, but it seems to be necessary; the up-flip
strategy that we are about to construct just doesn't work for a
square table.
The first difference between the up-down and the up-flip strategies
arises in the definition oftheJ probe patterns P 1 ,P2 , ... ,Pj . For the up-
down strategy, we were content to let these patterns be chosen arbitrarily,
subject only to the condition that Pi resolve the ith level of nesting. Here, we
will choose a particular sequence of patterns that possesses this property, but
also has a simple structure. In detail, we will specify that the pattern Pi
consist of 2j - i blocks of 2i - 1 consecutive hands separated by 2j - i blocks of
2i - 1 consecutive gaps. More formally, we can define the necklace Pi by the
relation
1
The pattern P 1 is made up of alternating g's and h's, while the pattern Pj
probes one semicircle and leaves the other unprobed. Consider the kth
hand in each block of h's in Pi' for some k in the range 1 ~ k ~ 2i - 1. These
2j - i hands together probe one of the two 2j - i -gons that constitute a
2j - i + 1_gon; furthermore, the other constituent 2j - i -gon goes unprobed,
since it falls under the kth gap in each block of g's in Pi' This demonstrates
that the probe pattern Pi defined by Equation 1 actually does resolve
the ith level of nesting.
The first pass of the up-flip strategy is the same as that of the up-down
strategy: we instruct our Player to probe each of the patterns Pi' and to
turn up every glass that she feels. By the same analysis as in the up-down
case, we know that our Player will either win the entire Game at some
point during this pass, or else, at the end of this pass, the table will have one
down glass and (n - 1) up glasses. Call the lone down glass the exceptional
glass, and the diagonal that contains it the exceptional diagonal. Since all of
the fun is going to occur during the second pass in this argument, we will
depart from our usual subscripting convention, and use the symbol Sl to
denote the state of the table at the beginning of the second pass. The first
pass forces the table into the state
Sl = <du 2j - 1 ).
The first probe of the second pass is also the same as it used to be. We
instruct our Player to probe with pattern Pl' If she feels a down glass, it
~303~
must be the exceptional glass; she can turn it up and win at once. So we can
assume that she feels only up glasses; we instruct her to turn them all down.
This forces the table into the state
2
Before our Player goes further in the second pass, we instruct her to
find a piece of scrap paper upon which she can keep notes about the state of
the table. She begins these notes with the state S2 given in Equation 2.
With this paper at her side, we can give our Player a set of instructions that
will guide her from the second probe up until, but not including, the last
probe of the second pass. These instructions are designed to maintain the
validity of certain conditions. Let Sj denote the state of the table
immediately before the ith probe of the second pass for 2 ~ i ~ j; we will
maintain the following conditions:
3 {The state Sj diffe.rs in exactly one position from a necklace

T j that is 21 - 2 -alternating.
4 {Just after the (i - 1)st probe of the second pass, our
Player can write the state Sj on her scrap paper.
Both of these conditions are satisfied for i = 2, since the state S2 is

given explicitly by Equation 2, and it does differ in only the exceptional
place from a I-alternating necklace. Our task during the interior of the
second pass is to arrange that these conditions continue to hold. Let us
consider the ith probe of the second pass for 2 ~ i <j, where we assume
that the conditions have held so far. For this probe, we instruct our Player
to use the probe pattern Pj. Since i <j, we can deduce by the same
reasoning as in the up-down case that the pattern P j is a perfect square;
that is, that it probes both ends of any diagonal of which it probes either
end. We first ask our Player to classify the diagonals that she finds.
By Condition 3, the state Si of the table that she is probing differs in
only the exceptional place from a necklace T j that is 2i - 2- alternating. The
necklace T j must also be 2 j - 1-periodic, and thus, since i <j, it must be at
least a perfect fourth power. We can conclude that all of the diagonals of Sj
other than the exceptional diagonal must be monotone. Hence, our Player
can determine at once if she is probing the exceptional diagonal. If she is,
we instruct her to make it monotone, by either turning its down glass up or its
up glass down; it doesn't matter. This will put the table into a state that is a
perfect square, and thus satisfy the demands of Lemma 3; the rest of the
second pass should be omitted.
Suppose instead that our Player does not probe the exceptional
diagonal; that is to say, she probes only monotone diagonals. In the up-
down strategy, we instructed her to turn down all of the monotone Prgons
that she felt; but in this strategy, we instruct her to turn the monotone up
diagonals down, and the monotone down diagonals up, that is, to flip each
.8304~
monotone diagonal that she feels. This explains the name of the up-flip
strategy, but what is it good for? Since Si was 2i - 2- alternating except for
the exceptional glass, it was also 2i - l-periodic except for the exceptional
glass. The effect of flipping every glass that can be felt with Pi is to
complement and leave alone alternating blocks of length i- 1 . Therefore,
this flipping will guarantee that the next state Si+ 1 is 2i - 1 -alternating
except for the exceptional glass. That is, the flipping guarantees the truth of
Condition 3 at time i + I.
But what about Condition 4? Recall that we are assuming that our
Player does not probe the exceptional diagonal, and hence, that she flips
every glass that she feels. Since the state Si is already written on her scrap
paper, it is enough to show that she can tell without ambiguity which of the
glasses in Si she is probing, that is, where her hands are located around the
table. Consider first what the situation would be if she were probing the
necklace Ti with pattern Pi. The results of the probes by each block of 2i - 1
hands would have the form XX for some string X of length 2i - 2, because
the necklace Ti is 2i - 2 -alternating. Now, suppose that the string XX
actually appears in two different places around T i , separated by a rotation
through l positions. Since Ti is also 2i - 1 -periodic, the entire necklace Ti
must then match itself when rotated l steps; that is, l must be a period of T i •
Finally, we can see the relevance of Lemma 5; from that result, we
deduce that 2i - 1 must actually be the basic period of T i , and hence that l
must be a multiple of2 i - 1 • Considering the structure of Pi' this implies that
there are only two different ways in which the necklaces Pi and ~ could
have been aligned that would have produced the probe results xx. In fact,
~ will consist precisely of2 j - i + 1 copies of the string XX, and the Players'
hands must be probing either all of the odd or all of the even copies. Any
other possibility would force ~ to have a nontrivial period, contrary to
Lemma 5. But our Player is really probing Si' not T i. Since we are
assuming that our Player does not feel the exceptional glass, one of the two
possible orientations for her hands around Si is eliminated; and this only
leaves one possibility. Therefore, our Player will be able to tell from the
results of the ith probe precisely where her hands are positioned around the
necklace Si. With this information, she can easily compute the necklace
Si+l' and write it on her scrap paper.
We have shown that our Player can maintain Conditions 3 and 4
throughout the interior of the second pass, the "flip" phase of the up-flip
strategy. We turn now to the last probe of the second pass, where she will
be probing one semicircle. First, we want to show that our Player will be
able to tell whether or not she is feeling the exceptional glass. This time, she
cannot distinguish the exceptional diagonal by looking for a diagonal that
is non-monotone, since she is only probing one end of each diagonal. But
from Condition 3 for i = j, we know that the state Sj of the table on inp~t to
thejth probe of the second pass will differ in only one position from a 2r2_
alternating necklace T j • Therefore, our Player can determine whether or
.8305~
not she is feeling the exceptional glass by checking to see whether or not the
results of her probes have the form XX for some string X of length 2j - 2. In
particular, if she misses the exceptional glass, then her probe results will
have the form XX; but if she is feeling the exceptional glass, her probe
results will have the form rdzfd<, where one of the two d's is the
exceptional glass.
Suppose that our Player does feel the exceptional glass. Her next task
is to determine which of the d's is the exception, so that it can be corrected.
And it is here that our assumption that j ~ 3 is critical. If we try and use
the up-flip strategy for a square table, we will be faced at this point with
trying to decide which of the d's in the block dd is the exceptional one and
there is just no way to tell. But we have already handled the case of n = 4
by a special argument. Therefore, we can conclude that at least one of the
two strings rand Z is non-empty, and hence that our Player can actually
feel some neighborhood of each of the two current candidates for the
exceptional glass. Note that those neighborhoods will be complementary.
Now, by Condition 4, we know that our Player's piece of scrap paper has
the exact necklace Sj written on it. From this necklace, it is easy to deduce
what states the left and right neighbors of the exceptional glass are in. This
observation allows our Players to determine which of the current can-
didates is actually the exceptional glass. We instruct her to flip just that
glass. This forces the exceptional diagonal into the monotone state <uu),
and hence satisfies the demands of Lemma 3.
There is only one more case to handle. We now suppose that, on the
last probe of the second pass, our Player does not feel the exceptional glass;
xx.
that is, the results of her probes take the form In this case, the necklace
T j from which the state Sj diff~rs ~n only the exceptional place must be
given by the formula T j =:. <XXXX). Arguing from Lemma S. again, we
can show that the string XX will appear as a substring of T j only in the two
obvious places. Hence, it only appears once as a substring of Sj; this means
that our Player will be able to determine uniquely the location of her hands
around Sj. Since she is probing a semicircle, and she doesn't feel the
exceptional glass, she must be probing the other vertex of the exceptional
diagonal. We instruct our Player to look at her scrap paper, and figure out
which of the up glasses that she can feel is the one diagonally opposite the
exceptional glass. Turningjust this up glass down will force the exceptional
diagonal into the monotone state <dd), and thus satisfy the demands of
Lemma 3.
We have completely determined the values of the function f (n). But
this does not mean that we can close the book on the Problem of the
Rotating Table. On the contrary, we mentioned in the Introduction
several other generalizations of the problem that are worthy of study, and
even the variant that we have been considering might repay further
analysis. For example, one might be interested in constructing short
winning strategies for a k-handed Player, where k either equals or exceeds
f(n).
.e3306~
PROBING THE ROT ATING TABLE
The strategy-building techniques that we have discussed can give us

some preliminary feel for the lengths of one family of strategies. In
particular, let n = ala2 ... aj be a factorization ofn in which each factor ai
satisfies a i 2 2; suppose for convenience that the ai have been numbered so
that all of the 2's, if any, come first. By using the up-down strategy if aj > 2,
or the up-flip strategy if a j = 2, we can reduce the strategy-building
problem of size n to the problem of size nj aj in at most 2) probes, using no
more than
K = max
l~i~j
(1 - ~)n
ai
hands. Therefore, we can construct a winning strategy for a size n table that
uses only Khands and takes at mostl + ) probes all told. There is a trade-
off here: either we can make the a/s large, which makes) small and the
resulting strategy short, but demands lots of hands; or we can make the a/s
small, which makes) large and the strategy long, but gets by with fewer
hands. Furthermore, it is relatively easy to shorten these strategies at least a
little. For example, it is often possible to save probes by carrying over
information from one up-down stage of a strategy to the next. But are there
any strategies that take substantially fewer probes in the worst case than
those that we have constructed?
Note added in Proof: Ted Lewis and Stephen Willard determined J(n)
independently, and they have published their results in the article The
Rotating Table in the May, 1980 issue of Mathematics Magazine (volume 53,
number 3, pages 174-179). When n is not a power of 2, they employ a
refinement of the up-down strategy in which carrying information over
from one up-down stage to the next allows the Player to get by with
roughly half as many probes. Furthermore, they discovered a close relative
of the up-down strategy that handles the cases in which n is a power of 2:
The first pass turns glasses up, and most of the second pass turns them
down. Flipping is confined to the next to the last probe of the second pass,
which flips a single diagonal.
eA307~
Numbers and
Coding Theory
-~-
Supernatural
Nuntbers
-~-
Donald E. Knuth
STANFORD UNIVERSITY
'G od," said Leopold Kronecker [10], "made the integers; everything
else is the work of man." If Kronecker was right, it would be heresy to
claim that any noninteger numbers are supernatural in the sense that they
have miraculous powers. On the other hand, mathematicians generally
refer to the nonnegative integers {O, 1,2, ... } as the set of natural numbers;
therefore if any numbers are supernatural, they are also natural. The
purpose of this essay is to discuss the representation of natural numbers that
are "super" in the sense that they are extremely large; many superscripts
are needed to express them in conventional notation.
It doesn't take a very big number to transcend the size of the known
universe. For example, if we consider a cube that's 40 billion light years
long on each edge, and if w~ pack that cube with tiny little 10- 13 cm
X 10- 13 cm X 10- 13 cm cubes (so that each tiny cube is much smaller
than a proton or neutron), the total number of little cubes is less than
10 125 . This quantity isn't really humungous; it has only 125 digits.
The great Archimedes seems to have been the first person to discuss
the existence of extremely large numbers; his famous essay "The Sand-
Reckoner" [1] concludes by demonstrating that fewer than 10 63 grains of
sand would be enough to fill the universe as defined in his day.
Furthermore he introduced a system of nomenclature by which he could
speak of numbers up to 1080,000,000,000,000,000
~310&
SUPERNATURAL NUMBERS
The English language doesn't have any names for such large
quantities, and it is instructive to consider how we could provide them
systematically. In these inflationary days, we may soon need new words to
express prices; during 1923, for example, Germany issued postage stamps
worth 50 billion marks each, but these stamps were almost valueless.
When we stop to examine our conventional names for numbers, it is
immediately apparent that these names are "Menschenwerk"; they could
have been designed much better. For example, it would be better to forget
about thousands entirely, and to make a myriad (10 4 ) the next unit after
hundreds. After all, numbers like 1984 are conventionally read as
"nineteen hundred eighty-four," not as "one thousand nine hundred
eighty-four." * The use of myriads would provide us with decent names for
everything up to 10 8. For example, the prime number 9999,9989 would be
called "ninety-nine hundred ninety-nine myriad ninety-nine hundred
eighty-nine. "
The next unit we need after myriads is 10 8 . Let's agree to call this a
myllion (pronounced mile-yun); similar words like myllionaire will, of
course, also be of use. After myllions come byllions; one byllion equals 10 16 .
It clearly is an improvement to have the unambiguous word byllion for this
next step, since people have never been able to agree about what a
"billion" is. (English and German" people think it is a million million, while
Americans and Frenchmen think it is a thousand million.) Note also the
comforting fact that our national debt is only a small fraction of a byllion
dollars. The nomenclature should now continue as follows:
1032 tryllion 102048 nonyllion 10131072 quindecyllion
10 64 quadryllion 104096 decyllion 10262144 sexdecyllion
10 128 quintyllion 10 8192 undecyllion 10524288 septendecyllion
10 256 sextyllion 1016384 duodecyllion 101048576 octodecyllion
10 512 septyllion 1032768 tredecyllion 102097152 novemdecyllion
10 1024 octyllion 1065536 quattuor- 104194304 vigintyllion
decyllion
For example, the total number of ways to shuffle a pack of cards is 52!
= 8065:: 81 75, 1709; 4387,8571 : 6606, 3685; 6403, 7669;; 7528,9505; 4408,
8327: 7824,0000; 0000, 0000, if we use punctuation marks to group the digits
by fours, eights, sixteens, and so on. The English name for this number
under the proposed system would be "eighty hundred sixty-five quadryllion
eighty-one hundred seventy-five myriad seventeen hundred nine myllion
forty-three hundred eighty-seven myriad eighty-five hundred seventy-one
byIIion sixty-six hundred six myriad thirty-six hundred eighty-five myllion
sixty-four hundred three myriad seventy-six hundred sixty-nine tryllion
seventy-five hundred twenty-eight myriad ninety-five hundred five myllion
forty-four hundred eight myriad eighty-three hundred twenty-seven byllion
seventy-eight hundred twenty-four myriad myllion." This number is
inexpressible in our present American language, unless one resorts to
.83ll&.
something like eighty thousand six hundred fifty-eight vigintillion one

hundred seventy-five nonillion ... eight hundred twenty-four trillion," since
unabridged dictionaries don't give names for any units past one vigintillion
( = 10 63 ) except centillion (= 10 303 ).
Of course we need to go past vigintyllions and even centyllions if we are
to have names in our language for all of God's creation. The next unit
presumably should be called "unvigintyllion," if we extrapolate the above
pattern and go further into Latin-like nomenclature. Thus, from the
formula
80 ,000 ,000 ,000 ,000 ,000 - + 2 52 + 2 51 + 2 50 + 2 45 + 244 + 242

2 56
+ 241 + 2 40 + 2 39 + 2 36 + 2 33 + 2 32 + 2 30
+ 2 29 + 2 28 + 227 + 2 26 + 2 25 + 2 19 ,
a reader who understands the proposed system will know the new name of
Archimedes' last number 10 80 ,000,000,000,000,000. (The answer appears at
the end of this article.)
Sooner or later we will run out of Latin names to give to the units, since
Romans never did count very high. Even if Roman scholars had adopted
Archimedes' scheme, we would have trouble naming a unit like
10 80,000,000,000,000,003
10 2
However, we can get around this problem by simply giving each basic unit
10 2 " + 2 the name "latin {the name of n with spaces deleted }yllion" for all
210000000000000002 . . .
large n. For example, 10 would be "latmbylhonylhon."
In this way we obtain English names for all the natural numbers, no matter
how super they are. For example, one of the names would be "latinlatinlat-
inbyllionyllionyllionyllion"; can the reader deduce its magnitude? (See the
end of this article for the answer.)
At this point the reader may be thinking, "So what? Who cares about
names for gigantic numbers, when ordinary decimal or binary notation
gives a simpler representation that is much more suited to calculations?"
Well, it's quite true that number names are more interesting to the linguistic
parts of our brains than to those little grey cells we use for calculation; the
mere fact that a great man like Archimedes wrote about some topic doesn't
necessarily prove that it has any scientific interest. Our discussion has not
been completely pointless, however, because it has prepared us to talk about
a more important problem whose solution involves somewhat similar
concepts.
The problem we shall discuss in the remainder of this article is this:
How can arbitrarily large natural numbers be represented as sequences of
O's and l's so that
.8312&
I The representation of each number is never a "prefix" of the

representation of any other number.
In other words, if some number is represented by the sequence "01101", no

other number will have a representation that begins with "0110 1".
Furthermore the representation should satisfy a second condition,
2 The representation of each number should be "lexico-

graphically smaller" than the representation of every larger
number.
Lexicographically smaller means that it would appear earlier in a

dictionary; for example, 01101 is lexicographically smaller than all
sequences beginning with "1" or with "0 Ill" or with "0110 10" or
"011011 ".
The importance of this problem is that we often want to communicate
a sequence of numbers to a computer. For example, such a sequence might
stand for instructions telling the computer what to do. The critical problem
is that the computer must know where one number stops and the next one
begins; condition I ensures that this will be the case, since the computer can
read the sequence from left to right and there will be no ambiguity.
Condition 2 is not quite so critical, but it is a nice property to have: it implies
that the representation is order-preserving in the sense that, whenever one
sequence of numbers is lexicographically smaller than another, the
representation of the first sequence will also be lexicographically smaller
than the representation of the second.
Conditions I and 2 are easy to achieve if we need to represent only a
few numbers. For example, if we happen to know that only the numbers 0
through 7 will ever be required, we can make all the representations of the
same length: 000, 001, 010, 011, 100, 10 1, 110, Ill. But we want to be able
to encode all natural numbers, not only the small ones.
Perhaps the first correct solution that comes to mind is a "unary"
representation, where we represent 0 as "0", 1 as "10", 2 as "110", and so
forth; the representation of n is "1 no", that is, n ones followed by a single
zero. The zero acts as an end-marker or "comma" between successive
numbers that the computer must read. This representation clearly satisfies I
and 2, but it is hardly a practical way to represent even medium-size
numbers, since it requires n + 1 bits to represent n. We really want to have a
representation that is as concise as possible, subject to conditions I and 2.
Perhaps the next simplest solution is based on the ordinary binary
notation for integers, namely {O, 1, 10, 11,100,10 I, ... }. This representation
as it stands doesn't satisfy either I or 2, but we can improve it by prefixing
the binary notation by a suitable sequence indicating its length, as follows:
eA 313&'
0-+00 8 -+ 1110000 64 -+ 1111110000000

1 -+ 01 9 -+ 1110001 127 -+ 1111110111111
2 -+ 100 10 -+ 1110010 128 -+ 111111100000000
3 -+ 101 15-+1110111 255 -+ 1111111 0 1111111
4 -+ 11000 16 -+ 111100000 256 -+ 11111111000000000
5 -+ 11001 31-+ 111101111 511-+ 11111111011111111
6-+11010 32 -+ 11111000000 512 -+ 1111111110000000000
7-+11011 63 -+ 11111011111 1000 -+ 1111111110111101000
In general for n ~ 2, if the binary representation of n is 10( where 0( is any

sequence of O's and l's, the new representation of n is
1 11%1 00(
where 10(1 denotes the length of 0(. If m bits are needed to write n in binary
notation, its new representation has 2m - 1 bits. We have roughly doubled
the number of bits needed to represent n, in order to indicate un-
ambiguously where each representation ends.
This solution can be further improved, but before we pursue the
investigation any further it is interesting to point out that our problem is
essentially equivalent to a guessing game, where one person has thought of an
arbitrary natural number and the other person tries to guess it. The rules of
this game are something like Twenty Questions: the one who guesses is only
allowed to ask "Is your number less than n?" for nonnegative integers n,
and the other player answers "yes" or "no." Of course the game might last
longer than twenty questions, because it is impossible to distinguish between
more than 2 20 numbers based on the answers to only 20 yes-no questions,
while infinitely many secret numbers are allowed. But the general idea is to
guess the number as quickly as possible.
The relation between this guessing game and the sequence repre-
sentation is not difficult to see. Any strategy that can be used by the guesser
to deduce all natural numbers leads to a solution to 1 and 2: We simply let
the representation ofn be the sequence of answers that would be given when
the secret number is n, with 0 standing for "yes" and 1 for "no." Con-
versely, given a solution to 1 and 2 we can construct a guessing strategy that
corresponds to it under this rule, if we also allow the guesser to ask the
stupid question "Is your number less than infinity?" (The answer will
always be "yes"; some solutions to 1 and 2 have every representation
beginning with 0.)
The unary solution that we discussed first corresponds to the strategy
under which the guesser simply asks
Is your number less than I?

Is your number less than 2?
~31H~
and so on, until receiving the first "yes." Our second solution IS more
clever: it begins with the questions
and so on, then it uses a "binary search" to pinpoint the secret number
once the first "yes" answer has revealed its order of magnitude.
Under this equivalence between the guessing game and the repre-
sentation problem, a good strategy for the guesser corresponds to a concise
representation for numbers; so our search for good representations boils
down to the same thing as the search for a good number-guessing scheme.
Incidentally, the number-guessing game is not simply frivolous, it
has important practical applications. For example, if a polynomial f(x)
is known to have exactly one positive root, whereJ(O) is negative andJ(x) is
positive for all sufficiently large x, the root is less than n if and only ifJ(n)
is positive; thus a good strategy for number guessing leads to an efficient
procedure for root location without evaluating derivatives off
We can improve on the second solution above by realizing that it is
essentially using the unary strategy to represent the length of n's binary
representation; the binary strategy can be substituted instead! In other
words, the sequence llal 0 used to encode the length of a can be replaced
by amore concise representation of Ial. By using this idea repeatedly, we
obtain progressively shorter and shorter representations for large numbers.
And we are eventually led to what might be called a recursive strategy, of the
following form: First guess the number of binary digits of n, then use binary
search to determine the exact value of n. To guess the number of digits of n,
the same strategy is used, recursively; thus we first guess the number of
digits in the number of digits of n, by guessing the number of digits in the
number of digits ofn, and so on. The first questions in the recursive strategy
are:
Is your number less than I?
and so on, until the answer is "yes." (Note that 2 = 21, 4 = 22 , 16 = 24 , and
65536 = 2 16 ; the next question wou1d refer to 265536.) The recursive
strategy proceeds in this way until finding an upper bound on the secret
number; then it proceeds to unwind the recursion. If the secret number is
really enormous, the guesser will soon be guessing really enormous values.
The recursive representation corresponding to this recursive guessing
scheme can be defined very simply: Let R(n) be the sequence ofO's and 1's
that represents n. Then
et315&'
R(O) = 0; and
R((lex)2) = 1 R(lexl)ex,
where (lex)2 is the number whose binary representation is lex and lexl is the
length of the sequence ex. Small integers are represented as follows:
0--+0 8 --+ 11101000 64 --+ 111100 10000000
1 --+ 10 9 --+ 11101001 127 --+ 11110010111111
2 --+ 1100 10 --+ 11101010 128 --+ 111100110000000
3 --+ 1101 15--+ 11101111 255--+ 111100111111111
4 --+ III 0000 16 --+ 111100000000 256 --+ 11110 100000000000
5 --+ 111000 1 31 --+ 11110000 1111 511--+ 11110100011111111
6 --+ 11100 10 32 --+ 1111000100000 512 --+ 11110 100 1000000000
7 --+ 1110011 63 --+ 1111000111111 1000 --+ 111101001111101000
It is easy to see that this representation satisfies 1 and 2. Computer scientists
interested in the transformation of recursive methods to iterative methods
will enjoy finding a simple nonrecursive procedure that evaluates n, given
its representation; a solution appears at the end of this article.
The recursive representation of a large number n will be about half as
long as its representation under the binary scheme. For example, the
representation of 265536 - 1 in the binary scheme is the sequence
1655350165535
of length 131071, while its representation in the recursive scheme is

150165554,
only 65560 bits long. On the other hand, it takes quite awhile for the
recursive scheme to show any payoff when n is small; the binary scheme
produces a representation shorter than the recursive one for all values of n
between 2 and 127 inclusive. (The recursive scheme wins out only when n
= 0 or when n is 512 or more.) Thus the binary scheme will be preferable
in many applications.
We can improve the recursive scheme's performance for small n by
observing that all sequences except for the one representing 0 begin with 1;
if we only want to represent strictly positive numbers, we can safely drop
the initial "1". Furthermore the natural numbers are in one-to-one
correspondence with the strictly positive integers, so there is a repre-
sentation Q(n) such that
lQ(n) = R(n + 1).
Under this modified recursive scheme we have
0--+0
1 --+ 100
2 --+ 101
3 --+ 110000
4 --+ 110001
.8316~
and so on; the binary scheme beats this one only for n = 1,3,4,5,6,7, and
for 15 ~ n ~ 63. The same transformation can be applied to the Qscheme,
for the same reasons, yielding a P scheme where
IP(n) = Q(n + 1),

so that we have
0 .... 00
1 .... 01
2 .... 10000
3 .... 10001
4 .... 10010
and so on; now the binary .scheme wins only when n = 2,3,6, 7,62,63, and
when 14 ~ n ~ 31.
In some sense we would like to find the best possible strategy, but it is
hard to say exactly what that means. The binary strategy looks a lot better
than the unary one, but even the unary strategy beats the binary one when
n = o. We will see below that the situation is inevitable: No strategy for the
guessing game dominates another-in the sense that the first needs no
more questions than the second does to determine n, for all n, and the first
sometimes needs fewer questions-unless, of course, the dominated strat-
egy asks stupid questions whose answer "yes" or "no" is already deducible
from the previous answers. If one nonstupid strategy is better than another
for some n, it will necessarily be worse for others. You win some, you lose
some.
In order to analyze just how good these schemes are, we ought to be
more quantitative. (Let the reader beware: The rest of this article involves
mathematical technicalities that are elementary but sometimes a bit
subtle.) If n ~ 1, let us write }..n for the unique natural number such that
2Aft ~ n < 21 + Aft. Thus ifn = (lexh in binary notation, the number}..n is lexl,
the length of ex. It is convenient to define }..O = 0, so that }..n is a natural
number whenever n is a natural number. We shall write }..}..n for }..(}..n), and
so on; furthermore }..mn will stand for the m-fold repetition of the }.. function,
so that }..°n = nand}" 3 n = }..}..}..n, etc. Finally, we also write}.. *n, to indicate
the least integer m such that }..mn = O.
It is easy to express the length of the representation of n in terms of
such functions. Let c(n) be this length, that is, the "cost" of representing n.
In the guessing game, c(n) is the number of questions needed to determine a
given secret number. The unary guessing strategy has a rather large cost,
cu(n) =n+ 1,
while the binary strategy reduces this to
2, ifn=Oorn=l;
cB(n) = {2hz+ I, ifn> 1.
.el317&.
The cost of the recursive strategy is
CR(n) =An+ AAn+ AAAn+ ... + A*n+ 1,

where the infinite series implied by the"···" is really finite, since Amn = 0
whenever m ~ A*n. Finally, the modified recursive strategies have costs
cQ(n) = cR(n+ 1) -1, cp(n) = cR(n + 2) - 2.
These formulas verify our earlier remark that the recursive strategies cost
about half as much as the binary scheme when n is large.
Let us say that a representation scheme is irredundant if the cor-
responding guessing game never asks any stupid questions (questions that
have answers already known or that could be deduced from those known).
This means that if IX is any sequence of O's and 1's such that IX is not
the representation of any n, but such that IX does occur as a prefix of some
representation of an integer, then both IXO and IXI occur as prefixes of
representations. All reasonable schemes will be irredundant; for if, say, IXO
occurs as a prefix but IXI doesn't, we can shorten the representation of some
integers without violating 1 or 2 merely by deleting the 0 following IX
whenever IX appears as a prefix.
The cost functions ofirredundant representations satisfy an important
arithmetical relation:
FACT 1 Let c(n) be the cost of representing n in an irredun-

dant representation scheme. Then
1 1 1
2c(O) + 2c(1) + 2 c(2) + ... = 1.
PROOF In fact, if IX is the representation of n, we have

1 1 1
2c(O) + 2c(1) + ... + 2 c(n-l) = (.IXh
in binary notation, for all n. This formula dearly holds when n

= 0, since there can be no 1's in an irredundant representation
ofO. Let 13 be the representation ofn + 1; we need to show that
I
(.IX) 2 + 21(11 = (.f3h-
No number is represented by a sequence of 1's only, since

the sequence 1m is followed in lexicographic order only by
sequences of which it is a prefix. Thus (.Q()2 + 2- 1(11 is less
than 1, and in fact (.Q(h + 2- 1(11 = (.ala2a3 ... h is the smallest
binary number greater than (.IX)2 not having Q( as a prefix. If
..6318~
(.f3)2=(.b 1b2b3... h is unequal to (.ala2a3 ... h, let j be

minimal such that bj =1= aj and let k be maximal such that ak = 1.
We have (. b1b2b3 ... h > (. ala2a3 ... h, hence bj = 1 and aj = O.
Ifk<j, the sequence b1 ... bj - 1 is a redundant prefix, because
b 1 ... bj - 1 l occurs as a prefix of the representation ofn+ 1 but
b1 ... bj-10 never occurs as a prefix. If k>j, the sequence
al ... ak-l is a redundant prefix, because al ... ak-lO occurs as
a prefix of the represen ta tion of n but a1 ... ak _ 11 never occurs
as a prefix. (Note that k is the position of the rightmost 0 in a.)
It is easy to verify that lk occurs as a prefix for all k. Now if
I k is a prefix of the representation of n, the sum 2- c(O) + 2 -c(l)
+ ... + 2- c(n-l) lies between (.1 kh = I - Tk and 1, hence the
infinite sum as n -H/J converges to 1. QE.D.
(This proof uses both properties 1 and 2, and indeed the result would
be false if only property 1 were assumed. For example, the following
procedure generates representations oflength n + k for each integer n in an
irredundant fashion, for any fixed k ~ 2: Represent 0 by Ok, then for n
= 1,2,3, ... , find a sequence a such that (a) a has occurred as a prefix of a
representation for some number < n but not as a representation; (b) aO and
al have not both occurred as prefixes; and (c) a is as short as possible
satisfying (a) and (b). Then the representation ofn can be any sequence of
length n + k containing aO or al as a prefix, whichever hasn't previously
occurred. For example, one such representation scheme for k = 2 begins
0-00 3 - 11000 6 - 11100000

1-100 4-011000 7 - 010100000
2 -0100 5 - 1010000 8 - 0111000000.
On the other hand, it is not difficult to prove the inequality
I I I
-+-+-+···<1
2c(O) 2c(1) 2c(2) -
for all representations which satisfy property 1; this relation is known as

"Kraft's inequality" [9].)
There is also a converse to Fact 1:
FACT 2 Let c(0),c(I),c(2), ... be a non decreasing sequence of

positive integers such that
1 1 1
-+-+-+···-1
2c(O) 2c 2c(2)
(1) -
Then there is an irredundant representation scheme

with cost function c(n).
~319&.
PROOF By lettinR the reRresentation ex of n be defined by

the formulas 2- C(0 + 2- c( )+ ... + 2- c(n-1) = (.exh and lexl
= c(n), we obtain a scheme with the desired properties. QE.D.
If X is an irredundant representation scheme whose cost function Cx is
not monotonic, we can permute the integers to obtain another repre-
sentation scheme with the same costs sorted into nondecreasing order.
Facts I and 2 now show that there exists an irredundant scheme having
these sorted costs. Representation schemes with monotonic cost functions
can be called standard, since most applications prefer to have cx(n) :$;
cx(n + I) for all n.
Let us conclude our investigations by trying to find the best possible
scheme, putting an emphasis on asymptotic questions (that is, on the
efficient representation of very large integers). Clearly the binary method is
more efficient than the unary method, and we also prefer the recursive
method to the binary method because of its superior performance on large
numbers. These examples suggest the following definition: "A repre-
sentation scheme X with cost function cx(n) dominates another scheme r
with cost function cy(n) if cx(n) :$; cy(n) for all large nand cx(n) < cy(n) for
infinitely many n." If X dominates rand r dominates Z by this definition,
then X dominates Z.
Clearly the recursive methods P, Q, R all dominate the binary
method B, and B dominates the unary method U. When we try to compare
the three recursive schemes with each other, however, it turns out that
none of them is dominant. Method P is best nearly all the time; in
particular, whenever An = A(n + 2) we havecp(n) = cQ(n) - I = cR(n) - 2.
But there are infinitely many n where Qis better than both P and R (namely
when n = 221< - 2 and k ;;::: I), and there are infinitely many n where R is
2k
better than both P and Q (namely when n = 22 - I). These facts suggest
that it might be impossible to dominate method R, and in that sense we
might conclude that method R is "optimal." However, any such hopes are
quashed by
FACT 3 If X is any standard representation scheme, there is

another standard representation scheme r that do-
minates X and satisfies cy(n) :$; cx(n) for all but one
value of n.
PROOF The general idea of the proof is to choose a sequence
of length c and replace it by infinitely many sequences of
respective lengths c + 1, c + 2, c + 3, etc., because 2- c = 2- c - 1 +
2- C - 2 + 2-c- 3+ ....
Proceeding more formally, we may assume that X is irre-
dundant. Fork> lletak be the numberofn'ssuch that cx(n) =k,
and let J be minimal such that aj > O. Let hj = aj -1 and let
~320~
bk = ak + I for all k > j. Then there is a unique nondecreasing

function cy(n) having exactly bk values ofn with cy(n) =k, and
this function satisfies 2- Cy (O)+ 2- Cy(1) + ... = 2- Cx(O) + 2-cx(1)
+ .... By Fact I and Fact 2, there is an irredundant representa-
tion scheme with cost function Cy. Furthermore it is easy to see
that cy(n) ~ cx(n) except for the single value of n such that cx(n)
= j and cx(n + I) > j; and cy(n) < cx(n) whenever j + I
<cx(n-l)<cx(n). Q.E.D.
I t is hopeless to find a strictly optimum scheme, because repeated

application of the construction in the proof of Fact 3 will give an infinite
family of better and better schemes, each dominating its predecessors.
However, it turns out that no scheme can be much better than our
recursive scheme R, in spite of all the apparent improvement guaranteed
by Fact 3.
FACT 4 Let An be the function defined by

An = An + AAn + AAAn + ....
Every cost function c(n) corresponding to a repre-
sentation scheme satisfies
c(n) > An + AA *n
for infinitely many n.
PROOF Let d(n) = An + AA *n; we shall show that the sum

Ln ~ 0 2 - dIn) diverges. This suffices to complete the proof, for
if c(n) ::=;;; d(n) for all n 2 m we have the impossible inequality
I = Ln~o2-c(n) 2 Losn<m(2-c(n)-2-d(n») + Ln~o2-d(n). In gen-
eral ifJ(n) = An + g(A *n) for any function g, we have
I I
L 2f(n)
n~O
= L 2g (n) ,
n~O
since the left-hand side can be written
and for m 2 I we have

I I I
L 2An = L 2).n+A).n = L 2k+Ak L I
).*n=m ).*).n=m-l )'*k=m-l ).n=k
I
= L Ak·
)'*k=m-l 2
~321~
Thus the sum Ln~o 2- d(n) diverges if and only if the sum
Ln~o 2- An diverges, and it does. QE.D.
Using the same proof technique, we can show in fact that

c(n) > An + AA *n + AA *A*n + ... + A(A *)mn + A(A *)m+ In
infinitely often, for any fixed representation scheme and any fixed m.
Let us close by showing how to improve scheme R so that we obtain
"ultimate" schemes that come very close to this lower bound. The
sequence R(n) begins with I A"no, a sequence oflength A*n + 1 that serves to
identify A*n, followed by a sequence oflength An that characterizes nonce
A*n is known. In this sense, the R method starts by using a unary approach
to guess A*n; and we know that we can do better. Let us therefore use the R
method to determine A*n, then continue to determine n as before, calling
this the RR method. The new cost function will be
cRR(n) ::::: An + cR(A*n) = An + AA*n + A*A*n + 1,
and the RR representation of n will begin with the sequence I A"A"no.
But wait, let's start by guessing A*A*n; then we obtain an RRR
method whose cost function is
cRRR(n) = An + AA *n + CR(A *A*n)
= An + AA *n + AA *A*n + A*A*A*n + l.
The R m + I method has a cost function equal to
c(n) =An+ AA*n+ ... + A(A*)mn+ (A*)m+l n + 1,
and this upper bound is almost the same as our lower bound.
AcknowledgeDlents
Philip Davis [4] has written a delightful introduction to the elementary
facts about large numbers. The problem of representation schemes was
originally treated by Levenshtein [11], who presented Method Rand
proved that c(n) > An - A*n for infinitely many n; he also discussed
representations with more than two symbols. The representation problem
has also been studied independently by Elias [5] and by Even and Rodeh
[6]; the guessing game for unbounded search was suggested by Bentley and
Yao [2]. Methods P and Qwere proposed by Jim Boyce and David Fuchs
during a recent conversation with the author.
Numbers much larger than those considered here are discussed in [8].
Using the notation of that article, we have A*(2jjm) = m + I and
A*(2iiim) = (2ijj(m - 1)) + 1, suggesting that the upper and lower
bounds we have derived are really not so close together after all; perhaps
the introduction of functions A**n and A***n, etc., will lead to further
clarification of "optimum" representation schemes when we imagine
ourselves dealing with supersupernatural numbers.
~ 322€:l..
An ancient Chinese mathematician named Hsii Yo (c. 200 A.D.)

discussed a nomenclature for large numbers containing the units wan
= 104 , i = 10 8 , chao = 1016 , and ching = 10 32 , in his interesting book Shu
Shu Chi I; see [12, p. 87]. I am indebted to Tung Yun-Mei for this reference.
There is a beautiful connection between the representation problem
discussed here and the information-theoretic concept of algorithmic
complexity originated independently by L. A. Levin in Russia and G. J.
Chaitin in Argentina. A certain function l(n), called KP(n) by Levin (cf.
[7]) and H(n) by qhaitin [3], has the following two properties: (a) There is
a representation scheme for natural numbers that satisfies condition 1 and
has cost l(n). (b) For every representation scheme satisfying condition 1
dnd having cost c(n), there is a constant C such that l(n) :::; c(n) + c.
Intuitively, I (n) represents the length of the "simplest" description of an
algorithm to compute n. This function I (n) is not computable, but it is
semicomputable from above, in the sense that, if l (n) is actually less than a
given number m, we can prove this fact in a finite amount of time. If c(n) is
the cost function for any representation scheme satisfying conditions 1 and
2 then there is a constant C such that An + l(An) :::; c(n) + C. Furthermore
there is a constant Co such that IAn + l(An) - max! :5k:5nl(k) I:::; Co. Thus the
constructions we have given provide bounds on max 1:5 k:5 n l (k). I am
indebted to Peter Gacs for these references.
References
I Archimedes. 1956. The Sand Reckoner. In The World if Mathematics,
vol. 1, ed. James R. Newman, pp. 420-31. New York: Simon and
Shuster.
2 Bentley, J. L. and Yao, A. C. 1976. An almost optimal algorithm for
unbounded searching. Inj. Proc. Letters 5: 82-87.
3 Chaitin, GregoryJ. 1975. A theory of program size formally identical to
information theory. Journal if the ACM 22: 329-340.
4 Davis, Philip J. 1966. The Lore if Large Numbers. New Mathematical
Library, vol. 6, New York: Random House.
5 Elias, P. 1975. Universal codeword sets and representations of the
integers. IEEE Trans. Iriform. Theory IT-2l: 194-203.
6 Even, S. and Rodeh, M. 1978. Economical encoding of commas
between strings. Comm. of the ACM 21: 315-317.
7 Gacs, Peter. 1974. On the symmetry of algorithmic information. Soviet
Math. Doklady 15, 5: 1477-1480,15,6: v.
8 Knuth, D. E. 1976. Mathematics and computer science: coping with
finiteness. Science 194: 1235-1242.
9 Kraft, L. G. 1949. A device for quantizing, grouping and coding
amplitude modulated pulses. M.S. thesis, Electrical Eng. Dept., Mass.
Inst. of Technology.
eJ323~
10 Kronecker, L. 1893. Remark in a lecture at the Berlin scientific

congress, 1886: "Die ganzen Zahlen hat der liebe Gott gemacht, alles
andere ist Menschenwerk." Quoted by H. Weber in Math. Annalen 43:
15.
II Levenshtein, V. E. 1968. On the redundancy and delay of decodable
coding of natural numbers. Systems Theory Research 20: 149-155.
12 Needham, Joseph. 1959. Science and Civilisation in China 3 Cambridge:
Cambridge University Press.
Solutions
1 Archimedes' last number: One septendecyllion trevigintyIlion quat-
tuorvigintyllion quinvigintyllion sexvigintyllion septenvigintyllion octo-
vigintyllion trigintyllion untrigintyllion quattuortrigintyllion sept-
en trigin tyllion octotrigin ty Ilion nov em trigin ty Ilion q uadragin tyllion
duoquadragintyllion trequadragintyllion octoquadragintyllion nov-
emq uadragintyllion q uinq uagin ty Ilion q uattuorq uinq uagin ty llion.
(His name for this number was much shorter, but the new system beats
his in lots of other cases.)
Incidentally, American names for large numbers are not based on
good Latin (sexdecillion should be "sedecillion" and novemdecillion
should be "undevigintillion"); so it is not clear how much further the
American system can be extrapolated.
2 One latinlatinlatinbyllionyllionyllionyllion is
210000000000000002
2(10 +2)
lO2(10 + 2)
We might call this one umptyllion, for short.

3 Iterative decoding procedure for Method R: Letthe input sequence be
where each Si is 0 or 1. Let k be the number of input bits
SI s2s3 ••• '
already read, let l be the length of a number to be input, letm be the
stack depth in a recursive implementation, and let n be the answer
returned. The following recursive method corresponds directly to the
definition, with k initially zero, using only global variables.
PROCEDURE R
Set k+-- k + 1.
If Sic = 0, set n'- 0 and exit from R.
Otherwise, call R (recursively).
Then set l .- n, n .- 1.
While l > 0, repeatedly set k .- k + 1, n .- 2n + sic' l.- l - 1.
Exit from R.
~324&.
By mechanical transformations this recursive program can be converted

into the following purely iterative procedure:
Set k +-1 and m +- O.

While Sk = I, repeatedly set m +- m + I and k +- k + 1.
Set n +- O.
While m > 0, repeatedly do the following operations:
Set m +- m - 1, I +- n, n +- 1.
While I > 0, repeatedly set k +- k + 1, n +- 2n + Sk, I +-l - 1.
The corresponding encoding procedure can be treated similarly.
* We might also consider changing "nineteen" to "onety-nine," etc.; but that

might make our nomenclature too logical.
~325~
The Graph Theorists
Who Count-and
What They Count
-~-
Ron Read
What is a graph? To the majority of people the word "graph" conjures

up a picture either of the kind of businessman's chart shown in Figure 1, or
of a smooth curve, like Figure 2, which displays the property of some
mathematical function. But to a large and growing number of mathema-
ticians this same word suggests something completely different. These are
the people who are concerned, in one way or another, with the branch of
mathematics known as graph theory. When they think of a graph they
10
J F M A J J A SO ND
FIGURE 1
~326~
WHAT THE GRAPH THEORISTS COUNT
FIGURE 2
have in mind a diagram consisting of a number of points, or nodes, and lines

or curves joining some of these nodes to others. These lines are often called
edges; an example is given in Figure 3. (Most of the other figures in this
article depict graphs of one sort or another.) These objects are what the
man in the street would tend to refer to as networks (a word that graph
theorists use in a somewhat special sense).
Graph theory-the study of these graphs and their properties-
started as a branch of topology. Indeed, it was once referred to as the
"slums of topology," but it has since become respectable (if it ever was
otherwise) and is now of considerable importance, finding application to
many problems throughout mathematics and the sciences.
Why should a subject which, at first sight, appears to be concerned
with nothing more substantial than paper and pencil doodling be of any
great use? We can answer this question by looking around us and seeing
how many things of importance in everyday life are essentially graphs, or
can be represented by graphs. Examples that come to mind are com-
munication systems of various kinds; telephone networks, where telephone
lines link certain cities with certain other cities; road, railroad or airline
FIGURE 3
~327~
• •
•
(a) (b) (c)
FIGURE 4
networks, and so on. In these examples we have actual physical objects that
are themselves graph-like in nature, with lines, roads, etc. being the edges of
the graph which join certain pairs of nodes-in this case, cities.
We can apply graph theory also to something that does not in itself
resemble a graph, by constructing a graph which summarizes some of its
important properties. Consider, for example, a group of people at a party.
We might construct a graph by representing each person by a node, and
joining two nodes by an edge if the two persons in question are acquainted.
The result will be a graph something like those shown in Figure 4. We can
tell a lot from such a graph; we can tell, for example, that the party of
Figure 4a will be deadly dull; the party of Figure 4b will perhaps be good,
but "cliquey"; and the party in Figure 4c will probably be wild!
We may note here that it may not be possible to draw a graph on a
plane surface without making some of the edges cross each other; the graph
of Figure 3 is a case in point. I t is for this reason that we have drawn the
nodes as small circles, so that, for example, the point where the edges AB
and CD cross, will not be mistaken for a node of the graph. Some graphs
can be drawn in the plane without any such accidental crossings. When
they are drawn in this fashion they are called plane graphs, and we shall
meet them again later on.
As an example of a practical problem in graph theory, consider a
saboteur, or an enemy agent, who wishes to disrupt a complicated
telephone network. He would like to know the minimum number of
telephone lines to cut, or the minimum number of exchanges to blow up, in
order to sever all communication between certain key points. This is a
problem concerning the "connectivity" of graphs. We may wish to extend
the concept of a graph by attaching a number to each edge, in which case
we obtain what is known as a network. An example is given in Figure 5.
These numbers could be the distances between points, or the cost of
transporting commodities from one depot to another, or any of a host of
other possibilities. So wide-spread are structures and situations that can be
represented by networks, that a large part of the important and growing
field of Operations Research is concerned with networks and their
properties.
~328~
FIGURE 5
For this reason almost all research in graph theory contains the seeds
of possible practical application, and is at least potentially useful. There
are some exceptions, however. Tucked away in a dark corner of the graph
theory edifice, with eyes carefully shielded from the light of the real world,
are some graph theorists busy solving problems which are not even
remotely practical. These are the devotees of graphical enumeration, a
subject which is concerned with answers to questions of the form "How
many different graphs are there of such and such a kind?" These graphical
enumerators spend their time either in counting graphs, or in devising
better and better ways of counting graphs. A strange occupation, but one
that is not without interest.
Who are these "Graph theorists who count" and what sort of things
do they count? By the end of this article I hope to have given at least a
rough answer to these questions.
Cayley and the

EnUDleration of Trees
The pioneer of graphical enumeration was Sir Arthur Cayley, who, from
1857 onwards, stated and solved many problems concerning the enu-
meration of those graphs called trees. Graph theorists define a tree to be a
connected graph (one that is all in one piece) which has no circuits. To
understand what a circuit is, think of a graph as a network of roads; a
circuit is then a walk on these roads that comes back to its starting node
and does not pass through any junction (node) more than once. If the
condition about returning to the starting node is dropped, we have what
graph theorists call a "path" between two nodes. Another way of defining
a tree is that it is a graph in which there is a unique path between any two
given nodes. The graph of Figure 3 is not a tree because it has the circuit
ABCDEFGA, among others; Figure 6 shows a graph that is a tree, and one
can easily see why these graphs are so called.
In a tree the number of edges is always one less than the number of
nodes (Problem No.1; Prove this.) Therefore the basic question of tree
enumeration is "How many trees are there with n nodes?" This is one of
the questions which Cayley solved.
~329~
FIGURE 6
In graphical enumeration problems it is very important to be clear as

to what exactly is being counted. When we say "How many different
graphs of such and such a kind are there?," we need to know in particular
what we mean by two graphs being the same or being different, and this may
vary from one problem to another. Usually we regard two graphs as the
same (or to use the technical term, isomorphic) if it is possible to label the
nodes of each with the integers 1,2, ... ,p so that for any two labels, the
nodes with those labels are either joined in both graphs or not joined in
either.
Thus the two graphs of Figure 7, which at first sight seem to be very
different, are in fact isomorphic. The figure shows one way oflabelling the
two graphs so as to show this. (Problem No.2: Figure 8 shows four graphs
on lO nodes. Three of them are isomorphic, while the other is different
from them. Which is the odd man out?)
Thus when we ask "How many different graphs are there?" we must
say what we mean by "different." Frequently it will mean "non-
isomorphic" but not necessarily; we may have something else in mind.
12
10 4
7
10
FIGURE 7
~330~
(a)
(b)
FIGURE 8
Counting Binary Trees

We shall now look in detail at a fairly simple problem, just to get a taste of
the kinds of methods that are used in graphical enumeration. We shall
enumerate binary trees. A binary tree is, first, a rooted tree. This means that
one of its nodes, called the root, is distinguished in some way from the
others. In drawing a diagram of a rooted tree, the root can be indicated by
circling it (as in Figure 6) or by drawing the tree so that the root is at the
bottom of the page. Next, at each node of the binary tree, there are, at
most, two edges going upwards, that is, away from the root. If there are
two, then one goes to the left and one to the right. If there is only one, it
may go to the left or to the right; these two possibilities are considered to be
different. If there is none, then the tree continues no further from that
node, and the node in question is an endnode or leaf of the tree. Figure 9
shows all the binary trees on 4 nodes. Note that because the distinction
between left and right is important-as part of the definition of binary
tree-all these trees are different. Had we not made this distinction, the
first two trees would have been isomorphic, as would many other pairs of
trees in the figure.
.a331~
FIGURE 9
To start out the enumeration of binary trees we note that from the
root of any given tree having, let us say n + I nodes, there will be two
upward branches-one to the left and one to the right. These are defined
as two subtrees-a left subtree and a right subtree, as shown in Figure 10.
Let Bn denote the number of binary trees with n nodes, and let us pretend
that we know the values of the numbers B 1 ,B 2 ••• Bn' To find the number
Bn+ 1 we consider how we could construct binary trees on n + I nodes. This
amounts to choosing two subtrees (having n nodes between them) as the left
and right subtrees, and joining their roots to a new node-the root of the
tree being constructed. In how many ways can we do this? If the left and
right subtrees have I nodes and r nodes respectively, then we can make our
choices in B( Br ways; for anyone of the BI trees on I nodes can be taken
with anyone of the Br trees on r nodes.
Ifwe repeat this with different values of I and r (remembering that I
+ r = n) and add up the results, we shall obtain the number of binary trees
with n + 1 nodes. But hold on! We have assumed that there are necessarily
two edges going upwards from the root, and therefore exactly two subtrees
every time; but there may be only one. Luckily we can avoid this difficulty
by means of a little trick. If the tree being counted has only one subtree,
we shall pretend that it really has two but that the other is a sort of
"phantom tree" -one that has no nodes at all!
By means of this stratagem we can say that every binary tree gives rise
to exactly two subtrees. Since there is only one such "phantom" tree we
take Bo = I, and allow 0 as a possible value for I and r. In this way we
arrive at the following result
which enables us to calculate the number Bn+ 1 if the numbers of trees up to

Bn are already known.
The reader may care to verify that these numbers begin as follows:
~332~
Root
FIGURE 10
n 0 1 2 3 4 5 6 7 8 9 10
B 1 1 2 5 14 42 132 429 1430 4862 16796
These numbers, the B n , are the well-known Catalan numbers, which

crop up in a large variety of combinatorial problems. Besides being the
number of binary trees on n nodes, Bn is also the number of ways of
dissecting an n-gon into triangles, the number of ways ofjoining in pairs 2n
points on the circumference of a circle by nonintersecting chords, and the
number of ways of properly nesting n pairs of parentheses. An interesting
account of these numbers and their properties was given by Martin
Gardner in the June 1976 issue ()f Scientific American.
It is a common and useful practice in enumeration to give the answer
to a problem in the form of a "generating function." To derive a
generating function from a sequence such as the Catalan numbers {Bn} we
make these numbers the coefficients in an infinite power series. The
generating function for the Catalan numbers is thus
B(x) = Bo + B1x + B2X2 + B3X3 + B4X4 + ....
Now the right-hand side of 1 is easily seen to be the coefficient of xn in the
product of B(x) with itself, and therefore the coefficient of xn+ 1 in x[B(x)F.
The left-hand side ofl is the coefficient ofxn + 1 in B(x). Hence we derive
the following result
2 B(x) = 1 + x[B(x)F
where the "1" is added to the right-hand side to take account of the
phantom tree, to which equation 1 does not apply.
Equation 2 is a quadratic equation in B(x). Solving it we find that
1 - )1 - 4x
B(x) = 2x
~333~
and from this, by a tedious but not difficult bit of algebraic manipulation,
we can get an explicit expression for the general Catalan number. It is
B = (2n)!
n (n+ I)! n!
The equation above illustrates a common technique for enumerating
rooted trees. Remove the root and look at the subtrees that we get. There
are many things that may happen when we do this, for there are many
different kinds of trees waiting to be counted. We may get more than two
subtrees-perhaps a variable number even; the subtrees may be ordered
(as with the distinction between left and right that we just made) or they
may not. Each subtree will, however, be rooted (at the node that was
joined to the original root), and provided they are of the same type as the
original tree, we may expect to be able to derive some kind of recurrence
for the required numbers. Almost all tree enumeration has followed this
kind of pattern.
Cayley enumerated, in particular, what we can call "ordinary"
rooted trees, for which there are no restrictions on how many edges meet at
a node, and where there is no importance attached to the order of these
edges or how the tree is depicted on the plane page. He showed that, if Tn is
the number of such trees on n nodes, then the following rather curious
equation holds
T1x+ T2x2+ T3x3 + ... =x(l-X)-Tl (l_x 2 )-TZ (l_x 3)-T3 ....
As before, if we know the values of Tn up to some given value of n,
then, by plugging these values into the right hand side of this equation we
can extract the next number in the sequence from the left-hand side.
Cayley also enumerated unrooted trees. This is a more difficult
problem because with no root at which to pull the graph apart (so to speak)
it is not as easy to express the problem in terms of smaller problems, as we
did above. Cayley also enumerated the numbers of different chemical
compounds of certain types, thus opening up one of the few regions of
enumerative graph theory that is possibly of practical use. The formulae for
these compounds (the alkanes-or paraffins-and some of their de-
rivatives) are essentially trees of rather special types.
The EnUDleration of Graphs

For graphs there is no relation between the numbers of nodes and edges as
there is for trees; a graph with p nodes can have any number of edges from 0
in the case of the "empty graph" up to p(p - 1) /2 for the "complete" graph
in which every node is joined to every other node. Thus the question to ask
about graphs is "How many different graphs are there with p nodes and q
edges" where p and q are given numbers .
.8334~
This is a problem which Cayley did not attempt to solve. Had he

tried, he would probably not have succeeded, since the tool for its solution
had not then been forged. This tool is a theorem which G. P6lya published
in 1938 (see [9]) though it is implicit in some of his earlier papers. P6lya's
theorem is fundamental to the solution of many combinatorial problems,
and although we cannot discuss it in detail, it will be useful to have a
general idea of its scope and its usefulness in solving graphical enumeration
problems.
Polya's Theorem.
Quite a number of combinatorial problems can be formulated in terms of
the idea of "putting things in boxes." Let us suppose that we have a
number of distinguishable boxes, and that in each box we may put an
object-usually referred to as afigure. Moreover, suppose that each figure
has associated with it, a nonnegative integer that is its value. For each box
there is a collection of figures, one of which is to be placed in the box. (It
frequently happens that we have the same range of choices for each box,
but this is not required). In each box we put exactly one of the figures
appropriate to the box, and in this way we obtain a configuration-the boxes
and the figures that they contain. The value of a configuration is then
defined to be the sum of the values of the figures that have been placed in
the boxes.
The first question we may ask is "Given all the relevant information
concerning which figures may be placed in which boxes, in how many ways
can we produce configurations having a given value?" As an example of a
problem of this type, suppose we have the task of placing 21 candles on a
rectangular birthday cake, the top having been divided into six sections as
shown in Figure II. We shall suppose that we may place any number of
candles in each of the sections and that the precise location of the candles
within a section is not important; all that matters is how many candles are
there. In this example a figure consists of a certain number (possibly zero)
of candles; and the value of a figure is the number of candles. The problem
then is to find the number of configurations that have the value 21.
It turns out that this is not a particularly difficult problem; so we shall
immediately make it harder. We now suppose that the boxes are not
••• • • •
• • • •
•• • • • • •
• • • •
FIGURE II
~335~
• • • •
•
FIGURE 12 FIGURE 13
entirely distinguishable, and that certain rearrangements of the boxes will

not be distinguishable from others.
For example, if, instead of the rectangular birthday cake, we have a
circular one, divided into six identical sectors, then it would not be realistic
to regard the two configurations of Figures 12 and 13 as different, since a
clockwise rotation through 60 degrees will convert the first into the second.
Two configurations that differ only in this sort of way will be regarded, for
the purposes of this problem, as being the same configuration. In the
general setting we suppose that we have a certain set of permutations
(which will form a group) and that any two configurations which can be
obtained, one from the other, by permuting the boxes (by some per-
mutation in this group) will be regarded as equivalent. When we ask for the
number of different configurations we are therefore asking for the number
of nonequivalent configurations. For the circular birthday cake the group
consists of the permutations of the boxes that correspond to rotations of the
cake through multiples of60 degrees. Note that the cake shown in Figure 14
is different from those in Figure 13, since no rotation through any angle
will make it look the same as them.
In this article we shall not go into P6lya's theorem in detail; indeed
we shall not even state it. We merely note that it provides a way of solving
• •
•
•
FIGURE 14
a336~
any problem of the general form just outlined. Given the data for such a
problem, (namely the information on what figures may be placed in the
boxes and their values) and given the group of permutations which
determines whether two configurations are equivalent or not, we can apply
P61ya's theorem and find the answer to the problem. Or rather the answers;
for P61ya's theorem gives us a generating function which contains the
number of configurations for any given value.
How can we reformulate the graph enumeration problem in terms of
putting things in boxes? We do it as follows. Consider graphs on p nodes,
labelled 1,2,3, ... ,p. There are p(p - 1) /2 pairs of nodes, and each of these
pairs is a possible site for an edge of the graph. Think of these pairs as the
"boxes" for the problem, and allow just two figures for each box, namely
the figure "no-edge" with value 0, and the figure "edge" with value I. The
result of putting one of these figures in each box corresponds, in an obvious
way, to a graph-nodes i andj are joined if the box (i,j) contains the figure
"edge" -and the value of a configuration (graph) is just the number of
edges in it. Graphs constructed in this way will, however, be labelled
graphs, since each node has its distinctive label from the set {1,2, ... ,p}.
This is not what we want. If two graphs are isomorphic we do not want to
count them as distinct just because they happen to be differently labelled.
So we must allow for the fact that since the nodes are not all distinct,
neither are the pairs of nodes-the boxes. Permuting the labels on the
nodes automatically permutes the boxes among themselves. Thus if we
relabel the nodes 1 and 2 so as to become nodes 5 and 9, the (1, 2)-box
becomes the (5,9)-box, and so on. Every permutation of the nodes gives
rise to a permutation of the boxes, and it is the set of these permutations of
the boxes that forms the group appropriate to this problem. By using
P61ya's theorem the answer can then be obtained.
It was in this way that Frank Harary of the University of Michigan,
now one of the most well-known people in the graph theory world,
enumerated graphs in 1955. The generating function which comes out as
the solution to this problem is quite complicated-fearsome even-but
can be used to compute the numbers of graphs, especially with the aid of an
electronic computer. C. A. King and E. M. Palmer have calculated the
total numbers of graphs up to p = 24. The number of nonisomorphic
graphs on 24 nodes is the following more-than-astronomica1 number:
195704906302078447922174862416726256004122075267063365754368.
(How's that for a really useful piece of information!)
The same technique will succeed in enumerating many variations on
graphs, of which the most important is that of "directed graphs," or (as
now usually abbreviated) digraphs, which are like graphs except that each
edge is given a direction (indicated by an arrow) going one way or the
other, or possibly both ways. Figure 15 shows a typical digraph. The
enumeration problem for digraphs differs only in details from that for
~337~
FIGURE 15
ordinary graphs and P6lya's theorem again gives the answer. This problem
was also solved by Harary, in the same paper [5] in which he enumerated
graphs.
Although P6lya's theorem is very powerful, it will not handle every
problem of graphical enumeration. A problem that is proof against it is the
enumeration of self-complementary graphs. The "complement" ofa graph
G is defined to be the graph G with the same set of nodes as G, having all the
edges that are not in G and none of the edges that are in G. Figure 16 shows
a graph (black edges) and its complement (grey edges) superimposed on
the same set of nodes so as to display how a graph and its complement have,
between them, all the p(p - 1) /2 edges that are possible in a graph with p
nodes.
Now it may happen that the complement of a graph is isomorphic to
the original graph. The simplest example of such a "self-complementary"
graph has 4 nodes and 3 edges and is shown in Figure 17. With 5 nodes
there are two different self-complementary graphs, shown in Figure 18.
These small examples are easily found by trial and error, but by the time
we get up to 8 nodes trial and error is no longer a reliable way of finding
out how many self-complementary graphs there are. (Problem No.3: How
many self-complementary graphs are there with 6 and 7 nodes?) What we
FIGURE 16 FIGURE 17
~338~
FIGURE 18
need is a theoretical result giving the number of these graphs for any value
ofp.
This problem requires something more powerful than P6lya's theo-
rem because it has an added complication. A graph and its complement
differ to the extent that we have interchanged the two figures "edge" and
"no-edge" in going from one to the other. Thus we have a different kind of
general problem-one in which we allow permutations ofthefigures as well
as permutations of the boxes to decide whether two configurations are
equivalent or not. We therefore have two permutation groups-one for the
boxes and one for the figures-and everything becomes much more
complicated. A method of coping with this more general type of problem
was first given by the Dutch mathematician N. G. deBruijn in a paper [1]
published in 1959, and it was not long before the enumeration of self-
complementary graphs was achieved (see [10]). Complementation for
digraphs is defined analogously to the corresponding graph concept.
Figure 19 shows a digraph that is isomorphic to its complement, and the
enmperation of such self-complementary digraphs came out along with
that of self-complementary graphs.
One very curious fact that emerged from these results is that the
number of self-complementary digraphs on an even number of nodes is
always the same as the number of self-complementary graphs on twice the
number of nodes! Such a direct relation suggests (intuitively, but very
strongly) that there ought to be some simple connection between the
FIGURE 19
~339~
graphs and the digraphs; that there should be some easy way of "seeing"
that the numbers must be the same, and why. Maybe there is; but ifso, no
one has yet discovered it.
As time went on, more powerful enumeration theorems were found-
theorems that generalized those of P61ya and de Bruijn, or which struck
out in quite different directions. Among the most important of these is the
"Power Group Enumeration Theorem" ofHarary and Palmer; but it is too
complex to be described here. The reader can find further information in
the book Graphical Enumeration [7] by these two authors-a compre-
hensive treatise on this specialized branch of graph theory.
The Strange Case of

J. H. Redfield
We should say something now about an unusual character who appeared
on the stage of graphical enumeration. Perhaps it would be more correct to
say "who reappeared" since ... but we shall see. In the early 1960's Frank
Harary drew the attention of his fellow enumerators to a remarkable paper
[11] called to his attention by one of his grad students. It was entitled "The
theory of group-reduced distributions" and had been published by a
hitherto unknown mathematician, J. H. Redfield, in 1927. In this paper
(as Harary discovered, and as the rest of us verified with astonishment)
Redfield had given a theorem equivalent to P61ya's theorem; he had
carried out the enumeration of graphs (at least in part-and it was clear
that he could have gone further); he had anticipated several other "recent"
discoveries in enumeration theory; and, to crown it all, he had proved a
powerful new theorem. I remember well this "rediscovery" of Redfield's
paper since one of the results that he had anticipated was a theorem
(concerning the enumeration of graphs obtained by superposing several
graphs on the same set of nodes) which had been the mainstay of my own
Ph.D. thesis a few years earlier! It seemed that for 30 years or so, graphical
enumerators, and others, had been painfully solving problems and proving
theorems that Redfield had known about and had published way back in
1927! And yet his paper had completely escaped notice. How was this
possible?
It was not that it had been published in an obscure journal; the
American Journal rif Mathematics, in which it appeared, is a well-known and
highly respected publication. Perhaps it was rather that several factors that
normally succeed in drawing attention to a paper, just happened to be
absent. Redfield was not a known worker in this field - in fact this was his
only mathematical paper; the title of the paper gave no inkling of the
nature of its contents; and Redfield's notation, including some rather
bizarre symbols, was such that even the printed page did not give an
immediate indication of what the paper was about. Above all, perhaps,
~340~
was a contributing factor best described by saying that "the time was not
ripe. "
Redfield was not primarily a mathematician-his field was, or had
been, linguistics. Nevertheless his one and only mathematical paper is a
remarkable piece of work, and had it been recognized earlier for what it
was, the course of enumerative graph theory would undoubtedly have been
very different. Certainly, since its rediscovery, Redfield's paper has
inspired several new advances. In particular R. W. Robinson of the
University of Newcastle in Australia has adapted and improved some of
Redfield's ideas, and hence solved several problems that were previously
considered intractable.
The Enumeration of
Plane Graphs
We have already seen what is meant by a plane graph; it is one that has
been drawn in the plane without any edges crossing. Let us look at some
enumeration problems pertaining to them. A natural question to ask is
"How many plane graphs are there with p nodes and q edges?," but this is
much too difficult. Let us consider "cubical plane graphs," by which we
mean graphs in which exactly 3 edges come together at each node. In a
cubical graph the numbers p and q of nodes and edges are related by
2q = 3p (Question 4: Why is this?), so that a single number determines the
size of the graph-the number of nodes (which must be even). We might
expect that this would make the problem easier, but it seems to be just as
hard. This is typical of plane graphs. The enumeration problems are
usually fiendishly difficult, however, when a problem can be solved, it often
happens that the solution is simpler than graphs are in general.
We can see this very well if we further restrict the problem, and ask
"How many plane cubical graphs are there with a fixed Hamiltonian
circuit?" A Hamiltonian circuit in a graph is a circuit-a "Grand Tour"
one might say-that visits every node of the graph. By being "fixed" we
mean that the Hamiltonian circuit has been drawn once and for all in the
plane, and only the remaining edges may be altered in position to form the
several graphs that we are counting. We are therefore looking at objects
like the one in Figure 20, in which the edges not in the Hamiltonian circuit
are drawn in grey. Now the grey edges lie either entirely inside or entirely
outside the circuit; for if they crossed it the graph would not be plane.
Further, each node is an end of exactly one thick edge. We can therefore
rephrase the enumeration problem for those graphs as follows: "Given a
polygon with 2n vertices, in how many ways can we join some pairs of
vertices by nonintersecting chords lying inside the polygon, and join the
remaining vertices in pairs by non-intersecting 'chords' (for want of a
better word) lying outside the polygon?"
~341&.
FIGURE 20
This problem was solved in a most elegant manner by W. T. Tutte,

Distinguished Professor at the University of Waterloo, who has made an
extensive study of plane graphs, and who is responsible for the bulk of the
work that has been done on their enumeration. Let us briefly look at his
result and its proof. (For full details see [12].)
First note that if we want all the chords to be on the inside of the
polygon we have a simpler problem. In fact, it is one of the many problems
for which the answer is the sequence of Catalan numbers, as already
mentioned. Thus if we stipulate that a certain set of 2r nodes shall be the
ends of the "internal" chords, then the number of ways of joining up these
nodes in pairs is the Catalan number
(2r)!
r!(r + I)!
A similar result holds for the remaining 2s nodes (where r + s = n) and the
"external" chords.
Now any way of disposing the internal chords can be taken with any
way of disposing the external chords, so we must multiply these two
numbers together. But we must also multiply by the number of ways of
choosing the two sets of nodes, and this, being just the number of ways of
choosing the 2r nodes in the first set from the total of 2n nodes, is the
binomial coefficient
2n) (2n)!
(
2r = (2r)!(2s)!'
Adding together these products for the various values of T and s we
obtain
L (2n)!. (2r)! • (2s)! .
r+s=n (2r)!(2s)! r!(r + I)! s!(s + I)! '
that is,
..a342~
L (2n)!
ds=n r!(r + I)! S!(S + I)!
A few simple manipulations now reduce this expression to the final,
very simple, formula
(2n)! (2n + 2)!
n!(n + 1)!2(n + 2)!
for the required number of graphs.
When n = 2 this formula gives the value 10. It may seem strange that
there should be so many possibilities for graphs on only 4 nodes, but we can
see from Figure 21 why this is so. The circuit is a square, and since we are
not allowed to rotate it, graphs 1 and 2 in the figure are counted as
different. Moreover, although graphs 3 and 5 are isomorphic (even taking
account of the colours of the edges) they are regarded as different here
because they are drawn differently in the plane.
Note also that we have allowed the possibility of two edges between
two nodes. Such "double edges" are sometimes allowed, sometimes not,
according to the nature of the problem. It would seem to be more
reasonable to stipulate here that double edges are not to occur; but if we do
that, we come across an example of the capriciousness of enumeration
problems. For this apparently minor change gives us a much more difficult
problem -one that, as far as I know, is still unsolved.
There are many unsolved problems in enumerative graph theory.
Some of these hover on the borders of possibility, seeming to need only the
right trick or the right generalization of existing theorems and techniques
to force them into submission. Others show all the signs of being stubbornly
recalcitrant, and resistant to all known forms of attack. Either way they
offer a fascinating challenge to one's mathematical ingenuity, and the
prospect of many hours of good clean fun-that is if you like that sort of
thing!
@~QOD
00000 FIGURE 21
9
~343&.
Solutions to the Problem.s

PROBLEM I The number of edges in a tree is one less than the
number of nodes.
We can draw a tree by starting with a single node and
adding edges one by one in such a way that each new edge joins
a new node to a node already drawn. At each stage we increase
by I the number of nodes and the number of edges. Hence, since
the result is true when we first start (with 0 edges and 1 node) it
is true when we finish drawing the tree.
PROBLEM 2 The graphs in Figure 9.

Graph (c) is the odd man out. The fact that the others are
isomorphic is seen from the labelling given in Figure 21. Graph
(c) contains a circuit of 4 nodes (ABeD), whereas the other
graphs have no circuit ofless than 5 nodes. This shows that (c)
cannot be the same as the other graphs.
PROBLEM 3 Self-complementary graphs on 6 or 7 nodes.

There are none. A graph and its complement have
between them all the p(p - I) /2 possible edges. Moreover, since
they are isomorphic they each have the same number of edges.
It follows that p(p - I) /2 must be an even number, and this is
5 2
10~--~-+~~--~8
10
FIGURE 22
~344~
not the case for p = 6 or 7. More generally, self-complementary

graphs can exist only when p(p - I) /4 is an integer, and this
happens only if p leaves remainder 0 or I when divided by 4.
PROBLEM 4 The equation 2q = 3p for cubical graphs.

Since each edge has 2 "ends" there are 2q ends in the
graph as a whole. Each node accounts for exactly 3 ends, from
which the result follows.
Bibliography
This bibliography includes some papers of general interest besides those
specifically referred to in the text.
1 deBruijn, N. G. Generalization of Palya's fundamental theorem in
enumerative combinatorial analysis. Indag. Math. 21: 59-69.
2 Cayley, A. 1857. On the theory of the analytical forms called trees. Phil.
Mag. 13: 172-176.
3 Cayley, A. 1874. On the mathematical theory of isomers. Phil. Mag. 47:
444-446.
4 Cayley, A. 1875. On the analytical forms called trees, with applications
to the theory of chemical compounds. Rep. Brit. Ass. 257-305.
5 Harary, F. 1955. The number oflinear, directed, rooted and connected
graphs. Trans. Amer. Math. Soc. 78: 445-463.
6 Harary, F. 1969. Graph Theory. Reading, Massachusetts: Addison-
Wesley.
7 Harary, F. and Palmer, E. M. 1973. Graphical enumeration. New
York: Academic Press.
8 King, C. A. and Palmer, E. M. Calculation of the number of graphs of
order p = 1(l)24. (unpublished.)
9 Palya, G. 1937. Kombinatorische Anzahlbestimmungen fUr Gruppen,
Graphen und chemische Verbindungen. Acta Math. 68: 145-254.
10 Read, R. C. 1963. On the number of self-complementary graphs and
digraphs. ]. London Math. Soc. 38: 99-lO4.
11 Redfield, J. H. 1927. The theory of group-reduced distributions. Amer.
]. Math. 49: 433-455.
12 Tutte, W. T. 1976. Hamiltonian circuits. Colloquio Internazionale
sulle Teorie Combinatorie. Atti de Convegni Lincei. 17: 194-199.
~345(2s.
Error-Correcting
Codes and
Cryptography
-~-
N. J. A. Sloane
BELL LABoRATORIES
This paper is intended to serve as an introduction to the excItmg

developments in secret codes that have taken place in the last ten years.
David Kahn's interesting book The Codebreakers appeared in 1967 [29],
which unfortunately was just before IBM described its Lucifer encryption
scheme [11], [20], [51] and triggered the developments that I am going to
describe.
The mathematical articles published in cryptography before 1967
were pretty dull (some say they were deliberately made dull, to discourage
research in this area). But by the end of this paper I think you will agree
that the new developments are very exciting indeed. These seem to me to
be some of the most fascinating ideas in communication theory to come
along in a long time. (Since most of this is not my own work I don't have to
apologize for being so enthusiastic about it.) In order to avoid interlacing
the text with superlatives let me say here once and for all that much of what
follows is based on six classic, brilliant, elegant papers:
Shannon (1949, [49]),
Feistel, Notz and Smith (1975, [13]),
~346~
ERROR-CORRECTING CODES AND CRYPTOGRAPHY
Wyner (1975, [56]),

Diffie and Hellman (1976, [5]),
Rivest, Shamir and Adelman (1978, [44]),
Merkle and Hellman (1978, [38]).
There will be nothing new here, except perhaps my way of looking at
things. On the other hand I am going to mention so many different kinds of
encryption schemes that most readers will find something that is
unfamiliar.
By the way, I should like to stress that I am an unclassified cryptog-
rapher: I know nothing about the classified work on this subject.
Fortunately those readers who do are under oath not to talk about it, and
will not be able to correct me if what I say is inaccurate. All the encryption
schemes I am going to describe have been published in the open literature.
Nevertheless the opinions expressed here are my own, not of the authors I
have referenced nor of the Bell System.
There are five parts to the article. The common theme is the problem
of transmitting information down a wire (see Figure 1). Let us assume that
Send 0 or 1 Receive 0 or 1
Noise
•
FIGURE I
we can send dots and dashes down the wire, or, to be more mathematical,
O's and l's. When we send a 0 usually a 0 comes out at the far end, but
sometimes (because of noise) a I is received. Conversely a 1 is usually
received as ai, occasionally as a o. This wire is sometimes called a binary
symmetric channel, and is described by the diagram ,shown in Figure 2.
99/100
99/100
FIGURE 2
A binary symmetric channel. When a 0 is transmitted, 99 times out of
a hundred a 0 is received, but I time in a hundred a I is received-
similarly if a I is transmitted. Thus the probability of error on this
channel is 1/100.
~347~
The data to be transmitted has already been converted into a string of

O's and l's (perhaps it is being produced by a computer). Our problem is
the following.
SEND AS MUCH INFORMATION AS

POSSIBLE DOWN THE WIRE, AS
QUICKLY AND AS RELIABLY AS
POSSIBLE. IT SHOULD ALSO BE
PROTECTED AGAINST EAVES-
DROPPERS.
I will describe five different kinds of solutions to this problem, under

progressively more and more difficult conditions (see Figure 3). The first
case (see Section 1) deals with the situation when the channel is simply that
shown in Figure 1, and makes use of error-correcting codes. In the next
part we have the same noisy channel, but now a second and suspicious-
looking wire has appeared (see also Figure 10 below): there is an
Error-correcting codes
Noise
~den er • IReceiver!
II Wire-tap channel
~f--------4I~~--N-01!~;~.ec~.e~iv~e~r~Bad guyl
III Conventional cryptography Key
Key
IV Codes which detect deception

Key
Bad guy
V Public key cryptosystem

No key!
Sender
FIGURE 3
The five different communication systems considered in this paper.
~348&
eavesdropper! To begin with we assume the eavesdropper is working with

poor quality equipment, and we will show how to defeat him easily
(Section 2). The third part deals with the case when we are up against a
well-equipped spy with the best equipment. This is the situation which
conventional cryptography deals with, by using a secret key which is known
to the sender and the receiver but not to the bad guy (Section 3). The
fourth part is an extreme situation. Now we actually give the message to
the bad guy to transmit for us. Yet we can still defeat him (Section 4). The
final part describes the so-called public-key cryptosystems: these are ways
of modifying the set-up of Section 3 so that the sender does not need to
know the key (Section 5). In all cases we are able to defeat the bad guy and
send our messages without him understanding them (or modifying them, in
Section 4).
Section 1
Error-Correcting Codes
The problem is to send information down the channel shown in Figures I
and 2 as quickly and as reliably as possible. The coding theorist's solution is
to allow not just any old sequence ofO's and 1's to be sent, but only certain
special sequences called codewords. The list of all the allowable codewords is
called the code; it is known to both sender and receiver, who are cooperating
in the attempt to eliminate transmission errors.
The messages are represented by the codewords. When a particular
message is to be sent, the corresponding codeword is transmitted. At the
receiver a (possibly distorted) version of the codeword arrives, and the
decoder must decide which codeword was sent and hence what the message
was. Figure 4 shows a simple example of a code: there are just two
codewords, 00000 and 11111, of length 5, corresponding say to the
messages Yes and No. Suppose the message is Yes, the codeword 00000 is
transmitted, and (because of noise) 01000 is received. The decoder knows
that errors are not very likely (see Figure 2) and so should decide that
00000 was transmitted and that the message is Yes. In general the decoder
Message Codeword
,
Yes
~EncoderH Channel
'N .
,00000, olse 01000
Ht '. .~-ec-o-d-e...,r~ Yes
No 11111
Receive
FIGURE 4
A simple error-correcting code in use. There are two codewords, 00000
and 11111. The sequence 01000 is received and is correctly decoded as
Yes.
~349~
should choose that codeword which in some sense is "closest" to the

received sequence.
To make this precise, let us define a distance between sequences.
The Hamming distance, abbreviated dist(u,v), between two sequences
u = (u 1 ,U 2 , ..• , un) and v = (V1'V 2 , ••• , vn) oflength n is the number of places
where they differ. For example
dist (0 1000, 00000) = 1,
dist (0 1000, 11111) =4.
We can now give the decoder more precise instructions: he should
decode a sequence as that codeword which is closest to it in Hamming
distance. This will be the most likely codeword to have been transmitted.
Figure 4 illustrates this procedure, and we see that in this example the
decoder was able to correct one error. In fact this code can even correct
any pair of errors. For example if 00000 is sent and 01100 is received (the
second and third digits are in error), 01100 is still closer to the true
codeword than to 11111. On the other hand it is obvious that this code
cannot correct three errors (if 11100 is received we would decode it
incorrectly as 11111). So this is a double-error-correcting code.
The reason for this is plain: the codewords themselves have Hamming
distance 5 apart. Precisely the same argument shows that for any code, the
most important parameter is the minimum Hamming distance between
codewords:
d = min dist(u,v);
utv
taken over all distinct codewords u and v. For then if not more than e
= [(d -1)/2] errors occur, the received vector will be closer to the true
codeword than to any other, and the decoder will make the correct
decision. (Here [x] denotes the largest integer not exceeding x.)
To sum up, a code with minimum distance d between codewords is an
e = [( d - 1) /2] error-correcting code. We wish therefore to find codes in
which d is large (to correct many errors) and the length n is small (to reduce
transmission time).
The following examples of codes are more interesting than the one in
Figure 4.
1 The even weight code En' consists of all binary words of

length n containing an even number of 1's-that is, having
even weight. Figure 5 shows E 4 . En contains 2n- 1 codewords
and has minimum distance d = 2 between codewords. This
code is of no use for correcting errors, although it is able to
detect if a single error has occurred. We shall meet it again in
Section 2.
08350&
0 0 0 0
0 0 1 1
0 1 0 1
1 0 0 1
0 1 1 0
1 0 1 0
1 1 0 0
1 1 1 1
FIGURE 5
The code E 4 , consisting of all eight vectors of length 4 that contain an
even number of I's.
2 The Hamming code of length 7. This is one of the many

codes with a geometric construction. A projective plane of order
p (a geometric object that will reappear in Section 4) is a
collection of p2 + P+ I points and p2 + P+ I lines, ar-
ranged so that there are p + 1 points on each line and p + 1
lines through each point. Figure 6 shows a projective plane
of order 2.
In order to describe a projective plane to someone else, it is
not necessary to draw a picture. Instead one can just specify
4~------~~~~--------~5
7
FIGURE 6
A projective plane of order 2. There are seven points, labeled
1,2, ... , 7, and seven lines (one of which is curved). Each line contains
three points and each point lies on three lines.
~351~
Points
2 3 4 5 6 7
o 000
00.1 0 0
o 0 o o
Lines 0 0 0 0
000 0
o 000
o 000
FIGURE 7
A matrix of O's and l's which describes the projective plane of
Figure 6. For example the first row of the matrix specifies that there is
a line containing the points I, 2 and 4.
which points lie on each of the lines (this could be done over the
telephone). For example Figure 6 contains the lines
1,2,4
2,3,5
3,4,6
4,5,7
1,5,6
2,6,7
1,3,7
Alternatively we could describe the picture by a matrix, in
which the rows represent lines and the columns represent the
points (see Figure 7).
The Hamming code we want consists of the complements
of the rows of this matrix (obtained by interchanging O's and
1's), together with the zero codeword. It is shown in Figure 8,
has minimum distance d = 4, and is a single-error-correcting
code.
All the codes we have described have a special property which makes
them easier to encode and decode: they are linear codes. That is, the
componentwise sum of two codewords, taken mod 2, is always a codeword.
For example the mod 2 sum of the second and third codewords in Figure 5,
0011
and
0101
~352~
0 0 0 0 0 0 0
0 0 1 0 1 1 1
1 0 0 1 0 1 1
1 1 0 0 1 0 1
1 1 1 0 0 1 0
0 1 1 1 0 0 1
1 0 1 1 1 0 0
0 1 0 1 1 1 0
FIGURE 8
The Hamming code of length 7, containing eight codewords. This is
obtained from the complements of the rows of Figure 7.
IS
0110,
which is indeed in the code.
A linear code can be concisely defined by giving a so-called parity-
check matrix for it. This is a matrix H with the property that a vector u
= (u 1 ,···, un) is in the code if and only if
HuT == 0 (mod 2),
where the T denotes transpose. It would require too much ofa digression to
explain why this is called a parity-check matrix-see [35, Chapter 1]. A
parity-check matrix for the code of Figure 5 is
[1111],
since (u 1,u2'u 3,u4 ) is in the code if and only if
u1 + u2 + u3 + u4 == 0 (mod 2).
For the code of Figure 8 we could use
1101000]
[ 01lO100
0011OlO
0001lOi
as a parity-check matrix.
In the thirty years or so that error-correcting codes have been in
existence a large number of good examples have been discovered, and
~353~
there is by now an extensive body oftheory. In fact, F.J. MacWilliams and

I have just written a 760-page book on the subject [35]. Many of these
codes are defined by an ingeniously chosen parity-check matrix. One
particularly powerful family are the Goppa codes. Unfortunately their
description requires some knowledge of finite fields, so the following
paragraph is optional reading.
3 Goppa codes of length n = 2m. Choose an irreducible poly-

nomial G(z) of degree t over the Galois field GF(2m).
Construct the parity check matrix
I
-- ... -I-
G( IXd G( IXn)
--
IX 1
... -IXn-
H= (I)
where 1X1,1X2' ••• , IXn are the elements ofGF(2 m). The codewords are all the
binary vectors u such that HUT = o. The properties of this code are
summarized in Figure 9.
Goppa codes (like BCH, Reed-Muller, and other codes we have not
mentioned) have an efficient decoding algorithm-there is an algebraic
process for taking the received sequence and finding the closest codeword.
this is certainly not the case for most codes, a fact which underlies one of
the public-key cryptosystems described in Section 5e. For further infor-
mation about error-correcting codes the reader is referred to [35].
Prop'erties of GOPP'a Code
Length n = 2 m
Number of codewords is at least 2n - ml
Minimum distance d ~ 21 + 1
This is a I-error-correcting code
FIGURE 9
~354&.
Send 0 or 1 Noise Receive 0 or 1
FIGURE 10
A wire-tap appears!
Section 2
The Wire-Tap Channel
The communication channel now changes from that shown in Figure 1 to
this (Figure 10).
The second wire is a tap leading to the Bad Guy, who is listening to
everything that is said. The problem has changed: the goal is now to send
information down the channel as quickly and as reliably as possible, and
simultaneously to minimize what the wire-tapper learns.
In this section we consider the case of a low-budget wire-tapper, using
imperfect equipment. That is, we assume that there is noise on the wire tap
itself-the eavesdropper is listening over a binary symmetric channel like
that in Figure 2. The main reference for this section is Wyner [56]; see also
[1], [31], [32], [55]. The simplest case is when there is no noise on the direct
wire-see Figure 11.
There is a beautiful and surprisingly simple solution to this problem.
The key idea is the following.
ENCRYPT 0 AS A LONG, RANDOM-

L Y CHOSEN STRING OF O'S AND 1'S
WITH AN EVEN NUMBER OF 1'So
ENCRYPT 1 AS A LONG, RANDOM-
LY CHOSEN STRING OF O'S AND 1'S
WITH AN ODD NUMBER OF 1'So
The length of the string is made great enough so that there are likely to be
several errors introduced by the wire leading to the eavesdropper. For
example, we might encrypt 0 as
r= 101110100000011111010100
~355&.
o or 1
,..-----,
o or 1
,..------,
FIGURE 11
The simplest version of the wire-tap channel. There is no noise on the
main wire, but the wire tap is a binary symmetric channel (see
Figure 2) with a certain error probability Po.
(a sequence oflength 24 containing an even number of 1's), but by the time

this reaches the eavesdropper this sequence might have been changed to
Z= 100110100110011111010100.
Even though the eavesdropper knows the encryption rule (but not of course
r
which sequence has been chosen) it is clear that he is completely baffled.
He knows that Z is a corrupted version of a sequence r, but has no way of
r
telling whether contained an even or an odd number of 1'so
On the other hand the legitimate receiver sees r
exactly as it was
transmitted, and has merely to count the 1's in r. He can do this very
simply by forming the sum (mod 2) of all the bits (= binary digits) in r.
I think you will agree this is a wonderfully clever idea. Of course it
has one obvious disadvantage-it is very slow. The rate of transmission of
r
information is almost zero, since we have to send all of to communicate
one bit to the receiver.
But this is easily fixed. First let me describe the scheme in another
way. Let F n denote the set of all binary vectors of length n. We divide F n
into two subsets: the subset En (see Section I) consisting of all vectors with
an even number of 1's, and the remaining vectors (call them Dn) which
contain an odd number of 1'so Thus
(2)
The language of group theory is helpful here. F n is a group (under
componentwise addition) and En is a subgroup (remember we observed in
Section I that it was a linear code). Also D n is a translate or coset of En:
Dn = 100 .. ·0+ En'
and Equation 2 is an illustration of the principle that it is always possible to
partition a group into translates of a subgroup. Our encryption scheme
reads as follows. Partition F n = En U Dn' To send a 0, transmit a random
vector from En' To send a 1, transmit a random vector from Dn.
We can now state the general solution to the wire-tap problem.
Choose a good linear error-correcting code C1 containing 2n -I< codewords
of length n (see Section 1). Partition F n into 2k cosets of Cl' say
~356~
F = C1 U C2 U C3 U ... U C2 k'
Number the possible messages to be sent from 1 to 2k. Then:
ENCRYPT THE i-th MESSAGE AS A

RANDOMLY CHOSEN VECTOR
FROM Ci .
I t is easy for the legitimate receiver to discover which message was

sent, since it is simple to decide to which coset of C1 a vector belongs (one
just computes the syndrome of the vector-see [35, p. 16]). But the wire-
tapper is still baiRed. The transmission rate of this scheme is kin (n bits are
transmitted in order to specify one of 2" messages), and Wyner [56] was
able to prove the following result.
THEOREM 1 If Po is the probability of error on the wire-tap,

then it is possible to transmit at any rate below
while keeping the eavesdropper in ignorance of

what is being said.
The way this theorem is proved is by introducing a function H called

entropy or uncertainry ([16, Chapter 1], [36, Chapter 2]). It is defined in such
a way that if X is the message, r is the encrypted message, and Z is the
distorted version of r heard by the eavesdropper, then H(X) measures the
uncertainty the eavesdropper had about X before hearing Z, and H(X1Z) is
a measure of his uncertainty about X after hearing Z. Of course
H(XIZ) ~ H(X).
To prove the theorem one shows that in fact
H(XIZ) ~ H(X),
H(X I Z) '" H(X)
FIGURE 12
Perfect secrecy is assured if the uncertainty H(X1Z) that the bad guy
has about the message X after learning Z is essentially equal to his
uncertainty H(X) about X bifore installing the wire-tap.
~357~
or in other words that the eavesdropper has essentially learned nothing by

installing the wire-tap (Figure 12). He might just as well have stayed
home.
For the proof itself the reader is referred to [56]. Wyner also analyzes
the more general situation when the direct wire from sender to receiver is
nOIsy.
Section 3
Conventional Cryptography
In the previous section we easily defeated a wire-tapper who was using
inferior equipment. But what if we are up against a professional, whose
equipment is so good that he overhears what is being said without
distortion? This is the situation to which conventional cryptographic
methods apply. To begin, we consider encryption schemes which make use
of a key that is known by the sender and receiver but not by the bad guy
(see Figure 13). Section 5 will describe some recently discovered schemes
in which only the receiver needs to know the key.
FIGURE 13
A conventional encryption scheme which makes use of a key known by
the sender and receiver but not by the bad guy.
3A The One-Time Pad

The simplest and most secure of all schemes is the one-time pad
(Figure 14). (Several names are associated with this invention, especially
that of G. S. Vernam, who at the time was working in American
Telephone and Telegraph's research department [29, Ch. 13], [54].)
A long random string of O's and l's is formed, perhaps by tossing a
coin many times. This is the key, which is punched onto paper tape and a
copy is given to both sender and receiver. The sender adds this string bit-
by-bit to the message to produce the encrypted text. For example if the
message is
···0100001101
.8358&.
... 0011 ... 1101

Message
... 1101
Message
FIGURE 14
The one-time pad. The key is added mod 2 to the message. The key
sequence is as long as the message sequence and is used only once.
and the key string is

···1101110011
then the encrypted message is their sum:
···1001111110
(The sum is formed mod 2, bit-by-bit, with no carries.) At the receiving
end the key string is again added mod 2 to the encrypted message, giving
the original message back again:
Encrypted message ... 1001111110
plus key string ... 110 11100 II
equals original message··· 0 10000 II 0 1.
The key string is used once and then destroyed. This is a perfect,
unbreakable cipher. For let X denote the message sequence, K the key
string, and
Y=X+K (3)
the encrypted message. The bad guy knows Y; but since all the different
key strings K are possible and equally likely, so are all the possible
messages. Again he has learned essentially nothing by installing his wire-
tap (see Figure 12).
The only disadvantage of this scheme is that it requires as much key
as there is data to be sent. Nevertheless it is widely used for important
messages.
However for more routine matters it is desirable to have a scheme
which uses a smaller amount of key. The art of designing a good encryption
scheme is to find a way of expanding the key, that is, of taking a small
amount of key and using it as the seed to produce a much longer key string.
A great many ways of doing this are known (see for example [8], [15], [29]),
but here we shall only mention some techniques using shift registers.
~359~
3B Linear Feedback Shift Registers

The first, simplest and weakest method is to use the output from a linear
feedback shift register as the key string (Figure 15).
If the shift register has m stages it is possible, by appropriate choice of
the feedback connections, to obtain an output sequence of period 2m - I.
(See for example [22] or [34].) It is worth mentioning that these output
sequences are codewords in a certain Reed-Muller error-correcting code
[35, p. 406]. If m is large (for example, 100) this period is astronomical
(2 100 - I ~ 10 30 ).
(a)
Shift register Shift Register
lc;QJo:J
0 1 0 1
... 100011110101
Output sequence
= key string
0 1 0
0 1
0
0 1
0 0
0 0 0
1 0 0 0
0 0 0
FIGURE 15
(a) The output from a linear feedback shift register is used as the key
string. (b) An example of a four-stage shift register showing the
successive contents of the register and the output sequence. The initial
contents of the register (0101) is a secret four-digit key, and produces
the output sequence
... 1110101100 1000 1111 010 1
of period 15 which is added bit-by-bit to the message.
~360~
In spite of this the cipher is still very easy to break. For to test the
strength of a cipher it must be assumed that the bad guy knows the
encryption algorithm (in this case the shift register), and has obtained a
number of pairs
(message X, corresponding encrypted message Y)
-that is, pairs of matching plain text and encrypted text. He is trying to
determine the key string K. In the case of a linear feedback shift register,
this is simple. Once m successive matching bits of the X and r sequences are
known, their difference gives the contents of the shift register and hence (by
letting the shift register run) the complete key string (cf. [18], [39]). In spite
of its weakness this encryption scheme is very popular, perhaps because the
large periods produce an illusion of strength.
3c Nonlinear Shift Registers

However as soon as we allow nonlinear components to be used in the
feedback circuit (Figure 16) the situation changes completely. Nonlinear
shift register sequences can be made as secure as we wish.
The number of possible feedback functions f that can be used is 22m ,
which when m is large is a very large number indeed. We wish to find a
subset of the possible j's which
1 efficiently scramble the message, that is, such that the output
sequence resembles the coin-tossing sequence of Section 3a,
Nonlinear feedback
shift register
Encrypted message
.. .x, + u" XI + UI
Message _ _ _---I
•• •X,x,xl
FIGURE 16
Encrypting by means ofa nonlinear shift register. Only the encryption
circuit is shown, although there is an identical circuit for decryption.
There are m stages in the shift register, initially containing a secret m-
digit key Um ,Um -t, ... ,u2 ,u t . The (m+ 1) - st digit Um + t which enters
from the left is a complicated nonlinear function of Ut, ... , Um , say Um + 1
= 1(U1,· .. , um )· Then Um +2 = 1(U2,···' Um + 1)' and so on .
..a361~
2 are simple to evaluate, and

3 are easily changed (so that by changing the key we can
change to a differentJ).
One way of doing this is to build complicated functions by combining

simple components. Encryption schemes of this type are called product
ciphers; they are'analogous to the classical algorithm for mixing dough:
ROLL IT
FOLD IT
ROLL IT
FOLD IT
ROLL IT
In other words, perform a sequence of non-commuting operations on the

dough! For example we might combine two digital operations such as a
permutation and a simple nonlinear function. A useful component for
constructing nonlinear functions is the two-state read-only memory device
shown in Figure 17.
(a)
4 Input bits ~ 4 Output bits
Control (or key) bit

(bl
Truth table
Output
Input
Key = 0 Key = 1
0000 10 10 01 1 1
0001 001 1 01 00
0010 1 001 1 1 1 1
001 1 0000 101 0
... ....
1 111 01 00 100 1
FIGURE 17
(a) A two-state read-only memory or 2-ROM with four input bits, four
output bits, and one control or key bit. (Widely available as an
inexpensive chip). (b) The output bits are specified as a function ofthe
input bits and the key bit by means of a truth table. (Of course vastly
more complicated components may be used: compare the circuit
shown on page 79 of the July 1979 Scientific American.)
~362~
8 Key
Bits Eight 2-state
read-only
'T--r--r-"'" memory chips
Permutation
~~~~~~~--~~~~
Register
123 4 567 8 29 30 31 32
FIGURE 18
Two non-commuting operations applied to 32 bits of data. The first is
a permutation, while the second is a nonlinear function formed from
eight chips of the type shown in Figure 17, and controlled by eight bits
of key.
To illustrate how these operations might be combined in an en-

cryption scheme, let us consider a shift register with 32 stages (bottom of
Figure 18). The basic operations are a permutation of the 32 bits (middle
of Figure 18) and a nonlinear function built out of eight different two-state
devices of the kind shown in Figure 17 (top of Figure 18).
These operations might be successively applied say 15 times, the
resulting 32 bits fed back into the register (Figure 19), and simultaneously
used as 32 bits of the key string and added to 32 message bits (Figure 20).
An alternative mode of operation is to first add the 32 message bits to
the contents of the shift register, then apply the fifteen permutations and
nonlinear functions, and finally take the resulting 32 bits as the encrypted
text.
Product ciphers and their analogy with mixing dough were perhaps
first described by Shannon [49]. The implementations using permutations
and nonlinear circuits were developed by several people at IBM -see for
example the papers by Feistel [11], [12], Feistel, Notz and Smith [13], [14],
Girsdansky [20], [21], and Smith [51]. The circuits shown in Figures 17-20
are just simple examples. The reader will have no difficulty in inventing
other, more complicated, circuits of his own.
Two actual encryption devices of this type that have received a lot of
attention are the IBM Lucifer Cipher and the National Bureau of
Standards Data Encryption Standard. The Lucifer cipher ([11], [20], [51])
uses a 128-bit key (rather than the 152 bits required in Figure 20), and
encrypts the data in blocks of 128 bits (instead of the 32-bit blocks in
Figure 20). The Data Encryption Standard ([3], [4], [40], [41]) uses a 56-bit
key and encrypts the data in blocks of 64 bits.
Unfortunately not much seems to be known about the security of
these ciphers, although some preliminary studies have been published ([2],
[24]). If the key size is small, the bad guy can break the cipher by simply
testing all possible keys. The Data Encryption Standard itself has been
..a363~
~------------'\ 32 Bits of
feedback
8 Bits
of key
8 Bits
of key
8 Bits Eight rom's

of key
Initial
contents
= 32 bits
of key
FIGURE 19
The operations shown in Figure 18 are applied fifteen times in
succession to the contents of the shift register, and the final 32 bits are
then fed back into the shift register. Each read-only memory requires
one bit of key, and the initial contents of the register are specified by 32
bits of key, so the whole circuit is specified by a total of 8 x 15 + 32
= 152 bits ofkey.
criticized on these grounds, since the number of different keys is only 2 56

~ 10 17 (see [6], [7], [27], [41], [53], [57]). Standards by which to judge the
security of this type of cipher are badly needed.
In all the encryption schemes described in this section it has been
assumed that the channel is free from errors. If channel errors do occur
their effect is disastrous, for good ciphers have the property that altering
one bit in the encrypted text changes about half the bits in the decrypted
message. To avoid this an error-correcting code (see Section 1) should be
used after the encrypting circuit, as shown in Figure 21 .
.a364~
32 Encrypted bits
32 Bits of message
FIGURE 20
Shows how the circuit in the previous figure might be incorporated
into a communication system. A fixed 152-bit key is used to specify the
shift register and the function 1, as in Figure 19. The data to be
transmitted is divided into blocks of32 bits each. To each block of data
we add the 32 bits in the shift register to produce 32 bits of encrypted
text. Then the fifteen permutations and nonlinear functions are
applied, giving the next 32 bits in the shift register; these are added to
the next 32 bits of data, and so on.
FIGURE 21
If the channel is noisy, an error-correcting code should be applied after
encryption, since even one channel error changes about half the bits in
the decrypted message.
Section 4
Codes Which Detect Deception
The channel continues to deteriorate: now it is actually controlled by the
bad guy (see Figure 3, part 4). He has promised to faithfully transmit our
messages, but we do not completely trust him. We therefore wish to sign or
authenticate our messages in such a way that he is unable to replace the true
message by a false one without being detected. E. N. Gilbert, F. J.
MacWilliams and I studied this problem in a paper published in 1974 [19].
.a365~
There are a number of situations where this problem arises. It was

originally presented to us by G. J. Simmons of Sandia Corporation in
connection with monitoring the production of certain materials in the
interests of strategic arms limitation. However it is simpler to describe the
problem in terms of a gambling casino.
The casino is managed by the Bad Guy, who is cheating the owner
(the Good Guy) by reporting the daily takings from the slot machines to be
less than they actually are and keeping the difference for himself. To
prevent this, the owner proposes to install in each slot machine a secret key
K and a device which takes as its input the day's takings X and the key K,
and produces as output a signature or authenticator
Z = «I>(X,K).
The device punches X and Z onto paper tape. The bad guy will then mail
the tape to the owner, who will read X, recalculate Z from X and K, and
check that this value of Z agrees with that punched on the tape. Ifit does,
he will assume that X is correct (Figure 22).
On the other hand the bad guy knows X, Z and «I> (that is, he knows
how the device works) but not K; and he wishes to replace X and Z by
another pair X' and Z'. If he can do this in such a way that
Z' = «I> (X',K)
then the substitution will not be detected, and he can pocket the difference
X'-x.
FIGURE 22
The good guy is planning to install a sealed box inside the slot
machine. This will record the day's takings X on paper tape, together
with a signature Z which is a function of X and the key K. Thus Z
= <I>(X,K). The bad guy, who knows everything except K, wishes to
replace (X,Z) by another pair (X',Zl If Z' satisfies Z' = <I>(X',K) he
will not be caught.
~366~
Figure 23 illustrates the kind of analysis the bad guy will perform. The
figure shows all possible messages Xl' X 2 , X 3 , ••. and the authenticators
Zt>Z2, ... corresponding to the different keys. Suppose the bad guy sees that
the message is Xl and the authenticator is Z2' By looking at the figure he sees
that the key is one of2, 3, 4, 5 or 6. He wishes to replace Xl by X 2 and must
decide which authenticator (Z4' Zs or Z6) to use. His chance of escaping
detection is greatest ifhe chooses Zs, for then he will not be caught if the key
turns out to be 2, 5 or 6-he has 3 chances in 5 of succeeding. But ifhe were
to choose Z4 or Z6 his chilnces would only be 1 in 5.
This argument suggests that a good way to design an authenticator
system is to ensure that there are a large number of possible keys
corresponding to each message-authenticator pair. Even so it is difficult to
make the bad guy's probability of success very small. In [19] we proved the
following result.
Message Key (Message, authenticator)
------<> (X"Z,)
X,~:-----~----iib (X"Z,)
~--i------:'" (X"Z,)
_---::===-"O (X" Z.)
X, ~:-------'7--~1IO (X"Z,)
-;;:----3'" (X"Z.)
_-===3100 (X"Z,)
X, . . : : : : : - - - - - - - - - : 3 1 1 0 (X"Z.)
--=====~"" (X"Z.)
FIGURE 23
Shows the correspondence between messages, keys and authenticators .
.8367~
THEOREM 2 Suppose there are M possible messages

Xl"'" X M and Npossible keys K 1 , ,K,N' lfthe
message and the key are picked at random, then
with optimal play the bad guy can guarantee
that his probability of success is at least 1I jN.
The proof uses techniques from information theory and will not be
given here. I t is possible that if the system is badly designed then the bad
guy's probability of success may be greater than II jN. However we also
showed how to design the system so that his probability of success is exact!J
I/jN, which by the theorem is best possible. (For simplicity we assume
that N is a perfect square.)
The construction makes use of a projective plane of order p = jN, as
defined in Section 1 (see Figure 24).
Recall that there are p2 + P + 1 points and p2 + P+ I lines in the
projective plane. We arbitrarily choose one of the lines and call it the
equator. The messages X will be represented by the p + I points on the
equator. The keys K are represented by the remaining p2 points. Finally
the authenticator Z = cD (X,K) is the unique line joining X and K. The
device will punch the coordinates of X and the equation to the line Z onto
the paper tape.
Now consider this from the bad guy's point of view. He knows X and
Z, but all he can say about the key is that it is one of the remainingp points
on Z. lfhe wants to replace X by X' he cannot do better than to pick one of
these p keys at random. For the possible keys (the points on Z) are in one-
to-one correspondence with the authenticators (the lines), given that one
Messages X
are all points
on the equator
Equator
Keys K are
all other
points
FIGURE 24
Optimal authentication system based on a projective plane of order p.
The authenticator Z = C1>(X,K) is (the equation to) the line through X
and K.
~368~
end of the line is fixed at X'. Thus his chance of success is lip = lifo, as
claimed.
This construction is mainly of theoretical interest, since it requires a
great deal of key, even more than the one-time pad described in Section 3a.
Nevertheless it is nice to see a scheme that can be proved to be secure (in
contrast to those described in Sections 3c and 5).
It is worth pointing out that this is one situation where the one-time
pad is useless. For in that case (see Equation 3) the authenticator would
simply be the sum of the message and the key, and the bad guy could find
the key immediately (since he knows the message).
In [19] we also described some other authentication schemes (based
on projective spaces and on random codes) which require a smaller
amount of key. All of our analysis (and in particular the proof of
Theorem 2) allows the bad guy to do as much computing as he needs. Of
course in practice he is restricted by the size of his computer. Because of this
limitation the conventional encryption schemes described in Section 3c-
the Lucifer cipher, for example-can also be used as authenticators. For
they produce an encrypted message r which is a complicated function of
the message X and the key K (see Figure 25), and this function is
specifically designed to make it hard to find K given X, rand $. This is a
much more practical solution than our construction using projective
planes, but suffers from the disadvantage that one does not know precisely
how secure it is.
Another solution to the authentication problem can be obtained by
using trapdoor functions-see Section 5f.
-6-
Key K
Message X Encrypted
Message
y = ~(X,K)
FIGURE 25
Section 5
Public-Key Cryptosystems
5A ONE.WAY FuNCTIONS
The predecessor of the schemes to be described in this section is the simple
and elegant one-way function. This is a function f which maps binary strings
of some fixed length (say 100) onto binary strings of some other fixed length
(say 120),
..a369~
Key = X
Input = Output
Constant Y = AX)
FIGURE 26
A conventional encryption scheme (such as the Lucifer cipher)
operated as a one-way function. The usual input is set equal to an
arbitrary constant, and the argument X is used as the key. The cipher
is explicitly designed so that it is essentially impossible to find X given
the output Y.
J: F 100 _ F 120 ,
and which has no known inverse! In other words, if you are told that
J(X) = 1010001···11011110
there is no way that you can find X, even knowing howJ is computed. This
is not quite accurate, since in theory one could find X by testing each
possible X in turn to see ifJ(X) is equal to the given string. Of course this is
essentially impossible, since there are too many different X's to try. So we
should say thatJ is a one-way function ifit is easy to computeJ(X) for any
X, while it is infeasible to computeJ- l (r) for almost all r in the range ofj
One-way functions certainly exist. For example we could operate one
of the conventional encryption schemes of Section 3c as shown in
Figure 26.
Such functions have a very pretty application to computers ([9], [42]).
I t is customary to allow the users of a computing system to choose their own
passwords. These identify the users when they log on. However it is
dangerous to store these passwords on a file in the computer since it is
difficult to keep such files private. Instead one can use a one-way functionJ
and store the values J (password) on the file. Since J cannot be inverted, a
bad guy cannot obtain the passwords from this file. On the other hand
when someone logs onto the computer he types in his password P, and the
computer calculatesJ(P) and compares it with the entry on the file. This is
a simple and (almost) foolproof scheme. (To make it more secure, users
must be encouraged not to choose passwords, such as short English names,
which are easy to guess ... )
5B TRAPDOOR FUNCTIONS
Of course a one-way function can't be used for encryption, for although
encrypting X into J(X) is secure, no one (including the legitimate receiver)
can recover X. As Diffie and Hellman pointed out in [5], the way around
..a370~
this is to use what they call a trapdoor function. This is a function

E: F lOO --+ F 120 (say),
which has an inverse
D: F 120 --+ F lOO ,
and both E and D are easy to compute. However the inverse cannot be
discovered by studying E : unless you are told what the inverse is, E
appears to be a one-way function. Then E can be used for encrypting and
D for decrypting, ~ince
D(E(X)) = X, for all X in F IOO •
Furthermore the trapdoor property implies that someone can know how to
encrypt without knowing how to decrypt. Examples of trapdoor functions
will be given in Section 5c-5e.
Once we know how to find trapdoor functions we can set up what
Diffie and Hellman call a public-key cryptosystem. Suppose there are a group
of people who wish to be able to talk to each other in privacy. Each person i
chooses a trapdoor function Ei with an inverse D i . The functions E l ,E2 , ...
are listed in a public directory, while each person keeps his inverse function
Di secret. When j wishes to send a message X to i he simply transmits
over a public wire. Since only i knows the inverse function D i , only i can
compute
and read the message. Thus we have a communication network which

ensures privacy without the use of keys.
5c TRAPDOOR FuNCTIONS BASED ON PRIME NUMBERS

Rivest, Shamir and Adelman were the first to find a really satisfactory
construction of trapdoor functions. This appeared in a paper [44] published
in 1978, although Martin Gardner had already outlined their method in
his August 1977 column [17]. In view of the extensive publicity this method
has received, only a brief description will be given here.
Choose two large prime numbers p and q (each for example about 50
digits long), and a number s which is relatively prime to both p - 1 and
q - 1. Calculate r = pq, and find t such that
st == 1 mod (p - 1) . (q - 1). (4 )
Then the encryption scheme is as follows .
..a371~
THE RIVEST-SHAMIR-ADELMAN
ENCRYPTION SCHEME
PUBLISH rAND s.
KEEP p, q, t SECRET.
ENCRYPT:
E(x) == xS(mod r) = y.
DECRYPT:
D(y) == l (mod r) = x.
Then D(y) = xst = x follows from Equation 4 and some elementary

number theory. E is (it is believed) a trapdoor function because the only
known way to find D, given rand s, is to factor r and so find p, q and hence t.
But r is a 100-digit number and at present it appears to be essentially
impossible to factor such large numbers. So the strength of this scheme rests
on the fact that while it is relatively easy to find large primes [51], it is
thought to be impossible to factor very large numbers [25], [46].
5D TRAPDOOR FUNCTIONS BASED ON KNAPSACK PROBLEMS

Shortly after Rivest, Adleman and Shamir found their trapdoor functions,
Merkle and Hellman [38] discovered another very simple family based on
what are called knapsack problems.
The prototype of a knapsack problem is this: given a knapsack and a
pile of objects that you would like to take on a hike, can you find a subset of
the objects that exactly fill the knapsack? A simple numerical problem of the
same type is:
A SIMPLE KNAPSACK PROBLEM
CAN YOU WRITE 31 AS THE SUM

OF CERTAIN OF THE NUMBERS
[10,17,9, 12,40,60]?
ANSWER: YES
31 = 10 + 9 + 12
Thus 31 can be represented as the sum of the first, third and fourth
numbers on the list, a fact which we indicate by writing
31 ~ (1,0, 1, 1,0,0).
Conversely given (1,0,1,1,0,0) we recover 31 by taking the sum of the first,
third and fourth numbers.
In this way we can use this list of numbers as the basis for a primitive
encryption scheme, as shown in Figure 27.
101.100
_ I
. Encrypt
31
• Decrypt
101100
•
'------'
FIGURE 27
A primitive encryption scheme in which both the sender and receiver
know the list of numbers [10, 17,9, 12,40,60].
Some numerical knapsack problems are very easy to solve, for

example if the numbers in the list are powers of two (or just increase very
rapidly).
A VERY EASY KNAPSACK

PROBLEM
EXPRESS 19 AS A SUM OF
DISTINCT NUMBERS FROM THE
LIST
[1,2,4,8,16,32]'
ANSWER: 19 = 1 + 2 + 16, OR
19 ~ (1, 1,0,0, 1,0).
For this just amounts to finding the binary expansion of the number.
On the other hand some are extremely difficult to solve. Suppose for
example the list contains 100 numbers ai' ... , a iOO , each about 40 digits
long, and y is a 42-digit number.
~373~
A HARD KNAPSACK PROBLEM
GIVEN [a1,a2,"" alOO]' a i Rj 1040 .

EXPRESS y( Rj 10 42 ) AS A SUM
OF THE ai
y = a 3 + all + a 12 + a 20 + ...
If the numbers on the list are chosen at random then there appears to be no
possible way of finding out which subset of them add up toy. For there are
2100 Rj 10 30 different subsets to be considered. If we used this list of
numbers in the communication system shown in Figure 27 it would
certainly be secure, but unfortunately even the legitimate receiver would
not be able to recover the message.
What we need is a trapdoor knapsack problem which is easy for the
legitimate receiver to solve, but which looks to anyone else like a hard
problem. Merkle and Hellman [38] described several methods of doing
this. The following variant of one of their constructions was discovered
independently by Graham [23] and Shamir (see [48]).
We first form an easy knapsack problem by choosing numbers
at> ... , a 1 00, say, which have powers of two embedded in their decimal
expansions (Figure 28 shows a smaller example).
Figure 29 shows how this easy knapsack problem could be used for
encryption. Of course this is very easy to break and is not how we will
actually use it.
a, = 8 3 0 0 5
a, = 2 0 2 0 6
a, = 7 8 0 4 0 9 0
a. = 3 5 0 8 0 4 9
a, = 2 4 6 0 3
a, = 3 3 3 2 0 7 8
T
Powers of 2
\
Column of zeros
FIGURE 28
Six numbers a1, ... ,a6' with embedded powers of two, to be used in a
trapdoor knapsack problem.
Instead we choose two large numbers rand s such that there exists a
number t with
st == I (mod r).
Then we scramble the a's by multiplying them by s and reducing mod r.
The resulting numbers
[h 1 , h2 , ••• , h100], hj == saj (mod r)
will appear to be random numbers in the range 0 to r - I, and look like a
very hard knapsack problem to anyone who does not know r, sand t. It is
this hard knapsack problem that will actually be used for encrypting. We
keep al, ... , alOO, r, sand t secret, and publish h1 , ••. , h lOO . The encryption
scheme works like this.
TRAPDOOR KNAPSACK
ENCRYPTION SCHEME
PUBLISH h1 ,h 2 , .•. ,hlOO.

ENCRYPT THE BINARY MESSAGE
AS THE NUMBER
100
E(x) = L
i= 1
xjh j = y.
DECRYPT BY FORMING
100
ry = L Xjthj
i=l
100
= L xja j (mod r)
j=l
AND SOLVING THIS EASY KNAP-

SACK TO GET Xl' ... 'XlOO.
Figure 30 shows an example. Anyone who does not know r, sand t is faced
with solving an extremely hard knapsack problem, but the legitimate
receiver recovers the message immediately.
~375&.
Easy
Kn!~
8 3 0 0 5 830 1 051
o 2 0 2 0 6
7 8 0 4 0 9 0 7 8 0 4 0 9 0
o 3 5 0 8 0 4 9
2 4 6 0 3 2 4 6 0 3
3 3 3 207 8 3 3 3 207 8
2 853232
T
53 = IOlOl1 in binary
which is the message!
FIGURE 29
Shows how the easy knapsack numbers from Figure 28 could be used
in an encryption scheme. The encrypted message is formed, just as in
Figure 27, by adding the numbers from the list corresponding to the
l's in the message. Decryption is immediate: the binary expansion of
the central pair of digits is the message!
Hard
WP.mt ~yp.t
8301051 5 1 5 8 629 6 7 5 1 5 8 6 296 7
o 20 2 0 6 1 6 7 927 3 397
7804090 5 1 343 8 305 5 1 343 8 305
o 3508049 402 1 6 1 783
2416013 42753 179 1 42753 179 1
3332078 3 7 605 264 1 3 7 605 264 1
1 8 328 8 5 704
Multiply by
s = 324358647 Cipher
and reduce mod
r = 786053315
FIGURE 30
The easy knapsack problem of Figure 28 is made into a hard problem
by multiplying the numbers by s = 324358647 and reducing them
mod r = 786053315. To decrypt we use the number t = 326072163,
which satisfies st == 1 (mod r). The encrypted message 1832885704 is
multiplied by t mod r, producing 21853232. The central digits, 53,
when expanded in binary (see Figure 29) give the message .
.8376~
5E TRAPDOOR FUNCTIONS BASED ON GOPPA CODES

McEliece [37] has constructed a family of trapdoor functions using Goppa
codes (see Example 3 of Section 1). We fix two numbers n = 2m and t,
choose an irreducible polynomial G(z) over GF(2m) of degree t, and form
the parity-check matrix H of the corresponding t-error-correcting Goppa
code (see Equation 1 above). From H we calculate a generator matrix for the
code: this is a k x n matrix M, where k = n - mt, such that if x
= (XI' ... , Xk) is the message vector, the corresponding codeword c
= (Cl"'" cn) is given by the matrix product
c=xM,
calculated mod 2. (It is easy to find M from H-see [35, Ch. 1].) The key
idea is to scramble M by choosing a random invertible k x k binary matrix
S and a random n X n permutation matrix P, and forming the new
generator matrix
M'=SMP.
Then M' is published while M, Sand P are kept secret. Here is the
encryption scheme:
ENCRYPTION SCHEME
USING GOPPA CODES
ENCRYPT k-BIT MESSAGE x AS

Y =xM'+ z
WHERE z IS A RANDOM VECTOR
CONTAINING t ONES, CHOSEN BY
THE SENDER.
TO DECRYPT, CALCULATE
yP- 1 = (xS)M + (zP- 1 ),
DECODE AS USUAL TO OBTAIN xS
HENCE x.
The sender obscures the codeword xM' by changing t randomly chosen

bits. Since the Goppa code is capable of correcting t errors, the legitimate
receiver can remove this distortion, for example by using the error-
correcting procedure described in [35, Chapter 12]. But an eavesdropper,
who does not know M, S or P, must try and decode the code defined by the
£Jt377~
generator matrix M'. This is a very large random-looking linear code, and
such codes are believed to be extremely difficult, if not impossible, to
decode.
As a numerical example we might take n = 1024 = 2 10 and t = 50.
There are about 10 149 possible choices for G(z), and an even larger number
of ways of choosing Sand P. The dimension of the code is at least k = 1024
- 10·50 = 524, and the eavesdropper is faced with the problem of
decoding an apparently random 50-error-correcting code of length 1024.
For further details see [37].
5F SIGNED MAIL
Some of the public-key cryptosystems make it possible to send signed mail,
that is, for i to be able to sendj a message encrypted in such a way that only
i could have sent it. This property is of course essential when these
encryption schemes are used for transferring money.
For this to be possible, the trapdoor functions E j and D j (see Section
5b) should satisfy
Ej(Dj(X)) = X for all X, (5)
as well as
Dj(Ej(X)) =X for all X.
The prime number schemes in Section 5c certainly satisfy this condition.
Then i simply encrypts the message X as
and sends it to j, who can recover X by calculating
Only i could have sent this message, since only i knows D j • Thus Z is a
signed version of X.
Trapdoor functions which satisfy Equation 5 can also be used to solve
the authentication problem described in Section 4. The message X can
simply be accompanied by the authenticator D(X). Everyone involved
knows E and can verify that X is genuine by checking that
E(D(X)) = x.
But since D itselfis secret, the bad guy cannot find the authenticator D (X')
needed to match his false message X'.
For more about signal mail see [30], [44] and [45]; the paper by
Shamir, Rivest and Adleman [47] in this volume contains a nice appli-
cation of these ideas to "mental poker."
~378~
5G CONCLUSION
Other public key cryptosystems have also been proposed ([26], [33]). But in
everyone of these schemes there is an important open question:
HOW SECURE ARE THEY? I

They appear secure, but up to now nothing has been proved about their
strength, and until that happens there is always the possibility that
someone will invent a clever way of breaking them. Some methods of
attack on both the prime number schemes and the knapsack schemes have
been proposed, but do not appear to pose a serious threat (see [28], [43],
[48], [50]).
I hope the reader has enjoyed this introduction to an exciting and
rapidly developing field. We have seen that, even under the most severe
conditions, ingenious schemes are available which enable us to com-
municate in privacy.
References
1 Carleial, A. B. and Hellman, M. E. 1977. A note on Wyner's wiretap
channel. IEEE Trans. Info. Theory IT-23: 387-390.
2 Coppersmith, D. and Grossman, E. 1975. Generators for certain
alternating groups with applications to cryptography. SIAM ]. Applied
Math. 29: 624-627.
3 Data Encryption Standard, Federal Information Processing Standard
Publication No. 46, National Bureau of Standards, U.S. Dept. of
Commerce, January 1977.
4 Davis, R. M. 1978. The Data Encryption Standard in perspective.
IEEE Communications Society Magazine, 16 (November): 5-9.
5 Diffie, W. and Hellman, M. E. 1976. New directions in cryptography,
IEEE Trans. lrifo. Theory IT-22: 644-654.
6 _ _ . 1976. A critique of the proposed Data Encryption Standard.
Comm. ACM 19: 164-165.
7 _ _ . 1977. Exhaustive analysis of the NBS data encryption standard.
Computer 10: (June) 74-84.
8 _ _ . 1979. Privacy and authentication: an introduction to cryptog-
raphy. Proc. IEEE 67: 397-427.
9 Evans, A. Jr., Kantrowitz, W., Weiss, E. 1974. A user authentication
scheme not requiring secrecy in the computer. Comm. ACM 17: 437-442.
10 Flik, V. 1979. Repeated use of codes which detect deception. IEEE
Trans. Info. Theory IT-25: 233-234.
~379~
11 Feistel, H. 1970. Cryptographic coding for data-bank privacy. Report

RC-2827, Yorktown Heights, N.Y.: IBM Watson Research Center.
12 - - . 1973. Cryptography and computer privacy. Scientific American
228 (May): 15-23.
13 Feistel, H., Notz, W. A. and Smith, ]. L. 1971. Cryptographic
techniques for machine to machine data communications. Report RC-
3663. Yorktown Heights, N.Y.: IBM Watson Research Center.
14 _ _ . 1975. Some cryptographic techniques for machine-to-machine
data communications. Proc. IEEE 63: 1545-1554.
15 Gaines, H. F. 1956. Cryptoanalysis. New York: Dover.
16 Gallager, R. 1968. Information Theory and Reliable Communication. New
York: Wiley.
17 Gardner, M. 1977. A new kind of cipher that would take millions of
years to break, Scientific American 237 (August): 120-124.
18 Geffe, P. R. 1967. An open letter to communication engineers. Proc.
IEEE 55: 2173.
19 Gilbert, E. N., MacWilliams, F. J. and Sloane, N. J. A. 1974. Codes
which detect deception. Bell Syst. Tech.]. 53: 405-424. For a sequel to
this paper see reference [10].
20 Girsdansky, M. B. 1971. Data privacy-Cryptology and the computer
at IBM Research. IBM Research Reports 7: (No.4), 12 pages.
21 _ _ . 1972. Cryptology, the computer and data privacy. Computers and
Automation 21 (April): 12-19.
22 Golomb, S. W. ed., 1964. Digital Communications with Space Applications,
Englewood Cliffs, N.J.: Prentice-Hall.
23 Graham, R. L. Personal communication.
24 Grossman, E. K. and Tuckerman, B. 1977. Analysis of a Feistel-like
cipher weakened by having no rotating key. Report RC-6375. Yorktown
Heights: N.Y.: IBM Watson Research Center.
25 Guy, R. K. 1975. How to factor a number. Proc. Fifth Manitoba
Conference on Numerical Math. pp. 49-89.
26 Hellman, M. E. 1978. An overview of public key cryptography. IEEE
Communications Socie{y Magazine 16 (November): 24-32.
27 - _ . 1980. A cryptanalytic time-memory tradeoff. IEEE Trans. Info.
Theory. IT-26 (July).
28 Herlestam, T. 1978. Critical remarks on some public-key cryptosys-
terns. BIT 18: 493-496.
29 Kahn, D. 1967. The Codebreakers. New York: Macmillan.
30 Kohnfelder, L. M. 1978. On the signature reblocking problem In
public-key cryptosystems. Comm. ACM 21: 179.
~380~
31 Leung-Van-Cheong, S. K. 1977. On a special class of wiretap channels.

IEEE Trans. Info. Theory. IT-23: 625-627.
32 Leung-Y an-Cheong, S. K. and Hellman, M. E. 1978. The Gaussian
wiretap channel. IEEE Trans. lifo. Theory IT-24: 451-456.
33 Leung-Yan-Cheong, S. K. and Vacon, G. V. A method for private
communication over a public channel, preprint.
34 MacWilliams, F . .J. and Sloane, N. J. A. 1976. Pseudo-random
sequences and arrays. Proc. IEEE 64: 1715-1729.
35 - - . 1977. The Theory of Error-Correcting Codes. New York: Elsevier.
36 McEliece, R . .J. 1977. The Theory of Information and Coding. Reading,
Mass.: Addison-Wesley.
37 _ _ . 1978. A public-key cryptosystem based on algebraic coding
theory. Deep Space Network Progress Report 42-44. Pasadena: Jet Propul-
sion Labs (January) pp. 114-116.
38 Merkle, R. C. and Hellman, M. E. 1978. Hiding information and
signatures in trapdoor knapsacks. IEEE Trans lifo. Theory IT -24: 525-
530.
39 Meyer, C. H. and Tuchman, W. L. 1972. Pseudorandom codes can be
cracked. Electronic Design 20 (November 9): 74-76.
40 Morris, R. 1978. The Data Encryption Standard-retrospective and
prospects. IEEE Communications Society Magazine 16 (November): 11-14.
41 Morris, R., Sloane, N . .J. A. and Wyner, A. D. 1977. Assessment of the
National Bureau of Standards Proposed Federal Data Encryption
Standard. Cryptologia I: 281-306.
42 Purdy, G. B. 1974. A high security log-in procedure. Communications
ACM 17: 442-445.
43 Rivest, R. L. 1978. Remarks on a proposed cryptanalytic attack on the
M.LT. public-key cryptosystem. Cryptologia 2: 62-65.
44 Rivest, R. L., Shamir, A. and Adelman, L. M. 1978. A method for
obtaining digital signatures and public-key cryptosystems. Comm. ACM
21: 120-126.
45 Shamir, A. 1978. A fast signature scheme. Report TM-I07. Laboratory
for Computer Science, M.LT.
46 _ _ . 1979. Factoring numbers in O(log n) arithmetic steps. lifo.
Processing Letters 8: 28-31.
47 Shamir, A., Rivest, R. L. and Adleman, L. M. Mental Poker. Intra., pp.
37-43.
48 Shamir, A. and Zippel, R. E. 1980. On the security of the Merkle-
Hellman cryptographic scheme. IEEE Trans. Info. Theory IT-26 (May).
49 Shannon, C. E. 1949. Communication theory of secrecy systems. Bell
Syst. Tech. J. 28: 656-715.
~381~
MATHEMATICAL GARIfNER
50 Simmons, G.J. and Norris, M.J. 1977. Preliminary comments on the

M.LT. public-key cryptosystem. Cryptologia 1: 406-414.
51 Smith, J. L. 1971. The design of Lucifer, a cryptographic device for
data communications. Report RC-3326. Yorktown Heights, N.Y.: IBM
Watson Research Center.
52 Solovay, R. and Strassen, V. 1977. A fast Monte-Carlo test for
primality. SIAM]. Computing 6: 84-85 and 7 (1978): 18.
53 Sugarman, R. et al., 1979. On foiling computer crime. IEEE Spectrum
16 (July): 31-4l.
54 Vernam, G. S. 1926. Cipher printing telegraph systems for secret wire
and radio telegraphic communications. ]. AlEE 45: 109-115.
55 Verriest, E. and Hellman, M. E. 1979. Convolutional encoding for
Wyner's wiretrap channel. IEEE Trans. Info. Theory, IT-25: 234-237.
56 Wyner, A. D. 1975. The wire-tap channel. Bell Syst. Tech.]. 54: 1355-
1387.
57 Yuval, G. 1979. How to swindle Rabin. Cryptologia 3: 187-189.
~382~

1981 Book TheMathematicalGardner PDF

Uploaded by

Copyright:

Available Formats

1981 Book TheMathematicalGardner PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1981 Book TheMathematicalGardner PDF

Uploaded by

Copyright:

Available Formats

What is the book about?

What is the book about?

What is Martin Gardner's role discussed in the introduction?

What is Martin Gardner's role discussed in the introduction?

THE

PRINDLE. WEBER & SCHMIDT WADSWORTH INTERNATIONAL

10 Davis Drive, Belmont, California

All rights reserved. No part of this book may be reproduced or transmitted in

Wadsworth International is a division of Wadsworth, Inc.

Library of Congress Cataloging in Publication Data

Main entry under title:

The Mathematical Gardner

1. Mathematical recreations - Addresses, essays,

ISBN-13: 978-1-4684-6688-1 e-ISBN-13: 978-1-4684-6686-7

T he articles in this book are dedicated to Martin Gardner, the world's

Anyone for Twopins? 2

Pretzel Solitaire as a Pastime for the Lonely Mathematician 16

Some Remarks about a Hex Problem 25

Cheap, Middling, or Dear 44

A Random Hopscotch Problem, or How to Make Johnny Read More 51

Wreaths of Tangent Circles 62

Bicycle Tubes Inside Out 75

Slicing It Thin 100

How Did Pappus Do It? 112

Fault-Free Tilings of Rectangles 120

Disections Into Equilateral Triangles 127

In Praise of Amateurs 140

Some Problems on Plane Tiling 167

Angels and Devils 197

Packing Problems and Inequalities 212

Can Cubes Avoid Meeting Face to Face? 226

Packing Handed Pentacubes 234

My Life Among the Polyominoes 243

FUN AND PROBLEMS

Noneuclidean Harmony 265

Magic Cuboctahedrons 270

Games, Graphs, and Galleries 274

Probing the Rotating Table 285

N UMBERS AND CODING THEORY

Supernatural Numbers 310

The Graph Theorists Who Count-and What They Count 326

Error-Correcting Codes and Cryptography 346

The bowling game of Twopins (pronounced "Tuppins") is played by

Take the mex of the nim-values of the options.

The player whose turn it is to move chooses one of the component

The nim-value of the sum of two games is the nim-sum of their

To find the nim-sum of two nonnegative integers, add them in binary

There is no legal move in which r beans may be taken.

The game of Twopins was discovered by Elwyn Berlekamp in the

where do = 0 or 4 (split a row or heap into two nonempty rows or heaps

1 O*iJk··· = *ijk ... = OOijk ... ,

where in is the Fibonacci number

However, it is unnecessary to analyze positions which are mere reflections

B ... * 0 0 0*··· ... * 0 000 0*··· C

C ... * 00 o 00* ... ... * 00 O?O 00* ... D

Thus, Sn satisfies the recurrence

and the total number, not counting reflections, is

was discussed in [1], in which 0 occurs in blocks oflength at least k. Here we

Formulas 7 and 8 were not given in [1].

1 OiJk··· = ijk ... = OOijk ... ,

B ... * 0 0 0··· ... 0 000 0*··· C

000······ ···000 * 000 000 *