Advance Praise

“This is a beautiful collection of mathematical puzzles, followed by hints and

solutions. No advanced mathematics is necessary, and yet many of the puzzles
are truly challenging, even for a professional mathematician. Peter Winkler has
led a Year of Puzzles at the National Museum of Mathematics in New York.
His superb exposition ensures that while reading the book and trying to solve the
puzzles you will have hours of fun, intrigue, joy and enlightenment.”
—Noga Alon, Princeton University

“This is an incredible collection of puzzles. While many books claim to be ‘for

all levels,’ here it is really true: the range of problems is truly remarkable. There
are simple ones for teens and pre-teens, right up to ones for grown-ups. I do not
know any book with such a spread.
“A particular feature of the book is that its easiest problems, aimed say
at youngsters, are so attractive. Thus, for example, each of ‘Phone Call’ and
‘Winning at Wimbledon’ are surprising and clever. And then there are problems
like ‘Watermelons,’ where the calculation is easy but the result itself is very un-
expected indeed. Such interest is rare among easy problems, and this feature by
itself would make this a must-buy.
“At the other end of the scale, there are problems like ‘Alternating Powers’
that lead on to some very deep mathematics.
“All of the book is written in Peter Winkler’s trademark fun and chatty style.”
—Imre Leader, Professor of Mathematics at Cambridge University

“Anyone, from novice to expert, with an interest in math or puzzles should have
this book. The puzzles are artfully chosen and lucidly explained in a way that
will boost the math IQ of readers at any level.
“As a reader with a lifelong interest in recreational math I was delighted with
these superb puzzles with many surprises, elegant solutions and detailed infor-
mation on their origins.”
—Dick Hess, author of Golf on the Moon and other puzzle books

“On average, who has more sisters, men or women? When a coin is rolled once
around another one of equal size, with no slipping, how many times does it
rotate? How can you get a 50-50 decision by flipping a bent coin? If you enjoy
challenges like these, you will be enthralled by this latest collection of delight-
ful dazzlers from Peter Winkler. From old classics to new gems, wrestling with
these brain ticklers will provide many weeks of fun and ‘Aha!’ moments.
 “Warning: these are addictive, and you’ll need to use a magic combination of
skill, intuition and insight to solve them all!”
—Colm Mulcahy, Professor of Mathematics at Spelman College and author
of Mathematical Card Magic: Fifty-Two New Effects

“Peter Winkler’s collection of attractive puzzles will appeal to readers at all

levels. Many famous puzzles are discussed, but there are also new and attractive
problems that readers will surely not be familiar with. I thoroughly enjoyed the
book and highly recommend it to those with an interest in the best that modern
puzzling has to offer.”
—Stan Wagon, Macalester College, author of Which Way Did the Bicycle
Go? and Bicycle or Unicycle?

“This is the third and by far the most substantial of Peter Winkler’s books on
mathematical puzzles, drawing from everyday life and from a wide spectrum
of mathematical topics. Mathematics’ answer to Ripley’s Believe It or Not, Peter
brings out the romantic side of mathematics rather than its utilitarian side.”
—Andy Liu, Leader of the Canadian team to the International
Mathematical Olympiad in 2000 and 2003

“Peter Winkler’s Mathematical Puzzles is thought-provoking, and I fell in love with

the puzzles. Their levels range from questions for smart high-school students to
rudiments of research problems for serious mathematicians. I was charmed by
the bar trivia sprinkled here and there on top of respectable mathematics. This
book is an ideal companion to pack in the vacation luggage for anyone inter-
ested in mathematics: it is bound to keep them entertained for hours.”
—Vašek Chvátal, Emeritus Professor at Concordia University

“Peter Winkler is the puzzle master’s puzzle master: a serious mathematician

not only with exquisite taste in what makes for a great puzzle but who also has
an unrivaled elegance of delivery and clarity of explication. This book presents
the crown jewels of mathematical puzzledom and will bring hours of joy – and
pain! – to all curious and playful minds.”
– Alex Bellos, author of Alex’s Adventures in Numberland

“This book may well be the best collection of mind-stretching teasers ever
assembled. You can’t help but be inspired, when Peter winks at you.”
—Donald E. Knuth, Emeritus Professor at Stanford University
and winner of the Turing Award
Mathematical Puzzles
Mathematical Puzzles
Revised Edition

Peter Winkler

Illustrations by Jess Johnson

First edition published 2024
by CRC Press
2385 NW Executive Center Drive, Suite 320, Boca Raton, FL 33431

and by CRC Press

4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN

© 2024 Peter Winkler

CRC Press is an imprint of Taylor & Francis Group, LLC

Reasonable efforts have been made to publish reliable data and information, but the author and
publisher cannot assume responsibility for the validity of all materials or the consequences of
their use. The authors and publishers have attempted to trace the copyright holders of all material
reproduced in this publication and apologize to copyright holders if permission to publish in this
form has not been obtained. If any copyright material has not been acknowledged please write and
let us know so we may rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information
storage or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, access www.
copyright.com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive,
Danvers, MA 01923, 978-750-8400. For works that are not available on CCC please contact
mpkbookspermissions@tandf.co.uk

Trademark notice: Product or corporate names may be trademarks or registered trademarks, and are
used only for identification and explanation without intent to infringe.

ISBN: 978-1-03270-953-6 (hbk)

ISBN: 978-1-03270-848-5 (pbk)
ISBN: 978-1-03270-965-9 (ebk)

DOI: 10.1201/9781032709659

Publisher's note: This book has been prepared from camera-ready copy provided by the authors.
for Miles, Sage, and Beatrix
 Contents
Acknowledgments ix

Preface xi

The Puzzles xv

The Hints lxxxix

1. Out for the Count 1

2. Achieving Parity 17

3. Intermediate Math 29

4. Graphography 41

5. Algebra Too 55

6. Safety in Numbers 67

7. The Law of Small Numbers 77

8. Weighs and Means 91

9. The Power of Negative Thinking 101

10. In All Probability 109

11. Working for the System 137

12. The Pigeonhole Principle 151

13. Information, Please 163

14. Great Expectation 181

vii
 Contents

15. Brilliant Induction 205

16. Journey into Space 223

17. Nimbers and the Hamming Code 235

18. Unlimited Potentials 253

19. Hammer and Tongs 273

20. Let’s Get Physical 295

21. Back from the Future 303

22. Seeing Is Believing 327

23. Infinite Choice 341

24. Startling Transformation 353

Notes and Sources 371

Bibliography 393

Author 399

Index 401

viii
 Acknowledgments
This volume owes its existence to two institutions and many individ-
uals. Dartmouth College has provided inspiring colleagues, a stim-
ulating environment, and steadfast encouragement; special thanks
are due to the endowers of the William Morrill Professorship. The
National Museum of Mathematics (with whom I spent the academic
year 2019–2020), through the generosity of the Simons Foundation,
provided a huge variety of outreach opportunities, with many audi-
ences on whom to test puzzles.
Some of the individual contributors (witting and otherwise) are
listed alphabetically below, with multiple contributors in italics and
the two most prolific in bold. Specific contributions appear in Notes &
Sources. Doubtless many deserving names have been left out owing to
gaps in my knowledge or memory, but I hope to rectify some of those
omissions as time goes on.

Dorit Aharonov · Arseniy Akopyan · David Aldous ·

Kasra Alishahi · Noga Alon · Dan Amir · A. V. Andjans ·
Titu Andreescu · Omer Angel · Gary Antonick · Steve Bab-
bage · Matt Baker · Yuliy Baryshnikov · Elwyn Berlekamp ·
Béla Bollobás · Christian Borgs · Bancroft Brown · Daniel
E. Brown · A. I. Bufetov · Joe Buhler · Caroline Calder-
bank · Neil Calkin · Teena Carroll · Amit Chakrabarti ·
Deeparnab Chakrabarty · Adam Chalcraft · A. S. Cheb-
otarev · William Fitch Cheney, Jr. · Vladimir Chernov ·
N. L. Chernyatyev · Vašek Chvátal · Barry Cipra · Bob
Connelly · John H. Conway · Tom Cover · Paul Cuff
· Robert DeDomenico · Oskar van Deventer · Randy
Dougherty · Peter Doyle · Ioana Dumitriu · Freeman
Dyson · Todd Ebert · Sergi Elizalde · Noam Elkies ·
Rachel Esselstein · R. M. Fedorov · Steve Fisk · Gerald
Folland · B. R. Frenkin · Ehud Friedgut · Alan Frieze · Anna
Gal · David Gale · A. I. Galochkin · Gregory Galperin ·
Martin Gardner · Bill Gasarch · Dieter Gebhardt · Giulio
Genovese · Carl Giffels · Sol Golomb · David Gontier ·

ix
 Acknowledgments

Bill Gosper · Ron Graham · Geoffrey Grimmett · Ori

Gurel-Gurevich · Olle Häggström · Sergiu Hart · Bob Hen-
derson · Dick Hess · Maya Bar Hillel · Iwasawa Hirokazu ·
Ander Holroyd · Ross Honsberger · Naoki Inaba · Svante
Janson · A. Ya. Kanel-Belov · Mark Kantrowitz · Howard
Karloff · Joseph Keane · Kiran Kedlaya · David Kempe
· Rick Kenyon · Tanya Khovanova · Guy Kindler · Mur-
ray Klamkin · Victor Klee · Danny Kleitman · V. A.
Kleptsyn · Niko Klewinghaus · Anton Klyachko · Don
Knuth · Maxim Kontsevich · Isaac Kornfeld · A. K. Ko-
valdzhi · Alex Krasnoshel’skii · Piotr Krason · Jeremy
Kun · Thomas Lafforgue · Michael Larsen · Andy Latto ·
V. N. Latyshev · Imre Leader · Tamás Lengyel · Hendrik
Lenstra · Anany Levitin · Jerome Lewis · Sol LeWitt ·
Michael Littman · Andy Liu · Po-Shen Loh · László Lovász
· Edouard Lucas · Russ Lyons · Marc Massar · David
McAllester · Peter Bro Miltersen · Grant Molnar · R. L.
Moore · Thierry Mora · Frank Morgan · Carl Morris ·
Lizz Moseman · Elchanon Mossel · Frederick Mosteller ·
Colm Mulcahy · Pradeep Mutalik · Muthu Muthukrish-
nan · Gerry Myerson · Assaf Naor · Girija Narlikar ·
Jeff Norman · Alon Orlitsky · Garth Payne · Yuval Peres
· Nick Pippenger · Dick Plotz · V. V. Proizvolov · Jim
Propp · Kevin Purbhoo · Anthony Quas · Dana Randall
· Lyle Ramshaw · Dieter Rautenbach · Sasha Razborov ·
Dan Romik · I. F. Sharygin · Bruce Shepherd · Raymond
Smullyan · Jeff Steif · Sara Robinson · Saul Rosenthal ·
Rustam Sadykov · Mahdi Saffari · Boris Schein · Christof
Schmalenbach · Markus Schmidmeier · Frederik Schuh ·
David Seal · Dexter Senft · Alexander Shapovalov · James
B. Shearer · S. A. Shestakov · Senya Shlosman · George
Sicherman · Vladas Sidoravicius · Sven Skyum · Laurie
Snell · Pablo Soberón · Emina Soljanin · Joel Spencer · A.
V. Spivak · Caleb Stanford · Richard Stanley · Einar Stein-
grimsson · Francis Su · Benny Sudakov · Andrzej Szulkin
· Karthik Tadinada · Bob Tarjan · Bridget Tenner · Prasad
Tetali · Mikkel Thorup · Mike Todd · Enrique Treviño ·
John Urschel · Felix Vardy · Tom Verhoeff · Balint Virag
· Stan Wagon · George Wang · Greg Warrington · Johan
Wästlund · Diana White · Avi Wigderson · Herb Wilf ·
Dana Williams · I. M. Yaglom · Steven Young · Paul Zeitz
· Leo Zhang · Yan Zhang · Barukh Ziv

x
 Preface
What's in This Book?
This book contains 300 of the very best mathematical puzzles in the
world. Some have appeared in previous books, including my own, or
in contests; others are brand new. The book is organized as follows:
First, all the puzzles are presented; then, for each puzzle, a hint to-
gether with the chapter number in which you can find the solution.
(The Hints section lists puzzles in alphabetical order by name, so may
be of use in locating a particular puzzle.)
The chapters themselves are classified by solution method, rather
than by general puzzle topic. Each chapter introduces a technique,
uses it to solve puzzles, and ends with an actual theorem of mathe-
matics that uses the technique in its proof. Additional information on
every puzzle is found in Notes and Sources at the end of the volume.

Who Is This Book For?

Puzzle lovers and math lovers. For background you’ll need no more
than high-school math. Calculus, abstract algebra, and higher college-
level material will not be used. Some math that you might not find in
every curriculum, like elementary graph theory and probability, will
be introduced as needed.
This does not mean the puzzles are easy; some are, others will
stump professional mathematicians. And, the existence of an easy so-
lution does not mean that a solution is easy to find!

How Were the Puzzles Chosen?

To me a great puzzle offers entertainment and enlightenment. It
should be fun to read and think about. It should make an interesting
mathematical point, perhaps by confounding your intuition, or per-
haps just by reminding you of the amazing things elementary mathe-
matics can do.

xi
 Mathematical Puzzles

Where Do the Puzzles Come From?

Some of the puzzles are original but most were contributed by puzzle
composers and puzzle mavens all over the world—they are the real
heroes of this book. Many of their names are found in the Acknowl-
edgments and Bibliography. Problems from contests are not usually ap-
propriate for this volume; they are designed to test, rather than to en-
tertain, and often require more advanced mathematics. Nonetheless
some irresistible gems have arisen in contests—especially, for some
reason, the Moscow Mathematical Olympiads. Often a puzzle has
multiple sources and rarely can its origin be identified, or even de-
fined, precisely; my meager efforts at doing so are reported in Notes
and Sources. Puzzles and solutions are presented in my own words,
which in some cases will be similar to a source’s, in others completely
different.

How Should Puzzles Be Used−or Not Used?

Mathematical puzzles should not be used for triage! Degree of ability
to solve puzzles does not tell you whether someone will make a good
mathematician, or even be good at mathematics.
Why not? Don’t mathematicians sit around all day solving puz-
zles?
Nothing could be farther from the truth. To begin with, most re-
search mathematicians are not problem-solvers but theory-builders.
They look for patterns, ask good questions, and put ideas together.
When mathematicians do solve problems, they take their time and
usually work together, each contributing different skills. And they of-
ten change the problem as they work! Moreover, the problems they
work on are not composed as clever traps, they arise naturally.
So it should not be a surprise that most mathematicians are not
great puzzle solvers—and many non-mathematicians are.
On top of all this, it is useful to remember that the ability to solve
puzzles has a lot to do with confidence in oneself—which in turn is
inextricably tied to social norms and prejudice, and perhaps testos-
terone as well. The last thing we want to do is discourage women and
minorities from pursuing mathematical careers.

xii
 Author's Preface

So what are puzzles good for, if the ability to solve them does not
tell us who should become a mathematician?
The key is that a good puzzle is a gem, a thing of beauty. I would
argue that the ability to appreciate a good puzzle, and the willingness to
put effort into solving it, tells us a great deal. (Smart employers know
this and if you are given a puzzle in a job interview, they are inter-
ested in your attitude and approach, not whether you get the “right”
answer.) Do you want to know what math can do? Or, equally impor-
tant, what math can’t do? Do you feel that you are desperate to know
the answer? Then, maybe, math is for you.
And if math is for you, then learning the techniques needed to
solve puzzles is a great way to dive in. As you will see in this volume,
those same techniques show up in proofs of theorems; and as (I hope)
you will also see, solving puzzles is eye-opening and a lot of fun.

Where Should You Start?

With the puzzles, of course. The puzzles are the stars of this show;
the rest of the book is just support. If you simply read the puzzles,
pick out the ones you like, and try to solve them, you are headed for
days, weeks, or months of intrigue, frustration, joy, and enlighten-
ment. Yes, the techniques proposed in the chapters are useful tools,
but be warned that no two people will classify puzzle-solving tech-
niques the same way or even assign puzzles the same way in a given
classification. Indeed, my own assignments are often distressingly ar-
bitrary.
Note also that for the harder puzzles, the solution given (which
may be one of many) is sometimes presented in rather compact fash-
ion. The heart symbol (♡) is used to signal the end of an argument or
proof; if you reach one unedified, go back!
Whatever you do, have fun and keep an open mind. Everything
in this volume will be free on line not more than 2 12 years after initial
publication, and even before then all readers are encouraged to send
me comments, corrections, complaints, additional solutions, and—
always!—new puzzles.

xiii
 The Puzzles
Bat and Ball
A bat costs $1 more than a ball; together they cost $1.10. How much
does the bat cost?

No Twins Today
It was the first day of class and Mrs. O’Connor had two identical-
looking pupils, Donald and Ronald Featheringstonehaugh (pro-
nounced “Fanshaw”), sitting together in the first row.
“You two are twins, I take it?” she asked.
“No,” they replied in unison.
But a check of their records showed that they had the same parents
and were born on the same day. How could this be?

Portrait
A visitor points to a portrait on the wall and asks who it is. “Brothers
and sisters have I none,” says the host, “but that man’s father is my
father’s son.” Who is pictured?

Half Grown
At what age is the average child half the height that he or she will be
as an adult?

xv
 Mathematical Puzzles

Shoes, Socks, and Gloves

You need to pack for a midnight flight to Iceland but the power is out.
In your closet are six pairs of shoes, six black socks, six gray socks,
six pairs of brown gloves, and six pairs of tan gloves. Unfortunately,
it’s too dark to match shoes or to see any colors.
How many of each of these items do you need to take to be sure
of getting a matched pair of shoes, two socks of the same color, and
matching gloves?

Phone Call
A phone call in the continental United States is made from a west
coast state to an east coast state, and it’s the same time of day at both
ends of the call. How is this possible?

Powers of Two
How many people are “two pairs of twins twice”?

Rotating Coin
While you hold a US 25-cent piece firmly to the tabletop with your
left thumb, you rotate a second quarter with your right forefinger all
the way around the first quarter. Since quarters are ridged, they will
interlock like gears and the second will rotate as it moves around the
first.

How many times will it rotate?

xvi
 The Puzzles

Squaring the Mountain State

Can West Virginia be inscribed in a square?

Natives in a Circle
An anthropologist is surrounded by a circle of natives, each of whom
either always tells the truth or always lies. She asks each native
whether the native to his right is a truth teller or a liar, and from their
answers, she is able to deduce the fraction of liars in the circle.
What fraction is it?

Winning at Wimbledon
As a result of temporary magical powers (which might need to be
able to change your gender), you have made it to the women’s singles
tennis finals at Wimbledon and are playing Serena Williams for all the
marbles. However, your powers cannot last the whole match. What
score do you want it to be when they disappear, to maximize your
chances of notching an upset win?

Polyhedron Faces
Prove that any convex polyhedron has two faces with the same num-
ber of edges.

Finding the Counterfeit

You have a balance scale and 12 coins, 11 of which are genuine and
identical in weight; but one is counterfeit, and is either lighter or heav-
ier than the others. Can you determine, in three weighings with a bal-
ance scale, which coin is counterfeit and whether it is heavy or light?

Signs in an Array
Suppose that you are given an m × n array of real numbers and per-
mitted, at any time, to flip the signs of all the numbers in any row or
column. Can you always arrange matters so that all the row sums and
column sums are non-negative?

xvii
 Mathematical Puzzles

Birthday Match
You are on a cruise where you don’t know anyone else. The ship an-
nounces a contest, the upshot of which is that if you can find someone
who has the same birthday as yours, you (both) win a beef Wellington
dinner.
How many people do you have to compare birthdays with in order
to have a better than 50% chance of success?

Meeting the Ferry

Every day at noon GMT a ferry leaves New York and simultaneously
another leaves Le Havre. Each trip takes seven days and seven nights,
arriving before noon on the eighth day. How many of these cross-
Atlantic ferries does one of them pass on its way across the pond?

Sinking 15
Carol and Desmond are playing pool with billiard balls number 1
through 9. They take turns sinking balls into pockets. The first to sink
three numbers that sum to 15 wins. Does Carol (the first to play) have
a winning strategy?

Monk on a Mountain
A monk begins an ascent of Mt. Fuji on Monday morning, reaching
the summit by nightfall. He spends the night at the summit and starts
down the mountain on the same path the following morning, reaching
the bottom by dusk on Tuesday.
Prove that at some precise time of day, the monk was at exactly
the same spot on the path on Tuesday as he was on Monday.

Mathematical Bookworm
The three volumes of Jacobson’s Lectures in Abstract Algebra sit in order
on your shelf. Each has 2′′ of pages and a front and back cover each
1 ′′ 1 ′′
4 , thus a total width of 2 2 .
A tiny bookworm bores its way straight through from page 1, Vol
I to the last page of Vol III. How far does it travel?

xviii
 The Puzzles

Other Side of the Coin

A two-headed coin, a two-tailed coin, and an ordinary coin are placed
in a bag. One of the coins is drawn at random and flipped; it comes
up “heads.” What is the probability that there is a head on the other
side of this coin?

Slicing the Cube

Before you is a circular saw and a 3 × 3 × 3 wooden cube that you
must cut into twenty-seven 1 × 1 × 1 cubelets. What’s the smallest
number of slices you must make in order to do this? You are allowed
to stack pieces prior to running them through the saw.

Rolling Pencil
A pencil whose cross-section is a regular pentagon has the maker’s
logo imprinted on one of its five faces. If the pencil is rolled on the
table, what is the probability that it stops with the logo facing up?

Watermelons
Yesterday a thousand pounds of watermelons lay in the watermelon
patch. They were 99% water, but overnight they lost moisture to evap-
oration and now they are only 98% water. How much do they weigh
now?

Boy Born on Tuesday

Mrs. Chance has two children of different ages. At least one of them
is a boy born on a Tuesday. What is the probability that both of them
are boys?

xix
 Mathematical Puzzles

Air Routes in Aerostan

The country of Aerostan has three airlines which operate routes be-
tween various pairs of Aerostan’s fifteen cities. What’s the smallest
possible total number of routes in Aerostan if bankruptcy of any of
the three airlines still leaves a network that connects the cities?

Tipping the Scales

A balance scale sits on the desk of the science teacher, Ms. McGregor.
There are weights on the scale, which is currently tipped to the right.
On each weight is inscribed the name of at least one pupil.
On entering the classroom, each pupil moves every weight carry-
ing his or her name to the opposite side of the scale. Must there be a
set of pupils that Ms. McGregor can let in that will tip the scales to
the left?

Chomp
Alice and Bob take turns biting off pieces of an m × n rectangular
chocolate bar marked into unit squares. Each bite consists of selecting
a square and biting off that square plus every remaining square above
and/or to its right. Each player wishes to avoid getting stuck with the
poisonous lower-left square.

Show that, assuming the bar contains more than one square, Alice
(the first to play) has a winning strategy.

Whose Bullet?
Two marksmen, one of whom (“A”) hits a certain small target 75%
of the time and the other (“B”) only 25%, aim simultaneously at that
target. One bullet hits. What’s the probability that it came from A?

xx
 The Puzzles

Poker Quickie
What is the best full house?
(You may assume that you have five cards, and you have just one
opponent who has five random other cards. There are no wild cards.
As a result of the Goddess of Chance owing you a favor, you are en-
titled to a full house, and you get to choose whatever full house you
want.)

Lines Through a Grid

If you want to cover all the vertices in a 10 × 10 square grid by lines
none of which is parallel to a side of the square, how many lines are
needed?

Bidding in the Dark

You have the opportunity to make one bid on a widget whose value
to its owner is, as far as you know, uniformly random between $0 and
$100. What you do know is that you are so much better at operating
the widget than he is, that its value to you is 80% greater than its value
to him.
If you offer more than the widget is worth to the owner, he will
sell it. But you only get one shot. How much should you bid?

Easy Cake Division

Can you cut a cubical cake, iced on top and on the sides, into three
pieces each of which contains the same amount of cake and the same
amount of icing?

Crossing the River

In eighth century Europe, it was considered unseemly for a man to
be in the presence of a married woman, even briefly, unless her hus-
band was there as well. This posed problems for three married cou-
ples who wished to cross a river, the only means being a rowboat that
could carry at most two people. Can they get to the other side with-
out violating their social norms? If so, what’s the minimum number
of crossings needed?

xxi
 Mathematical Puzzles

Bacterial Reproduction
When two pixo-bacteria mate, a new bacterium results; if the parents
are of different sexes the child is female, otherwise it is male. When
food is scarce, matings are random and the parents die when the child
is born.
It follows that if food remains scarce a colony of pixo-bacteria will
eventually reduce to a single bacterium. If the colony originally had
10 males and 15 females, what is the probability that the ultimate
pixo-bacterium will be female?

Sprinklers in a Field
Sprinklers in a large field are located at the vertices of a square grid.
Each point of land is supposed to be watered by exactly the three
closest sprinklers. What shape is covered by each sprinkler?

Bags of Marbles
You have 15 bags. How many marbles do you need so that you can
have a different number of marbles in each bag?

Men with Sisters

On average, who has more sisters, men or women?

Dots and Boxes Variation

Suppose you are playing Dots and Boxes, but with each player hav-
ing an option of whether to go again after making a box. Suppose the
board has an odd number of cells (thus, someone must win). Show
that the first player to make a box has a winning strategy.

Fair Play
How can you get a 50-50 decision by flipping a bent coin?

xxii
 The Puzzles

Salaries and Raises

Wendy, Monica, and Yancey were hired at the beginning of calendar
2020.
Wendy is paid weekly, $500 per week, but gets a $5 raise each
week; Monica is paid monthly, $2500 per month, and gets a $50 raise
each month; Yancey is paid yearly, $50,000 per year, and gets a $1500
raise each year.
Who will make the most money in 2030?

Two Runners
Two runners start together on a circular track, running at different
constant speeds. If they head in opposite directions they meet after a
minute; if they head in the same direction, they meet after an hour.
What is the ratio of their speeds?

Broken ATM
George owns only $500, all in a bank account. He needs cash badly
but his only option is a broken ATM that can process only with-
drawals of exactly $300, and deposits of exactly $198. How much cash
can George get out of his account?

Domino Task
An 8×8 chessboard is tiled arbitrarily with thirty-two 2×1 dominoes.
A new square is added to the right-hand side of the board, making the
top row length 9.
At any time you may move a domino from its current position
to a new one, provided that after the domino is lifted, there are two
adjacent empty squares to receive it.
Can you retile the augmented board so that all the dominoes are
horizontal?

Big Pairs in a Matrix

The sum of the greatest two numbers in each row of a certain square
matrix is r; the sum of the greatest two numbers in each column is c.
Show that r = c.

xxiii
 Mathematical Puzzles

Second Ace
What is the probability that a poker hand (5 cards dealt at random
from a 52-card deck) contains at least two aces, given that it has at
least one? What is the probability that it contains two aces, given that
it has the Ace of Spades? If your answers are different, then: What’s
so special about the Ace of Spades?

Fewest Slopes
If you pick n random points in, say, a disk, the pairs of points among
them will with probability 1 determine n(n−1)/2 distinct slopes. Sup-
pose you get to pick the n points deliberately, subject to no three being
collinear. What’s the smallest number of distinct slopes they can de-
termine?

Three Natives at the Crossroads

A logician is visiting the South Seas and as is usual for logicians in
puzzles, she is at a fork, wanting to know which of two roads leads to
the village. Present this time are three willing natives, one each from a
tribe of invariable truth-tellers, a tribe of invariable liars, and a tribe of
random answerers. Of course, the logician doesn’t know which native
is from which tribe. Moreover, she is permitted to ask only two yes-
or-no questions, each question being directed to just one native. Can
she get the information she needs? How about if she can ask only one
yes-or-no question?

Rolling All the Numbers

On average, how many times do you need to roll a die before all six
different numbers have turned up?

Uniform Unit Distances

Can it be that for any positive integer n, there’s a configuration of
points on the plane with the property that every point is at distance 1
from exactly n other points?

xxiv
 The Puzzles

Life Is a Bowl of Cherries

In front of you and your friend Amit are 4 bowls of cherries, contain-
ing, respectively, 5, 6, 7, and 8 cherries. You and Amit will alternately
pick a bowl and take one or more cherries from it. If you go first, and
you want to be sure to get the last cherry, how many cherries should
you take—and from which bowl?

Finding the Missing Number

All but one of the numbers from 1 to 100 are read to you, one every 10
seconds, but in no particular order. You have a good mind, but only a
normal memory, and no means of recording information during the
process. How can you ensure that you can determine afterward which
number was not called out?

Three-Way Duel
Alice, Bob, and Carol arrange a three-way duel. Alice is a poor shot,
hitting her target only 1/3 of the time on average. Bob is better, hitting
his target 2/3 of the time. Carol is a sure shot.
They take turns shooting, first Alice, then Bob, then Carol, then
back to Alice, and so on until only one is left. What is Alice’s best
course of action?

Splitting a Hexagon
Is there a hexagon that can be cut into four congruent triangles by a
single line?

xxv
 Mathematical Puzzles

Expecting the Worst

Choose n numbers uniformly at random from the unit interval [0,1].
What is the expected value of their minimum?

Belt Around the Earth

Suppose the earth is a perfect sphere and a belt is tightened around
the equator. Then the belt is loosened by adding 1 meter to its length,
and it now sits the same amount off the ground all the way around.
Is there enough room to slip a credit card underneath the belt?

Stopping After the Boy

In a certain country, a law is passed forbidding families to have an-
other child after having a boy. Thus, families might consist of one
boy, one girl and one boy, five girls and one boy. . . How will this law
affect the ratio of boys to girls?

Attic Lamp
An old-fashioned incandescent lamp in the attic is controlled by one
of three on-off switches downstairs—but which one? Your mission is
to do something with the switches, then determine after one trip to the
attic which switch is connected to the attic lamp.

Efficient Pizza-Cutting
What’s the maximum number of pieces you can get by cutting a
(round) pizza with 10 straight cuts?

Fourth Corner
Pegs occupy three corners of a square. At any time a peg can jump
over another peg, landing an equal distance on the other side. Jumped
pegs are not removed. Can you get a peg onto the fourth corner of the
square?

xxvi
 The Puzzles

Cutting the Necklace

Two thieves steal a necklace consisting of 10 red rubies and 14 pink
diamonds, fixed in some arbitrary order on a loop of golden string.
Show that they can cut the necklace in two places so that when each
thief takes one of the resulting pieces, he gets half the rubies and half
the diamonds.

Odd Run of Heads

On average, how many flips of a coin does it take to get a run of an
odd number of heads, preceded and followed by tails?

Subsets with Constraints

What is the maximum number of numbers you can have between 1
and 30, such that no two have a product which is a perfect square?
How about if (instead) no number divides another evenly? Or if no
two have a factor (other than 1) in common?

Spinning Switches
Four identical, unlabeled switches are wired in series to a light bulb.
The switches are simple buttons whose state cannot be directly ob-
served, but can be changed by pushing; they are mounted on the
corners of a rotatable square. At any point, you may push, simul-
taneously, any subset of the buttons, but then an adversary spins the
square. Is there an algorithm that will enable you to turn on the bulb
in at most a fixed number of spins?

Watches on the Table

Fifty accurate watches and a tiny diamond lie on a table. Prove that
there exists a moment in time when the sum of the distances from
the diamond to the ends of the minute hands exceeds the sum of the
distances from the diamond to the centers of the watches.

Bugging a Disk
Elspeth, a new FBI recruit, has been asked to bug a circular room
using seven ceiling microphones. If the room is 40 feet in diameter

xxvii
 Mathematical Puzzles

and the bugs are placed so as to minimize the maximum distance

from any point in the room to the nearest bug, where should the bugs
be located?

Points on a Circle
What is the probability that three uniformly random points on a circle
will be contained in some semicircle?

Two Different Distances

Find all configurations of four points in the plane that determine only
two different distances. (Note: there are more of these than you prob-
ably think!)

Same Sum Subsets

Amy asks Brad to pick 10 different numbers between 1 and 100, and
to write them down secretly on a piece of paper. She now tells him
she’s willing to bet $100 to $1 that his numbers contain two nonempty
disjoint subsets with the same sum! Is she nuts?

Players and Winners

Tristan and Isolde expect to be in a situation of severely limited com-
munication, at which time Tristan will know which two of 16 basket-
ball teams played a game, and Isolde will know who won. How many
bits must be communicated between Tristan and Isolde in order for
the former to find out who won?

Spaghetti Loops
The 100 ends of 50 strands of cooked spaghetti are paired at random
and tied together. How many pasta loops should you expect to result
from this process, on average?

xxviii
 The Puzzles

Swapping Executives
The executives of Women in Action, Inc., are seated at a long ta-
ble facing the stockholders. Unfortunately, according to the meeting
organizer’s chart, every one is in the wrong seat. The organizer can
persuade two executives to switch seats, but only if they are adjacent
and neither one is already in her correct seat.
Can the organizer organize the seat-switching so as to get every-
one in her correct seat?

Life Isn't a Bowl of Cherries?

In front of you and your friend Amit are 4 bowls of cherries, contain-
ing, respectively, 5, 6, 7, and 8 cherries. You and Amit will alternately
pick a bowl and take one or more cherries from it. If you go first, and
you want to be sure Amit gets the last cherry, how many cherries
should you take—and from which bowl?

Righting the Pancakes

An underchef of the great and persnickety Chef Bouillon has made
a stack of pancakes, but, alas, some of them are upside-down—that
is, according to the great Chef, they don’t have their best side up. The
underchef wants to fix the problem as follows: He finds a contiguous
substack of pancakes (of size at least one) with the property that both
the top pancake and the bottom pancake of the substack are upside-
down. Then he removes the substack, flips it as a block and slips it
back into the big stack in the same place.

Prove that this procedure, no matter which upside-down pancakes

are chosen, will eventually result in all the pancakes being right side
up.

xxix
 Mathematical Puzzles

Pie in the Sky

What fraction of the sky is occupied by a full moon?

Testing Ostrich Eggs

In preparation for an ad campaign, the Flightless Ostrich Farm needs
to test its eggs for durability. The world standard for egg-hardness calls
for rating an egg according to the highest floor of the Empire State
Building from which the egg can be dropped without breaking.
Flightless’ official tester, Oskar, realizes that if he takes only one
egg along on his trip to New York, he’ll need to drop it from (poten-
tially) every one of the building’s 102 floors, starting with the first, to
determine its rating.
How many drops does he need in the worst case, if he takes two
eggs?

Circular Shadows I
Suppose all three coordinate-plane projections of a convex solid are
disks. Must the solid be a perfect ball?

Next Card Red

Paula shuffles a deck of cards thoroughly, then plays cards face up one
at a time, from the top of the deck. At any time, Victor can interrupt
Paula and bet $1 that the next card will be red. He bets once and only
once; if he never interrupts, he’s automatically betting on the last card.
What’s Victor’s best strategy? How much better than even can he
do? (Assume there are 26 red and 26 black cards in the deck.)

Attention Paraskevidekatriaphobes
Is the 13th of the month more likely to be a Friday than any other day
of the week, or does it just seem that way?

Unanimous Hats
One hundred prisoners are offered the chance to be freed if they can
win the following game. In the dark, each will be fitted with a red or

xxx
 The Puzzles

black hat according to a fair coinflip. When the lights are turned on,
each will see the others’ hat colors but not his own; no communication
between prisoners will be permitted.
Each prisoner will be asked to write down his guess at his own hat
color, and the prisoners will be freed if they all get it right.
The prisoners have a chance to conspire beforehand. Can you
come up with a strategy for them that will maximize their probability
of winning?

Raising Art Value

Fans of Gallery A like to say that last year, when Gallery A sold to
Gallery B “Still Life with Kumquats,” the average value of the pic-
tures in each gallery went up. Assuming that statement is correct, and
that the value of each of the 400 pictures in the galleries’ combined
holdings is an integral number of dollars, what is the least possible dif-
ference between the average picture values of Gallery A and Gallery
B before the sale?

First Odd Number

What’s the first odd number in the dictionary? More specifically, sup-
pose that every whole number from 1 to, say, 1010 is written out in for-
mal English (e.g., “two hundred eleven,” “one thousand forty-two”)
and then listed in dictionary order, that is, alphabetical order with
spaces and hyphens ignored. What’s the first odd number in the list?

Lattice Points and Line Segments

How many lattice points (i.e., points with integer coordinates) can
you have in 3-space with the property that no two of them have a
third lattice point on the line segment between them?

Breaking a Chocolate Bar

You have a rectangular chocolate bar scored in a 6 × 4 grid of squares,
and you wish to break up the bar into its constituent squares. At each
step, you may pick up one piece and break it along any of its marked
vertical or horizontal lines.

xxxi
 Mathematical Puzzles

For example, you can break three times to form four rows of
six squares each, then break each row five times into its constituent
squares, accomplishing the desired task in 3 + 4 × 5 = 23 breaks.
Can you do better?

Meet the Williams Sisters

Some tennis fans get excited when Venus and Serena Williams meet
in a tournament. The likelihood of that happening normally depends
on seeding and talent, so let’s instead construct an idealized elimi-
nation tournament of 64 players, each as likely to win as to lose any
given match, with bracketing chosen uniformly at random. What is
the probability that the Williams sisters wind up playing each other?

Cards from Their Sum

The magician riffle-shuffled the cards and dealt five of them face
down onto your outstretched hand. She asked you to pick any number
of them secretly, then tell you their sum (jacks count as 11, queens 12,
kings 13). You did that and the magician then told you exactly what
cards you chose, including the suits!

After a while you figured out that her shuffles were defective—
they didn’t change the top five cards, so the magician had picked them
in advance and knew what they were. But she still had to pick the
ranks of those cards carefully (and memorize their suits), so that from
the sum she could always recover the ranks.
In conclusion, if you want to do the trick yourself, you need to find
five distinct numbers between 1 and 13 (inclusive), with the property
that every subset has a different sum. Can you do it?

Waiting for Heads

If you flip a fair coin repeatedly, how long do you have to wait, on
average, for a run of five heads?

xxxii
 The Puzzles

Half-right Hats
One hundred prisoners are offered the chance to be freed if they can
win the following game. In the dark, each will be fitted with a red or
black hat according to a fair coinflip. When the lights are turned on,
each will see the others’ hat colors but not his own; no communication
between prisoners will be permitted.
Each prisoner will be asked to write down his guess at his own hat
color, and the prisoners will be freed if at least half get it right.
The prisoners have a chance to conspire beforehand. Can you
come up with a strategy for them that will maximize their probability
of winning?

Finding a Jack
In some poker games the right of first dealer is determined by dealing
cards face-up (from a well-shuffled deck) until some player gets a jack.
On average, how many cards are dealt during this procedure?

Measuring with Fuses

You have on hand two fuses (lengths of string), each of which will
burn for exactly 1 minute. They do not, however, burn uniformly
along their lengths. Can you nevertheless use them to measure 45
seconds?

Rating the Horses

You have 25 horses and can race them in groups of five, but having
no stopwatch, you can only observe the order of finish. How many
heats of five horses do you need to determine the fastest three of your
25?

Red and Blue Hats in a Line

There are n prisoners this time, and again each will be fitted with a
red or black hat according to flips of a fair coin. The prisoners are to
be arranged in a line, so that each prisoner can see only the colors
of the hats in front of him. Each prisoner must guess the color of
his own hat, and is executed if he is wrong; however, the guesses are
made sequentially, from the back of the line toward the front. Thus,

xxxiii
 Mathematical Puzzles

for example, the ith prisoner in line sees the hat colors of prisoners
1, 2, . . . , i−1 and hears the guesses of prisoners n, n−1, . . . , i+1
(but he isn’t told which of those guesses were correct—the executions
take place later).
The prisoners have a chance to collaborate beforehand on a strat-
egy, with the object of guaranteeing as many survivors as possible.
How many players can be saved in the worst case?

Three Sticks
You have three sticks that can’t make a triangle; that is, one is longer
than the sum of the lengths of the other two. You shorten the long
one by an amount equal to the sum of the lengths of the other two,
so you again have three sticks. If they also fail to make a triangle, you
again shorten the longest stick by an amount equal to the sum of the
lengths of the other two.
You repeat this operation until the sticks do make a triangle, or
the long stick disappears entirely.
Can this process go on forever?

Spiders on a Cube
Three spiders are trying to catch an ant. All are constrained to the
edges of a cube. Each spider can move one third as fast as the ant can.
Can the spiders catch the ant?

Hopping and Skipping

A frog hops down a long line of lily pads; at each pad, he flips a coin
to decide whether to hop two pads forward or one pad back. What
fraction of the pads does he hit?

xxxiv
 The Puzzles

Divisibility Game
Alice chooses a number bigger than 100 and writes it down secretly.
Bob now guesses a number greater than 1, say k; if k divides Alice’s
number, Bob wins. Otherwise k is subtracted from Alice’s number
and Bob tries again, but may not use a number he used before. This
continues until either Bob succeeds by finding a number that divides
Alice’s, in which case Bob wins, or Alice’s number becomes 0 or neg-
ative, in which case Alice wins.
Does Bob have a winning strategy for this game?

Candles on a Cake
It’s Joanna’s 18th birthday and her cake is cylindrical with 18 candles
on its 18′′ circumference. The length of any arc (in inches) between
two candles is greater than the number of candles on the arc, exclud-
ing the candles at the ends.
Prove that Joanna’s cake can be cut into 18 equal wedges with a
candle on each piece.

Sums of Two Squares

On average, how many ways are there to write a positive number n as
the sum of two squares? In other words, suppose n is a random integer
between 1 and a zillion. What is the expected number of ordered pairs
(i, j) of integers such that n = i2 + j 2 ?

Pairs at Maximum Distance

Let X be a finite set of points on the plane. Suppose X contains n
points, and the maximum distance between any two of them is d.
Prove that at most n pairs of points of X are at distance d.

Comparing Numbers, Version I

Paula (the perpetrator) takes two slips of paper and writes an integer
on each. There are no restrictions on the two numbers except that
they must be different. She then conceals one slip in each hand.
Victor (the victim) chooses one of Paula’s hands, which Paula then
opens, allowing Victor to see the number on that slip. Victor must now

xxxv
 Mathematical Puzzles

guess whether that number is the larger or the smaller of Paula’s two
numbers; if he guesses right, he wins $1, otherwise he loses $1.
Clearly, Victor can at least break even in this game, for example,
by flipping a coin to decide whether to guess “larger” or “smaller”—
or by always guessing “larger,” but choosing the hand at random. The
question is: Not knowing anything about Paula’s psychology, is there
any way he can do better than break even?

Comparing Numbers, Version II

Now let’s make things more favorable to Victor: Instead of being cho-
sen by Paula, the numbers are chosen independently at random from
the uniform distribution on [0,1] (two outputs from a standard ran-
dom number generator will do fine).
To compensate Paula, we allow her to examine the two random
numbers and to decide which one Victor will see. Again, Victor must guess
whether the number he sees is the larger or smaller of the two, with $1
at stake. Can he do better than break even? What are his and Paula’s
best (i.e., “equilibrium”) strategies?

King's Salary
Democracy has come to the little kingdom of Zirconia, in which the
king and each of the other 65 citizens has a salary of one zircon. The
king cannot vote, but he has power to suggest changes—in particu-
lar, redistribution of salaries. Each person’s salary must be a whole
number of zircons, and the salaries must sum to 66. Each suggestion
is voted on, and carried if there are more votes for than against. Each
voter can be counted on to vote “yes” if his or her salary is to be in-
creased, “no” if decreased, and otherwise not to bother voting.
The king is both selfish and clever. What is the maximum salary
he can obtain for himself, and how many referenda does he need to
get it?

Adding, Multiplying, and Grouping

Forty-two positive integers (not necessarily distinct) are written in a
row. Show that you can put plus signs, times signs, and parentheses
between the integers in such a way that the value of the resulting ex-
pression is evenly divisible by one million.

xxxvi
 The Puzzles

Missing Card
Yola and Zela have devised a clever card trick. While Yola is out of the
room, audience members pull out five cards from a bridge deck and
hand them to Zela. She looks them over, pulls one out, and calls Yola
into the room. Yola is handed the four remaining cards and proceeds
to guess correctly the identity of the pulled card.
How do they do it? And once you’ve figured that out, compute
the size of the biggest deck of cards they could use and still perform
the trick reliably.

Ping-Pong Progression
Alice and Bob play table tennis, with Bob’s probability of winning
any given point being 30%. They play until someone reaches a score
of 21. What, approximately, is the expected number of points played?

Odd Light Flips

Suppose you are presented with a collection of light bulbs controlled
by switches. Each switch flips the state of some subset of the bulbs,
that is, turns on all the ones in the subset that were off, and turns off
those that were on. You are told that for any nonempty set of bulbs,
there is a switch that flips an odd number of bulbs from that subset (and
perhaps other bulbs as well).
Show that, no matter what the initial state of the bulbs, you can
use these switches to turn off all the bulbs.

Painting the Cubes

Can you paint 1000 unit cubes with 10 different colors in such a way
that for any of the 10 colors, the cubelets can be arranged into a 10 ×
10 × 10 cube with only that color showing?

Red Points and Blue

Given n red points and n blue points on the plane, no three on a line,
prove that there is a matching between them so that line segments
from each red point to its corresponding blue point do not cross.

xxxvii
 Mathematical Puzzles

Identifying the Majority

A long list of names is read out, some names many times. Your object
is to end up with a name that is guaranteed to be the name which was
a called a majority of the time, if there is such a name.
However, you have only one counter, plus the ability to keep just
one name at a time in your mind. Can you do it?

Returning Pool Shot

A ball is shot from a corner of a polygonal pool table (not necessarily
convex) with right-angle corners; let’s say all sides are aligned either
exactly east-west or exactly north-south.
The starting corner is a convex corner, that is, has an interior angle
of 90◦ . There are pockets at all corners, so that if the ball hits a corner
exactly it falls in. Otherwise, the ball bounces true and without energy
loss.
Can the ball ever return to the corner from which it began?

Pancake Stacks
At the table are two hungry students, Andrea and Bruce, and two
stacks of pancakes, of height m and n. Each student, in turn, must eat
from the larger stack a non-zero multiple of the number of pancakes
in the smaller stack. Of course, the bottom pancake of each stack is
soggy, so the player who first finishes a stack is the loser.
For which pairs (m, n) does Andrea (who plays first) have a win-
ning strategy?
How about if the game’s objective is reversed, so that the first
player to finish a stack is the winner?

Trapped in Thickland
The inhabitants of Thickland, a world somewhere between Edwin
Abbott’s Flatland and our three-dimensional universe, are an infi-
nite set of congruent convex polyhedra that live between two parallel
planes. Up until now, they have been free to escape from their slab,
but haven’t wanted to. Now, however, they have been reproducing
rapidly and thinking about colonizing other slabs. Their high priest is
worried that conditions are so crowded, no inhabitant of Thickland
can escape the slab unless others move first.
Is that even possible?

xxxviii
 The Puzzles

Hats and Infinity

Each of an infinite collection of prisoners, numbered 1, 2, . . . , is to be
fitted with a red or black hat. At a prearranged signal, all the prisoners
are revealed to one another, so that everyone gets to see all his fellow
prisoners’ hat colors—but no communication is permitted. Each pris-
oner is then taken aside and asked to guess the color of his own hat.
All the prisoners will be freed provided only finitely many guess
wrongly. The prisoners have a chance to conspire beforehand; is there
a strategy that will ensure freedom?

All Right or All Wrong

This time the circumstances are the same but the objective is different:
The guesses must either be all right or all wrong. Is there a winning
strategy?

Magnetic Dollars
One million magnetic “susans” (Susan B. Anthony dollar coins) are
tossed into two urns in the following fashion: The urns begin with
one coin in each, then the remaining 999,998 coins are thrown in
the air one by one. If there are x coins in one urn and y in the
other, magnetic attraction will cause the next coin to land in the first
urn with probability x/(x + y), and in the second with probability
y/(x + y).
How much should you be willing to pay, in advance, for the con-
tents of the urn that ends up with fewer susans?

Shoelaces at the Airport

You are walking to your gate at O’Hare and need to tie your shoelace.
Up ahead is a moving walkway that you plan to utilize. To minimize
your time to get to the gate, should you stop and tie your shoelace
now, or wait to tie it on the walkway?

Hazards of Electronic Coinflipping

You have been hired as an arbitrator and called upon to award a cer-
tain indivisible widget randomly to either Alice, Bob, or Charlie, each
with probability 13 . Fortunately, you have an electronic coinflipping

xxxix
 Mathematical Puzzles

device with an analog dial that enables you to enter any desired prob-
ability p. Then you push a button and the device shows “Heads” with
probability p, else “Tails.”
Alas, your device is showing its “low battery” light and warning
you that you may set the probability p only once, and then push the
coinflip button at most 10 times. Can you still do your job?

Handshakes at a Party
Nicholas and Alexandra went to a reception with 10 other couples;
each person there shook hands with everyone he or she didn’t know.
Later, Alexandra asked each of the other 21 partygoers how many
people they shook hands with, and got a different answer every time.
How many people did Nicholas shake hands with?

Area−Perimeter Match
Find all integer-sided rectangles with equal area and perimeter.

Prime Test
Does 49 + 610 + 320 happen to be a prime number?

Lost Boarding-Pass
One hundred people line up to board a full jetliner, but the first has
lost his boarding pass and takes a random seat instead. Each subse-
quent passenger takes his or her assigned seat if available, otherwise
a random unoccupied seat.
What is the probability that the last passenger to board finds his
seat unoccupied?

Lemming on a Chessboard
On each square in an n × n chessboard is an arrow pointing to one of
its eight neighbors (or off the board, if it’s an edge square). However,
arrows in neighboring squares (diagonal neighbors included) may not
differ in direction by more than 45 degrees.
A lemming begins in a center square, following the arrows from
square to square. Is he doomed to fall off the board?

xl
 The Puzzles

Packing Slashes
Given a 5 × 5 square grid, on how many of the squares can you draw
diagonals (slashes or backslashes) in such a way that no two of the
diagonals meet?

Peek Advantage
You are about to bet $100 on the color of the top card of a well-shuffled
deck of cards. You get to pick the color; if you’re right you win $100,
otherwise you lose the same.
How much is it worth to you to get a peek at the bottom card of
the deck? How much more for a peek at the bottom two cards?

Truly Even Split

Can you partition the integers from 1 to 16 into two sets of equal sizes
so that each set has the same sum, the same sum of squares, and the
same sum of cubes?

Line Up by Height
Yankees manager Casey Stengel famously once told his players to
“line up alphabetically by height.” Suppose 26 players, no two ex-
actly the same height, are lined up alphabetically. Prove that there
are at least six who are also in height order—either tallest to shortest,
or shortest to tallest.

Curves on Potatoes
Given two potatoes, can you draw a closed curve on the surface
of each so that the two curves are identical as curves in three-
dimensional space?

Falling Ants
Twenty-four ants are placed randomly on a meter-long rod; each ant
is facing east or west with equal probability. At a signal, they proceed
to march forward (that is, in whatever direction they are facing) at 1
cm/sec; whenever two ants collide, they reverse directions. How long
does it take before you can be certain that all the ants are off the rod?

xli
 Mathematical Puzzles

Polygon Midpoints
Let n be an odd integer, and let a sequence of n distinct points be given
in the plane. Find the vertices of a (possibly self-intersecting) n-gon
that has the given points, in the given order, as midpoints of its sides.

Rows and Columns

Prove that if you sort each row of a matrix, then each column, the
rows are still sorted!

Bugs on a Pyramid
Four bugs live on the four vertices of a triangular pyramid (tetrahe-
dron). Each bug decides to go for a little walk on the surface of the
tetrahedron. When they are done, two of them are back home, but
the other two find that they have switched vertices.
Prove that there was an instant when all four bugs lay on the same
plane.

Snake Game
Joan begins by marking any square of an n × n chessboard; Judy then
marks an orthogonally adjacent square. Thereafter, Joan and Judy
continue alternating, each marking a square adjacent to the last one
marked, creating a snake on the board. The first player unable to play
loses.
For which values of n does Joan have a winning strategy, and
when she does, what square does she begin at?

Three Negatives
A set of 1000 integers has the property that every member of the set
exceeds the sum of the rest. Show that the set includes at least three
negative numbers.

Numbers on Foreheads
Each of 10 prisoners will have a digit between 0 and 9 painted on his
forehead (they could be all 2’s, for example). At the appointed time
each will be exposed to all the others, then taken aside and asked to
guess his own digit.

xlii
 The Puzzles

In order to avoid mass execution, at least one prisoner must guess

correctly. The prisoners have an opportunity to conspire beforehand;
find a scheme by means of which they can ensure success.

Bias Test
Before you are two coins; one is a fair coin, and the other is biased
toward heads. You’d like to try to figure out which is which, and to
do so you are permitted two flips. Should you flip each coin once, or
one coin twice?

Non-Repeating String
Is there a finite string of letters from the Latin alphabet with the prop-
erty that there is no pair of adjacent identical substrings, but the ad-
dition of any letter to either end would create one?

Three-way Election
Alison, Bonnie, and Clyde run for class president and finish in a three-
way tie. To break it, they solicit their fellow students’ second choices,
but again there is a three-way tie. The election committee is stymied
until Alison steps forward and points out that, since the number of
voters is odd, they can make two-way decisions. She therefore pro-
poses that the students choose between Bonnie and Clyde, and then
the winner would face Alison in a runoff.
Bonnie complains that this is unfair because it gives Alison a better
chance to win than either of the other two candidates. Is Bonnie right?

Skipping a Number
At the start of the 2019 season WNBA star Missy Overshoot’s lifetime
free-throw percentage was below 80%, but by the end of the season it
was above 80%. Must there have been a moment in the season when
Missy’s free-throw percentage was exactly 80%?

Red and Black Sides

Bored in math class, George draws a square on a piece of paper and
divides it into rectangles (with sides parallel to a side of the square),

xliii
 Mathematical Puzzles

using a red pen after his black one runs out of ink. When he’s done,
he notices that every rectangular tile has at least one totally red side.
Prove that the total length of George’s red lines is at least the length
of a side of the big square.

Unbroken Lines
Can you re-arrange the 16 square tiles below into a 4 × 4 square—
no rotation allowed!—in such a way that no line ends short of the
boundary of the big square?

Baby Frog
To give a baby frog jumping practice, her four grandparents station
themselves at the corners of a large square field. When a grandpar-
ent croaks, the baby leaps halfway to it. In the field is a small round
clearing. Can the grandparents get the baby to that clearing, no matter
where in the field she starts?

Sequencing the Digits

In how many ways can you write the numbers 0 through 9 in a row,
such that each digit other than the leftmost is within one of some digit
to the left of it?

xliv
 The Puzzles

Recovering the Polynomial

The Oracle at Delphi has in mind a certain polynomial p (in the vari-
able x, say) with non-negative integer coefficients. You may query
the Oracle with any integer x, and the Oracle will tell you the value
of p(x).
How many queries do you have to make to determine p?

Faulty Combination Lock

A combination lock with three dials, each numbered 1 through 8, is
defective in that you only need to get two of the numbers right to open
the lock. What is the minimum number of (three-number) combina-

tions you need to try in order to be sure of opening the lock?

Prisoners and Gloves

Each of 100 prisoners gets a different real number written on his fore-
head, a black glove, and a white glove. Seeing the numbers on the
other prisoners’ foreheads, he must put one glove on each hand in
such a way that when the prisoners are later lined up in real-number
order and hold hands, the gloves match. How can the prisoners, who
are permitted to conspire beforehand, ensure success?

Gaming the Quilt

You come along just as a church lottery is closing; its prize is a quilt
worth $100 to you, but they’ve only sold 25 tickets. At $1 a ticket,
how many tickets should you buy?

xlv
 Mathematical Puzzles

Bracing the Grid

Suppose that you are given an n × n grid of unit-length rods, jointed
at their ends. You may brace some
√ subset S of the small squares with
diagonal segments (of length 2).
Which choices of S suffice to make the grid rigid in the plane?

Rectangles Tall and Wide

It is easy to check that a 6 × 5 rectangle can be tiled with 3 × 2 rect-
angles, but only if some tiles are horizontally oriented and others are
vertically oriented.
Show that this can’t happen if the big rectangle is a square—in
other words, if a square is tiled with congruent rectangles, then it can
be re-tiled with the same rectangles all oriented the same way.

Guarding the Gallery

A certain museum room is shaped like a highly irregular, non-convex,
11-sided polygon. How many guard-posts are needed in the room, in
the worst case, to ensure that every part of the room can be seen from
at least one guard-post?

Flying Saucers
A fleet of saucers from planet Xylofon has been sent to bring back
the inhabitants of a certain randomly selected house, for exhibition in
the Xylofon Xoo. The house happens to contain five men and eight
women, to be beamed up randomly one at a time.

xlvi
 The Puzzles

Owing to the Xylofonians’ strict sex separation policy, a saucer

cannot bring back earthlings of both sexes. Thus, it beams people up
until it gets a member of a second sex, at which point that one is
beamed back down and the saucer takes off with whatever it has left.
Another saucer then starts beaming people up, following the same
rule, and so forth.
What is the probability that the last person beamed up is a
woman?

Increasing Routes
On the Isle of Sporgesi, each segment of road (between one intersec-
tion and the next) has its own name. Let d be the average number of
road segments meeting at an intersection. Show that you can take a
drive on the Isle of Sporgesi that covers at least d road segments, and
hits those segments in strict alphabetical order!

Curve on a Sphere
Prove that if a closed curve on the unit sphere has length less than 2π,
then it is contained in some hemisphere.

Biased Betting
Alice and Bob each have $100 and a biased coin that comes up heads
with probability 51%. At a signal, each begins flipping his or her coin
once a minute and bets $1 (at even odds) on each outcome, against a
bank with unlimited funds. Alice bets on heads, Bob on tails. Suppose
both eventually go broke. Who is more likely to have gone broke first?
Suppose now that Alice and Bob are flipping the same coin, so that
when the first one goes broke the second one’s stack will be at $200.
Same question: Given that they both go broke, who is more likely to
have gone broke first?

Ascending and Descending

Fix a number n and call a list of the numbers from 1 to n “good” if
there is no descending subsequence of length 10. Show that there are
at most 81n good lists.

xlvii
 Mathematical Puzzles

Unbroken Curves
Can you re-arrange the 16 square tiles below into a 4 × 4 square—
no rotation allowed!—in such a way that no curve ends short of the
boundary of the big square?

Dot-Town Exodus
Each resident of Dot-town carries a red or blue dot on his (or her) fore-
head, but if he ever thinks he knows what color it is he leaves town
immediately and permanently. Each day the residents gather; one
day a stranger comes and tells them something—anything—nontrivial
about the number of blue dots. Prove that eventually Dot-town be-
comes a ghost town, even if everyone can see that the stranger’s statement
is patently false.

Drawing Socks
You have 60 red and 40 blue socks in a drawer, and you keep drawing
a sock uniformly at random until you have drawn all the socks of one
color. What is the expected number of socks left in the drawer?

Path Through the Cells

Cells in a certain cellular telephone network are assigned frequencies
in such a way that no two adjacent cells use the same frequency. Show
that if the number of frequencies used is as small as possible (subject to
this condition), then it’s possible to design a path that moves from cell
to adjacent cell and hits each frequency exactly once—in ascending
order of frequencies!

xlviii
 The Puzzles

Painting the Polyhedron

Let P be a polyhedron with red and green faces such that every red
face is surrounded by green ones, but the total red area exceeds the
total green area. Prove that you can’t inscribe a sphere in P .

Whim-Nim
You and a friend, bored with Nim and Nim Misére, decide to play a
variation in which at any point, either player may declare “Nim” or
“Misère” instead of removing chips. This happens at most once in a
game, and then of course the game proceeds normally according to
that variation of Nim. (Taking the whole single remaining stack in an
undeclared game loses, as your opponent can then declare “Nim” as
his last move.)
What’s the correct strategy for this game, which its inventor, the
late John Horton Conway, called “Whim”?

Bacteria on the Plane

Suppose the world begins with a single bacterium at the origin of the
infinite plane grid. When it divides its two successors move one vertex
north and one vertex east, so that there are now two bacteria, one at
(0,1) and one at (1,0). Bacteria continue to divide, each time with one
successor moving north and one east, provided both of those points
are unoccupied.
Show that no matter how long this process continues, there’s al-
ways a bacterium inside the circle of radius 3 about the origin.

Poorly Placed Dominoes

What’s the smallest number of dominoes one can place on a chess-
board (each covering two adjacent squares) so that no more fit?

Ants on the Circle

Twenty-four ants are randomly placed on a circular track of length 1
meter; each ant faces randomly clockwise or counterclockwise. At a
signal, the ants begin marching at 1 cm/sec; when two ants collide
they both reverse directions. What is the probability that after 100
seconds, every ant finds itself exactly where it began?

xlix
 Mathematical Puzzles

Chinese Nim
On the table are two piles of beans. Alex must either take some beans
from one pile or the same number of beans from each pile; then Beth
has the same options. They continue alternating until one wins the
game by taking the last bean.
What’s the correct strategy for this game? For example, if Alex is
faced with piles of size 12,000 and 20,000, what should he do? How
about 12,000 and 19,000?

Not Burning Brownies

When you bake a pan of brownies, those that share an edge with the
edge of the pan will often burn. For example, if you bake 16 square
brownies in a square pan, 12 are subject to burning. Design a pan and
divide it into brownie shapes to make 16 identically shaped brownies
of which as few as possible will burn. Can you get it down to four
burned brownies? Great! How about three?

Double Cover by Lines

Let Lθ be the set of all lines on the plane at angle θ to the horizontal.
If θ and θ′ are two different angles, the union of the sets Lθ and Lθ′
constitutes a double cover of the plane, that is, every point belongs to
exactly two lines.
Can this be done in any other way? That is, can you cover each
point of the plane exactly twice using a set of lines that contains lines
in more than two different directions?

Integral Rectangles
A rectangle in the plane is partitioned into smaller rectangles, each of
which has either integer height or integer width (or both). Prove that
the large rectangle itself has this property.

Losing at Dice
Visiting Las Vegas, you are offered the following game. Six dice are
to be rolled, and the number of different numbers that appear will be
counted. That could be any number from one to six, of course, but
they are not equally likely.

l
 The Puzzles

If you get the number “four” this way you win $1, otherwise you
lose $1. You decide you like this game, and plan to play it repeatedly
until the $100 you came with is gone.
At one minute per game, how long, on the average, will it take
before you are wiped out?

Even-Sum Billiards
You pick 10 times, with replacement (necessarily!), from an urn con-
taining 9 billiard balls numbered 1 through 9. What is the probability
that the sum of the numbers of the balls you picked is even?

Competing for Programmers

Two Silicon Valley startups are competing to hire programmers. Hir-
ing on alternate days, each can begin by hiring anyone, but every sub-
sequent hire by a company must be a friend of some previous hire—
unless no such person exists, in which case that company can again
hire anyone.
In the candidate pool are 10 geniuses; naturally each company
would like to get as many of those as possible. Can it be that the com-
pany that hires first cannot prevent its rival from hiring nine of the
geniuses?

Strength of Schedule
The 10 teams comprising the “Big 12” college football conference are
all scheduled to play one another in the upcoming season, after which
one will be declared conference champion. There are no ties, and each
team scores a point for every team it beats.
Suppose that, worried about breaking ties, a member of the gov-
erning board of the conference suggests that each team receive in addi-
tion “strength of schedule points,” calculated as the sum of the scores
achieved by the teams it beat.
A second member asks: “What if all the teams end up with the
same number of strength-of-schedule points?”
What, indeed? Can that even happen?

li
 Mathematical Puzzles

Locker Doors
Lockers numbered 1 to 100 stand in a row in the main hallway of
Euclid Junior High School. The first student arrives and opens all the
lockers. The second student then goes through and re-closes lockers
numbered 2, 4, 6, etc.; the third student changes the state of every
locker whose number is a multiple of 3, then the next every multiple
of 4, etc., until the last student opens or closes only locker number
100.
After the 100 students have passed through, which lockers are
open?

Gasoline Crisis
You need to make a long circular automobile trip during a gasoline
crisis. Inquiries have ascertained that the gas stations along the route
contain just enough fuel to make it all the way around. If you have an
empty tank but can start at a station of your choice, can you complete
a clockwise round trip?

Soldiers in a Field
An odd number of soldiers are stationed in a large field. No two sol-
diers are exactly the same distance apart as any other two soldiers.
Their commanding officer radios instructions to each soldier to keep
an eye on his nearest neighbor.
Is it possible that every soldier is being watched?

Home-field Advantage
Every year, the Elkton Earlies and the Linthicum Lates face off in a
series of baseball games, the winner being the first to win four games.
The teams are evenly matched but each has a small edge (say, a 51%
chance of winning) when playing at home.
Every year, the first three games are played in Elkton, the rest in
Linthicum.
Which team has the advantage?

lii
 The Puzzles

Zeroes and Ones

Show that every natural number has a non-zero multiple which, when
written out (base 10), contains only zeroes and ones. Even better,
prove that your telephone number, if it ends with a 1, 3, 7 or 9, has a
multiple whose decimal representation is all ones!

Random Intersection
Two unit-radius balls are randomly positioned subject to intersecting.
What is the expected volume of their intersection? For that matter,
what is the expected surface area of their intersection?

Profit and Loss

At a recent stockholders’ meeting of Widget Industrials Inc., the Chief
Financial Officer presented a chart of the month-by-month profits (or
losses) since the last meeting. “Note,” said she, “that we made a profit
over every consecutive eight-month period.”
“Maybe so,” complained a shareholder, “but I see we lost money
over every consecutive five-month period!”
What’s the maximum number of months that could have passed
since the last meeting?

Boarding the Manhole

An open manhole 4 meters in diameter has to be covered by boards
of total width w. The boards are each more than 4 meters in length,
so if w ≥ 4 it’s obvious that you can cover the manhole by laying the
boards side by side (see figure below). If w is only 3.9, say, you still
have plenty of wood and you are allowed to overlap the boards if you
want. Can you still cover the manhole?

liii
 Mathematical Puzzles

Pegs on the Half-Plane

Each grid point on the XY plane on or below the X-axis is occupied
by a peg. At any time, a peg can be made to jump over a neighbor peg
(horizontally, vertically, or diagonally adjacent) and onto the next grid
point in line, provided that point was unoccupied. The jumped peg is
then removed.
Can you get a peg arbitrarily far above the X-axis?

Unbreakable Domino Cover

A 6 × 5 rectangle can be covered with 2 × 1 dominoes, as in the figure
below, in such a way that no line between dominoes cuts all the way
across the board. Can you cover a 6 × 6 square that way?

Turning the Die

In the game of Turn-Die, a die is rolled and the number that appears
is recorded. The first player then turns the die 90 degrees (giving her
four options) and the new value is added to the old one. The second
player does the same and the two players alternate until a sum of 21
or higher is reached. If 21 was hit exactly, the player who reached it
wins; otherwise, the player who exceeded it loses.
Do you want to be first or second?

Protecting the Statue

Michelangelo’s David, in Florence, is protected (on the plane) by laser
beams in such a way that no-one can approach the statue or any of
the laser-beam sources without crossing a beam.

liv
 The Puzzles

What’s the minimum number of lasers needed to accomplish this?

(Beams reach 100 meters, say.)

Wild Guess
David and Carolyn are mathematicians who are unafraid of the infi-
nite and cheerfully invoke the Axiom of Choice when needed. They
elect to play the following two-move game. For her move, Carolyn
chooses an infinite sequence of real numbers, and puts each number
in an opaque box. David gets to open as many boxes as he wants—
even infinitely many—but must leave one box unopened. To win, he
must guess exactly the real number in that box.
On whom will you bet in this game, Carolyn or David?

Laser Gun
You find yourself standing in a large rectangular room with mirrored
walls. At another point in the room is your enemy, brandishing a laser
gun. You and she are fixed points in the room; your only defense is
that you may summon bodyguards (also points) and position them in
the room to absorb the laser rays for you. How many bodyguards do
you need to block all possible shots by the enemy?

Splitting the Stacks

Your job is to separate a stack of n objects into individual items. At
any time you may split a stack into two stacks, and get paid the prod-
uct of the sizes of the two stacks.
For example, for n=8, you could split into a stack of size 5 and
a stack of size 3, getting paid $15; then you can split the 5-stack into
2 and 3 to get $6 more. Then the two threes can each be split ($2
each) and the remaining three twos ($1 each) for a total income of
15 + 6 + (2 × 2) + (3 × 1) = 28 dollars.
What’s the most money you can get starting with n objects, and
how can you get it?

Chameleons
A colony of chameleons currently contains 20 red, 18 blue, and 16
green individuals. When two chameleons of different colors meet,

lv
 Mathematical Puzzles

each of them changes his or her color to the third color. Is it possible
that, after a while, all the chameleons have the same color?

Wires under the Hudson

Fifty identical wires run through a tunnel under the Hudson River,
but they all look the same, and you need to determine which pairs
of wire-ends belong to the same wire. To do this you can tie pairs of
wires together at one end of the tunnel and test pairs of wire-ends at
the other end to see if they close a circuit; in other words, you can
determine whether two wires are tied together at the other end.
How many trips across the Hudson do you need to accomplish
your task?

Two Round-robins
The Games Club’s 20 members played a round-robin checkers toura-
ment on Monday and a round-robin chess tournament on Tuesday.
In each tournament, a player scored one point for each other player
he or she beat, and half a point for each tie.
Suppose every player’s scores in the two tournaments differed by
at least 10. Show that in fact, the differences were all exactly 10.

Factorial Coincidence
Suppose that a, b, c, and d are positive integers, all different, all greater
than one. Can it be that a!b = c!d ?

Conversation on a Bus
Ephraim and Fatima, two colleagues in the Mathematics Department
at Zorn University, wind up seated together on the bus to campus.
Ephraim begins a conversation with “So, Fatima, how are your
kids doing? How old are they now, anyway?”
“It turns out,” says Fatima as she is putting the $1 change from the
bus driver into her wallet, “that the sum of their ages is the number of
this bus, and the product is the number of dollars that happen to be
in my purse at the moment.”

lvi
 The Puzzles

“Aha!” replies Ephraim. “So, if I remembered how many kids you

have, and you told me how much money you are carrying, I could
deduce their ages?”
“Actually, no,” says Fatima.
“In that case,” says Ephraim, “I know how much money is in your
purse.”
What is the bus number?

Coins on the Table

One hundred quarters lie on a rectangular table, in such a way that
no more can be added without overlapping. (We allow a quarter to
extend over the edge, as long as its center is on the table.)
Prove that you can start all over again and cover the whole table
with 400 quarters! (This time we allow overlap and overhang).

Roulette for the Unwary

Elwyn is in Las Vegas celebrating his 21st birthday, and his girlfriend
has gifted him with twenty-one $5 bills to gamble with. He saunters
over to the roulette table, noting that there are 38 numbers (0, 00, and
1 through 36) on the wheel. If he bets $1 on a single number, he will
win with probability 1/38 and collect $36 (in return for his $1 stake,
which still goes to the bank). Otherwise, of course, he just loses the
dollar.
Elwyn decides to use his $105 to make 105 one-dollar bets on the
number 21. What, approximately, is the probability that Elwyn will
come out ahead? Is it better than, say, 10%?

Uniformity at the Bakery

A baker’s dozen (thirteen) bagels have the property that any twelve of
them can be split into two piles of six each, which balance perfectly
on the scale. Suppose each bagel weighs an integer number of grams.
Must all the bagels have the same weight?

lvii
 Mathematical Puzzles

Slabs in 3-Space
A “slab” is the region between two parallel planes in three-
dimensional space. Prove that you cannot cover all of 3-space with
a set of slabs the sum of whose thicknesses is finite.

Pegs in a Square
Suppose we begin with n2 pegs on a plane grid, one peg occupying
each vertex of an n-vertex by n-vertex square. Pegs jump only hori-
zontally or vertically, by passing over a neighboring peg and into an
unoccupied vertex; the jumped peg is then removed. The goal is to
reduce the n2 pegs to only 1.
Prove that if n is a multiple of 3, it can’t be done!

Filling a Bucket
Before you are 12 two-gallon buckets and a 1-gallon scoop. At each
turn, you may fill the scoop with water and distribute the water any
way you like among the buckets.
However, each time you do this your opponent will empty two
buckets of her choice.
You win if you can get one of the big buckets to overflow. Can you
force a win? If so, how long will it take you?

Polygon on the Grid

A convex polygon is drawn on the coordinate plane with all its ver-
tices on integer points, but no side parallel to the x- or y-axis. Let h
be the sum of the lengths of the horizontal line segments at integer
height that intersect the (filled-in) polygon, and v the equivalent for
the vertical line segments. Prove that h = v.

Game of Desperation
On a piece of paper is a row of n empty boxes. Tristan and Isolde take
turns, each writing an “S” or an “O” into a previously blank box. The

lviii
 The Puzzles

winner is the one who completes an “SOS” in consecutive boxes. For

which n does the second player (Isolde) have a winning strategy?

Gluing Pyramids
A solid square-base pyramid, with all edges of unit length, and a solid
triangle-base pyramid (tetrahedron), also with all edges of unit length,
are glued together by matching two triangular faces.
How many faces does the resulting solid have?

Exponent upon Exponent

··
x·
Part I: If xx = 2, what is x?
·
··
xx
Part II: If x = 4, what is x?

Part III: How do you explain getting the same answer to Parts I and
II?

Random Intervals
The points 1, 2, . . . , 1000 on the number line are paired up at random,
to form the endpoints of 500 intervals. What is the probability that
among these intervals is one which intersects all the others?

North by Northwest
If you’ve never seen the famous Alfred Hitchcock movie North by
Northwest (1959), you should. But what direction is that, exactly? As-
sume North is 0◦ , East 90◦ , etc.

Missing Digit
The number 229 has 9 digits, all different; which digit is missing?

Coins in a Row
On a table is a row of 50 coins, of various denominations. Alix picks a
coin from one of the ends and puts it in her pocket; then Bert chooses

lix
 Mathematical Puzzles

a coin from one of the (remaining) ends, and the alternation continues
until Bert pockets the last coin.
Prove that Alix can play so as to guarantee at least as much money
as Bert.

Bugs on Four Lines

You are given four lines in a plane in general position (no two par-
allel, no three intersecting in a common point). On each line a ghost
bug crawls at some constant velocity (possibly different for each bug).
Being ghosts, if two bugs happen to cross paths they just continue
crawling through each other uninterrupted.
Suppose that five of the possible six meetings actually happen.
Prove that the sixth does as well.

First-grade Division
On the first day of class Miss Feldman divides her first-grade class into
k working groups. On the second day, she picks the working groups
a different way, this time ending up with k+1 of them.
Show that there are at least two kids who are in smaller groups on
the second day than they were on the first day.

Deterministic Poker
Unhappy with the vagaries of chance, Alice and Bob elect to play
a completely deterministic version of draw poker. A deck of cards
is spread out face-up on the table. Alice draws five cards, then Bob
draws five cards. Alice discards any number of her cards (the dis-
carded cards will remain out of play) and replaces them with a like
number of others; then Bob does the same. All actions are taken with
the cards face-up in view of the opponent. The player with the better
hand wins; since Alice goes first, Bob is declared to be the winner if
the final hands are equally strong. Who wins with best play?

lx
 The Puzzles

Precarious Picture
Suppose that you wish to hang a picture with a string attached at two
points on the frame. If you hang it by looping the string over two nails
in the ordinary way, as shown below, and one of the nails comes out,
the picture will still hang (albeit lopsidedly) on the other nail.

Can you hang it so that the picture falls if either nail comes out?

Find the Robot

At time t = 0 a robot is placed at some unknown grid point in 3-
dimensional space. Every minute, the robot moves a fixed, unknown
distance in a fixed, unknown direction, to a new grid point. Each
minute, you are allowed to probe any single point in space. Devise an
algorithm that is guaranteed to find the robot in finite time.

Early Commuter
A commuter arrives at her home station an hour early and walks to-
ward home until she meets her husband driving to pick her up at the
normal time. She ends up home 20 minutes earlier than usual. How
long did she walk?

Subtracting Around the Corner

A sequence of n positive numbers is written on a piece of paper. In one
“operation,” a new sequence is written below the old one according
to the following rules: The (absolute) difference between each num-
ber and its successor is written below that number; and the absolute
difference between the first and last numbers is written below the last
number. For example, the sequence 4 13 9 6 is followed by 9 4 3 2.

lxi
 Mathematical Puzzles

Try this for n = 4 using random numbers between, say, 1 and

100. You’ll find that after remarkably few operations, the sequence
degenerates to 0 0 0 0 where it of course remains. Why? Is the same
true for n = 5?

Powers of Roots
√ √
What is the first digit after the decimal point in the number ( 2+ 3)
to the billionth power?

An Attractive Game
You have an opportunity to bet $1 on a number between 1 and 6.
Three dice are then rolled. If your number fails to appear, you lose
your $1. If it appears once, you win $1; if twice, $2; if three times, $3.
Is this bet in your favor, fair, or against the odds? Is there a way to
determine this without pencil and paper (or a computer)?

Circular Shadows II
Show that if the projections of a solid body onto two planes are perfect
disks, then the two projections have the same radius.

Finding the Rectangles

Prove that any tiling of a regular 400-gon by parallelograms must con-
tain at least 100 rectangles.

Figure Eights in the Plane

How many disjoint topological “figure 8s” can be drawn on the plane?

lxii
 The Puzzles

Alternate Connection
100 points lie on a circle. Alice and Bob take turns connecting pairs
of points by a line, until every point has at least one connection. The
last one to play wins; which player has a winning strategy?

Tiling a Polygon
A “rhombus” is a quadrilateral with four equal sides; we consider two
rhombi to be different if you can’t translate (move without rotation)
one to coincide with the other. Given a regular polygon with 100 sides,
you can take any two non-parallel sides, make( )two copies of each and
translate them to form a rhombus. You get 50 2 different rhombi that
way. You can use translated copies of these to tile your 100-gon; show
that if you do, you will use each different rhombus exactly once!

Bindweed and Honeysuckle

Two vines, a bindweed and a honeysuckle, climb a tree trunk, start-
ing and ending at the same place. The bindweed circles the trunk
three times counter-clockwise while the honeysuckle circles five times
clockwise. Not counting the top and bottom, how many times do they
cross?

Uniting the Loops

Can you re-arrange the 16 square tiles below into a 4 × 4 square—no
rotation allowed!—in such a way that the curves form a single loop?
Lines that exit from the bottom are assumed to continue at the same
point on the top, and similarly right to left; in other words, imagine
that the square is rolled up into a doughnut.

lxiii
 Mathematical Puzzles

Splitting a Polygon
A chord of a polygon is a straight line segment that touches the poly-
gon’s perimeter at the segment’s endpoints and nowhere else.
Show that every polygon, convex or not, has a chord such that
each of the two regions into which it divides the polygon has area at
least 1/3 the area of the polygon.

Two Monks on a Mountain

Remember the monk who climbed Mt. Fuji on Monday and de-
scended on Tuesday? This time, he and a fellow monk climb the
mountain on the same day, starting at the same time and altitude,
but on different paths. The paths go up and down on the way to the
summit (but never dip below the starting altitude); you are asked to
prove that they can vary their speeds (sometimes going backwards) so
that at every moment of the day they are at the same altitude!

Alternative Dice
Can you design two different dice so that their sums behave just like
a pair of ordinary dice? That is, there must be two ways to roll a 3,
6 ways to roll a 7, one way to roll a 12, and so forth. Each die must
have 6 sides, and each side must be labeled with a positive integer.

Even Split
Prove that from every set of 2n integers, you can choose a subset of
size n whose sum is divisible by n.

Coconut Classic
Five men and a monkey, marooned on an island, collect a pile of
coconuts to be divided equally the next morning. During the night,
however, one of the men decides he’d rather take his share now. He
tosses one coconut to the monkey and removes exactly 1/5 of the

lxiv
 The Puzzles

remaining coconuts for himself. A second man does the same thing,
then a third, fourth, and fifth.
The following morning the men wake up together, toss one more
coconut to the monkey, and divide the rest equally. What’s the least
original number of coconuts needed to make this whole scenario pos-
sible?

Infected Checkerboard
An infection spreads among the squares of an n × n checkerboard
in the following manner: If a square has two or more infected neigh-
bors, then it becomes infected itself. (Neighbors are orthogonal only,
so each square has at most four neighbors.)
For example, suppose that we begin with all n squares on the main
diagonal infected. Then the infection will spread to neighboring diag-
onals and eventually to the whole board.
Prove that you cannot infect the whole board if you begin with
fewer than n infected squares.

Alternating Powers
Since the series 1 − 1 + 1 − 1 + 1 − · · · does not converge, the function
f (x) = x−x2 +x4 −x8 +x16 −x32 +· · · makes no sense when x = 1.
However, f (x) does converge for all positive real numbers x < 1. If
we wanted to give f (1) a value, it might make sense to let it be the
limit of f (x) as x approaches 1 from below. Does that limit exist? If
so, what is it?

Service Options
You are challenged to a short tennis match, with the winner to be
the first player to win four games. You get to serve first. But there are
options for determining the sequence in which the two of you serve:

1. Standard: Serve alternates (you, her, you, her, you, her, you).

2. Volleyball Style: The winner of the previous game serves the

next one.

3. Reverse Volleyball Style: The winner of the previous game re-

ceives in the next one.

lxv
 Mathematical Puzzles

Which option should you choose? You may assume it is to your ad-
vantage to serve. You may also assume that the outcome of any game
is independent of when the game is played and of the outcome of any
previous game.

Conway's Immobilizer
Three cards, an ace, a deuce, and a trey, lie face-up on a desk in some
or all of three marked positions (“left,” “middle,” and “right”). If they
are all in the same position, you see only the top card of the stack; if
they are in two positions, you see only two cards and do not know
which of them is concealing the third card.
Your objective is to get the cards stacked on the left with ace on
top, then deuce, then trey on bottom. You do this by moving one card
at a time, always from the top of one stack to the top of another (pos-
sibly empty) stack.
The problem is, you have no short-term memory and must, there-
fore, devise an algorithm in which each move is based entirely on
what you see, and not on what you last saw or did, or on how many
moves have transpired. An observer will tell you when you’ve won.
Can you devise an algorithm that will succeed in a bounded number
of steps, regardless of the initial configuration?

Same Sum Dice

You roll a set of n red n-sided dice, and a set of n black n-sided dice.
Each die is labeled with the numbers from 1 to n. Show that there
must always be a nonempty subset of the red dice, and a nonempty
subset of the black dice, showing the same sum.

Matching Coins
Sonny and Cher play the following game. In each round, a fair coin
is tossed. Before the coin is tossed, Sonny and Cher simultaneously
declare their guess for the result of the coin toss. They win the round if
both guessed correctly. The goal is to maximize the fraction of rounds
won, when the game is played for many rounds.
So far, the answer is obviously 50%: Sonny and Cher agree on
a sequence of guesses (for example, they decide to always declare
“heads”), and they can’t do any better than that. However, before

lxvi
 The Puzzles

the game begins, the players are informed that just prior to the first
toss, Cher will be given the results of all the coin tosses in advance!
She has a chance now to discuss strategy with Sonny, but once she
gets the coinflip information, there will be no further opportunity to
conspire. Show how Sonny and Cher can guarantee to get at least six
wins in 10 flips.
(If that’s too easy for you, show how they can guarantee at least
six wins in only nine flips!)

Next Card Bet

Cards are turned face up one at a time from the top of a well-shuffled
deck. You begin with a bankroll of $1, and can bet any fraction of your
current worth, prior to each revelation, on the color of the next card.
You get even odds regardless of the current composition of the deck.
Thus, for example, you can decline to bet until the last card, whose
color you will of course know, then bet everything and be assured of
going home with $2.
Is there any way you can guarantee to finish with more than $2? If
so, what’s the maximum amount you can assure yourself of winning?

Summing Fractions
Gail asks Henry to think of a number n between 10 and 100, but not
to tell her what it is. She now tells Henry to find all (unordered) pairs
of numbers j, k that are relatively prime and less than n, but add up
to more than n. He now adds all the fractions 1/jk.
Whew! Finally, Gail tells Henry what his sum is. How does she
do it?

Box in a Box
Suppose the cost of shipping a rectangular box is given by the sum of
its length, width, and height. Might it be possible to save money by
fitting your box into a cheaper box?

Option Hats
One hundred prisoners are told that at midnight, in the dark, each
will be fitted with a red or black hat according to a fair coinflip. The

lxvii
 Mathematical Puzzles

prisoners will be arranged in a circle and the lights turned on, enabling
each prisoner to see every other prisoner’s hat color. Once the lights
are on, the prisoners will have no opportunity to signal to one another
or to communicate in any way.
Each prisoner will then be taken aside and given the option of
trying to guess whether his own hat is red or black, but he may choose to
pass. The prisoners will all be freed if (1) at least one prisoner chooses
to guess his hat color, and (2) every prisoner who chooses to guess
guesses correctly.
As usual, the prisoners have a chance to devise a strategy before
the game begins. Can they achieve a winning probability greater than
50%?

Impressionable Thinkers
The citizens of Floptown meet each week to talk about town poli-
tics, and in particular whether or not to support the building of a
new shopping mall downtown. During the meetings each citizen talks
to his friends—of whom there are always an odd number, for some
reason—and the next day, changes (if necessary) his opinion regard-
ing the mall so as to conform to the opinion of the majority of his
friends.
Prove that eventually, the opinions held every other week will be
the same.

One-bulb Room
Each of n prisoners will be sent alone into a certain room, infinitely
often, but in some arbitrary order determined by their jailer. The pris-
oners have a chance to confer in advance, but once the visits begin,
their only means of communication will be via a light in the room
which they can turn on or off. Help them design a protocol which
will ensure that some prisoner will eventually be able to deduce that
everyone has visited the room.

Sphere and Quadrilateral

A quadrilateral in space has all of its edges tangent to a sphere. Prove
that the four points of tangency lie on a plane.

lxviii
 The Puzzles

Bluffing with Reals

Consider the following simple bluffing game. Louise and Jeremy ante
$1 each and each is given a secret random real number between 0 and
1. Louise may decide to pass in which case the $2 pot goes to the
player with the higher number. However, if she chooses, Louise may
add another dollar to the pot. Jeremy may then “call” by adding an-
other dollar himself, in which case the pot, now with $4 in it, goes
again to the player with the higher number. Or Jeremy may fold, ced-
ing the pot, with his $1 in it, to Louise.
Surely Louise has the advantage in this game, or at least equality,
since she can break even by always passing. How much is the game
worth to her? What are the players’ equilibrium strategies?

Tiling with Crosses

Can you tile the plane with 5-square crosses? Can you tile 3-space
with 7-cube crosses?

Y's in the Plane

Prove that only countably many disjoint Y’s can be drawn in the
plane.

Infected Cubes
An infection spreads among the n3 unit cubes of an n × n × n
cube, in the following manner: If a unit cube has 3 or more infected
neighbors, then it becomes infected itself. (Neighbors are orthogonal
only, so each little cube has at most 6 neighbors.)
Prove that you can infect the whole big cube starting with just n2
sick unit cubes.

Worst Route
A postman has deliveries to make on a long street, to addresses 2,
3, 5, 7, 11, 13, 17, and 19. The distance between any two houses is
proportional to the difference of their addresses.
To minimize the distance traveled, the postman would of course
make his deliveries in increasing (or decreasing) order of address. But
our postman is overweight and would like to maximize the distance

lxix
 Mathematical Puzzles

traveled making these deliveries, so as to get the most exercise he can.

But he can’t just wander around town; to do his job properly he is
obligated to walk directly from each delivery to the next one.
In what order should he make his deliveries?

Factorials and Squares

Consider the product 100! · 99! · 98! · · · · · 2! · 1!. Call each of the
100 factors k! a “term.” Can you remove one term and leave a perfect
square?

Halfway Points
Let S be a finite set of points in the unit interval [0,1], and suppose
that every point x ∈ S lies either halfway between two other points in
S (not necessarily the nearest ones) or halfway between another point
in S and an endpoint.
Show that all the points in S are rational.

Who Won the Series?

Two evenly-matched teams meet to play a best-of-seven World Series
of baseball games. Each team has the same small advantage when
playing at home. As is customary for this event, one team (say, Team
A) plays games 1 and 2 at home, and, if necessary, plays games 6 and
7 at home. Team B plays games 3, 4 and, if needed, 5 at home.
You go to a conference in Europe and return to find that the series
is over, and six games were played. Which team is more likely to have
won the series?

Tiling with L's

Can you tile the positive quadrant of the plane grid with unrotated tri-
ominoes each of which is either shaped like a letter L, or a backwards
L?

Dishwashing Game
You and your spouse flip a coin to see who washes the dishes each
evening. “Heads” she washes, “tails” you wash.

lxx
 The Puzzles

Tonight she tells you she is imposing a different scheme. You flip
the coin 13 times, then she flips it 12 times. If you get more heads than
she does, she washes; if you get the same number of heads or fewer,
you wash.
Should you be happy?

Random Judge
After a wild night of shore leave, you are about to be tried by your US
Navy superiors for unseemly behavior. You have a choice between ac-
cepting a “summary” court-martial with just one judge, or a “special”
court-martial with three judges who decide by majority vote.
Each possible judge has the same (independent) probability—
65.43%—of deciding in your favor, except that one officer who would
be judging your special court-martial (but not the summary) is noto-
rious for flipping a coin to make his decisions.
Which type of court-martial is more likely to keep you out of the
brig?

Angles in Space
Prove that among any set of more than 2n points in Rn , there are three
that determine an obtuse angle.

Wins in a Row
You want to join a certain chess club, but admission requires that you
play three games against Ioana, the current club champion, and win
two in a row.
Since it is an advantage to play the white pieces (which move first),
you alternate playing white and black.
A coin is flipped, and the result is that you will be white in the first
and third games, black in the second.
Should you be happy?

Chessboard Guess
Troilus is engaged to marry Cressida but threatened with deportation,
and Immigration is questioning the legitimacy of the proposed mar-
riage. To test their connection, Troilus will be brought into a room

lxxi
 Mathematical Puzzles

containing a chessboard, one of whose squares is designated as spe-

cial. On every square will be a coin, either heads-up or tails-up. Troilus
gets to turn over one coin, after which he will be ejected from the room
and Cressida brought in.
Cressida, after examining the chessboard, must guess the desig-
nated square. If she gets it wrong, Troilus is deported.
Can Troilus and Cressida save their marriage?

Split Games
You are a rabid baseball fan and, miraculously, your team has won
the pennant—thus, it gets to play in the World Series. Unfortunately,
the opposition is a superior team whose probability of winning any
given game against your team is 60%.
Sure enough, your team loses the first game in the best-of-seven
series and you are so unhappy that you drink yourself into a stupor.
When you regain consciousness, you discover that two more games
have been played.
You run out into the street and grab the first passer-by. “What
happened in games 2 and 3 of the World Series?”
“They were split,” he says. “One game each.”
Should you be happy?

Frames on a Chessboard
You have an ordinary 8 × 8 chessboard with red and black squares. A
genie gives you two “magic frames,” one 2 × 2 and one 3 × 3. When
you place one of these frames neatly on the chessboard, the 4 or 9
squares they enclose instantly flip their colors.
Can you reach all 264 possible color configurations?

Angry Baseball
As in Split Games, your team is the underdog and wins any given game
in the best-of-seven World Series with probability 40%. But, hold on:
This time, whenever your team is behind in the series, the players get
angry and play better, raising your team’s probability of winning that
game to 60%.
Before it all begins, what is your team’s probability of winning the
World Series?

lxxii
 The Puzzles

Bugs on a Polyhedron
Associated with each face of a solid convex polyhedron is a bug which
crawls along the perimeter of the face, at varying speed, but only in the
clockwise direction. Prove that no schedule will permit all the bugs to
circumnavigate their faces and return to their initial positions without
incurring a collision.

Serious Candidates
Assume that, as is often the case, no one has any idea who the next
nominee for President of the United States will be, of the party not
currently in power. In particular, at the moment no person has prob-
ability as high as 20% of being chosen.
As the politics and primaries proceed, probabilities change con-
tinuously and some candidates will exceed the 20% threshold while
others will never do so. Eventually one candidate’s probability will
rise to 100% while everyone else’s drops to 0. Let us say that after the

convention, a candidate is entitled to say he or she was a “serious”

candidate if at some point his or her probability of being nominated
exceeded 20%.
Can anything be said about the expected number of serious can-
didates?

Sums and Differences

Given 25 different positive numbers, can you always choose two of
them such that none of the other numbers equals either their sum or
their difference?

Two-Point Conversion
You, coach of the Hoboken Hominids football team, were 14 points
behind your rivals (the Gloucester Great Apes) until, with just a

lxxiii
 Mathematical Puzzles

minute to go in the game, you scored a touchdown. You have a choice

of kicking an extra point (95% success rate) or going for two (45% suc-
cess rate). Which should you do?

Swedish Lottery
In a proposed mechanism for the Swedish National Lottery, each par-
ticipant chooses a positive integer. The person who submits the low-
est number not chosen by anyone else is the winner. (If no number is
chosen by exactly one person, there is no winner.)
If just three people participate, but each employs an optimal, equi-
librium, randomized strategy, what is the largest number that has pos-
itive probability of being submitted?

Random Chord
What is the probability that a random chord of a circle is longer than
a side of an equilateral triangle inscribed in the circle?

Cube Magic
Can you pass a cube through a hole in a smaller cube?

Random Bias
Suppose you choose a real number p between 0 and 1 uniformly at
random, then bend a coin so that its probability of coming up heads
when you flip it is precisely p. Finally you flip your bent coin 100
times. What is the probability that after all this, you end up flipping
exactly 50 heads?

Gladiators, Version I
Paula and Victor each manage a team of gladiators. Paula’s gladiators
have strengths p1 , p2 , . . . , pm and Victor’s, v1 , v2 , . . . , vn . Gladi-
ators fight one-on-one to the death, and when a gladiator of strength
x meets a gladiator of strength y, the former wins with probability
x/(x+y), and the latter with probability y/(x+y). Moreover, if the
gladiator of strength x wins, he gains in confidence and inherits his

lxxiv
 The Puzzles

opponent’s strength, so that his own strength improves to x+y; simi-

larly, if the other gladiator wins, his strength improves from y to x+y.
After each match, Paula puts forward a gladiator (from those on
her team who are still alive), and Victor must choose one of his to face
Paula’s. The winning team is the one which remains with at least one
live player.
What’s Victor’s best strategy? For example, if Paula begins with
her best gladiator, should Victor respond from strength or weakness?

Gladiators, Version II
Again Paula and Victor must face off in the Colosseum, but this time,
confidence is not a factor, and when a gladiator wins, he keeps the
same strength he had before.
As before, prior to each match, Paula chooses her entry first. What
is Victor’s best strategy? Whom should he play if Paula opens with her
best man?

Rolling a Six
How may rolls of a die does it take, on average, to get a 6—given that
you didn’t roll any odd numbers en route?

Traveling Salesmen
Suppose that between every pair of major cities in Russia, there’s a
fixed one-way air fare for going from either city to the other. Trav-
eling salesman Alexei Frugal begins in St. Petersburg and tours the
cities, always choosing the cheapest flight to a city not yet visited (he
does not need to return to St. Petersburg). Salesman Boris Lavish also
needs to visit every city, but he starts in Kaliningrad, and his policy
is to choose the most expensive flight to an unvisited city at each step.
It looks obvious that Lavish’s tour costs at least as much as Fru-
gal’s, but can you prove it?

Napkins in a Random Setting

At a conference banquet for a meeting of the Association for Women
in Mathematics, the participants find themselves assigned to a big cir-
cular table. On the table, between each pair of adjacent settings, is

lxxv
 Mathematical Puzzles

a coffee cup containing a cloth napkin. As each mathematician sits

down, she takes a napkin from her left or right; if both napkins are
present, she chooses randomly. If the seats are occupied in random

order and the number of mathematicians is large, what fraction of

them (asymptotically) will end up without a napkin?

Lame Rook
A lame rook moves like an ordinary rook in chess—straight up, down,
left, or right—but only one square at a time. Suppose that the lame
rook begins at some square and tours the 8 × 8 chessboard, visiting
each square once and returning to the starting square on the 64th
move. Show that the number of horizontal moves of the tour, and
the number of vertical moves of the tour, are not equal!

Coin Testing
The Unfair Advantage Magic Company has supplied you, a ma-
gician, with a special penny and a special nickel. One of these is
supposed to flip “Heads” with probability 1/3, the other 1/4, but
UAMCO has not bothered to tell you which is which.
Having limited patience, you plan to try to identify the biases by
flipping the nickel and penny one by one until one of them comes up
Heads, at which point that one will be declared to be the 1/3-Heads
coin.
In what order should you flip the coins to maximize the probabil-
ity that you get the correct answer, and at the same time to be fair, that
is, to give the penny and the nickel the same chance to be designated
the 1/3-Heads coin?

Curve and Three Shadows

Is there a simple closed curve in 3-space, all three of whose projections
onto the coordinate planes are trees? This means that the shadows of

lxxvi
 The Puzzles

the curve, from the three coordinate directions, may not contain any
loops.

Majority Hats
One hundred prisoners are told that at midnight, in the dark, each
will be fitted with a red or black hat according to a fair coin-flip. The
prisoners will be arranged in a circle and the lights turned on, enabling
each prisoner to see every other prisoner’s hat color. Once the lights
are on, the prisoners will have no opportunity to signal to one another
or to communicate in any way.
Each prisoner is then taken aside and must try to guess his own
hat color. The prisoners will all be freed if a majority (here, at least
51) get it right.
As before, the prisoners have a chance to devise a strategy before
the game begins. Can they achieve a winning probability greater than
50%? Would you believe 90%? How about 95%?

Bugs on a Line
Each positive integer on the number line is equipped with a green,
yellow, or red light. A bug is dropped on “1” and obeys the following
rules at all times: If it sees a green light, it turns the light yellow and
moves one step to the right; if it sees a yellow light, it turns the light
red and moves one step to the right; if it sees a red light, it turns the
light green and moves one step to the left.
Eventually, the bug will fall off the line to the left, or run out to
infinity on the right. A second bug is then dropped on “1,” again fol-
lowing the traffic lights starting from the state the last bug left them
in; then, a third bug makes the trip.
Prove that if the second bug falls off to the left, the third will march
off to infinity on the right.

Prisoner and Dog

A woman is imprisoned in a large field surrounded by a circular fence.
Outside the fence is a vicious guard dog that can run four times as fast
as the woman, but is trained to stay near the fence. If the woman can
contrive to get to an unguarded point on the fence, she can quickly
scale the fence and escape. But can she get to a point on the fence
ahead of the dog?

lxxvii
 Mathematical Puzzles

Pegs on the Corners

Four pegs begin on the plane at the corners of a square. At any time,
you may cause one peg to jump over a second, placing the first on
the opposite side of the second, but at the same distance as before.
The jumped peg remains in place. Can you maneuver the pegs to the
corners of a larger square?

Circles in Space
Can you partition all of 3-dimensional space into circles?

Fifteen Bits and a Spy

A spy’s only chance to communicate with her control lies in the daily
broadcast of 15 zeros and ones by a local radio station. She does not
know how the bits are chosen, but each day she has the opportunity
to alter any one bit, changing it from a 0 to a 1 or vice-versa.
How much daily information can she communicate?

Flipping the Pentagon

The vertices of a pentagon are labeled with integers, the sum of which
is positive. At any time, you may change the sign of a negative label,
but then the new value is subtracted from both neighbors’ values so
as to maintain the same sum.
Prove that, inevitably, no matter which negative labels are flipped,
the process will terminate after finitely many flips, with all values non-
negative.

Love in Kleptopia
Jan and Maria have fallen in love (via the internet) and Jan wishes to
mail her a ring. Unfortunately, they live in the country of Kleptopia

lxxviii
 The Puzzles

where anything sent through the mail will be stolen unless it is sent
in a padlocked box. Jan and Maria each have plenty of padlocks, but
none to which the other has a key. How can Jan get the ring safely
into Maria’s hands?

Touring an Island
Aloysius is lost while driving his Porsche on an island in which ev-
ery intersection is a meeting of three (two-way) streets. He decides to
adopt the following algorithm: Starting in an arbitrary direction from
his current intersection, he turns right at the next intersection, then
left at the next, then right, then left, and so forth.
Prove that Aloysius must return eventually to the intersection at
which he began this procedure.

Badly Designed Clock

The hour and minute hands of a certain clock are indistinguishable.
How many moments are there in a day when it is not possible to tell
from this clock what time it is?

Fibonacci Multiples
Show that every positive integer has a multiple that’s a Fibonacci
number.

lxxix
 Mathematical Puzzles

Worms and Water

Lori is having trouble with worms crawling into her bed. To stop
them, she places the legs of the bed into pails of water; since the
worms can’t swim, they can’t reach the bed via the floor. But they
instead crawl up the walls and across the ceiling, dropping onto her
bed from above. Yuck!
How can Lori stop the worms from getting to her bed?

Light Bulbs in a Circle

In a circle are light bulbs numbered clockwise 1 through n, n > 1, all
initially on. At time t, you examine bulb number t (modulo n), and
if it’s on, you change the state of bulb t+1 (modulo n); i.e., you turn
off the clockwise-next bulb if it’s on, and on if it’s off. If bulb t is off,
you do nothing.
Prove that if you continue around and around the ring in this man-
ner, eventually all the bulbs will again be on.

Generating the Rationals

You are given a set S of numbers that contains 0 and 1, and contains
the mean of every finite nonempty subset of S. Prove that S contains
all the rational numbers between 0 and 1.

Emptying a Bucket
You are presented with three large buckets, each containing an inte-
gral number of ounces of some non-evaporating fluid. At any time,
you may double the contents of one bucket by pouring into it from a
fuller one; in other words, you may pour from a bucket containing x
ounces into one containing y ≤ x ounces until the latter contains 2y
ounces (and the former, x−y).

lxxx
 The Puzzles

Prove that no matter what the initial contents, you can, eventually,
empty one of the buckets.

Funny Dice
You have a date with your friend Katrina to play a game with three
dice, as follows. She chooses a die, then you choose one of the other
two dice. She rolls her die while you roll yours, and whoever rolls the
higher number wins. If you roll the same number, Katrina wins.
Wait, it’s not as bad as you think; you get to design the dice! Each
will be a regular cube, but you can put any number of pips from 1 to
6 on any face, and the three dice don’t have to be the same.
Can you make these dice in such a way that you have the advan-
tage in your game?

Picking the Athletic Committee

The Athletic Committee is a popular service option among the faculty
of Quincunx University, because while you are on it, you get free tick-
ets to the university’s sports events. In an effort to keep the committee
from becoming cliquish, the university specifies that no one with three
or more friends on the committee may serve on the committee—but,
in compensation, if you’re not on the committee but have three or
more friends on it, you can get free tickets to any athletic event of
your choice.
To keep everyone happy, it is therefore desirable to construct the
committee in such a way that even though no one on it has three or
more friends on it, everyone not on the committee does have three or
more friends on it.
Can this always be arranged?

Ice Cream Cake

On the table before you is a cylindrical ice-cream cake with chocolate
icing on top. From it you cut successive wedges of the same angle θ.
Each time a wedge is cut, it is turned upside-down and reinserted into
the cake. Prove that, regardless of the value of θ, after a finite number
of such operations all the icing is back on top of the cake!

lxxxi
 Mathematical Puzzles

Sharing a Pizza
Alice and Bob are preparing to share a circular pizza, divided by radial
cuts into some arbitrary number of slices of various sizes. They will
be using the “polite pizza protocol”: Alice picks any slice to start;
thereafter, starting with Bob, they alternate taking slices but always
from one side or the other of the gap. Thus after the first slice, there
are just two choices at each turn until the last slice is taken (by Bob if
the number of slices is even, otherwise by Alice).
Is it possible for the pizza to have been cut in such a way that Bob
has the advantage—in other words, so that with best play, Bob gets
more than half the pizza?

Moth's Tour
A moth alights on the “12” of a clock face, and begins randomly walk-
ing around the dial. Each time it hits a number, it proceeds to the next
clockwise number, or the next counterclockwise number, with equal
probability. It continues until it has been at every number.
What is the probability that the moth finishes at the number “6”?

Names in Boxes
The names of 100 prisoners are placed in 100 wooden boxes, one
name to a box, and the boxes are lined up on a table in a room. One
by one, the prisoners are led into the room; each may look in at most
50 boxes, but must then leave the room exactly as he found it and is
permitted no further communication with the others.
The prisoners have a chance to plot their strategy in advance, and
they are going to need it, because unless every single prisoner finds his
own name all will subsequently be executed.
Find a strategy that gives the prisoners a decent chance of survival.

Garnering Fruit
Each of 100 baskets contains some number (could be zero) of apples,
some number of bananas, and some number of cherries. Show that
you can collect 51 of those baskets that together contain at least half
the apples, at least half the bananas, and at least half the cherries!

lxxxii
 The Puzzles

Coin Game
You and a friend each pick a different heads-tails sequence of length
4 and a fair coin is flipped until one sequence or the other appears;
the owner of that sequence wins the game.
For example, if you pick HHHH and she picks TTTT, you win if
a run of four heads occurs before a run of four tails.
Do you want to pick first or second? If you pick first, what should
you pick? If your friend picks first, how should you respond?

Roulette for Parking Money

You’re in Las Vegas with only $2 and in desperate need of $5 to feed a
parking meter. You run in through the nearest door and find yourself
at a roulette table. You can bet any whole dollar amount on any al-
lowable set of numbers. What roulette-betting strategy will maximize
your probability of walking out with $5?

Invisible Corners
Can it be that you are standing outside a polyhedron and can’t see
any of its vertices?

Sleeping Beauty
Sleeping Beauty agrees to the following experiment. On Sunday she
is put to sleep and a fair coin is flipped. If it comes up Heads, she is
awakened on Monday morning; if Tails, she is awakened on Monday
morning and again on Tuesday morning. In all cases, she is not told
the day of the week, is put back to sleep shortly after, and will have
no memory of any Monday or Tuesday awakenings.

lxxxiii
 Mathematical Puzzles

When Sleeping Beauty is awakened on Monday or Tuesday,

what—to her—is the probability that the coin came up Heads?

Boardroom Reduction
The Board of Trustees of the National Museum of Mathematics has
grown too large—50 members, now—and its members have agreed
to the following reduction protocol. The board will vote on whether
to (further) reduce its size. A majority of ayes results in the immediate
ejection of the newest board member; then another vote is taken, and
so on. If at any point half or more of the surviving members vote nay,
the session is terminated and the board remains as it currently is.
Suppose that each member’s highest priority is to remain on the
board, but given that, agrees that the smaller the board, the better.
To what size will this protocol reduce the board?

Buffon's Needle
A needle one inch in length is tossed onto a large mat marked with
parallel lines one inch apart. What is the probability that the needle
lands across a line?

Seven Cities of Gold

In 1539 Friar Marcos de Niza returned to Mexico from his travels
in what is now Arizona, famously reporting his finding of “Seven
Cities of Gold.” Coronado did not believe the “liar friar” and when
subsequent expeditions came up empty, Coronado gave up the quest.
My (unreliable) sources claim that Coronado’s reason for disbe-
lieving de Niza is the latter’s claim that the cities were laid out in the
desert in a manner such that of any three of the cities, at least one pair
were exactly 100 furlongs apart.
Coronado’s advisors told him no such pattern of points existed.
Were they right?

lxxxiv
 The Puzzles

Life-Saving Transposition
There are just two prisoners this time, Alice and Bob. Alice will be
shown a deck of 52 cards spread out in some order, face-up on a table.
She will be asked to transpose two cards of her choice. Alice is then
dismissed, with no further chance to communicate to Bob. Next, the
cards are turned down and Bob is brought into room. The warden
names a card and to stave off execution for both prisoners, Bob must
find the card after turning over, sequentially, at most 26 of the cards.
As usual the prisoners have an opportunity to conspire before-
hand. This time, they can guarantee success. How?

More Magnetic Dollars

We return to Magnetic Dollars, but we strengthen their attractive power
just a bit.
This time, an infinite sequence of coins will be tossed into the two
urns. When one urn contains x coins and the other y, the next coin
will fall into the first urn with probability x1.01 /(x1.01 + y 1.01 ), other-
wise into the second urn.
Prove that after some point, one of the urns will never get another
coin!

Covering the Stains

Just as a big event is about to begin, the queen’s caterer notices, to his
horror, that there are 10 tiny gravy stains on the tablecloth. All he has
the time to do is to cover the stains with non-overlapping plates. He
has plenty of plates, each a unit disk. Can he do it, no matter how the
stains are distributed?

Zero-Sum Vectors
On a piece of paper, you have (for some reason) made an array whose
rows consist of all 2n of the n-dimensional vectors with coordinates
in {+1, −1}n —that is, all possible strings of +1’s and −1’s of length
n.
Notice that there are lots of nonempty subsets of your rows which
sum to the zero vector, for example any vector and its complement;
or the whole array, for that matter.

lxxxv
 Mathematical Puzzles

However, your 2-year-old nephew has got hold of the paper and
has changed some of the entries in the array to zeroes.
Prove that no matter what your nephew did, you can find a
nonempty subset of the rows in the new array that sums to zero.

Colors and Distances

In the town of Hoegaarden, Belgium, exactly half the houses are oc-
cupied by Flemish families, the rest by French-speaking Walloons.
Can it be that the town is so well mixed that the sum of the dis-
tances between pairs of houses of like ethnicity exceeds the sum of
the distances between pairs of houses of different ethnicity?

Two Sheriffs
Two sheriffs in neighboring towns are on the track of a killer, in a case
involving eight suspects. By virtue of independent, reliable detective
work, each has narrowed his list to only two. Now they are engaged
in a telephone call; their object is to compare information, and if their
pairs overlap in just one suspect, to identify the killer.
The difficulty is that their telephone line has been tapped by the
local lynch mob, who know the original list of suspects but not which
pairs the sheriffs have arrived at. If they are able to identify the killer
with certainty as a result of the phone call, he will be lynched before
he can be arrested.
Can the sheriffs, who have never met, conduct their conversation
in such a way that they both end up knowing who the killer is (when
possible), yet the lynch mob is still left in the dark?

Painting the Fence

Each of n industrious people chooses a random point on a circular
fence, and begins painting toward her farthest neighbor until she en-
counters a painted section. On average, how much of the fence gets
painted? How about if instead, each person paints toward her nearest
neighbor?

lxxxvi
 The Puzzles

Leading All the Way

Running for local office against Bob, Alice wins with 105 votes to
Bob’s 95. What is the probability that as the votes were counted (in
random order), Alice led the whole way?

Self-Referential Number
The first digit of a certain 8-digit integer N is the number of zeroes in
the (ordinary, decimal) representation of N . The second digit is the
number of ones; the third, the number of twos; the fourth, the number
of threes; the fifth, the number of fours; the sixth, the number of fives;
the seventh, the number of sixes; and, finally, the eighth is the total
number of distinct digits that appear in N . What is N ?

Filling the Cup

You go to the grocery store needing one cup of rice. When you push
the button on the machine, it dispenses a uniformly random amount
of rice between nothing and one cup. On average, how many times
do you have to push the button to get your cup?

Two Balls and a Wall

On a line are two identical-looking balls and a vertical wall. The balls
are perfectly elastic and friction-free; the wall is perfectly rigid; the
ground is perfectly level. If the balls are the same mass and the one
farther from the wall is rolled toward the closer one, it will knock the
close ball toward the wall; that ball will bounce back and hit the first
ball, which will then roll (forever) away from the wall. Three bounces
altogether.
Now assume the farther ball has mass a million times greater than
the closer one. How many bounces now? (You may ignore the effects
of angular momentum, relativity, and gravitational attraction.)

Bulgarian Solitaire
Fifty-five chips are organized into some number of stacks, of arbitrary
heights, on a table. At each tick of a clock, one chip is removed from
each stack and those collected chips are used to create a new stack.
What eventually happens?

lxxxvii
 The Hints
Below are hints and/or comments for each puzzle, together with the
number of the chapter where you can find a solution. (But to get the
most out of these puzzles, don’t go there until you’ve tried everything!)

Adding, Multiplying, and Grouping: Multiply sums of two num-

bers and of five numbers. (Chapter 12)
Air Routes in Aerostan: What can be said about the two-airline
networks? (Chapter 4)
All Right or All Wrong: Use both choice and parity. (Chapter 23)
Alternate Connection: If you number the points in the order they
get connected, which one do you not want to connect? (Chapter 1)
Alternating Powers: Note that f (x) = x − f (x2 ). (Chapter 9)
Alternative Dice: Represent a die by a polynomial in which the
coefficient of xk is the number of faces with k pips. (Chapter 5)
An Attractive Game: Look at the game from the bank’s point of
view. (Chapter 14)
Angles in Space: Translate the convex closure of the points.
(Chapter 16)
Angry Baseball: What happens after the last tie? (Chapter 10)
Ants on the Circle: Conservation of momentum applies! (Chap-
ter 20)
Area–Perimeter Match: Factoring a polynomial might help here.
(Chapter 5)
Ascending and Descending: Consider, for each number in a good
list, the length of the longest decreasing subsequence that ends there.
(Chapter 12)
Attention Paraskevidekatriaphobes: What is the cycle for our
Gregorian calendar? (Chapter 1)
Attic Lamp: You’ll need more than one bit of information. (Chap-
ter 13)
Baby Frog: Cover the field with a 2n × 2n grid. (Chapter 15)
Bacteria on the Plane: Give each child half the potential of its
parent. (Chapter 18)

lxxxix
 Mathematical Puzzles

Bacterial Reproduction: Experiment, and keep track of the num-

bers of females. (Chapter 18)
Badly Designed Clock: Note that you can tell what time it is when
the hands coincide. (Chapter 19)
Bags of Marbles: Nothing to prevent you from putting other
things in bags... (Chapter 1)
Bat and Ball: Do the math! (Chapter 5)
Belt Around the Earth: Use a letter for the earth’s circumference
(Chapter 5)
Bias Test: In which case can the second flip change your mind?
(Chapter 13)
Biased Betting: Fix a win–loss sequence that would result in going
broke. (Chapter 10)
Bidding in the Dark: Compute your expectation after a given bid.
(Chapter 14)
Big Pairs in a Matrix: Suppose not, and look separately at the
biggest and next-biggest entries in each row. (Chapter 9)
Bindweed and Honeysuckle: Untwist one while you twist the
other. (Chapter 1)
Birthday Match: Be careful; you want a match for yourself, not
just any pair of people. (Chapter 10)
Bluffing with Reals: Jeremy has one threshold, Louise two.
(Chapter 21)
Boarding the Manhole: Archimedes may be able to help you here.
(Chapter 16)
Boardroom Reduction: Start by thinking about what would hap-
pen if the board came down to just its three most senior members.
(Chapter 21)
Box in a Box: Assume you can, and compute the volume of the
set of all points within distance ε of the boxes. (Chapter 16)
Boy Born on Tuesday: Count the cases carefully. (Chapter 10)
Bracing the Grid: Make a graph whose vertices are the rows
and columns of the grid, with edges where the intersecting square
is braced. (Chapter 4)
Breaking a Chocolate Bar: Try this, and see the light. (Chapter
18)
Broken ATM: What’s the most George can hope to get, consid-
ering that the machine’s only options are multiples of $6? (Chapter
6)
Buffon’s Needle: What is expected number of lines crossed, per
unit length, by a randomly placed unit-diameter circle? (Chapter 14)

xc
 The Hints

Bugging a Disk: Assume the ceiling is flat, and inscribe a hexagon

in it. (Chapter 9)
Bugs on Four Lines: Add a time axis. (Chapter 16)
Bugs on a Line: First, check that the bug cannot wander forever
without going off to infinity. (Chapter 18)
Bugs on a Polyhedron: Draw an arrow from each face through its
bug to the next face. (Chapter 18)
Bugs on a Pyramid: Number the bugs and ask: When one looks
at the triangle made by the others, does she see them in clockwise or
counterclockwise order? (Chapter 3)
Bulgarian Solitaire: Rotate the configuration 45◦ and think about
gravity. (Chapter 18)
Candles on a Cake: What does the puzzle tell you about the
distance from a fixed point on the circumference to the ith candle?
(Chapter 7)
Cards from Their Sum: Big numbers work better. (Chapter 6)
Chameleons: Look not just at numbers of chameleons of a given
color, but differences between numbers. (Chapter 2)
Chessboard Guess: Assign a 6-digit nimber to each square of the
chessboard. (Chapter 17)
Chinese Nim: Starting from (1,2), find a pattern linking the posi-
tions you don’t want to be in. (Chapter 21)
Chomp: Someone must have a winning strategy; what if it’s Bob?
(Chapter 9)
Circles in Space: Note that you can tile a sphere, minus any two
points, with circles. (Chapter 22)
Circular Shadows I: Can you alter a ball slightly without affecting
its coordinate projections? (Chapter 22)
Circular Shadows II: Close in on the body with parallel planes.
(Chapter 16)
Coconut Classic: Start by allowing yourself negative numbers of
coconuts! (Chapter 7)
Coin Game: Who would you want to be if one of you picks
HHHH and the other THHH? (Chapter 10)
Coin Testing: You need to keep the number of flips of each coin
as equal as possible—but there’s more. (Chapter 10)
Coins in a Row: You are not asked for an optimal strategy, only
one that gets Alix half the money. (Chapter 7)
Coins on the Table: The shape of the table plays a part. (Chapter
7)

xci
 Mathematical Puzzles

Colors and Distances: Notice that the distance between two

points in the plane is proportional to the probability that a random
line, suitably constrained, passes between them. (Chapter 14)
Comparing Numbers, Version I: Have Victor pick a threshold.
(Chapter 10)
Comparing Numbers, Version II: What would Paula do if Victor
used 1/2 as his threshold? (Chapter 10)
Competing for Programmers: You might have good use for a
“friendship graph,” mentioned earlier in the same chapter. (Chapter
4)
Conversation on a Bus: One part of the conversation keeps the
bus number from being too small, while another part keeps it from
being too big. (Chapter 13)
Conway’s Immobilizer: A good start is to make sure your algo-
rithm does the right thing when it’s about to win. (Chapter 11)
Covering the Stains: How would you try to cover the stains if you
had left your glasses at home, and couldn’t see where the stains were?
(Chapter 14)
Crossing the River: In general: Get the men across first. (Chapter
19)
Cube Magic: Orient the cube to get a large projection. (Chapter
22)
Curve and Three Shadows: Try curves made from grid lines.
(Chapter 16)
Curve on a Sphere: Pick two points on the curve halfway around
from each other, and imagine that the midpoint on the great circle
connecting them is the North Pole. (Chapter 9)
Curves on Potatoes: Replace the potatoes by holograms. (Chapter
16)
Cutting the Necklace: Imagine the necklace laid out in a circle,
with a line through the center cutting it twice. (Chapter 3)
Deterministic Poker: Alice must, at the very least, stop Bob from
making an ace-high straight flush. (Chapter 21)
Dishwashing Game: Begin with 12 flips each. (Chapter 10)
Divisibility Game: Bob gets the most out of small numbers.
(Chapter 6)
Domino Task: Notice that a vertical domino can be turned hori-
zontal if one of its rows has a hole. (Chapter 7)
Dots and Boxes Variation: Think about your choice from your
opponent’s point of view. (Chapter 9)

xcii
 The Hints

Dot-Town Exodus: Reason carefully with a population of 3.

(Chapter 13)
Double Cover by Lines: What happens if you try to build your
double-cover one line at a time? (Chapter 23)
Drawing Socks: Order all the socks randomly and condition on
the color of the last one. (Chapter 14)
Early Commuter: How much time did the husband save? (Chap-
ter 1)
Easy Cake Division: Straight, vertical cuts can do the job. (Chap-
ter 16)
Efficient Pizza-cutting: How can you make the next cut produce
the maximum number of new pieces? (Chapter 1)
Emptying a Bucket: One way to do it, paradoxically, is to show
that the contents of one bucket can be increased until another is
empty. (Chapter 21)
Even Split: Prove it first when n is prime. (Chapter 6)
Even-sum Billiards: What would your answer be if there were no
9-ball? (Chapter 2)
Expecting the Worst: Bend the segment into a circle. (Chapter
24)
Exponent upon Exponent: What happens to the expression if you
cut off the bottom x? (Chapter 23)
Factorial Coincidence: Consider the power of a big prime that
factors each side of the expression. (Chapter 6)
Factorials and Squares: Use the fact that the product of perfect
squares is itself a perfect square. (Chapter 6)
Fair Play: Flip more than once, and look for equiprobable events.
(Chapter 8)
Falling Ants: Note that ants are interchangeable, for the purposes
of this puzzle. (Chapter 20)
Faulty Combination Lock: Try dividing the numbers from 1 to 8
into two groups, and using the fact that any combination must contain
at least two from one of them. (Chapter 1)
Fewest Slopes: Try regular polygons. (Chapter 11)
Fibonacci Multiples: Keep track of the remainders, modulo n, of
two successive Fibonacci numbers. (Chapter 21)
Fifteen Bits and a Spy: Let the “nimsum” of the broadcast be the
message. (Chapter 17)
Figure Eights in the Plane: Note that there are only countably
many rational points in the plane—or pairs of them, for that matter.
(Chapter 23)

xciii
 Mathematical Puzzles

Filling a Bucket: How far can you get trying to keep all the buck-
ets level? (Chapter 19)
Filling the Cup: Note that the fractional parts of the amount of
rice you have got so far also constitutes a sequence of independent
uniform random variables. (Chapter 14)
Find the Robot: How many ways are there to start and point the
robot? (Chapter 23)
Finding a Jack: The jacks divide the deck into five parts. (Chapter
8)
Finding the Counterfeit: Each weighing can have three possible
outcomes. (Chapter 13)
Finding the Missing Number: How much information must be
kept? (Chapter 19)
Finding the Rectangles: What do you encounter as you walk
across the tiled 400-gon? (Chapter 22)
First Odd Number: List the alphabetically early words that may
arise in a number. (Chapter 11)
First-grade Division: Imagine that each project requires the same
total effort. (Chapter 18)
Flipping the Pentagon: A reasonable thing to try as a potential
might be some measure of how close the numbers are to one another.
(Chapter 18)
Flying Saucers: It’s important that the person beamed back down
might not be the first one beamed up by the next saucer. (Chapter 7)
Fourth Corner: Each grid point can have one of four possible par-
ities. (Chapter 2)
Frames on a Chessboard: Look for a set of cells the parity of
whose number of red cells doesn’t change. (Chapter 18)
Funny Dice: How can you give one die an advantage over another,
even though they both have the same average roll? (Chapter 19)
Game of Desperation: What position can you put your opponent
in that will force her to let you win on your next move? (Chapter 21)
Gaming the Quilt: If you’ve already bought k tickets, when is one
more worth buying? (Chapter 5)
Garnering Fruit: Try it for just apples and bananas; adding cher-
ries brings a ham sandwich(!) into the picture. (Chapter 3)
Gasoline Crisis: Start by imagining a trip that starts with plenty
of extra gas aboard. (Chapter 7)
Generating the Rationals: Note that you easily get all fractions
whose denominators are powers of 2. (Chapter 19)

xciv
 The Hints

Gladiators, Version I: If you think of strength as money, each

bout becomes a fair game. (Chapter 24)
Gladiators, Version II: Think of each gladiator as a light bulb
trying to outlast his opponent. (Chapter 24)
Gluing Pyramids: How can faces disappear other than by hiding
them? (Chapter 22)
Guarding the Gallery: Try putting guard-posts at corners. (Chap-
ter 15)
Half Grown: Think about your own young relatives. (Chapter 1)
Half-right Hats: Prisoners don’t all need to have the same instruc-
tions. (Chapter 2)
Halfway Points: If S has n points, they satisfy n linear equations
with rational coefficients. (Chapter 9)
Handshakes at a Party: What numbers did Alexandra hear during
her survey? (Chapter 4)
Hats and Infinity: Have the prisoners choose a special hat se-
quence from each class of similar sequences. (Chapter 23)
Hazards of Electronic Coinflipping: Flip several times, and note
that some numbers of heads will allow you to decide between Alice
and Bob. Can you give the rest to Charlie? (Chapter 3)
Home-field Advantage: What if all seven games were always
played? (Chapter 10)
Hopping and Skipping: What’s the probability that the frog never
regresses from its current pad? (Chapter 5)
Ice Cream Cake: It’s easier to keep track if you cut the cake in the
same place each time, but rotate the cake between cuts. (Chapter 21)
Identifying the Majority: Use the counter for reinforcement of
your current name. (Chapter 19)
Impressionable Thinkers: Consider the number of current in-
stances where influence fails, that is, a voter thinks one way and her
acquaintance disagrees the next day. (Chapter 18)
Increasing Routes: Note that in a graph the average degree is
twice the number of edges divided by the number of vertices. (Chapter
8)
Infected Checkerboard: Try the process with various starting po-
sitions. What prevents the infected area from getting too “compli-
cated”? (Chapter 18)
Infected Cubes: You need some way to generalize the diagonals
that work in dimension 2. (Chapter 24)
Integral Rectangles: There are many ways to do this one; one of
them begins by nesting the big rectangle into the positive quadrant of

xcv
 Mathematical Puzzles

the plane grid, and making a graph whose vertices are grid points that
lie on corners of rectangles. (Chapter 24)
Invisible Corners: You can start by putting yourself in a room
made from six non-touching planks. (Chapter 22)
King’s Salary: Try to reduce the number of salaried citizens.
(Chapter 11)
Lame Rook: What happens if you try this on an n × n board for
various (even) n? (Chapter 15)
Laser Gun: Cover the plane with reflected copies of the room.
(Chapter 24)
Lattice Points and Line Segments: When does the line segment
between two lattice points contain another lattice point? (Chapter 12)
Leading All the Way: Try putting the ballots in random circular
order, instead of random linear order. (Chapter 10)
Lemming on a Chessboard: If the lemming stays on the board, it
must eventually cycle. (Chapter 9)
Life Is a Bowl of Cherries: Consider the general two-bowl game
first. (Chapter 17)
Life Isn’t a Bowl of Cherries?: Start with the nim strategy itself,
rather than its reverse. (Chapter 17)
Life-saving Transposition: If you can do Names in Boxes, you can
do this one. (Chapter 19)
Light Bulbs in a Circle: Don’t forget to keep track of which bulb
you’re looking at, as well as the state of all bulbs. (Chapter 21)
Line Up by Height: Keep track of the lengths of the descending
and ascending subsequences each player is at the end of, when lined
up alphabetically. (Chapter 12)
Lines Through a Grid: Nineteen parallel diagonals will do it. Can
you do better? (Chapter 12)
Locker Doors: What numbers have an odd number of divisors
(including 1 and themselves)? (Chapter 6)
Losing at Dice: You’ll need to consider two patterns to compute
the number of ways to “roll a 4” by this strange method. (Chapter 1)
Lost Boarding Pass: Try it with three passengers. What is the
probability that third passenger ends up in the second passenger’s
seat? (Chapter 7)
Love in Kleptopia: Can Jan somehow get one of his locks on the
box? (Chapter 19)
Magnetic Dollars: Try it with, say, six coins instead of a million.
(Chapter 24)

xcvi
 The Hints

Majority Hats: Can you beat the “preponderance” strategy,

where each prisoner guesses the color he sees most of ? (Chapter 17)
Matching Coins: Cher needs to use her coinflips to communicate
information about coming flips. (Chapter 13)
Mathematical Bookworm: Visualize! (Chapter 22)
Measuring with Fuses: Try lighting both fuses at once. (Chapter
11)
Meet the Williams Sisters: How many pairs of players play each
other? (Chapter 1)
Meeting the Ferry: Draw a picture! Note that all times are GMT
so you don’t have to worry about timezones. (Chapter 22)
Men with Sisters: What’s the answer among you and your sib-
lings? (Chapter 8)
Missing Card: Make use of the ordering of the four remaining
cards. (Chapter 13)
Missing Digit: Do you remember “casting out nines”? (Chapter
2)
Monk on a Mountain: Superimpose the days. (Chapter 3)
More Magnetic Dollars: Try to set this up so that each urn ac-
quires coins at random moments. (Chapter 24)
Moth’s Tour: Start from when the moth first gets to 5 o’clock or
7 o’clock. (Chapter 21)
Names in Boxes: Nothing prevents a prisoner from using what he
finds in a box to decide what box to open next. (Chapter 19)
Napkins in a Random Setting: It’s helpful to think of each diner
having flipped a coin to decide which napkin she would prefer to take.
(Chapter 14)
Natives in a Circle: What would happen to their answers if we
changed all the truth-tellers to liars and vice-versa? (Chapter 11)
Next Card Bet: The best guarantee, for given expectation, is when
the outcome is no longer random. (Chapter 14)
Next Card Red: Try it with a smaller deck. (Chapter 24)
No Twins Today: It’s not a trick. (Chapter 11)
Non-repeating String: Let the number of letters in the alphabet
be a variable, and use induction. (Chapter 15)
North by Northwest: An easy enough computation, if you know
what it means. (Chapter 1)
Not Burning Brownies: Try rectangles, then some sort of triangle.
(Chapter 22)
Numbers on Foreheads: Consider the sum of the numbers, mod-
ulo 10. (Chapter 6)

xcvii
 Mathematical Puzzles

Odd Light Flips: Note that the order in which switches are flipped
is irrelevant. (Chapter 15)
Odd Run of Heads: Use a variable to stand for the answer. (Chap-
ter 5)
One-bulb Room: Consider first the situation in which the room
is known to be dark at the start. (Chapter 19)
Option Hats: Since each guess has probability only 1/2 of being
correct, to improve the odds the prisoners need to arrange things so
that either very few guess and are right, or many guess and are wrong.
(Chapter 17)
Other Side of the Coin: Consider labeling the sides of the coins.
(Chapter 10)
Packing Slashes: In showing your solution is optimal, it may pay
to consider the 12 outer vertices of the inside 3 × 3 grid. (Chapter 11)
Painting the Cubes: Paint space and carve out the big cube!
(Chapter 16)
Painting the Fence: How does the question of whether an inter-
val between painters gets painted depend on its length, relative to the
lengths of its neighbor intervals? (Chapter 14)
Painting the Polyhedron: Assume you can inscribe a sphere, and
triangulate the sphere using the tangent points as vertices. (Chapter
16)
Pairs at Maximum Distance: Note that line segments connecting
two max-distance pairs must cross. (Chapter 9)
Pancake Stacks: Making the stacks close in size limits one’s op-
ponent. (Chapter 21)
Path Through the Cells: Recolor greedily with high-numbered
colors, and start your path with a cell that still has color 1. (Chapter
15)
Peek Advantage: Could the second peek change your mind?
(Chapter 13)
Pegs in a Square: Look for a useful way to two-color the holes.
(Chapter 18)
Pegs on the Corners: Notice that if the pegs begin at grid points,
they stay on grid points. (Chapter 21)
Pegs on the Half-Plane: Weigh the holes so that the entire lower
half-plane has finite weight. (Chapter 18)
Phone Call: Tackle this one hour at a time. (Chapter 19)
Picking the Athletic Committee: What happens if you just pick
a committee at random and try to fix it? (Chapter 18)

xcviii
 The Hints

Pie in the Sky: The moon is about one degree in diameter. (Chap-
ter 20)
Ping-Pong Progression: How many points does Bob win, on av-
erage, while Alice wins 21? (Chapter 14)
Players and Winners: Think about communication in both direc-
tions. (Chapter 13)
Points on a Circle: Pick random diameters first, then endpoints.
(Chapter 10)
Poker Quickie: Assume no wild cards, one opponent. (Chapter
11)
Polygon Midpoints: What happens if you pick some arbitrary
point on the plane and assume it is a vertex? (Chapter 22)
Polygon on the Grid: What is the relationship of h or v to the
polygon’s area? (Chapter 19)
Polyhedron Faces: Think about the face with the most sides.
(Chapter 12)
Poorly Placed Dominoes: You might need two tries. (Chapter 19)
Portrait: Who is “my father’s son”? (Chapter 21)
Powers of Roots: Try the 10th power, and see if you can guess
what’s going on. (Chapter 7)
Powers of Two: Start from the middle of the phrase. (Chapter 1)
Precarious Picture: Somehow you must arrange the string so that
it ultimately passes over each nail, when the other is ignored. (Chapter
22)
Prime Test: If you don’t find any small prime factors, what else
can you try? (Chapter 6)
Prisoner and Dog: When is the woman close enough to the fence
to make a straight run for it, if the dog is at the opposite point? (Chap-
ter 19)
Prisoners and Gloves: For this you might want to consider parity
of permutations. (Chapter 2)
Profit and Loss: Observe first that it can’t be as long as 40 months.
(Chapter 15)
Protecting the Statue: Observe that the protected area cannot be
convex. (Chapter 22)
Raising Art Value: Compare the picture’s value to the averages
for the galleries. (Chapter 8)
Random Bias: Can you somehow “do the flips” before the coin is
chosen? (Chapter 10)
Random Chord: Does it matter how the random chord is chosen?
(Chapter 10)

xcix
 Mathematical Puzzles

Random Intersection: What is the probability that a particular

point in one ball will end up inside the other as well? (Chapter 14)
Random Intervals: Try small numbers, and look for an explana-
tion for what you find. (Chapter 24)
Random Judge: Imagine that there’s one judge that would take
part either way. (Chapter 10)
Rating the Horses: You’ll certainly need to see all the horses.
(Chapter 1)
Recovering the Polynomial: Think of the coefficients as the ex-
pansion of a number in base x. (Chapter 5)
Rectangles Tall and Wide: Start by showing that if the tile’s sides
have irrational ratio, then any tiling must be all horizontal or all ver-
tical. (Chapter 6)
Red and Black Sides: What’s the total area of the rectangles?
(Chapter 5)
Red and Blue Hats in a Line: How can the first one to guess help
the others? (Chapter 2)
Red Points and Blue: What improves if you uncross two seg-
ments? (Chapter 18)
Returning Pool Shot: How many angles can the ball take after it
is shot? (Chapter 20)
Righting the Pancakes: Represent the state of the pancake stack
as a binary number. (Chapter 18)
Rolling All the Numbers: Think of the experiment in stages.
(Chapter 14)
Rolling Pencil: Needs more than one second of thought, but two
will do. (Chapter 22)
Rolling a Six: Watch out—this is not the same as ignoring odd
rolls. (Chapter 14)
Rotating Coin: Try it! (Chapter 20)
Roulette for Parking Money: Every bet loses money, so you want
speed as well as accuracy. (Chapter 14)
Roulette for the Unwary: How many wins does Elwyn need to
come out ahead? (Chapter 14)
Rows and Columns: You have to try this to see what’s happening.
(Chapter 1)
Salaries and Raises: How much does each person’s annual salary
go up, each year? (Chapter 1)
Same Sum Dice: It’s easier to prove a stronger statement about
ordered dice. (Chapter 12)

c
 The Hints

Same Sum Subsets: What if you found two overlapping sets with
the same sum? (Chapter 12)
Second Ace: Compute the conditional probabilities. (Chapter 10)
Self-referential Number: If you try a number and it doesn’t work,
what’s the simplest way to try to fix it? (Chapter 19)
Sequencing the Digits: There’s more than one way to go about
choosing such a sequence. (Chapter 1)
Serious Candidates: Imagine betting on each candidate as he or
she becomes serious. (Chapter 14)
Service Options: It may be convenient to assume that lots of
games are played, regardless of outcomes. (Chapter 10)
Seven Cities of Gold: Equilateral triangles are your best friends
for this puzzle. (Chapter 11)
Sharing a Pizza: The even case is easy; try the odd case with pieces
of size 0(!) and 1. (Chapter 19)
Shoelaces at the Airport: Think about time spent walking. (Chap-
ter 1)
Shoes, Socks, and Gloves: What are the worst choices you can
make? (Chapter 12)
Signs in an Array: Is there anything good that must happen when
you flip a line that had negative sum? (Chapter 18)
Sinking 15: You may assume the players can sink whatever balls
they want, so this is a deterministic game. Does it seem familiar?
(Chapter 24)
Skipping a Number: Missy’s free-throw percentage jumps up
when she hits (and down when she misses), so it should be easy to
set up a situation where it jumps over 80%, right? (Chapter 3)
Slabs in 3-Space: Infinity is tricky—try covering a big finite piece
instead. (Chapter 16)
Sleeping Beauty: Imagine repeating the whole experiment 100
times. (Chapter 10)
Slicing the Cube: How would you actually do it? Can you do bet-
ter? (Chapter 24)
Snake Game: Tile the board with dominoes! (Chapter 4)
Soldiers in a Field: Think about the two closest soldiers. (Chapter
9)
Spaghetti Loops: What’s the probability that your ith tying oper-
ation will make a loop? (Chapter 14)
Sphere and Quadrilateral: Weigh the vertices so as to balance the
edges at the tangent points. (Chapter 20)

ci
 Mathematical Puzzles

Spiders on a Cube: How many spiders would be needed if the

game were played on a tree? (Chapter 4)
Spinning Switches: Try the two-switch version first. (Chapter 7)
Split Games: Imagine that for some reason the split games might
not be counted. (Chapter 10)
Splitting a Hexagon: Start with the triangles. (Chapter 22)
Splitting a Polygon: Non-convex polygons are a problem, but you
can start by passing a line through that’s not of the same slope as any
side. (Chapter 3)
Splitting the Stacks: A little experimentation might help you see
the light. (Chapter 1)
Sprinklers in a Field: What points are closer to a given grid vertex
than other grid vertices? (Chapter 19)
Squaring the Mountain State: What is meant by inscribing West
Virginia in a square is drawing a square around the shape of the state,
on the plane, in such a way that the state is inside the square but
touches all four sides. (Chapter 3)
Stopping After the Boy: Assume twins are rare. (Chapter 10)
Strength of Schedule: Consider the strength-of-schedule scores
for the teams with the most wins, and for the teams with the fewest
wins. (Chapter 5)
Subsets with Constraints: Group the numbers so that each group
can supply only one number to your set. (Chapter 6)
Subtracting Around the Corner: Notice that odd numbers even-
tually disappear. (Chapter 2)
Summing Fractions: Try small values for n, then prove your con-
clusion by comparing gains with losses going from n to n+1. (Chapter
15)
Sums and Differences: Suppose otherwise and start by consider-
ing the highest number. (Chapter 19)
Sums of Two Squares: Think of pairs of integers as lattice points
in a circle. (Chapter 8)
Swapping Executives: Start by getting an executive who belongs
at one of the ends into her correct place. (Chapter 15)
Swedish Lottery: Think about betting the highest allowed num-
ber, versus “cheating” and betting higher. (Chapter 21)
Testing Ostrich Eggs: Think about Oskar’s situation after the first
egg breaks. (Chapter 21)
Three Natives at the Crossroads: You will get no information
from the random native. (Chapter 13)

cii
 The Hints

Three Negatives: Consider the lowest two numbers in the set.

(Chapter 5)
Three Sticks: Can you pick the lengths so that an operation leaves
their ratios unchanged? (Chapter 3)
Three-way Duel: Compare Alice’s survival probabilities when she
hits Bob, hits Carol, or misses both. (Chapter 21)
Three-way Election: Consider the majority second choice among
each candidate’s fans. (Chapter 1)
Tiling a Polygon: How do you find a tile with sides parallel to two
given sides of the 100-gon? (Chapter 22)
Tiling with Crosses: Do the plane with diagonals, and space with
plane slabs that have holes and pegs. (Chapter 22)
Tiling with L’s: Try it one row at a time, left to right. (Chapter
15)
Tipping the Scales: How much of the time does a given weight
spend on each side of the scale? (Chapter 8)
Touring an Island: What do you need to know about Aloysius’
position at a particular moment? (Chapter 22)
Trapped in Thickland: Try regular tetrahedra. (Chapter 22)
Traveling Salesmen: Compare Lavish’s kth most costly flight to
Frugal’s. (Chapter 15)
Truly Even Split: Observe that 16 is a power of 2. (Chapter 15)
Turning the Die: Your position depends on the current sum, and
also which numbers are available on the die. (Chapter 21)
Two Balls and a Wall: Plot the velocities of the balls, making use
of conservation of energy and momentum. (Chapter 20)
Two Different Distances: You need to be very systematic to get
them all! (Chapter 11)
Two Monks on a Mountain: You can assume each path consists
of a finite number of alternately ascending and descending line seg-
ments. (Chapter 4)
Two-Point Conversion: You may as well assume the Hominids
will score again, and that if the game goes to overtime, either team is
equally likely to win. (Chapter 10)
Two Round-Robins: Divide the players into the good chess play-
ers and the good checkers players. (Chapter 5)
Two Runners: Relate sums and difference of distances covered to
meeting time. (Chapter 5)
Two Sheriffs: Have Lew and Ralph agree on a list of partitions of
the suspect set into pairs, such that each pair of suspects appears once
on the list. (Chapter 13)

ciii
 Mathematical Puzzles

Unanimous Hats: What would the prisoners do if they knew the

number of red hats was even? (Chapter 2)
Unbreakable Domino Cover: Can you cross a line with just one
domino? (Chapter 19)
Curves: Start with the busiest tiles. (Chapter 11)
Unbroken Lines: Do the horizontal and vertical lines first. (Chap-
ter 11)
Uniform Unit Distances: How might you extend a configuration
that works for n, so that it could work for n+1? (Chapter 15)
Uniformity at the Bakery: Start with integer weights, and make
one odd. (Chapter 15)
Uniting the Loops: Look at the more general problem of getting
a single loop from tilings involving some of the given tiles, with some
repetitions. (Chapter 2)
Waiting for Heads: There’s a bit more to this than the probability
of flipping 5 heads. (Chapter 8)
Watches on the Table: Draw a picture and use geometry. (Chapter
8)
Watermelons: Try some numbers. (Chapter 1)
Whim-Nim: Notice that prior to any declarations, handing your
opponent either a Nim-0 position or a bunch of singletons is fatal.
(Chapter 17)
Who Won the Series?: What would the answer be if instead you
heard that at most six games were played? (Chapter 10)
Whose Bullet?: Check your intuition with math. (Chapter 10)
Wild Guess: You’ll need representative real sequences for this
one. (Chapter 23)
Winning at Wimbledon: How many chances to win the match
can you arrange to get? (Chapter 10)
Wins in a Row: Imagine playing four games instead of three.
(Chapter 10)
Wires under the Hudson: Link wires in pairs at each end. (Chap-
ter 4)
Worms and Water: You’ll need something over Lori that the
worms can’t crawl underneath. (Chapter 19)
Worst Route: Try it for just four houses; use various addresses,
and experiment. (Chapter 4)
Y’s in the Plane: Put small rational circles around each Y’s end-
points. (Chapter 23)
Zero-sum Vectors: Try to build a sequence of new rows whose
partial sums are tightly controlled. (Chapter 12)

civ
 The Hints

Zeroes and Ones: What happens when you subtract from a num-
ber another number with the same value modulo n? (Chapter 12)

cv
 1. Out for the Count
There’s nothing more basic in mathematics than counting—but at the
same time, counting mathematical objects can be dauntingly difficult.
We’ll start with some simple puzzles and gradually introduce some
of the special tools that we sometimes need for counting.

Half Grown
At what age is the average child half the height that he or she will be
as an adult?

Solution: Most people guess too high, maybe thinking that if full
height is reached around age 16, then half-height should be around 8.

There are two problems with that reasoning: (1) the rate of growth
for a human being is not constant, and (2) babies have a substantial
head start of 20 inches or so (when you prop them up).
The right answer: two years old! (For a girl, it’s actually about 2 14
years, for a boy 2 12 .)

Powers of Two
How many people are “two pairs of twins twice”?

Solution: There are four words in the phrase suggesting the number
“2” and most people rightly divine that these 2’s should be multiplied,
not added. So the answer is 16, right?

1
 Mathematical Puzzles

Not so fast—a twin is only one person, thus a pair of twins, only
two. So the correct answer is 8.

Watermelons
Yesterday a thousand pounds of watermelons lay in the watermelon
patch. They were 99% water, but overnight they lost moisture to evap-
oration and now they are only 98% water. How much do they weigh
now?

Solution: 500 lb. The watermelons contain 10 pounds of solid mat-

ter that now comprises 2% of their final weight; divide 10 pounds by
0.02, and there you are. Apparently, quite a bit of evaporation took
place!

Bags of Marbles
You have 15 bags. How many marbles do you need so that you can
have a different number of marbles in each bag?
Solution: Counting 0 as a number, you might reasonably deduce
that you should put no marbles in the first bag, 1 in the second, 2 in
the third, etc., and finally 14 in the last. How many marbles is that?
The quick way to answer that is to observe that the average num-
ber of marbles in one of your 15 bags is 7. Thus the total number of
marbles is 15 × 7 = 105.
But there’s a trick: You can put bags inside bags! If you put the
empty first bag inside the second along with one marble, then the
second inside the third along with another marble, etc., you end with
the last bag containing all the marbles. So you only need 14 marbles
in all.

Salaries and Raises

Wendy, Monica, and Yancey were hired at the beginning of calendar
2020.
Wendy is paid weekly, $500 per week, but gets a $5 raise each
week; Monica is paid monthly, $2500 per month, and gets a $50 raise
each month; Yancey is paid yearly, $50,000 per year, and gets a $1500
raise each year.
Who will make the most money in 2030?

2
 Out for the Count

Solution: Yancey gains $1500 per year, thus he makes $50,000 +

10 × $1500 = $65,000, 10 years later.
Monica gains 50 × 78(!) = $3900 per year, making $30,000 + 10 ×
$39,000 = $69,000 at the end of the 10th year (which is when 2030
is about to start). So Monica out-earns Yancey in 2030, without even
considering the additional raises she is due that year.
Wait, why 78? Because Monica’s monthly raises add up to $50 ×
(1 + 2 + 3 + · · · + 12) = $50 × 13 × 12/2 = $50 × 78.
By similar reasoning, figuring 52 weeks in a year, Wendy gains
$5 × 53 × 52/2 = $6890 per year, making $26,000 + 10 × $6890 =
$94,000 at end of her 10th year. Yay Wendy!
Here’s yet another chance to add up consecutive numbers.

Efficient Pizza-cutting
What’s the maximum number of pieces you can get by cutting a
(round) pizza with 10 straight cuts?

Solution: There are a number of ways to tackle this one, but per-
haps the easiest is to note that the nth cut can at best cross each of
the n−1 previous cuts, and between each pair of crossings, split a
previous piece in two. Since you also get a new piece before the first
crossing and after the last one, the cut ends up adding n new pieces.
It follows that with n cuts you can create at most 1 + 2 + · · · + n =
n(n+1)/2 new pieces, but remember you started with one piece (the
whole pizza), so the answer is 1 + n(n+1)/2 pieces. For 10 slices, that
works out to 56 pieces.

Wait, we haven’t actually shown that you can achieve that many
pieces—that would require having every two cuts cross, with never
any more than two cuts crossing at the same place. But if we just
mark 2n random points along the edge of the pizza, number them
clockwise (say) and cut from the first to the n+1st, then the second to
the n+2nd, etc., we’ll get what we want with probability 1.

3
 Mathematical Puzzles

You don’t like random cuts? I’ll leave it to you to come up with
a deterministic way to make your 10 (or n) cuts subject to the above
constraints.

Attention Paraskevidekatriaphobes
Is the 13th of the month more likely to be a Friday than any other day
of the week, or does it just seem that way?

Solution: Amazingly, it is true—in the sense that in the 400-year

cycle of our Gregorian calendar, more months have their 13th falling
on a Friday than on any other day of the week.
It turns out that in 688 out of 4800 months in the 400-year cycle,
the 13th falls on a Friday. Sunday and Wednesday claim 687 each,
Monday and Tuesday 685 each, and Thursday and Saturday only 684
each. To check this you need to remember that years which are mul-
tiples of 100 are not leap years unless (like 2000) they are divisible by
400.
Incidentally, with some practice you can quickly determine the
day of the week of any date in history—even accounting for past cal-
endar glitches. For lazier or more present-time-oriented mortals, a
useful fact to remember is that in any year 4/4, 6/6, 8/8, 10/10, 9/5,
5/9, 7/11, 11/7, and the last day of February all fall on the same day
of the week. (This is even easier to remember if you happen to play
craps daily from 9 to 5.) Note that is doesn’t matter whether you like
to put the month-number before the day-number, or vice-versa.
The late John H. Conway (whose name you will see often in this
volume) liked to call that day of the week “doomsday” for that year.
Doomsday advances one day each year, two before a leap year. For
2020 doomsday was Saturday, thus Sunday for 2021, Monday for
2022 Monday, Tuesday for 2023 but Thursday for 2024, a leap year.
For the next puzzle, we introduce some basic combinatorics—
techniques for the art of counting.
Most readers will know that if you construct something in two
steps, and you have n1 choices for the first step and n2 for the second,
then you have n1 × n2 ways to do it. This extends to multiple stages;
I call this the “multiplication rule.” For example, the number of
different license plates you can make that display three letters fol-
lowed by two digits is 26 × 26 × 26 × 10 × 10 = 263 × 100. The
number of subsets of an n-element set is 2n , because you can con-
struct the subset by listing elements of the big set and one by one,
deciding whether or not to put that element into the subset.

4
 Out for the Count

If you want to count the subsets of a particular size k, you’ll need

the division rule, which says that if you “accidentally” count each ob-
ject m times, then you have to divide by m at the end. To count the
subsets of size k of an n-element set, we first count ordered subsets—
subsets with a designated first element, second, etc. There are n ways
to pick the first element of your ordered subset, n−1 to pick the sec-
ond, down to n−k+1 for the last, so n(n−1)(n−2) · · · (n−k+1) in
all. But any subset of size k can be ordered k(k−1)(k−2) · · · 1 = k!
ways, so we’ve counted each subset that many times. It follows that
the number of subsets of size k is really
n(n−1)(n−2) · · · n−k+1 n!
= .
k(k−1)(k−2) · · · 1 k!(n−k)!
( )
The latter expression is written nk and usually pronounced “n choose
k.”
Here’s an example where k is just 2.

Meet the Williams Sisters

Some tennis fans get excited when Venus and Serena Williams meet
in a tournament. The likelihood of that happening normally depends
on seeding and talent, so let’s instead construct an idealized elimi-
nation tournament of 64 players, each as likely to win as to lose any
given match, with bracketing chosen uniformly at random. What is
the probability that the Williams sisters wind up playing each other?

Solution: This looks complicated and in fact, in Frederick Mosteller’s

delightful book Fifty Challenging Problems in Probability, the solution (of
an equivalent problem) is obtained by working it out for 2k players
when k is small, guessing the general answer, and then proving the
guess to be correct by induction on k. ( )
But there’s an easy way to think of it: There are 642 = (63×64)/2
pairs of players, out of which 63 will meet (since 63 players need to be
eliminated to produce the winner), so the probability that a particular
pair meets must be 63/((63 × 64)/2) = 1/32.

Rating the Horses

You have 25 horses and can race them in groups of five, but having
no stopwatch, you can only observe the order of finish. How many

5
 Mathematical Puzzles

heats of five horses do you need to determine the fastest three of your
25?

Solution: Hmm. It’s clear that all the horses need to be run, since
an untested horse could easily be one of the fastest three. You could
do that in five heats but that would not clinch the deal, so you’ll need
at least six heats for sure.
Well, suppose you do partition the horses into five heats; you’d
then need to test the winners against one another. What will you know
then?
Suppose (without loss of generality) that the winner of your sixth
heat came from Heat 1, the runner-up from Heat 2, and the placer
from Heat 3. Of course the fastest horse overall is the winner of Heat
6, but the second-fastest could either be the runner-up of Heat 6 or
the runner-up of Heat 1. The third-fastest horse could be the placer in
Heat 6, the runner-up or placer in Heat 1, or the runner-up in Heat 2.
So the only horses “in the running” (sorry about that) for second
and third place are the runner-up and placer from Heat 1, the winner
and runner-up from Heat 2, and the winner from Heat 3. That’s five
horses; run then in Heat 7, and you’re done. It’s not hard to see you
can’t beat this elegant solution.

Shoelaces at the Airport

You are walking to your gate at O’Hare and need to tie your shoelace.
Up ahead is a moving walkway that you plan to utilize. To minimize
your time to get to the gate, should you stop and tie your shoelace
now, or wait to tie it on the walkway?

Solution: You should tie your shoelace on the walkway. Either way,
you spend the same amount of time walking on solid ground, and the
same amount of time tying your shoelace. But if you tie your shoelaces
on solid ground, you spend more time walking on the walkway.

6
 Out for the Count

Rows and Columns

Prove that if you sort each row of a matrix, then each column, the
rows are still sorted!
Solution: To prove this rather surprising assertion, let us imagine
that the matrix has m rows and n columns, and that after each row is
sorted (smallest values to the left, say), the entry on the ith row and
jth column is aij ; after the columns are sorted, it is bij . We need to
show that if j < k then bij ≤ bik .
This is one of those things which alternates between being mys-
terious and obvious, each time you think about it. One way to argue
the point is to note that bik is the ith smallest entry in the old column
{a1k , a2k , . . . , amk }. For every entry ai′ k which ends up below aik ,
the entry ai′ j that was on the same row but in the jth column is no
larger; thus, counting aij as well, there are at least i entries from the
old jth column that are no bigger than bik . Thus, the ith smallest en-
try in the old jth column (namely, bij ) is itself no bigger than bik and
we are done.
Is this more convincing than just trying an example? You be the
judge.

Three-way Election
Alison, Bonnie, and Clyde run for class president and finish in a three-
way tie. To break it, they solicit their fellow students’ second choices,
but again there is a three-way tie. The election committee is stymied
until Alison steps forward and points out that, since the number of
voters is odd, they can make two-way decisions. She therefore pro-
poses that the students choose between Bonnie and Clyde, and then
the winner would face Alison in a runoff.
Bonnie complains that this is unfair because it gives Alison a better
chance to win than either of the other two candidates. Is Bonnie right?

Solution: Bonnie is correct—in fact, she has understated the case;

assuming no voters change their minds, Alison will win for sure! To
see this, suppose Alison’s supporters mostly prefer Bonnie to Clyde
(so that Bonnie would beat Clyde in the proposed two-candidate
race). Then Bonnie’s supporters must prefer Clyde to Alison, other-
wise Clyde would have garnered fewer than 1/3 of the second-place
votes; similarly Clyde’s supporters prefer Alison to Bonnie. Thus, in
this case, Alison will beat Bonnie in the runoff.

7
 Mathematical Puzzles

If Alison’s supporters prefer Clyde to Bonnie, a symmetric argu-

ment shows that Alison will beat Clyde in the runoff. ♡
This puzzle serves as a warning: There may be more to some
tiebreakers than meets the eye!
Let us return to the basic multiplication rule, but with some addi-
tional cleverness required.

Sequencing the Digits

In how many ways can you write the numbers 0 through 9 in a row,
such that each digit other than the leftmost is within one of some digit
to the left of it?

Solution: To apply the multiplication rule when you build some-

thing in stages, you need the number of choices in a stage not to de-
pend on choices made in previous stages. That seems not to be the
case here; for example, if you start your sequence with 8 you have
two choices for the next digit (7 or 9) but if you start with 9, you have
only one choice for the next digit (8) and in fact now the whole se-
quence is determined.
You can get around
( ) this with some fanciness involving binomial
coefficients (those nk guys mentioned above), but there’s a nicer way:
Create the sequence in reverse! The last digit must be 0 or 9; if it’s 9,
the next-to-last must be 0 or 8, while if it’s 0 the next-to-last must be 1
or 9. No matter what you do there are two choices at every stage until
you reach the leftmost digit, so the number of sequences is 29 = 512.

Faulty Combination Lock

A combination lock with three dials, each numbered 1 through 8, is
defective in that you only need to get two of the numbers right to open
the lock.
What is the minimum number of (three-number) combinations
you need to try in order to be sure of opening the lock?

Solution: Often the easiest way to tackle a problem of this sort is

geometrically. The space of all possible combinations is an 8 × 8 × 8
combinatorial cube, and each time you try a combination, you cover
all combinations on the three orthogonal lines which intersect at that
point.

8
 Out for the Count

Once you see the problem this way, you are likely to discover that
the best way to cover all the points in the cube is to concentrate all
your test points in just two of the eight 4 × 4 × 4 octants. You will
then arrive at a solution equivalent to the one described below.
Try all combinations with numbers from {1, 2, 3, 4} whose sum is
a multiple of 4; there are 16 of those, since if you pick the numbers
on the first two (or any two) of the dials, the number on the third is
determined. Now try all the combinations you get by adding (4,4,4),
that is, by adding 4 to each of the three numbers; there are 16 more of
those, and we claim that together the 32 choices cover all possibilities.
It is easy to see that this works. The correct combination must
have either two (or more) values in the set {1, 2, 3, 4}, or two or more
values in the set {5, 6, 7, 8}. If the former is the case, there is a unique
value for the third dial (the one whose number may not be among
{1, 2, 3, 4}) such that the three were among the first 16 tested combi-
nations. The other case is similar.
To see that we can’t cover with 31 or fewer test-combinations, sup-
pose that S is a cover and |S| = 31. Let Si = {(x, y, z) ∈ S : z = i}
be the ith level of S.
Let A be the set {1, 2, 3}, B = {4, 5, 6, 7, 8}, and C =
{2, 3, 4, 5, 6, 7, 8}. At least one level of S must contain 3 or fewer
points; we might as well assume S1 is this level and |S1 | = 3. (If
|S1 | ≤ 2, there’s an easy contradiction.) The points of S1 must lie in
a 3 × 3 × 1 subcube; we may assume that they lie in A × A × {1}.
The 25 points in B × B × {1} must be covered by points not in
S1 . No two of them can be covered by the same point in S, thus,
S \ S1 (that is, the elements of S that are not in S1 ) has a subset T

9
 Mathematical Puzzles

of size 25 that lies in the subcube B × B × C. Now consider the set

P = {(x, y, z) : z ∈ C, (x, y, 1) ̸∈ S0 , (x, y) ̸∈ B × B}. A quick
count shows that |P | = (64 − 3 − 25) × 7 = 252. The points in P are
not covered by S1 , and each point in T can cover at most 3 + 3 = 6
points in P . Therefore there are at least 252 − (6 × 25) = 102 points
in P that must be covered by points in S \ S1 \ T .
However, |S \ S1 \ T | = 31 − 3 − 25 = 3, and each point in the
cube covers exactly 22 points. Since 22 × 3 = 66 < 102, we have our
contradiction. ♡

Losing at Dice
Visiting Las Vegas, you are offered the following game. Six dice are
to be rolled, and the number of different numbers that appear will be
counted. That could be any number from one to six, of course, but
they are not equally likely.
If you get the number “four” this way you win $1, otherwise you
lose $1. You decide you like this game, and plan to play it repeately
until the $100 you came with is gone.
At one minute per game, how long, on the average, will it take
before you are wiped out?

Solution: This is a trick, of course; otherwise it would be in Chapter

10, In All Probability, not this one. On the average, it’ll take forever for
you to be wiped out—the game is in your favor, and thus you may
never even dip below your $10 stake.
There are 66 = 46,656 ways to roll the dice. For four differ-
ent numbers to appear, you need either the pattern AABBCD or
AAABCD. There are
( ) ( )
6 4
· /2 = 45
2 2

versions of the former pattern, keeping the equinumerous labels in

alphabetical order; e.g., AABBCD, ABABCD, ACDABB, but not
BBAACD or AABBDC. ()
For the latter pattern, there are 63 = 20 versions.
In either case, there are 6 × 5 × 4 × 3 = 360 ways to assign num-
bers to the letters, for a total of 360 × 65 = 23,400 rolls. Thus, the
probability of winning is 23400/46656 = 50.154321%. ♡
If you win some bets with this game, don’t forget to send 5% of
your profits to me c/o Taylor & Francis.

10
 Out for the Count

Splitting the Stacks

Your job is to separate a stack of n objects into individual items. At
any time you may split a stack into two stacks, and get paid the prod-
uct of the sizes of the two stacks.
For example, for n = 8, you could split into a stack of size 5 and
a stack of size 3, getting paid $15; then you can split the 5-stack into
2 and 3 to get $6 more. Then the two threes can each be split ($2
each) and the remaining three twos ($1 each) for a total income of
15 + 6 + (2 × 2) + (3 × 1) = 28 dollars.
What’s the most money you can get starting with n objects, and
how can you get it?
Solution: Try a few other ways to deal with eight objects, and you
will find that they all earn the same $28. Does it really not matter how
you split the stacks?
On the web, this problem has been used to demonstrate something
called “strong
( ) induction.” But there’s an easy way to see that the out-
come is n2 dollars no matter what you do: Each (unordered) pair of
objects earns $1!
Imagine that there is a string connecting every pair of objects, and
when you split a stack, you cut every string between pairs of objects
that straddle the split. If there are i objects on one side of the split
and j on the other, you cut ij strings and get paid $ij. Eventually
( ) you
cut every string and get paid $1 for each one, so you earn n2 dollars
regardless of strategy.

North by Northwest
If you’ve never seen the famous Alfred Hitchcock movie North by
Northwest (1959), you should. But what direction is that, exactly? As-
sume North is 0◦ , East 90◦ , etc.
Solution: Sounds like it should be halfway between north (0◦ ) and
northwest (315◦ ), thus 337.5◦ , but that’s actually “north–northwest,”
not “north by northwest.” The preposition “by” moves the needle just
11.25◦ in the direction of its object; thus “north by northwest” puts
you at (360 − 11.25)◦ = 348.75◦ . Northwest by north, on the other
hand, is (315 + 11.25)◦ = 326.25◦ . To make things even more con-
fusing, “north by northwest” is commonly designated more simply as
“north by west.” So the movie title was not actually a phrase in use;
in fact the movie’s original title was “In a Northwesterly Direction.”

11
 Mathematical Puzzles

Doesn’t roll off the tongue quite as well, though, does it?
The next puzzle is, really, just a bit of arithmetic.

Early Commuter
A commuter arrives at her home station an hour early and walks to-
ward home until she meets her husband driving to pick her up at the
normal time. She ends up home 20 minutes earlier than usual. How
long did she walk?

Solution: This kind of timing problem can drive you nuts, until you
think about it the “right way.” Here, you notice that the commuter’s
husband apparently saved 10 minutes each way, relative to normal,
so must have picked up his wife 10 minutes earlier than usual. So she
walked for 50 minutes.
To solve the next puzzle, we return to those binomial coefficients.

Alternate Connection
One hundred points lie on a circle. Alice and Bob take turns con-
necting pairs of points by a line, until every point has at least one
connection. The last one to play wins; which player has a winning
strategy?

Solution: The player who connects the 98th point is doomed, since
there will be either one or two unconnected points left( after
) her move,
which her opponent can then deal with. There are 97 2 moves that
can be made before the 98th needs to be connected, and since this
number is even, the second player can always win.

Bindweed and Honeysuckle

Two vines, a bindweed and a honeysuckle, climb a tree trunk, start-
ing and ending at the same place. The bindweed circles the trunk
three times counter-clockwise while the honeysuckle circles five times
clockwise. Not counting the top and bottom, how many times do they
cross?

Solution: By mentally twisting the tree trunk with the vines still on
it, we see that the answer is the same as it would be if the bindweed
went straight up while the honeysuckle went around 5 + 3 = 8 times.

12
 Out for the Count

Therefore there are 9 meetings in total, and the answer, with the top
and bottom not counted, is 7.

Do you remember phone trees? Suppose you belonged to an or-

ganization that needed a method of spreading information by phone.
It was once common in such a case to set up a phone tree. The idea is
this: As a member, you will have a special list of certain other mem-
bers (at least one), and everyone on your list will have your name on
their list. If you gain possession of outside information of interest to
the other members, you call the members on your list. If, instead, you
get information from another member, you pass it on to all the other
members on your list.
For the phone tree to work, it must be that there is a path between
any two members; in other words, if any member gets information,
then eventually the whole organization must hear it. If we think of
each member as a point, with a line between two points when the
corresponding members are on each others’ phone lists, then the col-
lection of all points and lines must form a connected graph. (There’ll be
more about graphs in Chapter 4, Graphography.)
Why is it called a phone tree and not a phone graph? Because if
it is efficiently constructed, there are only n−1 lines needed in the
graph—equivalently, members will never get their own information
relayed back to them.
Here is an example of a phone tree for a 10-member organization
whose names happen to be the numbers from 0 to 9.

13
 Mathematical Puzzles

The question we address is: How many ways are there to set up
a phone tree for an organization with n members? This question
was answered by German mathematician Carl Wilhelm Borchardt
in 1860, the same year Johann Philipp Reis built (arguably) the first
telephone prototype.
Borchardt’s contribution has been mostly forgotten and the name
of an English mathematician, Arthur Cayley, is associated with the
formula to come below.
Before we state and prove “Cayley’s formula,” roughly how many
phone trees might you guess would be possible if for a 10-member
organization? Hundreds? Thousands?
The answer is a hundred million. Exactly!
Theorem. The number of trees on a set of n labeled points is nn−2 .

Proof. We will prove the formula by showing that there is

a one-to-one correspondence between the set of trees on points
{0, 1, 2, . . . , n−1}, and the set of sequences of length n−2 of the num-
bers {0, 1, 2, . . . , n−1}.
Since there are n ways to choose each entry of the sequence, we
know from the multiplication rule above that there are nn−2 such se-
quences. Thus, if we can establish the aforementioned one-to-one cor-
respondence, we are done.
A leaf of a tree is a point (better, a “vertex”) of the tree that’s at-
tached to only one other vertex. To go from tree to sequence, we find
the lowest-numbered leaf and start the sequence not with the label of
that leaf, but with the label of its one neighbor.
Then that leaf is erased and the procedure is repeated to determine
the next entry in the sequence, and this continues until there are only
two vertices left in the tree. That’s it!
Let’s try it with the tree shown above. We get the following chain
of ever-longer sequences and ever-smaller trees; in each tree, all leaves
are circled and the vertex attached to the current lowest-numbers leaf
is boxed.
To show that this is a one-to-one correspondence, we need to get
from any sequence to the unique tree that would yield that sequence.
It turns out, that’s quite easy and logical; let’s do it for the sequence
we just got, and check that we get the same tree back.
Note that the leaves of a tree will not themselves appear in that
tree’s sequence. Thus, given a sequence, we look for the lowest number
that doesn’t appear, and start building our tree by attaching that vertex
to the vertex corresponding to the leftmost number in the sequence.

14
 Out for the Count

Then we erase that leftmost number and repeat, now looking for the
smallest number missing from the new, shorter sequence that has not
already been used as a leaf. This continues until the sequence is gone,
and then we connect the two vertices that were never used as leaves
to complete the tree.
We will not formally prove here that the two algorithms described
above are inverses of each other, but hopefully the example is con-
vincing. ♡

15
Mathematical Puzzles

16
 2. Achieving Parity
To a mathematician, “parity” often means just odd versus even. Typi-
cally, the concept of parity is not mentioned in a puzzle; but when you
start to play with the puzzle, the concepts of odd and even emerge.
And when they do, you are often half the way to solving it.
Here’s a simple example.

Bacterial Reproduction
When two pixo-bacteria mate, a new bacterium results; if the parents
are of different sexes the child is female, otherwise it is male. When
food is scarce, matings are random and the parents die when the child
is born.
It follows that if food remains scarce a colony of pixo-bacteria will
eventually reduce to a single bacterium. If the colony originally had
10 males and 15 females, what is the probability that the ultimate
pixo-bacterium will be female?

Solution: Here a bit of experimentation might lead to the observa-

tion that the parity of the number of females cannot change—in other
words, the number of females must remain odd.
Thus the probability that the last pixo-bacterium will be female
is 1.
For the next puzzle, you have to come up with the parity idea on
your own.

Fourth Corner
Pegs occupy three corners of a square. At any time a peg can jump
over another peg, landing an equal distance on the other side. Jumped
pegs are not removed. Can you get a peg onto the fourth corner of the
square?

17
 Mathematical Puzzles

Solution: The first thing to do here is think of the initial square as

a cell of the plane grid, for example, the points (0,0), (0,1), (1,0), and
(1,1) on the XY -plane. Then the pegs will always be on grid points.
Grid points, however, have four possible parities: Each coordinate
can be even or odd. When a peg jumps, its parity is preserved; its X-
coordinate goes up by even number (2, 0, or −2), and likewise its
Y -coordinate.
The points of the unit cell above have all four parities among them,
so the corner that starts without a peg can never be occupied.

Unanimous Hats
100 prisoners are offered the chance to be freed if they can win the
following game. In the dark, each will be fitted with a red or black hat
according to a fair coinflip. When the lights are turned on, each will
see the others’ hat colors but not his own; no communication between
prisoners will be permitted.

Each prisoner will be asked to write down his guess at his own hat
color, and the prisoners will be freed if they all get it right.
The prisoners have a chance to conspire beforehand. Can you
come up with a strategy for them that will maximize their probability
of winning?

Solution: The probability that Bob (a specified one of the prison-

ers) guesses correctly is exactly 12 , no matter what the general strategy
is, so the probability that they all get it right cannot exceed 12 . But,
amazingly, they can achieve 12 .
The key is: Suppose Bob knows there are an even number of red
hats. Then he is home free, because if he sees an even number of red
hats on his fellow prisoners’ heads, he knows his hat is black; if he
sees an odd number of red hats, he knows his hat is red.
So the prisoners simply agree to all guess that the number of red
hats will be even. Either they’re all right, or they’re all wrong; and the
probability that the number of red hats is even is exactly 12 . (That’s

18
 Achieving Parity

true even if only one hat is determined by coinflip, since no matter

what the other hat colors are, that hat color determines the parity.)
Once you’ve got that puzzle, the next one is easy.

Half-Right Hats
One hundred prisoners are offered the chance to be freed if they can
win the following game. In the dark, each will be fitted with a red or
black hat according to a fair coinflip. When the lights are turned on,
each will see the others’ hat colors but not his own; no communication
between prisoners will be permitted.
Each prisoner will be asked to write down his guess at his own hat
color, and the prisoners will be freed if at least half get it right.
The prisoners have a chance to conspire beforehand. Can you
come up with a strategy for them that will maximize their probability
of winning?

Solution: Yes, they just arrange for 50 prisoners to guess their color
by assuming the total number of red hats is even, while the other 50
assume the number of red hats is odd. Then one group or the other is
bound to be right, assuming no poor fellow screws up!

Red and Blue Hats in a Line

There are n prisoners this time, and again each will be fitted with a
red or black hat according to flips of a fair coin. The prisoners are to
be arranged in a line, so that each prisoner can see only the colors
of the hats in front of him. Each prisoner must guess the color of
his own hat, and is executed if he is wrong; however, the guesses are
made sequentially, from the back of the line toward the front. Thus,
for example, the ith prisoner in line sees the hat colors of prisoners
1, 2, . . . , i−1 and hears the guesses of prisoners n, n−1, . . . , i+1
(but he isn’t told which of those guesses were correct—the executions
take place later).
The prisoners have a chance to collaborate beforehand on a strat-
egy, with the object of guaranteeing as many survivors as possible.
How many prisoners can be saved in the worst case?

Solution: It’s clear that no more than n−1 prisoners can be sure of
surviving, because the first one to guess—namely, the last in line—
has no clue. But he can pass a clue to the prisoner in front of him,

19
 Mathematical Puzzles

by letting him know whether the number of red hats among prisoners
1 througn n−1 is even or odd. How does he communicate? By his
guess, of course. The prisoners can agree, for example, that the last
prisoner in line guesses “red” if he sees an odd number of red hats,
“black” otherwise.
So that saves the life of prisoner n−1; what about the rest? They’re
saved too! Everyone has heard the last prisoner, and knows that after
him, all guesses will be correct. For example, if the ith prisoner to
guess hears that there are an even number of reds among prisoners
1 through n−1, and after the first guess he hears 5 “red” guesses, he
concludes that the number of red hats from himself forward is odd. If
he actually sees an even number of red hats in front of him, he knows
that his own hat must be red and guesses accordingly.
The next one is a bit trickier. Luckily, the prisoners are getting
pretty good with these games.

Prisoners and Gloves

Each of 100 prisoners gets a different real number written on his fore-
head, a black glove, and a white glove. Seeing the numbers on the
other prisoners’ foreheads, he must put one glove on each hand in
such a way that when the prisoners are later lined up in real-number
order and hold hands, the gloves match. How can the prisoners, who
are permitted to conspire beforehand, ensure success?

Solution: This seems impossible: Each prisoner sees 99 different

real numbers and has no idea where the number on his own forehead
fits in, so how does he know what to do? What information has he
got?
The parity here is of a more subtle sort: It’s the parity of a per-
mutation. Permutations are classified as “even” or “odd” according
to the parity of the number of swaps needed to sort it. If we write
a permutation in “one-line form,” so that (p1 , p2 , . . . , pn ) represents
the permutation that sends i to pi , then (1,2,3,4,5) is even since it is
already the identity (requiring 0 swaps), but (4,2,3,1,5)—which can
be sorted by swapping “1” with “4”—is odd.
Assume the prisoners themselves are numbered from 1 to 100
(e.g., alphabetically). Then the numbers on the prisoners’ foreheads
induce a permutation of the numbers from 1 to 100: If the ith pris-
oner’s forehead shows the jth smallest real number, i is mapped to j
by the permutation.

20
 Achieving Parity

For example, suppose the prisoners are Able, Baker, Charlie, and
Dog, numbered 1 to 4 in that order. Say their numbers, respectively,
are 2.4, 1.3, 6.89, and π. Then the permutation is (2,1,4,3) where
the “2” at the front represents the fact that the first prisoner got the
second-lowest number, etc.
This permutation happens to have even parity, since interchanging
the 2 and the 1, then the 4 and the 3, restores it to the identity (1,2,3,4).
Of course, Charlie cannot see the 6.89 on his own forehead so he
doesn’t know the parity of the permutation. But suppose he assumes
that his own number is the lowest of all; then he can compute the
parity. He’ll be right if his number is the lowest, or the third lowest, or
the fifth lowest, and so forth; wrong if his number happens to be the
ith lowest, i being an even number.
Let’s have Charlie put the white glove on his left hand (and black
on right) if his assumption leads him to an even permutation, and
the reverse if he gets an odd permutation. (In the example, Charlie’s
assumption leads him to the permutation (3,2,1,4) which is odd, so
he puts the black glove on his left hand.)
That works! If the permutation actually is even, then the pris-
oner with lowest forehead number would have computed “even” and
would have a white glove on his left hand. The next-lowest-forehead-
numbered prisoner would have computed “odd” and his black-gloved
left hand would end up holding the first guy’s black-gloved right hand,
and similarly down the line. If the actual permutation is odd, the glove
colors will be the complement of these, and will still work.

Even-sum Billiards
You pick 10 times, with replacement (necessarily!), from an urn con-
taining nine billiard balls numbered 1 through 9. What is the proba-
bility that the sum of the numbers of the balls you picked is even?

Solution: If there were no ball number 9, an even sum would have

probability exactly 12 since no matter what happens in your first 9
picks, the parity of the tenth ball will determine whether the sum is
even or odd; and that ball is equally likely to be either.
Now put ball 9 back in. No matter what order you choose the balls
in, you can ask your assistant to reveal the numbers to you with all the
9’s first. Then the above argument still applies: The parity of the last
ball shown you, whose number is between 1 and 8, will determine the
parity of the sum; so again, the sum is equally likely to be even or odd.

21
 Mathematical Puzzles

Oops, wait, there’s just one possibility we haven’t considered: that

you got the 9 ball every time! That has probability (1/9)10 and results
in an even sum. So altogether, the probability of an even sum is
( )10 ( )10 ( )10
1 1 1 1 1 1
1− + = +
2 9 9 2 2 9
which is about 0.50000000014. Probably not enough bias to be worth
betting on!
Of course, “even” means that the remainder after dividing by 2 is
0; “odd,” that the remainer is 1. For the next puzzle, you’ll need to
stretch your definition of parity to the remainder after dividing by 3;
and for the puzzle after that, after dividing by 9.

Chameleons
A colony of chameleons currently contains 20 red, 18 blue, and 16
green individuals. When two chameleons of different colors meet,
each of them changes his or her color to the third color. Is it possible
that, after a while, all the chameleons have the same color?

Solution: The key thing is to observe that, after each encounter of

two chameleons, the difference between the number of individuals of
any two colors remains the same modulo 3. In symbols, letting NR
stand for the number of red chameleons and NB and NG similarly,
we claim that, for example, NR − NB has the same remainder after
division by 3, after any two chameleons meet, that it did before. This
is easy to verify by checking cases. Thus these differences remain the
same modulo 3 forever, and since in the given colony none of those
differences is equal to zero modulo 3, we can never get two of the
color populations to be zero.
If, on the other hand, the difference of two color populations (say,
NR − NB ) is a positive multiple of 3, we can lower that difference
by having a red chameleon meet with a green one (or if there are no
green ones, by first having a red one meet a blue one). We repeat until
NR = NB and then have reds meet blues until only green chameleons
remain.
Putting all this together, and noting that if two differences are mul-
tiples of 3 then the third must be as well, we see that:
• If all three differences are multiples of 3, then any color can take
over the colony;

22
 Achieving Parity

• If just one of the differences is a multiple of 3, then the remaining

color is the only one which can take over the colony; and finally,
• If none of the differences are multiples of 3, as in the given prob-
lem, the colony can never become monochromatic and will re-
main fluid until other circumstances (e.g., birth, death) inter-
vene.

Now for a quickie.

Missing Digit
The number 229 has 9 digits, all different; which digit is missing?

Solution: What to do? You can type “2^29” into your computer or
calculator and look at the number for yourself, but is there a way to
do it in your head without getting a headache?
Hmm . . . maybe you remember a technique from grade school
called “casting out nines,” where you keep adding up digits and al-
ways end up with a number’s value modulo 9 (that is, its remain-
der after dividing by 9). It uses the fact that 10 ≡ 1 mod 9, thus
10n ≡ 1n ≡ 1 mod 9 for every n. If we denote by x∗ the sum of the
digits of the number x, then we get (xy)∗ ≡ x∗ y ∗ mod 9 for every x
and y.
It follows in particular that (2n )∗ ≡ 2n mod 9. The powers of 2
mod 9 begin 2, 4, 8, 7, 5, 1, and then repeat; since 29 ≡ 5 mod 6, 2n
mod 9 ≡ the fifth number in this series, which happens to be 5.
Now the sum of all the digits is 10 × 4.5 = 45 ≡ 0 mod 9, so the
missing digit must be 4. Indeed, 229 = 536,870,912.

Subtracting Around the Corner

A sequence of n positive numbers is written on a piece of paper. In one
“operation,” a new sequence is written below the old one according
to the following rules: The (absolute) difference between each num-
ber and its successor is written below that number; and the absolute
difference between the first and last numbers is written below the last
number. For example, the sequence 4 13 9 6 is followed by 9 4 3 2.
Try this for n = 4 using random numbers between, say, 1 and
100. You’ll find that after remarkably few operations, the sequence
degenerates to 0 0 0 0 where it of course remains. Why? Is the same
true for n = 5?

23
 Mathematical Puzzles

Solution: Considering values modulo 2 solves both problems. In

the n = 4 case, up to rotation and reflection, 1 0 0 0 and 1 1 1 0
become 1 1 0 0, then 1 0 1 0, then 1 1 1 1, then 0 0 0 0. Since this
covers all cases, we see that when working with ordinary integers, at
most four steps are required to make all the numbers even; at that
point, we may as well divide out by the largest common power of two
before proceeding. Since the value M of the largest number in the
sequence can never increase, and drops by a factor of 2 or more at
least once every four steps, the sequence must hit 0 0 0 0 after at most
4(1 + ⌈log2 M ⌉) steps. (The symbol ⌈x⌉ stands for the “ceiling” of x,
that is, the least integer greater than or equal to x.)
On the other hand, for n = 5, the sequence 1 1 0 0 0 (considered
either as binary or ordinary numbers) cycles via 1 0 1 0 0, 1 1 1 1 0, 1
1 0 0 0. ♡
A little analysis, using polynomials over the integers modulo 2,
shows that the salient issue is whether n is a power of 2.
For our next puzzle, we return to even versus odd, but now you’ll
need some keen powers of observation to see how to apply this idea.

Uniting the Loops

Can you re-arrange the 16 square tiles below into a 4 × 4 square—no
rotation allowed!—in such a way that the curves form a single loop?
Lines that exit from the bottom are assumed to continue at the same
point on the top, and similarly right to left; in other words, imagine
that the square is rolled up into a doughnut.

24
 Achieving Parity

Solution: This puzzle (along with two others in the book) was in-
spired by the work of artist Sol LeWitt, who loved to construct paint-
ings (and sculptures) from pieces that consisted of every possible way
to do some combinatorial task.
In the case of the present puzzle, examination of the tiles shows
that each side of a tile has two entry/exit points. The four points be-
longing to two adjacent sides are paired up in one of two ways: with
the two points closest to the common corner paired, and the other two
paired (each by a quarter of a circle); or by pairing each point close to
the corner to the one farther from the corner on the other side (using
two crossing quarter-ellipses).
Thus, for each corner there’s a decision to be made: to cross, or not
to cross. These four binary decisions result exactly in the 16 different
tiles you see.
Since when we wrap around the doughnut, none of the curved
lines terminate, the curves are forced to form loops. The question is
whether we can arrange the tiles to form just one loop.
As with many puzzles, our initial strategy is just to try various
arrangements and count the loops. You will find that no matter how
you do it, there always seem to be an even number of loops. If that
holds in general, then, of course, we can never make a single loop.
But why should the number of loops always be even?
One natural way to try to prove it might be to show that if we
switch two adjacent tiles, the number of loops always changes by an
even number. Alas, there are a lot of curved lines that enter a pair
of adjacent tiles (12 in all) and it’s awfully hard to deal with all the
different things that can happen when two adjacent tiles are switched.
We need to consider some much smaller change. How about
crossing or uncrossing one pair of curves on a tile? Of course, if we do
that, it would effectively destroy one tile and duplicate another. We
would then be in a different universe where we have as many as we
like of each tile, and can choose any 16 tiles to arrange into a square.
A little experimentation shows that if we take any configuration
involving the 16 original tiles—for example, the one above—and cross
or uncross curved lines on one side of one tile, the number of loops
always goes up or down by one. If that’s universally true, we arrive at
a conjecture: If the total number of crossings in the picture is even,
the number of loops will be even; if it is odd, the number of loops will
be odd. More experimentation seems to confirm this; can we prove
it?

25
 Mathematical Puzzles

The figure below shows the three things that can happen when you
make or undo a crossing. The first two create or destroy one loop, thus
they add or subtract one from the loop total. The third is impossible;
if a loop is oriented (say, clockwise), and the tiles are colored like a
checkerboard, then a loop alternately enters gray tiles vertically and
white tiles horizontally, or vice-versa. The loops shown violate this
condition.

Now, if you start with the first figure, which has 16 crossings and
four loops, you can create or delete crossings to get any desired con-
figuration. But every time you create or delete a crossing, the number
of crossings and the number of loops remain the same parity—both
odd, or both even. Eventually when you get to any arrangement of
the 16 original tiles, you are back to an even number of loops. ♡

26
 Achieving Parity

Our theorem was an open problem for some time, until parity was
unleashed on it. It concerns partitions of an abstract combinatorial
construction.
Fix a positive number n. A box is a Cartesian product of n finite
sets; if the sets are A1 , . . . , An , then the box consists of all sequences
(a1 , . . . , an ) such that ai ∈ Ai for each i.
A box B = B1 × · · · × Bn is a proper sub-box of A = A1 × · · · × An
if Bi is a proper nonempty subset of Ai for each i.
If each factor Ai has at least two elements, so that each factor
Ai can be partitioned into two non-empty sets, then by taking every
product of one set from each Ai we could partition the box A into 2n
proper sub-boxes. However,

Theorem. A box can never be partitioned into fewer than 2n proper sub-
boxes.
Proof. Call a sub-box odd if it has an odd number of elements, and
let O be the set of all odd sub-boxes of the big box A, and if B is a
sub-box of A, let OB consist of the elements of O that intersect B in
an odd set.
If B = B1 × · · · × Bn then in order for a set C to intersect B in an
odd number of elements, it must intersect every factor Bi in an odd
number of elements (because |C ∩ B| = |C ∩ B1 | × |C ∩ B2 | × · · · ×
|C ∩ Bn |).
But the probability that a random odd subset Ci of Ai intersects
Bi in an odd number of elements is exactly 12 . Why? Because we can
flip a coin to determine whether each a ∈ Ai is in Ci or not, ending
with some element a′ that’s not in Ai (remember, Bi is supposed to
be a proper subset of Ai ). This last element a′ is put into Ci or not so
as to assure Ci is odd. The elements of Bi that we flipped for consti-
tute a random subset of BI , and since exactly half the subsets of any
nonempty finite set are odd, the probability that Bi ∩ Ci is odd is 12
as claimed.
Each C ∈ O is partitioned by the B’s, and since C is odd, at least
one of the parts of the partition must be odd. If there are m parts of
the partition, it follows that the probability that C ∈ OB is at least
1/m. But we just showed that probability is 1/2n , so m ≥ 2n ! ♡

27
 3. Intermediate Math
The Intermediate Value Theorem (IVT) states that if a real number
changes continuously as it moves from a to b, then it must pass
through every value between a and b. This simple and intuitive fact—
arguably, more of a property of the real numbers than a theorem—is
a powerful tool in mathematics, and shows up often in puzzle solu-
tions.
A useful variation of IVT addresses the case when one real number
is moving continuously from a to b while another goes from b to a;
they must cross. For example:

Squaring the Mountain State

Can West Virginia be inscribed in a square?
Solution: We are referring here to the shape of West Virginia as pro-
jected (any way you like) onto a plane. We begin by drawing a vertical
line west of the state and moving it east until it just touches the state
boundary. Then we draw another vertical line east of the state, mov-
ing it west until it too touches the state boundary. Similar operations
with two horizontal lines inscribe the state in a rectangle whose width
w is somewhat greater than its height h.

29
 Mathematical Puzzles

To get a square, we rotate the rectangle continuously, maintaining

its 90◦ angles while keeping all four sides tangent to the state. After
90◦ of rotation, the figure is the same as it was at the start, but now
w has become h and h has become w. It follows that at some angle,
h and w were equal; and at that point, the state was inscribed in a
square. ♡
Of course, this works for any state, and indeed for any bounded
shape in the plane.

The same technique has a practical application: You can always

rotate your square four-legged picnic table so as to have all four legs
in contact with the ground, assuming the spot you chose doesn’t have
any actual surface discontinuities. Just put the table down with three
legs touching and rotate until the fourth touches; this is guaranteed to
happen before you get to 90◦ , since when you do get to 90◦ , a ground-
contacting leg will be where your fourth leg began.
You might find the following classic puzzle even simpler.

Monk on a Mountain
A monk begins an ascent of Mt. Fuji on Monday morning, reaching
the summit by nightfall. He spends the night at the summit and starts
down the mountain on the same path the following morning, reaching
the bottom by dusk on Tuesday.

30
 Intermediate Math

Prove that at some precise time of day, the monk was at exactly
the same spot on the path on Tuesday as he was on Monday.
Solution: If we plot the monk’s position as a function of time of day,
we get two curves connecting opposing corners, and the conclusion
follows from IVT. But here’s another way to think of it, underscoring
the intuitiveness of IVT: Superimpose Monday and Tuesday, so now
the Monk climbs at the same time that his doppelgänger descends. They
must meet!

Cutting the Necklace

Two thieves steal a necklace consisting of 10 red rubies and 14 pink
diamonds, fixed in some arbitrary order on a loop of golden string.
Show that they can cut the necklace in two places so that when each
thief takes one of the resulting pieces, he gets half the rubies and half
the diamonds.

Solution: We could check all possible necklaces meeting the above

description, but there are more than 40,000 of them, and, anyway,
we’d like to solve the problem for any (even) number of rubies and of
diamonds.
To apply the IVT requires a bit of subtlety; we first have to set up
the problem geometrically, then make a couple of key observations.
The idea is to represent the necklace as a circle cut into 24 equal
arcs, each colored red or pink according to the pattern of rubies and
diamonds in the necklace. Then we draw a line through the center of
the circle, cutting the circle twice.
Our line will certainly cut the full set of gems into equal portions,
since any diameter cuts the circle into equal-length arcs (i.e., semicir-
cles). Thus, if a diameter happens to cut the rubies in half, it will cut
the diamonds in half as well.
The line pictured in the figure above does not work; it’s got more
red on one side than the other. But as we rotate the line, the fraction of

31
 Mathematical Puzzles

red on the redder side changes continuously, and after 180◦ of rotation
it has switched from redder to pinker. The IVT tells us that there is a
point when there was the same amount of red on both sides of the
line.
The remaining subtlety: What if that position involves cutting
through a gem? Well, it can’t cut a ruby on one end while it cuts a
diamond at the other end, because then the number of rubies on one
side of the line would not be an integer—and half of 10 is an integer.
So either both ends of the line cut through a ruby or both through
a diamond, and if we move the line a bit until it no longer cuts through
a gem, the number of rubies on one side will not change.
Another way to look at it: You can apply the IVT to integers, if
you replace the requirement of continuity by the requirement that the
integer value changes only by one. If the rotating line always points to
the boundary between two gems (at both ends), and moves one gem
at a time, the number of rubies on one side can never change by more
than one. We can thus conclude that as the line rotates 180◦ , it must
at some point have half the rubies on each side.
In some puzzles the need for IVT is not at first obvious.

Three Sticks
You have three sticks that can’t make a triangle; that is, one is longer
than the sum of the lengths of the other two. You shorten the long
one by an amount equal to the sum of the lengths of the other two,
so you again have three sticks. If they also fail to make a triangle, you
again shorten the longest stick by an amount equal to the sum of the
lengths of the other two.

32
 Intermediate Math

You repeat this operation until the sticks do make a triangle, or

the long stick disappears entirely.
Can this process go on forever?
Solution: You can certainly make the process go on indefinitely if
you design it so that the ratios of the lengths of the sticks are un-
changed after one operation. For that to happen with lengths a < b <
c you’ll need b/a = c/b = a/(c − a − b); let that ratio be r > 1, so that
the lengths are proportional to 1, r, r2 and satisfy 1/(r2 − r − 1) = r,
r3 − r2 − r − 1 = 0.
Writing f (r) = r3 − r2 − r − 1, we need an r > 1 for which
f (r) = 0. Does such an r exist? Yes, because f (1) = −2 < 0 while
f (2) = 1 > 0; since all polynomials are continuous functions, we can
apply the IVT to deduce that there is a root of f between 1 and 2.
It remains only to check that our root gives us sticks that don’t
make a triangle, but that’s easy, because r2 = r3/r = (r2 + r + 1)/r =
r + 1 + (1/r) > r + 1. ♡
Here’s another case where IVT saves you the trouble of finding a
root of a polynomial.

Hazards of Electronic Coinflipping

You have been hired as an arbitrator and called upon to award a cer-
tain indivisible widget randomly to either Alice, Bob, or Charlie, each
with probability 13 . Fortunately, you have an electronic coinflipping
device with an analog dial that enables you to enter any desired prob-
ability p. Then you push a button and the device shows “Heads” with
probability p, else “Tails.”
Alas, your device is showing its “low battery” light and warning
you that you may set the probability p only once, and then push the
coinflip button at most 10 times. Can you still do your job?

Solution: If you could set p twice, two flips would be enough: Set
p = 13 and on Heads give the widget to Alice, else reset to 21 and use
the result to decide between Bob and Charlie. Since you can only set
p once, it might make sense to try to design a scheme first, then pick
p to make the scheme work.
For example, you could flip your coin three times and if the flips
are not all the same, you could use the “different” flip to decide to
whom to award the widget (e.g., “HTT” means Alice gets it). But if
you get all heads or all tails, you’ll have to do it again, and maybe yet
again—so, not in any bounded number of flips.

33
 Mathematical Puzzles

But now suppose we flip the coin four times. There are four ways
to get one head, six to get two, four to get three: Even numbers all, so
anytime you get non-uniform results you can use the flips to decide
between Alice and Bob. If the coin is fair, the remaining probability—
that is, the probability q = p4 + (1 − p)4 of “HHHH or TTTT”—is
only 18 , too small for Charlie.
As you reduce p, however, q changes continuously and approaches
1. Thus, by IVT, there is some p for which the probability of “HHHH
or TTTT” is exactly the 13 you need. ♡

Bugs on a Pyramid
Four bugs live on the four vertices of a triangular pyramid (tetrahe-
dron). Each bug decides to go for a little walk on the surface of the
tetrahedron. When they are done, two of them are back home, but
the other two find that they have switched vertices.

Prove that there was an instant when all four bugs lay on the same
plane.

Solution: Let the bugs be A, B, C, and D. Initially when D looks

down on the ABC plane she sees the ABC triangle labeled clockwise,
but in final position she sees it counterclockwise (or vice-versa). Since
the bugs move continuously, there must either be some position where
she is on the plane ABC, or some position where ABC do not deter-
mine a plane. But in the latter case they are collinear and thus they are
coplanar with any other point, thus with D’s position in particular.
In the next puzzle, IVT’s continuity hypothesis is missing.

34
 Intermediate Math

Skipping a Number
At the start of the 2019 season WNBA star Missy Overshoot’s lifetime
free-throw percentage was below 80%, but by the end of the season it
was above 80%. Must there have been a moment in the season when
Missy’s free-throw percentage was exactly 80%?
Solution: Let X(t) be the fraction of Missy’s free-throws that have
gone in, up to time t. This value jumps up or down each time Missy at-
tempts a free throw, and is always a rational number. The IVT clearly
does not apply, so your immediate reaction is probably “no”—there’s
no reason why X(t) should have to be exactly 4/5 at some point.
But it seems annoyingly difficult to actually construct a “history”
for Missy that misses 80% (try it). We could certainly get her past
70%; perhaps she’s 2 for 3 at the start of the season, then sinks her
next free throw, jumping from 66 23 % directly to 75%. But 80% seems
impossible to skip. What’s going on here?
As is often the case, it pays to try simpler numbers. Can we get
Missy past 50% without hitting it? Ah—the answer is no, because if
H(t) is Missy’s number of hits up to time t and M (t) her number of
misses, then the difference H(t) − M (t) starts off negative and ends
positive. Since this difference is an integer which changes by only 1
when a free throw is attempted, it must (by applying integer-IVT) hit
0; at that moment, H(t) = M (t) and therefore Missy’s success rate is
exactly 50%.
Returning to 80%, we observe that at the moment when X(t) =
.8, if there is one, we have H(t) = 4M (t). Now, let t0 be the first time
at which Missy’s percentage hits or exceeds 80%. That moment was,
obviously, marked by a successful free throw; in other words, at time
t0 , H(t) went up by one while M (t) stayed fixed. But H(t) can’t have
skipped from below 4M (T ) to above 4M (t); again by integer-IVT, it
must hit 4M (t). So 80% can’t be skipped!
Observe: (1) Our argument required that Missy went from below
80% to above it. In fact, she can easily skip 80% going down. (2) The ar-
gument uses the fact that at 80%, the number of hits will be an integer
multiple of the number of misses. Thus, it works for any percentage
that represents a fraction of the form (k−1)/k, and you can readily
verify that any other fraction between 0 and 1 is skippable going up.
Going down, it is the fractions of the form 1/k that can’t be skipped.
Some processes are piecewise continuous, that is, continuous ex-
cept at a discrete set of points. To apply IVT in such a case, you might
have to “limit the damage” at jumping points.

35
 Mathematical Puzzles

Splitting a Polygon
A chord of a polygon is a straight line segment that touches the poly-
gon’s perimeter at the segment’s endpoints and nowhere else.
Show that every polygon, convex or not, has a chord such that
each of the two regions into which it divides the polygon has area at
least 13 the area of the polygon.
Solution: Call the polygon P and scale it so that it has area 1. If
P is convex we can always find a chord that cuts P into two pieces
of area 12 , as you, dear reader, now an expert with IVT, can readily
prove. In fact, you can choose arbitrarily the chord’s angle to the hor-
izontal. Just move a line at that angle across P ; since P is convex the
intersection of its interior with the line consists of a single chord. Be-
hind that chord is a piece of P whose area changes continuously from
0 to 1.
Try that when P is nonconvex, and things get messy: The line’s
intersection with P ’s interior may consist of several chords. Indeed,
some polygons simply can’t be cut into two equal-area parts by a
chord—see the figure below for an example.

When P is nonconvex, we simplify things by first choosing an

angle for our moving line L that doesn’t encounter any more trouble
than it needs to. We do that by ensuring that L is not parallel to any
line through two vertices of P . Then the worst that can happen as we
move our chord is that it encounters some concave vertex; it can’t hit
two vertices (or, therefore, any whole edge of P ) at the same time.
Suppose we’ve moved L across P but so far the area of the smallest
piece into which P has been cut has not reached 13 . If we hit a concave
vertex v head-on (as pictured on the left side of the figure), we may
be forced to split the chord. At that point we have cut P into three
pieces, with, say A (of area a) behind the chord and B and C (areas b
and c, respectively) ahead. At least one of b and c (say b, as in the figure
below) exceeds 13 ; we keep that part of the chord. If in fact 31 ≤ b ≤ 23 ,
we stop there; else we keep moving L forward.

36
 Intermediate Math

There is one other possibility: L hits v “from behind,” for example,

as pictured on the right, forcing the chord to suddenly jump in length.
If the region ahead (C) is of area at least 13 , we can stop here or plow
ahead as before. The fun case is when, as pictured, a and c are each
less than 13 : Then we must flip our chord over v and send it back the
other way into B!
In this fashion we continue pushing our chord through P , always
with more than area 23 ahead of us, until we have the desired split.

The last (and toughest) puzzle in this chapter doesn’t look at all
like a case for IVT, and even less like IVT’s three-dimensional ana-
logue. But let’s see how it plays out.

Garnering Fruit
Each of 100 baskets contains some number (could be zero) of apples,
some number of bananas, and some number of cherries. Show that
you can collect 51 of those baskets that together contain at least half
the apples, at least half the bananas, and at least half the cherries!

Solution: If we had just apples, then we could clearly get half the
fruit in only 50 baskets by just taking the 50 baskets with the most
apples. Or we could divide the baskets into two groups of 50 any
way we want, and just take the part with more apples. Or—more
geometrically—we could line up the baskets, divide the line (possibly
through the middle of a basket, maybe cutting apples in the basket)
in such a way that exactly half the apples are on either side of the
cut, and note that there at most 49 whole baskets on one side of the
cut or the other. Take those 49 plus all of the cut basket to get your

37
 Mathematical Puzzles

winning 50 baskets. (If no basket is cut, just take a side with at most
50 baskets.)
With two fruits, we can’t guarantee half of each with 50 baskets
out of 100; for example, one basket with just one apple in it, and the
remaining 99 with one lonely banana in each, would require 51. But
we can get 50 out of 99 in several ways.
For example, we could arrange the baskets by decreasing num-
ber of apples, then take basket 1, and from each pair (2,3), (4,5), ...,
(98,99) the basket with more bananas. This certainly gets at least half
the bananas since it gets at least half from baskets 2 through 99, plus
all in basket 1. And it gets at least half the apples since at worst it gets
all the apples from baskets 1, 3, 5, 7, etc. and basket k has as many
apples as basket k+1.
(Note that this algorithm uses only order information about the
numbers of apples or bananas in baskets, not the actual amount.)
Another way to do it is to space the baskets equally around a circle,
in any order. Now consider the 99 different “arc sets” you get by tak-
ing 50 baskets in a contiguous arc from the circle. We claim that most
(i.e., at least 50) of the arc sets contain at least half the apples. Why?
if they didn’t, then more than half of the complements of the arc sets
would contain more than half of the apples. But the complements are
arc sets of size 49, and each of those can be thought of as a subset of
a different arc set of size 50 (say, the one you get by adding the next
basket clockwise). Thus, if most of the 49-ers contained more than
half the apples, the same must be true of the 50’s, a contradiction.
But if at least 50 of the size-50 arc sets contain at least half the
apples, and (by a similar argument) at least 50 of them contain at
least half the bananas, then at least one arc set contains at least half
of both fruits.
It’s worth noting that this proof shows that even if the baskets are
arbitrarily numbered, there is a list of just 99 of the 5 × 1028 subsets
of size 50, one of which is guaranteed to have the desired property.
It’s even more instructive to wonder if the circular arrangement of the
baskets on the plane allows a straight-line cut, possibly right through
the middle of two baskets and through some apples and bananas, that
cuts the fruits exactly in two. If so, there are at most 48 baskets on one
side or the other of the cut to which we can add the cut baskets to get
50 with at least half of each fruit. And this is the kind of thing we need
to solve the problem when there are three (or more) types of fruit.

38
 Intermediate Math

The combinatorial methods don’t seem to generalize to the three-

fruit case, so let’s try the geometric ones. Suppose we place the baskets
in 3-space, and spread them out enough so that no plane cuts more
than three baskets. With three degrees of freedom, we can find a plane
that cuts the apples, bananas, and cherries each exactly in half. With
100 baskets, there must be at most 48 whole baskets on one side or
the other of the line. Put them together with the cut baskets, and we
have 51 that contain at least half the apples, half the bananas, and half
the cherries. (If fewer than three baskets were cut, it only gets easier.)
We’ve been a bit glib with this “degrees of freedom” business, but
it really is true (and known usually as the Stone–Tukey Theorem) that
in n-dimensional space, any n sets can be simultaneously bisected
by an n−1-dimensional hyperplane. The three-dimensional version is
known as the ham-sandwich theorem and says that any sandwich con-
sisting of a slice of ham and two slices of bread, however sloppily
piled, can be cut with a planar cut that exactly halves the ham and
each slice of bread. Such facts are usually proved using fancy mathe-
matics (algebraic topology).
Applying Stone–Tukey allows us to generalize to b baskets and
n kinds of fruit; the result is there are always the greatest integer in
b/2 + n/2 baskets with at least half of each fruit type. This is best
possible, because if b = m1 + · · · + mn + x where all mi are odd and
x = 0 or 1, then we can have mi boxes with nothing but one sample
of fruit i, and possibly one empty box. To get half of each fruit type
with this arrangement would costs us at least n/2 half-baskets beyond
b/2.

Of the many theorems that employ IVT in their proofs, the most
famous is the Mean Value Theorem of calculus, which says essentially
that if you drive 60 miles in an hour then there must have been a
moment when your speed was exactly 60 miles per hour. But we’re
avoiding calculus in this book and we don’t need it here: A simple
theorem about polynomials (simple compared to some of the above
puzzles, for sure!) will do.
For us a “polynomial in the variable x” is an expression p(x) of
the form a0 + a1 x + a2 x2 + · · · + ad xd where d is a non-negative
integer, the coefficients ai are real numbers, and ad is not zero. The
“degree” of p(x) is d, the exponent of the highest power of x having
nonzero coefficient.
Theorem. If a polynomial p(x) has odd degree, then it has a real root, that
is, there is a real number r such that p(r) = 0.

39
 Mathematical Puzzles

The oddness condition for the degree of p(x) is certainly neces-

sary; for example, no real number x satisfies 1 + xd = 0 when d is an
even non-negative integer.
The proof is a snap if you’ve gotten this far in the chapter. The
idea is to first assume that ad = 1 (dividing p(x) by ad does not affect
its roots). Then, for large enough values of x, p(x) will be positive; for
highly negative values of x, p(x) will be negative. Thus, since powers
and thus polynomials are always continuous, IVT tells us there will
be some x0 with p(x0 ) = 0.
How big does x have to be to guarantee that p(x) > 0? If we let
a = |a0 | + |a1 | + · · · + |ad−1 | be the sum of the absolute values of the
other coefficients, then x > max(a, 1) is good enough. The reason is
that then the powers of x are increasing, thus

p(x) ≥ xd − axd−1 = xd−1 (x − a) > 0.

Similarly,

p(−x) ≤ (−x)d + axd−1 = xd−1 (a − x) < 0.

This tells us a little more than the theorem, namely, that p(x) has
a root in the union of the intervals [−a, a] and [−1, 1], where a =
(|a0 | + |a1 | + · · · + |ad−1 |)/|ad |.

40
 4. Graphography
Among the simplest of mathematical abstractions is the one which
models objects with points (called “nodes” or “vertices”) and indi-
cates some relationship between a pair of nodes by the presence of a
line (called an “edge”) from one to the other. Such a structure is called
a graph, not be confused with the graph of a function like y = x2 − 4.
A graph is said to be connected if you can get from any node to any
other by following a path of edges. It’s not hard to see that you need
at least n−1 edges to connect n nodes; three ways to do that with five
nodes are shown in Fig below:

A graph with the minimum number of edges to be connected is

called a tree (we counted those at the end of Chapter 1). Note that
a tree cannot have a cycle (a closed path with three or more nodes)
because removal of an edge from a cycle cannot disconnect a graph.
Thus a tree can also be described as a graph with the maximal number
of edges to not have a cycle.
Consider the following connectivity problem.

41
 Mathematical Puzzles

Air Routes in Aerostan

The country of Aerostan has three airlines which operate routes be-
tween various pairs of Aerostan’s 15 cities. What’s the smallest pos-
sible total number of routes in Aerostan if bankruptcy of any one of
the three airlines still leaves a network that connects the cities?
Solution: Of course our graph will have 15 nodes, one for each city,
and we can add edges of three types (say, solid black, dashed black,
and solid pink) to represent the routes of the three airlines. From the
above remarks, it follows that any two of Aerostan’s airlines must op-
erate at least 14 routes, and thus together the airlines must operate at
least 21 routes. (Why? Add up the inequalities a + b ≥ 14, a + c ≥ 14,
and b + c ≥ 14).
Can 21 be achieved? If so, each airline will need exactly seven
routes, and it must be the case that every cycle of routes uses routes
from all three airlines (otherwise bankrupting the missing airline
would leave a cycle, wasting one of the 14 remaining edges). There
are lots of ways to do this; perhaps the simplest is to pick a hub city,
let Airline A connect the hub to a set of seven other cities, and let
Airline B connect the same hub to the remaining seven cities. Airline
C avoids the hub but connects A’s seven spoke cities pairwise to B’s.
The figure below shows this graph and another.

42
 Graphography

The second of these is known as a “windmill graph” for obvious

reasons. If the nodes represent people and the edges friendships, the
graph has the property that every two people have a unique friend in
common. (There is only one connected graph with this property and
a given odd number of nodes. If you want this property with an even
number of nodes, you’re out of luck!)

Spiders on a Cube
Three spiders are trying to catch an ant. All are constrained to the
edges of a cube. Each spider can move one third as fast as the ant can.
Can the spiders catch the ant?
Solution: If our protagonists were on a tree, then one spider—even
a super-slow old codger—could catch the ant by himself, just by mov-
ing steadily toward the ant. With no cycles to dash around, the ant is
doomed.
We take advantage of this idea by using two of our three spiders to
“patrol” one edge each, thereby reducing the edge-graph of the cube
to a tree. To patrol an edge P Q, a spider chases the ant off the edge if
necessary, then patrols the edge making sure he is at all times at most
1
3 the distance from P (respectively, Q) that the ant is. This is possible
since the distance from P to Q along the edges of the cube, if you are
not allowed to use the edge P Q, is three times the length of the edge.
If we choose for our two controlled edges two opposite edges (some
other choices are equally good), we find that with these edges—
including their endpoints—removed, the rest of the edge-network (in
black in the figure below) contains no cycles (so it’s just a collection
of trees). It follows that the third spider can simply chase the ant to
the end of a patrolled edge, where the ant will meet her sad fate.
In a graph, the number of edges ending at a particular node is
called the degree of the node. Since we’re not allowing edges from a
node to itself, or multiple edges between the same pair of nodes, the
degree of a node can never be more than the number of nodes in the
graph, minus one.

Handshakes at a Party
Nicholas and Alexandra went to a reception with ten other couples;
each person there shook hands with everyone he or she didn’t know.

43
 Mathematical Puzzles

Later, Alexandra asked each of the other 21 partygoers how many

people they shook hands with, and got a different answer every time.
How many people did Nicholas shake hands with?

Solution: It seems implausible that Alexandra’s survey somehow

tells us about Nicholas, but that’s what makes this a cute puzzle.
Here our graph has 22 nodes and we put an edge between two if
they shook hands during the party. The maximum possible degree of
a node is only 20, since each person knows his or her partner, and
of course the minimum is 0, so the numbers of handshakes recorded
by Alexandra’s survey must exactly be the 21 numbers 0, 1, . . . , 20.
The guest who shook all 20 hands of the other 10 couples must be the
partner of the one who shook 0 hands (otherwise, the two extremists
would have had to both shake hands and not shake hands). Similarly,
the one who shook 19 hands (all but the 0-guest’s) must be the partner
of the one who shook only 1 hand (only the 20-guest’s), and so on.
This leaves Nicholas as the one who shook 10 hands, since he’s not
partnered with anyone who was surveyed.
As sometimes happens with puzzles, there’s a cheap way to get
this answer, if you trust that the puzzle has a unique solution. You
could have set up the graph so that an edge represents two non-
coupled guests that did not shake hands. Whatever argument might
have enabled you to conclude that Nicholas shook hands with n peo-
ple in the original formulation would also allow you to conclude that
he failed to shake hands with n people, other than Alexandra, in the
second formulation. This is possible only if n = 10.

44
 Graphography

Snake Game
Joan begins by marking any square of an n × n chessboard; Judy then
marks an orthogonally adjacent square. Thereafter, Joan and Judy
continue alternating, each marking a square adjacent to the last one
marked, creating a snake on the board. The first player unable to play
loses.
For which values of n does Joan have a winning strategy, and
when she does, what square does she begin at?

Solution: If n is even, Judy has a simple winning strategy no matter

where Joan starts. She merely imagines a covering of the board by
dominoes, each domino covering two adjacent squares of the board.
Judy then plays in the other half of each domino started by Joan.
When n is odd, and Joan begins in a corner, Joan wins by imag-
ining a domino tiling that covers every square except the corner in
which she starts.
However, Joan loses in the odd-n case if she picks her starting
square badly, for example, if she begins at a square adjacent to a cor-
ner square. Suppose the corner squares are black in a checkerboard
coloring, so that Joan’s starting square is white. There is a domino
tiling of the whole board minus one black square; Judy wins by com-
pleting these dominoes. Joan can never mark the one uncovered black
square because all the squares she marks are white! ♡
There’ll be more about this game at the end of the chapter.

Bracing the Grid

Suppose that you are given an n × n grid of unit-length rods, jointed
at their ends. You may brace some
√ subset S of the small squares with
diagonal segments (of length 2).
Which choices of S suffice to make the grid rigid in the plane?

Solution: It helps to turn this into a graph-theoretical problem, but

not in the most obvious way (with vertices as joints and edges as rods).
Instead, suppose you have put in your braces; now envision a graph
G whose vertices correspond to the rows of squares and the columns
of squares. Each edge of G corresponds to a row and column whose
intersecting square is braced, so that the number of edges of G is the
same as the number of braces you used.

45
 Mathematical Puzzles

A partly-braced grid, with its associated graph, is pictured below.

If a row and a column are adjacent in G, the vertical rods in the

row are all forced to be perpendicular to the horizontal rods in the
column. If G is a connected graph, meaning that there is a path from
any vertex to any other, then all the horizontal rods must be perpen-
dicular to all the vertical ones. Thus all the horizontal rods are parallel
to one another, and similarly for the vertical ones, and now it is clear
that the grid is rigid.
On the other hand, suppose the graph is disconnected and let C
be a “component,” that is, a connected piece of G which has no edges
to the rest of G. Then there is nothing to prevent any vertical rod in
a C-row or any horizontal rod in a C-column from flexing relative to
the other rods in the grid.
Thus, the criterion for rigidity is exactly that the graph G be con-
nected. Since G has 2n vertices, it must have at least 2n−1 edges in
order to be connected (you can prove this easily by induction if you
haven’t seen it before), hence you must have at least 2n−1 braces to

46
 Graphography

make the grid rigid. Notice, however, that they cannot be “just any-
where.”
The next figure shows an efficiently braced 3 × 3 grid and its cor-
responding graph. For practice, you might want to compute the total
number of ways to rigidify the 3 × 3 grid with the minimum number
(five) of braces. A theorem of graph theory (to the effect that every
connected graph has a connected subgraph, called a “spanning tree,”
with the minimum number of edges) tells us that if more than 2n−1
squares are braced and the grid is rigid, then there are ways to remove
all but 2n−1 of the braces and preserve rigidity.

Competing for Programmers

Two Silicon Valley startups are competing to hire programmers. Hir-
ing on alternate days, each can begin by hiring anyone, but every sub-
sequent hire by a company must be a friend of some previous hire—
unless no such person exists, in which case that company can again
hire anyone.
In the candidate pool are 10 geniuses; naturally each company
would like to get as many of those as possible. Can it be that the com-
pany that hires first cannot prevent its rival from hiring nine of the
geniuses?

Solution: Yes. Imagine a friendship graph that contains a (non-

genius) hub Hubert with 10 friends each of whom knows a differ-
ent genius, plus one lonely friend (Lonnie) who knows only Hubert.
That’s it; the geniuses (black vertices on the periphery) each have only
one friend.
Suppose Company A goes first and hires Hubert. Then B hires
Lonnie and after that, every time A hires a friend of Hubert, B hires
that person’s genius friend.

47
 Mathematical Puzzles

If A hires Lonnie, B hires a genius. A must hire Hubert and then,

after A hires its genius’ friend, the game proceeds as before.
If A hires a genius, B hires another genius; three days later A hires
Hubert and then B hires Lonnie, after which the remaining geniuses
go to B.
Finally, if A hires a genius’ friend, B hires that genius. After A
takes Hubert and B takes Lonnie, A winds up with no genius at all.

Wires under the Hudson

Fifty identical wires run through a tunnel under the Hudson River,
but they all look the same, and you need to determine which pairs
of wire-ends belong to the same wire. To do this you can tie pairs of
wires together at one end of the tunnel and test pairs of wire-ends at
the other end to see if they close a circuit; in other words, you can
determine whether two wires are tied together at the other end.
How many trips across the Hudson do you need to accomplish
your task?
Solution: Suppose the wires are labeled w1 , w2 , . . . , w50 at the west
end of the tunnel and e1 , . . . , e50 at the east end. Supposing that you
start on the west side of the river, you tie w1 and w2 together, w3 and
w4 , w5 and w6 , etc. until all except w49 and w50 are paired.
You then test pairs of wire-ends at the east end of the tunnel until
you have identified all of the tied pairs. For example, you might find
that (say) e4 and e29 are tied, e2 and e15 are tied, e8 and e31 are tied,
and so forth, and finally that e12 and e40 are bachelors.

48
 Graphography

Next you return to the west end, untie all pairs, and instead tie w2
to w3 , w4 to w5 , etc. until all but w1 and w50 are paired.
Finally you test pairs at the east end until, as before, you have
identified all the paired wire-ends. To continue the example, the new
pairs might include e12 and e15 , e29 and e2 , and e4 and e31 , with e40
and e8 as bachelors.
This simple procedure suffices to identify all the wires.
Observe that the one east-wire-end which was paired the first time
but not the second (in our example, this is e8 ) must belong to w1 . The
east-wire-end with which e8 was paired the first time (here, e31 ) must
therefore belong to w2 . But then, w3 must belong to the east-wire-
end to which e31 was paired the second time, namely e4 . Proceeding
in this fashion, you find that w4 belongs to e29 (e4 ’s mate on the first
round), w5 belongs to e2 (e29 ’s mate on the second round), and so
forth. Eventually the sequence will end with w50 belonging to e40 .
If the number of wires (say, n) had been odd, you’d have left only
wn out of the pairings the first time, and w1 the second time; the rest
works pretty much the same way.

Two Monks on a Mountain

Remember the monk who climbed Mt. Fuji on Monday and de-
scended on Tuesday? This time, he and a fellow monk climb the
mountain on the same day, starting at the same time and altitude,
but on different paths. The paths go up and down on the way to the
summit (but never dip below the starting altitude); you are asked to
prove that they can vary their speeds (sometimes going backwards) so
that at every moment of the day they are at the same altitude!

Solution: It is convenient to divide each path into a finite sequence

of monotone “segments” within which the path is always ascending,
or always descending. (Level segments cause no problem as we can
have one monk pause while the other traverses such a segment). Then
we can assume that each such segment is just a straight ascent or de-
scent, since we can have the monks modulate their rates so that their
rate of altitude change is constant on any segment.

49
 Mathematical Puzzles

(There are mathematical curves of finite length that cannot be di-

vided into a finite number of monotone segments, but in our problem
we needn’t worry about segments that are shorter than a monk’s step
size.)
Label the X-axis on the plane by positions along the first monk’s
path, and the Y -axis by positions along the second monk’s path. Plot
all points where the two positions happen to be at the same altitude;
this will include the origin (where both paths begin) and the summit
(where they end, say at (1,1)). Our objective is to find a path along
plotted points from (0,0) to (1,1); the monks can then follow this path,
slowly enough to make sure that no monk is anywhere asked to move
faster than he can.
Any two monotone segments—one from each path—which have
some common altitude show up on the plot as a (closed) line segment,
possibly of zero length. If we regard as a vertex any point on the plot
which maps back to a segment endpoint (for either or both monks),
the plot becomes a graph (in the combinatorial sense); and an easy
checking of cases shows that except for the vertices at (0,0) and (1,1),
all the vertices are incident to either 0, 2, or 4 edges.
Once we begin a walk on the graph at (0,0), there is no place to
get stuck or be forced to retrace but at (1,1). Hence we can get to (1,1),
and any such route defines a successful strategy for the monks. ♡

The figure shows four possible landscapes, with one monk’s path
shown as a solid line, the other as dashed. Below each landscape is the
corresponding graph. Note that, as in the last case, there may be de-
tached portions of the graph which the monks cannot access (without
breaking the common altitude rule).

50
 Graphography

Worst Route
A postman has deliveries to make on a long street, to addresses 2,
3, 5, 7, 11, 13, 17, and 19. The distance between any two houses is
proportional to the difference of their addresses.
To minimize the distance traveled, the postman would of course
make his deliveries in increasing (or decreasing) order of address. But
our postman is overweight and would like to maximize the distance
traveled making these deliveries, so as to get the most exercise he can.
But he can’t just wander around town; to do his job properly he is
obligated to walk directly from each delivery to the next one.
In what order should he make his deliveries?
Solution: The “greedy algorithm” for maximizing the postman’s
distance would have him start at #2, go all the way to #19, then back
to #3, then to #17 and so forth, ending at #11. That would result in
total distance 17 + 16 + 14 + 12 + 8 + 6 + 4 = 77 units. Is this the best
he can do?
Here’s a useful way to look at the problem. Call the addresses A,
B, etc. up to H, in numerical order, and represent the trek from house
X to house Y by XY . No matter what the postman does, his trip can
be represented as a list of treks if the form AB, BC, CD, DE, EF ,
F G, and GH.
How many times can the postman perform the trek AB? At most
twice, once on his way to A and once back. A similar argument works
for GH. How about BC? That one he might be able to do four times:
to A and back (but not from B), and to B and back (not from A).
Continuing this reasoning, we see that the postman cannot possibly
do better than 2AB + 4BC + 6CD + 8DE + 6EF + 4F G + 2GH,
which adds up to 86 units in our puzzle.
Wait—the postman can’t actually do the EF trek eight times be-
cause he only makes seven trips between consecutive deliveries. So in
fact he can do no better than 2AB + 4BC + 6CD + 7DE + 6EF +
4F G + 2GH, which is 82 units in our puzzle. Can he achieve that
number?
Yes, he can, but only if he begins at D or E and ends at the other
one, after zigzagging between low-numbered houses (A, B, C, D) and
high (E, F, G, H). Example: D, H, C, G, B, F, A, E. There are 8×4×
3 × 3 × 2 × 2 × 1 × 1 = 1,152 ways to do this, and each one gives the
maximum possible length, 82 units.
The given addresses don’t matter a bit, nor even the number of
houses (provided that remains even). The reasoning above is quite

51
 Mathematical Puzzles

general, the conclusion being that the postman should always (a) be-
gin at one of the middle two houses and end at the other, and (b)
zigzag between the lower-half addresses and the upper-half addresses.

For our theorem, let’s take another look at the Snake Game puzzle.
We can play that game on any graph G: The rules are that the players
alternate marking vertices, and after Player 1 (Joan) marks her first
vertex, every subsequent vertex must be unmarked and adjacent to
the vertex just marked by the opponent. The first player unable to
move loses.
The proof above that the second player wins on an 8 × 8 chess-
board actually shows that Player 2 wins on any graph with a perfect
matching, that is, a set of disjoint edges that cover all the vertices (as
the dominoes do on the chessboard).
In fact, when G has a perfect matching, Player 2 wins even when
Player 1 is allowed to flout the main rule of the game, marking any
(unmarked) vertex instead of one adjacent to Player 1’s last move!
The strategy is the same: Whenever Player 1 marks a vertex v, Player
2 simply marks the vertex on the other end of the matching edge that
covers v.
So you might think that having a perfect matching is a far stronger
condition than is needed to assure a win for Player 2 in the original
game. But no.
Theorem. With best play, Player 2 wins on G if and only if G has a perfect
matching.
Proof. We’ve already proved the “if ” part of the theorem, so let us
suppose that G does not have a perfect matching, with the intent of
providing a winning strategy for Player 1.
The key is to pick some maximum-size matching M , that is, some
collection of disjoint edges which is as large as possible. Since there
is no perfect matching, there will be some vertex u1 of G that’s not
in M , and we’ll let Player 1 start there. After that, Player 1 intends to
follow the “matching strategy” from above, playing the other end of
whatever matching edge Player 2 marks a vertex of. If Player 1 can
always do that, she of course will win, but how do we know she can
always do that? What if Player 2 marks a vertex that’s not in an edge
of M ?
Well, Player 2 can’t immediately do that, because her first play (say,
v1 ) must be adjacent to v0 . If v1 were not in the matching, the edge
(u1 , v1 ) could have been added to M to make a bigger matching—and
M was supposed to be as big as possible.

52
 Graphography

And she can’t do it later, either. Suppose Player 1 has been able
to execute her matching strategy through her kth move, and now
must reply to Player 2’s kth move. The moves so far have been
u1 , v1 , u2 , v2 , . . . , uk , vk ; by assumption, all these vertices are distinct,
all the pairs (ui , vi ) are edges of G, and all the pairs (vi , ui+1 ) are not
only edges of G but also members of the matching M .
But now we claim that Player 2’s latest move, vk , belongs to an
edge of M . Why? Because if it didn’t, we could make M bigger
by replacing the k−1 edges (v1 , u2 ), (v2 , u3 ), . . . , (vk−1 , uk ) by the
k edges (u1 , v1 ), (u2 , v2 ), . . . , (uk , vk )—again contradicting the max-
imality of M .
So Player 1 is never stuck for a move, and the theorem is proved!
♡

53
 5. Algebra Too
The single most useful technique for solving puzzles that ask for a
quantity is simply to assign a letter (x and n are popular choices) to
the quantity, write down what the puzzle tells you, and then solve
using algebra.
Our first example is simple, but people get it wrong all the time.

Bat and Ball

A bat costs $1 more than a ball; together they cost $1.10. How much
does the bat cost?

Solution: Did you say $1? Oops! Let x be the price of the bat, so the
price of the ball is x − $1. Then we have the equation x + (x − $1) =
$1.10, 2x = $2.10, x = $1.05.
Note that we could have assigned another variable (y, perhaps) to
the price of the ball, then solved two equations in two unknowns. If
you can avoid that by expressing new unknowns in terms of old ones,
you can save yourself a lot of trouble.

Two Runners
Two runners start together on a circular track, running at different
constant speeds. If they head in opposite directions they meet after a
minute; if they head in the same direction, they meet after an hour.
What is the ratio of their speeds?

Solution: This is simultaneously confusing and easy. Let L be the

length of the track. Then, when going in opposite direcions, the run-
ners meet when, together, they have run a distance L. When going
the same direction, they meet when the difference of their achieved
distances is L.
Thus, (s + t)L/(s − t)L = 60, and solving gives s/t = 61/59.

55
 Mathematical Puzzles

A classic:

Belt Around the Earth

Suppose the earth is a perfect sphere and a belt is tightened around
the equator. Then the belt is loosened by adding 1 meter to its length,
and it now sits the same amount off the ground all the way around.
Is there enough room to slip a credit card underneath the belt?

Solution: The circumference of the earth is about 40 million me-

ters, but who cares? Call it C, and call the earth’s radius R, so that
C = 2πR, thus R = C/(2π). When we lengthen the belt its circum-
ference is C+1 so the new radius is (C+1)/(2π) = C/(2π)+1/(2π) =
R+16 cm, approximately. There’s room for a stack of about 200 credit
cards to pass under the belt.
There are dozens of published puzzles along the lines of “John is
twice as old as Mary will be when . . . ” that can be solved straightfor-
wardly as above. Let’s try something a little different.

Odd Run of Heads

On average, how many flips of a coin does it take to get a run of an
odd number of heads, preceded and followed by tails?

Solution: First, we need to get a tail, since starting with odd run
of heads won’t help us. If that takes time x on average, then, since
flipping H doesn’t help, we have
1
x=1+ ·x
2
giving x = 2. Now we need an odd run of heads; say that takes time y
on average. If we flip another T (probability 12 ) or HH (probability 14 ),
we’ve made no further progress; if we flip HT (probability 14 ) we’re
done. Thus,
1 1 1
y= · (1 + y) + · (2 + y) + · 2 ,
2 4 4
giving y = 6. So in all it takes an average of x + y = 2 + 6 = 8 flips
to get that run.
Randomness enters into the next puzzle as well. This is a toughie
but the general technique is the same.

56
 Algebra Too

Hopping and Skipping

A frog hops down a long line of lily pads; at each pad, he flips a coin
to decide whether to hop two pads forward or one pad back. What
fraction of the pads does he hit?
Solution: Let us label the lily pads with integers in order. One way
to begin solving the problem is to compute the probability p that a frog
starting at 1 regresses, at some point, to 0. To never regress, he must
jump ahead (probability 12 ) to lily pad #3 and not subsequently regress
three times from there (probability (1 − p3 ). Thus 1 − p = 12 (1 − p3 );
√
divide by (1 − p)/2 to get 2 = 1 + p + p2 , giving p = ( 5 − 1)/2 ∼
.618034, the familiar golden ratio.
It seems awkward to try to compute the probability that the frog
skips over a particular position (say, 1). You might want to start your
calculations from the first time the frog hits 0, if it does, but it might
already have hit 1. A better idea is to try to compute the probability
that during a particular jump the frog leaps over a pad that he never has,
and never will, hit.

For that to happen, he must (a) be jumping forward at the mo-

ment, (b) never drop back from where he lands, and (c) not have
reached the pad he is jumping over in the past.
We may as well assume the lily pad the frog is currently passing
over is #0. The key is to note that event (c) is an independent copy of
event (a) if you reverse both space and time. Looking at the frog back-
wards in time, and regarding lower-numbered lily pads as progress,
the frog appears to be behaving as before: jumping two forward or
one back with equal probability. Event (c) says that when he reaches
−1, he must not later “drop back” to 0.
The probability of these three events all occurring is thus 12 ·(1−p)·
(1−p) = (1−p)2 /2. We must not forget, however, that we have not
calculated the probability that 0 is skipped; only the probability that
a particular jump carries the frog over a skipped pad.

57
 Mathematical Puzzles

Since the frog travels, on the average, at rate 12 , he is creating

skipped lily pads at rate (1 − p)2 relative to his spatial progress.
√ It fol-
lows that the fraction of pads he hits is 1 − (1 − p)2 = (3 5 − 5)/2 ∼
0.854102. Whew! ♡
The next puzzle is an almost straightforward application of the al-
gebraic technique, but a little extra cleverness is needed—as is often
the case when the resulting equation or equations are Diophantine,
that is, they require integer solutions.

Area−perimeter Match
Find all integer-sided rectangles with equal area and perimeter.
Solution: There are two.
Let x and y be the sides, with x ≥ y. we need xy = 2x + 2y,
equivalently xy − 2x − 2y = 0, but the left-hand side of the latter
reminds us of the product (x−2)(y−2). Rewriting, (x−2)(y−2) = 4,
so either x−2 = y−2 = 2 or x−2 = 4 and y−2 = 1. This gives
x = y = 4, or x = 6 and y = 3.
The remaining puzzles involve algebra in other ways than solving
equations.

Three Negatives
A set of 1000 integers has the property that every member of the set
exceeds the sum of the rest. Show that the set includes at least three
negative numbers.

Solution: Let a and b be the only two negative numbers in the set,
and let the sum of all the numbers in the set be S. We are told that
a > S − a, and b > S − b; adding the two inequalities gives a + b >
2S − a − b, thus a + b > S which is impossible unless there are other
negative numbers in the set.

Red and Black Sides

Bored in math class, George draws a square on a piece of paper and
divides it into rectangles (with sides parallel to a side of the square),
using a red pen after his black one runs out of ink. When he’s done,
he notices that every rectangular tile has at least one totally red side.

58
 Algebra Too

Prove that the total length of George’s red lines is at least the length
of a side of the big square.
Solution: Why should this be true? Maybe each rectangle is long
and thin and its only red side is a short one. One thing we do know,
though, is that the total area of the rectangles is s2 , where s is the side
of the big square. If the rectangles are ri × b∑i , where each ri represents
a red side, then (using the convention that i xi means the sum of xi
for all salient i): ∑
ri bi = s2 .
i

But of course each bi ≤ s so

∑ ∑ ∑
ri bi ≤ ri s = s ri
i i i
∑
and it follows that i ri ≥ s. In fact, you get equality only if all the
rectangles span the square east–west (or, all north–south) with one
end of each rectangle colored red.

Recovering the Polynomial

The Oracle at Delphi has in mind a certain polynomial p (in the vari-
able x, say) with non-negative integer coefficients. You may query
the Oracle with any integer x, and the Oracle will tell you the value
of p(x).
How many queries do you have to make to determine p?

Solution: It only takes two queries, and the first is needed only to
get a bound on the size of the polynomial’s coefficients. The Oracle’s
answer to x = 1 (say, n) tells you that no coefficient can exceed n.
Then you can send in x = n+1, and when you expand the Oracle’s
answer base n+1, you have the polynomial!
For example, suppose (rather conveniently) that p(1) < 10, so you
know every coefficient is an integer between 0 and 9. You then send
in the number 10 and if the oracle says p(10) = 3,867,709,884, you
know that the polynomial is

3x9 + 8x8 + 6x7 + 7x6 + 7x5 + 9x3 + 8x2 + 8x + 4.

Here’s a puzzle whose solution is, potentially, a useful strategy.

59
 Mathematical Puzzles

Gaming the Quilt

You come along just as a church lottery is closing; its prize is a quilt
worth $100 to you, but they’ve only sold 25 tickets. At $1 a ticket,
how many tickets should you buy?
Solution: Clearly you should buy at least 1 ticket, assuming you
are trying to maximize your expected gain, because then for $1 you
have a 1/26 chance to win $100; you will gain on average almost $3
from the invesment.
On the other hand, it would be a losing proposition (except to the
church!) to buy 100 or more tickets, because then you could break
even at best.
So it looks like some intermediate number of tickets is right, but
how many? You can try to maximize your expectation (x/(25 + x)) ·
$100 − $x from buying x tickets using calculus, if you happen to know
your calculus, but what do you do in the (likely) event that the opti-
mum value of x is not an integer?
Often an effective way to tackle this kind of problem is to figure
out whether you gain or lose when you increment your variable by
1. If you increase your purchase from m tickets to m+1 tickets, your
expected return goes up by $100((m + 1)/(m + 26) − m/(m + 25)) −
1. This value goes steadily down and eventually becomes negative
(meaning that you shouldn’t buy another ticket). Thus, if you can
identify the least m for which this expression is negative, you’ll know
that the right number of tickets to buy is m.
That’s just algebra: You solve the (quadratic) equation $100((m +
1)/(m + 26) − m/(m + 25)) − 1 = 0 to get m = 24.5025. Thus 25 is
the least value of m with the property that you lose money by buying
another ticket, and it follows that 25 is the best number of tickets to
buy. Is it a coincidence that 25 was also the number of tickets sold
before you got there? Actually, it is.
And if you don’t care for solving quadratic equations? You can
still get the answer pretty quickly by plugging your guesses for m into
the expression $100((m + 1)/(m + 26) − m/(m + 25)) − 1, until you
zero in on the least m that makes the expression negative.

Strength of Schedule
The 10 teams comprising the “Big 12” college football conference are
all scheduled to play one another in the upcoming season, after which
one will be declared conference champion. There are no ties, and each

60
 Algebra Too

team scores a point for every team it beats.

Suppose that, worried about breaking ties, a member of the gov-
erning board of the conference suggests that each team receive in addi-
tion “strength of schedule points,” calculated as the sum of the scores
achieved by the teams it beat.
A second member asks: “What if all the teams end up with the
same number of strength-of-schedule points?”
What, indeed? Can that even happen?
Solution: No.
First, we observe that if all the teams end up with the same number
of strength-of-schedule points, they must also all have the same score.
Why? If not, let b be the biggest score and s the smallest. Each team
that won b games racked up at least bs strength-of-schedule points
(SSP’s, for short), because, at worst, all their wins came against teams
that won only s games. Similarly, teams that won only s games got at
most sb SSP’s. Whoops, that means both the b teams and the s teams
got exactly sb SSP’s, and therefore that all of the b teams’ wins came
against s teams and vice-versa. But that’s impossible unless there is
only one b team and one s team, because otherwise there would be a
game between two b teams or between two s teams.
It follows that the b team won only once. But then, any third team
would have to have beaten both the b team and the s-team, outscoring
the b-team and contradicting the definition of b.
We conclude (since there are more than two teams in the Big 12)
that if every team has the same number of SSP’s, then every team
has the same score. That’s not possible in a conference with an even
number of teams, because then each team would have to have won
half its games, and the number of games it plays is odd.
Round-robin tournaments are excellent puzzle fodder; here’s an-
other.

Two Round-robins
The Games Club’s 20 members played a round-robin checkers toura-
ment on Monday and a round-robin chess tournament on Tuesday.
In each tournament, a player scored 1 point for each other player he
or she beat, and half a point for each tie.
Suppose every player’s scores in the two tournaments differed by
at least 10. Show that in fact, the differences were all exactly 10.

61
 Mathematical Puzzles

Solution: Let’s divide the club members into the good checkers
players (who scored higher in the checkers tournament) and the good
chess players. One of these groups must have at least 10 members;
suppose there are k ≥ 10 good chess players, and that they scored
a total of t ≥ 10k points higher in the chess tournament than in the
checkers tournament.
This difference must have come entirely at the expense of the
good checkers players, because the number of points scored by the
good chess players among ( ) themselves had to be the same in both
tournaments—namely, k2 , A.K.A. k(k−1)/2. Since there are k(20−
k) intergroup matches in each tournament, the most they can con-
tribute to t is k(20 − k). So we have 10k ≤ k(20 − k), k ≤ 10.
But k was supposed to be at least 10, so we conclude that k = 10
and that all of the above inequalities must be equalities.
In particular, since t = 10k every chess player scored exactly 10
points higher in the chess tournament than in the checkers tourna-
ment. Moreover, since t = k(20 − k), every intergroup match in the
chess tournament must have been won by the good chess player, and
in the checkers tournament, by the good checkers player.

Alternative Dice
Can you design two different dice so that their sums behave just like
a pair of ordinary dice? That is, there must be two ways to roll a 3,
6 ways to roll a 7, one way to roll a 12, and so forth. Each die must
have six sides, and each side must be labeled with a positive integer.

Solution: The unique answer is that the labelings of the two dice
are {1, 3, 4, 5, 6, 8} and {1, 2, 2, 3, 3, 4}.

62
 Algebra Too

Perhaps you came up with this answer by trial and error, which is
a perfectly fine way of solving the problem. However, there is another
way in this case, involving a simple example of a powerful mathemat-
ical tool called generating functions.
The idea is to represent a die by a polynomial in the variable x, in
which the coefficient of the term xk represents the number of appear-
ances of the number k on the die. Thus, for example, an ordinary die
would be represented by the polynomial f (x) = x + x2 + x3 + x4 +
x5 + x 6 .
The key observation is that the result of rolling two (or more) dice
is represented by the product of their polynomials. For instance, if we
roll two ordinary dice, the coefficient of x10 in the product (namely,
f (x)2 ) is exactly the number of ways of picking two terms from f (x)
whose product is x10 . These ways are x4 · x6 , x5 · x5 , and x6 · x4 ; and
these represent the three ways to roll a sum of 10.
It follows that if g(x) and h(x) are the polynomials representing
our alternative dice, then g(x) · h(x) = f (x)2 . Now, polynomials,
like numbers, have unique prime factorizations; the polynomial f (x)
factors as x(x + 1)(x2 + x + 1)(x2 − x + 1). To get g(x) and h(x) to
multiply out to f (x)2 , we need to take each of these four factors and
assign either one copy to each of g(x) and h(x), or two copies to one
and none to the other. But there are some constraints: In particular,
we can’t have a non-zero constant term in g(x) or h(x) (which would
represent some sides labeled “0”) or any negative coefficient, and the
coefficients must sum to 6 since we have six sides to label.
The only way to do this (other than g(x) = h(x) = f (x)) is to
have
g(x) = x(x + 1)(x2 + x + 1)(x2 + x − 1)2 = x + x3 + x4 + x5 + x6 + x8
and
h(x) = x(x + 1)(x2 + x + 1) = x + 2x2 + 2x3 + x4 ,
or vice-versa.
This may seem a bit like trial and error after all, but with this tech-
nique you can solve problems that are much more complex than this
one. To begin with, you can invent alternatives for a pair of 8-sided
dice labeled 1 through 8 (there are three new ways to do it), or for
rolling three ordinary dice (many ways).
It will not be a surprise to readers that thousands, literally, of the-
orems have been proved using algebra. Here’s one that uses just the
elementary sort of algebra we have seen above.

63
 Mathematical Puzzles

Have you ever wondered: What is the probability that your fam-
ily surname will die out? Francis Galton posed this question in 1873
and an answer posted by the Reverend Henry William Watson led
to a joint paper. Galton was interested in the longevity of aristo-
cratic names, but he could (in more modern times) have been thinking
equivalently of the preservation of a Y-chromosome.
In Galton and Watson’s model of propagation, each individual
independently has a random number of offspring, given by a fixed
probability distribution: i offspring with probability pi . (In the case of
tracking patrilineal surnames, we would count only male offspring.)
We wish to know: Starting with (say) one individual, what is the
“extinction” probability x that the family name dies out? Of course
this depends on the probability distribution; as you might expect, if
the average number of offspring
∞
∑
µ= ipi
i=0

is less than 1, then the name will always die out, but if it’s greater
than 1, the name might live on forever. In the latter case, how do you
determine x?

Theorem. Let T be a Galton–Watson tree with offspring ∑ probabilities

p0 , p1 , . . . , pk . Then the extinction probability x satisfies x = p i xi .

Let’s do one example before we proceed to the proof. Suppose

each individual has zero offspring with probability 0.1, one with prob-
ability 0.6, and two with probability 0.3. Then µ = 1.2 and we ex-
pect the tree will sometimes be finite (i.e., die out) and sometimes
infinite (continue forever). The equation in the theorem becomes
x = 0.1 + 0.6x + 0.3x2 , which reduces to 3x2 − 4x + 1 = 0; fac-
toring gives (3x − 1)(x − 1) = 0, so x = 1 or x = 13 . Since the
probability of extinction is not 1, it must be 13 .
Proof. We observe that the tree is finite if either the starting indi-
vidual has no offspring, or he does have offspring but each child is
himself the start of a finite tree. If x is the extinction probability, then
the probability that each of i individuals leads only to a finite tree is
xi , thus:
∑ k
x= pi xi
i=0

as claimed. ♡

64
 Algebra Too

The equation in the theorem always has x = 1 as one solution,

since the probabilities sum to one. If µ > 1, there will be one other
solution—the correct one. If µ ≤ 1, except in the case where p1 = 1
and pi = 0 for all i = ̸ 1, the only solution will be x = 1 and thus
extinction is assured.

65
 6. Safety in Numbers
Numbers are everywhere, and endlessly fascinating. You probably
know that every positive integer can be expressed uniquely as the
product of primes (a prime number being one that cannot be divided
evenly by anything except itself and 1). Let’s use that and some other
elementary observations to solve some puzzles, shall we?

Broken ATM
George owns only $500, all in a bank account. He needs cash badly
but his only option is a broken ATM that can process only with-
drawals of exactly $300, and deposits of exactly $198. How much cash
can George get out of his account?

Solution: Since both the deposit and withdrawal amounts are

multiples of $6, George cannot hope to get more than $498 (the
largest multiple of 6 under $500). Using “−300” to indicate a with-
drawal and “+198” a deposit, we see that George can get $6 out by
−300, +198, −300, +198, +198. Doing this 16 times gets $96 out, after
which −300, +198, −300 finishes the job.

Subsets with Constraints

What is the maximum number of numbers you can have between 1
and 30, such that no two have a product which is a perfect square?
How about if (instead) no number divides another evenly? Or if no
two have a factor (other than 1) in common?

Solution: These puzzles can all be tackled the same way. The idea
is to try to cover the numbers from 1 to 30 by bunches of numbers with
the property that from each bunch, you can only take one number for
inclusion in your subset. Then, if you can construct a subset that con-
sists of one choice from each bunch, you have yourself a maximum-
sized subset.

67
 Mathematical Puzzles

In the first case, fix any number k which is square-free (in other
words, its prime factorization contains at most one copy of each
prime). Now look at the set Sk you get by multiplying k by all possible
perfect squares.
If you take two numbers, say kx2 and ky 2 , from Sk , then their
product is k 2 x2 y 2 = (kxy)2 so they can’t both be in our subset. On
the other hand, two numbers from different Sk ’s cannot have a square
product since one of the two k’s will have a prime factor not found in
the other, and that factor will appear an odd number of times in the
product.
Now, every number is in just one of these sets Sk —given n, you can
recover the k for which n ∈ Sk by multiplying together one copy of
each prime that appears with an odd exponent in the factorization of
n. Between 1 and 30, the choices for k (i.e., the square-free numbers)
are 1, 2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 2×3, 2×5, 2×7, 2×9, 2×11,
2×13, 3×5, 3×7, and finally 2×3×5: 20 in all. You can choose each k
itself as the representative from Sk , so a subset of size 20 is achievable
and best possible.
To avoid one number dividing another evenly, note that if you fix
an odd number j, then in the bunch Bj = {j, 2j, 4j, 8j, . . . }—that is,
j times the powers of 2—you can only take one. If you take for your
subset the top half of the numbers from 1 to 30, namely 16 through
30, you have got one from each Bj , and of course no member of this
subset divides another evenly since their ratios are all less than 2. So
the 15 numbers you get this way is best possible.
Finally, to get a maximum-sized subset all of whose members are
relatively prime, you naturally want to look at bunches consisting of
all multiples of a fixed prime p. You can take p itself as the repre-
sentative of its bunch, so you can do no better than to take as your
subset all the primes below 30, plus the number 1 itself, for a total of
11 members. ♡

Cards from Their Sum

The magician riffle-shuffled the cards and dealt five of them face
down onto your outstretched hand. She asked you to pick any number
of them secretly, then tell you their sum (jacks count as 11, queens 12,
kings 13). You did that and the magician then told you exactly what
cards you chose, including the suits!
After a while you figured out that her shuffles were defective—
they didn’t change the top five cards, so the magician had picked them

68
 Safety in Numbers

in advance and knew what they were. But she still had to pick the
ranks of those cards carefully (and memorize their suits), so that from
the sum she could always recover the ranks.
In conclusion, if you want to do the trick yourself, you need to find
five distinct numbers between 1 and 13 (inclusive), with the property
that every subset has a different sum. Can you do it?

Solution: If only you had a card worth 16, you’d be in fine shape:
The powers of two, in particular 1, 2, 4, 8, and 16, have the desired
property. In fact when you write a number in binary, the positions of
the ones tell you exactly which powers go into the sum. For example,
suppose that your victim reports a sum of 13; in binary that’s 1101,
thus 8 + 4 + 1, so his cards must consist of an eight, a four, and an
ace.
There are 25 = 32 ways to choose a subset of the five cards (if
you include the null set, which has sum 0). Can we get them all to be
different, without having a number bigger than 13?
The first thing to realize is that big numbers are better: It’s easier
to tell whether a big-numbered card was chosen or not. So let’s try
making our set using the greedy algorithm: Start with the king and
throw in the biggest cards possible without causing the same sum to
come up twice.
If you do that you’ll end up with K, Q, J, 9, 6; and that works! (It
turns out there’s only one other way to do it, the same cards but with
the 9 replaced by a 3).
What about suits? I recommend 6 of spades, 9 of diamonds, jack
of clubs, queen of hearts and king of spades. But pick anything that
looks random and that you can remember. Then practice (including
that fake riffle shuffle), and you’ll be the life of the party. Read Colm
Mulcahy’s book (see Notes & Sources) for more great mathematical
card tricks!

69
 Mathematical Puzzles

Divisibility Game
Alice chooses a whole number bigger than 100 and writes it down
secretly. Bob now guesses a number greater than 1, say k; if k di-
vides Alice’s number, Bob wins. Otherwise k is subtracted from Al-
ice’s number and Bob tries again, but may not use a number he used
before. This continues until either Bob succeeds by finding a num-
ber that divides Alice’s, in which case Bob wins, or Alice’s number
becomes 0 or negative, in which case Alice wins.
Does Bob have a winning strategy for this game?
Solution: The key is that there are enough small numbers around
so that by judicious use of them, Bob can find a divisor of Alice’s num-
ber. For example, Bob wins using 2, 3, 4, 6, 16, 12 (or 6, 4, 3, 2, 5, 12 or
others). To see this, we consider the possibilities for Alice’s number
modulo 12; in other words, the remainder when Alice’s number (let’s
call it A) is divided by 12.
Using the suggested sequence 2, 3, 4, 6, 16, 12, Alice is immedi-
ately lost if her number is even, that is, if A ≡ 0, 2, 4, 6, 8, or 10 mod-
ulo 12. If A ≡ 5 or 11 mod 12, it will be 3 or 9 mod 12 after 2 is
subtracted, allowing Bob to win on his second turn. If A ≡ 1 or 9,
Alice will succumb at turn 3; if A ≡ 3, at turn 4. The only value mod
12 that Alice can start with that has survived so far is 1, which after
2, 3, 4, and 6 have been subtracted, is now 4 modulo 12. Subtracting
16 knocks that down to 0 modulo 12, and now Bob’s 12 spells doom
for Alice. Of course, we must check that Bob’s guesses add up to at
most 100, and they do (2 + 3 + 4 + 6 + 16 + 12 = 43).

Prime Test
Does 49 + 610 + 320 happen to be a prime number?
Solution: Since the number is (29 )2 + 2 · 29 · 310 + (310 )2 = (29 +
310 )2 , it is a perfect square (so, it’s not prime). If you guessed correctly
that the number was composite but tried to prove it by finding a small
prime factor, you were in for a tough time, because 29 + 310 itself is
prime—so our number has no prime factor less than 59,561.
Yes, it is easy to forget that finding a divisor is not the only way to
show a number is composite. Showing that it is a square (or a cube,
etc.) is another; other, more sophisticated ways exist as well.

70
 Safety in Numbers

Numbers on Foreheads
Each of 10 prisoners will have a digit between 0 and 9 painted on his
forehead (they could be all 2’s, for example). At the appointed time
each will be exposed to all the others, then taken aside and asked to
guess his own digit.
In order to avoid mass execution, at least one prisoner must guess
correctly. The prisoners have an opportunity to conspire beforehand;
find a scheme by means of which they can ensure success.

Solution: It is useful in many problems to introduce probability,

even though none is present in the statement. Here, if we assume that
the numbers painted on foreheads are chosen independently and uni-
formly at random, we see that no matter what he does, each prisoner
has probability 1/10 of guessing correctly.
Let the prisoners be numbered from 0 to 9. Since we want the
probability that some prisoner guesses correctly to be 1, we need the
10 events “Prisoner k guesses correctly” to be mutually exclusive: In
other words, no two can occur. Otherwise, the probability of at least
1
one success would be strictly less than 10( 10 ) = 1.
To do this it would behoove us to separate the set of possible con-
figurations into 10 equally likely scenarios, then have each prisoner
base his guess on a different scenario. This reasoning may already
have led you to the easiest solution: Let s be the sum of the numbers
on all the prisoners’ foreheads, modulo 10 (in other words, the right-
most digit of the sum). Now let Prisoner k guess that s = k, in other
words, guess that his own number is k minus the sum of the numbers
he sees, modulo 10.
This will ensure that Prisoner s, whoever that may be, will be cor-
rect (and all others wrong).

Rectangles Tall and Wide

It is easy to check that a 6 × 5 rectangle can be tiled with 3 × 2 rect-
angles, but only if some tiles are horizontally oriented and others are
vertically oriented.
Show that this can’t happen if the big rectangle is a square—in
other words, if a square is tiled with congruent rectangles, then it can
be re-tiled with the same rectangles all oriented the same way.

Solution: First, we claim that if the sides of a tile are not commen-
surable (i.e., if their ratio is irrational) then a much stronger state-

71
 Mathematical Puzzles

ment holds: Any tiling of any rectangle either has all tiles horizontal
or all vertical. To see this, assume otherwise and draw a horizontal
line across the square at the top, and imagine moving it downwards.
Suppose that just below the top border of the square, our line cuts
a tiles vertically and b horizontally. Then, except when it coincides
with some tile boundary, our line must always cut a tiles vertically
and b horizontally; for otherwise, if the tiles are x × y, we would have
ax + by = cx + dy for (a, b) ̸= (c, d), giving x/y = (b − d)/(a − c), a
rational ratio.
If a or b is zero all the tiles are oriented the same way and we are
done. If the parts of our line that cross tiles vertically never change,
then the height of the big rectangle is both a multiple of x and of y, an
impossibility. Thus somewhere our line reaches a point where some
vertical crossings switch to horizontal while simultaneously some
horizontal crossings switch to vertical, but that again is impossible
because the distance from the top of the big rectangle to the place
where this happens would then be a multiple of both x and y.
If the ratio of the sides of a tile is rational, we may as well assume
the sides have integer lengths a and b, and that the tiled square C is
c × c. If there is no horizontal tiling, a and b do not both divide c. We
may as well suppose that a does not divide c, and write c = ma + r,
where 0 < r < a.
We want to show that in this case there is no tiling at all. For
example, there can’t be any tiling of a 10×10 square by 4×1 rectangles.
We cover C with a grid of c2 unit squares S(i, j), numbered as in
a matrix by pairs i, j, 1 ≤ i ≤ c, 1 ≤ j ≤ c. We now color all the
squares S(i, j) for which i − j ≡ 0 mod a.
This colors all c unit squares on the main diagonal, plus all the
squares on other diagonals a multiple of a diagonals away from the
main diagonal; and it colors every ath square along each row or col-
umn. As a consequence, suppose R is any rectangle carved out of the
grid. If either R’s height or R’s width is a multiple of a (or both) then
R is “balanced,” in the sense that the colored squares in R comprise
exactly the fraction 1/a of all the unit squares in R.
Thus the c × ma rectangle U containing S(1, 1) is balanced, as are
also the ma × c rectangle V containing S(1, 1) and their intersection,
the ma × ma square W again containing S(1, 1). Let L be the little
r ×r square containing S(c, c), that is, diagonally opposite the S(1, 1)
corner.
The figure shows the a = 4, c = 10 case, with the rectangles U
and V (and thus also the square W ) outlined in heavy black.

72
 Safety in Numbers

If |R| denotes the area of a rectangle R, then |C| = |U | + |V | −

|W | + |L| and a similar statement applies to the numbers of col-
ored squares in each of these rectangles. But L is not balanced: It
containes r colored squares on its main diagonal, so has a fraction
r/r@ = 1/r > 1/a of its squares colored. It follows that C is not
balanced.
But our tiles are balanced, thus they cannot tile C. ♡

Locker Doors
Lockers numbered 1 to 100 stand in a row in the main hallway of
Euclid Junior High School. The first student arrives and opens all the
lockers. The second student then goes through and re-closes lockers
numbered 2, 4, 6, etc.; the third student changes the state of every
locker whose number is a multiple of 3, then the next every multiple
of 4, etc., until the last student opens or closes only locker number
100.
After the 100 students have passed through, which lockers are
open?
Solution: The state of locker n is changed when the kth student
passes through, for every divisor k of n. Here, we make use of the fact
that divisors usually come in pairs {j, k} where j · k = n (including
the pair {1,n}); so the net effect of students j and k on this locker is
nil. The exception is when n is a perfect square,
√ in which case there
is no other divisor to cancel the effect of the nth student; therefore,
the lockers which are open at the end are exactly the perfect squares,
1, 4, 9, 16, 25, 36, 49, 64, 81, and 100. ♡

73
 Mathematical Puzzles

For the next puzzle you might like to know a certain handy rhyme,
attributed to the mathematician Nathan Fine but inspired by a lovely
proof by the great, late Paul Erdős:

Chebyshev said it and I say it again

There’s always a prime between n and 2n.

This fact is known as Bertrand’s Postulate, proved by Chebyshev in

1852, with later (neater) proofs by Ramanujan and Erdős.

Factorial Coincidence
Suppose that a, b, c, and d are positive integers, all different, all greater
than one. Can it be that a!b = c!d ?

Solution: Assume that such numbers exist, say with a < c, thus
b > d. Then c > 2, and by Bertrand’s Postulate, there is a prime p with
c/2 < p < c. This p appears exactly d times in the prime factorization
of c!d but either b times or not at all in the prime factorization of a!b .
The contradiction shows that the desired quartet (a, b, c, d) does not
exist.

Even Split
Prove that from every set of 2n integers, you can choose a subset of
size n whose sum is divisible by n.
Solution: Call a set “flat” if it sums to 0 modulo n. Let us note first
that the statement we want to prove implies the following seemingly
weaker statement: If S is a flat set of 2n numbers, then S can be split
into two flat sets of size n. However, that in turn implies that any set
of only 2n − 1 numbers contains a flat subset of size n because we
can add a 2nth number to make the original set flat, then apply the
previous statement to split this into two flat subsets of size n. One of
these (the one without the new number) will do the trick.
So all three of these statements are equivalent. Suppose we can
prove the second for n = a and for n = b. Then if a set S of size
2n = 2ab sums to 0 mod ab, it is, in particular, flat with respect to a,
and we can peel off subsets S1 , . . . , S2b of size a which are also flat
with respect to a. Each of these subsets Si has a sum we can write in
the form abi . The numbers bi now constitute a set of size 2b which
sums to 0 mod b, so we can split them into two sets of size b which

74
 Safety in Numbers

are flat with respect to b. The unions of the sets Si in each part are a
bipartition of the original S into sets of size ab which are ab-flat, just
what we wanted.
It follows that if we can prove the statement for n = p prime, then
we have it for all n. Let S be a set of size 2p, with the idea of creating
a p-flat subset of size p.
How can we create such a subset? One natural possibility is to
pair up the elements of S and choose one element from each pair. Of
course, if we do that, it will behoove us to ensure that the elements in
each pair are different mod p, so our choice will not be of Hobson’s
variety. Can we do that?
Yes, order the elements of S modulo p (say, 0 through p−1) and
consider the pairs (xi , xi+p ) for i = 1, 2, . . . , p. If xi were equivalent
to xi+p mod p for some i, then xi , xi+1 , . . . , xi+p would all be equiv-
alent mod p and we could take p of them to make our desired subset.
Now that we have our pairs, we proceed by “dynamic program-
ming.” Let Ak be the set of all sums (mod p) obtainable by adding one
number from each of the first k pairs. Then |A1 | = 2 and we claim
|Ak+1 | ≥ |Ak |, and moreover, |Ak+1 | > |Ak | as long as |Ak | ̸= p. This
is because Ak+1 = (Ak +xk+1 )∪(Ak +xk+1+p ); thus if |Ak+1 | = |Ak |,
these two sets are identical, implying Ak = Ak + (xk+1+p − xk+1 ).
This is impossible since p is prime and xk+1+p − xk+1 ̸≡ 0 mod p,
unless |Ak | = 0 or p.
Since there are p pairs, we must eventually have |Ak | = p for some
k ≤ p, hence |Ap | = p and, in particular, 0 ∈ Ap . The statement of
the puzzle follows. ♡

Factorials and Squares

Consider the product 100! · 99! · 98! · · · · · 2! · 1!. Call each of the
100 factors k! a “term.” Can you remove one term and leave a perfect
square?

Solution: Perfect squares have some nice properties. For example,

the product of any number of perfect squares is itself a perfect square:
e.g., A2 · B 2 · C 2 = (A · B · C)2 .
Let’s call our big product N and observe that it’s not far from al-
ready being a product of perfect squares. The product of N ’s first two
terms, for example, is 100 · 99!2 (which happens to be a perfect square
since 100 = 102 ).

75
 Mathematical Puzzles

In fact, we could pair up all the terms and write N in the following
form: 100 · 99!2 · 98 · 97!2 · 96 · 95!2 · · · · · 4 · 3!2 · 2 · 1!2 , which is a
perfect square times the number M = 100 · 98 · 96 · · · · · 4 · 2. But
we can rewrite M as 250 · 50 · 49 · 48 · · · · · 2 · 1 = (225 )2 · 50!.
So N is the product of a lot of perfect squares and the number 50!.
It follows that if we remove the 50! term from N , we’re down to a
perfect square. ♡
√
It’s time for our theorem. You probably know that 2 is irrational,
that is, it cannot be written as a fraction a/b for any integers a or b.
You might even √ know a proof, something like this:
Suppose 2 is a fraction, and write that fraction in lowest terms
as a/b. Then 2 = a2 /b2 , and therefore a2 = 2b2 , thus a2 is even and
so must be a; write a = 2k. Then b2 = 2k 2 so b is also even, but this
contradicts the assumption that a/b is in lowest terms.
√
Theorem. For any positive integer n, n is either an integer or irrational.

Proof. The proof for n = 2 extends automatically to the case where

√
n is a prime p: If p = a/b, reduced to lowest terms, then p = a2 /b2 ,
so a is a multiple of p and thus a2 = (pk)2 is a multiple of p2 , but then
b2 is a multiple of p thus so is b, and we again contradict a/b being
reduced to lowest terms.
This doesn’t work when n is not prime, though, because we cannot
deduce from a2 being a multiple of n that a is too. But wait, this does
work when n is square-free, that is, it is a product of distinct primes;
then, we can apply the√argument for each prime separately.
So we know that n is either an integer or irrational when n is
square-free. But every n is the product of a perfect square and a square-
free number (the latter obtained by taking the product of all primes
that appear an odd number of times in the factorization
√ √of n). So let’s
write n = k 2 m, where m is square-free. Then n = k m and using
the fact that an integer multiple of an integer is an integer, and a non-
zero-integer multiple of an irrational number is irrational, we have
our proof. ♡
With a little work you can extend the above to kth-roots of n,
where k and n are integers greater than 1.

76
 7. The Law of Small
Numbers
It sounds silly, but many otherwise-very-smart people, given a puz-
zle with numbers in it, treat those numbers as if underlined in a sa-
cred text. This is math—you’re allowed to change the numbers and
see what happens! If the numbers in a puzzle are dauntingly large,
replace them with small ones. How small? As small as possible, with-
out making the puzzle trivial; if that doesn’t give you enough insight,
make them gradually bigger.

Domino Task
An 8 × 8 chessboard is tiled arbitrarily with 32 2 × 1 dominoes. A
new square is added to the right-hand side of the board, making the
top row length 9.
At any time you may move a domino from its current position
to a new one, provided that after the domino is lifted, there are two
adjacent empty squares to receive it.
Can you retile the augmented board so that all the dominoes are
horizontal?

Solution: Yes. Let T be the “snake” tiling obtained by placing four

vertical tiles on the left column, three on the right column (missing the
top and bottom squares), and filling in all but the lower right square
with horizontal dominoes.
We proceed to construct the snake tiling one tile at a time, as fol-
lows: At each time there’s a hole at some square, say S, and adjacent
to it is the other square (call it S ′ ) covered by the snake tile that covers
S; move the domino currently covering S ′ so that it covers both S and
S ′ , thus becoming the snake tile.
In this way the hole zigzags across the board and back, finally
reaching the lower right-hand square when the snake tiling is com-
plete.

77
 Mathematical Puzzles

Now it’s easy to shift the snake tiles one at a time, beginning with
the rightmost one on the bottom row, to create a horizontal tiling.
How to find this curious two-part algorithm? Try the problem first
on a 4 × 4 board!

Spinning Switches
Four identical, unlabeled switches are wired in series to a light bulb.
The switches are simple buttons whose state cannot be directly ob-
served, but can be changed by pushing; they are mounted on the
corners of a rotatable square. At any point, you may push, simul-
taneously, any subset of the buttons, but then an adversary spins the
square. Is there an algorithm that will enable you to turn on the bulb
in at most a fixed number of spins?

Solution: Looking at a simpler version of this puzzle is crucial.

Consider the two-switch version, where all you’ve got are two but-
tons on diagonally opposite corners of the square. Pushing both but-
tons will ascertain whether the two switches were both in the same
state, since then the bulb will light (if it wasn’t already lit). Otherwise,
push one button, after which they will be in the same state, and at
worst one more operation of pushing both buttons will turn on the
bulb. So three operations suffice.

Back to the four-switch case. Name the buttons N, E, S, and W

after the compass directions, although of course the botton you’re call-
ing N now might be the button you will call E, W, or S after a spin.
Suppose that at the start, diagonally opposite switches (N and S, E
and W) are in the same state—both on or both off. Then you can
treat opposite pairs as a single button and use the two-button solu-
tion: Push both pairs (i.e., all four switches); then one pair (which
may as well be N–S); then both pairs again, and you’re done. So be-
gin with those three operations; if the light doesn’t go on, then one or
both of the opposite pairs must have been mismatched. Try flipping

78
 The Law of Small Numbers

two neighboring switches, say N and E, then going back through your
three-move two-button solution. Then you’re fine if both pairs were
mismatched. If not, push just one button; that’ll either make both op-
posite pairs match, or both mismatch. Run through the two-button
solution a third time. If bulb is still off, push N and E again and now
you know both opposite pairs match, and a fourth application of the
two-button solution will get that bulb turned on.
In conclusion, pushing buttons NESW, NS, NESW, NE, NESW,
NS, NESW, N, NESW, NS, NESW, NE, NESW, NS, NESW, will at
some point turn the light on—15 operations. No sequence of fewer
than 15 operations can be guaranteed to work because there are 24 =
16 possible states for the four switches, and they all must be tested;
you get to test one state (the starting state) for free.
Seeing the solution for four buttons, you can generalize to the case
where the number of buttons is any power of two; if there are 2k but-
k
tons, the solution will take, and require, 22 −1 steps. (When there
are n buttons, they are located at the corners of a spinnable regular
n-gon.)
The puzzle is insoluble when the number of buttons, n, is not a
power of 2. Let’s just prove that for three buttons, no fixed number of
operations can guarantee to get that bulb on. (For general n, write n
as m · 2k for some odd number m > 1; it is m which plays the role of
3 in what follows.)
You may as well assume the switches are spun before you even
make your first move. Suppose that before they are spun, the switches
are not all in the same state. Then it is easy to check that no matter what
move you planned, if you were unlucky with the spin, then after the spin
and your move, the switches will still not all be in the same state.
It follows that you can never be sure that you have ever had all
the switches in the same state, so no fixed sequence of moves can
guarantee to light the bulb. It’s curious that you can solve the problem
for 32 buttons (albeit in about 136 years, at one second per operation),
but not for just three buttons.

Candles on a Cake
It’s Joanna’s 18th birthday and her cake is cylindrical with 18 candles
on its 18′′ circumference. The length of any arc (in inches) between
two candles is greater than the number of candles on the arc, exclud-
ing the candles at the ends.

79
 Mathematical Puzzles

Prove that Joanna’s cake can be cut into 18 equal wedges with a
candle on each piece.
Solution: The conditions give some assurance that the candles are
fairly evenly spaced; one way to say that is that as we move around
the circumference from some fixed origin 0, the number of candles
we encounter is not far from the distance we have traveled. Accord-
ingly, let ai be the arc-distance from 0 to the ith candle, numbered
counterclockwise, and let di = ai − i.
We claim that for any i and j, di and dj differ by less than 1.
We may assume i < j; suppose, for instance, that dj − di ≤ −1.
Then j − i − 1 ≥ aj − ai , but j − i − 1 is the number of candles
between i and j, contradicting the condition. Similarly, if dj − di ≥ 1,
then di − dj ≤ −1 and the same argument applies to the other arc,
counterclockwise from j to i.
So the “discrepancies” di all lie in some interval of length less than
1. Let dk be the smallest of these, and let ε be some number between 0
and dk , so that all the di ’s lie strictly between ε and 1 + ε. Now cutting
the cake at ε, ε + 1, etc. gives the desired result.
How are you supposed to find this proof ? By trying two candles,
then three, instead of eighteen.

Lost Boarding Pass

One hundred people line up to board a full jetliner, but the first has
lost his boarding pass and takes a random seat instead. Each subse-
quent passenger takes his or her assigned seat if available, otherwise
a random unoccupied seat.
What is the probability that the last passenger to board finds his
seat unoccupied?

Solution: This is a daunting problem if you insist on working out

what happens with 100 passengers; the number of possiblities is astro-
nomical. So let’s reduce the number to something manageable. With
two passengers it’s obvious that the probability that the second (i.e.,
last) get her own seat is 21 . What about three passengers?
It’s useful to number the seats according to who was supposed to
sit there. If passenger 1 sits in seat 1, his assigned seat, then everyone
will be get his or her own seat. If he sits in seat 3, then passenger 2 will
get seat 2 and passenger 3 will get seat 1. Finally, if passenger 1 sits in
seat 2, then whether passenger 3 gets seat 3 will depend on whether

80
 The Law of Small Numbers

passenger 2 chooses seat 1 or seat 3. Altogether, the probability that

passenger 3 gets her own seat is 13 + 13 · 21 = 21 .
Interesting! Is it possible that the answer is always 12 ?
We notice that in the above analysis, the last passenger never ends
up in seat 2. In fact, now that we think of it, we see that with n pas-
sengers total, the last passenger never ends up in seat i for 1 < i < n.
Why? Because when passenger i came on board, either seat i was al-
ready taken, or it is taken now. Thus, seat i will never be available to
the last passenger. The only seats that passenger n could end up in are
seat 1 and seat n.
We can’t yet conclude that the probability that passenger n gets
seat n is 12 —we still need to argue that seat 1 and seat n are equally
likely to be available at the end. But that’s easy, because every time
someone took a random seat, they were equally likely to choose seat
1 or seat n. Putting it another way, seat 1 and seat n were treated
identically throughout the process; thus, by symmetry, each has the
same likelihood of being open when passenger n finally gets on board.

Flying Saucers
A fleet of saucers from planet Xylofon has been sent to bring back
the inhabitants of a certain randomly-selected house, for exhibition in
the Xylofon Xoo. The house happens to contain five men and eight
women, to be beamed up randomly one at a time.

Owing to the Xylofonians’ strict sex separation policy, a saucer

cannot bring back earthlings of both sexes. Thus, it beams people up
until it gets a member of a second sex, at which point that one is
beamed back down and the saucer takes off with whatever it has left.
Another saucer then starts beaming people up, following the same
rule, and so forth.

81
 Mathematical Puzzles

What is the probability that the last person beamed up is a

woman?
Solution: Let’s try some smaller numbers and see what happens.
Obviously if the house is all men or all women, the sex of the last per-
son beamed up will be determined. If there are equal numbers of men
and women, then by symmetry, the probability that the last person
beamed up is a woman would be 12 . So the simplest interesting case
is, say, one man and two women.
In that case, if the man is beamed up first (probability: 13 ), the last
one will be a woman. Suppose a woman is beamed up first; if she is
followed by a man (who is then beamed back down), we are down to
the symmetric case where the probability of ending with a woman is
2 . Finally, if a second woman follows the first (probability 3 · 2 = 3 ),
1 2 1 1

the man will be last to be beamed up. Putting the cases together, we get
probably 12 that the last person beamed up is a woman. Is it possible
that 12 is the answer no matter how many men and women are present,
as long as there’s at least one of each?
Looking more closely at the above analysis, it seems that the sex of
the last person beamed up is determined by the next-to-last saucer—the
one that reduces the house to one sex. To see why this is so, it is useful
to imagine that the Xylofonian acquisition process operates the fol-
lowing way: Each time a flying saucer arrives, the current inhabitants
of the house arrange themselves in a uniformly random permutation,
from which they are beamed up left to right.
For example, if the inhabitants at one saucer’s arrival consist of
males m1 and m3 and females f2 , f3 , and f5 , and they arrange them-
selves “f3 , f5 , m2 , f2 , m3 ,” then the saucer will beam up f3 , f5 , and
m2 , then will beam m3 back down again and take off with just the
females f3 and f5 . The remaining folks, m3 , m2 , and f2 will now
re-permute themselves in anticipation of the next saucer’s arrival.
We see that a saucer will be the next to last just when the per-
mutation it encounters consists of all men followed by all women,
or all women followed by all men. But no matter how many of each
sex are in the house at this point, these two events are equally likely!
Why? Because if we simply reverse the order of a such a permutation,
we go from all-men-then-all-women to all-women-then-all-men, and
vice-versa.
There’s just one more observation to make: If both men and
women are present initially, then one saucer will never do, thus there
always will be a next-to-last saucer. When that comes—even though

82
 The Law of Small Numbers

we do not know in advance which saucer it will be—it is equally likely

to depart with the rest of the men, or the rest of the women.

Gasoline Crisis
You need to make a long circular automobile trip during a gasoline
crisis. Inquiries have ascertained that the gas stations along the route
contain just enough fuel to make it all the way around. If you have an
empty tank but can start at a station of your choice, can you complete
a clockwise round trip?

Solution: Yes. The trick is to imagine that you begin at station 1

(say) with plenty of fuel, then proceed around the route, emptying each
station as you go. When you return to station 1, you will have the
same amount of fuel in your tank as when you started.
As you do this, keep track of how much fuel you have left as you
pull into each station; suppose that this quantity is minimized at sta-
tion k. Then, if you start at station k with an empty tank, you will not
run out of fuel between stations. ♡

Coins on the Table

One hundred quarters lie on a rectangular table, in such a way that
no more can be added without overlapping. (We allow a quarter to
extend over the edge, as long as its center is on the table.)
Prove that you can start all over again and cover the whole table
with 400 quarters! (This time we allow overlap and overhang).

Solution: Let us observe first that if we double the radius (say, from
1′′ to 2′′ ) of each of the original coins, the result will be to cover the
whole table. Why? Well, if a point P isn’t covered, it must be 2′′ or
more from any coin center, thus a (small) coin placed with its center
at P would have fit into the original configuration. (See the first two
figures below for an example of an original configuration, and what
happens when the coins are expanded.)
Now, if we could replace each big coin by four small ones that
cover the same area, we’d be done—but we can’t.
But rectangles do have the property that they can be partitioned
into four copies of themselves. So, let us shrink the whole picture (of
big coins covering the table) by a factor of two in each dimension, and
use four copies (as in the next figure) of the new picture to cover the
original table!

83
Mathematical Puzzles

84
 The Law of Small Numbers

Surprisingly (perhaps), this lovely but seemingly rather crude ar-

gument gives the best possible factor: Replace the factor 4 by anything
smaller, say 3.99, and the statement of the puzzle is no longer true.
To see this, we consider the limiting case where the table is very
large and the coins numerous, so that boundary effects are negligi-
ble. Replace the table by a honeycomb-pattern bathroom floor with
regular hexagonal tiles of diameter 2. Since each tile could then
√ parti-
tioned into six equilateral
√ triangles
√ of side 1 and thus area 3/4, the
tile itself has area 6 × 3/4 = 3 3/2.
We can cover the floor now by covering each tile with a coin whose
boundary is the tile’s circumscribing circle (see next figure).

Each coin then has radius 1 and thus area π. If the floor has area
A then,√ignoring boundary √ effects, the total area of the coins will be
πA/(3 3/2) = 2πA/(3 3).
Now, how thinly can we cover the floor without being able to add
a non-overlapping coin? Let us use the same tiling, but this time we
cover only every third tile (see next figure); and we cover it not with a
circumscribed circle, but a coin placed over the center of the hexagon
which is just a teensy bit bigger than the inscribed circle. This just barely
prevents us from adding any more coins; how much is the coin area
now?
Well, the coin radius is a tiny bit bigger than the altitude of √ one
of the six equilateral triangles making up√ a hexagon—namely, 3/2.
Hence the coin area just exceeds π × ( 3/2)2 = 3π/4.
It follows that the total
√coin area on the √ floor is as close as we like
to (1/3) × (3π/4) × A/(3 3/2) = πA/(6 3), one-fourth of what we
had before!
The result of all this is that we have proved not only the statement
of the puzzle, but two not-so-easy extremal properties of disks in the

85
 Mathematical Puzzles

plane. The first says that there is no better way to cover the plane by
unit disks than by circumscribing the tiles in a hexagonal tiling, as
we did above; the second says that there is no more efficient way to
prevent the addition of a non-overlapping unit disk than by centering
on every third hexagonal tile a disk slightly bigger than the inscribed
disk, again as we did above.
If you think these properties are obvious to begin with, consider
the (seemingly) even more obvious fact that the densest way to pack
unit disks in the plane is to use an inscribed disk in each hexagon.
This was not rigorously proved until the great Hungarian geometer
László Fejes Tóth (1915–2005) did it in 1972!

Coins in a Row
On a table is a row of 50 coins, of various denominations. Alix picks a
coin from one of the ends and puts it in her pocket; then Bert chooses
a coin from one of the (remaining) ends, and the alternation continues
until Bert pockets the last coin.
Prove that Alix can play so as to guarantee at least as much money
as Bert.

Solution: This puzzle resists the most obvious approaches. It’s easy
to check that Alix could do quite badly by always choosing the most
valuable coin, or the coin that exposes the less valuable coin to Bert,
or any combination of these. Basically, if she only looks a move or
two ahead, she’s in trouble.
In fact, for Alix to play optimally, she needs to analyze all the pos-
sible situations that may later arise. This can be done by a technique
called “dynamic programming.”

86
 The Law of Small Numbers

But we were not asked to provide an optimal strategy for Alix, just
a strategy that guarantees her at least half the money. Experimenting
with 4 or 6 coins instead of 50 might lead you to the following key
observation.
Suppose the coins alternate quarter, penny, quarter, penny, and so
forth, ending (since 50 is even) in a penny. Then Alix can get all the
quarters! In fact, no matter what the coins are, if we number the coins
from 1 to 50 left to right, Alix can take all the odd-numbered ones—or
all the even-numbered ones.
But wait a minute—one of those two groups of coins must contain
at least half the money! ♡

Powers of Roots
√ √
What is the first digit after the decimal point in the number ( 2+ 3)
to the billionth power?
√ √
Solution: If you try entering ( 2 + 3)1,000,000,000 in your com-
puter, you’re likely to find that you get only the dozen or so most
significant figures; that is, you don’t get an accurate enough answer
to see what happens after the decimal point.
But you can try smaller powers √ and√see what happens. For exam-
ple, the decimal expansion of ( 2 + 3)10 begins 95049.9999895. √ √
A bit of experimentation shows that each even power of ( 2 + 3)
seems to be just √ a hair
√ below some integer. Why? And by how much?
√ Let’s
2
√ try2 ( 2 + 3) , which is about √ 9.9. If√we play with 10 −
( 2 + 3) √ we √ discover
√ that√it’s equal to ( 3 − 2)2 . Aha!
Yes, ( 3+ 2)2n +( 3− 2)2n is always an integer, because when
you expand it, the odd binomial√ coefficients
√ 2n cancel and the even ones
are integers. But of course ( √ 3− √ 2) is very small, about 10−n , so
the first roughly n digits of ( 3 + 2)2n after the decimal point are
all 9’s.

Coconut Classic
Five men and a monkey, marooned on an island, collect a pile of
coconuts to be divided equally the next morning. During the night,
however, one of the men decides he’d rather take his share now. He
tosses one coconut to the monkey and removes exactly 15 of the re-
maining coconuts for himself. A second man does the same thing,
then a third, fourth, and fifth.

87
 Mathematical Puzzles

The following morning the men wake up together, toss one more
coconut to the monkey, and divide the rest equally. What’s the least
original number of coconuts needed to make this whole scenario pos-
sible?
Solution: You can solve this by considering two men instead of five,
then three, then guessing. But the following argument is irresistible,
once found.
There’s an elegant “solution” to the puzzle if you allow negative
numbers of coconuts(!). The original pile has −4 coconuts; when the
first man tosses the monkey a coconut, the pile is down to −5 but
when he “takes” 15 of this he is actually adding a coconut, restoring
the pile to −4 coconuts. Continuing this way, come morning there
are still −4 coconuts; the monkey takes one and the men split up the
remaining −5.
It’s not obvious that this observation does us any good, but let’s
consider what happens if there is no monkey; each man just takes 15
of the pile he encounters, and in the morning there’s a multiple of 5
coconuts left that the men can split. Since each man has reduced the
pile by the fraction 45 , the original number of coconuts must have been
a multiple of 56 (which shrinks to 45 · 5 by morning).
All we need to do now is add our two pseudo-solutions, by starting
with 56 − 4 = 15,621 coconuts. Then the pile reduces successively to
4·55 −4 coconuts, 42 ·54 −4, 43 ·53 −4, 44 ·52 −4, and 45 ·5−4. When
the monkey gets his morning coconut, we have 45 · 5 − 5 coconuts,
a multiple of 5, for the men to split. This is best possible because we
needed 55 ·k −4 coconuts to start with, just to have an integer number
come morning, and to get 45 · k − 5 to be a multiple of 5 we needed
k to be a multiple of 5.

Doubtless, many theorems in mathematics were “discovered”

when someone played around with small numbers and then saw a
pattern that turned out to be a provable phenomenon.
Here’s a theorem that could well have been found that way. Sup-
pose you are running a dojo with an even number n of students. Each
day you pair the students up for one-on-one sparring. Can you do this
in such a way that over a period of days, each student spars with each
other student exactly once?

Theorem. For any even positive integer n there is a set of pairings (“perfect
matchings”) of the numbers {1, 2, . . . , n} such that every pair {i, j} appears
in exactly one pairing.

88
 The Law of Small Numbers

A check of small numbers suggests that this seems to work: For

n = 2, for instance, there is just the one pairing consisting of the
pair {1, 2}, and for n = 4 we can (in fact, must) take the pairings
{{1, 2}, {3, 4}}, {{1, 3}, {2, 4}}, and {{1, 4}, {2, 3}}.
For n = 6, though, we have choices to make. Is there a nice way
to make them?
In fact, there are several; my favorite is the following. We know
student n has to be paired with every other student; let’s put her in
the middle of a circle, with the rest of the students spaced equally
around the circle. In the ith of our n−1 pairings, student n is paired
with student i; draw a radius from n to i. The rest of the students are
paired by line segments that run perpendicular to that radius (see the
figure below).

Since n−1 is odd, all the radii from n to other students are at dif-
ferent angles; it follows that no two students are paired
( ) twice, and
since the eventual number of pairs is (n−1) × n/2 = n2 , every pair
is accounted for. ♡

89
 8. Weighs and Means
You know how to compute the average of a set of numbers

x̄ := (x1 + x2 + · · · + xn )/n

or you wouldn’t even have picked up this book. In solving puzzles,

it’s often useful to reverse this process, and compute the sum from the
average instead.
For example, have you forgotten the formula for the sum of the
integers from 1 to n? No problem. Their average is (n+1)/2 (you can
pair them up to see that) and there are n of them, so their sum must
be n(n+1)/2. Similarly, the sum of the integers from a to b must be
(b−a+1)(a+b)/2.
Here’s a theorem about averages that’s helpful in problem-solving
and is almost too obvious to call a theorem:
Theorem. If x̄ is the average of {x1 , x2 , . . . , xn } then x̄ lies between the
minimum of the xi ’s and the maximum. Moreover, the minimum, maximum
and average are either all different or all the same.
Another way of saying the first part is that some xi must be larger
than or equal to x̄, and some xi must be smaller than or equal to x̄.
Obvious, but useful!
The above theorem applies also in the more general case of
weighted averages, also known as convex combinations. There, each el-
ement of the set to be averaged has some weight (a positive real num-
ber) wi , and

x̄ := (w1 x1 + w2 x2 + · · · + wn xn )/(w1 + · · · + wn ).

Often the weights already sum to 1, so you can skip the dividing. That
happens, for example, when the wi ’s constitute a probability distribu-
tion, in which case x̄ becomes the expected value of the xi ’s. (Much
more about expectation is coming in a later chapter.)
The following puzzle is fairly typical of those in which the above
theorem can come in handy.

91
 Mathematical Puzzles

Tipping the Scales

A balance scale sits on the desk of the science teacher, Ms. McGregor.
There are weights on the scale, which is currently tipped to the right.
On each weight is inscribed the name of at least one pupil.
On entering the classroom, each pupil moves every weight carry-
ing his or her name to the opposite side of the scale. Must there be a
set of pupils that Ms. McGregor can let in that will tip the scales to
the left?

Solution: To set ourselves up for averaging, consider all subsets of

students, including the empty set and the full set. Each weight will be
on the left exactly half the time (pick any name on that weight, say
Jay, and assume the decision of whether to include Jay in the set is
the last decision to be made.) Thus the total weight on the left for all
these subsets is the same as the total weight on the right. Another way
to say it is that on the average, the scale balances. Since the empty set
results in a tip to the right, some other set must tip to the left.
The “averaging” technique used in Tipping the Scales comes up
often: Watch for it!
Here’s a quickie.

92
 Weighs and Means

Men with Sisters

On average, who has more sisters, men or women?
Solution: This puzzle can seem confusing. If, like me, you were
raised in a family with both male and female children, you might have
observed that in such a family the boys have more sisters (by one)
than the girls have. So maybe men have more sisters than women on
average?
Wait, in a family with several children, all girls, there are several
girls with sisters; in a family of all boys, none of those boys has a sister.
So maybe that cancels out the effect of the families with children of
both sexes.
The real issue here is independence of gender. If you make the rea-
sonable (and pretty accurate) assumption that the genders of siblings
are not influenced by your own gender, then knowing the numbers of
brothers and sisters a person has tells you nothing about whether that
person is male or female. It follows that the reverse is true too, thus
men and women have the same number of sisters on average.
If you want to get down to detail, there’s actually a slight positive
correlation of gender between siblings; that is, the sibling of a girl is
slightly more likely to be a girl than the sibling of a boy is. This is
caused mostly by the fact that identical twins are the same sex. As a
result, women have very slightly more sisters on average than men do.
The next puzzle uses the continuum form of our theorem.

Watches on the Table

Fifty accurate watches and a tiny diamond lie on a table. Prove that
there exists a moment in time when the sum of the distances from
the diamond to the ends of the minute hands exceeds the sum of the
distances from the diamond to the centers of the watches.

Solution: Considering just one watch, we claim that during the

passing of an hour the average distance from the diamond D to the tip
M of the watch’s minute hand exceeds the distance from D to the
center W of the watch. This is so because if we draw a line L through
D perpendicular to the line from D to W , then the average distance
from L to M is clearly equal to LW which is in turn equal to DW .
But DM is at least equal to LM and usually bigger.

93
 Mathematical Puzzles

Of course, if we average over all the watches we reach the same

conclusion, and it follows that sometime during the hour the desired
inequality is achieved.
The requirement that the watches be accurate is to ensure that
each minute hand moves at constant speed. It doesn’t matter if some
watches are slow and some fast (assuming each minute hand rotates
at constant speed), unless our patience is limited to one hour.
One additional note: If the watches are accurate and you place
and rotate them carefully, you can ensure that the sum of the distances
from the center of the table to the ends of the minute hands is always
strictly greater than the sum of the distances from the center of the
table to the center of the watches.
Another nice property of averaging is that adding or subtracting
one datum from a list has a predictable effect on the average: If the
datum exceeds the average, then adding it will raise the average and
deleting it will lower the average. If the datum is below average, the
effect will be the opposite; and of course if the datum is exactly av-
erage, there is no effect either way. The next puzzle uses this idea to
examine a frequently-heard type of insult.

Raising Art Value

Fans of Gallery A like to say that last year, when Gallery A sold to
Gallery B “Still Life with Kumquats,” the average value of the pic-
tures in each gallery went up. Assuming that statement is correct, and
that the value of each of the 400 pictures in the galleries’ combined
holdings is an integral number of dollars, what is the least possible dif-
ference between the average picture values of Gallery A and Gallery
B before the sale?

94
 Weighs and Means

Solution: It’s a bit curious that we are asked for a dollar amount
when no quantity of money is specified in the puzzle, but let’s plunge
in anyway. Say there were p pictures in Gallery A before the sale, thus
400 − p in Gallery B. Suppose the average value of A’s pictures was a,
and B’s b; then the behavior of the averages says exactly that the value
s of “Still Life with Kumquats” satisfies a > s > b.
We know that a is a fraction that can be written as m/p where m is
a whole number (of dollars); similarly b can be written as n/(400 − p)
where n is a whole number. By how little can a exceed an integer?
Easy: By 1/p, which happens when m is one more than a multiple
of p. Similarly, b’s shortfall from an integer value can be as little as
1/(400 − p). Thus the least possible difference between a and b is
1/p + 1/(400 − p) which is minimized when the two denominators
are equal, giving a − b = 1/200 + 1/200 = 1/100. The upshot is that
the insult is not very damaging; we can conclude from it only that
the average value of Gallery A’s holdings exceeds the average value of
Gallery B’s holdings by at least one cent.
Averaging comes up a lot in probability theory—whole chapters
on probability and on expectation lie ahead. But we can do the follow-
ing two averaging puzzles without needing any special machinery.

Waiting for Heads

If you flip a fair coin repeatedly, how long do you have to wait, on
average, for a run of five heads?

Solution: Since the probability of seeing HHHHH in a particular

series of five coinflips is 1/32, you might think it would take 32 flips
on average to get HHHHH. Indeed, 32 flips is the average wait be-
tween occurrences of HHHHH, but this includes, for example, a wait
of length 1 between the first five heads in HHHHHH and the last five.

95
 Mathematical Puzzles

Waits of length 1 can’t help us, because we have no “head start” (OK,
pun intended) when we begin flipping.
The real answer is much greater. Between runs, half the time you
get the wait of 1 and the rest of the time 1+x, where x is the desired
quantity. Hence it is not x but the average of 1 and 1+x that is equal
to 32, which gives us x = 62.
What are best patterns (of fixed length, say 5) to choose if you
want to hit your pattern ASAP? HHHHT is one of them, because
seeing an HHHHT gives you no head start toward seeing another.
Therefore, starting fresh with your coinflips costs you nothing, and
you expect to hit your target in just 32 steps on average.

Finding a Jack
In some poker games the right of first dealer is determined by dealing
cards face-up (from a well-shuffled deck) until some player gets a jack.
On average, how many cards are dealt during this procedure?
Solution: 10.6. If you tried to compute for each k the probability
that the first jack would be the kth card, with the intent of computing
the expected value of k directly, you might have found it tough going.
But there’s an easy way.
Each non-jack is equally likely to be in any of the five regions
between successive jacks (to see this, insert a joker, permute circularly
at random, then cut at the joker and remove it; or, get a shuffled deck
by starting with the jacks, and inserting cards at random, so the next
card is equally likely to be in any of the five slots.) Thus the expected
number of cards before the first jack is 48/5 = 9.6; add 1 for the jack
itself.
Can we talk about the average of infinitely many things? Some-
times.

Sums of Two Squares

On average, how many ways are there to write a positive number n as
the sum of two squares? In other words, suppose n is a random integer
between 1 and a zillion. What is the expected number of ordered pairs
(i, j) of integers such that n = i2 + j 2 ?

Solution: Let f (n) be the number of ways to write n as the sum of

two squares. What we can’t do is add up f (n) for all integers n and

96
 Weighs and Means

then divide by (countable) infinity. But, as suggested in the puzzle, we

can choose some huge number Z, and compute

1 ∑
Z
f (n) .
Z n=1

If this value converges to some number r as Z → ∞, it seems reason-

able to call this r the average value of f (n) over all n.
∑Z
And in this case it does converge. Why? The sum n=1 f (n)
counts once each pair (i, j) of integers for which i2 + j 2 ≤ Z; these
√ the integer points in the circle x + y ≤ Z. This circle has radius
2 2
are
Z thus area πZ, and therefore the number q of integer points in it is
about πZ (more precisely, q/(πZ) → 1).
It follows that the number we want, q/Z, approaches π.
It is worth remarking that if we count just the unordered pairs of
positive integers whose squares sum to n, the answer is only π/8. The
reason is that most such pairs—in particular, pairs {i, j} for which i ̸=
j—show up eight times in the circle, namely as (i, j), (j, i), (i, −j),
etc.
We might also ask for the probability that a random n can be writ-
ten in at least one way as the sum of two squares. We can deduce from
the above calculation that this probability cannot exceed π/8, which
is what the answer would be if most numbers could be written as the
sum of two squares in at most one way, up to signs and swapping.
But this is not the case at all, and in fact the probability in question
approaches zero! A famous theorem says that n is representable as
the sum of two squares only if there is no prime which (a) appears an
odd number of times in the prime factorization of n, and (b) gives a re-
mainder of 3 when divided by 4. The probability that n passes this test
is about 0.764223653589220662990698731250092328116790541 di-
vided by the square root of the natural log of n, and since log n grows
without bound, we arrive at a curious fact: Although the average of
our function f (n) is π, its value is nearly always zero.

Increasing Routes
On the Isle of Sporgesi, each segment of road (between one intersec-
tion and the next) has its own name. Let d be the average number of
road segments meeting at an intersection. Show that you can take a
drive on the Isle of Sporgesi that covers at least d road segments, and
hits those segments in strict alphabetical order!

97
 Mathematical Puzzles

Solution: This puzzle restates a very general, yet surprisingly little-

known, theorem of graph theory. Recall that a graph is a (finite, for
us) set of vertices together with some pairs of vertices, known as edges.
The degree of a vertex is the number of edges it belongs to; if we sum
up the degrees of all the vertices each edge is counted twice, thus the
average degree of the graph is 2m/n, where m is the number of edges
and n is the number of vertices.
A walk of length k is a sequence of k edges, each one connecting
the end-vertex of the previous edge (if there is one) to the start-vertex
of the next one. In a walk, as opposed to a “path,” a vertex may be
used more than once. If the edges of the graph are ordered in some
fashion, then a walk is increasing if the edges are encountered in strict
increasing order—in particular, the edges of an increasing walk must
be distinct. The figure below shows an increasing walk in a graph with
edges ordered by letters.

Here, then, is the claimed theorem.

Theorem. Let G be an arbitrary graph with an arbitrary ordering of its
edges. Then there is an increasing walk in G of length at least the average
degree of G.
Proof. We denote by d (as in the puzzle statement) the average de-
gree of G. Since d is already an average, it’s natural to look for some
set of increasing walks whose average length is d; then we can ap-
ply our theorem to deduce that one of them has length at least d. But
where do we get such a set of walks? The set of all walks doesn’t work.

98
 Weighs and Means

Here’s an (admittedly brilliant) idea. Place a pedestrian at each

vertex, and at time 1, let the pedestrians at each end of the first edge
traverse that edge, thus changing places. At time 2, the pedestrians
who now find themselves at the ends of the second edge change
places, and so forth, until all m edges have been traversed.
Now, observe: Each pedestrian has taken an increasing walk! The
total length of all these walks is 2m and there are n of them, so their
average length is d = 2m/n and we are done. ♡
It’s worth noting that this theorem is tight, that is, there are graphs
with edges ordered so that no increasing walk exceeds d in length.
For example, let K2t denote the complete graph (all edges present)
on 2t vertices. A complete matching in K2t (or in any graph) is a set
of edges whose endpoints hit every vertex exactly once. The theorem
we proved in the last chapter shows that the edges of K2t can be par-
titioned into disjoint complete matchings (necessarily, 2t−1 of them,
since every vertex is incident to 2t−1 edges.)
Now, suppose we let the matchings be M1 , . . . , M2t−1 and we use
them to order the edges as follows. First, the edges of M1 are labeled
1 through t arbitrarily. Then the edges of M2 are labeled with t+1
through 2t, those of M3 from 2t+1 through 3t, etc.
A walk in K2t cannot step directly from one edge to another of the
same matching, because no two matching edges meet at a vertex. But
an increasing walk can’t return later to the same matching, on account
of the way we have numbered the edges. So an increasing walk hits
each of the matchings M1 , . . . , M2t−1 at most once, thus can have
length at most 2t−1 = d.

99
 9.The Power of Negative
Thinking
Contradiction is a ridiculously simple, yet somehow incredibly pow-
erful, tool. You want to prove statement A is true? Assume it isn’t, and
employ reductio ad absurdum: Derive an impossible conclusion. Voila!
The great mathematician Godfrey H. Hardy put it nicely: “Reduc-
tio ad absurdum, which Euclid loved so much, is one of a mathemati-
cian’s finest weapons. It is a far finer gambit than any chess play: A
chess player may offer the sacrifice of a pawn or even a piece, but a
mathematician offers the game.”
In fact, our first application of contradiction is to figure out who
wins a certain game even though we didn’t know how to play it well—
and still don’t.

Chomp
Alice and Bob take turns biting off pieces of an m × n rectangular
chocolate bar marked into unit squares. Each bite consists of selecting
a square and biting off that square plus every remaining square above
and/or to its right. Each player wishes to avoid getting stuck with the
poisonous lower-left square.

Show that, assuming the bar contains more than one square, Alice
(the first to play) has a winning strategy.

101
 Mathematical Puzzles

Solution: It’s easy to see that Alice can win if the bar is a square,
as she can bite off all but the left-hand and bottom edges, then when
Bob bites off a piece of one leg she bites of the same amount of the
other. But how does she win in general? The surprising answer is: No
one knows!
But we do know that either the first player (Alice) or the second
(Bob) must have a winning strategy; suppose it is Bob. Then, in partic-
ular, Bob must have a winning answer to Alice’s opening move when
she merely nibbles off the upper right-hand square.
However, whatever Bob’s reply is could have been made by Al-
ice as her own opening move, after which she could follow Bob’s al-
legedly winning strategy and chalk up the game. This of course con-
tradicts the assumption that Bob can always win. Hence, it must be
Alice that has a winning strategy.
This kind of proof is known as a strategy-stealing argument and does
not, unfortunately, tell you how Alice actually wins the game. A sim-
ilar argument shows that the first player must have a winning strategy
in the game of Hex (look it up if you haven’t heard of it), yet we don’t
even know a winning first move.
If you know how to play dots and boxes, you can try this one with
the same idea in mind.

Dots and Boxes Variation

Suppose you are playing Dots and Boxes, but with each player hav-
ing an option of whether to go again after making a box. Suppose the
board has an odd number of cells (thus, someone must win). Show
that the first player to make a box has a winning strategy.

Solution: This just amounts to noticing that by choosing to play

again, you are putting yourself in the position your opponent would
be in if you didn’t play again. One of these choices must be winning!

Big Pairs in a Matrix

Suppose that the sum of the greatest two numbers in each row of a
certain square matrix is r, and the sum of the greatest two numbers
in each column is c. Show that r = c.

Solution: Suppose otherwise; by symmetry we may assume r > c.

Circle the greatest number in each row, and draw a square around the

102
 The Power of Negative Thinking

next greatest. The circled numbers are all at least r/2, so they must
all be in different columns. Suppose the highest number with a square
around it, say x, is in column j along with column j’s circled number
y. Let z be the square-encased number in y’s row; then y + z = r but
y + x = c, impossible because c < r but x ≥ z.

Bugging a Disk
Elspeth, a new FBI recruit, has been asked to bug a circular room
using seven ceiling microphones. If the room is 40 feet in diameter
and the bugs are placed so as to minimize the maximum distance
from any point in the room to the nearest bug, where should the bugs
be located?

Solution: Assuming the ceiling is flat, the problem becomes two-

dimensional: Locate seven bugs in a disk of diameter 40 feet so that
every point in the disk is within distance d of some bug, where d is as
small as possible.
This means that the seven disks of radius d centered at the bugs
cover the room. Imagine a regular hexagon inscribed in the big circle;
each side will be 20 feet in length, and the centers of the sides, together
with the center of the room, will be within 10 feet of any point in the
room. That gives d = 10; can we do any better?

Suppose we could. Since the circles of radius 10 ft about the previ-

ous outer six points just barely cover the room’s circumference, each

103
 Mathematical Puzzles

using its full diameter, covering the room’s circumference with cir-
cles of radius less than 10 would require at least seven circles. But that
would leave the center of the room uncovered, and that is our contra-
diction.
Often using contradiction is tantamount to using mathematical
induction: You assume the statement you wish is false, and consider
a minimal counterexample.

Pairs at Maximum Distance

Let X be a finite set of points on the plane. Suppose X contains n
points, and the maximum distance between any two of them is d.
Prove that at most n pairs of points of X are at distance d.

Solution: To solve this puzzle, it’s useful to observe that if A, B and

C, D are two “max pairs” (pairs of points from X at distance d), then
the line segments AB and CD must cross (else one of the diagonals
of the quadrilateral ABDC would exceed d in length).
Now assume the statement of the puzzle is false, and let n be the
size of a smallest counterexample. Since there are more than n max
pairs and each has two points, there must be a point P which partici-
pates in three max pairs (say, with points A, B, and C). Every two of
the segments P A, P B, and P C must make at most a 60◦ angle at P ,
and one of them, say B, must lie between the others.
But this makes it pretty tough for B to be in any other max pair,
since if BQ were a max pair, it would have to intersect both P A and
P C—an impossibility. Thus we can drop B out of X altogether, los-
ing only one max pair and obtaining a smaller counterexample. This
contradiction completes the proof.

Lemming on a Chessboard
On each square in an n × n chessboard is an arrow pointing to one of
its eight neighbors (or off the board, if it’s an edge square). However,
arrows in neighboring squares (diagonal neighbors included) may not
differ in direction by more than 45 degrees.
A lemming begins in a center square, following the arrows from
square to square. Is he doomed to fall off the board?

104
 The Power of Negative Thinking

Solution: The lemming is indeed doomed. One way to see this is

to imagine that the lemming can move to any neighboring square, but
must turn to face in the direction of the arrow found there. Then the
lemming cannot turn 360 degrees around, because if he could you
could shrink the cycle on which he does that until it collapses into a
contradiction. But the real lemming, if he is to stay on the board, must
eventually cycle and when he does so, he will have to make that 360◦
turn.
The conclusion can also be reached using induction (thus also
contradiction). If the lemming stays on the board, he must, as we have
already noted, eventually settle into a cycle. Let C be the smallest-
area cycle (on any board) on which this can happen, and suppose it’s
a clockwise cycle. Cut the whole board down to C and its interior,
then rotate all the arrows 45◦ clockwise to force a smaller cycle!

Curve on a Sphere
Prove that if a closed curve on the unit sphere has length less than 2π,
then it is contained in some hemisphere.

Solution: Pick any point P on the curve, travel half way around the
curve to the point Q, and let N (standing for North Pole) be the point
half-way between P and Q. (Since the distance d(P, Q) from P to Q
is less than π, N is uniquely defined). Thinking of N as the North
Pole provides us with an equator, and if the curve lies entirely on N ’s
side, that is, in the northern hemisphere, we are done. Otherwise, the
curve crosses the equator, and let E be one of the points at which it
does so. Then, we observe that d(E, P ) + d(E, Q) = π, since if you
poke P through the equatorial plane to P ′ on the other side, P ′ is
antipodal to Q; hence, d(E, P ′ ) + d(E, Q) = π.
However, for any point X on the curve, d(P, X) + d(X, Q) must
be less than π, and this provides the desired contradiction. ♡

Soldiers in a Field
An odd number of soldiers are stationed in a large field. No two sol-
diers are exactly the same distance apart as any other two soldiers.
Their commanding officer radios instructions to each soldier to keep
an eye on his nearest neighbor.
Is it possible that every soldier is being watched?

105
 Mathematical Puzzles

Solution: The problem is most easily solved by considering the two

soldiers at shortest distance from each other. Each of these watches
the other; if anyone else is watching one of them, then we have a sol-
dier being watched twice and therefore another not being watched at
all. Otherwise, these two can be removed without affecting the others.
Since the number of soldiers is odd, this procedure would eventually
reduce to one soldier not watching anyone, a contradiction.
Our next puzzle is a more serious, and indeed classical, mathe-
matical conundrum.

Alternating Powers
Since the series 1 − 1 + 1 − 1 + 1 − · · · does not converge, the function
f (x) = x−x2 +x4 −x8 +x16 −x32 +· · · makes no sense when x = 1.
However, f (x) does converge for all positive real numbers x < 1. If
we wanted to give f (1) a value, it might make sense to let it be the
limit of f (x) as x approaches 1 from below. Does that limit exist? If
so, what is it?

Solution: This question began life in a paper by G. H. Hardy, the

same guy whose appreciation for the power of contradiction was
noted at the beginning of this chapter. Ironically, although Hardy
answers his question correctly, he comments that no completely el-
ementary proof seems to be known; a century or so later, we have
one, and it involves contradiction.
If you try to determine the answer to this puzzle by computing
f (x) for various x near 1, you might reach the wrong conclusion: It
appears to converge to 21 . But appearances can be deceptive, and in
fact the limit does not exist!
Suppose f (x) does have a limit, say c, as x approaches 1 from
below. Since f (x) = x − f (x2 ), that limit can only be 12 (because,
taking x as close to 1 as needed, we get c = 1 − c). But notice that
f (x4 ) − f (x) = x − x2 > 0, for all x strictly between 0 and 1. From
this we conclude that the sequence f (x), f (x4 ), f (x16 ), . . . is strictly
increasing.
It follows that if there is any x < 1 for which f (x) > 12 , there can
be no limit. In fact, f (0.995), for instance, exceeds 0.50088.
What actually happens is that f (x) oscillates more and more
rapidly inside an interval of length about 0.0055 centered at 12 . Seems
a bit capricious, no? The function g(x) = 1 − x + x2 − x3 + x4 − · · ·
is also defined for each positive x < 1 and has the same problem at

106
 The Power of Negative Thinking

x = 1 that f did. But this one is equal to 1/(x+1), as you can check
by adding xg(x) to g(x), thus it docilely approaches 12 as x → 1.

Halfway Points
Let S be a finite set of points in the unit interval [0,1], and suppose
that every point x ∈ S lies either halfway between two other points in
S (not necessarily the nearest ones) or halfway between another point
in S and an endpoint.
Show that all the points in S are rational.

Solution: Suppose S has n points, x1 < x2 < · · · < xn . By as-

sumption, they satisfy n linear equations with rational coefficients;
for example, if point xj lies halfway between xi and xk , it satis-
fies 2xj = xi + xk . If instead xj lies between xi and 1, it satisfies
2xj = xi + 1.
If this set of equations has just the one solution, they must be ratio-
nal. Why? You may already know from linear algebra that when a set
of linear equations over a field that has a unique solution, the values
in the solution lie in the field. We can prove that easily by induction:
Take any equation and solve it for one variable, say xn . Then substi-
tute in the remaining equations to get a uniquely solvable system with
one fewer variable.
So we need only prove that our system has only one solution; sup-
pose it has a second solution, y1 , . . . , yn . Letting zi = yi − xi and
subtracting corresponding equations, we find that the zi ’s satisfy sim-
ilar conditions except that (1) they may not all be distinct (but they’re
not all 0); (2) they lie between −1 and 1; and (3) the endpoint 1 is no
longer in play, that is, each zj either lies halfway between two other
zi ’s or halfway between another zi and zero.
We may assume one of the zi ’s is positive; let zj have the largest
value and among those with that maximum value, the largest index.
Obviously zj can’t be halfway between 0 and some other zi , so it must
be between zi and zk where zi = zj = zk . By choice of j, i, and k
must both be smaller indices than j, but that can’t be because in the
original solution xj must have lain halfway between xi and xk , con-
tradicting the fact that we indexed the original solution in size order.

Our theorem is quite a famous one, and the proof given below
was often cited by Paul Erdős as an example of a “book proof.”

107
 Mathematical Puzzles

(Although Erdős claimed not to believe in God, he spoke often of

a book maintained by God that contained the best proof of every the-
orem.)

Theorem. Let X be a finite set of points on the plane, not all on one line.
Then there is a line passing through exactly two points of X.

Proof. Assume that every line through two or more points in X

in fact contains at least three points of X. The idea is to find such a
line L, and a point P not on L, such that the distance from P to L is
minimized.

Since L contains at least three points of X, two of them, say Q and

R, lie on the same side of the perpendicular to L from P . But then, if
R is the farther one, the point Q is closer to the line through P and R
than P is to L—contradiction! ♡

108
 10. In All Probability
Dealing with probabilities is intuitive for some folks, less so for
others. But there are just a few basic ideas, which, if mastered, convey
a lot of power.
A thing that may or may not happen is called an event, and its
probability measures the likelihood that it will happen; if the “trial”
that could make the event happen is repeatable (like rolling a die),
then the event’s probability is the fraction of time that it occurs in
many trials.
If the trial has n possible outcomes that are equally likely (per-
haps because of symmetry) then the probability of the event A is the
number of outcomes that result in A occurring, divided by n. For ex-
ample, if we roll a die and A is the event that an even number appears
on top, then the probability P(A) that A occurs is 3/6 = 1/2, since
three of the outcomes are even numbers. The fact that the outcomes
are equally likely relies on the assumption that the die behaves like a
true, homogeneous, symmetrical, cube.
If we roll a pair of dice, there are 36 equally likely outcomes; note
that (2,3) is different from (3,2) because even if you can’t distinguish
between the two dice, Tyche (the goddess of chance) can. Thus, for
example, there are five ways to roll a sum of 6—(1,5), (2,4), (3,3),
(4,2), and (5,1)—so P(two dice roll a sum of 6) = 5/36.
If two events A and B are independent, that is, the occurrence of
one does not affect the probability of the other, then the probability
P(A ∧ B) that they both occur is equal to the product P(A)P(B) of
the probabilities of the two events. Without the assumption of inde-
pendence, we can only say P(A ∧ B) = P(A)P(B|A) where P(B|A),
read “the probability of B given A,” has intuitive meaning but is often
defined simply as the quantity that makes the equation true.
The probability P(A ∨ B) that at least one of A and B occurs is
always equal to P(A) + P(B) − P(A ∧ B), as you can see from a Venn
diagram. This reduces to P(A ∨ B) = P(A) + P(B) only in the case
where A and B are mutually exclusive—that is, they can’t both occur.
Let’s try a few amusing applications of these ideas.

109
 Mathematical Puzzles

Winning at Wimbledon
As a result of temporary magical powers (which might need to be
able to change your gender), you have made it to the women’s singles
tennis finals at Wimbledon and are playing Serena Williams for all the
marbles. However, your powers cannot last the whole match. What
score do you want it to be when they disappear, to maximize your
chances of notching an upset win?

Solution: Naturally, you want to have won the first set of the best-
of-three-sets match, and be far ahead in the second. Your first thought
might be that you’d like to be up five games to love and serving at 40-
love (or, if your serve resembles mine, up 5-0 in games and receiving at
love-40). Either gives you three bites at the apple, that is, if your proba-
bility of winning a point is some small quantity ε then your probability
of winning is about 3ε.
(Technically, if the results of the next three points are independent,
and you discount the possibility of winning the match despite losing
them all, your success probability is 1 − (1 − ε)3 = 3ε − 3ε2 + ε3 .)
But you can do better. Let the game score be 6-6, and suppose you
are ahead 6-0 in the seven-point tiebreaker. Then you have six chances
to win, almost doubling your probability when ε is small.
As I write this, you can do even better in the Australian Open,
where the final-set tiebreaker is played to 10 points. The lesson, either
way, is this: If you need a miracle, try to arrange for as many shots at
that miracle as possible.

Birthday Match
You are on a cruise where you don’t know anyone else. The ship an-
nounces a contest, the upshot of which is that if you can find someone
who has the same birthday as yours, you (both) win a beef Wellington
dinner.
How many people do you have to compare birthdays with in order
to have a better than 50% chance of success?

Solution: Let’s assume that (a) neither you (important!) nor any-
one else on board (less important) was born on February 29, (b) other
dates are equally likely to be a given shipmate’s birthday, and (c) there
are no sets of twins (or triplets etc.) on board that will cause you
to waste queries. Then the probability that a given shipmate fails to

110
 In All Probability

match your birthday is 364/365, thus if you ask n people, the proba-
bility that you have no luck is (364/365)n . You want this number to
dip below 50%, which it does when n reaches 253; in fact (364/365)253
= 0.49952284596... .
Is 253 higher than you guessed? You’d only need 183 queries to get
your success probability over 12 , if the people you asked were guaran-
teed to have different birthdays. The trouble is, you will probably hear
a lot of duplicated answers that don’t help you. (Well, it could help
you a bit if you demand a bite of their beef Wellington in return for
telling each about the other).
If you thought the answer to the puzzle was 23, you confused the
question with the more famous problem of how many people you
need in a room to have a better than 50% probability that some pair
of people share a birthday. In fact, you can get the answer( “23”
) from
our answer; the number of pairs of people in that room is 23 2 = 253.
(The pairs are not precisely independent, but they’re close enough to
make this work.)
By the way, how do you determine that 253 is the least n for which
(364/365)n < 12 without a lot of trial and error? My favorite way is
to use the approximation (1 − 1/m)m ∼ e−m for large m. If you
take this as an equality and solve (1 − 1/365)n = 1/2 you get n =
loge (2) · 365 ∼ 252.9987.
When the event A is decided before the event B, the conditional
probability of B given A is often easy to deduce but P(A|B) is less
intuitive. The laws of probability don’t care about the order of events,
though.

Other Side of the Coin

A two-headed coin, a two-tailed coin and an ordinary coin are placed
in a bag. One of the coins is drawn at random and flipped; it comes
up “heads.” What is the probability that there is a head on the other
side of this coin?

111
 Mathematical Puzzles

Solution: The temptation is to say that the drawn coin is either the
two-headed coin or the fair coin, so the probability that its other side is
heads is 12 . But is it? Imagine that the coin is flipped many times, and
comes up heads every time. It could still be either coin, but you can’t
help thinking it’s more likely to be the two-headed coin. That kind of
reasoning is basically how we learn from statistical observations.
In fact, the inference that the drawn coin is probably two-headed
exists already after one flip. We want to compute P(A|B) where A is
the event that the two-headed coin was drawn, and B the event that
heads was flipped. For any two events A and B, we calculate

P(A ∧ B) P(A)P(B|A)
P(A|B) = =
P(B) P((A ∧ B) ∨ (A ∧ B))

P(A)P(B|A)
=
P(A ∧ B) + P(A ∧ B)
P(A)P(B|A)
=
P(A)P(B |A) + P(A)P(B |A)
where A is the event that A does not occur. This equation is known as
Bayes’ Rule, and if we plug in the values from our experiment, we get
1
·1 2
P(A|B) = 3
= .
1
3 · 1 + 23 · 1
4
3

If all the above fractions leave you unconvinced, here’s another

way to get the answer. Steal a crayon or marker from your kid and
label each of the six coin-faces as follows: “1” and “2” on the two
sides of the two-headed coin, “3” on the heads side of the fair coin,
“4” on its tails side, and finally “5” and “6” on the two sides of the
two-tailed coin.
You will agree, I hope, that drawing a coin and flipping it has the
same effect as rolling a fair die: You are equally likely to come up with
any number from 1 to 6. Given that you flipped a head, then, you are
equally likely to have seen 1, 2, or 3. Two of those three faces have a
head on the other side.
“Reverse” conditional probabilities are crucial in statistics, where
you need to compute Pr(A|B) where A is the hypothesis you are test-
ing, and B the result of your experiment. Here are some more puzzles
involving that kind of calculation.

112
 In All Probability

Boy Born on Tuesday

Mrs. Chance has two children of different ages. At least one of them
is a boy born on a Tuesday. What is the probability that both of them
are boys?
Solution: We consider first the classic version, where the informa-
tion is that at least one of Mrs. Chance’s children is a boy. Here, the
solution seems simple: a priori, the children (in age order) are equally
likely to be boy–boy, boy–girl, girl–boy, or girl–girl. The last of these
is ruled out, so the probability both children are boys is 13 .
The difficulty is that in almost every way that you are likely to
find out that at least one of Mrs. Chance’s kids is a boy, there is an
additional inference which raises the probability of two boys from 13
to 12 .
For example, if you find that the older child is a boy, then the girl–
boy combination above is ruled out and it’s clear that the probability
that the younger child is a boy is 12 . But that reasoning also applies if
your information is that the taller child is a boy, or the one with darker
hair is a boy, or the child you saw Mrs. Chance with yesterday is a boy,
or even if you happen to catch Mrs. Chance in the store buying boys’
clothes.
Why the latter? Because if Mrs. Chance’s other child were a girl,
you’d be less likely to find her buying boys’ clothes. It’s quite similar to
the previous puzzle; the fact that if the coin were fair it might not have
come up “heads” affords an inference that the coin is the two-headed
coin, raising the probability of two heads from 12 to 23 . Here the effect
is to raise the probability that the children are both boys from 31 to 21 .
Why “almost”? Is there any way you could find out that at least
one of Mrs. Chance’s kids is a boy, that would lead to the conclusion
that the probability they are both boys is 13 ? As it happens, I have a
friend whose wife became pregnant with fraternal twins, and an am-
niotic test revealed the presence of Y chromosomes in the placenta.
The conclusion was that at least one of the twins was a boy, and nat-
urally my friend and his wife hoped that the other was a girl—but
assumed that the probability of this event was 12 . In fact, it really was
2
3 ; I inquired and was told that the test always finds Y chromosomes as
long as one or more fetuses is male. (In fact, the outcome was mixed
twins.)
Now back to the boy born on a Tuesday. This is a very delicate
question and you need to assume that the information has come to
you in with no additional inferences—for example, you actually chose

113
 Mathematical Puzzles

Mrs. Chance randomly among many mothers of two (non-twin) chil-

dren and asked her if at least one was a boy born on a Tuesday, and
she said yes.
Compared to the classic version, here there is a substantial built-in
inference that both kids are boys because that nearly doubles the prob-
ability that some boy was born on a Tuesday. But not quite, because
if they were both born on Tuesday they count only once. Thus you’d
expect the answer to be nearly 12 , and that is indeed the case. The cal-
culation is straightforward. If we classify all two-kid families accord-
ing to gender and birth-day-of-the-week we get 2 × 7 × 2 × 7 = 196
possibilities, of which 1 × 2 × 7 + 2 × 7 × 1 − 1 × 1 = 27 contain a boy
born on a Tuesday (the −1 comes from over-counting the case of two
boys born on Tuesdays). Of these, 1 × 1 × 7 + 1 × 7 × 1 − 1 × 1 = 13
involve two boys, so the probability we seek is 13/27.

Whose Bullet?
Two marksmen, one of whom (“A”) hits a certain small target 75%
of the time and the other (“B”) only 25%, aim simultaneously at that
target. One bullet hits. What’s the probability that it came from A?

Solution: You might think that since A is three times as good a shot
as B, he is three times as likely to have been the one whose bullet hit;
in other words, the probability that he deserves the credit is 75%.
But there are two things happening here: hitting and missing.
Since the probability that A hits and B misses is 3/4 × 3/4 = 9/16
while the probability that B hits but A misses is only 1/4×1/4 = 1/16,
the probability that the successful bullet was A’s is actually a full 90%.
Conditional probabilities can sometimes defy your intuition. An-
other example:

Second Ace
What is the probability that a poker hand (five cards dealt at random
from a 52-card deck) contains at least two aces, given that it has at
least one? What is the probability that it contains two aces, given that
it has the Ace of Spades? If your answers are different, then: What’s
so special about the Ace of Spades?
( ) ( ) (48)
Solution: There are( 52 poker hands of which 52
) (48) (4) (48)
5 5 − 5 have
at least one ace, and 5 − 5 − 1 · 4 have at least two aces;
52

114
 In All Probability

dividing the last of these expressions by the second gives our first an-
swer, 0.12218492854.
( ) (51)
(48)
There are 51 4 hands containing the Ace of Spades and 4 −
4 containing another ace as well, giving our second answer,
0.22136854741, quite a lot larger than the first answer.
Of course, there’s nothing special about the Ace of Spades; speci-
fying the ace held changes the conditional probabilities.
Here’s a simpler example that you don’t need a calculator for. Sup-
pose a third of the families in town have two children, a third have one,
and the remaining third none. Then the probability that a family has
a second child, given that it has at least one, is 12 . But the probability
that a family has a second child given that it has a girl is 35 , because
only half the families with one child have a girl, but 34 of those with
two children have a girl.

Stopping After the Boy

In a certain country, a law is passed forbidding families to have an-
other child after having a boy. Thus, families might consist of one
boy, one girl and one boy, five girls and one boy... How will this law
affect the ratio of boys to girls?

Solution: It might seem like limiting families to one boy should

cut down the boy–girl ratio, but assuming that gender is independent
among siblings, the child whose birth is prevented by having an older
brother is as likely to be a girl as a boy. Thus the sex ratio should not
be affected by this law.
One could make the case that, on account of the possibility of
identical twins, gender among siblings is not precisely independent.
However, in the statement of the puzzle, the law’s effect when twin
boys are born has not been made clear (and is perhaps a bit horrifying
to think about). Thus we solvers are pretty much forced to discount
that possibility.
Up until now we’ve been talking about discrete probability. Things
get a bit more subtle when a random value is chosen from an interval
of real numbers. Awkwardly, we have to assign probability 0 to the
event that any particular number is chosen, even though some event
of that kind is guaranteed to happen. Typically, we deal instead with
the probability that the chosen point lies in a particular subinterval.
For example, if we choose a point uniformly at random from the unit

115
 Mathematical Puzzles

interval, the probability that it falls between 0.3 and 0.4 is 1/10 (re-
gardless of whether we count 0.3 and 0.4 as part of the interval).

Points on a Circle
What is the probability that three uniformly random points on a circle
will be contained in some semicircle?

Solution: It’s not hard to see that no matter where we put the first
two points, assuming they don’t coincide, it is more likely that the
third point will be in some semicircle with the first two, than not.
Thus, in fact, the answer to the puzzle should be more than 12 . But
how do we compute it, without the bother of trying to condition on
the distance between the first two points?
The answer is (as is often the case in probability problems) to
choose the points in a different way—but, of course, one that still
results in uniformly random points. Here, let us choose three random
diameters of our circle. Each hits the circle at two (antipodal) points,
and we use three coinflips do decide which three points to use.
This may seem like an unnecessarily complicated way to choose
our points—why in six steps, when we can do it in three steps? To see
the answer, draw any three diameters and the six points where they
hit the circle. Observe that of those six, if three consecutive points are
chosen, they will certainly lie in a semicircle; otherwise, they won’t!
Well, there are 23 = 8 ways to pick the points, using our coin-
flips, and 6 of those result in consecutive points being chosen. So
the probability of their being contained in some semicircle is 6/8 =
3/4. ♡
(If you found that puzzle too easy, try to determine the probability
that four uniformly random points on a sphere are all contained in
some hemisphere!)

Comparing Numbers, Version I

Paula (the perpetrator) takes two slips of paper and writes an integer
on each. There are no restrictions on the two numbers except that
they must be different. She then conceals one slip in each hand.
Victor (the victim) chooses one of Paula’s hands, which Paula then
opens, allowing Victor to see the number on that slip. Victor must now
guess whether that number is the larger or the smaller of Paula’s two
numbers; if he guesses right, he wins $1, otherwise he loses $1.

116
 In All Probability

Clearly, Victor can at least break even in this game, for example,
by flipping a coin to decide whether to guess “larger” or “smaller”—
or by always guessing “larger,” but choosing the hand at random. The
question is: Not knowing anything about Paula’s psychology, is there
any way he can do better than break even?
Solution: Amazingly, there is a strategy which guarantees Victor a
better than 50% chance to win.
Before playing, Victor selects a probability distribution on the in-
tegers that assigns positive probability to each integer. (For example,
he plans to flip a coin until a “head” appears. If he sees an even num-
ber 2k of tails, he will select the integer k; if he sees 2k−1 tails, he will
select the integer −k.)
If Victor is smart, he will conceal this distribution from Paula, but
as you will see, Victor gets his guarantee even if Paula finds out. It
is perhaps worth noting that although Victor might like the idea of
assigning the same probability to every integer, this is not possible—
there is no such probability distribution on the integers!
After Paula picks her numbers, Victor selects an integer from his
probability distribution and adds 12 to it; that becomes his “threshold”
t. For example, using the distribution above, if he flips five tails before
his first head, his random integer will be −3 and his threshold t will
be −2 12 .
When Paula offers her two hands, Victor flips a fair coin to decide
which hand to choose, then looks at the number in that hand. If it
exceeds t, he guesses that it is the larger of Paula’s numbers; if it is
smaller than t, he guesses that it is the smaller of Paula’s numbers.
So, why does this work? Well, suppose that t turns out to be larger
than either of Paula’s numbers; then Victor will guess “smaller” re-
gardless of which number he gets, and thus will be right with prob-
ability exactly 12 . If t undercuts both of Paula’s numbers, Victor will
inevitably guess “larger” and will again be right with probability 12 .
But, with positive probability, Victor’s threshold t will fall between
Paula’s two numbers; and then Victor wins regardless of which hand
he picks. This possibility, then, gives Victor the edge which enables
him to beat 50%. ♡
Neither this nor any other strategy enables Victor to guarantee, for
some fixed ε > 0, a probability of winning greater than 50% + ε. A
smart Paula can choose randomly two consecutive multidigit integers,
and thereby reduce Victor’s edge to an insignificant smidgen.

117
 Mathematical Puzzles

Comparing Numbers, Version II

Now let’s make things more favorable to Victor: Instead of being cho-
sen by Paula, the numbers are chosen independently at random from
the uniform distribution on [0,1] (two outputs from a standard ran-
dom number generator will do fine).
To compensate Paula, we allow her to examine the two random
numbers and to decide which one Victor will see. Again, Victor must guess
whether the number he sees is the larger or smaller of the two, with $1
at stake. Can he do better than break even? What are his and Paula’s
best (i.e., “equilibrium”) strategies?
Solution: It looks like the ability to choose which number Victor
sees is paltry compensation to Paula for not getting to pick the num-
bers, but in fact this version of the game is strictly fair: Paula can pre-
vent Victor from getting any advantage at all.
Her strategy is simple: Look at the two random real numbers, then
feed Victor the one which is closer to 12 .
To see that this reduces Victor to a pure guess, suppose that the
number x revealed to him is between 0 and 12 . Then the unseen num-
ber is uniformly distributed in the set [0, x] ∪ [1−x, 1] and is, therefore,
equally likely to be smaller or greater than x. If x > 12 , then the set is
[0, 1−x] ∪ [x, 1] and the argument is the same.
Of course, Victor can guarantee probability 12 against any strategy
by ignoring his number and flipping a coin, so the game is completely
fair. ♡
This amusing game was brought to my attention at a restaurant in
Atlanta. Lots of smart people were present and were stymied, so if you
failed to spot this nice strategy of Paula’s, you’re in good company.

Biased Betting
Alice and Bob each have $100 and a biased coin that comes up heads
with probability 51%. At a signal, each begins flipping his or her coin
once a minute and bets $1 (at even odds) on each outcome, against a
bank with unlimited funds. Alice bets on heads, Bob on tails. Suppose
both eventually go broke. Who is more likely to have gone broke first?
Suppose now that Alice and Bob are flipping the same coin, so that
when the first one goes broke the second one’s stack will be at $200.
Same question: Given that they both go broke, who is more likely to
have gone broke first?

118
 In All Probability

Solution: In the first problem, Alice and Bob are equally likely to
have gone broke first. In fact, conditioned on Alice’s having gone
broke, they have exactly the same probability of having gone broke
at any particular time t.
To see this, pick any win–loss sequence s that ends by going broke;
suppose s has n wins, thus n + 100 losses. It will have probability
pn q n+100 for Alice, but pn+100 q n for Bob, where here p = 51% and
q = 49%. The ratio of these probabilities is the constant (q/p)100 ,
thus P(Alice goes broke) = (q/p)100 (which happens to be about
0.0183058708; call it 2%). Once we divide Alice’s probability of en-
countering s by this number, her probability is the same as Bob’s.
For the second problem, where Alice and Bob were flipping the
same coin, your intuition might tell you that Alice probably went
broke first. The reasoning is that if Bob went broke first, Alice would
have to have gone broke after building up to $200, an unlikely occur-
rence. But is this true? Can’t we argue as above, comparing win–lose
sequences that result in both going broke?
No. The proof above doesn’t work because here, Alice’s going
broke also affects Bob’s longevity. In fact, the same ratio (q/p)100 ap-
plies when comparing a both-broke win–loss sequence and the con-
clusion is that the probability that Bob goes broke first is (q/p)100
times the probability that Alice goes broke first. Intuition was correct;
Alice is more than 50 times more likely to have gone broke first!
It’s worth noting that, putting Bob aside, we can deduce from the
first calculation how long it takes, on average, for Alice to go broke
given that she goes broke. The reason is that under this condition, we
have shown Alice’s expected time to ruin is the same as Bob’s. But
Bob always goes broke and since he loses 2 cents per toss on average,
his expected time to ruin is 100/0.02 = 5000 tosses. At 60 tosses per
hour that comes to 83 13 hours, or about 3 12 days.

Home-Field Advantage
Every year, the Elkton Earlies and the Linthicum Lates face off in a
series of baseball games, the winner being the first to win four games.
The teams are evenly matched but each has a small edge (say, a 51%
chance of winning) when playing at home.
Every year, the first three games are played in Elkton, the rest in
Linthicum.
Which team has the advantage?

119
 Mathematical Puzzles

Solution: The series will last 4, 5, 6, or 7 games; it seems that the

average should be somewhere between 5 and 6 games, in which case
more games are played on average in Elkton than in Linthicum; there-
fore, Elkton has the advantage.
First issue: Are more games really played in Elkton, on average?
To verify that we’ll need to compute the probabilities that the series
will last 4, 5, 6, or 7 games. If all the games were 50-50, the probabil-
ities that a series will last 4, 5, 6, or 7 games would be, respectively,
1/8, 1/4, 5/16, and 5/16. Thus the average number of games played
would be
1 1 5 5 93 13
4· +5· +6· +7· = =5 = 5.8125
8 4 16 16 16 16
and since three games are always played in Elkton, on average only 2
13/16 would be played in Linthicum.
As one might expect, re-calculating the above odds taking into
account the 1% home-field advantage changes the numbers very little.
The average number of games played shifts upward a teensy bit to
5.81267507002.
So, indeed, more games are played in Elkton, on average. Over a
very long period of time, this will be reflected in the game statistics;
you can expect Elkton to have won about 50.03222697428% of the
games over a million-year period.
However, shockingly, you can expect Linthicum to have won most
of the series! Think of it this way: Imagine that all seven games are
played (with equal ferocity) regardless of the outcomes. This doesn’t
change the probability of winning the series (if it did, we would never
be able to quit after fewer than seven games). Those seven consist of
four games in Linthicum and only three in Elkton, giving the advan-
tage to Linthicum: Their probability of winning the series works out
to 50.31257503002%.
If it seems odd to you that the home-field advantage in games 5,
6, and 7, which might not get played, counts as much as home-field
advantage in earlier games, that’s understandable. But the point is
that the late home-field advantage is there when Linthicum needs it, and
something that’s there when you need it is just as good as its being
there all the time.
Have you ever wondered about the effect of different rules, from
sport to sport, concerning which player or team serves next? Maybe
the next puzzle will answer your questions.

120
 In All Probability

Service Options
You are challenged to a short tennis match, with the winner to be
the first player to win four games. You get to serve first. But there are
options for determining the sequence in which the two of you serve:
1. Standard: Serve alternates (you, her, you, her, you, her, you).
2. Volleyball Style: The winner of the previous game serves the
next one.
3. Reverse Volleyball Style: The winner of the previous game re-
ceives in the next one.
Which option should you choose? You may assume it is to your ad-
vantage to serve. You may also assume that the outcome of any game
is independent of when the game is played and of the outcome of any
previous game.

Solution: Again using the idea (from the previous puzzle’s solu-
tion) that it doesn’t hurt to assume extra games are played, assume
that you play lots of games—maybe more than are needed to deter-
mine the match winner. Let A be the event that of the first four served
by you and the first three served by your opponent, at least four are
won by you. Then, it is easily checked that no matter which service
option you choose, you will win if A occurs and lose otherwise. Thus,
your choice makes no difference. Notice that the independence as-
sumptions mean that the probability of the event A, since it always
involves four particular service games and three particular returning
games, does not depend on when the games are played, or in what
order.
If the game outcomes were not independent, the service option
could make a difference. For example, if your opponent is easily
discouraged when losing, you might benefit by using the volleyball
scheme, in which you keep serving if you win.
The lesson learned above from Home-Field Advantage will again
come in handy in the next puzzle.

Who Won the Series?

Two evenly-matched teams meet to play a best-of-seven World Se-
ries of baseball games. Each team has the same small advantage
when playing at home. As is customary for this event, one team (say,

121
 Mathematical Puzzles

Team A) plays games 1 and 2 at home, and, if necessary, plays games

6 and 7 at home. Team B plays games 3, 4 and, if needed, 5 at home.
You go to a conference in Europe and return to find that the series
is over, and six games were played. Which team is more likely to have
won the series?
Solution: You could work this about by assigning a variable p >
1/2 to the probability of winning a game at home, then doing a bunch
of calculations. But no calculations are actually necessary, if you re-
member that home-field advantage in potential games is as good as it
is in games that are always played.
It follows that if you know that at most six games were played, you
can conclude that teams A and B are equally likely to have won. On
the other hand, if you know that at most 5 had been played, then (as
host of 3 games) Team B is more likely to have won.
From these two facts we deduce that if exactly six games were
played, Team A is the more likely winner.
It sometimes happens that there’s alternative way to solve a puz-
zle if you trust the poser, and this is the case here. Since the degree of
advantage experienced by the home team is not specified, you might
choose to believe that the answer should be the same even if that de-
gree is huge—in other words, if it requires a virtual miracle to win an
away game.
In that case, with high probability there was just one upset (game
won by the visitor) in the six games that were played, but that upset
cannot have occurred in Game 1 or Game 2 else Team B would have
won in only five games. Thus it happened in Game 3, 4, 5, or 6 and
in the first three of those cases it is Team A who benefitted.
Next is the first of several probability-comparison puzzles.

Dishwashing Game
You and your spouse flip a coin to see who washes the dishes each
evening. “Heads” she washes, “tails” you wash.
Tonight she tells you she is imposing a different scheme. You flip
the coin 13 times, then she flips it 12 times. If you get more heads than
she does, she washes; if you get the same number of heads or fewer,
you wash.
Should you be happy?
Solution: If “should you be happy?” means anything at all, it
should mean “are you better off than you were before?”

122
 In All Probability

Here, the easy route is to imagine that first you and your spouse
flip just 12 times each. If you get different numbers of heads, the one
with fewer will be washing dishes regardless of the outcome of the
next flip; so those scenarios cancel. The rest of the time, when you tie,
the final flip will determine the washer. So it’s still a 50-50 proposition,
and you should be indifferent to the change in procedure (unless you
dislike flipping coins).
Now suppose that some serious experimentation with the coin in
question shows that it is actually slightly biased, and comes up heads
51% of the time. Should you still be indifferent to the new decision
procedure? It might seem that you would welcome it now: Heads are
good for you, so, the more flipping, the better.
But re-examination of the above argument shows the opposite.
When you and your spouse flip 12 times each, it’s still equally likely
that you flip more heads than she does, or the reverse. Only if the two
of you flip the same number of heads do you get your 51% advan-
tage from the final flip. Thus, overall, your probability of ducking the
dishwashing falls to somewhere between 50% and 51%.
The analysis suggests a third procedure: You and your spouse al-
ternate flipping the coin until you reach a point where you have both
flipped the same number of times but one of you (the winner) has
flipped more heads. The advantage of this scheme? It’s fair even if the
coin is biased!
Many puzzles that involve comparing two probabilities can be ap-
proached by a technique called coupling, in which two events associ-
ated with different conditions are somehow built into a single exper-
iment for which both events make sense.
Then, if you want to determine which of the two events A and B
is more probable, you can ignore outcomes where they both occur or
neither occurs, and just compare the probability of A happening with-
out B with the probability of B happening without A. This amounts
to comparing the areas of the two crescents in the Venn diagram be-
low.

123
 Mathematical Puzzles

OK, that’s a lot of abstract theory. Let’s see some examples.

Random Judge
After a wild night of shore leave, you are about to be tried by
your U.S. Navy superiors for unseemly behavior. You have a choice
between accepting a “summary” court-martial with just one judge,
or a “special” court-martial with three judges who decide by major-
ity vote.
Each possible judge has the same (independent) probability—
65.43%—of deciding in your favor, except that one officer who would
be judging your special court-martial (but not the summary) is noto-
rious for flipping a coin to make his decisions.
Which type of court-martial is more likely to keep you out of the
brig?

Solution: It’s not that hard to “do the math” on this one and sim-
ply compute your chance of getting off in the special court-martial,
then compare it to your 65.43% chance of surviving the summary
court-martial. (You might save yourself some arithmetic if you replace
65.43% by a variable, say p, and put the number back in later if nec-
essary.)
But let’s try coupling. You may as well suppose officer A would
be on both panels and would vote the same way on each; officer C
is the random one. If you go with the special court-martial you will
regret it precisely when A votes innocent and B and C vote guilty, but
thank your stars if A votes guilty while B and C vote innocent.
Without any calculations you know that these events have the
same probability, just by exchanging the roles of A and B and revers-
ing C’s coinflip. So your two choices are equally good.
Suppose there’s also an option of a “general” court-martial with
five judges, two of whom are coin-flippers. You can apply the coupling
method again, but this time it tells you to choose the five judges if p
(as here) is greater than 12 .
In fact, the Law of Large Numbers tells you that even if 90% of
the judges were coin-flippers, the remaining 10% voting in your favor
with some fixed probability p > 1/2, you could boost your chances
to 99% with enough judges.

124
 In All Probability

Wins in a Row
You want to join a certain chess club, but admission requires that you
play three games against Ioana, the current club champion, and win
two in a row.
Since it is an advantage to play the white pieces (which move first),
you alternate playing white and black.
A coin is flipped, and the result is that you will be white in the first
and third games, black in the second.
Should you be happy?
Solution: You can answer this question by assigning a variable w
to the probability of winning a game as White, and another variable
b < w to the probability of winning as Black; then doing some algebra.
Using coupling, you can get the answer without algebra—even if
the problem is modified so that you have to win two in a row out of
seventeen games, or m in a row out of n. (If n is even, it doesn’t matter
who plays white first; if n is odd, you want to be black first when m
is odd, white first when m is even.)
The coupling argument in the original 2-out-of-3 puzzle goes like
this. Imagine that you are going to play four games against Ioana,
playing white, then black, then white, then black. You still need to
win two in a row, but you must decide in advance whether to discount
the first game, or the last.
Obviously turning the first game into a “practice game” is equiv-
alent to playing BWB in the original problem, and failing to count
the last game is equivalent to playing WBW, so the new problem is
equivalent to the old one.
But now the events are in the same experiment. For it to make
a difference which game you discounted, the results must be either
WWLX or XLWW. In words: If you win the first two games, and
lose (or draw) the third, you will wish that you had discounted the
last game; if you lose the second but win the last two, you will wish
that you had discounted the first game.
But it is easy to see that XLWW is more likely than WWLX. The
two wins in each case are one with white and one with black, so those
cases balance; but the loss in XLWW is with black, more likely than
the loss in WWLX with white. So you want to discount the first game,
that is, start as black in the original problem.
A slightly more challenging version of this argument is needed if
you change the number of games played, and/or the number of wins
needed in a row.

125
 Mathematical Puzzles

Split Games
You are a rabid baseball fan and, miraculously, your team has won
the pennant—thus, it gets to play in the World Series. Unfortunately,
the opposition is a superior team whose probability of winning any
given game against your team is 60%.
Sure enough, your team loses the first game in the best-of-seven
series and you are so unhappy that you drink yourself into a stupor.
When you regain consciousness, you discover that two more games
have been played.
You run out into the street and grab the first passer-by. “What
happened in games 2 and 3 of the World Series?”
“They were split,” he says. “One game each.”
Should you be happy?

Solution: In my experience, presented with this puzzle, about half

of respondents answer “Yes—if those games hadn’t been split, your
team would probably have lost them both.”
The other half argue: “No—if your team keeps splitting games,
they will lose the series. They have to do better.”
Which argument is correct—and how do you verify the answer
without a messy computation?
The task at hand is to determine, after hearing that games 2 and
3 were split, whether you are better or worse off than before. In other
words, is your team’s probability of winning the series better now,
when it needs three of the next four games, than it was before, when
it needed four out of six?
Computing and comparing tails of binomial distributions is messy
but not difficult. Before the news, your team needed to win 4, 5, or
6 of the next six games. (Wait, what if fewer than seven games are
played? As in both Home-Field Advantage and Who Won the Series?, it
costs nothing to imagine that all seven games are played no matter
what.)
The probability that your team wins exactly four of six games is
“6 choose 4” (the number of ways that can happen) times 0.44 (the
probability that your team wins a particular four games) times 0.62
(the probability that the other guys win the other two). Altogether,
the probability that your team wins at least four of six is
( ) ( ) ( )
6 6 6
· 0.44 · 0.62 + · 0.45 · 0.6 + · 0.46
4 5 6
= 15 · 24 · 32 /56 + 6 · 25 · 3/56 + 26 /56 = 112/625.

126
 In All Probability

After the second and third games are split, your team needs at least
three of the remaining four. The probability of winning is now:
( ) ( )
4 4
· 0.4 · 0.6 +
3
· 0.44
3 4

= 4 · 23 · 3/54 + 24 /54 = 112/625(!)

So you should be exactly indifferent to the news! The two arguments
(one backward-looking, suggesting that you should be happy, the
other forward-looking, suggesting that you should be unhappy) seem
to have canceled each other precisely. Can this be a coincidence? Is
there any way to get this result “in your head”?
Of course there is—by coupling. To set this up a little imagination
helps. Suppose that, after game 3, it is discovered that an umpire who
participated in games 2 and 3 had lied on his application. There is a
movement to void those two games, and a counter-movement to keep
them. The commissioner of baseball, wise man that he is, appoints a
committee to decide what to do with the results of games 2 and 3; and
in the meantime, he tells the teams to keep playing.
Of course, the commissioner hopes—and so do we puzzle-
solvers—that by the time the committee makes its decision, the question will
be moot.
Suppose five(!) more games are played before the committee is
ready to report. If your team wins four or five of them, it has won
the series regardless of the disposition of the second and third game
results. On the other hand, if the opposition has won three or more,
they have won the series regardless. The only case where the commit-
tee can make a difference is where your team won exactly three of
those five new games.
In that case, if the results of games 2 and 3 are voided, one more
game needs to be played; your team wins the series if they win that
game, which happens with probability 2/5.
If, on the other hand, the committee decides to count games 2
and 3, the series ended before the last game, and whoever lost that
game is the World Series winner. Sounds good for your team, no?
Oops, remember that your team won three of those five new games,
so the probability that the last game is among the two they lost is again
2/5. ♡

127
 Mathematical Puzzles

Angry Baseball
As in Split Games, your team is the underdog and wins any given game
in the best-of-seven World Series with probability 40%. But, hold on:
This time, whenever your team is behind in the series, the players get
angry and play better, raising your team’s probability of winning that
game to 60%.
Before it all begins, what is your team’s probability of winning the
World Series?

Solution: It would be a pain in the neck to work out all the possi-
ble outcomes and their probabilities, but coupling again comes to our
rescue. Let k be the number of the first game played after the last time
the teams were tied. Since the teams were tied 0-0 at the start, k could
be 1, but it could also be 3, 5, or 7. (It turns out, you don’t care!)
Suppose team X wins game k. Let us represent the winners of the
rest of the games by a sequence of X’s and Y’s. Since there are no
more ties, X must have won the series, each win by X having had
probability 40% and each win by Y, 60%.
If it was Team Y that won game k, the same reasoning applies and
the same probability attaches to the subsequent win-sequence except
with X’s and Y’s exchanged. In effect, we couple win-sequences after
X wins game k with the complementary win-sequences after Y wins
game k.
We conclude that the probability of winning the series for Team X
is exactly the probability that Team X wins game k. That probability
is 40% when Team X is your team. Since that does not depend on the
value of k, your team’s probability of winning the series is 40% from
the start.
Note that unlike Split Games, this puzzle works with any game-
winning probability p (changing to 1−p when the team is behind).
The answer is then p.
Let’s move on to (American) football. Are you ready to use prob-
ability theory to be an effective coach?

Two-point Conversion
You, coach of the Hoboken Hominids football team, were 14 points
behind your rivals (the Gloucester Great Apes) until, with just a
minute to go in the game, you scored a touchdown. You have a choice

128
 In All Probability

of kicking an extra point (95% success rate) or going for two (45% suc-
cess rate). Which should you do?
Solution: You must assume the Hominids will be able to score a
second touchdown; and it’s reasonable to postulate that if the game
goes to overtime, either team is equally likely to win it.
Whatever you do, if it doesn’t work you’ll need to go for two points
after the second TD, winning with probability 0.45 · 12 = 0.225.
If you do succeed with the first conversion, you’ll want to go for
just one the next time, winning with probability 12 · 0.95 = 0.475 if
you had kicked last time but a full 0.95 if you had gotten two extra
points last time.
So, altogether (assuming that you did get that second touchdown)
your probability of winning with the “kick after first TD” option is

0.95 · 0.475 + 0.05 · 0.225 = 0.4625,

while your probability of winning if you went for two after the first
TD is
0.45 · 0.95 + 0.55 · 0.225 = 0.55125,
so going for two is much better.
It might be easier to see why this is so if you instead imagine that
the kick is a sure thing while the two-point conversion has probability
of success 12 . Then if you kick, you are headed for overtime for sure
(given, of course, the second touchdown) where you win with prob-
ability 12 . If you go for two and succeed you win for sure, so already
you get probability 12 of winning the game; but if you miss it ain’t
over. Altogether the two-point conversion strategy wins 5/8 of the
time.
So why do coaches frequently try to kick the extra point in these
situations? Do they think missing the conversion will demoralize their
team and make the second touchdown unlikely?
My own theory is that coaches tend to make conservative deci-
sions, making it harder for their managers to point at a coach’s de-
cision that cost the game. In this situation, for example, if the team
misses both two-point conversions, blame is likely to fall on the coach;
but if they kick both extra points and lose in overtime, most of the
blame will fall on whichever player screwed up in overtime.

129
 Mathematical Puzzles

Random Chord
What is the probability that a random chord of a circle is longer than
a side of an equilateral triangle inscribed in the circle?
Solution: Call a chord “long” if it is longer than the side of our
inscribed triangle.
Suppose we choose our random chord by fixing some direction—
horizontal is as good as any. Then we can draw a vertical diameter
in our circle, pick a point on it uniformly at random, and draw the
horizontal chord through that point.
In the left-hand third of the figure below, the inscribed triangle
has been drawn with its base down, so that the diameter we just drew
is perpendicular to the base. We claim the triangle’s base is exactly
halfway between the circle’s center and the bottom point of the diam-
eter; one way to see this is to draw another inscribed triangle, this one
upside-down, so that the two triangles make a “star of David.” The
hexagon in the middle of the star can be broken into six equilateral
triangles and once we do that, it is evident that the part of our diam-
eter that yields a long chord is the same length as the rest. Thus the
probability that the random chord is long is 12 .

But . . . we could instead fix one end of the chord, say at the trian-
gle’s apex, and choose its angle uniformly at random, as in the middle
third of the figure. Since the angle made by the triangle is 60◦ and our
random angle is between 0◦ and 180◦ , we deduce that the probability
the chord is long is 13 .
Wait, here’s yet another way to solve the problem. Except for di-
ameters (which have probability zero), every chord has a different
midpoint and every point inside the circle, except for the center, is
the midpoint of a unique chord. So let’s pick a point inside the cir-
cle uniformly at random, and use the chord with that midpoint. It’s
easy to see that the chord will be long exactly if our point is closer to
the center of the circle than the midpoint of a side of the triangle (see

130
 In All Probability

right-hand third of figure). Thus, we get a long chord if and only if

the chosen point lies inside the circle that’s inscribed in our triangle.
But this circle has half the radius, thus one-quarter the area, of the
original circle; therefore the probability that our chord is long is 14 .
What gives? What’t the right answer, 12 , 13 , 41 or something else?
The answer is that the concept of “uniform randomness” for an
object chosen from an infinite set is not automatically defined. True,
for a point chosen from a geometric shape (e.g., an interval or a poly-
gon) we generally take uniformity to mean that the probability that
the point lands in a particular subset is proportional to the measure
(e.g., length or area) of that subset.
For chords, we don’t have a standard notion of uniform random-
ness and as we’ve just seen, different attempts at defining uniformity
can lead to different answers.

Random Bias
Suppose you choose a real number p between 0 and 1 uniformly at
random, then bend a coin so that its probability of coming up heads
when you flip it is precisely p. Finally you flip your bent coin 100
times. What is the probability that after all this, you end up flipping
exactly 50 heads?

Solution: This puzzle again calls for a form of coupling; the idea in
this case is to couple the coinflips even though they involve different
coins, depending on the choice of p. To do this we choose, in advance,
a uniformly random threshold pi ∈ [0, 1] for the ith coinflip, indepen-
dently for each i. Then, after p is chosen, the ith coinflip is deemed
to be heads just when pi < p.
Now we need only observe that getting 50 heads is tantamount to
p being the middle number, that is, the 51st largest, of the 101 values
p1 , p2 , . . . , p100 , p. Since these numbers are all uniform and indepen-
dent, their order is uniformly random and the probability that p ends
up in the middle position is exactly 1/101.

Coin Testing
The Unfair Advantage Magic Company has supplied you, a magi-
cian, with a special penny and a special nickel. One of these is sup-
posed to flip “Heads” with probability 13 , the other 41 , but UAMCO
has not bothered to tell you which is which.

131
 Mathematical Puzzles

Having limited patience, you plan to try to identify the biases by

flipping the nickel and penny one by one until one of them comes up
Heads, at which point that one will be declared to be the 1/3-Heads
coin.
In what order should you flip the coins to maximize the probabil-
ity that you get the correct answer, and at the same time to be fair, that
is, to give the penny and the nickel the same chance to be designated
the 1/3-Heads coin?

Solution: From the point of view of accuracy, at any point when

you’ve flipped each coin the same number of times, it doesn’t matter
which one you flip next. At any point when you’ve flipped the penny
more times than the nickel, you need to flip the nickel next, and vice-
versa.
Thus, for the purpose of accuracy, the order doesn’t matter as long
as for every i, flips 2i−1 and 2i are of different coins.
However, if you want a fair test—one equally likely to damn either
coin—you need to be more careful. Imagine that you make an extra
flip when the first Head turns up at an odd flip, and declare the result
to be inconclusive if the succeeding flip (necessarily, of the other coin)
is again a Head. This would then certainly be fair to the two coins,
but it may mean more flipping. And if you’re reading this book, you’re
likely to be the kind of person that would rather spend an hour figuring
something out then an extra minute actually doing it.
It follows from considering this tie scenario that the bias toward
the penny (say) is the probability that the “tie” happens with the
penny having flipped Heads first, and the nickel immediately after;
and vice-versa.
To arrange these biases so that they cancel, we need to order the
flips so that if a tie “would have” occurred, it is equally like to have
been with the penny or the nickel coming up Heads first.
Suppose we flip the penny first, that is, we begin “PN”. The prob-
ability of getting Heads already on the first two flips is 13 · 14 = 12 1
.
What’s the probability that the modified game ends in a tie? You
will be flipping in pairs—odd flip, then even flip—until you either get
HT, TH, or HH, and the tie occurs in the last of these cases, which
has probability ( 13 · 14 )/( 23 · 14 + 34 · 31 + 31 · 41 ) = ( 12
1 6
)/( 12 ) = 61 .
Therefore, in order to cancel the penny’s first-flip advantage, all
the rest of the odd flips must be performed by the nickel!
In conclusion, the flip order should either be PNNPNPNPNPNP...
or NPPNPNPNPNPN... .

132
 In All Probability

Coin Game
You and a friend each pick a different heads–tails sequence of length
4 and a fair coin is flipped until one sequence or the other appears;
the owner of that sequence wins the game.
For example, if you pick HHHH and she picks TTTT, you win if
a run of four heads occurs before a run of four tails.
Do you want to pick first or second? If you pick first, what should
you pick? If your friend picks first, how should you respond?

Solution: At first it seems as if this must be a fair game; after all, any
particular sequence of four flips has the same probability of appearing
1
(namely, 16 ) as any other. Indeed, if the coin is flipped four times and
then the game begins anew, nothing of interest arises. But because
the winning sequence could start with any flip, including the second,
third or fourth, significant imbalances appear.
For example, if you pick HHHH (a miserable choice) as your se-
quence and your friend counters with THHH, you will lose unless the
1
game begins with four heads in a row (only a 16 probability). Why?
Because otherwise the first occurrence of HHHH will be preceded by
a tail, thus your friend will get there first with her THHH.
How can you determine how to play optimally? For that you need
a formula for the probability that string B beats string A. There is, in
fact, a very nice one, discovered by the late, indominable John Horton
Conway. Here’s how it works.
We use the expression “A·B” to denote the degree to which A and
B can overlap when A begins first (or they begin simultaneously). We
measure this with a four-digit binary number x4 x3 x2 x1 where xi is 1
if the first i letters of B match the last i of A, otherwise 0. For example,
HHTH·HTHT = 0101(binary) = 5 since H[HTH]T gives an overlap
of 3 while HHT[H]THT gives an overlap of 1. The left-most digit, x4 ,
can only be 1 if A = B; thus A·B is equal to or greater than 8 if and
only if A = B.
Now if your friend has chosen A as her sequence, how should you
choose B? You want B to have low waiting time, thus B·B should be
as low as possible. You want B·A to be high, so that you maximize
your probability of sneaking in ahead of your friend; and you want
A·B to be low to minimize the probability that she sneaks in ahead of
you. The formula says that the odds that B beats A are (A·A − A·B):
(B·B − B·A).
For example, suppose your friend picks A = THHT as her se-
quence (then A·A = 9). Your best choice is (always) to tack an H

133
 Mathematical Puzzles

or T to the front of her choice and drop the last letter; this ensures
that B·A is at least 4. If you pick B = HTHH you get B·B = 9, B·A
= 4, and A·B = 2, so Conway’s formula says the odds in your favor
are (9 − 2):(9−4) = 7:5. If instead you pick B = TTHH you get B·B
= 8, B·A = 4, and A·B = 1, and the formula says the odds in your
favor are (9 − 1):(8 − 4) = 2:1. So in this case B = TTHH is the better
choice.
Was A = THHT a poor choice by your friend? Actually, it was
among the best. Anything your friend chose would leave her at a 2:1
disadvantage or worse, with best play by you.
Conway’s formula continues to work beautifully for strings of
length k, for any integer k ≥ 2, and also to non-binary randomiz-
ers such as dice. In the case of dice, the overlap vectors still consist of
zeroes and ones, but are now interpreted base 6.

Sleeping Beauty
Sleeping Beauty agrees to the following experiment. On Sunday she
is put to sleep and a fair coin is flipped. If it comes up Heads, she is
awakened on Monday morning; if Tails, she is awakened on Monday
morning and again on Tuesday morning. In all cases, she is not told
the day of the week, is put back to sleep shortly after, and will have
no memory of any Monday or Tuesday awakenings.
When Sleeping Beauty is awakened on Monday or Tuesday,
what—to her—is the probability that the coin came up Heads?

Solution: This puzzle, introduced by philosopher Adam Elga in

2001, became the subject of major controversy in the philosophy com-
munity over the next 20 years. Many were persuaded by the following
argument: On Sunday SB (Sleeping Beauty) knows that the probabil-
ity of Heads is 12 . She knows she will be awakened, so when she is
awakened (on Monday or Tuesday), she has no new information and
thus no reason to change her mind. So the answer is 12 .
The flaw in this reasoning is that there is indeed new information:
that she is awake now. Many arguments show that the correct answer
is 13 ; one very nice example, offered by philosophers Cian Dorr and
Frank Arntzenius, runs as follows.
Suppose that SB is awakened both days, regardless of the coin,
but if the coin comes up Heads, then on Tuesday, 15 minutes after
she is awakened, she will be told “It’s Tuesday and the coin came up
Heads.” She knows all this in advance.

134
 In All Probability

When she is awakened, it’s equally likely to be Monday or Tues-

day (by symmetry) and equally likely to be Heads or Tails (again
by symmetry). Thus, the four events “Monday, Tails,” “Monday,
Heads,” “Tuesday, Tails,” and “Tuesday, Heads” are all equally
likely.
When, 15 minutes later, SB is not told that it’s “Tuesday, Heads”
she rules out that last option and the rest remain equally likely. ♡
You might, on the other hand, be more persuaded by a frequency
argument. If the experiment is run 100 times, you’d expect about
150 Monday/Tuesday awakenings, about 13 of which would be after
Heads was flipped.
The last puzzle will lead us to our theorem for this chapter.

Leading All the Way

Running for local office against Bob, Alice wins with 105 votes to
Bob’s 95. What is the probability that as the votes were counted (in
random order), Alice led the whole way?
Solution: Of course, what we’d really like to do is solve this prob-
lem in general (when Alice has won with a votes to Bob’s b, where a
and b are arbitrary positive integers with a > b).
The answer is given by what is often known as Bertrand’s Ballot
Theorem (although it was apparently first proved by W. A. Whitworth
in 1878, not J. L. F. Bertrand in 1887).

135
 Mathematical Puzzles

Theorem. Let a > b > 0 and let x = (x1 , x2 , . . . , xa+b ) be a uniformly

random string of 1’s and −1’s, containing a 1’s and b −1’s. For each k be-
tween 1 and a+b, let sk be the sum of the first k elements of x, that is,

∑
k
sk = xk .
i=1

Then the probability that s1 , s2 , . . . , sa+b are all positive is (a−b)/(a+b).

Let’s first check that this statement tells us what we want to know.
The string x codes the ballot-counting process; each 1 represents a
vote for Alice, each −1 a vote for Bob. The sum sk tells us by how
much Alice leads after the kth vote has been counted; for example,
if s9 = −3 we conclude that Alice was actually behind by three votes
after the ninth vote was counted.
If we apply the theorem to the puzzle above, we get the answer
(105 − 95)/(105 + 95) = 1/20 = 5%. So how can we prove the
theorem?
There’s a famous and clever proof that uses reflection of random
walk, but doesn’t really explain why the probability turns out to be
the vote difference divided by the total vote. Instead, consider the fol-
lowing argument.
One way to choose the random order in which the ballots are
counted is first to place the ballots randomly in a circle, then to choose
a random ballot in the circle and start counting (say, clockwise)
from there. Since there are a+b ballots, there are a+b places to start,
and the claim is that no matter what the circular order is, exactly a−b
of those starting points are “good” in that they result in Alice leading
all the way.
To see this we observe that any occurrence of 1,−1 (i.e., a vote for
Alice followed immediately by a vote for Bob in clockwise order) can
be deleted without changing the number of good starting positions.
Why? First, you can’t start with that 1 or −1, since if you start with
the 1 the candidates are tied after the second vote, and if you start
with the −1, Alice actually falls behind. Secondly, the 1,−1 pair has
no influence upon the the eligibility of any other starting point, since
all it does is raise the sum (Alice’s lead) by one and then lower it again.
So we can keep deleting 1,−1’s until all that’s left is a−b 1’s, each
of which is obviously a good starting point. We conclude that all along
there were a−b good starting points, and we are done! ♡

136
 11. Working for the
System
There’s no getting around it—many puzzles respond only to trying
out ideas until you find one that works. Even so, there are good ways
and bad ways to go about this process. Some people have a tendency
to circle back and try the same answer many times.
How can you avoid doing that? By classifying your attempts.
Make some choice and stick with it, until it either yields a solution
or runs out of steam; if the latter, you can then rule out one possibil-
ity, and progress has been made.
There are other benefits to this besides avoiding redundancy: For
example, you might actually see what’s going on and jump to the so-
lution!

No Twins Today
It was the first day of class and Mrs. O’Connor had two identical-
looking pupils, Donald and Ronald Featheringstonehaugh (pro-
nounced “Fanshaw”), sitting together in the first row.
“You two are twins, I take it?” she asked.
“No,” they replied in unison.
But a check of their records showed that they had the same parents
and were born on the same day. How could this be?

Solution: We have begun with a puzzle that might not seem to re-
spond to systematic search; you either get it or you don’t. But these
days many people train themselves to “think out of the box” and for-
get to first try thinking in the box. Accept that the boys are genetically
identical and born to the same parents at the same time, and ask how
this can happen if they are not twins. Then, at least, it might occur to
you that there was an additional simultaneous sibling (or more!).
The most likely possibility is that Donald and Ronald are triplets;
the third (Arnold, perhaps?) is in another class.

137
 Mathematical Puzzles

Natives in a Circle
An anthropologist is surrounded by a circle of natives, each of whom
either always tells the truth or always lies. She asks each native
whether the native to his right is a truth-teller or a liar, and from their
answers, she is able to deduce the fraction of liars in the circle.
What fraction is it?
Solution: Observe that if all truth-tellers are changed to liars and
vice-versa, none of answers would change. Therefore, if the fraction
x can be determined, it must have the property that x = 1−x, thus
x = 1/2.
It remains to check that there is a set of answers from which the an-
thropologist can deduce that half the natives are truth-tellers. That re-
quires, of course, that there be an even number of natives in the circle.
One possibility is that every native says his right-hand neighbor is a
liar, in which case natives must alternate between liar and truth-teller.

Poker Quickie
What is the best full house?
(You may assume that you have five cards, and you have just one
opponent who has five random other cards. There are no wild cards.
As a result of the Goddess of Chance owing you a favor, you are en-
titled to a full house, and you get to choose whatever full house you
want.)

Solution: All full houses with three aces are equally good, be-
cause there can only be one such hand in a deal from a single deck.
But there are other hands that can beat them: any four-of-a-kind, of
which there are always 11 varieties possible, and more relevantly any
straight flush. Since AAA99, AAA88, AAA77, and AAA66 kill the
most straight flushes (16—each ace kills only two, but each spot card
kills five) they are the best full houses. Note that AAA55 doesn’t quite
make the list, because one of the 5’s must be in the same suit as an
ace and the A2345 straight flush in that suit is thus doubly covered.
If you greedily insist on AAAKK, there are 40 − 9 = 31 possible
straight flushes that can beat you—even more if you haven’t got the
four suits covered—instead of just 40 − 16 = 24.
Although the argument above only works for head-to-head poker
(two players), the result has in fact been verified by computer for any
number of players.

138
 Working for the System

Fewest Slopes
If you pick n random points in, say, a disk, the pairs of points among
them will with probability 1 determine n(n−1)/2 distinct slopes. Sup-
pose you get to pick the n points deliberately, subject to no three being
collinear. What’s the smallest number of distinct slopes they can de-
termine?

Solution: Three points, since they are not permitted to lie on one
line, determine a triangle with three slopes. A little experimentation
will convince you that with four, you can’t have any fewer than the
four slopes you get from a square.
In fact the vertices of a regular n-sided polygon determine just n
different slopes. If n is odd, the polygon’s sides already have n different
slopes, but every diagonal is parallel to some side. If n is even, the sides
exhibit only n/2 different slopes but now the diagonals that connect
two vertices an even number of edges apart are not parallel to any
side, but they form n/2 parallel classes; so again we get n different
diagonals.
Can you do any better than the vertices of a regular n-gon? No.
Let X be any configuration of n points in the plane, no three on a line,
and let P be the southernmost point of X. Then you already have n−1
lines through P , each determining a different angle between −90◦ and
+90◦ (relative to north = 0◦ ). The line through the two points that give
the smallest and largest such angles gives you your nth slope.

Two Different Distances

Find all configurations of four points in the plane that determine only
two different distances. (Note: There are more of these than you prob-
ably think!)

Solution: It may pay to think of the points as vertices of a graph,

all of whose edges are the same length, and all of whose non-edges
are the same length. If there is a “3-clique” in the graph, that is, three
vertices of which any two are adjacent, then three of our points—call
them, say, A, B, and C—form an equilateral triangle.
The fourth point, D, can’t be connected to all the others, since in
the plane we can’t have four points determining only one distance. If
it is adjacent to two of the others, it forms another equilateral triangle
and we get the upper-left configuration in the figure.

139
 Mathematical Puzzles

If D is connected to one of the others, say A, it must be equidistant

from B and C; thus it must either stick out from A directly away from
the triangle, or directly toward it, giving the other two configurations
on the top row. Finally, if D is adjacent to none of the three it must be
equidistant from all three, thus in the middle of the triangle; this gives
the lower-left configuration.
What if there’s no triangle of edges (or non-edges) in the graph?
If there’s a 4-cycle (but no triangle), its diagonals must be of equal
length, thus the points are the corners of a square.
If there’s no triangle or 4-cycle in either the graph or its comple-
ment, the graph must be precisely a path three edges in length (hence
its complement is, too). This is the graph you get by deleting one edge
of a 5-cycle, and corresponds to the configuration you get by deleting
one vertex of a regular pentagon.

First Odd Number

What’s the first odd number in the dictionary? More specifically, sup-
pose that every whole number from 1 to, say, 1010 is written out in for-
mal English (e.g., “two hundred eleven,” “one thousand forty-two”)
and then listed in dictionary order, that is, alphabetical order with
spaces and hyphens ignored. What’s the first odd number in the list?

140
 Working for the System

Solution: In dictionary order, one word precedes another precisely

when the first letter in which the two words differ is, in the first word,
earlier in the alphabet. Spaces and punctuation are ignored, and if
there is no differing letter, the shorter word appears first.
So finding the dictionary-first word in a set of words is just a mat-
ter of carefully and systematically considering the successive words
involved in the description of a number.
The earliest actual digit is “eight,” and the earliest odd digit
“five.” We don’t need to consider any other digits, but other pos-
sibly useful words that appear in numbers include “billion,” “eigh-
teen,” “eighty,” “hundred,” “million,” and “thousand.” Our earliest
odd number must begin with a digit so should start with “eight bil-
lion.” After that, “eighteen” is the best we can do, and proceeding
along these lines, we eventually get the answer 8,018,018,885: “eight
billion, eighteen million, eighteen thousand, eight hundred eighty-
five.”
A little more work will get you to the first prime number in the
dictionary, 8,018,018,851.
Annoyingly, “dictionary order” is (according to Wikipedia) sub-
tly different from “alphabetical order,” and the difference makes a
difference. The issue is that in alphabetical order, which is used with
people’s names, spaces typically are not ignored, but instead precede
all other characters. Thus, alphabetically, “eight hundred” precedes
“eighteen,” and the first odd number in alphabetical order becomes
8,808,808,885. The first prime would in that case be the very next
odd number in alphabetical order, namely, 8,808,808,889.

Measuring with Fuses

You have on hand two fuses (lengths of string), each of which will
burn for exactly 1 minute. They do not, however, burn uniformly
along their lengths. Can you nevertheless use them to measure 45
seconds?

Solution: Simultaneously light both ends of one fuse and one end
of the other; when the first fuse burns out (after half a minute),
light the other end of the second. When it finishes, 45 seconds have
passed. ♡
If you allow midfuse ignition and arbitrary dexterity, you can do
quite a lot with fuses. For example, you can get 10 seconds from a sin-
gle 60-second fuse by lighting at both ends and at two internal points,

141
 Mathematical Puzzles

then lighting a new internal point every time a segment finishes; thus,
at all times, three segments are burning at both ends and the fuse ma-
terial is being consumed at six times the intended rate.
Bit of a mad scramble at the end, though. You’ll need infinitely
many matches to get perfect precision.

King's Salary
Democracy has come to the little kingdom of Zirconia, in which the
king and each of the other 65 citizens has a salary of one zircon. The
king can not vote, but he has power to suggest changes—in particu-
lar, redistribution of salaries. Each person’s salary must be a whole
number of zircons, and the salaries must sum to 66. Each suggestion
is voted on, and carried if there are more votes for than against. Each
voter can be counted on to vote “yes” if his or her salary is to be in-
creased, “no” if decreased, and otherwise not to bother voting.

The king is both selfish and clever. What is the maximum salary
he can obtain for himself, and how many referenda does he need to
get it?
Solution: There are two key observations: (1) that the king must
temporarily give up his own salary to get things started, and (2) that
the game is to reduce the number of salaried citizens at each stage.
The king begins by proposing that 33 citizens have their salaries
doubled to 2 zircons, at the expense of the remaining 33 citizens (him-
self included). Next, he increases the salaries of 17 of the 33 salaried
voters (to 3 or 4 zircons) while reducing the remaining 16 to no salary
at all. In successive turns, the number of salaried voters falls to 9, 5, 3,
and 2. Finally, the king bribes three paupers with one zircon each to
help him turn over the two big salaries to himself, thus finishing with
a royal salary of 63 zircons.

142
 Working for the System

It is not difficult to see that the king can do no better at any stage
than to reduce the number of salaried voters to just over half the pre-
vious number; in particular, he can never achieve a unique salaried
voter. Thus, he can do no better than 63 zircons for himself, and the
seven rounds above are optimal. ♡
More generally, if the original number of citizens is n, the king
can achieve a salary of n−3 zircons in k rounds, where k is the least
integer greater than or equal to log2 (2n−4).

Packing Slashes
Given a 5 × 5 square grid, on how many of the squares can you draw
diagonals (slashes or backslashes) in such a way that no two of the
diagonals meet?

Solution: You can get 15 easily (see figure below) using rows (or
columns) 1, 3, and 5, with all slashes or all backslashes on each line,
or by nested Ls. Is it possible to do any better than that?
There are 6 × 6 = 36 vertices (cell corners) in the grid and each
diagonal accounts for two, so you certainly will not be able to fit in
more than 18 diagonals. But you can see that, for example, it won’t
be possible to use all 12 corners of the top row of squares.

In fact, come to think of it, every diagonal touching one of the

20 outer vertices must also touch one of the 12 outer vertices of the
inside 3×3 grid—except for those in a corner cell that miss the corner
vertex of the big square, for example, a backslash in the upper right
cell. But there can’t be backslashes in both cells neighboring the upper
right corner cell, so putting a backslash in the upper right corner will
cause at least one other outside vertex to be missed. In conclusion, no
matter what we do, at least four outside vertices will be unmatched,
and we’re down to an upper bound of (36−4)/2 = 16 diagonals.

143
 Mathematical Puzzles

When we try to achieve 16, we notice that using up all the outer
vertices of the inside 3 × 3 grid to take care of outer vertices won’t
work—that leaves the four corners of the center cell unmatched, and
a diagonal in that cell takes care of only two, taking us back to 15
diagonals. We’ll need to leave at least two of the four corners of the
big square unmatched, and use two or four unmatched outer vertices
of the inside 3 × 3 grid to cover the corners of the central cell. A little
experimentation shows that this works only if you miss all the corners
of the big square, producing the picture below—the unique solution
to the puzzle, up to reflection.

Unbroken Lines
Can you re-arrange the 16 square tiles below into a 4 × 4 square—
no rotation allowed!—in such a way that no line ends short of the
boundary of the big square?

Solution: Notice that there are 24 = 16 different tiles that can be

made by choosing, for each line-type, whether to include that type or
not. Each tile is represented once in the puzzle.

144
 Working for the System

This is, again, a matter of intelligent search. The horizontal lines

must all be in two rows, the vertical lines in two columns; it costs
nothing to assume those are the top two rows and the leftmost two
columns, and already you have severe constraints.
The lengths of the SW–NE diagonals must sum to 8 squares
worth, and thus there are only a few possibilities: a 4 and two 2s,
a 4, a 3 and a 1, etc.
There are in fact several ways to do it of which one is illustrated
below. Curiously, all ways of doing it work on the torus as well as
on the square—in other words, if you imagine that the square wraps
around top-to-bottom and left-to-right, no line ends at all; all are, ul-
timately, loops.

[See Notes & Sources for the composer and inspiration behind this
and the next puzzle.]
Similar, but harder:

Unbroken Curves
Can you re-arrange the 16 square tiles below into a 4 × 4 square—
no rotation allowed!—in such a way that no curve ends short of the
boundary of the big square?

145
 Mathematical Puzzles

Solution: Again, there are 24 = 16 different tiles that can be made

by choosing, for each corner of the tile, whether to include a quarter-
circle at that corner or not. Each tile is represented once in the puzzle.
For this puzzle, the best route to a solution seems to be to guess
how the squares with the most quarter-circles fit together and (sys-
tematically!) chase down the possibilities.
Not all solutions to this one work on the torus as well as the
square, but the one illustrated below does work on both.

Conway's Immobilizer
Three cards, an ace, a deuce, and a trey, lie face-up on a desk in some
or all of three marked positions (“left,” “middle,” and “right”). If they
are all in the same position, you see only the top card of the stack; if
they are in two positions, you see only two cards and do not know
which of them is concealing the third card.
Your objective is to get the cards stacked on the left with ace on
top, then deuce, then trey on bottom. You do this by moving one card
at a time, always from the top of one stack to the top of another (pos-
sibly empty) stack.
The problem is, you have no short-term memory and must, there-
fore, devise an algorithm in which each move is based entirely on
what you see, and not on what you last saw or did, or on how many
moves have transpired. An observer will tell you when you’ve won.
Can you devise an algorithm that will succeed in a bounded number
of steps, regardless of the initial configuration?

Solution: Since there aren’t that many things you can do, it’s tempt-
ing to try some arbitrary set of rules and see if they work. Actually,
there’s a surprising number of different algorithms possible with the
above constraints—more than a quintillion. So some reasoning will
be needed. But it’s tricky to design an algorithm that makes progress,

146
 Working for the System

avoids cycling, and doesn’t do something stupid when it’s about to

win.
For example, it looks obvious to try to win by putting the A on
the 2 if you see 2,A,– or 2,–,A. Then if you see three cards with the
3 on the left, you can win in two moves by putting the 2 on the 3. So
far, so good.
Suppose the 2 is exposed on the left but no other card is seen; then
we may as well move the 2 to the center. Now, if you see 3,2,–, it’s
tempting to put the 2 on the 3. But if the A is hiding under the 3, you
will get caught in a loop, stuck forever going back and forth between
2,–,– and 3,2,–. So you’ll need to move the 2 to the right from 3,2,–.
But then you can’t put 2 on 3 from 3,–,2 without getting caught in a
loop of order 3. So you’ll have to move the 3 from that position. Darn!
It’s beginning to look like we should be filling holes by moving
cards in one direction, say to the right, to avoid loops. So let’s agree
that when we see only 2 cards, we move a card to the right (around the
corner from R to L, if necessary) to an open slot; the only exceptions
will be 2,A,– and 2,–,A where we place A on 2 to try to win.
Moving cards to the right will spread out the cards in two moves
at most, so we can see them. Seeing only one stack, sending the top
card to the left will then also spread the cards in two moves. It remains
only to decide what to do when we see all three cards, but the 3 is not
on the left. Clearly we can’t move a card left; it would come back
immediately. So we move a card to the right; which one? After two
moves whichever card we covered will be re-exposed, with the other
two switched.
Having got this far, several successful algorithms can now be de-
scribed. Here are the rules for one of them, in priority order:

1. Seeing 2,1,−, place 1 on 2;

2. Otherwise, seeing two cards, move a card right (around the cor-
ner if necessary) to the open place;
3. Seeing just one card, move that card to the left;
4. Seeing three cards, move the middle-rank card to the right.

This same algorithm works for any number of cards, numbered 1

through n. It takes quadratic time in the worst or random case, that is,
approximately a constant times n2 steps. A computer scientist would
conclude that a list of items can be sorted with just three LIFO (last-in,
first-out) stacks and no memory.

147
 Mathematical Puzzles

Here’s a second solution that’s slower but has a compensating

feature:

1. Seeing just one card, move a card to the right.

2. Seeing two cards:
(a) If the left is empty, move the center card to the right stack;
(b) If the center is empty, move the right card to the left stack;
(c) If the right is empty, move the left card to the right stack.
3. Seeing three cards, let the numbers be x, y, and z from left to
right.
(a) If z is the largest or smallest, move the smaller of x and y
onto z;
(b) If z is the middle, move the larger of x and y onto z.

This algorithm is cubic, requiring (2/3)n3 +(1/2)n2 −(7/6)n steps

if it is started with the cards stacked in reverse order on the left. Once
it gets started, it never has more than one card in the center. Thus,
we see that sorting can be done with only two LIFO stacks, a single
register, and no memory! If you can beat that, let me know.

Seven Cities of Gold

In 1539 Friar Marcos de Niza returned to Mexico from his travels
in what is now Arizona, famously reporting his finding of “Seven
Cities of Gold.” Coronado did not believe the “liar friar” and when
subsequent expeditions came up empty, Coronado gave up the quest.
My (unreliable) sources claim that Coronado’s reason for disbe-
lieving de Niza is the latter’s claim that the cities were laid out in the
desert in a manner such that of any three of the cities, at least one pair
were exactly 100 furlongs apart.
Coronado’s advisors told him no such pattern of points existed.
Were they right?

Solution: Putting ourselves in the position of Coronado’s advisors,

we need to decide whether there’s a way to position seven points in
the plane in such a way that among any three, there are at least two
that are a specified distance apart.
What could we try? What we want is a seven-vertex graph drawn
on the plane, with the property that every edge is a line segment of unit

148
 Working for the System

length (the unit being 100 furlongs = 12 12 miles), having the following
property: Every three vertices contain an edge.
If there were only three vertices, we could of course arrange them
as an equilateral triangle of side 1, and get more: an edge between
every pair of vertices. In fact, two of these equilateral triangles (say,
ABC and DEF ), no matter their relative position, would get us a
six-vertex solution. Why? Because any three vertices would include
at least two in one of the two triangles, and there’s our edge.
So let’s see if we can add a seventh vertex, call it G, to our two
triangles. Then we only have to worry about sets of three vertices that
contain vertex G.
We can’t connect G to all three vertices of one of our triangles, but
we can connect it to two (say, A and B from the first triangle, and D
and E from the second) by making diamonds AGBC and DGEF that
share vertex G. That leaves just one set of three vertices without and
edge: the set {G, C, F }. No problem: We just angle the two diamonds,
hinged at G, so that their far ends are a unit distance apart. Done!
Here’s what it looks like:

This graph is known as the “Moser spindle,” and it’s the only
graph with the desired property.

Theorem. Suppose seven distinct points in the plane have the property that
among any three of them, there are two that are exactly a unit distance apart.
Then the graph on those seven points, defined by inserting an edge between
each pair that are a unit distance apart, is the Moser spindle.

149
 Mathematical Puzzles

Proof. To prove this we again need to be systematic. We assume the

graph H satisfies the requisite properties, and gradually learn more
and more about it until we can show it actually is the Moser spindle.
First step: We observe that for any vertex v, the vertices not equal
to or adjacent to v must form a “clique,” that is, they must all be con-
nected to one another. Why? Because if two (say, x and y) are not,
then the set {v, x, y} would violate the condition. But four vertices
cannot all be connected to each other by unit-distance edges, so the
number of vertices not equal or adjacent to v is at most three; it follows
that v, and thus every vertex of H, has at least three neighbors.
Second step: No vertex can have as many as five neighbors. Rea-
son: If vertex u, say, has five neighbors, they all lie in a circle of unit
radius with center at u and cannot be connected among themselves
any better than by a path of four edges. Such a path has three vertices
no two of which are mutually adjacent, and that’s not allowed.
Third step: We now know all vertices have either three or four
neighbors, but they can’t all have just three neighbors—for then, the
sum of the degrees of the vertices of H would be 3 × 7 = 21 which
is an odd number, and each edge contributes 2 to that sum. Hence,
there is at least one vertex, say w, that has four neighbors.
Fourth step: How are w’s neighbors connected among themselves?
At least in two pairs, say p with q and r with s, because otherwise
there’d be three with no edge. The two remaining vertices (call them
x and y) must be adjacent to each other, else {w, x, y} would be a bad
set. Neither x nor y can be adjacent to three of p, q, r, and s because
there’s already a point (w) with that property and there can’t be two.
Wrap-up: Of p, q, r, and s, the ones x is not adjacent to must form
a clique, otherwise we’d have another bad set; and the same for y.
That is possible only if one of x and y is adjacent to precisely p and q,
and the other to r and s.
And that’s the Moser graph! ♡

150
 12. The Pigeonhole
Principle
The pigeonhole principle is nothing more than the evident fact that if n
pigeons are to be housed in m holes, and n > m, then some hole must
contain more than one pigeon. Yet this principle has wide-ranging
application, and is the key to proving some surprising statements and
some important mathematical truths.
First, a warm-up.

Shoes, Socks, and Gloves

You need to pack for a midnight flight to Iceland but the power is out.
In your closet are six pairs of shoes, six black socks, six gray socks,
six pairs of brown gloves, and six pairs of tan gloves. Unfortunately,
it’s too dark to match shoes or to see any colors.

How many of each of these items do you need to take to be sure

of getting a matched pair of shoes, two socks of the same color, and
matching gloves?
Solution: The easiest way to solve such puzzles is to figure out the
maximum number of items you could have without getting what you
want, then adding one. In this way you are implicitly invoking the
pigeonhole principle.

151
 Mathematical Puzzles

For the shoes, you could have one shoe from each pair, six in all;
so you need seven to guarantee a pair.
For the socks, one black and one gray will not work but the third
sock gets you a pair.
The gloves are yet another issue: You need a right- and a left-
hander, so the biggest disaster would be six brown all left or all right,
and six tan, also all left or all right. So 13 gloves are required, although
you would have to count yourself very unlucky indeed to take only 12
and not get a match!
The same idea is easily employed in certain poker calculations. A
standard deck of 52 playing cards contains 13 cards of each suit, and
within each suit, an ace, deuce, 3, 4, etc. up to 9, 10, jack, queen, king.
Thus there are four cards of each rank.
A poker hand is five cards. Two cards of the same rank constitute
a pair; three, trips (or “three of a kind”). Five cards of the same suit
make a flush; five cards of consecutive ranks (including 5432A as well
as AKQJ10) make a straight.

How many cards do you need to hold to guarantee a pair—that

is, to be sure that among your cards is a 5-card poker hand containing
two cards of the same rank? Easy: The worst holding (for this prob-
lem!) is 13 cards consisting of one card of each rank, therefore the
answer is 14 (but note that 12 cards are enough to guarantee a pair or
better).
If it’s trips you seek, the worst holding is two cards of each rank,
so you need 2 × 13 + 1 = 27 for your guarantee. For a flush, four of
each suit will stump you, so you need just 17 cards. The straight is
the trickiest: You must have a 5 or a 10 to make a straight, so you can
hold 52 − 2 × 4 = 44 cards without making a straight, hence it takes
a stunning 45 cards to guarantee one. Note the contrast with relative
probabilities of getting trips, flush or straight when dealt five cards:
The flush is the hardest, and indeed outscores the straight and trips in
poker, even though it takes the fewest cards to guarantee.
For the next puzzle, one preliminary bit of cleverness is needed.

152
 The Pigeonhole Principle

Polyhedron Faces
Prove that any convex polyhedron has two faces with the same num-
ber of edges.

Solution: The key is to pick a face with the maximum number of

edges, say E. Since it has at least E neighboring faces, the number of
faces of the polyhedron is at least E+1. But their numbers of edges
range only from 3 to E, comprising only E−2 different numbers, so
two must have the same number of edges. (In fact, since we have some
overkill here, you can deduce that there are either three pairs of faces
with the same number of edges, or a triple and a pair, or four faces
with the same number of edges.)
Sometimes you have to look around a bit to find the right items to
use for your pigeonholes, as in this geometric example.

Lines Through a Grid

If you want to cover all the vertices in a 10 × 10 square grid by lines
none of which is parallel to a side of the square, how many lines are
needed?

Solution: Your first idea here might be to use 19 diagonals running

45◦ running (say) north–west to south–east, but you immediately see
that the two extreme ones can be replaced by a single SW–NE diag-
onal covering the SW and NE corners, as in the figure below. This
reduces the count to 18; can you do any better?

No, because there are 36 vertices on the boundary of the grid, and
no non-vertical, non-horizontal line can cover more than two of them.

153
 Mathematical Puzzles

Now let us move on to an oddly counterintuitive fact about num-

bers.

Same Sum Subsets

Amy asks Brad to pick ten different numbers between 1 and 100, and
to write them down secretly on a piece of paper. She now tells him
she’s willing to bet $100 to $1 that his numbers contain two nonempty
disjoint subsets with the same sum! Is she nuts?

Solution: Of course, the answer is that Amy is perfectly sane and

in fact the bet is not merely favorable to her; it’s a sure thing. How can
you prove it?
There’s no unique maximum set-size Brad can pick that doesn’t
have the desired property, so the shoe-sock-glove technique won’t
work. But whenever you are asked to show that there are two objects
that share a property, you can try to show that there are more objects
than properties; then, by pigeonhole, you’re done.
Here, the number of nonempty subsets of Brad’s set of 10 numbers
is 210 − 1 = 1023. The possible sums range only from 1 to 100 + 99 +
· · · + 91 < 1000, so there are indeed fewer sums than subsets; by
the pigeonhole principle, there must be two different subsets with the
same sum.
But wait—the problem asked for disjoint subsets. That was a red
herring: Throw away the common elements of two sets with equal
sums to get disjoint sets with equal sums. ♡
The next problem looks geometric, but is really number-theoretic;
nor is it obvious at first glance that the pigeonhole principle has a role
to play here.

Lattice Points and Line Segments

How many lattice points (that is, points with integer coordinates) can
you have in 3-space with the property that no two of them have a third
lattice point on the line segment between them?

Solution: The first issue to be addressed here is: When do two lat-
tice points, say (a, b, c) and (d, e, f ), have another lattice point on the
line segment between them? A moment’s thought will convince you
that this happens when the numbers a−d, b−e, and c−f have a com-
mon divisor. The “easiest” common divisor for them to have is 2, and

154
 The Pigeonhole Principle

this will happen if a−d ≡ b−e ≡ c−f ≡ 0 mod 2, that is, if the two
vectors have the same parity coordinatewise. When that happens, the
midpoint of the two vectors is a lattice point.
But there are only 23 = 8 parities available, so by pigeonhole, we
can’t have nine points in our set. Can we have eight? Yes, just take the
corners of a unit cube.
Here’s another number problem.

Adding, Multiplying, and Grouping

Forty-two positive integers (not necessarily distinct) are written in a
row. Show that you can put plus signs, times signs, and parentheses
between the integers in such a way that the value of the resulting ex-
pression is evenly divisible by one million.
Solution: This puzzle has a pigeonhole flavor to it and indeed the
pigeonhole principle is useful in it, but not perhaps where you might
expect. Since a million is 26 · 56 , and 42 = 6 · 2 + 6 · 5, a logical
idea is to break up the numbers into six contiguous groups of size 2
and six more of size 5, with the multiplication symbols between the
groups; then all that is needed is to “persuade” each group of size 2
to represent an even number, and each group of size 5 to represent a
multiple of 5.
The groups of size 2 are easily handled: If the two numbers inside
a group are both even or both odd, add them; otherwise, multiply
them.
The groups of size 5 are only slightly trickier. If we can get a con-
tiguous substring of the five numbers that sums to a multiple of 5,
we’re done; we can add those and multiply the sum by the product of
the remaining numbers.
Here’s where pigeonhole comes in. Let the numbers be n1 , . . . , n5
and examine the partial sums s1 , . . . , s5 where si = n1 + · · · + ni . Let
ri be si mod 5, that is, ri is the remainder when si is divided by 5. If
any ri is 0, we use the substring n1 , . . . , ni ; otherwise, the five si ’s all
lie in the set {1, 2, 3, 4}. Thus two of them, say ri and rj with i < j,
are equal. Then sj − si = ni+1 + ni+2 + · · · + nj is a multiple of 5
and we are done. ♡
Notice that this argument works for 5 replaced by any n. We have
in effect proven a nice lemma which says that every string of n num-
bers has a nonempty substring that sums to a multiple of n.

155
 Mathematical Puzzles

Line Up by Height
Yankees manager Casey Stengel famously once told his players to
“line up alphabetically by height.” Suppose 26 players, no two ex-
actly the same height, are lined up alphabetically. Prove that there
are at least six who are also in height order—either tallest to shortest,
or shortest to tallest.
Solution: Let’s try small numbers to see what’s going on. Two play-
ers are, obviously, enough to get two that are also in height order one
way or the other. Three or four players won’t get you any more: For
example, their heights could be 73′′ , 75′′ , 70′′ , and 72′′ respectively.
But five players will get you three in height order; by the time you
reach that fifth player (say, Rizzuto) alphabetically, there’s a previous
player (say, Berra) who’s got both a taller and shorter player before
him. Then, whether Rizzuto is taller or shorter than Berra, he ends a
height-order subsequence of length 3.
This reasoning suggests a sort of “dynamic programming” ap-
proach to the 26-player problem. Give each player two numbers, an
“up” number and a “down” number. The up number records the
length of the longest height-increasing subsequence that ends with
that player, and the down number is defined analagously. For exam-
ple, in the height sequence 73′′ , 75′′ , 70′′ , 72′′ from above, the up-
down number pairs would be (1,1), (2,1), (1,2), and (2,2).
Notice these are all different—they have to be! Rizzuto and Berra,
for instance, can’t have the same pair because if the alphabetically
later player (Rizzuto) is taller his up number would be higher, and if
shorter, his down number would be higher.
Now we see why five players guarantees a subsequence of three
in height order: Only four different up-down pairs can be made of
numbers less than 3. Similarly, if there were no height-order subse-
quence of 26 players of length 6, all the up-down pairs would have to
be made up of numbers from 1 to 5 and there are only 25 of those. By
the pigeonhole principle, there must be a player with an up-number
or down-number of 6.
The key relationship is that if a subsequence of length k+1 is
sought, you need the number of players to be greater than k 2 . Still
more generally, if the number of players is at least jk+1, you are
guaranteed either a height-increasing subsequence of length j+1 or a
height-decreasing subsequence of length k+1. The argument is the
same: If you didn’t have either such subsequence, the number of

156
 The Pigeonhole Principle

up-down pairs would be limited to jk, and since the pairs are all dif-
ferent, you get a pigeonhole contradiction.
In fact, this last statement is known as the Erd˝os–Szekeres The-
orem, and can be even further generalized (with almost the same
pigeonhole proof) to a statement about height and width of par-
tially ordered sets—for those readers who know what they are.
The Erdős–Szekeres Theorem is “tight” in the sense that if there
are only jk players, you might indeed be limited to length j for
height-increasing subsequences and k for height-decreasing. An ex-
ample: If, say, j = 4 and k = 3, the heights (in inches) could be
76, 77, 78, 79, 72, 73, 74, 75, 68, 69, 70, 71.
Here’s a related problem where we can use a similar idea.

Ascending and Descending

Fix a number n and call a list of the numbers from 1 to n “good” if
there is no descending subsequence of length 10. Show that there are
at most 81n good lists.

Solution: Let x1 , . . . , xn be a good list and let di be the length of

the longest descending subsequence ending at xi . Then di is a number
between 1 and 9, which we call the “color” of the position i, and
also of the number xi . Essentially, di is the “down-number” from the
solution to the previous puzzle.
The numbers xi for which di is some particular color, say color
7, form an increasing subsequence. Suppose we know the positions
with color 7 and also the numbers of color 7. Then we can fill in all
the numbers that go in those positions, because we know they are in
increasing order.
For example, if the list begins 42, 68, 35, 50, then positions 1 and
2 get color 1, as do the numbers 42 and 68. Knowing that d1 = d2 = 1
and the set of numbers that got color 1, we can conclude that x1 = 42
and x2 = 68 because the first two positions are colored 1 and the two
lowest numbers that got the first color are 42 and 68.
Now, since there are at most 9n ways to assign colors to positions,
and 9n ways to assign colors to numbers, we get (using, again, the
Pigeonhole Principle) a maximum of (9n )2 = 81n different lists.
Note that this bound of 81n , though exponentially large, is a van-
ishingly small fraction of all the n! ways to list the numbers from
1 to n.

157
 Mathematical Puzzles

Sometimes the pigeonhole principle can be applied even though

a puzzle asks only for one item, not two similar ones. The trick is to
use subtraction to turn two similar items into one useful one.

Zeroes and Ones

Show that every natural number has a non-zero multiple which, when
written out (base 10), contains only zeroes and ones. Even better,
prove that your telephone number, if it ends with a 1, 3, 7, or 9, has a
multiple whose decimal representation is all ones!

Solution: There are infinitely many numbers whose decimal rep-

resentations contain only ones and zeroes—or, for that matter, only
ones. So some of them ought to be multiples of n unless there’s some
reason why they can’t be. But how can we prove it?
Perhaps we can make use of the fact that if we subtract two num-
bers each of which has the same remainder when divided by n, the
result will be a multiple of n. There are only n different remainders
possible when you divide by n, so pigeonhole tells us that n+1 num-
bers are enough to guarantee two with the same remainder.
Now we only have to find n+1 numbers with the property that
the difference between any two has only zeroes and ones as its dig-
its. That’s not hard: Take numbers of the form 111 . . . 111, where the
number of digits runs from 1 to n+1. Formally speaking, these num-
bers are (10k − 1)/9 for 1 ≤ k ≤ n+1. (Actually, we only need n
of them because if all their remainders modulo n differ, one of them
already has remainder 0, and we can use it directly.)
To wrap up, we find two distinct numbers all of whose digits
are ones, that have the same remainder when divided by n; sub-
tracting the smaller from the larger gives a number of the form
111 . . . 111000 . . . 000.
This means that we have actually proved a stronger statement
than required, namely that any n has a multiple of the form
111 . . . 111000 . . . 000. But, as often happens, asking for more would
have made the puzzle easier. We’ll see a more extreme example of
that phenomenon in the next puzzle.
Wait, what about your telephone number? We know now that it
has a multiple of the form N = 111 . . . 111000 . . . 000, ending in, say,
k zeroes. But if your number ends in 1,3, 7, or 9, it is not a multiple
of 5 or 2. It follows that N/(5 · 2)k , which is all ones, is still a multiple
of your phone number.

158
 The Pigeonhole Principle

The next puzzle looks like one in which pigeonhole refuses to

work. It takes some serious insight to see how it ultimately does.

Same Sum Dice

You roll a set of n red n-sided dice, and a set of n black n-sided dice.
Each die is labelled with the numbers from 1 to n. Show that there
must always be a nonempty subset of the red dice, and a nonempty
subset of the black dice, showing the same sum.

Solution: This seems like an ideal pigeonhole application: There

are lots of nonempty subsets (2n − 1 of each color) and sums can
only go from 1 to n2 . But it doesn’t do any good to have, say, two red
subsets with the same sum, and two black subsets with some other
sum.
Come to think of it, there may be very few distinct sums among,
say, the red dice; for example, they could all be showing the same
number of pips, say j. In the meantime the black dice could all be
agreeing on some other number, k. That does not give us a counterex-
ample, because then k of the reds and j of the blacks both sum to jk,
but that coincidence seems like something the pigeonhole principle
would be hard pressed to lay claim to.
Nonetheless, with help from a subtraction trick similar to the one
we used in Zeroes and Ones, pigeonhole saves us. The idea is the fol-
lowing: Suppose the subsets R (of the red dice) and B (of the black)
don’t work, because the sum of the dice in R exceeds that of B by
some non-zero number d. Let’s add some more red dice to R to get
a bigger set R+ , and similarly add some black dice to B to get B + .
Now we check the sums for R+ and B + and they still don’t agree, but
suppose that the discrepancy is the same as it was before: The dice in
R+ sum to exactly d more than the dice in B + . Then we’re done, be-
cause the dice we added to R must have had the same sum as those
we added to B.

159
 Mathematical Puzzles

To get this idea to work we put each set of dice in some fixed
left-to-right order. Amazingly, it doesn’t matter a bit what orders we
choose. Let ri be the sum of the first i red dice, reading the dice from
left to right, and similarly let bj be the sum of the leftmost j black dice.
We claim that there are contiguous subsets of the red and of the black
dice with the same sum.
A subset is “contiguous” if it is a substring of the dice as laid out,
for
(n+1)example, all the red dice from the third to the seventh. There are
2 = n(n+1)/2 such subsets of each color, because we can get one
by choosing two of the n+1 spaces between or at the ends of the dice
in which to put down supermarket dividers. It’s hard to believe that
we can get away with limiting ourselves, in this seemingly arbitrary
manner, to just these few of the exponentially many subsets of each
color. Can this really work no matter how the dice are rolled, and no
matter how we order them?
We can assume rn < bn ; if they are equal we’re done, and if the
inequality goes the other way we can just switch the roles of the colors.
Let us find, for each i, the j for which bj comes closest to ri without
exceeding it. More formally, let

j(i) = max{j : bj ≤ ri }

with the understanding that b0 = 0, and thus j(i) may be zero for
some small i’s if the red sequence starts with smaller numbers than
the black sequence.
Now let’s look at the discrepancies ri − bj(i) . All are non-negative
by definition of j(i); moreover, all are less than n, because if ri − bj(i)
were equal to n or more, we could have thrown in the next black die
without exceeding ri . (Since rn < bn , there always is a next black die.)
If any of the discrepancies ri − bj(i) is zero, we’re done, so we can
assume all the discrepancies are among the numbers 1, 2, . . . , n−1.
There are n such discrepancies, one for each i, so there most be two
that are equal; in other words, there are two distinct indices i < i′
with
ri′ − bj(i′ ) = ri − bj(i) .
Now we play the subtraction game, concluding that the red dice
from number i+1 up to number i′ must add up to the same value as
the black dice from number j(i)+1 up to number j(i′ ). Done!
Our last puzzle looks perhaps less like a pigeonhole problem than
its predecessors, but now that you’ve seen the subraction trick, you
won’t be fooled. Nonetheless, it’s not easy.

160
 The Pigeonhole Principle

Zero-sum Vectors
On a piece of paper, you have (for some reason) made an array whose
rows consist of all 2n of the n-dimensional vectors with coordinates
in {+1, −1}n —that is, all possible strings of +1’s and −1’s of length
n.
Notice that there are lots of nonempty subsets of your rows which
sum to the zero vector, for example any vector and its complement;
or the whole array, for that matter.
However, your 2-year-old nephew has got hold of the paper and
has changed some of the entries in the array to zeroes.
Prove that no matter what your nephew did, you can find a
nonempty subset of the rows in the new array that sums to zero.
Solution: The technique used in Same Sum Dice suggests that to get
a set of new rows that sum to zero, we might try to build a sequence
of new rows and use pigeonhole to argue that two of its partial sums
coincide; then the portion of the sequence between those two partial
sums will add to the zero vector.
To make that work, the sequence will have to be designed so that
the partial sums belong to some small set. How small can we make it?
Would you believe that we can force all the partial sums to be {0, 1}-
vectors?
We can certainly start the list with a {0, 1}-vector: We just take the
row u′ that used to be u = ⟨1, 1, 1, . . . , 1⟩. If our nephew happened to
have changed all of u’s coordinates to zeroes, we’re done. Otherwise,
there will be some 1’s in u′ , and we need to choose the next row v ′
in our sequence in such a way that u′ + v ′ has all of its coordinates
in {0, 1}. But we can do that by choosing v to be the vector that has
−1 in the places where u′ is 1, and +1 in the places where u′ is zero.
Then no matter how v was altered by your nephew, u′ + v ′ will have
the desired property.
We continue this procedure as follows: At time t, t altered rows
have been put on our sequence and all partial sums, including the sum
x of all t of them, are {0, 1}-vectors. Let z be the original row with −1
wherever t has a 1, and +1 wherever t has a 0; add z’s altered form z ′
to the sequence.
Since there are only finitely many {0, 1}-vectors of length n, we
will eventually hit the same sum twice; let us stop when this happens
for the first time. (We consider the 0th partial sum to be the zero vec-
tor, so one way this can happen is for the zero vector to arise again as
a partial sum.) Call the vectors w1 , w2 , . . . , so that our critical time is

161
 Mathematical Puzzles

the least k such that w1 + w2 + · · · + wj = x = w1 + w2 + · · · + wk ,

with 0 ≤ j < k.
Then wj+1 + wj+2 + · · · + wk is the zero vector, and we are done.
Note, very importantly, that we have not used any altered vector too
many times; the reason is that since up to time k all the partial sums
have been different, all the original vectors z whose altered forms we
appended to our sequence were different.

Here’s a more serious use of the pigeonhole principle in mathe-

matical analysis. Suppose we fix some real number r and let {nr} be
the fractional part of nr, that is, the unique number in the half-open
interval [0, 1) that can be expressed as nr minus an integer.
If r is a fraction (say, a/b) written in lowest terms, then the val-
ues {r}, {2r}, {3r}, . . . will cycle among the multiples of 1/b in [0, 1).
What if r is irrational? Then the {nr}’s will not repeat and it’s natural
to conjecture that they are in fact dense in the unit interval; in other
words, for any ε > 0 and any real x ∈ [0, 1), there is some positive
integer n for which |{nr} − x| < ε.

Theorem. For any irrational number r, if {nr} is the fractional part of nr,
then the numbers {r}, {2r}, {3r}, . . . are dense in the unit interval.
Proof. Fix some irrational r. Of course {nr} is never zero for any
positive integer n, but we can borrow an idea from the Zeroes and Ones
puzzle above to show that we can choose n to get nr is as close as we
want to an integer.
Let 1 ≤ i < j. If {ir} < {jr}, , then {(j−i)r} = {jr} − {ir};
if {ir} > {jr}, then {(j−i)r} = 1 − ({ir}−{jr}). Pick any ε > 0,
and let m be an integer bigger than 1/ε. Divide the unit interval into
m equal intervals; by pigeonhole, one of those little intervals (each of
length < ε) contains two of the values {r}, {2r}, . . . , {(m+1)r}. Let
those be {ir} and {jr} with i < j; then if k = j−i, we conclude that
either {kr} < ε or 1−{kr} < ε.
Suppose that the former inequality holds, that is, {kr} < ε. Given
any real value x ∈ [0, 1], let p be the greatest integer such that p{nr} <
x. Then {pkn} = p{kn} < x ≤ (p+1){kr} but the outside quatities
differ by {kr} which is less than ε, so in particular |x−{pkr}| < ε and
we have our approximation.
If instead 1−{kr} < ε, we take p to be the greatest integer
such that p(1−{kr}) > 1−x, and a similar argument again shows
|x−{pkr}| < ε. ♡

162
 13. Information, Please
The basis of this chapter is simple: If you want to distinguish among n
possibilities, you have to do something that results in at least n differ-
ent possible outcomes. A common puzzle application is to balance-
scale problems.
Each weighing by a balance scale can produce at most three out-
comes: tilt left (L), tilt right (R), or balance (B). If you are permitted
two weighings, you get 3×3 = 9 possible outcomes: LL, LR, LB, RL,
RR, RB, BL, BR, and BB. Similarly, with w weighings you could have
as many as 3w outcomes, but no more.
So, if a puzzle gives you a bunch of objects to test, k possible an-
swers, and allows w weighings on a balance scale, the first thing to
ask yourself is whether k ≤ 3w . If not, the task is impossible (or you
have missed some trick in the wording). If you do have k ≤ 3w , the
task may or may not be possible, and you can use the idea above to
help you construct a solution or to prove impossibility.
Here’s a classic example.

Finding the Counterfeit

You have a balance scale and 12 coins, 11 of which are genuine and
identical in weight; but one is counterfeit, and is either lighter or heav-
ier than the others. Can you determine, in three weighings with a bal-
ance scale, which coin is counterfeit and whether it is heavy or light?

Solution: There are 24 possible outcomes, less then 33 = 27, so the

task may be possible but needs to be pretty efficient—in other words,
most weighings will need to have all three possible outcomes. We cer-
tainly want to start by weighing k coins against k others, where k ≤ 6.
That would leave 12−2k suspect coins if the initial weighing balances,
so it better be possible to test that many in two weighings. Two weigh-
ings have at most nine possible outcomes, so we can’t afford to leave
more than four coins out of the first weighing; thus k is at least 4.

163
 Mathematical Puzzles

In fact k can’t be more than 4 because then “tilt left” would lead to
10 > 32 possible outcomes. So we have reached a firm conclusion:
If the task can be done, it must be right to put 4 coins against 4 (say,
coins ABCD against EFGH) in the first weighing.
Suppose they balance. Then the temptation is to put 1 against
1 or 2 against 2 of the remaining coins, but either is information-
theoretically deficient: If 1 balances against 1, four (so, more than
three) possibilities remain; if 2 tilts against 2, again four possibilities
remain. Either way one more weighing is insufficient. So we must use
the known-genuine coins A through H in the second weighing, and
the easiest thing to do is put three of them against three of the sus-
pect coins (say, IJK). If they balance, coin L is counterfeit and we can
weigh it against A to determine whether it is light or heavy. If IJK are
light or heavy, test I against J.
If the initial 4-against-4 weighing tilts left, so that either one of
ABCD is heavy or one of EFGH is light, put (say) ABE against CDF.
If they balance, the culprit is a light G or H and you can test either
against the genuine L. If they tilt left, you’re left with A or B heavy or
F light and you can determine which by putting A and F against two
genuine coins, say K and L; the solution is similar if CDF outweigh
ABE.
Now suppose you are given the same problem but with 13 coins.
There are now 26 possible outcomes, which is still less than or equal to
27, but our previous reasoning already showed that nonetheless three
weighings are not enough. Why? Because we saw that 5-against-5 in
the first weighing is no good, while 4-against-4, when they balance,
leaves five coins—too many—untested.

164
 Information, Please

Three Natives at the Crossroads

A logician is visiting the South Seas and as is usual for logicians in
puzzles, she is at a fork, wanting to know which of two roads leads to
the village. Present this time are three willing natives, one each from a
tribe of invariable truth-tellers, a tribe of invariable liars, and a tribe of
random answerers. Of course, the logician doesn’t know which native
is from which tribe. Moreover, she is permitted to ask only two yes-
or-no questions, each question being directed to just one native. Can
she get the information she needs? How about if she can ask only one
yes-or-no question?
Solution: We can dispose of the one-question case easily: If the
question is directed to the random answerer, the logician gets no in-
formation, thus can never guarantee to identify the right road. (This
argument does not apply if you assume the random answerer first flips
a mental coin to determine whether to lie or tell the truth; you could
then gain information with a well-chosen self-referential statement,
for example, “Out of the other two guys, if I pick the one whose re-
sponse’s truthfulness will least likely match your response’s truthful-
ness and ask him if Road 1 goes to the village, will he answer ‘Yes’?”
But we assume the random answerer just randomly answers “yes” or
“no” regardless of the question, so no information is imparted.)
Similarly, if the logician doesn’t know the right road after one
question, and her second question is directed to the random answerer, she
is in trouble. It follows that after the first answer, she must be able to
identify a native who is not the random answerer.
If she can do that, she’s in business, because she can then use a tra-
ditional one-native query as her second question: For example, some-
thing like “If I were to ask you whether Road 1 goes to the village,
would you say yes?”
To attain the objective, she’ll need to ask Native A something
about Native B or Native C, then use the answer to choose between
B and C. Here’s one that works: “Is B more likely than C to tell the
truth?”
Curiously, if A says “yes” she picks C, and if he says “no” she picks
B! If A is the truth-teller she wants next to query the companion who
is less likely to tell the truth, namely, the liar. If A is the liar she queries
the more truthful of his companions, namely the truth-teller.
Of course, if A is the random answerer it doesn’t matter which of
B and C she turns to next. ♡

165
 Mathematical Puzzles

If in a puzzle the information seems insufficient, we may have to

find extra sources.

Attic Lamp
An old-fashioned incandescent lamp in the attic is controlled by one
of three on–off switches downstairs—but which one? Your mission is
to do something with the switches, then determine after one trip to the
attic which switch is connected to the attic lamp.

Solution: This seems not to be doable: There are three possible an-
swers (as to which of three switches controls the bulb), but when you
get to the attic either the bulb is on or off. So only one bit of informa-
tion is available, and we need a “trit.”
Is there anything that can be observed about the bulb other than
whether it is on or off ? There’s a hint in the puzzle wording: It’s an
incandescent bulb, whose efficiency is limited by its habit of putting
out quite a lot of heat as well as light. So you could feel the bulb and
see if it’s hot; can that help?
Yes. Flip switches A and B on, wait 10 minutes or so, then flip
switch B off. Go quickly up to the attic: If the bulb is on, it’s switch
A; if off but warm, switch B; off but cold, switch C.
You could even do four switches: Flip A and B on, wait 10 min-
utes, then flip B off and C on and run up the stairs. The four possible
states of the bulb (on/warm, off/warm, on/cool, or off/cool) will tell
you what you need to know.

Information is customarily measured in bits; k bits gives you 2k

possibilities. The thriving field of “communication complexity” at-
tempts to determine the minimum number of bits of communication
needed between two or more parties, in order for them to accomplish
some task. Here is a toy example.

Players and Winners

Tristan and Isolde expect to be in a situation of severely limited com-
munication, at which time Tristan will know which two of 16 basket-
ball teams played a game, and Isolde will know who won. How many
bits must be communicated between Tristan and Isolde in order for
the former to find out who won?

166
 Information, Please

Solution: Tristan and Isolde can, in advance, label the 16 teams (in
alphabetic order, perhaps) by the binary numbers 0000 through 1111.
Then, when the time comes, Tristan can send Isolde eight bits: the
four bits corresponding to one team, and the four bits corresponding
to the other one. Isolde can then send back a 0 if the winning team
was the one with the smaller number, and a 1 otherwise. That comes
to nine bits, and we can trim that to 8 by noting that the number of
unordered pairs of teams is only 16 × 15/2 = 120 < 27 and thus
Tristan can code the matchup with only seven bits.
But an easier and faster solution is to just have Isolde send Tristan
the four bits corresponding to the winning team. Surely, we can’t do
any better than four bits, right?
Amazingly, we can! We can code those four positions (leftmost,
next-to-left, next-to-right, rightmost) by, say, 00, 01, 10, and 11 re-
spectively. The codes of the two playing teams must differ in at least
one of the four positions; Tristan picks such a position and sends its
code (two bits) to Isolde. She now only needs to look at the value
of the bit in that position of the winning team and sent it to Tristan.
Three bits total!
For example, suppose the playing teams are 0010 and 0110, and
0110 won. The only position where the playing teams differ is the
next-to-left, code 01, so Tristan sends “01” to Isolde. She checks the
next-to-left position of the winning team—it is a “1”—and sends it
back to Tristan. Done.
It is not hard to show three bits is unbeatable. Note that in the
most efficient communication scheme, more information is passed
from the student to the teacher than vice-versa. Is there a lesson to be
learned from that?
Here’s another example where we need to think about informa-
tion, and also to be clever.

Missing Card
Yola and Zela have devised a clever card trick. While Yola is out of the
room, audience members pull out five cards from a bridge deck and
hand them to Zela. She looks them over, pulls one out, and calls Yola
into the room. Yola is handed the four remaining cards and proceeds
to guess correctly the identity of the pulled card.
How do they do it? And once you’ve figured that out, compute
the size of the biggest deck of cards they could use and still perform
the trick reliably.

167
 Mathematical Puzzles

Solution: Let’s try to look at this trick information-theoretically.

The obvious (and correct, for this trick) way for Zela to communi-
cate information to Yola when handing her the four remaining cards
is to order the cards in some way. There are 4! = 24 ways to order four
objects, but that doesn’t seem like enough—there are 48 possible val-
ues of the missing card. We could, of course, try to sneak in another
bit of information, for example, whether the cards are passed face up
or face down.
We could do that pretty easily and we would in fact need to, if
the card to be removed were chosen by an innocent bystander. But
Zela gets to decide which card is removed, and thus can limit the
possibilities to much less than 48.
Since Zela gets to pick the card, the right information-theoretic
way to look at the problem is to ask whether seeing four cards and
their order is enough to deduce the original five cards. The number of
ways to pick five cards from a deck is
( )
52 52 · 51 · 50 · 49 · 48
= = 2,598,960,
5 5·4·3·2·1
while the number of ways to pick four cards where the order counts is

52 · 51 · 50 · 49 = 6,497,400,

so we are OK, at least from an information-theoretic standpoint.

What is needed is a human-implementable one-to-one map from a
subset of the ordered foursomes to all the unordered fivesomes, with
the property that each foursome is contained in the fivesome it is
mapped to.
That requires some cleverness; here’s how magician William Fitch
Cheney did it back in the 1930’s. Making use of the pigeonhole prin-
ciple, he noted that every five cards contain at least two of some suit.
One of those will be removed and a second of that suit left on top of

168
 Information, Please

the pile of four, so already Yola will know the suit of the removed
card. The remaining three of the four can be ordered six ways, and
Yola will use that info to determine the rank of the missing card.
Here’s how. Yola and Zela fix an ordering of all the cards, per-
haps A,2,3,... up to 10,J,Q,K, with ties broken by suit (the traditional
magician’s ordering of suits is the “CHaSeD” order, namely Clubs,
Hearts, Spades, Diamonds). Let the three cards in question be called
A, B, and C, according to the ordering above. If they are handed to
Yola in order ABC, that codes the number 1; if ACB, the number 2;
BAC = 3, BCA = 4, CAB = 5, and CBA = 6.
Yola looks at the rank of the top card and adds her code number
to it, going “around the corner” if necessary. For example, suppose
the top card is the queen of spades, and the order of the other three
cards is CAB. Since the code of CAB is 5 and the queen is rank 12 of
13, Yola adds 5 to 12: 13=K, 14=1=A, 2, 3, 4. She concludes that the
missing card is the 4 of spades.
Wait, suppose the missing card had been the 6 of spades; then
Yola would have had to add 8 to the Q, and there’s no code 8. Not
to worry: In that case, Zela would have removed not the 6 of spades
but the Q of spades; and adding 6 gets you from 6 to queen. In other
words, given two cards of the same suit, we can always pick one so
that you can get to the other by adding a number from 1 to 6. ♡
Yes, this will take some practice on the part of Yola and Zela.
There’s actually quite a bit of “room” in this trick. Information-
(n)
theoretically,
(n) it can be done with a deck of any size n satisfying 5 ≤
4! × 4 which is true up to n = 124. In fact, you really can do it with
a deck of size 124, but we will leave it to you, the reader, to figure out
how.

What is the value of information? We can quantify it the following

way: Subtract your expectation without the information from your
expectation with. Since you don’t have to act on the information, this
quantity can never be negative. Conversely, if the information can’t or
shouldn’t change your actions, it’s worthless.

Peek Advantage
You are about to bet $100 on the color of the top card of a well-shuffled
deck of cards. You get to pick the color; if you’re right you win $100,
otherwise you lose the same.

169
 Mathematical Puzzles

How much is it worth to you to get a peek at the bottom card of

the deck? How much more for a peek at the bottom two cards?
Solution: Without the peek, the bet is fair, that is, your expected
gain is $0. If your peek reveals a red card on the bottom, you will of
course bet on black and be right with probability 26/51; then your
expectation will be
26 25
· $100 + · (−$100) ∼ $1.96.
51 51
Since the calculation is the same if the card on the bottom is black,
the value of the information is $1.96.
We can do a similar calculation for peeking at two cards, but why
bother? The second peek is worthless! Suppose the bottom card is
red. If the next-to-bottom card is red, then of course you still bet that
the top card is black. If the next-to-bottom card is black, then you’re
back to a fair game, so you may as well keep your plan of betting on
black. Since the second peek does not have the power to change your
strategy, its additional value to you is exactly $0.
The same kind of reasoning is useful in the next puzzle.

Bias Test
Before you are two coins; one is a fair coin, and the other is biased
toward heads. You’d like to try to figure out which is which, and to
do so you are permitted two flips. Should you flip each coin once, or
one coin twice?

Solution: Suppose you only get one flip. If you flip coin A and it
comes up heads, you will of course guess that it is the biased coin; if
it comes up tails, that it is the fair coin.
Now if you flip coin B and get the opposite face, you will be even
happier with your previous decision. If you get the same face, you will
be reduced to no information and may as well stick with the same
choice. Thus, as with the card-peeking, flipping the second coin is
worthless.
Could flipping the same coin twice change your mind? Yes. If you
get heads the first time you are inclined to guess that coin is biased, but
seeing tails the next time will change your best guess to “unbiased.”
We conclude that flipping one coin twice is strictly better than flipping
each coin once.

170
 Information, Please

As it turns out, it’s a theorem that in trying to determine which

is which of two known probability distributions, it’s better to draw
twice from one than once from each. However, unlike in the puzzle
above, it may happen that after seeing the result of your first draw you
prefer to draw from the other distribution. We’ll leave it to the reader
to concoct an example.
Information about information—typically involving a chain of de-
ductions to the effect that seeing A, person X would have deduced
B—is a common puzzle theme. The following example is perhaps
the best variation of a classic.

Dot-Town Exodus
Each resident of Dot-town carries a red or blue dot on his (or her) fore-
head, but if he ever thinks he knows what color it is he leaves town
immediately and permanently. Each day the residents gather; one
day a stranger comes and tells them something—anything—nontrivial
about the number of blue dots. Prove that eventually Dot-town be-
comes a ghost town, even if everyone can see that the stranger’s statement
is patently false.

Solution: This puzzle, often in non-PC versions involving unfaith-

ful spouses, murder, or suicide, has been around for at least a century.
Usually the stranger is assumed to have made a true statement, but as
we shall see, that stipulation is not necessary. The condition of “non-
triviality” means that, a priori, the statement could be true or false; in
other words, there is some number of blue dots that would make it
true, and some other that would falsify it.
Example: All the dots are blue but the stranger says they are all
red. Then everyone knows that the stranger is lying and moreover, that
everyone else knows that the stranger is lying. How can the stranger’s
statement have any effect?
It’s easiest to see when the population of Dot-town is small, for
example, 3. Then you, if you are a Dot-town resident, know that the
stranger is lying. But if your own dot is red, then your friend Fred
is looking at a red dot and a blue dot and wondering if the third
townsperson, Emily, is seeing two red dots. If so, Emily will believe
the stranger and leave town. When she doesn’t, Fred will correctly
conclude that his own dot is blue, and will leave town the second
night. When this fails to happen, you—and Fred and Emily, as well—
can all conclude that your hats are blue, and all leave town on the third
night.

171
 Mathematical Puzzles

To do a proof of the general case, we need some notation. Let

S ⊂ {0, 1, . . . , n} be the set of numbers x with the property that if
there are x blue dots among the n residents of Dot-town, then the
stranger’s statement is true. In other words, the stranger’s statement
is equivalent to “The number of blue dots is in S.” Our non-triviality
assumption tells us that S is a proper, non-empty subset. Let b be the
actual number of blue dots, which may or may not be in S.
For resident i, let Bi be the set consisting of the possible numbers
of blue dots, from i’s point of view. Prior to the stranger’s visit, Bi =
{bi , bi +1} where bi is the number of blue dots resident i sees among
his compatriots.
If at any point Bi drops to one value, resident i is cooked. This
will happen immediately if |Bi ∩ S| = 1, but it will also happen the
night after any departures occur. To see this, we observe first that all
residents with the same dot color will behave identically, since they
all see the same number of dots. Thus if resident i sees that anyone
has left town, he deduces (correctly) that that person’s dot color is dif-
ferent from his own; he therefore knows his own color and is doomed
to exile.
Given S and b, let db be the number of steps (increments or decre-
ments by 1) needed to get from b across the border of S; in other
words, db is the least k such that b + k or b − k is in S if b is not, and
out of S (but still in {0, 1, . . . , n}) if b is in S. You can think of db as the
distance from the stranger’s statement to the truth, if the statement is
false; and the distance to falsehood, if it is true.
Example: If n = 10 and S = {0, 1, 2, 9, 10} then d0 = 3, d1 = 2,
d2 = d3 = 1, d4 = 2, d5 = d6 = 3, d7 = 2, d8 = d9 = 1, and d10 = 2.
We claim that the first departures will occur exactly on night db .
The proof is by induction on db . If db = 1, some residents will
find the stranger immediately credible, and will leave the first night.
Suppose t > 1 and assume that our claim is true whenever db <
t. The day after night t−1, since no departures have yet occurred,
everyone will deduce that db ≥ t. However, if db = t then either db−1
or db+1 must be equal to t−1. If the former, then those residents with
blue dots, who see that the number of blue dots is either b (the actual
number) or b−1, can rule out b−1 and are toast. If the former, it is
the red-dotted folks who can rule out b+1 and must pack their things.
Finally, if db−1 = db+1 = t−1 then nobody stays the night.
Since db is at most n, the proof tells us that everyone will have
exited by the n+1st night. But in the db = n case all residents have the

172
 Information, Please

same dot-color so they all leave together on the nth night. It follows
that Dot-town is always deserted within n days.
It is perhaps worth noting also that the definition of db makes no
distinction between S and its complement; from this it follows that
it makes no difference whether the stranger says “X” or “not X,” the
residents of Dot-town will behave exactly the same way in either case.
You might reasonably wonder whether the Dot-town residents,
knowing that a stranger is coming and liable to break the manifestly
justifiable no-talk-about-dot-colors taboo, can organize some defense.
For example, everyone who knows the stranger is lying jumps up and
says so. Alas, a little thought will show you that neither this nor any
similar strategy can save the town. A fragile lot, these Dot-towners.
A whole class of puzzles is built around conversations in which
the speakers take several rounds to make a deduction. Your job is to
deduce something of your own from the conversation, even though
you have less information than the speakers!
My favorite of these is the following. Part of the beauty of this one
is that no numbers are named in the puzzle statement.

Conversation on a Bus
Ephraim and Fatima, two colleagues in the Mathematics Department
at Zorn University, wind up seated together on the bus to campus.
Ephraim begins a conversation with “So, Fatima, how are your
kids doing? How old are they now, anyway?”
“It turns out,” says Fatima as she is putting the $1 change from the
bus driver into her wallet, “that the sum of their ages is the number of
this bus, and the product is the number of dollars that happen to be
in my purse at the moment.”
“Aha!” replies Ephraim. “So, if I remembered how many kids you
have, and you told me how much money you are carrying, I could
deduce their ages?”
“Actually, no,” says Fatima.
“In that case,” says Ephraim, “I know how much money is in your
purse.”
What is the bus number?

Solution: We will assume all numbers in the puzzle are positive

integers—and that Ephraim knows the number of the bus he’s on.
That number (call it n) must allow two different sets of kids’ ages,
with the same number of kids, the same sum n, and the same product;

173
 Mathematical Puzzles

otherwise Ephraim would have been able to deduce the kids’ ages
from sum, product and number of kids.
One would expect that large values of n would tend to have this
property; what prevents any large number from being the answer? Ah,
the last line of the conversation shows that there can’t be two pairs of
kids’-age-sets with sum n, one pair sharing one product, the other
sharing a different product. If there were two such pairs, Ephraim
would not have been able to deduce the amount of money in Fatima’s
purse at the conversation’s end.
Bus number n = 13 is such an example. Fatima could have five
kids of ages 6, 2, 2, 2, and 1, or of ages 4, 4, 3, 1, and 1; either would
put $48 in her purse. But she could instead have three kids of ages 9,
2, and 2, or of ages 6, 6, and 1; those possibilities would both give her
purse $36. So even with the information that knowing the number of
kids and purse contents (in addition to the bus number) would not
have been sufficient to deduce the kids’ ages, Ephraim would still not
be in a position to deduce the purse contents if the bus number had
been 13.
What about higher numbers? Here’s the crucial observation: Once
you’ve hit a number with 13’s “ambiguity” property, every higher
number has it too! Why? Because you can take those four sets of kids’
ages and add one kid of age 1 to each one of them. That changes no
products but raises all the sums by 1, so if n is ambiguous, so is n+1.
Now it’s time to do the math, checking values of n below 13 to
try to find one that boasts exactly one pair of kids’-age-sets with the
same number of kids, sum n, and the same product. That’s not as time
consuming as it sounds, because (a) you only need consider sets of size
at least three (one or two numbers can always be deduced from their
sum and product), and (b) many products yield the kids’ ages readily
on account of unique prime factorization.
Turns out the puzzle-composer has not misled you: There is just
one bus number that qualifies, n = 12. The critical kids’-age-sets are
{4, 4, 3, 1} and {6, 2, 2, 2}.
The next example illustrates the complexities of trying to commu-
nicate and play a game at the same time.

Matching Coins
Sonny and Cher play the following game. In each round, a fair coin
is tossed. Before the coin is tossed, Sonny and Cher simultaneously

174
 Information, Please

declare their guess for the result of the coin toss. They win the round if
both guessed correctly. The goal is to maximize the fraction of rounds
won, when the game is played for many rounds.

So far, the answer is obviously 50%: Sonny and Cher agree on a

sequence of guesses (e.g., they decide to always declare “heads”), and
they can’t do any better than that. However, before the game begins,
the players are informed that just prior to the first toss, Cher will be
given the results of all the coin tosses in advance! She has a chance
now to discuss strategy with Sonny, but once she gets the coinflip
information, there will be no further opportunity to conspire. Show
how Sonny and Cher can guarantee to get at least 6 wins in 10 flips.
(If that’s too easy for you, show how they can guarantee at least
six wins in only nine flips!)

Solution: Guaranteeing 5 wins in 10 flips is a cinch: Cher chooses

her first five guesses to match the last five coinflips. Then Sonny and
Cher both repeat that sequence and collect their five wins. With luck
they might catch one or two additional wins among the first five flips,
but that can’t be guaranteed.
Is there a better way? Suppose Cher tries to communicate three
flips at a time to Bob. The nice thing about three flips is that they are
either mostly heads or mostly tails. If flips 2 through 4 are mostly
heads, Cher calls “heads” at flip 1 and this tells Sonny to guess
“heads” at flips 2, 3, and 4. Suppose flips 2 and 4 are heads but flip 3
is a tail. Then Cher will call “heads” at flips 2 and 4, garnering two
wins, and will use flip 3 to tell Sonny what to do in flips 5 through 7.
What if flips 2, 3, and 4 are all tails? Then, by agreement, we just
use Cher’s last call (on flip 4) to tell Sonny what to guess at the next
three flips.
Flips 5 through 7 are treated similarly, with the minority flip (or
the last flip, if all flips are the same) used to tell Sonny which way to
guess on flips 7 through 9.
In this way Sonny and Cher are guaranteed to get at least two wins
out of three in each of the flip-groups 2-4, 5-7, and 8-10.

175
 Mathematical Puzzles

Readers who like to imagine life in asymptopia will want to know

what fraction of correct flips can be guaranteed when the total number
of flips is large. It turns out Sonny and Cher can do much better than
2/3, in fact they can beat 80%(!) if the number of flips is huge and
they work hard to squeeze the last drop of blood out of their scheme.
As for 6 out of 9, that takes some rather careful work and I
wouldn’t want to deprive the ambitious reader of the opportunity to
devise a successful scheme. (Advice: Do 3 out of 5 first.) Eventually,
I’ll post a solution on the web.
In the science of cryptography, communication must be done with
an eye to keeping information from falling into the wrong hands. This
is most easily done if the parties that wish to communicate have a
common secret.
For example, suppose Alice is expecting to see Bob at a party and
hoping to tell him “yes” or “no” to a certain business proposition, pri-
vately, even though others at the party may be nearby and interested.
No problem; she and Bob arrange beforehand that “Hi” means yes
and “Hello” means no. This shared secret enables them to communi-
cate secretly.
The last puzzle in this chapter was designed to show that a shared
secret can sometimes be created over an open channel, provided that
the participants have some shared knowledge—that is, when put to-
gether they could deduce things that an eavesdropper couldn’t know.

Two Sheriffs
Two sheriffs in neighboring towns are on the track of a killer, in a case
involving eight suspects. By virtue of independent, reliable detective
work, each has narrowed his list to only two. Now they are engaged
in a telephone call; their object is to compare information, and if their
pairs overlap in just one suspect, to identify the killer.
The difficulty is that their telephone line has been tapped by the
local lynch mob, who know the original list of suspects but not which
pairs the sheriffs have arrived at. If they are able to identify the killer
with certainty as a result of the phone call, he will be lynched before
he can be arrested.
Can the sheriffs, who have never met, conduct their conversation
in such a way that they both end up knowing who the killer is (when
possible), yet the lynch mob is still left in the dark?

176
 Information, Please

Solution: Label the suspects ABCDEF GH and organize them

into lists so that each list consists of four disjoint (unordered) pairs,
and each pair appears in exactly one list. For example:

AB CD EF GH
AC BD EG FH
AD BC EH FG
AE BF CG DH
AF BE CH DG
AG BH CE DF
AH BG CF DE

Let’s name the sheriffs Lew and Ralph. Lew can share the above array
with Ralph over the phone and then tell Ralph which row his pair (i.e.,
the pair of suspects to which Lew has narrowed down his search) lies
in. If Ralph’s pair lies in the same row, he tells Lew; that means the
two sheriffs have arrived at the same pair, so they may as well hang
up.
Otherwise, Ralph announces a partition of the row into sets of two
pairs each, with the property that Ralph’s two final suspects are in the
same part.
For example, suppose Lew’s pair is EF . Then he announces that
his pair is in the first row. If Ralph’s pair is F G, he partitions that row
into one part consisting of the pairs AB and CD, and another con-
taining the pairs EF and GH. (At this point, Lew and Ralph already
have established a shared secret: They both know which part the killer
is in.)
Lew now reveals whether his pair is the leftmost or rightmost in its
part of the row (in this case, it’s the leftmost). That tells Ralph which
pair is Lew’s, and hence who the killer is (here, F ). To communicate
that safely to Lew, he just tells Lew whether the killer is the left or the
right member of Lew’s pair (here, the rightmost).
The conversation would be identical if the pairs had been AB and
BC, making B the killer instead of F . We hope that the mob either
refuses to chance lynching an innocent person, or doesn’t have enough
rope for two.

Cryptography is the study of means for sending private messages

over an open channel. Classically, the sender (Alice, say) and receiver
(Bob) are equipped in advance with a “key” that enables Alice to en-
crypt her message and Bob to decrypt it, while eavesdropper Eve, not
possessing the key, can make no sense of the intercepted stream.

177
 Mathematical Puzzles

Suppose Eve has the time and computing power to test every pos-
sible key, looking for one which, when used to decrypt the message
she intercepted, yields a credible possibility for the message Alice in-
tended Bob to get. When might this work?
If the number of possible keys is small, Eve can expect that only
one will yield an intelligible message; but if it’s large, many might do
so. Information theory can give us an idea of how many possible keys
there should be, in order for Alice and Bob to be safe.
Suppose that for every choice of message Alice might have wanted
to send Bob on this occasion, and every “cyphertext” x1 , . . . , xn that
Eve might intercept, there is a unique key that would encrypt that
message as x1 , . . . , xn . Then, assuming the key was chosen uniformly
at random and Eve doesn’t know it, we have an easy conclusion.
Theorem. In the above situation, Eve can gain no information whatever
about Alice’s intended message.

Proof. Since any message could have been intended, each associ-
ated with one key, and all keys are equally likely, Eve learns nothing.
♡
There actually is a cryptographic scheme that achieves this goal—
assuming accurate random generation and perfect secrecy in the key
choice. It’s called a “one-time pad,” and here’s an example of how it
could work. Alice translates her intended message into a stream of n
0s and 1s (by some means that Bob needs to know and Eve may also
know). Then she extracts the next n bits from the one-time pad that
she and Bob share, and adds those bits one by one, modulo 2, to the
message bits. For example, if the message is 00101110 and the bits
from the one-time pad are 01001001, then the encrypted message is
01100111 and this is what Alice sends (and Eve intercepts).
Bob, who knows where he and Alice are in the one-time pad, takes
his received message 01100111 and, one by one, adds the one-time
pad bits 01001001 to it to get the original message, 00101110, in-
tended by Alice. (This works nicely because modulo 2, addition and
subtraction are equivalent. Much more about this useful kind of arith-
metic is found in Chapter 17.)
Eve is kept completely in the dark, because looking only at
01100111, any eight-bit message could have been intended by Alice.
As its name suggests, the security of the one-time pad is heavily
dependent on each bit of the pad being used only once, and in only
one message. Otherwise a “depth” is created and Eve may be able to
crack the code.

178
 Information, Please

The difficulty with one-time pads is that if Alice and Bob are plan-
ning any long conversations, they need to share (and keep private) a
great many bits. In most applications, keys are much shorter than
messages, and the security of encryption relies on Eve not having the
time and/or computing power to test enough keys.

179
 14. Great Expectation
The expectation or expected value of a random number is the average
value that is anticipated (via the Law of Large Numbers) when the
random number is “redrawn” independently many times. If the ran-
dom number can take only finitely or countably many values (e.g.,
if it takes only integer values), we can define its expectation as the
average of the values, weighted by their probabilities.
Numbers whose values are determined by chance are called by
probabilists random variables, and denoted typically by boldface capital
letters from the end of the Latin alphabet, such as X. If X’s possible
values are x1 , x2 , . . . with corresponding probabilities p1 , p2 , . . . , then
the expectation of X is
∑
EX := p i xi .
i

The terminology is a bit misleading; for example, the expectation

when you roll a fair die is 3 12 , but you don’t “expect” the result 3 12 . It
would be more accurate to call 3 12 the “expected average” rather than
the “expectation” or “expected value.” But convenience triumphs.
Expectation is often easier to work with than probability. Suppose
X and Y are two random variables. If we want to know the probability
that X + Y > 5 (say) it is not enough to know everything there is to
know about each of X and Y individually; you will also need to know
how X and Y interact with each other.
But if you only want the expectation of X + Y, it is guaranteed to
satisfy
E(X + Y) = EX + EY
so all you need to know are the expectation of X and the expectation
of Y. For example, suppose X and Y each take values 1, 2, 3, 4, 5,
or 6 with equal probability, so that each has expectation 3 21 . What
does the random variable X + Y look like? Well, if X and Y are
independent—for example, if X is the first roll of a die and Y the
second—then X + Y takes on values 2, 3, . . . , 12 with probabilities

181
 Mathematical Puzzles

1/36, 2/36, . . . , 5/36, 6/36, 5/36, . . . , 1/36. But if X and Y are both
the outcome of the first roll, that is, if X = Y, then X+Y = 2X takes
on only even values from 2 to 12, each with probability 16 . And if Y
happens to be defined as 7 − X, we find that X + Y is the constant
value 7. All very different, but notice that in every case E(X + Y) =
3 12 + 3 12 = 7.
This miracle is a special case of what mathematicians call “linear-
ity of expectation,” whose general formulation is
∑ ∑
E ci Xi = ci EXi ,
i i

that is, the expectation of a weighted sum of random variables is the

weighted sum of their expectations.
Let’s start with a calculation of expectation in the service of mak-
ing a strategic decision.

Bidding in the Dark

You have the opportunity to make one bid on a widget whose value
to its owner is, as far as you know, uniformly random between $0 and
$100. What you do know is that you are so much better at operating
the widget than he is, that its value to you is 80% greater than its value
to him.
If you offer more than the widget is worth to the owner, he will
sell it. But you only get one shot. How much should you bid?

Solution: You should not bid. If you do bid $x, then the expected
value to the widget’s owner, given that he sells, is $x/2; thus, its expected
value to you, if you get it, is 1.8 · $x/2 = $0.9x. Thus you lose money
on average if you win, and of course, you gain or lose nothing when
you do not, so it is foolish to bet. ♡
One common setting for expectation is waiting time; in particu-
lar, the average number of times you expect to have to try something
until you succeed. For example, how many times, on average, do you
need to roll a die until you get a 6? We encountered this sort of prob-
lem in Odd Run of Heads in Chapter 5. Let x be the expected number
of rolls. “If at first you don’t succeed” (probability 56 ) you still have
an average of x rolls to “try, try again,” so x = 1 + 56 · x giving x = 6.
The same argument shows that in general, if you conduct a series of
“trials” each of which has probability of success p, then the average
number of trials you need to get your first success is 1/p.
Let’s apply this observation to a slightly trickier dice puzzle.

182
 Great Expectation

Rolling All the Numbers

On average, how many times do you need to roll a die before all six
different numbers have turned up?

Solution: Often the most important step in applying linearity of ex-

pectation is simply recognizing that what you are being asked for is
the expectation of a sum of random variables. The key for this puzzle
is to think of rolling all the numbers as a 6-stage process, where at
Stage i you are going for your ith new number. In other words, let Xi
be the number of rolls needed to get your ith new number, given that
you have already seen i−1 different numbers.
Then the total number of rolls needed is X := X1 + X2 + · · · + X6
and what we are asked for is EX = EX1 + EX2 + · · · + EX6 . We
already know that EX1 = 1 since it always takes exactly one roll to
get the first new number. The second will come with probability 5/6
from any given roll, so takes 6/5 rolls on average. Continuing in this
manner, we have
6 6 6 6 6
EX = 1 + + + + + = 14.7
5 4 3 2 1
and we are done.
Warning: Some experimentation will remind you that X is by no
means guaranteed to be near 14.7; the last new number, in particular,
could take a frustratingly long time to roll. Later we’ll be reminded
that the Law of Large Numbers sometimes takes a while to kick in.
Next, see if you can set up a sum of random variables for this
puzzle.

Spaghetti Loops
The 100 ends of 50 strands of cooked spaghetti are paired at random
and tied together. How many pasta loops should you expect to result
from this process, on average?

183
 Mathematical Puzzles

Solution: The key here is to recognize that you start with 100
spaghetti-ends and each tying operation reduces the number of
spaghetti-ends by two. The operation results in a new loop only if the
first end you pick up is matched to the other end of the same strand.
At step i there are 100−2(i−1) ends, thus 100−2(i−1)−1 = 101−2i
other ends you could tie any given end to. In other words, there are
101−2i choices of which only one makes a loop; the others just make
two strands into one longer one.
It follows that the probability of forming a loop at step i is 1/(101−
2i). If Xi is the number of loops (1 or 0) formed at step i, then the
expected value of Xi is 1/(101 − 2i). Since X := X1 + · · · + X50 is
the final number of loops, we apply linearity of expectation to get
1 1 1 1 1
EX = EX1 +· · ·+EX50 = + + +· · ·+ + = 2.93777485 . . . ,
99 97 95 3 1
less than three loops! ♡
For the last two puzzles, the summands were independent random
variables—we didn’t actually need to know that independence is not
required for linearity of expectation. Independence of two events A
and B means that the probability they both occur is the product of
their individual probabilities, so, for example, if we wanted to know
the probability of getting all six numbers in only six rolls of a die, we
could have obtained it as
5 4 3 2 1 6!
1· · · · · = 6 ∼ 1.5%.
6 6 6 6 6 6
For all the spaghetti to end up in one big loop, you must never
tie a strand to its other end except at the last step; in other words,
X1 , . . . , X49 must all be zero. This happens with probability
98 96 94 2
· · · ··· ·
99 97 95 3
49! · 249 49!2 · 298
= = = 0.12564512901 . . . ,
99!/(49! · 249 ) 99!
so about one time in eight.

184
 Great Expectation

Ping-Pong Progression
Alice and Bob play table tennis, with Bob’s probability of winning
any given point being 30%. They play until someone reaches a score
of 21. What, approximately, is the expected number of points played?

Solution: With high probability, Alice will win, and thus the ex-
pected number of points won by Bob is (3/7)21 = 9 and therefore the
expected total number of points is close to 30. Another way to make
that calculation: Since Alice wins any given point with probability
70%, it takes on average 1/0.7 = 10/7 rallies to earn Alice a point,
thus 21 × 10/7 = 30 rallies on average for her to get to 21.
The actual answer is a bit less than 30, because playing until Alice
scores 21 points is sufficient, but not always necessary, to finish the
game. How much less?
To find out, we employ a bit of statistics (readers not familiar with
variance and normal approximations are invited to skip to the next
puzzle).
The probability of Bob winning is equal to the probability that he
wins the majority of the first 41 points. Let Xi = 1 if Bob wins the
ith point, 0 otherwise. The variance of Xi is 0.3 × 0.7 = 0.21, thus
the√standard deviation of the number of points Bob wins out of 41
is 41 × 0.21 = 2.9343. His expected score being 0.3 × 41 = 12.3,
we apply the normal approximation with the intent of computing the
probability that Bob’s score exceeds 20.5; that amounts to (20.5 −
12.3)/2.9343 ∼ 2.795 sigmas, which works out to a probability of
about 0.0000387. If Bob does win, he’ll win by about 2 points on
average, leading to 40 points played instead of the approximately 43
points needed if we wait for Alice to hit 21. Thus, the cost to our 30-
point estimate is around 3 × 0.0000387 ∼ 0.0001.
So the result of all that work is revising our estimate to 29.9999.
Do you recall a puzzle called “Finding a Jack” from a Chapter 8?
You can use its solution technique—twice!—for the next puzzle.

Drawing Socks
You have 60 red and 40 blue socks in a drawer, and you keep drawing
a sock uniformly at random until you have drawn all the socks of one
color. What is the expected number of socks left in the drawer?

Solution: It’s useful to imagine that the sock-drawing is done by

arranging all the socks in a uniformly random order, then removing

185
 Mathematical Puzzles

them, starting from the beginning of the order, until all of one color
have been removed. If the color of the last sock is blue (probability
0.4), then it must have been the red socks that were all removed, and
the number of socks “left in the drawer” is the number of blue socks
encountered starting from the end of the arrangement before any red
sock is found.
If indeed the last sock is blue, there are 39 other blue socks dis-
tributed in 61 slots (before the first red sock, between the first and sec-
ond red socks, and so forth, finally after the 60th red sock). Thus, we
expect 39/61 blue socks in that last slot; adding the final blue sock
gives on average of 100/61 blue socks left in the drawer. Similarly,
if the last sock is red (probability 0.6), we have 59 other red socks in
41 slots so on average 59/41 in the last slot; that leaves 1 + 59/41
= 100/41 red socks behind. Putting these together, we have that the
expected number of socks left is 0.4 × 100/61 + 0.6 × 100/41 ∼ 2.12.
If there are m socks of one color and n of the other, the above
argument gives you an average of m/(n+1) + n/(m+1) socks left in
the drawer.
Linearity of expectation applies even when there are continuum-
many random variables. Modern mathematicians would replace the
sum by an integral, but sometimes an Archimedean approach will do
the trick.

Random Intersection
Two unit-radius balls are randomly positioned subject to intersecting.
What is the expected volume of their intersection? For that matter,
what is the expected surface area of their intersection?

Solution: We can assume that the center of the first ball is at the
origin of a three-dimensional coordinate system; the first ball thus
consists of all points in space within distance 1 of the origin. The cen-
ter C of the second ball must be somewhere in the ball of radius two
about the origin, in order for it to intersect the first ball. The condi-
tions of the problem imply that C is uniformly random subject to this
constraint.
(Note that this is not at all the same as taking the center of the
second ball to be a random point between −2 and 2 on the x-axis.
The latter assumption would give a much-too-big expected intersec-
tion size of π/2.)

186
 Great Expectation

Now consider a point P inside the first ball: What’s the probability
that P will be in the intersection of the two balls? That will happen
just when C falls in the unit ball whose center is P . This ball is inside
the ball of radius 2 about the origin that C is chosen from, with 1/23 =
1/8 of the volume. Thus the probability that P lies in our intersection
is 1/8, and it follows from linearity of expectation that the expected
volume of the intersection is 1/8 times the volume of the first ball,
that is, 1/8 × 4π/3 = π/6.
If it seems fishy to you to be adding up points P like this, good
for you! Think of P not as a point but a small cell, or “voxel,” inside
the ball. If the voxel’s volume is (4π/3)/n with the ball divided into
n voxels, n being some large integer, then an ordinary finite applica-
tion of linearity of expectation gives approximately the same answer,
π/6. It’s not exact owing to the fact that a few voxels will lie partly
inside and partly outside the intersection. As the voxels get smaller
the approximation gets better.
For the surface area of the intersection, the point P —now repre-
senting a two-dimensional pixel—is taken on the surface of the unit
ball centered about the origin. This will lie on the surface of the inter-
section if it’s inside the second ball, that is, again, if C lies in the unit
ball centered at P . We’ve already seen that this happens with prob-
ability 1/8, but note that the intersection of the two spheres has two
surfaces of the same area, one that was part of the surface of the first
ball—that’s the part that could contain P —and one that was part of
the surface of the second ball. Thus the expected surface area of the
intersection is 2 × 18 × 4π = π.
You may have noticed, by the way, that all this can be done in
any dimension; with disks on the plane, for example, you get ex-
pected area π/4 and perimeter π. You can even do it with (some)
other shapes.
One place where linearity of expectation arises frequently is in
gambling. A game is called “fair” if the expected return for each player
is 0, and we can deduce that any combination of fair games is fair.
This applies even if your choice of game, or strategy within a game,
is made “as you play” (as long as you can’t foresee the future).
Of course, in a gambling casino, the games are not fair; your ex-
pectation is negative, the casino’s positive. For example, in (Amer-
ican) roulette there are 38 numbers (0, 00, and 1 through 36) on
the wheel, but if you bet $1 on a single number and win, you earn
only $36, not $38, in return for your dollar. The result is that your

187
 Mathematical Puzzles

expectation is
1 37 1
· $35 + · (−$1) = −$ ,
38 38 19
so that you lose, on average, about a nickel per dollar bet. This same
negative expectation applies also to bets made on groups of numbers
or colors.
However:

Roulette for the Unwary

Elwyn is in Las Vegas celebrating his 21st birthday, and his girlfriend
has gifted him with twenty-one $5 bills to gamble with. He saunters
over to the roulette table, noting that there are 38 numbers (0, 00, and
1 through 36) on the wheel. If he bets $1 on a single number, he will
win with probability 1/38 and collect $36 (in return for his $1 stake,
which still goes to the bank). Otherwise, of course, he just loses the
dollar.
Elwyn decides to use his $105 to make 105 one-dollar bets on the
number 21. What, approximately, is the probability that Elwyn will
come out ahead? Is it better than, say, 10%?

Solution: We know from linearity of expectation that on average,

19 ∼ $5.53. But this says very little
Elwyn should expect to lose $ 105
about the probability that Elwyn comes out ahead. A random variable
with expectation −5.53 could be anywhere from always negative, to
almost never negative!
In this case, we need to work out the probability that Elwyn wins
three or more of his 105 bets, because if he wins three times he pulls
in 3 · $36 = $108 and comes out ahead.
That takes a little work. The probability that Elwyn loses every
time is (37/38)105 ∼ 0.06079997242. To compute the probability that
he wins exactly once, we multiply the number of bets by the probabil-
ity of winning a given bet and losing the rest; this gives us
( )104
1 37
105 · · ∼ 0.17254046227.
38 38

The probability that Elwyn wins exactly twice is the number of pairs
of bets times the probability of winning two specified bets and losing

188
 Great Expectation

the rest, which works out to

( )2 ( )103
105 · 104 1 37
· · ∼ 0.24248929833.
2 38 38

Adding these three numbers and subtracting from 1 gives the proba-
bility that Elwyn comes out ahead: 0.52417026698, better than 50%!
No, this does not mean that Elwyn has figured out how to beat
’Vegas. He comes out slightly ahead rather often but far behind quite
often as well, since winning only twice leaves him with just $72 to
show for his $105 stake. Elwyn’s negative expectation is the more rel-
evant statistic here.
If you want to go to Vegas and be able to tell your friends that you
made a profit at roulette, here’s a suggestion: Pick a number of dollars
that’s one less than a power of 2, and that you can afford to lose.
Suppose it’s $63. Then bet $1 on “red.” (Half the positive numbers
are red, so you get your dollar back and win an additional $1 with
probability 18/38, but lose your dollar otherwise.) If you do win, go
home. If you lose, bet $2 on red; if you win, go home having made a
$1 profit. If you lose, bet $4 next time, etc., until you win or have lost
your sixth straight bet ($32). At this point, you must give up or risk
ruin; fortunately the probability that you will be this unlucky is only
(20/38)6 , about 2%. Your expectation for the whole experiment is of
course still negative.

An Attractive Game
You have an opportunity to bet $1 on a number between 1 and 6.
Three dice are then rolled. If your number fails to appear, you lose
your $1. If it appears once, you win $1; if twice, $2; if three times, $3.
Is this bet in your favor, fair, or against the odds? Is there a way to
determine this without pencil and paper (or a computer)?

Solution: This game seems pretty decent, maybe even favorable.

You have three chances to get your number; if you were guaranteed
to roll three different numbers, the probability that one would be yours
would be 50%, making the game perfectly fair. If you don’t roll three
different numbers, you could end up with a bonus. So what’s not to
like?
In fact this is a well known game, often called “Chuck-a-Luck” or
“Birdcage” (the latter because the dice are typically kept inside a cage

189
 Mathematical Puzzles

and rolled by shaking the cage). Like any gambling game it is designed
to look attractive but to make money for the house, and indeed that
is the case.
The easiest way to see this is to imagine that there are six players
each of whom bets on a different number. The expectation for each
player is of course the same, since the rules don’t discriminate. If that
expectation is x then the house’s expectation must be −6x by linearity
of expectation—this is a “zero-sum game,” since no money leaves or
enters from outside.
But now if the three rolls are different, three players win and three
lose so the house breaks even. If two numbers are the same, the house
pays $2 to one player and $1 to another while picking up $1 from each
of the remaining four players, so makes a $1 profit. If all three dice
come up the same, the house pays out $3 and collects $5, making a $2
profit. So the house’s expectation is positive and therefore the players’
is negative.
(It’s not hard to work out that the probability the three rolls are
different is (6·5·4)/63 = 5/9, the probability that they are all the same
is 1/62 = 1/36, and thus the probability that just two of them match is
the remaining 15/36 = 5/12. Thus the house’s expectation is (5/12) ·
$1+(1/36)·$2 = $17/36 and each player’s expectation is thus (−1/6)·
$17/36 ∼ −$0.0787037037, so a player loses on the average about 8
cents per dollar bet—worse even than American roulette, about which
more later.)
Here’s a game which is definitely favorable; the question is, to
what extent can you turn positive expectation into a sure thing?

Next Card Bet

Cards are turned face up one at a time from the top of a well-shuffled
deck. You begin with a bankroll of $1, and can bet any fraction of your
current worth, prior to each revelation, on the color of the next card.
You get even odds regardless of the current composition of the deck.
Thus, for example, you can decline to bet until the last card, whose
color you will of course know, then bet everything and be assured of
going home with $2.
Is there any way you can guarantee to finish with more than $2? If
so, what’s the maximum amount you can assure yourself of winning?
Solution: With a little thought you can improve on $2 by waiting
to bet until three cards remain. If they’re all one color, great—you can

190
 Great Expectation

bet everything for the remaining three rounds and go home with $8.
If one is red and the others black, bet 1/3 of your money on black;
if you’re right you now have $1 13 which you can double at the last
card for a final tally of $2 23 . If you’re wrong you’re down to $ 32 but
the remaining cards are both black, so you can double this twice to
again end with $2 23 . Since a similar strategy works when the remain-
ing cards are two red and one black, you have found a way to guar-
antee $2 23 .
To extend that backward and guarantee even more starts to look
messy, so let’s look at the problem another way. Whatever betting
strategy you use, you’ll certainly want to bet everything at every round
as soon as the remaining deck becomes monochromatic; let’s call
strategies with this property “sensible.”
It is useful first to consider which of your strategies are optimal
in the sense of expectation, that is, which maximize your expected
return. It is easy to see that all such strategies are sensible.
Surprisingly, the converse is also true: No matter how crazy your
betting is, as long as you come to your senses when the deck becomes
monochromatic, your expectation is the same! To see this, consider
first the following pure strategy: Imagine some fixed specific distribu-
tion of red and black in the deck, and bet everything you have on that
distribution at every turn.
Of course, you will nearly always go broke with this strategy, but
( )is then 2 × $1,
52
if you win you can buy the earth—your take-home
around 4.5 quadrillion dollars. Since there are 52 ways the colors
26 ( )
52 52
can be distributed in the deck, your expected return is $2 / 26 =
$9.0813.
Of course, this strategy is not realistic, but it is “sensible” by our
definition, and, most importantly, every sensible strategy is a combination
( )
of pure strategies of this type. To see this, imagine that you have 52 26
assistants working for you, each playing a different one of the pure
strategies.
We claim that every sensible strategy amounts to distributing your
original $1 stake among these assistants, in some way. If at some point
your collective assistants bet $x on red and $y on black, that amounts
to you yourself betting $x − $y on red (when x > y) and $y − $x on
black (when y > x).
Each sensible strategy can be implemented by a distribution of
money to the assistants, as follows. Say you want to bet $0.08 that the
first card is red; this means that the assistants who are guessing “red”
first get a total of $0.54 while the others get only $0.46. If, on winning,

191
 Mathematical Puzzles

you plan next to bet $0.04 on black, you allot $0.04 more of the $0.54
total to the “red–black” assistants than to the “red–red” assistants.
Proceeding in this manner, eventually each individual assistant has
his assigned stake.
Now, any mix of strategies with the same expectation shares that
expectation, hence every sensible strategy has the same expected re-
turn of $9.08 (yielding an expected profit of $8.08). In particular, all
sensible strategies are optimal.
But one of these strategies guarantees $9.08; namely, the one in
which the $1 stake is divided equally among the assistants. Since we
can never guarantee more than the expected value, this is the best
possible guarantee.
This strategy is actually quite easy to implement (assuming as we
do that US currency is infinitely divisible). If there are b black cards
and r red cards remaining in the deck, where b ≥ r, you bet a fraction
(b − r)/(b + r) of your current worth on black; if r > b, you bet
(r − b)/(b + r) of your worth on red. You will be entirely indifferent
to the outcome of each bet, and can relax and collect your $8.08 profit
at the end.
Sometimes the notion of fair game comes in handy even when
there is no mention of a game.

Serious Candidates
Assume that, as is often the case, no one has any idea who the next
nominee for President of the United States will be, of the party not
currently in power. In particular, at the moment no person has prob-
ability as high as 20% of being chosen.
As the politics and primaries proceed, probabilities change con-
tinuously and some candidates will exceed the 20% threshold while
others will never do so. Eventually one candidate’s probability will
rise to 100% while everyone else’s drops to 0. Let us say that after the

convention, a candidate is entitled to say he or she was a “serious”

192
 Great Expectation

candidate if at some point his or her probability of being nominated

exceeded 20%.
Can anything be said about the expected number of serious can-
didates?
Solution: It seems that you’d need to know a lot about the political
process, and the current circumstances, to answer this question. For
example, doesn’t it seem plausible that under some conditions, the
first candidate to exceed 20% is likely to continue rising and eventu-
ally secure the nomination? Wait, that can’t be right; if he/she really
had probability only 20% at the critical point, then most likely some-
one else would end up with the nomination.
In fact, the expected number of “serious” candidates has to be 5.
To see this, imagine that you enter the prediction market with the
following strategy: Anytime a candidate hits 20% for his or her first
time, you bet $1 on that candidate. If the market’s probabilities are
“correct” and the market is fair, you should win $5 if that candidate
is eventually nominated.
In this game, you are bound to win $5; no matter which candidate
gets nominated, you will have bet on him or her at some point. So
the only variable is how many bets you made, and since the game is
presumed to be fair, your expected expenditure must be $5.
Why is the word “correct” in quotes above? The problem is that
when everything is done and (say) Smith has won the nomination,
one might argue that a good political prognosticator would, or should,
have known all along that that would be the outcome. Thus when
Smith hit 20% he or she should probably have been rated at 100%, or
at least a lot higher than 20%. Without getting into the philosophy of
probability, let’s just say that a candidate’s rating represents some sort
of consensus which you, the bettor, have no reason to doubt one way
or the other.

Rolling a Six
How many rolls of a die does it take, on average, to get a 6—given
that you didn’t roll any odd numbers en route?
Solution: Obviously, three. Right? It’s the same as if the die had
only three numbers, 2, 4, and 6, each equally likely to appear. Then
the probability of rolling a 6 is 1/3, and as we have seen in the puz-
zle Rolling All the Numbers, in an experiment that has probability of
success p, it takes, on average, 1/p trials to get a success.

193
 Mathematical Puzzles

Except that it’s not the same at all. Imagine that you arrange to
compute your answer to this puzzle via multiple experiments, mak-
ing use of the law of large numbers. You start rolling your die, count-
ing trials as you go. If you hit a “6” before you get an odd number,
you record the number of trials and repeat. (For example, if you roll
“4,4,6” you record a 3.) What happens if you do roll an odd number?
Then you can’t simply throw out that roll and continue counting trials
(if you did that, you would eventually converge to 3 as your answer).
Instead, you must throw out that whole experiment, recording noth-
ing, and start anew, beginning with roll #1 of your new experiment.
That will produce a smaller average than 3, but how much smaller?
An intuitive way to look at it: When you’ve rolled a 6 with-
out rolling an odd number, there’s an inference that you succeeded
quickly—since if you took a long time to get your 6, you’d probably
have hit an odd number.
To figure out the correct answer, it helps to think about
what you’re doing in your multi-experiment. Basically, each sub-
experiment consists of rolling the die until you get a 6 or an odd num-
ber. If you counted rolls in all of these, you’d discover their average
length to be 3/2, since the probability of rolling a 6, 1, 3, or 5 is 2/3.
In your gedankenexperiment you’re only counting the sub-experiments
that end in a 6, but does that matter? The finishing roll (be it 1, 3, 5,
or 6) is independent of all that has gone before, and independence is
a reciprocal notion; thus, the number of rolls “doesn’t care” what the
sub-experiment ended with. It follows that the average length of the
subexperiments that ended with a 6 is that same 3/2.

Napkins in a Random Setting

At a conference banquet for a meeting of the Association for Women
in Mathematics, the participants find themselves assigned to a big cir-
cular table. On the table, between each pair of adjacent settings, is
a coffee cup containing a cloth napkin. As each mathematician sits
down, she takes a napkin from her left or right; if both napkins are
present, she chooses randomly. If the seats are occupied in random
order and the number of mathematicians is large, what fraction of
them (asymptotically) will end up without a napkin?

Solution: We want to compute the probability that the diner in po-

sition 0 (modulo n) is deprived of a napkin. The limit of this quantity,
as n → ∞, is the desired limiting fraction of napkinless diners.

194
 Great Expectation

We may assume that everyone decides in advance whether to go

for her right or left napkin, in case both are available; later, of course,
some will have to change their minds or go without.
Say that diners 1, 2, . . . , i−1 choose “right” (away from 0) while
i chooses “left”; and diners −1, −2, . . . , −j +1 choose “left” (again,
away from 0) while diner −j chooses “right.” Note that i and j might
both be equal to 1, meaning that the diners on both sides of diner 0
are planning to go for the napkin between them and diner 0.
If k = i + j + 1, then as long as k ≤ n (which is the case with high
probability), the probability of this configuration is 21−k .
Observe that diner 0 loses out only when she is last to pick among
−j, . . . , i and none of the diners −j +1, . . . , −2, −1; 1, 2, . . . , i−1 get
the napkins they wanted. If t(x) is the time at which diner x makes
her grab, then this happens exactly when t(0) is the unique local max-
imum of t in the range [−j, −j +1, . . . , 0, . . . , i−1, i].
If t is plotted on this interval, it looks like a mountain with (0, t(0))
on top; more precisely, t(−j) < t(−j +1) < · · · < t(−1) < t(0) >
t(1) > t(2) > · · · > t(i).
Instead of evaluating the probability of this event for fixed i and j,
it is convenient to lump all pairs (i, j) together which satisfy i+j+1 =
k for fixed k. Altogether there are k! ways the values t(−j), . . . , t(i)
can be ordered. If T is the set of all k grabbing times and tmax is
the last of these, then each mountain-ordering is uniquely identified
by the nontrivial subset of T \ {tmax } which constitutes the values
{t(1), . . . , t(i)}. Thus, the number of orderings that make a valid
mountain is 2k−1 − 2.
Finally, the total probability that diner 0 is deprived of her napkin
is
∑∞
21−k · 2k−1 − 2 ( √ )2
= 2 − e ≈ 0.12339675. ♡
k!
k=3

195
 Mathematical Puzzles

Roulette for Parking Money

You’re in Las Vegas with only $2 and in desperate need of $5 to feed a
parking meter. You run in through the nearest door and find yourself
at a roulette table. You can bet any whole dollar amount on any al-
lowable set of numbers. What roulette-betting strategy will maximize
your probability of walking out with $5?

Solution: There are two major considerations in trying to maxi-

mize your probability of boosting your fortune to $5. One is to avoid
waste: That is, try to hit $5 exactly. Overshooting gains nothing and
since your expectation is limited by your initial $2, minus what the
“house” is expecting to take from you, you don’t want to spend it on
overshooting your goal.
The second is speed. As we saw above in Roulette for the Unwary, it
costs you about a nickel in expectation every time you bet $1 playing
roulette, no matter what you bet on.
Of course, you must do some betting. Ideal would be if there were
a roulette bet that paid off 5:2, so you could just put your $2 down
once and win what you need or lose everything in one shot.
Unfortunately there is no such bet in roulette, so it looks like you
are doomed to either risk overshooting, or risk having to make many
bets. Example: You could bet $1 on the numbers 1 through 8. If you
win you’re home, otherwise you could bet your remaining $1 on red
and if it wins, start over. That gives you probability p of success where
8 30 18
p= + · · p.
38 38 38
That gives p = 38 8
/(1 − 30
38 · 38 ) = 0.33628318584, about 1/3.
18

But you can do better by making correlation work for you. How?
By making simultaneous bets of $1 each of the right kind. Specifically:
Bet one of your dollars on the numbers 1 through 12, and simultane-
ously, the other on 1 through 18. Then, if you do get a number between
1 and 12, you hit your $5 on the nose. If you hit a high number, you
lose it all (but at least you’ve lost it all quickly). If you happen to hit
13, 14, 15, 16, 17, or 18, you again have $2 and can repeat your double
bet.
The probability p of success for this scheme is given by
12 6
p= + · p,
38 38
which yields p = 1238 /(1 − 38 ) = 3/8 = 0.375, quite a bit better. In
6

fact this scheme can be shown to be optimal.

196
 Great Expectation

Next is a puzzle that you might find familiar. Posed and solved by
Georges-Louis Leclerc, Comte de Buffon in 1733, it is said to be the
very first solved problem in geometric probability.
But to solve it we will use almost no geometry, relying instead on
linearity of expectation.

Buffon's Needle
A needle one inch in length is tossed onto a large mat marked with
parallel lines one inch apart. What is the probability that the needle
lands across a line?

Solution: Let C be a “noodle”: any smooth plane curve—self-

intersections permitted—of length, say, ℓ. Toss the noodle randomly
onto the mat, and let XC be the number of crossings you get, that is,
points where C crosses a line. We claim that the average number of
crossings, EXC , is proportional to ℓ, irrespective of the choice of C!
The figure below shows the results of five tosses of a particular noodle,
each accompanied by its number of crossings.

To verify the claim, imagine that C is composed of many short

line segments, each of length ε. Throwing just one line segment L of
length ε < 1 onto the mat produces a crossing with some probability
p, but never more than one crossing, thus EXL = p. At the moment,
we don’t know what p is.
But, by linearity of expectation, we know that if C is composed of
n such segments, then EXC = np. It follows, by letting ε shrink, that
EXC = αℓ for some constant α, since a smooth C is approximable
by a string of short line segments. If we can determine what α is, we
can infer that EXN = α, since Buffon’s needle N has length 1.
To find α we only have to pick C cleverly, and this is how: Let C
be a circle of diameter 1! Then no matter how C lands on the mat, it
produces exactly 2 crossings. Thus EXC = 2 = α · π, and we solve to

197
 Mathematical Puzzles

get α = 2/π. We deduce that EXN = 2/π, and since (with probability
1) the needle either produces one crossing or none, the probability that
Buffon’s needle crosses a line is that same 2/π. ♡
This wonderful proof was published by T. F. Ramaley in 1969, in
a paper entitled “Buffon’s Noodle Problem.”
Here are a couple of more modern geometry problems, that seem
not to have any probability in them.

Covering the Stains

Just as a big event is about to begin, the queen’s caterer notices, to his
horror, that there are ten tiny gravy stains on the tablecloth. All he
has the time to do is to cover the stains with non-overlapping plates.
He has plenty of plates, each a unit disk. Can he do it, no matter how
the stains are distributed?
Solution: The naked mathematical question being asked here is:
Can any set of 10 points in the plane be covered by disjoint unit disks?
Sure, if the points are close together, you can cover them with one
disk; if they’re spread widely apart, you can cover each with its own
disk. It’s the medium distances that could cause problems.
It’s a good idea, as advertised in Chapter 7, to think about replac-
ing 10 by other numbers. It’s pretty obvious that any one or two points
can be covered, and easy to show that any three can be covered.
On the other hand, the answer is certainly “no” if 10 is replaced by
10,000. The reason is that you can’t “tile” the plane with unit disks;
in other words, you can’t arrange disjoint unit disks so as to cover the
whole plane (as you could, say, with unit squares). If you put down
10,000 points in a 100 × 100 grid pattern covering a square somewhat
bigger than a unit disk, it won’t be possible to cover every grid point
with disjoint unit disks.
So it’s a question of quantity—is 10 more like 3, or more like
10,000? Here’s a thought—how much of the plane can be covered by
disjoint unit disks? You would probably guess that the best way to
pack unit disks is in a hexagonal array, as shown in the figure below.
In fact, it is provable that this is the best packing.
How good is it—in other words, if you choose a large region of the
plane (say, a 1,000,000 × 1,000,000) square) and pack it hexagonally
with disks, what fraction of the region will be covered? In the limit,
that fraction must be the fraction of the area of one of those hexagonal
cells that is occupied by an inscribed unit disk. Of course, the disk,

198
 Great Expectation

having radius 1, has area π. The hexagon

√ is made up of six√equilateral
√
triangles of altitude 1 and area 1/ 3, thus total area 6/ 3 = √ 12.
It follows that the fraction of area covered by the disk is π/ 12 ∼
0.9069.
Hmm. The disks cover more than 90% of the plane, that is, more
than 9 points out of 10. Can we make use of that? Trouble is, the gravy
stains are not random points, they are (in the worst case) arranged as
if by an adversary.
But we can shift the hexagonal disk-packing randomly. To do this,
just fix some hexagonal cell and pick a point in it uniformly at ran-
dom; then shift (without rotating) the whole picture by moving the
center of the cell to the selected point. Now we can say that for any
fixed point of the plane, the probability that it is covered by some disk
of the randomly shifted configuration is 0.9096.
It follows by linearity of expectation that the expected number of
our 10 gravy stains covered by disks of the randomly shifted hexago-
nal packing is 9.096 stains, that is, on average more than 9 of the 10
points are covered. But then some configurations must cover all 10
points, and the problem is solved. ♡
This technique, in which we prove that something can be done by
showing that if we do it randomly it succeeds with probability > 0,
is often known by the fancy name “the probabilistic method” and asso-
ciated with Paul Erdős and Alfréd Rényi. We’ll see another example
in the theorem at the end of this chapter.

199
 Mathematical Puzzles

The next puzzle provides another example where introduction of

randomness is useful.

Colors and Distances

In the town of Hoegaarden, Belgium, exactly half the houses are oc-
cupied by Flemish families, the rest by French-speaking Walloons.
Can it be that the town is so well mixed that the sum of the distances
between pairs of houses of like ethnicity exceeds the sum of the dis-
tances between pairs of houses of different ethnicity?

Solution: No. Represent each house in Hoegaarden as a distinct

point in the plane, and fix a large disk that contains all the points.
Now consider a random line that intersects the disk. (To get such a
line, choose a radius of the disk uniformly at random, then take the
line perpendicular to the radius that crosses it at a uniformly random
point along the radius.)
Your random line will cut the house-points into two sets (one of
which could be empty); we claim it separates at least as many house-
pairs of different ethnicity as pairs of the same ethnicity. Why? Sup-
pose there are n houses of each ethnicity, and that f Flemish houses
and w Walloon houses lie on one side of the line. Then the number
of heterogeneous crossings is (n − w)f + (n − f )w while the number
of homogeneous crossings is (n − w)w + (n − f )f ; the former minus
the latter is f 2 + w2 − 2wf = (f − w)2 ≥ 0.
Now we make use of the critical fact that for any two particular
points, the probability that they are separated by our random line is
proportional to the distance between the points. Adding expectations,
we deduce that the total distance between heterogeneous (resp., ho-
mogeneous) pairs of points is proportional to the expected number
of separated heterogeneous (resp., homogeneous) pairs. Since every
line cuts at least as many heterogeneous as homogeneous pairs, we
conclude that the former distances add up to at least as much as the
latter.
Note that for distinct points the inequality is strict, because as long
as there is one point which is not simultaneously of both ethnicities,
there is positive probability that a line will result in f ̸= w causing an
imbalance in favor of heterogeneous pairs.
The next puzzle also has a geometric component, but here it is
less surprising that we use linearity of expectation.

200
 Great Expectation

Painting the Fence

Each of n industrious people chooses a random point on a circular
fence, and begins painting toward her farthest neighbor until she en-
counters a painted section. On average, how much of the fence gets
painted? How about if instead, each person paints toward her nearest
neighbor?

Solution: For n = 2, the answers are easily determined to be 3/4

and 1/4, respectively. For larger n, we consider any interval I be-
tween painters and compare it to its two adjacent intervals. In Case
(A), where each painter paints toward her farthest neighbor, I will
go unpainted exactly when it is the shortest of the three; in Case (B),
when it is the longest. Each of these events happens with probability
1/3.
It seems like we’re going to need to know the expected length of the
shortest (resp., longest) of three intervals chosen by cutting a given in-
terval J at two random points. You can get this noting that choosing
two random points from the unit interval is equivalent to choosing
a uniformly random point in an equilateral triangle; the barycentric
coordinates of the point give your interval lengths. Then, if you iden-
tify the regions of the triangle which result in the middle third being
shortest (resp. longest) of the three intervals and compute conditional
expectations, you get 1/9 and 11/18, respectively.
That calculation was for the unit interval, so if the length of I
together with its two neighbor intervals is x, then the expected length
of I given that it is the shortest is x/9 and given that it is the longest,
11x/18.
On average, the total length of I and its two neighbors is 3/n since
adding up this figure for all I counts every interval three times. But in
both versions, I is unpainted only a third of the time (in expectation).
It follows that in Case (A) the expected amount of unpainted fence is
n 3/n n 11·3/n
3 · 9 = 1/9, and in Case (B), 3 · 18 = 11/18.
For the next puzzle, it is useful to have an alternative formula for
expectation. Suppose X is a “counting” random variable, that is, a
random variable that takes values in the non-negative integers. Then
∞
∑
EX = P(X > i).
i=0

To see that this formula is equivalent to the definition at the top of

the chapter, we just need to verify that for each j, the probability that

201
 Mathematical Puzzles

X = j is counted j times in the sum. But that’s easy: It’s counted once
for each i between 0 and j−1, so j times as required.

Filling the Cup

You go to the grocery store needing one cup of rice. When you push
the button on the machine, it dispenses a uniformly random amount
of rice between nothing and one cup. On average, how many times
do you have to push the button to get your cup?

Solution: We know that at least two pushes are necessary, and in

fact two will suffice just half the time; and sometimes more than three
will be required. So the answer must be more than 2 12 . Would you
believe 2.718281828459045?
Let X1 , X2 etc., be the amount dispensed, and let Yi be the frac-
tional part of X1 + · · · + Xi . Then the Yi ’s, like the Xi ’s, are inde-
pendent and uniformly random in the unit interval.
The first decrease in the values of the Yi ’s signals the point when
the amount of rice exceeds 1 cup. Thus, the question at hand is equiv-
alent to: What is the expected value of Z, the length of the longest
increasing initial sequence of the Yi ’s?
The probability that the first j values are increasing is 1/j!, since
the increasing order is just one of j! permutations. Here’s where we
use the special formula above for the expected value of a counting
random variable:
∞
∑ ∞
∑
EZ = P(Z > j) = 1/j!
j=0 j=0

which is just exactly e, Euler’s number (also known as Napier’s

constant, but, naturally, discovered by yet someone else—Jacob
Bernoulli).

It’s time for our theorem, which, as promised, will be proved using
the probabilistic method.
Arguably the most famous theorem in combinatorics, the follow-
ing remarkable fact was proved in 1930 by Frank Plumpton Ramsey,
scion of a famous family of British intellectuals (and brother of Arthur
Michael Ramsey, Archbishop of Canterbury). In its simplest finite
form, Ramsey’s Theorem states that for any positive integer k there
is a number n such that if you color every unordered pair of numbers

202
 Great Expectation

from the set {1, . . . , n} either red or green, then there must be a set
S of k numbers which is “homogeneous” in the sense that every pair
from S is the same color.
The least n for which this is true is called the “kth Ramsey num-
ber” and denoted R(k, k). For example, R(4, 4) turns out to be 18;
that means that for any two-coloring of the pairs of numbers between
1 and 18, there’s a homogeneous set of size 4; but this does not hold
for all colorings of pairs from the set {1, . . . , 17}. But R(5, 5) is not
known! Paul Erdős, who wrote more papers than any other mathe-
matician in history and was fascinated by Ramsey’s theorem, enjoyed
giving the following advice. If a powerful alien force demands, on
penalty of Earth’s annihilation, that we tell them the value of R(5, 5),
we should put all the computing power on the planet to work and
perhaps succeed in computing that number.
But if they ask for R(6, 6), we should attack them before they at-
tack us.
It is not hard to prove—and, ( indeed,
) we will prove it in Chap-
ter 15—that R(k, k) is at most 2r−2 2k
r−1 , which in turn is less than 2 .
What we’re going to do is use linearity of expectation to get an expo-
nential lower bound for R(k, k).
Theorem. For all k > 2, R(k, k) > 2k/2 .
The idea of the proof, published by Erdős in 1947, is to show that
a random coloring of the pairs of numbers from 1 to n = 2k/2 will,
with positive probability, have no monochromatic set of size k. Thus,
such colorings exist (although the proof does not exhibit one).
By a random coloring, we mean this: For each pair {i, j} of
numbers between 1 and n, we flip a fair coin to decide whether to
color
(k) it red or green. If we pick a fixed set S of size k, it contains
2 = k(k−1)/2 pairs and thus will be homogeneous with probabil-
2
ity 2/2k(k−1)/2 = 2(k+1)/2−k /2 .
The number of candidates for( the ) set S, that is, the2 number of
subsets of {1, . . . , n} of size k, is nk < (2k/2 )k /k! = 2k /2 /k!, thus
by linearity of expectation, the expected number of homogeneous sets
is less than
2 2 2(k+1)/2
2(k+1)/2−k /2 · 2k /2 /k! =
k!
which is equal to 1 when k = 2 and declines as k increases.
So the expected number of homogeous sets is less than 1, and we
claim that means the probability that there are no homogeneous sets
is greater than zero. One way to see this is to use the formula for

203
 Mathematical Puzzles

expectation of a counting random variable from Filling the Cup above.

The expected value of X, if X is the number of homogeneous sets,
equals Pr(X > 0) plus more non-negative terms; so if EX is less than
1, so is Pr(X > 0). ♡
The theorem tells us that R(5, 5) is at least ⌈2(5/2)⌉ = 6 and our
later induction proof implies an upper bound of 84 = 70. In fact
R(5, 5) is, at the time of this writing, known to be between 43 and 48
inclusive.

204
 15. Brilliant Induction
Induction (from the verb induce, not induct!) is one of the most elegant
and effective tools in mathematics. In its simplest form, you want to
prove something for all positive integers n; you do it by proving the
statement for n = 1 and then showing that for all n > 1, if it’s true
for n−1, it’s true for n.
More generally, when proving your statement for n, you may as-
sume it’s already true for all positive integers m < n. That assumption
is known as the induction hypothesis, or “IHOP” for short (see Notes &
Sources).
I also like the following formulation, which is even more general.
You want to show that some statement is true in all instances satis-
fying certain conditions. You do this by assuming that there is some
instance in which it is false, and focusing on such an instance which
is in some sense minimal. Then you show how this bad instance can
be turned into another bad instance that is even smaller, contradicting
your assumption that the instance you began with was minimal. You
conclude that there are no bad instances, therefore your statement is
always true.
So in a sense, induction is a special case of reductio ad absurdum—in
other words, showing that a contrary assumption leads to a contra-
diction.
Is this all a bit too abstract for you? That’s what the puzzles are
for.

Uniform Unit Distances

Can it be that for any positive integer n, there’s a configuration of
points on the plane with the property that every point is at distance 1
from exactly n other points?

Solution: For n = 1, the endpoints of a unit-length line segment

will do. For n = 2, we can translate the segment by a unit distance,
and add the result to the original segment, to get four points in a di-
amond. In what direction? Almost any direction will do (as long as

205
 Mathematical Puzzles

it’s not at an angle to the original segment that is multiple of 60◦ ). So

if we make it a random direction, it’ll work with probability 1.
We now proceed in that manner by induction. Assume the set Sn
of points (containing 2n points, but we don’t care) has the property
that every point in it is at unit distance from exactly n others. Now
translate Sn by a unit distance in a random direction to form the set
Sn′ , and let Sn+1 be the union of Sn′ and Sn . Each point of Sn+1 is
now at unit distance from n others plus the translated copy of itself,
for a total of n + 1. (See figure below for the construction in the n = 2
case.) Only finitely many translation angles could cause an “acciden-
tal” unit distance, so with probability one this works. ♡

Swapping Executives
The executives of Women in Action, Inc., are seated at a long ta-
ble facing the stockholders. Unfortunately, according to the meeting
organizer’s chart, every one is in the wrong seat. The organizer can
persuade two executives to switch seats, but only if they are adjacent
and neither one is already in her correct seat.
Can the organizer organize the seat-switching so as to get every-
one in her correct seat?

Solution: Yes. Let the positions (and the executives that belong in
them) be numbered 1 to n from left to right. By induction on n, it
suffices to get executives m through n into their correct seats (for some
m ≤ n) provided the rest are still all incorrectly seated.
To do that we move executive n to the right by successive swaps
until she encounters an executive, say the ith, who’s sitting in posi-
tion i+1 and therefore would be swapped into her correct seat if we

206
 Brilliant Induction

proceeded. It may also be that executive i+1 is in seat i+2, executive

i+2 in seat i+3, etc. forming what we will call a “blockade” contain-
ing at least executive i.
If the blockade goes all the way to the end of the row, we go ahead
and perform all the swaps, ending with executives i through n all in
their correct seats while the rest remain to be sorted.
If the blockade terminates before the end, let executive j be the
executive immediately to the right of the blockade. Instead of swap-
ping n with i, we swap j toward the left until j swaps with n; at no
point in this process will j be in her correct seat because all the seats
she is swapped into belong to members of the blockade.
Now we redirect our attention to n, who has moved one step to the
right, and continue swapping her to the right until the next encounter
with some executive that starts a shifted block, handling that situation
as before.
Eventually executive n reaches the right-hand end and the induc-
tive proof is complete. ♡
How many swaps might be needed for this? While executive n is
being dealt with, no executive i can be in a blockade twice, because
when she is in a blockade she is swapped to the right (from seat i+1
to seat i+2) and then passed by executive n. We conclude that it takes
no more than 2(n−1) steps to handle executive n, thus at most 2(1 +
2 + · · · + n−1) = n(n−1) swaps to get everyone into her correct seat.

Odd Light Flips

Suppose you are presented with a collection of light bulbs controlled
by switches. Each switch flips the state of some subset of the bulbs,
that is, turns on all the ones in the subset that were off, and turns off
those that were on. You are told that for any nonempty set of bulbs,
there is a switch that flips an odd number of bulbs from that subset (and
perhaps other bulbs as well).
Show that, no matter what the initial state of the bulbs, you can
use these switches to turn off all the bulbs.

Solution: The proof is by induction on number n of bulbs. Note

that the order of switch-flipping does not matter; all that counts is
whether each switch is flipped an odd or even number of times. The
statement is trivially true for n = 1, so suppose n ≥ 2 and we are
given a configuration of n bulbs and a set of switches for them. By the
IHOP, for every bulb i there is a set Si of switches that can turn off

207
 Mathematical Puzzles

all the bulbs except possibly bulb i. If for any i the set Si does turn off
bulb i we are done, so let’s assume no i has this property.
Choose a pair of distinct bulbs, i and j, and perform Si followed by
Sj . The result is that except for i and j, every bulb that was initially
off stays off, while every bulb that was initially on was flipped off
and then on again—thus, remains in the state it was in before. The
exceptions are bulbs i and j, which are both flipped.
It follows that if bulbs i and j are both on, we can use Si ∪ Sj to
turn them both off without changing the state of any other bulb. So if
an even number of bulbs are on, we can turn them off pairwise and
thus turn all the bulbs off.
What if an odd number of bulbs are on? Then we use the given
condition, the relevant set of bulbs being all of them, to change the
number of lit bulbs to an even number. ♡
There’s something a bit mysterious about this proof—didn’t we
use the given condition only in one special case, where the set of bulbs
that we’re trying to flip an odd number from is the whole set of bulbs?
Maybe that’s all we needed!? No, that would fail even for two bulbs,
when there’s just one switch and it flips one bulb; then there’d be no
way to turn the other bulb off.
The issue here is that to execute the induction, we needed a con-
dition that “persists downward”—in this case, one that continues to
hold when we remove a bulb. Otherwise, we could not apply the
IHOP after removing bulb i.
In fact we really do need the condition for all subsets. If there’s a
set S of bulbs we can’t flip an odd number from, and we begin with
just one bulb from S on, we’re stuck.

Truly Even Split

Can you partition the integers from 1 to 16 into two sets of equal sizes
so that each set has the same sum, the same sum of squares, and the
same sum of cubes?
Solution: There is indeed such a partition: One set is {1, 4, 6, 7, 10,
11, 13, 16}, the other {2, 3, 5, 8, 9, 12, 14, 15}.
To see this, you might think to yourself: Hmm, 16 is a power of 2;
is it possible that this is an example of a more general statement? Can I
partition 1 through 8 into two equal-sized sets with the same sum and
sum of squares, for example? How about partitioning 1 through 4 into
two equal-sized sets with the same sum? The latter is certainly easy:

208
 Brilliant Induction

{1, 4} versus {2, 3}. Then of course {5, 8} versus {6, 7} also works for
the numbers from 5 to 8, and if you put these together cross-wise, you
get {1, 4, 6, 7} versus {2, 3, 5, 8} which works perforce for sums and
now also seems to work for sums of squares.
In general, you can prove by induction that the integers from 1 to
2k can be partitioned into sets X and Y so that each part has the same
sum of jth powers, where j runs from 0 to k − 1; equivalently, such
that for any polynomial
∑ P of degree less than k, the number P (X),
which we define as {P (x) : x ∈ X}, is equal to P (Y ).
To move up to 2k+1 , take X ′ = X ∪(Y +2k ) (where you get Y +2k
by adding 2k to each element of Y ) and Y ′ = Y ∪ (X+ 2k ). We need
to prove that for any polynomial P of degree at most k,

P (X) + P (Y + 2k ) = P (Y ) + P (X + 2k ).

If P has degree less than k, then by induction P (X) = P (Y ) and we

also get P (X + 2k ) = P (Y + 2k ) since each term is merely another
polynomial (Q, say) in xi or yi . Thus X ′ and Y ′ certainly agree for
polynomials of degree less than k, but what if P has degree k?
But we’re OK here too, because the kth power terms on both sides
of the displayed equation are the same as the kth power terms of
P (X) + P (Y ). ♡

Non-Repeating String
Is there a finite string of letters from the Latin alphabet with the prop-
erty that there is no pair of adjacent identical substrings, but the ad-
dition of any letter to either end would create one?

Solution: Yes, by induction on the number n of letters in the al-

phabet (then taking n = 26). Let L1 , . . . , Ln be the letters; for
n = 1 the one-letter string S1 = ⟨L1 ⟩ does the trick. We can take
S2 = ⟨L1 L2 L1 ⟩, S3 = ⟨L1 L2 L1 L3 L1 L2 L1 ⟩, and in general Sn =
Sn−1 Ln Sn−1 .
Proof that this works: Assume Sn−1 works for the alphabet
{L1 , . . . , Ln−1 }; we want to show Sn = Sn−1 Ln Sn−1 works for
the alphabet {L1 , . . . , Ln }. First we note that Sn can’t have a re-
peated substring, because such a substring could not contain the once-
appearing letter Ln , and would thus have to occur entirely inside one
of the Sn−1 ’s, contradicting our IHOP.
Now suppose the letter Lk , 1 ≤ k ≤ n, is added to the right-hand
end of Sn ; we want to show that this does create a repeated substring.

209
 Mathematical Puzzles

Certainly it does if k = n, as our whole string is now Sn−1 Ln followed

by itself. Otherwise, our IHOP tells us that there is already a repeated
substring inside Sn−1 Lk . The argument when Lk is added to the left-
hand end of Sn is essentially the same. ♡
For the Latin alphabet, the length of our string is 226 − 1 =
67,108,863 letters.

Baby Frog
To give a baby frog jumping practice, her four grandparents station
themselves at the corners of a large square field. When a grandpar-
ent croaks, the baby leaps halfway to it. In the field is a small round
clearing. Can the grandparents get the baby to that clearing, no matter
where in the field she starts?

Solution: Yes. Cover the field with a 2n × 2n square grid; we will

show by induction on n that from any starting point, the baby frog can
be guided to any grid cell we want. Suppose the target cell is in the SW
quadrant. Double that smaller grid in both directions and position it
so that it covers the whole field as a 2n−1 × 2n−1 grid. Our IHOP
tells us that the baby frog can be croaked into the doubled target cell
(which is the union of four cells of the original grid).
Now one croak from the SW corner grandparent will get the baby
to the original target cell.
That concludes our induction proof, and it remains only to take n
large enough so that some cell of the 2n × 2n grid lies entirely inside
the clearing that the baby frog is meant to visit.

Guarding the Gallery

A certain museum room is shaped like a highly irregular, non-convex,
11-sided polygon. How many guard-posts are needed in the room, in
the worst case, to ensure that every part of the room can be seen from
at least one guard-post?

210
 Brilliant Induction

Solution: Three guard-posts are sufficient (and sometimes neces-

sary, as for the 11-gon illustrated below).

In general, in any n-gon, ⌊n/3⌋ (that is, the greatest integer in n/3)
guard-posts suffice, and this number is the best bound possible. The
next figure shows, by extrapolation, that ⌊n/3⌋ posts may be neces-
sary; the following proof shows that they suffice, and in fact can be
placed in corners.

The first move is to triangulate the polygon. This can done by draw-
ing non-crossing diagonals until no more will fit; a diagonal is a line
segment between two vertices of the polygon whose interior lies en-
tirely in the interior of the polygon.
Next we show by induction on n that the polygon’s vertices can be
colored with three colors so that every triangle’s vertices get all three
colors. Choose any diagonal D in the triangulation and cut through it
“lengthwise” to form two polygons, each with fewer than n vertices,
and each having D as an edge. Color both using the IHOP and per-
mute the colors of one so that the colors at the endpoints of D match,
then put the two small polygons back together to get a coloring of the
original polygon (as is done in the figure below).

211
 Mathematical Puzzles

The least-frequently used color is found on at most ⌊n/3⌋ vertices,

and stationing guards there covers all the triangles and thus the whole
room. ♡

Path Through the Cells

Cells in a certain cellular telephone network are assigned frequencies
in such a way that no two adjacent cells use the same frequency. Show
that if the number of frequencies used is minimal (subject to this con-
dition), then it’s possible to design a path that moves from cell to ad-
jacent cell and hits each frequency exactly once—in ascending order
of frequencies!

Solution: This is really a problem in abstract graph theory; an as-

signment of frequencies, or “colors,” to vertices in a graph in such a
way that no two adjacent nodes get the same color is called a proper
coloring. The statement to be proved is that in any graph properly col-
ored with the minimum number of colors, there is a path that hits ev-
ery color exactly once; and moreover, if the colors are ordered (e.g.,
they correspond to the numbers from 1 to k) we can choose the path
so as to hit the colors in ascending order. The graph G in our case is
the one whose vertices are cells, two vertices constituting an edge if
the corresponding cells are adjacent (thus susceptible, presumably, to
frequency interference).
The idea is to re-color G in a “greedy” fashion, trying to use high-
numbered colors. First, re-color any vertex colored k−1 by color k,

212
 Brilliant Induction

if it isn’t adjacent to a vertex of color k. Then recolor any vertex of

color k−2 by color k−1, if it isn’t adjacent to a vertex of color k−1.
Continue in this way to get a new coloring which, by our minimality
assumption, still uses all k colors.
Now start the path at any vertex v1 whose new color (and there-
fore whose old color as well) is 1; we know there’s an adjacent vertex
v2 of new color 2, since otherwise the color of v1 would have been
incremented. Note v2 ’s old color is also 2; it cannot have been recol-
ored from color 1 since it is adjacent to v1 . Similarly, there’s a vertex
v3 adjacent to v2 of color 3, which could not have been recolored
from color 2 on account of v2 . Proceeding in this manner we get a
path v1 , v2 , . . . that hits every color in both the new coloring and the
old in increasing order.
This puzzle is one of many examples where asking for a proof
of a weaker statement (e.g., just that there is a path that hits every
frequency exactly once) can make the puzzle harder.

Profit and Loss

At a recent stockholders’ meeting of Widget Industrials Inc., the Chief
Financial Officer presented a chart of the month-by-month profits (or
losses) since the last meeting. “Note,” said she, “that we made a profit
over every consecutive eight-month period.”
“Maybe so,” complained a shareholder, “but I see we lost money
over every consecutive five-month period!”
What’s the maximum number of months that could have passed
since the last meeting?

Solution: What’s needed, of course, is a maximum-length se-

quence of numbers such that every substring of length 8 adds up to
more than 0, but every substring of length 5 adds up to less than 0.
The string must certainly be finite, in fact less than 40 in length, else
you could express the sum of the first 40 entries both as the (positive)
sum of 5 substrings of length 8 and the (negative) sum of 8 substrings
of length 5.
Let’s tackle the problem more generally and let f (x, y) be the
length of the longest string such that every x-substring has positive
sum and every y-substring negative; we may suppose x > y. If x is a
multiple of y, then f (x, y) = x−1 and we must accept vacuous truth
with respect to the x-substrings.

213
 Mathematical Puzzles

What if y = 2 and x is odd? Then you can have a string of length

x itself, with entries that alternate between, say, x−1 and −x. But you
can’t have x+1 numbers, because in each x-substring the odd entries
must be positive (since you can cover it with 2-substrings leaving out
any odd entry). But there are two x-substrings and together they imply
that the middle two numbers are both positive, a contradiction.
Applying this reasoning more generally suggests that f (x, y) ≤
x + y − 2 when x and y are relatively prime, that is, they have no com-
mon divisor other than 1. We can prove this by induction as follows.
Suppose to the contrary that we have a string of length x+y −1 which
satisfies the given conditions. Write x = ay + b where 0 < b < y, and
look at the last y + b − 1 numbers of the sequence. Observe that any
consecutive b of them can be expressed as an x-substring of the full
string, with a y-substrings removed; therefore, it has positive sum. On
the other hand, any (y − b)-substring of the last y + b − 1 can be ex-
pressed as a+1 y-strings with an x-string removed, hence has negative
sum. It follows that f (b, y − b) ≥ y + b − 1, but this contradicts our
induction assumption because b and y − b are relatively prime.
To show that f (x, y) is actually equal to x+y −2 when x and y are
relatively prime, we construct a string which has the required proper-
ties and more: It takes only two distinct values, and it is periodic with
periods both x and y. Call the two values u and v, and imagine at first
that we assign them arbitrarily as the first y entries of our string.
Then these assignments are repeated until the end of the string,
making the string perforce periodic in y. To be periodic in x as well,
we only need to ensure that the last y − 2 entries match up with the
first y − 2, which entails satisfying y − 2 equalities among the original
y choices we made. Since there are not enough equalities to force all
the choices to be the same, we can ensure that there is at least one u
and one v.
Let us do this, for example, with x = 8 and y = 5. Call the
first five string entries c1 , . . . , c5 , so the string itself will be c1 c2 c3 c4 c5
c1 c2 c3 c4 c5 c1 . To be periodic with period 8, we must have c4 = c1 ,
c5 = c2 , and c1 = c3 . This allows us to have c1 = c3 = c4 = u, for
example, and c2 = c5 = v; the whole sequence is thus uvuuvuvuuvu.
Getting back to general x and y, we note that a string which is
periodic in x automatically has the property that every x-substring
has the same sum; because, as you shift the substring one step at a
time, the entry picked up at one end is the same as the entry dropped
at the other end. Of course the same applies to y-substrings if the string
is periodic in y.

214
 Brilliant Induction

Let Sx be the x-substring sum and Sy similarly; we claim Sx /x = ̸

Sy /y. The reason is that if there are, say, p copies of u in each x-
substring and q copies of v in each y-substring, then Sx /x = Sy /y
would imply y(pu + (x − p)v) = x(qu + (y − q)v) which reduces to
yp = xq. Since x and y are relatively prime, this cannot happen for
0 < p < x and 0 < y < q.
It follows that we can adjust u and v so that Sx is positive and Sy
is negative. In the case above, for example, each 8-substring contains
5 copies of u and 3 of v, while each 5-substring contains 3 copies of
u and 2 of v. If we take u = 5 and v = −8, we get Sx = 1 and
Sy = −1. The final sequence, solving the original problem, is then
5, −8, 5, 5, −8, 5, −8, 5, 5, −8, 5. ♡

Uniformity at the Bakery

A baker’s dozen (thirteen) bagels have the property that any 12 of
them can be split into 2 piles of 6 each, which balance perfectly on
the scale. Suppose each bagel weighs an integer number of grams.
Must all the bagels have the same weight?

Solution: Yes. Suppose otherwise, and let b1 , . . . , b13 be a set of

weights, not all the same, with the stated property. We can assume
(by shifting all the weights down) that some bagel weighs 0 grams.
Let 2k be the greatest power of 2 that divides all the weights evenly,
and suppose our counterexample was chosen to minimize k. There
can’t be a bagel of odd weight, because to be balanced the remain-
ing bagels must have an even total weight; but leaving out the 0-gram
bagel must also leave even weight behind, an impossibility. But if all
the weights are even we can divide them all by 2, reducing k and con-
tradicting our IHOP. ♡
It is worth noting that the requirement that the weighings have
six bagels on a side is necessary. Otherwise, for example, a baker’s

215
 Mathematical Puzzles

dozen consisting of 7 bagels of 50 grams each and 6 bagels of 70 grams

each would have the stated property. The place where the proof breaks
down is where we shift all the weights down; that relies on there being
the same number of items on each side during the weighings.

Summing Fractions
Gail asks Henry to think of a number n between 10 and 100, but not
to tell her what it is. She now tells Henry to find all (unordered) pairs
of numbers j, k that are relatively prime and less than n, but add up
to more than n. He now adds all the fractions 1/jk.
Whew! Finally, Gail tells Henry what his sum is. How does she
do it?

Solution: For n = 2, the only eligible pair is {1, 2} so the sum is

1/2. For n = 3, we have pairs {1, 3} and {2, 3} for a total of 1/3 +
1/6 = 1/2. For n = 4, the pairs are {1, 4}, {2, 3}, and {3, 4} giving
1/4 + 1/6 + 1/12 = 1/2. Hmm. Is it possible that the sum is always
1/2?
Let’s try to prove it by induction, starting from n = 2. Moving
from n to n+1, we gain 1/jn for each j with gcd(j, n) = 1, and lose
1/jk for each j and k with gcd(j, k) = 1 and j + k = n. But if j + k =
n, then any pair of j, k, and n being relatively prime implies that all
three pairs are. Thus, each pair j, k that sum to n with gcd(j, k) = 1
signifies a loss of 1/jk but a gain of 1/jn + 1/kn = 1/jk, neatly
canceling. ♡

Tiling with L's

Can you tile the positive quadrant of the plane grid with unrotated tri-
ominoes each of which is either shaped like a letter L, or a backwards
L?
Solution: The answer is perhaps a bit surprising: You can indeed
do it, but in only one way, and that way is not periodic.
If you try to do it by hand, starting at the origin with a full row
along the X-axis and proceeding upward row by row, you will find
that a little forethought (thinking one row ahead) is all you need.
That first row must start (at the left) with a forward L, which
we’ll call an “el.” To its right we’ll need to put a backward L, which
we’ll call a “le,” so that the space above these two triominoes will be

216
 Brilliant Induction

fillable. Then you’ll need another el, another le, another el, and so
forth, alternating out to infinity.
The next row is already half-covered; to cover the rest of those
cells, you’ll need to start with a le. Each successive choice of el or le
must leave an even-length gap on row 3, otherwise you’ll be stumped
later. The good news is that the choice of el or le always changes the
length of the previous gap by one, thus you can always make that gap
even.

This sets us up nicely for a proof by induction that the construction

is possible. What should the induction hypothesis be? We’d like to say
that all rows are covered up to and including row k (say), and the next
row, row k+1, is covered up to position m (which could be zero). After
position m, row k+1 has even-length gaps out to infinity.
What else? Well, we also need to say that row k+2 has even-length
gaps out to position m, and is empty after that; and no higher row is
covered at all.
That’s a lot of hypotheses, but we need them all and once we have
them, the argument is easy. We insert a new triomino in such a way
that its bottom covers cells m+1 and m+2 in row k+1, choosing el or
le so as to make the new gap, above in row k+2, even. That will also
have the effect of shortening the next gap on row k+1 by two, which
kills the gap or leaves it of even length, either of which is OK.
Everything seems to be fine, but hold on a minute—what, exactly,
are we inducing on? It can’t be the number of triominoes placed, be-
cause that’s already infinite after the first row goes down. What we’re
really doing is a series of inductions, one for each row. Each induction
ends with a tiling in which rows 1 through k are completely covered,
row k+1 partly covered, no higher row covered at all, and all gaps on

217
 Mathematical Puzzles

row k+1 of even length. This is just an “m = 0” case of our induction

hypothesis.
Uniqueness of our tiling follows by assuming we have any valid
tiling, choosing any k and m, and looking at that part of the tiling
corresponding to the k, m case of our induction hypothesis. It must
satisfy the hypothesis in order to be extendible, and since our con-
struction leaves no choice at any point, the tiling must be identical to
ours.

Traveling Salesmen
Suppose that between every pair of major cities in Russia, there’s a
fixed one-way air fare for going from either city to the other. Trav-
eling salesman Alexei Frugal begins in St. Petersburg and tours the
cities, always choosing the cheapest flight to a city not yet visited (he
does not need to return to St. Petersburg). Salesman Boris Lavish also
needs to visit every city, but he starts in Kaliningrad, and his policy
is to choose the most expensive flight to an unvisited city at each step.
It looks obvious that Lavish’s tour costs at least as much as Fru-
gal’s, but can you prove it?

Solution: One way is to show that for any k, the kth cheapest flight
(call it f ) taken by Lavish is at least as costly as the kth cheapest flight
taken by Frugal. This seems like a stronger statement than what was
requested, but it really isn’t; if there were a counterexample, we could
adjust the flight costs, without changing their order, in such a way that
Lavish paid less than Frugal.
For convenience, imagine that Lavish ends up visiting the cities in
west-to-east order. Let F be the set of Lavish’s k cheapest flights, X
the departure cities for these flights, and Y the arrival cities. Note that
X and Y may overlap.
Call a flight “cheap” if its cost is no more than f ’s; we want to
show that Frugal takes at least k cheap flights. Note that every flight
eastward out of a city in X is cheap, since otherwise it would have
been taken by Lavish instead of the cheap flight in F that he actually
took.
Call a city “good” if Frugal leaves it on a cheap flight, “bad”
otherwise. If all the cities in X are good, we are done; Frugal’s de-
partures from those cities constitute k cheap flights. Otherwise, let
x be the westernmost bad city in X; then when Frugal gets to x, he
has already visited every city to the east of x, else Frugal could have

218
 Brilliant Induction

departed x cheaply. But then every city to the east of x, when visited
by Frugal, had its cheap flight to x available to leave on, so all are
good. In particular, all cities in Y east of x are good, as well as all
cities in X west of x; that is k good cities in all. ♡

Lame Rook
A lame rook moves like an ordinary rook in chess—straight up, down,
left, or right—but only one square at a time. Suppose that the lame
rook begins at some square and tours the 8 × 8 chessboard, visiting
each square once and returning to the starting square on the 64th
move. Show that the number of horizontal moves of the tour, and
the number of vertical moves of the tour, are not equal!

Solution: You should be asking yourself: Is this true for n × n

boards as well? Some experimentation will remind you that there is
no such tour if n is odd, because (for one thing) such a board has more
squares of one color than another but a tour that returns to its starting
position must visit the same number of squares of each color.
Moreover, for n = 2 and n = 6, it is easy to construct tours where
the numbers of horizontal and vertical steps are equal. So you might
guess that the puzzle works on n × n boards when n ≡ 0 mod 4, and
moreover that the issue could be that the number of (say) horizontal
moves must be equivalent to 2 mod 4 and therefore can’t be 64/2. All
this is correct, but how to prove it?
Assume the rook is a point in the middle of a square, and let P be
the polygon thus traced. Note that the vertices of P lie on the dual grid
of the chessboard, whose vertices lie at the centers of the chessboard
squares.
The polygon P outlines a tree T (pink in the figure below) on the
grid lines of the chessboard.
We will show by induction that the following holds when P is the
outline of any tree on the chessboard grid: If there are n0 dual-grid
vertices on P ’s perimeter that lie on even columns, and n1 on odd
columns, and h is the sum of the lengths of the horizontal borders of
P , then n ≡ n1 − n0 + 2 mod 4.
The base case is easy: If the tree is a single vertex, we have n1 =
n0 = 2 and h = 2. Now let u be any leaf of T ; if it’s attached ver-
tically, cutting it off decrements both n1 and n0 by 1 and does not
change h. If u is attached horizontally, it changes n1 − n0 by two
while decrementing h by two.

219
 Mathematical Puzzles

Since the polygon P generated by the lame rook has n1 = n0 , the

number of horizontal moves is 2 mod 4 and thus cannot be 32. ♡

For our theorem we return to Ramsey theory, using induction to

get an upper bound for Ramsey numbers (and incidentally to show
that they exist).
Given any two positive integers s and t, Ramsey’s Theorem claims
that for sufficiently large n, for every red–green coloring of the un-
ordered pairs of numbers in {1, 2, . . . , n} there is either a set S of size
s all of whose pairs are colored red, or a set T of size t all of whose
pairs are colored green. The least n for which this holds is, by defini-
tion, the Ramsey number R(s, t).
( )
Theorem. For any s and t, R(s, t) exists and is no more than s+t−2 s−1 .

We note first that since there are no pairs inside a set of size 1,
R(1, k) = R(k, 1) = 1 for any k. Further, to get (say) a red set (i.e., a
set all of whose pairs are red) of size 2 we only need one red edge. It
follows that R(2, k) = R(k, 2) = k, because (in the case of R(2, k)) if
there’s any red pair its members give us a red set of size 2, else all the
pairs are green and we can use all the numbers to get our green set of
size k.
These observations conform to the theorem. Now we will proceed
by induction on the sum s+t.
( ′ ′ −2)
Fix s and t and suppose R(s′ , t′ ) ≤ s s+t for all s′ , t′ with
′ ′
′ −1
( )
s +t < s+t. Let U be the set {1, . . . , n} where n = s+t−2 s−1 , and

220
 Brilliant Induction

suppose some adversary has colored all the unordered pairs in U red
or green. We want to find in U either a red subset S of size s or a green
subset T of size t. That would ( show ) that the conclusion of Ramsey’s
theorem holds when n = s+t−2 s−1 , therefore that R(s, t) is at most
this number. ( )
Recall that (from Pascal’s triangle) n = j+k, where j = s+t−3
(s+t−3) s−2
and k = s−1 . Note that j is our “n” when s′ = s−1 and t′ = t,
while k is our “n” when s′ = s and t′ = t−1.
Let’s focus our attention on the pairs containing the number 1.
There are n−1 of those, and we claim that either at least j of these
pairs are red, or at least k of them are green. Why? Because otherwise
there are at most j−1 red pairs and at most k−1 green pairs, and
that adds up to only j+k−2 = n−2. (Yes, you can think of it as an
application of the pigeonhole principle.)
Suppose the former, that is, there are j red pairs of the form {1, i},
and let V be the set of numbers i that appear ( in them
) (not including
1 itself). Since (s−1) + t < s+t, and j = (s−1)+t−2 (s−1)−1
, we know from
′
our IHOP that there’s either a red set S ⊂ V of size s−1 or a green
set T ′ ⊂ V of size t. If it’s the latter, we are happy; and if the former,
we can add the number 1 to our set—since all of the pairs involving 1
are red—and get a red set S = S ′ ∪ {1} of size s.
The other case is similar: If there are k green pairs of the form
{1, i}, let W be the set of numbers i that appear in them; since s +
( )
(t−1) < s + t, and k = s+(t−1)−2 (s−1
, we know from our IHOP that
there’s either a red set S ′ ⊂ V of size s or a green set T ′ ⊂ V of size
t−1. If it’s the former, we have our S; if the latter, we can add the
number 1 to T —since all of the pairs involving 1 are green—and get
a green set T = T ′ ∪ {1} of size t. Done! ♡
( ) ()
As it happens, R(3, 3) is equal to 3+3−2 = 24 = 6, but
3−1 (4+4−2) (6)
R(4, 4) = 18 which is strictly less than our bound of 4−1 = 3 =
20. Putting our theorem together with the one proved in the previous
chapter, we know that R(k, k) lies somewhere between 2k/2 and 22k .
Most combinatorialists would guess that there’s some number α be-
tween 12 and 2 such that R(k, k) behaves like 2αk as k grows—more
precisely, that the logarithm base 2 of the kth root of R(k, k) tends to
some α as k goes to infinity. Finding α would have earned you $5000
from “Uncle Paul” Erdős, and he’d have paid happily.

221
 16. Journey into Space
We all live in three spacial dimensions, and are attuned to thinking
three-dimensionally even though we see, draw and paint, for the most
part, only in two. Thus, geometrical problems in three dimensions,
while potentially much harder than problems in the plane, can appeal
to our intuition in pleasant and even useful ways.
Some of the puzzles that follow are set naturally in three dimen-
sions; others can be placed there to good effect.

Easy Cake Division

Can you cut a cubical cake, iced on top and on the sides, into three
pieces each of which contains the same amount of cake and the same
amount of icing?

Solution: You can get any number k of pieces with the same
amount of cake and the same amount of icing in the following neat
way: Looking at the cake from above, divide the perimeter of the
square top of the cake into k equal parts and make straight, verti-
cal cuts to the middle of the square from each boundary point on the
perimeter.

That works because each piece, looked at from the top, is the
union of triangles with the same altitude (namely, half the side of the
cake) and the same total base (namely, 1/k times the perimeter), as
in the figure below.
Speaking of cubes, here’s a question about painting them.

223
 Mathematical Puzzles

Painting the Cubes

Can you paint 1000 unit cubes with 10 different colors in such a way
that for any of the 10 colors, the cubelets can be arranged into a 10 ×
10 × 10 cube with only that color showing?
Solution: Note that the numbers work out perfectly: 6000 faces
need to be colored, and the number of faces showing on the big cube
is 6 × 100 = 600. Thus we’ll need to paint exactly 600 faces with each
color, and when we put the big cube together, we can’t afford to bury
even a single face of the designated color inside.
It turns out that despite these restrictions, there are many ways to
color the unit cubes that will do the trick. But finding one by hammer
and tongs is a pain in the neck.
Instead, number the colors 0 through 9. Apply paint to all the
cubes in the infinite 3-dimensional cubic grid, as follows: Paint both
sides of all the faces on the plane x = 0 with color 0, all faces on x = 1
with color 1, etc., rotating back to color 0 for the plane x = 10, and
so forth. Do the same for the y-planes and z-planes.
This scheme colors every face of every unit cube with a color cor-
responding to the rightmost digit of that face’s u-coordinate, where
u is x, y or z according to which axis is perpendicular to the face.
In particular, the scheme colors unit cubes in 103 different ways. But
now, if you want your big cube to display color i on its outside, cut
out your big cube from space via planes x = i, x = i+10, y = i,
y = i+10, z = i, and z = i+10. This will use each type of unit cube
once, and you are done! ♡
Of course, you can replace 10 by any positive integer n (that is, you
can color n3 unit cubes with which you can then make an n × n × n
cube with any given color the only one showing). The above argument
still works, and in fact versions of it work in other dimensions, as well.
Just don’t think too hard about what it means to “paint” a 3-
dimensional face of a 4-dimensional hypercube.

Curves on Potatoes
Given two potatoes, can you draw a closed curve on the surface
of each so that the two curves are identical as curves in three-
dimensional space?

Solution: Yes, you can: intersect the potatoes!

Well, maybe a few more words of explanation would not go amiss.

224
 Journey into Space

Think of the potatoes as holograms and bring them together so that

they overlap in space. Then the intersection of their surfaces will de-
scribe a curve (possibly more than one curve) that lives on both sur-
faces, and solves the puzzle.

Painting the Polyhedron

Let P be a polyhedron with red and green faces such that every red
face is surrounded by green ones, but the total red area exceeds the
total green area. Prove that you can’t inscribe a sphere in P .

Solution: Assume the sphere is inscribed, and triangulate the faces

of P using the points where the sphere is tangent. Then the triangles
on either side of any edge of P are congruent, thus have the same
area; at most, one of every such pair of triangles is red. It follows
that the red area is at most equal to the green area, contradicting our
assumptions. ♡

The illustration shows a two-dimensional version, where sides

and vertices of a polygon take the place of faces and edges of P .

Boarding the Manhole

An open manhole 4 meters in diameter has to be covered by boards
of total width w. The boards are each more than 4 meters in length,
so if w ≥ 4 it’s obvious that you can cover the manhole by laying the
boards side by side (see figure below). If w is only 3.9, say, you still
have plenty of wood and you are allowed to overlap the boards if you
want. Can you still cover the manhole?

Solution: Intuitively speaking, even though you have plenty of

wood, it seems that you have to waste a lot of it either by covering

225
 Mathematical Puzzles

the center of the manhole many times, or laying boards too close to
the edge to be of much use. Putting it another way, it’s relatively ex-
pensive to cover parts of the manhole far from the center. Can you
quantify that statement in a useful way?
It turns out you can, using a famous fact known sometimes as
Archimedes’ hat box theorem. Archimedes used his “method of ex-
haustion” (these days, we would use calculus) to show that if a sphere
of radius r is intersected by two parallel planes a distance d apart, as
in the figure below, the surface area of the sphere between the planes
is 2πrd. In particular, it doesn’t depend on where the planes cut the
sphere, but only on how far apart the planes are. (The “hat box” pre-
sumably alludes to a cylinder of radius r perpendicular to the planes;
the cylinder’s area between the planes is that same 2πrd.)

Where do we get our planes from? Let’s replace the manhole by

a sphere, also of diameter w, whose equator is the circumference of
the manhole. Each board is replaced by a vertical slab consisting of

226
 Journey into Space

two parallel vertical planes whose distance d apart is the width of the
board. Now the key observation: Arranging the boards to cover the
manhole is equivalent to arranging the slabs to cover the sphere.
But by Archimedes’ theorem, the slabs cannot cover more than
2π w2 d = πwd of the surface of the sphere, which has area πw2 . So if
d < w, you’re stuck. ♡
Now we are set up perfectly to conquer the next puzzle.

Slabs in 3-Space
A “slab” is the region between two parallel planes in three-
dimensional space. Prove that you cannot cover all of 3-space with
a set of slabs the sum of whose thicknesses is finite.
Solution: In a way the conclusion seems obvious; if you can’t cover
space when the slabs are parallel and disjoint, surely you can’t cover
when they waste space by overlapping. But infinity is a tricky concept.
The slabs all have infinite volume, so what’s to prevent them from
covering anything they want?
Actually, we know from the previous puzzle that they have trou-
ble covering a large ball. If the total thickness of the slabs is T , then,
by Archimedes’ theorem, they can’t cover the surface of a ball of di-
ameter greater than T , thus they can’t cover the ball itself.
Thus they certainly can’t cover all of 3-space. But it’s odd that to
show the latter, we seem to need to reduce the problem to just a finite
piece of space.

Bugs on Four Lines

You are given four lines in a plane in general position (no two par-
allel, no three intersecting in a common point). On each line a ghost
bug crawls at some constant velocity (possibly different for each bug).
Being ghosts, if two bugs happen to cross paths they just continue
crawling through each other uninterrupted.
Suppose that five of the possible six meetings actually happen.
Prove that the sixth does as well.

Solution: As the title of this chapter suggests, lifting the puzzle into
space provides an elegant solution. How? By means of a time axis.
Suppose every pair of bugs meets except bug 3 with bug 4. Construct
a time axis perpendicular to the plane of the bugs, and let gi be the

227
 Mathematical Puzzles

graph (this time, in the sense of graphing a function) of the ith bug in
space. Since each bug travels at constant speed, each such graph is a
straight line; its projection onto the plane of the bugs is the line that
bug travels on. If (and only if) two bugs meet, their graphs intersect.
The lines g1 , g2 , and g3 are coplanar since all three pairs inter-
sect, and the same applies to g1 , g2 , and g4 . Hence all four graphs are
coplanar. Now g3 and g4 are certainly not parallel, since their projec-
tions onto the original plane intersect, thus they intersect on the new
common plane. So bugs 3 and 4 meet as well.

Circular Shadows II
Show that if the projections of a solid body onto two planes are perfect
disks, then the two projections have the same radius.
Solution: The conclusion seems eminently reasonable, yet it’s not
completely obvious how to prove it.
An easy way to make your intuition rigorous is to select a plane
which is simultaneously perpendicular to the two projection planes,
and move parallel copies of it toward the body from each side. They
hit the body at the opposite edges of each projection, so that the dis-
tance between the parallel planes at that moment is the common di-
ameter of the two projected circles. ♡

Box in a Box
Suppose the cost of shipping a rectangular box is given by the sum of
its length, width, and height. Might it be possible to save money by
fitting your box into a cheaper box?

228
 Journey into Space

Solution: The issue here is that if packed diagonally, the inside box
could have some dimensions that exceed the greatest dimension of
the outside
√ box. For example, a narrow needle-like box of length al-
most 3 could be packed inside a unit cube. This particular example
wouldn’t provide a way to cheat, because in this case the total cost of
the outside cube would be 3, much more than the 1.7 or so you might
have to pay for the inside box. But how do we know that some better
example won’t turn up?
Here’s a delightful argument that lets ε go to infinity(!) in order to
show that you can’t cheat the system.
Suppose your box is a × b × c, and let R stand for the region of
space occupied by your box including its interior. Suppose R can be
packed into an a′ × b′ × c′ box R′ with a + b + c > a′ + b′ + c′ . Let
ε > 0 and consider the region Rε consisting of all points in space that
are within distance ε of R. This Rε region will be a rounded convex
shape containing your box R. The volume of Rε is equal to your box’s
volume plus 2ε(ab + bc + ac)s for the volume added to the sides, plus
πε2 (a + b + c) for the volume added to the edges, plus 43 πε3 for the
volume (consisting of the eight octants of a ball) added to the corners.
If you do the same to the outside box R′ you get a region Rε′ which of
course must contain Rε . But that’s impossible! In the expression for
vol(Rε′ ) − vol(Rε ), the ε3 terms cancel, so if you take ε to be large, the
dominant term is the ε2 term, namely πε2 ((a′ + b′ + c′ ) − (a + b + c)),
which is negative. ♡

Angles in Space
Prove that among any set of more than 2n points in Rn , there are three
that determine an obtuse angle.
Solution: It makes sense that the 2n corners of a hypercube repre-
sent the most points you can have in n-space without an obtuse angle.
But how to prove it?
Let x1 , . . . , xk be distinct points (vectors) in Rn , and let P be their
convex closure. We may assume P has volume 1 by reducing the di-
mension of the space to the dimension of P , then scaling appropri-
ately; we may also assume x1 is the origin (i.e., the 0 vector). If there
are no obtuse angles among the points, then we claim that for each
i > 1, the interior of the translate P + xi is disjoint from the interior
of P ; this is because the plane through xi perpendicular to the vector
xi separates the two polytopes.

229
 Mathematical Puzzles

Furthermore, the interiors of P + xi and P + xj , for i = ̸ j, are

disjoint as well; this time via the separating plane running through
xi + xj perpendicular to the edge xj − xi of P . We conclude that the
volume of the union of the P + xi , for 1 ≤ i ≤ k, is k.
However, all these polytopes lie inside the doubled polytope 2P =
P + P , whose volume is 2n . Hence, k ≤ 2n as claimed!

Curve and Three Shadows

Is there a simple closed curve in 3-space, all three of whose projections
onto the coordinate planes are trees? This means that the shadows of
the curve, from the three coordinate directions, may not contain any
loops.

Solution: No one knew the answer to this question until such

a curve was actually found, by John Terrell Rickard of Advanced
Telecommunication Modules Ltd. in Cambridge, UK. His beautiful,
symmetrical solution is shown below with its three shadows. I have
no idea how Rickard found it.

Later, Donald Knuth, author of the multi-volume classic The Art

of Computer Programming, programmed a computer to see if there were
any other solutions that, like Rickard’s, could be carved from a 3×3×3
region of a cubical grid. In the time it took to push his ENTER but-
ton (according to Knuth), the computer found that there was just
one other solution, shown below. Knuth’s solution, in contrast to
Rickard’s, has no symmetry at all.
If you found either one of these curves, more power to you!
Our theorem is perhaps the world’s best-known example of prov-
ing a theorem about plane geometry by moving into 3-dimensional
space. But there’s an ironic twist.

230
 Journey into Space

Given any two circles in the plane of different radii, neither con-
tained in the other, we can construct two lines that are tangent to the
circles such that for each line, the circles are on the same side of the
line. (See picture below). These lines will intersect at some point on
the plane, known as the Monge point of the two circles.

Theorem. Suppose three circles are given, all of different radii, none con-
tained in another. Then the three Monge points determined by the three pairs
of circles all lie on a line.
This is what the picture looks like (we have taken the circles to be
disjoint, but in fact they’re allowed to overlap).
The theorem is attributed to Gaspard Monge, 1746–1818, a dis-
tinguished French mathematician and engineer. On Wikipedia (as I
write this) you will find a proof which is elegant, famous, and wrong!
Here’s how it goes:
Replace each circle by a sphere whose equator is that circle, so
now you have three spheres (of different diameters) intersecting the
original plane at their equators. Pick two of them: Instead of two

231
 Mathematical Puzzles

tangent lines you now have a whole tangent cone whose apex is the
Monge point of the circles.
Now, take a plane tangent to the three spheres, thus also tangent
to the three cones. The three Monge points will all lie on that plane,
as well as on the original plane. But the two planes intersect in a line,
so the Monge points are all on that line, QED!
It’s not easy to spot the hole in this proof. The difficulty is that
there may not be a plane tangent to the three spheres; for example, if
two of the circles are large and the third is between them and smaller.
The following proof, attributed by cut-the-knot.org to Nathan
Bowler of Trinity College, Cambridge, works by erecting cones instead
of spheres on top of the circles. Call them C1 , C2 , and C3 , and let them
all be “right” cones—that is, they support 90◦ angles at their apices.
(Actually, we only need all the cones to have the same angle.) Each
pair of cones determines two (outside) tangent planes, say P1 and Q1
(for cones C2 and C3 ), P2 and Q2 (for cones C1 and C3 ), and finally
P3 and Q3 (for cones C1 and C2 ).
Each pair of planes Pi , Qi intersect in a line Li which passes
through the apex of both tangent cones, as well as through the point
where the corresponding circle tangents meet. Thus, in particular, L1

232
 Journey into Space

and L2 both meet at the apex of C3 , L1 , and L3 at the apex of C2 ,

and L2 and L3 at the apex of C1 . Hence the three lines of intersection
are coplanar (all lie on the plane determined by the three apices); the
intersection of that plane with the original plane of the circles is a line
through the three foci, and this time we are really done. ♡
There are other ways to prove Monge’s theorem and as you might
suspect, one of them is to transform the plane in a circle-preserving
way so as to escape annoying initial configurations that screw up the
spheres proof. But it’s curious, and a bit sobering, that the unfixed
sphere argument lives on as the “standard” proof of Monge’s theorem.

233
 17. Nimbers and the
Hamming Code
Probably the most famous equation in mathematics is not E = mc2
or eiπ = −1 but 1 + 1 = 2. But if, like many computer scientists, you
are fond of binary arithmetic, you might prefer 1 + 1 = 10; if you are
an algebraist or logician who works with numbers modulo 2, you
might prefer 1 + 1 = 0.
It turns out to be surprisingly useful, in certain circumstances, to
work in a world where any number added to itself is zero. Such is the
world of nimbers.
Nimbers are a lot like numbers, and in fact you can think of them
as non-negative integers, written in binary. But their rules for addition
and subtraction are different from the usual ones. In particular, you
add nimbers without carrying.
Here’s another way to say the same thing: To compute the ith bit
(from the right) of the sum of a bunch of nimbers, you only need to
know the ith bit of each nimber. If an odd number of these are 1’s, the
ith bit of the sum will be a 1; otherwise it will be a 0.
Let’s try this. To avoid confusion, we’ll write the decimal expres-
sion of a nimber with an overline, so, for example, 7 = 111 and
10 = 1010. To distinguish nimber addition from ordinary addition
of binary numbers, we’ll use the symbol ⊕ instead of +.
Thus, 7 ⊕ 10 = 111 ⊕ 1010 = 1101 = 13. Of course, 1 ⊕ 1 =
1 ⊕ 1 = 0 = 0 and in fact, as we said, n ⊕ n = 0 for any n.
Nimber arithmetic, once you get used to it, is really rather nice. It
obeys the usual laws of arithmetic—it’s commutative, associative, and
satisfies x ⊕ 0 = x for any x. It has additive inverses, as we saw: Any
nimber is the additive inverse of itself! Thus, we have no use for minus
signs in nimber arithmetic, and subtraction is the same as addition. If
we have two nimbers m and n and we want to know which nimber to
add to m to get n, why, it’s the nimber m ⊕ n, since m ⊕ (m ⊕ n) =
(m ⊕ m) ⊕ n = 0 ⊕ n = n.

235
 Mathematical Puzzles

If you’ve studied linear algebra, you may recognize nimbers as

vectors over a two-element field. When you add vectors in any vector
space, the coordinates are added independently; that’s what is hap-
pening with our nimbers, since we don’t carry. One difference is that
for nimbers we don’t bother to specify the dimension (number of co-
ordinates) in advance. We can always add more zeroes to the left of a
binary number if we need more coordinates (bits), without changing
the number’s value or its nimber properties.
Whence the name? As we shall see, nimbers are numbers that
arise in the game of Nim, so the term “nimber” was irresistible to
the himself irresistible mathematician John H. Conway. Understand-
ably, “nimber” has largely replaced an old term, “Grundy number.”
Consider:

Life Is a Bowl of Cherries

In front of you and your friend Amit are 4 bowls of cherries, contain-
ing, respectively, 5, 6, 7, and 8 cherries. You and Amit will alternately
pick a bowl and take one or more cherries from it. If you go first, and
you want to be sure to get the last cherry, how many cherries should
you take—and from which bowl?

Solution: In the traditional game of Nim the bowls of cherries are

replaced by piles of stones or (my preference) stacks of chips, and the
problem is how best to play from any starting position.
We denote positions in Nim by listing the stack sizes from largest
to smallest; for example, in the cherries puzzle, the given position is
8|7|6|5. The rules of Nim specify that from a given position you can
reach any position with the same numbers except that one number has
been either lowered or (if all the chips from that stack were removed)
deleted. If, for example, you remove two cherries from the bowl of
seven, you reach the new position 8|6|5|5.
The goal is to be the player to achieve the empty position. This
can be done immediately if there is only one stack; just take all of it.

236
 Nimbers and the Hamming Code

If there are just two stacks, and they are of different sizes, you (as first
player) have the following winning strategy: Take chips from the big
stack to reduce it to exactly the size of the smaller one, and repeat.
If there are more than two stacks, things start to get compli-
cated. Nimbers make it simple—they enable you to win whenever
you should, and often (in practice) when you shouldn’t!
Here’s how it works. We define the nimsum of a position to be the
nimber-sum of the stack sizes. For example, the nimsum of the initial
position in the cherries puzzle is

8 ⊕ 7 ⊕ 6 ⊕ 5 = 1000 ⊕ 111 ⊕ 110 ⊕ 101 = 1100 = 12.

The fact is that you (as first player) can force a win from any posi-
tion whose nimsum is not zero. How do you do it? Easy: Change it to
a position whose nimsum is zero. Your opponent will have to change
the position to one whose nimsum is again non-zero, and then you
repeat. Eventually you get to the empty position, whose nimsum is of
course zero, and you win!
There are some things to verify here. First of all, how do you know
that your opponent can’t change a zero-nimsum position to another
zero-nimsum position? Well, your opponent (like you) must remove
one or more chips from just one stack, so he reduces one of the stack-
sizes by some positive number k. That will have the effect of nimber-
subtracting k from the previous position nimsum, but remember, nim-
ber subtraction is the same as nimber addition. So the nimsum of the
new position will be 0 ⊕ k = k ̸= 0.
A bit more subtle is showing that from any position whose nim-
sum is not zero, you can get to one whose nimsum is zero. Here’s
how you can do it. Suppose the current nimsum is s, and suppose the
leftmost 1 in the binary representation of s is in the ith bit from the
right (that is, the bit representing 2i−1 ). Then there must be at least
one stack whose size (sj , say) also has a 1 as its ith bit—otherwise,
the ith bit of s would have been zero. We claim we can remove some
chips from that stack to change its size to sj ⊕ s. Doing so will change
the position’s nimsum to s ⊕ s = 0. The only thing we need to check
is that sj ⊕ s is a smaller number than sj (since we’re not allowed to
add chips to a stack). But that is easy because nimber-adding s to sj
will change the ith bit from 1 to 0, and won’t affect any bits of sj that
are to the left of the ith. It follows that s ⊕ sj is smaller than sj , so
reducing the sj -stack to size s ⊕ sj is a legal move.
Let’s try that with the cherries. We saw that the nimsum of the
given position was 12 = 1100, whose leftmost 1 is in the fourth

237
 Mathematical Puzzles

position from the right. Here there happens to be only one stack
whose size (in binary) has a 1 in the fourth position from the right:
the stack of size 8. Nimber-adding 1100 to 1000 gives 100 = 4 so we
reduce the number of cherries in the bowl of 8 to 4.
The resulting position, 7|6|5|4, has nimsum 111 ⊕ 110 ⊕ 101 ⊕
100 = 0 and we have Amit where we want him.
In some positions there are several stacks whose sizes, in binary,
have a 1 in the same position as the leftmost 1 in the position’s nim-
sum; in such cases, there will be one winning move in each of those
stacks. For our initial cherries position, there is only one such stack,
thus only one winning move. After any other first move, it is Amit
who can force a win.
Even within the specialized field of combinatorial games (where
two players alternate and the first one who is unable to move loses),
nimbers play a much bigger role than just their part in Nim. It turns
out that any position in an “impartial” combinatorial game—that
is, one in which both players have the same move choices from any
given position—can be represented by a nimber! This is the content of
the celebrated Sprague-Grundy Theorem, about which the interested
reader can read in Winning Ways or any other book on combinatorial
games.
In fact, here’s another impartial game to which we can apply
nimbers—but perhaps not in the way you would think.

Life Isn't a Bowl of Cherries?

In front of you and your friend Amit are 4 bowls of cherries, contain-
ing, respectively, 5, 6, 7, and 8 cherries. You and Amit will alternately
pick a bowl and take one or more cherries from it. If you go first, and
you want to be sure Amit gets the last cherry, how many cherries
should you take—and from which bowl?

Solution: So, in effect, in this second puzzle you want your friend
to win. (In such a situation you are said to be playing the misère version
of the game.) It looks like since your objective is the opposite of the
previous puzzle’s, you should make any move but the one that reduces
the number of cherries in the bowl of eight to four.
That’s fine if Amit cooperates, that is, if he himself wants the last
cherry, but it will not do if you want to insist that Amit get the last
cherry. Surprisingly, your unique winning first move for the second
puzzle is the same as the one for the first!

238
 Nimbers and the Hamming Code

The philosophy behind this counterintuitive solution is that play-

ing for zero-nimber positions allows you to control the game, until
the crucial moment when you decide whom you want to get the last
cherry. That moment comes when you first encounter a position con-
taining exactly one bowl with multiple cherries, that is, in general Nim
terms, a position in which every remaining stack except one contains
only a single chip.
Note that such a position never has a zero nimber, because the
big stack’s size, in binary, has at least one 1 in it that can’t be can-
celled. Thus, when such a position arises, as it inevitably must, it will
be you facing it, not your opponent. Now you need consider only two
moves: Removing all of the big stack, and removing all but one chip
from it. Either way there will only be “singleton” stacks left for your
opponent. If there are an even number of them, you will get the last
chip; if there are an odd number of them, he will. Since one of the
above two moves will leave an even number of stacks and the other
an odd number, you control who gets the last chip.
Let’s do a game of cherries to see an example of how this works.
In the following pictures, we zigzag through time with your move
(hollow cherries are those to be removed) on the left and Amit’s on
the right. The nimsum for each position (prior to the indicated move)
and the nimbers for each stack are indicated in binary.
At the last position shown, you will take all the cherries in the 4-
cherry bowl if you want to end up with the last cherry; if you want
Amit to get it, you’ll take just three of them.

Whim-Nim
You and a friend, bored with Nim and Nim Misère, decide to play a
variation in which at any point, either player may declare “Nim” or
“Misère” instead of removing chips. This happens at most once in a
game, and then of course the game proceeds normally according to
that variation of Nim. (Taking the whole single remaining stack in an
undeclared game loses, as your opponent can then declare “Nim” as
his last move.)
What’s the correct strategy for this game, which its inventor, the
late John Horton Conway, called “Whim”?

Solution: Owing to our previous analyses, the game is effectively

over once someone declares. So we concentrate on how to play the
game pre-declaration.

239
 Mathematical Puzzles

It’s easy to see that if you’re in a 0-position you win by simply

declaring “Nim.” Thus, you must not create a 0-position for your
opponent. Moreover, you don’t want to be handing your opponent
nothing but singleton stacks, because then he will declare appropri-
ately (“Nim” if the number of singletons is even, else “Misère”) and
win. On the other hand, you’d love to give your opponent a single
stack of size 2. For that matter, you can give him any odd number of
stacks of size 2, together with any even number of stacks of size 1 or
3. Then eventually he’ll be the one to reduce to all singletons, and you
will declare and win.
John Conway had a nice way to look at this: Imagine an addi-
tional stack, the “whim stack,” of size 2; then play the normal Nim
strategy of handing your opponent 0-positions. If the only way for
you to achieve a 0-position is to remove the whim stack, you declare
appropriately instead.
The only trouble is: What if achieving a 0-position requires re-
moving just 1 chip from the whim stack? That won’t happen if there’s
a stack of size 2 or another stack of size 3, and you’re all set if all the

240
 Nimbers and the Hamming Code

stacks are singletons, so you can be in trouble only if there’s one or

more “tall” stack of size at least 4.
Conway’s solution to that problem is to imagine that the whim
stack has size only 1, when there’s a tall stack in play. We need to
check that the transition causes no problem, but that turns out to be
easy. Your opponent will never be the one to kill the only tall stack
because, being in a 0-position with the whim stack (thus a 1-position
without), he can’t be looking at just one tall stack. If it is you who
are facing one tall stack, and leaving your opponent with a 0-position
requires reducing that stack to size 0, 1, 2, or 3, then you can instead
reduce it to 3, 2, 1, or 0 respectively and have a 0-position with whim
stack of size 2.
In summary: If it is your move, and there’s no tall stack, you can
win as long as the position’s Nim value, counting an imaginary stack
of size 2, is not 0; if there is a tall stack, then the Nim value, counting
an imaginary stack of size only 1, must not be 0. ♡
Let’s put nimbers to work on hats.

Option Hats
One hundred prisoners are told that at midnight, in the dark, each
will be fitted with a red or black hat according to a fair coinflip. The
prisoners will be arranged in a circle and the lights turned on, enabling
each prisoner to see every other prisoner’s hat color. Once the lights
are on, the prisoners will have no opportunity to signal to one another
or to communicate in any way.
Each prisoner will then be taken aside and given the option of
trying to guess whether his own hat is red or black, but he may choose to
pass. The prisoners will all be freed if (1) at least one prisoner chooses
to guess his hat color, and (2) every prisoner who chooses to guess
guesses correctly.
As usual, the prisoners have a chance to devise a strategy before
the game begins. Can they achieve a winning probability greater than
50%?

Solution: The prisoners have virtually no chance of winning if they

all guess, so it looks reasonable to pick one prisoner (say, Joe) to guess
and have everyone else decline; then, at least, the prisoners are all
freed if Joe gets it right. In fact, it almost looks like there’s a proof that
they can’t do better: Every individual guess has only a 50% probability

241
 Mathematical Puzzles

of being right, so what’s the advantage of ever having more than one
person guess?
Suppose we fix some strategy S that tells each prisoner whether to
guess black (“B”), guess red (“R”), or decline to guess (“D”), depend-
ing on the 99 hats he sees. Imagine that we write down a huge matrix
indicating, for each of the 2100 possible hat configurations, what each
prisoner does according to S. If each guess (B or R) is boldfaced if
it is correct, then the matrix will have exactly the same number of
boldfaced guesses as lightfaced; because, if prisoner i’s guess is (say)
a boldface R in a particular configuration, it will be a lightface R for
the configuration that is exactly the same except that prisoner i’s hat
is changed to black.
In other words, no matter what the strategy, over all possible con-
figurations half the guesses are correct. That suggests an idea: What
if we crowd as many wrong guesses as possible into a few configura-
tions, and try to put just one correct guess into each remaining config-
uration?
Let’s do a little bit of arithmetic. Fixing a strategy S, let w be the
number of winning configurations (that is, those in which all pris-
oners who guess are correct). Let x be the average number of (cor-
rect) guesses in winning configurations, and y the average number of
wrong guesses in losing configurations. Then wx ≤ (2100 −w)y with
equality only when all guesses in the losing configurations are wrong.
To maximize w, we would like to have every prisoner guess and all
be wrong, in losing configurations; in winning configurations, to have
just one prisoner guess and be correct. If we could achieve that for n
prisoners, we would get a winning probability of n/(n+1). We call
this number the count bound. Achieving the count bound would be
very good news indeed for our 100 prisoners, as they would all be
freed with probability 100/101—better than 99%.
But let’s not get ahead of ourselves. For one thing, we could never
get 100/101 probability of winning for our 100 prisoners, because
that would require that the number of losing configurations be ex-
actly 1/101 times the total number of configurations, and 2100 is not
divisible by 101.
But n+1 does divide evenly into 2n for n = 3; let’s see if we can
get probability 3/4 of winning for three prisoners. We need two of the
23 = 8 hat configurations to be losers; the logical ones to try are all
red, and all black. If we instruct each prisoner to guess “red” when he
sees two black hats and “black” when he sees two red, that will have
the desired effect: They’ll all guess wrong in the losing configurations.

242
 Nimbers and the Hamming Code

Let’s see: If we ask them to decline to guess otherwise, we get just what
we want: In any configuration with both colors present, the prisoner
with the odd hat will guess correctly while the other two decline!
Here’s the matrix for this strategy:

Notice that we have numbered the prisoners in binary—actually,

we have nimbered them. If we define the nimsum of a configuration to
be the nimber-sum of the nimbers of the red-hatted prisoners, we find
that the two losing configurations, RRR and BBB, are the ones with
nimsum 0. Every other configuration has a non-zero nimsum which
corresponds to some prisoner—namely, the one who guesses (and is
right) when that configuration occurs.
Thought of this way, the above strategy generalizes to any number
of prisoners that is one less than a power of two—that is, whenever
n = 2k − 1 for some positive integer k. The prisoners are nimbered
from 1 to k, that is, from 000 . . . 001 to 111 . . . 111. The losing config-
urations are those for which the nimsum of the red-hatted prisoners
is 0. The strategy is as follows: Each prisoner assumes that the con-
figuration is not a losing one, and if that assumption tells him what
his own hat color is, he guesses accordingly. Otherwise he declines.
What that amounts to is this. Prisoner i computes the nimber-
sum mi of the nimbers of the red-hatted prisoners he sees; we call mi
prisoner i’s “personal” nimsum. If mi = 0, prisoner i reasons that
his hat must be red (else the nimsum of the whole configuration is 0,
which every prisoner is assuming is not the case). If mi = i, prisoner
i’s own nimber, then his hat must be black (else the nimsum of the

243
 Mathematical Puzzles

whole configuration would be i ⊕ i = 0). If mi is anything other than

0 or i, prisoner i knows the configuration is not a losing one. That is
great news, but it doesn’t tell prisoner i what his hat color is, so he
declines to guess.
This works because: (1) If the configuration has nimsum 0, every
prisoner will guess, and they will all be wrong; (2) If the configura-
tion’s nimsum is s ̸= 0, then prisoner s will guess correctly while ev-
ery other prisoner declines. Thus the prisoners will achieve exactly
the count bound for their success probability, namely n/(n+1) =
(2k − 1)/2k .
We have exploited the following property of the set L of configura-
tions with zero nimsum: Every configuration not in L can be changed
to one in L by flipping the color of someone’s hat. (Here, there is
just one such hat: the one belonging to the prisoner whose nimber is
the configuration’s nimsum.) Interpreted as a set of binary strings of
length n (red=1, black=0), L is said to be an error-correcting code, and
the above bit-changing property makes this particular code (called
the Hamming code of length n) a “perfect 1-error correcting code.”
The Hamming code and other error-correcting codes are objects of
enormous practical as well as theoretical interest—we’ll identify some
more serious applications later.
What can the prisoners do if their number n is not one less than a
power of two? Well, they can find the biggest k such that 2k − 1 ≤ n,
and have prisoners 1 through 2k − 1 run the above protocol among
themselves, while the others decline to guess no matter what they see.
This will not generally be optimal but it won’t be far off; the probabil-
ity of failure will always be less than 2/(n+1), thus small (for large n)
and never as much as twice the failure probability associated with the
count bound. Our group of 100 prisoners can have 63 prisoners run
the Hamming code protocol among themselves while the remaining
37 stand by, declining to guess. That will give the prisoners a pretty
impressive 63/64 = 98.4375% probability of winning.
To get the precise optimal winning probability for values of n that
aren’t one less than a power of 2, the prisoners would need to find
the smallest set L of configurations with the property that any con-
figuration is at most one color-flip away from a configuration in L.
We know from the count bound that such a set, sometimes called a
“radius-1 covering code,” must be of size at least 2n /(n+1) rounded
up to an integer.
Let’s see this reasoning in action for a couple of small numbers.
For n = 4, the count bound says that we need |L| ≥ 4, and we

244
 Nimbers and the Hamming Code

already get winning probability 1 − 4/16 = 3/4 by having three pris-

oners employ the Hamming code, so that solution is optimal. For
n = 5, though, the count bound allows a radius-1 covering code of
size ⌈32/5⌉ = 6. In fact, 6 is impossible but there’s a radius-1 covering
code of size 7:

{00000, 00111, 01111, 10111, 11001, 11010, 11100}.

With this the prisoners can get winning probability 1−7/64 = 57/64,
beating the Hamming code. Notice here that if, for example, the ac-
tual configuration is 11111 (all hats red), two prisoners—numbers 1
and 2, counting from the left—will guess (correctly) that their hats
are red, the first to avoid the codeword 01111, the second to avoid
10111.
It turns out that for large n, radius-1 covering codes that are (rela-
tively) close in size to 2n /(n+1) can always be found. They are much
harder to describe than Hamming codes, but if the number of pris-
oners is large and awkward (e.g., equal to 2k − 2 for some k), hard
work will enable the prisoners to cut down their losing probability by
a factor of almost 2.

Chessboard Guess
Troilus is engaged to marry Cressida but threatened with deportation,
and Immigration is questioning the legitimacy of the proposed mar-
riage. To test their connection, Troilus will be brought into a room
containing a chessboard, one of whose squares is designated as spe-
cial. On every square will be a coin, either heads-up or tails-up. Troilus
gets to turn over one coin, after which he will be ejected from the room
and Cressida brought in.
Cressida, after examining the chessboard, must guess the desig-
nated square. If she gets it wrong, Troilus is deported.
Can Troilus and Cressida save their marriage?

Solution: Troilus must somehow impart 6 bits worth of informa-

tion (since the chessboard has 64 = 26 squares) with his one turned
coin. He does have 64 different things he can do, so this is not a priori
impossible, but he and Cressida are going to need to be very efficient.
The natural approach—if you know about nimbers—is to assign
a nimber of length 6 to each square of the chessboard. When Cres-
sida comes in, she sums the nimbers of the squares whose coin shows

245
 Mathematical Puzzles

heads; that gives a nimber and she now guesses that its associated
square is the designated special square.
So Troilus needs to do is make sure the sum points to the right
square, but that’s easy. He computes the sum of the nimbers of the
heads-squares of the board as presented to him, and compares it with
the nimber of the designated special square. If their sum (i.e., differ-
ence!) is the nimber b, he turns the coin on square b. That will add
(i.e., subtract) b from the sum and leave the sum pointing to the spe-
cial square.

Majority Hats
One hundred prisoners are told that at midnight, in the dark, each
will be fitted with a red or black hat according to a fair coinflip. The
prisoners will be arranged in a circle and the lights turned on, enabling
each prisoner to see every other prisoner’s hat color. Once the lights
are on, the prisoners will have no opportunity to signal to one another
or to communicate in any way.
Each prisoner is then taken aside and must try to guess his own
hat color. The prisoners will all be freed if a majority (here, at least
51) get it right.
As before, the prisoners have a chance to devise a strategy before
the game begins. Can they achieve a winning probability greater than
50%? Would you believe 90%? How about 95%?

Solution: Too bad that a tie is insufficient to free the prisoners:

They could force a tie (see “Half-Right Hats” in Chapter 2) by having
50 of them guess “black” when they see an even number of red hats
and “red” otherwise, while the other 50 do the reverse. That guaran-
tees that exactly 50 prisoners will guess correctly. Unfortunately, the
way the problem is phrased, that guarantees that the prisoners lose.
But maybe we can use this idea some other way.
Before we begin our analysis, though, it’s worth checking one pos-
sibility that may already have occurred to you. What if each prisoner
simply guesses that his hat color is the same as the majority of those he
sees? This works as long as the number of red-hatted prisoners is not
exactly 50; in the latter case, the plan comes spectacularly a cropper,
with every prisoner guessing incorrectly. But the likelihood of exactly
50 red hats is not very high: There are 100 choose 50 ways this can
happen, out of 2100 ways( to)pick the hat colors, so the probability
of exactly 50 red hats is 10050 /2
100
= 0.07958923738... . Thus, this

246
 Nimbers and the Hamming Code

“preponderance” strategy frees the prisoners with probability better

than 92%; not bad!
The fact that when it goes wrong, the preponderance scheme pro-
duces 100 wrong guess, is a feature, not a bug. What we have is a
problem very similar to the one whose solution recalls the so-called
St. Petersburg paradox. You are in Las Vegas, and determined to go
home with more money than you came with; how can you maximize
your probability of doing so, even though you expectation from gam-
bling is always negative? The answer was discussed in Chapter 14 in
connection with the puzzle Roulette for the Unwary. Set aside 2k − 1
dollars that you can afford to lose, with k as large an integer as pos-
sible; then make successive even-money (and as fair as possible) bets
of sizes $1, $2, $4, etc. until you win a bet. At that point you quit, $1
ahead.
You end up probably ahead $1 but possibly broke, and that’s the
hallmark of a good scheme: You either just barely accomplish your
goal, or miss it by a mile.
The same applies to Majority Hats. Since every guess is a 50%
proposition, the ideal scheme would have just two possible outcomes:
Win barely (meaning, in this case, have exactly 51 correct guesses)
or lose spectacularly (every guess wrong). If you could do that for
100 prisoners, their probability p of success would satisfy 51p =
49p + 100(1−p), since for any scheme, the expected number of right
guesses has to equal the expected number of wrong guesses. That
works out to p = 100/102 > 98%, very impressive—if it were possi-
ble.
In general, with n prisoners, this calculation shows that when n
is even their success probability cannot exceed n/(n+2); when n is
odd, and the prisoners need only one more right guess than wrong,
they can hope for a success probability as high as n/(n+1). We will
show that this “count bound” is achievable when n is one or two less
than a power of 2.
Suppose, for example, that there are 31 = 25 −1 prisoners, and let’s
assign them nimbers from 1 to 11111 (the binary representation of 31).
We divide the prisoners into five groups according to the position of
the leftmost 1 in their nimbers; thus Group 1 consists of just prisoner
1, Group 2 contains prisoners 2 and 3, and in general each group has
twice as many prisoners as the previous group.
Prisoner 1 guesses “black.” If he’s right—and, remember, every-
one else can see if he’s going to be right or not—every other group
behaves as if it’s playing Half-Right Hats. The agreement could be, for

247
 Mathematical Puzzles

example: All prisoners with even nimbers guess as if they know the
total number of red hats in their group is even, while all those with odd
nimbers assume the opposite. Then groups 2 through 5 will each be
exactly half right, and the prisoners end up with a 16-to-15 majority
correct.
If prisoner 1’s hat is red, Group 2 plays the Unanimous Hats
strategy—they all guess as if they knew the number of red hats in
their group were even. If they’re right (probability 1/2), then each re-
maining group splits itself as before and again the prisoners win by
one.
If Group 2 is doomed to failure as well as Group 1—which, again,
everyone else can see—then there are three wrong guesses coming,
but they are rescued by Group 3 which now attempts to be all cor-
rect. This pattern continues and always produces a 16-15 majority
unless every one of the five groups contains an odd number of red
hats, which happens with probability only 1/32. So the 31 prisoners
achieve success probability 1 – 1/32 = 96.875%. We know this is un-
beatable because it achieves the ideal: The prisoners either enjoy a
1-guess victory, or everyone is wrong.
We can achieve this ideal (though probability of success will be
only 1 − 1/16) with 30 prisoners as well; here there’s no group of size
1, and the group of size two plays the Unanimous Hats strategy to try
to achieve the 2-guess advantage.
What about 100 prisoners? The count bound is never exactly
achievable unless n+1 or n+2 is a power of 2, because there are 2n hat
configurations and thus any protocol’s probability of winning must be
a multiple of 1/2n . One hundred is not a good number, but we could
let prisoners 1 through 62 use our “St. Petersburg protocol,” while the
remaining 38 prisoners just split themselves using the Half-Right Hats
strategy, so as not to get in the way of the others. That frees the pris-
oners with probability 1 – 1/32 = 96.875%, pretty good. But it seems
as if this strategy wastes 38 prisoners.
Let’s go back to small numbers again. The smallest number that’s
not one or two less than a power of 2 is 4. Can 4 prisoners win with
probability better than 1/2? Yes: Have prisoner 1 guess black. If he’s
wearing a black hat, the other three prisoners can run their St. Peters-
burg protocol to get 2-out-of-3 right with probability 3/4. If prisoner
1 has a red hat, the other three play the Unanimous Hats strategy and
thus all three are right with probability 1/2. Altogether the 4 prisoners
achieve their 2-guess majority with probability 12 · 34 + 12 · 21 = 5/8.

248
 Nimbers and the Hamming Code

With this observation we can split our 100 prisoners into groups
of sizes 4, 6, 12, 24, and 48, with just 6 left over. The group of 4 plays
as above, succeeding with probability 5/8. As usual, if they are going
to succeed (which everyone else can see), all other prisoners play the
Half-Right Hats strategy. If they are going to fail, the next group (size 6)
plays Unanimous Hats, and if they fail the group of 12 plays Unanimous
Hats, etc. The final group of 6 always plays Half-Right Hats.
Result? The prisoners win unless the first group of 4 fails and the
next four groups all suffer an odd number of red hats. This strategy
thus frees the prisoners with probability 1− 38 · 12 · 12 · 12 · 12 = 1−3/128 =
97.65625%.
Can we tweak this any further? Yes, by recursively computing
the best strategies of this kind for achieving k correct guesses out
of n, for all n and k up to 100. This “dynamic programming” ap-
proach can reach a success probability of 1129480068741774213/
1152921504606846976, about 97.96677954%; very close indeed to
the count bound of 100/102 = 98.039215686%.
That last solution didn’t end up making much use of nimbers, but
the next puzzle brings us back to the Hamming code.

Fifteen Bits and a Spy

A spy’s only chance to communicate with her control lies in the daily
broadcast of 15 zeros and ones by a local radio station. She does not
know how the bits are chosen, but each day she has the opportunity
to alter any one bit, changing it from a 0 to a 1 or vice-versa.
How much daily information can she communicate?

Solution: Since there are 16 things the spy can do (change any bit
or none), she can in principle communicate as much as four bits of
information each day to her control. But how?
The answer is easy once you have nimbers in your arsenal. The spy
and her control assign the 4-bit nimber corresponding to the number

249
 Mathematical Puzzles

k to the kth bit of the broadcast, and their “message” is defined to be

the sum of the nimbers of the ones in the broadcast—the “nimsum”
of the broadcast.
The claim is that the spy can send any of the 16 possible nimsums
at will, thus achieving a full four bits of communication. Suppose she
wishes to send the nimsum n but the sum of the nimbers correspond-
ing to the ones in the station’s intended broadcast is m ̸= n. Then she
flips the n ⊕ mth bit. The resulting nimsum is m ⊕ (n ⊕ m) = n. Of
course, of the broadcast’s nimsum already happens to be n, she leaves
it alone. ♡

We will finish this chapter with a brief introduction to the raison

d’etre of error-correcting codes: sending messages over an imperfect
channel.
Almost every means of communication suffers from some proba-
bility of error, whether it be voice, telegraph, radio, TV, typed letters,
hand-written words, photographs, or bit streams over the internet. It’s
pretty amazing that you can send a billion bits to your friend in an
email attachment, and not a single one is accidentally flipped!
This comes about through the use of error-correcting codes. If bits
had to be sent so slowly that every one was guaranteed to arrive intact
with probability 1 − 1/109 , we could forget about streaming movies.
No, the individual electical, radio and light pulses that zip around
the internet are imperfect and are mis-sent or mis-read with alarming
frequency. What saves us is that information is coded in such a way
that errors can be detected and corrected.
The simplest error-detection method is one you may already know
about, and is typical for, say, credit card numbers. By choosing the last
“check” digit correctly, your credit card company can ensure that the
sum of the digits in your credit card number is (say) a multiple of 10.
Then, if you mis-enter one digit of your credit card during a purchase,
the digit-sum will be thrown off, and a warning will appear.
Error-correction is trickier but you’ve seen it above in the person
of the Hamming code. Simplest case: You want to send a string of
bits and be confident that they all arrive intact, despite some small
probability that any given bit will be corrupted. Send each bit three
times! Only in the unlikely event that two bits in the same group of
three are corrupted will your message get unfixably mangled.
In fact, this is the Hamming code of length 3 that we used when
three prisoners were faced with either of our hat problems. The code-
words were 000 (all hats black) and 111 (all hats red).

250
 Nimbers and the Hamming Code

Of course, tripling the lengths of our messages is rather a heavy

price to pay. If we are confident that the likelihood of more than one
error in a given group of 2k − 1 bits is small, we can use the Hamming
code of length n = 2k − 1, as follows.
Suppose the message to be sent is (or could be represented by)
a long string of N bits. Break up the message into “blocks” of size
m = n−k. Before transmission, each block is replaced by a codeword
of length n; at the receiving end, these are “decoded,” and the original
blocks recovered correctly, provided the codeword was received with
at most one flipped bit.
We can state this as a theorem:
Theorem. At the cost of adding k bits to the transmission, a block of bits of
length n−k, where n = 2k − 1, can be sent in such a way that as long as at
most one bit is corrupted, the original block can be faithfully recovered at the
receiving end.
Proof. We will do a little more than is necessary for the proof, in order
to show that all this can be done with very little computational effort
by transmitter and receiver.
First, we nimber the bits in the n-bit codeword, using k-bit
non-zero nimbers from 000 . . . 001 to 111 . . . 111. There are many
ways to do this, but it’s especially convenient to assign nimbers
20 = 000 . . . 001, 21 = 000 . . . 010, 22 = 000 . . . 100, . . . , 2n−1 =
100 . . . 000 to the rightmost k positions, reading from right to left.
The same nimbering is used for a block of length n−k, regarded
as the leftmost n−k bits of a codeword. For example, if k = 3, the bits
of our 7-bit codewords, and the corresponding 4-bit blocks, could be
nimbered as in the following figure:

The nimsum of a block or codeword is the nimber sum of the po-

sitions in which a “1” is found. The nimsum of the codeword in the
figure is 0 = 000; indeed, we define our codewords to be the exactly
the n-bit strings with zero nimsum. The nimsum of the block in the
figure happens to be 3 = 011.

251
 Mathematical Puzzles

The easiest way to transform a block into a codeword is simply

to append k “check bits” to the right-hand end of the block, in such
a way that the nimsum of the resulting n-bit string is zero. That is
particularly easy to do with the nimbering we recommend, because
the check bits will then be exactly the nimsum of the block!
The reason for this pleasant coincidence is that by choosing the
“basis” nimbering of the check bits, we ensure that the nimsum of
the check bits is the string of check bits itself. Thus, if the nimsum of
the block is s, when we add the check bits the new nimsum will be
s ⊕ s = 0.
In this way the “encoding” part of our Hamming code is made
as easy as possible; what about the “decoding” part? No problem. As
receiver of an n-bit string, we compute its nimsum; if we get 0, we
assume it was transmitted correctly and accept its leftmost n−k bits
as the intended message block.
If we get nimsum s ̸= 0, we assume that one, and only one, error
has been made. If so that error must be in the bit nimbered s; so we flip
that bit, and only now take the leftmost n−k bits as the intended mes-
sage block. (The transmission error could have occurred in a check
bit, in which case the extracted message will not change.)

We have seen that the property we needed from codewords in

the hat puzzles—that each non-codeword is one bit-flip away from a
unique codeword—is used to correct transmission errors. The Ham-
ming code is one of many codes, some of which detect or correct
many errors, and have lots of other neat properties. (Example: Reed-
Muller codes, Reed-Solomon codes, and Golay codes were used to
transmit photos of Mars and Jupiter from various Mariner and Voy-
ager spacecraft.) Error-correcting codes have played a major role in
the theory of computing, as well, leading (among other things) to
miraculous strings called “probabilistically checkable proofs” and
thence to huge advances in our understanding of how well com-
puters can do approximate calculations. And error-correcting codes
will need to play a major role if quantum computing is to become
practical.

252
 18. Unlimited Potentials
Puzzles often involve processes—systems that evolve in time randomly,
deterministically, or under your control. Popular questions include:
Can the system reach a certain state? Does it always reach a certain
state? If so, how long does it take?
The key to such a puzzle may be to identify some parameter of the
current state of the system that measures progress toward your goal.
We call such a parameter a potential.
For example, suppose you can show that the potential never in-
creases, but starts below the potential of the desired end state. Then
you have proved that the end state is unreachable.
On the other hand, perhaps the end state has potential zero, the
beginning state has positive potential p > 0, and every step of the
process reduces the potential by at least some fixed amount ε > 0.
Then you must reach the end state, and moreover must do so in time
at most p/ε.
If the loss per step in potential is not bounded, it could be that the
potential never reaches zero and may not even approach zero. But
if the system has only finitely many states, you don’t need to worry
about this possibility.
Can you find an appropriate potential for the following puzzle?

Signs in an Array
Suppose that you are given an m × n array of real numbers and per-
mitted, at any time, to flip the signs of all the numbers in any row or
column. Can you always arrange matters so that all the row sums and
column sums are non-negative?

Solution: We note first that there are only finitely many states here:
Each number in the array has at most two states (positive or negative)
so there can’t be more than 2mn possible states of the array.
Actually, there’s a better bound: If a line (a row or column) is
flipped twice, it’s as if it had never been flipped. Thus the state of

253
 Mathematical Puzzles

each line is either “flipped an odd number of times” or “flipped an

even number of times,” and it follows that the number of states of the
array can’t exceed 2m+n .
Back to the problem at hand. The obvious algorithm is to find a
line with negative sum and flip it; does this eventually work? After
all, flipping a row (say) with negative sum could result in changing
several column sums from positive to negative. Thus, the number of
lines with negative sum would not work as a potential.
Instead, let’s take the sum s of all the entries in the array! That
number goes up by 2c if you flip a line whose sum was −c. Since
there are only finitely many possible states, the potential cannot keep
going up, and you must reach a point where no line sum is negative.
Fair warning: From here on, it’s going to be trickier to spot the
right potential.

Righting the Pancakes

An underchef of the great and persnickety Chef Bouillon has made
a stack of pancakes, but, alas, some of them are upside-down—that
is, according to the great Chef, they don’t have their best side up. The
underchef wants to fix the problem as follows: He finds a contiguous
substack of pancakes (of size at least one) with the property that both
the top pancake and the bottom pancake of the substack are upside-
down. Then he removes the substack, flips it as a block and slips it
back into the big stack in the same place.

Prove that this procedure, no matter which upside-down pancakes

are chosen, will eventually result in all the pancakes being right side
up.

254
 Unlimited Potentials

Solution: The idea is that even though a flip might result in turning
more pancakes upside-down, it rights some pancake that’s relatively
high in the stack. Let’s penalize the underchef 2k points for an upside-
down pancake k pancakes from the bottom, and keep track of the total
of the penalties at any given time.
Every flip decreases the total penalty, because the penalty of the
target pancake exceeds the sum of all possible penalties below it. Thus
the penalty total must reach 0, at which time the task is finished.
An equivalent way of thinking about this potential function is to
code a stack of pancakes by a binary number whose left-most digit
is 1 if the top pancake is upside-down, and 0 otherwise. The next-to-
top pancake determines the next digit, and so forth. Flipping always
decreases the code until the code reaches 0.

Breaking a Chocolate Bar

You have a rectangular chocolate bar scored in a 6 × 4 grid of squares,
and you wish to break up the bar into its constituent squares. At each
step, you may pick up one piece and break it along any of its marked
vertical or horizontal lines.
For example, you can break three times to form 4 rows of 6 squares
each, then break each row five times into its constituent squares, ac-
complishing the desired task in 3 + 4 × 5 = 23 breaks.
Can you do better?

Solution: This delightful problem has embarrassed some brilliant

professional mathematicians, as well as many amateurs, all looking
for geometrical insight.
But geometry is a red herring: All you need for a potential function
is the number of pieces! You start with one piece, increase the number
of pieces by one every time you break, and end with 24 pieces. So there
are 23 breaks no matter how you do it.

Red Points and Blue

Given n red points and n blue points on the plane, no three on a line,
prove that there is a matching between them so that line segments
from each red point to its corresponding blue point do not cross.

255
 Mathematical Puzzles

Solution: Suppose you pair the red and blue points up arbitrarily
and draw the connecting line segments. If the line segment connect-
ing r1 and b1 intersects the segment connecting r2 and b2 , then these
segments are the diagonals of a quadrilateral and by the triangle in-
equality, the sum of the lengths of the non-crossing segments from r1
to b2 and from r2 to b1 is less then sum of the lengths of the diagonals.
The trouble is, if you re-pair the points as suggested above, you
may have created new intersections between these segments and oth-
ers. So “number of intersections” is not the right potential; how about
“total length of line segments?”
That works beautifully! Just take any matching that minimizes the
sum of the lengths of its line segments, and by the above argument, it
contains no crossings.

Bacteria on the Plane

Suppose the world begins with a single bacterium at the origin of the
infinite plane grid. When it divides its two successors move one vertex
north and one vertex east, so that there are now two bacteria, one at
(0,1) and one at (1,0). Bacteria continue to divide, each time with one
successor moving north and one east, provided both of those points
are unoccupied.

Show that no matter how long this process continues, there’s al-
ways a bacterium inside the circle of radius 3 about the origin.

Solution: It’s worth trying this to see how things get clogged up
near the origin, preventing that area from clearing. How can we
demonstrate this using a potential function?

256
 Unlimited Potentials

Since each bacterium divides into two, it’s natural to give each
child half the potential of its parent. We can do this by penalizing its
potential by a factor of 2 for each additional edge in its shortest path
to the origin.
This can be done by assigning value 2−x−y to a bacterium at (x,y),
so that the total assigned value begins at 1 and never changes.
The total potential of all bacteria on the half line y = 0 (i.e., the
positive X-axis) cannot exceed 1 + 1/2 + 1/4 + · · · = 2, nor can the
potential of the points on the half-line y = 1 exceed 1/2 + 1/4 +
1/8 + · · · = 1; continuing in this manner, we see that even if the
whole north-east quadrant were full of bacteria, the total potential
would only add up to 4.
But the potentials of bacteria at the nine non-negative integer
points inside the circle x2 + y 2 = 9 already add up to 49/16, more
than 3. Thus the potentials of all the bacteria not inside this circle add
up to less than 1, and it follows that there must be bacteria left inside
the circle no matter how long the process continues.

Pegs on the Half-Plane

Each grid point on the XY plane on or below the X-axis is occupied
by a peg. At any time, a peg can be made to jump over a neighbor peg
(horizontally, vertically, or diagonally adjacent) and onto the next grid
point in line, provided that point was unoccupied. The jumped peg is
then removed.
Can you get a peg arbitrarily far above the X-axis?

Solution: The difficulty is that as pegs rise higher, grid points be-
neath them are denuded. What is needed is a parameter P which is re-
warded by highly placed pegs, but compensatingly punished for holes
left behind. A natural choice would be a sum over all pegs of some
function of the peg’s position. Since there are infinitely many pegs,
we must be careful to ensure that the sum converges.
We could, for example, assign value ry to a peg on (0, y), where
r is some real number greater than 1, so that∑the values of the pegs
0
on the lower Y -axis sum to the finite number y=−∞ ry = r/(r−1).
Values on adjacent columns will have to be reduced, though, to keep
the sum over the whole plane finite; if we cut by a factor of r for each
step away from the Y -axis, we get a weight of ry−|x| for the peg at

257
 Mathematical Puzzles

(x, y), and a total weight of

r 1 1 1 1 r2 + r
+ + + + + ··· = <∞
r−1 r−1 r−1 r(r −1) r(r−1) (r−1)2

for the initial position.

If a jump is executed, then at best (when the jump is diagonally
upward and toward the Y -axis), the gain to P is vr4 and the loss v +
vr2 , where v is the previous value of the jumping peg.√As long as r is
at most the square root of the “golden ratio” θ = (1+ 5)/2 ≈ 1.618,
which satisfies θ2 = θ + 1, this gain can
√ never be positive.
If we go ahead and assign r = θ, then the initial value of P
works out to about 39.0576; but the value of a peg at the point (0, 16)
is θ8 ≈ 46.9788 by itself. Since we cannot increase P , it follows that
we cannot get a peg to the point (0, 16). But if we could get a peg to
any point on or above the line y = 16, then we could get one to (0, 16)
by stopping when some peg reaches a point (x, 16), then redoing the
whole algorithm shifted left or right by |x|. ♡

Pegs in a Square
Suppose we begin with n2 pegs on a plane grid, one peg occupying
each vertex of an n-vertex by n-vertex square. Pegs jump only hori-
zontally or vertically, by passing over a neighboring peg and into an
unoccupied vertex; the jumped peg is then removed. The goal is to
reduce the n2 pegs to only 1.
Prove that if n is a multiple of 3, it can’t be done!

Solution: Color the points (x, y) of the grid red if neither x nor y is
a multiple of 3, otherwise white. This leaves a regular pattern of 2 × 2
red squares (as in the figure).

258
 Unlimited Potentials

If two pegs are (orthogonally) adjacent on the grid, both on red

points or both on white, the peg remaining after the jump will be on
white. If one is on red and the other on white, however, the peg re-
maining after the jump will be on red. It follows that if you start with
any configuration having an even number of pegs on red squares, then
this property will persist forever regardless of what jumps are made.
It is easy to check that a 3 × 3 square of pegs, no matter where it
is placed on the plane grid, hits an even number of red points. Since
an n × n square with n a multiple of 3 is composed of such squares, it
too will always hit an even number of red points. Suppose, however,
that it were possible to reduce such a square to a single peg. Then we
could shift the original square so that the surviving peg ended up on
a red point, and this contradiction concludes the proof.
It is routine, but not particularly easy or enlightening, to show that
if n is not a multiple of 3, you can reduce an n × n square to a single
peg.

First-Grade Division
On the first day of class Miss Feldman divides her first-grade class into
k working groups. On the second day, she picks the working groups
a different way, this time ending up with k+1 of them.
Show that there are at least two kids who are in smaller groups on
the second day than they were on the first day.

Solution: Most people find this simple-seeming combinatorial

statement frustratingly difficult to prove, unless they hit upon the
“magic” potential. Think of each working group as a task force for
a job that requires a total of one unit of effort; then, assume that each
child in a working group of size s contributes an amount 1/s worth
of effort.
The sum of these “effort” fractions is evidently k on the first day,
k+1 on the second. A given child’s contribution to these numbers
cannot increase by as much as 1 from one day to the next (since his
contribution is greater than 0 on day 1, and at most 1 on day 2). So
at least two kids must be contributing more to the total effort on their
second day—meaning that both have moved to smaller groups.

259
 Mathematical Puzzles

Infected Checkerboard
An infection spreads among the squares of an n × n checkerboard
in the following manner: If a square has two or more infected neigh-
bors, then it becomes infected itself. (Neighbors are orthogonal only,
so each square has at most four neighbors.)
For example, suppose that we begin with all n squares on the main
diagonal infected. Then the infection will spread to neighboring diag-
onals and eventually to the whole board.
Prove that you cannot infect the whole board if you begin with
fewer than n infected squares.
Solution: This puzzle was presented to me (by NYU’s Joel
Spencer) as having a “one-word solution.” That’s an exaggeration,
perhaps, but not a huge one.
At first it might seem that to infect the whole board, you need to
start with a sick square in every row (and in every column). That
would imply the conclusion, but it’s not true. For example, sick
squares in alternating positions in the leftmost column and bottom
row can infect the whole board.
What we really need is a potential function, and the one that works
like a charm—and the one word Spencer had in mind—is “perime-
ter.”
The perimeter of the infected region is just the total length of its
boundary. We may as well assign each grid edge length 1, so the
perimeter could as well be defined as the number of edges that have
a sick square on one side and either a well square or nothing on the
other side.
Here is the key observation: When a square becomes infected (by
two or more neighbors), the perimeter cannot increase! Indeed, it is
easy to check that when just two neighbors administer the dose, the
perimeter remains the same; if the well square had three or four sick
neighbors, the perimeter actually goes down.
If the whole board gets infected, the final perimeter is 4n; since
the perimeter never increased, the initial perimeter must have been at
least 4n. But to start with perimeter 4n there must have at least n sick
squares.

Impressionable Thinkers
The citizens of Floptown meet each week to talk about town poli-
tics, and in particular whether or not to support the building of a

260
 Unlimited Potentials

new shopping mall downtown. During the meetings each citizen talks
to his friends—of whom there are always an odd number, for some
reason—and the next day, changes (if necessary) his opinion regard-
ing the mall so as to conform to the opinion of the majority of his
friends.
Prove that eventually, the opinions held every other week will be
the same.
Solution: To prove that the opinions eventually either become fixed
or cycle every other week, think of each acquaintanceship between
citizens as a pair of arrows, one in each direction. Let us say that an
arrow is currently “bad” if the opinion of the citizen at its tail differs
from the next week’s opinion of the citizen at its head.
Consider the arrows pointing out from citizen Clyde at week t−1,
during which (say) Clyde is pro-shopping mall. Suppose that m of
these are bad. If Clyde is still (or again) pro on week t+1, then the
number n of bad arrows pointing toward Clyde at week t will be exactly
m.
If, however, Clyde is anti-shopping mall on week t+1, n will be
strictly less than m since it must have been that the majority of his
friends were anti on week t. Therefore a majority of the arrows out of
Clyde were bad on week t−1 and now only a minority of the arrows
into Clyde on week t are bad.
The same observations hold, of course, if Clyde is anti on week
t−1.
But, here’s the thing: Every arrow is out of someone on week t−1,
and into someone on week t. Thus, the total number of bad arrows
cannot rise between weeks t−1 and t and, in fact, will go strictly down
unless every citizen had the same opinion on week t−1 as on week
t+1.
But, of course, the total number of bad arrows on a given week
cannot go down forever and must eventually reach some number k
from which it never descends. At that point, every citizen will either
stick with his opinion forever or flop back and forth every week.

Frames on a Chessboard
You have an ordinary 8 × 8 chessboard with red and black squares. A
genie gives you two “magic frames,” one 2 × 2 and one 3 × 3. When
you place one of these frames neatly on the chessboard, the 4 or 9
squares they enclose instantly flip their colors.

261
 Mathematical Puzzles

Can you reach all 264 possible color configurations?

Solution: There are 249 ×236 ways to select frame locations, enough
in theory to get all 264 color configurations. But of course there may
be many ways to achieve a given configuration, and maybe we can
find a potential that shows some configurations are unobtainable.
In fact, you might recall that the puzzle Odd Light Flips, of Chapter
15, offered a useful criterion: the ability to change the color (or bulb
status) of an odd number of items in any set. Is there a troublesome
set of chessboard squares?
A bit of experimenting might lead you to the following set. Label
squares as “special” if they are in row 3 or 6 or file c or f but not both
(see figure below), Then every 2 × 2 or 3 × 3 frame covers an even
number of special squares.

If the board starts with an even number of red special squares

(as the usual chessboard does), you cannot reach any configuration
where the number of red special squares is odd, and vice-versa. Either
way, you cannot reach all color configurations.

Bugs on a Polyhedron
Associated with each face of a solid convex polyhedron is a bug which
crawls along the perimeter of the face, at varying speed, but only in the
clockwise direction. Prove that no schedule will permit all the bugs to
circumnavigate their faces and return to their initial positions without
incurring a collision.

262
 Unlimited Potentials

Solution: Let us observe first that we may assume no bug begins

on a vertex (by advancing or retarding bugs slightly). We may also
assume that the bugs move one at a time, crossing a vertex each time.
At any time, we may draw an imaginary arrow from the center of
each face F , through F ’s bug, to the center of the face on the other
side of the bug. If we start at any face and follow these arrows, we
must eventually hit some face a second time, completing a cycle of
arrows on the polyhedron.

This cycle divides the surface of the polyhedron into two portions;
let us define the “inside” of the cycle to be that portion surrounded
clockwise by the cycle. Let P be the number of vertices of the polyhe-
dron inside the cycle.
Initially, P could be anything from 0 to all of the polyhedron’s
vertices; the extremes occur if there are two bugs on the same edge,
causing a cycle of length 2. In the P = 0 case, the two bugs are facing
each other, and doomed to collide.
When a bug on the cycle moves to its next edge, the arrow through
it rotates clockwise. The vertex through which it passed, previously
on the inside of the cycle, is now outside; other vertices may also have
passed from inside to outside the cycle, but there is no way for a vertex
to move inside. To see this, note that the new arrow now points inside
the cycle. The chain of arrows emanating from its head has no way
to escape the cycle so must hit the tail of some cycle arrow, creating a
new cycle with smaller interior. In particular, P has now dropped by
at least 1.
Since we can never restore P to its starting value, there is nothing
to do but hope that the bugs are carrying collision insurance. ♡

263
 Mathematical Puzzles

Bugs on a Line
Each positive integer on the number line is equipped with a green,
yellow, or red light. A bug is dropped on “1” and obeys the following
rules at all times: If it sees a green light, it turns the light yellow and
moves one step to the right; if it sees a yellow light, it turns the light
red and moves one step to the right; if it sees a red light, it turns the
light green and moves one step to the left.
Eventually, the bug will fall off the line to the left, or run out to
infinity on the right. A second bug is then dropped on “1,” again fol-
lowing the traffic lights starting from the state the last bug left them
in; then, a third bug makes the trip.
Prove that if the second bug falls off to the left, the third will march
off to infinity on the right.

Solution: We first need to convince ourselves that the bug will either
fall of to the left, or go to infinity on the right; it cannot wander forever.
To do so, it would have to visit some numbers infinitely often; let n
be the least of those numbers, but now observe that every third visit
to n will find it red and thus will incur a visit to n − 1, contradicting
the assumption that n − 1 was visited only finitely often.
With that out of the way, it will be useful to think of a green light
as the digit 0, red as 1, and yellow, perversely, as the “digit” 12 . The
configuration of lights can then be thought of as a number between 0
and 1 written out in binary,

x = .x1 x2 x3 . . . ,

where, numerically,
( 1 )1 ( 1 )2
x = x1 · + x2 · + ··· .
2 2
Think of the bug at i as an additional “1” in the ith position, defining
( 1 )i
y =x+ .
2
The point of this exercise is that y is an invariant, that is, it does not
change as the bug moves. When the bug moves to the right from point
i, the digit upon which it sat goes up in value by 12 ; therefore, x in-
( )i+1
creases by 12 , but the bug’s own value diminishes by the same
( )i
amount. If the bug moves to the left from i, it gains in value by 12 ,
but x decreases by a whole digit in the ith place to compensate.

264
 Unlimited Potentials

The exception is when the bug falls off to the left, in which case
both x and the bug’s own value drop by 12 , for a loss of 1 overall. When
the next bug is added, y goes up by 12 . To put it another way, the value
of x goes up by 12 if a bug is introduced and disappears to the right;
and drops by 12 if a bug is introduced and falls off to the left.
Of course, x must always lie in the unit interval. If its initial value
lies strictly between 0 and 12 , the bugs must alternate right, left, right,
left; if between 12 and 1, the alternation will be left, right, left, right.
The remaining cases can be checked by hand. If x = 1 initially (all
points red) the first bug turns point 1 green and drops off to the left;
the second wiggles off to infinity leaving all points red again, so the
alternation is left, right, left, right. If x = 0 initially (all points green),
the bugs will begin right, right (as the points change to all yellow, then
all red), and then left, right, left, right as before.
The x = 12 case is the most interesting because there are several
ways to represent 12 in our modified binary system: x can be all 12 ’s,
or it can start with any finite number (including 0) of 12 ’s, followed
either by 0111 . . . or 1000 . . . . In the first case, the leadoff bug turns
all the yellows to red as it zooms off to the right; thus, we get a right,
left, right, left alternation. The second case is similar, the first bug
wiggling off to the right, but again leaving all points red behind it. In
the third case, the bug changes the yellows to red as it marches out,
but when it reaches the red point, it reverses and heads left, turning
reds to green on its way to dropping off the left end. Thereafter, we
are in the x = 0 case, so the final pattern is left, right, right, left, right,
left, right.
Checking back all the cases, we see that indeed, whenever the sec-
ond bug went left, the third went right. ♡

Flipping the Pentagon

The vertices of a pentagon are labeled with integers, the sum of which
is positive. At any time, you may change the sign of a negative label,
but then the new value is subtracted from both neighbors’ values so
as to maintain the same sum.
Prove that, inevitably, no matter which negative labels are flipped,
the process will terminate after finitely many flips, with all values non-
negative.

Solution: Even with experimentation (highly recommended), it’s

a little hard to see what can be used as a potential. The number of

265
 Mathematical Puzzles

negative entries doesn’t necessarily go down, nor does the size of the
largest negative entry.
Similarly, neither the sum of the absolute values of the differences
between adjacent numbers, nor the sum of the squares of the differ-
ences, seems to go reliably down (or up).
But it turns out, if you take the sum of the squares of the differ-
ences between non-adjacent numbers, it works!
Suppose the numbers are, reading around the pentagon, a, b, c, d
and e. Then the sum in question is s = (a − c)2 + (b − d)2 + (c − e)2 +
(d − a)2 + (e − b)2 . Suppose b is negative and we flip it, replacing b
by −b, a by a + b, and c by c + b. Then the new s is equal to the old s
plus 2b(a + b + c + d + e), and since b is negative and a + b + c + d + e
is positive, s goes down. And since s is an integer, it goes down by at
least 1.
But s is non-negative (being a sum of squares) so it can’t go below
zero. That means we must reach a point where there’s no negative
entry left to flip, and we are done.
But there’s an even better potential—better because it works for
all polygons, not just the pentagon. It’s the sum, over all sets of con-
secutive vertices, of the absolute value of the sum of the numbers in
the set.
An even more remarkable solution was later found by Prince-
ton computer scientist Bernard Chazelle. Construct a doubly-infinite
sequence of numbers whose successive differences are the pentagon
(or polygon) values, reading clockwise around the figure. If the pen-
tagon’s labels began as 1, 2, −2, −3, 3, the sequence might be
. . . , −1, 0, 2, 0, −3, 0, 1, 3, 1, −2, 1, 2, 4, 2, −1, 2, 3, 5, . . . .
Notice that the sequence climbs gradually (since the sum of the
numbers around the polygon is positive) but not steadily (since some
of the numbers around the polygon are negative).
Now the key observation: Flipping a vertex has the effect of trans-
posing pairs of entries of the above sequence that were in the wrong
order—in other words, the flipping process turns into a sorting pro-
cess for our infinite sequence!
For example, if we flip the −2 on our pentagon to get 1, 0, 2, −5,
3, the sequence changes to
. . . , −1, 0, 0, 2, −3, 0, 1, 1, 3, −2, 1, 2, 2, 4, −1, 2, 3, 3, . . . .
It turns out to be easy to find a potential for the sorting progress
that declines by exactly one per turn, and to conclude that not only

266
 Unlimited Potentials

does the process always terminate, but it terminates in the same num-
ber of steps and in the same final configuration, no matter which la-
bels you flip!

Picking the Athletic Committee

The Athletic Committee is a popular service option among the faculty
of Quincunx University, because while you are on it, you get free tick-
ets to the university’s sports events. In an effort to keep the committee
from becoming cliquish, the university specifies that no one with three
or more friends on the committee may serve on the committee—but,
in compensation, if you’re not on the committee but have three or
more friends on it, you can get free tickets to any athletic event of
your choice.
To keep everyone happy, it is therefore desirable to construct the
committee in such a way that even though no one on it has three or
more friends on it, everyone not on the committee does have three or
more friends on it.
Can this always be arranged?

Solution: What’s the dumbest way to try to find a good committee?

How about this: Start with an arbitrary set of faculty members, as a
prospective Athletic Committee. Oops, Fred is on the committee and
already has three friends on the committee? Throw Fred out. Mona
is not on the committee, but has fewer than three friends on it? Put
Mona on. Continue fixing in this haphazard manner.
Now, why in the world would you expect this to work? Clearly,
the above actions could make things worse; for example, throwing
Fred off the committee might create many more Monas; maybe we
should have thrown off one of Fred’s on-committee friends instead.
So there doesn’t seem to be anything to prevent cycling back to the
same bad committee. Moreover, even if you don’t cycle back, there
are exponentially many possible committees and you can’t afford to
consider every one. Suppose there are 100 faculty members in all; then
the number of possible committees is 2100 > 1030 which, even if you
spent only a nanosecond considering each committee, would take a
thousand times longer than all the time that has passed since the Big
Bang.
But if you try it, you will find that after shockingly few corrections,
you end up with a valid committee. And this happens in situations
where there is only one valid committee, as well as where there are
many.

267
 Mathematical Puzzles

How can this be? Sounds like there must be some potential func-
tion at work here, something that is improving each time you throw
someone off or add someone to the current prospective committee.
Let’s see: When you throw someone off, you destroy at least three
on-committee friendships; when you put someone on, you add at
most two. Let F (t) be the number of friendships on the committee
minus 2 12 times the number of people on the committee at time t. Then
when Fred is thrown off, F (t) goes down by at least 12 . When Mona
is put on, F (t) again goes down by at least 12 . But F (0) can’t be more
than (100 × 99)/2 − 250 = 245 and F (t) can never dip below −250,
so there can’t be more than 2 × (245 − (−250)) = 990 steps total. (A
computer scientist would say that the number of steps in the process
is at worst quadratic in the number of faculty members.)
In practice, the number of steps is so small that if there are 100
faculty members and you start with (say) the empty committee, you
will reach a solution easily by hand. Of course, you’ll need access to
the friendship graph, so you might need to do some advance polling.
It might be interesting to see who claims friendship with whom that
isn’t reciprocated.
For the last puzzle, we use the mother of all potentials—potential
energy!

Bulgarian Solitaire
Fifty-five chips are organized into some number of stacks, of arbitrary
heights, on a table. At each tick of a clock, one chip is removed from
each stack and those collected chips are used to create a new stack.
What eventually happens?

Solution: It’s worth actually trying this—for example, with a deck

of cards (which often comes with two jokers and an order card, mak-
ing 55 cards in all.) If all you have is 52 cards, remove seven and try
the experiment with 45 cards.
You will discover, if you have sufficient patience, that you even-
tually reach the “staircase” configuration where the stack heights are
10, 9, 8, 7, 6, 5, 4, 3, 2, and 1; and it’s easy to see that this configura-
tion never changes. Proving that the staircase is always reached is a
typical job for a potential function, but what potential is minimized
by the staircase?
It helps to think of each chip as a square block, and each stack as
a column of square blocks. We can arrange the columns side-by-side
and descending in height, as in the figure below.

268
 Unlimited Potentials

Then one “operation” can be thought of as stripping the bottom

blocks from all the columns, turning them 90◦ , then adding the re-
sulting new column to the picture in the proper position to keep de-
scending order of heights, as in the next figure.
The configuration we’re aiming for still does not seem to minimize
anything natural—until we rotate the whole picture 45◦ to counter-
clockwise, as shown in the next figure. The staircase now represents
the “lowest” way to pack our 55 square blocks into a V-shaped cradle.

By “lowest” we mean, or at least should mean, the configuration

with least potential energy under gravity. Let the stacks be numbered

269
 Mathematical Puzzles

in descending order of height order, and the blocks in each stack num-
bered from bottom to top; then let sij be the ith block in the jth col-
umn. We can size the V-frame and blocks so that the height of the
center of square sij is i+j, and then the potential energy of a con-
figuration will be proportional to the sum over all the blocks of the
quantities i+j.
Now, what happens when we perform an operation? The blocks
along the right-hand arm of the V move to the left-hand arm; their
potential energy does not change. Nor does the potential of the other
blocks, since each of those just moves one position to the right. But
wait, there’s one more step: We might have to rearrange blocks to
get the (original) columns in descending order. That’s equivalent to
letting gravity push blocks downward (diagonally to the left), as in
the next figure, reducing the potential energy of the configuration.
If we’re already in the staircase configuration, we are at minimum
potential energy and rearrangement of the stacks is never again re-
quired. Suppose we’re not in the staircase configuration; what then?
The staircase configuration is possible, to begin with, only because
55 is a “triangular number,” equal to 1 + 2 + 3 + · · · + ( k for
) some k;
these are just the numbers that can be expressed as k+1 2 for some
positive integer k. In the staircase configuration, there’s a block in
every position (i, j) with 2 ≤ i+j ≤ k+1, and no block any higher. It
follows that if the number of blocks is triangular but we’re not in the
triangular configuration, there’s a “hole” somewhere at height k+1
and a block somewhere at height k+2.
As long as no rearrangement is needed, the hole and the block
each move one square to the right at every operation, regularly hitting
the right arm of the V and cycling back to the left. But the hole’s cycles
are of length k and the block’s of length k+1, so eventually the block
will find itself directly above and to the right of the hole. At that point
the block will fall into the hole, decreasing the potential energy of the
configuration.
We have shown, therefore, that if the number of blocks is 55 or
any other triangular number, and the current configuration is not a
staircase, that its potential energy will eventually go down. There are
only finitely many possible configurations, thus only finitely many
values for this potential. It follows that eventually we must reach the
minimum-potential triangular configuration.

Our theorem is a relatively easy application of potentials to graph

theory. An unfriendly partition of a graph G = ⟨V, E⟩ is a partition of

270
 Unlimited Potentials

the vertex set V into two sets with the property that every vertex in
V has at least as many neighbors in the other part as it has in its own
part.

Theorem. Every finite graph has an unfriendly partition.

Proof. Start with any partition V = X ∪Y , and suppose it is not un-

friendly; then there is some vertex v, say in X, that has more neighbors
in X than in Y . Move v to Y . What happens? Unfortunately it may
not be the case that the number of “friendly” vertices decreases, be-
cause it could be that other vertices are negatively affected by moving
v. But the number of edges that cross the partition increases, because
the only affected edges are those emanating from v and the majority
of those didn’t cross before but now do.
That makes things easy: The number of crossing edges cannot
keep increasing indefinitely, so we must eventually come to an un-
friendly partition. Equivalently: Start with a partition that maximizes
the number of crossing edges, and you’re done! ♡
It’s worth noting that this argument fails for infinite graphs, even
though it seems like when a vertex has infinitely many neighbors, it’s
easier to ensure that there are as many on the opposite side (now con-
sidering infinite cardinalities) as there are at home. In fact the infinite
case is tied up with optional axioms of set theory, and currently it’s
not even known whether every countably infinite graph has an un-
friendly partition.

271
 19. Hammer and Tongs
Often, to solve a puzzle (or prove a theorem), you need to try things,
see where problems arise, then fix them. We call this the “hammer-
and-tongs” approach, and sometimes it works wonders.

Phone Call
A phone call in the continental United States is made from a west
coast state to an east coast state, and it’s the same time of day at both
ends of the call. How is this possible?

Solution: Generally speaking, the west coast of the United States is

on Pacific Time, 3 hours later than Eastern Time. But you may know
that Florida, an east coast state, has a panhandle that extends quite
far west; in fact Pensacola, FL, is on Central Time. So that picks up
one of the three hours we need to make up.
If you check you will find that there is a small part of Oregon that’s
on Mountain Time, including, for example, the town of Ontario. That
picks up our second hour.
To get the third, we time the call so that Daylight Saving has just
clicked off in Pensacola. An instant after 1:59:59 a.m. in Pensacola,
typically on the first Sunday in November, the clock jumps back to 1
a.m. while it’s still just the “first” 1 am in Ontario, OR. So you’ve got
an hour to make that phone call.
As I write this I note that a revolt against Daylight Savings is brew-
ing among the states. So before you attempt to earn drinks at the bar
with this tidbit of a puzzle, you might want to check that it still works!

Crossing the River

In eighth century Europe, it was considered unseemly for a man to
be in the presence of a married woman, even briefly, unless her hus-
band was there as well. This posed problems for three married cou-
ples who wished to cross a river, the only means being a rowboat that

273
 Mathematical Puzzles

could carry at most two people. Can they get to the other side with-
out violating their social norms? If so, what’s the minimum number
of crossings needed?
Solution: This is just a question of trying it out. Call the women
1, 2, and 3, the corresponding husbands A, B, and C. The ideal plan
would require nine crossings, with the boat carrying two people across
and one back with each round trip. If 1 and A cross, it must be A
that returns with the boat (else 1 would find herself on the near shore
with B and C but no husband). Now 2 and 3 must cross with one
woman returning; two men can then join their wives on the far side.
Now, however, a married couple must come back together in the boat,
costing us a couple of crossings and making the nine-crossing ideal
impossible.
Now the remaining men cross the river for good and the women
finish the job. The final plan has eleven crossings, the minimum pos-
sible, and is illustrated below.

274
 Hammer and Tongs

Sprinklers in a Field
Sprinklers in a large field are located at the vertices of a square grid.
Each point of land is supposed to be watered by exactly the three
closest sprinklers. What shape is covered by each sprinkler?
Solution: We first must ask ourselves: What are the three sprinklers
closest to a given point in the field? Suppose the point is in the grid
square bounded by sprinklers at a, b, c, and d, labeled clockwise. Di-
vide the square into four congruent subsquares, and call them A, B,
C, and D respectively, according to whether the corner shared with
the big square is a, b, c, or d. It’s not hard to see that for any point in A
the closest three sprinklers are at a, b, and d, and similarly for points
in the other small squares.
It follows that the sprinkler at a has to reach all the points in A,
B, and D, plus the points in corresponding subsquares of the other
three grid squares incident to a. This amounts to 12 little subsquares
in the shape of a fat Greek cross.

Fair Play
How can you get a 50-50 decision by flipping a bent coin?
Solution: Flip the bent coin twice hoping to get a head and a tail; if
the head comes first, call the result HEADS; if the tail comes first, call
it TAILS. If the result is two heads or two tails, repeat the experiment.
This solution is often attributed to the late, great mathemati-
cian and computing pioneer John von Neumann, and called “von

275
 Mathematical Puzzles

Neumann’s trick.” It relies on the fact that even if the coin is bent, suc-
cessive flips are (or at least should be) independent events. Of course,
it also relies on it being at least possible for the bent coin to land on
either side!
If you want to minimize the number of flips to get your decision,
the above scheme can be improved upon. For example, if you get HH
for the first pair of flips and TT for the second, you can quit and call
the result HEADS (TT followed by HH would then be called TAILS).

Finding the Missing Number

All but one of the numbers from 1 to 100 are read to you, one every ten
seconds, but in no particular order. You have a good mind, but only a
normal memory, and no means of recording information during the
process. How can you ensure that you can determine afterward which
number was not called out?

Solution: Easy—you keep track of the sum of the numbers being

called out, adding each one in turn to your accumulated total. The
sum of all numbers from 1 to 100 is 100 times the average number
(50 12 ), namely 5,050; that minus your final sum will be the missing
number.
There is, by the way, no need to keep the hundreds digit or thou-
sands digit during the process. Addition modulo 100 is good enough.
At the end you subtract the result from 50 or 150 to get an answer in
the correct range.
Dealing with streams of data, when handicapped by limited com-
puting and memory resources, is a major subject of study in modern
computer science. Here’s another streaming puzzle.

Identifying the Majority

A long list of names is read out, some names many times. Your object
is to end up with a name that is guaranteed to be the name which was
a called a majority of the time, if there is such a name.
However, you have only one counter, plus the ability to keep just
one name at a time in your mind. Can you do it?

Solution: It makes sense to use the counter to keep track of how

well the name currently in memory is doing. Then, whenever that
name is heard again, you can increment the counter; when some other

276
 Hammer and Tongs

name is heard, decrement it. If the counter reaches 0, where it began,

you replace whatever name was in memory by the currently-heard
name and then increment.
You could, of course, finish with a name in mind which occurred
only once (e.g., if the list were “Alice, Bob, Alice, Bob, Alice, Bob,
Charlie”). However, if a name occurs more than half the time, it’s
guaranteed to be the one in your memory at the end. The reason is that
when this name is in memory, the counter is more often incremented
than decremented. ♡

Poorly-Placed Dominoes
What’s the smallest number of dominoes one can place on a chess-
board (each covering two adjacent squares) so that no more fit?

Solution: The goal is to place dominoes so that the number h of

“holes” (uncovered squares of the chessboard) is as large as possible,
without having two adjacent holes. The number of dominoes used
would then be d = (64 − h)/2.
It seems natural to arrange dominoes in diagonal lines, as in the
first figure below; worst fit is achieved by switching from horizontal
in half the board to vertical in the other half, then filling in adjacent
holes as necessary. But alternating, gapped horizontal lines work bet-
ter provided you switch to vertical (or vice-versa), as in the second
figure, when the edge of the board causes problems. This gives 20
holes, thus 22 dominoes.

277
 Mathematical Puzzles

Can we do any better? Note that except for holes on the top row,
every hole has a domino directly above it—and no domino can be
directly above two holes. Moreover, dominoes that touch the bottom
row (of which there must be at least 3) aren’t helping.
In the best case, therefore, where there are four holes on the top
row and three dominoes touching the bottom, d ≥ h−1 and therefore
(64 − h)/2 ≥ h − 1, 33 ≥ 3h/2, h ≤ 22.
But wait, if we really did have four holes on the top row and only
three dominoes touching the bottom, then we could up-end the above
argument to show that h ≤ 23. But this is impossible since the number
of holes must be even. Conclusion: we can’t beat our 22-hole solution.
Dominoes appear in many intriguing puzzles.

Unbreakable Domino Cover

A 6 × 5 rectangle can be covered with 2 × 1 dominoes, as in the figure
below, in such a way that no line between dominoes cuts all the way
across the board. Can you cover a 6 × 6 square that way?

278
 Hammer and Tongs

Solution: No. You need two dominoes, not just one, to cross each
interior line—with just one, there’s be an odd number of squares on
each side of that line, but that number is a multiple of 6. Since there
are 10 interior lines, you’ll need 20 crossing dominoes, but there’s
room for only 62 /2 = 18 dominoes on the board. The 6 × 5 rectan-
gle is in fact the only one with both dimensions under 7 that can be
“unbreakably” covered.

Filling a Bucket
Before you are 12 two-gallon buckets and a 1-gallon scoop. At each
turn, you may fill the scoop with water and distribute the water any
way you like among the buckets.
However, each time you do this your opponent will empty two
buckets of her choice.
You win if you can get one of the big buckets to overflow. Can you
force a win? If so, how long will it take you?

Solution: It’s safe to assume your opponent will always empty the
fullest two buckets. To force your opponent to pour out as little water
as possible, it’s natural to begin by keeping all the buckets at the same
level. How far will this get you?
You’ll start by putting 1/12 of a gallon in each bucket; your op-
ponent will empty two buckets, reducing the total amount of water
to 10/12. You can then add a gallon total to bring each bucket up
to (1 + 10/12)/12 = 11/72. Continuing in this manner, you make
progress as long as the amount in each bucket is less than half a gal-
lon (since then your opponent is pouring off less than you’re adding).
You can get close to, but never quite reach, half a gallon per bucket
by this method. Then what?
Then you’re going to need to give up on keeping all the buckets
level. Suppose you build up to x gallons per bucket, then give up on
the two buckets your opponent just emptied, and fill the rest evenly.
Then you can get x + 1/10 gallons in those; she empties two of them
and you build up to x + 1/10 + 1/8 in the remaining eight, and so on.
You end with x+1/10+1/8+1/6+1/4+1/2 = x+1.141666 in the
last two buckets, not good enough to cause overflow, since x < 1/2.
But wait—you don’t need two buckets overflowing, only one. So,
you start by building up only 11 buckets, with the idea of later re-
ducing to 9, 7, 5, 3, and finally, 1 bucket. That will get you up to
x + 1/9 + 1/7 + 1/5 + 1/3 + 1 ∼ x + 1.7873015873. That’s more like
it! So it’s enough to get x ≥ 0.2127.

279
 Mathematical Puzzles

You start with 1/11 of a gallon in each of 11 buckets, abandoning

the twelfth forever. You build that up to (1 + 9 · 1/11)/11 = 20/121 ∼
0.1653 gallons in each of the 11 buckets after the second round, and
(1 + 9 · 20/121)/11 ∼ 0.2261 after the third. That’s enough to go on
to the reduction phase. Altogether it takes you 3 + 5 = 8 rounds, and
with some effort it can be shown that you can’t do any better.

Polygon on the Grid

A convex polygon is drawn on the coordinate plane with all its ver-
tices on integer points, but no side parallel to the x- or y-axis. Let h
be the sum of the lengths of the horizontal line segments at integer
height that intersect the (filled-in) polygon, and v the equivalent for
the vertical line segments. Prove that h = v.

Solution: Both are equal to the area of the polygon. One way to
see that is to let L1 , . . . , Lk be the integer-height horizontal lines that
intersect the interior of the polygon, dividing it into two triangles (at
the top and bottom ends) and k−1 trapezoids. If the length of the
segment of Li that intersects the polygon is ℓi , then the sum of the
areas of the triangles and trapezoids is
1 1 1 1 1
ℓ1 + (ℓ1 + ℓ2 ) + (ℓ2 + ℓ3 ) + · · · + (ℓk−1 + ℓk ) + ℓk ,
2 2 2 2 2
which equals h, and a similar argument holds for v.

One-Bulb Room
Each of n prisoners will be sent alone into a certain room, infinitely
often, but in some arbitrary order determined by their jailer. The pris-
oners have a chance to confer in advance, but once the visits begin,
their only means of communication will be via a light in the room
which they can turn on or off. Help them design a protocol which
will ensure that some prisoner will eventually be able to deduce that
everyone has visited the room.

Solution: It will, of course, be necessary to assume that no one

fools with the room’s light between visits by prisoners; but prisoners
do not need to know the initial state of the light. The idea is that one
prisoner (say, Alice) repeatedly tries to turn the light on, and each of
the others turns it off twice.

280
 Hammer and Tongs

More precisely, Alice always turns on the light if she finds it off,
otherwise she leaves it on. The rest of the prisoners turn it off the first
two times they find it on, but otherwise leave the light alone.
Alice keeps track of how many times she finds the room dark after
her initial visit; after 2n−3 dark revisits she can conclude that everyone
has visited. Why? Every dark revisit signals that one of the other n−1
prisoners has visited. If one of them, say Bob, hasn’t been in the room,
then the light cannot have been turned off more than 2(n−2) = 2n−4
times. On the other hand, Alice must eventually achieve her 2n − 3
dark revisits because eventually the light will have been turned off
2(n−1) = 2n−2 times and only one of these (caused by a prisoner
darkening an initially light room before Alice’s first visit) can fail to
cause a dark revisit by Alice.

Sums and Differences

Given 25 different positive numbers, can you always choose two of
them such that none of the other numbers equals either their sum or
their difference?

Solution: Yes. Let the numbers be x1 < x2 < · · · < x25 , and sup-
pose that every two of them have either their sum or their difference
represented among the other numbers. For i < 25, x25 + xi can’t be
on the list, so x25 − xi must be there; it follows that the first 24 num-
bers are paired with xi + xn−i = x25 . Now consider x24 together with
any of x2 , . . . , x23 ; these pairs sum to more than x25 = x24 + x1 and
so x2 , . . . , x23 must also be paired, and in particular x2 + x23 = x24 .
But we just had x2 + x23 = x25 , a contradiction.

Prisoner and Dog

A woman is imprisoned in a large field surrounded by a circular fence.
Outside the fence is a vicious guard dog that can run four times as fast
as the woman, but is trained to stay near the fence. If the woman can
contrive to get to an unguarded point on the fence, she can quickly
scale the fence and escape. But can she get to a point on the fence
ahead of the dog?

Solution: We may as well take our unit of distance to be the radius

of the field. If the prisoner were constrained to a smaller, concentric
circle of radius r, where r < 14 , she would be able to maneuver her-
self to the farthest available point from the dog (point P in the figure

281
 Mathematical Puzzles

below); this is because the circumference of the small circle would be

less than 14 of the circumference of the field. But if r is close enough
to 14 , the woman can then make a run for it straight to the fence. Her
distance to the fence would only be a hair over 34 of a unit, but the dog
has to go half way around the field, a distance of π units. Since π > 3,
this is more than four times further than the woman has to run.

Love in Kleptopia
Jan and Maria have fallen in love (via the internet) and Jan wishes to
mail her a ring. Unfortunately, they live in the country of Kleptopia
where anything sent through the mail will be stolen unless it is sent
in a padlocked box. Jan and Maria each have plenty of padlocks, but
none to which the other has a key. How can Jan get the ring safely
into Maria’s hands?

Solution: In one solution, Jan sends Maria a box with the ring in
it and one of his padlocks on it. Upon receipt Maria affixes her own
padlock to the box and mails it back with both padlocks on it. When

282
 Hammer and Tongs

Jan gets it he removes his padlock and sends the box back to Maria,
who can now open her own padlock and enjoy the ring. This solu-
tion is not just play; the idea is fundamental in Diffie-Hellman key
exchange, an historic breakthrough in cryptography.
Depending on one’s assumptions, other solutions are possible as
well. A nice one was suggested by several persons at the Gathering
for Gardner VII, including origami artist Robert Lang: It requires that
Jan find a padlock whose key has a large hole, or at least a hole which
can be sufficiently enlarged by drilling, so that the key can be hooked
onto a second padlock’s hasp.
Jan uses this second padlock, with the aforementioned key
hooked on its hasp, to lock a small empty box which he then sends
to Maria. When enough time has passed for it to get there (perhaps
he awaits an email acknowledgment from Maria) he sends the ring in
another box, locked by the first padlock. When Maria gets the ring
box, she picks up the whole first box and uses the key affixed to it to
access her ring.

Badly designed Clock

The hour and minute hands of a certain clock are indistinguishable.
How many moments are there in a day when it is not possible to tell
from this clock what time it is?

Solution: Let us first note that for the problem to make sense, we
must assume that the hands move continuously, and that we are not
tasked with deciding whether a time is a.m. or p.m. (We assume there
is no “second hand.”) Note that we can tell what time it is when the
hour and minute hands coincide, even though we can’t tell which
hand is which; this happens 22 times a day, since the minute hand
goes around 24 times while the hour hand goes around twice, in the
same direction.
This reasoning turns out to be good practice for the proof. Imag-
ine that we add to our clock a third “fast” hand, which starts at 12
midnight and runs exactly 12 times as fast as the minute hand.
Now we claim that whenever the hour hand and the fast hand
coincide, the hour and minute hands are in an ambiguous position.
Why? Because later, when the minute hand has traveled in all 12 times
as far as it had moved since midnight, it will be where the fast hand
(and thus also the hour hand) is now, while the hour hand is where the
minute hand is now. That’s the definition of an ambiguous moment.

283
 Mathematical Puzzles

Conversely, by the same reasoning, all ambiguous positions occur

when the hour hand and fast hand coincide.
So, we need only compute the number of times a day this coinci-
dence of fast hand and hour hand occurs. The fast hand goes around
122 × 2 = 288 times a day, while the hour goes around just twice, so
this happens 286 times.
Of these 286 moments, 22 are times when the hour hand and
minute hand (thus all three hands) are coincident, leaving 264 am-
biguous moments.

Worms and Water

Lori is having trouble with worms crawling into her bed. To stop
them, she places the legs of the bed into pails of water; since the
worms can’t swim, they can’t reach the bed via the floor. But they
instead crawl up the walls and across the ceiling, dropping onto her
bed from above. Yuck!
How can Lori stop the worms from getting to her bed?

Solution: This curious problem is arguably more about engineering

than mathematics.
Lori can indeed keep the worms off her bed by (in addition to
putting the bed’s legs in water buckets) hanging a large canopy from
the ceiling, extending well over the bed. But the canopy must curve
underneath itself at its edges, creating a ring-gutter underneath that
is filled with water. (See the figure below for a cross section of the
canopy.)
The under-slung gutter prevents the worms from dropping onto
the canopy’s edge, crawling under the canopy’s surface to some point
over the bed, and dropping onto Lori.

284
 Hammer and Tongs

If the worms have no high-up way to get in her bedroom, Lori can
more easily accomplish this same task by encircling the room itself
with a water-filled gutter.

Generating the Rationals

You are given a set S of numbers that contains 0 and 1, and contains
the mean of every finite nonempty subset of S. Prove that S contains
all the rational numbers between 0 and 1.
Solution: First note that S contains all the “dyadic” rationals, that
is, rationals of form p/2n ; we can obtain all those with denominator
2n and odd numerator by averaging two adjacent ones with lower-
powered denominators.
Now any general p/q is of course the average of p ones and q −p
zeros. We choose n large and replace the zeros by 1/2n , −1/2n , 2/2n ,
−2/2n , 3/2n , etc. including one 0 if p is odd. Similarly, we replace the
ones by 1 − 1/2n , 1 + 1/2n , 1 − 2/2n , and so forth. Of course, some
of these numbers lie outside the unit interval, but we can rescale the
procedure to fit some dyadic interval containing p/q and lying strictly
between 0 and 1. ♡

Funny Dice
You have a date with your friend Katrina to play a game with three
dice, as follows. She chooses a die, then you choose one of the other
two dice. She rolls her die while you roll yours, and whoever rolls the
higher number wins. If you roll the same number, Katrina wins.
Wait, it’s not as bad as you think; you get to design the dice! Each
will be a regular cube, but you can put any number of pips from 1 to
6 on any face, and the three dice don’t have to be the same.
Can you make these dice in such a way that you have the advan-
tage in your game?

Solution: What you’d like to do is design the dice so that no matter

which of the three Katrina chooses, you can pick one of the remaining
two which will roll a higher number more than half the time. Such a
trio is known as a “set of non-transitive dice,” and there are many
ways to design them.

285
 Mathematical Puzzles

The key is that if die A’s roll has the same average as die B’s, that
doesn’t mean A beats B half of the time. It could be that when A beats
B, it wins by a lot (e.g., 6 versus 1) but when it loses, it does so by only
a little bit (e.g., 5 versus 4). Then it would follow that B would win
more often.
To take an extreme example, put a 6 and a five 3’s on die A (av-
eraging 3.5), and a 1 and five 4’s on die B (same average). Then with
probability 56 × 56 = 25/36 > 1/2, there’ll be a 3 on die A and a 4 on
die B. And you’ve eliminated ties by using different numbers on the
two dice.
Let’s use the remaining numbers, 2 and 5, on die C—three of each,
so that again the average roll is 3.5. It remains only to check that,
indeed, die C is a favorite over die B (winning anytime C rolls a 5,
and also occasionally 2 to 1, altogether with probability 7/12), but die
A is the expected winner over die C. Since we already know B beats
A, your strategy against Katrina is clear.

Sharing a Pizza
Alice and Bob are preparing to share a circular pizza, divided by radial
cuts into some arbitrary number of slices of various sizes. They will
be using the “polite pizza protocol”: Alice picks any slice to start;
thereafter, starting with Bob, they alternate taking slices but always
from one side or the other of the gap. Thus after the first slice, there
are just two choices at each turn until the last slice is taken (by Bob if
the number of slices is even, otherwise by Alice).
Is it possible for the pizza to have been cut in such a way that Bob
has the advantage—in other words, so that with best play, Bob gets
more than half the pizza?

Solution: If the number of slices is even (as with most pizzas), the
puzzle Coins in a Row from Chapter 7 applies and Alice can always
get at least half the pizza. In fact that is so even if Bob gets to start
by choosing one radial cut and insisting that Alice’s choice of first
slice be on one side or the other of that cut. That makes the problem
exactly the same as Coins in a Row; Alice can just number the slices 1,
2, etc. starting clockwise (say) from the cut, and play so as to take all
the even-numbered slices or all the odd, whichever is better for her.
The argument fails if the number of slices is odd. But the odd case
sounds even better for Alice since then she ends up with more slices.
How can we get a handle on the odd case?

286
 Hammer and Tongs

Taking another cue from the Chapter 7, let’s limit the slice sizes
to 0 or 1. (Can a slice be of size zero? Mathematically, no problem;
gastronomically, think of a slice of size ε = one one-hundredth of a
pizza, say.) To try to give the advantage to Bob, we somehow need to
organize these so that no matter where Alice begins, Bob can use his
parity advantage to overcome Alice’s one-slice head start.
That takes a fair number of slices, but it turns out that 21 of these
{0, 1}-sized slices is enough to turn the advantage to Bob. With a
bit more fooling around, you will find that you can combine some
of the slices to get a 15-slice pizza with slice-sizes 0, 1, and 2 of
which nothing can stop Bob from acquiring 5/9. (Two schemes work:
010100102002020 and 010100201002020.) Pictured below is one of
the Bob-friendly 15-slice schemes, with pepperoni distributed to in-
dicate the slice sizes. No matter how she plays, Alice can never get
more than 4 of the 9 pepperoni chunks against a smart, hungry Bob.

It has been shown that this pizza is best possible for Bob. Sum-
mary: If the number of slices is even, or odd but at most 13, Alice can
get at least half the pizza; if the number of slices is odd and at least
15, she can guarantee at least 4/9 but no larger fraction.

287
 Mathematical Puzzles

Names in Boxes
The names of 100 prisoners are placed in 100 wooden boxes, one
name to a box, and the boxes are lined up on a table in a room. One
by one, the prisoners are led into the room; each may look in at most
50 boxes, but must then leave the room exactly as he found it and is
permitted no further communication with the others.
The prisoners have a chance to plot their strategy in advance, and
they are going to need it, because unless every single prisoner finds his
own name all will subsequently be executed.
Find a strategy that gives the prisoners a decent chance of survival.

Solution: To solve it, the prisoners must first agree on a random

labeling of the boxes by their own names. (The point of making it
random is that it makes it impossible for the warden to place names
in boxes in such a way as to foil the protocol described below.) When
admitted to the room, each prisoner inspects his own box (that is,
the box with which his own name has been associated). He finds a
name in that box, probably not his own. He then looks into the box
belonging to the name he just found, and then into the box belonging
to the name he found in the second box, etc., until he either finds his
own name, or has opened 50 boxes.
That’s the strategy; now, why on earth should it work? Well, the
process which assigns to a box’s owner the name found in his box is
a permutation of the 100 names, chosen uniformly at random from
the set of all such permutations. Each prisoner is following a cycle of
the permutation, beginning with his box and (if he doesn’t run over the
50-box limit) ending with his name on a piece of paper. If it happens
that the permutation has no cycles of length greater than 50, this process
will work every time and the prisoners will be spared.
In fact, the probability that a uniformly random permutation of
the numbers from 1 to 2n contains no cycle of length greater than n
always exceeds 1 minus the natural logarithm of 2—about 30.6853%.
To see this, let k > n and count( the
) permutations having a cycle
C of length exactly k. There are 2n k ways to pick the entries in C,
(k −1)! ways to order them, and (2n−k)! ways to permute the rest;
the product of these numbers is (2n)!/k. Since at most one k-cycle
can exist in a given permutation, the probability that there is one is
exactly 1/k.
It follows that the probability that there is no long cycle is
1 1 1
1− − − ··· − = 1 − H2n + Hn ,
n+1 n+2 2n

288
 Hammer and Tongs

where Hm is the sum of the reciprocals of the first m positive integers,

approximately ln m. Thus our probability is about 1 − ln 2n + ln n =
1 − ln 2, and in fact is always a bit larger. For n = 50 we get that the
prisoners survive with probability 31.1827821%.
Let’s put these new skills to work on a variation.

Life-Saving Transposition
There are just two prisoners this time, Alice and Bob. Alice will be
shown a deck of 52 cards spread out in some order, face-up on a table.
She will be asked to transpose two cards of her choice. Alice is then
dismissed, with no further chance to communicate to Bob. Next, the
cards are turned down and Bob is brought into the room. The warden
names a card and to stave off execution for both prisoners, Bob must
find the card after turning over, sequentially, at most 26 of the cards.
As usual the prisoners have an opportunity to conspire before-
hand. This time, they can guarantee success. How?

Solution: This is just a variation on the names-in-boxes puzzle. We

may assume the cards are numbered 1 to 52. Bob’s algorithm, when
asked to search for card i, will be to flip the ith card in the display. If
it’s value is j, he then flips the jth card; if the value of the jth card is k,
he flips the kth card, etc. Thus, he follows a cycle in the permutation
that maps the nth card to its value, and if no cycle in this permutation
is of length > 26, he will succeed.
Alice can guarantee this condition. If she sees a cycle of length
> 26 (of course, there cannot be more than one of these) she bisects
it by transposing two antipodal or nearly antipodal cards of the cycle.
Otherwise she makes a harmless move, for example, bisects a short
cycle or if the ordering happens to be the identity, does any transpo-
sition.
We conclude this chapter’s puzzles with one that is mind-
numbingly easy if you try the right thing.

Self-Referential Number
The first digit of a certain 8-digit integer N is the number of zeroes in
the (ordinary, decimal) representation of N . The second digit is the
number of ones; the third, the number of twos; the fourth, the number
of threes; the fifth, the number of fours; the sixth, the number of fives;

289
 Mathematical Puzzles

the seventh, the number of sixes; and, finally, the eighth is the total
number of distinct digits that appear in N . What is N ?

Solution: It’s pretty tough to work out N by reasoning. But there’s

an easy way!
Take any 8-digit number N0 , and transform it according to the
puzzle conditions: that is, make a new number N1 by letting the first
digit of N1 be the number of 0’s in N0 , etc.
If by chance N1 = N0 , you’ve solved the puzzle. But you’d have
to be astonishingly lucky for that to happen.
Not to worry. Repeat the procedure with N1 to get N2 , then N3 ,
until you reach a time t when Nt = Nt−1 . Then you’re done.
Let’s try this. For fun we’ll start with 31415926, the first 8 digits of
the decimal expansion of π, but you should try it yourself with your
favorite 8-digit number.
We obtain:

It worked! But why should that happen in a reasonable number of

steps? Indeed, why should it happen at all? Couldn’t we start to cycle
at some point, with Nt duplicating some previous Nt−k with k > 1?

290
 Hammer and Tongs

Sadly, I don’t know. This method does not work with all problems
of this type. But it is surprisingly useful—a great weapon to have in
your armory.

We conclude this chapter with a marvelous theorem that’s well

known to geometers and to folks who do recreational math, but de-
serves to be better known among serious mathematicians in other ar-
eas. Described by Georg Alexander Pick in 1899, it tells you how to
compute the area of a lattice polygon just by counting lattice points.
A lattice polygon is a closed figure made from line segments that
terminate at vertices of a plane grid, like the one shown in the figure
below.

Theorem. The area of a lattice polygon is exactly I + B/2 − 1 times the

area of a single square cell of the grid, where I is the number of lattice points
strictly inside the polygon, and B the number on its boundary.
Proof. We will assume that each cell of the grid has unit area, so
that the theorem simply says the area of a lattice polygon P is I +
B/2 − 1.
The formula carries some reasonable intuition: Inside points
(filled circles in the figure) are fully counted, and boundary points
(empty circles) are arguably half-in and half-out, so should be counted
at half value—except that corners are really somewhat less than half-
in, so we subtract one to account for them.

291
 Mathematical Puzzles

The formula certainly works when the polygon is a single grid cell,
since there I = 0 and B = 4.
Unfortunately we can’t generally build lattice polygons out of grid
cells, but it’s worth thinking about how we could build them out of
smaller units. Suppose P is the union of two lattice polygons R and
S that share one side from each, as in the next figure.

If the formula is correct for R and S, does it follow that it’s correct
for P ? Indeed, that is the case: If there are k lattice points on the com-
mon edge, not counting its endpoints, then we have IP = IR + IS + k
and BP = BR + BS − k − 2, so that the area of P is IR + IS + k +
(BR + BS − k − 2)/2 − 1 = (IR + BR /2 − 1) + (IS + BS /2 − 1)
which is indeed the area of R plus the area of S. So far so good.
Note that this calculation also shows that if the formula is cor-
rect for any two of the three polygons P , R, and S, it works for the
third—in other words, the formula works with subtraction as well as
addition.
That’s great—and because we can cut any lattice polygon into lat-
tice triangles, as in the next figure, it’s enough to prove the theorem for
triangles. This is something of a hammer-and-tongs type operation.

292
 Hammer and Tongs

First we note that right triangles are easy, if their sides are vertical
and horizontal line segments. The reason is that we can copy such a
triangle, rotate the copy 180◦ , then match its hypotenuse to the orig-
inal to get an “aligned” lattice rectangle (i.e., one with sides parallel
to the axes).
The formula holds for aligned rectangles, because they’re made
out of unit cells. Since our two triangles have the same values for I and
B, Pick’s formula assigns them the same area which must therefore
be half the area of the aligned rectangle, as it should be.
Finally, we only need observe that any triangle can be made into
an axis-aligned rectangle by adding (at most) three right triangles of
the above sort, as in the final figure. ♡

293
 20. Let's Get Physical
Rotating Coin
While you hold a US 25-cent piece firmly to the tabletop with your
left thumb, you rotate a second quarter with your right forefinger all
the way around the first quarter. Since quarters are ridged, they will
interlock like gears and the second will rotate as it moves around the
first.

How many times will it rotate?

Solution: A classic. It rotates twice, relative to the table: once rela-

tive to the stationary quarter, and once more owing to its revolution
around the stationary quarter. Try it!

Pie in the Sky

What fraction of the sky is occupied by a full moon?

Solution: The moon is about half a degree in diameter (actually,

between 29.43 arc minutes at apogee and 33.5 minutes at perigee),
hence its area in “square degrees” is πr2 ∼ π(1/4)2 = π/16.
How many square degrees is the sky? Its circumference is 360 so
its radius in degrees (!) is 180/π and the sky’s area is half the area of
the celestial sphere, about (1/2)4πr2 = 2(180)2 /π square degrees.
Thus the fraction of sky covered by the moon is about (π/16)/
(2(180)2 /π) = π 2 /(32 · 1802 ) ∼ 1/105,050.

295
 Mathematical Puzzles

If we repeat this calculation with the more accurate estimate of 31

minutes for the diameter of the moon, we get 1/101,661 of the sky—
so it is remarkably accurate to say the moon occupies ten millionths
of the sky. The sun, by the way, is about the same size.

Returning Pool Shot

A ball is shot from a corner of a polygonal pool table (not necessarily
convex) with right-angle corners; let’s say all sides are aligned either
exactly east-west or exactly north-south.
The starting corner is a convex corner, that is, has an interior angle
of 90◦ . There are pockets at all corners, so that if the ball hits a corner
exactly it falls in. Otherwise, the ball bounces true and without energy
loss.
Can the ball ever return to the corner from which it began?

Solution: Suppose it is possible, and imagine such a shot. We may

assume the ball starts off going NE; then when it finally returns it is
going SW. At each bounce one of the two direction letters changes;
for example, the second lap will be either SE (if it bounces off an E-W
cushion) or NW (if it bounces off a N-S cushion). It follows that it
takes an even number of bounces, thus an odd number of laps.

All NE and SW laps are parallel (likewise all NW and SE

paths), thus the initial and final laps follow the same line in opposite

296
 Let's Get Physical

directions. Since reflection works as well backwards as forwards, the

whole path out from the original corner and back is a palindrome—
the sequence of bounce points reads the same backwards and for-
wards. It follows that the two middle bounce points are the same
point, which is ridiculous. ♡

Falling Ants
Twenty-four ants are placed randomly on a meter-long rod; each ant
is facing east or west with equal probability. At a signal, they proceed
to march forward (that is, in whatever direction they are facing) at 1
cm/sec; whenever two ants collide, they reverse directions. How long
does it take before you can be certain that all the ants are off the rod?

Solution: The key to this puzzle is that, ants being interchangeable

as far as we’re concerned, it would make no difference to the process if
they passed one another “like ships in the night” instead of bouncing.
Then it’s clear that each ant is simply walking straight ahead and must
fall off within 100 seconds.

You can’t afford to wait any shorter amount of time because if

there’s an ant starting at one end of the rod and facing the other end,
some ant won’t exit the far end until 100 seconds later.

Ants on the Circle

Twenty-four ants are randomly placed on a circular track of length 1
meter; each ant faces randomly clockwise or counterclockwise. At a
signal, the ants begin marching at 1 cm/sec; when two ants collide
they both reverse directions. What is the probability that after 100
seconds, every ant finds itself exactly where it began?

297
 Mathematical Puzzles

Solution: This time we have to be a bit more careful about the ants’
anonymity; the argument that we can replace bouncing by passing
tells us only that the ants’ set of locations will be exactly the same after
100 seconds, but any particular ant might end up in some other ant’s
starting spot.
In fact, since the ants cannot pass one another, their final locations
will be some rotation of their initial locations. Putting it another way,
the whole collection will rotate by some number of ants, and we are in
effect being asked to determine the probability that that number will
be a multiple of 24.
In fact it could be 24 (clockwise or counterclockwise) only if the
ants are all facing the same way, thus each walks once around the track
without any collisions. There are 224 ways to choose how the ants
face of which only two have this property, so the probability of one of
these outcomes is a minuscule 1/223 .
Much more likely is that the net rotation will be zero. When does
that happen? Well, conservation of angular momentum tells us that the
rate of rotation of the ant collection as a whole is constant. Thus the
net rotation will be zero if and only if the initial rate of rotation is zero,
meaning that exactly the same number ants start off facing (counter- ) 24
clockwise as clockwise. The probability of that happening is 24 12 /2
which is about 16.1180258%. Adding 1/223 to that boosts the final
answer to about 16.1180377%.

Sphere and Quadrilateral

A quadrilateral in space has all of its edges tangent to a sphere. Prove
that the four points of tangency lie on a plane.
Solution: The idea is to weight the vertices of the quadrilateral so
that the point where each edge touches the sphere is exactly the center
of gravity of the edge. We can do this because each vertex is equidis-
tant from the tangency points of the two edges emanating from it; if
we give that vertex weight equal to the reciprocal of that distance, we
get the desired weighting.
So what? Well, if we draw a line between opposite points of tan-
gency, the center of gravity of the quadrilateral will have to be on that
line. There are two such lines, though, so the center of gravity is on
both; thus, the lines intersect. Thus the quadrilateral formed by the
points of tangency lies on a plane! ♡

298
 Let's Get Physical

Two Balls and a Wall

On a line are two identical-looking balls and a vertical wall. The balls
are perfectly elastic and friction-free; the wall is perfectly rigid; the
ground is perfectly level. If the balls are the same mass and the one
farther from the wall is rolled toward the closer one, it will knock the
close ball toward the wall; that ball will bounce back and hit the first
ball, which will then roll (forever) away from the wall. Three bounces
altogether.
Now assume the farther ball has mass a million times greater than
the closer one. How many bounces now? (You may ignore the effects
of angular momentum, relativity, and gravitational attraction.)
Solution: You can expect quite a lot of bouncing now, because the
heavy ball will hardly notice that it has been hit the first time, and
proceed undaunted toward the wall. But the light ball will bounce off
the wall and hit the heavy one many more times, until the heavy ball
slows down, reverses direction, and eventually disappears.
In fact the number of collisions will be 3141, the first four digits
of the decimal expansion of π.
Coincidence? No indeed. Make the heavy ball a googol (that is,
10100 ) times heavier than the light one, and the number of bounces
will be the first 51 digits of π, namely 314,159,265,358,979,323,846,
264,338,327,950,288,419,716,939,937,510. It doesn’t matter how far
apart the three objects are, or how hard the heavy ball is pushed.
The expectation is that for any integer k ≥ 0, the number of
bounces when the heavy ball is 102k times heavier than the light one
will be the first k+1 digits of π. Almost as remarkable as the statement
itself is the fact that it is not a theorem, that is, no one can prove it.
But following is an explanation of why we think it’s true.
The key is to make use of conservation of energy and momentum.
Let v be the velocity of the heavy ball and y be the velocity of the light
ball. The mass of the heavy ball is m = 102k , and√conservation of
energy says mv 2 + y 2 is a constant. Letting x = m · v = 10k v
and starting at x = −1, y = 0 puts the system on the unit circle
x2 + y 2 = 1.
Total momentum is mv + y = 10k x + y = −10k , so y =
−10k x − 10k , putting us initially on a line of slope s = −10k . So
when the light ball first hits the wall, the “system point”—that is, the
point on the XY-plane that describes the state of the system—moves
from (−1,0) south-east along the momentum line to the second point
where that line intersects the circle. But upon hitting the wall the light

299
 Mathematical Puzzles

ball’s momentum flips, reflecting the new point across the X-axis back
to the positive side.
In this way the system point zigzags down and up, moving slightly
eastward on the down-zigs, until finally it reaches a point just a bit
northwest of (1,0) from which moving southeast along a line of slope
s would either miss the circle entirely or hit it above the X-axis. That
indicates there will be no more collisions.
The figure below shows our circle, with four system states corre-
sponding to the first two and last two system points.

The total number of collisions is the number of zigzags; how

many were there? Well, the momentum line’s angle to the vertical
is tan(10−k ) and thus the arc of the unit circle between successive
points above the X-axis is 2 tan(10−k ). It follows that the number of
collisions is the greatest integer in 2π/(2 tan(10−k ) = π/ tan(10−k ).
Since 10−k is a small angle, the approximation tan(10−k ) ∼ 10−k
is a very good one (the difference is about the square of the angle, that
is, of order only 10−2k ). Thus when k is large we expect the number
of collisions to be the greatest integer in π/10−k = 10k π, but that is
precisely the first k+1 digits of π!

300
 Let's Get Physical

We’re a bit lucky that this already works for k = 1 (if we change
from base 10 to some other base, it sometimes fails for k = 1). But as
k goes up the approximation above gets better so rapidly that there’s
only a wisp of a chance that it every fails.
What would it take to fail? Roughly speaking, it would miss by 1 if
for some k, digits k+1 through 2k of the decimal expansion of π were
all 9’s. If, as most mathematicians believe, the digits of π behave like
a random sequence, this is ludicrously unlikely. (In 2013 A.J. Yee and
S. Kondo computed the first 12.1 trillion digits of π, with no more
than a dozen 9’s in a row anywhere, about what you’d expect.)
But no one has been able to prove that the digits of π continue to
behave randomly, and even if they did the above glitch could occur.
So it’s not a theorem. But you could bet your life on it.

We conclude with something that really is a theorem.

Let’s think of a polyhedron as any solid in three-dimensional space
with flat faces. It doesn’t have to be convex and could even have holes
in it.
To each face F , we associate a vector vF perpendicular to the face,
pointing outward, and with length proportional to the area of the face.

Theorem. For any polyhedron, the sum over all faces F of the vectors vF is
zero.
Proof. Pump the polyhedron full of air! The pressure on each face
F will be proportional to the area of F , and is exerted outwardly

301
 Mathematical Puzzles

perpendicular to F ; in other words, it is represented by vF . If the sum

of these forces were not zero, the polyhedron would move of its own
accord.
And we can’t have that, can we? ♡

302
 21. Back from the Future
Nothing says you have to try to solve a puzzle by starting with the
premise, then working straight toward the answer. Often starting from
the end—“retrograde analysis” is the fancy term for it—will make
things much easier. (Are you among those that think solving a maze
any way other than by drawing a line from “start” to “finish” is tanta-
mount to cheating? That’s fine if that makes it more fun for you, but
I advise against applying this strategy to every puzzle.)
Consider the following classic.

Portrait
A visitor points to a portrait on the wall and asks who it is. “Brothers
and sisters have I none,” says the host, “but that man’s father is my
father’s son.” Who is pictured?

Solution: Where to start? At the end. Who is “my father’s son”?

Could be the host, if male. Wait, the host has no brothers and sisters,
so it is the host.
Substituting, we get “that man’s father is me,” so the face in the
picture is the host’s son. Simple!
Another classic, but maybe with a twist to the solution that you
haven’t seen:

Three-way Duel
Alice, Bob, and Carol arrange a three-way duel. Alice is a poor shot,
hitting her target only 13 of the time on average. Bob is better, hitting
his target 23 of the time. Carol is a sure shot.
They take turns shooting, first Alice, then Bob, then Carol, then
back to Alice, and so on until only one is left. What is Alice’s best
course of action?

303
 Mathematical Puzzles

Solution: Let’s think about the situation after Alice takes her turn.
If only Carol survives Alice’s turn, Alice is doomed. If only Bob sur-
vives, Alice has probability p of survival where p = 13 ( 13 + 23 p), giving
p = 1/7, not very good. If both Bob and Carol survive, things are
much better because they will aim at each other, with only one sur-
vivor, at whom Alice gets to shoot first. Thus, in that situation, her
survival probability is more than 13 .
It follows that Alice doesn’t want to kill anyone, and her best
course of action is to shoot to miss!
But wait. We’e been assuming the others wouldn’t do that, but
now that we have allowed Alice the option of abstaining, we must
surely allow it to the others. They can work out that whenever three
duelers are still alive, Alice will never aim to kill. Again applying ret-
rograde analysis, if it gets to Carol’s turn with no one dead, should she
kill Bob? If Bob tried to shoot her, then yes. But if Bob shot into the
air, thereby suggesting a willingness to do so indefinitely, then Carol
should do the same—that way, no one’s life is at risk and when the
ammunition runs out, everyone can go home and do math puzzles.
Going back one turn, we deduce that Bob should indeed shoot into
the air and the whole duel will be a dud.
Thinking back on it, it’s pretty obvious that if the highest prior-
ity for all three parties is to stay alive—which is what we’ve been
assuming—then the duel should never have been arranged.
Here’s an application of retrograde analysis to experimental de-
sign.

Testing Ostrich Eggs

In preparation for an ad campaign, the Flightless Ostrich Farm needs
to test its eggs for durability. The world standard for egg-hardness calls
for rating an egg according to the highest floor of the Empire State
Building from which the egg can be dropped without breaking.
Flightless’s official tester, Oskar, realizes that if he takes only one
egg along on his trip to New York, he’ll need to drop it from (poten-
tially) every one of the building’s 102 floors, starting with the first, to
determine its rating.
How many drops does he need in the worst case, if he takes two
eggs?
Solution: Let’s look ahead at the position Oskar will be in when
he’s down to one egg. At that point there will be some minimum

304
 Back from the Future

possible egg rating m (namely, the highest floor from which the first
egg survived) and some maximum possible egg rating M (one less
than the floor from which the first egg went splat). That adds up to
M −m drops.
The plan for dropping the first egg will be some increasing se-
quence of floors, say f1 , f2 , . . . , fk . The spaces between the fi ’s will
be decreasing, since (assuming you want to prevent having to make
more than some fixed number d of drops) the more drops made with
the first egg, the fewer remain for the second.
For example, suppose Oskar’s first drop of the first egg is from the
10th floor. If it breaks, he’s got 9 floors to try with the 2nd egg, for a
total of 10 drops. But then to prevent having to make 11 drops, he’d
have to make his second drop of egg number one from the 19th floor,
allowing for 8 more drops of the 2nd egg, the 3rd from the 27th floor,
4th from the 34th, 5th from the 40th, 6th from the 45th, 7th from the
49th, 8th from the 52nd, 9th from the 54th, and 10th from the 55th.
That’s his ten drops, so he’d be OK only if the building had at most
55 floors.
How high must Oskar start, to make it to floor 102? Easy way to
figure this out: Start at the other end! Add 1 + 2 + 3 + · · · until you
reach or exceed 102, and your last addend, which proves to be 14, is
Oskar’s starting floor—and also his guaranteed maximum number of
drops. Later drops of the first egg, until it breaks, are at floors 27, 39,
50, 60, 69, 77, 84, 90, 95, 99, and 102.

305
 Mathematical Puzzles

You can get 14 another way by deriving or remembering the for-

mula 1+2+3+· · ·+k = k(k+1)/2. (To derive this, note that there are
k addends and the average size of an addend is (k+1)/2.) Then solve
the equation k(k+1)/2 = 102 and round k up to the next integer.
If there are more than two eggs, you can extend this retrograde
analysis in a similar way. For example, for three eggs, you compute the
numbers k(k+1)/2 (directly or by adding) and then sum these to reach
102: 1 + 3 + 6 + 10 + 15 + 21 + 28 + 36 = 120 where 35 = 8(8+1)/2.
So Oskar drops the first egg at floor 36, then 64, then 85. If it breaks
at (say) 64, he’s then playing the two-egg game in a 28-story building
(floors 36 through 63 of the Empire State Building) which requires a
maximum of 7 more drops. The total maximum number of drops for
three eggs is nine.
How about when the number of eggs is unlimited? If you’re a com-
puter scientist, then you may already have guessed that binary search is
the answer: The next drop is made at the midpoint of the current range
of ratings. For example, if the building originally had 63 floors (so
that the rating could be anything from 0 to 63), the first drop should
be from floor 32. If the egg breaks the new range is 0 to 31, else 32 to
63. Since 102 is between 64 = 26 and 128 = 27 , Oskar might need as
many as 7 drops.
There’s a whole class of puzzles where retrograde analysis is the
first thing you should try: namely, game puzzles. Suppose you are
presented with a two-person, alternating move, deterministic, full-
information game. That’s a lot of adjectives but they describe an enor-
mous number of popular games, from chess, checkers, go, hex, and
reversi to nim and tic-tac-toe.
We’ll look at some games of that sort in which no draw is possible.
In such a game, a “P” position is one that, if you can create it with
your move, enables you to force a win. (Think of “P” as standing for
“Previous player wins.”) The other positions are “N” positions (“Next
player wins”) and those are the ones you would like passed to you by
your opponent.
It follows that from an N position there is always a move that
creates a P position, and from a P position, all moves lead to an N
position. This completely defines the N and P positions and tells you
how to play optimally. But these rules operate backwards in time; to
classify positions this way you must start from the end of the game
and work backward.
Let’s try this on the next puzzle.

306
 Back from the Future

Pancake Stacks
At the table are two hungry students, Andrea and Bruce, and two
stacks of pancakes, of height m and n. Each student, in turn, must eat
from the larger stack a non-zero multiple of the number of pancakes
in the smaller stack. Of course, the bottom pancake of each stack is
soggy, so the player who first finishes a stack is the loser.
For which pairs (m, n) does Andrea (who plays first) have a win-
ning strategy?
How about if the game’s objective is reversed, so that the first
player to finish a stack is the winner?
Solution: Let m and n be the two stack sizes. You (if you are play-
ing) are immediately stuck just when m = n; this is therefore a P
position. If m > n and m is a multiple of n, you are in a winning
N position because you can then eat m−n pancakes from the n-stack
and reduce your opponent to the above P position. In particular, if
the short stack has just one pancake, you are in great shape.
What if m is close to a multiple of n? Suppose, for example, that
m = 9 and m = 5. Then you are forced to reduce to m = 4, n = 5
but your opponent must now give you a 1-stack. If m = 11 and n = 5
you have a choice but reducing to m = 6, n = 5 wins for you.
It’s beginning to look like you want to make the ratio of the stacks
small for your opponent, forcing her to make the ratio big for you.
Let’s see. Suppose the current ratio r = m/n is strictly between 1 and
1
2; then the next move is forced and the new ratio is 1−r . These ratios
√
are equal only for r = ϕ = (1+ 5)/2 ∼ 1.618, the golden mean; since
1
ϕ is irrational, one of the two ratios r and 1−r must exceed ϕ while the
other is smaller than ϕ. Aha! Thus, when you present your opponent
with r < ϕ she must make it bigger, then you make it smaller, etc.,
until she’s stuck with r = 1 and must lose!
We conclude that Andrea wins exactly when the initial ratio of
larger to smaller stack exceeds ϕ. In other words, the P positions of
the game are exactly those with m/n < ϕ, where m ≥ n. To see this,
suppose m > ϕn, but m is not a multiple of n. Write m = an + b,
where 0 < b < n. Then either n/b < ϕ, in which case Alice eats an,
or n/b > ϕ, in which case she eats only (a−1)n. This leaves Bruce
with a ratio below ϕ, and faced with a forced move which restores a
ratio greater than ϕ.
Eventually, Andrea will reach a point where her ratio m/n is an
integer, at which point she can reduce to two equal stacks and stick
Bruce with a soggy pancake. But note that she can also, if desired, grab

307
 Mathematical Puzzles

a whole pile for herself, thus winning the alternative game where the
eater of a soggy pancake is the victor.
Of course, if Andrea is instead faced with a ratio m/n which is
strictly between 1 and ϕ, she is behind the eight-ball and it is Bruce
who can force the rest of the play.
We conclude that no matter which form of Pancakes is played, if
the stacks are at heights m > n, Andrea wins precisely when m/n >
ϕ. Only in the trivial case when the stacks are initially of equal height
does it matter what the game’s objective is!
Try the above approach on this similar game.

Chinese Nim
On the table are two piles of beans. Alex must either take some beans
from one pile or the same number of beans from each pile; then Beth
has the same options. They continue alternating until one wins the
game by taking the last bean.
What’s the correct strategy for this game? For example, if Alex is
faced with piles of size 12,000 and 20,000, what should he do? How
about 12,000 and 19,000?

Solution: As in classical Nim and many other games, it’s easiest to

try to characterize the P positions, as there are fewer of those. Once
you know the P positions, the correct strategy is automatic. If a player
is in an N position, he or she makes a move that puts his or her op-
ponent in a P position; that opponent must then move to another N
position.
In Chinese Nim, the empty position (no beans left) is a P posi-
tion since the previous player has just won the game. Any position
with one pile empty, or both piles of the same size, is an N position,
since the empty position is reachable in one move. It’s not hard to de-
duce that the simplest non-empty P position is {1, 2}. After that, you
can work out that {3, 5}, {4, 7}, and {6, 10} are P positions as well.
What’s the pattern?
Let {x1 , y1 }, {x2 , y2 }, . . . be the non-empty P positions, with xi <
yi and xi < xj for i < j. Notice you cannot have xi = xj for i = ̸
j because then a player could reduce the larger of yi and yj to the
smaller, leaving another P position.
Some thought will lead you to conclude that given {x1 , y1 }
up to {xn−1 , yn−1 }, xn is the least positive number not among
{x1 , . . . , xn−1 } ∪ {y1 , . . . , yn−1 }, and yn = xn + n. Notice that this

308
 Back from the Future

forces yn to be a higher number than any in the set {x1 , . . . , xn−1 } ∪

{y1 , . . . , yn−1 }.
The proof is by induction on n. You have already seen that xn
can’t be among the numbers in {x1 , . . . , xn−1 } ∪ {y1 , . . . , yn−1 } and
also that there can’t be more than one yn to go with this xn , so all you
need to do is show that this {xn , yn } really is a P position.
If {xn , yn } were an N position (for Alex, say), it must be that he
can reduce it to {xi , yi } for some i < n; but he cannot get to this
position by reducing the smaller pile or by reducing both piles by the
same amount, because that would leave the difference of the two piles
at n or more. Nor can he get there by reducing the larger pile, because
then he would get another y for the same x. Thus {xn , yn } is indeed
a P position.
You now have the means to generate as long a list as you like of
P positions. From this Alex’s strategy is easy to work out. If he is
faced with {xi , yi } he removes a bean or two and hopes for an error.
If he sees {xi , z} for z > yi , he reduces z to yi . If he sees {xi , z} with
xi < z < yi , the difference d = z − xi is less than i; he takes from
both piles to get down to {xd , yd } (if z = yj for some j < i he also
has the option of just reducing xi to xj ). If he sees {yi , z} with yi ≤ z
he can reduce z all the way down to xi , and may have other options
as well.
But it might take a while to generate enough P positions to decide
what to do with thousands of beans in each pile. Is there a more direct
way to characterize the P positions?
Well, you know that for each n, xn is somewhere between n and
2n, because it is preceded by all the xi ’s for i < n and some of the
yi ’s. It is reasonable to guess that xn is approximately equal to rn,
for some ratio r between 1 and 2. If so, yn would be approximately
rn + n = (r+1)n.
If this holds up, it follows that the n xi ’s between 1 and xn are
more or less evenly distributed, and therefore a fraction r/(r +1) of
them will have their corresponding yi below xn . Thus there are about
nr/(r + 1) yi ’s below xn , together with the n xi ’s, adding up to xn
numbers in all; making an equation out of this gives
r
n+n = nr,
r+1
√
which gives us r +1 = r2 , r = (1+ 5)/2, the familiar “golden ratio”
(again!).
If we were really lucky, it would turn out that for each n, xn is
exactly ⌊rn⌋ (the greatest integer in rn), and yn = ⌊r2 n⌋. In fact, this

309
 Mathematical Puzzles

is the case, and you can prove it using the theorem at the end of this
chapter. The key is that because r and r2 are irrational numbers that
sum to 1, every positive integer can be uniquely represented either as
⌊rj⌋ for some integer j, or ⌊r2 k⌋ for some integer k.
Let’s use this to find Alex’s move in the example positions. Note
that 12000/r is a fraction under 7417, and 7417r = 12000.9581 . . . so
12000 is an xi , namely x7417 . The corresponding y7417 is ⌊7417r⌋ =
19417 so if the other pile has 20000 beans, Alex can win by taking
20000 − 19417 = 583 beans away from it. If there are only 19000
beans in the other pile, Alex can win instead by reducing the piles
simultaneously to {x7000 , y7000 } = {11326, 18326}. Since 1900 hap-
pens to be a yj , namely y2674 , Alex can also win by reducing the x-pile
to x2674 = ⌊2674r⌋ = 4326.

Turning the Die

In the game of Turn-Die, a die is rolled and the number that appears
is recorded. The first player then turns the die 90 degrees (giving her
four options) and the new value is added to the old one. The second
player does the same and the two players alternate until a sum of 21
or higher is reached. If 21 was hit exactly, the player who reached it
wins; otherwise, the player who exceeded it loses.
Do you want to be first or second?

Solution: This game, which I call Turn-Die, yields nicely to retro-

grade analysis using P and N positions—but first you have to figure
out what, exactly, “positions” are.
In Turn-Die, to know your position you must certainly know the
current sum; but you also need to know what numbers are available
when you turn the die. A standard die has numbers that sum to 7 on
opposite faces; that is, 1 opposite 6, 2 opposite 5, and 3 opposite 4.
Thus, for example, if the die is now showing 1 or 6, the options are to
turn it to 2, 3, 4, or 5 only.
Thus, what you need to know for strategic purposes is what the
current sum is, and whether the die is showing 1 or 6 (case A), 2 or 5
(case B), or 3 or 4 (case C).
To do the retrograde analysis, it helps to make a picture of the posi-
tions. Write the numbers from 1 to 21 three times in parallel columns
labeled A, B, and C; let’s say with the 1’s on the bottom, and the 21’s
on top. You then have a picture of 63 positions (not all reachable),
and can assign P’s and N’s to them from the top down.

310
 Back from the Future

Let’s see how this process would start (in other words, how the
game ends). The 21’s are all P positions, since the previous player has
now won the game. Position 20A is also a P position, since the 1 is
unavailable to turn to; the player facing 20A will have to overshoot
and lose. The other 20’s, 20B and 20C, are N positions.
Continuing in this way, you eventually determine the state of the
six possible opening positions, namely 1A, 2B, 3C, 4C, 5B, and 6A
(boxed in the figure below). It turns out that of these only 3C and 4C
are P positions, so if you go first, you will with probability 23 start with
an N position and be able to force a win.

311
 Mathematical Puzzles

It’s perhaps worth noting that in some vague sense, you usually
want to be the first player in a combinatorial game that begins in a
random position. The reason is that the condition for being an N po-
sition (that some move leads to a P) is milder than the condition for
being a P position (that every move leads to an N position). This is
why you usually find more N positions than P. When the game starts
in an N position, you want to be first to play.

Game of Desperation
On a piece of paper is a row of n empty boxes. Tristan and Isolde take
turns, each writing an “S” or an “O” into a previously blank box. The
winner is the one who completes an “SOS” in consecutive boxes. For
which n does the second player (Isolde) have a winning strategy?

Solution: The game seems confusing and impenetrable until you

realize that the only way you can force a win on your next move is to
oblige your opponent to play into the configuration S-blank-blank-S
(henceforth to be called a “pit”). Thus, for example, Tristan can win
when n = 7 by placing an S in the middle, then another at the end
farther from Isolde’s response, to make a pit. After a move by each
player outside the pit, Isolde must play into the pit and lose.
The same applies for any odd n greater than 7, as Tristan can play
an S anywhere at least 4 spaces from the ends, then form a pit on one
side or the other and wait.
When n is even Tristan has no chance, as there will never be a
time when Isolde has only pits to play in; when she moves, there is
always an odd number of blank squares to play in. Instead, when n
is even and large, Isolde wins by playing an S far from the ends and
from Tristan’s first move. However, if Tristan begins with an O, Isolde
cannot put an S next to it, so she needs extra room.
In the n = 14 case, if Tristan writes an O in place 7 (of 1 through
14), Isolde’s best response is an S in position 11 (threatening to make
a pit with an S in 14). Tristan can counter this with an O in 13 or 14
(or an S in 12), and now Isolde would like to make a pit with S in
position 8 but can’t, as Tristan would then win with S in 6.
Thus n = 14 is a draw; Isolde needs n to be even and at least 16.
To wrap up, Tristan wins when n is odd and at least 7, Isolde when n
is even and at least 16; all other values of n lead to a draw with best
play.

312
 Back from the Future

Deterministic Poker
Unhappy with the vagaries of chance, Alice and Bob elect to play
a completely deterministic version of draw poker. A deck of cards
is spread out face-up on the table. Alice draws five cards, then Bob
draws five cards. Alice discards any number of her cards (the dis-
carded cards will remain out of play) and replaces them with a like
number of others; then Bob does the same. All actions are taken with
the cards face-up in view of the opponent. The player with the better
hand wins; since Alice goes first, Bob is declared to be the winner if
the final hands are equally strong. Who wins with best play?
Solution: You need to know a little about the ranking of poker
hands for this puzzle: Namely, that the best type of hand is the straight
flush (five cards in a row of the same suit), and that an ace-high
straight flush (known as a “royal flush”) beats a king-high straight
flush and on down.
That means if Bob is allowed to draw a royal flush, Alice’s goose
is cooked. For Alice to have a chance, her initial hand must contain a
card from each of the four possible royal flushes.
The best card of each suit, for that purpose, is the 10, since it stops
all straight flushes which are 10-high or better. Indeed, a moment’s
thought will convince you that any hand of Alice’s containing the
four 10s will win. Bob cannot now hope to get a straight flush better
than 9-high. To stop Alice getting a royal flush, he must draw at least
one high card from each suit, leaving room for only one card below
a 10. Alice can now turn in four cards and make herself a 10-high
straight flush in a suit other than the suit of Bob’s low card, and Bob
is helpless.
Alice has other winning hands as well—see if you can find them
all!
The next puzzle is a difficult one, asking you to solve a bluffing
game (even the simplest of bluffing games need some machinery to
analyze; imagine what poker might require!). For this we do need the
concept of equilibrium—a pair of strategies, one for each player, with
the property that neither player can improve her results by changing
strategy, if the other player doesn’t change hers. Example: In Rock-
Paper-Scissors, the unique equilibrium is achieved when both players
choose each of the three options with equal probability.
For many games, and all bluffing games, equilibrium strategies
are (like the ones for Rock-Paper-Scissors) randomized. To find these

313
 Mathematical Puzzles

strategies (which, by the way, are often but not always unique) you
can take advantage of the following fact.
Suppose the equilibrium strategy for Player I calls for her to
choose among several options, say A1 , . . . , Ak , each with some posi-
tive probability. Then each option must give her the same expectation
against Player II’s strategy. Why? because if the expectation of, say,
option A3 were among the highest and better than that of some other
options, she could improve her results by choosing A3 instead of ran-
domizing; and this contradicts the definition of equilibrium.

Bluffing with Reals

Consider the following simple bluffing game. Louise and Jeremy ante
$1 each and each is given a secret random real number between 0 and
1. Louise may decide to pass in which case the $2 pot goes to the
player with the higher number. However, if she chooses, Louise may
add another dollar to the pot. Jeremy may then “call” by adding an-
other dollar himself, in which case the pot, now with $4 in it, goes
again to the player with the higher number. Or Jeremy may fold, ced-
ing the pot, with his $1 in it, to Louise.
Surely Louise has the advantage in this game, or at least equality,
since she can break even by always passing. How much is the game
worth to her? What are the players’ equilibrium strategies?

Solution: For each value x that she could get, Louise has to decide
whether to pass or raise; and for each value y that Jeremy could get,
he has to decide whether to call or fold if Louise raises. Thus, there are
in principle infinitely many strategies for each—a pretty big infinity,
too (technically, two to the power of the continuum).
So to find equilibrium strategies, we’ll need to restrict our search
space. A little thought will convince you that Jeremy needn’t consider
any strategy not of the form “call just when y > q” for some fixed
threshold q. The reason is that no matter what Louise does, a higher
value of y cannot increase Jeremy’s incentive to fold.
So let’s assume that Jeremy has picked such a threshold q and con-
sider Louise’s best response when holding the value x. If she passes,
she wins $1 when y < x (probability x) and otherwise loses $1, so her
expectation is

x · $1 + (1−x) · (−$1) = $(2x − 1).

Suppose x > q. Then when y < q Louise wins $1 whatever she does;

314
 Back from the Future

when y > q the game is in Louise’s favor just when x is more than
halfway from q to 1, that is, when x > (q+1)/2, so that’s when she
should raise.
What about when x < q? Should Louise ever raise? If she doesn’t,
we know her expectation is $(2x − 1), which is dismally low when x
is small. If she does raise, she wins $1 when her bluff works, that is,
when y < q, otherwise she loses $2, for a net expectation of

q · $1 + (1−q) · (−$2) = $(3q − 2).

Setting 2x−1 = 3q −2 gives us x = (3q −1)/2 and tells us that Louise

should raise whenever x < (3q − 1)/2, as well as when x > (q+1)/2.
We could at this point compute (as a function of q) what this strat-
egy brings, on average, to Louise’s coffers, and then minimize that
(using calculus—ugh) to find Bob’s choice for his threshold q. But
we know that at equilibrium, if Bob actually drew the number q, he
should be indifferent to the choice of folding or calling. Of course if
he folds he loses $1. If he calls, he wins $2 when Louise was bluffing
and loses $2 otherwise. For this to cost Bob $1 on average, Louise’s
probability of bluffing given that she raised has to be 1/4; that means
3q − 1 (1 − q)/2
= ,
2 3
giving 9q − 3 = 1 − q, q = 0.4.
We conclude that Louise raises when x < 0.1 or x > 0.7, and Bob
calls when y > 0.4. The value of the game is computed as follows:

• When x < 0.1 Louise bluffs, winning $1 when y < 0.4 and
losing $2 otherwise for a net gain of −$0.80.
• When 0.1 < x < 0.7 Louise passes and, with her average hold-
ing of x = 0.4, earns $0.40 − $0.60 = −$0.20.
• Finally, when x > 0.7 Louise breaks even on average when
y > 0.7 as well, but she picks up $1 when y < 0.4, and $2
when 0.4 < y < 0.7 and Bob calls her raise. This gives her a net
average profit of $1 even.

Bottom line: Louise makes 0.1·(−$0.80)+0.6·(−$0.20)+0.3·$1 =

−$0.08 − $0.12 + $0.30 = $0.10. So the game is worth exactly a dime
to Louise.
The next puzzle, also involving equilibrium strategies, is equally
hard to fully analyze—but we’re being asked for much less than that.

315
 Mathematical Puzzles

Swedish Lottery
In a proposed mechanism for the Swedish National Lottery, each par-
ticipant chooses a positive integer. The person who submits the low-
est number not chosen by anyone else is the winner. (If no number is
chosen by exactly one person, there is no winner.)
If just three people participate, but each employs an optimal, equi-
librium, randomized strategy, what is the largest number that has pos-
itive probability of being submitted?

Solution: Suppose k is the highest number any player is willing to

play. If a player chooses k, he wins anytime the other two players
agree, except if they agree on k. But if he chooses k + 1, he wins
anytime they agree, period. Hence, k + 1 is a better play than k, and
we cannot be in equilibrium. The contradiction shows that arbitrarily
high submissions must be considered—sometimes one should choose
1,487,564.
The actual equilibrium strategy calls for each player to submit the
number j with probability (1−r)rj−1 , where
√3
√
1 2 17 + 3 33
r=− − 3 √ √ + ,
3 17 + 3 33 3

which is about 0.543689. The probabilities for choosing 1, 2, 3, and 4

are, respectively, about 0.456311, 0.248091, 0.134884, and 0.073335.
I do not know if our “Swedish lottery” was ever implemented or
even seriously considered for any official lottery, but don’t you think
it should have been?
Suppose a puzzle presents a process, that is, one that moves in time
from one state to another—deterministically, randomly, or under your
control. It’s often helpful to determine whether the process is reversible,
in the sense that you can tell from the state you’re in at the moment
what the previous state must have been.
If there are only a finite number of possible states, then reversibility
tells you that no matter what state you start in, you must return to that
state. Why? You must return to some state, surely, but why the state
you started with?
The proof is by contradiction. Let S be the first repeated state, and
say that this repetition—that is, the second occurrence of state S—
takes place at time t. Reversibility says that the previous state must be
some particular state R. But if S was not the starting state, then S’s

316
 Back from the Future

first occurrence was also preceded by R—so R was repeated before S

was, contradicting our choice of S.
To apply this “reversibility theorem” the trickiest part is often fig-
uring out what information should go into the “state.” Here’s a typical
example.

Pegs on the Corners

Four pegs begin on the plane at the corners of a square. At any time,
you may cause one peg to jump over a second, placing the first on
the opposite side of the second, but at the same distance as before.
The jumped peg remains in place. Can you maneuver the pegs to the
corners of a larger square?

Solution: Note first that if the pegs begin on the points of a grid (i.e.,
points on the plane with integer coordinates), then they will remain
on grid points.
In particular, if they sit initially at the corners of a unit grid square,
then they certainly cannot later find themselves at the corners of a
smaller square since no smaller square is available on the grid points.
But why not a larger one?
Here’s the key observation: The jump step is reversible! If you
could get to a larger square, you could reverse the process and end
up at a smaller square, which we now know is impossible.
Slightly trickier:

Touring an Island
Aloysius is lost while driving his Porsche on an island in which ev-
ery intersection is a meeting of three (two-way) streets. He decides to
adopt the following algorithm: Starting in an arbitrary direction from
his current intersection, he turns right at the next intersection, then
left at the next, then right, then left, and so forth.
Prove that Aloysius must return eventually to the intersection at
which he began this procedure.

Solution: Between intersections Aloysius’ current state can be

characterized by a triple consisting of the edge he’s on, the direction
he’s going on the edge, and the type of his last turn (right or left).
From this information you can determine Aloysius’ last move, and
an application of our reversibility theorem does the rest.
Trickier still:

317
 Mathematical Puzzles

Fibonacci Multiples
Show that every positive integer has a multiple that’s a Fibonacci
number.
Solution: We need to show that for any n, as we generate the Fi-
bonacci numbers (1, 1, 2, 3, 5, 8, 13, 21, 34, etc.) we eventually
find one that is equal to zero modulo n. We normally define the Fi-
bonacci numbers by specifying that F1 = F2 = 1 and for k > 2,
Fk = Fk−1 + Fk−2 .
As it stands, this does not seem like a reversible process; given the
value (modulo n) of, say, Fk+1 , we can’t immediately determine the
value (again, modulo n) of Fk . But if we keep track of the value mod-
ulo n of two consecutive Fibonacci numbers, then we can go backwards
by just subtracting.
For example, if n = 9 and F8 ≡ 3 mod 9 and F7 ≡ 4 mod 9
then we know F7 ≡ 3 − 4 ≡ 8 mod 9. Putting it another way, if we
define Dk = (Fk mod n, Fk+1 mod n), then the Dk ’s constitute a
reversible process.
But so what? We’d like to find a k such that one of the coordinates
of Dk is zero, but how do we know it doesn’t cycle among pairs that
don’t contain a zero?
Aha, a second trick: Start the Fibonacci numbers with F0 = 0,
that is, start one step ahead with 0, 1, 1, 2, 3, etc. Then D0 = (0, 1)
and therefore eventually (after at most n2 steps, in fact) Dk will cycle
back to (0,1), which means that Fk is a multiple of n.
If we take n = 9 as above, for instance, our Dk ’s, beginning with
D0 , are (0,1), (1,1), (1,2), (2,3), (3,5), (5,8), (8,4), (4,3), (3,7), (7,1),
(1,8), (8,0) and we can stop there with F12 = 144 ≡ 0 mod 9. The
D sequence continues (0,8), (8,8), (8,7), (7,6), (6,4), (4,1), (1,5), (5,6),
(6,2), (2,8), (8,1), (1,0), (0,1) so it cycles back in this case after only 24
steps.

Light Bulbs in a Circle

In a circle are light bulbs numbered clockwise 1 through n, n > 1, all
initially on. At time t, you examine bulb number t (modulo n), and if
it’s on, you change the state of bulb t+1 (modulo n); that is, you turn
off the clockwise-next bulb if it’s on, and on if it’s off. If bulb t is off,
you do nothing.

318
 Back from the Future

Prove that if you continue around and around the ring in this man-
ner, eventually all the bulbs will again be on.
Solution: We observe first that there is no danger of turning all the
lights off; if a change is made at time t, bulb t (modulo n) is still on.
Moreover, if we look at the circle just after time t, we can deduce the
state of the bulbs before t (by changing the state of bulb t+1 if bulb t
is on). Thus the process is reversible, provided we take care to include
in the state information not only which bulbs are on and which off,
but which bulb is the one whose state determined the last action.
The number of possible states is less than n×2n , thus is finite, and
we can apply the above argument to conclude that we will eventually
cycle back to the initial all-on state—moreover, we can even insist that
that we reach such a point at a moment when we are slated to again
examine bulb #1.
In the next puzzle, reversibility is just a part of the solution.

Emptying a Bucket
You are presented with three large buckets, each containing an inte-
gral number of ounces of some non-evaporating fluid. At any time,
you may double the contents of one bucket by pouring into it from a
fuller one; in other words, you may pour from a bucket containing x
ounces into one containing y ≤ x ounces until the latter contains 2y
ounces (and the former, x−y).

Prove that no matter what the initial contents, you can, eventually,
empty one of the buckets.

319
 Mathematical Puzzles

Solution: There is more than one way to solve this puzzle, but our
strategy here will be to show that the contents of one of the buckets
can always be increased until one of the other buckets is empty.
To do this, we first note that we can assume there is exactly one
bucket containing an odd number of ounces of fluid. This is true be-
cause if there are no odd buckets, we can scale down by a power of 2;
if there are two or more odd buckets, one step with two of them will
reduce their number to one or none.
Second, note that with an odd and an even bucket we can always
do a reverse step, that is, get half the contents of the even bucket into
the odd one. This is because each state of this pair of buckets can be
reached from at most one state, thus if you take enough steps, you
must cycle back to your original state; the state just before you return
is the result of your “reverse step.”
Finally, we argue that as long as there is no empty bucket, the
odd bucket’s contents can always be increased. If there is a bucket
whose contents are divisible by 4, we can empty half of it into the odd
bucket; if not, one forward operation between the even buckets will
create such a bucket. ♡
Our last “reversibility” puzzle is—there’s no other word for it—
unbelievable.

Ice Cream Cake

On the table before you is a cylindrical ice-cream cake with chocolate
icing on top. From it you cut successive wedges of the same angle θ.
Each time a wedge is cut, it is turned upside-down and reinserted into
the cake. Prove that, regardless of the value of θ, after a finite number
of such operations all the icing is back on top of the cake!

Solution: If you think you have a proof that no finite number of

operations can restore all the icing when x is an irrational angle, you
have justification. When θ is irrational, that is, is not a rational mul-
tiple of 2π radians, every cut will be at a different angle. But the very
first inverted wedge will create a boundary-line on top (and bottom)
of the cake between iced and un-iced areas. If we never again cut at
that point, how can we erase that boundary-line?
In fact, we can cut again at the boundary-line, because when a
wedge is inverted, its iced/un-iced pattern is not only complemented
but reversed. Thus, boundary-lines move.

320
 Back from the Future

In analyzing this puzzle, and indeed many serious algorithmic

problems as well, it helps to redefine the operation so that it is only the
“state space”—here, the icing pattern on the cake—and not the op-
eration itself, that changes from step to step. In this case, that means
rotating the cake after each operation so that you are always cutting
in the same place.
We will use standard mathematical notation, reading angles coun-
terclockwise around the cake counting “east” as 0 radians. One oper-
ation will consist of cutting the cake at −θ and 0, flipping that piece
over, then rotating the whole cake clockwise by angle θ.

Suppose that k operations is the minimum needed to get all the

way around the cake; that is, kθ ≥ 2π but (k−1)θ < 2π. Putting it
another way, k is the greatest integer in 2π/θ. Let δ be the amount by
which you overshoot the first cut, that is, δ = kθ−2π; then 0 ≤ δ < θ.
At this point we advise the reader to try some reasonable angle,
say a bit more than π/4 (making k = 4). You will find that after only
seven cuts (2k−1 in general), all new cuts are in the same places as
old ones!
Let S be the following set of 2k−1 angles, listed in counterclock-
wise order around the cake:
S = {0, θ−δ, θ, 2θ−δ, 2θ, 3θ−δ, 3θ, . . . , (k−1)θ−δ, (k−1)θ}.
The figure below shows the cuts in the case θ = 93.5◦ , where k = 4
and δ = 4 · 93.5◦ − 360◦ = 14◦ .

321
 Mathematical Puzzles

We claim that these are the angles of all cuts that you will ever
make. In fact, it is not hard to see that the set S is closed under our
operation. The angle 0 maps to itself, while the angle −θ = 2π − θ =
(k−1)θ − δ maps to θ. The only line on the piece thus cut is at angle
(k−1)θ, which then moves to θ − δ. The rest of the angles are shifted
up by θ as the cake rotates, so mθ moves to (m+1)θ and mθ−δ moves
to (m+1)θ − δ for 1 ≤ m ≤ k−2.
All the angles in S are indeed cut, in just two passes around the
cake, so S represents our set of cuts exactly. And here’s where we use
reversibility: Since there are only 2k−1 potential cake slices and each
can only have either all its icing on top or all on bottom, there are
only 22k−1 possible states of the cake. Our operation is completely re-
versible, so we must return to the all-icing-on-top state that we started
with.
In fact, we return after many fewer than 22k−1 steps. To see this,
notice that the portions into which the cuts in S divide the cake come
in only two sizes, δ and θ − δ, with k−1 of the former and k of the
latter. (These sizes could be equal, but the types are still distinct.)
A θ-wedge of cake consists of two of these portions, one of each
type. One operation flips the next two portions, one of each type,
reading clockwise around the cake. Notice that the portions of a given
type remain in order after a flip, since the flip involves only two por-
tions of different type.
To get all the icing back on top, each portion must be flipped an
even number of times; thus the number of operations required must
be an even multiple of k−1 (to get the δ portions right) and simulta-
neously an even multiple of k (to get the θ − δ portions right). The
smallest number of operations that fits the bill is 2k(k−1), and that is
the precise answer—except when δ = 0, that is, when θ = 2π/k for
some integer k, in which case only 2k operations are required.
Notice that except in the δ = 0 case, to get all the icing on the
bottom you’d need the number of flips to be simultaneously an odd
multiple of k−1 and an odd multiple of k, which is impossible since
either k−1 or k must be an even number. So it never happens that all
the icing is on the bottom.
A certain very well known mathematician’s reaction, upon hear-
ing the ice cream cake puzzle, was: “I find it hard to believe that the
icing ever returns to the top. But, one thing I’m sure of: If it does,
there must also be a time when it’s all on the bottom!”

322
 Back from the Future

Moth's Tour
A moth alights on the “12” of a clock face, and begins randomly walk-
ing around the dial. Each time it hits a number, it proceeds to the next
clockwise number, or the next counterclockwise number, with equal
probability. It continues until it has been at every number.

What is the probability that the moth finishes at the number “6”?
Solution: No surprise: It pays to think about how the process ends.
Let’s generalize slightly and think about the probability that the
moth’s tour ends at i, where i is any number on the dial except 12.
Consider the first time that the moth gets within one number of i.
Suppose this happens when the moth reaches i − 1 (the argument is
similar if it happens at i + 1). Then i will be the last number visited if
and only if the moth gets all the way around to i+1 before it gets to i.
But the probability of this event doesn’t depend on i. So the prob-
ability of ending at 6 is the same as the probability of ending at any
other number, other than 12 itself. Therefore the probability that the
moth’s tour ends at 6 is 1/11.

Boardroom Reduction
The Board of Trustees of the National Museum of Mathematics has
grown too large—50 members, now—and its members have agreed
to the following reduction protocol. The board will vote on whether
to (further) reduce its size. A majority of ayes results in the immediate
ejection of the newest board member; then another vote is taken, and
so on. If at any point half or more of the surviving members vote nay,
the session is terminated and the board remains as it currently is.
Suppose that each member’s highest priority is to remain on the
board, but given that, agrees that the smaller the board, the better.
To what size will this protocol reduce the board?

323
 Mathematical Puzzles

Solution: We start from the end, of course. If the board gets down
to two members they will certainly both survive, as member 2 (num-
bered from most senior to newest) will vote to retain herself. Thus,
member 3 will not be happy if the board comes down to members 1,
2, and 3; members 1 and 2 would then vote “aye” and eject him.
It follows that at size 4, the board would be stable; member 3
would vote nay to prevent reduction to 1, 2, and 3, and member 4
would vote nay to save him/herself immediately.
The pattern suggests that perhaps the stable board sizes are exactly
the powers of 2; if so, the board will reduce to 32 members and then
stop. Is this right? Suppose not. Let n be the least number for which
the claim fails, and let k be the greatest power of 2 strictly below n.
Since n is our least counterexample, the board is stable when it gets
down to members 1 through k. Thus those k will vote “aye” until then,
outvoting the remaining n−k unless n = 2k. Thus when n = 2k the
k newer members must vote “nay” to save themselves, stopping the
process when n is itself a power of 2. Thus n is not a counterexample
after all, and our claim is correct.
Note we have used not just retrograde analysis, but also induction
and contradiction to solve this puzzle—not to mention consideration
of small numbers. So this puzzle could easily have gone into at least
three other chapters of the book.

For our theorem we will look at a surprising (to some) fact about
numbers, which we can formulate in terms of two steadfast blinkers.
Suppose that two regular blinkers begin with synchronized blinks
at time 0, and afterwards there is an average of one blink per minute
from the two blinkers together. However, they never blink simultane-
ously again (equivalently, the ratio of their frequencies is irrational).
The theorem then tells us that for every positive integer t, there
will be exactly one blink between time t and time t+1.
Theorem. Let p and q be irrational numbers between 0 and 1 that sum to 1.
Let P be the set of real numbers of the form n/p, where n is a positive integer,
and Q the set of reals of the form n/q. Then for every positive integer t, there
is exactly one element of P ∪ Q in the interval [t, t+1).
Let’s first see why the theorem applies. Let p be the rate of the first
blinker; that is, it blinks p times per second. Since the first blink will
be at time 0, the next will be at time 1/p, then 2/p, and so forth; that
is, its blinking times after time 0 will be the set P .
Similarly, the second blinker will blink at rate q and its blinking
times will be 0 together with times in the set Q.

324
 Back from the Future

Saying that together they blink an average of once a second is the

same as saying p+q = 1, and if p/q (which we could write as p/(1−p)
or (1−q)/q) is irrational than so are p and q.
The conclusion of the theorem is then that there is exactly one
blink in each interval between any positive integer and its successor.
Why did I think that might be surprising? We know that we’ll see one
blink on average in such intervals, but exactly one every time? Since the
two rates are not rationally related, we know that there will be some
moment when there are two blinks less than (say) one millionth of a
minute apart. Yet, if you believe the theorem, wedged between those
two blinks there must be an integer!
But the theorem is true; let’s prove it.
Proof. Of course, we start from the end. The key is to notice that the
theorem’s conclusion is equivalent to the following statement: For any
positive integer t, the number of blinks (by both blinkers together) after
time 0 and before time t is exactly t−1. In symbols, |(P ∪Q)∩(0, t)| =
t−1.
Why? Because then the number of blinks between time t and time
t+1 has to be (t + 1 − 1) − (t − 1) = 1.
The number of times the first blinker blinks after time 0 and before
time t is ⌊pt⌋, the greatest integer less than or equal to pt. Similarly, the
second blinker blinks ⌊qt⌋ times in that period. The total number of
blinks in the open interval (0, t) is thus ⌊pt⌋+⌊qt⌋. We know pt+qt =
(p + q)t = t is an integer, but pt and qt aren’t integers; so what can
⌊rt⌋ + ⌊st⌋ be? It’s strictly less than the integer pt + qt but can’t be less
by as much as 2, because we only cut off two fractions between 0 and
1. Since it’s an integer itself, it must be exactly rt + st − 1 = t − 1,
and we are done! ♡

325
 22. Seeing Is Believing
Your mind’s eye is a powerful tool. It’s not a coincidence that when
you suddenly understand something, you say “I see it now.” Draw-
ing a picture—paper is allowed, too, or laptop or electronic pad—can
be the magic ingredient that helps you solve a puzzle, even one that
doesn’t seem at first to call for it.

Meeting the Ferry

Every day at noon GMT a ferry leaves New York and simultaneously
another leaves Le Havre. Each trip takes seven days and seven nights,
arriving before noon on the eighth day. How many of these cross-
Atlantic ferries does one of them pass on its way across the pond?

Solution: Draw it! Put New York on the left of your page and Le
Havre on the right; imagine time running down the page. Then each
ferry trip is a slanted line across the page, and you can verify that each
line meets 13 others on the way.

327
 Mathematical Puzzles

Mathematical Bookworm
The three volumes of Jacobson’s Lectures in Abstract Algebra sit in order
on your shelf. Each has 2′′ of pages and a front and back cover each
1 ′′ 1 ′′
4 , thus a total width of 2 2 .
A tiny bookworm bores its way straight through from page 1, Vol
I to the last page of Vol III. How far does it travel?

Solution: Sounds like the answer should be 3 × 2 12 ′′ − 2 × 14 ′′ = 7′′ ,

that is, the total width of the three volumes less the front cover of
Volume I and the back cover of Volume III.
But visualize the three volumes lined up on your shelf. Where is the
first page of Volume I? And the last page of Volume III? That’s right,
the worm passes only through the middle volume and two additional
′′ ′′
covers, a total of 2 12 + 2 × 14 = 3′′ .

Yes, this means that in a sense the correct way to order multi-
volume books on your shelf is right-to-left, not left-to-right. If you pull
the books off the shelf together with the intent of reading them in one
sitting (not recommended), they are then stacked correctly.

Rolling Pencil
A pencil whose cross-section is a regular pentagon has the maker’s
logo imprinted on one of its five faces. If the pencil is rolled on the
table, what is the probability that it stops with the logo facing up?

Solution: A mental picture is all you need for this one. The pencil
will end up with one of the five faces down, therefore none facing
straight up—so the answer is zero. If you count facing partially up,
make that 25 .

328
 Seeing Is Believing

Splitting a Hexagon
Is there a hexagon that can be cut into four congruent triangles by a
single line?

Solution: Drawing random hexagons and trying to split them this

way will give you a headache. Better: Start with four copies of some
triangle, and try to abut them all to a single line segment, to make
a hexagon. The triangles have a total of 12 sides; uniting three pairs
of sides will get you down to nine sides, so you’ll need to get some
additional pairs of triangle-sides to line up.
This is most easily done if you use right triangles, in which case
you want each side of the segment to support one long leg and one
short leg, while the remaining legs join up perpendicular to the seg-
ment, as pictured below.
Not exactly the kind of hexagon that first springs to mind, though!

Circular Shadows I
Suppose all three coordinate-plane projections of a convex solid are
disks. Must the solid be a perfect ball?
Solution: No, in fact you can take a ball and slice off a piece from
it without affecting its coordinate-plane projections. Just pick a point
on its surface that’s far from any coordinate plane through the ball’s

329
 Mathematical Puzzles

center—for example, the point ( √13 , √13 , √13 ) on the surface of the
unit-radius ball centered at the origin.
Mark a small circle on the ball’s surface around your point and
shave off the cap bounded by the circle. That creates a little flat spot
on the sphere that will go undetected by the projections.
If you really want to avoid wasting material when you 3D-print
your solid, you could do this with eight points, one centered in each
octant, making the circles just touch the coordinate planes.
But you can also go the other way and construct a big object that
contains every other object, convex or not, with the same disk-shaped
coordinate projections: Intersect three long cylinders of the same ra-
dius, each with its axis along a different coordinate axis.

Trapped in Thickland
The inhabitants of Thickland, a world somewhere between Edwin
Abbott’s Flatland and our three-dimensional universe, are an infi-
nite set of congruent convex polyhedra that live between two parallel
planes. Up until now, they have been free to escape from their slab,
but haven’t wanted to. Now, however, they have been reproducing
rapidly and thinking about colonizing other slabs. Their high priest is
worried that conditions are so crowded, no inhabitant of Thickland
can escape the slab unless others move first.
Is that even possible?
Solution: Yes, even if the inhabitants are regular tetrahedra. First,
note that if you view a regular tetrahedron edge-on, as shown on the
left of the figure below, you see a square. Arrange light and dark tetra-
hedra as shown, with the near edge of the dark ones running SW to
NE, and of the light ones NW to SE. One of the two parallel planes
enclosing Thickland contains the near edge of each tetrahedron, the
other the far edge.

330
 Seeing Is Believing

At the moment the tetrahedra have lots of freedom; each can pop
straight out of Thickland by just traveling perpendicular to the planes.
But now let’s squash the tetrahedra together as shown. Now every
dark tetrahedron is locked in by two light ones (its NW and SE neigh-
bors) from above and two other light ones (its SW and NE neighbors)
from below, and similarly for light tetrahedra. So no one can escape
unless at least two other tetrahedra move over. In fact, if you jam
them in so that their faces touch, infinitely many tetrahedra will need
to move just to get one out.

Polygon Midpoints
Let n be an odd integer, and let a sequence of n distinct points be
given in the plane. Find the vertices of a (possibly self-intersecting)
n-gon that has the given points, in the given order, as midpoints of its
sides.

Solution: Let the midpoints given by M1 , . . . , Mn , and suppose

that indeed there are points V1 , . . . , Vn which are the vertices of a
polygon such that M1 is the midpoint of the side V1 , V2 , M2 the mid-
point of A2 , A3 , and so forth, ending with Mn being the midpoint of
An , A1 .
It’s clear that if we know any of the Ai (say, A1 ), the rest of the
vertices are determined; we could reflect A1 180◦ around M1 to get
A2 , and continue. That will work find if, when we finally reflect An
around Mn , we find ourselves back at A1 . But why should we?
Let’s try it: Take an arbitrary point P in the plane, reflect it around
M1 to get P2 , then M2 to get P3 , etc., ending, say, at the point Q. You
won’t find that Q = P1 unless you’re fantastically lucky. But what
happens if you do the procedure again, starting at Q? Then the vectors
QP , P1 Q1 , Q2 P2 , P3 Q3 etc., are all the same, and since n is odd, you
end back at P .
Aha! In that case, let’s do the procedure one more time, starting
at the midpoint R between P and Q. Now you finish at R, and the
polygon has been constructed!
It turns out that this solution is unique. When n is even, generally
there is no solution; but if some starting point P works (by getting
back to P after reflections) then any point will work, giving infinitely
many solutions.

331
 Mathematical Puzzles

Not Burning Brownies

When you bake a pan of brownies, those that share an edge with the
edge of the pan will often burn. For example, if you bake 16 square
brownies in a square pan, 12 are subject to burning. Design a pan and
divide it into brownie shapes to make 16 identically shaped brownies
of which as few as possible will burn. Can you get it down to four
burned brownies? Great! How about three?

Solution: You can beat the squares with sixteen 14 × 1 rectangular

brownies, arranged as in the figure below, only four of which lie on
the boundary.
It seems that you ought to be able to get down to three burnt
brownies, but ordinary triangles won’t do the trick. What if you curve
them?

The 3-burn shape in the figure is constructed using circular arcs

with the property that the straight line distance between the end points
is equal to the radius of the circle.

Protecting the Statue

Michelangelo’s David, in Florence, is protected (on the plane) by laser
beams in such a way that no-one can approach the statue or any of
the laser-beam sources without crossing a beam.
What’s the minimum number of lasers needed to accomplish this?
(Beams reach 100 meters, say.)

Solution: You’ll need a non-convex protected area around David,

in order to protect the lasers themselves as well as David; in fact you’ll
want every edge of your polygon to be incident to a concave corner.

332
 Seeing Is Believing

You can get this with a hexagon shaped like a three-pointed star.
Arrange the lasers at the vertices of a regular hexagon around David,
beamed outward to cross at three corners, as shown below.

To see that six laser-beam sources are needed, let C be the convex
closure of the protected area; it has at least three sides. At each vertex
of C, at least two beams must cross, but no beam can contribute to two
vertices because then the laser source for that beam would lie outside
the protected area. So there must be at least 3 × 2 = 6 lasers.

Gluing Pyramids
A solid square-base pyramid, with all edges of unit length, and a solid
triangle-base pyramid (tetrahedron), also with all edges of unit length,
are glued together by matching two triangular faces.
How many faces does the resulting solid have?
Solution: This problem showed up in 1980 on the Preliminary
Scholastic Aptitude Exam (PSAT), but, to the embarrassment of the
Educational Testing Service (ETS), the answer they marked as cor-
rect was wrong. A confident student called the ETS to task when his
exam was returned to him. Luckily for us, the correct answer boasts a
marvelous, intuitive proof.
The square-base pyramid has five faces and the tetrahedron four.
Since the two glued triangular faces disappear, the resulting solid has
5 + 4 − 2 = 7 faces, right? This, apparently, was the intended line
of reasoning. It may have occurred to the composer that in theory,

333
 Mathematical Puzzles

some pair of faces, one from each pyramid, could in the gluing process
become adjacent and coplanar. They would thus become a single face
and further reduce the count. But, surely, such a coincidence can be
ruled out. After all, the two solids are not even the same shape.
In fact, this does happen (twice): The glued polyhedron has only
five faces.
You can see this in your own head. Imagine two square-based
pyramids, sitting side-by-side on a table with their square faces down
and abutting. Now, draw a mental line between the two apices; ob-
serve that its length is one unit, the same as the lengths of all the
pyramid edges.
Thus, between the two square-based pyramids, we have in effect
constructed a regular tetrahedron. The two planes, each of which con-
tains a triangular face from each square-based pyramid, also contain
a side of the tetrahedron; the result follows. ♡ (Check the figure below
if you find this hard to visualize.)

Precarious Picture
Suppose that you wish to hang a picture with a string attached at two
points on the frame. If you hang it by looping the string over two nails
in the ordinary way, as shown below, and one of the nails comes out,
the picture will still hang (albeit lopsidedly) on the other nail.

334
 Seeing Is Believing

Can you hang it so that the picture falls if either nail comes out?
Solution: One of several ways to hang the picture is illustrated be-
low, with slack so you can see better how it works. This solution re-
quires passing the string over the first nail, looping it over the second,
sending it back over the first nail, then looping it again over the second
nail but counter-clockwise.

There are also some non-topological solutions: For example, you

can pinch a loop of the string between two closely spaced nails,
assuming the nail-head width is not much larger than the string di-
ameter. But why rely on friction when you can use mathematics?

Finding the Rectangles

Prove that any tiling of a regular 400-gon by parallelograms must con-
tain at least 100 rectangles.

Solution: Walk across the tiling from any edge to the opposite edge.
That path must cross a similar path from the 90-degree-away edges,
at a rectangle. Since the sides of that rectangle are parallel or perpen-
dicular to only four edges of the polygon, there must be at least 100
such rectangles.

Tiling a Polygon
A “rhombus” is a quadrilateral with four equal sides; we consider two
rhombi to be different if you can’t translate (move without rotation)
one to coincide with the other. Given a regular polygon with 100 sides,

335
 Mathematical Puzzles

you can take any two non-parallel sides, make( )two copies of each and
translate them to form a rhombus. You get 50 2 different rhombi that
way. You can use translated copies of these to tile your 100-gon; show
that if you do, you will use each different rhombus exactly once!

Solution: Let ⃗u be one of the sides of the 2n-gon; a ⃗u-rhombus is

any of the n−1 rhombi using ⃗u as one of its two vectors. In a tiling,
the tile next to a ⃗u-side must be a ⃗u-rhombus, as must the tile on the
other side of that one, and so forth until we reach the opposite side
of the 2n-gon. Notice that each step of this path proceeds in the same
direction (i.e., right or left) with respect to the vector ⃗u, as must any
other path of ⃗u-rhombi; but then there can be no other ⃗u-rhombi, since
they would generate paths with no way to close and nowhere to go.

The similarly defined path for a different side ⃗v must cross the ⃗u,
and the shared tile is of course made up of ⃗u and ⃗v . Can they cross
twice? No, because a second crossing would have ⃗u and ⃗v meeting at
an angle greater than π inside the common rhombus. ♡

Tiling with Crosses

Can you tile the plane with 5-square crosses? Can you tile 3-space
with 7-cube crosses?

Solution: Tiling the plane with 5-square crosses is a snap if you

organize them into diagonals, as shown on the left side of the figure
below.

336
 Seeing Is Believing

To tile 3-space with 7-cube crosses, use five of the seven cubes of
each of a lot of crosses to tile a unit-thickness flat slab, much as the
flat crosses were used to tile the plane. However, between each pair
of diagonal strips, leave blank a strip of cattycorner dominoes.
Each of these is filled in one cube from below the slab, and one
from above. ♡

Cube Magic
Can you pass a cube through a hole in a smaller cube?
Solution: Yes. To pass a unit cube through a hole in a second unit
cube, it suffices to identify a projection of the (second) cube which
contains a unit square in its interior. A square cylindrical hole of side
slightly more than 1 can then be made in the second cube, leaving
room through which to pass the first cube.
You can then do the same, with even smaller tolerances, if the
second cube is just a bit smaller than a unit cube.
The easiest (but not the only) projection you can try this with is the
regular hexagon you get by placing three vertices and the centroid on
a plane perpendicular to your line of sight. You can see this hexagon
by viewing the cube so that one of its vertices is centered.
Letting A be the projection of one of the visible faces onto√ the
plane, we observe that its long diagonal is the same length ( 2) as a
unit square’s, since that line has not been foreshortened. If we slide a
copy of A over to the center of the hexagon, then widen it to form a
unit square B, B’s widened corners will not reach the vertices of the
hexagon (since the distance between opposing vertices of the hexagon
exceeds the distance between opposing sides).

337
 Mathematical Puzzles

It follows that if we now tilt B slightly, all four of its corners will
lie strictly inside the hexagon. ♡

Circles in Space
Can you partition all of 3-dimensional space into circles?

Solution: It’s an odd thing to ask—partitioning space into objects

of lower dimension. But there’s nothing to prevent you from trying to
do that, and in fact, in this case it is indeed possible.
But how? We can certainly partition 3-space minus a point into
spheres, namely all spheres having the missing point as their center.
Can we partition a sphere into circles? If we leave out the poles, you
can certainly partition the rest into circles of lattitude. In fact we can
omit any two points P and Q, and partition the rest into circles whose
centers lie on the plane through P , Q and the sphere’s center; for ex-
ample, circles whose ratio of distance to P and Q is the same along
the short route from P to Q as it is on the long route.
With these observations we can construct a partition of all of 3-
space into circles as follows. We start with a line of unit-radius circles
on the XY-plane, one centered at (4n+1, 0) for each integer n. The key

338
 Seeing Is Believing

is that every sphere centered at the origin hits the union of those circles
in exactly two points. And, very importantly, the origin is covered (by
the unit circle centered at (1,0)).

Thus, if we partition every sphere centered at the origin, with its

two intersection points omitted, into circles, then all those circles plus
our line of circles on the plane partition space perfectly! ♡

Invisible Corners
Can it be that you are standing outside a polyhedron and can’t see
any of its vertices?
Solution: Yes. Imagine six long planks arranged so that they meet
in their middles to form the walls of a cubical room, but do not quite
touch. From the middle of this room you won’t be able to see any
vertices. These planks can easily be hooked up far from the room to
form a polygon, with the necessary additional vertices well hidden.
[If there is a plane separating you from the polyhedron, this can’t
happen.]

For our theorem, I can’t resist using one of the most famous of all
theorems with no-word proofs. A picture is all you need!
A lozenge is a rhombus formed from a pair of unit-edge triangles
glued together along an edge. Suppose a regular hexagon with integer
sides is tiled with lozenges. The tiles come in three varieties, depend-
ing on orientation.

339
 Mathematical Puzzles

Theorem. Any tiling of a regular hexagon with integer sides by unit lozenges
uses precisely the same number of lozenges of each of the three orientations.
Proof.

340
 23. Infinite Choice
Need an axiom? As often as not, the Axiom of Choice is the axiom
of choice.
The Axiom of Choice (abbreviated AC here and in logic books)
says that given any collection of nonempty sets, you can choose an
element from each set. Seems pretty reasonable, right? What’s to stop
you?
Indeed, if the collection is finite, there’s no problem (you can prove
by induction that one item can be selected from each set.) If the sets
have distinguished objects, you can use them; for instance, among
infinitely many pair of shoes, you could choose the left shoe from
each pair. But suppose you have infinitely many pairs of socks?
There is a name, actually, for the set of all ways to choose an el-
ement from each set: It’s called the product of the sets. The size of
the product, if both the sets and the collection are finite, is the ordi-
nary product of the sizes of the sets. It seems unlikely that the product
of an infinite collection of nonempty sets would suddenly decide to
be empty, but that can happen if you don’t have AC. And if you’re
a mathematician, you might find it handy to have tools like Zorn’s
Lemma, the well-ordering theorem, or the Hausdorff maximal prin-
ciple (each equivalent to AC), as well as myriad other theorems that
require AC without being obvious about it.
On the other hand, AC has some pretty weird consequences. The
most famous of these is the Banach–Tarski paradox: A unit ball in
3-space can be partitioned into five subsets which can then be re-
assembled (using only rotations and translations) to form two unit
balls! Mathematicians are used to having to get around problems like
this by restricting attention to “measurable” sets. But, as you will see
later, there are other bizarre consequences of AC that mathematicians
are less familiar with.
So is AC true or false? I like to tell my students that although
neither the axiom of choice nor its negation can be disproved, either
can be made to look ridiculous. Both AC and “not AC” are consis-
tent with the usual axioms of set theory, so you have a choice; as a

341
 Mathematical Puzzles

puzzle solver you will mostly want to choose AC. (Assumptions like
the Axiom of Determinacy that are not consistent with AC can oc-
casionally come in handy, but we will not pursue such contingencies
here.)
You might reasonably ask: How do collections of sets that don’t
have obvious “choice functions” arise? One way is in the considera-
tion of equivalence relations.
A (binary) relation is technically a set of pairs, but you can think
of it as a property that may or may not be true of a given pair of items.
For example, among people, “x knows y” is a relation; so is “x and
y are siblings.” The latter of these is a symmetric relation, one that is
satisfied by the pair (x, y) if it is satisfied by (y, x). It is also a transitive
relation: If (x, y) and (y, z) satisfy the relation, so must (x, z).
Suppose we change the relation to “x and y are either full siblings,
or are the same person.” Then the relation is also reflexive, meaning
that (x, x) satisfies the relation for any x. Relations that are symmet-
ric, transitive, and reflexive are said to be equivalence relations. If a set
has an equivalence relation, the set partitions neatly into equivalence
classes: subsets whose elements are all related to one another but not
to anyone outside the subset. For example, for the above relation, the
equivalence classes are the sets of people that share a given pair of
parents.
Suppose humans are forced to settle a new planet and it’s decided
that exactly one person from each set of siblings will go. Such a col-
lection is called a set of representatives of the equivalence relation. In
this case we don’t need to apply AC to get a set of representatives; for
example, we could always send the eldest of the siblings (and anyway
there are only finitely many sibling sets that we know of).
But, generally, we may indeed need AC to get a set of representa-
tives. Consider, for example, the set of all (countably) infinite binary
sequences. Suppose (a1 , a2 , . . . ) is one such sequence and (b1 , b2 , . . . )
is another. Say that these two sequences are related if for all but
finitely many indices i, ai = bi . You can easily verify that this relation
is symmetric, transitive, and reflexive, therefore is an equivalence re-
lation. Thus the set of all infinite binary sequences breaks up into (a
lot of) equivalence classes. One of these, for example, is the set of all
binary sequences that have only finitely many 1’s.
According to AC, it is possible to pick one sequence from each
equivalence class. From the aforementioned class of sequences with
only finitely many 1’s, you might want to pick the all-zero sequence.
But for most classes there won’t be an obvious choice, so it’s hard to

342
 Infinite Choice

see how you can avoid using AC. Let’s go ahead and invoke AC to get
our set of representatives. What good is it? Well, if you’re a prisoner. . .

Hats and Infinity

Each of an infinite collection of prisoners, numbered 1, 2, . . . , is to be
fitted with a red or black hat. At a prearranged signal, all the prisoners
are revealed to one another, so that everyone gets to see all his fellow
prisoners’ hat colors—but no communication is permitted. Each pris-
oner is then taken aside and asked to guess the color of his own hat.
All the prisoners will be freed provided only finitely many guess
wrongly. The prisoners have a chance to conspire beforehand; is there
a strategy that will ensure freedom?

Solution: If we associate “red” with the digit 1 and “black” with

0, assignments of hats to prisoners correspond precisely to infinite
binary sequences. Suppose that the prisoners believe in AC and have
the wherewithal to agree in advance on a set of representatives for the
equivalence relation described above. Thus, for each set S of related
sequences, they have identified a particular member.
When the curtain goes up, prisoner i sees the whole binary se-
quence except for its ith entry. That’s enough for him to identify the
equivalence class S to which that sequence belongs. He then looks up
(or remembers) the representative that everyone has agreed upon for
the set S, and guesses that his own hat color corresponds to the ith
entry of that representing sequence.
Consequently, everyone’s guess will correspond to this same rep-
resenting sequence, and since that sequence is itself in S, only finitely
many prisoners will have guessed wrong. ♡
It can be shown that under certain assumptions that contradict
AC, the prisoners have essentially no chance, even if the warden is
obligated to assign colors uniformly at random.
Try this similar puzzle.

All Right or All Wrong

This time the circumstances are the same but the objective is different:
The guesses must either be all right or all wrong. Is there a winning
strategy?

343
 Mathematical Puzzles

Solution: This version perhaps sounds even tougher, as only two

successful outcomes are possible. But it does admit a cute solution if
there are only finitely many prisoners (stop here and see if you can do
the finite version before reading further).
Yes, if there are only finitely many prisoners, they can simply de-
cide to assume that the number of red hats is even. In other words, a
prisoner who sees an odd number of red hats guesses that his own hat
is red, otherwise that it is black. Then if, indeed, the number of red
hats is even, the prisoners are all correct; if odd, all wrong.
In the infinite case, this won’t work unless the warden assigns only
finitely many hats of one color. But the prisoners’ agreed-upon set of
representatives makes things easy. Each prisoner again identifies the
equivalence class of the hat assignment and finds its representative;
he then assumes that (say) the number of discrepancies between the
actual sequence and the representative is even, and guesses his own
color accordingly.
However, there’s a simpler solution: Everyone guesses “green”! ♡
The Axiom of Choice has the powerful consequence that any set
can be well-ordered. (A set X is said to be well-ordered if it has an
order relation on it with the property that all non-empty subsets of
X have a lowest element.) In addition, it’s possible to ensure that
the set of elements below any element of X has “cardinality” (i.e.,
size) less than the cardinality of X itself. We won’t prove these con-
sequences here, but we’ll use them to solve the following innocent-
sounding problem.

Double Cover by Lines

Let Lθ be the set of all lines on the plane at angle θ to the horizontal.
If θ and θ′ are two different angles, the union of the sets Lθ and Lθ′
constitutes a double cover of the plane, that is, every point belongs to
exactly two lines.
Can this be done in any other way? That is, can you cover each
point of the plane exactly twice using a set of lines that contains lines
in more than two different directions?

Solution: Since this problem appears here, you have a big clue that
the answer is yes if you have AC. You can prove it by well-ordering
all the points on the plane in such a way that the set of points below
any point has cardinality less than 2ℵ0. (2ℵ0 stands for the cardinality

344
 Infinite Choice

of the set of all real numbers, also the set of angles, the set of points
on the plane, and many other sets; it’s strictly greater than ℵ0 , the
cardinality of the set of whole numbers. The symbol ℵ, pronounced
“aleph,” is the first letter of the Hebrew alphabet.) We start by picking
three lines that cross one another, so that we already have our three
directions. Let P be the lowest point in our well-ordering that isn’t
already double-covered, and pick a line through P that misses the
three points that are currently double-covered.
Now repeat the process with a new point Q, this time avoiding
a larger, but still finite, set of points that are already double-covered.
We can do this until we have a countably infinite number of lines in
our set. But why stop there? Until we’re done, there’s always a lowest
point of the plane that hasn’t been double-covered yet. And because
the number of points so far considered is less than 2ℵ0, there’s always
an angle available that misses all points that are currently double-
covered.
This construction is perhaps disappointing in that it does not leave
you with any geometry you can wrap your head around. No better
construction is known.
Let us return to direct applications of AC. If you were unim-
pressed by AC’s power to free prisoners, perhaps the next puzzle will
get your attention.

Wild Guess
David and Carolyn are mathematicians who are unafraid of the infi-
nite and cheerfully invoke the Axiom of Choice when needed. They
elect to play the following two-move game. For her move, Carolyn
chooses an infinite sequence of real numbers, and puts each number
in an opaque box. David gets to open as many boxes as he wants—
even infinitely many—but must leave one box unopened. To win, he
must guess exactly the real number in that box.
On whom will you bet in this game, Carolyn or David?
Solution: I can hear you thinking: “I’m betting on Carolyn! She
can just put, say, a random real number between 0 and 1 in each box.
David has no clue what’s in the unopened box, hence his probability
of winning is zero.”
In fact David has an algorithm that will guarantee him at least a
99% chance of winning, regardless of Carolyn’s strategy. You don’t

345
 Mathematical Puzzles

believe me? That shows good judgment on your part, but it’s nonethe-
less true given AC.
Here’s how it works. Before the game even starts, David proceeds
much as the prisoners in Hats and Infinity did, but this time with se-
quences of real numbers instead of bits. (If you state the problem using
bits instead of reals, David can use the prisoners’ set of representa-
tives.) So, two sequences of reals x1 , x2 , . . . and y1 , y2 , . . . are related
if xi ̸= yi for only finitely many indices i. Again this is an equivalence
relation, and David invokes AC to choose from each equivalence class
of real sequences one representative sequence. He can even show his
set of representatives to Carolyn; it won’t help her!
Now Carolyn picks her sequence, let’s call it c1 , c2 , . . . , and boxes
up the numbers. David takes the boxes and breaks them up into 100
infinite rows. We may as well assume the boxes in row 1 contain c1 ,
c101 , c201 , etc., while row 2 begins with c2 and c102 ; finally row 100
contains c100 , c200 and so forth. Each row is itself a sequence of real
numbers.

In a crucial step, David now chooses, uniformly at random, an

integer between 1 and 100. Say it’s 44.
David first opens all the boxes in row 1. This sequence belongs
to some equivalence class, say S1 , whose representative he now looks
up. The representative, say d1 , d2 , . . . , will differ from row 1 in finitely
many positions, the last of which is, say, the 289th entry in each.
David writes on a piece of paper “n1 = 289.” (If the sequences hap-
pen to be identical, he writes “n1 = 0.”)

346
 Infinite Choice

He repeats the procedure with row 2, this time getting a (probably)

different class S2 with its own representative, and ends up writing
down perhaps “n2 = 4,183,206.”
This continues with every row except the 44th. When he’s done
with row 100, David has written down values for the 99 numbers
n1 , n2 , . . . , n43 and n45 , . . . , n100 . He sets m equal to one more than
the maximum of these numbers.
Now he finally turns to row 44, opening not all the boxes, but
only those after the mth. This is enough for him to determine the
equivalence class S44 of that row, and to retrieve its representative
sequence, which we’ll denote by f1 , f2 , . . . . David now guesses that
the mth box in row 44 contains the real number fm .
So why does this work? Well, ultimately there is some largest
index (n44 ) which represents the last position where row 44 and
f1 , f2 , . . . disagree. Unless n44 ≥ m, David’s guess will be correct.
But m was greater than the maximum of n1 , n2 , . . . , n43 and
n45 , . . . , n100 , so in order for n44 to equal or exceed m, n44 would
have to be the unique largest of the 100 numbers n1 , . . . , n100 .
What’s the probability of that? Since 44 was random, and only one
of these numbers can be the unique largest, it’s at most 1/100.
Bingo!
If you find that you have to read this proof more than once to ap-
preciate it, you are not alone. But it’s legitimate. The Axiom of Choice
really does imply that David can win this game with any desired prob-
ability p < 1. Is there something wrong with your intuition, or is the
Axiom of Choice suspect? You be the judge. If it’s any help, you can
also use AC to predict the future (see Notes & Sources).
I hear the following objection. Suppose Carolyn chooses her num-
bers independently, each uniformly random between 0 and 1. Then
the content of any box, in particular the mth box of row 44, is random
and there’s no way David can guess it with probability 99% or indeed
any probability greater than zero.
The counter to this objection is that when we say infinitely many
random numbers are independent, we usually mean that each value
is independent of any finite number of the others. If you want each
number to be independent of all of the others, and you believe in AC,
you’re stuck! It’s interesting, in my view, that AC—which we usually
think of as an aid to constructing infinite things—prevents you from
constructing an infinite sequence of random reals, or even random
bits, that are independent in this strong sense.

347
 Mathematical Puzzles

We now move on to some other puzzles involving infinity (but

we’ll be assuming the axiom of choice holds, in everything that fol-
lows). We start with a an infinite expression.

Exponent upon Exponent

··
x·
Part I: If xx = 2, what is x?
··
x·
Part II: If xx = 4, what is x?
Part III: How do you explain getting the same answer to Parts I and
II?
· ·
x·
Solution: If it means anything at all, xx must be the limit of the
sequence
x xx
x, xx , xx , xx , . . . ,
assuming that limit exists. Note that this expression is not the same
as
(. . . (((xx )x )x ) . . . ).
··
x·
The exponent of the bottom x in the expression xx is the same as
x·
·· √
the expression itself; thus, if xx = 2 then x2 = 2, x = 2.
That’s one part of the puzzle out of the way.
·
··
xx
For the second part,√similar
√ reasoning tells you that if x = 4,
then x4 = 4, thus x = 4 2 = 2.
√ ··
√ √2 2 ·
Aha. Something’s wrong here. What exactly is 2 ? It can’t
be both 2 and 4. Which is it, if either? Or is it something else, or noth-
ing at all? √
√ √ √2 √ √ 2 2
We need to look at the sequence 2, 2 , 2 , . . . , and
determine whether it has a limit. In fact, it does; the sequence is in-
creasing and bounded above.
To show the former, we name the sequence s1 , s2 , . . . and prove
by induction that√1 < si < si+1 for each i ≥ 1. This is easy: Since
√ √ 2 √ √ si √ si−1
2 > 1, s2 = 2 > 2 = s1 ; and si+1 = 2 > √2 = si .
To get the bound, observe that if we replace the top 2 in any si by
the larger value 2, the whole expression collapses to 2. (We could sim-
ilarly show the sequence is bounded by 4, but of course if it’s bounded
by 2 it’s automatically bounded by anything bigger than 2.)

348
 Infinite Choice

Now√ that we know the limit exists, let us call it y; it must indeed
y
satisfy 2 = y. Looking at the equation x = y 1/y , we observe (per-
haps using elementary calculus—sorry) that x is strictly increasing in
y up to its maximum at y = e and strictly decreasing thereafter. Thus,
there are at most
√ two values of y corresponding to any given value of
x, and for x = 2, we know the values: y = 2 and y = 4.
Since our sequence is bounded by 2, we rule out 4 and conclude
that y = 2. ♡
··
x·
Generalizing the above argument, we see that xx is meaningful
and equal to the lower root of x = y 1/y , as long as x ≤ e1/e . For
x = e1/e , the expression is equal to e, but as soon as x exceeds e1/e ,
the sequence diverges to infinity. This is why Part II of the puzzle fails:
··
x·
There is no x such that xx = 4.
We have already noted that there are different infinities; the small-
est is countable infinity, which we called ℵ0 . A set is said to be count-
able if you can put its members into one-to-one correspondence with
the positive integers. The set of rational numbers (fractions) is count-
able; reference to a particularly nice way of putting them into corre-
spondence with the positive integers is given in Notes & Sources. The
set of real numbers, however, is not countable, as observed by the bril-
liant Georg Cantor in 1878.
Here’s a simple puzzle for which you can make use of countability.

Find the Robot

At time t = 0 a robot is placed at some unknown grid point in 3-
dimensional space. Every minute, the robot moves a fixed, unknown
distance in a fixed, unknown direction, to a new grid point. Each
minute, you are allowed to probe any single point in space. Devise an
algorithm that is guaranteed to find the robot in finite time.

349
 Mathematical Puzzles

Solution: There are only countably many integer vectors (a, b, c, x,

y, z), each coding a starting position (a, b, c) and displacement
(x, y, z). List them as v1 , v2 , etc., and at time t, probe the point
(a, b, c) + t(x, y, z) where (a, b, c, x, y, z) = vt .

Figure Eights in the Plane

How many disjoint topological “figure 8s” can be drawn on the plane?

Solution: We could draw concentric circles on the plane with all

possible positive real diameters, hence if the puzzle asked for circles
instead of figure eights, the answer would be “uncountably many,” or
more precisely, “the cardinality of the reals.”
However, we can only draw countably many eights. Associate
with each 8 a pair of rational points (points of the plane with both
coordinates rational numbers), one in each loop; no two figure eights
can share a pair of points. Hence, the cardinality of our set of 8s is
no greater than the set of pairs of pairs of rational numbers, which is
countable. ♡
Was that too easy? Try the next one!

Y's in the Plane

Prove that only countably many disjoint Y’s can be drawn in the
plane.
Solution: Here is a particularly neat proof (there is more than one
way to do this). Associate with each Y three rational circles (rational
center and radius) containing the endpoints, and small enough so that
none contains or intersects any other arm of the Y. We claim that no

350
 Infinite Choice

three Ys can all have the same triple of circles; for, if that were so,
you could connect the hub of each Y to the center of each circle by
following the appropriate arm until you hit the circle, then following a
radius to the circle’s center. This would give a planar embedding of the
graph K3,3 , sometimes known as the “gas-water-electricity network.”
In other words, we have created six points in the plane, divided
into two sets of three each, with each point of one set connected by a
curve to each point of the other set, and no two curves crossing. This
is impossible; in fact, readers who know Kuratowski’s Theorem will
recognize this graph as one of the two basic nonplanar graphs.
To see for yourself that K3,3 cannot be embedded in the plane
without crossings, let the two vertex sets be {u, v, w} and {x, y, z}.
If we could embed it without crossings the sequence u, x, v, y, w, z
would represent consecutive vertices of a (topological) hexagon. The
edge uy would have to lie inside or outside the hexagon (let us say
inside); then vz would have to lie outside to avoid crossing uy, and
wx has no place to go. ♡

There’s no shortage of serious mathematical theorems that use

some form of AC in their proofs; they’re rampant in algebra, analysis
and topology. Here’s one that concerns only basic graph theory, which
we make use of elsewhere in this volume.
Recall that a graph is a set of points, called vertices, together with a
symmetric relation between pairs of distinct vertices. Two related ver-
tices are said to be adjacent and to constitute an edge of the graph, usu-
ally illustrated by connecting the vertices by a line segment or curve.
A graph is two-colorable if its vertices can be partitioned into two sets
X and Y so that no two vertices in the same one of these sets are ad-
jacent. Below is an example of a graph with its vertices two-colored
(by black and white).

351
 Mathematical Puzzles

The graph whose vertices are the digits 0 through 9 with 0 adjacent
to 1, 1 to 2, 2 to 3, . . ., 8 to 9 and 9 to 0, can be colored with two
colors by putting the even digits in X and the odd digits in Y. But if
we try two-coloring the same construction but with an odd number
of vertices—an “odd cycle”—it won’t work.
In fact, no graph containing an odd cycle can be two-colorable. A
famous, elementary theorem of graph theory says the converse.
Theorem. If a graph contains no odd subset of vertices whose adjacencies
include those of an odd cycle, then the graph is two-colorable.

The idea of the proof is this: Pick some vertex v and put it in X,
then put all the vertices that are adjacent to v in Y , then all the vertices
adjacent to them in X again, and so forth. The absence of odd cycles
assures you that you will never find yourself trying to put the same
vertex in both X and Y . If at any stage you find that you are finding
no new adjacencies but haven’t used up all the vertices of the graph,
pick an unassigned vertex v ′ , put it in X, and continue as before.
That’s all very fine, but if the graph is infinite and has infinitely
many connected components, the part of the proof asking you to
“pick an unassigned vertex” requires the Axiom of Choice. Here’s
an example of such a graph: Let the vertices be all real√numbers, with
x adjacent to y if either |x − y| = 1 or |x − y| = 2. This graph
has no odd cycles, because if you start a cycle at x and return to x,
you will necessarily use an even number of 1-sized steps (some num-
ber up, and an equal number down) and likewise
√ √ an even number of
2-sized steps; this is because no multiple of 2 is an integer.
Nonetheless, under some axioms that contradict AC, this graph
has no two-coloring! Indeed, if you try to construct a set of vertices
that could serve as X, you will encounter only frustration. Three-
colorings are easily found, though; it’s really quite reasonable, in
graph-theoretic terms, to postulate that this graph has chromatic num-
ber 3, that is, can be properly colored with three colors but no fewer.

352
 24. Startling
Transformation
Often a puzzle that looks impenetrable becomes suddenly transpar-
ent if you just look at it a different way. Yes, that may require some
creativity and imagination on your part; but great ideas, no matter
how brilliant they appear at first, don’t arise out of nothing. A little
experience may go a long way to sparking your next bit of genius.

Sinking 15
Carol and Desmond are playing pool with billiard balls number 1
through 9. They take turns sinking balls into pockets. The first to sink
three numbers that sum to 15 wins. Does Carol (the first to play) have
a winning strategy?

Solution: If Carol and Desmond record their sunk balls on the

magic square below, then their objective is to fill all three squares in
a row, column, or diagonal. Thus, they are playing Tic-Tac-Toe!

As a consequence, Carol does not have a winning strategy; best

play leads to a draw.

Slicing the Cube

Before you is a circular saw and a 3 × 3 × 3 wooden cube that you
must cut into twenty-seven 1 × 1 × 1 cubelets. What’s the smallest

353
 Mathematical Puzzles

number of slices you must make in order to do this? You are allowed
to stack pieces prior to running them through the saw.
Solution: If you hold the cube together (carefully!) throughout, you
can make two horizontal cuts all the way through the cube, then two
vertical North–South cuts and two vertical East–West cuts—six cuts
in all—to reduce the cube to 27 cubelets. Can you do it with fewer
than six cuts?
No, because the center cubelet has to be cut out on all six sides,
and no cut can do more than one.

Expecting the Worst

Choose n numbers uniformly at random from the unit interval [0,1].
What is the expected value of their minimum?

Solution: The answer is 1/(n+1). (And the average value of the

next-smallest is 2/(n+1), and so forth.) This kind of problem can be
solved using calculus, but when the solution is so nice, you might sus-
pect there is a simpler way—and there is. We make use of the interval’s
more symmetrical cousin, the circle.
Choose n+1 numbers x0 , . . . , xn uniformly and independently
from a circle of circumference one. By symmetry, the expected dis-
tance between neighboring numbers is 1/(n+1). Cut the circle at x0
and open it up to form a line segment of length one. The remaining
numbers x1 , . . . , xn will be uniformly distributed on this line, and the
least of these is the next xi to the right of x0 (assuming you opened
up the circle clockwise). Thus its expected value is the same as the
expected value of the distance between x0 and the next selected point
to its right, which is 1/(n+1).
The idea of bending a line segment around to make a circle, then
exploiting the symmetry of the circle, will be useful later for proving
our theorem. In the meantime, let’s look at a betting game.

Next Card Red

Paula shuffles a deck of cards thoroughly, then plays cards face up one
at a time, from the top of the deck. At any time, Victor can interrupt
Paula and bet $1 that the next card will be red. He bets once and only
once; if he never interrupts, he’s automatically betting on the last card.

354
 Startling Transformation

What’s Victor’s best strategy? How much better than even can he
do? (Assume there are 26 red and 26 black cards in the deck.)
Solution: We know that Victor’s expectation in this game, if he
plays it well, is at least 0—because he can just bet on the first card,
which is equally likely to be one of the 26 red cards or to be one of
the 26 black cards. Or, he could wait for the last card, with the same
result.
But Victor is a smart guy and knows that if he waits for a point
when he’s seen more black cards go by than red, he can bet at that
moment and enjoy odds in his favor. Of course, he could wait until
he’s seen all the black cards go by and then bet knowing he’s going
to win, but that might not ever happen. Perhaps best is the following
conservative strategy: He waits until the first time that the number of
black cards he’s seen exceeds the number of red, and takes the result-
ing (probably small) odds in his favor. For example, if the very first
card is black, he bets on the next one, accepting a positive expectation
of (26/51) · $1 + (25/51) · (−$1) which is about 2 cents.
Of course, though unlikely, it possible that the black cards seen
never get ahead of the red ones, in which case Victor is stuck betting
on the last card, which will be black. It seems like it might be difficult
to compute Victor’s expectation from this or any moderately complex
strategy.
But there is a way—because it’s a fair game! Not only has Victor
no way to earn an advantage, he has no way to lose one either: All
strategies are equally ineffective.
This fact is a consequence of the martingale stopping time theo-
rem, and can also be established by induction on the number of cards
of each color in the deck. But there is another proof, which I will de-
scribe below, and which must surely be in “the book.” (As readers
may know or recall from Chapter 9, the late, great mathematician
Paul Erdős often spoke of a book owned by God in which is written
the best proof of each theorem. I imagine Erdős is reading the book
now with great enjoyment, but the rest of us will have to wait.)
Suppose Victor has elected a strategy S, and let us apply S to a
slightly modified variation of the game. In the new variation, Victor
interrupts Paula as before, but this time he is betting not on the next
card in the deck, but instead on the last card of the deck.
In any given position, the last card has precisely the same proba-
bility of being red as the next card. Thus, the strategy S has the same
expected value in the new game as it did before.

355
 Mathematical Puzzles

But, of course, the astute reader will already have observed that
the new variation is a pretty uninteresting game; Victor wins if and
only if the last card is red, regardless of his strategy. ♡

Magnetic Dollars
One million magnetic “susans” (Susan B. Anthony dollar coins) are
tossed into two urns in the following fashion: The urns begin with
one coin in each, then the remaining 999,998 coins are thrown in the
air one by one. If there are x coins in one urn and y in the other, mag-
netic attraction will cause the next coin to land in the first urn with
probability x/(x + y), and in the second with probability y/(x + y).

How much should you be willing to pay, in advance, for the con-
tents of the urn that ends up with fewer susans?
Solution: You might very reasonably worry that one urn will “take
over” and leave very little for the other. In my experience, faced with
this problem, most people are unwilling to offer $100 in advance for
the contents of the lesser urn.
In fact the lesser urn is worth, on average, a quarter of a million
dollars. This is because the final content of (say) Urn A is precisely
equally likely to be any whole number of dollars from $1 all the way
to $999,999. Thus, the average amount in the lesser urn is
1
(2 · $1 + 2 · $2 + . . . 2 · $499,999 + 1 · $500,000)
999,999

which is $250,000 plus change.

How can we see this? There are several proofs but here’s my fa-
vorite. Imagine that we shuffle a fresh-from-the-box deck of cards in
the following careful manner. We place the first card, say the Ace of
Spades, face down on the table. Then we take the second card, maybe
the Deuce of Spades, and randomly place it over or under the ace.
Now there are three possible slots for the Three of Spades; choose

356
 Startling Transformation

one at random, putting it on top of the current pile, under the current
pile, or between the ace and deuce, each with probability 13 .
We continue in this manner; the nth card is inserted in one of n
possible slots, each with probability 1/n. The result eventually is a
perfectly random permutation of the cards.
Now think of the cards that go in above the Ace of Spades as the
coins (not counting the first) that go into Urn A, and the ones that
go under as coins going into Urn B. The magnetic rule above de-
scribes the process perfectly, because if there are currently x−1 cards
above the Ace of Spades (thus x slots) and y−1 cards under the Ace
of Spades (thus y slots), the probability that the next card goes above
the Ace of Spades is x/(x+y).
Since in the final shuffled deck the Ace of Spades is equally likely
to be in any position, the number of slots ending up above it is equally
likely to be anything from 1 to the number of cards in our deck, which
for our puzzle is 999,999.

Integral Rectangles
A rectangle in the plane is partitioned into smaller rectangles, each of
which has either integer height or integer width (or both). Prove that
the large rectangle itself has this property.
Solution: This problem is famous for its many proofs, each a kind
of transformation. Here’s one proof that turns the picture into a graph.
Place the big rectangle with its lower left corner at the origin of
a Cartesian grid, and sides parallel to the coordinate axes. Let G be
the graph whose vertices are all the points with integer coordinates
that are also corners of small rectangles (indicated by shaded circles),
plus the small rectangles themselves (indicated by white circles at their
centers).

357
 Mathematical Puzzles

Put an edge between a shaded vertex and a white vertex if the

shaded vertex is a corner of the white vertex’s rectangle, as shown in
the example.
Every white vertex has 0, 2, or 4 neighbors (since one of its dimen-
sions is an integer), and every shaded vertex has two or four neigh-
bors, except that shaded vertices at the corners of the big rectangle
have just one neighbor each.
In any graph there must be an even number of vertices that have
an odd number of neighbors, since every edge contributes two to the
total number of neighbors of all vertices. Thus the origin can’t be the
only vertex of G with an odd number of neighbors, and it follows that
at least one of the other corners of the big rectangle must be an integer
point as well, and we are done. ♡

Laser Gun
You find yourself standing in a large rectangular room with mirrored
walls. At another point in the room is your enemy, brandishing a laser
gun. You and she are fixed points in the room; your only defense is
that you may summon bodyguards (also points) and position them in
the room to absorb the laser rays for you. How many bodyguards do
you need to block all possible shots by the enemy?

Solution: You can certainly protect yourself if you have continuum

many bodyguards, for example, by arranging them in a circle around
yourself with, of course, your enemy outside the circle.
That number can be reduced to a countably infinite number of
bodyguards, because there are actually only countably many direc-
tions your enemy can fire in that will hit you. One way to see this is
to view the room as a rectangle in the plane, with you at P and your
enemy at Q. You can now tile the plane with copies of the room, by
repeatedly reflecting the room about its walls, each copy containing
a new copy of your enemy (see figure below).

358
 Startling Transformation

Every possible accurate shot by the enemy can be represented on

this picture by a straight line from some copy of Q to P ; every time
such a line crosses a boundary between rectangles, the real laser beam
would be bouncing off a wall. In the figure, one such (dotted) line is
indicated; a solid line shows the corresponding path of the laser beam
back in the original room.
Thus, you can pick a point on every shot line and station a guard
there and so protect yourself with countably many guards. But back
in the original room the shot-lines cross each other often; maybe if
you place guards at the intersections, you can get away with a finite
number.
Would you believe 16?
Your object is going to be to intercept every shot at its halfway
point. To do this, you first trace a copy of the above plane tiling, pin
it to the plane at your position P , and shrink the copy by a factor of
two vertically and horizontally. The many copies of Q on the shrunk
copy will be our bodyguard positions; they serve our purpose because
each copy of Q on the original tiling appears halfway between it and
you on the shrunk copy.
In the second figure, the shrunk copy is in gray, and some virtual
laser paths are indicated; you can see that they pass, at their half-way
points, through the corresponding smaller dots in the gray grid.
Of course, there are infinitely many such points, but we claim that
they are all reflections of the right set of 16 points in the original room.
Four of the points will already be in the original room; the four points
in the room to the left of the original room can be reflected back to
give four new points, and similarly for the room above the original
one. Finally, the four points in the room above and to the left of the

359
 Mathematical Puzzles

original room can be reflected twice to provide the last four points in
the original room. In the third figure, the 12 new points (centers filled
in gray) have been added in black to the original rectangle. A virtual
laser path is added, with its corresponding real path crossing one of
the new points.

Since every room looks exactly like the original or one of the other
three we just examined, all the bodyguard points in the plane are re-
flections of the 16 points we have now identified in the original room.
Since every line from a copy of Q passed through a reflected body-
guard, the actual shot hits a “real” bodyguard at its halfway point (if
not sooner) and is absorbed.

360
 Startling Transformation

If the locations of P and Q are carefully chosen, some of the

16 bodyguard locations will coincide; but in general the full 16 are
needed.

Random Intervals
The points 1, 2, . . . , 1000 on the number line are paired up at random,
to form the endpoints of 500 intervals. What is the probability that
among these intervals is one which intersects all the others?

Solution: Amazingly, the answer is exactly 23 , regardless of the

number of intervals (as long as there are at least two).
Suppose the interval endpoints are chosen from {1, 2, . . . , 2n}. We
will label the points A(1), B(1), A(2), B(2), . . . , A(n−2), B(n−2)
recursively as follows. Referring to points {n+1, . . . , 2n} as the right
side and {1, . . . , n} as the left side, we begin by setting A(1) = n and
letting B(1) be its mate. Suppose we have assigned labels up to A(j)
and B(j), where B(j) is on the left side; then A(j +1) is taken as the
left-most point on the right side not yet labeled, and B(j + 1) as its
mate. If B(j) is on the right side, A(j +1) is the right-most unlabeled
point on the left side and again B(j +1) is its mate.
If A(j) < B(j), we say that the jth interval “went right,” other-
wise it “went left.” Points labeled A(·) are said to be inner endpoints,
the others outer.

It is easily checked by induction that after the labels A(j) and B(j)
have been assigned, either an equal number of points have been la-
beled on each side (in case A(j) < B(j)) or two more points have
been labeled on the left (in case A(j) > B(j)).
When the labels A(n−2) and B(n−2) have been assigned, four un-
labeled endpoints remain, say a < b < c < d. Of the three equiprob-
able ways of pairing them up, we claim two of them result in a “big”
interval which intersects all others, and the third does not.

361
 Mathematical Puzzles

In case A(n−2) < B(n−2), we have a and b on the left and c and d
on the right, else only a is on the left. In either case, all inner endpoints
lie between a and c, else one of them would have been labeled. It
follows that the interval [a, c] meets all others, and likewise [a, d], so
unless a is paired with b, we get a big interval.
Suppose, on the other hand, that the pairing is indeed [a, b] and
[c, d]. Neither of these can qualify as a big interval since they do not
intersect each other; suppose some other interval qualifies, say [e, f ],
labeled by A(j) and B(j).
When a and b are on the left, the inner endpoint A(j) lies between
b and c, thus [e, f ] cannot intersect both [a, b] and [c, d], contradicting
our assumption.
In the opposite case, since [e, f ] meets [c, d], f is an outer endpoint
(so f = B(j)) and [e, f ] went right; since the last labeled pair went
left, there is some k > j for which [A(k), B(k)] went left, but [A(k−
1), B(k−1)] went right. Then A(k) < n, but A(k) < A(j) since A(k)
is a later-labeled, left-side inner point. But then [A(j), B(j)] does not,
after all, intersect [B(k), A(k)], and this final contradiction proves the
claim.

Infected Cubes
An infection spreads among the n3 unit cubes of an n × n × n cube,
in the following manner: If a unit cube has three or more infected
neighbors, then it becomes infected itself. (Neighbors are orthogonal
only, so each little cube has at most six neighbors.)
Prove that you can infect the whole big cube starting with just n2
sick unit cubes.
Solution: You might recall the Infected Checkerboard puzzle where
you were asked to show that you couldn’t infect all of an n×n checker-
board if you began with fewer than n sick squares. Essentially the
same proof works in higher dimension to show that you need at least
nd−1 initially sick squares to infect all of a d-dimensional hypercube
of side n (where d infected neighbors infect a unit hypercube).
But this time, it’s not obvious how to choose the initial nd−1 unit
hypercubes (“cells”) to infect.
In what follows we will consider only the 3-dimensional case, but
the construction easily generalizes. Label the cells by vectors (x, y, z),
with x, y, z ∈ {1, 2, . . . , n}, so that two cells will be neighbors if all
their coordinates are the same except in one position, where their val-
ues differ by 1.

362
 Startling Transformation

Start by infecting all cells such that x + y + z ≡ 0 mod n. These

cells form a “diagonal subspace” that is cut up into many pieces. From
these initially sick cells, the infection fills up the big cube in a weird
way. It barely succeeds, relying on many apparent coincidences that
allow the growth to continue. Very different from the 2-dimensional
case! It seems remarkable that the process manages to infect the whole
big cube.
To prove it really works, we transform the puzzle to a game. The
players are you and your adversary—a doctor, say—who is trying to
trap you, the infector. Let us choose a k and start with the infection
as above.
In the game, the doctor begins by putting you at some cell C =
(x, y, z). Now repeat: The doctor chooses a coordinate (first, second,
or third) and you move either forward or backward in that dimension
(if the corresponding coordinate is 1 or n, you won’t have a choice).
You win if you can reach some cell the sum of whose coordinates is a
multiple of n; the doctor wins if she can keep you wandering forever.
We now claim that if you have a winning strategy, then the big
cube will indeed be fully infected.
To see this, we first refine the claim to state that if you can win
starting from cell C, then C itself will become infected. Since from C,
the doctor has three choices of coordinate, a winning strategy for you
must work for all three possibilities. This implies that your strategy
also wins if started at any of the three neighbors of C that you were
preparing to move to. By induction (over the number of steps to a
win), all three of these neighbors of C will be infected, thus C will
as well. The base case of the induction is when the starting point C
has coordinates summing to k modulo n, in which case it is of course
already sick.
Now all we need is to provide you with a winning strategy; For
any cell C = (x, y, z), let c be the number x + y + z + 12 mod n.
After the doctor chooses a coordinate, say the first, you increment x
if it’s less than c (thus c goes down as well, though possibly around
the corner from 12 to n− 12 ) and decrement x if x > c.
If you ever get to a cell where c = 12 , you’ve won the game. So if
x = 1 the algorithm will not ask you to decrement it. If x = n then x
will always be greater than c, so you will not be asked to increment it.
It follows that the move prescribed by the algorithm is always legal,
unless you have already won the game.
We now assert that the doctor can’t force you to cycle. Suppose to
the contrary that starting from C the algorithm cycles forever. Let I

363
 Mathematical Puzzles

be the set of coordinates (among the first, second, and third) chosen
infinitely often by the doctor. We may as well assume that you are
past the point where any index not in I will ever be chosen. Let m be
the biggest value ever encountered at a coordinate in I. Let J be the
set of indices in I which are at the moment equal to m.
If it ever happens that c > m, then you will be incrementing at
every step, pushing c up until it snaps around the corner to 12 and
you win. Therefore it must be that c is always below m. But then,
whenever the doctor chooses a coordinate j ∈ J, the jth coordinate
must decrease to m−1. It follows that eventually J will disappear,
leaving you forever with a smaller maximum value m. This can’t go
on forever, so we have our contradiction.
We conclude that the above algorithm will win the game for you,
no matter where the doctor starts you off or how she chases you
around. The existence of a winning strategy means that the infection
really does capture the whole big cube, and we are done. ♡

Gladiators, Version I
Paula and Victor each manage a team of gladiators. Paula’s gladiators
have strengths p1 , p2 , . . . , pm and Victor’s, v1 , v2 , . . . , vn . Gladi-
ators fight one-on-one to the death, and when a gladiator of strength
x meets a gladiator of strength y, the former wins with probability
x/(x+y), and the latter with probability y/(x+y). Moreover, if the
gladiator of strength x wins, he gains in confidence and inherits his
opponent’s strength, so that his own strength improves to x+y; simi-
larly, if the other gladiator wins, his strength improves from y to x+y.
After each match, Paula puts forward a gladiator (from those on
her team who are still alive), and Victor must choose one of his to face
Paula’s. The winning team is the one which remains with at least one
live player.
What’s Victor’s best strategy? For example, if Paula begins with
her best gladiator, should Victor respond from strength or weakness?

Solution: All strategies for Victor are equally good. To see this,
imagine that strength is money. Paula begins with P = p1 + · · · + pm
dollars and Victor with V = v1 +· · ·+vn dollars. When a gladiator of
strength x beats a gladiator of strength y, the former’s team gains $y
while the latter’s loses $y; the total amount of money always remains
the same. Eventually, either Paula will finish with $P +$V and Victor
with zero, or the other way ’round.

364
 Startling Transformation

The key observation is that every match is a fair game. If Victor

puts up a gladiator of strength x against one of strength y, then his
expected financial gain is
x y
· $y + · (−$x) = $0.
x+y x+y
Thus, the whole tournament is a fair game, and it follows that
Victor’s expected worth at the conclusion is the same as his starting
stake, $V . We then have

q · ($P + $V ) + (1 − q) · $0 = $V,

where q is the probability that Victor wins. Thus, q = V /(P + V ),

independent of anyone’s strategy in the tournament. ♡
Here’s another, more combinatorial proof. Using approximation
by rationals and clearing of denominators, we may assume that all the
strengths are integers. Each gladiator is assigned x balls if his initial
strength is x, and all the balls are put into a uniformly random ver-
tical order. When two gladiators battle, the one whose topmost ball
is highest wins (this happens with the required x/(x+y) probability)
and the loser’s balls accrue to the winner.
The surviving gladiator’s new set of balls is still uniformly ran-
domly distributed in the original vertical order, just as if he had started
with the full set; hence, the outcome of each match is independent of
previous events, as required. But regardless of strategy, Victor will
win if and only if the top ball in the whole order is one of his; this
happens with probability V /(P +V ).

Gladiators, Version II
Again Paula and Victor must face off in the Colosseum, but this time,
confidence is not a factor, and when a gladiator wins, he keeps the
same strength he had before.
As before, prior to each match, Paula chooses her entry first. What
is Victor’s best strategy? Whom should he play if Paula opens with her
best man?

Solution: Obviously, the change in rules makes strategy consider-

ations in this game completely different from the previous one—or
does it? No, again the strategy makes no difference!
For this game, we take away each gladiator’s money (and balls),
and turn him into a light bulb.

365
 Mathematical Puzzles

The mathematician’s ideal light bulb has the following property:

Its burnout time is completely memoryless. That means that knowing
how long the bulb has been burning tells us absolutely nothing about
how long it will continue to burn.
You may know that the unique probability distribution with this
property is the exponential; if the expected (average) lifetime of the
bulb is x, then the probability that it is still burning at time t is e−t/x .
However, no formula is necessary for this puzzle. You only need to
know that a memoryless probability distribution exists.
Given two bulbs of expected lifetimes x and y, respectively, the
probability that the first outlasts the second is x/(x+ y). To see this
without calculus, consider a light fixture that uses one “x-bulb” (with
average lifetime x) and one “y-bulb”; every time a bulb burns out, we
replace it with another of the same type. When a bulb does burn out,
the probability that it is the y-bulb is a constant independent of the
past. But that constant must be x/(x+y), because over a long period
of time, we will use y-bulbs and x-bulbs in proportion x:y.
Back in the Colosseum, we imagine that the matching of two glad-
iators corresponds to turning on their corresponding light bulbs until
one (the loser) burns out, then turning off the winner until its next
match; since the distribution is memoryless, the winner’s strength in
its next match is unchanged. Substituting light bulbs for the gladiators
may be less than satisfactory for the spectators, but it’s a valid model
for the fighting.
During the tournament, Paula and Victor each have exactly one
light bulb lit at any given time; the winner is the one whose total light-
ing time (of all the bulbs/gladiators on her/his team) is the larger.
Since this has nothing to do with the order in which the bulbs are
lit, the probability that Victor wins is independent of strategy. (Note:
That probability is a more complex function of the gladiator strengths
than in the previous game.)

More Magnetic Dollars

We return to Magnetic Dollars, but we strengthen their attractive power
just a bit.
This time, an infinite sequence of coins will be tossed into the two
urns. When one urn contains x coins and the other y, the next coin
will fall into the first urn with probability x1.01 /(x1.01 + y 1.01 ), other-
wise into the second urn.

366
 Startling Transformation

Prove that after some point, one of the urns will never get another
coin!
Solution: The really neat way to show that one urn gets all but
finitely many coins is to employ those continuous memoryless wait-
ing times that proved so useful in the previous puzzle.
Look just at the first urn and suppose that it acquires coins by
waiting an average of 1/n1.01 hours between its nth coin and its
(n+1)st coin, where the waiting time is memoryless. Coins will ar-
rive slowly
∑∞ and sporadically at first, then faster and faster; since the
series n=1 1/n1.01 converges, the urn will explode with infinitely
many coins at some random moment (typically about 4 days after the
process begins).
Now suppose we start two such processes simultaneously, one
with each urn. If at some time t, there are x coins in the first urn
and y in the second, then (as we saw with the gladiator-light bulbs)
the probability that the next coin goes to the first urn is

1/y 1.01 x1.01

= ,
1/x1.01 + 1/y 1.01 x1.01 + y 1.01

exactly what it should be. Nor does it matter how long it’s been since
the xth coin in the first urn (or yth in the second) arrived, since the
process is memoryless. It follows that this accelerated experiment is
faithful to the puzzle.
However, you can see what happens now; with probability 1, the
two explosion times are different. (For this, you only need to know
that our memoryless waiting time has a continuous distribution.) But
the experiment ends at the first explosion, at which time the other urn
is stuck with whatever finite number of coins it had. ♡
Seems like kind of a scary experiment, doesn’t it? The slow urn
never got to finish because, in effect, time ended.

For our theorem, we look at a combinatorial object that arises with

surprising frequency called a “parking function.”
Imagine that there are n parking spaces 1, 2, . . . , n (in that order)
marked on a one-way street. Cars numbered 1 through n enter the
street in some order, each with a preferred parking spot (a random
number between 1 and n). Each car drives to its preferred spot and
parks there if the spot is available; otherwise, it takes the next available
spot. If there are no empty spots beyond the filled preferred spot, the
car must leave the street without parking; tough luck!

367
 Mathematical Puzzles

We want to know: What is the probability that all n cars will be

able to park?
Theorem. The order in which the cars arrive makes no difference, that is,
the issue of whether all cars find a spot depends only on what the cars’ pre-
ferred parking spots are. If each car’s preferred spot is independently chosen
and uniformly random, the probability of success is (n+1)n−1 /nn .

The total number of ways to assign preferred spots to the n cars

is nn ; we need to figure out how many of these will lead to everyone
finding a spot. For example, if everyone prefers a different spot, all will
be well. Or if everyone prefers slot #1. But not if everyone prefers slot
#2. In general, we’re in trouble if there’s some k for which fewer than
k cars want slots from 1 through k; with some thought, you’ll be able
to see that if there is no such k, everyone will get parked, no matter
how the cars are ordered.
Another way to say the same thing: If we order the preferred spots
xi so that x1 ≤ x2 ≤ · · · ≤ xn , then everyone gets parked (in which
case the mapping from i to xi is called a parking function) provided
xi ≤ i for each i. (If the spot i ends up left open, then there must
have been fewer than i cars wanting a spot between 1 and i.) This is a
pretty natural mathematical condition; it would be nice to know the
probability that it holds.
But it turns out we don’t need to know the condition to prove the
theorem.
We transform the situation by adding an n+1st slot and imagining
that the slots now close up to form a circle. We’ll let each car chose
a preferred slot from 1 all the way to n+1; each car in turn goes to
its preferred slot. Then, as before, if the slot is occupied it continues
(clockwise, say) around the circle and parks in the first available slot.
In this scenario every car will find a slot, and there will be one
empty slot left over. If that empty slot happens to be slot n+1, the list
of preferences would have worked in the original problem; otherwise,
it would have failed.

368
 Startling Transformation

Now we take advantage of the circle’s symmetry: We notice that

if all the car’s preferences were advanced by k, modulo n+1, then
the whole processes would simply take place k slots clockwise from
where it was; consequently, the final empty slot would also advance
by k. But that implies that for every assignment of preferred slots,
exactly 1 of its n+1 rotations works.
We conclude that the probability that the final empty slot is slot
n+1 is 1/(n+1). The total number of ways to assign preferences in our
circular case is (n+1)n , hence the total number of good assignments
is (n+1)n /(n+1) = (n+1)n−1 . Note that, necessarily, none of those
good assignments has any car preferring slot n+1.
We conclude that the probability of success in the original problem
is (n+1)n−1 /nn . ♡
How big is this number? For large n, we can write
1 n
(n+1)n−1 /nn = (1 + ) /n ∼ e/n.
n
For example, if there are 100 cars, the theorem pegs the probability
of success at about 2.678033494%, close to e/100 ∼ 2.718281828%.
One might say that it’s a pleasant surprise, but not shockingly lucky,
if all the cars find parking spots.

369
 Notes and Sources
1. Out for the Count
Half Grown: Sent by Jeff Steif, of Chalmers University in Sweden.
Powers of Two: Classic.
Watermelons: Adapted from Gary Antonick’s New York Times “Number Play” blog,
11/7/12; but the puzzle appeared earlier in Kvant, June/August 1998, contributed
by I. F. Sharygin.
Bags of Marbles: Contributed by pathologist Dick Plotz, of Providence RI.
Salaries and Raises: Idea from Rouse Ball [9].
Efficient Pizza-Cutting: Subject of a MoMath (National Museum of Mathematics)
event of 10/30/19, led by Paul Zeitz.
Attention Paraskevidekatriaphobes: As far as I can tell this remarkable fact was first
observed by Bancroft Brown (a Dartmouth math professor, like your author), who
published his calculation in the American Mathematical Monthly Vol. 40 (1933), p.
607. My present-day colleague Dana Williams is the one who brought all this to
my attention.
The origin of superstition concerning Friday the 13th is usually traced to the date of an
order given by King Philip IV of France (Philip the Fair), dismantling the Knights
Templar.
Meet the Williams Sisters: See also Hess [62].
Rating the Horses: Suggested by Saul Rosenthal of TwoSigma.
Shoelaces at the Airport: Told to me by Dick Hess at IPP35 (the 35th Annual Inter-
national Puzzle Party).
Rows and Columns: This is a classical theorem, simple and surprising; I was reminded
of it by Dan Romik, a math professor at the University of California, Davis. Don-
ald Knuth, in Vol. III of The Art of Computer Programming, traces the result to a
footnote in a 1955 book by Hermann Boerner. Bridget Tenner, a student of famed
combinatorialist Richard Stanley, wrote a paper called “A Non-Messing-Up Phe-
nomenon for Posets” generalizing this theorem.
Three-way Election: Dreamed up by Ehud Friedgut of the Weizmann Institute, Israel.
Sequencing the Digits: Remembered from the Putnam Exam, many years ago.
Faulty Combination Lock: Suggested to me by Amit Chakrabarti of Dartmouth; it
had been proposed for the International Mathematical Olympiad in 1988 by East
Germany. The optimality proof given (supplied by Amit) doesn’t quite work for 8
replaced by 10, and fails increasingly thereafter. Thus the general puzzle, when the
lock has n numbers per dial, is unsolved as far as I know.
Losing at Dice: I discovered this curious fact accidently, 40 or so years ago, while
constructing homework problems for an elementary probability course at Emory
University.
Splitting the Stacks: From Einar Steingrimsson, University of Strathclyde, Glasgow.
North by Northwest: Original.
Early Commuter: From the one and only Martin Gardner.

371
 Mathematical Puzzles

Alternate Connection: From the First Iberoamerican Mathematical Olympiad [7].

Bindweed and Honeysuckle: From Karthik Tadinada, but recast by me in recollection
of a delightful tragicomic song by Michael Flanders & Donald Swann.
theorem: The sequence we associated with each tree is known as the Prüfer code [92] for
the tree. Besides helping us prove the theorem, the Prüfer code gives us additional
information. For example, you can easily check that the number of neighbors of
a vertex of the tree is always one more than the number of times it appears in
the code; leaves do not appear at all. From this we could deduce that when n is
large, the number of leaves in a random tree is about n/e, that is, about 37% of the
vertices are leaves.

2. Achieving Parity
Bacterial Reproduction: Based on Supplementary Problem C/4, [68], p. 14.
Fourth Corner: Brought to my attention by Paul Zeitz.
Unanimous Hats: My version of a classic.
Half-right Hats: My version of a classic.
Red and Blue Hats in a Line: Passed on to me by Girija Narlikar of Bell Labs, who
heard it at a party. Note that an analogue of the given solution works with any
number k of hat colors. The colors are assigned numbers from 0 to k−1, say, and
the last prisoner in line guesses the color corresponding to the sum, modulo k, of
the color-numbers of the hats in front of him. As before, all other prisoners are
saved, if they’re careful with their modular arithmetic.
Prisoners and Gloves: Suggested by game theorist Sergiu Hart of the Hebrew Univer-
sity.
Even-sum Billiards: Inspired by a problem from John Urschel’s column,
https://www.theplayerstribune.com/author/jurschel.
Chameleons: Boris Schein, an algebraist at the University of Arkansas, sent me this
puzzle; it may be quite old. On one occasion it was given to an 8th-grader in
Kharkov, on another to a young Harvard grad interviewing at a big finance firm;
both solved it.
Missing Digit: Lifted from Elwyn Berlekamp and Joe Buhler’s Puzzles Column
in Emissary, Spring/Fall 2006; they heard it from the number theorist Hendrik
Lenstra.
Subtracting Around the Corner: When I was in high school, a substitute math teacher
told my class that some WWII prisoner of war entertained himself by trying various
sequences of four numbers to see how long he could get them to survive under
the above operations. The process (with four numbers) is sometimes known as a
difference box or diffy box. Among real numbers, there is an (essentially) unique
quad that never terminates [12]. That quad is (0, 1, q(q − 1), q), where q is the
unique real root of q 3 − q 2 − q − 1 = 0, namely,
( √ √ )
1 3 √ 3 √
q= 1 + 19 + 3 33 + 19 − 3 33 ∼ 1.8393,
3
so q(q − 1) ∼ 1.5437. Remarkably, this number arises also in the solution to
another, seemingly unrelated, puzzle in this volume.
Uniting the Loops: This is one of a number of ingenious puzzles devised by
writer/mathematician Barry Cipra, of Northfield Minnesota.
theorem: Conjectured in [67] and proved at Danny Kleitman’s 65th Birthday conference
in August 1999 by Noga Alon, Tom Bohman, Ron Holzman and Danny himself
[2].

372
 Notes and Sources

3. Intermediate Math
Squaring the Mountain State: My version of a classic.
Monk on a Mountain: Ancient.
Cutting the Necklace: Suggested by Pablo Soberón of Baruch College, City University
of New York. For more information see [4].
Three Sticks: Moscow Mathematical Olympiad 2000, Problem D4, contributed by A.
V. Shapovalov [40].
Hazards of Electronic Coinflipping: See also Fair Play
√√in Chapter 19, Hammer and
Tongs. The actual value of p works out to 12 ± 2
3
− 34 . But there’s no need
to limit yourself to three people; you can choose uniformly among n people, for
any positive integer n, in a bounded number of flips. Johan Wästlund suggests
the following method. Suppose you have enough power for f flips. You allot as
many of the 2f outcomes as can be divided into n−1 equivalent sets; left over are
at most n−2 of each level, that is, for each k at most n−2 outcomes involving
exactly k heads. Now the IVT can be applied as in the n = 3 case as long as the
total probability of the left-out outcomes, at p = 12 , is at most n 1
. The latter is
achieved when f is just over 2 log2 (n) flips—only a factor of two away from the
information-theoretic lower bound.
Bugs on a Pyramid: Devised by Dr. Kasra Alishahi, of Sharif University; sent to me
by Mahdi Saffari.
Skipping a Number: Based on the Putnam exam problem (#A1, from 2004) that in-
troduced the now-famous (in certain circles) Shanille O’Keal.
Splitting a Polygon: From a Moscow Mathematical Olympiad in the 1990’s.
Garnering Fruit: Sent to me by Arseniy Akopyan, author of the unique and delightful
book Geometry in Figures. The first of the two solutions to the two-fruit version came
from Pablo Soberón of Baruch College, City University of New York.
theorem: Known since ancient times.

4. Graphography
Air Routes in Aerostan: Moscow Mathematical Olympiad 2003, Problem A5, con-
tributed by R. M. Fedorov [40].
Spiders on a Cube: Passed to me by Matt Baker, Georgia Tech.
Handshakes at a Party: Classic.
Snake Game: From the 12th All Soviet Union Mathematical Competition, Tashkent,
1978.
Bracing the Grid: Passed to me by geometer Bob Connelly of Cornell, based on [29].
Competing for Programmers: Moscow Mathematical Olympiad 2011.
Wires under the Hudson: This is a variation of a puzzle publicized by Martin Gard-
ner, and sometimes called the Graham-Knowlton Problem. To electricians it is the
“WIP” (wire identification problem). In Gardner’s version, you can tie any num-
ber of wires together at either end, and test at either end, as well. Our solution,
found in [99, 57], satisfies our additional constraints and involves only two oper-
ations at each end (thus three river crossings, not counting additional crossings to
untie and perhaps actually use the wires). The solution is not unique, however, so
even if your three-crossing solution is different, it may be just as good.
Two Monks on a Mountain: Brought to my attention by Yuliy Baryshnikov of the
University of Illinois at Urbana-Champaign.
Worst Route: Adapted from a puzzle found in [84]. If the number of houses were
odd, distances would matter slightly: The postman would then begin or end at the

373
 Mathematical Puzzles

middle house, zigzag between high-numbered and low-numbered houses, and end
or begin at the closest house to the middle house. Example: If the addresses were
1,2,4,8,16 one of the maximum-distance routes would be 4,8,1,16,2.
theorem: The idea of switching odd and even edges used in the proof is known as the
“alternating path method” and is an important tool in graph theory.

5. Algebra Too
Bat and Ball: An oldie, as you can see from the prices!
Two Runners: From [62].
Belt Around the Earth: Origin unknown, but Enrique Treviño of Lake Forest College
has informed me that a version appears in the notes to a 1702 translation of Euclid
by William Whiston.
Odd Run of Heads: A variation of this puzzle was suggested, but not used, for an IMO
in the early ’80s (see [68]).
Hopping and Skipping: Devised by mathematician James B. Shearer of IBM, this puz-
zle appeared in the April 2007 edition of IBM’s puzzle site “Ponder This,”
http://domino.research.ibm.com/Comm/wwwr\_ponder.nsf/challenges/
April2007.html.
Area–Perimeter Match: Problem 176, p. 71 of [75].
Three Negatives: Due to Mark Kantrowitz of Carnegie Mellon University.
Red and Black Sides: Moscow Mathematical Olympiad 1998, Problem C2 [39], con-
tributed by V. V. Proizvolov. A nice non-algebraic approach is instead to project
the red lines onto, say, the left edge of the square. If the projection covers the whole
left edge, we’re done; suppose it misses an interval I. Then every rectangle that in-
tersects the horizontal band through I has a red horizontal side, so projecting these
downward covers the bottom edge of the square.
Recovering the Polynomial: Sent to me by Joe Buhler of Reed College, who believes it
may be quite old (but, maybe not as old as Delphi). Note that without the restriction
that the Oracle must be fed integers, the polynomial can be determined in one step
with (say) x = π. Of course, the Oracle would have to find a way to convey p(π)
in a finite amount of time; if she gives a decimal expansion digit by digit, there’d
be no way to know when to cut her off.
Gaming the Quilt: Sent to me by Howard Karloff of Goldman Sachs.
Strength of Schedule: Moscow Mathematical Olympiad 1997, Problem C5 [39], con-
tributed by B. R. Frenkin.
Two Round-robins: Moscow Mathematical Olympiad 1997, Problem B5 [39], con-
tributed by B. R. Frenkin.
Alternative Dice: This problem is famous enough so that its solution has a name:
“Sicherman dice.” Martin Gardner wrote in Scientific American magazine, in 1978,
about their discovery by one Colonel George Sicherman, of Buffalo NY.
theorem: The model described, often now called the “Galton–Watson tree,” had already
been studied independently by I. J. Bienaymé.

6. Safety in Numbers
Broken ATM: Moscow Mathematical Olympiad 1999, Problem A4 [39]. Hidden in
this puzzle is a form of the Chinese Remainder Theorem.
Subsets with Constraints: The first part of the puzzle was presented by the late, long-
time puzzle maven Sol Golomb (University of Southern California) at the Seventh

374
 Notes and Sources

Gathering for Gardner; the second part was suggested by Prasad Tetali of Georgia
Tech; the third, an obvious variation.
Cards from Their Sum: The trick is a version of Colm Mulcahy’s “Little Fibs.” The
two solutions are described in the Coda of Mulcahy’s book [87], and were supplied
by combinatorialist Neil Calkin of Clemson University.
Divisibility Game: Moscow Mathematical Olympiad 2003, Problem A6 [40], con-
tributed by A. S. Chebotarev.
Prime Test: Moscow Mathematical Olympiad 1998, Problem B1 [39], contributed by
A. K. Kovaldzhi.
Numbers on Foreheads: Sent to me by Noga Alon of Princeton University.
Rectangles Tall and Wide: There are other proofs; see, for example,
http://www-math.mit.edu/~rstan/transparencies/tilings3.pdf.
Locker Doors: Classic.
Factorial Coincidence: From [96], sent to me by Christof Schmalenbach of IBM.
Even Split: This puzzle, with n replaced by 100, appeared in the 4th All Soviet Union
Mathematical Competition, Simferopol, 1970. It is elegant enough to be called a
theorem and in fact it is a theorem in [37].
Factorials and Squares: Both the puzzle and the solution given are due to Jeremy Kun,
author of “A Programmer’s Introduction to Mathematics.” Passed to me by Diana
White of the National Association of Math Circles. (There are other equally valid
ways to get the answer.)
theorem: This probably goes back to ancient Greece, but proofs of this fundamental fact
are still of interest in the 21st century; see, for example, [15].

7. The Law of Small Numbers

Domino Task: Moscow Mathematical Olympiad 2004, Problem A6 [40], contributed
by A. V. Shapovalov.
Spinning Switches: This puzzle reached me via Sasha Barg of the University of Mary-
land, but seems to be known in many places. Although no fixed number of steps
can guarantee turning the bulb on in the three-switch version, a smart randomized
algorithm can get the bulb on in at most 5 57 steps on average, against any strategy
by an adversary who sets the initial configuration and turns the platform.
Candles on a Cake: Moscow Mathematical Olympiad 2003, Problem D4 [40], con-
tributed by G. R. Chelnokov. For a somewhat similar idea, see Gasoline Crisis be-
low.
Lost Boarding Pass: I first heard the puzzle at the Fifth Gathering for Gardner; the ver-
sion here, from probabilist Ander Holroyd. Later the puzzle’s popularity zoomed
when it appeared on Car Talk.
Flying Saucers: My version of a puzzle suggested by Sergiu Hart.
Gasoline Crisis: This puzzle has been around for a long time and can be found, for
example, in [79].
Coins on the Table: This nice puzzle came to me by way of computer scientist Guy
Kindler, during a marvelous visiting year by each of us at the Institute for Advanced
Study in Princeton.
Coins in a Row: Passed to me by mathematician Ehud Friedgut, this puzzle was al-
leged to have been used by a high-tech company in Israel to test job candidates.
Powers of Roots: Idea from 29th Annual Virginia Tech Mathematics Contest, 2007,
and from Using Your Head is Permitted, brand.site.co.il/riddles, October 2008.
Coconut Classic: This is the “Williams version” of an ancient classic, where the mon-
key gets a sixth coconut in the morning. With the original version, you need the

375
 Mathematical Puzzles

ending number to be 1 mod 5, achieved by beginning with 56 − 4 coconuts in-

stead of 55 − 4. That ends with 5 × 45 − 4 = 5116 coconuts, giving each man
1023 and the monkey 1.
theorem: Known now as a “1-factorization of the complete graph on an even number
of vertices,” such constructions have been around at least since 1859 [94].

8. Weighs and Means

Tipping the Scales: The Second All Soviet Union Mathematical Competition,
Leningrad 1968.
Men with Sisters: Classic.
Watches on the Table: The 10th All Soviet Union Mathematical Competition,
Dushanbe, 1976.
Raising Art Value: Original.
Waiting for Heads: Original, more or less.
Finding a Jack: Idea from Dexter Senft of Lehman Bros; appears as Problem 40 in
Fred Mosteller’s delightful book [86].
Sums of Two Squares: Ross Honsberger [64].
Increasing Routes: This puzzle, and its elegant solution, were passed to me by Ehud
Friedgut, of Hebrew University. With his help the puzzle was traced back to the
fourth problem in the second round of the “Bundeswettbewerb Mathematik 1994.”
The Bundeswettbewerb is one of two big old German mathematical competitions.
theorem: Shockingly little-known, this result is recommended for inclusion in the next
new graph theory text!

9. The Power of Negative Thinking

Chomp: Due to the late David Gale of U.C. Berkeley (although a roughly equiva-
lent number-theoretic game had been studied earlier by the Dutch mathematician
Frederik Schuh).
Dots and Boxes Variation: From Dartmouth professor Sergi Elizalde.
Big Pairs in a Matrix: 2011 Moscow Olympiad; from MIT puzzle maven Tanya Kho-
vanova’s blog, http://www.tanyakhovanova.com.
Bugging a Disk: This puzzle was published by CMU math professor Alan Frieze on
CMU’s puzzle website http://www.cs.cmu.edu/puzzle/solution26.pdf; the
proof above was suggested by George Wang and Leo Zhang.
Pairs at Maximum Distance: From the Putnam Exam of 1957.
Lemming on a Chessboard: Devised by Kevin Purbhoo when he was a high school
student at Northern Secondary School in Toronto. Purbhoo is now on the faculty
at the University of Waterloo.
Curve on a Sphere: Passed to me by physicist Senya Shlosman, who heard it from Alex
Krasnoshel’skii. The solution given is Senya’s. Omer Angel, of the University of
British Columbia, has a different proof that is less elementary, but still elegant and
educational.
Let C be our closed curve and C ˆ its convex closure in 3-space, that is, the smallest
convex set containing C. If C is not contained in a hemisphere, then Ĉ contains
the origin 0; otherwise, there would be a plane through 0 cutting off Ĉ from 0.
Thus, by Carathéodory’s Theorem, there is a set of four points on C some convex
combination of which is 0. Putting it another way, the tetrahedron whose vertices
are those four points contains the origin.

376
 Notes and Sources

Now, move the points continuously toward one another along the curve. When the
points merge their tetrahedron will no longer contain the origin, so somewhere
during the process, there was a time when the origin lay on a face of the tetrahedron.
The three points that determine that face lie on a great circle, and each pair has a
shortest route along that circle not containing the third point. Hence, the sum of
the pairwise distances of the three points is 2π, impossible since they all lie on C.
Soldiers in a Field: From the 6th All Soviet Union Mathematical Competition in
Voronezh, 1966.
Alternating Powers: Spotted in Emissary, Fall 2004; see [61]. The proof is from Noam
Elkies, Harvard mathematician and (musical) composer.
Halfway Points: Moscow Mathematical Olympiad 1998, Problem B6 [39], contributed
by V. V. Proizvolov.
theorem: The statement was originally conjectured by James Joseph Sylvester (arguably,
the first great mathematician to leave Europe for America) back in 1893. It was first
proved by the Hungarian mathematician Tibor Gallai, but the elegant proof given
was found by L. M. Kelly and published in 1948 [28].

10. In All Probability

Winning at Wimbledon: Suggested by Dick Hess.
Birthday Match: Problem 32 of [86]. You may have wondered, since the answer 253
just barely works, whether accounting for the possibility that one or more of your
shipmates might have been born on February 29 could push the answer up to 254;
in fact, it does. (The answer, 23, to the better-known version of the puzzle is unaf-
fected.) If you yourself were born on February 29, you’d need to query a daunting
1013 people to be a favorite to find a birthmate.
Other Side of the Coin: A similar calculation shows that you should switch doors in the
notorious “Monty Hall” problem, provided you assume Monty chooses randomly
which door to open when two are available.
Boy Born on Tuesday: Created by Thomas Starbird of JPL and unleashed by puzzle
designer Gary Foshee at the Ninth Gathering for Gardner.
Whose Bullet?: Original.
Second Ace: Suggested by a puzzle in Littlewood’s Miscellany [82].
Stopping After the Boy: Classic, promulgated by Martin Gardner.
Points on a Circle: Suggested by combinatorics legend Richard Stanley, of the Uni-
versity of Miami. For k points instead of 3, similar reasoning will lead to proba-
bility 2k/2k of being containable in a semicircle. One can also do this in higher
dimension; for example, the probability that four random points on a sphere can
be contained in a hemisphere turns out to be 1/4.
Comparing Numbers, Version I: To the best of my knowledge, this problem originated
the late, supremely creative Tom Cover of Stanford University [26].
Comparing Numbers, Version II: Source unknown.
Biased Betting: Original.
Home-field Advantage: Original. For a faster way to do some of the computations, see
[78].
Service Options: Adapted from independent suggestions by authors Hirokazu Iwa-
sawa (Iwahiro) and Dick Hess, and Cambridge University’s Geoffrey Grimmett.
Who Won the Series?: Heard from Pradeep Mutalik, mathematics writer for the online
magazine QUANTA, at the Eleventh Gathering for Gardner. The solution given
is my own.
Dishwashing Game: Source unknown.

377
 Mathematical Puzzles

Random Judge: Problem 3 of [86].

Wins in a Row: Problem 2 of [86], and popularized by Martin Gardner in his “Math-
ematical Games” column in Scientific American. Gardner gave a (correct) algebraic
proof of the answer (“no”) but acknowledging the value of a proof by reasoning,
he also provided two arguments that you’d be better off playing black first: (1) You
must win the crucial middle game, thus you want to be playing white second; and
(2) You must win a game as black, so you’re better off with two chances to do so.
In fact, neither argument is convincing, and even together they are not a proof.
Split Games: Original (although my Dartmouth colleague Peter Doyle came up with
something similar, independently).
Angry Baseball: Suggested, at the author’s Charles River Probability Lecture entitled
“Probability in Your Head,” by Po-Shen Loh. Po-Shen is on the faculty at Carnegie
Mellon and is the current coach of the U.S. Mathematical Olympic Team.
Two-point Conversion: Adapted from a problem in John Urschel’s column,
https://www.theplayerstribune.com/author/jurschel. John is a former
starting guard with the NFL’s Baltimore Ravens, now a PhD student in mathe-
matics at MIT.
Random Chord: A classic paradox that I learned as a teenager from Martin Gardner’s
column.
Random Bias: Brought to me by Paul Zeitz of the University of San Francisco. Paul
is the founder of San Francisco’s Proof School, and a frequent contributor at the
National Museum of Mathematics.
Coin Testing: Original.
Coin Game: This game is often called “Penney Ante,” a pun on the name of its inventor
[90]. Conway’s formula appears in [52]; see also [41, 51]. The quantity A·B is
sometimes called the “correlation” of A and B, but I avoided that terminology
as it suggests that A·B = B·A which is not generally the case.
Sleeping Beauty: For Arntzenius and Dorr’s argument see [8, 33]; more arguments
and many references can be found in my article [105].
Leading All the Way and theorem: Bertrand’s original paper [16] was followed by a
nice proof by Désiré André [8].

11. Working for the System

No Twins Today: Classic.
Natives in a Circle: Moscow Mathematical Olympiad 1998, Problem B3 [39], con-
tributed by B. R. Frenkin. Note that in fact anytime the proportion of liars is 1/2,
the anthropologist will be able to deduce this from their answers.
Poker Quickie: Sent to me by puzzle maven Stan Wagon of Macalester College, who
found it in [47].
Fewest Slopes: Moscow Mathematical Olympiad 1993, Problem D3 [39], contributed
by A. V. Andjans.
Two Different Distances: Goes back at least to Einhorn and Schoenberg [35].
First Odd Number: The prime version was brought to me by the late Herb Wilf, of the
University of Pennsylvania. I’ve heard it attributed to Don Knuth.
Measuring with Fuses: This and other fuse puzzles seem to have spread like wild-
fire some years ago. Recreational mathematics expert Dick Hess has put together
a miniature volume called Shoelace Clock Puzzles devoted to them; he first heard
the one above from Carl Morris of Harvard University. (Hess considers multiple
fuses—shoelaces, for him—of various lengths, but lights them only at ends.)

378
 Notes and Sources

King’s Salary: This puzzle was devised by Johan Wästlund of Chalmers University,
(loosely!) inspired by historical events in Sweden.
Packing Slashes: Heard from Vladimir Chernov of Dartmouth. Lyle Ramshaw, a re-
searcher at Hewlett-Packard, has devised schemes for the general n × n case, but
as of this writing the optimal solutions for odd n are not known to me. See [18].
Unbroken Lines: Devised by mathematics writer Barry Cipra, this one is directly in-
spired by a work of the late Sol LeWitt.
Unbroken Curves: Again by Cipra, but with less direct connection to LeWitt.
Conway’s Immobilizer: I don’t know the composer of this puzzle but the name came
from the late John Horton Conway’s claim that it immobilized one solver in his
chair for six hours. The first solution given here was suggested by my former PhD
student Ewa Infeld; the second was devised by Takashi Chiba in response to a
puzzle column in Japan, and sent to me by Ko Sakai of the Graduate School of
Pure and Applied Sciences, University of Tsukuba. See [24] for yet more solutions.
Seven Cities of Gold: Adapted from a suggestion by Frank Morgan of Williams Col-
lege.
theorem: The given uniqueness proof is the author’s, but there are many others. The
Moser Spindle (discovered by brothers William and Leo Moser in 1961) cannot be
vertex-colored with three colors without having adjacent vertices of the same color.
As a consequence, the “unit-distance graph,” whose vertices are all points on the
plane with two adjacent when they are distance 1 apart, has chromatic number at
least 4. This lower bound stood until 2018 when computer scientist and biologist
Aubrey de Grey found a 1581-vertex graph that raised the lower bound to 5. The
upper bound remains at 7.

12. The Pigeonhole Principle

Shoes, Socks & Gloves: Original.
Polyhedron Faces: 1973 Moscow Mathematical Olympiad, but with a solution that
used Euler’s formula.
Lines Through a Grid: Moscow Mathematical Olympiad 1996, Problem C2 [39], con-
tributed by A. V. Shapovalov.
Same Sum Subsets: Based on a problem from the 1972 International Mathematical
Olympiad. There doesn’t seem to be any easy way to actually find two disjoint sub-
sets of Brad’s numbers with the same sum; the number of pairs of disjoint nonempty
subsets of a 10-element set is a daunting (310 − 2 · 210 + 1)/2 = 28,501. In-
deed, given n numbers (not necessarily distinct), even the problem of determining
whether you can divide all of them into two disjoint sets with the same sum is no-
toriously difficult; it’s one of the “original” NP-complete problems from around
1970. That means there’s probably no way to decide this efficiently. Yet, the pi-
geonhole principle tells us that if the n numbers are distinct integers between 1
and 2n /n, and disjoint subsets (not necessarily a partition) with the same sum are
required, then the answer to the existence question is easy—it’s always “yes”!
Lattice Points and Line Segments: Source unknown; best guess, some Moscow Math-
ematical Olympiad.
Adding, Multiplying, and Grouping: Moscow Mathematical Olympiad 2000, Prob-
lem B2 [40], contributed by S. A. Shestakov.
Line Up by Height: The theorem first appeared in [38].
Ascending and Descending: Moscow Mathematical Olympiad 1998, Problem C6 [39],
contributed by A. Ya. Kanel-Belov and V. N. Latyshev.
Zeroes and Ones: The final idea was passed to me by the late David Gale.

379
 Mathematical Puzzles

Same Sum Dice: Brought to me by David Kempe of the University of Southern Cali-
fornia. Similar results can be found in a paper by the very notable mathematicians
Persi Diaconis, Ron Graham and Bernd Sturmfels [31]. Pointed out by Greg War-
rington of the University of Vermont: The puzzle (and proof) still works with m
n-sided dice versus n m-sided dice.
Zero-sum Vectors: Sent to me by Noga Alon of Princeton University. The problem
appeared as D6 in the 1996 Moscow Mathematical Olympiad [39]. The puzzle’s
statement is “tight” in at least two senses. First, if any of the 2n {−1, +1}-vectors
is left off the original list, the result fails—even if the others are represented ar-
bitrarily often! Suppose, for example, that n = 5 and that the missing vector is
y = ⟨+1, +1, +1, −1, −1⟩. In all other vectors, change all +1’s among the first
three coordinates to zeroes, and all −1’s among the last two coordinates to zero.
Since y is missing, the zero vector will not appear among the altered vectors; and
any sum of altered vectors will have a negative number among its first three coor-
dinates, or a positive number among its last two, or both.
Second, as noted by mathematician and hacker Bill Gosper, for any n there is a way
to alter the vectors in such a way that there is no way to get a zero sum short of
adding up the entire list. We leave it to the reader to verify this fact.
theorem: In fact the numbers {nr}, for r irrational, are not only dense but remarkably
(and usefully) evenly spaced: Roughly speaking, if you run n through the integers
from 1 to any large number, the fraction of the values {nr} that appear in some
subinterval of [0, 1) will be close to the length of that subinterval. This observation
is fundamental in a branch of mathematics sometimes called “discrepancy theory,”
for which a wonderful source is [11].

13. Information, Please

Finding the Counterfeit: Classical; suggested by Robert DeDomenico. David Gontier
found a solution in a 1946 article [34] by the late, great physicist Freeman Dyson.
When I distributed the puzzle for the National Museum of Mathematics’ “Mindben-
ders for the Quarantined,” subscriber Markus Schmidmeier pointed out that there
are non-adaptive solutions such as (with numbered coins) 1,2,3,4 versus 5,6,7,8;
1,2,9,10 versus 3,4,8,11; and 1,3,5,9 versus 2,6,10,12. Another subscriber, Bob
Henderson, pointed out that this solution allows you to diagnose also the case
of no counterfeit—in which case all weighings balance.
Three Natives at the Crossroads: Puzzles of this sort were studied and popularized
by Martin Gardner and Raymond Smullyan, among others; this particular ver-
sion came to me from two mathematical physicists, Vladas Sidoravicius and Senya
Shlosman.
Attic Lamp: Classical, but losing ground as incandescent bulbs disappear even from
memory. By the way, with enough patience you can extend the solution to five
switches: Flip switch E on a year or two in advance, to acquire a fifth possible
state, “burnt out”!
Players and Winners: Devised and communicated to me by Alon Orlitsky, of the Uni-
versity of California, San Diego.
Missing Card: The trick first appeared in Wallace Lee’s book Math Miracles, in which
he credits its invention to William Fitch Cheney, Jr., a.k.a. “Fitch,” around 1950.
See also [69, 87].
Peek Advantage: Brought to my attention by professional gambler Jeff Norman, who
was offered the chance to bet a lot more than $100 on the “color of the flop” prior
to a deal of Texas hold ’em, with the extra inducement of being allowed to peek at
one of his own hole cards before deciding which color to bet on.

380
 Notes and Sources

Bias Test: Original. The theorem mentioned in the text was proved by Gheorghe
Zbǎganu [22, 109].
Dot-Town Exodus: Classical. Steve Babbage, a manager and cryptographer with
Vodafone, points out that if the residents of Dot-town begin to worry that a
suicide was not caused by knowing one’s dot color—but perhaps because some
Dot-towner “has finally cracked under the strain of living in such a ludicrous
environment”—then under certain circumstances the rest of the town may yet sur-
vive the stranger’s incursion.
Conversation on a Bus: A creation of John H. Conway’s. Variations can be found in a
paper of Tanya Khovanova’s posted at https://arxiv.org/pdf/1210.5460.pdf.
Matching Coins: Brought to my attention by Oded Regev of the Courant Institute,
NYU. In [56] the authors show that with more sophisticated versions of this
scheme, Sonny and Cher can get as close as they want to a fraction x of success,
where x is the unique solution to the equation
−x log2 x − (1−x) log2 (1−x) + (1−x) log2 3 = 1,
about 0.8016, but they can do no better. Moreover, this applies whether the coin-
flips are random or adversarial.
Two Sheriffs: This puzzle was devised and presented in [10] to give a toy example of
how two parties that share information, but have no common secret, can establish
a common secret over an open channel, and then use it to communicate in secret.
theorem: Properly speaking, the study of secret codes is cryptology, which divides into de-
signing codes (cryptography) and breaking them (cryptanalysis). In practice “cryp-
tography” is often used to cover the whole subject. A particularly entertaining
source is [65].

14. Great Expectation

Bidding in the Dark: Sent by Maya Bar Hillel, University of Jerusalem.
Rolling All the Numbers: Classic.
Spaghetti Loops: Classic.
Ping-Pong Progression: Adapted from a problem posed on 2/28/17 at the MoMath
Masters, an annual puzzle-solving contest put on by the National Museum of
Mathematics.
Drawing Socks: Suggested by Yan Zhang of San Jose State University.
Random Intersection: Observed and proved by Jim Propp, of the University of
Massachusetts, Lowell. He cannot be blamed for the calculus-avoiding “proof ”
given here, however.
Roulette for the Unwary: My variation on Problem 7 from [86].
An Attractive Game: Problem 6 from [86]. Our proof shows that the game becomes
fair if the payoff for two of your number is $3 instead of $2, and for three $5 instead
of $3.
Next Card Bet: I heard this problem from probabilist Russ Lyons, of Indiana Univer-
sity, who heard it from Yuval Peres, who heard it from Sergiu Hart; but the problem
goes back apparently to a paper of Tom Cover [25]. If the original $1 stake is not di-
visible, but is composed of 100 indivisible cents, things become more complicated
and it turns out that Victor does a dollar worse. A dynamic program (written by
Ioana Dumitriu, now a professor at UCSD) shows that optimal play guarantees
ending with at least $8.08. Warning: you need to be a bit more conservative in the
“100 cents” game than in the continuous version; if instead you always bet the

381
 Mathematical Puzzles

nearest number of cents to the fraction (b−r)/(b+r) of your current worth, you
could go bankrupt before half the deck is gone.
Serious Candidates: Devised by probabilist David Aldous of U.C. Berkeley, who was
inspired by the prediction market and the Republican party’s presidential nomina-
tion process in 2012.
Rolling a Six: Devised and told to me by MIT probabilist Elchanon Mossel, who
dreamed it up as an easy problem for his undergraduate probability students and
then realized the answer was not 3. The solution given is your author’s.
Napkins in a Random Setting: Posed on the spot by the late, great John Horton Con-
way at a math conference banquet where the circular table, coffee cups and napkins
were as described.
Roulette for Parking Money: Due to Dick Hess, who was inspired by a note written
by Bill Cutler. It’s called “Bus Ticket Roulette” in [63]. Here we have assumed that
the table has both a “0” and a “00” but the same scheme is optimal in Monte Carlo
where there’s only the single zero.
Buffon’s Needle: Ramaley’s paper [93].
Covering the Stains: Devised by Naoki Inaba and sent to me by Iwasawa Hirokazu,
known also as Iwahiro; both are prolific puzzle composers. The actual maximum
number of points in the plane that can always be covered by disjoint open disks
is not known (see [5]); embarrassingly, the best results currently are that 12 points
are always coverable and 42 are not. My guess? 25.
Colors and Distances: Sent by probabilist Ander Holroyd, who heard it from Russ
Lyons. My setting is a real town which, the last time I looked, was pretty evenly
split.
Painting the Fence: From the Spring ’17 Emissary Puzzles Column by Elwyn
Berlekamp and Joe P. Buhler, to which it was contributed by Paul Cuff who recalls
it from one of Tom Cover’s seminars at Stanford.
Filling the Cup: Classic.
theorem: For (much) more on the probabilistic method, the reader is enthusiastically
referred to [3].

15. Brilliant Induction

IHOP: Yes, I know, “IHOP” doesn’t really work as an abbreviation for “induction
hypothesis.” Lizz Moseman, a former PhD student of mine, had an earlier math
teacher who used that term and Lizz and I both found it irresistible.
Uniform Unit Distances: Posed in the (now defunct) journal Mathematical Spectrum
and solved by David Seal of Winchester College (UK) in Vol. 6, Issue 2, 1973–4.
As you may have noticed, the sets Sn we constructed look like projections onto
the plane of an n-dimensional hypercube.
Swapping Executives: Moscow Mathematical Olympiad 2002, Problem C4 [40], con-
tributed by A. V. Shapovalov.
Odd Light Flips: Moscow Mathematical Olympiad 1995, Problem C6 [39], con-
tributed by A. Ya. Kanel-Belov. A proof using linear algebra is also possible.
Truly Even Split: This puzzle was suggested by Muthu Muthukrishnan who heard it
from eminent computer theorist Bob Tarjan.
Non-repeating String: Moscow Mathematical Olympiad 1993, Problem A5 [39], con-
tributed by A. V. Spivak.
Baby Frog: Moscow Mathematical Olympiad 1999, Problem D4 [39], contributed by
A. I. Bufetov. Note that the√given proof provides a pretty efficient scheme. Since a
circular clearing of radius 2 times a cell side (thus area 2π/22n , relative to the

382
 Notes and Sources

big square) must contain a grid cell, the baby frog can be guided to a circle of area
A with only ⌈ 12 log2 (2π/A)⌉ croaks.
Guarding the Gallery: The question (with 11 replaced by n) was asked by the late
Victor Klee, a brilliant geometer who spent most of his career at the University of
Washington, and answered by combinatorialist Václav Chvátal. The lovely proof
in this volume was found by Steve Fisk [45].
Path Through the Cells: Moscow Mathematical Olympiad 1999, Problem D5 [39],
contributed by N. L. Chernyatyev. Notice that the fact that the cells of a telephone
network most likely constitute a planar graph is not needed for this result.
Profit and Loss: This puzzle is adapted from one which appeared on the 1977 Inter-
national Mathematical Olympiad, submitted by a Vietnamese composer. Thanks
to Titu Andreescu for telling me about it. The solution given is my own; the
industrious reader will not find it difficult to generalize it to the case where x
and y have a greatest common divisor gcd(x, y) other than 1. The result is that
f (x, y) = x + y − 1 − gcd(x, y).
Uniformity at the Bakery: This lovely puzzle is from a Russian competition and ap-
pears in [97]. The given argument also works if all the weights are rational num-
bers, since we can just change units so that the weights are integers. But what if
the weights are irrational? Regard the real numbers R as a vector space over the ra-
tionals Q, and let V be the (finite-dimensional) subspace generated by the weights
of the bagels. Let α be any member of a basis for V , and let qi be the rational co-
efficient of α when the weight of the ith bagel is expressed in this basis. Now the
same argument used in the rational case shows that all the qi ’s must be 0, but this
is a contradiction since then α was not in V to begin with.
Summing Fractions: From the 3rd All Soviet Union Mathematical Competition, Kiev,
1969.
Tiling with L’s: Told to me by Rick Kenyon of Yale University, an expert on random
tilings.
Traveling Salesmen: From the 11th All Soviet Union Mathematical Competition,
Tallinn, 1977; thanks to Barukh Ziv of Intel for sending me the intended solu-
tion. The solution given was devised by me and Bruce Shepherd of the University
of British Columbia; another nice solution was sent to me by Emmanuel Boussard
of Boussard & Gavaudan Asset Management.
Lame Rook: Composed by Rustam Sadykov and Alexander Shapovalov for a 1998
Olympics, and told to me by Rustam. I love that unexpected question at the end
of the puzzle statement! Two additional remarks: (1) there’s a non-inductive proof
using Pick’s Theorem, and (2) in fairy chess, the lame rook is called the “wazir.”
theorem: A good source on Ramsey theory is [58].

16. Journey Into Space

Easy Cake Division: Martin Gardner, in [53], attributes the puzzle to Coxeter. The
solution given works for any regular polygon and also for any triangle, provided
you cut to the incenter.
Painting the Cubes: Puzzle and solution both told to me by Paul Zeitz.
Curves on Potatoes: Communicated to me by Dick Hess, who heard it from Dieter
Gebhardt; appears in [14].
Painting the Polyhedron: Told to me by Emina Soljanin (formerly a Distinguished
Member of Technical Staff at Bell Labs, now at Rutgers).
Boarding the Manhole: Classic.
Slabs in 3-Space: From an early Putnam Exam.

383
 Mathematical Puzzles

Bugs on Four Lines: This puzzle was passed to me by Matt Baker, of Georgia Tech. It
is sometimes called “the four travellers’ problem” and appears on the website “In-
teractive Mathematics Miscellany and Puzzles” at http://www.cut-the-knot.
org/.
Circular Shadows II: From the 5th All Soviet Union Mathematical Competition, Riga,
1971.
Box in a Box: Told to me by Anthony Quas (University of Victoria) who heard it, and
the solution given, from Isaac Kornfeld, a professor at Northwestern University;
Kornfeld had heard the puzzle many years ago in Moscow. Additional very nice
solutions were sent to me by Mike Todd of Cornell University and by Marc Mas-
sar. Mike’s uses vectors and the triangle inequality, while Marc’s is based on the
observation that for an a × b × c box, (a + b + c)2 is equal to the area of the box
plus the square of its diagonal length.
Angles in Space: I was tested on this puzzle during a visit to MIT, and was stumped.
It turns out that the question was for some time (since the late 1940’s) an open
problem of Paul Erdős and Victor Klee, then was solved by George Danzig and
Branko Grünbaum in 1962.
The question of the maximum number of points in n-space that determine only angles
that are strictly acute remained open much longer, for a long time known only to
be between 2n−1 and 2n − 1. Only recently [54] has it been shown that you can
get halfway to the upper bound: There’s a set of 2n−1 + 1 points with all angles
strictly acute.
Curve and Three Shadows: The question was raised by number theorist Hendrik
Lenstra, after hearing about Oskar van Deventer’s puzzle “Oskar’s Cube” in which
three orthogonal sticks poke through the sides of a cube, each of which has a maze
cut out of it. Since the mazes can’t have loops without a piece falling out, it wasn’t
clear whether the puzzle could be designed so that the stick intersection can follow
a closed curve in space.
theorem: The Monge Circles theorem, with its famous sphere proof, was first brought
to my attention by computer theorist Dana Randall of Georgia Tech. The flaw
was pointed out to me by Jerome Lewis, Professor of Computer Science at the
University of South Carolina Upstate.

17. Nimble Nimbers

Life Is a Bowl of Cherries: The game is ancient; the analysis goes back at least as far
as 1935 [98].
Life Isn’t a Bowl of Cherries?: For more on this and the previous puzzle, [13] is the
fun place to go.
Whim-Nim: Devised by John H. Conway, who called it simply “Whim.”
Option Hats: This puzzle dates back to the U.C. Santa Barbara 1998 PhD thesis of
Todd Ebert (now at Cal State Long Beach). It was the subject of a New York Times
article entitled “Why Mathematicians Now Care About Their Hat Color,” by Sara
Robinson, April 10, 2001 (see http://web.csulb.edu/~tebert/hatProblem/
nyTimes.htm). The radius-1 covering code of size 7, for the n = 5 case, was taken
from [21].
Chessboard Guess: From [81] and sent to me by one of that book’s authors, Anany
Levitin.
Majority Hats: Posted by Thane Plambeck of Counterwave, Inc. Many insights, plus
the dynamic program and its results, were contributed by Johan Wästlund of
Chalmers University. Wästlund points out that there is an application of the re-
sult to gambling at a “responsible” casino where you buy chips once in advance,

384
 Notes and Sources

and are paid in cash, so that you can’t gamble with your winnings and perhaps get
addicted. The dynamic program maximizes your probability of coming out ahead
given that you bought 100 chips that can be bet at even odds on fair coinflips.
(Yes, you’d’ve been better off had you bought only 99 chips.) But one can some-
times do better in Majority Hats than in the responsible casino; with computer help,
Wästlund now has a protocol that has pushed the majority probability for 100 pris-
oners to 1156660500373338319469/1180591620717411303424, or approximately
97.9729552603863%.
Fifteen Bits and a Spy: Told to me by László Lovász, then of Microsoft Research, now
president of the Hungarian Academy of Sciences. Laci is uncertain of its origin.
theorem: A nice source for error-correcting codes is [91].

18. Unlimited Potentials

Signs in an Array: Classic. We were not asked how long the process takes, but using
the facts that (1) every line needs to be flipped at most once, and (b) flipping every
line that some scheme doesn’t flip gives the same result, it can be deduced that you
can always get all the sums non-negative in at most ⌊(m + n)/2⌋ flips.
Righting the Pancakes: Moscow Mathematical Olympiad 2000, Problem A5 [40], con-
tributed by A. V. Shapovalov.
Breaking a Chocolate Bar: I’d love to know who first came up with this one.
Red Points and Blue: From a Putnam Exam of the 1960’s. Another nice solution was
sent to me by tax expert Carl Giffels. The claim is that any group of n red and n
blue points, with no three collinear, can be divided by a line so that each half-plane
contains the same number of red points as blue. To see this, pick any point P that
doesn’t lie on any line containing more than one of the 2n points. Draw a line
through P and rotate it slowly about P . The line may start with (say) more red
points than blue on its right-hand side, but 180◦ later the reverse will be true; thus,
at some point, the numbers of red and blue points on the “right” side of the line
(and therefore also on the left) must be equal. Now, each half-plane can be further
sub-divided the same way (this time by a half-line ending at the previous line), and
the sub-divisions divided similarly, until every region contains only a single red and
a single blue point. The line segments connecting those pairs cannot cross.
The puzzle’s conclusion is true in higher dimensions as well, as pointed out to me by
Pablo Soberón of Baruch College, City University of NY. In three dimensions, you
have n points each of colors red, blue, and green, no three collinear. You want
disjoint triangles connecting a 3-way matching, and you get it by induction via
the Ham Sandwich Theorem (a.k.a. the Stone–Tukey Theorem). If n is even, use
the theorem to find an appropriate plane, missing all the points, then apply the
theorem to each half. If n is odd, it will cut a point of each color and you can
use that triangle along with the ones obtained by induction. And if the plane cuts
through 3 or more red (say) points? No problem, split them and include the right
number in each induction piece.
Bacteria on the Plane: Variation of a puzzle posed by Fields Medal winner Maxim
Kontsevich (see [39], p. 110).
Pegs on the Half-plane: Variation of a problem described in [13], Vol II; I am told the
problem was invented originally by the second author, Conway. In his problem,
diagonal jumps were not permitted; one can nonetheless get a peg to the line y = 4
without much difficulty, but an argument like the one below shows that no higher
position can be reached.

385
 Mathematical Puzzles

Returning to my variation, Dieter Rautenbach wrote to me that a summer school stu-

dent, Niko Klewinghaus, came up with a proof that one can get a peg up eight
units and that this is the highest possible value.
Pegs in a Square: There is more than one way to solve this puzzle, which is part of a
problem which appeared at the 1993 International Mathematical Olympiad. The
proof given was communicated to me by combinatorialist Benny Sudakov of ETH
Zurich. At the Olympiad, contestants were asked to determine precisely for which
n the squares were reducible—pretty tough to do on the spot!
First-grade Division: Passed to the third author by Ori Gurel-Gurevich, a mathemati-
cian at the Hebrew University of Jerusalem. Gurel-Gurevich heard it from his army
friend Alon Amit, but the puzzle may go back much further in time. For general-
izations, see [46].
Infected Checkerboard: The puzzle seems to have started life somewhere in the old
Soviet Union, then migrated to Hungary where Spencer heard it. Its generalization
contributed to a whole new field of study called “bootstrap percolation,” initiated
by Béla Bollobás [17] in 1968.
Impressionable Thinkers: Suggested to me by Sasha Razborov, of the University of
Chicago; he tells me that it was considered for an International Mathematics
Olympiad, but rejected as too hard. It was posed and solved in [55].
The puzzle can be generalized considerably, for example, by adding weights to ver-
tices (meaning that some citizens’ opinions are more highly prized than others’),
allowing loops (citizens who consider their own current opinions as well), allow-
ing tie-breaking mechanisms, and even allowing different thresholds for opinion
changes.
Frames on a Chessboard: Sent by Ehud Friedgut of Hebrew University, it is a varia-
tion of a problem which appeared on an Israeli youth mathematics contest. In the
contest, the frames were 3 × 3 and 4 × 4, and a counting argument says that you
cannot reach every color configuration. The point is that the order in which the
frames are laid down is irrelevant; all you need to know is which of the 52 ways to
put the 4 × 4 frame down are utilized, and which of the 62 ways to put the 3 × 3
frame. Altogether there are thus 225 × 236 = 261 ways to try to get color configu-
rations; not enough. With Friedgut’s modification, however, we need some more
subtle argument, such as the one given.
Bugs on a Polyhedron: Presented in a paper by Anton Klyachko [70].
Bugs on a Line: The analysis given was done by Ander Holroyd (then at the University
of British Columbia) and Jim Propp, at a meeting of the Institute for Elementary
Studies in Banff, Alberta, 2003. The bug was proposed by Propp as a way to sim-
ulate deterministically a random walk on the non-negative integers in which steps
are made, independently, to the left with probability 1/3 and to the right with prob-
ability 2/3. In such a walk, a given bug drops off to the left or proceeds to infinity
on the right with equal probability; as we saw, the deterministic model gives a strict
alternation instead, after the first couple of trips. The argument can be extended to
other random walks.
Flipping the Pentagon: First brought to my attention by Andy Liu, this now-notorious
puzzle was Problem #3 at the 1986 International Mathematical Olympiad, held
in Poland. Joseph Keane, a US contestant, received a special prize for his sets-of-
consecutive-vertices solution.
Picking the Athletic Committee: Brought to me by my Dartmouth computer science
colleague Deeparnab Chakrabarty, who needed the general result (with the num-
ber 3 replaced by an arbitrary integer k) for a research paper. Recently Iwasawa
Hirokazu has generalized the result even further.

386
 Notes and Sources

Bulgarian Solitaire: See [1]. For non-triangular numbers of chips, you end up cycling
through triangular formations with extra chips shifting along the new diagonal.
theorem: See [85] for more on the infinite case.

19. Hammer and Tongs

Phone Call: I cannot recall from whom I first heard this, many years ago.
Crossing the River: Appears in the medieval Latin manuscript Propositiones ad Acuendos
Juvenes (Problems to Sharpen the Young), attributed to the polymath Alcuin of
York, 735–804.
Sprinklers in a Field: Moscow Mathematical Olympiad 1996, Problem A3 [39], con-
tributed by I. F. Sharygin.
Fair Play: I was reminded of this puzzle by Tamás Lengyel, of Occidental College.
More improvements are possible and [88] shows how to get the last drop of blood
out of the process, minimizing the expected number of flips to get a decision, re-
gardless of the coin’s probability of landing heads up. The problem of extracting
unbiased, random bits from various tainted random sources is of major importance
in the theory of computing, and the subject of many research papers and significant
breakthroughs in recent years.
Finding the Missing Number: An alternative for those who are weak on addition but
strong on dexterity is to use fingers and toes to keep track of which ones digits, and
which tens digits (respectively), have appeared odd numbers of times.
Identifying the Majority: The algorithm given is described in [44].
Poorly Placed Dominoes: Original, but inspired by Problem 429 in [83].
Unbreakable Domino Cover: Sent by aeronautical engineer Bob Henderson.
Filling the Bucket: Passed to me by Dartmouth PhD student Grant Molnar. The
puzzle was composed by Caleb Stanford and appeared on the 2017 Utah Math
Olympiad. Note that if the buckets contain k gallons instead of just two, you
can still win—but you’ll need a lot of buckets and a lot of time, since the series
1 + 1/3 + 1/5 + . . . takes exponentially many terms to get to k. (The two-gallon
case can be done with 7 buckets, but for k = 3 you already need 43 buckets and
lots of time.)
Polygon on the Grid: Moscow Mathematical Olympiad 2000, Problem C3 [40], con-
tributed by Gregory Galperin, now of Eastern Illinois University.
One-bulb Room: I heard this puzzle from Adam Chalcraft, who has the distinction of
having represented Great Britain internationally in unicycle hockey. The puzzle
has also appeared on IBM’s puzzle site and was reprinted in Emissary, the newslet-
ter of the Mathematical Sciences Research Institute in Berkeley, California. A ver-
sion even appeared on the justly famous public radio program Car Talk in 2003.
Closely related, but much more challenging, is the puzzle of the Two-Bulb Room where
the conditions are similar but there is a second bulb; the catch is that all prisoners
must use the same protocol. An elegant solution known as the “see-saw protocol”
is presented in [43].
Sums and Differences: From the 5th All Soviet Union Mathematical Competition,
Riga, 1971. The argument given works for any odd number of numbers greater
than 3; for even numbers, the contradiction is reached even more quickly. There
is, however, a set of three numbers with no pair satisfying the requirements of the
puzzle: {1, 2, 3}!
Prisoner and Dog: Brought to my attention by Giulio Genovese, a mathematician at
Harvard’s McCarroll Lab. The prisoner’s strategy described, though good enough
for speed ratio 1:4, can be improved upon.

387
 Mathematical Puzzles

Love in Kleptopia: This puzzle was passed to me by Caroline Calderbank, daughter

of mathematicians Ingrid Daubechies and Rob Calderbank.
Badly Designed Clock: Posed by Andy Latto, a Boston-area software engineer, at the
Gathering for Gardner IV. It can be solved algebraically or geometrically, with suf-
ficient care and patience; the proof here was supplied to Andy by Michael Larsen,
a mathematics professor at Indiana University. The idea of a third hand (instead
of a second clock) came to me from David Gale.
Worms and Water: Told to me by Balint Virag, a probabilist at the University of
Toronto.
Generating the Rationals: From The 13th All Soviet Union Mathematical Competi-
tion, Tbilisi, 1979.
Funny Dice: Non-transitive (or “intransitive”) sets of dice have been around for a long
time; a recent article [23] showed that in fact randomly designed dice often have
this property.
Sharing a Pizza: Devised by Daniel E. Brown in 1996, this puzzle has attracted a lot
of attention, in part because of your author’s conjecture that Alice could always
get at least 4/9 of the pizza—proved independently by two groups [20, 72].
Names in Boxes: This puzzle has a short but fascinating history. Devised by Danish
computer scientist Peter Bro Miltersen, a version appeared in a prize-winning pa-
per of his and Anna Gal’s [48]. But Miltersen didn’t think there was a solution
until one was pointed out to him over lunch by colleague Sven Skyum. The puzzle
reached me via quantum computing expert Dorit Aharonov.
Eugene Curtin and Max Warshauer [30] have recently proved that the solution given
cannot be improved upon.
Lambert Bright and Rory Larson, and independently Richard Stanley of MIT, pro-
posed the following variation. Suppose each prisoner must look in 50 boxes, and
the requirement for survival is that every prisoner not find his own name. Despite
having the diametrically opposite objective from before, it seems that the prisoners
can do no better than to follow exactly the same strategy. Here, though, they sur-
vive only if every cycle has more than 50 boxes in it, which can only happen if there
is just one big cycle—for which their chances are precisely 1 in 100. Not great, but
a lot better than 1 in 2100 . They do just as well even if every prisoner is required to
look in 99 boxes—again, they follow the strategy and win just when the random
permutation has just one big cycle. In this case it is immediately obvious that no
better strategy is available, because the very first prisoner, no matter what he does,
has only a 1% probability of avoiding his own name. The amazing thing is that fol-
lowing the strategy, if that first prisoner succeeds, then automatically every other
prisoner will succeed as well.
Life-Saving Transposition: Sent to Emissary via Kiran Kedlaya by Piotr Krason.
Self-referential Number: Devised and told to me by probabilist Ander Holroyd.
theorem: Due to Georg Alexander Pick in 1899. There are many proofs; the one pre-
sented is an amalgam of several.

20. Let’s Get Physical

Rotating Coin: Another classic popularized by Martin Gardner.
Pie in the Sky: Original (but of course, I may be the millionth person to have thought
of this question).
Returning Pool Shot: Moscow Mathematical Olympiad 2004, Problem B3 [40], con-
tributed by A. Ya. Kanel-Belov.
Falling Ants: As far as I know the first publication of the puzzle was in Francis Su’s
“Math Fun Facts” web-column at Harvey Mudd College; Francis recalls hearing it

388
 Notes and Sources

in Europe from someone he can’t trace named Felix Vardy. The puzzle then showed
up in the Spring/Fall 2003 issue of Emmissary. Dan Amir, a former Rector of Tel
Aviv University, read the puzzle in Emmissary and posed it to Noga Alon, who
brought it to the Institute for Advanced Study; I first heard it from Avi Wigderson
of the I.A.S., in late 2003.
Ants on the Circle: Elwyn Berlekamp and I came up with this one together (but of
course others may have thought of it as well—turning a line into a circle is, as we
have seen, often a useful thing to do).
Sphere and Quadrilateral: Told to me by Tanya Khovanova, who lists it among the
“coffin puzzles” on her blog. These are puzzles with simple solutions that are dif-
ficult to find, especially for someone taking a timed exam. According to Tanya
and others, such puzzles were used in the Soviet Union to keep “undesirables,”
for example, Jewish students, out of the best schools.
Two Balls and a Wall: Disseminated by Dick Hess and Gary Antonick at the Eleventh
Gathering for Gardner, based on a discovery [49] by Gregory Galperin.
theorem: Brought to my attention by Yuval Peres.

21. Back from the Future

Portrait: Goes back at least to a puzzle submitted by G. H. Knight to an Oxford journal
in 1872 [71], p. 240.
Three-way Duel: Classical, and Problem 20 of [86]. The argument that Bob should also
shoot in the air was advanced, convincingly, to me by Gerry Myerson of Macquarie
U., Sydney.
Testing Ostrich Eggs: Sent to me by Tamás Lengyel, it appears in [73] as Problem
#166.
Pancake Stacks: Brought to my attention by Bill Gasarch of the University of Mary-
land, this puzzle appeared in the 12th All Soviet Union Mathematical Olympiad,
Tashkent, 1978.
Chinese Nim: Also known as Wythoff ’s Game, introduced in 1907 [106].
Turning the Die: I was challenged by a superior officer in the U.S. Navy to play this
game for drinks, at a stateside bar during the Vietnam War. I declined, but of course
analyzed the game so I would be prepared on the next occasion—which, perhaps
luckily, never materialized.
Game of Desperation: Brought to my attention by my PhD student, Rachel Esselstein,
the puzzle appeared at the 28th Annual USA Mathematical Olympiad in 1999. See
also [42].
Deterministic Poker: From an early Martin Gardner “Mathematical Games” column
in Scientific American.
Bluffing with Reals: Origin unknown.
Swedish Lottery: This nice lottery idea was brought to my attention by Olle Häggström
of Chalmers University (Göteborg, Sweden). I’ve tried it a few times with groups
of twenty or so students; twice, the winning number was 6.
Pegs on the Corners: Told to me by computer theorist Mikkel Thorup of the University
of Copenhagen, who heard it from Assaf Naor (now at Princeton), who heard it
from graduate students at the Hebrew University of Jerusalem.
Touring an Island: Seen on Carnegie Mellon University’s “Puzzle Toad” web page,
and found in [50].
Fibonacci Multiples: Told to me by Richard Stanley.
Light Bulbs in a Circle: From the International Mathematical Olympiad of 1993. Puz-
zle maven Tom Verhoeff tells me that by applying the theory of linear feedback

389
 Mathematical Puzzles

shift registers, he has determined that the bulbs first return to all-on at time t =
181,080,508,308,501,851,221,811,810,889, much greater than the age of the uni-
verse in seconds. Check out his demo [102].
Emptying a Bucket: From the 5th All Soviet Union Mathematical Olympiad, Riga,
1971. The puzzle showed up again, minus the hardware, on the Putnam Exam
in 1993, and reached me via Christian Borgs, then at Microsoft Research. The
solution given is mine, but there is also an elegant number-theoretic solution found
independently by Svante Janson of Uppsala University, Sweden, and Garth Payne
of Penn State.
Ice Cream Cake: Told to me by French graduate student Thierry Mora, who heard it
from his prep-school teacher Thomas Lafforgue; Stan Wagon tells me that it comes
from a Moscow Mathematical Olympiad (which, if you’ve read this far in Notes &
Sources, should not be a surprise.) The puzzle originally involved a second angle
as well, indicating the amount of cake passed over between wedge-cuttings; it still
requires only finitely many operations to get all the icing back on top, as ambitious
readers will verify. But, the puzzle as stated here (where the second angle is 0) is
already surprising and challenging.
Moth’s Tour: Suggested to me by Richard Stanley. Thirty years ago, at a Chinese
restaurant in Atlanta, Laci Lovász and I [80] managed to prove that the cycle is
the only graph, apart from the complete graph, with this nice property (that the
last new vertex visited by a random walk is equally likely to be any vertex other
than the starting point).
Boardroom Reduction: Original. Many complex versions abound of a similar puzzle
involving pirates and gold coins, but it seemed to me that the essence of that puzzle
could be achieved without the coins.
theorem: This marvelous fact (in another context) was observed by Lord Rayleigh, and
probably many others since—and maybe before. A nice modern reference is [95].
Sometimes called “Beatty’s problem” (after Samuel Beatty 1881–1970), the puzzle
appeared as Problem 3117 in The American Mathematical Monthly 34 (1927), pp.
159–159, and again on the 20th Putnam Exam, November 21, 1959.

22. Seeing Is Believing

Meeting the Ferry: Due to Edouard Lucas, 19th Century French mathematician.
Mathematical Bookworm: From Martin Gardner.
Rolling Pencil: My late colleague Laurie Snell caught me on this one, which appears
in [104].
Splitting a Hexagon: From Tanya Khovanova’s blog.
Circular Shadows I: Moscow Mathematical Olympiad 1997, Problem C1 [39], con-
tributed by A. Ya. Kanel-Belov.
Trapped in Thickland: Moscow Mathematical Olympiad 2000, Problem D6 [40], con-
tributed by A. Ya. Kanel-Belov.
Polygon Midpoints: Sent by Barukh Ziv who found it in a book by I. M. Yaglom
[107]. The solution given is Ziv’s. Incidentally, if you don’t care for self-intersecting
polygons, note that any finite set of points in the plane can be taken to be the
vertices of a non-self-intersecting polygon: Just pick a point C inside the convex
closure of your set of points, then connect your points in clockwise order around
C.
Not Burning Brownies: Told to me by Brown University’s Michael Littman, devised
by him and fellow computer scientist David McAllester.
Protecting the Statue: Moscow Mathematical Olympiad 2001, Problem C2 [40], con-
tributed by V. A. Kleptsyn.

390
 Notes and Sources

Gluing Pyramids: The argument given, sometimes called the “pup tent” solution, ap-
peared in a 1982 article by Steven Young [108]. After the PSAT debacle, the Educa-
tional Testing Service created a panel, on which your author served, for reviewing
the questions on their mathematics aptitude tests.
Precarious Picture: Contributed by Giulio Genovese, who heard it from more than
one source in Europe.
Finding the Rectangles: Sent by Yan Zhang.
Tiling a Polygon: Sent by Dana Randall.
Tiling with Crosses: Sent by Senya Schlosman.
Cube Magic: I was reminded of this puzzle, which appeared in a Martin Gardner col-
umn, by Gregory Galperin. A polytope has the “Rupert” property if a straight
tunnel can be made in it large enough to pass an identical polytope through (and,
therefore, can be enlarged to pass an even bigger copy through). The cube case was
originally proved, allegedly, by Prince Rupert of the Rhine in the late 17th century.
As of now nine of the Archimidean polytopes are known to have the Rupert prop-
erty, the last added recently in [77].
Circles in Space: Told to me by computer theorist Nick Pippenger of Harvey Mudd
College; the construction given is due to Andrzej Szulkin of Stockholm University
[100]. As pointed out to me by Johan Wästlund, this and some other tiling prob-
lems can also be solved using transfinite induction, as in Double Cover by Lines
from Chapter 23.
Invisible Corners: Moscow Mathematical Olympiad 1995, Problem D7 [39], con-
tributed by A. I. Galochkin.
theorem: This classic appears in [89] as “The Problem of the Calissons.”

23. Infinite Choice

Hats and Infinity: Devised (to the best of my knowledge) by Yuval Gabay and Michael
O’Connor, then graduate students at Cornell; the solution was already implicit in
the work of Fred Galvin of the University of Kansas. Christopher Hardin (Smith
College) and Alan D. Taylor (Union College) then included it their article [60].
Stan Wagon wrote it up as a Macalester College Problem of the Week; additional
nice observations about this and the next version were made by Harvey Friedman
(Ohio State), Hendrik Lenstra (Universiteit Leiden) and Joe Buhler (Reed College).
It was the last of these, and (independently) Matt Baker of Georgia Tech, who
communicated the puzzle to me.
All Right or All Wrong: Similar history to Hats and Infinity above. The “all green”
solution was suggested to me by Teena Carroll of St. Norbert College.
Double Cover by Lines: Sent by Senya Shlosman, who is unaware of its origin.
Wild Guess: Sent by Sergiu Hart.
Exponent upon Exponent: Observed in 1964 by me and my college classmate Gerald
Folland, now a professor at the University of Washington.
Find the Robot: Suggested by Barukh Ziv. The nice way of counting the rationals men-
tioned above the puzzle can be found in [19].
Figure Eights in the Plane: An oldie. I once heard it attributed to the late topologist
R. L. (Robert Lee) Moore of the University of Texas.
Y’s in the Plane: Supplied by Randy Dougherty of Ohio State University, a three-time
winner of the U.S.A. Mathematical Olympiad.
theorem: Found in most elementary books on graph theory.

391
 Mathematical Puzzles

24. Startling Transformation

Sinking 15: Mentioned in Vol. II of [13], where it is attributed to E. Pericoloso Sporg-
ersi. However, rather suspiciously, this “name” is found also on Italian railroad
trains, warning passengers not to lean out of the window.
Slicing the Cube: Classic.
Expecting the Worst: Suggested by Richard Stanley.
Next Card Red: There is a discussion of this game in [27]. The modified version in
our proof reminds me of a game which was described—for satiric purposes—in
the Harvard Lampoon of March 30, 1967. The issue is dubbed “Games People Play
Number” and the game in question appears to have been composed by D. C. Ken-
ney and D. C. K. McClelland. Called “The Great Game of Absolution and Re-
demption,” it required that the players move via dice rolls around a Monopoly-
like board, until everyone has landed on the square marked “DEATH.” So who
wins? Well, at the beginning of the game, you were dealt a card face down from
the Predestination Deck. At the conclusion, you turn your card face up, and if it
says “damned,” you lose.
Magnetic Dollars: Probabilists call the mechanism of this puzzle the Pólya urn, after the
late George Pólya, a Hungarian-born professor at Stanford who famously wrote
about problem-solving in mathematics.
Integral Rectangles: Our proof is #8 of fourteen in Stan Wagon’s delightful article
[103], and there are more!
Laser Gun: Brought to my attention by Giulio Genovese, who got it from Enrico Le
Donne; they traced it to a 1990 Mathematics Olympiad in St. Petersburg. This
puzzle sparked some more general research; see, for example, [76] and work by
Keith Burns and Eugene Gutkin.
Random Intervals: This problem has a curious history. A colleague (Ed Scheinerman
of The Johns Hopkins University) and I needed to know the answer in order to
compute the diameter of a random interval graph, and we at first computed an
asymptotic value of 2/3. Later, using a lot of messy integrating, we found that the
probability of finding an interval which intersects all others is exactly 2/3, for any
number of intervals (from 2 on up). The combinatorial proof given was found by
Joyce Justicz, then taking a graduate reading course with me at Emory University.
Infected Cubes: The construction is due to Matt Cook and Erik Winfree of Caltech,
the proof that it works to their colleague Len Schulman.
Gladiators, Version I: I have a theory that the “confidence” condition of Version I
came about when someone tried to reconstruct Version II from the fair-game so-
lution. The second proof of Version I is from Graham Brightwell of the London
School of Economics.
Gladiators, Version II: From [66].
More Magnetic Dollars: This variation of Pólya’s urn was posed and solved by Joel
Spencer of New York University and his student Roberto Oliveira.
theorem: Suggested (as a puzzle) by Richard Stanley. The proof given is due to Henry
Pollak of Bell Labs and now Columbia Teachers College. The theorem was origi-
nally stated and proved by Konheim and Weiss [74].

392
 Bibliography
[1] Ethan Akin and Morton Davis, Bulgarian solitaire, Amer. Math. Monthly 92 #4
(1985), 310–330.
[2] Noga Alon, Tom Bohman, Ron Holzman, and Daniel J. Kleitman, On parti-
tions of discrete boxes, Discrete Math. 257 #2–3 (28 November 2002), 255–258.
[3] Noga Alon and Joel H. Spencer, The Probabilistic Method, Fourth Edition, Wiley
Series in Discrete Mathematics and Optimization, Hoboken, NJ, 2015.
[4] Noga Alon and D. B. West, The Borsuk-Ulam theorem and bisection of neck-
laces, Proc. Amer. Math. Soc. 98 #4 (December 1986), 623–628.
[5] Greg Aloupis, Robert A. Hearn, Hirokazu Iwasawa, and Ryuhei Uehara, Cov-
ering points with disjoint unit disks, 24th Canadian Conference on Computational
Geometry, Charlottetown, PEI (2012).
[6] D. André, Solution directe du problème résolu par M. Bertrand, Comptes Rendus
de l’Académie des Sciences Paris 105 (1887), 436–437.
[7] Titu Andreescu and Zuming Feng, eds., Mathematical Olympiads, MAA Press,
Washington DC, 2000.
[8] Frank Arntzenius, Some problems for conditionalization and reflection, J. Phi-
los. 100 #7 (2003), 356–370.
[9] W. W. Rouse Ball, Mathematical Recreations and Essays, Macmillan & Co., Lon-
don, 1892.
[10] D. Beaver, S. Haber, and P. Winkler, On the isolation of a common secret,
in The Mathematics of Paul Erd˝os Vol. II, R. L. Graham and J. Neˇ
setˇ
ril, eds.,
Springer-Verlag, Berlin, 1996, 121–135. (Reprinted and updated in 2013.)
[11] József Beck and William W. Chen, Irregularities of Distribution, Cambridge Uni-
versity Press, Cambridge, UK, 1987.
[12] Antonio Behn, Christopher Kribs-Zaleta, and Vadim Ponomarenko, The con-
vergence of difference boxes, Amer. Math. Monthly 112 #5 (May 2005), 426–
439.
[13] Elwyn R. Berlekamp, John H. Conway, and Richard K. Guy, Winning Ways
for Your Mathematical Plays, Volumes 1–4, Taylor & Francis, Abingdon-on-
Thames, UK, 2003.
[14] E. R. Berlekamp and T. Rodgers, eds., The Mathemagician and Pied Puzzler, AK
Peters, Natick, MA, 1999.
[15] Geoffrey C. Berresford, A simpler proof of a well-known fact, Amer. Math.
Monthly 115 #6 (June–July 2008), 524.
[16] J. Bertrand, Solution d’un problème, Comptes Rendus de l’Académie des Sciences,
Paris 105 (1887), 369.

393
 Mathematical Puzzles

[17] B. Bollobás, Weakly k-saturated graphs, Beiträge zur Graphentheorie (Kollo-

quium, Manebach, May 1967), 25–31. Teubner-Verlag, Leipzig, 1968.
[18] Peter Boyland, Ivan Roth, Gabriella Pintér, István Laukó, Jon E. Schoenfield,
and Stephen Wasielewski, Non-intersecting diagonals in an array, J. Integer Seq.
20 (2017), Article 17.2.4, 1–24.
[19] Neil Calkin and Herbert Wilf, Recounting the rationals, Amer. Math. Monthly
107 #4 (2000), 360–363.
[20] J. Cibulka, J. Kynˇ
cl, V. Mészáros, R. Stolaˇr, and P. Valtr, Solution of Peter
Winkler’s pizza problem, in Fete of Combinatorics and Computer Science,
Bolyai Society Mathematical Studies 20 (2010), Janos Bolyai Mathematical
Society and Springer-Verlag, Berlin, 63–93.
[21] G. Cohen, I. Honkala, S. Litsyn, and A. Lobstein, Covering Codes, North-
Holland, Amsterdam, 1997.
[22] Joel E. Cohen, Johannes H. B. Kemperman, and Gheorghe H. Zbǎganu, El-
ementary inequalities that involve two nonnegative vectors or functions, Proc.
Natl. Acad. Sci. USA 101 #42 (October 19 2004), 15018–15022.
[23] Brian Conrey, James Gabbard, Katie Grant, Andrew Liu, and Kent Morrison,
Intransitive dice, Math. Mag. 89 #2 (April 2016), 133–143.
[24] J. H. Conway and B. Heuer, All solutions to the immobilizer problem, Math.
Intell. 36 (2014), 78–86. https://doi.org/10.1007/s00283-014-9503-z.
[25] Thomas M. Cover, Universal Gambling Schemes and the Complexity Mea-
sures of Kolmogorov and Chaitin, Technical Report No. 12, October 14, 1974,
Dept. of Statistics, Stanford University, Stanford, CA.
[26] T. M. Cover, Pick the largest number, in Open Problems in Communication and
Computation, T. Cover and B. Gopinath, eds., Springer Verlag, Berlin,1987, 152.
[27] T. Cover and J. Thomas, Elements of Information Theory, Wiley, Hoboken, NJ,
1991.
[28] H. S. M. Coxeter, A problem of collinear points, Amer. Math. Monthly 55 #1
(1948), 26–28.
[29] Henry Crapo and Ethan Bolker, Bracing rectangular frameworks I, SIAM J.
Applied Math. 36 #3 (June 1979), 473–490.
[30] Eugene Curtin and Max Warshauer, The locker puzzle, Math. Intell. 28 #1
(2006), 28–31.
[31] P. Diaconis, R. L. Graham, and B. Sturmfels, Primitive partition identities, in
Volume II of Paul Erdős is 80, János Bolyai Society, Budapest, 1986.
[32] D. Djuki´c, V. Jankovi´
c, I. Mati´c, and N. Petrovi´
c, The IMO Compendium, Second
Edition, Springer, New York, 2011.
[33] Cian Dorr, Sleeping beauty: In defence of Elga, Analysis 62 (2002), 292–296.
[34] F. J. Dyson, The problem of the pennies, The Mathematical Gazette 30 #291
(October 1946), 231–234.
[35] Sheldon J. Einhorn and I. J. Schoenberg, On Euclidean sets having only two
distances between points II, Indagationes Mathematicae (Proceedings) 69 (1966),
489–504.
[36] A. Engel, Problem-Solving Strategies, Springer, New York, 1998.
 Bibliography

[37] P. Erdős, A. Ginzburg, and A. Ziv, Theorem in the additive number theory,
Bull. Res. Council of Israel 10F (1961), 41–43.
[38] P. Erdös and G. Szekeres, A combinatorial problem in geometry, Compos. Math.
2 (1935), 463–470.
[39] R. Fedorov, A. Belov, A. Kovaldzhi, and I. Yashchenko, eds., Moscow Mathe-
matical Olympiads, 1993–1999, MSRI/AMS 2011.
[40] R. Fedorov, A. Belov, A. Kovaldzhi, and I. Yashchenko, eds., Moscow Mathe-
matical Olympiads, 2000–2005, MSRI/AMS 2011.
[41] D. Felix, Optimal Penney Ante strategy via correlation polynomial identities,
Electron. J. Comb. 13 R35 (2006), 1–15.
[42] Thomas S. Ferguson, A Course in Game Theory, published by the author, 2020.
Online at https://www.math.ucla.edu/~tom/Game_Theory.
[43] M. J. Fischer, S. Moran, S. Rudich, and G. Taubenfeld, The wakeup problem,
Proc. 22nd Symp. on the Theory of Computing, Baltimore, MD, May 1990.
[44] M. J. Fischer and S. L. Salzberg, Finding a majority among n votes, J. Algo-
rithms 3 4 (December 1989), 362–380.
[45] Steve Fisk, A short proof of Chvátal’s watchman theorem, J. Comb. Theory B
24 (1978), 374.
[46] M. L. Fredman, D. J. Kleitman, and P. Winkler, Generalization of a puzzle
involving set partitions, Barrycades and Septoku; Papers in Honor of Martin Gard-
ner and Tom Rodgers, Thane Plambeck and Tomas Rokicki, eds., AMA/MAA
Spectrum Vol. 100, MAA Press, Providence, RI, 2020, pp. 125–129.
[47] Aaron Friedland, Puzzles in Math and Logic, Dover, Mineola, NY, 1971.
[48] Anna Gal and Peter Bro Miltersen, The cell probe complexity of succinct
data structures, International Colloquium on Automata, Languages and Program-
ming (ICALP), Eindhoven, The Netherlands. Jos C. M. Baeten, ed., Lecture
Notes in Computer Science 2719, Springer-Verlag, Berlin, 2003.
[49] G. A. Galperin, Playing pool with π: The number π from a billiard point of
view, Regul. Chaotic Dyn. 8 (2003), 375–394.
[50] G. A. Galperin and A. K. Tolpygo, Moscow Mathematical Olympiads, Moscow,
Prosveshchenie, 1986.
[51] Martin Gardner, The Colossal Book of Short Puzzles and Problems, W. W. Norton
& Co., New York, NY, 2006.
[52] Martin Gardner, On the paradoxical situations that arise from nontransitive
relations, Sci. Am. 231 4, (October 1974), 120–125.
[53] Martin Gardner, Time Travel and Other Mathematical Bewilderments, W. H. Free-
man and Company, New York, NY, 1988. ISBN 0-7167-1925-8.
[54] Balász Gerencsér and Viktor Harangi, Too acute to be true: The story of acute
sets, Amer. Math. Monthly 126 #10 (December 2019), 905–914.
[55] E. Goles and J. Olivos, Periodic behavior of generalized threshold functions,
Discrete Math. 30 (1980), 187–189.
[56] Olivier Gossner, Penélope Hernández, and Abraham Neyman, Optimal use of
communication resources, Econometrica 74 #6 (November 2006), 1603-1636.
[57] N. Goyal, S. Lodha, and S. Muthukrishnan, The Graham-Knowlton problem
revisited, Theory of Computing Systems 39 #3 (2006), 399–412.
 Mathematical Puzzles

[58] Ronald L. Graham, Bruce L. Rothschild, and Joel H. Spencer, Ramsey The-
ory, Second Edition, Wiley Series in Discrete Mathematics and Optimization,
Hoboken, NJ, 2013.
[59] L. J. Guibas and A. M. Odlyzko, String overlaps, pattern matching, and non-
transitive games, J. Comb. Theory Ser. A. 30 (1981), 183–208.
[60] Christopher S. Hardin and Alan D. Taylor, A peculiar connection between the
axiom of choice and predicting the future, Amer. Math. Monthly 115 #2 (2008),
91–96.
[61] G. H. Hardy, On certain oscillating series, Q. J. Math. 38 (1907), 269–288.
[62] Dick Hess, Golf on the Moon, Dover, Mineola, NY, 2014.
[63] Dick Hess, The Population Explosion and Other Mathematical Puzzles, World Sci-
entific, Singapore, 2016. https://doi.org/10.1142/9886.
[64] Ross Honsberger, Ingenuity in Mathematics, New Mathematical Library Series
23, Mathematical Association of America, Providence, RI, 1970.
[65] David Kahn, The Codebreakers, Scribner & Sons, New York, 1967 and 1996.
[66] K. S. Kaminsky, E. M. Luks, and P. I. Nelson, Strategy, nontransitive dom-
inance and the exponential distribution, Austral. J. Statist. 26 #2 (1984), 111–
118.
[67] K. A. Kearns and E. W. Kiss, Finite algebras of finite complexity, Discrete Math.
207 (1999), 89–135.
[68] Murray Klamkin, International Mathematical Olympiads 1979–1985, Mathemati-
cal Association of America, Providence, RI, 1986.
[69] Michael Kleber, The best card trick, Math. Intell. 24 #1 (Winter 2002), 9–11.
[70] Anton Klyachko, A funny property of sphere and equations over groups, Com-
mun. Algebra 21 #7 (1993), 2555–2575.
[71] G. H. Knight, in Notes and Queries, Oxford Journals (Firm), Oxford U. Press, Ox-
ford, UK, 1872.
[72] K. Knauer, P. Micek, and T. Ueckerdt, How to eat 4/9 of a pizza, Discrete Math.
311 #16 (2011), 1635–1645.
[73] Joseph D. E. Konhauser, Dan Velleman, and Stan Wagon, Which Way Did the
Bicycle Go?, Cambridge University Press, 1996.
[74] Alan G. Konheim and Benjamin Weiss, An occupancy discipline and applica-
tions, SIAM J. Appl. Math. 14 #6 (November 1966), 1266–1274.
[75] Boris Kordemsky, The Moscow Puzzles, Charles Scribner’s Sons, New York,
1971.
[76] J.-F. Lafont and B. Schmidt, Blocking light in compact Riemannian manifolds,
Geom. Topol. 11 (2007), 867–887.
[77] Gerárd Lavau, The truncated tetrahedron is Rupert, Amer. Math. Monthly 126
#10 (2019), 929–932.
[78] Tamas Lengyel, A combinatorial identity and the world series, SIAM Rev. 35
#2 (June 1993), 294–297.
[79] László Lovász, Combinatorial Problems and Exercises, North Holland, Amster-
dam, 1979.
 Bibliography

[80] L. Lovász and P. Winkler, A note on the last new vertex visited by a random
walk, J. Graph Theory 17 #5 (November 1993), 593–596.
[81] Anany Levitin and Maria Levitin, Algorithmic Puzzles, Oxford U. Press, Oxford,
UK, 2011.
[82] J. E. Littlewood, The dilemma of probability theory, Littlewood’s Miscellany,
Béla Bollobás, ed., Cambridge University Press, Cambridge, UK, 1986.
[83] A. Liu and B. Shawyer, eds., Problems from Murray Klamkin, The Canadian Col-
lection, MAA Problem Book Series, MAA Press, Providence, RI, 2009.
[84] Chris Maslanka and Steve Tribe, The Official Sherlock Puzzle Book, Woodland
Books, Salt Lake City, UT, 2018. ISBN 978-1-4521-7314-6.
[85] E. C. Milner and S. Shelah, Graphs with no unfriendly partitions, in A Tribute
to Paul Erdős, A. Baker et al., eds., Cambridge University Press, Cambridge,
UK, 1990, 373–384.
[86] Frederick Mosteller, Fifty Challenging Problems in Probability, Addison-Wesley,
Boston, MA, 1965.
[87] Colm Mulcahy, Mathematical Card Magic: Fifty-Two New Effects, CRC press,
Boca Raton, FL, 2013. ISBN: 978-14-6650-976-4.
[88] Serban Nacu and Yuval Peres, Fast simulation of new coins from old, Ann.
Appl. Probab. 15 #1A (2005), 93–115.
[89] Roger B. Nelsen, Proofs Without Words, Mathematical Association of America,
Providence, RI, 1993.
[90] W. Penney, Problem: Penney-ante, J. Recreat. Math. 2 (1969), 241.
[91] Vera Pless, Introduction to the Theory of Error-Correcting Codes, John Wiley &
Sons, Hoboken, NJ, 1982.
[92] H. Prüfer, Neuer Beweis eines Satzes über Permutationen, Arch. Math. Phys. 27
(1918), 742–744.
[93] T. F. Ramaley, Buffon’s noodle problem, Amer. Math. Monthly 76 #8 (October
1969), 916–918.
[94] M. Reiss, Über eine Steinersche combinatorische Aufgabe, welche im 45sten
Bande dieses Journals, Seite 181, gestellt worden ist, J. Reine Angew. Math. 56
(1859), 226–244.
[95] I. J. Schoenberg, Mathematical Time Exposures, Mathematical Association of
America, Providence, RI, 1982.
[96] Harold N. Shapiro, Introduction to the Theory of Numbers, Dover, Mineola, NY,
1982.
[97] D. O. Shklarsky, N. N. Chentov, and I. M. Yaglom, The USSR Problem Book, W.
H. Freeman and Co., San Francisco, 1962.
[98] R. P. Sprague, Über mathematische Kampfspiele, Tohoku Math. J. 41 (1935–
1936), 438–444.
[99] Roland Sprague, Recreation in Mathematics, Blackie & Son Ltd., London, 1963.
[100] Andrzej Szulkin, R3 is the union of disjoint circles, Amer. Math. Monthly 90 #9
(1983), 640–641.
[101] P. Vaderlind, R. Guy, and L. Larson, The Inquisitive Problem Solver, Mathemat-
ical Association of America, Providence, RI, 2002.
 Bibliography

[102] Tom Verhoeff, Circle of Lamps, Wolfram Demonstrations Project 2020.

https://demonstrations.wolfram.com/CircleOfLamps/.
[103] Stan Wagon, Fourteen proofs of a result about tiling a rectangle, Amer. Math.
Monthly 94 #7 (August–September 1987), 601–617.
[104] Chamont Wang, Sense and Nonsense of Statistical Inference, Marcel Dekker, New
York, 1993.
[105] P. Winkler, The sleeping beauty controversy, Amer. Math. Monthly 124 #7
(August–September 2017) 579–587; also in The Best Writing on Mathematics
2018, M. Pitici, ed., Princeton University Press, Princeton, NJ and Oxford,
UK, 2019, 117–129.
[106] W. A. Wythoff, A modification of the game of Nim, Nieuw Arch. Wisk. 7 (1907)
199–202.
[107] I. M. Yaglom, Geometric Transformations I, Translated by Allen Shields. Math-
ematical Association of America, Providence, RI, 1962.
[108] S. C. Young, The mental representation of geometrical knowledge, J. Math.
Behav. 3 #2 (1982), 123–144.
[109] G. Zbǎganu, A new inequality with applications in measure and information
theories, Proc. Rom. Acad. Series A 1 #1 (2000), 15–19.
[110] Yao Zhang, Combinatorial Problems in Mathematical Competitions (Vol. 4, Math-
ematical Olympiad Series), East China Normal University Press, Shanghai,
2011.

398
The Author

Peter Winkler is the William Morrill Professor of Mathematics and

Computer Science at Dartmouth College, and for 2019–2020, the Dis-
tinguished Visiting Professor for the Public Dissemination of Mathe-
matics at the National Museum of Mathematics. He is the author of
160 research papers, a dozen patents, two previous puzzle books, a
book on cryptographic techniques in the game of bridge, and a port-
folio of compositions for ragtime piano.

399
 Index
A B
Adding, multiplying, and Baby frog, xlvi, xci, 210
grouping, xxxviii, xci, Bacterial reproduction, xxiv,
155 xcii, 17
Air routes in Aerostan, xxii, xci, Bacteria on the plane, li, xci,
42 256–257
Algebra, 55–65 Badly designed clock, lxxxi, xcii,
linear, 107, 236 283–284
topology, 39 Bags of marbles, xxiv, xcii, 2
Algorithm, lxxxi, 15, 146–148 Bat and ball, xvii, xcii, 55
Bob’s, 289 Belt around the earth, xxviii,
greedy, 51, 69 xcii, 56
All right or all wrong, xli, xci, Bertrand’s postulate, 74
343–344 Biased betting, xlix, xcii,
Alternate connection, lxv, xci, 12 118–119
Alternating powers, lxvii, xci, Bias test, xlv, xcii, 170–171
106–107 Bidding in the dark, xxiii, xcii,
Alternative dice, lxvi, xci, 62–65 182
An attractive game, lxiv, xci, Big pairs in a matrix, xxv, xcii,
189–190 102–103
Angles in space, lxxiii, xci, Bindweed and honeysuckle, lxv,
229–230 xcii, 12–16
Angry baseball, lxxiv, xci, 128 Birthday match, xx, xcii,
Ants on the circle, li, xci, 110–111
297–298 Bluffing with reals, lxxi, xcii,
Area–perimeter match, xlii, 58 314–315
Ascending and descending, l, Boarding the manhole, lv, xcii,
xlix, xci, 157–158 225–227
Attention paraskevidekatria- Boardroom reduction, lxxxvi,
phobes, xxxii, xci, xcii, 323–325
4–5 Borchardt, C. W., 14
Attic lamp, xxviii, xci, 166 Box in a box, lxix, xcii, 228–229
Average, 91 Boy born on tuesday, xxi, xcii,
Axiom of choice (AC), 341–352 113–114

401
 Index

Bracing the grid, xlviii, xcii, Chomp, xxii, xciii, 101–102

45–47 Circle
Breaking a chocolate bar, ants on the, li, xci, 297–298
xxxiii–xxxiv, xcii, 255 light bulbs in a, lxxxii,
Broken ATM, xxv, xcii, 67 xcviii, 318–319
Buffon’s needle, lxxxvi, xcii, natives in a, xix, xcix, 138
197–198 points on a, xxx, ci, 116
Bugging a disk, xxix–xxx, xciii, in space, lxxx, xciii,
103–104 338–339
Bugs on a line, lxxix, xciii, Circular
264–265 shadows I, xxxii, xciii,
Bugs on a polyhedron, lxxv, 329–330
xciii, 262–263 shadows II, lxiv, xciii, 228
Bugs on a pyramid, xliv, xciii, 34 Coconut classic, lxvi–lxvii, xciii,
Bugs on four lines, lxii, xciii, 87–89
227–228 Coin
Bulgarian solitaire, lxxxix, xciii, game, lxxxv, xciii, 133–134
268–271 in a row, lxi–lxii, xciii,
86–87, 286
C on the table, lix, xciii,
Cake 83–86
candles on a, xxxvii, xciii, testing, lxxviii, xciii,
79–80 131–132
division, xxiii, xcv, 223 Colors and distances, lxxxviii,
ice cream, lxxxiii, xcvii, xciv, 200
320–322 Comparing numbers
Candles on a cake, xxxvii, xciii, version I, xxxvii–xxxviii,
79–80 xciv, 116–117
Cards from their sum, xxxiv, version II, xxxviii, xciv, 118
xciii, 68–69 Competing for programmers,
Cayley, A., 14 liii, xciv, 47–48
Chameleons, lvii–lviii, xciii, Conversation on a bus, lviii–lix,
22–23 xciv 173–174
Chessboard Conway, J. H., 236
frames on a, lxxiv, xcvi, Conway’s immobilizer, lxviii,
261–262 xciv 146–148
guess, lxxiii–lxxiv, xciii, Corner
245–246 fourth, xxviii, xcvi, 17–18
lemming on a, xlii, xcviii, invisible, lxxxv, xcviii,
104–105 339–340
Chinese Nim, lii, xciii, 308–310 pegs on the, lxxx, c, 317

402
 Index

subtracting around the, losing at, lii–liii, xcviii, 10

lxiii–lxiv, civ, 23–24 same sum, lxviii, cii,
Coupling, 123 159–160
Covering the stains, lxxxvii, xciv, set of non-transitive, 285
198–200 Discrete probability, 115
Crossing Dishwashing game, lxxii–lxxiii,
heterogeneous, 200 xciv, 122–124
homogeneous, 200 Divisibility game, xxxvii, xciv, 70
horizontal, 72 Division
the river, xxiii, xciv, easy cake, xxiii, xcv, 223
273–274 first-grade, lxii, xcvi, 259
vertical, 72 rule, 5
Cryptography, 176–178 Domino task, xxv, xciv, 77–78
Cube Dots and boxes variation, xxiv,
infected, lxxi, xcvii, xciv, 102
362–364 Dot-town exodus, 1, xcv,
magic, lxxvi, xciv, 337–338 171–173
painting, xxxix, c, 224 Double cover by lines, lii, xcv,
slicing the, xxi, ciii, 344–345
353–354 Drawing socks, 1, xcv, 185–186
spiders on a, xxxvi, civ, 43
Curve(s) E
on potatoes, xliii, xciv, Early commuter, lxiii, xcv, 12
224–225 Easy cake division, xxiii, xcv,
on a sphere, xlix, xciv, 105 223
and three shadows, Efficient pizza-cutting, xxviii,
lxxviii–lxxix, xciv, xcv, 3–4
230–233 Emptying a bucket,
unbroken, cvi, 145–146 lxxxii–lxxxiii, xcv,
Cutting 319–320
necklace, xxix, xciv, xcv, Equilibrium, 313–314
31–32 Erdős–Szekeres theorem, 157
pizza, 3–4 Even split, lxvi, xcv, 74–75
Even-sum billiards, liii, xcv,
D 21–22
Deterministic poker, lxii, xciv, Expectation, 181–204
313–314 Expecting the worst, xxviii, xcv,
Dice 354
alternative, lxvi, xci, 62–65 Exponent upon exponent, lxi,
funny, lxxxiii, xcvi, 285–286 xcv, 348–349

403
 Index

F xcvi, 261–262
Factorial(s) Functions
coincidence, lviii, xcv, 74 choice, 342
and squares, lxxii, xcv, generating, 63
75–76 parking, 367, 368
Fair play, xxiv, xcv, 275–276 potential, 255, 256, 260, 268
Falling ants, xliii, xcv, 297 Funny dice, lxxxiii, xcvi,
Faulty combination lock, xlvii, 285–286
xcv, 8–10
Fewest slopes, xxvi, xcv, 139 G
Fibonacci multiples, lxxxi, xcv, Galton, F., 64
318 Game/gaming
Fifteen bits and a spy, lxxx, xcv, attractive, lxiv, xci, 189–190
249–252 coin, lxxxv, xciii, 133–134
Figure eights in the plane, lxiv, of desperation, lx–lxi, xcvi,
xcv, 350 312
Filling dishwashing, lxxii–lxxiii,
a bucket, lx, xcvi, 279–280 xciv, 122–124
the cup, lxxxix, xcvi, divisibility, xxxvii, xciv, 70
202–204 snake, xliv, ciii, 45
Finding split, lxxiv, civ, 126–128
the counterfeit, xix, xcvi, quilt, xlvii, xcvi, 60
163–164 Garnering fruit, lxxxiv, xcvi,
a Jack, xxxv, xcvi, 96 37–40
the missing number, xxvii, Gasoline crisis, liv, xcvi, 83
xcvi, 276 Generating
the rectangles, lxiv, xcvi, functions, 63
335 the rationals, lxxxii, xcvi,
Find the robot, lxiii, xcvi, 285
349–350 Gladiators
First-grade division, lxii, xcvi, version I, lxxvi–lxxvii,
259 xcvii, 364–365
First odd number, xxxiii, xcvi, version II, lxxvii, xcvii,
140–141 365–366
Flipping the pentagon, lxxx, Gluing pyramids, lxi, xcvii,
xcvi, 265–267 333–334
Flying saucers, xlviii–xlix, xcvi, Graphography, 41–53
81–83 Greedy algorithm, 51
Fourth corner, xxviii, xcvi, Guarding the gallery, xlviii,
17–18 xcvii, 210–212
Frames on a chessboard, lxxiv,

404
 Index

H Intermediate Value Theorem

Half grown, xvii, xcvii, 1 (IVT), 29–40
Half-right hats, xxxv, xcvii, 19, Invisible corners, lxxxv, xcviii,
247–249 339–340
Halfway points, lxxii, xcvii,
107–108 K
Hammer-and-tongs approach, King’s salary, xxxviii, xcviii,
273–293 142–143
Hamming code, 235–252 Kuratowski’s theorem, 351
Ham-sandwich theorem, 39
Handshakes at a party, xlii, L
xcvii, 43–44 Lame rook, lxxviii, xcviii,
Hardy, G. H., 101, 106 219–221
Hats and infinity, xli, xcvii, 343 Laser gun, lvii, xcviii,
Hazards of electronic 358–361
coinflipping, xli–xlii, Lattice points, xxxiii, xcviii,
xcvii, 33–34 154–155
Home-field advantage, liv, xcvii, Law of large numbers, 124
119–120 Leading all the way, lxxxix,
Hopping and skipping, xxxvi, xcviii, 135–136
xcvii, 57–58 Lectures in Abstract Algebra
(Jacobson), xx, 328
I Lemming on a chessboard, xlii,
Ice cream cake, lxxxiii, xcvii, xcviii, 104–105
320–322 Life is a bowl of cherries, xxvii,
Identifying the majority, xl, xcviii, 236–238
xcvii, 276–277 Life isn’t a bowl of cherries?,
Impressionable thinkers, lxx, xxxi, xcviii, 238–239
xcvii, 260–261 Life-saving transposition,
Increasing routes, xlix, xcvii, lxxxvii, xcviii, 289
97–99 Light bulbs in a circle, lxxxii,
Induction, 205–221 xcviii, 318–319
hypothesis, 205 Line(s)
Infected bugs on a, lxxix, xciii ,
checkerboard, lxvii, xcvii, 264–265
260 bugs on four, lxii, xciii,
cubes, lxxi, xcvii–xcviii, 227–228
362–364 double cover by, lii, xcv,
Information, 163–179 344–345
Integral rectangles, lii, xcviii, through a grid, xxiii, xcviii,
357–358 153–154

405
 Index

up by height, xliii, xcviii, Moser spindle, 149

156–157 Moth’s tour, lxxxiv, xcix, 323
unbroken, xlvi, cvi, Multiples
144–145 fibonacci, lxxxi, xcv, 318
Line segments, xxxiii, xcviii, zeroes and ones, lv, cvii,
154–155 158–159
Locker doors, liv, xcviii, 73–74
Losing at dice, lii–liii, xcviii, 10 N
Lost boarding pass, xlii, xcviii, Names in boxes, lxxxiv, xcix,
80–81 288–289
Love in Kleptopia, lxxx–lxxxi, Napkins in a random setting,
xcviii, 282–283 lxxvii–lxxviii, xcix,
Lozenge, 339 194–195
Natives in a circle, xix, xcix, 138
M Negative thinking, 101–108
Magnetic dollars, xli, xcviii, Next card
356–357 bet, lxix, xcix, 190–192
Majority hats, lxxix, xcix, red, xxxii, xcix, 354–356
246–249 Nimbers, 235–252
Match Non-repeating string, xlv, xcix,
area–perimeter, xlii, xci, 58 209–210
birthday, xx, xcii, 110–111 North by northwest, lxi, xcix,
coins, lxviii–lxix, xcix, 11–12
174–176 Not burning brownies, lii, xcix,
Mathematical bookworm, xx, 332
xcix, 328 No twins today, xvii, xcix, 137
Measuring with fuses, xxxv, Numbers. See also Comparing
xcix, 141–142 numbers
Meeting the ferry, xx, xcix, 327 first odd, xxxiii, xcvi,
Meet the Williams sisters, xxxiv, 140–141
xcix, 5 on foreheads, xliv–xlv, xcix,
Men with sisters, xxiv, xcix, 93 71
Missing law of large, 124
card, xxxix, xcix, 167–169 missing, xxvii, xcvi, 276
digit, lxi, xcix, 23 rolling all the, xxvi, cii, 183
Monge point, 231, 232 safety, 67–76
Monge’s theorem, 233 self-referential, lxxxix, ciii,
Monk on a mountain, xx, xcix, 289–293
30–31 skipping a, xlv, ciii, 35
More magnetic dollars, lxxxvii, small, 77–89
xcix, 366–369

406
 Index

O halfway, lxxii, xcvii,

Odd light flips, xxxix, c, 207–208 107–108
Odd run of heads, xxix, c, 56 lattice, xxxiii, xcviii,
One-bulb room, lxx, c, 280–281 154–155
Option hats, lxix–lxx, c, 241–245 Monge, 231, 232
Other side of the coin, xxi, c, red and blue, xxxix, cii,
111–112 255–256
system, 299
P two-point conversion,
Packing slashes, xliii, c, 143–144 lxxv–lxxvi, cv, 128–129
Painting Poker quickie, xxiii, ci, 138
cubes, xxxix, c, 224 Polite pizza protocol, 286
fence, lxxxviii, c, 201–202 Polyhedron
polyhedron, li, c, 225 bugs on, lxxv, xciii, 262–263
Pairs at maximum distance, faces, xix, ci, 153
xxxvii, c, 104 painting, li, c, 225
Pancake stacks, xl, c, 307–308 Polygon
Parity, 17–27 on the grid, lx, ci, 280
Path through the cells, l, c, midpoints, xliv, ci, 331
212–213 Poorly placed dominoes, li, ci,
Peek advantage, xliii, c, 169–170 277–278
Pegs Portrait, xvii, ci, 303
on the corners, lxxx, c, 317 Potentials, 253–271
on the half-plane, lvi, c, Powers
257–258 alternating, lxvii, xci,
in a square, lx, c, 258–259 106–107
Pentagon, flipping the, lxxx, of roots, lxiv, ci, 87
xcvi, 265–267 of two, xviii, ci, 1–2
Permutations, 20–21 Precarious picture, lxiii, ci,
Phone call, xviii, c, 273 334–335
Picking the athletic committee, Prime test, xlii, ci, 70
lxxxiii, c, 267–268 Prisoner(s)
Pie in the sky, xxxii, ci, 295–296 and dog, lxxix, ci, 281–282
Pigeonhole principle, 151–162 and gloves, xlvii, ci, 20–21
Ping-pong progression, xxxix, ci, Probabilistically checkable
185 proofs, 252
Players and winners, xxx, ci, Probability, 109–136
166–167 discrete, 115
Points reverse conditional, 112
on a circle, xxx, ci, 116 whose bullet? xxii, cvi, 114

407
 Index

Profit and loss, lv, ci, 213–215 all the numbers, xxvi, cii,
Protecting the statue, lvi–lvii, ci, 183
332–333 pencil, xxi, cii, 328
Pyramid a six, lxxvii, cii, 193–194
bugs on, xliv, xciii, 34 Rotating coin, xviii, cii, 295
gluing, lxi, xcvii, 333–334 Roulette
for parking money, lxxxv,
R cii, 196–197
Raising art value, xxxiii, ci, for the unwary, lix, cii,
94–95 188–189
Ramsey’s theorem, 202, 203, Rows and columns, xliv, cii, 7
220, 221
Random S
bias, lxxvi, ci, 131 Salaries and raises, xxv, cii, 2–3
chord, lxxvi, ci, 130–131 Same sum
intersection, lv, cii, 186–188 dice, lxviii, cii, 159–160
intervals, lxi, cii, 361–362 subsets, xxx, ciii, 154
judge, lxxiii, cii, 124 Second ace, xxvi, ciii, 114–115
variables, 181 Self-referential number, lxxxix,
Rating the horses, xxxv, cii, 5–6 ciii, 289–293
Rationals, generating, lxxxii, Sequencing the digits, xlvi, ciii, 8
xcvi, 285 Serious candidates, lxxv, ciii,
Recovering the polynomial, 192–193
xlvii, cii, 59 Service options, lxvii–lxviii, ciii,
Rectangles tall and wide, xlviii, 121
cii, 71–73 Set of non-transitive dice, 285
Red Seven cities of gold, lxxxvi, ciii,
and black sides, xlv–xlvi, 148–150
cii, 58–59 Sharing a pizza, lxxxiv, ciii,
and blue hats in a line, 286–287
xxxv–xxxvi, cii, 19–20 Shoelaces at the airport, xli, ciii,
points and blue, xxxix, cii, 6
255–256 Shoes, socks, and gloves, xviii,
Returning pool shot, xl, cii, ciii, 151–152
296–297 Signs in an array, xix, ciii,
Reverse conditional 253–254
probabilities, 112 Sinking 15, xx, ciii, 353
Reversibility theorem, 317 Skipping a number, xlv, ciii, 35
Righting the pancakes, xxxi, cii, Slabs in 3-space, lx, ciii, 227
254–255 Sleeping beauty, lxxxv–lxxxvi,
Rolling ciii, 134–135

408
 Index

Slicing the cube, xxi, ciii, Subtracting around the corner,

353–354 lxiii–lxiv, civ, 23–24
Small numbers, 77–89 Summing fractions, lxix, civ, 216
Snake game, xliv, ciii, 45 Sums
puzzle, 52 cards from their, xxxiv,
Soldiers in a field, liv, ciii, xciii, 68–69
105–106 and differences, lxxv, civ,
Space 281
angles in, lxxiii, xci, of two squares, xxxvii, civ,
229–230 96–97
circles in, lxxx, xciii, Swapping executives, xxxi, civ,
338–339 206–207
slabs in 3-space, lx, ciii, 227 Swedish lottery, lxxvi, civ,
three-dimensional, 223–233 316–317
Spaghetti loops, xxx, ciii, System point, 299
183–184
Sphere and quadrilateral, lxx, T
ciii, 298 Testing ostrich eggs, xxxii, civ,
Spiders on a cube, xxxvi, civ, 304–306
42–43 Three-dimensional space,
Spinning switches, xxix, civ, 223–233
78–79 Three natives at the crossroads,
Split games, lxxiv, civ, 126–128 xxvi, civ, 165–166
Splitting Three negatives, xliv, cv, 58
hexagon, xxvii, civ, 329 Three sticks, xxxvi, cv, 32–33
polygon, lxvi, civ, 36–37 Three-way
stacks, lvii, civ, 11 duel, xxvii, cv, 303–304
Sprague-Grundy theorem, 238 election, xlv, cv, 7–8
Sprinklers in a field, xxiv, civ, Tiling
275 a polygon, lxv, cv, 335–336
Squaring the mountain state, with crosses, lxxi, cv,
xix, civ, 29–30 336–337
Stone–Tukey theorem, 39 with L’s, lxxii, cv, 216–218
Stopping after the boy, xxviii, Tipping the scales, xxii, cv, 92
civ, 115–116 Touring an island, lxxxi, cv, 317
Strategy-stealing argument, 102 Transformation, 353–369
Strength of schedule, liii, civ, Trapped in thickland, xl, cv,
60–61 330–331
Subsets with constraints, xxix, Traveling salesmen, lxxvii, cv,
civ, 67–68 218–219

409
 Index

Truly even split, xliii, cv, Vertices, 351

208–209 Von Neumann’s trick, 275–276
Turning the die, lvi, cv, 310–312
Two balls and a wall, lxxxix, cv, W
299–302 Waiting for heads, xxxiv, cvi,
Two different distances, xxx, cv, 95–96
139–140 Watches on the table, xxix, cvi,
Two monks on a mountain, lxvi, 93–94
cv, 49–50 Watermelons, xxi, cvi, 2
Two-point conversion, Watson, H. W., 64
lxxv–lxxvi, cv, 128–129 Weighs and means, 91–99
Two round-robins, lviii, cv, Whim-Nim, li, 239–241
61–62 Whose bullet?, xxii, cvi, 114
Two runners, xxv, cv, 55 Who won the series?, lxxii, cvi,
Two sheriffs, lxxxviii, cv, 121–122
176–179 Wild guess, lvii, cvi, 345–348
Winning at Wimbledon, xix, cvi,
U 110
Unanimous hats, xxxii–xxxiii, Wins in a row, lxxiii, cvi, 125
cvi, 18–19, 248, 249 Wires under the Hudson, lviii,
Unbreakable domino cover, lvi, cvi, 48–49
cvi, 278–279 Worms and water, lxxxii, cvi,
Unbroken 284–285
curves, l, cvi, 145–146 Worst route, lxxi–lxxii, cvi,
lines, xlvi, cvi, 144–145 51–53
Unfriendly partition, 270–271
Uniformity at the bakery, lix, Y
cvi, 215–216 Y’s in the plane, lxxi, cvi,
Uniform unit distances, xxvi, 350–352
cvi, 205–206
Uniting the loops, lxv, cvi, Z
24–27 Zeroes and ones, lv, cvii,
158–159
V Zero-sum vectors,
Variation, dots and boxes, xxiv, lxxxvii–lxxxviii, cvi,
102 161–162

410