monthly

THE AMERICAN MATHEMATICAL

VOLUME 123, NO. 10 DECEMBER 2016

A Letter from the Editor: The Long and Winding Road 955
Scott T. Chapman

Sets of Lengths 960

Alfred Geroldinger

Balanced Factorizations 989

Anton A. Klyachko and Anton N. Vassilyev

Cantor-Polynomials and the Fueter-Pólya Theorem 1001

Melvyn B. Nathanson

Optimizing the Video Game Multi-Jump: Player Strategy,

AI, and Level Design 1013
Aaron M. Broussard, Martin E. Malandro, and Abagayle Serreyn

NOTES
Dilated Floor Functions that Commute 1033
Jeffrey C. Lagarias, Takumi Murayama, and D. Harry Richman

Quotients of Fibonacci Numbers 1039

Stephan Ramon Garcia and Florian Luca

Parking Cars of Different Sizes 1045

Richard Ehrenborg and Alex Happ

PROBLEMS AND SOLUTIONS 1050

BOOK REVIEW
A First Course in the Calculus of Variations 1058
by Mark Kot
Chris A. Marx

Referee Thank You 1062

MATHBITS
959, Single Digit NFL Scores; 1049, Corrigendum to “A New
Proof of the Change of Variable Theorem for the Riemann
Integral”

An Official Publication of the Mathematical Association of America

The MAA is pleased to announce the release of a NEW

Carus Mathematical Monograph

Linear Inverse Problems and
Tikhonov Regularization
Mark S. Gockenbach

Inverse problems occur frequently in

science and technology, whenever we need
to infer causes from effects that we can
measure. Mathematically, they are difficult
problems because they are unstable: small
bits of noise in the measurement can completely throw off the solu-
tion. Nevertheless, there are methods for finding good approximate
solutions.
Linear Inverse Problems and Tikhonov Regularization examines
one such method: Tikhonov regularization for linear inverse prob-
lems defined on Hilbert spaces. This is a clear example of the power
of applying deep mathematical theory to solve practical problems.
Beginning with a basic analysis of Tikhonov regularization, this
book introduces the singular value expansion for compact opera-
tors, and uses it to explain why and how the method works. Tikhon-
ov regularization with seminorms is also analyzed, which requires
introducing densely defined unbounded operators and their basic
properties. Some of the relevant background is included in appendi-
ces, making the book accessible to a wide range of readers.
Catalog Code: CAM32 Print ISBN: 978-0-88385-141-8
List Price: $63.00 e-ISBN: 978-1-61444-029-1
Member Price: $47.25 334 pp., Hardbound & ePDF, 2016
ebook: $28.00 Carus Mathematical Monographs
To order a print book, visit store.maa.org or call 1-800-331-1622.
To order the electronic book, visit www.maa.org/ebooks/CAM32.
monthly
THE AMERICAN MATHEMATICAL

VOLUME 123, NO. 10 DECEMBER 2016

EDITOR
Scott T. Chapman
Sam Houston State University

EDITOR-ELECT NOTES EDITOR BOOK REVIEW EDITOR

Susan Colley Sergei Tabachnikov Jeffrey Nunemacher
Oberlin College Pennsylvania State University Ohio Wesleyan University

PROBLEM SECTION EDITORS

Douglas B. West Gerald Edgar Doug Hensley
University of Illinois Ohio State University Texas A&M University

ASSOCIATE EDITORS
William Adkins Jeffrey Lawson
Louisiana State University Western Carolina University
David Aldous C. Dwight Lahr
University of California, Berkeley Dartmouth College
Elizabeth Allman Susan Loepp
University of Alaska, Fairbanks Williams College
Jonathan M. Borwein Irina Mitrea
University of Newcastle Temple University
Jason Boynton Bruce P. Palka
North Dakota State University National Science Foundation
Edward B. Burger Vadim Ponomarenko
Southwestern University San Diego State University
Minerva Cordero-Epperson Catherine A. Roberts
University of Texas, Arlington College of the Holy Cross
Allan Donsig Rachel Roberts
University of Nebraska, Lincoln Washington University, St. Louis
Michael Dorff Ivelisse M. Rubio
Brigham Young University Universidad de Puerto Rico, Rio Piedras
Daniela Ferrero Adriana Salerno
Texas State University Bates College
Luis David Garcia-Puente Edward Scheinerman
Sam Houston State University Johns Hopkins University
Sidney Graham Anne Shepler
Central Michigan University University of North Texas
Tara Holm Frank Sottile
Cornell University Texas A&M University
Lea Jenkins Susan G. Staples
Clemson University Texas Christian University
Daniel Krashen Daniel Ullman
University of Georgia George Washington University
Ulrich Krause Daniel Velleman
Universität Bremen Amherst College
Steven Weintraub
Lehigh University

ASSISTANT MANAGING EDITOR MANAGING EDITOR

Bonnie K. Ponce Beverly Joy Ruedi
NOTICE TO AUTHORS
The MONTHLY publishes articles, as well as notes and other fea-
tures, about mathematics and the profession. Its readers span
a broad spectrum of mathematical interests, and include pro-
fessional mathematicians as well as students of mathematics
at all collegiate levels. Authors are invited to submit articles
and notes that bring interesting mathematical ideas to a wide
audience of MONTHLY readers.
The MONTHLY’s readers expect a high standard of exposition;
they expect articles to inform, stimulate, challenge, enlighten,
and even entertain. MONTHLY articles are meant to be read, en-
joyed, and discussed, rather than just archived. Articles may
be expositions of old or new results, historical or biographical
essays, speculations or definitive treatments, broad develop-
ments, or explorations of a single application. Novelty and
generality are far less important than clarity of exposition
and broad appeal. Appropriate figures, diagrams, and photo-
graphs are encouraged.
Notes are short, sharply focused, and possibly informal. They
are often gems that provide a new proof of an old theorem, a
novel presentation of a familiar theme, or a lively discussion
of a single issue.
Submission of articles, notes, and filler pieces is required via the
MONTHLY’s Editorial Manager System. Initial submissions in pdf or
LATEX form can be sent to the Editor-Elect Susan Colley at
www.editorialmanager.com/monthly
The Editorial Manager System will cue the author for all re-
quired information concerning the paper. The MONTHLY has in-
stituted a double blind refereeing policy. Manuscripts which
contain the author’s names will be returned. Questions con-
cerning submission of papers can be addressed to the Editor-
Elect at monthly@maa.org. Authors who use LATEX can find
our article/note template at www.maa.org/monthly.html.
This template requires the style file maa-monthly.sty, which
can also be downloaded from the same webpage. A format-
ting document for MONTHLY references can be found there too.
Letters to the Editor on any topic are invited. Comments, criti-
cisms, and suggestions for making the MONTHLY more lively,
entertaining, and informative can be forwarded to the Editor-
Elect at monthly@maa.org
The online MONTHLY archive at www.jstor.org is a valuable
resource for both authors and readers; it may be searched
online in a variety of ways for any specified keyword(s). MAA
members whose institutions do not provide JSTOR access
may obtain individual access for a modest annual fee; call
800-331-1622 for more information.
See the MONTHLY section of MAA Online for current informa-
tion such as contents of issues and descriptive summaries of
forthcoming articles:
www.maa.org/monthly.html
Proposed problems or solutions should be sent to:
DOUG HENSLEY, MONTHLY Problems
Department of Mathematics
Texas A&M University
3368 TAMU
College Station, TX 77843-3368
In lieu of duplicate hardcopy, authors may submit pdfs to
monthlyproblems@math.tamu.edu.
Advertising correspondence should be sent to:
MAA Advertising
1529 Eighteenth St. NW
Washington DC 20036
Phone: (202) 319-8461
E-mail: advertising@maa.org
Further advertising information can be found online at www.
maa.org.
Change of address, missing issue inquiries, and other sub-
scription correspondence can be sent to:
maaservice@maa.org
or
The MAA Customer Service Center
P.O. Box 91112
Washington, DC 20090-1112
(800) 331-1622
(301) 617-7800
Recent copies of the MONTHLY are available for purchase
through the MAA Service Center at the address above.
Microfilm Editions are available at: University Microfilms In-
ternational, Serial Bid coordinator, 300 North Zeeb Road, Ann
Arbor, MI 48106.
The AMERICAN MATHEMATICAL MONTHLY (ISSN 0002-9890) is
published monthly except bimonthly June-July and August-
September by the Mathematical Association of America at
1529 Eighteenth Street, NW, Washington, DC 20036 and Lan-
caster, PA, and copyrighted by the Mathematical Association
of America (Incorporated), 2015, including rights to this jour-
nal issue as a whole and, except where otherwise noted, rights
to each individual contribution. Permission to make copies
of individual articles, in paper or electronic form, including
posting on personal and class web pages, for educational and
scientific use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage
and that copies bear the following copyright notice: [Copy-
right 2016 Mathematical Association of America. All rights
reserved.] Abstracting, with credit, is permitted. To copy
otherwise, or to republish, requires specific permission of the
MAA’s Director of Publications and possibly a fee. Periodicals
postage paid at Washington, DC, and additional mailing offic-
es. Postmaster: Send address changes to the American Math-
ematical Monthly, Membership/Subscription Department,
MAA, 1529 Eighteenth Street, NW, Washington, DC 20036-1385.
 A Letter from the Editor:
The Long and Winding Road
Scott T. Chapman

With this issue, my journey as Editor of the M ONTHLY comes to a close. It began
over 6 years ago in a McDonald’s. While I was eating something that both my wife
and physician would tell me is bad for my cholesterol, my cell phone rang. John Ewing
was on the other end with the news that I would be the next Editor of the M ONTHLY. At
the time, it was impossible for me to know how life altering that phone call was. This
became much more apparent several weeks later when I met with John in New York
City. From across a table in a very Seinfeldian coffee shop, John told me that the job
of editing the M ONTHLY was perhaps one of the most difficult in all of mathematics.
Anyone who has held this position would quickly agree with John, but hard work has
its rewards. I have learned that such rewards at the M ONTHLY are almost countless.
Six years (1 as Editor-Elect and 5 as Editor) is a long period, and yes, the road has
been not only long but winding:
• Over 25,000 miles in travel to both Sectional and National meetings;
• 4,669 submissions (as of the writing of this letter, all but 2 have received a final
decision);
• 5,319 reviewers invited;
• 2,942 reviews received;
• 1,406 revisions received (and yes, on one paper we requested 7 revisions);
• During this period, our Editorial Manager System handled over 80,000 electronic
mail messages.
I am happy to report 2 positive statistics related to this dizzying list of numbers:
• For our 4,669 submissions, the average number of days between the receipt of the
manuscript and our first decision was 37.6 days;
• For our 2,942 reviews received, the average number of days between the date that
the reviewer agreed to the review and date that the review was submitted was 28.4
days.
The terms of every past M ONTHLY Editor have wound through challenges. I am
lucky; my challenges have paled in comparison to some of my predecessors. A quick
review of [3] (and its cited references) shows that our founder Benjamin Finkel spent
a great deal of his time during the years 1894–1909 looking for funding to keep this
publication afloat. In 1912, Hebert Slaught was so distressed about the current state
of the M ONTHLY that he wrote to a colleague “I cannot long stand the pressure. I
must either put the M ONTHLY on a different basis or stand from under.” [3]. I was
particularly struck by a parting letter from Lester Ford [2] which describes an attempt
by the government during World War II to cut from 2 to 1 the number of staples holding
each issue together.
My first challenge was to assemble an Editorial Board that could not only handle
papers throughout the breadth of the Mathematical Sciences Classification Index, but
http://dx.doi.org/10.4169/amer.math.monthly.123.10.955

December 2016] A LETTER FROM THE EDITOR 955

also reflect the diversity of the MAA membership. I cannot say enough good things
about my Editorial Board. Their devotion, patience, and hard work (above and beyond
the call of duty) have led us to whatever success we have enjoyed. While space does
not allow me to thank each one individually, there are several that deserve special
thanks: Sergei Tabachnikov (Notes Editor), Jeff Nunemacher (Book Reviews Editor),
and Doug Hensley, Doug West, and Gerald Edgar (Problem Section Editors). I am
happy to see that Editor-Elect Susan Colley has retained a large portion of my Board
for the 2017–2021 Editorial Board, and I am also happy to accept her invitation to
remain on the Board myself.
My second challenge was to pull the M ONTHLY into the 21st century and move
submissions and all major editorial processes to a web based system. Our use of Aries
Systems Editorial Manager began on January 1, 2012. While the end result was suc-
cessful, the road was not without bumps—I distinctly remember during the first two
months of my editorship receiving an email that began “Chapman, turn off your robot.”
My third challenge involved the MAA’s adoption of “double-blind” reviewing.
After much general discussion within the MAA, I appointed a subcommittee of the
Editorial Board to examine the possible adoption of the double-blind system for
the M ONTHLY. The committee’s report, forwarded to the Board of Governors on
June 29, 2012, recommended against the adoption of the double-blind system for the
M ONTHLY, and the M ONTHLY Editorial Board voted 33-2 (with 3 abstentions) to
support the report. At the August 2012 Board of Governors meeting, the Governors
voted overwhelmingly to move all MAA journals to the double-blind system. The
wording of the actual proposal passed by the Governors allowed the current MAA
journal editors to keep their current refereeing system, and hence double-blind was
not actually mandated until there was a change in Editor. During 2013 and 2014, I
kept the M ONTHLY on a traditional manuscript reviewing system. While I was under
no obligation to do so, I did move the M ONTHLY during 2015 to the double-blind
platform. My hope was that this would make the transition to a new Editor-Elect (and
eventually Editor) a much smoother process. While I found that many of the theorized
problems with moving to double-blind were actually easy to solve, others raised eth-
ical issues which require much more thought and debate. I am happy to report that
my predecessor Dan Velleman is chairing an MAA Taskforce which will consider
policies designed to resolve such ethical matters. While one year of data surrounding
the M ONTHLY’s double-blind excursion is likely not enough to draw any hard conclu-
sions, I do note that submission and acceptance statistics from 2015 are completely in
line with what we saw in 2011, 2012, 2013, and 2014.
My fourth challenge was completely unexpected, but consumed a great deal of the
last two years of my editorship. Sometime during the Spring of 2014 our longtime
typesetters, managed by Pearson Education, terminated their production agreement
with the MAA. Not only did this end a long business relationship, but it forced us
to quickly find a replacement. Early in the summer of 2014, the MAA reached an
agreement with Cenveo Publishing Services to replace Pearson. Due to the unique
nature of the M ONTHLY, our production processes are much different than those of
a traditional research journal, and needless to say, our transition to Cenveo has not
exactly been smooth. The M ONTHLY (as well as its two MAA sister journals) fell well
behind our normal publication schedule. I was able in late 2015 to get the release dates
for the usual issues of the M ONTHLY back to normal, but Cenveo again fell behind and
it was not until late this year that the M ONTHLY started to arrive to its readers again
on time.
My final challenge is one that I have shared with each of my recent predecessors—to
not only keep the M ONTHLY as the world’s preeminent source for mathematical expo-

956 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
sition, but to improve it and expand our reader base. Most of the new things I tried
were well received. We produced 2 special issues. The first in March 2013 contained
papers written by speakers at the 2011 International Summer School for Students at
Jacobs University in Bremen, Germany. Three of this volume’s papers were finalists
for the 2013 Ford–Halmos Award. The second in November 2014 was dedicated to
Mathematical Biology. The M ONTHLY has traditionally published very few papers in
applied mathematics, and with the almost explosive amount of interest in Mathemat-
ical Biology, I saw this as an opportunity for the M ONTHLY to open new doors. To
honor Lloyd Shapley’s 2013 Nobel Prize in Economics, we reprinted in May of 2013
his classic M ONTHLY paper (co-authored with David Gale) College Admissions and
the Stability of Marriage, which had been cited by the Nobel Committee. As of the
writing of this letter, the Shapley–Gale paper has been cited over 4,300 times. At the
Centennial MathFest in Washington, I organized with the 4 other living M ONTHLY
Editors a special session titled “Generations of M ONTHLY Gems,” which celebrated
the M ONTHLY’s history and its impact on the development of our Association.
I have thanked my Editorial Board, but this only scratches the surface of the
acknowledgments which I owe. I begin at the top and thank the 4 MAA Presidents
under which I served: David Bressoud, Paul Zorn, Bob Devaney, and Francis Su.
Special thanks go to Tina Straley and Michael Pearson, the MAA Executive Directors
during my term. Ivars Peterson, Bev Ruedi, and the entire publications staff (past and
present) at the MAA deserve special recognition for their tireless efforts to make the
MAA journals so outstanding. Sam Houston State University was more than wel-
coming to the M ONTHLY, and among those I should thank are Jaimie Hebert, John
Pascarella, and Brian Loft. Due to her outstanding performance as my Editorial Assis-
tant, Bonnie Ponce has been hired permanently by the MAA as Assistant Managing
Editor for Journals. Thank you Bonnie, as I could not have done this without you. Last
and certainly not least, I thank my wife Lenora and sons Jonathan and Cameron for
putting up with everything associated with the scary list of numbers you saw on the
first page.
Challenges have been a theme of this letter, and I part with a challenge to the
M ONTHLY, its readers, and the general membership of the MAA. I recently spoke
at the Intermountain Section Spring meeting, and during a session of Math Jeopardy,
the following question came up: How many issues a year does the American Mathe-
matical M ONTHLY publish? None of the contestants knew the answer (which is 10).
With a little more digging, it became clear to me that many MAA members (especially
our younger ones) know little about the M ONTHLY. In many respects, the M ONTHLY
is the MAA; the Association was founded to run the M ONTHLY. The M ONTHLY is the
crown jewel of the MAA—if you don’t know much about it, then learn more. Go to
your library and find a copy of John Ewing’s tribute to the M ONTHLY’s first hundred
years [1]. Revel in the history of the M ONTHLY and its unique standing in the math-
ematical community. Its pages are filled with papers by not only Nobel Prize winners
and Field’s medalists, but faculty and students from every conceivable level of the
educational spectrum.
I am excited about the future of the M ONTHLY. I hand the baton off to Susan Colley,
who has already demonstrated that she will be an outstanding editor. I look forward to
working closely with her throughout the years of her term.
It has been an incredible ride and an honor for me to guide the reins of the
M ONTHLY. I guess one never knows what might start in a McDonald’s.

December 2016] A LETTER FROM THE EDITOR 957

REFERENCES

1. J. Ewing, A Century of Mathematics: Through the Eyes of the M ONTHLY. Cambridge University Press,
Cambridge, 1996.
2. L. R. Ford, Retrospect, Amer. Math. M ONTHLY 53 (1946) 582–585.
3. T. H. Straley, A History of the American Mathematical M ONTHLY, http://www.maa.org/
history-of-the-american-mathematical-monthly.

958 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 Single Digit NFL Scores
Traditional NFL scoring dictates that a team cannot exclusively score 1 point (this
can occur only after a touchdown). While a team can score any number of points
equal to a linear combination of 2 and 3 over the nonnegative integers, what single
digit combinations have actually occurred? The following table answers this question
with data complied from [2]. In the table an “X” indicates that the score has never
occurred and an integer indicates the frequency with which it has occurred (since the
inception of the NFL in 1920). For instance, the score 3-2 has occurred twice.

0 2 3 4 5 6 7 8 9
0 73
2 5 X
3 59 2 13
4 X X X X
5 3 X 1 X X
6 77 2 34 X X 16
7 93 3 30 X 1 56 31
8 5 X 1 X 1 1 X X
9 29 1 16 X 2 28 39 X X

This table might be a bit disappointing for fans of the digit 4, but do not despair, an
NFL team has once scored a total of 4 points in a game. On November 25, 1923,
the game between the Racine Legion and Chicago Cardinals ended in a 10-4 Racine
victory. There is also hope for you M ONTHLY readers who like the digit 1. It is now
possible for an NFL team to score a single point (see [1]), but as of the submission
of this piece, it has never occurred. — submitted by Franco Bradshaw.

REFERENCES

1. N. F. L. Team Scoring Just 1 Point? Now It’s Possible, The New York Times, September 10, 2015.
2. All Game Scores in Pro Football History, Pro Football Reference, http://www.
pro-football-reference.com/boxscores/game-scores.htm.

http://dx.doi.org/10.4169/amer.math.monthly.123.10.959
MSC: Primary 00A99

December 2016] A LETTER FROM THE EDITOR 959

 Sets of Lengths
Alfred Geroldinger

Abstract. Oftentimes the elements of a ring or semigroup can be written as finite products
of irreducible elements. An element a can be a product of k irreducibles and a product of 
irreducibles. The set L(a) of all possible factorization lengths of a is called the set of lengths
of a, and the system consisting of all these sets L(a) is a well-studied means of describing the
nonuniqueness of factorizations of a ring or semigroup. We provide a friendly introduction,
which is largely self-contained, to what is known about systems of sets of lengths for rings
of integers of algebraic number fields and for transfer Krull monoids of finite type as their
generalization.

1. INTRODUCTION. We all know that every positive integer can be written as a

finite product of irreducibles (primes) and that such a factorization is unique up to
the order of appearance. Similar to factorizations in the positive integers, in many
rings and semigroups, elements can also be written as finite products of irreducible
elements, but unlike the case of the integers, such factorizations need not always be
unique. It is the main objective of factorization theory to describe the various aspects
of nonuniqueness and to classify them in terms of invariants of the underlying alge-
braic structure. Before it was extended to commutative ring and semigroup theory,
factorization theory had its origin in algebraic number theory, and only in recent years
has been extended to noncommutative settings [4, 50]. For further background, we
refer the reader to several monographs and conference proceedings [1, 11, 16, 23, 28].
It is no surprise that this development has been chronicled over the years by a series
of Monthly articles (from [38] to [5, 6]). While the focus of this interest has been on
commutative domains and their semigroups of ideals, such studies range from abstract
semigroup theory to the factorization theory of motion polynomials with application
in mechanism science [37].
Sets of lengths are the most investigated invariants in factorization theory. To fix
notation, if an element a in a semigroup can be written as a product of irreducible
elements, say a = u 1 · . . . · u k , then k is called the length of the factorization, and the
set L(a) ⊂ N of all possible factorization lengths of a is called the set of lengths of a.
Under a mild condition on the semigroup, sets of lengths are finite nonempty subsets
of N0 , and if there is an element a in the semigroup with |L(a)| > 1 (meaning that a
has factorizations of distinct lengths), then sets of lengths can get arbitrarily long (a
precise statement of this is in Lemmas 1 and 2).
The goal of this paper is to give a friendly introduction to factorization theory.
Indeed, we take the reader on a tour through sets of lengths which is highlighted by
two structure theorems (Theorems 5 and 15), two open problems (Problem 14 and
the characterization problem at the beginning of Section 6), and a conjecture (Con-
jecture 24). In Section 2 we introduce sets of distances and unions of sets of lengths.
We provide a full and self-contained proof for the structure theorem for unions of
sets of lengths (Theorem 5), and outline an argument that finitely generated commuta-
tive monoids satisfy all assumptions of that structure theorem. In Section 3 we discuss
commutative Krull monoids. This class includes Dedekind domains and hence rings of
http://dx.doi.org/10.4169/amer.math.monthly.123.10.960
MSC: Primary 20M13, Secondary 11B30; 11R27

960 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
integers of algebraic number fields, and we will provide an extended list of examples
stemming from a variety of mathematical areas. The central strategy for studying sets
of lengths in a given class of semigroups is to construct homomorphisms (called trans-
fer homomorphisms) which can be used to transfer analogous results in a simpler
class of semigroups directly back to the more complex class. In Section 4 we discuss
transfer homomorphisms, show that they preserve sets of lengths, and provide a self-
contained proof for the fact that there is a transfer homomorphism from commutative
Krull monoids to monoids of zero-sum sequences (which are monoids having a com-
binatorial flavor that will often be the simpler class of semigroups to which the more
complex semigroup is reduced). We provide an extended list of transfer Krull monoids
(these are the monoids allowing a transfer homomorphism to a monoid of zero-sum
sequences), and then we restrict our discussion to this class of monoids. In Section 5
we discuss the structure theorem for sets of lengths and provide examples showing that
all aspects addressed in the structure theorem occur naturally. In Section 6 we discuss
sets of lengths of the monoid of zero-sum sequences over a finite abelian group (the
transfer machinery of Section 4 guarantees that these sets of lengths coincide with the
sets of lengths of a ring of integers). They can be studied with methods from additive
combinatorics and their structure is by far the best understood (among all classes of
monoids). In this setting, unions of sets of lengths and the set of distances are intervals
and they have natural upper bounds (Proposition 18). In spite of the fact that almost all
(in a certain sense) sets of lengths are intervals (Theorem 20) we conjecture that the
system of sets of lengths is a characteristic for the group (Conjecture 24). In order to
keep this article as self-contained as possible, we do not mention arithmetical concepts
beyond sets of lengths (such as catenary and tame degrees), or factorization theory in
rings with zero-divisors, or divisibility theory in nonatomic rings.

2. BASIC NOTATION AND UNIONS OF SETS OF LENGTHS. We denote by

N the set of positive integers and set N0 = N ∪ {0}. For integers a, b ∈ Z, we denote
by [a, b] = {x ∈ Z | a ≤ x ≤ b} the (discrete) interval between a and b, and by an
interval we always mean a set of this form. Let L , L  ⊂ Z be subsets of the integers.
Then L + L  = {a + b | a ∈ L , b ∈ L  } is the sumset of L and L  . Thus we have
L + ∅ = ∅, and we set −L = {−a | a ∈ L}. For an integer m ∈ Z, m + L = {m} + L
is the shift of L by m. For k ∈ N, we denote by k L = L + · · · + L the k-fold sumset of
L and by k · L = {ka | a ∈ L} the dilation of L by k. A positive integer d ∈ N is called
a distance of L if there are k,  ∈ L with  − k = d and the interval [k, ] contains no
further elements of L. We denote by (L) ⊂ N the set of distances of L. By definition
we have (L) = ∅ if and only if |L| ≤ 1, and L is an arithmetical progression if and
only if |(L)| ≤ 1 . For L ⊂ N, we denote by ρ(L) = sup L/ min L ∈ Q≥1 ∪ {∞} the
elasticity of L and we set ρ({0}) = 1.
By a semigroup we always mean an associative semigroup, and if not stated other-
wise, we use multiplicative notation. Let S be a semigroup. We say that S is cancelative
if for all elements a, b, c ∈ S, the equation ab = ac implies b = c and the equation
ba = ca implies b = c. All rings and semigroups are supposed to have an identity,
and all ring and semigroup homomorphisms preserve the identity. By a monoid, we
mean a cancelative semigroup. Clearly subsemigroups of groups are monoids, and
finite monoids are groups. If aS ∩ bS = ∅ and Sa ∩ Sb = ∅ for all a, b ∈ S, then S
has a (unique left and right) quotient group which will be denoted by q(S). If R is
a ring, then the set of cancelative elements R • is a monoid. A domain D is a ring in
which zero is the only zero-divisor (i.e., D • = D \ {0}). We use the abbreviation ACC
for the ascending chain condition on ideals.

December 2016] SETS OF LENGTHS 961

 Let P be a set. We denote by F ∗ (P) the free monoid with basis P, the elements of
which may be viewed as words on the alphabet P. We denote by | · | : F ∗ (P) → N0
the function which maps each word onto its length. The free abelian monoid with basis
P will be denoted by F (P). Every a ∈ F (P) has a unique representation of the form

a= pν p , where ν p ∈ N0 and ν p = 0 for almost all p ∈ P.
p∈P

Therefore, for every p ∈ P, there is a homomorphism (called the p-adic exponent)

v p : F (P) → N0 defined by v p (a) = ν p . Similar to the case of free monoids, we
denote by | · | : F (P) → N0 the usual length function, and we observe that |a| =

p∈P v p (a) for all a ∈ F (P).
Let H be a monoid and let a, b ∈ H . The element a is said to be invertible if there
exists an element a  ∈ H such that aa  = a  a = 1. The set of invertible elements of H
will be denoted by H × , and we say that H is reduced if H × = {1}.
The element a ∈ H is called irreducible (or an atom) if a ∈ / H × and, for all u, v ∈
× ×
H , a = uv implies that u ∈ H or v ∈ H . The monoid H is said to be atomic if every
a ∈ H \ H × is a product of finitely many atoms of H . While most integral domains
introduced in elementary courses are atomic, not all such algebraic objects are. An
elementary example of a nonatomic monoid can be found in [10, p. 166]. If a ∈ H and
a = u 1 · . . . · u k , where k ∈ N and u 1 , . . . , u k ∈ A(H ), then we say that k is the length
of the factorization. For a ∈ H \ H × , we call

L H (a) = L(a) = {k ∈ N | a has a factorization of length k} ⊂ N

the set of lengths of a. For convenience, we set L(a) = {0} for all a ∈ H × . By defini-
tion, H is atomic if and only if L(a) = ∅ for all a ∈ H . Furthermore, it is clear that the
following conditions are equivalent: (1) L(a) = {1}; (2) a ∈ A(H ); and (3) 1 ∈ L(a).
If a, b ∈ H , then L(a) + L(b) ⊂ L(ab). If H is commutative, then Hred = H/H × =
{a H × | a ∈ H } is the associated reduced monoid, and H is called factorial if Hred is
free abelian. We say that H is a BF-monoid (or a bounded factorization monoid) if
L(a) is finite and nonempty for all a ∈ H . We call

L(H ) = {L(a) | a ∈ H }

the system of sets of lengths of H . So if H is a BF-monoid, then L(H ) is a set of

finite nonempty subsets of the nonnegative integers. We say that H is half-factorial if
|L| = 1 for every L ∈ L(H ). Clearly, factorial monoids are half-factorial, and in 1960
Carlitz showed that the ring of integers of an algebraic number field is half-factorial
if and only if the class group has at most two elements (see Propositions 10 and 12).
Since then, half-factoriality has been a central topic in factorization theory (see, for
example, [13, 25, 41]). We focus in this paper on the structure of sets of lengths in
non-half-factorial BF-monoids. We start with two simple lemmas.

Lemma 1. Let H be a monoid.

1. If H satisfies the ACC on principal left ideals and the ACC on principal right
ideals, then H is atomic.
2. Suppose that H is atomic. Then H is either half-factorial or for every N ∈ N
there is an element a N ∈ H such that |L(a N )| > N .

962 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
Proof. For ease of discussion, suppose that a H = H a for all a ∈ H (see [49, Propo-
sition 3.1] for details in the general case). Assume to the contrary that the set  of
all nonunits a ∈ H that are not products of atoms is nonempty. If a ∈ , then a = bc
with nonunits b, c ∈ H , and either b ∈  or c ∈ . Thus for any a ∈ H , there is some
a  ∈  with a H  a  H . Starting from an arbitrary a ∈ H , this gives rise to a properly
ascending chain of principal ideals, a contradiction.
To verify the second statement, suppose that H is atomic but not half-factorial.
Then there exist an element a ∈ H , integers k,  ∈ N with k < , and atoms u 1 , . . . , u k ,
v1 . . . , v ∈ A(H ) such that a = u 1 · . . . · u k = v1 · . . . · v . Then for every N ∈ N we
have

a N = (u 1 · . . . · u k )ν (v1 · . . . · v ) N −ν for all ν ∈ [0, N ],

and hence {N − ν( − k) | ν ∈ [0, N ]} ⊂ L(a N ).

Let H be a monoid. A function λ : H → N0 is called a right length function (resp.

a left length function) if λ(a) < λ(b) for all a ∈ H and all b ∈ a H \ a H × (resp. all
b ∈ H a \ H × a).

Lemma 2. Let H be a monoid.

1. H is a BF-monoid if and only if there is a left length function (or a right length
function) λ : H → N0 .
2. If H is a BF-monoid, then H satisfies the ACC on principal left ideals and on
principal right ideals.
3. Submonoids of free monoids and of free abelian monoids are BF-monoids.

Proof. 1. If H is a BF-monoid, then the function λ : H → N0 , defined by a →

max L(a) for every a ∈ H , is a right length function and a left length function. Now,
if there is a right length function λ : H → N0 , then we have to show that H is a
BF-monoid (the proof is completely analogous for left length functions). First we
observe that, if b ∈ H × and c ∈ H \ H × = bH \ bH × , then λ(c) > λ(b) ≥ 0. We
claim that every a ∈ H \ H × can be written as a product of finitely many atoms, and
that sup L(a) ≤ λ(a). If a ∈ A(H ), then L(a) = {1}, and the assertion holds. Suppose
that a ∈ H is neither an atom nor a unit. Then a has a product decomposition of the
form

a = u1 · . . . · uk where k ≥ 2 and u 1 , . . . , u k ∈ H \ H × .

For i ∈ [0, k], we set ai = u 1 · . . . · u i , with a0 = 1, and hence ai+1 ∈ ai H \ ai H × for

all i ∈ [0, k − 1]. This implies that λ(a) = λ(ak ) > λ(ak−1 ) > · · · > λ(a1 ) > 0 and
thus λ(a) ≥ k. Therefore, there is a maximal k ∈ N such that a = u 1 · . . . · u k where
u 1 , . . . , u k ∈ H \ H × , and this implies that u 1 , . . . , u k ∈ A(H ) and k = max L(a) ≤
λ(a).
2. Suppose that H is a BF-monoid. Let λ : H → N0 be a right length function and
assume to the contrary that there is a properly ascending chain of principal right ide-
als a0 H  a1 H  a2 H  · · · . Then λ(a0 ) > λ(a1 ) > λ(a2 ) > · · · , a contradiction.
Similarly, we can show that H satisfies the ACC on principal left ideals.
3. Use 1., and note that the restriction of a length function is a length function.

December 2016] SETS OF LENGTHS 963

 Next we introduce a main parameter describing the structure of sets of lengths,
namely

(H ) = (L),
L∈L(H )

which is the set of distances of H (also called the delta set of H ). We open by showing
that (H ) satisfies a fundamental property.

Proposition 3. Let H be an atomic monoid with (H ) = ∅. Then min (H )

= gcd (H ).

Proof. We set d = gcd (H ). Clearly it suffices to show that d ∈ (H ). There are
t ∈ N, d1 , . . . , dt ∈ (H ) and m 1 , . . . , m t ∈ Z \ {0} such that d = m 1 d1 + · · · +
m t dt . After renumbering if necessary, there is some s ∈ [1, t] such that m 1 , . . . , m s ,
−m s+1 , . . . , −m t are positive. For every ν ∈ [1, t], there are xν ∈ N and aν ∈ H such
that
{xν , xν + dν } ⊂ L(aν ) for every ν ∈ [1, s] and
{xν − dν , xν } ⊂ L(aν ) for every ν ∈ [s + 1, t].

This implies that

{m ν xν , m ν xν + m ν dν } ⊂ L(aνm ν ) for every ν ∈ [1, s] and

{−m ν xν + m ν dν , −m ν xν } ⊂ L(aν−m ν ) for every ν ∈ [s + 1, t].
−m
We set a = a1 1 · . . . · asm s as+1s+1 · . . . · at−m t , and observe that L(a) ⊃
m

 

s 
t 
s 
t
k := m ν xν − m ν xν ,  := m ν (xν + dν ) − m ν (xν − dν ) .
ν=1 ν=s+1 ν=1 ν=s+1

Since  − k = d ≤ min (H ), it follows that d ∈ (L(a)) ⊂ (H ).

We now introduce unions of sets of lengths. Let H be an atomic monoid, and k, 

∈ N. In the extremal case where H = H × it is convenient to set Uk (H ) = {k}. Now
suppose that H = H × . We define Uk (H ) to be the set of all  ∈ N such that we have
an equation of the form

u 1 · . . . · u k = v1 · . . . · v where u 1 , . . . , u k , v1 , . . . , v ∈ A(H ).

In other words, Uk (H ) is the union of sets of lengths containing k. Clearly we have

k ∈ Uk (H ), and 1 ∈ Uk (H ) if and only if k = 1, and U1 (H ) = {1}. Furthermore, we
have k ∈ U (H ) if and only if  ∈ Uk (H ). Now we define

ρk (H ) = sup Uk (H ) and λk (H ) = min Uk (H ),

and we call ρk (H ) the kth elasticity of H . Since Uk (H ) + U (H ) ⊂ Uk+ (H ), we infer

that

λk+ (H ) ≤ λk (H ) + λ (H ) ≤ k +  ≤ ρk (H ) + ρ (H ) ≤ ρk+ (H ), (1)

964 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
and we will use these inequalities without further mention. The invariant

ρ(H ) = sup{ρ(L) | L ∈ L(H )} ∈ R≥1 ∪ {∞}

is called the elasticity of H . If k ∈ N and L ∈ L(H ) with min L ≤ k, then sup L

≤ ρ(H ) min L ≤ kρ(H ) and hence ρk (H ) ≤ kρ(H ). Thus if the elasticity ρ(H ) is
finite, then H is a BF-monoid and ρk (H ) < ∞ for all k ∈ N.
The kth elasticities ρk (H ) were first studied for rings of integers of algebraic num-
ber fields and then in the setting of abstract semigroups [26]. Unions of sets of lengths
have been introduced by Chapman and Smith [12] in the setting of commutative
Dedekind domains, and the structure theorem for unions of sets of lengths (as given in
Theorem 5) has first been proved in [19] (in a commutative setting).

Proposition 4. Let H be an atomic monoid with H = H × .

1. For every k ∈ N we have ρk (H ) = sup{sup L | L ∈ L(H ), min L ≤ k}
≥ sup{sup L | L ∈ L(H ), k = min L}, and equality holds if ρk (H ) < ∞.
2.


ρk (H ) ρk (H )
ρ(H ) = sup k ∈ N = lim and
k k→∞ k


1 λk (H ) λk (H )
= inf k ∈ N = lim .
ρ(H ) k k→∞ k

3. Suppose that ρ(H ) < ∞. Then the following statements are equivalent.
(a) There is an L ∈ L(H ) such that ρ(L) = ρ(H ).
(b) There is an N ∈ N such that k Nρ(H ) = ρk N (H ) for all k ∈ N.
(c) There is some k ∈ N such that kρ(H ) = ρk (H ).
If one of the above statements holds, then there is some M ∈ N such that
ρk (H ) − ρk−1 (H ) ≤ M for all k ≥ 2.

Proof. 1. Let k ∈ N. We define ρk (H ) = sup{sup L | L ∈ L(H ), min L ≤ k},

ρk (H ) = sup{sup L | L ∈ L(H ), k = min L}, and obtain by definition ρk (H )
≥ ρk (H ) ≥ ρk (H ). Hence we must prove that ρk (H ) ≥ ρk (H ), and if ρk (H ) < ∞,
then also ρk (H ) ≥ ρk (H ). Let a ∈ H with min L(a) =  ≤ k and u ∈ A(H ). Then
k ∈ L(au k− ) implies ρk (H ) ≥ sup L(au k− ) ≥ sup L(a) + k −  ≥ sup L(a), and
therefore ρk (H ) ≥ ρk (H ). Assume now that ρk (H ) < ∞, and let a be as above such
that sup L(a) = ρk (H ). Then min L(au k− ) ≤ min L(a) + (k − ) = k and therefore

ρk (H ) ≥ sup L(au k− ) ≥ sup L(a) + (k − ) ≥ ρk (H ).

Thus k =  = min L(a) and ρk (H ) ≥ sup L(a) = ρk (H ).

2. If there is a k ∈ N such that ρk (H ) = ∞, then all three terms of the first equation
are equal to infinity. Suppose that ρk (H ) < ∞ for all k ∈ N. Then the first equality
follows from 1. To verify the first limit assertion, let ρ  < ρ(H ) be arbitrary. We must
prove that ρk (H ) ≥ kρ  for all sufficiently large k. If ρ  < ρ  < ρ(H ), then there
exists some q0 ∈ N such that

qρ  + 1
> ρ for all q ≥ q0 ,
q +1

December 2016] SETS OF LENGTHS 965

and there exists some N ∈ N such that ρ N (H ) > Nρ  . If k > N q0 , then k = N q + r
for some q ≥ q0 and r ∈ [1, N ], and ρk (H ) ≥ qρ N (H ) + ρr (H ) ≥ qρ N (H ) + r by
Inequality (1). Since ρ N (H ) ≥ N , it follows that

ρk (H ) qρ N (H ) + r qρ N (H ) + N q Nρ  + N
≥ ≥ > > ρ.
k qN + r qN + N qN + N
 

Since ρ(H ) = sup mn m, n ∈ L , {0} = L ∈ L(H ) , we have

m 
1
= inf m, n ∈ L , {0} = L ∈ L(H ) ,
ρ(H ) n

with 1/ρ(H ) = 0 if ρ(H ) = ∞. The verification of the second limit assertion runs
along the same lines as the proof of the first one (just replace ρk (H ) by λk (H ) and
reverse all inequality signs).
3. In order to show the implication (a) ⇒ (b), let a ∈ H with L = L(a) ∈ L(H )
such that ρ(L) = ρ(H ) and set N = min L. Then k N ∈ L(a k ) for all k ∈ N, and thus

ρk N (H ) sup L(a k ) k sup L(a) ρk N (H )

≥ ≥ = ρ(L) = ρ(H ) ≥ .
kN kN kN kN
The implication (b) ⇒ (c) is obvious. Suppose that (c) holds and let k ∈ N such
that kρ(H ) = ρk (H ). Let L ∈ L(H ) such that min L = k and ρk (H ) = max L. Then
kρ(H ) = max L = kρ(L) and hence ρ(L) = ρ(H ).
Suppose that the equivalent statements hold and let N ∈ N such that m Nρ(H )
= ρm N (H ) for all m ∈ N. Let k ∈ N be given and set k = i N + j with i ∈ N0 and
j ∈ [1, N ]. Then we infer that i Nρ(H ) + j ≤ ρi N (H ) + ρ j (H ) ≤ ρi N + j (H ) and

(i N + j)ρ(H ) − N (ρ(H ) − 1) ≤ (i N + j)ρ(H ) − j (ρ(H ) − 1) ≤ ρi N + j (H ).

Thus kρ(H ) − ρk (H ) ≤ M  := N (ρ(H ) − 1) for all k ∈ N. If k ≥ 2, then ρk (H )

≤ kρ(H ), (k − 1)ρ(H ) − ρk−1 (H ) ≤ M  , and hence ρk (H ) − ρk−1 (H ) ≤ M 
+ ρ(H ).

Unions of sets of lengths do have—in a variety of settings—a structure which is

only slightly more general than that of arithmetical progressions. In order to show this,
we introduce the concept of almost arithmetical progressions.

Definition. Let d ∈ N and M ∈ N0 . A subset L ⊂ Z is called an almost arithmetical

progression (AAP for short) with difference d and bound M if

L = y + (L  ∪ L ∗ ∪ L  ) ⊂ y + dZ,

where y ∈ Z and L ∗ is a nonempty (finite or infinite) arithmetical progression with

difference d such that min L ∗ = 0, L  ⊂ [−M, −1], and L  ⊂ sup L ∗ + [1, M] (with
the convention that L  = ∅ if L ∗ is infinite).

Clearly every single finite set is an AAP with a particularly trivial choice of M
and d. Suppose we have an atomic monoid with nonempty set of distances. Then,
by Lemma 1, sets of lengths become arbitrarily large, whence the unions Uk (H ) are

966 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
growing as k is growing. The next theorem states (under the given assumptions) that
all unions Uk (H ) are AAPs with the critical point that a single choice of M works for
all sufficiently large k. Once this M is chosen, it says that all unions have a “middle”
piece that becomes larger and larger as k grows and that middle part has a very rigid
structure. The poorly behaved “end” pieces are bounded in size by a constant and in
their distance from the middle.

Theorem 5 (Structure Theorem for Unions of Sets of Lengths). Let H be an

atomic monoid with finite nonempty set of distances (H ) and d = min (H ). Sup-
pose that either ρk (H ) = ∞ for some k ∈ N, or that there is an M ∈ N such that
ρk (H ) − ρk−1 (H ) ≤ M for all k ≥ 2.
1. There exist constants k ∗ and M ∗ ∈ N such that for all k ≥ k ∗ , Uk (H ) is an AAP
with difference d and bound M ∗ . Moreover, if ρk (H ) < ∞ for all k ∈ N, then
the assertion holds for k ∗ = 1.
2. We have

|Uk (H )| 1 1 
lim = ρ(H ) − .
k→∞ k d ρ(H )

Proof. 1. For every k ∈ N, we have λk (H ), k ∈ Uk (H ), and since every d  ∈ (H )

is a multiple of d by Proposition 3, it follows that Uk (H ) ⊂ λk (H ) + dN0 . Thus it
remains to show that there exist constants k ∗ , M ∗ ∈ N such that, for all k ≥ k ∗ ,

Uk (H ) ∩ [k, ρk (H ) − M ∗ ] and Uk (H ) ∩ [λk (H ) + M ∗ , k] (∗)

are arithmetical progressions with difference d (note, if ρk (H ) = ∞, then [k, ρk (H )

− M ∗ ] = N≥k ). We break the proof into three steps.
1.(a) First, we show that the left set in (∗) is an arithmetical progression. Since
d ∈ (H ), there is an element a ∈ H and m ∈ N such that {m, m + d} ⊂  L(a).Since
(H ) is finite and min (H ) = gcd (H ) by Proposition 3, ψ := ρ (H ) ∈ N.
Then the ψ-fold sumset satisfies the containment

U ∗ := ψ{m, m + d} = {ψm, ψm + d, . . . , ψm + ψd} ⊂ L(a ψ ).

We set k0 = ψm and observe that U ∗ ⊂ Uk0 (H ), say Uk0 (H ) = U   U ∗  U  , where

U  = {x ∈ Uk0 (H ) | x < k0 } and U  = {x ∈ Uk0 (H ) | x > k0 + ψd}. Let k ∗ ∈ N with
k ∗ ≥ 2k0 . If there is some  ∈ N with ρ (H ) = ∞, then let 0 denote the smallest such
 ∈ N, and we suppose further that k ∗ − k0 ≥ 0 . Now let k ≥ k ∗ be given. Then
Uk0 (H ) + Uk−k0 (H ) =
  
U  + Uk−k0 (H ) ∪ U ∗ + Uk−k0 (H ) ∪ U  + Uk−k0 (H ) ⊂ Uk (H ).

 
Clearly
 we have k ∈ U ∗ + Uk−k0 (H ). Since max  Uk−k0 (H ) ≤ max (H ) and
 Uk−k0 (H ) ⊂ dN, it follows that U ∗ + Uk−k0 (H ) is an arithmetical progression
with difference d. If there is some  ∈ N such that ρ (H ) = ∞, then ρk−k0 (H )
= ρk (H ) = ∞ and
 ∗ 
U + Uk−k0 (H ) ∩ N≥k = k + dN0 = Uk (H ) ∩ N≥k .

December 2016] SETS OF LENGTHS 967

Suppose that ρ (H ) < ∞ for all  ∈ N. Then

max Uk (H ) − max U ∗ + Uk−k0 (H ) = ρk (H ) − max U ∗ − ρk−k0 (H ) ≤ k0 M,

and hence
 ∗ 
U + Uk−k0 (H ) ∩ [k, ρk (H ) − k0 M] = Uk (H ) ∩ [k, ρk (H ) − k0 M]

is an arith. progression with difference d. Thus the assertion follows with M ∗ = k0 M.

1.(b) By 1.(a) there are k ∗ , M ∗ ∈ N such that for all k ≥ k ∗ , the set Uk (H ) ∩
[k, ρk (H ) − M ∗ ] is an arithmetical progression with difference d. Without restriction
we may suppose that M ∗ ≥ k ∗ .
Let k ≥ k ∗ and  = λk (H ). We show that Uk (H ) ∩ [ + M ∗ , k] is an arithmetical
progression with difference d. Let m ∈ [ + M ∗ , k] be such that k − m is a multiple
of d. In order to show that m ∈ Uk (H ), we verify that k ∈ Um (H ). Since

k ≤ ρ (H ) and  + M ∗ ≤ m,

it follows that k ≤ ρ (H ) ≤ ρm−M ∗ (H ) and hence

k + M ∗ ≤ ρm−M ∗ (H ) + M ∗ ≤ ρm−M ∗ (H ) + ρ M ∗ (H ) ≤ ρm (H ).

Since k ∈ m + dN0 with k ≤ ρm (H ) − M ∗ and Um (H ) ∩ [m, ρm (H ) − M ∗ ] is an

arithmetical progression with difference d, it follows that k ∈ Um (H ).
1.(c) Suppose that ρk (H ) < ∞ for all k ∈ N. In this step we show that for all k ∈ N
the sets Uk (H ) are AAPs with difference d and some bound M.  Suppose that the
assertion holds with the constants k ∗ and M ∗ ∈ N. Since for all k ∈ [1, k ∗ − 1],
   
Uk (H ) = [λk (H ), k − 1] ∩ Uk (H ) ∪ {k} ∪ [k + 1, ρk (H )] ∩ Uk (H )

is an AAP with bound M  = max{k − λk (H ), ρk (H ) − k | k ∈ [1, k ∗ − 1]}, it fol-

lows that for all k ∈ N the sets Uk (H ) are AAPs with difference d and bound
M = max{M ∗ , M  }.
2. If there is some k ∈ N such that ρk (H ) = ∞, then both the left- and the right-
hand side of the asserted equation are infinite. Suppose that ρk (H ) < ∞ for all k ∈
N. By 1. there are k ∗ ∈ N and M ∗ ∈ dN such that, for all k ≥ k ∗ , Uk (H ) ∩ [λk (H )
+ M ∗ , ρk (H ) − M ∗ ] is an arithmetical progression with difference d. Thus for all
k ≥ k ∗ we obtain that
   
ρk (H ) − M ∗ − λk (H ) + M ∗ + d |Uk (H )| ρk (H ) − λk (H ) + d
≤ ≤ .
dk k dk
Since, by Proposition 4.2,

ρk (H ) λk (H ) 1
lim = ρ(H ) and lim = ,
k→∞ k k→∞ k ρ(H )

the assertion follows.

We end this section with a discussion of finitely presented monoids and of commu-
tative finitely generated monoids. Let H be a monoid. For every generating set P of
H , there is an epimorphism ψ : F ∗ (P) → H and F ∗ (P)/ ker(ψ) ∼ = H , where

968 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 ker(ψ) = {(x, y) ∈ F ∗ (P) × F ∗ (P) | ψ(x) = ψ(y)}

is a congruence relation on F ∗ (P). If there is a finite generating set P and a finite

set of relations R ⊂ ker(ψ) which generates ker(ψ) as a congruence relation, then H
is said to be finitely presented (by P and R). If R = {(x1 , y1 ), . . . , (xt , yt )}, then we
write (as usual) H = P | R = P | x1 = y1 , . . . , xt = yt .

Proposition 6. If H = A(H ) | R is a reduced atomic monoid with a finite set of

relations R, then the set of distances (H ) is finite.

∗
Proof. We set A = A(H ), ψ : F (A) → H , R = {(x1 , y1 ), . . . , (xt , yt )},

M = max{ |x1 | − |y1 | , . . . , |xt | − |yt | }, and assert that (H ) ⊂ [1, M]. Let a ∈ H .
Then L(a) = {|x| | x ∈ F ∗ (A) with ψ(x) = a}. We choose two words v, w ∈ F ∗ (A)
with ψ(v) = ψ(w) = a. Since F ∗ (P)/ ker(ψ) ∼ = H and ker(ψ) is generated by R
(as a congruence), there is a sequence of words v = v0 , . . . , vn = w in ψ −1 (a) ⊂
F ∗ (A) where vν arises from vν−1 by replacing xi by yi for some i ∈ [1, t] and all
ν ∈ [1, n].
 We set L = {|v0 |, . . . , |vn |} and obtain (L) ⊂ [1, M]. Thus it follows that
 L(a) ⊂ [1, M] and hence (H ) ⊂ [1, M].

There are atomic finitely presented monoids H such that ρ(H ) = ρk (H ) = ∞ for
all k ≥ 2. To provide an example, consider the monoid H = a, b | a 2 = ba 2 b (note
that H is an Adyan semigroup and hence cancelative, see [4, Section 2] for details).
Obviously, H is finitely presented and atomic with A(H ) = {a, b}, and ρ2 (H ) = ∞
which implies that ρk (H ) = ∞ for all k ≥ 2. Since H a 2 bi  H a 2 bi+1 for all i ∈ N,
H does not satisfy the ACC on principal left ideals. As another example, the monoid
H = a, b | a = bab is a finitely presented monoid which is not atomic (note that
a is not a finite product of atoms). However, this behavior cannot occur in the case
of commutative monoids. The next result shows in particular that finitely generated
commutative monoids satisfy all assumptions of Theorem 5, and hence they satisfy
the structure theorem for unions of sets of lengths.

Proposition 7. Let H be a reduced commutative monoid.

1. For a subset A ⊂ A(H ) the following statements are equivalent.
(a) H is atomic and A = A(H ).
(b) A is the smallest generating set of H (with respect to set inclusion).
(c) A is a minimal generating set of H .
2. H is finitely generated if and only if H is atomic and A(H ) is finite.
3. Suppose that H is finitely generated. Then H is a BF-monoid with finite set of
distances and finite elasticity. Moreover, there is an L ∈ L(H ) such that ρ(L)
= ρ(H ), and there is an M ∈ N such that ρk (H ) − ρk−1 (H ) ≤ M for all k ≥ 2.

Proof. 1.(a) ⇒ (b) Since H is atomic, A is a generating set of H and it remains to

show that it is the smallest one. Let A be any generating set of H . If u ∈ A, then
u = v1 · . . . · vk with k ∈ N and v1 , . . . , vk ∈ A \ {1}. Since u is an atom, it follows
that k = 1 and u = v1 ∈ A . The implication (b) ⇒ (c) is obvious.
(c) ⇒ (a) It suffices to verify that A = A(H ). Since A is a minimal generating set, it
follows that 1 ∈ / A. If u ∈ A(H ), then u = v1 · . . . · vk with k ∈ N and v1 , . . . , vk ∈ A.
This implies k = 1, u = v1 ∈ A, and thus A(H ) ⊂ A. Assume to the contrary that
there is some u ∈ A \ A(H ), say u = vw with v, w ∈ H \ {1}. Then v = u 1 · . . . · u m

December 2016] SETS OF LENGTHS 969

and w = u m+1 · . . . · u n where n ∈ N≥2 , m ∈ [1, n − 1], and u 1 , . . . , u n ∈ A. Therefore
we obtain that u = u 1 · . . . · u n and u ∈
/ {u 1 , . . . , u n }. Thus A \ {u} is a generating set
of H , a contradiction.
2. This assertion follows directly from 1.
3. By 1. and 2., H is atomic and A = A(H ) is the smallest generating set. By
Redei’s theorem, every finitely generated commutative monoid is finitely presented.
Thus (H ) is finite by Proposition 6. Next we show that there is an L ∈ L(H ) such
that ρ(H ) = ρ(L) < ∞. This implies that H is a BF-monoid and by Proposition 4.3
it follows that there is an M ∈ N such that ρk (H ) − ρk−1 (H ) ≤ M for all k ≥ 2.
Let π : F (A) → H be the canonical epimorphism. We set

S = {(x, y) ∈ F (A) × F (A) | π(x) = π(y)} and S ∗ = S \ {(1, 1)},

 

and we observe that ρ(H ) = sup |x| |y|
(x, y) ∈ S ∗ . Clearly it is sufficient to
show that this supremum is attained for some pair (x, y) ∈ S ∗ . There is an isomor- 
 
phism f : F (A) × F (A) → (N0A × N0A , +), defined by u∈A u mu
, u∈A u nu
→
  ∗
(m u )u∈A , (n u )u∈A . By Dickson’s theorem [23, Theorem 1.5.3], the set f (S ) has
only finitely many minimal points, and let T ⊂ S ∗ denote the inverse image of the set
of minimal points. Therefore it suffices to prove that
 

|x| |x |  
≤ max (x , y ) ∈ T for all (x, y) ∈ S ∗ .
|y| |y  |

We proceed by induction on |x| + |y|. If (x, y) ∈ T , then there is nothing to do. Sup-
pose that (x, y) ∈ / T . Then there exist (x1 , y1 ) ∈ T such that (x, y) = (x1 x2 , y1 y2 )
with (x2 , y2 ) ∈ F (A) × F (A). It follows that (x2 , y2 ) ∈ S ∗ , and clearly we have |x j |
+ |y j | < |x| + |y| for j ∈ {1, 2}. Then

 

|x| |x1 | + |x2 | |x1 | |x2 | |x |  

= < max , ≤ max (x , y ) ∈ T
|y| |y1 | + |y2 | |y1 | |y2 | |y  |

by the induction hypothesis.

A most interesting class of finitely generated commutative monoids are numerical

monoids. Their study was initiated by Frobenius in the 19th century and they are still
a topic of much research due to their intrinsic relationship with a wide area of mathe-
matics. A monoid H is said to be numerical if it is a submonoid of (N0 , +) such that
the complement N0 \ H is finite. Clearly numerical monoids are reduced. Let H be
a numerical monoid with H = N0 . Since N0 \ H is finite, H has a finite generating
set and hence a smallest generating set. Thus Proposition 7 implies that the small-
est generating set is the set of atoms, and that the elasticity and the set of distances
are both finite. Suppose that A(H ) = {n 1 , . . . , , n t } with t ∈ N and n 1 < · · · < n t .
We encourage the reader to check that ρ(H ) = n t /n 1 and that min (H ) = gcd{n 2 −
n 1 , . . . , n t − n t−1 } (compare with Proposition 18.2). These results were the starting
points of detailed investigations of the arithmetic of numerical monoids initiated by
Chapman and Garcı́a-Sánchez.
Clearly there are natural connections between the arithmetical invariants of factor-
ization theory and the presentations of a monoid. This point has been emphasized
by Garcı́a-Sánchez and it opened the way to an algorithmic approach towards the
computational determination of arithmetical invariants. Many algorithms have been

970 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
implemented in GAP (see the GAP Package [14] and a survey by Garcı́a-Sánchez
[20]).

3. COMMUTATIVE KRULL MONOIDS. It was the observation of the mathe-

maticians of the 19th century that a ring of integers in an algebraic number field need
not be factorial (in other words, it need not satisfy the fundamental theorem of arith-
metic). This led to the development of ideals (every nonzero ideal in a ring of integers
is a unique product of prime ideals whence the fundamental theorem of arithmetic
holds for ideals) and subsequently to the development of “divisor theories” from the
elements to the ideals. A divisor theory is a divisibility preserving homomorphism to
an object which fulfills the fundamental theorem of arithmetic. Semigroups allowing
a divisor theory are now called Krull monoids.
Commutative Krull monoids can be studied with divisor theoretic and with ideal
theoretic tools. We start with divisor theoretic concepts. Let H and D be commutative
monoids. A monoid homomorphism ϕ : H → D is said to be:
• a divisor homomorphism if a, b ∈ H and ϕ(a) | ϕ(b) (in D) implies that a | b (in
H );
• cofinal if for every α ∈ D there is an a ∈ H such that α | ϕ(a) (in D);
• a divisor theory if D is free abelian, ϕ is a divisor homomorphism, and  for every
α ∈ D there are a1 , . . . , am ∈ H such that α = gcd ϕ(a1 ), . . . , ϕ(am ) .
In particular, every divisor theory is a cofinal divisor homomorphism. Let ϕ : H → D
be a cofinal divisor homomorphism. The group
 
C (ϕ) = q(D)/q ϕ(H )
 
is called the class group of ϕ. For a ∈ q(D) we denote by [a] = aq ϕ(H ) ∈ C (ϕ)
the class containing a. We use additive notation for the class group and observe that
[1] is the zero element of the abelian group C (ϕ). Divisor theories of a given monoid
are unique up to isomorphism. If H has a divisor theory, then there is a free abelian
monoid F = F (P) such that the inclusion ϕ : Hred → F is a divisor theory, and the
class group

C (ϕ) = C (H ) = q(F)/q(Hred )

is called the (divisor) class group of H and G 0 = {[ p] | p ∈ P} ⊂ C (H ) is the set of

classes containing prime divisors. We continue with the most classical example of a
cofinal divisor homomorphism and a divisor theory.

Proposition 8. Let R be a commutative domain, I ∗ (R) the monoid of invertible ideals

where the operation is the usual multiplication of ideals, and let ϕ : R • → I ∗ (R) be
the homomorphism mapping each element onto its principal ideal.
1. The map ϕ is a cofinal divisor homomorphism and C (ϕ) is the Picard group
Pic(R) of R.
2. If R is a commutative Dedekind domain, then ϕ is a divisor theory and C (ϕ) is
the usual ideal class group of R.
3. If R is the ring of integers of an algebraic number field, then C (ϕ) is finite and
every class contains infinitely many prime ideals.

Proof. 1. A short calculation shows that for two invertible ideals I, J  R we have
J | I in I ∗ (R) if and only if I ⊂ J . To show that ϕ is a divisor homomorphism, let

December 2016] SETS OF LENGTHS 971

a, b ∈ R • be given and suppose that b R | a R in I ∗ (R). Then there is a J ∈ I ∗ (R)
such that (a R)J = b R whence a −1 b R = J ⊂ R and a | b in R • . To show that ϕ is
cofinal, let I ∈ I ∗ (R) be given. If a ∈ I , then a R ⊂ I and hence I | a R in I ∗ (R). The
definition of C (ϕ) coincides with the definition of Pic(R).
2. Suppose that R is a commutative Dedekind domain. Then every nonzero ideal is
invertible and a product of prime ideals. Thus I ∗ (R) is free abelian. Let I ∈ I ∗ (R).
Then I is generated by two elements a, b ∈ R, whence I = a, b = a R + b R
= gcd(a R, b R) = gcd(ϕ(a), ϕ(b)). Therefore ϕ is a divisor theory.
3. This can be found in many textbooks on algebraic number theory (see, for
example, [23, Theorem 2.10.14] for a summary).

The previous proposition shows that in case of commutative Dedekind domains the
embedding in a monoid of ideals establishes a divisor theory. This holds true in much
greater generality and in order to outline this we mention briefly some key notions on
divisorial ideals (see [36] for a thorough treatment of divisorial ideals).
Let H be a commutative monoid and let A, B ⊂ q(H ) be subsets. We denote by
(A : B) = {x ∈ q(H ) | x B ⊂ A}, by A−1 = (H : A), and by Av = (A−1 )−1 . By an
ideal of H we always mean an s-ideal (thus AH = A holds), and an s-ideal A is a
divisorial ideal (or a v-ideal) if Av = A. We denote by Fv (H ) the set of all fractional
divisorial ideals and by Iv (H ) the set of all divisorial ideals of H . Furthermore, Iv∗ (H )
is the monoid of v-invertible
  divisorial ideals (with v-multiplication) and its quotient
group Fv (H )× = q Iv∗ (H ) is the group of fractional invertible divisorial ideals. By
X(H ), we denote the set of all minimal nonempty prime s-ideals of H and
 = {x ∈ q(H ) | there is a c ∈ H such that cx n ∈ H for all n ∈ N} ⊂ q(H )
H

is called the complete integral closure of H . We say that H is completely integrally

closed if H = H . Straightforward arguments show that every factorial monoid is com-
pletely integrally closed and that a noetherian commutative domain is completely inte-
grally closed if and only if it is integrally closed.

Theorem 9 (Commutative Krull monoids).

Let H be a commutative monoid. Then the following statements are equivalent.
(a) H is completely integrally closed and satisfies the ACC on divisorial ideals.
(b) The map ϕ : H → Iv∗ (H ), a → a H for all a ∈ H , is a divisor theory.
(c) H has a divisor theory.
(d) There is a free abelian monoid F such that the inclusion Hred → F is a divisor
homomorphism.
If one of the equivalent statements holds, then H is called a Krull monoid, and Iv∗ (H )
is free abelian with basis X(H ).

For a proof of Theorem 9 we refer to [23, Section 2.5]. Note, since H is factorial if
and only if Hred is free abelian, it follows that H is factorial if and only if it is Krull
with trivial class group. In the remainder of this section we present a list of examples
of commutative Krull monoids stemming from quite diverse mathematical areas.

Commutative domains. Let R be a commutative domain and H = R • . Then the maps

 
Fv (R) → Fv (H ) Fv (H ) → Fv (R)
ι• : and ι◦ : (2)
a → a \ {0} a → a ∪ {0}

972 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
are inclusion preserving isomorphisms which are inverse to each other. In particular,
if a is a divisorial semigroup theoretical ideal of H , then a ∪ {0} is a divisorial ring
theoretical ideal of R. Thus R satisfies the ACC on (ring theoretical) divisorial ideal
of R if and only if H satisfies the ACC on (semigroup theoretical) divisorial ideals of
H . Since, by definition, R is completely integrally closed if and only H is completely
integrally closed, we obtain that R is a commutative Krull domain if and only if H is
a commutative Krull monoid.
Property (a) in Theorem 9 easily implies that noetherian integrally closed commu-
tative domains are Krull. Furthermore, a commutative Krull domain is Dedekind if and
only if it is at most one dimensional. If R is Dedekind, then every ideal is divisorial
and Iv∗ (R) = I ∗ (R) (confer Theorem 9.(b) and Proposition 8.2).
Submonoids of commutative domains. Let R be a commutative Krull domain,
{0} = f  R an ideal, and  ⊂ (R/f)× a subgroup. Then the monoid

H = {a ∈ R • | a + f ∈ }

is a Krull monoid, called the (regular) congruence monoid defined in R modulo f by

. We refer the reader to [23, Section 2.11] for more on congruence monoids.
Monadic submonoids of rings of integer-valued polynomials. Let us consider the
classical ring of integer-valued polynomials over the integers. This is the ring

Int(Z) = { f ∈ Q[X ] | f (Z) ⊂ Z} ⊂ Q[X ].

We refer the reader to the Monthly article by Cahen and Chabert [9] for a friendly intro-
duction to integer-valued polynomials and to their monograph [8] for a deeper study.
It is well known that Int(Z) is an integrally closed two-dimensional Prüfer domain. It
is a BF-domain but it is not Krull. However, every divisor-closed submonoid of Int(Z),
which is generated by one element, is a Krull monoid [44, Theorem 5.2]. We refer to
recent work of Frisch and Reinhart [18, 43].
Monoids of regular elements in commutative rings with zero divisors. By a com-
mutative Krull ring we mean a completely integrally closed commutative ring which
satisfies the ACC on regular divisorial ideals. The isomorphisms (as given in Equation
(2)) between monoids of divisorial ideals carry over from the setting of commutative
domains to the setting of commutative rings with zero divisors. Thus, if a commutative
ring R is Krull, then the monoid of cancelative (regular) elements is a Krull monoid,
and the converse holds for v-Marot rings [27, Theorem 3.5].
Monoids of modules. Let R be a ring and let C be a class of right R-modules which
is closed under finite direct sums, direct summands, and isomorphisms. For a module
M in C , let [M] denote the isomorphism class of M. Let V (C) denote the set of iso-
morphism classes of modules in C (we assume here that V (C) is indeed a set). Then
V (C ) is a commutative semigroup with operation defined by [M] + [N ] = [M ⊕ N ]
and all information about direct-sum decomposition of modules in C can be studied
in terms of factorization of elements in the semigroup V (C ). In particular, the direct-
sum decompositions in C are (essentially) unique (in other words, the Krull–Remak–
Schmidt–Azumaya theorem holds) if and only if V (C) is a free abelian monoid. This
semigroup-theoretical point of view was justified by Facchini [15] who showed that
V (C ) is a reduced Krull monoid provided that the endomorphism ring End R (M) is
semilocal for all modules M in C . This result allows one to describe the direct-sum
decomposition of modules in terms of factorization of elements in Krull monoids.
We refer the reader to the M ONTHLY article by Baeth and Wiegand [5].

December 2016] SETS OF LENGTHS 973

  of a commuta-
Finitely generated monoids and affine monoids. The root closure H
tive monoid H is defined as

 = {x ∈ q(H ) | x n ∈ H for some n ∈ N} ⊂ q(H ),

H

and H is said to be root closed (also the terms normal, full, and integrally closed are
used) if H = H . If H is finitely generated, then H = H . Since finitely generated
monoids satisfy the ACC on ideals, they are Krull if and only if they are root closed
(see Theorem 9.(a)).
A monoid is called affine if it is a finitely generated submonoid of a finitely gener-
ated free abelian group. It is easy to check that the concepts of normal affine monoids
and of reduced finitely generated commutative Krull monoids coincide (a variety of
further characterizations are given in [23, Theorem 2.7.14]). (Normal) affine monoids
play an important role in combinatorial commutative algebra.
Monoids of zero-sum sequences. Let G be an additively written abelian group and
G 0 ⊂ G a subset. By a sequence over G 0 , we mean a finite sequence of terms from G 0
where repetition is allowed and the order is disregarded. Clearly, the set of sequences
forms a semigroup, with concatenation as its operation and with the empty sequence
as its identity element. We consider sequences as elements of the free abelian monoid
with basis G 0 . This algebraic point of view has turned out to be quite convenient from
a notational point of view. But there is much more which we start to outline here and
later in Proposition 12. Let

S = g1 · . . . · g = g vg (S) ∈ F (G 0 ),
g∈G 0

where  ∈ N0 and g1 , . . . , g ∈ G. Then |S| =  is the length of S, supp(S)

= {g1 , . . . , g } is the support of S, −S = (−g1 ) · . . . · (−g ), and σ (S) = g1 + · · ·
+ g is the sum of S. We say that S is a zero-sum sequence if σ (S) = 0, and clearly
the set

B (G 0 ) = {S ∈ F (G 0 ) | σ (S) = 0} ⊂ F (G 0 )

of all zero-sum sequences is a submonoid, called the monoid of zero-sum sequences

(also called block monoid) over G 0 . Obviously the inclusion B (G 0 ) → F (G 0 ) is a
divisor homomorphism and hence B (G 0 ) is a reduced commutative Krull monoid by
Theorem 9.(d). Monoids of zero-sum sequences form a powerful link between the the-
ory of (general) Krull monoids and additive combinatorics [28, 34]. Thus all methods
from the later area are available for the study of sets of lengths in Krull monoids, and
we will make heavily use of this in Section 6.

Proposition 10. Let G be an additive abelian group and G 0 ⊂ G a subset.

1. If G 0 is finite, then B (G 0 ) is finitely generated.
2. The following statements are equivalent.
(a) |G| ≤ 2.
(b) B (G) is factorial.
(c) B (G) is half-factorial.
3. If |G| ≥ 3, then the inclusion B (G) → F (G) is a divisor theory with class
group isomorphic to G and every class contains precisely one prime divisor.

974 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 4. Let G  be an abelian group. Then the monoids B (G) and B (G  ) are isomorphic
if and only if the groups G and G  are isomorphic.

Proof. 1. The map f : N0 0 → F (G 0 ), defined by m = (m g )g∈G 0 → g∈G 0 g m g , is a
G

monoid isomorphism. The embedding  := f −1 (B (G 0 )) → N0 0 is a divisor homo-

G

morphism (i.e, m, n ∈  and m ≤ n implies that n − m ∈ ). By Dickson’s lemma

[23, Theorem 1.5.3],  is generated by the finite set of minimal points Min(), and
hence B (G 0 ) is generated by f (Min()).
2. If G = {0}, then B (G) = F (G) ∼ = (N0 , +) is free abelian. If G = {0, g}, then
B (G) is free abelian with basis A(G) = {0, g 2 }. Thus (a) ⇒ (b), and obviously (b)
⇒ (c). In order to verify that (c) ⇒ (a), it suffices to show that |G| ≥ 3 implies that
B (G) is not half-factorial. Suppose that |G| ≥ 3. If there is some element g ∈ G with
ord(g) = n ≥ 3, then U = g n , −U , and V = (−g)g are atoms of B (G) and (−U )U =
V n shows that B (G) is not half-factorial. If there are two distinct elements e1 , e2 ∈ G
of order two, then U = e1 e2 (e1 + e2 ), V0 = (e1 + e2 )2 , V1 = e12 , and V2 = e22 are atoms
of B (G) and U 2 = V0 V1 V2 shows that B (G) is not half-factorial.
3. Let |G| ≥ 3. Clearly the inclusion is a cofinal divisor homomorphism. To
show that it is a divisor
 theory, let g ∈ G \ {0} be given. If ord(g) = n ≥ 3, then
g = gcd g n , g(−g) . If ord(g) = 2, then there is an element h ∈ G \ {0, g}, and we
obtain that g = gcd g 2 , gh(−g − h) . It is easy to check that the map
       
 : C B (G) = q F (G) /q B (G) = {[S] = Sq B (G) | S ∈ F (G)} → G,

defined by ([S]) = σ (S) is a group isomorphism. Since for every S ∈ F (G), [S] ∩
G = {σ (S)}, every class of C B (G) contains precisely one prime divisor.
4. This follows from 2., 3., and the fact that a reduced commutative Krull monoid
is uniquely determined by its class group and the distribution of prime divisors in its
classes ([23, Theorem 2.5.4]).

4. TRANSFER HOMOMORPHISMS AND TRANSFER KRULL MONOIDS.

A central method to study the arithmetic of a given class of monoids H is to con-
struct simpler auxiliary monoids B (and such constructions are often based on the
ideal theory of H ) and homomorphisms θ : H → B (called transfer homomorphisms)
which allow us to pull back arithmetical results from B to H . The concept of transfer
homomorphisms was introduced by Halter–Koch in the commutative setting [35]) and
recently generalized to the noncommutative setting ([4, Definition 2.1]).

Definition. Let H and B be atomic monoids. A monoid homomorphism θ : H → B

is called a weak transfer homomorphism if it has the following two properties.
(T1) B = B × θ(H )B × and θ −1 (B × ) = H × .
(WT2) If a ∈ H , n ∈ N, v1 , . . . , vn ∈ A(B) and θ(a) = v1 · . . . · vn , then there
exist u 1 , . . . , u n ∈ A(H ) and a permutation τ ∈ Sn such that a = u 1 · . . . ·
u n and θ(u i ) ∈ B × vτ (i) B × for each i ∈ [1, n].

Property (T1) says that θ is surjective up to units and that only units are mapped
onto units. Property (WT2) says that factorizations can be lifted up to units and up to
order. We do not discuss equivalent formulations or variants of the definition and we
do not give the definition of a transfer homomorphism, but note that the two concepts
coincide if H and T are both commutative.

December 2016] SETS OF LENGTHS 975

Lemma 11. Let H and B be atomic monoids, and let θ : H → B be a weak transfer
homomorphism.
 
1. For every a ∈ H , we have L H (a) = L B θ(a) . In particular, an element a ∈ H
is an atom in H if and only if θ(a) is an atom in B.
2. L(H ) = L(B). In particular, (H ) = (B), Uk (H ) = Uk (B), and ρk (H )
= ρk (B) for every k ∈ N.
 
Proof. Since 2. follows directly from 1., we prove 1. Let a ∈ H . If n ∈ L B θ(a) , then
θ(a) = v1 · . . . · vn with v1 , . . . , vn ∈ A(B), and thus (WT2) implies that n ∈ L H (a).
Conversely, let n ∈ L H (a). Then there are u 1 , . . . , u n ∈ A(H ) such that a = u 1 · . . . ·
u n . Thus θ(a) = θ(u 1 ) · . . . · θ(u n ), and we have to verify that θ(u 1 ), . . . , θ(u n ) ∈
A(B). Let i ∈ [1, n]. Property (T1) implies that θ(u i ) is not a unit. Since B is atomic,
there are m ∈ N and w1 , . . . , wm ∈ A(B) such that θ(u i ) = w1 · . . . · wm . Since this
factorization can be lifted and u i is an atom, it follows that m = 1 and that θ(u i ) =
w1 ∈ A(B). Since an element of an atomic monoid is an atom if and only if its set of
lengths equals {1}, the statement follows.

Next we discuss the most classic example of a transfer homomorphism and its appli-
cation. This is the homomorphism from a commutative Krull monoid to an associated
monoid of zero-sum sequences. If H is a commutative monoid, then H is Krull if and
only if Hred is Krull, and if this holds, then the canonical epimorphism π : H → Hred
is a transfer homomorphism. Thus, in the following proposition we may restrict to
reduced Krull monoids for technical simplicity, but without loss of generality.

Proposition 12. Let H be a reduced commutative Krull monoid, F = F (P) a free

abelian monoid such that the embedding H → F is a cofinal divisor homomorphism
with class group G, and let G 0 = {[ p] | p ∈ P} ⊂ G = q(F)/q(H ) denote the set of
classes containing prime divisors. Then there
 is a transfer homomorphism β : H →
B (G 0 ). In particular, we have L(H ) = L B (G 0 ) .

Proof. Let 
β : F → F (G 0 ) be the unique epimorphism defined by 
β ( p) = [ p] for all
p ∈ P. We start with the following assertion.

A1. For every a ∈ F, we have  β (a) ∈ B (G 0 ) if and only if a ∈ H . Thus 

β (H )
−1  
= B (G 0 ) and 
β B (G 0 ) = H .
Proof of A1. Let a = p1 · . . . · p ∈ F where  ∈ N0 and p1 , . . . , p ∈ P. Then
 

β (a) = [ p1 ] · . . . · [ p ] ∈ F (G 0 ) and σ [ p1 ] · . . . · [ p ] = [ p1 ] + · · · + [ p ] = [a].

Since H → F is a divisor homomorphism, we have [a] = 0 ∈ G if and only if a ∈ H .

Therefore all assertions follow and we have proved A1.
Therefore we can define the homomorphism β =  β |H : H → B (G 0 ), and we assert
that it is a transfer homomorphism. Clearly H and B (G 0 ) are reduced and β is surjec-
tive. Thus (T1) reads as

B (G 0 ) = β(H ) and β −1 ({1}) = {1},

which holds true by A1. We continue with the following assertion.

976 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 A2. If a ∈ H , B, C ∈ B (G 0 ) and β(a) = BC, then there exist b, c ∈ H such that
a = bc, β(b) = B, and β(c) = C.
Proof of A2. Let a = p1 · . . . · p ∈ H , where  ∈ N0 and p1 , . . . , p ∈ P, and suppose
that β(a) = BC, say B = [ p1 ] · . . . · [ pk ] and C = [ pk+1 ] · . . . · [ p ] for some k ∈
[0, ]. By A1, we infer that b = p1 · . . . · pk ∈ H , c = pk+1 · . . . · p ∈ H and clearly
we have a = bc. This completes the proof of A2.
Clearly (WT2) follows from A2 by a straightforward induction,  and hence β is a
transfer homomorphism. Then Lemma 11 implies that L(H ) = L B (G 0 ) .

Definition. A monoid H is said to be a transfer Krull monoid (over G 0 ) if there exists

a weak transfer homomorphism θ : H → B (G 0 ) for a subset G 0 of an abelian group
G. If G 0 is finite, then we say that H is a transfer Krull monoid of finite type.

By Proposition 12, every commutative Krull monoid is a transfer Krull monoid. If

a monoid H ∗ has a weak transfer homomorphism to a commutative Krull monoid, say
θ : H ∗ → H , then the composition β ◦ θ : H ∗ → B (G 0 ) is a weak transfer homomor-
phism (with the notation of Proposition 12) and hence H ∗ is a transfer Krull monoid.
Thus a monoid is a transfer Krull monoid if and only if it allows a weak transfer homo-
morphism to a commutative Krull monoid.
Since monoids of zero-sum sequences are BF-monoids (this can be checked directly
or by using Lemma 2), Lemma 11 shows that transfer Krull monoids are BF-monoids.
However, the examples given below reveal that transfer Krull monoids need neither
be commutative nor completely integrally closed nor Mori (i.e., they do not neces-
sarily satisfy the ACC on divisorial ideals). Before we provide a list of transfer Krull
monoids, we briefly discuss general, not necessarily commutative Krull monoids (for
details we refer to [22]). This concept was introduced by Wauters in 1984 in complete
analogy to the ideal theoretic definition of commutative Krull monoids (compare with
Theorem 9.(a)).
Suppose that H is a monoid such that a H ∩ bH = ∅ and H a ∩ H b = ∅ for all
a, b ∈ H . Then H is called a Krull monoid (or a Krull order) if it is completely inte-
grally closed and satisfies the ACC on two-sided divisorial ideals. The isomorphisms
in Equation (2) between monoids of divisorial ideals carry over from the setting of
commutative domains to the setting of prime Goldie rings. Thus, in analogy to the
commutative setting, we have that a prime Goldie ring is a Krull ring if and only if
its monoid of cancelative elements is a Krull monoid [22, Proposition 5.1]. Moreover,
Krull monoids play a central role in the study of noetherian semigroup algebras. We
refer to [11] for recent surveys on noncommutative Krull rings and monoids.

Examples. 1. As outlined above, commutative Krull monoids (hence all the examples
given is Section 3) are transfer Krull monoids. But more generally, every normalizing
Krull monoid is a transfer Krull monoid by [22, Theorems 4.13 and 6.5] (a monoid H
is said to be normalizing if a H = H a for all a ∈ H ).
2. Let H be a half-factorial monoid. Since the map θ : H → B ({0}), defined by
θ(ε) = 1 for all ε ∈ H × and θ(u) = 0 for every u ∈ A(H ), is a transfer homo-
morphism, H is a transfer Krull monoid (over the trivial group {0}). Only recently
M. Roitman showed that commutative half-factorial domains need not be Mori [45].
Thus, transfer Krull monoids satisfy the ACC on principal left ideals and on principal
right ideals (since they are BF-monoids; see Lemma 2) but they do not necessarily
satisfy the ACC on divisorial ideals.
3. Let O be the ring of integers of an algebraic number field K , A a central simple
algebra over K , and R a classical maximal O-order of A. Then R • is a Krull monoid.

December 2016] SETS OF LENGTHS 977

If every stably free left R-ideal is free, then R • is a transfer Krull monoid over a ray
class group of O (note that this group is finite). If there is a stably free left R-ideal
that is not free, then R • is not a transfer Krull monoid. This is due to Smertnig [49,
Theorem 1.1 and 1.2], and for related results in a more general setting we refer to [4].
4. Let R be an order in an algebraic number field K , R the integral closure of R
(thus R is the ring of integers of K ), and let π : spec(R) → spec(R) be the natural
map defined by π(P) = P ∩ R for all nonzero prime ideals P  R.
4.(a) If R is seminormal, π is bijective, and there is an isomorphism δ : Pic(R)
→ Pic(R), then R • is a transfer Krull monoid over Pic(R) ([25, Theorem 5.8]).
4.(b) Suppose that π is not bijective. Since ρ(R • ) = ∞ by [23, Corollary 3.7.2],
R is not a transfer Krull monoid of finite type by Theorem 13. Moreover, R • is not
•

a transfer Krull monoid over an infinite abelian group G (compare Theorems 13 and
17).
5. Let D be a commutative Krull domain, R ⊂ D a subring having the same quo-
tient field such that D = R D × , D × ∩ R = R × , and (R : D) = m ∈ max(R). Then
the inclusion R • → D • is a transfer homomorphism and hence R • is a transfer Krull
monoid [23, Proposition 3.7.5]. Note that K + M-domains satisfy the above assump-
tions. Indeed, let R  D be commutative domains, m a nonzero maximal ideal of D,
and let K  L  D be subfields such that D = L + m and R = K + m. If D is Krull,
then the above assumptions are satisfied.
6. Let R be a bounded HNP (hereditary noetherian prime) ring, and note that a
commutative domain is an HNP ring if and only if it is a Dedekind domain. If every
stably free left R-ideal is free, then R • is a transfer Krull monoid [48, Theorem 4.4].
7. In [3], the authors study monoids of modules over HNP rings and thereby
monoids of the following type occur. Let H0 be a commutative Krull monoid but not
a group, D be a commutative monoid with D = {1 D }, and define H = (H0 \ H0× ) ×
D ∪ H0× × {1 D }. Then H is a transfer Krull monoid which is not completely integrally
closed [3, Proposition 6.1].
8. In [2], the authors study conditions under which monoids of upper triangular
matrices over commutative domains allow weak transfer homomorphisms to the under-
lying domain. Thus, in case of commutative Krull domains we obtain transfer Krull
monoids. Smertnig established characterizations on the existence of transfer homo-
morphisms from full matrix rings over commutative noetherian rings with no nonzero
nilpotent elements to commutative Krull domains [50, Theorem 5.18].

Sets of lengths in transfer Krull monoids (hence in all above examples) can be stud-
ied successfully with the strategy using transfer homomorphisms. Indeed combining
Lemma 11 and Proposition 10 we are able to apply the structural results for finitely
generated monoids (derived in Section 2) to transfer Krull monoids. This is done in
Theorem 13 whose proof follows from Propositions 7, 10, and from Lemma 11.

Theorem 13. Let H be a transfer Krull monoid of finite type. Then the set of distances
(H ) is finite, the elasticity ρ(H ) is finite, the unions Uk (H ) of sets of lengths are
finite for all k ∈ N, and they satisfy the structure theorem for unions of sets of lengths,
as given in Theorem 5.

We end this section by posing the following problem (see [30]).

Problem 14. Let R be an order in an algebraic number field. Characterize when the
monoid of nonzero elements R • and when the monoid of invertible ideals I ∗ (R) are
transfer Krull monoids, resp. transfer Krull monoids of finite type.

978 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
5. THE STRUCTURE THEOREM FOR SETS OF LENGTHS. In transfer Krull
monoids of finite type, not only do unions of sets of lengths do have a well-defined
structure (as given in Theorem 13), but the same is true for sets of lengths. We start
with a set of examples which demonstrate that the structure of sets of lengths is richer
than that of their unions.

Examples. Let G be a finite abelian

 group
 and G 0 ⊂ G a subset such that B (G 0 )
is not half-factorial. Since min  B (G 0 ) = gcd B (G 0 ) by Proposition 3, it follows
that for every B ∈ B (G 0 ) and every y ∈ L(B) we have
 
L(B) ⊂ y + dZ where d = min  B (G 0 ) .

Clearly every set of lengths  in B (G 0 ) is an arithmetical progression with difference

d if and only if  B (G 0 ) = {d}. We will demonstrate that arithmetical progressions
(of arbitrary lengths) actually occur as sets of lengths, but also several variants of
arithmetical progressions do occur naturally.
1. Arithmetical progressions. Let g ∈ G with ord(g) = n ≥ 3. Then U = g n,
−U = (−g)n , and V = (−g)g are atoms, (−U  )U = V , and clearly L (−U )U
n

= {2, n}. For every k ∈ N, we have L (−U ) U = 2k + {ν(n − 2) | ν ∈ [0, k]}.

k k

2. Sumsets of arithmetical progressions. Let r, k1 , . . . , kr ∈ N and n 1 , . . . , n r ∈

n
N≥3 . For every i ∈ [1, r ], let gi ∈ G with ord(gi ) = n i and we define Bi = (−gi )ni gi i .
k1
If g1 , . . . , gr  = g1  ⊕ · · · ⊕ gr , then by 1., L(B1 · . . . · Brkr ) =

k

r
L(B1 1 ) + · · · + L(Brkr ) = 2(k1 + · · · + kr ) + {ν(n i − 2) | ν ∈ [0, ki ]}
i=1

is the sum of r arithmetical progressions. Clearly the sum of r long arithmetical pro-
gressions with differences d1 , . . . , dr is an almost arithmetical progression with differ-
ence d = gcd(d1 , . . . , dr ).
3. Almost arithmetical progressions (AAPs, see Definition 2). We sketch the  argu-
ment that large sets of lengths in B (G 0 ) are AAPs with difference d = min  B (G 0 )
(for a formal statement and proof we refer to [23, Theorem 4.3.6]).
We proceed as at the beginning of the proof of Theorem 5. Clearly there exist an ele- 
ment C0 ∈B (G 0 )and m ∈ N such that {m, m + d} ⊂ L(C0 ). Since d = gcd  B (G 0 ,
ψ = ρ  B (G 0 − 1 ∈ N. Then L 0 = {k0 , k0 + d, . . . , k0 + ψd} ⊂ L(C) where
ψ
C = C0 and k0 = ψm. Now pick any large element A ∈ B (G 0 ), where by large we
mean that A is divisible by C. Thus, for some B ∈ B (G 0 ), we have

A = BC and L 0 + L(B) ⊂ L(C) + L(B) ⊂ L(A).

Since L(B) can be viewed as an arithmetical progression with difference d which has
gaps (whose number is controlled by ψ), the sumset L0 + L(B) is an arithmetical
progression with difference d. Thus, if A is large (with respect to the parameters d and
ψ depending on G 0 ), the set of lengths L(A) contains a long arithmetical progression
with difference d as the central part, whereas the initial and end parts may have gaps.
4. Almost arithmetical multiprogressions (the definition is given below). Let G 1 ⊂
G 0 be a subset and let B ∈ B (G 1 ) be such that L(B) is an AAP with difference d, say

L(B) = y + (L  ∪ L ∗ ∪ L  ) ⊂ y + dZ,

December 2016] SETS OF LENGTHS 979

where L ∗ is a long arithmetical progression with difference d (the long central part
of L(B)) such that min L ∗ = 0. It is not difficult to show that every finite subset of
N≥2 can be realized as a set of lengths (e.g., [23, Proposition 4.8.3]). Thus for any
set D ⊂ [0, d] with min D = 0 and max D = d, there is a zero-sum sequence C with
L(C) = x + D for some x ∈ N. Suppose that C ∈ B (G 2 ) for a subset G 2 ⊂ G 0 with
G 1  ∩ G 2  = {0}. Then

(x + y)+ (L  + D)  (L ∗ + D)  (L  + D)

⊂(x + y) + L(B) + L(C) ⊂ (x + y) + L(BC) ⊂ (x + y) + D + dZ.

Note that the long central part L ∗ + D repeats the set D periodically, whereas
the short initial and end parts L  + D and L  + D may contain gaps. Indeed, if
L ∗ = {0, d, 2d, . . . , d}, then

L ∗ + D = D ∪ (d + D) ∪ . . . ∪ (d + D) ⊂ D + dZ.

Consider transfer Krull monoids of finite type. Then their sets of lengths coincide
with sets of lengths of the monoid of zero-sum sequences. Moreover, the structure
theorem for sets of lengths for these monoids states that no other phenomena besides
those which we have described in the above examples can occur. We make this more
precise with the following definition.

Definition. Let d ∈ N, , M ∈ N0 and {0, d} ⊂ D ⊂ [0, d]. A subset L ⊂ Z is called

an almost arithmetical multiprogression (AAMP for short) with difference d, period
D, length  and bound M, if

L = y + (L  ∪ L ∗ ∪ L  ) ⊂ y + D + dZ,

where y ∈ Z is a shift parameter,

• the central part L ∗ satisfies min L ∗ = 0, L ∗ = [0, max L ∗ ] ∩ (D + dZ), and  ∈ N
is maximal such that d ∈ L ∗ ,
• the initial part L  satisfies L  ⊂ [−M, −1], and
• the end part L  satisfies L  ⊂ max L ∗ + [1, M].

Note that AAMPs are finite subsets of the integers, that an AAMP with period
D = {0, d} is an AAP, and that an AAMP with period D = {0, d} and bound M = 0 is
a usual arithmetical progression with difference d. As it was with AAPs (see Definition
2), every single finite set is an AAMP with a trivial choice of parameters (let L ∗ be
a singleton and set M = max L). To discuss one example of an AAMP (with natural
k
parameters), let n = p11 · . . . · prkr , where r, k1 , . . . , kr ∈ N and p1 , . . . , pr are distinct
primes. We consider the set A = {a ∈ [0, n] | gcd(a, n) > 1} ∪ {0} and observe that
A = ∪ri=1 pi N0 ∩ [0, n]. Setting d = p1 · . . . · pr and D = A ∩ [0, d], we obtain that

A = D + {0, d, 2d, . . . , (n/d − 1)d} ⊂ D + dZ

is an AAMP with difference d, period D, and bound M = 0.

Consider an atomic monoid with nonempty set of distances. Lemma 1 shows that
sets of lengths become arbitrarily large. The structure theorem for sets of lengths (for-
mulated below) states that the set of distances is finite (whence there are only finitely

980 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
many periods D with differences in (H )) and there is one global bound M for all sets
of lengths. Thus (with the above notation) long sets of lengths have a highly structured
central part L ∗ , and L ∗ is the only part of the set of lengths that can become arbitrarily
large, whereas the initial and end parts are universally bounded.

Theorem 15 (Structure Theorem for Sets of Lengths). Let H be a transfer Krull

monoid of finite type. Then the set of distances is finite and there is some M ∈ N0 such
that every L ∈ L(H ) is an AAMP with some difference d ∈ (H ) and bound M.

The above theorem was first proved in [21] (in a slightly weaker version), and a
detailed proof can be found in [23, Chapter 4]. To provide an additional example for
the validity of the structure theorem, take a commutative Mori domain R with complete
 and with nontrivial conductor f = (R : R).
integral closure R, 
 If the class group C ( R)
and the residue class ring R/f are both finite, then the structure theorem holds true [23,
Theorems 2.11.9 and 4.6.6] (this setting includes orders in algebraic number fields).
It is an open problem whether the assumption on the finiteness of R/f is necessary
for the validity of the structure theorem [24, 39]. On the other hand, for transfer Krull
monoids of finite type the description given by the above structure theorem is best
possible as the following realization theorem by Schmid [46] shows.

Theorem 16 (A Realization Theorem). Let M ∈ N0 and ∗ ⊂ N be a finite

nonempty set. Then there exists a commutative Krull monoid H with finite class
group such that the following holds: for every AAMP L with difference d ∈ ∗ and
bound M there is some y H,L ∈ N such that

y + L ∈ L(H ) for all y ≥ y H,L .

We end this section with results which are in sharp contrast to the structure theorem.
Indeed, they offer monoids where every finite subset of N≥2 occurs as a set of lengths.
Moreover, there is a transfer Krull monoid H1 and a monoid H2 , which is not a transfer
Krull monoid, whose systems of sets of lengths coincide (the first class is due to a
theorem of Kainrath [40] and the second example due to Frisch [17].

Theorem 17. For the following classes of monoids we have

 
L(H ) = {L ⊂ N≥2 | L is finite and nonempty} ∪ {0}, {1} .
• H is a transfer Krull monoid over an infinite abelian group G.
• H = Int(Z)• is the monoid of nonzero integer-valued polynomials over Z.
Moreover, Int(Z)• is not a transfer Krull monoid.

6. THE CHARACTERIZATION PROBLEM FOR SYSTEMS OF SETS OF

LENGTHS. Let H be a transfer Krull monoid of finite type. As we have seen in
Theorems 13 and 15, the finite type property implies the finiteness of the set of dis-
tances and the structural results on unions of sets of lengths and on sets of lengths.
In this final section we will always suppose that H is a transfer Krull monoid over a
finite abelian group, and this assumption will imply even stronger results.
Thus let H be a transfer Krull
 monoid
 over a finite abelian group G.
 Then Lemma
11 implies that L(H ) = L B (G) , and as usual we set L(G) := L B (G) . Recall
all the examples discussed in Section 3 and in Examples 4. In particular, rings of
integers of algebraic number fields are the prototypical examples for transfer Krull

December 2016] SETS OF LENGTHS 981

monoids over finite abelian groups. Classical philosophy in algebraic number theory
(dating back to the 19th century) states that the class group determines their arithmetic.
This idea can be justified (see [23, Section 7.1]), and concerning lengths of factor-
izations it holds true by Proposition 12. In the 1970s Narkiewicz posed the inverse
question of whether or not arithmetical behavior (in other words, behavior describing
the nonuniqueness of factorizations) characterize the class group. The first affirmative
answers ([23, Sections 7.1 and 7.2]) have an artificial flavor because the given charac-
terizations are based on rather abstract arithmetical properties which are designed to
do the characterization and play only a small role in other parts of factorization theory.
Since on the other hand sets of lengths are of central interest in factorization theory, it
has been natural to ask whether their structure is rich enough to force characterizations,
and this question is known as the characterization problem.
The Characterization Problem. Given two finite abelian groups G and G  with
D(G) ≥ 4 such that L(G) = L(G  ). Does it follow that G ∼ = G ?
Clearly a necessary condition for an affirmative answer is that G and G  are iso-
morphic if and only if the associated monoids B (G) and B (G  ) are isomorphic. This
necessary condition is guaranteed by Proposition 10.4. Answering the characterization
problem is a long-term goal in the study of sets of lengths of transfer Krull monoids
over finite abelian groups. We start with two elementary results (Propositions 18 and
19). Then we will be in a position to analyze the characterization problem in greater
detail. As usual we set
       
A(G) := A B(G) , (G) :=  B(G) , Uk (G) = Uk B(G) , and ρk (G) := ρk B(G)

for every k ∈ N. Since A(G) is finite (see Propositions 7 and 9), the Davenport
constant

D(G) = max{|U | | U ∈ A(G)}

is finite. Clearly D(G) is the smallest integer  ∈ N such that every sequence S over G
of length |S| ≥  has a zero-sum subsequence T of length |T | ≥ 1. The significance of
D(G) for the study of sets of lengths will become clear in ournext result. If |G| ≤ 2,
then D(G) = |G| and Proposition 10.2 implies that L(G) = {k} | k ∈ N0 , whence
(G) = ∅, and Uk (G) = {k} for every k ∈ N. Thus we suppose that 2 < |G| < ∞.

Proposition 18. Let G be a finite abelian group with |G| ≥ 3.

1. For every k ∈ N, Uk (G) is an interval, ρ(G) = D(G)/2, ρ2k (G) = kD(G), and
 
D(G)
1 + kD(G) ≤ ρ2k+1 (G) ≤ kD(G) + .
2

2. (G) is an interval with min (G) = 1 and max (G) ≤ D(G) − 2.

Proof. 1. Let k ∈ N. First, we show that Uk (G) is an interval. Note that it suf-
fices to prove that [k, ρk (G)] ⊂ Uk (G). Indeed, suppose that this is done, and
let  ∈ [min Uk (G), k]. Then  ≤ k ≤ ρ (G), hence k ∈ U (G) and consequently
 ∈ Uk (G).
Thus let  ∈ [k, ρk (G)] be minimal such that [, ρk (G)] ⊂ Uk (G) and assume to the
contrary that  > k. Let  be the set of all A ∈ B (G) such that {k, j} ⊂ L(A) for some
j ≥ , and let B ∈  be such that |B| is minimal. Then B = U1 · . . . · Uk = V1 · . . . ·
V j , where j ≥  and U1 , . . . , Uk , V1 , . . . , V j ∈ A(G). Since j > k, we have B = 0|B| ,

982 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
and (after renumbering if necessary) we may assume that Uk = g1 g2 U  and V j−1 V j
= g1 g2 V  , where g1 , g2 ∈ G and U  , V  ∈ F (G). Then Uk = (g1 + g2 )U  ∈ A(G),

and we suppose that V j−1 = (g1 + g2 )V  = W1 · . . . · Wt , where t ∈ N and W1 , . . . , Wt
∈ A(G). If B = U1 · . . . · Uk−1 Uk , then |B  | < |B| and B  = V1 · . . . · V j−2 W1 · . . . ·


Wt . By the minimal choice of |B|, it follows that j − 2 + t < , hence t = 1, j = ,

and  − 1 ∈ Uk (G), a contradiction.
Second, we study ρk (G). We start with the following assertion.
A. If A = 0m B ∈ B (G), with m = v0 (A) ∈ N0 and B ∈ B (G), then

D(G)
2 max L(A) − m ≤ |A| ≤ D(G) min L(A) − m(D(G) − 1) and ρ(A) ≤ .
2
Proof of A. Let A = 0m U1 · . . . · U where , m ∈ N0 and U1 , . . . , U ∈ A(G) \ {0}.
Then 2 ≤ |Uν | ≤ D(G) for all ν ∈ [1, ] and hence

m + 2 ≤ |A| ≤ m + D(G).

Choosing  = min L(B) and  = max L(B) we obtain the first inequalities, and then
we obtain
max L(A) m + max L(B) max L(B) D(G)
ρ(A) = = ≤ ≤ .
min L(A) m + min L(B) min L(B) 2

The proof of A and Proposition 4.2 imply that ρk (G) ≤ kρ(G) ≤ kD(G)/2. If U =
D(G)  k
g1 · . . . · gD(G) ∈ A(G), then (−U )k U k = i=1 (−gi )gi , whence kD(G) ≤ ρ2k (G)
and thus ρ2k (G) = kD(G). Furthermore, it follows that

(2k + 1)D(G)
1 + kD(G) = ρ1 (G) + ρ2k (G) ≤ ρ2k+1 (G) ≤ .
2
Finally, Proposition 4.2 implies that ρ(G) = D(G)/2.
2. The proof that (G) is an interval is similar but trickier than that of 1., and
we refer to [31]. It is easy to verify that 1 ∈ (G), and we encourage the reader
to do so. Next we prove that max (G) ≤ D(G) − 2. If A = 0k A with k ∈ N0 and
A ∈ B (G) with 0  A, then L(A ) = k + L(A) and (L(A )) = (L(A)). Thus we
have to prove that max (L(A)) ≤ D(G) − 2 for all A ∈ B (G) with 0  A, and we
proceed by induction on |A|. Suppose that

A = U1 · . . . · Ui = V1 · . . . · Vk , where i < k, U1 , . . . , Ui , V1 , . . . , Vk ∈ A(G),

and L(A) ∩ [i, k] = {i, k}. If |A| ≤ 2D(G), then k ≤ D(G) and k − i ≤ D(G) − 2.
Suppose that |A| > 2D(G) and that max (L(A )) ≤ D(G) − 2 for all A with |A |
< |A|. If |V j | ≥ i for all j ∈ [1, k], then

ki ≤ |V1 · . . . · Vk | = |U1 · . . . · Ui | ≤ iD(G) and hence k − i ≤ D(G) − 2.

Suppose that there is a j ∈ [1, k] such that |V j | < i, say j = 1, V1 | U1 · . . . · Ui−1 , and
let U1 · . . . · Ui−1 = V1 W2 · . . . · W with  ∈ N and W2 , . . . , W ∈ A(G). We distin-
guish two cases.
Case 1:  ≥ i. Since L(U1 · . . . · Ui ) ∩ [i, k] = {i, k}, it follows that L(U1 · . . . · Ui−1 ) ∩
[i, k − 2] = ∅ and hence  ≥ k − 1. We may suppose that  ≥ k − 1 is minimal such

December 2016] SETS OF LENGTHS 983

that U1 · . . . · Ui−1 satisfies such an equation. Then the induction hypothesis implies
that  − (i − 1) ≤ D(G) − 2 and hence k − i ≤ ( + 1) − i ≤ D(G) − 2.
Case 2:  ≤ i − 1. Since V1 · . . . · Vk = U1 · . . . · Ui = V1 W2 · . . . · W Ui , it follows
that V2 · . . . · Vk = Ui W2 · . . . · W . Note that L(V2 · . . . · Vk ) ∩ [i, k − 2] = ∅, and
suppose that  ≤ i − 1 is maximal such that V2 · . . . · Vk satisfies such an equation.
Then the induction hypothesis implies that (k − 1) −  ≤ D(G) − 2 and hence k − i
= (k − 1) − (i − 1) ≤ k − 1 −  ≤ D(G) − 2.

The state of the art on ρ2k+1 (G) is discussed in [47]. For some small groups G the
system L(G) can be written down explicitly.

Proposition 19.
 
1. L(C3 ) = L(C2 ⊕ C2 ) = y + 2k + [0, k] y, k ∈ N0 .
   
2. L(C4 ) = y +k + 1+[0, k] | y, k ∈ N0 ∪ y +2k +2 · [0, k] | y, k ∈ N0 .
 
3. L(C23 ) = y + (k + 1) + [0, k]  y ∈ N0 , k ∈ [0, 2] ∪ 
y + k + [0, k] y ∈ N0 , k ≥ 3 ∪ y + 2k + 2 · [0, k] y, k ∈ N0 .
4. L(C32 ) = {[2k, ] | k ∈ N0 ,  ∈ [2k, 5k]}
∪ {[2k + 1, ] | k ∈ N,  ∈ [2k + 1, 5k + 2]} ∪ {{1}}.

Proof. We prove the first statement. The proofs of the remaining statements are
similar but more lengthy (details can be found in [29, Proposition 4.2]). Suppose
that G is cyclic of order three, say G = {0, g, −g}. Then A(G) = {0, U = g 3 , −U,
V = (−g)g}, D(G) = 3, and (−U )U = V 3 is the only minimal relation. Clearly
L(V 3 ) = {2, 3}, L(V 3k ) = {2k, 2k + 1, . . . , 3k} = 2k + [0, k], and L(0 y V 3k ) = y
+ 2k + [0, k] for all y, k ∈ N0 . Since (G) = {1}, ρ2k (G) = 3k, and ρ2k+1 (G) =
3k + 1 for every k ∈ N by Proposition 18, there are no further sets of lengths.
Suppose that G is an elementary 2-group of rank two, say G = {0, e1 , e2 , e1 + e2 }.
Then A(G) = {0, U = e1 e2 (e1 + e2 ), V1 = e12 , V2 = e22 , V3 = (e1 + e2 )2 }, and hence
U 2 = V1 V2 V3 is the only minimal relation. Now the proof runs along the same lines as
above.

One big difficulty in all work on the characterization problem stems from the fact
that most sets of lengths over any finite abelian group are intervals. To make this pre-
cise we mention two deep results without proof.

Theorem 20 (Sets of lengths which are intervals). Let G be a finite abelian group
with |G| ≥ 3.
1. If A is a zero-sum sequence whose support supp(A) ∪ {0} is a subgroup of G,
then L(A) is an interval.
2. If R is the ring of integers of an algebraic number field K with class group G,
then

#{a R | N K /Q (a R) ≤ x, L(a) is an interval}

lim = 1.
x→∞ #{a R | N K /Q (a R) ≤ x}

The first statement is a result in additive combinatorics which can be found in [23,
Theorem 7.6.8]. The limit formula is based on the first statement and on the analytic

984 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
machinery of counting functions [23, Theorem 9.4.11]. In the 1960s Narkiewicz ini-
tiated a systematic study of the asymptotic behavior of counting functions associated
with nonunique factorizations. We refer to the monographs [42, Chapters 7 and 9], [23,
Chapters 8 and 9], and to [39] (analytic monoids, introduced in [39], are Krull monoids
which have an abstract norm function satisfying axioms which allow to develop a the-
ory of L-functions).
In spite of Theorem 20 and having Propositions 18 and 19 at our disposal, we start
with a more detailed analysis of the characterization problem. We have seen that

L(C1 ) = L(C2 ) and L(C3 ) = L(C2 ⊕ C2 ).

An abelian group G has Davenport constant D(G) ≤ 3 if and only if it is either cyclic
of order |G| ≤ 3 or isomorphic to C2 ⊕ C2 . Thus we focus on groups whose Davenport
constant is at least four. Let G be a finite abelian group with D(G) ≥ 4, say


r
G∼
= Cn1 ⊕ · · · ⊕ Cnr with 1 < n 1 | · · · | n r and set D∗ (G) = 1 + (n i − 1).
i=1

Clearly the system L(G) depends only on G and hence on the group invariants
(n 1 , . . . , n r ). Thus L(G) as a whole as well as the invariants controlling L(G)—such
as the set of distances (G) and the kth elasticities ρk (G)—allow a description in
(n 1 , . . . , n r ). We demonstrate the complexity of such problems by considering ρ2 (G).
By Proposition 18, we have ρ2 (G) = D(G). The Davenport constant D(G) is one
of the most classical zero-sum invariants which has been studied since the 1960s. If
(e1 , . . . , er ) is a basis of G with ord(ei ) = n i for each i ∈ [1, r ], then


r
n −1
A = (e1 + · · · + er ) ei i ∈ A(G),
i=1

and hence D∗ (G) = |A| ≤ D(G). It has been known since the 1960s that equality
holds for p-groups and groups of rank at most two [23, Theorem 5.8.3]. It is an open
problem whether equality holds for groups of rank three, but for every r ≥ 4 there
are infinitely many groups G having rank r and for which D∗ (G) < D(G) holds. We
refer to [28, 47] for a survey of what is known on parameters controlling L(G). We
first show a simple finiteness result and then present one result (a proof can be found
in [23, Theorem 6.6.3]) revealing characteristic phenomena of L(G) for cyclic groups
and elementary 2-groups.

Lemma 21. Let G be a finite abelian group with D(G) ≥ 4. Then there are only
finitely many abelian groups G  (up to isomorphism) such that L(G) = L(G  ).

Proof. If G  is an abelian group such that L(G  ) = L(G), then Proposition 18 implies
D(G) = ρ2 (G) = ρ2 (G  ) = D(G  ) and hence D∗ (G  ) ≤ D(G). Since there are only
finitely many G  (up to isomorphism) such that D∗ (G  ) is bounded above by a con-
stant, there are only finitely many groups G  for which L(G  ) = L(G) can hold.

Proposition 22. Let G be a finite abelian group with D(G) ≥ 4. Then {2, D(G)} ∈
L(G) if and only if G is either cyclic or an elementary 2-group.

The next theorem gathers what is known on the characterization problem.

December 2016] SETS OF LENGTHS 985

Theorem 23. Let G be a finite abelian group with D(G) ≥ 4, and let G  be an abelian
group with L(G) = L(G  ). Then G and G  are isomorphic in each of the following
cases.
1. G∼= Cn1 ⊕ Cn2 where n 1 , n 2 ∈ N with n 1 | n 2 and n 1 + n 2 > 4.
2. G is an elementary 2-group.
3. G∼= Cnr where r, n ∈ N with n ≥ 2 and 2r < n − 2.
4. D(G) ≤ 11.

Proof of a special case. We give a sketch of the proof for cyclic groups and for
elementary 2-groups. Let G be either cyclic or an elementary 2-group with D(G) ≥ 4,
and let G  be any abelian group with L(G) = L(G  ). Since D(G) = ρ2 (G) = ρ2 (G  )
= D(G  ) by Proposition 18 and since {2, D(G)} ∈ L(G) by Proposition 22, it follows
that {2, D(G  )} ∈ L(G  ). Again Proposition 22 implies that G  is either cyclic or an
elementary 2-group. There are two proofs showing that the system of sets of lengths
of cyclic groups and that of elementary 2-groups (with the same Davenport constant)
are distinct ([28, Corollary 5.3.3, page 77] or [23, Theorem 7.3.3]), and neither of them
is elementary. Both proofs use the Savchev–Chen structure theorem for long zero-sum
free sequences over cyclic groups ([28, Theorem 5.1.8, page 61], [34, Chapter 11]), or
related statements. To discuss one approach, let k ∈ N and consider the inequality for
ρ2k+1 (G) given in Proposition 18. Elementary examples show in case of elementary
2-groups that we have equality on the right side, whereas for cyclic groups we have
equality on the left side ([28, Theorem 5.3.1, page 75]. Detailed proofs can be found
in [28, Corollary 5.3.3, page 77] and [23, Theorem 7.3.3].
Now suppose that G has rank two, say G ∼ = Cn1 ⊕ Cn2 where n 1 , n 2 ∈ N with
1 < n 1 | n 2 and n 1 + n 2 > 4, and let G  be any abelian group such that L(G) = L(G  ).
The proof that G and G  are isomorphic has two main ingredients. First, it is based on
the characterization of all minimal zero-sum sequences over G of length D(G). This
has been done in a series of papers by Gao, Geroldinger, Grynkiewicz, Reiher, and
Schmid (see [7] for the characterization and detailed references). Second, it is based
on the structure theorem for sets of lengths (Theorem 15), on an associated inverse
result ([23, Proposition 9.4.9]), and on a detailed study of the set of minimal distances
∗ (G) ([33]), which is defined as

∗ (G) = {min (G 0 ) | G 0 ⊂ G with (G 0 ) = ∅} ⊂ (G).

Detailed proofs of 1., 3., and 4. can be found in [29, 32, 51].

We end this survey with the conjecture stating that the characterization problem has
a positive answer for all finite abelian groups G having Davenport constant D(G) ≥ 4.

Conjecture 24. Let G be a finite abelian group with D(G) ≥ 4. If G  is an abelian

group with L(G) = L(G  ), then G and G  are isomorphic.

ACKNOWLEDGMENT. I would like to thank Daniel Smertnig, Salvatore Tringali, Qinghai Zhong, and all
the anonymous referees for their careful reading. Their comments helped me to eliminate a couple of flaws,
to improve the presentation of the paper, and to increase its readability. Furthermore, I would like to thank the
Editor, Scott Chapman, for his encouragement to write this survey and all his patience, and Moshe Roitman
for showing to me that transfer Krull monoids need not be Mori.
This work was supported by the Austrian Science Fund FWF, Project Number P 28864-N35.

986 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
REFERENCES

1. D. D. Anderson, Ed., Factorization in Integral Domains. Lect. Notes Pure Appl. Math., Vol. 189, Dekker,
New York, 1997.
2. D. Bachman, N. Baeth, J. Gossell, Factorizations of upper triangular matrices, Linear Algebra Appl. 450
(2014) 138–157.
3. N. R. Baeth, A. Geroldinger, D. J. Grynkiewicz, D. Smertnig, A semigroup-theoretical view of direct-
sum decompositions and associated combinatorial problems, J. Algebra Appl. 14 (2015) 1550016 (60
pages).
4. N. R. Baeth, D. Smertnig, Factorization theory: From commutative to noncommutative settings, J. Alge-
bra 441 (2015) 475–551.
5. N. R. Baeth, R. Wiegand, Factorization theory and decomposition of modules, Amer. Math. Monthly 120
(2013) 3–34.
6. P. Baginski, S. T. Chapman, Factorizations of algebraic integers, block monoids, and additive number
theory, Amer. Math. Monthly 118 (2011) 901–920.
7. P. Baginski, A. Geroldinger, D. J. Grynkiewicz, A. Philipp, Products of two atoms in Krull monoids and
arithmetical characterizations of class groups, Eur. J. Comb. 34 (2013) 1244–1268.
8. P.-J. Cahen, J.-L. Chabert, Integer-Valued Polynomials, Vol. 48, Amer. Math. Soc. Surveys and Mono-
graphs, Washington, DC, 1997.
9. , What you should know about integer-valued polynomials, Amer. Math. Monthly 123 (2016)
311–337.
10. S. T. Chapman, A tale of two monoids: A friendly introduction to nonunique factorizations, Math. Mag.
87 (2014) 163–173.
11. S. T. Chapman, M. Fontana, A. Geroldinger, B. Olberding, Eds., Multiplicative Ideal Theory and Factor-
ization Theory. Proceedings in Mathematics and Statistics, Vol. 170, Springer, New York, 2016.
12. S. T. Chapman, W. W. Smith, Factorization in Dedekind domains with finite class group, Isr. J. Math. 71
(1990) 65–95.
13. J. Coykendall, W. W. Smith, On unique factorization domains, J. Algebra 332 (2011) 62–70.
14. M. Delgado, P. A. Garcı́a-Sánchez, J. Morais, “numericalsgps”: A gap package on numerical semi-
groups, (http://www.gap-system.org/Packages/numericalsgps.html).
15. A. Facchini, Direct sum decomposition of modules, semilocal endomorphism rings, and Krull monoids,
J. Algebra 256 (2002) 280–307.
16. M. Fontana, E. Houston, T. Lucas, Factoring Ideals in Integral Domains, Lecture Notes of the Unione
Matematica Italiana, Vol. 14, Springer, New York, 2013.
17. S. Frisch, A construction of integer-valued polynomials with prescribed sets of lengths of factorizations,
Monatsh. Math. 171 (2013) 341–350.
18. , Relative polynomial closure and monadically Krull monoids of integer-valued polynomials, in
Multiplicative Ideal Theory and Factorization Theory. Eds. S. T. Chapman, M. Fontana, A. Geroldinger,
B. Olberding. Springer, New York, 2016. 145–157.
19. W. Gao, A. Geroldinger, On products of k atoms, Monatsh. Math. 156 (2009) 141–157.
20. P. A. Garcı́a-Sánchez, An overview of the computational aspects of nonunique factorization invariants, in
Multiplicative Ideal Theory and Factorization Theory. Eds. S. T. Chapman, M. Fontana, A. Geroldinger,
B. Olberding. Springer, New York, 2016. 159–181.
21. A. Geroldinger, Über nicht-eindeutige zerlegungen in irreduzible elemente, Math. Z. 197 (1988)
505–529.
22. , Non-commutative Krull monoids: A divisor theoretic approach and their arithmetic, Osaka J.
Math. 50 (2013) 503–539.
23. A. Geroldinger, F. Halter-Koch, Non-Unique Factorizations. Algebraic, Combinatorial and Analytic
Theory, Chapman & Hall/CRC, Boca Raton, FL, 2006.
24. A. Geroldinger, F. Kainrath, On the arithmetic of tame monoids with applications to Krull monoids and
Mori domains, J. Pure Appl. Algebra 214 (2010) 2199–2218.
25. A. Geroldinger, F. Kainrath, A. Reinhart, Arithmetic of seminormal weakly Krull monoids and domains,
J. Algebra 444 (2015) 201–245.
26. A. Geroldinger, G. Lettl, Factorization problems in semigroups, Semigroup Forum 40 (1990) 23–38.
27. A. Geroldinger, S. Ramacher, A. Reinhart, On v-Marot Mori rings and C-rings, J. Korean Math. Soc. 52
(2015) 1–21.
28. A. Geroldinger, I. Ruzsa, Combinatorial Number Theory and Additive Group Theory, Advanced Courses
in Mathematics, CRM Barcelona, Birkhäuser, 2009.
29. A. Geroldinger, W. A. Schmid, A characterization of class groups via sets of lengths, http://arxiv.
org/abs/1503.04679.

December 2016] SETS OF LENGTHS 987

30. A. Geroldinger, W. A. Schmid, Q. Zhong, Systems of sets of lengths: Transfer Krull monoids versus
weakly Krull monoids, http://arxiv.org/abs/1606.05063.
31. A. Geroldinger, P. Yuan, The set of distances in Krull monoids, Bull. Lond. Math. Soc. 44 (2012)
1203–1208.
32. A. Geroldinger, Q. Zhong, A characterization of class groups via sets of lengths II, J. Théor. Nombres
Bordx. (forthcoming).
33. , The set of minimal distances in Krull monoids, Acta Arith. 173 (2016) 97–120.
34. D. J. Grynkiewicz, Structural Additive Theory. Developments in Mathematics, Springer, New York, 2013.
35. F. Halter-Koch, Finitely generated monoids, finitely primary monoids and factorization properties of
integral domains, in Factorization in Integral Domains. Ed. D. D. Anderson. Lect. Notes Pure Appl.
Math., Vol. 189, Dekker, New York, 1997. 73–112.
36. , Ideal Systems. An Introduction to Multiplicative Ideal Theory. Marcel Dekker, New York, 1998.
37. G. Hegedüs, Z. Li, J. Schicho, H. P. Schröcker, From the fundamental theorem of algebra to Kempe’s
universality theorem, Int. Math. Nachrichten 229 (2015) 13–26.
38. B. Jacobson, Matrix number theory: An example of non-unique factorization, Amer. Math. Monthly 72
(1965) 399–402.
39. F. Kainrath, Arithmetic of Mori domains and monoids: The Global Case, Multiplicative Ideal Theory and
Factorization Theory. Springer, 2016. 183–218.
40. F. Kainrath, Factorization in Krull monoids with infinite class group, Colloq. Math. 80 (1999) 23–30.
41. P. Malcolmson, F. Okoh, Half-factorial subrings of factorial domains, J. Pure Appl. Algebra 220 (2016)
877–891.
42. W. Narkiewicz, Elementary and Analytic Theory of Algebraic Numbers. Third ed. Springer, New York,
2004.
43. A. Reinhart, On the divisor-class group of monadic submonoids of rings of integer-valued polynomials,
Commun. Korean Math. Soc. (forthcoming).
44. , On monoids and domains whose monadic submonoids are Krull, in Commutative Algebra.
Recent Advances in Commutative Rings, Integer-Valued Polynomials, and Polynomial Functions. Eds.
M. Fontana, S. Frisch, S. Glaz. Springer, New York, 2014. 307–330.
45. M. Roitman, Half-factorial domains, manuscript.
46. W. A. Schmid, A realization theorem for sets of lengths, J. Number Theory 129 (2009) 990–999.
47. , Some recent results and open problems on sets of lengths of Krull monoids with finite class
group, in Multiplicative Ideal Theory and Factorization Theory. Eds. S. T. Chapman, M. Fontana,
A. Geroldinger, B. Olberding. Springer, New York, 2016. 323–352.
48. D. Smertnig, Factorizations in bounded hereditary noetherian prime rings, http://arxiv.org/abs/
1605.09274.
49. , Sets of lengths in maximal orders in central simple algebras, J. Algebra 390 (2013) 1–43.
50. , Factorizations of elements in noncommutative rings: A Survey, in Multiplicative Ideal Theory
and Factorization Theory. Eds. S. T. Chapman, M. Fontana, A. Geroldinger, B. Olberding. Springer, New
York, 2016. 353–402.
51. Q. Zhong, Sets of minimal distances and characterizations of class groups of Krull monoids, http://
arxiv.org/abs/1606.08039.

ALFRED GEROLDINGER received his MSc in mathematics from the University of Vienna, his MSc in
computer science from the Vienna University of Technology, and his Ph.D. from the University of Graz. He
is professor of mathematics at the University of Graz, and his research interests include commutative algebra
and number theory.
University of Graz, NAWI Graz, Institute of Mathematics and Scientific Computing
Heinrichstrasse 36, 8010 Graz, Austria
alfred.geroldinger@uni-graz.at

988 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 Balanced Factorizations
Anton A. Klyachko and Anton N. Vassilyev

Abstract. Any rational number can be factored into a product of several rationals whose
sum vanishes. This simple but nontrivial fact was suggested as a problem on a mathematical
olympiad for high school students. We completely solve similar questions in all finite fields
and in some other rings, e.g., in the complex and real matrix algebras. Also, we state several
open questions.

1. INTRODUCTION. “Prove that any rational number can be factored into a prod-
uct of several rationals whose sum vanishes.”
This problem was invented by the second author and suggested at the Kazakhstan
republican mathematical olympiad for high-school students in 2013 [8]. A similar
question about arbitrary fields of characteristic not two was suggested at the Alge-
bra olympiad for university students at Moscow university in 2014 [9]. Afterwards,
we learned that the problem had been considered earlier [1] (also in an educational
context).
The existence of such balanced factorizations is easy to prove in any field of charac-
teristic not two (see Theorem 1 below). However, the question on the possible numbers
of factors in such factorizations is much more difficult. This question is the main sub-
ject of our paper. For example, any rational admits a balanced factoring into a product
of five factors, but some rationals do not admit balanced factorings into products of
three factors [1]; the question about four factors is open and seems to be difficult.1 For
instance, the author of [1] reproduced the following letter by M. A. Tsfasman to him:

3 = (363/70) · (20/77) · (−49/110) · (−5). Uf. . . Vax M.A.

This is in Russian, but no translation is needed—the letter contains the first discov-
ered balanced decomposition of 3 into four factors in the field of rationals (along
with an interjection and signature). Such decompositions of 1 and 2 are less impres-
sive: 1 = 1 · 1 · (−1) · (−1) and 2 = 16 · 92 · (− 32 ) · (−4). Computer-generated balanced
decompositions of first fifty positive integers into products of four rational factors can
be found in [1].
A similar problem for finite fields seems to be easier. Indeed, in each given finite
field, we can use a brute-force search and find all element admitting balanced decom-
positions into any given number of factors. We later realized that some more advanced
algebra gives a complete and computer-free solution to the problem in all finite fields.
One of our main results (Theorem 2) describes all pairs (q, k) such that every ele-
ment of the q-element field Fq admits a balanced decomposition into a product of k
factors. The answer is nontrivial and rather complicated. For instance, it turns out that,
in all finite fields except exactly one, each element admits a balanced factoring into
a product of at most three factors. The role of the unique exception is played by the
1 When this paper was written, we learned that this question has a positive answer [2].
http://dx.doi.org/10.4169/amer.math.monthly.123.10.989
MSC: Primary 12E12, Secondary 14H52; 16U99

December 2016] BALANCED FACTORIZATIONS 989

seven-element field F7 . The main tool of our study of finite fields is Hasse’s estimate
of the number of rational points on an elliptic curve over a finite field.
In Section 2 we prove also that, in each field of characteristic not 2, there is a
“universal” formula allowing us to obtain a balanced factorization of almost any ele-
ment. For example, formula (1) gives a balanced factorization into five factors for any
nonzero element (in any field of characteristic not two). We prove that similar formulae
exist for 6, 7, and any larger numbers of factors but do not exist for three factors. This
fact is deduced from the Mason–Stothers theorem (the abc-theorem for polynomials).
In Section 3 we show that the question about balanced factorizations in finite-
dimensional algebras is in essence reduced to a similar question about fields. This
allows us to solve the problem completely for some natural algebras, e.g., the matrix
algebras over C and R. The last section contains a list of questions remaining open.

2. GENERAL RESULTS AND REMARKS. The formal definition of our main

definition is as follows. Suppose that an element a of a ring is factored
 into a product
of the form a = a1 a2 · · · ak ; we call this factorization balanced if ai = 0.
First note that over algebraically closed fields the problem is trivial.
For any k  2 any element of any algebraically closed field admits a
balanced decomposition into a product of k factors.
Indeed, for a given a and k, we need to find a1 , . . . , ak such that a = a1 . . . ak and
a1 + · · · + ak = 0. The solution is straightforward; take, e.g., a3 = · · · = ak = 1 and
find a1 and a2 = 2 − k − a1 from the quadratic equation a1 (2 − k − a1 ) = a.
The second observation is that the problem is easy (though nontrivial) for any field
of characteristic not equal to two provided the number of factors is at least five.

Theorem 1. For each k  5, in any field of characteristic not equal to two, every
element decomposes into a product of k factors whose sum vanishes. For each k < 5,
there exists a field of characteristic not equal to two where a similar assertion is false.

Proof. Let us prove the first assertion. For zero element, we have nothing to prove; for
nonzero element a, we can write
 
a a 2 2
a = · · (−a) · · − . (1)
2 2 a a

This gives a balanced decomposition into five factors. A slight modification of (1)
gives a balanced decomposition of any element b into six factors:
 
a a 2 2
b = c − ca = · · (c − a) · · −
2
· (−c),
2 2 a a
2
where c is an element such that 0 = c2 = b, and a = c c−b . (Such a c exists, except
in the case where the field is F3 and b = 1; in this exceptional case we can take the
factoring b = 1 = 16 .)
A balanced decomposition into k  7 factors can be obtained by multiplying one
of the above decompositions and a balanced decompositions of minus one into two
factors: −1 = (−1) · 1. For example, we obtain the following balanced decomposition
of any element into a product of 100 factors:
 
a a 2 2
−b = −(c − ca) = · · (c − a) · · −
2
· (−c) · (−1)47 · 147 .
2 2 a a

990 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
(Note that −b is an arbitrary element if b is an arbitrary element.) This completes the
proof of the first assertion.
The second assertion follows from Theorem 2 (see the next section). For k  3, the
field F7 is a required example; for k = 4, we can take F3 . This completes the proof of
Theorem 1.

Note that, in this study, we consider no decompositions as “trivial.” We allow factors

to be 1, or −1, or anything, and the problem remains nontrivial. Actually, the role of
bad decompositions is played by so-called power decompositions, i.e., decompositions
with all factors equal. This is not a priori clear why such factorizations are useless, but
look at Theorem 4.

3. FIELDS.

Theorem 2. Suppose that k  2 is an integer and F is a finite field. Then, in F, any

element can be decomposed into a product of k factors whose sum vanishes if and only
if
either |F| = 2 and k is even,
or |F| = 4 and k = 3,
or |F| is a power of two but neither two nor four (and k is arbitrary),
or |F| ∈ {3, 5} and k ∈/ {2, 4},
or |F| = 7 and k ∈ / {2, 3},
or |F| is neither a power of two nor three nor five nor seven and k = 2.

In other words, the situation in finite fields is the following:

k = 2 k = 3 k = 4 k = 5, 7, 9, . . . k = 6, 8, 10, . . .
F2 yes no yes no yes
F3 no yes no yes yes
F4 yes no yes yes yes
F5 no yes no yes yes
F7 no no yes yes yes
F8 , F16 , F32 , F64 , . . . yes yes yes yes yes
F9 , F11 , F13 , F17 , . . . no yes yes yes yes.

Proof. Let us arrange the proof by the columns of this table.

Case 1: k = 2. In a finite field of characteristic two, any element is a square (because
the order of the multiplicative group of such a field is odd), i.e., each element is a
product of two equal factors whose sum vanishes, because the characteristic is two.
If the characteristic of a finite field is not two, then not every element is a square
and, therefore, not every element decomposes into a product of two factors whose
sum vanishes (because the equality a = x · (−x) implies that −a is a square; so, if
each element has a balanced decomposition into a product of two factors, then each
element is a square).
Case 2: k = 3. If the characteristic is three, then the order of the multiplicative group
q − 1 = 3k − 1 is not divisible by three and, therefore, each element is a cube and the
decomposition a = bbb is as required (because b + b + b = 0 in a field of character-
istic three).
To study the fields of other characteristics, we need the well-known Hasse’s estimate
(also known as the Hasse–Weil bound).

December 2016] BALANCED FACTORIZATIONS 991

Hasse’s estimate. (See, e.g., [5], Theorem V.1.1). The number of points of an elliptic
curve (i.e., a nonsingular and irreducible over the closure of the field projective curve
√
of genus one) over a finite q-element field Fq is at least q + 1 − 2 q. In particular,
this is true for nonsingular and irreducible (over the closure of the field) cubic curves
in the projective plane over Fq .

Let us continue the proof assuming that the characteristic is not three. We have to
show that the system of equations

x+y+z =0
(2)
x yz = a

over a finite field Fq has at least one solution for any a ∈ Fq . In other words, we have
to show that the cubic (affine) curve defined by the equation

x y(x + y) = −a

has at least one point over Fq . In homogeneous coordinates, the corresponding projec-
tive curve has the equation

XY(X + Y ) = −a Z 3 , (3)

and singular points of this curve are the solutions of the system of equations consisting
of equation (3) and its partial derivatives with respect to X , Y , and Z :
⎧
⎪
⎪ XY(X + Y ) = −a Z 3
⎪
⎨2XY + Y 2 = 0
. (4)
⎪2XY + X 2 = 0
⎪
⎪
⎩
−3a Z 2 = 0

We assume that a = 0, because if a = 0, then system (2) has an obvious solution

(zero). Therefore, (and since the characteristic is not three) the last equation of
(4) implies that Z = 0. The difference of the second and third equations shows
that X = ±Y ; now, the second equation shows that X and Y are zero (recall that
char Fq = 3). Thus, system (4) has no nonzero solutions, i.e., our projective curve has
no singular points over the closure of the field (if char Fq = 3). This automatically
implies that our curve is irreducible (and, therefore, elliptic), because a reducible
cubic curve always has a singular point (over the closure of the field). This is a point
of intersection of components.
Thus, we can apply Hasse’s estimate and conclude that projective cubic (3) has
more than three points over the field Fq if the characteristic of this field is not three
√
and q + 1 > 2 q + 3. This inequality holds for q  8. Thus, for q  8, the projective
curve contains more than three points. Hence, the corresponding affine curve contains
at least one point, because the intersection of an irreducible cubic with the line at infin-
ity cannot contain more than three points1 , i.e., system (2) has a solution as required.
It remains to investigate the fields F2 , F4 , F5 , and F7 .
In F2 , the unity obviously has no balanced decompositions into a product of three
factors (because factors cannot be zero, but 1 + 1 + 1 = 0).
1 In the case under consideration, the curve contains precisely three points at infinity: (1, 0, 0), (0, 1, 0), and

(1, −1, 0) (in homogeneous coordinates).

992 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 In F4 , any nonzero balanced product x yz of three factors cannot contain equal fac-
tors (because x + x = 0), therefore, there is exactly one such product. This is the
product of all nonzero elements of the field and it equals one; hence, elements differ-
ent from one and zero do not admit balanced decompositions into products of three
factors.
In F5 , system (2) has a solution: x = y = b, z = −2b, where b is a cubic root of − a2
(in F5 , any element is a cube).
The seven-element field indeed is an exception: ±3 have no balanced decomposi-
tions into products of three factors: if

x+y+z =0
,
x yz = ±3

then x, y, and z must be pairwise different. Indeed, if y = x, then z = −2x and ∓3

= 2x 3 , but ∓3 is not twice a cube (cubes in F7 are 0 and ±1). Certainly, no two
from x, y, and z can be opposite. So, only one possibility remains (up to signs
and permutations): x = ±1, y = ±2, and z = ±3. But the product of such
three numbers is ±1, not ±3. (The element 3 has a shorter balanced decomposition:
3 = 2 · (−2) but −3 has no such factorizations.)
Case 3: k = 4. Let us try to obtain a balanced decomposition of an element a ∈ F
into a product of four factors, where one factor is 1. The following argument (up to a
point) are similar to the proof in the case k = 3. We want to show that the system of
equations

x +y+z+1=0
(2 )
x yz = a

over a finite field Fq has at least one solution for any a ∈ Fq . In other words, we want
show that the cubic (affine) curve defined by the equation

x y(x + y + 1) = −a

has at least one point over Fq . In homogeneous coordinates, the corresponding projec-
tive curve has the equation

XY(X + Y + Z ) = −a Z 3 (3 )

and the singular points of this curve are the solutions of the system consisting of equa-
tion (3 ) and its partial derivatives with respect to X , Y , and Z :
⎧
⎪
⎪ XY(X + Y + Z ) = −a Z 3
⎪
⎨2XY + Y 2 + YZ = 0
. (4 )
⎪
⎪ 2XY + X 2
+ XZ = 0
⎪
⎩
XY = −3a Z 2

The difference of the second and third equations is Y 2 − X 2 + Z (Y − X ) = 0. Thus,

either X + Y + Z = 0 or X = Y .
If X + Y + Z = 0, then the first equation of (4 ) gives Z = 0. Now, the last equation
of (4 ) gives XY = 0, and, therefore all unknowns vanish (because we assume that
X + Y + Z = 0).

December 2016] BALANCED FACTORIZATIONS 993

 If X = Y , then the second equation of (4 ) shows that 3X 2 + XZ = 0. Here, if
X = 0, then Y = 0 and, therefore, Z = 0 (from the first equation of (4 )). If X = 0,
we obtain 3X + Z = 0. Then, the last equation of (4 ) gives 27a = −1. Thus, if 27a
= −1 and |F|  8, then we can apply Hasse’s estimate and conclude that a ∈ F has a
balanced decomposition into a product of four factors (one of which is 1). If 27a = −1,
we have a balanced decomposition of a:
     
1 1 1 1
− = − · − · − · 1.
27 3 3 3

(Actually, if 27a = −1, the curve is singular but the singular point itself is not a point
at infinity and, hence, gives a balanced factorization of a.)
It remains to consider small fields F with |F| < 8. In F2 and in F4 (as well as in any
finite field of characteristic two) any element is the fourth power of another element
and this gives a balanced decomposition into a product of four (equal) factors.
In F3 , the only nonzero balanced product of four factors is 1 · 1 · (−1) · (−1) and it
equals 1; therefore, −1 does not admit such decompositions.
In F5 , a product of four nonzero factors can be one of the following:

(±1) (±1) (±1) (±1) , (±1) (±1) (±1) (±2) , (±1) (±1) (±2) (±2) ,
(±1) (±2) (±2) (±2) , (±2) (±2) (±2) (±2) .

The first, third, and fifth products equal ±1, because all squares equal ±1. In the sec-
ond and fourth product, there are only two arrangements of signs making the sum of
factors zero:

1 · 1 · 1 · 2, (−1) · (−1) · (−1) · (−2),

(−1) · 2 · 2 · 2, 1 · (−2) · (−2) · (−2).

All these products equal two; therefore, −2 ∈ F5 admits no balanced decomposition

into a product of four factors.
In F7 , we find explicit balanced decompositions:

0 = 04 , 1 = 12 · (−1)2 , − 1 = 1 · 1 · 2 · 3,
2 = 2 · (−2) ,
2 2
−2 = 1 · (−2) · 3,
2
3 = (−1) · 2 · 32 , −3 = (−1)3 · 3.

Case 4: k > 4 is even. If each element a has a balanced decomposition into a product
of k factors, say a = a1 · · · ak , then each element has a balanced decomposition into
a product of (k + 2) factors, −a = a1 · · · ak · 1 · (−1) (because −a is an arbitrary
element if a is an arbitrary element). Therefore, it suffices to prove the assertion for
k = 6. Moreover, for all finite fields, except F3 and F5 , the assertion is true, because
we have constructed a balanced decomposition of each element into a product of four
factors.
In F3 , we have 0 = 06 , 1 = 16 , −1 = 13 · (−1)3 . In F5 , the balanced product x · x ·
(−2x) · 1 · 1 · (−2) equals −x 3 which is any element, because all elements are cubes.
Case 5: k > 4 is odd. The same induction as in the case of even large k makes it
possible to reduce the problem to the case k = 5. Moreover, for all finite fields, except
F2 , F4 , and F7 , the assertion is true, because we have already constructed a balanced
decomposition of each element into a product of three factors.

994 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 In F7 , the desired decomposition exists by Theorem 1. In F4 , we can write
a = b2 x yz, where b is a square root of a and x, y, z are all nonzero elements of
the field (their product is one and their sum is zero). In F2 , there are no balanced
decompositions of 1 into products of odd number of factors (because the factor cannot
be zero and the sum of an odd number of unities is not zero). This completes the
proof.

Formula (1) can be considered as a “universal formula” making it possible to factor

in a balanced manner almost any element of any field of characteristic not equal to two
into a product of five factors (where almost any means any, except a finite number of
elements). Theorem 1 shows that such a universal formula exists for each k  5. The
proof of Theorem 1 gives explicit formulae:
 
t t 2 2
t= · · (−t) · · − =
2 2 t t


5 f actor s

1−t 1−t 2 2
= · ·t · · · (−1) =

2 2 1
− t t − 1
6 f actor s
     
t t 2 2
= − · − ·t · − · · (−1) · 1 = · · · .
2 2 t t


7 f actor s

The following theorem shows that no “universal formula” for balanced decompositions
into three factors exists (a universal balanced decomposition into two factors do not
exist either for an obvious reason; the question about four factor remains open, see the
last section).

Theorem 3. For any field F, the element t of the field of rational fractions F(t) does
not admit a balanced decomposition into a product of three factors.

Proof. Assuming the contrary (and finding a common denominator), we obtain the
identity

x(t) y(t) z(t)

ts = · · , where x, y, z ∈ F[t] and x + y + z = 0.
v(t) v(t) v(t)

We have to show that s = 1, but we prefer to prove a stronger fact:

the above equalities imply that s is a multiple of 3.

The polynomials x, y, and z can be assumed to be coprime, because the equal-

ity x yz = t s v 3 shows that an irreducible common divisor of x, y, and z must either
divide v or be t; in both cases, the equation can be canceled. In addition, we may
assume that v(0) = 0 (increasing s if needed).
Let us recall the well-known Mason–Stothers theorem ([4], [7]) that can be found
in many books (see, e.g., [3]). We prefer to use the version due to Snyder, which works
in any characteristic.

December 2016] BALANCED FACTORIZATIONS 995

Mason–Stothers theorem. (In the form of Snyder [6]). If 3 polynomials x, y, z ∈ F[t]
over a field F are coprime and x + y + z = 0, then either the degrees of all these poly-
nomials are strictly less than the number of different roots of the product x yz in the
algebraic closure of F or all three derivatives x  , y  , and z  vanish (as polynomials).

In the case under consideration, x yz = t s v 3 and the number of different roots of

this polynomial is at most deg v + 1; therefore, the Mason–Stothers theorem says that
either the degree of each of x, y, z is at most the degree of v or x  = y  = z  = 0.
In the first case, deg(x yz)  3 deg v and, hence, s = 0 (because x yz = t s v 3 ) as
required. In the second case, the derivative of the product vanishes:

0 = (x yz) = (t s v 3 ) = st s−1 v 3 + 3t s v 2 v  = v 2 t s−1 (sv + 3tv  );

canceling v 2 t s−1 , we obtain sv = −3tv  . This means that s is divisible by char F,

since v(0) = 0. Therefore, either char F = 3 and s is a multiple of three as required,
or v  = 0.
If v  = 0, let us recall that an equality f  = 0 means that the polynomial f has the
form

f (x) = f 1 (x p ),

where p is the characteristic of the field, and f 1 is a polynomial. Therefore, substituting

x(t) = x1 (t p ), y(t) = y1 (t p ), z(t) = z 1 (t p ), v(t) = v1 (t p ),

to the initial identity, we obtain that s is divisible by p and, putting t p = τ , we arrive

at a similar equality for polynomials of lower degree:

x1 (τ ) y1 (τ ) z 1 (τ )
τ s/ p = · · , where x1 , y1 , z 1 ∈ F[τ ] and x1 + y1 + z 1 = 0.
v1 (τ ) v1 (τ ) v1 (τ )

An obvious induction completes the proof.

4. ALGEBRAS.

Lemma 1. Suppose that the value of a one-variable polynomial over an associative

commutative ring with unity at some point d is nilpotent and the value of the derivative
at this point is invertible. Then the polynomial has a root in this ring. Moreover, for
some root b, the difference d − b is divisible by f (d).

Proof. An obvious change of variables reduces the situation to case, where d = 0.

Suppose that the polynomial over a ring R has the form

f (x) = a0 + a1 x + · · · + an x n ,

where a1 is invertible and a0s = 0. We argue by induction on s and have to prove that
f has a root, divisible by a0 .
In the quotient ring R̄ = R/(a0s−1 R), the image f¯ of f has a root c̄a¯0 by the induc-
tion hypothesis. Take some preimage c ∈ R of the element c̄ ∈ R̄ and let us try to find

996 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
a root of f in the form b = ca0 + ta0s−1 , where t is an (unknown) element of R. Since
a0s = 0, we have

f (b) = a0 + a1 (ca0 + ta0s−1 ) + · · · + an (ca0 + ta0s−1 )n =

(5)
= f (ca0 ) + a1 ta0s−1 .

Now, ca0 is a root of f modulo the ideal a0s−1 R and, hence, f (ca0 ) ∈ a0s−1 R, i.e.,
f (ca0 ) = ra0s−1 for some r ∈ R. It remains to note that, in (5), f (b) vanishes if we
take t = −r/a1 . This completes the proof.

The following theorem reduces the question on balanced factorizations in finite-

dimensional algebras to a similar question in fields if we take into account only non-
power factorizations, i.e., factorizations having at least two nonequal factors.

Theorem 4. Let F be a field and let n be an integer larger than two. If, in all finite
extensions of F, each element has a nonpower balanced decomposition into a prod-
uct of n elements, then the same is true for each element of each finite-dimensional
associative unital algebra over F.

Proof. Clearly it suffices to prove the assertion for finite-dimensional one-generator

unital algebras (because any element of any algebra lies in a one-generated subalge-
bra). Thus, we assume that an algebra A over F has the form A = F[x]/( f ), where
f ∈ F[x]. Such algebra A decomposes into a direct sum


m
A Fi [x]/(x ki ), where fields Fi are finite extensions of F
i=1

 ki
(Fi  F[x]/( pi ) if f = pi is the decomposition of f into a product of irreducible
(over F) factors). It suffices to obtain a balanced decomposition for each direct term.
Therefore, we assume that A = G[x]/(x k ), where the field G is a finite extension of
F. Such algebra A is local, i.e., it has a unique maximal ideal I (generated by x),
A/I  G and all elements not lying in I are invertible.
We want to decompose any element a ∈ A into a product of n elements with zero
sum.
Case I. a ∈ / I . In this case we find a nonpower balanced decomposition of a modulo
ideal I , i.e., in the field G. Thus, we obtain elements a1 , . . . , an ∈ A such that

a − a1 a2 · · · an ∈ I, a1 + · · · + an ∈ I,
and (without loss of generality) a1 − an ∈
/ I.

This means that, for the quadratic polynomial

g(t) = a + ta2 a3 · · · an−1 (t + a2 + a3 + · · · + an−1 ),

(6)
we have g(a1 ) ∈ I.

For the derivative of g, we obtain g  (a1 ) = a2 a3 · · · an−1 (a1 + a2 + a3 + · · · + an−1 )

+ a1 a2 a3 · · · an−1 ∈ a2 a3 · · · an−1 (a1 − an ) + I. The ideal I consists of nilpotent ele-
ments and all elements of A \ I are invertible. Therefore, the conditions of Lemma 1

December 2016] BALANCED FACTORIZATIONS 997

are satisfied, because a1 = an (mod I ). Applying Lemma 1, we find a root 
t ∈ A of g
and obtain a decomposition:

ta2 a3 · · · an−1 (−

a = t − a2 − a3 − · · · − an−1 ) with zero sum of factors. (7)

This decomposition is nonpower, because 

t ≡ a1 (mod I ) by Lemma 1 and a1 ≡ an
(mod I ) by the assumption.
Case II. a ∈ I . Let us choose invertible (i.e., not lying in I ) elements a2 , . . . , an−1 ∈ A
in such a way that their sum is also invertible. This is possible if the field G = A/I
has more than two elements. If the field G is two-element, then the unit element has in
G no nonpower decomposition that contradicts the condition.
For the polynomial g(t) (see formula (6)) we obtain that g(0) = a is a nilpotent
element and

g  (0) = a2 a3 · · · an−1 (a2 + a3 + · · · + an−1 ) is an invertible element.

Therefore, by Lemma 1, g has a root  t ∈ A as required (see (7)). Decomposition (7)

cannot be power, because a2 is invertible but a is not. This completes the proof.

Corollary 1. Each element of a finite-dimensional associative unital algebra (over a

field) decomposes into a product of
a) three elements whose sum vanishes if the field is algebraically closed;
b) five elements whose sum vanishes if the characteristic of the field is not two.

Proof. The first assertion follows immediately from Theorem 4, because, in an alge-
braically closed field, each element has a nonpower balanced decomposition into a
product of three factors (to obtain a nonpower balanced decomposition a = a1 a2 a3 of
a given element a, we can choose any element a1 such that a13 = a and then a2 and a3
can be found from a quadratic equation).
To prove the second assertion, it suffices to apply Theorems 4 and 1 and note that
formula (1) always gives a nonpower decomposition.

Corollary 2. For any k  3, any complex or real matrix can be decomposed into a
product of k matrices (over the same field) whose sum vanishes.

Proof. The assertion follows immediately from Theorem 4, because each real or com-
plex number a admits a nonpower balanced decomposition

a = x · (x + 1) · 1k−3 · (2 − k − 2x),

as this equality is a cubic equation with respect to x.

Now, we give examples showing that no conditions of Theorem 4 and its corollaries
can be omitted.

Example 1. Each element of the tree-element field F3 has a balanced decompo-

sition into a product of three factors: 0 = 0 · 0 · 0, 1 = 1 · 1 · 1, 2 = 2 · 2 · 2.
However, in the two-dimensional algebra A = F3 [x]/(x 2 ) over this field, the ele-
ment 1 + x does not admit balanced decompositions into a product of three fac-
tors, because the decomposition 1 = 1 · 1 · 1 is the unique balanced decomposi-
tion of 1 in F3 ; therefore, the balanced decomposition of 1 + x ∈ A must have the

998 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
form 1 + x = (1 + kx)(1 + lx)(1 + mx) (where k, l, m ∈ F3 ), whence we obtain
k + l + m = 1 and the decomposition is not balanced. This example shows that
Theorem 4 become false if we omit the words nonpower.

Example 2. In the algebra of polynomials F[x] over any field, the element x has
no balanced decompositions. This example shows that finite-dimensionality condition
cannot be omitted in Theorem 4 and Corollary 1.

In algebras with zero multiplication, no nonzero element has balanced decomposi-

tions. This shows that the condition that the algebra has a unit also cannot be omitted
in Theorem 4 and Corollary 1.
The condition n > 2 can be omitted in Theorem 4, because this condition follows
from other conditions: in any field, any balanced decomposition of zero into a product
of two factors must be power. On the other hand, in any nonzero ring, zero has non-
power decompositions into products of three and any larger numbers of factors, e.g.,
0 = 02016 · b · (−b), where b is a nonzero element. However, there is the following
simple example.

Example 3. In the field of complex numbers, any nonzero element has a nonpower
balanced decomposition into a product of two factors, but the nilpotent Jordan block
obviously has no balanced decomposition into a product of two factors (for any field),
because such a decomposition of J would mean that −J is a square, but it is not.

Example 3 also shows that, in Corollary 2, we cannot omit the condition k > 2 and,
in Corollary 1(a), it is impossible to replace three with two. The following example
shows that, in Corollary 1(b), we cannot replace five with a lower number.

Example 4. As mentioned above (see Example 1), in the two-dimensional algebra

A = F3 [x]/(x 2 ), the element 1 + x does not admit a balanced decomposition into three
factors. In the same algebra (as well as in the field F3 ), minus one admits no balanced
decomposition into a product of four factors and 1 has no balanced decompositions
into products of two factors.

Example 5. In the field F2 , the identity element does not admit balanced decompo-
sitions into products of five factors. This simple example shows that the condition on
characteristic cannot be omitted in Corollary 1(b) (and in Theorem 1).

5. OPEN QUESTIONS.
Question 1 (A. V. Ivanishchuk [1]). Can any rational number be decomposed into a
product of four rational numbers whose sum vanishes?2
Question 2. Can any element of any field be decomposed into a product of at most
four factors whose sum vanishes?
Question 3. Does there exist a universal formula for balanced decomposition into four
factors? More precisely, does the element t of the field of rational fractions C(t) (or
even Q(t)) admit a balanced decomposition into a product of four factors?2
Question 4. What does occur in characteristic 2? Does there exist a universal formula?
Does any element of any field admit a balanced factorization?
2 When this paper was written, we learned that the answers to Questions 1 and 3 are positive [2].

December 2016] BALANCED FACTORIZATIONS 999

ACKNOWLEDGMENT. The authors thank Yu. G. Prokhorov, M. A. Tsfasman, and also (undergraduate stu-
dents) Evgenia Kosheleva, Alisa Pikulina, Nadira Shoketaeva, and (a high-school student) Rauan Zhakypbek
for useful discussions. We are also grateful to anonymous referees for valuable remarks that allowed us to
improve the paper.
The work of the first author was supported by the Russian Foundation for Basic Research, project
no. 15-01-05823. The work of the second author was supported by Science committee of Ministry of Educa-
tion and Science of Republic of Kazakhstan, project no. 4-0816.

REFERENCES

1. A. V. Ivanishchuk, The experience of learning and research activity of students in Lyceum 1511 (MEPhI),
a chapter in the book Research Problems for Beginners by A. I. Sgibnev, MCCME, Moscow, 2013. 21–25,
http://www.mccme.ru/free-books/ (in Russian).
2. A. A. Klyachko, A. M. Mazhuga, A. N. Ponfilenko, Balanced factorisations in some algebras (2016),
http://arxiv.org/abs/1607.01957.
3. S. Lang, Algebra. Graduate Texts in Mathematics, Vol. 211. Springer-Verlag, New York, 2002.
4. R. C. Mason, Diophantine Equations over Function Fields. London Mathematical Society Lecture Note
Series, Vol. 96. Cambridge Univ. Press, Cambridge, 1984.
5. J. H. Silverman, The Arithmetic of Elliptic Curves. Springer-Verlag, New York, 1986.
6. N. Snyder, An alternate proof of Mason’s theorem, Elem. Math. 55 no.3 (2000) 93–94.
7. W. W. Stothers, Polynomial identities and hauptmoduln, Quart. J. Math. 32 no. 3 (1981) 349–370.
8. A. N. Vassilyev, Problem 4 (Final stage, 9th grade), in Kazakhstan republican olympiad in mathematics
(2013), http://matol.kz/olympiads/151 (in Russian).
9. , Problem 3, in IX Algebra olympiad for students in MSU (2014),
http://halgebra.math.msu.su/Olympiad/ (in Russian).

ANTON ALEKSANDROVICH KLYACHKO works at Moscow State University. His main interests lie in
Group Theory although he is probably wider known for his Car-crash lemma (aka Ants-on-a-ball puzzle).
Faculty of Mechanics and Mathematics of Moscow State University, Moscow 119991, Leninskie gory, MSU
klyachko@mech.math.msu.su

ANTON NIKOLAEVICH VASSILYEV received his M.Sc. in 2009 and his Ph.D. in 2014 from Moscow
State University. Currently he teaches students at the Kazakhstan branch of Moscow State University in Astana.
He attended the International Mathematical Olympiad in 2011–2016 as the Deputy Leader of the Kazakhstan
national team.
Department of Mathematics and Computer Science, Kazakhstan branch of Moscow State University, Astana
010010, ul. Kazhimukana 11, MSU
antonvassilyev@mail.ru

1000 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 Cantor Polynomials and the Fueter-Pólya
Theorem
Melvyn B. Nathanson

Abstract. A packing polynomial is a polynomial that maps the set N20 of lattice points with
nonnegative coordinates bijectively onto N0 . Cantor constructed two quadratic packing poly-
nomials, and Fueter and Pólya proved analytically that the Cantor polynomials are the only
quadratic packing polynomials. The purpose of this paper is to present a beautiful elementary
proof of Vsemirnov of the Fueter–Pólya theorem. It is a century-old conjecture that the Cantor
polynomials are the only packing polynomials on N20 .

1. STORING AND PACKING LATTICE POINTS. Let Zm be the group of lattice

points in Rm . Let N0 = {0, 1, 2, . . .} be the set of nonnegative integers, and let
Nm0 = {(x1 , . . . , xm ) ∈ Zm : xi ∈ N0 for i = 1, . . . , m}
be the additive semigroup of m-dimensional nonnegative lattice points.
In computer science, there is the problem of saving and retrieving matrices and
other multidimensional data in linear memory. We can think of computer memory as
an infinite sequence of “storage boxes” numbered 0, 1, 2, 3, . . .. Suppose that we want
to store the matrix (ai, j ) in memory. A simple procedure would be to choose an easily
computed one-to-one function F : N20 → N0 and store the matrix element ai, j in the
memory box with the number F(i, j). More generally, for any subset S of Zm , a one-
to-one function from S into N0 is called a storing function on S. A function that maps S
bijectively onto N0 is called a packing function on S. Computer scientists are interested
in finding polynomial or other “elementary” storing and packing functions on S.
For example, let m = 2 and consider the function C1 : N20 → N0 that enumerates,
from lower right to upper left, the nonnegative lattice points on the consecutive par-
allel line segments x + y = k for k = 0, 1, 2, . . .. Thus, C1 (0, 0) = 0, C1 (1, 0) = 1,
C1 (0, 1) = 2, C1 (2, 0) = 3, C1 (1, 1) = 4, C1 (0, 2) = 5, C1 (3, 0) = 6, . . . . For k ∈ N0 ,
the number of nonnegative lattice points (x, y) with x + y < k is 1 + 2 + · · · + k =
k(k + 1)/2. It follows that C1 (k, 0) = k(k + 1)/2. If (x, y) ∈ N20 and x + y = k, then
C1 (x, y) = C1 (k, 0) + y = C1 (x + y, 0) + y
(x + y)(x + y + 1)
= +y
2
1 1
= (x + y)2 + (x + 3y) .
2 2
The function C2 : N20 → N0 that enumerates, from upper left to lower right, the non-
negative lattice points on the consecutive parallel line segments x + y = k is
1 1
C2 (x, y) = (x + y)2 + (3x + y) = C1 (y, x).
2 2
The functions C1 (x, y) and C2 (x, y) are called the Cantor packing polynomials.
http://dx.doi.org/10.4169/amer.math.monthly.123.10.1001
MSC: Primary 11C08, Secondary 11L99; 11A15

December 2016] 1001

 The classical applications of polynomial packing functions are due to Cauchy and
Cantor. Cantor [2, page 494], [3, page 107] used the quadratic polynomials C1 (x, y)
and C2 (x, y) to prove that the set N20 is countable. Much earlier, to reduce double sums
to single sums, Cauchy [4, Part 1, Chapter 6, Theorem 6] used the identity


∞ 
∞ 
∞
ai b j = ck
i=0 j=0 k=0

where


k
ck = ai b j = ak b0 + ak−1 b1 + ak−2 b2 + · · · + a0 bk .
i+ j=k

This is exactly the enumeration of the nonnegative lattice points that Cantor subse-
quently applied to set theory.
The following simple argument implies that a packing function cannot be linear.

Lemma 1. Let , m ∈ N0 with m ≥ 2. For x = (x1 , . . . , xm ) ∈ Nm0 , let min(x)

= min(xi : i = 1, . . . , m), and let S be the set of all lattice points x ∈ Nm0 such
that min(x) ≥ . No linear polynomial is a storing function on S.

Proof. For x = (x1 , . . . , xm ) ∈ Rm , consider the linear polynomial

F(x) = F(x1 , x2 , . . . , xm ) = a1 x1 + a2 x2 + · · · + am xm + c.

Let A = max(|ai | : i = 1, . . . , m). For i = 1, . . . , m, let ei ∈ Nm0 be the standard basis

vector whose ith coordinate is 1 and whose other coordinates are 0.
Suppose that F is a storing function on S, that is, F(S) ⊆ N0 and F is one-to-one
on S. If x ∈ S, then x + ei ∈ S and

F(x + ei ) − F(x) = ai ∈ Z
m m
for all i = 1, . . . , m. It follows that i=1 ai xi ∈ Z, and so c = F(x) − i=1 ai xi ∈ Z.
Suppose that F : S → N0 is a storing function. Choose x ∈ S with min(x) ≥  + A.
For all i, j ∈ {1, 2, . . . , m} with i = j, we have x + ai e j − a j ei ∈ S and

F(x + ai e j − a j ei ) = F(x) + ai a j − a j ai = F(x).

Because F is one-to-one on S, it follows that x = x + ai e j − a j ei , and so ai e j = a j ei .

This implies that ai = a j = 0, and so ai = 0 for all i ∈ {1, . . . , m}. Thus, the storing
function F(x) = c is constant, which is absurd. This means that no linear polynomial
can be a storing function on S.

Conjecture. The Cantor packing polynomials C1 (x, y) and C2 (x, y) are the only
polynomial bijections from N20 to N0 .

This conjecture is nearly 100 years old. In 1923, Fueter and Pólya obtained the first
result about the uniqueness of packing polynomials in two variables.

Theorem (Fueter-Pólya [6]). Every quadratic packing polynomial on N20 is a Cantor

polynomial.

1002 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 In 1978, Lew and Rosenberg [8, 9] proved that no polynomial packing function on
N20 has degree three or four. It is not known if there exists a packing polynomial on N20
of degree greater than four.
There exist many packing polynomials for lattice points of dimension m ≥ 3.
The simplest are constructed by composing Cantor polynomials. For example, the
function (x, y, z) → (C1 (x, y), z) is a bijection from N30 to N20 , and so (x, y, z) →
C1 (C1 (x, y), z) is a packing polynomial on N30 . There also exist packing polynomials
on Nm0 that are not compositions of packing polynomials in lower dimensions. Much
work has been done on this problem, e.g., [5, 7, 10, 12, 13, 14, 16, 17].
Fueter and Pólya used methods from analytic number theory to prove their theorem.
In 2001, Vsemirnov [19, 20] gave a beautiful proof of the Fueter–Pólya theorem that
uses only Gauss’s law of quadratic reciprocity and Dirichlet’s theorem on primes in
arithmetic progressions. The purpose of this paper is to present Vsemirnov’s proof.

2. ELEMENTARY PROOF OF THE FUETER-PÓLYA THEOREM. We start

with an exercise in elementary number theory.

Lemma 2. If D and  are nonzero integers and D is not a square, then there exists a
prime p such that D is a quadratic nonresidue modulo p and p does not divide .

Proof. We can write


k
α β
D = (−1) 2 m 2
qi
i=1

where α, β ∈ {0, 1}, m ∈ N, and q1 , . . . , qk are distinct odd primes.

Suppose that k = 0 and D = (−1)α 2β m 2 . If β = 0, then α = 1 and D = −m 2 .
Choosing a prime p ≡ 3 (mod 4), we obtain
   
D −1
= = −1.
p p

If β = 1, then D = ±2m 2 . Choosing a prime p ≡ 5 (mod 8), we obtain

   
D ±2
= = −1.
p p

By Dirichlet’s theorem on primes in arithmetic progressions, each of these congruence

classes contains infinitely many primes. Choosing p > , we obtain a prime p that
does not divide .
Suppose that k ≥ 1. For every prime number such that p ≡ 1 (mod 8), we have
   
−1 2
= = 1.
p p

Applying the multiplicativity of the Legendre symbol and quadratic reciprocity,

we obtain
  k 
  k  
D qi p
= = .
p i=1
p i=1
qi

December 2016] 1003

Choose integers r1 , . . . , rk such that
 
r1
= −1
q1

and
 
ri
=1 for i = 2, . . . , k.
qi
By the Chinese remainder theorem, there is an integer s such that
s≡1 (mod 8)
and
s ≡ ri (mod qi ) for i = 1, 2, . . . , k.
Moreover,
 

k
s, 8 qi = 1.
i=1

If p is a prime number such that


k
p≡s (mod 8 qi ), (1)
i=1

then
  k    k    k    
D p s ri r1
= = = = = −1.
p i=1
qi i=1
qi i=1
qi q1

By Dirichlet’s theorem on primes in arithmetic progressions, there are infinitely many

primes p that satisfy the congruence (1). Choosing a prime p in this arithmetic
progression such that p >  completes the proof.

We can now begin the proof of the Fueter–Pólya theorem. There are four lemmas.

Lemma 3. If F(x, y) is a quadratic packing polynomial N20 , then there exist nonneg-
ative integers a, c, f and integers b, d, e such that

a≡d (mod 2)

c≡e (mod 2)

and
1 2
1
F(x, y) = ax + 2bx y + cy 2 + (d x + ey) + f. (2)
2 2
Moreover, if a = c = 0, then b ≥ 1.

1004 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
Proof. Every quadratic polynomial F(x, y) with complex coefficients can be written
in the form (2). Because F is a function from N20 to N0 , we have

a = F(2, 0) − 2F(1, 0) + F(0, 0) ∈ Z

c = F(0, 2) − 2F(0, 1) + F(0, 0) ∈ Z

and

f = F(0, 0) ∈ N0 .

For all x, y ∈ N0 , the inequalities

a 2 d
F(x, 0) = x + x+ f ≥0
2 2
and
c 2 e
F(0, y) = y + y+ f ≥0
2 2
imply that a and c are nonnegative integers. The identities

a+d
F(1, 0) − F(0, 0) = ∈Z
2

c+e
F(0, 1) − F(0, 0) = ∈Z
2
imply that d and e are integers such that

a≡d (mod 2) and c≡e (mod 2).

It follows that
a+d c+e
F(1, 1) = b + + + f ∈ N0 ,
2 2
and so b is an integer.
If a = c = 0, then b = 0 because F is quadratic. For all x ∈ N0 ,

d +e
F(x, x) = bx 2 + x + f ≥ 0,
2
and so b ≥ 1. This completes the proof.

Lemma 4. If F(x, y) is a quadratic packing polynomial of the form (2), then the
quadratic form

1 2

Q(x, y) = ax + 2bx y + cy 2
2

is positive-definite on N20 . Moreover, a ≥ 1 and c ≥ 1.

December 2016] 1005

Proof. The quadratic form Q(x, y) is nonzero because the polynomial F(x, y) is
quadratic. Defining the linear form

1
L(x, y) = (d x + ey)
2
we can write

F(x, y) = Q(x, y) + L(x, y) + f

If a ≥ 1 and r > |d|, then Q(r, 0) > |L(r, 0)|. If c ≥ 1 and s > |e|, then Q(0, s)
> |L(0, s)|. If a = c = 0 and r > |d + e|/2, then b ≥ 1 by Lemma 3, and

|d + e|r
Q(r, r ) = br 2 ≥ r 2 > = |L(r, r )|.
2

Thus, there exists (r, s) ∈ N20 \ {(0, 0)} such that

Q(r, s) > |L(r, s)|. (3)

It is easy to show that Q is nonnegative-definite. For all (x, y) ∈ N20 and t ∈ N0 ,

F(xt, yt) = Q(xt, yt) + L(xt, yt) + f = Q(x, y)t 2 + L(x, y)t + f.

If Q(x, y) < 0 for some (x, y) ∈ N20 , then F(xt, yt) < 0 for all sufficiently large t,
which is absurd. Therefore, Q(x, y) ≥ 0 for all (x, y) ∈ N20 .
Suppose that Q(u, v) = 0 for some (u, v) ∈ N20 \ {(0, 0)}. For all t ∈ N0 , we have

F(ut, vt) = Q(u, v)t 2 + L(u, v)t + f = L(u, v)t + f = wt + f

where w = L(u, v) is a positive integer because F(ut, vt) is nonconstant and nonneg-
ative for t ∈ N0 . Choosing (r, s) ∈ N20 that satisfies inequality (3), we have

F(r w, sw) = Q(r, s)w 2 + L(r, s)w + f = wm + f

where

m = Q(r, s)w + L(r, s) ≥ Q(r, s) − |L(r, s)| > 0.

Because F is one-to-one on N20 and

F(um, vm) = wm + f = F(r w, sw),

it follows that

(um, vm) = (r w, sw),

and so

0 = Q(u, v)m 2 = Q(um, vm) = Q(r w, sw) = Q(r, s)w 2 > 0,

1006 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
which is absurd. Therefore, the quadratic form Q(x, y) is positive-definite on N20 . This
implies that
a = 2Q(1, 0) ≥ 1
and
c = 2Q(0, 1) ≥ 1.
This completes the proof.

Lemma 5. If F(x, y) is a quadratic packing polynomial of the form (2), then b ≤ 1.

Proof. Let m ≥ max(2, |d|, |e|) and (x, y) ∈ N20 . If b ≥ 2, then

1 2
1
F(x, y) = ax + 2bx y + cy 2 + (d x + ey) + f
2 2
1 2
1
≥ x + 4x y + y 2 − (|d|x + |e|y)
2 2
1 m
≥ (x + y)2 + x y − (x + y) .
2 2
We obtain the following inequalities:
1
F(x, y) ≥ (x + y)(x + y − m) + x y (4)
2
and, for every positive integer k,
1 k−1 1 1
F(x, y) ≥ (x + y)2 + x y + (x − km)y + (y − km)x. (5)
2 k 2k 2k
There are three cases.
Case 1. If max(x, y) ≥ 25m, then x + y ≥ 25m. Applying inequality (4), we obtain
1 1
F(x, y) ≥ (x + y)(x + y − m) ≥ (25m)(24m) = 300m 2 .
2 2
Case 2. If min(x, y) ≥ 10m, then x + y ≥ 20m. Applying inequality (5) with k = 10,
we obtain
1 9
F(x, y) ≥ (20m)2 + (10m)2 = 290m 2 .
2 10
Case 3. If min(x, y) ≥ m and x + y ≥ 24m, then, applying inequality (5) with k = 1,
we obtain
1
F(x, y) ≥ (24m)2 = 288m 2 .
2
It follows that if (x, y) ∈ N20 and F(x, y) < 288m 2 , then

max(x, y) < 25m

min(x, y) < 10m

December 2016] 1007

and

min(x, y) < m or x + y < 24m.

Equivalently, the lattice point (x, y) must belong to exactly one of the following five
sets:
 
Z 1 = (x, y) ∈ N20 : 0 ≤ x < m and 0 ≤ y < 25m
 
Z 2 = (x, y) ∈ N20 : m ≤ x < 10m and 0 ≤ y < 24m − x
 
Z 3 = (x, y) ∈ N20 : 10m ≤ x < 14m and 0 ≤ y < 10m
 
Z 4 = (x, y) ∈ N20 : 14m ≤ x < 23m and 0 ≤ y < 24m − x
 
Z 5 = (x, y) ∈ N20 : 23m ≤ x < 25m and 0 ≤ y < m .

For i = 1, . . . , 5, let Ni denote the number of lattice points in the set Z i . We have


m−1
N1 = 25m = 25m 2
x=0

and


10m−1
333 2 9
N2 = (24m − x) = m + m.
x=m
2 2

Similarly,

N3 = 40m 2
99 2 9
N4 = m + m
2 2
N5 = 2m 2 .

Therefore, the number of nonnegative integers n < 288m 2 represented by the polyno-
mial F(x, y) with (x, y) ∈ N20 is at most


5
Ni = 283m 2 + 9m < 288m 2
i=1

because m ≥ 2. It follows that F : N20 → N0 is not surjective, which is absurd. There-

fore, b ≤ 1. This completes the proof.

Lemma 6. If F(x, y) is a quadratic packing polynomial of the form (2), and if

D = b2 − ac

then

8a D F(x, y) = Du 2 − v 2 + r (6)

1008 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
where

u = 2ax + 2by + d (7)

v = 2Dy + bd − ae (8)

and

r = (bd − ae)2 − Dd 2 + 8a D f. (9)

Moreover, D is a perfect square.

Proof. We begin with an exercise in “completing the square.” We have

8a D F(x, y) − 8a D f = 4D(a 2 x 2 + 2abx y + acy 2 + ad x + aey)

= 4D (ax + by)2 − Dy 2 + ad x + aey

= 4D (ax + by)2 + d(ax + by) − Dy 2 − (bd − ae)y

= D 4(ax + by)2 + 4d(ax + by) − 4D 2 y 2 + 4D(bd − ae)y

= D (2ax + 2by + d)2 − (2Dy + bd − ae)2 + (bd − ae)2 − Dd 2
= Du 2 − v 2 + r − 8a D f.

This proves (6) – (9).

Suppose that D = b2 − ac is a not a square. Then D = 0. If x and y are integers,
then there are integers u = u(x, y), v = v(y), and r such that

8a D F(x, y) = Du 2 − v 2 + r.

Because 8a = 0 and D is not a square, we apply Lemma 2 with  = 8a and obtain a

prime number p such that D is a quadratic nonresidue modulo p and (8a D, p) = 1.
There is an integer s such that

8a Ds ≡ r (mod p2 ).

Because F(x, y) is a packing polynomial, there exist infinitely many lattice points
(x, y) ∈ N20 such that F(x, y) ≡ s (mod p) and

Du 2 − v 2 + r = 8a D F(x, y) ≡ 8a Ds ≡ r (mod p).

Therefore,

Du 2 ≡ v 2 (mod p).

If u ≡ 0 (mod p), then

2
D ≡ vu −1 (mod p)

which is impossible because D is a quadratic nonresidue modulo p. It follows that

u ≡ 0 (mod p), and so v ≡ 0 (mod p) and

Du 2 − v 2 ≡ 0 (mod p2 ).

December 2016] 1009

Thus, if (x, y) ∈ N20 and F(x, y) ≡ s (mod p), then

8a D F(x, y) = Du 2 − v 2 + r ≡ r ≡ 8a Ds (mod p2 ).

Because (8a D, p) = 1, we see that the congruence F(x, y) ≡ s (mod p) implies that
F(x, y) ≡ s (mod p 2 ). This means that there do not exist integers x and y such that
F(x, y) ≡ s + p (mod p 2 ), and so the polynomial F(x, y) is not surjective from N20
onto N0 , which is absurd. Therefore, D is a square.

We can now complete Vsemirnov’s proof of the Fueter–Pólya theorem.

Proof. By Lemma 6, D = b2 − ac is a square; hence, D = t 2 for some nonnegative

integer t. We have

Q(t − b, a) = a(t − b)2 + 2b(t − b)a + ca 2

2
a 2 a
= (t − b2 + ac) = (t 2 − D) = 0.
2 2
Recall that a and c are positive integers. By Lemma 4, the quadratic form Q(x, y) is
positive-definite for (x, y) ∈ N20 , and so t − b < 0. By Lemma 5, we have 0 ≤ t <
b ≤ 1, and so b = 1. Therefore, 1 − ac = b2 − ac = D = t 2 ≥ 0 and 0 < ac ≤ 1.
This implies that a = c = 1, and so

(x + y)2 1
F(x, y) = + (d x + ey) + f.
2 2
Moreover, d ≡ a ≡ 1 (mod 2) and e ≡ c ≡ 1 (mod 2), that is, d and e are odd inte-
gers.
If d = e, then F(x, y) = F(y, x) for all (x, y) ∈ N20 and F(x, y) is not one-to-one.
Therefore, d = e. If d > e, then d − e = 2g for some positive integer g, and

(x + y)(x + y + e)
F(x, y) = + gx + f.
2
We have e = 0 because e is odd, and

F(0, −e) = F(0, 0) = f.

Because F(x, y) is a storing function, we conclude that e ≥ 1. Therefore, F(x, y) ≥ f

for all (x, y) ∈ N20 , and so f = 0.
If e ≥ 3, then for all (x, y) ∈ N20 \ {(0, 0)} we have x + y ≥ 1 and F(x, y) ≥
(1 + e)/2 ≥ 2. This means that F(x, y) = 1 for all (x, y) ∈ N20 , which is absurd.
Therefore, e = 1 and

(x + y)(x + y + 1)
F(x, y) = + gx
2
for some positive integer g. We have F(0, 1) = 1, F(1, 0) = 1 + g, and F(x, y) ≥
3 for all (x, y) ∈ N20 with x + y ≥ 2. If g ≥ 2, then F(x, y) = 2 for all (x, y) ∈
N20 , which is absurd. Therefore, g = 1 and F(x, y) = C1 (x, y) is the first Cantor
polynomial.

1010 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 Similarly, if e > d, then F(x, y) = C2 (x, y) is the second Cantor polynomial. This
completes the proof of the Fueter–Pólya theorem.

3. PACKING POLYNOMIALS IN SECTORS. For every positive real number α,

we construct the real sector

S(α) = {(x, y) ∈ R2 : 0 ≤ y ≤ αx}

and the integer sector

I (α) = S(α) ∩ N20 = {(x, y) ∈ N20 : 0 ≤ y ≤ αx}.

The real sector S(α) is the cone with vertex at (0, 0) generated by the points (1, 0) and
(1, α). A sector is called rational if α is a rational number and irrational if α is an
irrational number. Recent work on Cantor polynomials has concentrated on packing
polynomials in rational sectors.

Theorem (Nathanson [15]). Let r and s be relatively prime positive integers such
that 1 ≤ r < s and r divides s − 1. Let d = (s − 1)/r . The polynomials

r (x − dy)2 (2 − r )x + (dr − 2d + 2)y

Fr/s (x, y) = +
2 2
and

r (x − dy)2 (r + 2)x − (2d + s + 1)y

G r/s (x, y) = +
2 2
are quadratic packing polynomials for the integer sector I (r/s). Moreover, for s ≥ 2,
the polynomials

(x − (s − 1)y)2 x + (3 − s)y
F1/s (x, y) = +
2 2
and

(x − (s − 1)y)2 3x + (1 − 3s)y
G 1/s (x, y) = +
2 2
are the unique quadratic packing polynomials for the integer sector I (1/s).

Recent work by Stanton [18] and Brandt [1] has determined explicitly the quadratic
packing polynomials for all rational sectors. Morales [11] has generalized Nathanson’s
theorem to packing polynomials on multidimensional integer sectors.
It is an open problem to understand storing and packing polynomials on irrational
sectors.

REFERENCES

1. M. Brandt, Quadratic packing polynomials on sectors of R2 , arXiv:1409.0063, 2014.

2. G. Cantor, Beiträge zur Begründung der transfiniten Mengenlehre, Math. Annalen 46 (1895) 481–512.
3. , Contributions to the Founding of the Theory of Transfinite Numbers. Dover Publications, New
York, 1952, Translated, and provided with an introduction and notes, by Philip E. B. Jourdain.

December 2016] 1011

 4. A.-L. Cauchy, Cours d’analyse de l’École Royale Polytechnique. Cambridge Library Collection,
Cambridge Univ. Press, Cambridge, UK, 2009, Reprint of the 1821 original, Previously published by
Cooperativa Libraria Universitaria Editrice Bologna, Italy, 1992.
5. H. L. Fetter, J. H. Arredondo R., L. B. Morales, The diagonal polynomials of dimension four, Adv. Appl.
Math. 34 no. 2 (2005) 316–334.
6. R. Fueter, G. Pólya, Rationale Abzählung der Gitterpunkte. Vierteljschr Naturforsch. Gesellsch. Zurich
58 (1923) 380–386.
7. J. S. Lew, Polynomials in two variables taking distinct integer values at lattice-points, Amer. Math.
Monthly 88 (1981) 344–346.
8. J. S. Lew, A. L. Rosenberg, Polynomial indexing of integer lattice-points. I. General concepts and
quadratic polynomials, J. Number Theory 10 no. 2 (1978) 192–214.
9. , Polynomial indexing of integer lattice-points. II. Nonexistence results for higher-degree poly-
nomials, J. Number Theory 10 no. 2 (1978) 215–243.
10. L. B. Morales, Diagonal polynomials and diagonal orders on multidimensional lattices, Theory Comput.
Syst. 30 no. 4 (1997) 367–382.
11. , Packing polynomials on multidimensional integer sectors, Electronic J. Combinatorics 23 no. 4
(2016).
12. L. B. Morales, J. S. Lew, An enlarged family of packing polynomials on multidimensional lattices, Math.
Systems Theory 29 no. 3 (1996) 293–303.
13. A. Sánchez-Flores, A family of (n − 1)! diagonal polynomial orders of Nn , Order 12 no. 2 (1995)
173–187.
14. L. B. Morales, A. Sánchez-Flores, Erratum: “Diagonal polynomials and diagonal orders on multidimen-
sional lattices,” [Theory Comput. Syst. 30 no. 4 (1997) 367–382], Theory Comput. Syst. 33 no. 1 (2000)
107.
15. M. B. Nathanson, Cantor polynomials for semigroup sectors, J. Algebra Appl. 13 no. 5 (2014) 1350165,
14.
16. A. L. Rosenberg, Data graphs and addressing schemes, J. Comput. System Sci. 5 (1971) 193–238.
17. , Storage Mmappings for Extendible Arrays. Current Trends in Programming Methodology,
Vol. IV, Ed. by R. T. Yeh. Prentice-Hall, Englewood Cliffs, NJ, 1978, 263–311.
18. C. Stanton, Packing polynomials on sectors of R2 , Integers 14 (2014) Paper No. A67, 13.
19. M. A. Vsemirnov, Two elementary proofs of the Fueter-Pólya theorem on matching polynomials, Algebra
i Analiz 13 no. 5 (2001) 1–15.
20. , Errata: “Two elementary proofs of the Fueter–Pólya theorem on matching polynomials,”
(Russian) [.Algebra i Analiz 13 no. 5(2001) 1–15; MR1882861 (2003a:11021)], Algebra i Analiz 14
no. 5 (2002) 240.

MEL NATHANSON is a professor of mathematics at the City University of New York (Lehman College and
the CUNY Graduate Center).
Department of Mathematics, Lehman College (CUNY), Bronx, NY 10468 and CUNY Graduate Center, New
York, NY 10016
melvyn.nathanson@lehman.cuny.edu

1012 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 Optimizing the Video Game Multi-Jump:
Player Strategy, AI, and Level Design
Aaron M. Broussard, Martin E. Malandro, and Abagayle Serreyn

Abstract. This article initiates the mathematical study of multi-jumping in video games.
We begin by proving a necessary, and frequently sufficient, condition for a multi-jump to
be optimal, i.e., achieve the highest possible height after traveling a given horizontal distance.
We then give strategies that can be used by human players and by AI to select successful multi-
jumps in real time. We also show how a video game designer can build the ground around a
platform to guarantee that the platform is reachable—or unreachable—by a multi-jump begin-
ning at any point on the ground.

1. INTRODUCTION. In many platform-based video games the player is able to per-

form a double jump, which is a normal jump followed by a second jump initiated in
midair without the aid of a platform. The arc of the second jump might be identical
to that of the first, it might be a smaller version of the first and otherwise be subject
to the same gravitational pull, or it might have its own arc and obey a completely dif-
ferent gravitational law. For example, Capcom R
’s Devil May CryTM1 features the first
 
kind of double jump, Konami ’s Castlevania : Symphony of the NightTM features
R R

the second, and Klei Entertainment R

’s Mark of the NinjaTM features the third. In this
paper we study multi-jumps, which generalize double jumps. A multi-jump is a finite
sequence of jumps where the first jump is initiated from the ground and the rest are ini-
tiated in midair. The number of jumps in a multi-jump is the length of the multi-jump,
so a double jump is a multi-jump of length two. Several video games, such as Chair
Entertainment Group R
’s Shadow ComplexTM and Nintendo R
’s Super Smash Bros.TM
Melee, feature triple jumps or multi-jumps of even longer length.
The basic problem we consider in this paper is the following. Suppose that a char-
acter in a two-dimensional side-scrolling video game wishes to use a multi-jump to
jump to the right from a fixed starting point across a gap and land on a fixed platform.
By a platform we always mean an impenetrable horizontal platform (so the charac-
ter cannot pass through the bottom of the platform) that begins at a point and extends
indefinitely to the right. In most games the character can control the horizontal velocity
component of her jumps. We assume that the target platform begins far enough to the
right from the character that she will utilize the maximum horizontal velocity possible
for each jump. We therefore assume that the character has a known finite sequence of
jump arcs available to her and faces the problem of selecting when to jump in midair,
i.e., to switch from the arc of one jump to the next, so as to land on the platform. See
Figures 1–2, in which the character can jump twice in midair.
Provided the platform is reachable by a multi-jump, we give strategies for solv-
ing this problem on the fly for both player-controlled and artificial intelligence (AI)-
controlled characters. In the simplest situation all jumps available to the character are
equal and fully concave (Definition 5). In this situation we give a simple strategy (the
line method) that is usable by both players and AI. In our experience the majority of
http://dx.doi.org/10.4169/amer.math.monthly.123.10.1013
MSC: Primary 90C27, Secondary 91A55; 97A20
1 All products, company names, brand names, and trademarks are properties of their respective owners.

December 2016] OPTIMIZING MULTI-JUMPS 1013

 Figure 1. Character wishes to jump to distant platform

(A) An unsuccessful multi-jump (B) A successful multi-jump

Figure 2. Multi-jumps

games featuring multi-jumps are covered by this situation. We give two further strate-
gies for AI-controlled characters in more-complicated situations. Our first AI strategy
is very general, in that it applies to any collection of standard jump functions (Defini-
tion 1). We also give a faster (less computationally intensive) AI strategy for collec-
tions of standard jump functions whose derivative inverses are known and computable
exactly.
We have evidence that our AI results are new, or at least previously unknown to
game developers: We have observed that the multi-jumping AI in Super Smash Bros.TM
Melee is not optimal, in that there are situations in the game where the AI will con-
sistently select a multi-jump that fails to cross a gap even though such a multi-jump is
possible. The more recent games in Nintendo R
’s Super Smash Bros.TM series (Super
Smash Bros. Brawl and Super Smash Bros. for Wii U TM ) feature better, but still
TM TM

not optimal, multi-jumping AI. Due to the online nature of Super Smash Bros.TM for
Wii U TM it is possible that the AI in this game could be improved in a future update.
While multi-jumping is a common feature in video games, the only games we could
find that feature real-time multi-jumping AI are the games in the Super Smash Bros.TM
series.
In reality, platforms have finite length. We use the assumption of infinite-length plat-
forms only to justify the correctness of our player and AI strategies, and if our strate-
gies would cause a character to overshoot a finite-length platform, the same strategies
could be applied to land successfully on such a platform by either lowering the hori-
zontal velocity of the character or by initiating the multi-jump earlier, i.e., farther to
the left.
We also consider applications to game level design. Given a fixed platform and a
fixed sequence of jumps available to the player, we consider the problem of how
to design the ground around the platform so that the platform is reachable—or
unreachable—by a multi-jump starting at any point on the ground. Platform-based
adventure games where the player gains new abilities as she explores the map and
uses these new powers to reach previously inaccessible areas are frequently referred

1014 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
to in the gaming community as Metroidvanias [9]. (This term is a combination of
MetroidTM and Castlevania R
, which are two famous series of games featuring similar
gameplay.) In a typical Metroidvania the player eventually gains the ability to jump
a second or third time in midair. Our results can be used, for example, to place a
platform just tantalizingly out of reach of a player who is only able to jump once (or
twice), but which is easily reached when the player gains the ability to jump a second
(or third) time.
We note that while we carry out our analysis in two dimensions, our results are
directly applicable to multi-jumps occurring in a two-dimensional plane in any three-
dimensional game.
The study of the complexity of games and the development of AI for playing
games have rich histories. For instance, computers have been playing games of strat-
egy against humans for over 35 years, and have become sophisticated enough to
challenge the world’s best players. IBM R
’s Deep BlueR
famously bested former
World Chess Champion Garry Kasparov in a 1997 series [4]. More recently, Google R

TM
DeepMind ’s AlphaGo AI [6] beat Lee Sedol, the world’s top Go player over the pre-
vious decade, four to one in a five-game series [2]. There is also a wealth of research
directly applicable to video game AI. For instance, pathfinding algorithms such as the
A∗ algorithm [3] are important for AI in real-time strategy and first-person shooter
games. As for complexity, a number of video games, including generalized versions of
MetroidTM , have been shown to be NP-hard—see [1] and the references therein. To our
knowledge this paper marks the first time that the problem of optimal multi-jumping
in video games has been studied.

2. JUMP FUNCTIONS AND MULTI-JUMPS. Jump functions are the building

blocks of multi-jumps. The graph of a jump function captures the trajectory that a
character (as represented by a point in space) would follow by starting at the origin
and jumping to the right.

Definition 1. A jump function is a continuous function f : R≥0 :→ R for which

f (0) = 0 and ∃c > 0 such that:
• f is strictly increasing on [0, c], and
• f is continuously differentiable on [c, ∞), with f  (x) ≤ 0 for x ∈ (c, ∞) (so f is
weakly decreasing on [c, ∞)) and f + (c) = 0.
We say that f peaks at c. Additionally if f is concave down (i.e., f  is strictly decreas-
ing) on [c, ∞) and limx→∞ f  (x) = −∞, then we say f is a standard jump function.

This terminology is our own, as jumps in video games have not been studied for-
mally before. Note that if f is a standard jump function, then f is automatically strictly
decreasing on [c, ∞) and limx→∞ f (x) = −∞. By definition, if f is a jump function
we have f + (c) = 0. For convenience we will write f  (c) = 0, which will make our
central result (Theorem 2) easier to state.
Examples of graphs of both standard and nonstandard jump functions may be found
in Figure 3. It is easy to generate jump functions. For instance, let a, c, k > 0, r > 1,
and let g(x) be any strictly increasing continuous function for which g(0) = 0 and
g(c) = k. Then

g(x) if 0 ≤ x < c,
f (x) = (1)
−a(x − c)r + k if x ≥ c

December 2016] OPTIMIZING MULTI-JUMPS 1015

 y y

x x
c c

(A) Standard jump function (B) Standard jump function

y y

x x
c c

(C) Nonstandard jump function (D) Nonstandard jump function

Figure 3. Examples of jump function graphs

is a standard jump function that peaks at c. Jump functions of the form (1) will appear
again in Sections 4, 5, and 6.
Unless otherwise stated, without loss of generality we assume that the character
initiates her multi-jump at the origin (0, 0). Fix a sequence F = ( f 1 , . . . , f n ) of jump
functions.

Definition. If (x1 , . . . , xn−1 ) is a sequence of nonnegative real numbers, then the

piecewise function F(x1 ,...,xn−1 ) : R≥0 → R given by
⎧
⎪
⎪ f 1 (x) if 0 ≤ x ≤ x1 ,
⎪
⎪
⎪
⎪ f 2 (x − x 1 ) + f 1 (x 1 ) if x1 ≤ x ≤ x1 + x2 ,
⎪
⎪
⎨ f 3 (x − x1 − x2 ) + f 2 (x2 ) + f 1 (x1 ) if x1 + x2 ≤ x ≤ x1 + x2 + x3 ,
F(x1 ,...,xn−1 ) (x) = ..
⎪
⎪ .
⎪
⎪ n−2 n−2 n−2 n−1
⎪
⎪ (x − ) + f (x ) xi ≤ x ≤ i=1 xi ,
⎪
⎪ f n−1
n−1
x i
n−1i=1 i i
if i=1
n−1
⎩ i=1
f n (x − i=1 xi ) + i=1 f i (xi ) if i=1 xi ≤ x

is called the multi-jump defined by (x1 , . . . , xn−1 ).

The graph of F(x1 ,...,xn−1 ) captures the trajectory a character would follow by starting
at the origin and jumping to the right n times, following the arcs of f 1 through f n in
sequence, where the arc of the first jump is followed for x1 horizontal units, the arc of
the second is followed for x2 horizontal units, and so on, and the arc of f n is followed
indefinitely.
n xn ≥ 0, we denote by F(x1 ,...,xn ) the restriction of F(x1 ,...,xn−1 ) to the domain
Given
[0, i=1 xi ]. That is, the graph of F(x1 ,...,xn ) is the same as the graph of F(x1 ,...,xn−1 ) ,
except that in F(x1 ,...,xn ) the final arc of the multi-jump is followed for only xn units.
The x1 , . . . , xn are called jump points, and (x1 + · · · + xn , F(x1 ,...,xn ) (x1 + · · · + xn ))
is called the ending point of the multi-jump. Note that the final jump point xn does
not initiate a new jump, and that the ending point of a multi-jump might be in midair.

1016 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 4.5 y 4 y y

2.25
x x x
3 6 2 4 1.5 3
y = f1 (x) y = f2 (x) y = f3 (x)

9.75
8.375
7.75

4.375 4

0 0
0 3.5 5.5 0 4 6.5 8.5
y = F (3.5,2) (x) y = F (4,2.5,2) (x)

Figure 4. Graphs of F(x1 ,x2 ) and F(x1 ,x2 ,x3 ) for the jump functions in the Main Example, together with the
platform beginning at (8.5, 9.25)

Also note that F(x1 ,...,xn ) (x1 + · · · + xn ) = f (x1 ) + · · · f (xn ), and that the maximum
possible height achievable by a multi-jump is f 1 (c1 ) + · · · + f n (cn ), achievable by
jumping at peaks—that is, by selecting the jump points (c1 , . . . , cn ). In summary, given
a sequence of n jump functions, a sequence of n − 1 jump points defines a multi-jump
with no ending point, while a sequence of n jump points defines a multi-jump with an
ending point. In general n can be any fixed integer n ≥ 1. In this paper we will draw
all of our figures with n = 3.

Main Example, Part 1. For the Main Example let us take the following sequence of
jump functions. Let

f 1 (x) = −0.5(x − 3)2 + 4.5,

f 2 (x) = −(x − 2)2 + 4,
f 3 (x) = −(x − 1.5)2 + 2.25.

These functions are the building blocks for Figure 2, in which the multi-jumps begin
at (0, 0) and the distant platform begins at (8.5, 9.25). The graph in Figure 2(a) is the
graph of F(3.5,2) , while the graph in Figure 2(b) is the graph of F(4,2.5,2) , with domain
extended slightly to the right so that the final arc touches the platform. See Figure 4.
We note that there is no requirement for the f i to be quadratics (or even polynomi-
als) in general. We have chosen them to all be quadratics for this example simply for
ease of hand calculations later in the paper.

3. THE FUNDAMENTAL THEOREM. In this section we state and prove what

we call the fundamental theorem of multi-jumping optimization. This theorem applies
to all jump functions and has far-reaching consequences, as it and its primary con-
sequence (Theorem 3) form the basis for the design of all of our player and AI

December 2016] OPTIMIZING MULTI-JUMPS 1017

 9.75
8.375
7.75

4.375 4

0 0
0 3.5 5.5 8.5 0 4 6.5 8.5
F (3.5,2,3) is a non-optimal F (4,2.5,2) is the unique opti-
multi-jump ending at x = 8.5. mal multi-jump ending at x =
8.5; hmax (8.5) = 9.75.

Figure 5. Optimal and nonoptimal multi-jumps, drawn together with the platform beginning at (8.5, 9.25) and
tangent lines at jump points

strategies and are integral ingredients in our proofs throughout the rest of the paper.
Let ( f 1 , . . . , f n ) be a sequence of jump functions, where f i peaks at ci ∈ R>0 . Let
C = c1 + · · · + cn . Recall our assumption that multi-jumps begin at the origin (0, 0).

Theorem 2 (The fundamental theorem of multi-jumping optimization). Let d ∈ R

such that d ≥ C. Then there exists a largest value hmax(d) ∈ R such that the platform
beginning at (d, hmax(d)) is reachable by a multi-jump, and if (x1 , . . . , xn ) is any
sequence of jump points for which F(x1 ,...,xn ) ends at (d, hmax(d)), then xi ≥ ci for
i ∈ {1, . . . , n}, and

f 1 (x1 ) = f 2 (x2 ) = · · · = f n (xn ).

Remark. In summary, the fundamental theorem says that for a multi-jump to be

optimal—that is, achieve the maximum possible height after traveling a specified
horizontal distance d—the slope of the tangent line to the character’s trajectory at
the ending point of the multi-jump must equal the slope of the tangent line to her tra-
jectory at every jump point along the way. See Figure 5, which revisits the multi-jumps
from Figure 4. (In Figure 5, that F(4,2.5,2) is in fact optimal and is the unique optimal
multi-jump ending at x = 8.5 will be shown in the Main Example, Part 2 at the end of
this section.)
Optimal multi-jumps corresponding to a sequence of increasing d values are shown
in Figure 6. This figure demonstrates the interplay between d and hmax(d), in that
hmax(d) is a (generally weakly) decreasing function of d. This figure also shows
that hmax(d) can be negative when d is large enough. From a gameplay perspec-
tive hmax(d) being negative makes perfect sense—sometimes the target platform is
at a lower height than the character’s starting position. Also note from this figure that
when hmax(d) = 0, the jump points for an optimal multi-jump do not have to occur
along the line y = 0.

Proof of Theorem 2. First we show the existence of hmax(d) using a bit of topology.
Let X = [0, d]n ⊂ Rn and let Y = {(x1 , . . . , xn ) ∈ X : x1 + · · · + xn = d}.
Y is an n − 1-simplex, so Y is compact. Since the function H : [0, ∞)n → R given
by H (x1 , . . . , xn ) = f 1 (x1 ) + · · · + f n (xn ) is continuous, and the continuous image of

1018 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 10.75
6.75

0 0

0 3 5 6.5 0 5 8 10.5
hmax (6.5) = 10.75 hmax (10.5) = 6.75

0 0

− 5.25
p1 p2 p3
0 0 7 11 14.5
hmax(6.5 + √43) = 0 hmax (14.5) = –5.25
√ 43
p1 = (3 + 2
, − .875)
3 √ 43
p2 = (5 + 4
, .4375)

p3 = (6.5 + √ 43, 0)

Figure 6. Optimal multi-jumps corresponding to selected d-values; jump functions from the Main Example

a compact set is compact, H attains a maximum value on Y . This maximum value is

by definition hmax(d).
Let (x1 , . . . , xn ) be a sequence of jump points such that F(x1 ,...,xn ) ends at (d, hmax(d)),
i.e., for which x1 + · · · + xn = d and f 1 (x1 ) + · · · + f n (xn ) = hmax(d).
Next we show that xi ≥ ci for all i. Suppose not, so for some i we have 0 ≤ xi < ci .
Since x1 + · · · + xn = d ≥ C = c1 + · · · + cn , for some j we have x j > c j . Let  > 0
such that xi +  < ci and x j −  > c j . Then the sequence of jump points (x1 , . . . , xn )
given by
⎧
⎪
⎨xi +  if k = i,
xk = x j −  if k = j,
⎪
⎩x
k otherwise

has the property that x1 + · · · + xn = d. Furthermore, since f i (xi ) > f i (xi ) and
f j (x j ) ≥ f j (x j ) we have f 1 (x1 ) + · · · + f n (xn ) > f 1 (x1 ) + · · · + f n (xn ) = hmax(d),
contradicting the maximality of hmax(d). Therefore xi ≥ ci for all i.
Finally we show that f 1 (x1 ) = f 2 (x2 ) = · · · = f n (xn ).
If xi > ci for all i we may apply the method of Lagrange multipliers in Rn (see,
e.g., [7, Ex. 5-16]) to the domain (c1 , ∞) × · · · × (cn , ∞) ⊂ Rn of the objective func-
tion H (x1 , . . . , xn ) = f 1 (x1 ) + · · · + f n (xn ), subject to the constraint g(x1 , . . . , xn ) =

December 2016] OPTIMIZING MULTI-JUMPS 1019

0, where g(x1 , . . . , xn ) = x1 + · · · + xn − d. Applying Lagrange multipliers, we
immediately obtain the system of equations

f 1 (x1 ) = λ,
f 2 (x2 ) = λ,
..
.
f n (xn ) = λ

for some λ ∈ R, so λ = f 1 (x1 ) = f 2 (x2 ) = · · · = f n (xn ).

On the other hand, if xi = ci for some i, then f i (xi ) = 0. Suppose for the sake of
contradiction that f j (x j ) = 0 for some j. Then x j > c j and f j (x j ) < 0. For the sake
of clarity we give the idea of the rest of the argument, which can be made rigorous
in a straightforward, albeit somewhat lengthy manner. Since f i and f j are continuous
on [ci , ∞) and [c j , ∞), f i (xi ) = 0, and f j (x j ) < 0, it is possible to bump xi slightly
to the right and x j slightly to the left to arrive at a contradiction. In particular, due
to the continuity of the derivatives it is possible to choose points xi and x j such that
xi > xi , c j < x j < x j , xi + x j = xi + x j , and f i (xi ) + f j (x j ) > f i (xi ) + f j (x j ). It
then immediately follows that for the sequence of jump points given by (x1 , . . . , xn )
(where xk = xk for k = i, j), we have x1 + · · · + xn = d and f 1 (x1 ) + · · · + f n (xn ) >
f 1 (x1 ) + · · · + f n (xn ) = hmax(d), contradicting the maximality of hmax(d). It fol-
lows that 0 = f 1 (x1 ) = f 2 (x2 ) = · · · = f n (xn ), completing the proof.

The rest of the paper will focus on standard jump functions. While the fundamental
theorem provides a necessary condition for a multi-jump to be optimal, for standard
jump functions it is also sufficient, as the following very useful consequence demon-
strates.

Theorem 3. Suppose f 1 , . . . , f n are standard. Write f i−1 for the inverse of the restric-
tion of f i to [ci , ∞). Let d ∈ R such that d ≥ C. Then for some unique x1 ≥ c1 we
have

n
d = x1 + f i−1 ( f 1 (x1 )),
i=2


n
hmax(d) = f 1 (x1 ) + f i ( f i−1 ( f 1 (x1 ))),
i=2

and the unique multi-jump ending at (d, hmax(d)) is defined by the sequence of jump
points

(x1 , f 2−1 ( f 1 (x1 )), . . . , f n−1 ( f 1 (x1 )))

(i.e., the sequence (x1 , x2 , . . . , xn ) for which f 1 (x1 ) = f 2 (x2 ) = · · · = f n (xn )).

Proof. Since f i is standard, f i restricted to [ci , ∞) is invertible, and as a function of

n
x1 ∈ [c1 , ∞), the function x1 + i=2 f i−1 ( f 1 (x1 )) is increasing, continuous, and has
n
range [C, ∞). Hence there exists unique x1 ≥ c1 such that d = x1 + i=2 f i−1 ( f 1 (x1 )).

1020 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 Suppose the sequence of jump n points (z 1 , . . . , z n ) defines
n a multi-jump ending
at (d, hmax(d)). Then d = i=1 z i and hmax(d) = i=1 f i (z i ). By the funda-
mental theorem z i ≥ ci for all i, and for i ≥ 2 we have f 1 (z 1 ) = f i (z i ). Hence
n
z i = f i−1 ( f 1 (z 1 )). Since i=1 z i = d, we have z 1 = x1 , so

(z 1 , . . . , z n ) = (x1 , f 2−1 ( f 1 (x1 )), . . . , f n−1 ( f 1 (x1 )))

is the unique sequence of jump points

n defining a multi-jump ending at (d, hmax(d)).
Furthermore, we have hmax(d) = i=1 n
f i (z 1 ) = f 1 (x1 ) + i=2 f i ( f i−1 ( f 1 (x1 )).

Remark 4. For standard jump functions, Theorem 3 reduces the problem

n of finding
optimal multi-jumps to the problem of analyzing the function x1 + i=2 f i−1 ( f 1 (x1 ))
of the single variable x1 .

Main Example, Part 2. The functions in the Main Example are all standard. Let us
use Theorem 3 to find hmax(8.5) as well as the jump points x1 , x2 , x3 such that the
multi-jump F(x1 ,x2 ,x3 ) ends at (8.5, hmax(8.5)). We compute

f 1 (x) = −x + 3,
x
f 2 (x) = −2x + 4, f 2−1 (x) = − + 2,
2
x 3
f 3 (x) = −2x + 3, f 3−1 (x) = − + .
2 2

We seek x1 such that x1 + f 2−1 ( f 1 (x1 )) + f 3−1 ( f 1 (x1 )) = 8.5, that is,

−x1 + 3 −x1 + 3 3
x1 + − +2 + − + = 8.5,
2 2 2

which yields x1 = 4. We then obtain x2 = f 2−1 ( f 1 (4)) = 2.5 and x3 = f 3−1 ( f 1 (4))
= 2. Finally we get

hmax(8.5) = f 1 (4) + f 2 (2.5) + f 3 (2) = 9.75,

so F(4,2.5,2) is the unique multi-jump ending at (8.5, hmax(8.5)) = (8.5, 9.75), as

shown in Figure 5. The jump points for the graphs in Figure 6 were computed in an
analogous fashion.

4. PLAYER STRATEGY.

Definition 5. A jump function f is fully concave if f is standard and f is concave

down on [0, ∞).

Of the jump functions appearing in Figure 3, only (A) is fully concave. Let
( f 1 , . . . , f n ) be a sequence of jump functions, where f i peaks at ci . Let C = c1 +
· · · + cn . Fix d ≥ C and h ∈ R. Consider the problem of selecting a multi-jump to
reach a platform beginning at (d, h), assuming the platform is reachable. In this sec-
tion we give a simple and complete solution to this problem, implementable in real
time by human players (and AI), in the case that f 1 = · · · = f n are fully concave jump
functions. This is the case for a majority of games that feature multi-jumps, probably

December 2016] OPTIMIZING MULTI-JUMPS 1021

 (0, 0)
(0, 0)
(d, h) (0, 0) (d, h) (d, h)

Figure 7. The line method selects successful multi-jumps.

because this scenario is the easiest to implement by a game programmer. In Section 5

we will address this problem for AI-controlled characters in the relaxed situation that
f 1 , . . . , f n are standard and not necessarily equal.
Suppose f 1 is a fully concave jump function. Write f = f 1 , c = c1 , and consider
the sequence of jump functions ( f, . . . , f ). Here is our strategy, which we call the
  
n
line method. Consider the line connecting (0, 0) and (d, h). Whenever the character’s
trajectory intersects this line, jump. See Figure 7.

Theorem 6. If the platform beginning at (d, h) is reachable, then the line method will
select a multi-jump that causes the character to land on the platform.

Proof. By Theorem 3 the unique sequence of jump points (z 1 , . . . , z n ) defining the

multi-jump ending at (d, hmax(d)) has z 1 = · · · = z n . Let z = z 1 , so nz = d. Since
d ≥ C = nc we have z ≥ c. Since f is concave down it lies above any of its secant
lines, so the following “strategy” selects the jump points (z, . . . , z): Consider the line
L̂ connecting (0, 0) and (d, hmax(d)). Whenever the character’s trajectory intersects
L̂, jump. (This is not a strategy a player could follow because hmax(d) is not known
to the player.)
Since the platform beginning at (d, h) is reachable we have h ≤ hmax(d). The
slope of L̂ is hmax(d)
d
, while the slope of the line L for the line method is dh . Since
h ≤ hmax(d) and d > 0 we have dh ≤ hmax(d) d
. In particular since L̂ intersects the down-
ward (and not the upward) trajectory of the character’s first jump, L also intersects
the downward (and not the upward) trajectory of the character’s first jump. Induction
establishes that the line method selects jump points (a, . . . , a) with a ≥ z.
Let F = F(a, . . . , a ) and consider the graph y = F(x) of the multi-jump selected
  
n
by the line method. Using again the fact that the graph of f lies above any of its
secant lines and that na ≥ nz = d we have that the graph y = F(x) from x = 0 to
x = d is always above (or touching) L. Therefore, d horizontal units into this trajec-
tory the character will be at the point (d, F(d)), with F(d) ≤ hmax(d) (by definition
of hmax(d)) and F(d) ≥ h (since L intersects (d, h)). That is, d horizontal units into
the trajectory selected by the line method the character will either be above or touch-
ing the platform. Therefore, since the platform extends indefinitely to the right and
limx→∞ f (x) = −∞, continuing to follow this trajectory to the right will eventually
cause the character to land on the platform.

Remark. We require the assumption that f is concave down throughout its domain to
ensure that the trajectory selected by the line method will not collide with the underside
of the platform. For example, if

1022 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 
x3 if x ≤ 1,
f (x) =
−(x − 1) + 1 if x > 1,
2

then for the sequence of jump functions ( f, f ) we have C = 2, and the platform
beginning at (2, 1) is easily reachable by a double jump (by jumping at the peak, for
instance). However one may check that jumping according to the line method—that is,
jumping in midair when the line connecting (0, 0) and (2, 1) intersects the downward
trajectory of the first jump—would cause the character to collide with the underside of
the platform.

Remark. Suppose the platform beginning at (d, h) is reachable. When the jump func-
tions f 1 , . . . , f n available are not fully concave or not all equal, we have been unable
to give a strategy that a human player could follow that guarantees that the player will
reach the platform. However in this situation the fundamental theorem still provides a
useful heuristic—try to imagine a multi-jump that ends at or above (d, h), and jump
whenever the slope of the tangent to your current jump matches the slope of the tangent
of the ending point of the imagined multi-jump.

5. AI STRATEGY. Let ( f 1 , . . . , f n ) be a sequence of standard jump functions,

where f i peaks at ci . Let C = c1 + · · · + cn . Fix d ≥ C and h ∈ R. Consider the prob-
lem of deciding whether the platform beginning at (d, h) is reachable by a multi-jump,
and if it is, deciding what multi-jump to use to reach the platform.
We now give a numerical solution to this problem which is efficient enough to be
used in real time by an AI-controlled character. Following Remark 4, our strategy is
to analyze the function g(x1 ) = x1 + i=2 n
f i−1 ( f 1 (x1 )) to find the sequence of jump
points (x1 , . . . , xn ) defining the multi-jump that ends at (d, hmax(d)). Then the plat-
form beginning at (d, h) is reachable if and only if h ≤ hmax(d), and if the platform
is reachable, then continuing to follow the multi-jump ending at (d, hmax(d)) to the
right will eventually land the character on the platform (since the platform extends
indefinitely to the right and limx→∞ f n (x) = −∞).
Algorithm 1 implements this approach. It accepts  > 0, f 1 , . . . , f n , c1 , . . . , cn ,
and d ≥ c1 + · · · + cn , and computes (x1 , . . . , xn ) such that, with error less than , the
multi-jump defined by (x1 , . . . , xn ) ends at (d, hmax(d)). Specifically, the output of
Algorithm 1 is guaranteed to satisfy |d − i=1 n
xi | <  and | f i (xi ) − f 1 (x1 )| <  for
all i.
We assume we have access to a function root( f (x), a, b, e), which accepts a con-
tinuous function f (x) on [a, b] and returns z ∈ [a, b] such that f (r ) = 0 for some
r ∈ [a, b] with r ∈ [z − e, z + e], and throws an exception if no such z ∈ [a, b] exists.
Such a function is available in any number of scientific computational packages—see,
e.g., scipy [5]. We also assume that the f i and f i can be evaluated exactly.

Algorithm 1. Algorithm for computing the jump points (x1 , . . . , xn ) defining the
multi-jump ending at (d, hmax(d)).
1 # I n p u t : ( f 1 , . . . , f n ), (c1 , . . . , cn ), d, 
2 # O u t p u t : (x1 , . . . , xn )
3
4 d e f NumericalInverse( f, y, a, b, e) :
5 g(x) = f (x) − y
6 while True :
7 try :
8 r e t u r n root(g(x), a, b, e)

December 2016] OPTIMIZING MULTI-JUMPS 1023

 9 e x c e p t R u n t i m e E r r o r ( no r o o t ) : # i f t h e r e i s no r o o t on [a, b]
10 b = b × 10 # t h e n e x p a n d t h e s e a r c h r a n g e t o t h e r i g h t
11
12 d e f OptimalJumpPoints(( f 1 , . . . , f n ), (c1 , . . . , cn ), d, ) :
13 e=
14 d e f DerivativeCheck((x1 , . . . , xn )) :
15 f o r i = 2, . . . , n :
16 i f | f i (xi ) − f 1 (x1 )| ≥  :
17 return False
18 r e t u
r n True
19 F(x) = x − d + i=2 n
NumericalInverse( f i , f 1 (x), ci , d, e/n)
20 while True :
21 x1 = root(F(x), c1 , d, e/n)
22 f o r i = 2, . . . , n :
 
23 xn i = NumericalInverse( f i , f 1 (x1 ), ci , d, e/n)
24 i f |d − i=1 xi | <  and DerivativeCheck((x1 , . . . , xn )) :
25 r e t u r n True
26 else :
27 e = e × 0.1
28
29 r e t u r n OptimalJumpPoints(( f 1 , . . . , f n ), (c1 , . . . , cn ), d, )

In Algorithm 1 we assume that the f i−1 must be evaluated numerically. If they can
be evaluated exactly, a handful of simple changes makes the algorithm significantly
more efficient:

Algorithm 2. Modification of Algorithm 1 when the f i−1 can be computed exactly.

• The functions NumericalInverse and DerivativeCheck are no longer neces-
sary.
• NumericalInverse( f i , f 1 (x), ci , d, e/n) is replaced with f i−1 ( f 1 (x)) on line 19.
• NumericalInverse( f i , f 1 (x1 ), ci , d, e/n) is replaced with f i−1 ( f 1 (x1 )) on line
23.
• The phrase “and DerivativeCheck((x1 , . . . , xn ))” is removed from line 24.

We have implemented Algorithms 1 and 2 in Sage, a free and open-source computer

algebra system [8], and have found our implementations to be quite fast in practice
across a wide range of standard jump functions.
For purposes of comparison we tested Algorithms 1 and 2 with functions of the
form given by (1). Each of our two tests consisted of 2500 samples, where a sample
consisted of a sequence of jump functions ( f 1 , f 2 , , . . . , fn ) (with each f i of the form
n n
f i (x) = −ai (x − ci )ri + ki ), d a random real number in i=1 ci , 1.5 i=1 ci , and
a set value of . We ran each sample 50 times in each of our algorithms and took the
time for the 40th-slowest run. That is, 80% of calls to our algorithms ran at least as
quickly as the numbers reported here. We chose to report at the 80th percentile level
(as opposed to the slowest-seen level) to give reasonable running time estimates—
the slowest calls are sometimes slowed by system-specific scheduling issues that have
nothing to do with the intrinsic speed of our algorithms. We ran our tests in a single
thread on an Intel R
CoreTM i5-6600 processor at stock speed.
To establish a baseline for comparison, our first test was n = 3, with ai ∈ [0.5, 10], ci
∈ [1, 10], ki ∈ {1, . . . , 10} (each selected uniformly at random), ri = 2, and  = 10−9 .

1024 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 Observed frequency (out of 2500)
Observed frequency (out of 2500)
800
400
600

200 400

200

0 0
0 100 200 0 10 20 30
Algorithm 1 time (ms) Algorithm 2 time (ms)

Figure 8. Algorithm test 2

In this test both algorithms were extremely fast, with Algorithm 1 always completing
in less than 10.1 milliseconds and Algorithm 2 in less than 1.7 milliseconds.
Algorithm 1 running time (ms)

300 Algorithm 2 running time (ms) 30

200 20

100 10

0 0
0 2,000 4,000 0 20 40 60 80
Number of function evaluations Number of function evaluations

Figure 9. Algorithm test 2

= 10−1 = 10−4 = 10−9

300 300 300
Time (ms)

Time (ms)

Time (ms)

200 200 200

100 100 100
0 0 0
1 1.25 1.5 1 1.25 1.5 1 1.25 1.5
d/C d/C d/C
Figure 10. Algorithm 1 speeds for the Main Outlier: n = 4 with f 1 (x) = −9.76(x − 5.81)2 + 7, f 2 (x)
= −8.17(x − 7.83)5 + 5, f 3 (x) = −8.44(x − 7.59)5 + 1, f 4 (x) = −6.16(x − 7.64)4 + 3

In our second test we went beyond parabolas, taking n = 4 and selecting

ri ∈ {2, 3, 4, 5} uniformly at random. We continued to take ai ∈ [0.5, 10], ci ∈
[1, 10], ki ∈ [1, 10] (each selected uniformly at random), and  = 10−9 . In this test
Algorithm 1 always completed within 242 milliseconds, while Algorithm 2 always

December 2016] OPTIMIZING MULTI-JUMPS 1025

 = 10−1 = 10−4 = 10−9
300 300 300

Time (ms)

Time (ms)

Time (ms)
200 200 200
100 100 100
0 0 0
1 2 3 1 2 3 1 2 3
d/C d/C d/C
Figure 11. Algorithm 1 speeds for n = 4 for a typical sample: f 1 (x) = −5.05(x − 8.66)4 + 8, f 2 (x)
= −5.02(x − 3.30)3 + 4, f 3 (x) = −5.95(x − 3.75)2 + 8, f 4 (x) = −5.35(x − 2.75)4 + 6

completed within 29 milliseconds. Median completion times were 27 milliseconds

and 4 milliseconds, respectively. See Figure 8.
The primary factor affecting the running times of our algorithms is the number of
function evaluations made during root finding. Scatter plots of the number of function
evaluations versus overall algorithm speed in our second test are given in Figure 9,
together with their lines of best fit. For each algorithm we see three outliers (which
are in fact the same three samples), the worst of which required over 4000 function
evaluations in Algorithm 1. That sample was, to two decimal places (which is enough
to replicate this behavior),

f 1 (x) = −9.76(x − 5.81)2 + 7, f 2 (x) = −8.17(x − 7.83)5 + 5,

f 3 (x) = −8.44(x − 7.59)5 + 1, f 4 (x) = −6.16(x − 7.64)4 + 3.

We will call this the Main Outlier. For each of our outliers we found that d was only
very slightly larger than C, and that either increasing d slightly or decreasing  was
enough to make our algorithms run with times comparable to the other samples in
our tests. In Figure 10 we show the running time of Algorithm 1 for the Main Outlier
across a range of d and  values. The type of information in Figure 10 is the most perti-
nent type of information for a game’s AI developer, who works with a fixed collection
of jump functions. This information can vary considerably depending on the specific
collection of jump functions. For instance, in Figure 11 we display the running times
of a typical sample in our tests across a range of d and  values. If that sample were
the collection of jump functions in our video game, we could be confident that Algo-
rithm 1 would complete quickly given any reasonable combination of d and  input
values.

6. LEVEL DESIGN. In this section we consider multi-jumps beginning at points

other than the origin. Let ( f 1 , . . . , f n ) be a sequence of jump functions and (x1 , . . . , xn )
a sequence of nonnegative real numbers. The multi-jump beginning at (a, b) defined
by (x1 , . . . , xn ), denoted F(x(a,b)
1 ,...,x n )
, is the translation of F(x1 ,...,xn ) horizontally by |a|
units (left if a < 0 and right if a > 0) and vertically by |b| units (up if b > 0 and down
if b < 0). The graph of F(x(a,b)1 ,...,x n )
captures the trajectory a character would follow by
starting at (a, b) and jumping to the right n times, following the arcs of f 1 through
f n in sequence, where the arc of the first jump is followed for x1 horizontal units, the
arc of the second is followed for x2 horizontal units, and so on, and the arc of f n is
followed for xn units. The multi-jump F(x(a,b) 1 ,...,x n )
covers a total horizontal distance of
x1 + · · · + xn units and ends at a height of b + f (x1 ) + · · · + f n (xn ).

1026 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 Suppose now that f 1 , . . . , f n are standard, xi ≥ ci for all i, and f 1 (x1 ) = f 2 (x2 ) =
 n
· · · = f n (xn ). Let d = i=1 xi and h = hmax(d). By Theorem 3, of all multi-jumps
beginning at (0, 0), the multi-jump defined by (x1 , . . . , xn ) is the unique multi-jump
ending at (d, h). In this section we address the question: How can we design the ground
G so that for any point (x, G(x)) on the ground, the unique optimal multi-jump starting
at (x, G(x)) and ending at x = d also has the property that it ends at y = h? Answer-
ing this question allows a game designer to place a platform which is guaranteed to be
reachable from any point on the ground (by having the platform begin at or just below
(d, h)), or more interestingly, to place a platform which is guaranteed to be just barely
unreachable from every point on the ground (by having the platform begin just above
(d, h)).
Let C = c1 + · · · + cn . Suppose that x ∈ R with d − x ≥ C, and that the unique
optimal multi-jump beginning at (x, y) and ending at (d, h) is defined by jump points
(x1 , . . . , xn ). We have f 1 (x1 ) = f 2 (x2 ) = · · · = f n (xn ),


n
x1 + f i−1 ( f 1 (x1 )) = d − x, (2)
i=2

and

n
f 1 (x1 ) + f i ( f i−1 ( f 1 (x1 ))) = h − y. (3)
i=2

By (2) x1 is determined by x, so by (3) y is determined by x. That is, there is only

one possible design for the ground having the required property. Write y = G(x). One
can deduce G(x) by solving (2) for x1 and then using (3) to find y = G(x). For most
collections of jump functions this will require a numerical solution, but when the f i are
all of the form given by (1) and have the same degree, for instance, an exact solution
can be given.

Theorem 7. Suppose f i (x) = −ai (x − ci )r + ki for x ≥ ci for all i, with ai , ci , ki > 0

and r > 1. Then
 n 1−r  
  n
1/(1−r )
G(x) = ai (d − C − x) + h −
r
ki
i=1 i=1

defines the unique ground shape {(x, G(x)) : d − x ≥ C} for which the optimal multi-
jump beginning at (x, G(x)) and ending at x = d also ends at y = h.

Before we prove Theorem 7 we give several remarks.

Remark. In Theorem 7 the shape of the ground is a polynomial of the same degree
as each of the f i . In the case that r = 2, the shape of the ground is a parabola and the
formula
   
1 n
G(x) = n 1 (d − C − x) + h − 2
ki
i=1 ai i=1

is given in vertex form.

December 2016] OPTIMIZING MULTI-JUMPS 1027

 The ground G (x) = 0.25 (x − 2)2 − 1,
x≤2

9.75

0
−1
0 2 8.5
Figure 12. The optimal multi-jump beginning at (0, 0) and ending at (8.5, hmax(8.5) = 9.75), drawn together
with the ground G and the platform beginning at (8.5, 9.75); jump functions from the Main Example

Main Example, Part 3. Using the functions in the Main Example with d = 8.5 and
h = hmax(8.5) = 9.75, we obtain
 
1
G(x) = 1 (8.5 − (3 + 2 + 1.5) − x)2 + [9.75 − (4.5 + 4 + 2.25)]
1/2
+ 1
1
+ 1
1

1
= (2 − x)2 − 1,
4
with x ≤ 2. In Figure 6 we graph G together with F(4,2.5,2) , the unique optimal multi-
jump beginning at (0, 0) and ending at (8.5, 9.75). In Figure 6 we add to this the graphs
of several other optimal multi-jumps beginning along the ground G, all of which end
at x = 8.5 and hence at (8.5, 9.75).

Remark. The reader may have noticed that in Figure 6, the slope of the tangent line
to the ground at the starting point of any multi-jump appears to be equal to the slope of
the tangent line at every jump point on that multi-jump. Indeed they are equal, and this
is no coincidence—in general, as long as the ground G is continuously differentiable
with G  (x) ≤ 0 for x ≤ d − C, one may regard a character walking along G as follow-
ing the decreasing portion of the first jump of a multi-jump and apply the fundamental
theorem to conclude that if F(x(x,G(x))
1 ,...,x n )
is an optimal multi-jump that begins at (x, G(x))
and ends at (d, h), then G  (x) = f 1 (x1 ) = · · · = f n (xn ).

Remark. By definition, any ground that lies strictly below G is a ground shape from
which the platform beginning at (d, h) is unreachable using any multi-jump. See
Figure 14.

Proof of Theorem 7. Let G be the shape of the ground having the desired property.
Since f i (x) = −rai (x − ci )r −1 we have

1/(r −1)
−x
f i−1 (x) = + ci ,
rai

1028 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 The ground G (x) = 0.25 (x − 2)2 − 1,
x≤2

9.75

0
−1
0 2 8.5
Figure 13. All optimal multi-jumps beginning along G and ending at x = 8.5 end at (8.5, hmax(8.5)); jump
functions from the Main Example

G given d = 8.5 and

hmax(8.5) = 9.75
H
9.75

0
−1

0 2 8.5

Figure 14. A ground design H from which no multi-jump can reach the platform beginning at (8.5, 9.75)

and hence

1/(r −1)
a1
f i−1 ( f 1 (x1 )) = (x1 − c1 ) + ci . (4)
ai

By (2), then, we have



n
a1 1/(r −1)
x1 + (x1 − c1 ) + ci = d − x.
i=2
ai

December 2016] OPTIMIZING MULTI-JUMPS 1029

Solving for x1 we obtain
n  a1 1/(r −1) n
d − x + c1 i=2 ai
− i=2 ci
x1 =  1/(r −1)
n
1+ i=2
a1
ai

n  a1 1/(r −1) n
d − x + c1 i=1 ai
− i=1 ci
=  1/(r −1)
n a1
i=1 ai

n  a1 1/(r −1)
d − C − x + c1 i=1 ai
=
  1/(r −1) n a1
i=1 ai

d −C −x
= + c1 .
n  a1 1/(r −1)
i=1 ai

That is,
d −C −x
x1 − c1 = .
n  a1 1/(r −1)
(5)
i=1 ai

r/(r −1)
−a1
It follows from (4) that f i ( f i−1 ( f 1 (x1 ))) = ai 1/(r −1)
(x1 − c1 )r + ki , so by (5) we have
r/(r −1)
−a1
f i ( f i−1 ( f 1 (x1 ))) =
(d − C − x)r + ki
n  a1 1/(r −1) r
ai 1/(r −1) j=1 aj

−1
=
(d − C − x)r + ki
  1/(r −1) r
n
ai 1/(r −1) j=1
1
aj

−ai 1/(1−r )
=  r (d − C − x)r + ki .
n 1/(1−r )
j=1 a j

Putting this together with (3), we have


n
G(x) = y = h − f i ( f i−1 ( f 1 (x1 )))
i=1
⎛ ⎞
n
⎜ −ai 1/(1−r )
⎟
=h− ⎝  r (d − C − x)r + ki ⎠
n 1/(1−r )
i=1 j=1 a j
⎡ ⎤
 
⎢ 
n n
ai 1/(1−r ) ⎥
=⎣  r ⎦ (d − C − x) + h −
r
ki
n 1/(1−r )
i=1 j=1 j a i=1

1030 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 ⎡ ⎤
n  
⎢ ai
1/(1−r )
⎥ n
= ⎣  i=1 r ⎦ (d − C − x)r + h − ki
n 1/(1−r )
j=1 a j i=1

 1−r  

n 
n
1/(1−r )
= ai (d − C − x) + h −
r
ki
i=1 i=1

as claimed.

7. SUMMARY AND CONCLUSIONS. We have introduced jump functions, stan-

dard jump functions, and fully concave jump functions, and have studied how to com-
bine them optimally into multi-jumps. We have also given strategies for reaching a
distant platform with a multi-jump that can be followed by any player or AI (if the
multi-jump is formed from a sequence of equal fully concave jump functions), or by
an AI after some quick numerical computation (if the multi-jump is formed from a
sequence of standard jump functions). We have also shown how a game designer can
build the ground around a platform so that the platform is reachable (or unreachable)
from any point on the ground by a multi-jump.
With the exception of the fundamental theorem, our results only apply to multi-
jumps formed by standard jump functions. When the jumps available to a character are
nonstandard it is possible for there to be several multi-jumps ending at (d, hmax(d)),
and the condition f 1 (x1 ) = · · · = f n (xn ) is only a necessary condition for the sequence
of jump points (x1 , . . . , xn ) to define an optimal multi-jump. This situation should be
explored further. It would be interesting to give an AI strategy for selecting multi-
jumps ending at (d, hmax(d)) when the jump functions are nonstandard. It would also
be interesting to give a faster AI strategy than Algorithm 1 for standard jump functions
when their derivative inverses cannot be evaluated exactly. We are currently working
on an AI strategy that will apply when the jump functions are nonstandard, which we
hope will also be faster than Algorithm 1.

ACKNOWLEDGMENT. The authors are grateful to the College of Sciences at Sam Houston State University
for the undergraduate research award that partially supported this work and to the anonymous referees for their
helpful comments.

REFERENCES

1. G. Aloupis, E. D. Demaine, A. Guo, G. Viglietta, Classic Nintendo games are (computationally) hard,
Theoret. Comput. Sci. 586 (2015) 135–160, http://dx.doi.org/10.1016/j.tcs.2015.02.037.
2. AlphaGo: Google DeepMind, http://deepmind.com/alpha-go.html. Accessed March 23, 2016.
3. P. E. Hart, N. J. Nilsson, B. Raphael, A formal basis for the heuristic determination of minimum
cost paths, IEEE Trans. Syst. Sci. Cybern. 4 (1968) 100–107, http://dx.doi.org/10.1109/TSSC.
1968.300136.
4. IBM 100—Deep Blue, http://www-03.ibm.com/ibm/history/ibm100/us/en/icons/deepblue/.
Accessed March 23, 2016.
5. E. Jones, T. Oliphant, P. Peterson et al., SciPy: Open source scientific tools for Python (2001–),
http://www.scipy.org/. Accessed Dec. 3, 2015.
6. D. Silver et al., Mastering the game of Go with deep neural networks and tree search, Nature 529 (2016)
484–489, http://dx.doi.org/10.1038/nature16961.
7. M. Spivak, Calculus On Manifolds: A Modern Approach To Classical Theorems Of Advanced Calculus.
Westview Press, Boulder, CO, 1965 and 1998.
8. W. A. Stein et al., Sage Mathematics Software (Version 5.6.0), The Sage Development Team (2013),
http://www.sagemath.org. Accessed Dec. 3, 2015.

December 2016] OPTIMIZING MULTI-JUMPS 1031

9. What is Metroidvania? (2015) http://www.youtube.com/watch?v=LfEOEqnYiM4. Accessed March
23, 2016.

AARON M. BROUSSARD received his B.S. in mathematics from Sam Houston State University in 2013.
He then became a Simulation Software Engineer at Lockheed Martin Aeronautics and Space Systems. His free
time usually involves working on puzzles with his wife, reading xkcd, math, programming, and video games.
abroussard11@gmail.com

MARTIN E. MALANDRO received his B.S. in mathematics from Texas Tech University in 2003 and his
Ph.D. in mathematics from Dartmouth College in 2008. He then joined the faculty at Sam Houston State
University, where he is now Associate Professor of Mathematics. In his spare time he enjoys programming,
music, and the occasional video game.
Department of Mathematics and Statistics, Box 2206, Sam Houston State University, Huntsville, TX 77341-
2206
malandro@shsu.edu

ABAGAYLE SERREYN is an undergraduate chemistry major and math minor at College of the Ozarks. She
plans to graduate in May of 2018 and become certified to teach high school. When she is not busy doing lab
work she relaxes by reading speculative fiction, solving math and logic puzzles, and singing along with her
favorite songs.
abagayleserreyn@gmail.com

100 Years Ago This Month in The American Mathematical Monthly

Edited by Vadim Ponomarenko
Recent papers read before the ASSOCIATION and the SOCIETY indicate that renewed
interest is apparent in all phases of mathematical history. Hence, no apology is
needed for the publication of notes such as the following:
In Nature, December 3, 1914, p. 363, Professor CAJORI showed that the cross
× as a symbol of multiplication, which is said in histories to occur first in William
Oughtred’s Clavis mathematicae (1631), is given in form of the letter x and X
in Edward Wright’s translation of John Napier’s Mirifici logarithmorum canonis
descriptio, second edition, London, 1618, where we read, page 4: “The note of addi-
tion is (+), of subtracting (−), of multiplying (×).” This is taken from a part of the
book headed “Appendix to the Logarithmes,” the authorship of which is not given
but is believed now most probably to be attributed to William Oughtred.
In 1902 Professor W. W. BEMAN pointed out (L’Intermédiaire des mathémati-
ciens, T. 9, Paris, p. 229, question 2424) that the colon (:) occurs as the symbol
for geometric ratio at the end of the tables in Oughtred’s Trigonometria of 1657.
Professor CAJORI has found that the colon was so used by the astronomer Vincent
Wing in 1651, 1655, and 1656 and by a Suffolk schoolmaster with the initials
“R. B.” in 1655.
The first designation of the sides of a triangle by the same letters, respectively, as
the angles opposite, one group of letters being capitals A, B, C, and the other group
small letters a, b, c, has been attributed to Leonhard Euler (Histoire de l’académie
de Berlin, année, 1753, p. 231), but Professor CAJORI finds that it occurs in a pam-
phlet containing trigonometric formulas published by Richard Rawlinson of Queen’s
College, Oxford, sometime between 1655 and 1668.
—Excerpted from “Notes and News” 23 (1916) 399–404.

1032 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 NOTES
Edited by Sergei Tabachnikov

Dilated Floor Functions That Commute

Jeffrey C. Lagarias, Takumi Murayama, and D. Harry Richman

Abstract. We determine all pairs of real numbers (α, β) such that the dilated floor functions
αx and βx commute under composition, i.e., such that αβx = βαx holds for all
real x.

1. INTRODUCTION. The floor function x rounds a real number down to the near-
est integer. The ceiling function x, which rounds up to the nearest integer, satisfies

x = −−x. (1)

These two fundamental operations discretize (or quantize) real numbers in different
ways. The names floor function and ceiling function, along with their notations, were
coined in 1962 by Kenneth E. Iverson [5, p. 12], in connection with the programming
language APL. Graham, Knuth, and Patashnik [4, Chap. 3] note this history and give
many interesting properties of these functions.
We study the floor function applied to a linear function α (x) = αx, yielding the
dilated floor function f α (x) = αx, where α is a real number. Dilated floor functions
arise in constructing digital straight lines, which are “lines” drawn on two-dimensional
graphic displays using pixels, and are discussed further below. This note addresses the
question: When do two dilated floor functions commute under composition of func-
tions? Linear functions always commute under composition and satisfy the identities

α ◦ β (x) = β ◦ α (x) = αβ (x) for all x ∈ R. (2)

However, discretization generally destroys such commutativity. We have the following.

Theorem 1. The complete set of all (α, β) ∈ R2 such that

αβx = βαx

holds for all x ∈ R consists of:

(i) three continuous families (α, α), (α, 0), (0, α) for all α ∈ R;
(ii) the infinite discrete family
   
1 1
(α, β) = , : m, n ≥ 1 ,
m n

where m, n are positive integers. (The families overlap when m = n.)

http://dx.doi.org/10.4169/amer.math.monthly.123.10.1033
MSC: Primary 11A00

December 2016] NOTES 1033

 The interesting feature of this classification is the existence of the infinite discrete
family (ii) of solutions where commutativity survives. The family (ii) fits together to
form an infinite family of pairwise commuting functions Tm (x) := f 1/m (x) =  m1 x
for integers m ≥ 1. Moreover, these functions satisfy for all m, n ≥ 1 the further rela-
tions

Tm ◦ Tn (x) = Tn ◦ Tm (x) = Tmn (x) for all x ∈ R,

which are the same relations satisfied by composition of linear functions (2).
One can ask an analogous question for dilated ceiling functions: When do two
dilated ceiling functions commute? The resulting classification turns out to be iden-
tical. To see this, set gα (x) := αx. Using the identity (1), we deduce that for any
α, β,

f α ◦ f β (x) = −gα ◦ gβ (−x), for all x ∈ R.

Since x → −x is a bijection of the domain R to itself, we see that gα and gβ commute

under composition if and only if f α and f β commute under composition.
The commuting family (ii) was noted by Cardinal [3, Lemma 6] in a number-
theoretic context. He studied certain semigroups of integer matrices, constructed using
the floor function, from which he constructed a family of symmetric integer matrices
that he related to the Riemann hypothesis. Also from this number-theoretic perspec-
tive, symmetry properties of the solutions may be important. Both sets of solutions (i)
and (ii) are invariant under exchange (α, β) to (β, α). However:
(1) The set of all continuous solution parameters (i) is invariant under the reflection
symmetry taking (α, β) to (−α, −β), while the discrete solutions (ii) break this
symmetry.
(2) If one restricts to strictly nonzero parameters, then the continuous solution
parameters (i) are invariant under the symmetry taking (α, β) to ( α1 , β1 ), while
the discrete solutions (ii) break this symmetry.
In the next section, we prove Theorem 1, and in the final section, we discuss the
problem in the general context of digital straight lines.

2. PROOF OF THEOREM 1. Two immediate cases where commutativity holds are

when α = 0 or β = 0: In these cases, the functions f α and f β commute since their
composition is the zero function. In what follows, we suppose that αβ = 0, and then
we reparameterize the problem in terms of inverse parameters (1/α, 1/β), which will
simplify the resulting formulas.
We prove Theorem 1 by a case analysis that depends on the signs of α and β. The
proofs analyze the jump points in the graphs of f 1/α ◦ f 1/β (x). We define for real y the
upper level set at level y:

S1/α,1/β (y) := {x : f 1/α ◦ f 1/β (x) ≥ y} = ( f 1/α ◦ f 1/β )−1 [y, ∞).

The commutativity property asserts the equality S1/α,1/β (n) = S1/β,1/α (n) of upper
level sets for all n ∈ Z, and the converse holds because the range of f 1/α ◦ f 1/β is a
subset of Z. The key formulas are identities determining these upper level sets given
in Lemmas 1 and 4, leading to formulas characterizing commutativity when α, β > 0
and α, β < 0 given in Lemmas 2 and 5, respectively.

1034
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
Case 1. Both α and β are positive. We begin with a formula for the upper level sets
at integer points.

Lemma 1. For α, β > 0 and each n ∈ Z, the upper level set is

S1/α,1/β (n) = [βnα, ∞) .

Proof. We have the following implications:

  
1 1
x ∈ S1/α,1/β (n) ⇔ x ≥n (by definition)
α β
 
1 1
⇔ x ≥n (the right side is in Z)
α β
 
1
⇔ x ≥ nα (since α > 0)
β
 
1
⇔ x ≥ nα (the left side is in Z)
β
1
⇔ x ≥ nα (the right side is in Z)
β
⇔ x ≥ βnα (since β > 0).

Lemma 2. For α, β > 0, the function f 1/α commutes with f 1/β if and only if the
equality

βnα = αnβ (3)

holds for all integers n ∈ Z.

Proof. By Lemma 1, we have x ∈ S1/α,1/β (n) if and only if x ≥ βnα. Similarly,

x ∈ S1/β,1/α (n) if and only if x ≥ αnβ, so that commutativity of the functions is
equivalent to the desired equality of ceiling functions.

Lemma 3. For α, β > 0, the function f 1/α commutes with f 1/β if and only if either
α = β or if α and β are both positive integers.

Proof. If α = β, then commutativity clearly holds. If α, β are both (positive) integers,

then the relation (3) holds for all n ∈ Z since the ceiling functions have no effect.
Hence, commutativity holds.
The remaining case is that where at least one of α, β is not an integer; without loss
of generality, assume α is not an integer. We write α = A ≥ 1, with A > α, and
β = B ≥ 1. We show that commutativity occurs only if α = β.
Starting from Lemma 2, the relation (3) can be rewritten

α nα
= , (4)
β nβ

whenever the term nβ is nonvanishing; here, nβ ≥ 1 holds for n ≥ 1.

December 2016] NOTES 1035

 Since α < A, there exists a finite n ≥ 2 such that kα = k A for 1 ≤ k ≤ n − 1,
while nα = n A − 1. Now, (4) requires

α A kα
= = for all k ≥ 1.
β B kβ

By induction on k ≥ 1, this relation implies kβ = k B for 1 ≤ k ≤ n − 1. It also

implies that nβ = n B or n B − 1. The relation (4) for k = n becomes

α A nα nA − 1
= = = ,
β B nβ nβ

which rules out nβ = n B. Thus, nβ = n B − 1, and we now have

A nA − 1
= .
B nB − 1

Clearing denominators yields nAB − A = nAB − B, whence A = B. Thus, we have

α
β
= BA = 1, so that α = β as asserted.

Case 2. Both α and β are negative. We obtain a criterion that parallels Lemma 2 in
the positive case.

Lemma 4. For α, β < 0 and each n ∈ Z, the upper level set is

S1/α,1/β (n) = (βnα + β, ∞) .

Proof. We have the following implications:

  
1 1
x ∈ S1/α,1/β (n) ⇔ x ≥n (by definition)
α β
 
1 1
⇔ x ≥n (the right side is in Z)
α β
 
1
⇔ x ≤ nα (since α < 0)
β
 
1
⇔ x ≤ nα (the left side is in Z)
β
1
⇔ x < nα + 1 (the right side is in Z)
β
⇔ x > βnα + β (since β < 0).

Lemma 5. For α, β < 0, the function f 1/α (x) commutes with f 1/β (x) if and only if the
equality

βnα + β = αnβ + α

holds for all integers n ∈ Z.

1036
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
Proof. By Lemma 4, we have x ∈ S1/α,1/β (n) if and only if x > βnα + β. Similarly,
we have x ∈ S1/β,1/α (n) if and only if x > αnβ + α so that commutativity of the
functions is equivalent to the desired equality.

Lemma 6. For α, β < 0, the function f 1/α commutes with f 1/β if and only if α = β.

Proof. Choose n = 0 in Lemma 5. We obtain that α = β is a necessary condition for

commutativity. But this condition is obviously sufficient.

Case 3. α and β are of opposite signs.

Lemma 7. For (α, β) with αβ < 0, the function f 1/α (x) never commutes with f 1/β (x).

Proof. Without loss of generality, we may consider α > 0 and β < 0. It suffices to
show S1/α,1/β (n) = S1/β,1/α (n). We will see that both of these upper level sets start at
−∞ and have a finite right endpoint.
We first compute S1/α,1/β (n). We can follow the same steps as in Lemma 1, except
in the last step where we have instead that x ∈ S1/α,1/β (n) if and only if x ≤ βnα
since β < 0. We obtain for α > 0 and β < 0 that

S1/α,1/β (n) = (−∞, βnα]

is a closed interval.
Next, we compute S1/β,1/α (n). We can follow the same steps as in Lemma 4, except
in the last step where we have instead that x ∈ S1/β,1/α (n) if and only if x < αnβ + α
since α > 0. We find in this case that

S1/α,1/β (n) = (−∞, αnβ + α)

is an open interval. It follows that the two functions cannot commute.

The case analysis is complete, and Theorem 1 follows.

3. DIGITAL STRAIGHT LINES. The mathematical study of digital straight lines,

which are “lines” drawn on two-dimensional graphic displays represented by pixels,
was initiated by A. Rosenfeld [9] in 1974. For more recent work, see Klette and Rosen-
feld [7] and Kiselman [6]. In drawing a digital image of the line α,γ (x) := αx + γ , a
simple recipe is to associate to the abscissa n = x the pixel (x, αx + γ ) ∈ Z2
(more complicated recipes are used in practice). Bruckstein [2] noted self-similar fea-
tures of digital straight lines, relating them to the continued fraction expansion of their
slopes; see also McIlroy [8]. In contrast, our proof of Theorem 1 does not require
continued fractions.
From the digital straight line viewpoint, one can view αβx as a step func-
tion approximation to the straight line αβ (x) := αβx in the sense that the difference
function

h α,β (x) := αβx − αβx

is a bounded function. This difference function is explicitly given by a combination of

iterated fractional part functions h α,β (x) = −α{βx} − {α(βx − {βx})} so is a bounded

December 2016] NOTES 1037

generalized polynomial in the sense of Bergelson and Leibman [1]. The commuta-
tivity problem studied here is that of determining when the generalized polynomial
h α,β (x) − h β,α (x) is identically zero.
Commutativity questions under composition can be considered for general dig-
ital straight lines such as f α,γ (x) := αx + γ . However, general linear functions
α,γ (x) = αx + γ with distinct nonzero γ do not commute under composition. We
do not know whether any interesting new commuting pairs occur in this more general
context.

ACKNOWLEDGMENT. The authors thank S. Mori for a helpful comment. The first author received financial
support from NSF grant DMS-1401224.

REFERENCES

1. V. Bergelson, A. Leibman, Distribution of values of bounded generalized polynomials, Acta Math. 198
no. 2 (2007) 155–230, http://dx.doi.org/10.1007/s11511-007-0015-y.
2. A. M. Bruckstein, Self-similarity properties of digitized straight lines, in Vision Geometry (Hoboken, NJ,
1989), Contemp. Math., Vol. 119, Ed. by R. A. Metter, A. Rosenfeld, P. Bhatacharya, American Mathe-
matical Society, Providence, RI, 1991. 1–20, http://dx.doi.org/10.1090/conm/119/1113896.
3. J.-P. Cardinal, Symmetric matrices related to the Mertens function, Linear Algebra Appl. 432 no. 1 (2010)
161–172, http://dx.doi.org/10.1016/j.laa.2009.07.035.
4. R. L. Graham, D. E. Knuth, O. Patashnik, Concrete Mathematics: A Foundation for Computer Science.
Second ed. Addison-Wesley, Reading, MA, 1994.
5. K. E. Iverson, A Programming Language. John Wiley and Sons, New York, 1962.
6. C. O. Kiselman, Characterizing digital straightness and digital convexity by means of difference operators,
Mathematika 57 no. 2 (2011) 355–380, http://dx.doi.org/10.1112/S0025579311001318.
7. R. Klette, A. Rosenfeld, Digital straightness—A review, Discrete Appl. Math. 139 no. 1–3 (2004)
197–230, http://dx.doi.org/10.1016/j.dam.2002.12.001.
8. M. D. McIlroy, Number theory in computer graphics, in The Unreasonable Effectiveness of Number The-
ory, Proc. Sympos. Appl. Math., Vol. 46, Ed. by S. Burr, American Mathematical Society, Providence, RI,
1991. 105–121, http://dx.doi.org/10.1090/psapm/046/1195844.
9. A. Rosenfeld, Digital straight line segments, IEEE Trans. Computers C-23 no. 12 (1974) 1264–1269,
http://dx.doi.org/10.1109/t-c.1974.223845.

Dept. of Mathematics, University of Michigan, Ann Arbor, MI 48109–1043

lagarias@umich.edu
takumim@umich.edu
hrichman@umich.edu

1038
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 Quotients of Fibonacci Numbers
Stephan Ramon Garcia and Florian Luca

Abstract. There have been many articles in the M ONTHLY on quotient sets over the years.
We take a first step here into the p-adic setting, which we hope will spur further research. We
show that the set of quotients of nonzero Fibonacci numbers is dense in the p-adic numbers
for every prime p.

The nth Fibonacci number Fn is defined by the recurrence relation Fn+2 = Fn+1
+ Fn with initial conditions F0 = 0 and F1 = 1. Let F = {1, 2, 3, 5, 8, 13, 21, . . .}
denote the set of nonzero Fibonacci numbers and let R(F) = {Fm /Fn : m, n ∈ N} be
the set of all quotients of elements of F. Binet’s formula

1
Fn = √ (ϕ n − ϕ̃ n ), (1)
5

in which
√ √
1+ 5 1− 5
ϕ= = 1.618 . . . and ϕ̃ = = −0.618 . . . ,
2 2
√
implies that |Fn − ϕ n / 5| tends to zero exponentially fast as n → ∞ [37, p.57]. So
as a subset of R, R(F) accumulates only at the points ϕ k for k ∈ Z [8, Ex.17].
On the other hand, R(F) is a subset of the rational number system Q, which can be
endowed with metrics other than the one inherited from R. For each prime p, there is
a p-adic metric on Q, with respect to which Q can be completed to form the set Q p of
p-adic numbers.
Our aim here is to prove the following theorem.

Theorem 2. R(F) is dense in Q p for every prime p.

By this we mean that the closure of R(F) in Q p with respect to the p-adic metric is
Q p . Although there have been many papers in the M ONTHLY on quotient sets over the
years [1, 4, 8, 9, 15, 17, 26, 33], we are unaware of similar work being undertaken in
the p-adic setting. We hope that this result will spur further investigations.

Definition 3. If p is a prime number, then each r ∈ Q\{0} has a unique representation

a
r = ± pk , (4)
b
in which k ∈ Z, a, b ∈ N, and a, b, p are pairwise relatively prime. The p-adic val-
uation ν p : Q → Z is defined by ν p (0) = +∞ and, for r as in (4), by ν p (r ) = k.
It satisfies ν p (x y) = ν p (x) + ν p (y) for x, y ∈ Q. The p-adic absolute value on Q is
r  p = p −ν p (r ) and the p-adic metric on Q is d p (x, y) = x − y p .
http://dx.doi.org/10.4169/amer.math.monthly.123.10.1039
MSC: Primary 11B39, Secondary 11A41; 11S99

December 2016] NOTES 1039

 The p-adic number system Q p is the completion of Q with respect to the p-adic
metric. Ostrowski’s theorem asserts that Q p , for p prime, and R are essentially the
only completions of Q that arise from an absolute value [20]. Thus, it is natural to seek
results like Theorem 2, since similar questions in R have long since been explored.
In fact, Q p is a field that contains Q as a subfield. Its elements can be 
represented
as infinite series, convergent with respect to the p-adic metric, of the form ∞ n
n=k an p ,
in which k ∈ Z and an ∈ {0, 1, . . . , p − 1}. The p-adic valuation, norm, and metric
extend uniquely to Q p in a manner that respects these series representations.

Example 5. In Q2 we have 1 + 2 + 22 + 23 + 24 + · · · = −1 since

 N −1   
   1 − 2N     
    = 1 − (1 − 2 N ) = 2 N  = 2−N
 2n
− (−1)  =  + 1
  1−2 2
2 2
n=0 2

tends to zero as N → ∞.

Manipulating p-adic numbers is analogous to handling decimal expansions of real

numbers. Instead of powers 10n with n running from some nonnegative N to −∞, we
have powers pn with n running from N (potentially negative) to +∞. Further details
on Q p can be found in [11, 20].
To prove our theorem, we require a few preliminary results. In what follows, we
write a|b to denote that the integer b is divisible by the integer a. Proofs of the follow-
ing assertions can be found in [37, p.82, p.73, & p.81].

Lemma 6.
(a) If j|k, then F j |Fk .
(b) For each m ∈ N, there is a smallest index z(m) so that m|Fkz(m) for all positive
integers k.
(c) ν p (Fz( p) p j ) = ν p (Fz( p) ) + j for any odd prime p and j ∈ N.

We also need a convenient dense subset of Q p . We freely make use of the fact that
d p (x, y) ≤ 1/ pk if and only if ν p (x − y) ≥ k.

Lemma 7. For each prime p, the set { p−r n : n, r ∈ N} is dense in Q p .

Proof. Since Q p is the closure of Q with respect to the p-adic metric, it suffices to
show that each rational number can be arbitrarily well approximated, in the p-adic
metric, by rational numbers of the form p−r n with n, r ∈ N. If x = p k ∞ j=0 a j p ∈
j
 N −k−1
Q p , then let N ≥ k + 1 and n = j=0 a j p j so that

 
∞   
∞ 
ν p (x − p n) = ν p p
k k
aj p j
= k + νp a j p j ≥ k + (N − k) = N .
j=N −k j=N −k

An algebraic
√ integer is a root of a monic polynomial with integer coefficients. For
instance, 2, 5, and ϕ are algebraic integers. They are roots of x − 2, x 2 − 5, and
x 2 − x − 1, respectively. The following is [24, Cor.1, p.15].

Lemma 8. A rational algebraic integer is an integer.

1040
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 √
We also require a few facts about arithmetic in the field K = Q( 5). Let OK denote
the set of algebraic integers in K. It √ is a ring; in fact, OK = {a + bϕ : a, b ∈ Z} [24,
Cor.2, p.15]. In particular, ϕ, ϕ̃, and 5 each belong to OK .
An ideal in OK is a additive subgroup i of OK so that αi ⊆ i for all α ∈ OK . A
prime in OK refers to a prime ideal in the ring OK . A prime ideal is an ideal p ⊆ OK
with the property that, for any α, β ∈ OK , the condition αβ ∈ p implies that α ∈ p or
β ∈ p. To avoid confusion, we refer to the primes 2, 3, 5, 7, . . . as rational primes.
The product of two ideals i, j in OK is defined by ij = {αβ : α ∈ i, β ∈ j}; it is
also an ideal in OK . Positive powers of ideals are defined inductively by i1 = i and
in = i(in−1 ) for n = 2, 3, . . .. We say that i divides j if i ⊆ j.
For an ideal i in OK , we write α ≡ β (mod i) to mean that β − α ∈ i. The famil-
iar properties of congruences hold when working modulo an ideal. For instance,
α ≡ β (mod i) implies that α j ≡ β j (mod i) for j ∈ N. Similarly, if p and q are
distinct prime ideals, α ≡ β (mod p), and α ≡ β (mod q), then α ≡ β (mod pq).

Lemma 9. Let a, b ∈ Z and j ∈ N. If a ≡ b (mod p j ), then p j divides b − a in Z.

Proof. Suppose that a, b ∈ Z and a ≡ b (mod p j ). Then, b − a = p j α for some α ∈

OK . Since α = (b − a)/ p j is a rational algebraic integer, it is an integer by Lemma 8.
Consequently, p j divides b − a in Z.

Each rational prime p gives rise to an ideal p = pOK in OK This ideal √ factors
uniquely as a product of prime ideals [24, Th.16, p.59]. In fact, since Q( 5) is a
quadratic extension of Q we have [24, p.74]:
√
Lemma 10. Let K = Q( 5) and let p be a rational prime. Then p = pOK is the
product of at most two prime ideals (not necessarily distinct).

We can be a little more specific. The only rational prime p√for which p is a square
of a prime ideal is p = 5; this reflects the factorization 5 = ( 5)2 [24, Th.24, p.72].

Proof of Theorem 2. There are two cases, according to whether p is odd or p = 2. Let
p be an odd rational √ prime and let p = p OK . For each j ∈ N, Lemma 6 ensures that
p2 j |Fz( p) p2 j . Since 5 ∈ OK , (1) reveals that

2j 2j √
ϕ z( p) p − ϕ̃ z( p) p = 5Fz( p) p2 j ≡ 0 (mod p2 j ),

and hence
2j 2j
ϕ z( p) p ≡ ϕ̃ z( p) p (mod q2 j )

for each prime q in OK that divides p. Since ϕ̃ = −1/ϕ, we have

2j 2j
ϕ 2z( p) p ≡ ϕ −2z( p) p (mod q2 j ),

so that
2j 2j
ϕ 4z( p) p ≡ 1 (mod q2 j ) and ϕ̃ 4z( p) p ≡ 1 (mod q2 j ). (11)

December 2016] NOTES 1041

 2j 2j
Setting x = ϕ 4z( p) p and y = ϕ̃ 4z( p) p in the identity

x m − y m = (x − y)(x m−1 + x m−2 y + · · · + y m−1 ), m ∈ N,

implies that
2j 2j
F4z( p) p2 j m ϕ 4z( p) p m − ϕ̃ 4z( p) p m
=
F4z( p) p2 j ϕ 4z( p) p2 j − ϕ̃ 4z( p) p2 j
2j 2j 2j 2j
= (ϕ 4z( p) p )m−1 + (ϕ 4z( p) p )m−2 (ϕ̃ 4z( p) p ) + · · · + (ϕ̃ 4z( p) p )m−1
≡ m (mod q2 j )

by (11). Now Lemma 10 ensures that

F4z( p) p2 j m
≡ m (mod p j )
F4z( p) p2 j

for any m ∈ N. However, F4z( p) p2 j m /F4z( p) p2 j is a rational integer by Lemma 6, so

F4z( p) p2 j m
≡ m (mod p j ) (12)
F4z( p) p2 j

by Lemma 9. Appealing to Lemma 6, we have

ν p (F4z( p) p2(k+2r ) ) = ν p (F4z( p) p2(k+r ) ) + 2r, (13)

so
F4z( p) p2(k+2r )
= p2r ,  ∈ Z, gcd(, p) = 1. (14)
F4z( p) p2(k+r )

Set m = pr n and j = k + r in (12) and use (14) to conclude that

F4z( p) p2(k+r ) pr n F4z( p) p2(k+r ) pr n

p2r  · = ≡ pr n (mod p k+r ).
F4z( p) p2(k+2r ) F4z( p) p2(k+r )

Since gcd(, p) = 1, the preceding yields

F4z( p) p2k+3r n
pr ≡ n (mod p k ),
F4z( p) p2k+4r

and hence
 
F4z( p) p2k+3r n
νp − p−r n ≥ k.
F4z( p) p2k+4r

This holds for all n, r ∈ N and all k > r , so R(F) is dense in Q p by Lemma 7.
If p = 2 we must replace the exponents 4z( p) p j in (11) with 3 · 2 j+2 since z(2) = 3
and 4 = 22 . The proof can proceed as before if we replace (13) with the corresponding
statement for p = 2. It suffices to show that ν2 (F3·2 j ) = j + 2 for j ∈ N; see [22] for

1042
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
a more general result. We proceed by induction on j. The base case j = 1 is ν2 (F6 )
= ν2 (8) = 3. Now suppose that ν2 (F3·2 j ) = j + 2 for some j ≥ 2. Let L n denote the
nth Lucas number; these satisfy the recurrence L n+2 = L n+1 + L n with initial condi-
tions L 0 = 2 and L 1 = 1. Since L 2n = 2(−1)n + 5Fn2 [37, p.177] and since 2 = F3
divides F3·2 j−1 , we have
j−1

ν2 (L 3·2 j ) = ν(L 2(3·2 j−1 ) ) = ν2 2(−1)3·2 + 5F3·2

2
j−1 = 1.

Because F2n = Fn L n [37, p.177], we conclude that

ν2 (F3·2 j+1 ) = ν2 (F2(3·2 j ) ) = ν2 (F3·2 j ) + ν2 (L 3·2 j ) = ( j + 2) + 1 = ( j + 1) + 2,

which completes the induction.

ACKNOWLEDGMENT. This work was partially supported by National Science Foundation Grant DMS-
1265973. We thank the anonymous referees for numerous helpful suggestions.

REFERENCES

1. B. Brown, M. Dairyko, S. R. Garcia, B. Lutz, M. Someck, Four quotient set gems, Amer. Math. Monthly
121 no. 7 (2014) 590–599.
2. J. Bukor, P. Erdős, T. Šalát, J. T. Tóth, Remarks on the (R)-density of sets of numbers. II, Math. Slovaca
47 no. 5 (1997) 517–526.
3. J. Bukor, T. Šalát, J. T. Tóth, Remarks on R-density of sets of numbers, Tatra Mt. Math Publ. 11 (1997)
159–165.
4. J. Bukor, J. T. Tóth, On accumulation points of ratio sets of positive integers, Amer. Math. Monthly 103
no. 6 (1996) 502–504.
5. S. A. Burr, On moduli for which the Fibonacci sequence contains a complete system of residues,
Fibonacci Quart. 9 no. 9 (1971) 497–504.
6. J.-M. De Koninck, A. Mercier, 1001 Problems in Classical Number Theory. American Mathematical
Society, Providence, RI, 2007.
7. B. Fine, G. Rosenberger, Number Theory: An Introduction via the Distribution of Primes. Birkhäuser,
Boston, 2007.
8. S. R. Garcia, V. Selhorst-Jones, D. E. Poore, N. Simon, Quotient sets and Diophantine equations, Amer.
Math. Monthly 118 no. 8 (2011) 704–711.
9. S. R. Garcia, Quotients of Gaussian Primes, Amer. Math. Monthly 120 no. 9 (2013) 851–853.
10. R. Gelca, T. Andreescu, Putnam and Beyond. Springer, New York, 2007.
11. F. Q. Gouvêa, p-adic Numbers: An Introduction. Second ed. Springer-Verlag, Berlin, 1997.
12. R. Gupta, M. R. Murty, A remark on Artin’s conjecture, Invent. Math. 78 no. 1 (1984) 127–130.
13. G. H. Hardy, E. M. Wright, An Introduction to the Theory of Numbers. Sixth ed. Oxford Univ. Press,
Oxford, 2008.
14. G. Harman, Metric Number Theory. London Mathematical Society Monographs, Vol. 18. Clarendon
Press, New York, 1998.
15. S. Hedman, D. Rose, Light subsets of N with dense quotient sets, Amer. Math. Monthly 116 no. 7 (2009)
635–641.
16. D. R. Heath-Brown, Artin’s conjecture for primitive roots, Quart. J. Math. Oxford Ser. (2) 37 no. 145
(1986) 27–38.
17. D. Hobby, D. M. Silberger, Quotients of primes, Amer. Math. Monthly 100 no. 1 (1993) 50–52.
18. C. Hooley, On Artin’s conjecture, J. Reine Angew. Math. 225 (1967) 209–220.
19. E. T. Jacobson, Distribution of the Fibonacci numbers mod 2k , Fibonacci Quart. 30 no. 3 (1992)
211–215.
20. N. Koblitz, p-adic Numbers, p-adic Analysis, and Zeta-Functions. Second ed. Graduate Texts in Mathe-
matics. Springer-Verlag, New York, 1984.
21. T. Koshy, Fibonacci and Lucas Numbers with Applications. Pure and Applied Mathematics. Wiley-
Interscience, New York, 2001.
22. T. Lengyel, The order of the Fibonacci and Lucas numbers, Fibonacci Quart. 33 no. 3 (1995) 234–239.
23. F. Luca, C. Pomerance, S. Wagner, Fibonacci integers, J. Number Theory 131 no. 3 (2011) 440–457.

December 2016] NOTES 1043

24. D. A. Marcus, Number Fields. Springer-Verlag, New York, 1977.
25. I. Niven, H. S. Zuckerman, H. L. Montgomery, An Introduction to the Theory of Numbers. Fifth ed. John
Wiley & Sons, New York, 1991.
26. A. Nowicki, Editor’s endnotes, Amer. Math. Monthly 117 no. 8 (2010) 755–756.
27. P. Paullack, Not Always Buried Deep: A Second Course in Elementary Number Theory. American Math-
ematical Society, Providence, RI, 2009.
28. P. Ribenboim, Classical Theory of Algebraic Numbers. Springer-Verlag, New York, 2001.
29. ———, The Book of Prime Number Records. Second ed. Springer-Verlag, New York, 1989.
30. I. Steward, D. Tall, Algebraic Number Theory and Fermat’s Last Theorem. Third ed. AK Peters, Natick,
MA, 2002.
31. T. Šalát, On ratio sets of sets of natural numbers, Acta Arith. 15 (1968) 273–278.
32. ———, Corrigendum to the paper “On ratio sets of sets of natural numbers,” Acta Arith. 16 (1969) 103.
33. P. Starni, Answers to two questions concerning quotients of primes, Amer. Math. Monthly 102 no. 4
(1995) 347–349.
34. O. Strauch, J. T. Tóth, Asymptotic density of A ⊂ N and density of the ratio set R(A), Acta Arith. 87
(1998) 67–78.
35. ———, Corrigendum to Theorem 5 of the paper: “Asymptotic density of A ⊂ N and density of the ratio
set R(A),” Acta Arith. 87 no. 1  (1998) 67–78.
36. C. Vanden Eynden, Proofs that 1/ p diverges, Amer. Math. Monthly 87 no. 5 (1980) 394–397.
37. S. Vajda, Fibonacci & Lucas Numbers, and the Golden Section. Ellis Horwood Series: Mathematics and
its Applications. Theory and applications, with chapter XII by B. W. Conolly. Ellis Horwood, Chichester;
Halsted Press John Wiley & Sons, 1989.

Department of Mathematics, Pomona College, Claremont, CA 91711

stephan.garcia@pomona.edu
School of Mathematics, University of the Witwatersrand, Private Bag X3, Wits 2050, Johannesburg, South
Africa
Florian.Luca@wits.ac.za

1044
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 Parking Cars of Different Sizes
Richard Ehrenborg and Alex Happ

Abstract. We extend the notion of parking functions to parking sequences, which include cars
of different sizes, and prove a product formula for the number of such sequences.

1. THE RESULT. Parking functions were first introduced by Konheim and Weiss [5].
The original concept was that of a linear parking lot with n available spaces, and n cars
with a stated parking preference. Each car would, in order, attempt to park in its pre-
ferred spot. If the car found its preferred spot occupied, it would move to the next
available slot. A parking function is a sequence of parking preferences that would
allow all n cars to park according to this rule. This notion is equivalent to the follow-
ing formal definition.

Definition 1. Let a = (a1 , a2 , . . . , an ) be a sequence of positive integers, and let

b1 ≤ b2 ≤ · · · ≤ bn be the increasing rearrangement of a . Then the sequence a is
a parking function if and only if bi ≤ i for all indexes i.

It is well known that the number of such parking functions is (n + 1)n−1 . This
is Cayley’s formula for the number of labeled trees on n + 1 nodes and Foata and
Riordan found a bijective proof [3]. Stanley discovered the relationship between park-
ing functions and noncrossing partitions [10]. Further connections have been found
to other structures, such as priority queues [4], Gončarov polynomials [6], and hyper-
plane arrangements [11].
The notion of a parking function has been generalized in myriad ways; see the
sequence of papers [2, 6, 7, 8, 12]. We present here a different generalization, returning
to the original idea of parking cars. This time the cars have different sizes, and each
takes up a number of adjacent parking spaces.

Definition 2. Let there be n cars C1 , . . . , Cn of sizes y1 , . . . , yn , where y1 , . . . , yn are

n
positive integers. Assume there are i=1 yi spaces in a row. Furthermore, let car Ci
have the preferred spot ci . Now let the cars in the order C1 through Cn park according
to the following rule.

Starting at position ci , car Ci looks for the first empty spot j ≥ ci . If the spaces j through
j + yi − 1 are empty, then car Ci parks in these spots. If any of the spots j + 1 through
j + yi − 1 is already occupied, then there will be a collision, and the result is not a parking
sequence.

Iterate this rule for all the cars C1 , C2 , . . . , Cn . We call (c1 , . . . , cn ) a parking
sequence for y = (y1n, . . . , yn ) if all n cars can park without any collisions and
without leaving the i=1 yi parking spaces.

As an example, consider three cars of sizes y = (2, 2, 1) with preferences

c = (2, 3, 1). Then there are 2 + 2 + 1 = 5 available parking spaces, and the final
configuration of the cars is as follows.
http://dx.doi.org/10.4169/amer.math.monthly.123.10.1045
MSC: Primary 05A15

December 2016] NOTES 1045

 C3 C1 C2
1 2 3 4 5

All cars are able to park, so this yields a parking sequence.

There are two ways in which a sequence can fail to be a parking sequence. Either
a collision occurs, or a car passes the end of the parking lot. As an example, consider
three cars with y = (2, 2, 2) and preferences c = (3, 2, 1). Then we have 2 + 2 + 2
= 6 parking spots, and the first car parks in its desired spot.

C1
1 2 3 4 5 6

However, the second car prefers spot 2, and since spot 2 is open, he tries to take spots 2
and 3, but collides with C1 in the process. Hence, this is not a parking sequence.
If, instead, we had y = (2, 2, 2) and c = (2, 5, 5), then again the first two cars are
able to park with no difficulty.

C1 C2
1 2 3 4 5 6

But car C3 will pass by all the parking spots after his preferred spot without seeing an
empty spot. Hence, this also fails to be a parking sequence.
The classical notion of a parking function is obtained when all the cars have size 1;
that is, y = (1, 1, . . . , 1). Note in this case that there are no possible collisions.
In the classical case, any permutation of a parking function is again a parking func-
tion. This is not true for cars of larger size. As an example, note for y = (2, 2) that
c = (1, 2) is a parking sequence. However, the rearrangement c  = (2, 1) is not a
parking sequence. This shows that the notion of parking sequence differs from the
notion of parking function in the papers [2, 6, 7, 8, 12].
The classical result is that the number of parking functions is given by (n + 1)n−1 ;
see [5]. For cars of bigger sizes we have the following result.

Theorem 3. The number of parking sequences f (y ) for car sizes y = (y1 , . . . , yn ) is
given by the product

f (y ) = (y1 + n) · (y1 + y2 + n − 1) · · · (y1 + · · · + yn−1 + 2).

2. CIRCULAR PARKING ARRANGEMENTS. Consider M = y1 + y2 + · · · +

yn + 1 parking spaces arranged in a circle. We will consider parking cars on this circu-
lar arrangement, without a cliff for cars to fall off. Observe that when all the cars have
parked, there will be one empty spot left over. We claim that there are

M · f (y ) = (y1 + n) · (y1 + y2 + n − 1) · · · (y1 + · · · + yn + 1) (2.1)

such circular parking sequences. The first car C1 has M ways to choose its parking
spot.
The next step is counterintuitive. After car C1 has parked, erase the markings for
the remaining y2 + · · · + yn + 1 spots and put in n + 1 dividers. These dividers cre-
ate n + 1 intervals on the circle, where one interval is taken up by C1 . Furthermore,

1046 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
these dividers are on wheels and can freely move along the circle. Each interval
will accept one (and only one) car. For example, consider the case where n = 5 and
y = (2, 5, 1, 3, 2) so that M = 2 + 5 + 1 + 2 + 3 + 1 = 14, and c1 = 5.

14 1 2
13 3
12 4

11 5

C1
C1
10 6
9 8 7

We will now create a circular parking sequence, but only at the end do we obtain
the exact positions of cars C2 through Cn+1 . That is, instead of focusing on the number
of specific spot preferences each car could have, we keep track of the order the cars
park in, which will then determine the exact locations of the cars.
The second car has two options. The first is that it has a desired position already
taken by C1 . In this case, it will cruise until the next empty spot. This can happen
in y1 ways, and then car C2 obtains the next open interval after the interval C1 is in.
Otherwise, the car C2 has a preferred spot not already taken. In this case C2 has n open
intervals to choose from. The total number of options for C2 is y1 + n.
The third car C3 has the same options. First, it may desire a spot that is already
taken, in which case it will have to cruise until the next open interval. This can happen
in y1 + y2 ways. Note that this count applies to both the case when C1 and C2 are
parked next to each other, and when C1 and C2 have open intervals between them.
Otherwise, C3 has n − 1 open intervals to pick from.
In general, car Ci has y1 + · · · + yi−1 + n + 2 − i choices. This pattern continues
up to Cn , which has y1 + · · · + yn−1 + 2 possibilities. For example, suppose C2 and C3
in our above example have parked as below.
C3
C2

C1

Then C4 may either cruise on C1 and C3 (in y1 + y3 ways), it may cruise on C2 (in y2
ways), or it can pick one of the three available intervals directly. In total, C4 has
(y1 + y3 ) + y2 + 3 = 11 ways to park.
One can imagine that when we park a car, we do not set the parking brake, but put
the car in neutral, so that the car and the dividers can move as necessary to make room
for future cars.

December 2016] NOTES 1047

 Thus the total number of circular parking arrangements of this type is

M · (y1 + n) · (y1 + y2 + n − 1) · · · (y1 + · · · + yn−1 + 2),

where the ith factor is the number of options for the car Ci . This proves the claim
about the number of circular parking sequences in (2.1).
Hence, to prove Theorem 3 we need only observe that the circular parking
sequences with spot M empty are the same as our parking sequences. This follows
from the observation that no car in the circular arrangement has preference M, since
otherwise this spot would not be empty. Furthermore, no car would cruise by this
empty spot.
Observe that the set of circular parking sequences is invariant under rotation.
That is, if (c1 , c2 , . . . , cn ) is a parking sequence, then so is the sequence (c1 + a, c2
+ a, . . . , cn + a), where all the additions are modulo M. In particular, the number of
circular parking sequences with spot M empty is given by 1/M · M · f (y ) = f (y ).

3. CONCLUDING REMARKS. The idea of considering a circular arrangement

goes back to Pollak; see [9]. In fact, when all the cars have size 1, this argument
reduces to his argument that the number of classical parking functions is (n + 1)n−1 .
The idea of not using fixed coordinates when placing cars in the circular arrange-
ment is reminiscent of the argument Athanasiadis used to compute the characteristic
polynomial of the Shi arrangement [1].

ACKNOWLEDGMENT. The authors thank two referees for their comments as well as Margaret Readdy
for her comments on an earlier draft of this note. Both authors were partially supported by National Security
Agency grant H98230-13-1-0280. The first author wishes to thank the Mathematics Department of Princeton
University where this work was carried out.

REFERENCES

1. C. A. Athanasiadis, Characteristic polynomials of subspace arrangements and finite fields, Adv. Math.
122 (1996), 193–233.
2. D. Chebikin, A. Postnikov, Generalized parking functions, descent numbers, and chain polytopes of
ribbon posets, Adv. Appl. Math. 44 (2010), 145–154.
3. A. D. Foata, J. Riordan, Mappings of acyclic and parking functions, Aequationes Math. 10 (1974), 10–22.
4. J. D. Gilbey, L. H. Kalikow, Parking functions, valet functions, and priority queues, Discrete Math. 197–
198 (1999), 351–373.
5. A. G. Konheim, B. Weiss, An occupancy discipline and applications, SIAM J. Appl. Math. 14 (1966),
1266–1274.
6. J. P. S. Kung, C. Yan, Gončarov polynomials and parking functions, J. Combin. Theory Ser. A 102 (2003),
16–37.
7. ———, Exact formulas for moments of sums of classical parking functions, Adv. Appl. Math. 31 (2003),
215–241.
8. ———, Expected sums of general parking functions, Ann. Comb. 7 (2003), 481–493.
9. J. Riordan, Ballots and trees, J. Comb. Theory 6 (1969), 408–411.
10. R. P. Stanley, Parking functions and noncrossing partitions, Electron. J. Combin. 4 (1997), Research
Paper 20, 14 pp.
11. ———, Hyperplane arrangements, parking functions and tree inversions, Mathematical essays in honor
of Gian-Carlo Rota. Eds. B. E. Sagan, R. P. Stanley. Birkhäuser, Boston, 1998. 359–375.
12. C. Yan, Generalized parking functions, tree inversions, and multicolored graphs, Special issue in honor
of Dominique Foata’s 65th birthday Adv. Appl. Math. 27 (2001) 641–670.

Department of Mathematics, University of Kentucky, Lexington, KY 40506

richard.ehrenborg@uky.edu, alex.happ@uky.edu

1048 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 Corrigendum to “A New Proof of the Change of Variable Theorem
for the Riemann Integral”
Paragraph 3 of [1, Proof of Theorem 1] should be as follows. Suppose that (g ◦ F) f
is integrable on [a, b]. Since F  = f = 0 a.e. on B, it can be shown using properties
(i)–(iii) that F(B) is of measure 0 (so we do not need to assume property (v) for the
proof of the theorem). By the lemma, g is integrable on F(Ai ), hence continuous
a.e. on F(Ai ) for all i. Therefore, g is continuous a.e. on F(∪i=1 ∞ A ) = ∪∞ F(A ),
i i=1 i
∞
thus continuous a.e., hence integrable, on F([a, b]) = F(∪i=1 Ai ) ∪ F(B).
Paragraph 4 of [1, Proof of Theorem 1] should be as follows. Suppose that
(g ◦ F) f and g are both integrable on [a, b] and F([a, b]), respectively. For
 x f n := f on ∪i=1 Ai and f n := 0 on [a, b] \ ∪i=1 Ai . Define Fn as
each n, let n n

Fn (x) := a f n (t) dt + F(a), for x ∈ [a, b]. For i ∈ {1, . . . , n}, let (u i , vi ) := Ai
and {[s j , t j ]}kj=1
n
be such that ∪kj=1
n
[s j , t j ] := [a, b] \ ∪i=1
n (u , v ). Since F is Lip-
i i n

 vi  Fn (vi )
schitz and Fn = f n a.e., it follows from the lemma that u i (g ◦ Fn ) f n = Fn (u i ) g.
t  F (t )
Since f n = 0 on [s j , t j ], Fn (s j ) = Fn (t j ), and thus s jj (g ◦ Fn ) f n = 0 = Fnn(s jj) g.
By the additivity of integrals over subintervals, (g ◦ Fn ) f n and g are integrable on
b  F (b)
[a, b] and F([a, b]), respectively, and a (g ◦ Fn ) f n = Fnn(a) g. Without loss of
generality, assume f is 0 off ∪i=1 ∞ A (observe that the functions f and (g ◦ F) f ,
i 0 0
∞
where f 0 = f on ∪i=1 Ai and f 0 = 0 on [a, b] \ ∪i=1 ∞ A , are both integrable
i
on [a, b]). Since f n → f , it follows from properties (ii) and (iii) that Fn (x)
x x
= a f n + F(a) → a f + F(a) = F(x), for x ∈ [a, b]. By the continuity of
b  Fn(b)  Fn(b)  F(b)
definite integrals, a (g ◦ Fn ) f n = Fn (a) g = F(a) g → F(a) g. By property
b b
(ii), since (g ◦ Fn ) f n → (g ◦ F) f , a (g ◦ Fn ) f n → a (g ◦ F) f , and so, by the
b  F(b)
uniqueness of limit, a (g ◦ F) f = F(a) g.

REFERENCE

1. H. Tandra, A new proof of the change of variable theorem for the Riemann integral, Amer. Math.
Monthly 122 (2015) 795–799.

http://dx.doi.org/10.4169/amer.math.monthly.123.10.1049
MSC: Primary 26A42

December 2016] NOTES 1049

 PROBLEMS AND SOLUTIONS
Edited by Gerald A. Edgar, Doug Hensley, Douglas B. West
with the collaboration of Itshak Borosh, Paul Bracken, Ezra A. Brown, Randall
Dougherty, Tamás Erdélyi, Zachary Franco, Christian Friesen, Ira M. Gessel, László
Lipták, Frederick W. Luttmann, Vania Mascioni, Frank B. Miles, Steven J. Miller,
Mohamed Omar, Richard Pfiefer, Dave Renfro, Cecil C. Rousseau, Leonard Smiley,
Kenneth Stolarsky, Richard Stong, Walter Stromquist, Daniel Ullman, Charles Vanden
Eynden, and Fuzhen Zhang.

Proposed problems should be submitted online via www.americanmathematical

monthly.submittable.com. Proposed solutions to the problems below should be sub-
mitted on or before April 30, 2017 at the same link. More detailed instructions are
available online. Solutions to problems numbered 11921 or lower should continue
to be submitted via email to monthlyproblems@math.tamu.edu. Proposed problems
must not be under consideration concurrently to any other journal and must not be
posted to the internet before the deadline date for solutions. An asterisk (*) after the
number of a problem or a part of a problem indicates that no solution is currently
available.

PROBLEMS
11943. Proposed by Keith Kearnes, University of Colorado, Boulder, CO, and Greg
Oman, University of Colorado, Colorado Springs, CO. Let X be a set, and let F be
a collection of functions f from X into X . A subset Y of X is closed under F if
f (y) ∈ Y for all y ∈ Y and f in F . With the axiom of choice given, prove or disprove:
There exists an uncountable collection F of functions mapping Z+ into Z+ such that
(a) every proper subset of Z+ that is closed under F is finite, and
(b) for every f ∈ F , there is a proper infinite subset Y of Z+ that is closed under
F \{ f }.
11944. Proposed by Yury Ionin, Central Michigan University, Mount Pleasant, MI. Let
n be a positive integer, and let [n] = {1, . . . , n}. For i ∈ [n], let Ai , Bi , Ci be disjoint
sets such that Ai ∪ Bi ∪ Ci = [n] − {i} and |Ai | = |Bi |. Suppose also that
|Ai ∩ B j | + |Bi ∩ C j | + |Ci ∩ A j | = |Bi ∩ A j | + |Ci ∩ B j | + |Ai ∩ C j |
for i, j ∈ [n]. Prove that i ∈ A j if and only if j ∈ Ai and, likewise, for the Bs and Cs.
11945. Proposed by Martin Lukarevski, University “Goce Delcev,” Stip, Macedonia.
Let a, b, and c be the lengths of the sides of triangle ABC opposite A, B, and C,
respectively, and let wa , wb , wc be the lengths of the corresponding angle bisectors.
Prove
a b c √
+ + ≥ 2 3.
wa wb wc

11946. Proposed by Moubinool Omarjee, Lycée Henri IV, Paris, France. Let f be
a twice differentiable function from [0, 1] to R with f  continuous on [0, 1] and
2/3
1/3
f (x) d x = 0. Prove
http://dx.doi.org/10.4169/amer.math.monthly.123.10.1050

1050 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
  1 2  1
4860 f (x) d x ≤ 11 f  (x) d x.
0 0

11947. Proposed by George Stoica, University of New Brunswick, Saint John, Canada.
Let n be a positive integer, and let z 1 , . . . , z n be the zeros in C of z n + 1. For a > 0,
prove

1
n
1 1 + a 2 + · · · + a 2(n−1)
= .
n k=1 |z k − a|2 (1 + a n )2

11948. Proposed by Navid Safaei, Sharif University of Technology, Tehran, Iran. Find
all surjective functions f : R → R+ such that (1) f (x) ≤ x + 1 for f (x) ≥ 1, (2)
f (x) = 1 for x = 0, and (3) for x, y ∈ R,
f (x f (y) + y f (x) − x y) = f (x) f (y).

11949. Proposed by Eugen J. Ionascu, Columbus State University, Columbus, GA.

Show that there exists a unique function f from R to R such that f is differentiable,
2 cos(x + f (x)) − cos x = 1 for all real x, and f (π/2) = −π/6.

SOLUTIONS

Flett’s Mean Value Theorem

11814 [2015, 76]. Proposed by Cezar Lupu, University of Pittsburgh, Pittsburgh, PA.
Let φ be a continuously differentiable function from [0, 1] into R, with φ(0) = 0 and
φ(1) = 1, and suppose that φ  (x) = 0 for 0 ≤ x ≤ 1. Let f be a continuous function
1 1
from [0, 1] into R such that 0 f (x) d x = 0 φ(x) f (x) d x. Show that there exists t
t
with 0 < t < 1 such that 0 φ(x) f (x) d x = 0.
Solution by New York Math Circle, NY. Define
 s  s
h(s) = φ(x) f (x) d x − φ(s) f (x) d x.
0 0

Note that h(0) = 0 = h(1). From Rolle’s theorem, we obtain h  (c) = 0 for some c ∈
(0, 1). Also, we compute
 s
 
h (s) = −φ (s) f (x) d x
0
 
and, in particular, h (0) = 0. Since φ (s) = 0 for all s ∈ (0, 1), the inverse function
φ −1 (s) exists and is differentiable on (0, 1). Letting H (s) = h(φ −1 (s)), we see that
h  (φ −1 (s))
H  (s) = .
φ  (φ −1 (s))
Applying Flett’s mean value theorem [Math. Gaz. 42 (1958) 38–39] to the function H
on the interval [0, φ(c)], we have
H (T ) − H (0)
= H  (T )
T −0

December 2016] PROBLEMS AND SOLUTIONS 1051

for some T ∈ (0, φ(c)) ⊂ (0, 1). Setting t = φ −1 (T ) ∈ (0, 1), this becomes
 t  t  t
1 h(t) h  (t)
φ(x) f (x) d x − f (x) d x = =  =− f (x) d x.
φ(t) 0 0 φ(t) φ (t) 0
t
Thus 0 φ(x) f (x) d x = 0, as desired.
Also solved by K. F. Andersen (Canada), R. Bagby, M. W. Botsko, B. S. Burdick, R. Chapman (U. K.),
P. P. Dályay (Hungary), M. Dan-Ştefan & F. Cătălin-Emil & O. Alexandru (Romania), L. Giugiuc (Romania),
D. Hancock, E. A. Herman, O. Kouba (Syria), J. K. Lindsey II, O. P. Lossers (Netherlands), A. Mingarelli &
J. M. Pacheco & Á. Plaza (Spain), P. Perfetti (Italy), I. Pinelis, D. Ritter, B. Schmuland (Canada), R. Stong, R.
Tauraso (Italy), T. P. Turiel, T. Wiandt, M. Wildon (U. K.), and the proposer.

Pascal’s Theorem
11816 [2015, 76]. Proposed by Sabin Tabirca, University College Cork, Cork, Ireland.
Let ABC be an acute triangle, and let B1 and C1 be the points where the altitudes from
B and C intersect the circumcircle. Let X be a point on arc BC, and let B2 and C2
denote the intersections of XB1 with AC and XC1 with AB, respectively. Prove that the
line B2 C2 contains the orthocenter of ABC.
Solution by Adnan Ali, A.E.C.S.-4, Mumbai, India. The claim holds not only for the cir-
cumcircle of ABC but also for any circumconic of the triangle—i.e. a conic circum-
scribing the triangle—as this problem is a special case of Pascal’s theorem, according
to which, if ABCDEF is a hexagon with vertices on a conic, then the intersections of
lines AB with ED, AF with CD, and EF with CB are collinear. (A point at infinity is
allowed.)
Also solved by M. Atasever (Turkey), M. Bataille (France), B. S. Burdick, J. Cade, R. B. Campos (Spain), R.
Chapman (U. K.), P. P. Dályay (Hungary), M. Dan-Ştefan & O. Alexandru & F. Cătălin-Emil (Romania), P.
De (India), O. Faynshteyn, D. Fleischman, O. Geupel (Germany), M. Goldenberg & M. Kaplan, J.-P. Grivaux
(France), J. G. Heuver (Canada), S. Hong (Korea), E. J. Ionaşcu, Y. J. Ionin, I. M. Isaacs, O. Kouba (Syria),
G. Lord, O. P. Lossers (Netherlands), J. Minkus, M. A. Shayib, N. Stanciu & T. Zvonaru (Romania), R. Stong,
T. Viteam (India), Z. Vörös (Hungary), T. Wiandt, GCHQ Problem Solving Group (U. K.), Missouri State
University Problem Solving Group, and the proposer.

Cycle Covers for Infinite Complete Graphs

11817 [2015, 175]. Proposed by Mohammed Jahaveri, Siena College, Loudonville, NY.
A cycle double cover of a graph is a collection of cycles that, counting multiplicity,
includes every edge exactly twice. Let X be an infinite set and let K X be the complete
graph on X . Construct a cycle double cover for X .
Solution I by I. M. Isaacs, Berkeley, CA. We construct a set of triangles covering
each edge exactly once. Taking each triangle twice produces a cycle double cover.
We may replace X by another set of the same cardinality, so we consider the set S of
all nonempty finite subsets of X . For each edge AB in K S , let C = A B (the sym-
metric difference), and use the triangle with vertices A, B, C. Since B C = A and
C A = B, each edge lies in exactly one such triangle.
Solution II by Jerrold Grossman and László Lipták, Oakland University, Rochester,
MI. We partition the edges into triangles. Use the axiom of choice to well order the
edges of K X using an ordinal that is also a cardinal. Transfinitely perform the fol-
lowing operation as long as there remains an edge not yet covered: For the least such
uncovered edge uv, choose a vertex w not yet in any triangle, and add uvw to the set
of triangles. Such a vertex w exists because the cardinality of the set of vertices used
so far in the process is less than the cardinality of X .

1052 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
Editorial comment. The method of Grossman and Lipták can be used to partition edges
of K X into copies of G for any finite graph G.
Also solved by E. Bojaxhiu (Albania) & E. Hysnelaj (Australia), B. Burdick, R. Chapman (U. K.), B. Karaivanov
& T. Vassilev (Canada), R. Stong, FAU Problem Solving Group, and the proposer.

Some Trig for the Nagel Cevians

11818 [2015, 175]. Proposed by Oleh Faynshteyn, Leipzig, Germany. Let ABC be a tri-
angle and let A1 , B1 , and C1 be the points on sides opposite A, B, and C respectively
at which the ecircles of the triangle are tangent to those sides. Let R and r be the cir-
cumradius and inradius of the triangle. Let the name of a vertex of ABC or of A1 B1 C1
also stand for the radian measure of the corresponding angle. Prove that, wherever the
expression is defined,
cot A1 + cot(A/2) cot B1 + cot(B/2) cot C1 + cot(C/2) 6R
+ + = .
cot A cot B cot C r

Solution by P. Nüesch, Switzerland. In fact, more is true: Each term on the left side
of the identity equals 2R/r . Write a, b, c for the side lengths of ABC, s for the
semiperimeter, and F for the area. Write u, v, w for the side lengths of A1 B1 C1 and
F1 for the area. Now
F 2R
= . (1)
F1 r
By the law of cosines,
cot(A/2) 1 4s(s − a)
= +1 = 2 . (2)
cot A cos A b + c2 − a 2
The modified cosine laws
b2 + c2 − a 2 v 2 + w2 − u 2
cot A = , and cot A1 =
4F 4F1
together with (1) give us
cot A1 2R v 2 + w2 − u 2
= . (3)
cot A r b2 + c2 − a 2
Now we have to prove (3) + (2) = 2R/r , or equivalently,
2R  2 
(b − v 2 ) + (c2 − w2 ) − (a 2 − u 2 ) = 4s(s − a).
r
Observe that b2 − v 2 = 2(s − c)(s − a)(1 + cos B) = 4(s − c)(s − a) cos2 (B/2) =
bF/R and similarly for the other two sides. Therefore, as required,

2R bF cF aF 2R
+ − = [b + c − a] = 2s · 2(s − a) = 4s(s − a).
r R R R r

Editorial comment. Lines AA1 , BB1 , and CC1 are the Nagel cevians of the triangle.
Also solved by A. Alt, R. Bagby, R. Chapman (U. K.), H. Y. Far, M. E. Kuczma (Poland), J. C. Smith, R.
Stong, H. Widmer, and the proposer.

December 2016] PROBLEMS AND SOLUTIONS 1053

 Twin Hölders
11819 [2015, 175]. Proposed by Cezar Lupu, University of Pittsburgh, Pittsburgh, PA.
Let f be a continuous, nonnegative function on [0, 1]. Show that
 1  1   1 
f (x) d x ≥ 4
3
x f (x) d x
2
x f (x) d x .
2
0 0 0

Solution by Radouan Boukharfane, Université du Poitiers, Chasseneuil, France. We

apply Hölder’s inequality twice
 1  1 2/3  1 1/3
x 2 f (x) d x ≤ x3 dx f 3 (x) d x
0 0 0
 1  1 1/3  1 2/3
x f 2 (x) d x ≤ x3 dx f 3 (x) d x .
0 0 0

Now multiply the inequalities

 1  1  1   1   1 
1
x f (x) d x x f (x) d x ≤
2 2
x dx3
f (x) d x =
3
f (x) d x .
3
0 0 0 0 4 0

Editorial comment. Several solvers proved generalizations. For example,

 the argument
above,
 b using the conjugate
 a+b exponents (a
 a+b + b)/a and (a + b)/b, yields f a
(x)g b
(x) d x
f (x)g (x) d x ≤ f (x) d x g (x) d x.
a

Also solved by R. A. Agnew, A. Alt, T. Amdeberhan & V. H. Moll, K. F. Andersen (Canada), R. Bagby,
M. Bataille (France), P. Bracken, M. A. Carlton, R. Chapman (U. K.), H. Chen, L. V. P. Cuong (Vietnam),
P. J. Fitzsimmons, W. R. Green, N. Grivaux (France), E. A. Herman, B. Karaivanov (USA) & T. S. Vazzilev
(Canada), O. Kouba (Syria), M. E. Kuczma (Poland), K.-W. Lau (China), J. H. Lindsey II, P. W. Lindstrom,
M. Omarjee (France), X. Oudot (France), P. Perfetti (Italy), Á. Plaza & F. Perdomo (Spain), K. Schilling, J. G.
Simmonds, J. C. Smith, A. Stenger, R. Stong, R. Tauraso (Italy), J. Vinuesa (Spain), H. Wang & J. Wojdylo,
G. White, Q. Zhang (China), Z. Zhang (China), NSA Problems Group, and the proposer.

Noetherian Subrings
11820 [2015, 175]. Proposed by Alborz Azarang, Shahid Chamran University of
Ahvaz, Ahvaz, Iran. Let K be a field and let R be a subring of K [X ] that contains K .
Prove that R is noetherian, that is, that every ascending chain of ideals in R terminates.
Solution by the National Security Agency Problems Group, Fort Meade, MD. Since a
finitely generated K -algebra is a quotient of K [x1 , . . . , xn ] for some n and, hence, is
noetherian, it suffices to show that R is finitely generated as a K -algebra. We use a
lemma of independent interest.
Lemma. Any set S of nonnegative integers that is closed under addition is finitely
generated: Thatis, there are elements d1 , . . . , dn ∈ S such that every s ∈ S can be
written as s = nk=1 ek dk for some nonnegative integers e1 , . . . , en .
Proof. This is clear if S is empty or equals {0}. Otherwise, let n be the least positive
integer in S. For 1 ≤ i < n, let di be the least element of S congruent to i (modulo
n), or di = 0 if S has no such element. We claim that {d1 , . . . , dn−1 , n} generates S.
If s ∈ S, then s ≡ di mod n for some i. Also, s ≥ di . Hence, s = di + kn for some
nonnegative integer k.
Now let R be a subring of K [X ] that contains K . Let S be the degrees of the
elements of R; note that S is closed under addition. By the lemma, there are inte-
gers d1 , . . . , dn that generate S. For a ≤ i ≤ n, let f i be a monic polynomial in

1054 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
R of degree di . Using induction on the degree m of a polynomial in R, we prove
R = K [ f 1 , . . . , f n ]. For m = 0, note that the constant polynomials are in R. For
m > 1, let a be the leading  coefficient of a polynomial f in R. There are integers
e e
ei ≥ 0 such that m = nk=1 ek dk . Let g = f − a f 1 1 f 2 2 · · · f nen . Note that g is in R and
has degree less than m. By the induction hypothesis, g ∈ K [ f 1 , . . . , f n ]. Hence, also,
f ∈ K [ f 1 , . . . , f n ], as desired.
Editorial comment. Various solvers used theorems from commutative algebra such as
the Eakin–Nagata theorem, the Artin–Tate theorem, and the Hilbert basis theorem, as
well as the chicken McNugget theorem, which is also known as the Frobenius coin
problem from number theory.
Also solved by A. J. Bevelacqua, T. Borislav (Canada) & V. Karaivanov, N. Caro (Brazil), R. Chapman (U. K.),
I. M. Isaacs, J. H. Lindsey II, F. Perdomo & A. Francisco (Spain), J. C. Smith, R. Stong, D. Ware, and the
proposer.

When a Composition of Polynomials Is Real

11822 [2015, 176]. Proposed by George Stoica, University of New Brunswick, Saint
John, Canada. Call a polynomial real if all its coefficients are real. Let P and Q be
polynomials with complex coefficients such that the composition P ◦ Q is real. Show
that if the leading coefficient of Q and its constant term are both real then P and Q are
real.
Solution by Borislav Karaivanov, Sigma Space, Lanham, MD & Tzvetalin Vassilev,
Nipissing University, North
 p q Bay, iON, Canada. Let P and Q be defined as P(x) =
a
i=0 i x i
and Q(x) = i=0 bi x with bq and b0 real. Since both bq and the coeffi-
cient a p bqp of x pq in P(Q(x)) are real, it can be concluded that a p is real.
We first claim that Q is real. If not, then let k be the largest index for which bk is
not real. The coefficient of x q( p−1)+k in P(Q(x)) is real and has the form a p bqp−1 bk +
M, where M is a polynomial expression in a p and bi , for k < i ≤ q, with integer
coefficients, and thus real. Hence, bk must be real. This contradiction shows that Q is
real.
Next, we claim that P is real. If not, then let k be the largest index for which ak is not
real. Consider the coefficient of x qk in P(Q(x)). By the premise of the problem, it is
real. On the other hand, it has the form ak bqk + N , where N is a polynomial expression
in the real variables ak+1 , . . . , a p and b0 , . . . , bq with integer coefficients . Therefore,
there is no such ak , and P is real.
Also solved by B. Bekker (Russia) & Y. J. Ionin (USA), A. J. Bevelacqua, N. Caro (Brazil), R. Chapman
(U.K.), P. De, B. Sury (India) & N. V. Tejaswi (Netherlands), D. Fleischman, J.-P. Grivaux (France), E. A.
Herman, P. W. Lindstrom, R. Stong, R. Tauraso (Italy), N. Thornber, J. Van Hamme (Belgium), T. Viteam
(Japan), and the proposer.

Inversion in a Circle?
11823 [2015, 176]. Proposed by Sabin Tabirca, University College Cork, Cork,
Ireland. Let P be a point inside a circle C.
(a) Prove that there exists a point P  outside C such that, for all chords XY of C
through P, (|XP | + |YP |)/|XY| is the same. (Here, |UV| denotes the distance from U
to V .)
(b) Is P  unique?
Solution by Ahmad Habil, Damascus University, Damascus, Syria. Let O denote the
center of C, r the radius of C, and p the distance of P from O. We must exclude the

December 2016] PROBLEMS AND SOLUTIONS 1055

case P = O (i.e., p = 0) because in that case there is no suitable P  at all. We (1)
show this then (2) show that, for any other P inside C, the image of P under inversion
in C can serve as P  and then finally (3) show that no other point can play the role of
P  (so in answer to (b), P  is unique).
(1) Suppose P = O, and let Q be any “trial” point outside C. We will demon-
strate that Q cannot meet the requirement for P  by finding two chords X 1 Y2 and X 2 Y2
through P for which the ratios (|X j Q| + |Y j Q|)/|X j Y j | are unequal. Let q = |OQ|.
Let X 1 , Y1 be the diameter lying along O Q and let X 2 , Y2 be the perpendicular
 diam-
eter. Now (|X 1 Q| + |Y1 Q|)/|X 1 Y1 | = qr , but (|X 2 Q| + |Y2 Q|)/|X 2 Y2 | = r 2 + q 2 /r
> qr .
(2) Consider P = O, so p = 0. Let P  be the image of P under inversion in C.
This is the point on ray OP such that pp  = r 2 , where p  = |OP |. When T is any
point on C (but not on line OP), triangles TOP and P  O T are similar, since they have

a common angle at O and |TP|/|OP| = |OP |/|OT| or rp = pr from the definition of
p  . Thus, |TP |/|OT| = |TP|/|OP|, or |TP | = rp |TP|.
Now let XY denote any chord through P. We have both |XP | + rp |XP| and |YP | =
r
p
|YP|. Adding, |XP | + |YP | = rp (|XP| + |YP|). Since XPY is a straight line, |XP| +
|YP| = |XY|. We conclude (|XP | + |YP |)/|XY| = rp .

Note that the constant value of (|XP | + |YP |)/|XY| must be pr , which equals rp and
is the cosecant of half the angle intercepted by C at P  .
(3) Now let Q  be a point outside C such that (|XQ | + |YQ |)/|XY| is constant for
all chords XY containing P. We must show that Q  is the inversion image of P in C,
that is, the point P  from part (2). Let Q be the image of Q  under inversion in C.
Thus, qq  = r 2 , where q = |QO| and q  = |Q  O|. Inversion is self-dual, so our claim
is that Q = P.
Applying part (2) starting with Q, we obtain that (|XQ | + |YQ |)/|XY| is constant
for all chords through Q. Consider the diameter X 1 Y1 containing Q and the chord
X 2 Y2 lying along line PQ . Label them so that X i is closer than Yi to Q  in each case.
(These chords may be the same; in fact, we prove that they are.)
Now consider a chord X  Y  through both P and Q. (If P = Q, as we will show,
then there are infinitely many such chords, but in any case, there is at least one.)
Because both X  Y  and X 1 Y1 include Q, we have (|X  Q  | + |Y  Q  |)/|X  Y  |
= (|X 1 Q  | + |Y1 Q  |)/|X 1 Y1 |. Next, since both X  Y  and X 2 Y2 include P, we have
(|X  Q  | + |Y  Q  |)/|X  Y  | = (|X 2 Q  | + |Y2 Q  |)/|X 2 Y2 |. Therefore,

|X 1 Q  | + |Y1 Q  | |X 2 Q  | + |Y2 Q  |
= .
|X 1 Y1 | |X 2 Y2 |

Using |Y1 Q  | + |X 1 Q  | + |X 1 Y1 | and |Y2 Q  | = |X 2 Q  | + |X 2 Y2 |, we get (2|X 1 Q  |

+ |X 1 Y1 |)/|X 1 Y1 | = (2|X 2 Q  | + |X 2 Y2 |)/|X 2 Y2 |. Subtracting 1 and dividing by 2
yields |X 1 Q  |/|X 1 Y1 | = |X 2 Q  |/|X 2 Y2 |. Inverting these ratios and adding 1 gives
(|X 1 Q  | + |X 1 Y1 |)/|X 1 Q  | = (|X 2 Q  | + |X 2 Y2 |)/|X 2 Q  |. Since Q  X 1 Y1 and Q  X 2 Y2
are straight lines, |Y1 Q  |/|X 1 Q  | = |Y2 Q  |/|X 2 Q  |, or equivalently, |Y1 Q  |/|Y2 Q  |
= |X 1 Q  |/|X 2 Q  |.
By the concurrent chords theorem, |Y2 Q  |/|Y1 Q  | = |X 1 Q  |/ X 2 Q  |. Hence, |Y1 Q  |/
|Y2 Q  | = |Y2 Q  |/|Y1 Q  |, so |Y2 Q  | = |Y1 Q  |, and in turn |X 2 Q  | = |X 1 Q  |. This
implies that X 1 Y1 and X 2 Y2 are the same chord and, therefore, that P and Q are the
same point. Finally, Q  = P  , and so the image of P under inversion is the unique
point with the desired constant ratio property.

1056 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
Editorial comment. Several solvers noted that circle C is the Circle of Apollonius
determined by P and its image under inversion in C, using the ratio (r − p)/(r + p)
to get started. This observation provides another way to prove (b), that the image of P
under inversion in C is a suitable P  .
Also solved by R. Bagby, M. Bataille (France), E. Bojaxhiu (Albania)& E. Hysnelaj (Australia), J. Cade, R.
Chapman (U. K.), W. J. Cowieson, E. A. Herman, L. R. King, M. E. Kuczma (Poland), G. Lord, J. Schlosberg,
J. C. Smith, N. Stanciu & T. Zvonaru (Romania), R. Stong, E. A. Weinstein, and the proposer.

A Binomial Coefficient Inequality

11826 [2015, 284]. Proposed by Michel Bataille, Rouen, France. Let m and n be
positive integers with m ≤ n. Prove that
   n  
m+k−1 2  m+n 2
n
4n+1−k ≥ .
k=m
m−1 k=m
k

Solution by Timothy Woodcock,

 2m−1Stonehill
 College, Easton, MA. Equality holds
when n = m since 2m = 2 . Now suppose n > m, and inductively assume
n−1 n−k m+k−12 n−1 m+n−12
m m−1

k=m 4 m−1
≥ k=m k
. We have


n  2   
n−1  2
m+k−1 m+n−1 2 n−k m + k − 1
4n+1−k
=4 +4 4
k=m
m−1 m−1 k=m
m−1
  n−1 
  n−1 
 
m+n−1 2 m+n−1 2 m+n−1 2
≥4 +4 =4 .
m−1 k=m
k k=m−1
k
 m+n−12  2
It now suffices to prove 4 n−1 k=m−1 k
≥ nk=m m+n
k
. Since (x + y)2 ≤
2(x 2 + y 2 ) for x, y ∈ R,
n   n    
m+n 2  m+n−1 m+n−1 2
= +
k=m
k k=m
k−1 k
   
n
m+n−1 2 m+n−1 2
≤ 2 +
k=m
k−1 k
  n−1 
  n  
m+n−1 2 m+n−1 2 m+n−1 2
=4 +4 =4 .
n k=m
k k=m
k

Editorial comment. Allen Stenger notes that (x + y) p ≤ 2 p−1 (x p + y p ), valid for

x, y > 0 and p > 1, may be used in place of (x + y)2 ≤ 2(x 2 + y 2 ) to yield the gen-
eralization
n  p  n  
p(n+1−k) m + k − 1 m+n p
2 ≥ .
k=m
m−1 k=m
k

Also solved by R. Chapman (U. K.), J. H. Lindsey II, J. C. Smith, A. Stenger, R. Stong, R. Tauraso (Italy), and
the proposer.

December 2016] PROBLEMS AND SOLUTIONS 1057

 REVIEWS
Edited by Jeffrey Nunemacher
Mathematics and Computer Science, Ohio Wesleyan University, Delaware, OH 43015

A First Course in the Calculus of Variations. By Mark Kot. Student Mathematical Library
Vol 72, American Mathematical Society, Providence, Rhode Island, 2014, x + 298 pp., ISBN
978-1-4704-1495-5, $ 50.

Reviewed by Chris A. Marx

“I have with one blow solved two fundamental problems, one optical and the other
mechanical and have accomplished more than I have asked of others: I have shown
that the two problems, which arose from totally different fields of mathematics, never-
theless possess the same nature.”
This was Johann Bernoulli’s own summary of his solution of the two problems he
had posed in 1696 “to the acutest mathematicians of the world” [1, 4]1 . Bernoulli had
thereby demonstrated that the two problems in question, the famous brachistochrone
of finding the shortest path along which a particle moves in gravitation, and Fermat’s
principle of propagation of light, could be understood within a common framework.
His groundbreaking work marked the starting point of what has since become known
as the calculus of variations.
In the three centuries since Bernoulli’s work, it has been shown that in fact most
fundamental equations of physics, from classical physics to quantum field theory and
general relativity, as well as many diverse questions in mathematics can be formulated
within this framework. For a particularly beautiful and accessible account of the role
of calculus of variations in physics we refer the reader to Richard Feynman’s lecture
on the least action principle, published in [5]. Nevertheless, despite the richness of its
applications, it is probably fair to say that the calculus of variations does not appear
prominently in modern-day undergraduate education. Of course, it has its place in
standard physics courses on advanced mechanics, but in the mathematics curriculum,
if present at all, it is often deferred to specialized graduate level classes on partial
differential equations or functional analysis.
Inspired by the book under review, I want to make a case for including the subject in
the standard introductory course on differential equations offered in most undergradu-
ate mathematics programs in the United States.
Comparing current textbooks on differential equations to, say, those of twenty years
ago, a major shift in focus becomes apparent immediately. Aided in parts by the easy
accessibility of mathematical software, a course that merely focuses on teaching vari-
ous tricks to solve different types of differential equations has become obsolete. Con-
sequently, various approaches have been proposed for a more modern course, ranging
from a greater emphasis on modeling and numerics to perspectives influenced by qual-
itative analysis and dynamical systems.
It seems to me that including an introduction to the calculus of variations is another
very natural direction to take in a differential equations course. First and foremost,
1 An English translation from the original Latin can be found in, e.g., [9].

http://dx.doi.org/10.4169/amer.math.monthly.123.10.1058

1058 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
exposing students to the basics of the calculus of variations allows them to see the
origin of many important differential equations. From the wave equation to the Euler–
Lagrange equations in mechanics, all can be derived from the common perspective of
finding stationary points of appropriate functionals.
In addition, the calculus of variations is a natural source for many interesting prob-
lems. Kot’s book presents a great variety of problems accessible to undergraduates,
ranging from “classics” like Queen Dido’s isoperimetric problem or finding geodesics,
to the design of pneumatic trains (“terrestrial brachistrochrone”) or the path light trav-
els in optical fibers. Many of these case studies and exercises in the book lead to sepa-
rable or linear equations and hence are within reach for students at the beginning level.
The subject thus naturally complements and can easily follow the usual discussion of
first order and linear equations that typically starts courses on differential equations.
With an audience often consisting of a large percentage of students from the sci-
ences and engineering, calculus of variations also lends itself to showing naturally the
close ties between mathematics and the sciences.
To my mind, one of the most fundamental insights that the calculus of variation
provides with regard to the sciences is to shed a clear light on the nature of conserved
quantities. Conservation of energy, for instance, manifests itself as time independence
of the integrand (Lagrangian) of the action functional. This deep connection between
conserved quantities and invariances of the Lagrangian culminates in the famous
Emmy Noether theorem of mechanics.
From a more general perspective, the importance of using first integrals to sim-
plify the Euler–Lagrange equations is emphasized consistently throughout Kot’s book.
As conservation laws also serve as starting points for qualitative analysis of differen-
tial equations, having a clear understanding of the mathematical origin and implica-
tions of conserved quantities seems beneficial even to students primarily interested in
mathematics.
From a more practical point of view, knowledge of the basics of variational calcu-
lus considerably simplifies problems from mechanics. Probably because the subject is
accessible to students with diverse backgrounds and interests, mechanics is a particu-
larly popular source for problems in differential equations courses.
The constraints, however, imposed by the geometry of many of these problems,
even in such simple examples as the coupled or spherical pendulum, can make the
derivation using Newtonian mechanics somewhat tedious. The variational approach
(Hamilton’s principle of least action) giving rise to Lagrangian mechanics handles
constraints naturally; within Lagrange’s formalism the equations of motions can be
derived in a simple and consistent way, which then allows focus to be directed to ana-
lyzing the properties of solutions instead. The sliding rod, in Chapter 5 of Kot’s text,
pointedly illustrates the advantages that a variational approach gives, by contrasting
different ways in which it allows constraints to be treated, thereby shedding light on
different physical aspects of the problem.
My final motivation for proposing to include a discussion of calculus of variations
in a course on differential equations comes from my teaching of analysis. Possibly
because the first proof-based classes mathematics majors in the US take are typically
more algebraic, the transition to “thinking like an analyst” is often rather challenging
for students. In many conversations, students have expressed a disconnect to earlier
classes in the curriculum as one reason for their difficulties. Calculus itself as a moti-
vation is often perceived as too remote, especially since more and more students take
calculus before even starting college.
Including a topic such as the calculus of variations in an intermediate level course
like differential equations can provide an opportunity to help bridging the gap by

December 2016] REVIEWS 1059

motivating some important ideas in analysis without the technical preliminaries that
an introductory course in analysis necessitates.
Chapter 2 of Kot’s text provides a wonderful resource for such a discussion. Here,
the author lays the theoretical foundations for the basic problem of variational calculus,
that is, to find local extrema of a functional
 b
J [y] = f (x, y(x), y  (x)) dx
a

subject to the boundary conditions y(a) = ya and y(b) = yb .

Two question immediately arise: “What should the properties of admissible solu-
tions y(x) be, e.g., should y(x) be C 2 , C 1 , or piecewise-C 1 ?” and “How can one carry
out variations of the functional J [y]?” Both questions in turn lead to thinking about
how to quantify “largeness” of variations, i.e., what the appropriate choice of norm
should be.
The difference between the two natural choices, the C 1 -norm y1 := y∞ +

y ∞ , giving rise to weak variations, and the C 0 -norm, y∞ , giving rise to strong
variations, are explained beautifully and in a way accessible to students with only a
calculus background. The discussion can be nicely supplemented by further examples
provided at the beginning of Chapter 11, which illustrate how simple sawtooth-like
modifications of weak minima can produce strong, global minima.
The remainder of Chapter 2 lays the foundations of the theory of weak variations by
proving the Euler–Lagrange equations as the most basic necessary condition for local
extrema of J [y]. The discussion first presents Euler’s original heuristic argument based
on discretization of the variational problem, and then proceeds to present Lagrange’s
proof, which has since become the standard argument.
It is a refreshing feature of the entire text that the author does not shy away from
pointing out flaws or inconsistencies that were made in the historical development of
the subject. This not only makes certain technical subtleties more interesting to the
reader but, from a more general perspective, underlines “constructive failures” as a
crucial source for progress in mathematics.
For instance, in the context of proving the Euler–Langrange equations mentioned
above, a gap in Lagrange’s original proof led to the fundamental lemma of the calcu-
lus of variations of du Bois–Reymond, which eventually enabled them to show that
Euler–Lagrange equations are necessary for extrema in C 1 (and not merely in C 2 as
Lagrange’s proof implicitly assumes).
In conclusion, Mark Kot’s book has been a more than inspiring read. Of course
there are many excellent texts available on the subject, e.g., Gelfand–Fomin [6], Sagan
[8], or Carathéodory [2], to name some of the classics. With the book under review,
Kot’s achievement was however to present a contemporary introduction to the subject
which is accessible to undergraduate students with only a background in multivariable
calculus and some basic ordinary differential equations. As argued above, it can thus
easily be used to supplement the standard sophomore or junior level course on differ-
ential equations. While many of the more recent introductions to the field, e.g., [7],
[3], [10], require some prior knowledge of analysis and/or PDEs, the text under review
does not assume these prerequisites, without however sacrificing mathematical rigor.
Particularly noteworthy in Kot’s book are also the sections on recommended readings,
which, at the end of each chapter, beautifully bridge the gap from relevant historical
references to modern research articles.
Finally, on a more personal level, Kot’s book gave me a fresh perspective on teach-
ing a classical subject like differential equations. I hope the fruits of this process laid

1060 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
out above will provide useful food for thought for the reader of this review. As fortune
would have it, I will have ample opportunity to put my arguments to a test in a section
of differential equations that I am teaching in Fall 2016. Meanwhile, Kot’s book proved
to be a great source to revisit a beautiful subject in mathematics. I whole-heartedly
recommend the book for self-study and courses alike!

REFERENCES

1. J. Bernoulli, Curvatura radii in diaphanis non uniformibus, solitioque problematis a se in Actis 1696,
p. 269, proposti, de invenienda linea brachystochrona, id est, in qua grave a dato puncto ad datum punc-
tum brevissimo tempore decurrit, & de curva synchrona seu readiorum unda construenda, Acta Erudito-
rum 16 (1697) 206–211.
2. C. Carathéodory, Calculus of Variations and Partial Differential Equations of First Order. Third ed. AMS
Chelsea Publishing, American Mathematical Society, Providence, RI, 1999.
3. B. Dacorogna, Introduction to the Calculus of Variations. Third ed. World Scientific Publishing, Hong
Kong, 2014.
4. J. Ferguson, A Brief Survey of the History of the Calculus of Variations and its Applications (2004),
arXiv:math/0402357 [math.HO].
5. R. Feynman, The Feynman Lectures on Physics, Vol. II. Basic Books, New York, 2011.
6. I. M. Gelfand, S. V. Fomin, Calculus of Variations. Dover Books on Mathematics, Dover Publications,
Mineola, NY, 2000.
7. D. Liberzon, Calculus of Variations and Optimal Control Theory: A Concise Introduction. Princeton
Univ. Press, Princeton, NJ, 2012.
8. H. Sagan, Introduction to the Calculus of Variations, Dover Books on Mathematics, Dover Publications,
Mineola, NY, 1992.
9. D. J. Struik, A Source Book of Mathematics, 1200–1800. Harvard Univ. Press, Cambridge, MA, 1969.
10. B. van Brunt, The Calculus of Variations. Springer-Verlag New York, 2004.

Oberlin College, Oberlin, OH 44074

cmarx@oberlin.edu

December 2016] REVIEWS 1061

 M ONTHLY R EFEREES FOR 2016
The M ONTHLY expresses its appreciation to the following people for their help in
refereeing during the past year (July 2015 to July 2016). We could not function
successfully without such people and their hard work.

Stephen D. Abbott David Applegate Jacque Benatar

Malek Abdesselam Federico Ardila Manuel Benito
Ulrich Abel Masoud Ariannejad Arno Berger
Aaron Abrams Enrique Arrondo François Bergeron
Daniel Abrams Michael Aschbacher Alexander Berkovich
Gene Abrams Shamil Asgarli Leah Berman
Lowell Abrams William Aspray Bruce C. Berndt
Emad Abu Osba Michael Atiyah Dennis Bernstein
Peter Aczel Dan Aulicino Mladen Bestvina
Colin Adams Christian Avart Andras Bezdek
Warren Adams Athanassios N. Avramidis Therese Biedl
William Adkins Steve Awodey Elizabeth Billington
Anurag Agarwal Sheldon Axler Yuri Bilu
Robert A. Agnew Nathan Axvig David Bindel
Takashi Agoh Jean-Louis Ayme Dario Andrea Bini
Ian Agol Haris Aziz Robert Bix
Keith Agre Károly Böröczky Dakota Blair
Scott David Ahlgren Yuliya Babenko Greg Blekherman
Oswin Aichholzer Liljana Babinkostova Bill Goldbloom Bloch
Martin Aigner Nicholas R. Baeth Harold P. Boas
Amir Akbary Paul Baginski Matthew Boelkins
Ethan Akin David H. Bailey Alexander Bogomolny
Arseniy Akopyan Joseph Bak Michael Bolt
Michael Albert Michel Balinski Alexander Borisov
Stephanie Alexander Cristina M. Ballantine Alexei Borodin
Victor Alexandrov John Bamberg Alexandre V. Borovik
Krishnaswami Alladi Thomas F. Banchoff Jonathan Michael Borwein
Daniel Allcock Paul Bankston Michael W. Botsko
J.-P. Allouche Erkao Bao Amin Boumenir
Ricardo Almeida Edward J. Barbeau A. Bowers
Noga Alon Alexander Barg Khristo N. Boyadzhiev
Roger C. Alperin Marcy M. Barge Steven J. Brams
Dale E. Alspach Michael Barnsley David A. Brannan
Dean Alvis Wayne Barrett Simina Branzei
Tewodros Amdeberhan Alexander Barvinok Andrea Breard
Fredric Ancel Pierre Bayard David M. Bressoud
Dan Anderson Matthias Beck Kevin A. Broughan
David F. Anderson Gerald Beer Robert F. Brown
Titu Andreescu Ehrhard Behrends Lance Bryant
George E. Andrews Jason Behrstock Maribel Bueno
Omer Angel sarah-marie belcastro Jorge Buescu
Moa Apagodu Iddo Ben-Ari Ralf Bundshuh
http://dx.doi.org/10.4169/amer.math.monthly.123.10.1062

1062 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
Yurii Burman James Coykendall Jordan Ellenberg
James Butler Keenan Crane Jesse C. Elliott
Steve Butler Peter Cromwell Peter Elliott
Paul-Jean Cahen Rod Cross Mark Embree
Grant Cairns Péter Csikvári Ali Enayat
Jack S. Calcut Julia Cufi Shlomo Engelberg
Jeff Calder John Cullinan Kimmo Eriksson
David Callan Branko Curgus John Ewing
Antonio Campello Eugene Curtin Geoffrey Exoo
Michele Campiti Daniel Daners Marián Fabian
Emmanuel Candes Cecile Dartyge Kenneth Falconer
Mahai Caragiu Jean-Marie Shaun Fallat
Daniel Carando De Koninck Hershel Farkas
David A. Cardon Jesśs A. De Loera David Farmer
Carsten Carstensen Oswaldo Rio Branco Frank Farris
J. Scott Carter de Oliveira Wlodzimierz Fechner
Francis Castro Benne de Weger Timothy Feeman
Minerva Catral F. Michel Dekking Daniela Ferrero
Jean-Luc Chabert Jean-Paul Delahaye R. M. Fewster
Marc Chamberland Erik Demaine M. Gloria Fiestras-Janeiro
Robin Chapman Elizabeth Denne Michael Filaseta
Charalambos Harm Derksen James Fill
Charalambides Sebastien Destercke Jessica Fintzen
Bernard Chazelle Alice Devillers Patrick Mike Fitzpatrick
Bin Cheng Michel Marie Deza Valerie Flammang
Carmen Chicone Harold G. Diamond Cristina Flaut
Stephen Karl Dilcher Robbert Fokkink
Kwok-kwong Choi David E. Dobbs Gerald Folland
Jean Christianidis Brandy Doleshal James Forbes
Julianne Chung Feng Ming Dong Aviezri Fraenkel
Javier Cilleruelo David Drasin Natalie Priebe Frank
Mihai Cipu Greg Dresden Mikael Andersson Franko
Pete L. Clark Lobna Dridi John Franks
W. Edwin Clark Bau-Sen Du Craig Franze
Aaron Clauset Rui Alves Duarte Greg Frederickson
Todd Cochrane John Duncan Michael Freeze
Adam Coffman Charles Dunkl Chris Freiling
Stephen Cohen Grégoire Dupont Christopher Frenzen
Henry Cohn Bruno Durand Roland Freund
Vincent Coll Aaron Dutle Alan Frieze
William Wistar Comfort Herbert Edelsbrunner Marléne Frigon
Keith Conrad Tom Edgar Augustin Fruchard
Curtis Cooper Allan Edmonds Jason Fulman
Joshua Cooper Bruce Edwards Daniel Galicer
Bruce Cooperstein Gove Effinger David Galvin
Robert Corless Costas John Efthimiou Moubariz Garaev
Juan Carlos Cortéz López Richard Ehrenborg Ignacio Garcı́a Jurado
Ricardo Cortez Friedrich Eisenbrand Stephan Ramon Garcia
John Cosgrave Ivar Ekeland Thomas Garrity
Carl Cowen Shalom Eliahou Jaime Gaspar
David A. Cox Noam Elkies Meinolf Geck

December 2016] 1063

Stefan Geschke Aimo Hinkkanen Richard Kenyon
Ira M. Gessel Andreas Hinz Eric S. Key
Peter Giblin Jean-Baptiste Mohamed Khamsi
Juan Gil Hiriart-Urruty Tanya Khovanova
Paolo Giordano Alexander Hirsch Jerry King
Benjamin Girard Leslie Hogben James Kirkwood
Darren Glass Judy A. Holdener Jiri Klaska
Edray Herber Goins Alexander Holroyd Sandi Klavzar
David A. Goldberg Lars Holst Anthony Knapp
Luis Gonzalez Sanchez Roger Horn Oliver Ralph Knill
Julio Gonzalez-Diaz Leon Horsten Arnold Knopfmacher
L. Gonzalez-Vega Xiang-Dong Hou Alan Koch
Chaim Goodman-Strauss James Howard Alex Kontorovich
Mike T. Goodrich Cory Howk Arthur Kosowsky
Gary Gordon John Hubbard Mark Kot
Russell Arthur Gordon Thierry E. Huillet Irwin Kra
Pamela Gorkin Klaus Hulek Steven G. Krantz
Daniel Gottlieb Thomas Hull Ulrich Krause
Fernando Gouvêa Alexander Mike Krebs
Branko Grünbaum Jorg Hulpke Allan Kroopnick
Jeremy J. Gray Steve Humble Fred Kuczmarski
Robert E. Greene Jeffrey Hunter Angel Kumchev
Philip T. Gressman Piotr Indyk Ken Kunen
Robert Griffiths Ron Irving Dave Kung
J. Eli Grigsby Ivan Izmestiev Joseph Kung
David Grynkiewicz Michael J. Jacobson Greg Kuperberg
Robert Guralnick Graham James Jeffrey C. Lagarias
Ron Gurantz Oscar Jameson Evelyn Lamb
Tony Guttmann Sophia Jang Eric Landquist
Ollie Haggestrom Svante Janson Alessandro Languasco
Leon Hall Florian Jarre Dionysios Lappas
Franz Halter-Koch Mohammad Javaheri Niels Lauritzen
James Hamblin Thomas Jech Gregory F. Lawler
Emily Hamilton Paul Jenkins Sami Lazaar
Harald Hanche-Olsen David Jerison Imre Leader
Brian Harbourne Charles R. Johnson Chris Leary
Kathryn E. Hare Keith Johnson Franz Lemmermeyer
Kevin Hare Peter Johnson Melvin Leok
James D. Harper William Johnston Matthew Lettington
Robin E. Harte Lenny Jones Mark Levi
Russle Hendel Vladimir Kadets Lionel Levine
David W Henderson Jeff Kahn Leo Liberti
Allison Henrich Dan Kalman Daniel Liberzon
Kevin Henriot Masanobu Kaneko John B. Little
Russell Herman N. Karcanias Lance Littlejohn
Andrew Hetzel Hidefumi Katsuura Hailiang Liu
Andrew Hicks Mikhail Katz Yunan Liu
Philip Hieronymi Louis Kauffman Charles Livingston
Nicholas J. Higham Hermanshu Kaul Leo Livshits
Ted Hill Kiran Kedlaya Susan Loepp
A. J. W. Hilton Gary Kennedy Po-Shen Loh

1064 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
Victor Lomonosov Allison Moore Jonathon Peterson
Sergey Lototsky Justin Tatch Moore Conrad Piaut
Stephen Lovett luis Morales Oleg Pikhurko
Florian Luca Boris Mordhukovich Damian Pinasco
David Lyons Pieter Moree Rom Pinchasi
Russ Lyons Kent E. Morrison Nicholas Pippenger
Neal Madras Cristinel Mortici Mircea Pitici
Endre Makai M. Ãngeles Mulero James Pitman
Martin Malandro Gary Mullen Alaine Plagne
Toufik Mansour Laura Munteanu Leszek Plaskota
Eli Maor Marius Munteanu Robert Pless
Milton Maritz Ram Murty Kim Plofker
Diego Marques Derege Mussa John Polhill
Susan Hammond James Nagy John Polking
Marshall Ramin Naimi Paul Pollack
Donald A. Martin Takashi Nakamura Carl Pomerance
Greg Martin Nguyen Mau Nam Gerald Porter
Keith Matthews Alan Nathan Cheryl Praeger
Ernst W. Mayer Tristan Needham Victor Prasolov
Peter Mayr Stelios Negrepontis Arule Procaccia
John McCleary Roger B. Nelsen James Propp
Darryl McCullough Mark Newman Charles C. Pugh
Judith McDonald Lenny Ng Kari Pulli
Luis A. Medina Pace P. Nielsen Patrick Rabier
Robert Mena Sam Northshield David Rader
Anthony Mendes Richard J. Nowakowski Bala Rajaratnam
Franklin Mendivil Hiroshi Nozaki Jorge Luis
Mircea Merca Kevin O’Bryant Ramı́rez Alfonsı́n
Michael Ignacio Ojeda Alexander Ramm
Mesterton-Gibbons Tomas Oliveira e Silva Stuart Alexander
Fred Metcalf Peter Olofsson Rankin
Jeff Meyer Peter Olver Matthew Rave
William Meyerson Greg Oman Ana J. Reguera
Istvan Mezo Christopher O’Neill Victor Reiner
Piotr Micek Joseph O’Rourke Marc Renault
Brian Miceli Brad Osgood Jacques Resing
Dorel Mihet Thomas Osler Bruce Reznick
Dragan Milicic Alexander Ostermann Lovas Rezsö László
Alan D. Miller Valentin Ovsienko John Rhodes
Arnold Miller Péter Pach William Richardson
Russell Miller Nathan Paldor David Richeson
Steven J. Miller Zsolt Pales Thomas Richmond
David Minda Katrina Palmer Daniele Ritelli
Cleve Moler John Parker John Roe
Kenneth Monks Joseph Pasciak Pedro Roitman
Paul Monsky Mike Paterson Stefania Gabelli Roma
Luis Montejano Steen Pedersen José Carlos Rosales
Juan Monterde Maria Cristina Pereyra Peter Rosenthal
Vicente Montesinos David Perkinson Kim Roth
Richard Montgomery Ronald Perline David Rowe
J. W. Moon Kate Petersen Eric Rowland

December 2016] 1065

Javier S ánchez-Reyes Robert S. Strichartz Glen Van Brummelen
Donald G. Saari Walter Stromquist Charles Van Loan
Franco Saliola David Strong Hendrik van Maldeghem
David Sanchez Jeffrey Stuart Janet Vassilev
Altino Santos Bernd Sturmfels Jerry Vaughan
Stephen Schecter Francis Edward Su Robert C. Vaughan
Dierk Schleicher Brendan W. Sullivan Gerard A. Venema
James R. Schott John M. Sullivan Suresh
Peter Schreder Seth Sullivant Venkatasubramanian
Egon Schulte Masatoshi Suzuki Domenico Vicinanza
Peter D. Schumer Irena Swanson Sujith Vijay
Richard Schwartz Piotr Sworowski Stan Wagon
Armin Schwartzman Mario Szegedy Grace Wahba
Karl Schwede Daniel B. Szyld Gary Walsh
Carl Seaquist Hanjo Täubig Mark Walters
James Sellers Eitan Tadmore James Wan
Brigitte Servatius Bit-Shun Tam Changyou Wang
Jeffrey O. Shallit Roberto Tamassia Gerhard Wanner
Corey Shanbrom Christino Tamon Lawrence C. Washington
Ruslan Sharipov Martin Tancer Greg Wasilkowski
Donald Sherbert Wee Kee Tang John J. Watkins
Thomas Shermer Abdelhamid Tayebi D. J. Webb
Ian Short Alan Taylor Jeff Weeks
Igor Shparlinsky Holger Teismann Philip Welch
Tasos Sidiropoulos Alexander Teplyaev Elisabeth Werner
Andrew Simoson Cem Tezer Nicholas Werner
Michael Singer Willem F. D. Theron Ursula Whitcher
Yaron Singer Lola Thompson Roger A. Wiegand
Reyer Sjamaar Brian S. Thomson Harry Wiggins
Matthew Skerritt Hermann Thorisson Norman Wildberger
Gordon Slade Frank Thorne Hugh Williams
Martin Slowik Jög Thuswaldner Lauren Williams
Anthony Sofo Adam Timar Peter Winkler
Slawomir Solecki Todor D. Todorov James Wiseman
Jonathan Sondow Anthony Tongen Gerhard Woeginger
Michele Soria Alberto Torchinsky Julia Wolf
Mark Spivakovsky Cynthia Traub Zhaohua Wu
Renzo Sprugnoli Will Traves Shuji Yamamoto
Suvrit Sra Ann Trenk Catherine Yan
Katherine E. Stange Enrique Trevi ño Yong Yang
Caitlin King Stanton Ognian Trifonov Ae Ja Yee
Michael Steele Emmanuel Tsukerman Matt Young
Cameron L. Stewart Hirofumi Tsumura Liping Yuan
John Stillwell Kevin Tucker Doron Zeilberger
Kenneth B Stolarsky Caroline Paul Zeitz
Alexander L. Stolyar Turnage-Butterbaugh Fuzhen Zhang
Jeffery Stopple Brooke Ullery Ilze Ziedens
Joseph Straight David C. Ullrich Wadim Zudilin
Armin Straub Abraham A. Ungar William Seymour Zwicker

1066 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
 Enjoy math contests?

12 3 4 Take a look at our latest …

Euclidean Geometry in Mathematical Olympiads
Evan Chen

This is a challenging problem-solving book in Euclidean geom-

etry. Each chapter contains carefully chosen worked examples,
which explain not only the solutions to the problems but also de-
scribe in close detail how one would invent the solution to begin
with. The text contains a selection of 300 practice problems of

6
varying di culty from contests around the world, with exten-

5
sive hints and selected solutions. The exposition is friendly and
relaxed, and accompanied by over 248 beautifully drawn gures.

7
e-ISBN: 9781614444114 ebook: $28.00 311 pages, 2016
To order visit www.maa.org/ebooks/EGMO

A Gentle Introduction to the American Invitational

Mathematics Exam

9
Scott Annin

This book celebrates mathematical problem solving at the level

8 10
of the American Invitational Mathematics Examination. There
are more than 250 fully-solved problems in the book, containing
examples from AIME competitions of the 1980’s, 1990’s, 2000’s,
and 2010’s. In some cases, multiple solutions are presented to
highlight variable approaches. To help problem-solvers with the
exercises, the author provides two levels of hints to each exercise
in the book, one to help get an idea how to begin, and another to
provide more guidance in navigating an approach to the solution.

e-ISBN: 9781614444077 ebook: $27.00 398 pages, 2015

To order visit www.maa.org/ebooks/GIA
 New from the MAA
Beyond Lecture
Resources and Pedagogical
Techniques for Enhancing the
Teaching of Proof-Writing
Across the Curriculum
Rachel Schwell, Aliza Steurer and
Jennifer F. Vasquez, Editors

The challenges students can face in the transition from computational

mathematics to proof-writing lead many instructors to seek pedagogi-
cal techniques that extend beyond standard lecture. This Notes volume
unites a wide variety of such techniques, along with resources to aid in
incorporating them. Written with the busy instructor in mind, the
articles present practical methods in a “nuts-and-bolts” fashion, for
easy access to the details of each technique. Courses throughout the
entire undergraduate mathematics curriculum are represented: this
includes a variety of proof-based courses and also non-traditional ones
such as calculus and mathematics for liberal arts
Beyond Lecture should appeal to both novice and seasoned instructors,
while also hopefully providing a springboard for experimentation in
readers’ own classrooms.

Catalog Code: NTE85 332 pp., Paperbound or ebook, 2016

Print on Demand: $56.00 Print ISBN: 9780883851951
ebook PDF: $28.00 Electronic ISBN: 9781614443209

MAA members can download the ebook

for FREE from the Member Library.
 New MAA eBook
Introduction to the
Mathematics of
Computer Graphics
Nathan Carter

This text, by an award-winning MAA author, was designed to accompany his

first-year seminar in the mathematics of computer graphics. Readers learn the
mathematics behind the computational aspects of space, shape, transforma-
tion, color, rendering, animation, and modeling. The software required is freely
available on the Internet for Mac, Windows, and Linux.
The text answers questions such as these: How do artists build up realistic
shapes from geometric primitives? What computations is my computer doing
when it generates a realistic image of my 3D scene? What mathematical tools
can I use to animate an object through space? Why do movies always look
more realistic than video games?
Containing the mathematics and computing needed for making their own 3D
computer-generated images and animations, the text, and the course it sup-
ports, culminates in a project in which students create a short animated movie
using free software. Algebra and trigonometry are prerequisites; calculus is
not, though it helps. Programming is not required. Includes optional advanced
exercises for students with strong backgrounds in math or computer science.
Instructors interested in exposing their liberal arts students to the beautiful
mathematics behind computer graphics will find a rich resource in this text.

Catalog Code: IMCG 480 pp., ebook, 2016

ebook PDF: $35.00 Electronic ISBN: 9781614441229

Order now at www.maa.org/ebooks/IMCG

MATHEMATICAL ASSOCIATION OF AMERICA

1529 Eighteenth St., NW O Washington, DC 20036

Need a text for a QR course?

Common Sense
Mathematics
Ethan D. Bolker and Maura B. Mast

&RPPRQ6HQVH0DWKHPDWLFVLVD
WH[WIRUDRQHVHPHVWHUFROOHJHOHYHO
FRXUVHLQTXDQWLWDWLYHOLWHUDF\7KH
WH[WHPSKDVL]HVFRPPRQVHQVHDQG
FRPPRQNQRZOHGJHLQDSSURDFKLQJ
UHDOSUREOHPVWKURXJKSRSXODUQHZVLWHPVDQG¿QGLQJXVHIXOPDWK
HPDWLFDOWRROVDQGIUDPHVZLWKZKLFKWRDGGUHVVWKRVHTXHVWLRQV

Catalog Code: CSM Print ISBN: 9781939512109

E-ISBN: 9781614446217 List: $60.00
ebook: $30.00 MAA Member: $45.00
MAA Textbooks 328 pages, Hardbound, 2016

To order a print book visit www.store.maa.org or call 800-331-1622.

To order an electronic book visit www.maa.org/ebooks/CSM.