Professional Documents
Culture Documents
A Letter from the Editor: The Long and Winding Road 955
Scott T. Chapman
NOTES
Dilated Floor Functions that Commute 1033
Jeffrey C. Lagarias, Takumi Murayama, and D. Harry Richman
BOOK REVIEW
A First Course in the Calculus of Variations 1058
by Mark Kot
Chris A. Marx
EDITOR
Scott T. Chapman
Sam Houston State 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
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
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.
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
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.
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.
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 }
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 = u1 · . . . · uk where k ≥ 2 and u 1 , . . . , u k ∈ H \ 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.
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].
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
u 1 · . . . · u k = v1 · . . . · v where u 1 , . . . , u k , v1 , . . . , v ∈ A(H ).
964
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
and we will use these inequalities without further mention. The invariant
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.
qρ + 1
> ρ for all q ≥ q0 ,
q +1
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
L = y + (L ∪ L ∗ ∪ L ) ⊂ y + dZ,
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.
|Uk (H )| 1
1
lim = ρ(H ) − .
k→∞ k d ρ(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 .
and hence
∗
U + Uk−k0 (H ) ∩ [k, ρk (H ) − k0 M] = Uk (H ) ∩ [k, ρk (H ) − k0 M]
k ≤ ρ (H ) and + M ∗ ≤ m,
k + M ∗ ≤ ρm−M ∗ (H ) + M ∗ ≤ ρm−M ∗ (H ) + ρ M ∗ (H ) ≤ ρm (H ).
ρk (H ) λk (H ) 1
lim = ρ(H ) and lim = ,
k→∞ k k→∞ k ρ(H )
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)}
∗
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.
|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
970
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
implemented in GAP (see the GAP Package [14] and a survey by Garcı́a-Sánchez
[20]).
C (ϕ) = C (H ) = q(F)/q(Hred )
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
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
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.
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 ∈ }
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].
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
B (G 0 ) = {S ∈ F (G 0 ) | σ (S) = 0} ⊂ F (G 0 )
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
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]).
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.
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.
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.
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 ) .
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.
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.
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.
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
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,
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.
L = y + (L ∪ L ∗ ∪ L ) ⊂ y + D + dZ,
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
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.
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.
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].
for every k ∈ N. Since A(G) is finite (see Propositions 7 and 9), the Davenport
constant
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| < ∞.
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 · . . . ·
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
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
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
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
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
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.
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
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.
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.
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:
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
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.
3. FIELDS.
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
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
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
(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:
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.
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
f (x) = f 1 (x p ),
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 )
4. ALGEBRAS.
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
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.
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.
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.
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),
Now, we give examples showing that no conditions of Theorem 4 and its corollaries
can be omitted.
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.
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 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].
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 .
∞
∞
∞
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.
F(x) = F(x1 , x2 , . . . , xm ) = a1 x1 + a2 x2 + · · · + am xm + c.
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
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.
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.
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 .
k
α β
D = (−1) 2 m 2
qi
i=1
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
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
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
and
f = F(0, 0) ∈ N0 .
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
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
1
L(x, y) = (d x + ey)
2
we can write
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
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
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
where
it follows that
and so
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.
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
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
and
8a D F(x, y) = Du 2 − v 2 + r.
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
Therefore,
Du 2 ≡ v 2 (mod p).
Du 2 − v 2 ≡ 0 (mod p2 ).
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.
(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
(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.
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
(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
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.
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.
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
x x
c c
y y
x x
c c
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.
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
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.
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).
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
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 ) =
f 1 (x1 ) = λ,
f 2 (x2 ) = λ,
..
.
f n (xn ) = λ
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
(i.e., the sequence (x1 , x2 , . . . , xn ) for which f 1 (x1 ) = f 2 (x2 ) = · · · = f n (xn )).
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
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
4. PLAYER STRATEGY.
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
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.
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.
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)
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:
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)
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)
200 20
100 10
0 0
0 2,000 4,000 0 20 40 60 80
Number of function evaluations Number of function evaluations
Time (ms)
Time (ms)
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
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.
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
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.
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
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
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
n
a1 1/(r −1)
x1 + (x1 − c1 ) + ci = d − x.
i=2
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
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.
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.
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
1032
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
NOTES
Edited by Sergei Tabachnikov
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
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
αβx = βαx
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,
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 2. For α, β > 0, the function f 1/α commutes with f 1/β if and only if the
equality
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.
α nα
= , (4)
β nβ
α A kα
= = for all k ≥ 1.
β B kβ
α A nα nA − 1
= = = ,
β B nβ nβ
A nA − 1
= .
B nB − 1
Case 2. Both α and β are negative. We obtain a criterion that parallels Lemma 2 in
the positive case.
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
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.
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
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
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.
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.
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.
tends to zero as N → ∞.
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.
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].
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).
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 )
so that
2j 2j
ϕ 4z( p) p ≡ 1 (mod q2 j ) and ϕ̃ 4z( p) p ≡ 1 (mod q2 j ). (11)
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 )
F4z( p) p2 j m
≡ m (mod p j )
F4z( p) p2 j
F4z( p) p2 j m
≡ m (mod p j ) (12)
F4z( p) p2 j
so
F4z( p) p2(k+2r )
= p2r , ∈ Z, gcd(, p) = 1. (14)
F4z( p) p2(k+r )
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
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.
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.
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.
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.
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
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.
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 ).
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.
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
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).
SOLUTIONS
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
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.
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.
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.
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.
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
|X 1 Q | + |Y1 Q | |X 2 Q | + |Y2 Q |
= .
|X 1 Y1 | |X 2 Y2 |
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.
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
Also solved by R. Chapman (U. K.), J. H. Lindsey II, J. C. Smith, A. Stenger, R. Stong, R. Tauraso (Italy), and
the proposer.
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.
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
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.
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
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
1066
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 123
Enjoy math contests?
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
9
Scott Annin
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.
Common Sense
Mathematics
Ethan D. Bolker and Maura B. Mast
&RPPRQ6HQVH0DWKHPDWLFVLVD
WH[WIRUDRQHVHPHVWHUFROOHJHOHYHO
FRXUVHLQTXDQWLWDWLYHOLWHUDF\7KH
WH[WHPSKDVL]HVFRPPRQVHQVHDQG
FRPPRQNQRZOHGJHLQDSSURDFKLQJ
UHDOSUREOHPVWKURXJKSRSXODUQHZVLWHPVDQG¿QGLQJXVHIXOPDWK
HPDWLFDOWRROVDQGIUDPHVZLWKZKLFKWRDGGUHVVWKRVHTXHVWLRQV