Bruce Landman - Combinatorial Number Theory

De Gruyter Proceedings in Mathematics
Combinatorial Number
Theory
Proceedings of the Integers Conference 2011
Carrollton, Georgia, October 2629, 2011
edited by
Bruce Landman
Melvyn B. Nathanson
Jaroslav Neetril
Richard J. Nowakowski
Carl Pomerance
Aaron Robertson
De Gruyter
Mathematics Subject Classification 2010: 11-06.
Editors
Prof. Dr. Bruce Landman Prof. Dr. Melvyn B. Nathanson
University of West Georgia The City University of New York
Department of Mathematics Lehman College (CUNY)
1601 Maple Street Department of Mathematics
Carrollton, GA 30118, USA 250 Bedford Park Boulevard West
landman@westga.edu Bronx, NY 10468, USA
nathansn@alpha.lehman.cuny.edu
Prof. Dr. Jaroslav Neetril Prof. Dr. Richard J. Nowakowski

Charles University Dalhousie University
Department of Applied Mathematics Department of Mathematics & Statistics
Malostransk nam. 25 Chase Building
118 00 Prague, Czech Republic Halifax, NS B3H 3J5, Canada
nesetril@kam.ms.mff.cuni.cz rjn@mathstat.dal.ca
Prof. Dr. Carl Pomerance Prof. Dr. Aaron Robertson

Dartmouth College Colgate University
Department of Mathematics Department of Mathematics
6188 Kemeny Hall 13 Oak Drive
Hanover, NH 03755-3551; USA Hamilton, 13346, USA
carl.pomerance@dartmouth.edu aaron@math.colgate.edu
ISBN 978-3-11-028048-7
e-ISBN 978-3-11-028061-6
Set-ISBN 978-3-11-028062-3
Library of Congress Cataloging-in-Publication Data

A CIP catalog record for this book has been applied for at the Library of Congress.
Bibliographic information published by the Deutsche Nationalbibliothek

The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographic data are available in the internet at http://dnb.dnb.de.
2013 Walter de Gruyter GmbH, Berlin/Boston
Typesetting: PTP-Berlin Protago-TEX-Production GmbH, www.ptp-berlin.eu

Printing and binding: Hubert & Co. GmbH & Co. KG, Gttingen
Printed on acid-free paper
Printed in Germany
www.degruyter.com
Preface
The Integers Conference 2011 was held October 2629, 2011, at the University of
West Georgia in Carrollton, Georgia, United States. This was the fifth Integers Con-
ference, held bi-annually since 2003. It featured plenary lectures presented by Ken
Ono, Carla Savage, Laszlo Szekely, Frank Thorne, and Julia Wolf, along with 60 other
research talks.
This volume consists of ten refereed articles, which are expanded and revised ver-
sions of talks presented at the conference. These ten articles will appear as a special
volume of the journal Integers. They represent a broad range of topics in the areas
of number theory and combinatorics including multiplicative number theory, additive
number theory, game theory, Ramsey theory, enumerative combinatorics, elementary
number theory, the theory of partitions, and integer sequences.
The conference was made possible with the generous support of the Number Theory
Foundation and the University of West Georgia.
Carrollton, Georgia, USA, March 2013 The Editors

Contents
Preface v
1 Rebecca Milley, Richard J. Nowakowski, and Paul Ottaway

The Misre Monoid of One-Handed Alternating Games 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Equivalences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 The Misre Monoid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2 Neil Hindman and John H. Johnson

Images of C-Sets and Related Large Sets under Nonhomogeneous
Spectra 15
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 The Various Notions of Size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 The Functions f and h . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 Preservation of J -Sets, C -Sets, and C -Sets . . . . . . . . . . . . . . . . . . . 27
2.5 Preservation of Ideals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3 Daniel A. Goldston and Andrew H. Ledoan

On the Differences Between Consecutive Prime Numbers, I 37
3.1 Introduction and Statement of Results . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 The HardyLittlewood Prime k-Tuple Conjectures . . . . . . . . . . . . . . 38
3.3 InclusionExclusion for Consecutive Prime Numbers . . . . . . . . . . . . 39
3.4 Proof of the Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4 Pr Kurlberg, Jeffrey C. Lagarias, and Carl Pomerance

On Sets of Integers Which Are Both Sum-Free and Product-Free 45
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2 The Upper Density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
viii Contents
4.3 An Upper Bound for the Density in Z=nZ . . . . . . . . . . . . . . . . . . . . . 50

4.4 Examples With Large Density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5 Frank Thorne
Four Perspectives on Secondary Terms in the DavenportHeilbronn
Theorems 55
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.2 Counting Fields in General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.2.1 Counting Torsion Elements in Class Groups . . . . . . . . . . . . . 59
5.3 DavenportHeilbronn, DeloneFaddeev, and the Main Terms . . . . . . 60
5.3.1 The Work of Belabas, Bhargava, and Pomerance . . . . . . . . . . 61
5.4 The Four Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.5 The Shintani Zeta-Function Approach . . . . . . . . . . . . . . . . . . . . . . . . 63
5.5.1 Nonequidistribution in Arithmetic Progressions . . . . . . . . . . . 66
5.6 A Refined Geometric Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.6.1 Origin of the Secondary Term . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.6.2 A Correspondence for Cubic Forms . . . . . . . . . . . . . . . . . . . . 69
5.7 Equidistribution of Heegner Points . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.7.1 Heegner Points and Equidistribution . . . . . . . . . . . . . . . . . . . 71
5.8 Hirzebruch Surfaces and the Maroni Invariant . . . . . . . . . . . . . . . . . . 73
5.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6 Brian Hopkins
Spotted Tilings and n-Color Compositions 79
6.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.2 n-Color Composition Enumerations . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.3 Conjugable n-Color Compositions . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
7 Aviezri S. Fraenkel and Yuval Tanny
A Class of Wythoff-Like Games 91
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.2 Constant Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.2.1 A Numeration System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
7.2.2 Strategy Tractability and Structure of the P -Positions . . . . . . 98
7.3 Superadditive Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Contents ix
7.4 Polynomial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

7.5 Further Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
8 Takao Komatsu, Florian Luca, and Yohei Tachiya
On the Multiplicative Order of FnC1 =Fn Modulo Fm 109
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
8.2 Preliminary Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
8.3 Proof of Theorem 8.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
8.4 Comments and Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
9 Neil A. McKay and Richard J. Nowakowski
Outcomes of Partizan Euclid 123
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
9.2 Game Tree Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
9.3 Reducing the Signature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
9.3.1 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
9.4 Outcome Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
9.5 Open Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
10 Thomas W. Pensyl and Carla D. Savage
Lecture Hall Partitions and the Wreath Products Ck o Sn 137
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
10.2 Lecture Hall Partitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
10.3 Statistics on Ck o Sn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
10.4 Statistics on s-Inversion Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
10.5 From Statistics on Ck o Sn to Statistics on In;k . . . . . . . . . . . . . . . . . . 141
10.6 Lecture Hall Polytopes and s-Inversion Sequences . . . . . . . . . . . . . . . 143
10.7 Lecture Hall Partitions and the Inversion Sequences In;k . . . . . . . . . . 145
10.8 A Lecture Hall Statistic on Ck o Sn . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
10.9 Inflated Eulerian Polynomials for Ck o Sn . . . . . . . . . . . . . . . . . . . . . 150
10.10 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
Index 155
Combinatorial Number Theory, 113 De Gruyter 2013
Chapter 1
The Misre Monoid of One-Handed Alternating

Games
Rebecca Milley, Richard J. Nowakowski, and Paul Ottaway
Abstract. Misre games are notoriously more difficult to analyze than their normal-play coun-
terparts. To deal with the inherent lack of structure, Plambeck (2005) [7] considered restricted
universes of games and developed the concept of indistinguishability, modulo a given uni-
verse; the resulting quotient semigroup is called the misre monoid. Results for only three
general universes are known. We introduce a fourth: the universe of alternating games and
their sums. We find that the misre monoid of one-handed alternating games is isomorphic to
.Z; C/.
Keywords. Combinatorial Games, Misre Games.
Mathematics Subject Classification 2010. 91A46.
1.1 Introduction
Two players Left and Right play with a stack of pennies, which are either all heads-up
or all tails-up; Left can play on a tails-up stack and Right can play on a heads-up stack
by removing at least one penny and then inverting any remaining coins. The player who
removes the last penny loses. This is an alternating game, since neither Left nor Right
can make two consecutive moves in one stack, and it is being played under misre
rules. In the present chapter, we analyze individual alternating components (e.g., one
stack of pennies) in order to answer the following question: under misre play, who
wins a set of alternating games when they are played in a disjunctive sum? Note that a
disjunctive sum of alternating components is not necessarily itself an alternating game.
For a given game G, we use G L to denote a general left option and G L to denote
the set of all such options. We may refer to a single left option from the position G R
as G RL and to the set of all left options from the position G R as G RL . A game H
is a follower of G if H can be reached from G by some sequence of (not necessarily
alternating) moves.
Definition 1.1. A game G is said to be alternating if G LL D ; and G RR D ; for all

left and right options G L ; G R of G, and if every follower of G is also alternating.
In PENNY NIM, each component is one-handed, i.e. only one player has an op-
tion from the initial position. We can specify which player, Left or Right, has the
first move by the terms left-handed or right-handed, respectively. Games can also be
two-handed alternating, where both players have at least one option from the start
(G L 6D ;; G R 6D ;). If we placed a heads-up or tails-up stack of pennies on its side
and allowed either player to move by taking some pennies and orienting the stack ap-
propriately, then PENNY NIM would become a two-handed alternating game. This chap-
ter considers one-handed positions; however, the majority of our one-handed results
are true in the universe of all alternating games. Since every two-handed game has
exclusively one-handed followers, an understanding of one-handed games will help
us to subsequently analyze all alternating games. Readers familiar with normal-play
games will notice that one-handed alternating games are restricted to the normal-play
values 1, 0, and 1.
It is interesting to note that one-handed or end games in general can be very prob-
lematic for misre play Aaron Siegel calls them significant pathologies [9] and
yet there is simple, elegant structure in one-handed alternating games.
1.1.1 Background
We begin with an overview of prerequisite material. Conventionally, the players Left
and Right are female and male, respectively. We have already encountered an instance
of misre-play, where the first player unable to move wins; the more standard ending
condition is normal-play, where the first player unable to move loses. Games or posi-
tions are defined in terms of their options: G D G L j G R . The simplest game is the
zero game, 0 D j , where the dot indicates an empty set of options.
In both play conventions, the outcome classes next (N ), previous (P ), left (L), and
right (R) are partially ordered as shown in Figure 1.1, with Left preferring moves
toward the top and Right preferring moves toward the bottom. In our analysis of al-
ternating games, we often simultaneously consider the misre outcome and normal
outcome of a game; to distinguish between the two, we introduce the superscripts
and C , respectively. Thus a game in N \ RC is a next-win under misre play and
a right-win under normal play. When convenient we also use the outcome functions
o .G/ and oC .G/ to identify the misre or normal outcome of a game G.
L

P N

R
Figure 1.1. The partial order of outcome classes.
Chapter 1 The Misre Monoid of One-Handed Alternating Games 3
Many definitions from normal-play game theory1 are used without modification for
misre games, including disjunctive sum, equality, and inequality. Thus, for misre
games,
G D H if o .G C X / D o .H C X / for all games X;
G H if o .G C X / o .H C X / for all games X:
In normal-play, the negative of a game is defined recursively as G D G R jG L ,
and is so called because G C .G/ D 0 for all games G. Under misre play, however,
this property holds only if G is the zero game [5]. To avoid confusion and inappropriate
cancellation, we write G instead of G and refer to this game as the conjugate of G.
In normal-play games, there is an easy test for equality: G D 0 if and only if G 2
P C , and so G D H if and only if G H 2 P C . In misre-play, no such test exists.
Equality of misre games is difficult to prove and, moreover, is relatively rare; for
example, besides j itself, there are no games equal to the zero game under misre
play [5]. Within the last ten years (see, e.g., [79]), a partial solution to these challenges
has been presented: redefine equality by restricting the game universe. This method
has been used with much success by [24]. Given a set of games X, misre equivalence
(modulo X) is defined by
G H (mod X) if o .G C X / D o .H C X / for all games X 2 X;
while misre inequality (modulo X) is defined by
G = H (mod X) if o .G C X / o .H C X / for all games X 2 X:
We use the words equivalent and indistinguishable interchangeably, and if G 6 H

(mod X) then G and H are said to be distinguishable.
Given a universe X, we can determine the equivalence classes under .mod X/
and form the quotient semi-group X= . This quotient, together with the tetra-partition
of elements into the sets P , N , R , and L , is called the misre monoid of the
universe X, denoted MX . The misre monoid is by convention written multiplica-
tively, and its identity element, denoted e, is the equivalence class of zero. As an ex-
ample, the misre monoid of the games 0, 0 j , and all possible sums, is given by
he; j 2 D i, with N D e, R D , P D L D ; [2]. The closure of a set of
games is the set of all disjunctive sums of those games and their followers. We denote
the closure of the set of all alternating games by A and the closure of one-handed al-
ternating games by O. Note that if X Y then G H (mod Y) implies G H
(mod X); in particular, equivalence in A immediately gives equivalence in O A.
Following Ottaways PhD thesis [6], we consider the universe A under misre play
and find the misre monoid MO of one-handed alternating games. We would like to
acknowledge Meghan Allen for suggesting the more difficult problem of constructing
the monoid for all alternating games, of which this paper is a first step.
1A complete overview of normal-play game theory can be found in [1].
In Section 1.2, we determine the equivalence classes of one-handed alternating

games inside the larger universe A, and in Section 1.3, we calculate outcomes and
identify the equivalence classes of O alone. In Section 1.4 we present the misre
monoid MO . Future goals for this research naturally include analyzing two-handed
games and determining MA . Some progress has been made, but the equivalence class-
es of two-handed games are more complicated than those for O. It would also be inter-
esting to classify the nonalternating games that can be written as a disjunctive sum of
alternating positions; for example, the game j can be decomposed into C . In
this way we could extend our knowledge of the alternating universe to a broader fam-
ily of games. Interested readers should refer to Ottaways thesis [6] for an extended
discussion of alternating2 games, including an analysis of subtraction games on coins
(a generalization of PENNY NIM).
1.2 Equivalences
In analyzing alternating games, we find it useful to classify a position by both its mis-
re outcome and normal outcome. In general all 42 D 16 pairs of outcomes can be
attained [5] (for example, there exist games which are Left-win under both normal-
and misre-play), but more than half of these pairs do not occur among one-handed
games. Since either Left or Right has no first move in a one-handed game G, that player
wins immediately under misre rules and loses immediately under normal rules, and
so o .G/ 6D P , oC .G/ 6D N C , and o .G/ 6D oC .G/. The remaining seven out-
come pairs are each attained in the one-handed universe, as demonstrated in Figure 1.2
with the zero game and the games
A D 0 j ; B D 0; A j ; C D B j ;
along with their corresponding right-handed conjugates3 .
R \ LC N \ LC R \ P C N \ P C L \ RC N \ RC L \ P C
A B C 0 A B C
Figure 1.2. The possible outcomes pairs for one-handed alternating games.
2 In
[6] and elsewhere, alternating games are referred to as consecutive move ban games.
3 Wewill find that 0 j is incomparable with zero in our universe; for this and other reasons we
chose not to use the normal-play name 1.
Most of the positions in Figure 1.2 appear in PENNY NIM; if we consider stacks on
which Left can play, then a one-penny stack is the game A, a two-penny stack is the
game B, and all other nonzero stacks are equivalent (modulo A) to B.
Throughout the chapter we make repeated use of the fact that a single one-handed
game alternates between being left-handed and right-handed as players alternate turns.
Furthermore, if Left has a good (misre) first move in a left-handed game, then it is to
a position in L , because no one-handed game is in P . Likewise, if Right as a good
first move in an alternating game, it is to a position in R .
We show in Theorem 1.4 that all one-handed games in N \ P C are indistinguish-
able from the zero game, an incredibly useful result that actually holds in the larger
universe A of all alternating games. We then extend this result by establishing that
a single one-handed game is indistinguishable (modulo A) from any other with the
same outcome pair; in particular, the seven games in Figure 1.2 are equivalent to all
others in their respective classes.
We first require the following two lemmas.
Lemma 1.2. If X is a sum of right-handed alternating games with at least one com-
ponent in L , then X 2 L .
Proof. Left wins trivially playing first on X . Assume, for followers of X , Left wins
playing second on a sum of right-handed games when at least one of them is in L .
Let G be a component of X in L . If Right plays in G to some G R then Left either
has no response and wins immediately, or responds with a good move G RL 2 L
and wins by induction on the sum. If Right plays in some other component of X then
Left either has no response and wins, or can play in that component to bring it back to
a right-handed game, and then wins playing second on the sum, by induction, since G
is still in L .
Lemma 1.3. If G is a single right-handed game in P C then G = 0 (mod A), and if

G is a single left-handed game in P C then G 5 0 (mod A).
Proof. We prove the first statement, and the second will follow by symmetry. Let
G be a right-handed game with normal outcome P C and X be a sum of alternating
games. We need to show that o .G C X / o .X /. If o .X / D R then trivially
o .G C X / o .X /. It remains to show that if Left can win X going first (second),
then she can win G CX going first (second). Assume inductively that these statements
hold for all followers of G C X of the form G 0 C X 0 with G 0 2 P C .
Suppose now that Left can win X going first. If she has no move in X , then she has
no move in G C X , and thus wins immediately. Otherwise, she has a move to some
X L from which she wins moving second, and by induction she then wins G C X L
moving second. Thus Left can win G C X moving first.
Suppose Left wins X moving second. Note that Right always has a move at the
outset. If Right moves in G, then Left has a (normal-play) winning local response to
some G RL , which as a right-handed game must be in P C and not LC . Then Left wins
G RL C X by induction. If Right moves in X to X R , then either Left wins because
she has no move, or she responds with her good second move X RL and by induction
wins G C X RL moving second.
We are now set to prove our earlier claim that every one-handed game in N \ P C
is indistinguishable among all alternating games from zero.
Theorem 1.4. If G is any one-handed alternating game in N \ P C , then G 0

(mod A).
Proof. Assume G is left-handed (the other case is symmetric). By Lemma 1.3 we

already have G 5 0, and so it suffices to show G = 0. Let X be any sum of alternating
games. Assume that for all followers X 0 of X , Left can win G CX 0 going first (second)
when she can win X 0 going first (second).
Suppose Left wins X playing first. If she has no move in X , then in G C X she can
make a good first misre move in G 2 N , say G L 2 L , to bring the whole position
to a sum of right-handed games, one of which is in L . Lemma 1.2 then shows Left
wins playing first on G C X . Otherwise, Left has a move in X to some X L , from
which she wins playing second; Left then wins G C X L playing second, by induction,
and so wins G C X playing first.
Suppose Left wins X playing second. Then Left can win X R playing first for any
Right move X R , and thus by induction wins G C X R playing first. Since Right has
no choice but to play in X on the sum G C X , this shows Left wins playing second on
G C X.
We want to show that every one-handed alternating game not in N \ P C is also

equivalent to every other game with the same pair of outcomes. We first establish
two instances of domination among left options: Left should move to a position in
L \ P C over one in N \ P C , and should choose N \ P C over a position in
N \ RC . Naively, Left prefers one option over another if the misre or normal out-
come is more favourable for Left and the other outcome is just the same. As with
normal-play, if G D G L j G R , then we can remove dominated options from G L
and G R to obtain a game equivalent to G under misre-play [9].
Theorem 1.5. If G 2 N \ RC and H 2 L \ P C are one-handed alternating

games, then G 5 0 5 H (mod A).
Proof. Note that G and H are right-handed games. Since H 2 P C , Lemma 1.2 gives
us 0 5 H . For G 5 0, let X be a sum in A. Suffices to show that if Right can win X
playing first (second), then Right can win G C X playing first (second); assume that
this is true for all followers X 0 of X . Now suppose Right can win X playing first. If
Right has no moves in X , he can make a good first misre move in G to G R 2 R
and win on G R C X by Lemma 1.2. Otherwise, Right makes his good misre move
in X to X R ; since Right wins X R playing second, he wins G C X R playing second
by induction. This shows Right wins G C X playing first.
If Right wins X playing second then Left has no good first move in X and cannot
play in the right-handed game G. Left must move X to X L , which Right can win
playing first, and then by induction Right can win G C X L playing first. Thus Right
wins G C X playing second.
The symmetric result obviously holds as well, and together these give us a partial
partial order of moves. In particular, Theorem 1.5 tells us that C 5 0 5 B and B 5
0 5 C . We will be able to establish another chain of domination (C 5 A 5 B,
B 5 A 5 C ) after we prove Theorem 1.6.
Theorem 1.6. If G is a one-handed alternating game, then the following statements

are true modulo A:
(i) If G 2 R \ LC then G A. (ii) If G 2 L \ RC then G A.

(iii) If G 2 N \ LC then G B. (iv) If G 2 N \ RC then G B.
(v) If G 2 R \ P C then G C . (vi) If G 2 L \ P C then G C .
Proof. (i)(ii): If G 2 R \ LC then G is left-handed, and since Left has no good first
misre move, every Left option must be in N or R . But all Left options are right-
handed (or zero), and thus cannot be in R . So every Left option is in N . Similarly,
since G 2 LC , there is at least one Left option in P C or LC ; but right-handed (or
zero) games cannot be in LC , so at least one Left option is in P C . Together this
shows that there exists a G L 0 2 N \ P C . By Theorem 1.5, any other options
(necessarily in N \ RC ) are dominated by 0, and so G 0 j A. Case (ii)
follows by symmetry.
(iii)(vi): We combine the remaining results and use an inductive proof. The reader
can check that B is the smallest one-handed game in N \ LC (i.e., no one-handed
game with a shorter game tree is in N \ LC ), and B, C , and C are the smallest games
in R \ P C , N \ RC , and L \ P C , respectively.
We now proceed by induction on the followers. Let G and H be games in R \ P C
and N \ LC (note that G and H are left-handed; the right-handed results follow
similarly). As argued in (i)(ii), every Left option of G must be in N because Left
has no good first misre move in G 2 R . Additionally, G 2 P C forces every Left
option to be in RC or N C ; but one-handed games cannot be in N C , so every option
is in N \ RC . By induction each of these options is indistinguishable from B. Thus
G B j D C .
For H there is a bit more to show. Since H 2 N , Left has a good misre move
to a right-handed game in L ; since H 2 LC , Left has a good normal-play move
to a right-handed game in P C . This shows that Left either has a move to L \ P C
( C , by induction) or has moves to both L \ RC ( A by (i)) and N \ P C ( 0).

If there are any moves to N \ RC , they are dominated by 0 or C , by Theorem 1.5.
Using this and the domination of C over 0, we reduce the possibilities for H to
H A; 0 j ; C j ; or A; C j :
In the first case we have H B immediately. It remains to show that C j B

and A; C j B.
Claim 1. C j B. Let J D C j . We need to show o .B C X / D o .J C X /
for all sums X in A. Assume true for all followers of X . If Right wins playing first
in B C X , then he must win second from some B C X R , which by induction has the
same outcome as J C X R , and then Right wins first from J C X with the same move
in X . If Left wins playing second in B C X , the same argument shows that Left wins
playing second in J C X , since for every X R , o .B C X R / D o .J C X R /.
If Right wins playing second on B C X , and Lefts first move in J C X is to some
J C X L , then since Right can win playing first on B C X L he can win playing first
from J C X L , by induction. Otherwise, Lefts first move in J C X is to C C X , and
Right should respond with B C X , from which he wins playing second.
Finally, suppose Left wins B C X playing first. If her good move is in X , then
as above Left wins J C X with the same first move. If Lefts good first move is
to B L C X , then Left must be able to win playing second from any position B L C
X RL:::RL obtained from optimal Left play (including B L C X ). Now, in J C X , Left
should move first to C C X and play her original strategy in X . When Right chooses
to play in C , to B C X RL:::RL for some (not necessarily proper) follower of X , Left
plays to B L C X RL:::RL and wins from there playing second.
Claim 2. A; C j B. let K D A; C j . Note that when at least one left option
already exists, introducing another left option cannot create a position worse for Left;
Left simply ties her hand and ignores the extra option. In this way we see that K
J B. It remains to show that K 5 B; i.e., that Right wins K C X whenever he
wins B C X .
If Right wins B CX playing first, then he wins K CX playing first, by induction, as
in Claim 1. Suppose Right wins B C X playing second. If Left moves first in K C X
to C C X , then Right wins by moving C to B and then playing second on B C X . If
Left moves first in K C X to A C X , then Right can win first from that position, since
A C X is a possible first Left move from B C X .
Theorem 3.5 is a powerful result. Since the equivalences hold modulo all alternat-
ing games, and since every alternating game becomes one-handed after the first move,
these equivalences are pertinent to future research in the universe A. As promised,
Theorem 3.5 also allows us to prove the missing chain of domination; we can then
combine the results of Theorems 1.5 and 1.7 (below) to obtain the partial orders illus-
L \ P C N \ LC

N \P
C L \ RC N \P
C R \ LC

N \ RC
R \ PC

C B

0 A 0 A

B C
Left options (right-handed games) Right options (left-handed games)
Figure 1.3. The partial orders (modulo A) given by Theorems 1.5 and 1.7.
trated in Figure 1.3. As indicated by the figure, the game 0 is incomparable or fuzzy
with both A and A. To see this, note that 0 C 2 P while A C 2 N .
Theorem 1.7. If G 2 N \ RC ; H 2 L \ P C ; and K 2 L \ RC are one-handed

alternating games, then G 5 K 5 H modulo A.
Proof. By Theorem 1.6 we need only show B 5 A 5 C . As in the proof of Claim 2

above, we immediately have B 5 A, since Right can tie his hand in B D j 0; A
and pretend he is playing with the position A D j 0; that is, B is at least as good as
A for Right.
To see that A 5 C , let X be any sum of alternating games, and suppose Left wins
A C X playing first. Then we know Left must be able to win playing second from
A C X LR:::L , for any X LR:::L obtained from X under optimal play. In C C X , Left
plays as usual until Right moves from C CX LR:::L to B CX LR:::L ; Left then moves to
ACX LR:::L and wins playing second from there. A similar argument shows Left wins
C C X playing second whenever she wins A C X playing second, and so A 5 C .
We have nearly finished our analysis of one-handed games in the alternating uni-
verse. The final result of this section addresses one of the main difficulties in misre
game theory: the lack of inverses under disjunctive sum. As mentioned in the introduc-
tion, we do not generally have G D G in misre games as we do for normal games.
Luckily for us, one-handed alternating games do possess this very convenient property.
Theorem 1.8. If G is a one-handed alternating game, then G C G 0 (mod A).

Proof. This is trivially true for games in N \ P C . For any other (say, left-handed)
game G, we have G is equivalent (modulo A) to A; B; or C . Thus, it suffices to show
A C A 0, B C B 0, and C C C 0. Let X be be any sum of alternating
games, and suppose without loss of generality that Left wins X . Then in A C A C X ,
Left plays his winning strategy in X , responding in A C A (and thereby bringing it to
zero) only if Right plays there. If this happens before Left runs out of move in X , then
play resumes in X and Left wins as usual. Otherwise Right does not play in A C A
before Left wins X , at which point Left plays in A, bringing the game to a sum of
right-handed components including A 2 L . Left then wins by Lemma 1.2. Notice
this argument works whether Left wins X playing first or playing second or both.
For B C B C X , the argument works as above; if Right plays in B, then Left copies
her in B, bringing those components to 0 C 0 or A C A 0, and resumes play in X .
If Left runs out of moves in X before this happens, she plays in B to A and wins by
Lemma 1.2. Finally, in C C C C X , if Right plays in C , then Left copies in C to get
B C B 0. If Right does not play in C , then after Left runs out of moves in X he
plays in C to B. The game is now a sum of right-handed positions including B 2 L ,
so once again Left wins by Lemma 1.2.
Theorem 1.8 shows that a disjunctive sum forms the set of one-handed alternating
games into a group; G is in fact G. We use the notations interchangeably for the
remainder of the paper and write kG to represent k copies of the game G.
Together, Theorems 1.6 and 1.8 show that any sum of one-handed games in the al-
ternating universe A can be written as aACbB CcC for integers a; b; c. In Section 1.3
we determine the outcome of such a sum, and in Section 1.4 we use this information
to determine the misre monoid of O.
1.3 Outcomes
If G is a sum of one handed games in A; then G aA C bB C cC for integers a; b; c.
We can therefore represent any such sum as an ordered triple .a; b; c/. Figure 1.4
illustrates the possible Left and Right options from G (with the conditions for each
indicated below the horizontal line), and Theorem 1.9 establishes the misre outcome
of G.
Theorem 1.9. Let G D aA C bB C cC . Then

8

<L ; if a C c < 0;

o .G/ D N ; if a C c D 0;

:
R ; if a C c > 0:
Proof. The zero game is a next-player win under misre rules and has a C c D 0.
Assume the outcomes above hold for all followers of G D .a; b; c/.
.a; b; c/
.a1;b;c/ .a;b1;c/; .a;b1;c1/ .aC1;b;c/ .a;bC1;c/; .a;bC1;cC1/
.a1;b1;c/ .aC1;bC1;c/
if a>0; if b>0; if c>0; if a<0; if b<0; if c<0;
A7!0 B7!0;A C 7!B A7!0 B7!0;A C 7!B
Figure 1.4. Left and Right options from G D aA C bB C cC .
If a C c < 0; then there exists at least one copy of A or C in the sum, so Right has
a move going first to .a C 1; b; c/ or .a; b C 1; c C 1/. At best then, Right can increase
the value of a C c to zero, giving Left a next-win position by induction. If Left plays
first, she can play to either .a 1; b; c/ or .a; b 1; c 1/, and so can guarantee a C c
remains negative and by induction leaves a Left-win position. Since Left wins playing
first or second, G 2 L when a C c < 0. A symmetric argument shows G 2 R
when a C c > 0.
If a C c D 0; then Left going first either has no moves and wins immediately, or has
a move in a copy of A; B, or C . From Figure 1.4 we see that Left can always decrease
either a or c by 1, thereby moving to a C c < 0 and winning by induction. Right wins
playing first by symmetry, and so G 2 N when a C c D 0.
This theorem will be very useful in analyzing the two-handed universe. For the
present chapter, it shows us that the equivalence classes among one-handed games
collapse even further when the universe is restricted to O. Notice that the integer b
does not influence the outcome of G D aACbB CcC . Thus, for X a sum of only one-
handed alternating games, we have o .X / D o .B C X /. This gives the following
corollaries.
Corollary 1.10. If O is the closure of one-handed alternating games, then B 0

(mod O) and C A (mod O).
Corollary 1.11. If G is a sum of one-handed games then G aA (mod O) for some

integer a.
Note that we do have to restrict the universe to O to get these results. In the two-
handed universe, B is distinguished from 0 with the game D 0 j 0: Left can win
B C going first, but loses playing first on .
The last thing to note in our analysis of O is that the positions aA and a0 A are
equivalent if and only if a D a 0 . If a 6D a0 ; then the games are distinguished by aA,
since aA C .aA/ is in N while a0 A C .aA/ is in L or R .
1.4 The Misre Monoid

The previous section gives us everything we need to describe the one-handed alternat-
ing universe O. Let represent the equivalence class of the game A; and let e represent
the equivalence class of the game 0. The misre monoid of O is given by
MO D he; ; 1 j 1 D ei;
N D 0; P D ;; R D a j a 2 N; L D a j a 2 N:
As noted earlier, this monoid is actually a group; the mapping aA 7! a for a 2 Z

shows MO .Z; C/.
Acknowledgments. The first two authors would like to thank NSERC for partial sup-
port of this research.
References
[1] M. H. Albert, R. J. Nowakowski, and D. Wolfe, Lessons in Play. Peters, Ltd., Nattick,
MA, 2007.
[2] M. R. Allen, An Investigation of Misre Partizan Games. PhD thesis, Dalhousie Univer-
sity, 2009.
[3] M. R. Allen, Peeking at partizan misre quotients, in: Games of No Chance 4, to appear.
[4] N. A. McKay, R. Milley, and R. J. Nowakowski, Hackenbush sprigs, preprint, 2012.
[5] G. A. Mesdal and P. Ottaway. Simplification of partizan games in misre play, Integers 7
(2007), #G06.
[6] P. Ottaway, Combinatorial Games with Restricted Options under Normal and Misre
Play, PhD Thesis, Dalhousie University, 2009.
[7] T. E. Plambeck, Taming the wild in impartial combinatorial games, Integers 5 (2005),
#G5, Comb. Games Sect.
[8] T. E. Plambeck and A. N. Siegel, Misre quotients for impartial games, Journal of Com-
binatorial Theory, Series A 115(4) (2008, 593 622.
[9] A. N. Siegel. Misre canonical forms of partizan games. a.arXiv:math/0703565v1, to ap-
pear in Games of No Chance 4.
Author information
Rebecca Milley, Department of Mathematics and Statistics, Dalhousie University, Halifax,
Nova Scotia, Canada.
Email: rkmilley@mathstat.dal.ca
Richard J. Nowakowski, Department of Mathematics and Statistics, Dalhousie University,

Halifax, Nova Scotia, Canada.
Email: rjn@mathstat.dal.ca
Paul Ottaway, Department of Mathematics, Thompson Rivers University, Kamloops, British

Columbia, Canada.
Email: pottaway@tru.ca
Chapter 2
Images of C-Sets and Related Large Sets under

Nonhomogeneous Spectra
Neil Hindman and John H. Johnson
Abstract. Let > 0 and 0 < < 1. Define g; W N ! N by g; .n/ D b n C c. The set
g; .n/ W n 2 N is called the nonhomogeneous spectrum of and . By extension, we refer
to the maps g; as spectra. Bergelson, Hindman, and Kra showed that if A is an IP -set, a
central set, an IP -set, or a central -set, then g; A is the corresponding object. We extend
this result to include several other notions of largeness: C -sets, J -sets, strongly central sets,
and piecewise syndetic sets. Of these, C -sets are particularly interesting, because they are the
sets which satisfy the conclusion of the central sets theorem (and so have many of the strong
combinatorial properties of central sets) but have a much simpler elementary description than
do central sets.
Keywords. Central Set, Spectrum, IP-Set.
Mathematics Subject Classification 2010. 05D10.
2.1 Introduction
Given a positive real number , the set bnc W n 2 N is called the spectrum of .
(We write N for the set of positive integers and ! for the set of nonnegative integers.)
Numerous results about spectra were derived by Skolem [18] and Bang [1]. (For a nice
presentation of these results see [16].) In [19] Skolem introduced the more general sets
bn C c W n 2 N, determining for example when two such sets can be disjoint.
In terminology introduced by Graham, Lin, and Lin [9], the set bn C c W n 2 N
is called a nonhomogeneous spectrum. By extension, we refer to the function g; W
N ! N defined by g; .n/ D bn C c as a nonhomogeneous spectrum.
In [4], V. Bergelson, B. Kra, and the first author of this paper showed that if > 0,
0 < < 1, and a subset A of N is large in any of several senses, then so is g; A. The
main reason anyone should care about this fact is that it provides explicit nontrivial
examples of sets with these largeness properties. For example, a set A is an IP -set
P hxn inD1 is a sequence in N, F S.hxn inD1 / \ A ;, where
1 1
if and only if whenever
F S.hxn inD1 / D n2F xn W F 2 Pf .N/. Given any set X , Pf .X / is the set of
1
finite nonempty subsets of X . It is trivial that for each n 2 N, nN is an IP -set

just take n terms of any sequence which are congruent mod n and add them. It is not
p
so trivial that b 7 bn C 1e c C 23 c W n 2 N is an IP -set. This is because it is
gp7;2=3 g;1=e N.
The notions of largeness which were shown in [4] to be preserved by spectra were,
in addition to the IP -sets mentioned above, IP -sets, central sets, central* sets, -
sets, and -sets. The set A is an IP -set if and only if it contains F S.hxn i1 nD1 / for
some sequence hxn i1 nD1 . (Thus a set is an IP -set if and only if it has nontrivial
intersection with each IP -set.) The set A is a -set if and only if it contains y x W
x; y 2 B and x < y for some infinite set B N. And a -set is one which has
nontrivial intersection with each -set.
Central sets in N were introduced by Furstenberg in [8] and were defined using no-
tions from topological dynamics. They have a much simpler characterization in terms
of the algebra of N. This characterization was obtained in [2] with the assistance
of B. Weiss. (It was shown to be equivalent in arbitrary semigroups by H. Shi and
H. Yang [17].) In order to discuss this we give a very brief overview of that algebraic
structure. The reader is referred to [13] for an elementary introduction to the subject.
Detailed historical references can also be found in that book. Most of the basic facts
mentioned in the next two paragraphs are not due to either of the authors of [13].
Let .S; C/ be a semigroup, not necessarily commutative, with the discrete topology.
(We denote the operation by C because we are primarily interested in the semigroup
.N; C/.) The StoneCech compactification S of S can be taken to be the set of ultra-
filters on S , with the points of S identified with the principal ultrafilters. The topology
on S has a basis consisting of A W A S where A D p 2 S W A 2 p. The
operation extends to S making .S; C/ a right topological semigroup (so for each
p 2 S the function p W S ! S defined by p .q/ D q C p is continuous) with
S contained in its topological center (so for each x 2 S , the function x W S ! S
defined by x .q/ D x C q is continuous). Given p; q 2 S and A S , A 2 p C q
if and only if x 2 S W x C A 2 q 2 p where x C A D y 2 S W x C y 2 A.
The reader should be cautioned that .S; C/ is very unlikely to be commutative; the
center of .N; C/ is N.
Any compact Hausdorff right topological semigroup T has a smallest two-sided
ideal, K.T /, which is the union of all of the minimal right ideals of T and is also the
union of all the minimal left ideals of T . The intersection of any minimal right ideal
and any minimal left ideal is a group, and any two such groups are isomorphic. (In
.N; C/ there are 2c minimal left ideals and 2c minimal right ideals.) In particular
there are idempotents in K.T /. An idempotent in T is in K.T / if and only if it is
minimal with respect to the ordering of idempotents, wherein p q if and only if
p D p C q D q C p. Idempotents in K.T / are referred to as minimal idempotents.
A subset A of S is central if and only if A is a member of a minimal idempotent of
S . And A is central* if and only if it has nontrivial intersection with each central set;
equivalently, A is central* if and only if it is a member of every minimal idempotent
in S . It is a trivial consequence of the definition that central sets are partition regular.
That is, if A is central and A is partitioned into finitely many cells, then one of these
Chapter 2 Images of C -Sets and Related Large Sets under Nonhomogeneous Spectra 17
cells is central. In particular, whenever S is partitioned into finitely many parts, one of
these must be central. From a combinatorial point of view, what is probably the most
important fact about central sets is that they satisfy the central sets theorem. We state
here the commutative version thereof because it is much simpler to state and we are
primarily concerned with .N; C/. We write NS for the set of sequences in S .
Theorem 2.1. Let .S; C/ be a commutative semigroup, and let A be a central subset
of S . There exist functions W Pf .NS / ! S and H W Pf .NS / ! Pf .N/ such that
(1) if F; G 2 Pf .NS / and F G, then max H.F / < min H.G/; and
(2) whenever m 2 N, G1 ; G2 ; : : : ; Gm 2 Pf .NS /, G1 G2 : : : Gm , and for
P P
each i 2 1; 2; : : : ; m, fi 2 Gi , one has m
iD1 .Gi /C t2H.Gi / fi .t / 2 A.
Proof. [7, Theorem 2.2].
The following is the original central sets theorem as proved by Furstenberg [8,
Proposition 8.21].
Corollary 2.2. Let l 2 N and for each i 2 1; 2; : : : ; l, let hyi;n i1

nD1 be a sequence
in Z. Let A be a central subset of N. Then there exist sequences han i1 nD1 in N and
hHn i1
nD1 in Pf .N/ such that
(1) for all n, max Hn < min HnC1 , and
P P
(2) for all F 2 Pf .N/ and all i 2 1; 2; : : : ; l, n2F .an C t2Hn yi;t / 2 A.
Proof. We first note that A is central in Z because by [13, Theorem 1.65 and Exer-
cise 4.3.5], K.N/ D K.Z/ \ N. Pick W Pf .NZ/ ! Z and H W Pf .NZ/ !
Pf .N/ as guaranteed by Theorem 2.1.
Let G1 D 0; hy1;n i1 1 1
nD1 ; hy2;n inD1 ; : : : ; hyl;n inD1 , where 0 is the function con-
stantly equal to 0. For n 2 N, pick f 2 NN n Gn and let GnC1 D Gn [ f .
For n 2 N, let Hn D H.Gn / and let an D .Gn /. To see that each an 2 N,
P
note that 0 2 Gn so an D .Gn / C t2H.Gn / 0 .t / 2 A N. Conclusion (1) holds
directly. To verify conclusion (2), let F 2 Pf .N/ and i 2 1; 2; : : : ; l. Then for
P P P
each n 2 F , hyi;n i1 2 Gn so n2F .an C t2Hn yi;t / D .Gn / C
P nD1 n2F
y
t2H.Gn / i;t 2 A.
From Corollary 2.2, Furstenberg deduced several facts about central subsets of N,
including the fact that any partition regular system of homogeneous linear equations
with coefficients from Q has solutions in any central set; see [13, Chaps. 1416] for
many other examples of the strong properties enjoyed by central subsets of N.
The algebraic description of central sets is very simple and quite easy to work with.
There is an elementary description of central sets which was produced in [12] (or
see [13, Theorem 14.25]). However, that description is based on the notion of a collec-
tionwise piecewise syndetic family of sets a notion which is extremely complicated.
The masochistic reader is referred to [13, Definition 14.19].
For many of the facts about central sets, what is important is that they satisfy the cen-
tral sets theorem. We define a set to be a C -set if and only if it satisfies the conclusion
of the central sets theorem. We have not stated the general version of the central sets
theorem. In the case of commutative semigroups, the definition becomes as follows.
Definition 2.3. Let .S; C/ be a commutative semigroup. A set A S is a C -set if

and only if there exist functions W Pf .NS / ! S and H W Pf .NS / ! Pf .N/ such
that
(1) if F; G 2 Pf .NS / and F G, then max H.F / < min H.G/; and
(2) whenever m 2 N, G1 ; G2 ; : : : ; Gm 2 Pf .NS /, G1 G2 : : : Gm , and for

P P
each i 2 1; 2; : : : ; m, fi 2 Gi , one has miD1 .Gi / C t2H.Gi / fi .t / 2 A.
There is an elementary characterization of C -sets [14] which is very similar in form

to the elementary description of central sets. The crucial distinction is that the notion
of a family of sets being collectionwise piecewise syndetic is replaced by the notion
of a single set being a J -set.
Definition 2.4. Let .S; C/ be a commutative semigroup. A set A S is a J -set if

and only if whenever F 2 Pf .NS /, there exist a 2 S and H 2 Pf .N/ such that for
P
each f 2 F , a C t2H f .t / 2 A.
Definition 2.5. Let S be a semigroup. Then J.S / D p 2 S W .8A 2 p/.A is a

J -set/.
Of course one defines C -sets and J -sets analogously to IP -sets and central*
sets. That is, a set is a C -set if and only if it has nonempty intersection with every
C -set, and a set is a J -set if and only if it has nonempty intersection with every J -set.
Other notions of size with which we will be concerned in this chapter include strongly
central sets (abbreviated by S C ), piecewise syndetic sets (abbreviated by PS ), AP -
sets, syndetic sets, and thick sets. A subset of N is strongly central if and only if it is a
member of some idempotent in every minimal left ideal of N. (Since every left ideal
contains a minimal left ideal, this is equivalent to being a member of an idempotent
in every left ideal of N.) In .N; C/, a set A is piecewise syndetic if and only if there
is a bound b such that there are arbitrarily long blocks of N in which A has no gaps
longer than b. That is .9b 2 N/.8k 2 N/.9x 2 N/.x C 1; x C 2; : : : ; x C k
Sb
tD1 t C A/. In .N; C/, a set A is an AP -set if and only if it contains arbitrarily
long arithmetic progressions, is syndetic if and only if it has bounded gaps, and is thick
if and only if it contains arbitrarily long integer intervals. We say that a set is PS if
and only if it has nonempty intersection with each piecewise syndetic set and is AP
if and only if it has nonempty intersection with each AP -set.
Of the notions of size that we are considering, IP -sets, central sets, C -sets, J -sets,
piecewise syndetic sets, AP -sets, and -sets are all partition regular.
Notice that any J -set in N is an AP -set. In fact, it must contain arbitrarily long
arithmetic progressions with increment taken from the finite sums of any prespecified
sequence in N. To see this, let a sequence hxn i1 nD1 in N be given and let l 2 N. Let

F D htxn i1 nD1 W t 2 1; 2; : : : ; l :
P
Pick a 2 N and H 2 Pf .N/ such that for each t 2 1; 2; : : : ; l, a C t2H txn 2 A.
P
If d D t2H x t , then

a C t d W t 2 1; 2; : : : ; l A:
Definition 2.6. AP D p 2 N W .8A 2 p/.A is an AP -set/.

In Section 2.2 we will describe the relationship among the various notions of size
that we are considering and also present elementary proofs regarding the preservation
of these notions by g; for those notions for which we have such elementary proofs.
In Section 2.3 we introduce the functions f W N ! T and h D g;1=2 and present
some of the basic facts about these functions.
In Section 2.4 we present some background material on C -sets and J -sets, including
the fact in Theorem 2.20 that in a commutative semigroup, the translate a may be
taken to be a member of the J -set itself. We then present positive results about the
preservation of J -sets, C -sets, and C -sets under spectra. We then show that if P is
any property which is preserved by all spectra (including any of the properties in the
left two columns of Figure 2.1), > 1, and 0 < < 1, then a set A has property P
if and only if g; A has property P . Consequently, if A is an example of something
with one of those properties but not others, so is g; A.
e
In Section 2.5 we show the extent to which g; preserves certain ideals of N,
e
where g; W N ! N is the continuous extension of g; .
2.2 The Various Notions of Size

We begin with a theorem providing characterizations in terms of .N; C/ of each of
the notions of size that we are considering. (Most of these characterizations hold for
arbitrary semigroups.)
We let N D N n N and given T N, E.T / D p 2 T W p C p D p. Given
p 2 N, p is the member of Z generated by A W A 2 p.
Theorem 2.7. Let A N.

(a) A is a -set if and only if there is some p 2 N such that A 2 p C p.
(b) A is an AP -set if and only if A \ AP ;.
(c) A is an IP -set if and only if A \ E.N/ ;.
(d) A is a J -set if and only if A \ J.N/
;.
(e) A is a C -set if and only if A \ E J.N/ ;.
(f) A is a PS -set if and only if A \ K.N/ ;.
(g) A is a central set if and only if A \ E K.N/ ;.
(h) A is a syndetic set if and only if for every left ideal L of N, A \ L ;.
(i) A is a S C -set if and only if for every left ideal L of N, A \ E.L/ ;.
(j) A is a thick set if and only if there exists a left
ideal L of N such that L A.
(k) A is a central* set if and only if E K.N/ A.
(l) A is a PS -set if and only if K.N/
A.
(m) A is a C -set if and only if E J.N/ A.
(n) A is a J -set if and only if J.N/ A.
(o) A is an IP set if and only if E.N/ A.
(p) A is an AP set if and only if AP A.
(q) A is a -set if and only if for all p 2 N , A 2 p C p.
Proof. Statements (g) and (i) are the definitions of central and S C respectively.
Statements (a) and (q) follow from [3, Lemma 1.9(a) and (j)] respectively.
Statement (b) follows from [13, Theorem 3.11] and the fact that AP -sets are parti-
tion regular.
Statements (c), (d), (e), and (f) follow from [13, Theorems 5.12, 14.14.7, 14.15.1,
and 4.40], respectively.
Statements (h) and (j) follow from [5, Theorem 2.9(d) and (c)], respectively.
Statements (k), (l), (m), (n), (o), and (p) follow from statements (g), (f), (e), (d), (c),
and (b), respectively.
Figure 2.1 shows the relations that hold among the notions of size that we are con-
sidering. After noting that, by [13, Theorems 14.4.4 and 14.5], J.N/ and AP are
ideals of .N; C/ and, as we have already observed, each J -set is an AP -set so that
J.N/ AP , each of the implications follows quickly from the characterizations
in Theorem 2.7 except for the fact that IP ) and the corresponding fact that
) IP . (To see that central ) S C and thick ) central, one needs to note
that by [13, Corollary 2.6], each left ideal of N contains a minimal left ideal with an
idempotent, which is therefore an idempotent in K.N/.)
The easiest way to see that any IP -set is a -set (and thus any -set
Pn is an IP -

set) is to note that if F S.hxn inD1 / A and for each n 2 N, yn D tD1 x t , then
1
yn ym W n; m 2 N and m < n A.
AP

IP J

C PS
|
central* |

SC thick

syndetic central

PS C

J IP

AP
Figure 2.1. Diagram of implications
That none of the missing implications is valid is established by the following table
which lists for each property, a subset of N with that property that does not have any
of the other properties except those that it is forced to have because of the implications
in the diagram. In that table, the set D is the set produced in [11] which is a C -set but
has Banach density 0. In particular D \ K.N/ D ;. Since D 2N, D C 1 is not a
-set.
Property Set with no extra properties

2n 2m W n; m 2 N and m < n
AP 22n C m2n C 1 W n; m 2 N and m < n
IP n2F 22n W F 2 Pf .N/
J DC1
C D
PS 2n C 2m 1 W n; m 2 N and m < n
central 2n C 2m W n; m 2 N and m < n
syndetic 2N C 1
p p S
SC 2N \ x 2 N W . 2x b 2x C 1=2c/ 2 1 nD1 1=.2n C 1/; 1=.2n/
thick 2n C m W n; m 2 N and m < n
central* 2N n D
PS N nD
Property Set with no extra properties

C 2N n n2F 22n W F 2 Pf .N/
J N n .22n C m2n C 1 W n; m 2 N and m < n [
n2F 22n W F 2 Pf .N//
IP 2N n 2n 2m W n; m 2 N and m < n
AP N n n2F 22n W F 2 Pf .N/
2N
In most cases it is at least routine, if not obvious, that the set has the specified
property. We will explain now why this is true for J , central, S C , C , and J .
The set D is a member of an idempotent p 2 J.N/, so D C 1 2 p C 1 and
p C 1 2 J.N/. The set 2n C 2m W n; m 2 N and m < n is the intersection of
a thick
set with pa central*pset, and so is central.
S It is a consequence of [6,
Theorem 3.2]
that x 2 N W . 2x b 2x C 1=2c/ 2 1 nD1 1=.2n C 1/; 1=.2n/ , and its com-
plement are both strongly central, and thus its intersection with 2N, which is IP , is
strongly central. (We shall be giving a short P proof of this consequence of [6, Theo-
rem 3.2] in Theorem 2.15 below.) Since n2F 22n W F 2 Pf .N/ is not an AP -set,
it is not a J -set, and so
P
2N n n2F 22n W F 2 Pf .N/
is a C -set. And, by Lemma 2.22 below, 22n C m2n C 1 W n; m 2 N and m < n is

P
not a J -set, and so N n .22n C m2n C 1 W n; m 2 N and m < n [ n2F 22n W
F 2 Pf .N// is a J -set.
In all cases but two it is routine to verify that the listed set does not have any proper-
ties except those it is forced to have by the implications in the diagram. The exceptions
are AP and S C . One needs that the set listed for AP is not J and not . It is trivially
not a -set. The fact that it is not a J -set is Lemma 2.22. One needs that the set listed
for S C is not central*
and notp thick. It p
is trivially not thick.
S It is not central* because

the complement of x 2 N W . 2x b 2x C 1=2c/ 2 1 nD1 1=.2n C 1/; 1=.2n/
is strongly central and thus central.
It was shown in [4, Theorem 6.1] that if > 0, 0 < < 1, P is any of , IP ,
central, central*, IP , or , and A N is a P -set, then so is g; A. We shall
show in the remainder of this chapter that the same assertion holds if P is any of the
properties in the left two columns of Figure 2.1. And we shall show that as long as
1, the properties in the right column of Figure 2.1 are also preserved by g; , but
are not if > 1.
In the remainder of this section, we establish those results for which we have ele-
mentary proofs, not requiring the algebra of N.
Theorem 2.8. Let A N, let > 0, and let 0 < < 1.

(1) If A is syndetic, then so is g; A.
(2) If 1 and A is thick, then so is g; A.
Proof. (1) Pick b 2 N such that for all x 2 !, x C 1; x C 2; : : : ; x C b \ A ;

and pick k 2 N such that < k. We shall show that for all y 2 N, there exists
m 2 g; A such that 0 m y bk. So let y 2 N be given and let x be the largest
element of ! such that g; .x/ < y. Pick t 2 1; 2; : : : ; b such that x C t 2 A and
let m D g; .x C t /. Then m g; .x C 1/ y. Also m .x C t / C and
x C < y so m y < t < bk.
(2) Let b 2 N. We need to show that there is some m 2 N such that m; m C
1; : : : ; m C b g; A. Pick r 2 N such that > 1r , and pick n 2 N such that
n; nC1; : : : ; nClr A. Let m D g; .n/. Since 1, we have that for all x 2 N,
g; .x C 1/ g; .x/ C 1, and so it suffices to show that g; .n C lr/ m C l.
Since m n C and l < lr, this is true.
Theorem 2.9. Let > 1 and let 0 < < 1. Then g; N is not thick. In particular,
none of the properties in the right column of Figure 2.1 are preserved by g; .
Proof. Pick r 2 N such that 1r < 1. Given n 2 N, let k D g; .n/. Then
g; .n C r/ k C r C 1 so at least one of k C 1; k C 2; : : : ; k C r is not in
g; N.
Theorem 2.10. Let A be a piecewise syndetic subset of N, let > 0, and let 0 <
< 1. Then g; A is piecewise syndetic.
Proof. Pick b 2 N such that for all w 2 N there exists x 2 N such that x; x C
S
1; : : : ; x C w btD0 .t C A/. Let m D db C e. We claim that for all y 2 N
S
there exists z 2 N such that z; z C 1; : : : ; z C y m tD0 .t C g; A/. To this
1
end, let y 2 N and let w D d ye. Pick x 2 N such that x; x C 1; : : : ; x C w
Sb Sm
tD0 .t C A/. Let z D g; .x/. We claim that z; z C 1; : : : ; z C y tD0 .t C
g; A/. To this end, let k 2 0; 1; : : : ; y be given. Now y w < y C so
z Cy .x Cw/C < z C1Cy C. Thus z Cy g; .x Cw/. Pick the first l 2 !
such that z C k g; .x C l/, and note that l 2 0; 1; : : : ; w. Pick t 2 0; 1; : : : ; b
such that x C l C t 2 A. Let v D g; .x C l C t / .z C k/. Then z C k C v 2 g; A
so it suffices to show that 0 v m. Since g; .x C l C t / g; .x/ D z, we
have v 0. By the choice of l, z C k > g; .x C l 1/ so x C l < z C k.
Consequently xCl Ct C < zCkCt C zCkCbC so g; .xCl Ct / <
z C k C b C z C k C m as required.
Theorem 2.11. Let A be an AP -set, let > 0, and let 0 < < 1. Then g; A is
an AP -set.
Proof. Let r 2 N n 1; 2. We need to show that g; A contains a length r arith-

metic progression. By van der Waerdens theorem, pick m 2 N such that whenever a
length m arithmetic progression is two-colored it contains a length r monochromatic
arithmetic progression. For i 2 0; 1, let Bi D x 2 N W g; .x/ C 2i x C <
g; .x/ C iC1
2 . Pick a length m arithmetic progression P contained in A; and pick
i 2 0; 1 and a length r arithmetic progression a; a Cd; : : : ; a C.r 1/d contained
in P \ Bi . We claim that
g; .a/; g; .a C d /; : : : ; g; .a C .r 1/d /
is an arithmetic progression. (It is contained in g; P and is therefore contained in
g; A.)
To see this, let k D g; .a/ and let c D g; .a C d / k. Now,
i i C1
kC a C < k C and
2 2
i i C1
k C c C .a C d / C < k C c C ; so
2 2
1 1
./ c 2 < d < c C 2 :
We claim that for each t 2 0; 1; : : : ; r 1, g; .aCt d / D kCt c. This is trivially true
for t 2 0; 1, and so assume that t 2 1; 2; : : : ; r 2 and that g; .a Ct d / D k Ct c.
Let m D g; .a C .t C 1/d /. Then,
i i C1
k C tc C .a C t d / C < k C t c C and
2 2
i i C1
m C .a C .t C 1/d / C < m C ; so
2 2
1 1
./ m k tc 2 < d < m k t c C 2 :
Combining ./ and ./ we have
1 1
c < m k tc C and
2 2
1 1
m k tc < c C :
2 2
Thus we have that k C .t C 1/c 1 < m < k C .t C 1/c C 1 and so k C .t C 1/c
1 D m.
e
Corollary 2.12. Let > 0 and let 0 < < 1. Then g; AP AP .
e
Proof. Let p 2 AP and let A 2 g; .p/. Pick B 2 p such that g; B A. Then B
is an AP -set, and by Theorem 2.11, g; B is an AP -set, and thus A is an AP -set.
2.3 The Functions f and h

We consider the circle group T to be R=Z and represent the points of T by points in
the interval 12 ; 12 /. (That is, if x 2 12 ; 12 / we write x to represent the coset x C Z.)
The reason we do this is that we will be dealing extensively with intervals in R, and
so it is convenient to have members of T represented by elements of R.
Given > 0, we let h D g;1=2 , so that for x 2 N, h is the nearest integer
to x (with ties broken by rounding up). We define f W N ! T by, for x 2 N,
f .x/ D x h .x/; and we let f f W N ! T be the continuous extension of f .
e
Similarly, g; W N ! N and h f W N ! N are the continuous extensions of
g; and h , respectively.
f
Definition 2.13. Let > 0. Then Z D p 2 N W f .p/ D 0.
We gather some basic facts about the objects we have just defined.
Lemma 2.14. Let > 0; and let 0 < < 1. The function f f is a homomorphism
from N into T , and consequently each idempotent is in Z The restriction of h f
e
.
to Z is an isomorphism onto Z1= , whose inverse is the restriction of h1= to Z1= .
e f
For all p 2 Z , g; .p/ D h .p/.
Proof. The first conclusion follows from [13, Corollary 4.22]. The second and third
conclusions are [4, Theorem 5.10] and [4, Theorem 5.8], respectively.
As promised earlier, we provide a short proof of the consequence of [6, Theo-

rem 3.2] which we used in the previous section.
Theorem 2.15. If U is an open subset of .0; 12 / with 0 2 c`U and > 0 is irrational,
then f1 U is strongly central.
Proof. Let L be a minimal left ideal of N. Pick sequences han i1 1

nD1 and hbn inD1 in
S1
.0; 12 / converging to 0 with anC1 < bn < an for each n such that nD1 .bn ; an / U .
For each n 2 N pick cn and dn such that bn < dn < cn < an and pick by Kroneckers
theorem xn such that dn < f .xn / < cn . Pick p 2 N \ xn W n 2 N; and note that
p 2 Z . Pick by [13, Theorem 1.61] an idempotent q 2 L \ .p C N/; and pick
r 2 N such that q D p C r. By Lemma 2.14 f e .q/ D fe .p/ C f
e .r/; and thus
r 2 Z .
We claim that xn W n 2 N y 2 N W y C f1 U 2 r; so that f1 U 2
p C r as required. Let n 2 N, and let D mindn bn ; an cn . Then z 2 N W
f .z/ 2 .; / 2 r and z 2 N W f .z/ 2 .; / .xn C f1 U .
A major contrast between the preservation of largeness notions established in [4]

and the results of our next section is that each of the notions considered in [4], namely
IP -sets, -sets, central sets, IP -sets, central* sets, and -sets, is determined by
e
members of Z , so that g; agrees with an isomorphism on Z at those points. The
notion of strongly central had not yet been introduced at the time of the writing of [4].
We pause to observe that the methods of that paper are sufficient to show that it is
preserved by the functions g; .
Theorem 2.16. Let A be a strongly central subset of N, let > 0, and let 0 < < 1.
Then g; A is strongly central.
Proof. Let L be a minimal left ideal of N. We need to show that there is an idem-
potent in L \ g; A. Now L \ Z1= ; because any idempotent in L is in this
intersection, and so L\Z1= is a minimal left ideal of Z1= by [13, Theorem 1.65(2)].
e
Therefore h1= L \ Z1= is a minimal left ideal of Z by Lemma 2.14. Again by [13,
Theorem 1.65(2)], we may pick a minimal left ideal M of N such that h1= L \ e
Z1= D M \ Z . Pick an idempotent p 2 M \ A. Then p 2 Z so h f .p/ 2
f
h e e f
h1= L \ Z1= D L \ Z1= . Since A 2 p, g; A 2 g; .p/ D h .p/. Thus
f
h f
.p/ is an idempotent in L and g; A 2 h .p/.
In the rest of this chapter we need to consider points of N which f can take
anywhere in T . The following lemma will simplify some of the computations. In this
lemma we give specific values for , but we only really care that exists.
Lemma 2.17. Let > 0, let p 2 N, and let

D f f 1 3
.p/. There exist 2 2 ; 4 and
> 0 such that for all 2 .0; , x 2 N W g; .x/ C
< x < g; .x/ C
C
2 p.
Proof. Assume first that

12 . Let D 12 and let D min 12
; 12 C
. Let
2 .0; and let A D x 2 N W f .x/ 2 .
;
C /. Then by the continuity of
f
f , A 2 p. We note that A D x 2 N W g; .x/ C
< x < g; .x/ C
C .
Now assume that
D 12 . Let D 34 and let D 14 . Let 2 .0; and let A D x 2
N W f .x/ 2 12 ; 12 C / [ x 2 N W f .x/ 2 . 12 ; 12 /. Then by the continuity
f
of f , A 2 p. We claim that A x 2 N W g; .x/C
< x < g; .x/C
C.
(Actually, equality holds, but we do not care.) To this end, let x 2 A and let k D h .x/.
Assume first that f .x/ 2 12 ; 12 C /. Then 12 x k < 12 C , and so k <
x C 34 < k C1. Thus, k D g; .x/; and so g; .x/C
< x < g; .x/C
C.
Now assume that f .x/ 2 . 12 ; 12 /. Then 12 < xk < 12 . Thus, kC1 D g; .x/;
and so g; .x/ C
< x < g; .x/ C
C .
Lemma 2.18. Let F 2 Pf .NN/, let > 0, and let > 0. There exists H 2 Pf .N/
P
such that for all k 2 F , f t2H k.t / 2 .; /.
Proof. Let A D x 2 N W f .x/ 2 . 2 ; 2 /. Since f f is a homomorphism, for any

idempotent in N, f f .p/ D 0; and thus by Theorem 2.23, A is a C -set, and therefore
a J -set. Pick by Theorem 2.20 some a 2 A and H 2 Pf .N/ such that for all k 2 F ,
P P
a C t2H k.t / 2 A. Then, given k 2 F , f .a/ C f k.t / D f a C
P
P t2H
t2H k.t / 2 . 2 ; 2 / and f .a/ 2 . 2 ; 2 /, so f t2H k.t / 2 .; /.
nD1 is an increasing sequence in Pf .N/; we mean that

When we write that hHn i1
for each n 2 N, max Hn < min HnC1 .
Lemma 2.19. Let F 2 Pf .NN/, let > 0, and let > 0. There is an increas-
ing sequence hHn i1 in P .N/ such that, for each n 2 N and each k 2 F ,
P nD1 f
f t2Hn k.t / 2 . 2n ; 2n /.
Proof. Pick H1 as guaranteed by Lemma 2.18 for 2 . Now let n 2 N and assume that
H1 ; H2 ; : : : ; Hn have been chosen. Let l D max Hn . For k 2 F define rk 2 NN
by, for t 2 N, rk .t / D k.l C t /. Pick by Lemma 2.18, L 2 Pf .N/ such that for all
P
k 2 F , f t2L rk .t / 2 . 2n ; 2n /. Let HnC1 D l C L.
2.4 Preservation of J -Sets, C -Sets, and C -Sets

We begin by showing as promised that in a commutative semigroup the translate in
the definition of J -sets can be taken to be in the given J -set.
Theorem 2.20. Let .S; C/ be a commutative semigroup, and let A be a J -set in S .

For each F 2 Pf .NS /, there exist a 2 A and H 2 Pf .N/ such that for each f 2 F ,
P
a C t2H f .t / 2 A.
Proof. Let F 2 Pf .NS / be given and pick c 2 S . Denote by c the sequence con-
stantly equal to c. For f 2 F define gf 2 NS by, for each n 2 N, gf .n/ D c C f .n/.
Let K D c [ gf W f 2 F . Then K 2 Pf .NS / so pick b 2 S and H 2 Pf .N/
P P
such that b C t2H c 2 A and for each f 2 F , b C t2H gf .t / 2 A. Let
P P
a D b C t2H c. Then for each f 2 F , a C t2H f .t / 2 A.
Lemma 2.21. Let A D 22n C m2n C 1 W n; m 2 N and m < n. If a; d 2 N and

a; n 2 N such that a; a C d; a C 2d
2na C d; an C 2d A, then there is some
2 C m2 C 1 W m 2 1; 2; : : : ; n 1 .
Proof. Suppose first we have some k; n 2 N with k < n, some r 2 1; 2; : : : ; k 1,

and some s 2 1; 2; : : : ; n1 such that a D 22k Cr2k C1 and aCd D 22n Cs2n C1.
Then a 22n2 C .n 2/2n1 C 1 22n2 C 2n1 2n1 D 22n1 and a C d >
22n so d > 22n1 . Thus a C 2d > 22n C 22n1 > 22n C .n 1/2n C 1. Also,
d < a C d 22nC1 so a C 2d < 22nC2 . Since 22n C .n 1/2n C 1 < a C 2d <

22nC2 , a C 2d A.
Thus we must have some n 2 N and r; s 2 1; 2; : : : ; n 1 such that a D 22n C
r2 C 1 and a C d D 22n C s2n C 1. So
n
d D .s r/2n and a C 2d D 22n C .2s
r/2n C 1 < 22nC2 . Therefore a C 2d 2 22n C m2n C 1 W m 2 1; 2; : : : ; n 1 .
Lemma 2.22. Let A D 22n C m2n C 1 W n; m 2 N and m < n. Then A is not a

J -set.
Proof. Suppose that A is a J -set. Let
F D h2t i1
tD1 ; h2
tC1 1
i tD1 ; h22t i1
tD1 ; h2
2tC1 1
i tD1 :
Pick by Theorem 2.20 some a 2 A and H 2 Pf .N/ such that for all f 2 F , a C
P
t2H f .tP/ 2 A. Pick n 2 N and r 2 1; 2; : : : ; n 1 such that a D 22n C r2n C 1.
t
P 2t
Let b D t2H 2 and let d D t2H 2 . Then a; a C b; a C 2b A and
a; a C d; a C 2d A, and so by Lemma 2.21 we have a; a C b; a C 2b; a C
d; a C 2d 22n C m2n C 1 W m 2 1; 2; : : : ; n 1 . Pick m 2 1; 2; : : : ; n 1
P
such that a C b D 22n C m2n C 1. Then b D .m r/2n so 2n divides t2H 2t ,
and thus min H n. But then d 22n , and so a C d > 22n C .n 1/2n C 1,
a contradiction.
Theorem 2.23. Let S be a semigroup. Then J.S / is a closed two-sided ideal of S ,

and a subset A of S is a C -set if and only if there is an idempotent p 2 J.S / \ A.
Proof. [7, Theorems 3.5 and 3.8].
Thus the relationship between C -sets and J -sets is very similar to the relation-
ship between central sets and piecewise syndetic sets: however, there are a few con-
trasts. Firstly, J.S / is closed, while K.S / is commonly not closed. (In particular
K.N/ is not closed.) By [13, Corollary 4.41], p 2 S W .8A 2 p/.p is piecewise
syndetic/ D c`K.S /. Secondly, K.T / makes sense in an arbitrary compact Haus-
dorff right topological semigroup. We know of no reasonable meaning for J.T / in
that generality.
Theorem 2.24. Let S be a semigroup. If B [ C is a J -set, then either B or C is a

J -set. Consequently, a set A S is a J -set if and only if A \ J.S / ;.
Proof. The first conclusion is [15, Theorem 2.14]. The second then follows from [13,
Theorem 3.11].
We show now that spectra preserve J -sets. As we remarked previously, this proof
f
is complicated by the fact that f can take members of J.N/ anywhere in T .
Theorem 2.25. Let A be a J -set in N, let > 0, and let 0 < < 1. Then g; A is
a J -set.
Proof. Let F 2 Pf .NN/ be given. Pick by Theorem 2.24 some p 2 J.N/ such that
A 2 p. Let
D f f .p/. Pick and as guaranteed by Lemma 2.17. Pick > 0 such
that and, if
C 0; 1, then .
C ;
C C / \ 0; 1 D ;.
Pick by Lemma 2.19 an increasing sequence hHn i1
P nD1 in Pf .N/ such that for each

n 2 N and each k 2 F , f1= t2Hn k.t / 2 . 32n ; 32n /. For k 2 F and n 2 N,
P
let rk .n/ D t2Hn k.t /. We claim that
P
./ for each k 2 F and each L 2 Pf .N/, we have h n2L .h1= rk /.n/ D
P P
n2L rk .n/ and f n2L .h1= rk /.n/ 2 . 3 ; 3 /.
To see this, let k 2 F and L 2 Pf .N/ be given. For n 2 L, let xn D .h1=

rk /.n/. Then f1= rk .n/ D 1 rk .n/ xn so xn 32 1
n < rk .n/ < xn C 32n so
P P
xn 32 n < rk .n/ < xn C 32n . Therefore, n2L xn 3 < n2L rk .n/ <
P P P P
n2L xn C 3 and consequently n2L rk .n/ 3 < n2L xn < n2L rk .n/C 3
as required for ().
We have that x 2 N W g; .x/ C
3 < x < g; .x/ C
C 3 2 p and, since
J.N/ N n N, x 2 N W x > 3 2 p. Let

C1 D x 2 N W x > 3 and g; .x/ C
3 < x < g; .x/ C
and let
C2 D x 2 N W x > 3 and g; .x/ C
x < g; .x/ C
C 3 :
Pick j 2 1; 2 such that Cj 2 p. Now A \ Cj 2 p so A \ Cj is a J -set. Pick by
Theorem 2.20 some a 2 A \ Cj and some L 2 Pf .N/ such that for each k 2 F ,
P
a C n2L .h1= rk /.n/ 2 A \ Cj .
Let x D g; .a/; and for k 2 F let
P
zk D g; a C n2L .h1= rk /.n/ , let
P
wk D g; a C n2L .h1= rk /.n/ , and let
P
y k D h n2L .h1= rk /.n/ :
We claim that it suffices to show that there is some b 2 N such S that zk D yk C b

for each k 2 F . Assume thatP we have done this, and let M D n2L Hn . We claim
that for each k 2 F , b C t2M k.t / D zk which will suffice since zk 2 g; A.
P
Thus, let k 2 F be given. By () zk b D yk D h .h1= rk /.n/ D
P P P P n2L
n2L rk .n/ D n2L t2Hn k.t / D t2M k.t / as required.
We show first that there is some i 2 1; 0; 1 such that for all k 2 F , zk D wk Ci .
Case 1. j D 1. Then,
P
wk C
3 < a C n2L .h1= rk /.n/ < wk C
; and thus
P
wk C
C 3 < a C n2L .h1= rk /.n/ C < wk C
C :
If
C 0, then zk D wk 1. If 0 <
C 1, then 0 <
C 3 so zk D wk .
If 1 <
C , then 1 <
C 3 so zk D wk C 1.
Case 2. j D 2. Then
P
wk C
a C n2L .h1= rk /.n/ < wk C
C 3 ; and thus
P
wk C
C a C n2L .h1= rk /.n/ C < wk C
C C 3 :
If
C < 0, then
C C 3 < 0 so zk D wk 1. If 0
C < 1, then
C C 3 < 1 so zk D wk . If 1
C , then zk D wk C 1.
Now to complete the proof that there is some b 2 N with zk D yk C b for each
k 2 F , we show that x > 2 and wk D yk C x for each k 2 F . We know that
P
wk C
3 < a C n2L .h1= rk /.n/ < wk C
C 3
P
and by (), 3 < 2
n2L .h1= rk /.n/ yk < 3 so wk yk C
3 < a <
wk yk C
C 2
3 . Also xC
3 < a < xC
C 3 . Thus xC
3 < wk yk C
C 3
2
and wk yk C
2
3 < x C
C 3 so x < wk yk < x C so wk yk D x.
Also, 3 < a < x C
C 3 , so x 3.
The proof of Theorem 2.25 uses the fact from Theorem 2.24 that any J -set is a
member of an ultrafilter in J.N/, and therefore uses the axiom of choice. We suspect
that with a bit more work, the proof can be rewritten to only depend on the fact that
J -sets are partition regular. The proof of that fact in [15, Theorem 2.14] is elementary,
using the HalesJewett theorem, but is rather complicated.
e
Corollary 2.26. Let > 0, and let 0 < < 1. Then g; J.N/ J.N/.
e
Proof. Let p 2 J.N/ and let A 2 g; .p/. Pick B 2 p such that g; B A. Then
B is a J -set and by Theorem 2.25, g; B is a J -set and thus A is a J -set.
Corollary 2.27. Let A be a C -set in N, let > 0, and let 0 < < 1. Then g; A
is a C -set.
Proof. Pick by Theorem 2.23 an idempotent p 2 A \ J.N/. Then g; A 2 g; .p/.e
e
By Corollary 2.26, g; .p/ 2 J.N/, while by Lemma 2.14, since p is an idempotent
e f
and is therefore in Z , g; .p/ D h f
.p/ and h .p/ is an idempotent.
While we know that ge is a homomorphism on Z , it is not a homomorphism on

;
N or even on N n N. The following lemma is a partial result in that direction.
Lemma 2.28. Let > 0, let 0 < < 1, let q 2 N n N, and let r 2 Z . Then
e e e
g; .q C r/ D g; .q/ C g; .r/.
e
Proof. We shall show that there is some X 2 q such that for all x 2 X , g; .x Cr/ D
e e
g; .x/ C g; .r/. This will suffice since then the continuous functions g; r and
e e
g .r/ g; agree on a member of q and therefore agree at q. Once we have defined
;
X , we will let x 2 X and produce Y 2 r such that for all y 2 Y , g; .x C y/ D
e e
g; .x/Cg; .y/, so that the continuous functions g; x and g; .x/ g; agree
on a member of r.
Let
D f f 1 3 1
.q/. Pick by Lemma 2.17, 2 2 ; 4 and 2 .0; 2 / such that for all
2 .0; , x 2 N W g; .x/ C
< x < g; .x/ C
C 2 q. We consider
two cases (with three subcases each). Only in cases 1b and 1c does the choice of the
set Y depend on the choice of x.
Case 1.
C 2 0; 1. Let i D
C .
Case 1a. x 2 N W g; .x/C
D x 2 q. Let X D x 2 N W g; .x/C
D x.
Then is rational, since otherwise jX j 1, and thus Y D y 2 N W f .y/ D 0 2 r.
Let x 2 X and y 2 Y and let k D g; .x/ and l D h .y/. Then k C
D x and
l D y. Then k C i D x C , k C i C l D .x C y/ C , and l < y C < l C 1
so g; .x C y/ D k C i C l D g; .x/ C g; .y/.
Case 1b. x 2 N W g; .x/ C
< x < g; .x/ C
2 q. Let X D x 2 N W
g; .x/ C
< x < g; .x/ C
and let x 2 X be given. Let k D g; .x/ and
let D min; 1 ; k C
x. Let Y D y 2 N W f .y/ 2 .; /, let y 2 Y ,
and let l D h .y/.
Now kCi 1 < kC
C < xC < kC
C D kCi so g; .x/ D kCi 1.
Also l l C < y C < l C C l C 1 so g; .y/ D l. Finally, note
that y < l C k C l C
x so that k C l C i 1 < k C l C
C <
.x C y/ C < k C l C
C D k C l C i and thus g; .x C y/ D k C l C i 1.
Case 1c. x 2 N W g; .x/ C
< x < g; .x/ C
C 2 q. Let X D x 2 N W
g; .x/ C
< x < g; .x/ C
C and let x 2 X be given. Let k D g; .x/ and
let D min; 1 ; x k
. Let Y D y 2 N W f .y/ 2 .; /, let y 2 Y ,
and let l D h .y/.
Now k Ci D k C
C < x C < k C
C C < k Ci C1 so g; .x/ D k Ci .
Exactly as in case 1b, g; .y/ D l. Observe that y > l l x C k C
so that
k C l C i D k C l C
C < .x C y/ C < k C l C i C C < k C l C i C 1
and thus g; .x C y/ D k C l C i .
Case 2.
C 0; 1. Pick 2 .0; / such that min; 1 and .
C
;
C C/\0; 1 D ;. Let X D x 2 N W g; .x/C
2 < x < g; .x/C
C 2
and let Y D y 2 N W f .y/ 2 . 2 ; 2 /. Let k D g; .x/ and let l D h .y/. Then
l < l 2 C < y C < l C 2 C < l C 1 so g; .y/ D l.
Case 2a.
C < 0. Then k 1 < k C
C 2 < x C < k C
C C 2 < k
so g; .x/ D k 1. Also k C l 1 < k C l C
C < .x C y/ C <
k C l C
C C < k C l so g; .x C y/ D k C l 1.
Case 2b. 0 <

C < 1. Then k < kC
C 2 < xC < kC
C C 2 < kC1
so g; .x/ D k. Also k Cl < k Cl C
C < .x Cy/C < k Cl C
C C <
k C l C 1 so g; .x C y/ D k C l.
Case 2c. 1 <
C. Then k C1 < k C
C 2 < x C < k C
C C 2 < k C2
so g; .x/ D k C 1. Also k C l C 1 < k C l C
C < .x C y/ C <
k C l C
C C < k C l C 2 so g; .x C y/ D k C l C 1.
e
Theorem 2.29. Let > 0 and let 0 < < 1. Then g; K.N/ K.N/, and
e
consequently g; c`K.N/ c`K.N/.
Proof. Of course the second conclusion follows from the first by continuity. To see
e
that g; K.N/ K.N/, let p 2 K.N/. Pick a minimal left ideal L of N
such that p 2 L and pick an idempotent r 2 L. Then r 2 Z (since f f is a homo-
morphism). Now K.Z / D Z \ K.N/ by [13, Theorem 1.65]. By Lemma 2.14,
e
g; .r/ D h f e
.r/ is an idempotent in K.Z1= / and so g; .r/ 2 K.N/. Now
p 2 L D L C r so there is some q 2 L such that p D q C r. Thus, by Lemma 2.28,
e e e
g; .p/ D g; .q/ C g; .r/ 2 K.N/.
We see now that spectra preserve C -sets. We saw in Theorem 2.9 that they need
not preserve thick, PS , J , or AP -sets. The reason for the distinction seems to be
that C -sets depend on idempotents, which are in particular members of Z . (All of
the properties shown to be preserved in [4] also depended on points which are members
of Z for each > 0.)
Theorem 2.30. Let A be a C -set in N, let > 0, and let 0 < < 1. Then g; A
is a C -set.
Proof. We need to show that g; A is a member of every idempotent in J.N/, so

let an idempotent p 2 J.N/ be given. By Lemma 2.14 and Corollary 2.26, h1= .p/ e
e
e e
is an idempotent in J.N/ so A 2 h1= .p/. Again by Lemma 2.14, g; h1= .p/ D
f
h

e
h1= .p/ D p and so g; A 2 p.
In [13, Corollary 16.43] it was shown that if 1, 0 < < 1, P is any of the
properties central, central*, IP , or IP , and A N, then g; A has property P
if and only if A has property P . The arguments given there depended on the specific
property. We see now that in fact, all that is required is the preservation of the property
by all spectra.
Theorem 2.31. Let A be a set of subsets of N which is closed under passage to

supersets and assume that whenever > 0, 0 < < 1, and A 2 A, one has that
g; A 2 A. Then whenever 1, 0 < < 1, and A N, one has g; A 2 A if
and only if A 2 A.
Proof. Let 1, 0 < < 1, and A N. The sufficiency holds by assumption.

If D 1, then g; is the identity function so we can assume > 1. Notice that

for all x 2 N, g1=;.1/= g; .x/ D x. (If y D g; .x/, then 1 y x C <
1 1
y C so x < 1 y C 1 1
x C < x C 1.) Consequently, for any A N,

g1=;.1/= g; A A. Thus if g; A 2 A, one has A 2 A.
There are a few basic constructions which have been used to produce examples of
C -sets which are not central. By Theorem 2.31, if A is such an example, > 1, and
0 < < 1, then g; A is another such example.
2.5 Preservation of Ideals

We show in this section that if 1 and 0 < 1, then the ideals K.N/,
e
c`K.N/, and J.N/ are preserved by g; but fail dramatically to be preserved if
> 1.
Lemma 2.32. Let p 2 N, let 0 < < 1, and let 0 < < 1. There exist l 2 Z,
2 12 ; 34 , and A 2 p such that for all x 2 A, g; l C g1=; .x/ D x.
e
Proof. Let
D f1= .p/, and pick by Lemma 2.17 2 12 ; 34 and > 0 such that
for all 2 .0; , x 2 N W g1=; .x/ C
< 1 x < g1=; .x/ C
C 2 p. Let
D min 1 1
2 ; , let A D x 2 N W g1=; .x/ C
< x < g1=; .x/ C
C ,
and let x 2 A.
Now .
C 1 / .
C / D 1 2 1 2 1
2 D 1; and so pick l 2 Z
1
such that
C l <
C . Then
C l <
C 1.
Let k D g1=; .x/. We show that g; .l C k/ D x.
Now k C
< x < k C
C so x
Cl C < .k Cl/C <
x
C C l C , 0 l
C , and l
C C 1 so
x < .l C k/ C < x C 1 as required.
Theorem 2.33. Let I be a subset of N which is a left ideal of .Z; C/ and assume
e
that whenever 0 < and 0 < < 1, one has g; I I . Let 0 < 1
e e
and let 0 < < 1. Then g; I D I . In particular, g; K.N/ D K.N/,
e e e
g; c`K.N/ D c`K.N/, g; J.N/ D J.N/, and g; AP D AP .
Proof. If D 1, then g; is the identity function, and thus we assume that < 1.
e
We need to show that I g; I . Thus, let p 2 I . By Lemma 2.32 we may pick

l 2 Z, 2 12 ; 34 , and A 2 p such that for all x 2 A, g; l C g1=; .x/ D x. By
A A
assumption g1=; .p/ 2 I and since I is a left ideal of Z, l C g1=; .p/ 2 I . Since
e A e
A
g; l g1=; agrees with the identity on A, we have p D g; l C g1=; .p/

e
and so p 2 g; I .
To verify the in particular" conclusions, by Corollary 2.26 and Theorem 2.29, we

only need verify that each of K.N/, c`K.N/, J.N/, and AP are left ideals of
.Z; C/. By [13, Lemma 1.43(c)] each minimal left ideal of N is a left ideal of Z
and so K.N/ is the union of left ideals of Z and is thus a left ideal of Z. Then
by [13, Theorem 2.17], c`K.N/ is a left ideal of Z. To see that J.N/ is a left ideal
of Z, let p 2 J.N/. Then Z C p D c`.Z C p/; and thus it suffices to show that
Z C p J.N/. Then, let m 2 Z and let A 2 m C p. We need to show that A \ N
is a J -set in N, and so let F 2 Pf .NN/ be given. We have that m C A 2 p and
x 2 N W x > m 2 p, and thus pick by Theorem 2.20 some P a 2 .m C A/ \ x 2
N W x > m and H 2 Pf .N/ such that for all k 2 F , a C t2H k.t / 2 .m C A/.
P
Then m C a 2 N, and for all k 2 F , m C a C t2H k.t / 2 A. The proof that AP is
a left ideal of Z is nearly identical. Given p 2 AP , m 2 Z, k 2 N, and A 2 m C p,
.m C A/ \ x 2 N W x > m 2 p so it contains a length k arithmetic progression
a; a C d; : : : ; a C .k 1/d . Then m C a; m C a C d; : : : ; m C a C .k 1/d is a
length k arithmetic progression in A \ N.
We saw in Theorem 2.8 that if 1 and 0 < < 1, then g; preserves thick sets.
Corollary 2.34. Let 0 < 1, let 0 < < 1, and let A N. If A is a PS -

set, then g; A is a PS -set. If A is a J -set, then g; A is a J -set. If A is an
AP -set, then g; A is an AP -set.
Proof. The proofs are essentially the same. We will do the proof for J -sets. Let A
e e
be a J -set. Then J.N/ A so J.N/ D g; J.N/ g; A D g; A.
We saw in Theorem 2.9 that if > 1, then the conclusion of Corollary 2.34 fails
badly. We see now that if > 1, then the conclusion of Theorem 2.33 also fails badly.
Theorem 2.35. Let > 1 and let 0 < < 1. There exists x 2 N such that for
every p 2 Z1= , g; N x C p. Consequently, if I is a left ideal of N, then
e
I n g; N ;.
Proof. Assume first that is rational and pick m; n 2 N such that m > n, D m n,
i iC1
and m and n are relatively prime. Pick i 2 0; 1; : : : ; n 1 such that n < n .
Pick t 2 N such that mt i 1 .mod n/. (See for example [10, Theorem 56].)
Pick x 2 N such that mt D x n i 1. Then m m iC1
n t C < n t C n D x. Also,
m m m i m1
n .t C 1/ C n t C n C n D x C n x C 1. Therefore
m m
./ if l 2 N and n l C x; then n l C x C 1:
Now, let p 2 Z1= D Zn=m D mN, and suppose that g; N 2 x C p. Then
.x C g; N/ \ mN 2 p, and so pick k 2 N such that x C mk 2 g; N and pick
r 2 N such that x C mk D g; .r/. Then x C mk m n r C < x C mk C 1 so

m
x n .r nk/ C < x C 1, contradicting ().
Now assume that is irrational, so that t W t 2 N is dense mod 1. (See for
example [10, Theorem 36].) Pick x; t 2 N such that x . 1/ < t < x .
Then x . 1/ < t C < x and .t C 1/ C > x C 1. Let D minx .t C
/; .t C 1/ C .x C 1/; 12 . We claim that for any integer l,
./ if l C > x , then l C x C 1 C :
To see this, let l 2 Z. If l t , then l C t C x . Thus, assume that

l t C 1. Then l C t C C x C 1 C .
Now let p 2 Z1= and suppose that g; N 2 x C p. Then .x C g; N/ \ y 2
N W f1= .y/ 2 . ; / 2 p, and so pick y 2 N such that f1= .y/ 2 . ; / and
x C y 2 g; N. Pick r 2 N such that x C y D g; .r/. Let k D h1= .y/. Then
k < 1 y < k C , and so k < y < k C . Therefore x C k <
x C y r C < x C y C 1 < x C k C C 1 so x < .r k/ C < x C C 1,
which is a contradiction to ().
To verify the last conclusion, pick x as guaranteed and let I be a left ideal of
N. Then I contains a minimal left ideal hence contains an idempotent p which
e
is necessarily in Z1= . Then x C p 2 I . Now g; N D c`g; N, and so
e
x C p g; N.
Acknowledgments. The first author acknowledges support received from the Na-
tional Science Foundation via grant DMS-0852512. Some of the results in this paper
are from the second authors Ph.D. thesis.
References
[1] T. Bang, On the sequence n; n D 1; 2; : : : ; Math. Scand. 5 (1957), 6976.
[2] V. Bergelson and N. Hindman, Nonmetrizable topological dynamics and Ramsey Theory,
Trans. Amer. Math. Soc. 320 (1990), 293320.
[3] V. Bergelson and N. Hindman, Partition regular structures contained in large sets are
abundant, J. Comb. Theory (Series A), 93 (2001), 1836.
[4] V. Bergelson, N. Hindman, and B. Kra, Iterated spectra of numbers elementary, dynam-
ical and algebraic approaches, Trans. Amer. Math. Soc. 348 (1996), 893912.
[5] V. Bergelson, N. Hindman, and R. McCutcheon, Notions of size and combinatorial prop-
erties of quotient sets in semigroups, Topology Proc. 23 (1998), 2360.
[6] V. Bergelson, N. Hindman, and D. Strauss, Strongly central sets and sets of polynomial
returns mod 1, Proc. Amer. Math. Soc., to appear.
[7] D. De, N. Hindman, and D. Strauss, A new and stronger Central Sets Theorem, Funda-
menta Mathematicae 199 (2008), 155-175.
[8] H. Furstenberg, Recurrence in ergodic theory and combinatorical number theory,

Princeton University Press, Princeton, 1981.
[9] R. Graham, S. Lin, and C. Lin, Spectra of numbers, Math. Mag. 51 (1978), 174176.
[10] G. Hardy and E. Wright, An introduction to the theory of numbers, 5th edn., Oxford
University Press, London, 1979.
[11] N. Hindman, Small sets satisfying the Central Sets Theorem, in: B. Landman, M.
Nathanson, J. Neetril, R. Nowakowski, C. Pomerance, and A. Robertso (eds.), Combina-
torial Number Theory, pp. 5764, DeGruyter, Berlin, 2009. Also published in Integers 9
(Supplement) (2007), #A5. http://www.integers-ejcnt.org/vol9supp.html.
[12] N. Hindman, A. Maleki, and D. Strauss, Central sets and their combinatorial characteri-
zation, J. Comb. Theory (Series A) 74 (1996), 188208.
[13] N. Hindman and D. Strauss, Algebra in the StoneCech compactification: theory and
applications, 2nd edn., de Gruyter, Berlin, 2011.
[14] N. Hindman and D. Strauss, A simple characterization of sets satisfying the Central Sets
Theorem, New York J. Math. 15 (2009), 405413.
[15] N. Hindman and D. Strauss, Cartesian products of sets satisfying the Central Sets Theo-
rem, Topology Proceedings 35 (2010), 203223.
[16] I. Niven, Diophantine approximations, Interscience Publishers, New York, 1963.
[17] H. Shi and H. Yang, Nonmetrizable topological dynamical characterization of central
sets, Fund. Math. 150 (1996), 19.
[18] T. Skolem, On certain distributions of integers in pairs with given differences, Math.
Scand. 5 (1957), 5768.
[19] T. Skolem, ber einige Eigenschaften der Zahlenmengen n C bei irrationalem
mit einleitenden Bemerkungen ber einige kombinatorische probleme, Norske Vid. Selsk.
Forh. 30 (1957), 4249.
Author information
Neil Hindman, Department of Mathematics, Howard University, Washington, DC, USA.
Email: nhindman@aol.com
John H. Johnson, Department of Mathematics, James Madison University, Harrisonburg, Vir-

ginia, USA.
Email: john.j.jr@gmail.com
Chapter 3
On the Differences Between Consecutive Prime

Numbers, I
Daniel A. Goldston and Andrew H. Ledoan
Abstract. We show by an inclusion-exclusion argument that the prime k-tuple conjecture of

Hardy and Littlewood provides an asymptotic formula for the number of consecutive prime
numbers which are a specified distance apart. This refines one aspect of a theorem of Gallagher
that the prime k-tuple conjecture implies that the prime numbers are distributed in a Poisson
distribution around their average spacing.
Keywords. Primes, k-Tuple Conjecture.
Mathematics Subject Classification 2010. Primary: 11N05; Secondary: 11P32, 11N36.
3.1 Introduction and Statement of Results

In 1976, Gallagher [5,6] showed that a uniform version of the prime k-tuple conjecture
of Hardy and Littlewood implies that the prime numbers are distributed in a Poisson
distribution around their average spacing. Specifically, let Pr .h; N / denote the number
of positive integers n less than or equal to N such that the interval .n; n C h contains
exactly r prime numbers. Gallagher then proved that an appropriate form of the prime
k-tuple conjecture implies, for any positive constant and h log N as N ! 1,
that
r
Pr .h; N / e N:
r
In particular, if r D 0, then we obtain by an argument using the prime number
theorem that, as N ! 1,
X N
1 e : (3.1)
log N
pnC1 N
pnC1 pn log n
Here, pn is used to denote the n-th prime number. The purpose of the present chapter
is to obtain a refinement of (3.1) which shows that the Poisson distribution of the
prime numbers in short intervals extends down to the individual differences between
consecutive prime numbers. To obtain this result, we employ a version of the prime
k-tuple conjecture formulated as Conjecture H in Section 3.2, which is equivalent to

the form of the conjecture used by Gallagher.
Theorem 3.1. Assume Conjecture H. Let d be any positive integer, and let p be a
prime number. Let, further,
8 Y p 1

< 2C2 p2
; if d is even;
S.d / D pjd

p>2
:
0; if d is odd;
Y
where
1
C2 D 1 D 0:66016 : : : ;
.p 1/2
p>2
and define X
N.x; d / D 1;
pnC1 x
pnC1 pn Dd
where pn denotes the n-th prime number. Then for any positive constant and d even
with d log x as x ! 1, we have
x
N.x; d / e S.d / : (3.2)
.log x/2
Here we note that S.d / is the singular series in the conjectured asymptotic formula
for the number of prime pairs differing by d . Our theorem shows that for consecutive
prime numbers the Poisson density is superimposed onto this formula for prime pairs.
Our theorem as well as its proof are implicitly contained in the 1999 paper by
Odlyzko, Rubinstein, and Wolf [11] on jumping champions.1 Without claiming any
originality, we think it is worthwhile to explicitly state and prove (3.2). More precise
results when d= log x ! 0 will be addressed in a second paper.
3.2 The HardyLittlewood Prime k-Tuple Conjectures

Let H D h1 ; : : : ; hk be a set of k distinct integers. Let .xI H / denote the number
of positive integers n less than or equal to x for which nCh1 ; : : : ; nChk are simultane-
ously prime numbers. Then the prime k-tuple conjecture of Hardy and Littlewood [8]
is that, for x ! 1,
.xI H / S.H /lik .x/; (3.3)
1 An integer d is called a jumping champion for a given x if d is the most frequently occurring
difference between consecutive prime numbers up to x.

Chapter 3 On the Differences Between Consecutive Prime Numbers, I 39
where
Y k
1 H .p/
S.H / D 1 1 ;
p
p p
H .p/ denotes the number of distinct residue classes modulo p occupied by the ele-
ments of H , and Z x
dt
lik .x/ D k
: (3.4)
2 .log t /
Note in particular that, if H .p/ D p for some prime number p, then S.H / D 0.
However, if H .p/ < p for all prime numbers p, then S.H / 0, in which case the
set H is called admissible. In (3.3), H is assumed to be admissible, since otherwise
.xI H / is equal to 0 or 1.
The prime k-tuple conjecture has been verified only for the prime number theorem.
That is to say, for the case of k D 1. It has been asserted that, in its strongest form,
the conjecture holds true for any fixed integer k with an error term that is Ok .x 1=2C" /
at most and uniformly for H
1; x. (See Montgomery and Soundararajan [9, 10].)
However, we do not need such strong conjectures here. Using

x kx
lik .x/ D CO ; (3.5)
.log x/k .log x/kC1
obtained from integration by parts, we replace lik .x/ by its main term and make the
following conjecture.
Conjecture H. For each fixed integer k 2 and admissible set H , we have

x
.xI H / D S.H / .1 C ok .1//;
.log x/k
uniformly for H
1; h, where h log x as x ! 1 and is a positive constant.
3.3 InclusionExclusion for Consecutive Prime Numbers

The prime k-tuple conjecture for the case when k D 2 provides an asymptotic for-
mula for the number of prime numbers with a given difference d . We need to find
a corresponding formula where we restrict the count to prime numbers that are con-
secutive, and for this one can use the prime k-tuple conjecture with k D 3; 4; : : :
and inclusionexclusion to obtain upper and lower bounds for the number of consecu-
tive prime numbers with difference d . This method has appeared in a series of papers
by Brent [13] and was used by Erdos and Strauss [4] and Odlyzko, Rubinstein, and
Wolf [11] in their study of jumping champions.
We consider a special type of tuple Dk for which
D2 D 0; d
and, for k 3,
Dk D 0; d1 ; : : : ; dk2 ; d :
Here, we require d to be even. We want to count the number of consecutive prime
numbers which do not exceed x and have difference d , namely N.x; d /, and for this
we do inclusionexclusion with
X
2 .x; d / D 1;
px
pp 0 Dd
where p 0 is also a prime number and, for k 3,

X
k .x; d1 ; : : : ; dk2 ; d / D 1:
px
pp 0 Dd
ppj Ddj ; 1j k2
Inserting the expected main term, we obtain

2 .x; d / D S.d /li2 .x/ C R2 .x; d / (3.6)
and, for k 3,
k .x; d1 ; : : : ; dk2 ; d / D S.Dk /lik .x/ C Rk .x; Dk /: (3.7)
We now carry out the inclusion-exclusion. We trivially have
N.x; d / 2 .x; d /:
The consecutive prime numbers that differ by d are those prime numbers p and p 0
satisfying p p 0 D d such that there is no third prime number p 00 with p 0 < p 00 < p.
We can exclude these nonconsecutive prime numbers differing by d by removing
all triples of this form, although this will exclude the same nonconsecutive pair of
prime numbers more than once if there are quadruples of prime numbers such that
p 0 < p 00 < p 000 < p. Hence, writing p p 00 D d 0 , we obtain the lower bound
X
N.x; d / 2 .x; d / 3 .x; d 0 ; d /:
1d 0 <d
We next obtain an upper bound by including the quadruples eliminated in the previous
step and continue in this fashion to get, for R 1,
Q2RC1 .x; d / N.x; d / Q2R .x; d /; (3.8)
where, for N 2,
X
N X
QN .x; d / D 2 .x; d / C .1/k k .x; d1 ; : : : ; dk2 ; d /:
kD3 1d1 <<dk2 <d
We use the convention here that an empty sum has the value zero.
To evaluate QN .x; d /, we require a special type of singular series average consid-

ered by Odlyzko, Rubinstein, and Wolf [11]. Let, for k 3,
X
Ak .d / D S.Dk /: (3.9)
1d1 <:::<dk2 <d
Odlyzko, Rubinstein and Wolf [11] proved that, for k 3,
d k2
Ak .d / D S.d / C Ek .d /; (3.10)
.k 2/
where
!
d k2
Ek .d / D Ok : (3.11)
log log d
(See also Goldston and Ledoan [7].) Thus, on substituting (3.6), (3.7), (3.9), and (3.10),
we find that, for N 2,
X
N
QN .x; d / D S.d /li2 .x/ C .1/k Ak .d /lik .x/ C R2 .x; d /
kD3
X
N X
C .1/k Rk .x; Dk /
kD3 1d1 <<dk2 <d
" 2 #
ZX 1 d k
x N
dt
Z xX
N
dt
D S.d / 2
C .1/k Ek .d /
2 k log t .log t / 2 .log t /k
kD0 kD3
X
N X
C R2 .x; d / C .1/k Rk .x; Dk /;
kD3 1d1 <<dk2 <d
where we used (3.4) in the second line.

We can extract a main term independent of N out of the first term on the far right-
hand side above by using Taylors theorem. With the remainder expressed in La-
granges form, we have that, for M 0 and x > 0,
X
M
1 e
e x D .x/k C .x/M C1 ;
k .M C 1/
kD0
where lies in the open interval joining 0 and x. Hence, we have

Z x "NX 2 #
1 d k dt
2 k log t .log t /2
kD0
Z x Z x !
d dt 1 d N 1 dt
D exp CO
2 log t .log t /2 2 .N 1/ log t .log t /2
Z x N 1 !
d dt 1 3d x
D exp CO p ;
2 log t .log t /2 N N log x .log x/2
by the Stirling formula

p 1
.M 1/ D 2M M 1=2 e M 1 C O :
M
Therefore, we have proved the following lemma.
Lemma 3.2. For N 2, we have

Z x X
N
dt
QN .x; d / D S.d /I.x; d / C .1/k Ek .d / C R2 .x; d /
2 .log t /k
kD3
X
N X
C .1/k Rk .x; Dk /
kD3 1d1 <<dk2 <d
N 1 !
1 3d x
CO p ;
N N log x .log x/2
where Z
x
d dt
I.x; d / D exp :
2 log t .log t /2
3.4 Proof of the Theorem

If we had imposed the additional condition that n C hj is less than or equal to x,
for j 2 1; : : : ; k, in the definition of .xI H / in Section 3.2, we would have that
2 .n; d / D .xI D2 / and, for k 3, k .x; d1 ; : : : ; dk2 ; d / D .xI Dk /. However,
this condition has no effect on Conjecture H, since with H
1; h the condition
removes at most h tuples, which are absorbed into the error term. Thus, assuming that
Conjecture H holds true for k N , we have that R2 .x; d / D o.S.d /x=.log x/2 /,
Rk .x; Dk / D ok .S.Dk /x=.log x/k / and, by (3.11), Ek .d / D ok .d k2 /. Then by
the lemma and since S.d / 1 for d even, we have, for x ! 1 and d log x,

2 x x
QN .x; d / D S.d /I.x; d / C oN e C o S.d /
.log x/2 .log x/2
XN X
x
C ok S.Dk /
.log x/k
kD3 1d1 <<dk2 <d
N 1 !
1 3 x
CO p
N N .log x/2
N !
x 4 x
D S.d /I.x; d / C oN e 2 S.d / CO :
.log x/2 N .log x/2
Finally, on letting N tend to infinity sufficiently slowly, the theorem follows from the
estimate
x
I.x; d / e
.log x/2
and (3.8).
To prove this last estimate, we let dN D d= log x and apply (3.5) to obtain the upper
bound
Z x
N dt dN x x
I.x; d / e d 2
D e C O ;
2 .log t / .log x/2 .log x/3
and the lower bound
Z x
d dt
I.x; d / exp 2
log x log log x x= log x .log t /
h

log log x i h
x i
exp dN 1 C O li2 .x/ li2
log x log x
" !#
N x N
d log log x
e d 1CO :
.log x/2 log x
Hence, we have
" !#
dN x dN log log x
I.x; d / D e 1CO ;
.log x/2 log x
and the required estimate now follows since, if d log x, dN as x ! 1. Hence,

the proof of the theorem is completed.
Acknowledgments. During the preparation of this work, the first author received sup-
port from the National Science Foundation Grant DMS-1104434. The authors would
like to express their sincere gratitude to the referee for his comments on the earlier
version of this paper.
References
[1] R. P. Brent, The distribution of small gaps between successive primes, Math. Comp. 28
(1974), 315324.
[2] R. P. Brent, Irregularities in the distribution of primes and twin primes, in: Collection of
articles dedicated to Derrick Henry Lehmer on the occasion of his seventieth birthday,
Math. Comp. 29 (1975), 4356.
[3] R. P. Brent, Correction to: Irregularities in the distribution of primes and twin primes
(Math. Comp. 29 (1975), 4356), Math. Comput. 30(133) (1976), 198.
[4] P. Erdos and E. G. Straus, Remarks on the differences between consecutive primes, Elem.
Math. 35(5) (1980), 115118.
[5] P. X. Gallagher, On the distribution of primes in short intervals, Mathematika 23 (1976),
49.
[6] P. X. Gallagher, Corrigendum: On the distribution of primes in short intervals (Mathe-
matika 23 (1976), 49), Mathematika 28(1) (1981), 86.
[7] D. A. Goldston and A. H. Ledoan, The jumping champion conjecture, submitted for pub-
lication (2012), available at: http ://arxiv.org/pdf/1102.4879v1.
[8] G. H. Hardy and J. E. Littlewood, Some problems of Partitio numerorum; III: On the
expression of a number as a sum of primes, Acta Math. 44(1) (1922), 170. Reprinted as
pp. 561630 in :Collected Papers of G. H. Hardy, vol. I (including joint papers with J. E.
Littlewood and others; edited by a committee appointed by the London Mathematical
Society), Clarendon Press, Oxford University Press, Oxford, 1966.
[9] H. L. Montgomery and K. Soundararajan, Beyond pair correlation, in: Erdos and his
mathematics, I, pp. 507514, Budapest, 1999; Bolyai Soc. Math. Stud. 11, Jnos Bolyai
Math. Soc., Budapest, 2002.
[10] H. L. Montgomery and K. Soundararajan, Primes in short intervals, Comm. Math. Phys.
252(13) (2004), 589617.
[11] A. Odlyzko, M. Rubinstein, and M. Wolf, Jumping champions, Experiment. Math. 8(2)
(1999), 107118.
Author information
Daniel A. Goldston, Department of Mathematics, San Jos State University, San Jos, Cali-
fornia, USA.
Email: daniel.goldston@sjsu.edu
Andrew H. Ledoan, Department of Mathematics, University of Tennessee at Chattanooga,

Chattanooga, Tennessee, USA.
Email: andrew-ledoan@utc.edu
Chapter 4
On Sets of Integers Which Are Both Sum-Free

and Product-Free
Pr Kurlberg, Jeffrey C. Lagarias, and Carl Pomerance
Abstract. We consider sets of positive integers containing no sum of two elements in the set
and also no product of two elements. We show that the upper density of such a set is strictly
smaller than 12 and that this is best possible. Further, we also find the maximal order for the
density of such sets which are also periodic modulo some positive integer.
Keywords. Sum-Free Set, Product-Free Set, Upper Density.
Mathematics Subject Classification 2010. 11B05, 11B75.
4.1 Introduction
If A, B are sets of integers, we let A C B denote the set of sums a C b with a 2 A,
b 2 B, and we let A B denote the set of products ab with a 2 A, b 2 B. The sum-
product problem in combinatorial number theory is to show that if A is a finite set of
positive integers, then either A C A or A A is a much larger set than A. Specifically,
Erdos and Szemerdi [2] conjecture that if > 0 is arbitrary and A is a set of N
positive integers, then for N sufficiently large depending on the choice of , we have
jA C Aj C jA Aj N 2 :
This conjecture is motivated by the cases when either jA C Aj or jA Aj is un-

usually small. For example, if A D 1; 2; : : : ; N , then A C A is small, namely,
jA C Aj < 2N . However, A A is large since there is some c > 0 such that
jA Aj > N 2 =.log N /c . And if A D 1; 2; 4; : : : ; 2N 1 , then jA Aj < 2N , but
jA C Aj > N 2 =2. The best that we currently know towards this conjecture is that it
holds with exponent 4=3 in the place of 2, a result of Solymosi [8]. (In fact, Solymosi
proves this when A is a set of positive real numbers.)
In this paper we consider a somewhat different question: how dense can A be if
both A C A and A A have no elements in common with A? If A \ .A C A/ D ;, we
say that A is sum-free, and if A \ .A A/ D ;, we say that A is product-free. Before
stating the main results, we give some background on sets that are either sum-free or
product-free.
If a 2 A and A is sum-free, then aCA is disjoint from A, and so we immediately

have that the upper asymptotic density
1
d .A/ WD lim sup jA \ 1; nj
n!1 n
is at most 12 . Density 12 can be achieved by taking A as the set of odd natural numbers.
Similarly, if A is a set of residues modulo n and is sum-free, then D.A/ WD jAj=n is
at most 12 , and this can be achieved when n is even and A consists of the odd residues.
The maximal density for D.A/ for A a sum-free set in Z=nZ was considered in [1].
In particular, the maximum for D.A/ is 13 3n 1
if n is divisible solely by primes that
1 1
are 1 modulo 3, it is 3 C 3p if n is divisible by some prime that is 2 modulo 3 and p
is the least such, and it is 13 otherwise. Consequently, we have D.A/ 25 if A is a
sum-free set in Z=nZ and n is odd. It is worth noting that maximal densities of subsets
of arbitrary finite Abelian groups are determined in [3]. For generalizations to subsets
of finite non-Abelian groups, see [4].
The problem of the maximum density of product-free sets of positive integers, or of
subsets of Z=nZ, only recently received attention. For subsets of the positive integers,
it was shown in [5] that the upper density of a product-free set must be strictly less
than 1; in fact, it cannot exceed 1 2a1 0 , where a0 is the least member of the set. Let
D.n/ denote the maximum value of D.A/ as A runs over product-free sets in Z=nZ.
In [7] it was shown that D.n/ < 12 for the vast majority of integers, namely for every
integer not divisible by the square of a product of six distinct primes. Moreover, the
density of integers which are divisible by the square of a product of six distinct primes
was shown to be smaller than 1:56 108 .
Somewhat surprisingly, D.n/ can in fact be arbitrarily close to 1 (see [5]), and thus
there are integers n and sets of residues modulo n consisting of 99% of all residues,
with the set of pairwise products lying in the remaining 1% of the residues. However, it
is not easy to find a numerical example that beats 50%. In [5], an example of a number
n with about 1:61 108 decimal digits was given with D.n/ > 12 ; it is not known if
there are any substantially smaller examples, say, with fewer than 108 decimal digits.
In [6] the maximal order of D.n/ was essentially found: there are positive constants
c; C such that for all sufficiently large n, we have
c
D.n/ 1 1 2e log 2
;
.log log n/ .log log log n/1=2
and there are infinitely many n with
C
D.n/ 1 1 2e log 2
:
.log log n/ .log log log n/1=2
In this chapter we consider two related questions. First, if A is a set of integers
which is both sum-free and product-free, how large may the upper asymptotic density
Chapter 4 On Sets of Integers Which Are Both Sum-Free and Product-Free 47
2
d .A/ be? Here a sum-free and product-free set with natural density 5 is achievable
by taking
A D n W n 2; 3 .mod 5/:
1
Our first result shows that for a set A to achieve upper asymptotic density close to 2
it must omit all small integers.
Theorem 4.1. Let A be a set of positive integers that is both product-free and sum-
free, and let a0 be the smallest element of A. If d .A/ > 25 ; then necessarily

1 1
d .A/ 1 :
2 2a0
Second, set
D .n/ WD maxD.A/ W A is a sum-free, product-free subset of Z=nZ:
What is the maximal order of D .n/? We prove the following complementary results,
showing that density 12 can be approached, and quantifying the rate of approach.
Theorem 4.2. There is a positive constant such that for all sufficiently large num-
bers n,
1
D .n/ e :
2 .log log n/ 2 2 .log log log n/1=2
1 log
Theorem 4.3. There is a positive constant and infinitely many integers n with
1
D .n/ e :
2 .log log n/1 2 log 2 .log log log n/1=2
Note that D .5/ D 25 and, if 5jn, then D .n/ 25 . A possibly interesting com-
putational problem is to numerically exhibit some n with D .n/ > 25 . Theorem 4.3
assures us that such numbers exist, but the least example might be very large.
One might also ask for the densest possible set A for which A, A C A, and A A
are pairwise disjoint. However, Proposition 4.9 below implies immediately that any
sum-free, product-free set A
Z=nZ with D.A/ > 25 also has A C A and A A
disjoint. Thus, from Theorem 4.3, we may have these three sets pairwise disjoint with
D.A/ arbitrarily close to 12 .
4.2 The Upper Density

Here we prove Theorem 4.1. We begin with some notation which we use in this section.
For a set A of positive integers and a positive real number x, we write A.x/ for A \
1; x. Set
jA.x/j 1
x WD 1 2 ; so that jA.x/j D .1 x /x:
x 2
Note that x 0 for jA.x/j 12 x: If a is an integer, we write a C A for a C A and

we write aA for a A.
Lemma 4.4. Suppose that A is a sum-free set of positive integers and that a1 ; a2 2
A. Then for all x > 0,
1
j.a1 C A.x a1 // \ .a2 C A.x a2 //j .1 3x / .a1 C a2 /:
2
Proof. We have the sets A.x/; a1 C A.x a1 /; a2 C A.x a2 / all lying in 1; x and
the latter two sets are disjoint from the first set (since A is sum-free). Thus,
j.a1 C A.x a1 // \ .a2 C A.x a2 //j
D ja1 C A.x a1 /j C ja2 C A.x a2 /j
j.a1 C A.x a1 // [ .a2 C A.x a2 //j
ja1 C A.x a1 /j C ja2 C A.x a2 /j .x jA.x/j/
.jA.x/j a1 / C .jA.x/j a2 / C .jA.x/j x/
D 3jA.x/j x .a1 C a2 /:
But 3jA.x/j x D 12 .1 3x /x, so this completes the proof.
For a set A of positive integers, define the difference set
A WD a1 a2 W a1 ; a2 2 A:
Further, for an integer g, let
Ag WD A \ .g C A/ D a 2 A W a C g 2 A:
Corollary 4.5. If A is a sum-free set of positive integers and g 2 A then, for any
x > 0,
1
jAg .x/j .1 3x / x C O.1/;
2
in which the implied constant depends on both g and A.
Proof. Suppose that g 2 A, so that there exist a1 ; a2 2 A such that a1 a2 D g. If

a 2 A.x a1 / and a C a1 2 a2 C A.x a2 /, then a C g D a C a1 a2 2 A, so that
a 2 Ag . That is, Ag .x a1 / contains a1 C .a1 C A.x a1 // \ .a2 C A.x a2 //.
Thus, by Lemma 4.4,
1
jAg .x a1 /j j.a1 C A.x a1 // \ .a2 C A.x a2 //j .1 3x / .a1 C a2 /;
2
from which the corollary follows.
Proposition 4.6. If A is a sum-free set of positive integers with upper density greater
than 25 , then A is the set of all even integers and A consists solely of odd numbers.
Proof. We first show that A is a subgroup of Z. Since A is closed under multipli-

cation by 1, it suffices to show that if g1 ; g2 2 A, then g1 Cg2 2 A. If g1 CAg1
contains a member a of Ag2 , then a g1 2 A and a Cg2 2 A, so that g1 Cg2 2 A.
Note that g1 CAg1 and Ag2 are both subsets of A. Now, by Corollary 4.5, if g1 CAg1
and Ag2 were disjoint, we would have for each positive real number x,
1
.1 3x /x C O.1/ jA.x/j D .1 x /x;
2
so that x 15 CO. x1 /. Hence lim inf x 15 , contradicting the assumption that A has
upper density greater than 25 . Thus, g1 C Ag1 and Ag2 are not disjoint, which as we
have seen, implies that g1 C g2 2 A. Thus, A is a subgroup of Z, say A D gZ
for some positive integer g.
Since each a1 a2 0 .mod g/ for all a1 ; a2 2 A, all members of A are in one
residue class modulo g. Since A has upper density greater than 25 , it follows that g D 1
or 2. But A must be disjoint from A. Indeed, if a1 a2 D a3 with a1 ; a2 ; a3 2 A,
then a1 D a2 C a3 , violating the condition that A is sum-free. Thus, if g D 1, A D ;,
and if g D 2, A consists solely of odd numbers. The first case violates our hypothesis,
so our proof is complete.
Remark 4.7. Proposition 4.6 is best possible, as can be seen by taking A as the set of
positive integers that are either 2 or 3 modulo 5.
We now prove the following result, which immediately implies Theorem 4.1.
Proposition 4.8. Suppose that A is a sum-free set of positive integers with least mem-
ber a0 . Suppose in addition that a0 A is disjoint from A. Then the upper density of A
is at most max 25 ; 12 .1 2a1 0 /.
Proof. If the upper density of A is at most 25 , the result holds trivially, so we may
assume the upper density exceeds 25 . It follows from Proposition 4.6 that A consists
solely of odd numbers. Thus, for any real number x a0 , both a0 A.x=a0 / and A.x/
consist solely of odd numbers, they are disjoint, and they lie in 1; x. Thus,

x
x 1
jA.x/j C A D jA.x/j C a A x C O.1/:
a0 a0 2
0
Further A.x/ n A.x=a0 / is contained within the odd numbers in .x=a0 ; x, so that

x 1 x
jA.x/j A x C O.1/:
a0 2 a0
Adding these two inequalities and dividing by 2 gives that jA.x/j . 12 1

4a0 /x C
1 1
O.1/, so that A has upper density at most 2 4a0 , giving the result.
4.3 An Upper Bound for the Density in Z=nZ

In this section we prove Theorem 4.2. We begin by noting the following simple con-
sequence of Proposition 4.6.
Proposition 4.9. Suppose that n is a positive integer and A

Z=nZ is sum-free. If
D.A/ > 25 , then n is even and A is a subset of the odd residues classes in Z=nZ.
Proof. Replace A with A, N the set of positive numbers in the residue classes in A. Then
N
A has density D.A/ and is sum-free. It follows from Proposition 4.6 that all members
of AN are odd. If n were odd, then AN would contain both odd and even members, so we
must have n even and A a subset of the odd residue classes in Z=nZ. This completes
the proof.
We are now ready to prove Theorem 4.2. For those n with D .n/ 25 , the result
holds for any number , so assume that D .n/ > 25 . Let A
Z=nZ be a product-
free, sum-free set with D.A/ D D .n/. By Proposition 4.9 we have that n is even
and that A is a subset of the odd residues modulo n. Suppose that k is an integer with
n 2k < 2n. Let N D 22k n and let B be the set of positive integers of the form
2j b where j k and b N=2j D 22kj n, such there is some a 2 A with b a
.mod n/. Then the members of B are in 1; N and
X
k
2kj k kC1 1 2kC1
jBj D 2 jAj D 2 .2 1/jAj > 1 2 jAj: (4.1)
n
j D0
We note that B is product-free as a set of residues modulo N . Indeed, suppose

2ji bi 2 B, for i D 1; 2; 3 and
2j1 b1 2j2 b2 2j3 b3 .mod N /:
Let ai 2 A be such that bi ai .mod n/ for i D 1; 2; 3. We have that a1 ; a2 ; a3

are odd, and since n is even, this implies that b1 ; b2 ; b3 are odd. Using j1 C j2 2k,
j3 k, and 22k jN , we have j1 C j2 D j3 . Hence a1 a2 a3 .mod n/, a violation of
the assumption that A is product-free modulo n. We conclude that B is product-free
modulo N .
It now follows from Theorem 1.1 in [6] that for n sufficiently large,

c
jBj N 1 e :
.log log N /1 2 log 2 .log log log N /1=2
Further, since N is of order of magnitude n3 , we have that log log N D log log n C
O.1/, and so for any fixed choice of c0 < c we have for n sufficiently large that

c0
jBj N 1 e :
.log log n/1 2 log 2 .log log log n/1=2
Thus, from our lower bound for jBj in (4.1) we have

N 1 1 c0
jAj < 2kC1 1 1 e :
2 n .log log n/1 2 log 2 .log log log n/1=2
Since N=22kC1 D n=2, it follows that for any fixed c1 < c0 and n sufficiently large,
we have
n c1
jAj < 1 e :
2 .log log n/1 2 log 2 .log log log n/1=2
We thus may choose as any number smaller than c=2. This concludes the proof of
Theorem 4.2.
4.4 Examples With Large Density

In this section we prove Theorem 4.3. We follow the argument in [5] with a supplemen-
tary estimate from [6]. Let x be a large number, let `x be the least common multiple
of the integers in 1; x, and let nx D `2x . Then nx D e.2Co.1//x as x ! 1 so that
log log nx D log x C O.1/. For a positive integer m, let .m/ denote the number of
prime factors of m counted with multiplicity. Let k D k.x/ D b 4e log log nx c, let
Dx0 D d j`x W d odd; k < .d / < 2k ;
and let A be the set of residues a modulo nx with gcd.a; nx / 2 Dx0 . Then A is product-
free (cf. Lemma 2.3 in [5]), and since nx is even and every residue in A is odd, we have
that A is sum-free as well. We shall now establish a sufficiently large lower bound on
D.A/ to show that D .nx / satisfies the inequality in the theorem with n D nx .
For d 2 Dx0 , the number of a .mod nx / with gcd.a; nx / D d is '.nx /=d , so that
!
'.nx / X 1 '.nx / X 1 X 1
D.A/ D D : (4.2)
nx 0
d nx d d
d 2Dx d j`x d j`x
d odd d odd
d 62Dx0
We have
X 1 Y p Y 1
Y p

1 .x/
D 1 aC1 1 ;
d p1 p p1 x
d j`x 2<px 2<px 2<px
d odd p a k`x
Q
and, since '.nx /=nx D 21 pjnx ; p>2 .1 1=p/, we find that

'.nx / X 1 1 1 .x/ 1 .x/
1 : (4.3)
nx d 2 x 2 x
d j`x
d odd
We now use Equation (6.2) in [6], which is the assertion that

X 1
e
.log x/ 2 log 2
:
d .log log x/1=2
P .d /x
.d /62.k;2k/
Here P .d / denotes the largest prime factor of d . Since this sum includes every odd
integer d j`x with d 62 Dx0 , we have
'.nx / X 1
e
'.nx / .log x/ 2 log 2 1
1=2
e ;
nx
d j`
d nx .log log x/ .log x/ 2 2 .log log x/1=2
1 log
x
d odd
d 62Dx0
Q
where we use Mertens theorem in the form '.nx /=nx D px .11=p/ 1= log x
for the last step. Putting this estimate and Inequality (4.3) into Equation (4.2), we get
1 .x/ c0
D.A/ e
2 x .log x/1 2 log 2 .log log x/1=2
for some positive constant c 0 . Using .x/=x 1= log x and log x D log log nx C
O.1/, we have
1
D.A/ e
2 .log log nx /1 2 log 2 .log log log nx /1=2
for any fixed constant > c 0 and x sufficiently large. Thus, D .nx / satisfies the
condition of Theorem 4.3 for x sufficiently large, completing the proof.
Acknowledgments. We thank Albert Bush, Chris Pryby, and Joseph Vandehey for
raising the question of sets which are both sum-free and product-free. We also thank
Imre Ruzsa for several nice suggestions that simplified our arguments. P. K. was sup-
ported in part by grants from the Gran Gustafsson Foundation, and the Swedish Re-
search Council. J. C. L. was supported in part by NSF grant DMS-1101373. C. P. was
supported in part by NSF grant DMS-1001180.
References
[1] P. H. Diananda and H. P. Yap, Maximal sum-free sets of elements of finite groups, Proc.
Japan Acad. 45(1) (1969), 15.
[2] P. Erdos and E. Szemerdi, On sums and products of integers: To the memory of Paul
Turn, in: P. Erdos (ed.), Studies in pure mathematics, pp. 213218, Birkhuser, Basel,
1983.
[3] B. Green and I. Z. Ruzsa, Sum-free sets in abelian groups, Israel J. Math. 147 (2005),
157188.
[4] K. S. Kedlaya, Product-free subsets of groups, then and now, in: R. Y. Chow and D. C.
Isaksen (eds.), Communicating Mathematics, pp. 179187, Contemp. Math. 479, Amer.
Math. Soc., Providence, RI, 2009.
[5] P. Kurlberg, J. C. Lagarias, and C. Pomerance, Product-free sets with high density, Acta
Arith. 155 (2012), 163173.
[6] P. Kurlberg, J. C. Lagarias, and C. Pomerance, The maximal density of product-free
sets in Z=nZ, IMRN 2013 (2013), 827845. First published online February 14, 2012,
doi:10.1093/imrn/ms014.
[7] C. Pomerance and A. Schinzel, Multiplicative properties of sets of residues, Moscow J.
Combinatorics and Number Theory 1 (2011), 5266.
[8] J. Solymosi, Bounding multiplicative energy by the sumset, Adv. Math. 222 (2009), 402
408.
Author information
Pr Kurlberg, Department of Mathematics, KTH Royal Institute of Technology, Stockholm,
Sweden.
Email: kurlberg@math.kth.se
Jeffrey C. Lagarias, Department of Mathematics, University of Michigan, Ann Arbor, Michi-

gan, USA.
Email: lagarias@umich.edu
Carl Pomerance, Mathematics Department, Dartmouth College, Hanover, New Hampshire,

USA.
Email: carl.pomerance@dartmouth.edu
Chapter 5
Four Perspectives on Secondary Terms

in the DavenportHeilbronn Theorems
Frank Thorne
Abstract. This paper is an expanded version of the authors lecture at the Integers Conference
2011. We discuss the secondary terms in the DavenportHeilbronn theorems on cubic fields
and 3-torsion in class groups of quadratic fields. Such secondary terms had been conjectured by
DatskovskyWright and Roberts, and proofs of these or closely related secondary terms were
obtained independently by Bhargava, Shankar, and Tsimerman, Hough, Zhao, and Taniguchi
and the author.
In this chapter we discuss the history of the problem and highlight the diverse methods used
in several papers to address it.
Keywords. Quadratic Field, Cubic Field, Class Group.
Mathematics Subject Classification 2010. 11R16, 11R29.
5.1 Introduction
This chapter concerns the following two theorems:
Theorem 5.1. Let N3 .X / count the number of cubic fields K with 0 < Disc.K/ <
X . We have
1 4.1=3/
N3 .X / D C X C K X 5=6 C o.X 5=6 /; (5.1)
12.3/ 5.2=3/3 .5=3/
p
where C D 3, C C D 1, K D 3, and K C D 1.
Theorem 5.2. For any quadratic field with discriminant

p D, let Cl3 .D/ denote the
3-torsion subgroup of the ideal class group Cl.Q. D//; we have
X
#Cl3 .D/ D
0<D<X
!
3 C C 8.1=3/ Y p 1=3 C 1
X C K 1 X 5=6 C o.X 5=6 /; (5.2)
2 5.2=3/3 p p.p C 1/
56 Frank Thorne
and where the sum ranges over fundamental discriminants D, the product is over all
primes, and the constants are as before.
The main terms are due to Davenport and Heilbronn [15], and the secondary term
in Theorem 5.1 was conjectured by Roberts [32] and implicitly by Datskovsky and
Wright [14].
The secondary terms were expected to be difficult to prove. However, progress has
recently been made in four independent works, using a variety of methods. Both results
above have been proved by Bhargava, Shankar, and Tsimerman [7], using the geom-
etry of numbers, and by Taniguchi and the present author [37], using Shintani zeta
functions. Hough [24] has proved a variation of Theorem 5.2 by studying the distribu-
tion of Heegner points in the complex upper half-plane, and Zhao [45] has obtained a
variation of Theorem 5.1 for function fields, using algebraic geometry.1
The authors lecture at the 2011 Integers Conference explained (briefly) how each
of these four approaches sheds light on these secondary terms, and this chapter is an
expanded version of our lecture. We begin with some background on counting fields
and on counting cubic fields in particular. The reader might skip to Section 5.4 for our
discussion of the secondary terms.
5.2 Counting Fields in General

How does one count number fields? This question has been addressed in many ex-
cellent expository accounts, such as that by Cohen, Diaz y Diaz, and Olivier [10], or
Chapter 6 of Bhargavas ICM Proceedings article [5]. Therefore, our overview will be
brief. To a large extent we concentrate on the trivial problem of counting quadratic
fields, as some observations on this problem anticipate the methods developed to count
cubic fields with good error terms.
Two early important theorems in the subject were proved by Minkowski and Her-
mite:
Theorem 5.3 (Hermite). There are only finitely many number fields of bounded dis-
criminant.
Theorem 5.4 (Minkowski). The discriminant of a number field K of degree n satisfies

n 2 n
n
jDisc.K/j : (5.3)
n 4
If we write n D r C 2s, where r is the number of real embeddings of K, and s is
the number of pairs of complex embeddings, we may replace .=4/n with .=4/2s .
1 The
error terms in [24] and [45] are larger than the secondary terms at present, but both of these
approaches naturally explain the secondary term, and both authors are currently working on refining
their methods.
Chapter 5 Secondary Terms in the Davenport-Heilbronn Theorems 57
p n
Moreover, by Stirlings formula n 2 n ne and therefore

n n
jDisc.K/j e 2 C o.1/ D 5:803 C o.1/ : (5.4)
4
2
The constant e 4 can be improved; see Odlyzko [29]. Conversely, this bound is sharp
apart from the constant; Golod and Shafarevich [21] proved the existence of an infinite
class field tower of fields K for which jDisc.K/j1=n is constant.2
To prove these theorems3 , use the r embeddings K ,! R and the s pairs of em-
beddings K ,! C to embed the ring of integers of K as a lattice in n-dimensional
space. As proved by Minkowski, any convex body centered around the origin whose
volume is greater than 2rCs jDisc.K/j1=2 must contain a lattice point other than zero.
However, this point must have norm at least 1, and geometric considerations allow
one to conclude Minkowskis lower bound.
Hermites theorem can be proved in a similar way. We again use Minkowskis con-
vex body theorem to find an element 2 OK , the sizes of whose complex embeddings
Qwhich is guaranteed to generate K over Q. The minimal polynomial
are all small, and
of is equal to .X .//, where ranges over all real and complex embeddings
of K, and the bounds obtained from Minkowskis theorem yield bounds on the coeffi-
cients of the minimal polynomial of . For a fixed discriminant bound, there are only
finitely many such polynomials and thus only finitely many possibilities for K.
Remark 5.5. Although this proof of Hermites theorem relies on Minkowskis work,
Hermites paper [23] preceded Minkowskis by over thirty years. Hermite claimed
in [23] that there are finitely many fields of bounded discriminant with fixed degree;
the stronger claim of Theorem 5.3 seems to depend also on Minkowskis work.
These proofs yield results about fields of each fixed degree, and so far this has been
a natural way to count them. In what follows, let Nn .X / be the number of fields K of
degree n with jDisc.K/j < X . (Results are also known when the sign of the discrim-
inant, or more generally the number of real embeddings, is specified.)
We have the following results (excluding n D 3):
n = 1. There is only Q.
n = 2. We have the equality of Dirichlet series
X
.s/
jDisc.K/js D 1 C 1 2s C 2 4s ; (5.5)
.2s/
KWQ
D2
and it follows that
6 p
N2 .X / D 2
X C O. X/: (5.6)

p
1 ; 46/, which he proved to have an infinite
2 Martinet [27] gave the example of F D Q.11 C 11
2-class field tower, for which jDisc.K/j1=n D 92:2 : : : for each field K. We refer to Lemmermeyer [25]
for a discussion of related issues, along with a very thorough bibliography.
3 See [28, Chap. III.2] for complete proofs.
58 Frank Thorne
The typical proof of (5.6) estimates, for each squarefree odd q, the number of inte-
gers n divisible by q 2 satisfying the 2-adic conditions implied by Equation (5.6), and
then uses inclusionexclusion to sieve for squarefreeness. The details are commonly
left as an exercise for beginning analytic number theory students. However, it was
not until the recent work of Belabas, Bhargava, and Pomerance [2] that the analogous
method was developed to yield power-saving error terms for N3 .X /.
How does one estimate the number of integers n with 0 < n < X and q 2 jn?
(For simplicity of explanation, we ignore the 2-adic conditions.) This is a lattice-point
counting problem in a one-dimensional vector space, so the answer is q12 times the
length of the interval .0; X / plus an error of O.1/. This is trivial, but its generaliza-
tion to the cubic case, where one counts GL2 .Z/-orbits on a four-dimensional lattice,
satisfying local conditions .mod q 2 / and a global condition 0 < Disc.f / < X , is
not. (The higher dimensional analogues are even less so!)
These integers may also be counted using analytic number theory. By Perrons for-
mula, we have Z 2Ci1
X ds
1D q 2s .s/X s ; (5.7)
2i1 s
0<n<X
q 2 jn
and these sums may be estimated by shifting the contour and using the functional
equation of the zeta function. This functional equation relies on Poisson summation,
and so again relies on the questions interpretation as a lattice point counting problem.
Even the most diehard fan of contour integration would likely prefer the trivial
proof, but this analytic method also generalizes usefully to the cubic case, where the
Riemann zeta function is replaced with a Shintani zeta function.
These two counting techniques, in combination with the inclusionexclusion sieve,
are the starting points of the geometric and analytic proofs of Theorems 5.1 and 5.2,
respectively!
We note one additional point related to Equation (5.5). Wright [42] observed that
this Dirichlet series has the beautiful representation
X Y 1 X
jDisc.K/js D jDisc.Kv /jps ; (5.8)
p
2
KWQ
D2 Kv WQp
2
and proved this in a much more general framework. (He obtained similar formulas for
degree n cyclic extensions of any number field.) His results follow from considering
a nontrivial twist of the adelic zeta function of Tates thesis [39]. This twist makes
the affine line into a prehomogeneous vector space, with the action of GL.1/ given
by .t /x D t n x.
Essentially, a vector space is prehomogeneous if it has an action of an algebraic
group G, which is transitive over C apart from the vanishing locus of an irreducible
polynomial. This prehomogeneous property is essential both in Bhargavas work
and in the zeta-function approach pioneered by Sato and Shintani [34]; cubic, quartic,
and quintic fields are parameterized by lattice points4 up to the action of G.Z/, and
these may be counted geometrically or analytically.
Wrights work on (5.8) and its generalizations mirrors his work with Datskovsky
[13,14,41] on the Shintani zeta function associated to cubic fields, and this latter work
is at the heart of the analytic approach to counting cubic fields. In follow-up work,
Wright and Yukie [43] proposed a program to enumerate quartic and quintic fields by
studying the zeta functions associated to appropriate prehomogeneous vector spaces.
So far their program has not succeeded. However:
n = 4, 5. Bhargava [3, 6] proved the asymptotic formulas
5 Y
N4 .X; S4 / 1 C p 2 p 3 p 4 X; (5.9)
24 p
13 Y
N5 .X / 1 C p 2 p 4 p 5 X: (5.10)
120 p
The count N4 .X; S4 / includes only quartic fields with Galois group S4 , and (5.9) im-
plies an asymptotic for N4 .X / in combination with work of Baily [1], Cohen, Diaz y
Diaz, and Olivier [9], and Wong [40]. Bhargavas geometric approach may be consid-
ered an extension of the methods described in Sections 5.3 and 5.6.
n > 5. For n > 5, Bhargava conjectured [4] that Nn .X; Sn / Cn X . The explicit
constants Cn have natural interpretations as Euler products; each Euler factor repre-
sents the probability that a field K of degree n has a certain localization. Ellenberg
and Venkatesh
p
[19] have proved the best upper bound to date, namely Nn .X /
X exp.C log n/ .
It is unclear whether prehomogeneous vector spaces may be used to obtain formu-
las for Nn .X / for n > 5. By a classification theorem of Sato and Kimura [33], it is
known that there are no reduced, irreducible, and reductive prehomogeneous vector
spaces parameterizing rings of rank > 5; it remains to be seen whether there are any
prehomogeneous vector spaces not satisfying these assumptions which parameterize
such rings.
5.2.1 Counting Torsion Elements in Class Groups

The problem of counting 3-torsion elements in class groups of quadratic fields is re-
lated to that of counting cubic fields. If D 1 is a fundamental discriminant,
p then class
field theory provides a bijection betweenp subgroups of Cl.Q. D// of index 3, and
unramified cyclic cubic extensions of Q. D/. Such extensions, in turn, correspond
to (non-Galois) cubic extensions of Q whose discriminant is D. Therefore, counting
3-torsion elements in quadratic fields is equivalent to counting cubic fields whose dis-
criminant is fundamental, which are those not totally ramified at any prime.
4 Not all of the G.Z/-orbits correspond to fields, as we will see in the cubic case.
60 Frank Thorne
Therefore, this problem may be treated as a variation of the problem of counting

cubic fields. This is the approach of Sections 5.3, 5.5, and 5.6; however, there are
other approaches as well, such as that of Section 5.7. A general heuristic for statistics
of torsion elements in class groups was proposed by Cohen and Lenstra [11]; their
heuristic has inspired a great deal of interesting follow-up work, but for the most part
the CohenLenstra heuristics remain wide open.
5.3 DavenportHeilbronn, DeloneFaddeev, and the Main

Terms
In this section we discuss the proofs of the main terms in Theorems 5.1 and 5.2. To
summarize, the idea is as follows. We count cubic fields by counting their maximal or-
ders, and so we count all cubic orders and then sieve for maximality. We count cubic
orders by counting cubic rings and then subtracting the contribution of the reducible
rings, and these correspond to binary cubic forms. (The space of binary cubic forms
is prehomogeneous in the sense described previously.) Thus at last we have a geomet-
ric lattice-point counting problem, which can be attacked using brute force. This
interpretation in terms of lattice points also opens the door to the use of zeta functions.
We recommend the paper of Bhargava, Shankar, and Tsimerman [7] for the simplest
self-contained account of the DavenportHeilbronn theorems, with complete proofs.
We begin by defining cubic rings and cubic forms. A cubic ring is a commutative
ring which is free of rank 3 as a Z-module. The discriminant of a cubic ring is defined
to be the determinant of the trace form hx; yi D Tr.xy/, and the discriminant of the
maximal order of a cubic field is equal to the discriminant of the field.
The lattice of integral binary cubic forms is defined by

VZ WD au3 C bu2 v C cuv 2 C dv 3 W a; b; c; d 2 Z ; (5.11)
and the discriminant of such a form is given by the usual formula
Disc.f / D b 2 c 2 4ac 3 4b 3 d 27a2 d 2 C 18abcd: (5.12)
There is a natural action of GL2 .Z/ (and also of SL2 .Z/) on VZ , given by
1
. f /.u; v/ D f ..u; v/ /: (5.13)
det
We call a cubic form f irreducible if f .u; v/ is irreducible as a polynomial over Q,

and nondegenerate if Disc.f / 0.
Cubic rings are related to cubic forms by the following correspondence of Delone
and Faddeev, as further extended by Gan, Gross, and Savin [20]:
Theorem 5.6 ([16, 20]). There is a natural, discriminant-preserving bijection be-

tween the set of GL2 .Z/-equivalence classes of integral binary cubic forms and the
set of isomorphism classes of cubic rings. Furthermore, under this correspondence,
irreducible cubic forms correspond to orders in cubic fields.
Finally, if x 2 VZ is a cubic form corresponding to a cubic ring R, we have
StabGL2 .Z/ .x/ ' Aut.R/.
It is therefore necessary to exclude reducible and nonmaximal rings. The nonmaxi-

mality condition is the more difficult of the two to handle, and for this Davenport and
Heilbronn established the following criterion:
Proposition 5.7 ([7, 15]). Under the DeloneFaddeev correspondence, a cubic ring
R is maximal if any only if its corresponding cubic form f belongs to the set Up
VZ
for all p, defined by the following equivalent conditions:
the ring R is not contained in any other cubic ring with index divisible by p;
the cubic form f is not a multiple of p, and there is no GL2 .Z/-transformation of
f .u; v/ D au3 C bu2 v C cuv 2 C dv 3 such that a is a multiple of p 2 and b is a
multiple of p.
In particular, the condition Up only depends on the coefficients of f modulo p 2 .

Therefore one can count cubic fields as follows. One obtains an asymptotic formula
for the number of cubic rings of bounded discriminant by counting lattice points in fun-
damental domains for the action of GL2 .Z/, bounded by the constraint jDisc.x/j < X .
The fundamental domains may be chosen so that almost all reducible rings correspond
to forms with a D 0, and so these may be excluded from the count.
One then multiplies this asymptotic by the product of all the local densities of the
sets Up . This gives a heuristic argument for the main term in Equation (5.1), and one
incorporates a sieve to obtain a proof.
The main term of Theorem 5.2 may be provedp similarly. By class field theory, there
is a bijection between subgroups of Cl3 .Q. D// of index 3 and cubic fields of dis-
criminant D, which are precisely those not totally ramified at any prime. This ramifi-
cation condition may also be detected by reducing cubic forms modulo p 2 , and thus
may be incorporated into Proposition 5.7. The remainder of the proof is the same.
5.3.1 The Work of Belabas, Bhargava, and Pomerance

In [2], Belabas, Bhargava, and Pomerance (BBP) introduced improvements to Daven-
port and Heilbronns method and obtained an error term of O.X 7=8C / in Equations
(5.1) and (5.2). They began by observing that
X
N3 .X / D
.q/N .q; X /; (5.14)
q1
62 Frank Thorne
where N .q; X / counts the number of cubic orders of discriminant 0 < D < X
which are nonmaximal at every prime dividing q. This simple observation has proved
to be quite useful! For large q, BBP prove that N .q; X / X 3!.q/ =q 2 using rea-
sonably elementary methods. Therefore, one has
X
N3 .X / D
.q/N .q; X / C O.X=Q1 /: (5.15)
qQ
For small q, BBP estimate N .q; X / with explicit error terms using geometric
methods. These error terms are good enough to allow them to take the sum in Equation
(5.14) up to Q D .X log X /1=8 , which yields a final error term of O.X 7=8C /.
In addition, their methods extend to counting S4 -quartic fields, where they obtain
a main term of C4 X with error X 23=24C .
5.4 The Four Approaches

This, then, brings us to Theorems 5.1 and 5.2. Theorem 5.1 was conjectured in print by
Roberts [32], and implicitly in the foundational work of Datskovsky and Wright [13,
14,41]. Roberts gave substantial and convincing numerical evidence for his conjecture,
as well as the heuristic argument described in Section 5.5.
The four approaches discussed in this paper are the following:
Shintani Zeta Functions (Taniguchi-T. [37, 38]). Taniguchi and the author further
developed the theory of Shintani zeta functions and proved Theorems 5.1 and 5.2
with error terms of O.X 7=9C / and O.X 18=23C /, respectively. These proofs closely
follow the heuristic arguments of Roberts and DatskovskyWright.
The Shintani zeta-function approach is quite flexible, and [37] contains several gen-
eralizations of the main theorems. Perhaps the most interesting of these is a general-
ization to counting cubic field discriminants (or 3-torsion elements of class groups)
in arithmetic progressions, where a certain nonequidistribution phenomenon arises.
For example, when counting cubic field discriminants a .mod 7/, the secondary
term in Theorem 5.1 is different for every residue class a .mod 7/.
A Refined Geometric Approach (Bhargava, Shankar, and Tsimerman [7]). Bhar-
gava, Shankar, and Tsimerman gave the first proof of Theorem 5.1, with an error
term of O.X 13=16C /. Their proof follows Davenport and Heilbronns original work,
counting lattice points via geometric arguments, with substantial simplifications and
improvements.
Their proof of the secondary term is based on a slicing argument which we de-
scribe briefly in Section 5.6. Their treatment of the error terms introduces a certain cor-
respondence for nonmaximal cubic rings; this correspondence has also proved useful
in the context of Shintani zeta functions, and a combined approach ([8], in progress)
has yielded an error term of O.X 2=3C /.
Equidistribution of Heegner Points (Hough [24]). Hough obtained a statement close-

ly related to Theorem 5.2.
pHis scope is limited to counting 3-torsion elements in imagi-
nary quadratic fields Q. D/ with D 2 .mod 4/; in this context, he proves a version
of Theorem 5.2, but with an error term larger than X 5=6 . In follow-up work (in prepa-
ration), he proves the secondary term for a smoothed variation of this sum.
Hough derives this result as a consequence of an equidistribution theorem for the
Heegner points associated to the 3-part of class groups of imaginary quadratic fields;
his proof of the DavenportHeilbronn theorem appears as a bonus, and without ref-
erence to binary cubic forms. His techniques also apply to the k-parts of these class
groups, for odd k 5, where he obtains secondary terms (of order X 1=2C1=k ). His
error terms are currently larger than both the secondary and the main terms for k 5,
but this work sheds light on a notoriously difficult problem.
Trigonal Curves and the Maroni Invariant (Zhao [45]). Zhao enumerates cubic ex-
tensions of the rational function field Fq .t / using algebraic geometry. Such extensions
are in bijection with isomorphism classes of smooth trigonal curves, that is, smooth
3-fold covers of P 1 , and these may be counted by embedding them in certain sur-
faces Fk . He obtains a secondary term as a consequence of a bound for the integer
k, called the Maroni invariant; for now his error terms are larger than this secondary
term.
In all of these approaches, the secondary term is easier to see than to prove. In
particular, each of these approaches yield the secondary term in a natural way, but it
is not a priori evident that the error terms can be made smaller than X 5=6 !
In what follows we will describe, insofar as we can in a couple of pages, why each
of these approaches naturally yields a secondary term. For the Shintani zeta-function
approach, we will say only a little about the error term, and for the other approaches
we will say nothing. Indeed, we will use the notation O. / whenever the details are
best left to the respective papers.
5.5 The Shintani Zeta-Function Approach

Taniguchi and the author [37] proved Theorems 5.1 and 5.2 using the analytic theory
of Shintani zeta functions. The Shintani zeta functions associated to the space of binary
cubic forms are defined5 by the Dirichlet series
X 1
.s/ WD jDisc.x/js

jStab GL 2 .Z/ .x/j
x2GL2 .Z/nVZ
X 1
D jDisc.R/js : (5.16)
jAut.R/j
Disc.R/>0
5 The Shintani zeta functions are commonly defined in terms of SL .Z/ rather than GL .Z/, multi-
2 2
plying (5.16) by a factor of 2.
64 Frank Thorne
In the former sum, VZ is the lattice defined in Equation (5.11), the sum is over elements
of positive or negative discriminant respectively, and Stab.x/ is the stabilizer of x in
GL2 .Z/. The latter sum is over isomorphism classes of cubic rings. The equality of
these two series follows from the DeloneFaddeev correspondence (Theorem 5.6).
This is an example of a zeta function associated to a prehomogeneous vector space.
The definition of prehomogeneous concerns the GL2 .C/ action on VC , but here the
important (related) fact is that GL2 .R/ acts transitively on the positive- and negative-
discriminant loci VR . Sato and Shintani [34] developed a general theory of zeta func-
tions associated to prehomogeneous vector spaces, and for the space of binary cubic
forms Shintani proved [35] that these zeta functions enjoy analytic continuation and an
explicit functional equation. These zeta functions have poles at s D 1, as is common
for zeta functions, and at s D 5=6, which is much more unusual.
Shintanis work opens the door to the study of cubic fields using analytic number
theory. In particular, by Perrons formula and standard techniques we have
X Z 2Ci1
1 Xs
D .s/ ds (5.17)
jStab.x/j 2i1 s
x2GL2 .Z/nVZ
Disc.x/<X 6
D RessD1 .s/X C RessD5=6 .s/X 5=6 C O.X 3=5C /:
5
Although the left side is not the counting function of cubic fields, for the first time we
see the X 5=6 secondary term. This, therefore, gives an explanation for the secondary
terms in Theorems 5.1 and 5.2, because the Shintani zeta functions have secondary
poles.6
To count cubic fields, the most important step is sieving for maximality. This is
possible by work of Datskovsky and Wright [13,14,41], who gave Shintanis work an
adelic formulation along the lines of Tates thesis [39]. This allowed them to incorpo-
rate a variety of conditions into the definition of the Shintani zeta functions, including
the DavenportHeilbronn maximality conditions (given in Proposition 5.7).
A bogus proof of Roberts conjecture is as follows. For a set of primes P , define
the P -maximal Shintani zeta function by the Dirichlet series (5.16), with the added
condition that the only cubic forms counted are those in the set Up for each p 2 P .
Therefore, letting P be the set of primes < X , it follows that
X0 Z 2Ci1
1 Xs
D P .s/ ds (5.18)
jAut.R/j 2i1 s
Disc.x/<R
6
D RessD1 P .s/X C RessD5=6 P .s/X 5=6 C O.X 3=5C /;
5
where the sum is over all maximal cubic rings, corresponding to fields of degree 3,
with Disc.R/ < X . The quadratic fields (and Q) can easily be counted and sub-
tracted, but the error term in (5.18) depends on the particular zeta function P , and
6 Of
course this begs the question of why the Shintani zeta functions have poles at s D 5=6. This can
of course be explained by the various calculations in Shintanis work, but for now a satisfying highbrow
argument is unknown to the author.
therefore the implied constant depends on X (indeed, when standard techniques are
used, exponentially so!).
However, this flawed argument enabled Datskovsky and Wright [14] to give a rigor-
ous analytic proof of the DavenportHeilbronn theorem. The key step7 is to use (5.18),
but with P equal to the set of all primes < Y , where Y tends to infinity very slowly
with X . Given their adelic framework, their work yielded an asymptotic formula for
the number of cubic extensions of any global field of characteristic not 2 or 3.
Moreover, Roberts [32] computed the limit of the secondary terms in (5.18) and
found an excellent match with numerical data. This led him to conjecture that (5.18)
(equivalently, (5.1)) was correct with some undetermined error term smaller than
X 5=6 .
He closed his paper with the following (paraphrased very slightly):
The pessimistic discussion in [14] suggests to us that the way may be difficult. However, one

ingredient of a proof might be the functional equation of P .s/ with respect to s ! 1 s,
studied in [13, 41]. Another ingredient might be [34, Theorem 3], which concerns growth
of arithmetic functions whose associated Dirichlet series satisfy such a functional equation.
This is actually a good summary of our proofs of Theorems 5.1 and 5.2! To get the
ball rolling, we needed to incorporate the sieve used by Belabas, Bhargava, and Pomer-
ance described in Section 5.3. In place of the P -maximal zeta function we introduced
the q-nonmaximal zeta function, which counts only cubic rings not maximal at each
prime dividing q. This allowed us to count maximal cubic rings, and the irreducible
such rings correspond to cubic fields.
Remark 5.8. As described at the end of Section 5.2, we count 3-torsion elements of
class groups by a variation of this argument. Such elements correspond to cubic fields
not totally ramified at any prime, and so we expand the definition of the q-nonmaximal
zeta function to count cubic rings either nonmaximal or totally ramified at every prime
dividing q.
Equation (5.18) holds for each q-nonmaximal zeta function, and the q-dependence
of the error term can be bounded in terms of the exponential sum
X X 1 X

bq .x/j WD
j q .y/ exp.2 i x; y=q /; (5.19)
2
q8
x2VZ=q 2 Z x2VZ=q 2 Z y2VZ=q 2 Z
where q .x/ is the characteristic function of the Davenport-Heilbronn q-nonmaxi-

mality condition, and x; y D x4 y1 13 x3 y2 C 13 x2 y3 x1 y4 : Each inner sum
appears naturally when we shift the contour in Equation (5.18) to the left of <.s/ D 0
and apply the functional equation. Trivially, the sum in Equation (5.19) is bounded
7 We have adapted their argument somewhat to our point of view. Datskovsky and Wright also need,
as did Davenport and Heilbronn, the bound (5.15) for the number of cubic rings nonmaximal at any
prime > Y .
66 Frank Thorne
by q 8 ; this is enough to allow us to take Q D X 1=25 in (5.15) and obtain an error term
of X 24=25C in the Davenport-Heilbronn theorems. We could improve this using the
CauchySchwarz inequality, but not below X 5=6 .
In [37] we incorporated several improvements to lower the error terms in Theo-
rems 5.1 and 5.2 to O.X 7=9C / and O.X 18=23C /, respectively. Among these is a
careful analysis of the sum in Equation (5.19), carried out in [38]. For each x we ob-
tained exact formulas for the inner sum in Equation (5.19), proving that the outer sum
is q 1C .
5.5.1 Nonequidistribution in Arithmetic Progressions

The Shintani zeta-function approach can also be used to study the distribution of cubic
field discriminants in arithmetic progressions. To do this, one twists the Shintani zeta
functions by Dirichlet characters, obtaining L-functions which are also proved to enjoy
analytic continuation and a functional equation.
The details are complicated, so here we simply present representative numerical
data. The tables below list the number of cubic fields K with 0 < Disc.K/ < 2 106
and Disc.K/ a .mod m/, for m D 5 and m D 7, and each residue class a.
Discriminant modulo 5 0 1 2 3 4
Actual count 21277 22887 22751 22748 22781
Theoretical result 21307 22757 22757 22757 22757
Difference 30 130 6 9 24
Discriminant modulo 7 0 1 2 3 4 5 6
Actual count 15330 17229 14327 15323 17027 18058 15150
Theoretical result 15316 17209 14277 15316 17024 18063 15131
Difference 14 20 50 7 3 5 19
The data modulo 5 is essentially equidistributed, apart from the a D 0 column. (The
relative deficit of cubic fields divisible by 5, or by any other modulus, can be explained
by a careful reading of Davenport and Heilbronns original paper.) However, the data
modulo 7 shows a striking lack of equidistribution.8
The theoretical results above come from a generalization of Theorem 5.1 to arith-
metic progressions; the secondary term now involves a sum of residues of twisted
Shintani zeta functions. These residues are evaluated in [38], and for characters for
8 The counting functions for cubic field discriminants 0 .mod 7/ and 3 .mod 7/ are not the same.
The main term for 0 .mod 7/ is smaller than that for 3 .mod 7/, the secondary term is larger, and the
equality in the second row of the table is a coincidence of roundoff error.
which 6 1, the associated residue at s D 5=6 is zero. However, when 6 D 1 this

residue is often nonzero, and nonzero residues explain biases in arithmetic progres-
sions such as the one seen above.
5.6 A Refined Geometric Approach

To give an overview of Bhargava, Shankar, and Tsimermans proof [7], we start with
their treatment of the main term in Theorem 5.1. Later, we will describe how the sec-
ondary term appears in their work, but only in the context of irreducible cubic forms
(corresponding to cubic orders).
The argument in [7] largely follows Davenport and Heilbronns original work. Fix
a sign, and let Virr denote the irreducible points x 2 VZ with Disc.x/ > 0. Write
nC D 6, n D 2 for the order of the stabilizers of the action of GL2 .R/ on VR . The
equality nC D 3n is reflected in the fact that cubic fields of negative discriminant
are three times as common as those of positive discriminant.
Write F for a certain fundamental domain for GL2 .Z/nGL2 .R/ in GL2 .R/ (see
Equation (12) of [7]), originally constructed by Gauss. For any v 2 V , consider the
multiset F v, where the multiplicity of an element x 2 V is equal to the number
of g 2 F for which gv D x. It is readily checked that each element x 2 GZ nVZ
is represented in this multiset n =m.x/ times, where m.x/ denotes the size of the
stabilizer of x in GL2 .Z/. We conclude that
X 1 1
N.V I X / WD D x 2 F v \ Virr W jDisc.x/j < X :
jStab.x/j n
0<Disc.x/<X
x irred.
(5.20)
The reducible points can be handled without too much difficulty: The number of
reducible integral binary cubic forms au3 C bu2 v C cuv 2 C dv 3 in the multiset
RX .v/ WD w 2 F v W jDisc.w/j < X , satisfying a 0, is X 3=4C . Conversely,
all cubic forms with a D 0 are reducible. This condition a 0 meshes well with the
1
geometric arguments that follow. In addition, the weight jStab.x/j occurring in (5.20)
can be neglected, as there are X 1=2 points x with jDisc.x/j < X with nontrivial
stabilizer.
The formula in Equation (5.20) does not depend on v, and one innovation of [7]
(previously used by Bhargava in [3]) is to average over many v. In particular, define
B WD w D .a; b; c; d / 2 V W 3a2 C b 2 C c 2 C 3d 2 10; jDisc.w/j 1, and we
have
R 1
v2B\V n x 2 F v \ Virr W jDisc.x/j < X jDisc.v/j dv
1

N.V I X / D R :
1
v2B\V jDisc.v/j dv
(5.21)
68 Frank Thorne
The authors describe this step as thickening the cusp. As jDisc.v/j1 dv is GL2 .R/-
invariant, we may rewrite this as
Z
1
N.V I X / D #x 2 Virr \ gB W jDisc.x/j < X dg; (5.22)
M g2F
n R
where M WD 2 1
v2B\V jDisc.v/j dv: This is rewritten as
Z
1
N.V ; X / D #x 2 Virr \ B.n; t; ; X /t 2 d nd t d (5.23)
M 0
g2N .a/A 0
using a standard decomposition of F ((12) of [7]), where B.n; t; ; X / is the region

x 2 gB W jDisc.x/j < X . A proposition of Davenport establishes that in this
situation, the count of lattice points above is well approximated by the volume of
B.n; t; ; X /, provided that t is not too large. Conversely, when t is large all the lat-
tice points are in the cusp a D 0, and hence reducible and not counted. Thus, the above
is reduced to
Z
1
N.V ; X / D Vol.B.n; t; ; X //t 2 d nd t d C O. /:
M g2N 0 .a/A0
t<C 1=3 1=3 (5.24)
The condition on t is now removed, subject to an error term, and this integral is
equal to
Z
1
N.V ; X / D Vol.RX .v//jDisc.v/j1 dv C O. /: (5.25)
2M v2B\V
This volume does not in fact depend on v, so that we have (excluding error terms)
1 2
N.V ; X / D Vol.RX .v// D X: (5.26)
n 12n
5.6.1 Origin of the Secondary Term

We now explain how Bhargava, Shankar, and Tsimerman refine these calculations to
obtain a secondary term. Our brief explanation will necessarily be somewhat vague,
which should encourage the reader to read [7]. We focus on their count of irreducible
binary cubic forms; as in [2, 37], they also incorporate a sieve to count cubic fields.
They begin by incorporating several tweaks to Equation (5.23). For technical rea-
sons, they count discriminants in dyadic intervals X=2; X . In addition, they introduce
a smooth function 0 .t / on R0 , such that 0 .t / D 0 for t 2 and 0 .t / D 1 for
t 3. They thus rewrite Equation (5.23) (restricted to a dyadic interval X=2; X ) as
Z
1 t t
N.V ; X=2; X / D 0 1=3 C 1=3
M g2N 0 .a/A0
#x 2 Virr \ B.n; t; ; X=2; X /t 2 d nd t d ; (5.27)
where WD 1 0 , and is a parameter to be chosen later (to minimize error terms).

The decomposition C 0 D 1 splits this integral into two, and we restrict our
attention to the 0 part, which yields the secondary term.
They now slice the count of binary cubic forms by the first coordinate. In par-
ticular, for a 2 Z, let Ba .n; t; ; X=2; X / denote the set of binary cubic forms in
Ba .n; t; ; X=2; X / whose u3 coefficient is equal to a. Then we have

# x 2 Virr \ B.n; t; ; X=2; X / D
X
# x 2 Virr \ Ba .n; t; ; X=2; X / : (5.28)
a2Z
a0
Skipping ahead in their work, the 0 -contribution to (5.27) is equal to

1 Z Z 1=3 10=3 1=3
2 X X
1=4
u u
p 0 1=3 1=3
3M aD1 D. 3=2/3 =C u>0 a a

Vol Bu .X=.24 /; X=4 / d ud C O. /; (5.29)
where Ba .Y =2; Y / denotes the set of cubic forms in B with first coordinate a and
discriminant in Y =2; Y . By Mellin inversion, we have
X
1 1=3
1=3 u e0 .2/. 3 u/2=3 C O. /;
a 0 1=3
D .1=3/ C 3 (5.30)
aD1
a
and it is this (negative) .1=3/ which contributes the secondary term. In contrast, the
second term of Equation (5.30) (which looks larger, especially in light of the even-
tual choice D X 1=12 ) is reincorporated into Equation (5.29) and then combined with
the term of Equation (5.27) to obtain the main term in Equation (5.26). When the
.1=3/ term is plugged into Equation (5.29), the resulting integral can be evaluated
without undue difficulty, and it has order X 5=6 .
Remark 5.9. As Shankar has explained to the author, a more geometric argument
can also be given which does not rely on the analytic continuation of the Riemann zeta
function. However, it is unclear whether the resulting error term would be smaller than
X 5=6 when combined with the sieve.
5.6.2 A Correspondence for Cubic Forms

To count cubic fields one must sieve for maximality, using the DavenportHeilbronn
conditions of Proposition 5.7. These could be treated in a naive manner, but the authors
improve their error terms by introducing a useful correspondence for nonmaximal cu-
bic forms.9
9 As Shankar explained to me, in the case of cubic fields it is this correspondence which allows them
to cross the X 5=6 barrier.

70 Frank Thorne
We introduce some notation. Let N .VZ I X / count all cubic forms v with 0 <
Disc.v/ < X . For a prime p, let N .Upc I X / count those cubic forms which are
nonmaximal at p, i.e., which do not lie in the set Up defined in Proposition 5.7. (In
referring to cubic forms as nonmaximal we are implicitly appealing to the Delone
Faddeev correspondence.) Finally, for any 2 P 1 .Z=pZ/, let N .Vp; I X / count
those cubic forms v such that the reduction of v modulo p has as a root.
With this notation, we have the following correspondence:
Proposition 5.10 ([7]). We have the identity

X X
N .Upc I X / D N .Vp; I X=p 2 / N .Vp; I X=p 4 /CN .VZ I X=p 4 /:
2P 1 .Fp / 2P 1 .Fp /
(5.31)
This identity generalizes in a straightforward way from prime p to squarefree q,
where we count cubic forms not in Up for any pjq. The identity amounts to a combi-
natorial argument, enumerating ways in which a nonmaximal cubic ring is contained
in a larger ring (with smaller discriminant), and then counting these larger rings.
This identity is crucial in [7], and it can also be translated into an identity for the
q-nonmaximal Shintani zeta function! In work in progress by Bhargava, Taniguchi,
and the author [8] we are developing this combined approach. We have proved an
error term of O.X 2=3C / in Theorem 5.1 as well as a secondary term for relative cubic
extensions of quadratic base fields, and we are currently working on further extensions
and generalizations.
5.7 Equidistribution of Heegner Points

Remarkably, Hough [24] demonstrated that the DeloneFaddeev correspondence is
not the only path to the DavenportHeilbronn theorem. Hough studied the distribution
of Heegner points associated to 3-torsion elements of ideal class groups of quadratic
fields, and he obtained a result related to Theorem 5.2 as a consequence.
For simplicity, Hough worked only with imaginary discriminants D 2 .mod 4/;
notepthat any such discriminant is not fundamental, and the discriminant of the field
Q. D/ is 4D.10 He proves that
X 4
#Cl3 .D/ D 2 X C O.X 19=20C /; (5.32)

0<D<X
D2 .mod 4/
10 It seems that the restrictions that D 2 .mod 4/ and D < 0 could both be lifted with some
p
effort. The restriction on the sign is the more serious of the two, as elements of class groups of Q. D/
for positive D correspond to closed geodesics rather than Heegner points. However, Dukes result holds
for either sign, so in [24] Hough expresses optimism that the positive discriminant case of Theorem 5.2
could be handled as well.
and for compactly supported, infinitely differentiable test functions , he proves that
X 4 b
#Cl3 .D/.D=X / D .1/X C O.X 7=8C /: (5.33)
2
0<D
D2 .mod 4/
The correct secondary term of order X 5=6 appears in both formulas, in spite of the
larger error terms, and in follow-up work (in progress) he obtains an error less than
X 5=6 in (5.33), thereby obtaining a proof of a related secondary term.
Houghs work has very different prospects for generalization from the approaches
described previously. In particular, his approach can be used to study k-torsion in class
groups for odd k > 3, and he makes the following conjecture:
Conjecture 5.11 (Hough [24]). For good test functions and odd k 3, we have
X
#Clk .D/.D=X / D
0<D
D2 .mod 4/
4b b 1 1
.1/X C C1;k C X 1=2C1=k C o.X 1=2C1=k /; (5.34)
2 2 k
where
1 .1 k2 / . 12 /. 12 k1 /
1=k 11=k

C1;k WD 1 2 C 2
6k .2/ .1 k1 /
Y
1 1=k 1C2=k 1C1=k 1

1C p p p p : (5.35)
pC1
p>2
Indeed he proves his conjecture, but with error terms larger than X . Even a proof
of the main term for the single case k D 5 would be a major achievement. Indeed, the
average size of Clp .D/ was conjectured by Cohen and Lenstra [11] to be 2 for each
prime p 3, and Cohen and Lenstra further conjectured that the p-part of Clp .D/ is
1
isomorphic to a fixed p-group H with probability proportional to jAut.H /j
. No case of
these conjectures is currently known for any p 5, and essentially nothing is known
for p D 3 beyond Theorem 5.2.
5.7.1 Heegner Points and Equidistribution

Houghs main result concerns the equidistribution of Heegner points associated to the
the relevant background. If a D x1 ; x2 is an ideal
3-part of the class group. We recall p
of an imaginary quadratic field Q. D/, where by reordering if necessary we have
x1 =x2 2 H WD z 2 C W =.z/ > 0, we call x1 =x2 the Heegner point za associated
to the ideal a. A change of basis for a corresponds to a linear fractional transformation
72 Frank Thorne
aza Cb
za ! cz a Cd
with ad bc D 1, and therefore the Heegner point za is well-defined in
p
the quotient SL2 .Z/nH. One checks thatpza D za for any 2 Q. D/, and hence
za depends only on the class ofpa in Cl. D/. One similarly checks that za zb if
a and b are inequivalent in Cl. D/. p
Therefore, there are h.D/ Heegner points associated to Q. D/ in SL2 .Z/nH,
and in [17] Duke proves that they equidistribute with respect to the hyperbolic measure
dx dy
y2
as D ! 1. Hough proves the same for the Heegner points associated to the
3-part of the ideal class group:
Theorem 5.12 (Hough [24]). Let B be a bounded Borel measurable subset of H

having boundary of measure zero. Then as D ! 1,
X p 6X
# a primitive, nonprincipal W a 2 C l3 .Q. D//; za 2 B 3 vol.B/:

0>D>X
d 2 .mod 4/ (5.36)
squarefree
Here we take B H rather than B SL2 .Z/nH, so that each Heegner point
is counted once for each representative of its SL2 .Z/ orbit in B. (We recall that a
primitive ideal is one without any rational integer divisors other than 1.)
Straightaway one can see that this is an equidistribution statement. But it also im-
plies the DavenportHeilbronn theorem! Taking B to be the standard (or any other)
fundamental domain for the action of SL2 .Z/pon H, one counts each Heegner point
(and thus each nontrivial element of Cl3 .Q. D//) exactly once. The Davenport-
Heilbronn theorem follows from the classical computation vol.B/ D 3 .
We will follow Houghs approach sufficiently far as to see the secondary term.
His work begins with the following parameterization of ideals (and therefore Heegner
points), essentially due to Soundararajan [36].
Proposition 5.13 ([24, 36]). Let D 2 .mod 4/ be squarefree and let k 3 be odd.
The set
.l; m; n; t / 2 .ZC /4 W lmk D l 2 n2 C t 2 D; ljD; .m; ntD/ D 1 (5.37)
p bijection with primitive ideal pairs .a; a/ with a 1 and a principal in

is in k
Q. D/. Explicitly, the ideal a is given as a Z-module by

p
a D lm; lnt 1 C D
where Na D lm and t 1 is the inverse of t modulo m.
p straightforward. For example, if a kis principal

k and coprime
The proof is relatively
to D, then a D .n C t D/ and taking norms yields .Na/ D n2 C t 2 D, a solution
k
to Equation (5.37). Ideals a not coprime to D yield solutions to Equation (5.37) with
l > 1.
As one of the two main steps in his proof, Hough now applies this parameterization
to count Heegner points in the region

1 1 1
RY WD z 2 H W < <.z/ < ; =.z/ > : (5.38)
2 2 Y
This establishes the vertical distribution of Heegner points, and he separately estab-
lishes that they are equidistributed horizontally. He proves that the number of Heegner
points in RY is equal to
6
YX C C5=6 X 5=6 C O. /; (5.39)
3
for a fairly complicated error term depending on both X and Y . The secondary term
appears here for the following
p reason: By the parameterization above, we may write
an ideal a as lm; lnt 1 C D. We have lm3 D l 2 np2 C t 2 D, so that lm D 1=3 ,
implying that the Heegner point za has imaginary part lmD D 1=6 X 1=6 . In other
words, the Heegner points equidistribute as X ! 1, but for fixed X they do not
appear high in the cusp.
The hyperbolic volume of the subset of RY with =.z/ > X 1=6 is equal to
Z 1=2 Z 1
dx dy
D X 1=6 : (5.40)
xD1=2 yDX 1=6 y2
Accounting (much more carefully than done here!) for the fact that this subset con-
tains no Heegner points yields the negative secondary term in Equation (5.39), and an
analogous argument explains the second term in Equation (5.34).
Remark 5.14. Houghs methods also extend to arithmetic progressions, where he

finds the same lack of equidistribution observed at the end of Section 5.5; some re-
sults along these lines are in preparation.
5.8 Hirzebruch Surfaces and the Maroni Invariant

Zhaos work [45] is still in preparation, so we offer only a very brief overview here.
Using algebraic geometry, Zhao estimates the number of cubic extensions of the ra-
tional function field Fq .t /.
Fix a finite field Fq of characteristic not equal to 2 or 3, and let S.2N / be the
number of isomorphism classes of cubic extensions K=Fq .t / of degree 3, such that
jDisc.F /j D q 2N . In [14], Datskovsky and Wright proved that
RessD1 Fq .t/ .1/

S.2N / D 2 q 2N C o.q 2N /; (5.41)
Fq .t/ .3/
74 Frank Thorne
where Fq .t/ .s/ is the Dedekind zeta function of Fq .t /, or equivalently the zeta func-
tion of the algebraic curve PF1q .
Datskovsky and Wrights proof uses the adelic Shintani zeta function; in [45] Zhao
obtains another proof of Equation (5.41) using algebraic geometry. A trigonal curve
is a threefold cover of P 1 , and there is a bijection between isomorphism classes of
smooth trigonal curves and cubic rational function fields. Already this perspective is
enlightening; for example, the RiemannHurwitz formula implies that the discriminant
has even degree, explaining why Equation (5.41) is an equation for S.2N / and not
S.N /.
The problem, then, is to count smooth trigonal curves. These may be counted by
embedding the curves into Hirzebruch surfaces Fk ; for each trigonal curve C there is
a unique nonnegative integer k, called the Maroni invariant of C , such that C embeds
into Fk WD PP 1 .O O.k//:
Zhao now counts curves on each surface Fk and sieves for smoothness, using a
sieve method related to one developed by Poonen [31]. However, the main technique
is different and new. The main term in Equation (5.41) comes from estimating the
number of trigonal curves C with each Maroni invariant k and then summing the
results over all integers k. However, the Maroni invariant of a smooth trigonal curve
C is at most N=3, provided that the field corresponding to C has absolute discriminant
q 2N . Limiting our sum to k N=3 naturally introduces a secondary term of order
q 5N=3 into Equation (5.41).
Zhaos method, like all other methods described in this paper, comes with error
terms; at present he has obtained an error term of O.N 2 q 5N=3 / in Equation (5.41),
narrowly missing a proof of the secondary term. He has some optimism that this error
term can be reduced below q 5N=3 , and also that some of his methods may be adapted
to say something interesting in the number field case.
5.9 Conclusion
Cubic fields have seen a great deal of recent study recently, and we recommend the pa-
pers of CohenMorra [12] and MartinPollack [26] (among others) for recent studies
of cubic fields from other perspectives.
Many interesting open questions remain in the subject. For example, what is the
correct error term in the DavenportHeilbronn theorems? Even if we cannot prove
it, it would still be interesting to determine the expected order of magnitude. To get
the best error terms it seems that we should count fields K of degree at most 3, each
1
weighted by jAut.K/j . This is suggested both by numerical data and also by consid-
ering the space of binary cubic forms, where reducible maximal forms correspond to
quadratic fields or to Q. (Note that this heuristic does not generalize directly to higher
degree fields.) The data in [32] suggests that the true error may be smaller than X 1=2 ,
and a comparison with the divisor problem suggests that the error might be on the
order of X 3=8 . Nevertheless, the comparison with the divisor problem is not exact,
and the numerical data is inconclusive.
Perhaps the most compelling open question concerns quartic and quintic fields.
Asymptotic formulas were proved by Bhargava [3,6]; should these formulas have sec-
ondary terms? It is generally believed that they should. However, so far we lack even
a good conjecture. The zeta-function approach seems likely to work, and Yukie [44]
has made some progress in this direction; the geometric approach seems likely to work
as well. However, for now the difficulties with either approach appear rather severe.
Another open question concerns the multiplicity of cubic fields. It is believed that
there should be n cubic fields of discriminant n, but the best bound known is
O.n1=3C /, due to Ellenberg and Venkatesh [19]. Nontrivial bounds were also ob-
tained by Helfgott and Venkatesh [22] and Pierce [30], using a variety of methods.
Any of the methods described in the present paper could potentially yield improve-
ments, but none of them have succeeded to date. (The present author has attempted
improvements by means of Shintani zeta functions, which have led to a number of
instructive and interesting failures.)
Finally, this paper raises the question of whether or not unexpected connections
might be found between the four perspectives presented here, and some promising
preliminary results have been obtained in this direction. We anticipate that this and
related questions will be addressed in the near future.
Acknowledgments. I would like to thank the many people who have shared their in-
sights on these secondary terms, especially Manjul Bhargava, Jordan Ellenberg, Bob
Hough, Arul Shankar, Kannan Soundararajan, Takashi Taniguchi, Melanie Matchett
Wood, and Yongqiang Zhao. I would further like to thank Bhargava, Shankar,
Taniguchi, Zhao, and an anonymous referee for comments on this paper in particu-
lar. Most of all I would like to thank Akshay Venkatesh who originally brought this
topic to my attention.
In addition, I would like to thank the organizers of Integers 2011 for hosting an
outstanding conference.
References
[1] A. M. Baily, On the density of discriminants of quartic fields, J. Reine Angew. Math. 315
(1980), 190210.
[2] K. Belabas, M. Bhargava, and C. Pomerance, Error estimates in the Davenport-Heilbronn
theorems, Duke Math. J. 153(1) (2010), 173210.
[3] M. Bhargava, The density of discriminants of quartic rings and fields, Ann. of Math. 162(2)
(2005), 10311063.
[4] M. Bhargava, Mass formulae for extensions of local fields, and conjectures on the density
of number field discriminants, Int. Math. Res. Not. (2007), no. 17, Art. ID rnm052.
76 Frank Thorne
[5] M. Bhargava, Higher composition laws and applications, in: Proceedings of the Interna-
tional Congress of Mathematicians, vol. II, 271294, Eur. Math. Soc., Zrich, 2006.
[6] M. Bhargava, The density of discriminants of quintic rings and fields, Ann. of Math.
172(3) (2010), 1559-1591.
[7] M. Bhargava, A. Shankar, and J. Tsimerman, On the Davenport-Heilbronn theorem and
second order terms, preprint; available at http://arxiv.org/abs/1005.0672, last accessed
6/19/2013.
[8] M. Bhargava, T. Taniguchi, and F. Thorne, work in preparation.
[9] H. Cohen, F. Diaz y Diaz, and M. Olivier, Enumerating quartic dihedral extensions of Q,
Compositio Math. 133 (2002), 6593.
[10] H. Cohen, F. Diaz y Diaz, and M. Olivier, Counting discriminants of number fields,
J. Thor. Nombres Bordeaux 18(3) (2006), 573593.
[11] H. Cohen and H. Lenstra, Heuristics on class groups of number fields, in: Number theory,
Noordwijkerhout 1983, pp. 3362, Lecture Notes in Math. 1068, Springer, Berlin, 1984.
[12] H. Cohen and A. Morra, Counting cubic extensions with given quadratic resolvent, J. Al-
gebra 325 (2011), 461478.
[13] B. Datskovsky and D. Wright, The adelic zeta function associated to the space of binary
cubic forms. II. Local theory, J. Reine Angew. Math. 367 (1986), 2775.
[14] B. Datskovsky and D. Wright, Density of discriminants of cubic extensions, J. Reine
Angew. Math. 386 (1988), 116138.
[15] H. Davenport and H. Heilbronn, On the density of discriminants of cubic fields. II, Proc.
Roy. Soc. London Ser. A 322(1551) (1971), 405420.
[16] B. N. Delone and D. K. Faddeev, The theory of irrationalities of the third degree (in
English translation), AMS, Providence, 1964.
[17] W. Duke, Hyperbolic distribution problems and half-integral weight Maass forms, Invent.
Math. 92(1) (1988), 7390.
[18] J. Ellenberg and A. Venkatesh, The number of extensions of a number field with fixed
degree and bounded discriminant, Ann. of Math. 163(2) (2006), 723741.
[19] J. Ellenberg and A. Venkatesh, Reflection principles and bounds for class group torsion,
Int. Math. Res. Not. (2007), no. 1, Art. ID rnm002.
[20] W. T. Gan, B. Gross, and G. Savin, Fourier coefficients of modular forms on G2 , Duke
Math. J. 115 (2002), 105169.
[21] E. Golod and I. Shafarevich, On the class field tower, Izv. Akad. Nauk SSSR Ser. Mat. 28
(1964), 261272, in Russian.
[22] H. Helfgott and A. Venkatesh, Integral points on elliptic curves and 3-torsion in class
groups, J. Amer. Math. Soc. 19(3) (2006), 527550.
[23] C. Hermite, Extrait dune lettre de M. C. Hermite M. Borchardt sur le nombre limit
dirrationalits auxquelles se rduisent les racines des quations coefficients entiers
complexes dun degr et dun discriminant donns, J. Reine Angew. Math. 53 (1857),
182192.
[24] B. Hough, Equidistribution of Heegner points associated to the 3-part of the class group,
preprint, available at http://arxiv.org/abs/1005.1458, last accessed 6/19/2013.
[25] F. Lemmermeyer, Class field towers, http://www.rzuser.uni-heidelberg.de/hb3/publ/

pcft.pdf, last accessed 6/19/2013.
[26] G. Martin and P. Pollack, The average least character nonresidue and further variations
on a theme of Erdos, preprint, available at http://arxiv.org/abs/1112.1175, last accessed
6/19/2013.
[27] J. Martinet, Petits discriminants, Ann. Inst. Fourier 29 (1979), 159170.
[28] J. Neukirch, Algebraic number theory, Springer-Verlag, Berlin, 1999.
[29] A. Odlyzko, Bounds for discriminants and related estimates for class numbers, regulators
and zeros of zeta functions: a survey of recent results, Sm. Thor. Nombres Bordeaux
(2) 2(1) (1990), 119141.
[30] L. Pierce, The 3-part of class numbers of quadratic fields, J. London Math. Soc. (2) 71(3)
(2005), 579598.
[31] B. Poonen, Bertini theorems over finite fields, Ann. of Math. (2) 160(3) (2004), 1099
1127.
[32] D. Roberts, Density of cubic field discriminants, Math. Comp. 70(236) (2001), 1699
1705.
[33] M. Sato and T. Kimura, A classification of irreducible prehomogeneous vector spaces
and their relative invariants. Nagoya Math. J. 65 (1977), 1155.
[34] M. Sato and T. Shintani, On zeta functions associated with prehomogeneous vector
spaces, Ann. of Math. (2) 100 (1974), 131170.
[35] T. Shintani, On Dirichlet series whose coefficients are class numbers of integral binary
cubic forms, J. Math. Soc. Japan 24 (1972), 132188.
[36] K. Soundararajan, Divisibility of class numbers of imaginary quadratic fields, J. London
Math. Soc. (2) 61 (2000), 681690.
[37] T. Taniguchi and F. Thorne, Secondary terms in counting functions for cubic fields, sub-
mitted, available at http://arxiv.org/abs/1102.2914, last accessed 6/19/2013.
[38] T. Taniguchi and F. Thorne, Orbital L-functions for the space of binary cubic forms,
submitted, available at http://arxiv.org/abs/1112.5030, last accessed 6/19/2013.
[39] J. Tate, Fourier analysis in number fields and Heckes zeta-functions, thesis (Princeton,
1950), reprinted in J. W.S. Cassels and A. Frhlich, eds., Algebraic number theory, Aca-
demic Press, London, 1986.
[40] S. Wong, Automorphic forms on GL.2/ and the rank of class groups, J. Reine Angew.
Math. 515 (1999), 125153.
[41] D. Wright, The adelic zeta function associated to the space of binary cubic forms. I.
Global theory, Math. Ann. 270(4) (1985), 503534.
[42] D. Wright, Twists of the Iwasawa-Tate zeta function, Math. Z. 200 (1989), 209231.
[43] D. Wright and A. Yukie, Prehomogeneous vector spaces and field extensions, Invent.
Math. 110(2) (1992), 283314.
[44] A. Yukie, Shintani zeta functions, London Mathematical Society Lecture Note Series
183, Cambridge University Press, Cambridge, 1993.
[45] Y. Zhao, doctoral thesis, University of Wisconsin, in preparation.
78 Frank Thorne
Author information
Frank Thorne, Department of Mathematics, University of South Carolina, Columbia, South
Carolina, USA.
Email: thorne@math.sc.edu
Chapter 6
Spotted Tilings and n-Color Compositions

Brian Hopkins
Abstract. Spotted tilings are presented as a new combinatorial interpretation of n-color com-
positions. Previous and new results are proven using this tool, an interpretation of the Fibonacci
numbers, and a case of Terquems problem. The spotted tilings allow for MacMahons zig-zag
graphs to be applied to n-color compositions, addressing a question of Agarwal in the article
where these compositions were introduced in 2000.
Keywords. Integer Compostions, Tilings, Zig-Zag Graph.
6.1 Background
For a given positive integer n, the compositions of n are ordered t -tuples .c1 ; : : : ; c t /
of positive integers with c1 C C c t D n. For instance, there are four compositions
of 3, namely .3/; .2; 1/; .1; 2/; and .1; 1; 1/. The individual ci are called parts of the
composition. Compositions are sometimes referred to as ordered partitions.
Compositions of n may be represented graphically as tilings of a 1 n board, where
a part k corresponds to a 1 k rectangle. This is essentially MacMahons construction
using nodes on a line [7]. Figure 6.1 shows the tilings for the compositions of 3. Notice
that each tiling has a vertical bar at the leftmost edge, possibly vertical bars separating
the rectangles corresponding to parts, and a vertical bar at the rightmost edge.
3 21 12 111
Figure 6.1. The 4 tilings representing the compositions of 3.
Agarwal [1] introduced the concept of n-colored compositions, where a part k has
one of k possible colors, denoted by a subscript 1; : : : ; k. There are eight n-colored
compositions of 3, namely
.31 /; .32 /; .33 /; .21 ; 11 /; .22 ; 11 /; .11 ; 21 /; .11 ; 22 /; .11 ; 11 ; 11 /:
Let C C.n/ denote the number of n-color compositions of n, so C C.3/ D 8. These
compositions, inspired by the analogous concept of n-color partitions, have been fur-
ther considered in several articles and monographs.
80 Brian Hopkins
We introduce a combinatorial tool, spotted tilings, to provide a graphic representa-

tion of n-colored compositions. A part ki corresponds to a 1 k rectangle with a spot
in position i . Spotted tilings for the n-color compositions of 3 are shown in Figure 6.2.
Notice that the bars and spots alternate, since there is one spot per part.
31 u 21 11 u u 11 21 u u 11 11 11 u u u
32 u 22 11 u u 11 22 u u
33 u
Figure 6.2. The 8 spotted tilings for C C.3/.
We will use this new tool to prove previous and new results about colored compo-
sitions (Section 6.2) and to develop an analog of MacMahons zig-zag graphs (Sec-
tion 6.3), addressing a question of Agarwal [1].
There are two combinatorial results we will use in the sequel. The Fibonacci num-
bers are given by F0 D 0, F1 D 1 and the recurrence Fn D Fn1 C Fn2 for n 2.
The following theorem is proven using generating functions in [6].
Theorem 6.1. Let C o.n/ denote the number of compositions of n whose parts are all
odd. Then C o.n/ D Fn .
Proof. Proceed by induction. Certainly C o.0/ D 0, as 0 cannot be written as a sum of

positive odd integers. Also, C o.1/ D 1, since the single composition of 1 consists of
an odd part. Assume the claim is true for n1 and n2. The compositions counted by
C o.n/ include the odd-part compositions of n 1 with an additional part 1 included at
the end, and the odd-part compositions of n2 with the last part increased by 2 (which
results in another odd part). Since any odd-part composition of n ends in either 1 or a
larger odd number, C o.n 1/ C C o.n 2/ counts all odd-part compositions of n by
the induction hypothesis.
Terquem [9] considered the problem of counting subsequences of 1; : : : ; n where

the terms alternate parity. For instance, such 3-term subsequences of 1; : : : ; 5 with
first term odd are 1; 2; 3, 1; 2; 5, 1; 4; 5, and 3; 4; 5. Church and Gould [4] give
a combinatorial proof of Terquems general problem. We give a different proof of the
special case needed below.
Theorem 6.2. The number of subsequences x1 ; : : : ; x2j 1 of 1; :::; 2k 1 where

Ck1
the parity of xi matches the parity of i is given by j2j 1 .
Proof. Let T .j; k/ denote the number of desired subsequences. Since there are k odd
numbers in 1; : : : ; 2k 1, we have T .1; k/ D k. To determine T .j; k/ inductively,
there are two cases to consider. If a subsequence counted by T .j; k/ does not include
Chapter 6 Spotted Tilings and n-Color Compositions 81
the number 2k 1, then it is among the subsequences counted by T .j; k 1/ D

j Ck2
2j 1 . If a subsequence counted by T .j; k/ does end with the number 2k 1, then
the second to last entry is one of the even numbers 2j 2; : : : ; 2k 2 and the number
of possible x1 ; : : : ; x2j 3 are counted by T .j 1; j 1/; : : : ; T .j 1; k 1/,
respectively. Together, this gives
T .j; k/ D T .j; k 1/ C T .j 1; k 1/ C C T .j 1; j 1/
! ! !
j Ck2 j Ck3 2j 3
D C C C
2j 1 2j 3 2j 3
! !
j Ck2 j Ck2
D C
2j 1 2j 2
!
j Ck1
D ;
2j 1
using the binomial coefficient identities known as the hockey-stick theorem and
Pascals lemma (identities 135 and 127 of [3]).
6.2 n-Color Composition Enumerations

In this section, we use spotted tilings to give new proofs of enumeration results for
C C.n/ and two special classes, C C.n; k/, the number of n-color compositions with
k parts, and C C e.n/, the number of n-color compositions with each part even. We
also prove a new enumeration result for C C o.n/, the number of n-color compositions
with each part odd.
Theorem 6.3 ([1, Theorem 1b, d]). The number of n-color compositions of n satisfies
the formula C C.n/ D F2n and the recurrence relation C C.n/ D 3 C C.n 1/
q
C C.n 2/. The corresponding generating function is 13qCq 2.
Proof. We establish C C.n/ D C o.2n/ by a bijection between n-color compositions

of n and (normal) compositions of 2n with all parts odd.
Consider a spotted tiling for an n-color composition of n. Recall that the bars and
spots alternate. Consider the spot signifying part ki to be positioned at i 12 within
the 1 k rectangle. Then the distance between a spot and an adjacent bar has the
form m C 12 for some integer m 0. Create a (normal) tiling of 2n by replacing
spots with bars and doubling all distances, which are now of the form 2m C 1, odd
numbers. Figure 6.3 shows the transition from spotted tilings counted by C C.3/ to
tilings counted by C o.6/.
The process is invertible because the bars in an all-odd composition of 2n alternate
even/odd, considering the leftmost bar to be at 0. Because 2n is even, there are at least
82 Brian Hopkins
u
u
u
u u
u u
u u
u u
u u u
Figure 6.3. Row by row correspondence between C C.3/ and C o.6/.
bars at 0, some odd position, and 2n. To create the corresponding spotted tiling, halve
all distances and replace the bars at any half-integer positions with spots.
By Theorem 6.1, C C.n/ D F2n . The recurrence relation for F2n is known, but
we demonstrate the recurrence in terms of spotted tilings to introduce ideas used in
subsequent proofs. We establish a bijection between the n-color compositions counted
by 3 C C.n 1/ and by C C.n/ C C C.n 2/.
Given three copies of the spotted tilings counted by C C.n 1/, perform the fol-
lowing operations:
(a) to the first set, add 11 at the end of each composition;

(b) to the second set, for each composition, replace its last part kj with .k C 1/j ;
(c) To the third set, for each composition,
(c1) if the last part is k1 , replace it with .k C 1/kC1 ;
(c2) if the last part is kj with j > 1, replace it with .k 1/j 1 .
In terms of spotted tilings, operation (a) adds a box at the right hand side. Operations
(b) and (c1) both extend the final rectangle by length one; (b) keeps the spots fixed,
while (c1) moves the last spot from the first position of a 1 k rectangle to the last
position of the new 1 .k C 1/ rectangle. The results of (a), (b), and (c1) are spotted
tilings counted by C C.n/ with no repetition notice that (b) produces tilings where
the spot is not in the rightmost position, while tilings produced by (a) and (c1) all have
a spot in the rightmost position. Operation (c2) moves the spot of the final rectangle
one position to the left and decreases the rectangle by length one, producing a tiling
counted by C C.n 2/. Both (c1) and (c2) can be considered to move the last spot one
position to the left, allowing wrap around from the leftmost to rightmost position.
We demonstrate this with 3 C C.3/ D C C.4/ C C C.2/. Since operations (a) and
(b) are less complicated, Figure 6.4 shows them only applied to 32 , while all 8 spotted
tilings counted by C C.3/ are shown under (c1) and (c2).
For the reverse direction, there are three cases for n-color compositions counted
by C C.n/. If the last part is 11 , remove it; this is the inverse of operation (a). If the
last part is kj for k 2 and j < k, replace it with .k 1/j ; this is the inverse of
operation (b). If the last part is kk for k 2, replace it with .k 1/1 ; this is the
(a) u u u
(b) u u
(c1) u u
(c2) u u
(c2) u u
(c1) u u u u
(c1) u u u u
(c1) u u u u
(c2) u u u u
(c1) u u u u u u
Figure 6.4. For three copies of C C.3/, examples of operations (a) and (b), and row by row
details of (c1) and (c2).
inverse of operation (c1). For an n-color composition counted by C C.n 2/ with last
part kj , replace it with .k C 1/j C1 ; this is the inverse operation of (c2).
We have established C C.n/ D 3 C C.n 1/ C C.n 2/. The generating function
follows from the recurrence (denominator) and initial values of the sequence (numer-
ator); see [10] for more details.
Theorem 6.4 ([1, Theorem 1a, c]). The number of n-color compositions of n with

m parts satisfies the formula C C.n; m/ D nCm1
2m1 . The corresponding generating
qm
function is .1q/2m
.
Proof. Consider the spotted tiling associated to an n-color composition of n with m

parts. The spotted tiling will include m spots and m C 1 bars, more specifically bars at
0, 2n, and m 1 other positions. As in the previous proof, consider the spot signifying
part ki to be positioned at i 12 within the 1k rectangle. Doubling all distances places
spots at odd positions and bars at even positions. Since spots and bars alternate, the
positions of the m spots and m 1 internal bars create a subsequence x1 ; : : : ; x2m1
of 1; : : : ; 2n 1 where the parity of xi matches the parity of i . From Theorem 6.2,

we know there are nCm1 2m1 such subsequences.
For generating functions associated with binomial coefficients, see [10].
P
Since C C.n/ D m C C.n; m/, combining Theorems 6.3 and 6.4 gives
!
X
n
nCm1
D F2n
2m 1
mD1
which is equivalent to the even index case of an identity connecting Fibonacci numbers
to diagonals in Pascals triangle (identity 4 in [3]).
84 Brian Hopkins
Theorem 6.5 ([5, Theorems 1.2, 5.1]). The number of n-color compositions of n
having only even parts satisfies the formula C C e.n/ D 4 C C e.n 2/ C C e.n 4/
with initial values C C e.0/ D 0 and C C e.2/ D 2. The corresponding generating
2q 2
function is 14q 2 Cq 4
.
Proof. Note that C C e.n/ D 0 for all odd n. We establish a bijection between the
n-color compositions counted by C C e.n/ C C C e.n 4/ and by 4 C C e.n 2/.
Similar to the recurrence relation proof of Theorem 6.3, we perform the following
operations on four copies of the spotted tilings counted by C C.n 2/:
(a) to the first set, add 21 at the end;

(b) to the second set, add 22 at the end;
(c) for the third set, replace the final part kj with .k C 2/j ;
(d) For the fourth set,
(d1) if the last part is k1 , replace it with .k C 2/kC1 ;
(d2) if the last part is k2 , replace it with .k C 2/kC2 ;
(d3) if the last part is kj with j > 2, replace it with .k 2/j 2 .
Notice that the three (d) operations each move the spot in the last rectangle two to
the left, with the convention that the rightmost position is to the left of the leftmost
position. Adding parts 2j and changing between part lengths k, k C 2, and k 2
maintains the parity restriction. Figure 6.5 shows the results on four copies of the
spotted tilings counted by C C e.4/.
(a) u u u
(b) u u u
(c) u u
(d1) u u
(d2) u u
(d3) u u
(d3) u u
(d1) u u u u
(d2) u u u u
(d1) u u u u
(d2) u u u u
Figure 6.5. For three copies of C C e.4/, examples of operations (a), (b), and (c), and row by
row details of (d1), (d2), and (d3).
As in the proof of Theorem 6.3, operations (a), (b), (c), (d1), and (d2) produce n-
color compositions counted by C C e.n/ with no repetition, by inspection of the last
part. Operation (d3) produces n-color compositions counted by C C e.n 4/.
For the inverse map, there are again several cases for an n-color composition counted
by C C e.n/. If the last part is 21 or 22 , remove it. If the last part is kj with j k 2,
replace it with .k 2/j . If the last part is kk1 or kk , replace it with .k 2/1 or .k 2/2 ,
respectively. These are inverse operations for (a), (b), (c), (d1), and (d2), respectively.
Given an n-color composition counted by C C e.n 4/, replace the last part kj with
.k C 2/j C2 , the inverse of operation (d3).
We have established C C e.n/ D 4C C e.n2/C C e.n4/. The initial conditions
are easily checked. The sequence begins 0; 0; 2; 0; 8; 0; 30; 0; 112; 0; 418. Removing
the zeros, this is double the sequence A001353 in [8]. The generating function follows
from the recurrence relation and the initial values.
Theorem 6.6. The number of n-color compositions of n having only odd parts sat-
isfies the recurrence C C o.n/ D C C o.n 1/ C 2 C C o.n 2/ C C C o.n 3/
C C o.n 4/ with initial values 0; 1; 1; 4. The corresponding generating function is
qCq 3
1q2q 2 q 3 Cq 4
.
Proof. We establish a bijection between the n-color compositions counted by

C C o.n/ C C C o.n 4/ and by C C o.n 1/ C 2 C C o.n 2/ C C C o.n 3/.
Similar to the previous proofs, we perform the following operations on spotted
tilings counted by C C o.n 1/ C 2 C C o.n 2/ C C C o.n 3/:
(a) given an n-color composition counted by C C o.n 1/, add 11 at the end;
(b) for the first set of n-color composition counted by C C o.n 2/, replace the final
part kj with .k C 2/j ;
(c) for the second set of n-color composition counted by C C o.n 2/;
(c3)] if the last part is kj with j 3, replace it with .k 2/j 2 ;
(d) given an n-color composition counted by C C o.n 3/, add 33 at the end.
As in the proof of Theorem 6.5, the three (c) operations move the spot in the last
rectangle two to the left, with the wrap around convention. Adding parts 11 and 33 and
changing between part lengths k, k C 2, and k 2 maintains the parity requirement.
Figure 6.6 shows the results on C C o.5/ C 2 C C o.4/ C C C o.3/.
Notice that neither operations (c1) nor (c2) produces an n-color composition of
n C 2 with last part 33 . As in the previous proofs, operations (a), (b), (c1), (c2), and (d)
produce n-color compositions counted by C C o.n/ with no repetition, by inspection of
the last part. Operation (c3) produces n-color compositions counted by C C o.n 4/.
86 Brian Hopkins
(a) u u u
(b) u u u u
(c1) u u u u
(c1) u u u u
(c1) u u u u
(c1) u u u u
(c2) u u u u
(c3) u u u u
(c1) u u u u u u u u
(d) u u u
Figure 6.6. For n D 6, examples of operations (a), (b), and (d), and row by row details of (c1),
(c2), and (c3).
For the inverse map, consider an n-color composition counted by C C o.n/. If the
last part is 11 or 33 , remove it. If the past part is kj with j k 2, replace it with
.k 2/j . If the last part is kk1 or kk , replace it with .k 2/1 or .k 2/2 , respectively.
There are inverse operations for (a), (d), (b), (c1), and (c2), respectively. Given an n-
color composition counted by C C o.n 4/, replace the last part kj with .k C 2/j C2 ,
the inverse operation of (c3).
We have established C C o.n/ D C C o.n 1/ C 2 C C o.n 2/ C C C o.n 3/
C C o.n 4/. The initial conditions are easily checked. The sequence begins 0, 1, 1, 4,
7, 15, 32, 65, 137, 284, 591, which is A119749 in [8]. The generating function follows
from the recurrence relation and the initial values.
6.3 Conjugable n-Color Compositions

In his foundational work Combinatory Analysis, MacMahon introduced the zig-zag
graph of a composition as a way of defining conjugacy. A composition is broken up
with one rectangular part per row, such that the first segment of a part is directly be-
neath the last segment of its predecessor; see Figure 6.7 for an example. MacMahon
explains, Whereas the composition is read from left to right in successive rows from
top to bottom, the conjugate is read from top to bottom in successive columns from
left to right. [7, Sect. IV, Chap. 1, 129]
1 3 1 1
2
1
3
Figure 6.7. Mahonian zig-zag graphs showing that the conjugate of (2,1,3) is (1,3,1,1).
In the paper where Agarwal introduced n-color compositions [1], he concluded with
two questions, including What will be the shape of MacMahons zig-zag graph in the
case of n-colour compositions? The notion of spotted tilings introduced here allows
a possible answer to this question simply carry the spots into the zig-zag graph. But
for which n-color compositions does conjugation of the spotted zig-zag graph lead to
a valid spotted tiling?
Consider the composition in Figure 6.7. Can .2; 1; 3/ be colored so that it is con-
jugable? In order for the first column to include a spot, the 2 must become 21 . The
second column automatically has a spot since 1 can only become 11 . To avoid another
spot in the second column, the 3 must become 32 or 33 , but either possibility leaves
some column without a spot. So no coloring of .2; 1; 3/ gives a conjugable n-color
composition. In contrast, Figure 6.8 shows that .21 ; 11 ; 32 ; 11 / is conjugable.
11 32 11 22
21 u u
11 u u
32 u u
11 u u
Figure 6.8. Spotted zig-zag graphs showing that the conjugate of .21 ; 11 ; 32 ; 11 / is .11 , 32 ,
11 , 22 /.
Definition. An n-color composition is conjugable if its zig-zag graph has exactly one
spot per column.
Our final theorem characterizes and counts conjugable n-color compositions.
Theorem 6.7.
(a) Conjugable n-color compositions have the form
a
21 ; 11 ; 2b2 ; .32 ; 2a1 ; 11 ; 2b2 /c
where exponents denote repetition, a; b; c 0, and a; b may vary in each occur-
rence.
(b) The number of conjugable n-color compositions of n is 0 if n is even, 2.n1/=2 if
n is odd.
Proof. The analysis of conjugable n-color compositions is most easily done in terms
of the zig-zag graphs. By construction, there is one spot per row. By definition, there
is one spot per column. Since a subsequent part is positioned with its first segment
under the last segment of its predecessor, its spot must appear to the right of the spot
above it. In order to have one spot per column, it is exactly one position to the right.
The spots, therefore, lie on the diagonal.
The parts 31 , 33 , and kj for any k 4 cannot occur in a conjugable n-color compo-
sitions, since each would leave at least one column without a spot. As spots must ap-
pear on the diagonal, the allowed parts 11 ; 21 ; 22 ; 32 may occur within the constraints
shown in Figure 6.9. The description in part (a) follows.
88 Brian Hopkins
u u u u
u u u u
Figure 6.9. Parts 11 and 22 can only be followed by some allowed k2 while parts 21 and 32
can only be followed by some allowed k1 .
The spotted zig-zag graph of a conjugable n-color composition of n with m parts

fits in an m m square with m spots on the main diagonal and exactly m 1 more
segments in the two adjacent diagonals to connect the spots. Thus n D 2m 1 and
there are no conjugable n-color compositions of even n.
To count the conjugable n-color compositions of odd n D 2m 1, consider the
original representation as a tiling of a 1 n board. Spots on the main diagonal of the
zig-zag graph here means that the first, third, : : : ; .2m 1/-st positions contain spots,
leaving m 1 gaps where there are two choices for the position of a vertical bar; see
Figure 6.10. Thus there are 2m1 D 2.n1/=2 such compositions.
z z z
u u u u u u u u u u u u
Figure 6.10. Between adjacent spots, there are two choices for the position of a vertical bar,
so there are 4 conjugable n-color compositions of 5.
In conclusion, spotted tilings are a combinatorial interpretation of n-color composi-

tions in the spirit of MacMahon that allow for bijective proofs of various enumeration
results and an analog of the zig-zag graph and conjugacy. The interested reader will
want to compare the spotted tilings introduced here to the combinatorial interpretation
of n-color compositions using lattice paths given in [2].
Acknowledgments. Thanks to the organizers for putting together this fifth Integers
Conference and to the anonymous referee for helpful comments.
References
[1] A.K. Agarwal, n-colour compositions, Indian J. Pure Appl. Math. 31(11) (2000) 1421
1437.
[2] A.K. Agarwal and G. Narang, Lattice paths and n-colour compositions, Discrete Math.
308(9) (2008) 17321740.
[3] A. Benjamin and J. Quinn, Proofs that Really Count: The Art of Combinatorial Proof,
Mathematical Association of America, 2003.
[4] C. Church Jr. and H. Gould, Lattice point solution of the generalized problem of Terquem
and an extension of Fibonacci numbers, Fibonacci Quart. 5(1) (1967) 5968.
[5] G. Yu-Hong, n-colour even self-inverse compositions, Proc. Indian Acad. Sci. (Math.
Sci.) 120(1) (2010) 2733.
[6] V. Hoggatt Jr. and D. Lind, Fibonacci and binomial properties of weighted compositions,
J. Combin. Theory 4(2) (1968) 121124.
[7] P. A. MacMahon, Combinatory Analysis, volume 1, Cambridge University Press, 1915.
[8] The On-Line Encyclopedia of Integer Sequences, published electronically at http://oeis.
org, 2012, last accessed 12/12/2012.
[9] O. Terquem, Sur un symbole combinatoire dEuler et son utilit dans LAnalyse, J. Math.
Pures Appl. 4 (1839) 177184.
[10] H. Wilf, generatingfunctionology, Academic Press, 1990.
Author information
Brian Hopkins, Department of Mathematics, Saint Peters University, Jersey City, New Jer-
sey, USA.
Email: bhopkins@saintpeters.edu
Chapter 7
A Class of Wythoff-Like Games

Aviezri S. Fraenkel and Yuval Tanny
Abstract. We present a class of two-player Wythoff game variations we dub Wyt(f ) that
depends on a given function f .k/. In this class a move consists of removing either a positive
number of tokens from precisely one of two given piles, or k tokens from one pile and `
from the other, subject to the constraint 0 < k ` < f .k/. We analyze three classes of
integer-valued functions f .k/: constant, superadditive, and polynomial of degree > 1 with
nonnegative integer coefficients. The nature of the winning positions in the games is essentially
unique for each class.
Keywords. Combinatorial Games, Wythoff Games.
7.1 Introduction
We propose and analyze a family of two-player combinatorial take-away games played
on two piles, dubbed Wyt(f ). The two players move alternately by selecting one of
the following moves:
move of the first type: take any positive number of tokens from a single pile, possibly
the entire pile;
move of the second type: take k > 0 tokens from one pile and ` > 0 tokens from
the other. This move is restricted by the condition
0 < k ` < f .k/; (7.1)
where f .k/ W Z0 ! Z0 is a function that distinguishes the games from each
other.
A position in the game can be represented by a pair .a; b/ (a; b 2 Z0 ) which de-
notes the number of tokens in each pile. Without loss of generality we assume through-
out 0 a b. The games we consider here are played under the normal play conven-
tion, that is, the player making the last move wins the game and the opponent loses.
In the misre play convention the player making the last move loses and the opponent
wins.
Note that these are impartial games the set of possible moves for a player de-
pends only on the game position, not on the player. They are deterministic (no chance
moves) and acyclic (the number of tokens decreases at every move until it becomes 0).
Therefore we can partition the positions of the game into Next player winning posi-
tions (denoted N -positions, or by the set N WD N .f / Z20 ) and Previous player
winning positions (denoted P -positions, or by the set P WD P .f / Z20 ).
A game G is tractable if
for every position, its state (P - or N -position) can be decided in polynomial time;
the next move from any N -position to a P -position can be computed in polynomial
time;
the winner can consummate a win in at most an exponential number of moves.
To the run of the mill algorithmicians the last item dooms the game as intractable. It
may be quite a surprise to them that this is not the case: whereas we dislike computing
in exponential time, the human race relishes observing some of its members being
tormented for an exponential length of time! In fact, the easy game Nim and similar
take-away games do require that amount of time. For games there are also notions of
polynomiality and efficiency; see [12], especially Section 4.
In the next sections of this chapter, we define sequences An and Bn to analyze the
P -positions. We use the notation
A D [1 1
iD0 Ai ; B D [iD0 Bi :
In addition, when analyzing the P -positions we frequently use the mex function: Let
S be a finite set of nonnegative integers. Then mex.S / is defined to be the least non-
negative integer not in S . Note that the mex of the empty set is 0.
This chapter presents characterizations of the P -positions and tractable winning
strategies for the following types of functions:
f .k/ D t; t > 0 an integer;
f .k/ is a strictly increasing superadditive function;
P
f .k/ D niD0 ai k i is a polynomial of degree n > 1 with integer coefficients ai 0
and a0 > 0.
Note that a function f W Z0 ! Z0 is called superadditive if it satisfies: f .k/ k
and f .k C `/ f .k/ C f .`/ for all k; ` 2 Z0 .
Wyt(f ) is a general class of games that encapsulates a number of previously an-
alyzed games and new games. In its simplest case, where f .k/ 2 0; 1, the move
of the second type cannot be done and the game reduces to Nim on 2 piles (in nor-
mal play). For the case f .k/ D k C 1 the game reduces to the classical Wythoff
game [24], where in a move of the second type the player has to take the same pos-
itive number of tokens from both piles. An analysis of the constant difference class
f .k/ D k C t , t 1 integer and later for the linear class f .k/ D sk C t s; t 1
Chapter 7 A Class of Wythoff-Like Games 93
integers was done by Fraenkel [8, 11]. For other examples of variations of 2-piles
Wythoff, see [3, 5, 10, 13, 15, 18, 20].
In Section 7.2 we deal with the case f .k/ D t , and in Section 7.3 with superaddi-
tivity. In Section 7.4 we handle polynomials, where we resort to real analysis for the
proofs. Further possible work is indicated in the final Section 7.5.
7.2 Constant Function

Considering Wyt(f ) for f .k/ D t; t > 0 an integer, using the move of the second
type, a player can take k tokens from one pile and ` tokens from the other as long as
0 < k ` < t . As implied by the penultimate paragraph of the previous section, we
may assume t 2. The class of games with the restriction k C ` < t for moves of the
second type is called Cyclic Nimhoff and was settled by Fraenkel and Lorberbom [14]
for a general n-piles game; see also [5]. Duchne and Gravier [7] examined (among
other geometrical extensions of Wythoff) a bounded Wythoff game in which it is only
possible to take k < t tokens from one pile or k < t tokens from both piles.
Given fixed t 2 Z2 , we define g W Z0 ! Z0 such that for all m 2 Z0 ,
g.m/ D t m .t 2 1/bm=.t C 1/c

and
An D mexAi ; Bi W 0 i < n; Bn D g.An /; n 0:
In this section we show that [1
iD0 .Ai ; Bi / are the P -positions of Wyt(f ) when
f .k/ D t for any constant integer t 2. To prove this, we begin by showing the
relation between the An and Bn sequences using a specific numeration system. This
numeration system will also help us to derive a tractable strategy for winning the game.
The first few P -positions for the games where t D 3 and t D 10 are shown in Ta-
ble 7.1 and Table 7.2, respectively. After Theorem 7.13 (Section 7.2.2) we point out
in Remark 7.14 that Table 7.1 exhibits certain periodicities which help in exhibiting
the structure of the sequences A and B.
We begin with an auxiliary result.
Lemma 7.1.
(i) Let t 2. For every m 2 Z0 , g.m/ m, with equality if and only if m
0 .mod .t C 1//.
(ii) The function g is an injection.
Proof. (i) Write m D s.t C 1/ C r, 0 r < t C 1, s; r 2 Z0 . Then

g.m/ m D .t 1/.m .t C 1/b.s.t C 1/ C r/=.t C 1/c/:
Thus .g.m/ m/=.t 1/ D m s.t C 1/ 0 with equality if and only r D 0.
(ii) An elementary algebra exercise.
Table 7.1. t D 3:
n 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
An 0 1 2 4 5 8 9 10 12 13 16 17 18 20 21 24 25
Bn 0 3 6 4 7 8 11 14 12 15 16 19 22 20 23 24 27
Table 7.2. t D 10:
n 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
An 0 1 2 3 4 5 6 7 8 9 11 12 13 14 15 16
Bn 0 10 20 30 40 50 60 70 80 90 11 21 31 41 51 61
Lemma 7.1 immediately implies:
Corollary 7.2. For every n 0, Bn D An if and only if An 0 .mod .t C 1//.
7.2.1 A Numeration System

Lemma 7.3. Let t 2 Z2 . Every nonnegative integer m can be represented uniquely
in the form
m D m2 .t 2 1/ C m1 t C m0 ; (7.2)
where the digits m0 ; m1 ; m2 are integers satisfying:
0 m 0 ; m1 < t I m2 0I m0 D m1 D t 1 is not permitted: (7.3)
Proof. Theorem 1 of [9] states: let 1 D u0 < u1 < u2 < be any finite or infinite
sequence of integers. Every nonnegative integer m has precisely one representation
in the system S D u0 ; u1 ; u2 ; : : : of the form m D niD0 mi ui , where the mi are
nonnegative integers satisfying
mi ui C mi1 ui1 C C m0 u0 < uiC1 for i 0:
By this theorem, every nonnegative integer m can be represented uniquely in the
form (7.2) where m0 < t and m0 C m1 t < t 2 1. It follows immediately that m1 < t
and that m0 D m1 D t 1 is not permitted.
Throughout this section, we use the convention that for every nonnegative number m
we let m0 ; m1 ; m2 be the appropriate digits of m in the above numeration system.
Corollary 7.4. With hypotheses as in Lemma 7.3, g.m/ D m if and only if m0 D m1 .
Proof. m D m2 .t 2 1/ C m1 .t C 1/ C m0 m1 0 .mod .t C 1// if and only if

m0 D m1 , since 0 m0 ; m1 < t . The proof is complete by Lemma 7.1.
Using our numeration system we can present the g function in a more intuitive way.
Lemma 7.5. Let t 2 Z2 , and m 2 Z0 which is represented uniquely in the form
(7.2) with digits m0 ; m1 ; m2 . If m1 m0 , then
g.m/ D m2 .t 2 1/ C m0 t C m1 :
Otherwise,
g.m/ D .m2 C 1/.t 2 1/ C m0 t C m1 :
Proof. We have,
g.m/ D g.m2 .t 2 1/ C m1 t C m0 /
D m2 .t 2 1/t C m1 t 2 C m0 t .t 2 1/
.b.m1 t C m0 /=.t C 1/c C m2 .t 1//
D m2 .t 2 1/ C m1 t 2 C m0 t .t 2 1/b.m1 t C m0 /=.t C 1/c:
If m1 m0 , then b.m1 t C m0 /=.t C 1/c D b.m1 .t C 1/ C .m0 m1 //=.t C 1/c D

m1 C b.m0 m1 /=.t C 1/c D m1 since 0 m0 ; m1 < t . Therefore, the g function
for this case becomes:
g.m/ D m2 .t 2 1/ C m1 t 2 C m0 t m1 .t 2 1/
D m2 .t 2 1/ C m0 t C m1 :
Otherwise, b.m1 t C m0 /=.t C 1/c D m1 1 and the g function becomes:
g.m/ D m2 .t 2 1/ C m1 t 2 C m0 t .m1 1/.t 2 1/

D .m2 C 1/.t 2 1/ C m0 t C m1 :
Remark 7.6. Notice that in the two displayed formulas of Lemma 7.5 the digits m0 ,
m1 of the unique representation (7.2) of m have been switched. Furthermore, m has a
unique representation of the form (7.2) if and only if m0 WD m2 .t 2 1/ C m0 t C m1
has a unique representation. Their value difference is m m0 D .m1 m0 /.t 1/.
We presented the relation between An and Bn using the special numeration system.
In the next couple of lemmas we show that an independent simple characterization of
the Ai and Bi sets follows from the use of this numeration system. This charac-
terization will enable us to prove that the An and Bn pairs constitute the P -positions
of the game.
Lemma 7.7. Let t 2 Z2 , and m 2 Z0 which is represented uniquely in the form
(7.2) with digits m0 ; m1 ; m2 . Then m 2 A if and only if m1 m0 .
Proof. Induction on m. Assume that the statement is true for all n in K D

0; 1; : : : ; m 1. Suppose first that m1 m0 . Using the induction hypothesis, we
show that then m Bn D g.n/ W n 2 K \ A. It then follows from the mex defini-
tion that m 2 A.
Let n 2 K \ A. Then n1 n0 by the induction hypothesis. If n1 D n0 , then n D
g.n/ < m by Corollary 7.4. So we may assume n1 < n0 . Let r D r2 .t 2 1/Cr1 t Cr0
be the unique representation of g.n/. Lemma 7.5 and Remark 7.6 then imply that
r1 D n0 and r0 D n1 . Thus r0 < r1 . Assuming m D r, the uniqueness of the
representation (Lemma 7.3) and Remark 7.6 then imply r1 D m1 and r0 D m0 ,
contradicting m1 < m0 . Therefore, m g.n/.
Secondly, suppose that m 2 A. Let m0 WD m2 .t 2 1/ C m0 t C m1 . Assume that
m1 > m0 . Then m m0 D .m1 m0 /.t 1/ > 0, so m0 2 K \ A by the induction
hypothesis. By Lemma 7.5, m D g.m0 / 2 B. In fact, if, say, m D An , then m0 D Ai
for some i < n, and m D g.m0 / D Bi . But this contradicts the definition of the mex
function. Hence m1 m0 .
Consequently we have:
Lemma 7.8. m D m2 .t 2 1/ C m1 t C m0 2 B if and only if m0 m1 :
Proof. If m1 > m0 then m A by Lemma 7.7 and therefore m 2 B by the mex prop-
erty. If m1 D m0 then m 2 A by Lemma 7.7 and m D g.m/ 2 B by Corollary 7.4.
Conversely, if m 2 B, then by definition there exists a 2 A such that m D g.a/. Let
a0 ; a1 ; a2 be the digits of a represented in form (7.2). Then by Lemma 7.7, a1 a0 ;
and by Lemma 7.5, m0 D a1 and m1 D a0 , so m0 m1 .
Corollary 7.9.
(i) If An D m2 .t 2 1/ C m1 t C m0 then Bn D m2 .t 2 1/ C m0 t C m1 .
(ii) Bn An D .m0 m1 /.t 1/ 0. In particular, Bn D An if and only if m0 D m1 .
Proof. (i) By Lemma 7.7, m D An if and only if m1 m0 ; and if m1 m0 , then

Lemma 7.5 implies that g.m/ D Bn D m2 .t 2 1/ C m0 t C m1 .
(ii) This also follows from m1 m0 .
Theorem 7.10. For f .k/ D t a constant, the set of P -positions of Wyt.f / is given
by W WD [1
iD0 .Ai ; Bi /.
Proof. Since Wyt(f ) is an acyclic game, it suffices to show two things:

(1) every move from any position in W lands in a position outside W ;
(2) for every position outside W , there is a move to some position in W .
(1) A move of the first type from .An ; Bn / 2 W leads to a position not in W ,
because there are no repeating terms in A, nor in B: (i) The mex property implies that
the sequence A is strictly increasing thus has no repeating terms. (ii) There are also no
repeating terms in B, since the function g is an injection (Corollary 7.1 (ii)) and there
are no repeating terms in A.
Suppose that a move of the second type from .An ; Bn / 2 W produces a position
.Am ; Bm / 2 W . Then m < n. This move can be made either in the form
(i) An ! Am ; Bn ! Bm I or in the form (ii) An ! Bm ; Bn ! Am :
(i) An ! Am , Bn ! Bm . Then

2 An Am
Bn Bm D t .An Am / .t 1/ :
t C1 t C1
If bAn =.t C1/c D bAm =.t C1/c then Bn Bm D t .An Am / t , contradicting
the move rule (7.1). Otherwise, bAn =.t C 1/c bAm =.t C 1/c 1. Since An
Am D k < t , we get
Bn Bm t .t 1/ .t 2 1/ D 1 t < 0;
again contradicting (7.1).
(ii) An ! Bm , Bn ! Am . Let r WD Am . Then r1 r0 by Lemma 7.7. We may

assume r0 > r1 , because r0 D r1 implies Bm D Am (Corollary 7.9(ii)), whence
case (ii) reverts back to case (i). The move (ii) can only be made if An > Bm .
Then Bn Am D g.An /Am An Am > Bm Am D .r0 r1 /.t 1/ t 1,
contradicting move rule (7.1).
(2) Let .x; y/ with x y be any position not in W . The construction of An by the
mex rule implies that the set [1 iD0 Ai ; Bi D Z0 . Therefore, x D Bn or x D An
for some n 0. We consider several cases.
(i) x D Bn . Then move y ! An using the first move rule.
(ii) x D An and y > Bn , then move y ! Bn using the first move rule.
(iii) x D An y < Bn . Recall that we can represent x and y uniquely in our
numeration system: x D x2 .t 2 1/ C x1 t C x0 and y D y2 .t 2 1/ C y1 t C y0 .
Let
x 0 WD x2 .t 2 1/ C x1 t C s; y 0 WD x2 .t 2 1/ C st C x1 ;
where
s WD y1 C b.y0 x1 /=t /c:
We show the following:
(a) (a1) If y0 x1 then s D y1 , x0 > y1 ; if y0 < x1 then s D y1 1, x0 y1 .
(a2) 0 x1 s < x0 < t .
(b) There exists m such that x 0 D Am and y 0 D Bm .
(c) The move x ! x 0 D Am , y ! y 0 D Bm is a legal move, that is, 0 < x x 0 < t ,

0 y y0 < t .
(a) Notice that s D y1 C b.y0 x1 /=t /c 2 y1 1; y1 , since 0 x1 ; y0 ; y1 < t .

By Corollary 7.9(i), An and Bn have the same coefficient (digit) multiplier of t 2 1,
so a fortiori y2 D x2 . Hence x D An y < Bn D g.An / implies:
x1 t C x0 y1 t C y0 < x0 t C x1 : (7.4)
The right-hand side of Equation (7.4) is equivalent to .x0 y1 /t > y0 x1 . If y0 x1 ,
then s D y1 and x0 > y1 , so s < x0 < t . If y0 < x1 , then s D y1 1, yet
x0 y1 , and we have again s < x0 < t . The left-hand side of Equation (7.4) is
equivalent to y0 x0 .x1 y1 /t . By Lemma 7.7, x1 x0 , since x D An . Hence
s y1 C b.y0 x0 /=t /c x1 0:
(b) By (a2), x1 s < t , so by Lemma 7.7, x 0 D Am for suitable m. Again by (a2),
0 s < t , so y 0 D Bm by Corollary 7.9(i).
(c) From (a), 0 < x x 0 D x0 s < t . Now y y 0 D .y1 s/t C y0 x1 , since
y2 D x2 . By (a1), if y0 x1 then s D y1 , so 0 y y 0 D y0 x1 < t ; and if
y0 < x1 then s D y1 1, so 0 < y y 0 D t .x1 y0 / < t .
7.2.2 Strategy Tractability and Structure of the P-Positions

Theorem 7.11. Given game position .x; y/ with 0 x y, there is a tractable
linear-time algorithm to decide whether or not .x; y/ 2 P .
Proof. We wish to compute whether or not .x; y/ 2 P in time linear in the input size
.log xy/. Expand x and y in our numeration system, which can be done linearly by
using the simple greedy algorithm from [9]:
x2 D bx=.t 2 1/c
x1 D b.x x2 .t 2 1//=t c
x0 D x x2 .t 2 1/ x1 t:
Then, check whether x0 x1 . If negative, .x; y/ 2 N . If positive, check whether

y2 D x2 and y1 D x0 and y0 D x1 . If positive, then .x; y/ 2 P (Corollary 7.9(i));
otherwise, .x; y/ 2 N .
Corollary 7.12. Let t 2 Z2 and f .k/ D t a constant function. Then there is a
tractable strategy for winning Wyt(f ).
Proof. Using the instructions in the second part of the proof of Theorem 7.10, the
greedy algorithm in Theorem 7.11, and Lemmas 7.7 and 7.8, it is clear how to construct
a tractable strategy for winning Wyt(f ) for any given N -position.
We now turn to the structure of the sequences A and B.

A sequence sn n0 with limn!1 sn D 1 is called arithmetically periodic if there

exist integers 1 < q p (the periods), and distinct nonnegative integers r1 ; : : : ; rq 2
0; p 1, such that si ri .mod p/ whenever n i .mod q/ .i D 1; : : : ; q/.
This is a variation of the definitions of arithmetic (or additive) periodicity given or
implied in [1,4,17]. It accommodates sequences that are not necessarily monotonically
increasing.
Theorem 7.13. Let t 2.

(i) Each of the sequences A and B is arithmetically periodic with periods q D
.t 1/.t C 2/=2 and p D t 2 1.
(ii) Each of the sequences A and B contains .t 1/.t C2/=2 distinct residues mod p,
t 1 of which are common to both sequences.
(iii) A [ B contains all the t 2 1 residues mod p, t 1 of which appear in both
sequences.
(iv) For all n 0, 0 Bn An .t 1/2 .
Proof. (i) Lemmas 7.7 and 7.8 and Corollary 7.9 imply that the sequences A and B
have residues that are periodic mod p. Indeed, membership of x in A and B depends
only on the relative size of x0 and x1 , not on the size of x2 .
(ii) Using (i) we can assume x2 D 0. By the mex property, all the residues of A in
0; p 1 are increasing, so they are distinct. Their number is precisely the number of
x0 , x1 satisfying 0 x1 x0 t 1. Since x0 D x1 D t 1 is forbidden, this
number is precisely .t 1/.t C 2/=2, t 1 of which are common to both sequences,
namely those for which 0 x0 D x1 < t 1. The proof for the residues of B is
similar.
(iii) Follows immediately from (ii).
(iv) Follows immediately from Corollary 7.9(ii).
Remark 7.14. Notice that the preceding propositions (i)-(iv) of Theorem 7.13 can be
seen in Table 7.1.
Almost all game positions .x; y/, 0 x y are N -positions see [22, 23]. There-
fore .x; y/ P with high probability. Theorem 7.13 can help to dispose of them
quickly by computing the residues of x and y mod p and searching for them in a
pre-constructed table of the p D t 2 1 residue pairs of .An ; Bn / mod p.
7.3 Superadditive Functions

In this section we examine Wyt(f ), when f is a strictly increasing superadditive func-
tion. We show that the P -positions of this game have a simple recursive formula. As
an example, the first few P -positions for the game for which f .x/ D x 2 are shown
in Table 7.3.
Table 7.3. f .x/ D x 2 .
n 0 1 2 3 4 5 6 7 8 9 10 11 12 13
An 0 1 2 3 5 6 7 8 10 11 12 13 14 15
Bn 0 1 4 9 25 36 49 64 100 121 144 169 196 225
Remark 7.15. The function f in this example is a polynomial, and polynomials are
considered in the next section. However, a polynomial of degree t 1 of the form
X
t
f .x/ D ai x i ; ai 2 Z0 for 0 i t; a t > 0; (7.5)
iD0
is clearly superadditive if and only if a0 D 0. In the next section we consider the case
a0 > 0.
Remark 7.16. Notice that if f W Z0 ! Z0 is strictly increasing, then f .k/ k.
Let
An D mexAi ; Bi W 0 i < n; Bn D f .An /; n 0: (7.6)
Theorem 7.17. Let f W Z0 ! Z0 be a strictly increasing superadditive function.

Then the set of P -positions of Wyt.f / is given by W WD [1
iD0 .Ai ; Bi /.
Proof. As in the proof of Theorem 7.10, it suffices to show:

(1) Every move from any position in W lands in a position outside W .
(2) For every position outside W , there is a move to some position in W:
(1) The mex definition implies that An is a strictly increasing sequence. Thus, for
every i > j we have Ai > Aj . Hence Bi D f .Ai / > f .Aj / D Bj since f is a
strictly increasing function. Therefore, there are no repeating terms in An nor in Bn .
It follows that a move of the first type from .An ; Bn / 2 W leads to a position W .
Suppose that a move of the second type from .An ; Bn / 2 W produces a position
.Am ; Bm / 2 W . Then m < n. Let k WD An Am and ` WD Bn Bm . Since An is a
strictly increasing sequence we have k > 0. The superadditivity of f then implies,
` D f .An / f .Am / f .An Am / D f .k/:
By Remark 7.16 we have ` k. Thus 0 < k ` f .k/, contradicting move

rule (7.1).
The move .An ; Bn / ! .Bm ; Am / where m < n is impossible too: Let k 0 WD An
Bm and `0 WD Bn Am . Since Bn D f .An / An , we have: 0 < k 0 k ` `0 .
As was shown above, ` f .k/, so `0 ` f .k/ f .k 0 / since f is a strictly

increasing function. Hence 0 k 0 `0 f .k 0 /, again contradicting move rule (7.1).
(2) Let .x; y/ with 0 x y be any position not in W . It follows from the mex
iD0 Ai ; Bi D Z0 . Therefore, x D Bn or x D An for some n 0.
definition that [1
Case (i). x D Bn . Then move to the position .An ; Bn / by subtracting y An tokens
from one pile, using a move of the first type.
Case (ii). x D An and y > Bn . Then subtract y Bn tokens from one pile, using a
move of the first type.
Case (iii). x D An y < f .An / D Bn . Then move to .0; 0/ using a move of the
second type.
The time complexity for deciding whether a given position is a P -position or N -

position depends on the time complexity of the function f . Assume f to be a polyno-
mial time computable function. The nave algorithm for deciding whether a position
.x; y/ is a winning position consists of calculating An and Bn using their recursive
definition (7.6) until it is known whether x 2 An or not. This method takes exponen-
tial time and space since the input position is given in succinct form. In the previous
section and in other Wythoff-like games (e.g., [11]) constructing a special numeration
system assisted in building a tractable strategy to win the game. We show here that
this is not necessary for this case.
Remark 7.18. For all superadditive functions f; f .0/ D 0. This follows from f .0/ D
f .0 C 0/ f .0/ C f .0/.
Lemma 7.19. Let f W Z0 ! Z0 be a strictly increasing superadditive function.

For every x 2 Z0 , if f .x/ D x then x D Ax D Bx .
Proof. Induction on x. Since A0 D 0 by the mex property, B0 D f .0/ D 0 (Re-

mark 7.18). Assume that the statement is true for all m < x, m 2 Z0 . If f .x/ D x
then since f is strictly increasing, for every m < x, f .m/ D m. The induction hy-
pothesis then implies A0 ; : : : Ax1 ; B0 ; : : : Bx1 D 0; 1; : : : ; x 1. Therefore,
x D Ax by the mex property. If f .x/ D x then also Bx D f .x/ D x.
Lemma 7.20. Let f W Z0 ! Z0 be a strictly increasing superadditive function. If

f is not the identity function f .x/ D x, then there exist N 2 Z1 and a > 1 such
that for every integer n N , f .n/=n a and for every integer n < N , f .n/ D n.
Proof. Since f is not the identity function, there exists a minimum integer N 2 Z1
such that f .N / N . Thus for every integer n < N , f .n/ D n. Notice that N 0
because f .0/ D 0 (Remark 7.18). By Remark 7.16, we have f .N / N C 1.
Let a WD 1 C 1=.2N 1/. Since N 1 it is clear that a > 1. We can write every
integer n N in the form n D kN C r, where 0 r < N and k 2 Z1 . The
superadditivity of f and f .N / N C 1 then imply that

f .n/ D f .kN C r/ f .kN / C f .r/ kf .N / C f .r/ k.N C 1/ C r:
Since r < N and k 1, we have
f .n/ k.N C 1/ C r k k 1
D1C 1C 1C a:
n kN C r kN C r kN C kr N Cr
Theorem 7.21. Let f W Z0 ! Z0 be a polynomial time computable strictly in-
creasing superadditive function. Let An , Bn be given by (7.6), and let .x; y/ with
0 x y be any game position. Then there is a polynomial-time algorithm to
decide whether or not x 2 A.
Proof. We apply the following recursive

Procedure 1. Input: x. Output: 1 if x 2 A; 0 otherwise.
Find z such that f .z/ is as close as possible to x, beginning with z D x=2 and then
proceeding in the form of a binary search on the interval 0; x (O.log x/ operations).
There are three possible outcomes:
(1) there is no z such that x D f .z/: then return 1 since x 2 A by the mex definition;
(2) there exists z such that x D f .z/ and z D x: then return 1 since x 2 A by
Lemma 7.19;
(3) There exists z such that x D f .z/ and z < x: then x 2 A if and only if z A. So
apply Procedure 1 with z as input and return 1 minus the returned value from this
new call of Procedure 1.
Suppose the recursive search algorithm stops with x t , for which f t .x t / D x, where
f t denotes f composed with itself t -times. We show that t D O.log x/. Conse-
quently, the algorithm is polynomial because each recursive call is polynomial in log x.
By Lemma 7.20 there exist a > 1 and minimal N 2 Z1 , such that for every integer
n N , f .n/=n a. In particular, f .N /=N a. Since f .N / N it follows that
f .f .N //=f .N / a. Hence,
f 2 .N / f 2 .N / f .N /
D a2 I
N f .N / N
and by induction we have
f t .N /
at :
N
Therefore f t .N / at N . We assume f .xn / xn . Otherwise, f .x/ D x D
f t .x t / D x t and then t D 0. By Lemma 7.20, x t N . Since f is an increas-
ing function, f .x t / f .N /, and by simple induction, f t .x t / f t .N /. Hence
x D f t .x t / f t .N / at N . Then loga .x/ t loga .aN / t , since a > 1 and
N 1. Consequently, t D O.log x/.
Corollary 7.22. Let f be a polynomial time computable strictly increasing superad-

ditive function. Then there is a tractable strategy for winning Wyt(f ).
Proof. Using the algorithm in Theorem 7.21 and the instructions in the second part
of the proof of Theorem 7.17, it is clear how to construct a tractable strategy for win-
ning Wyt(f ) for any given N -position. Note that by using the binary search described
in the beginning of Theorem 7.21, it is possible to calculate f 1 .x/, if it exists, in
polynomial time.
7.4 Polynomial
In this section we consider the game Wyt(f ) for which f W R0 ! R0 is a polyno-
mial of the form (7.5) with t 2 and a0 > 0.
Thus f is convex although not superadditive (Remark 7.15). The analysis of this
class of games is similar to that of the previous section, since the function which defines
the B sequence is superadditive, as we show presently.
We define here, as before, the sequences that compose the P -positions of this class
of games. For x 2 R0 , we define

maxf .x/; f .1/x if x 2 1; 1/
g.x/ D
f .1/x if x 2 0; 1/;
and
An D mexAi ; Bi W 0 i < n; Bn D g.An /; n 0: (7.7)
The function g is defined on R0 in order to enable us to use basic calculus for proving
that g is a superadditive function. Note that g.Z0 / Z0 .
As an example, the first few P -positions for the game for which f .x/ D x 2 C 9
are displayed in Table 7.4.
Table 7.4. f .x/ D x 2 C 9:
n 0 1 2 3 4 5 6 7 8 9 10 11 12 13
An 0 1 2 3 4 5 6 7 8 9 11 12 13 14
Bn 0 10 20 30 40 50 60 70 80 90 130 153 178 205
Lemma 7.23. The function g.x/ is a strictly increasing superadditive function.
Proof. If g.x/ is a nonnegative continuous convex function which vanishes at the

origin, then g.x/ is superadditive see [2, Theorem 5]. Since g.x/ is a nonnegative
continuous function and g.0/ D 0, it suffices to show that g.x/ is a convex function
0
in 0; 1/. By [16, Corollary 1.1.9], if the right derivative gC .x/ is nondecreasing in
0; 1/ then g.x/ is a convex function in this interval. We show here that gC 0 .x/ is
nondecreasing by showing: (i) gC 0 .x/ is nondecreasing in 0; 1/. (ii) g 0 .x/ is nonde-

C
0
creasing in .1; 1/. (iii) gC .1/ gC 0 .x/ for x 2 0; 1/.
0 .x/ D g 0 .x/ D f .1/. Hence, g 0 .x/ nondecreasing in 0; 1/.
(i) In 0; 1/, gC C
(ii) In 1; 1/, g.x/ is a convex function, since a maximum of convex functions is
0 .x/ is nondecreas-
convex [16, Theorem 1.1.3]. Therefore by [16, Corollary 1.1.6], gC
ing in 1; 1/.
(iii) For h > 0,
g.1 C h/ g.1/ f .1/.1 C h/ f .1/
D f .1/:
h h
0 0
Hence gC .1/ f .1/ D gC .x/ for x 2 0; 1/.
0 .x/ is nondecreasing in 0; 1/ and thus g.x/ is superadditive there.
Therefore, gC
Since g.x/ is continuous and gC 0 .x/ f .1/ > 0 for all x 2 0; 1/, g.x/ is strictly
increasing there.
Let S D n 2 Z0 W f .An / f .1/An :
Remark 7.24. Since t 2, f .An / > f .1/An for all sufficiently large n, so S is
finite, containing at most the first few nonnegative integers. The next lemma throws
some light on the set S .
Lemma 7.25. The set S is nonempty, and for every n 2 S we have An < f .1/ and
An D n.
Note 7.26. For the example f .x/ D x 2 C 9 we have An < f .1/ and f .An /
f .1/An for 0 n 9, but f .An / > f .1/An for all n > 9. Thus An D n for
0 n 9, but An > n for all n > 9 (see Table 7.4).
Proof. By the mex property, A0 D 0. Hence B0 D g.0/ D f .1/0 D 0. Again

by mex, A1 D 1. Thus f .A1 / D f .1/A1 , so S ;. Also, Bn D g.An / D
maxf .An /; f .1/An for all n 1. In particular, B1 D f .1/ 2.
Next we show An < f .1/ for all n 2 S . Suppose An f .1/ for some n 2 S . We
note that n 1, since 0 S : f .0/ > f .1/0. Let k WD An f .1/ 0. Since t 2
and a0 > 0 we have,
f .An / D f .f .1/ C k/ a t .f .1/ C k/t C a0 > .f .1/ C k/2 f .1/.f .1/ C k/
D f .1/An :
Thus, f .An / > f .1/An D Bn D g.An /, contradicting the definition of g. Therefore
An < f .1/ D B1 for all n 2 S .
We already verified that An D n for n 2 0; 1. Suppose that Am D m for all m < n
for which An < f .1/ D B1 . Then An D mex A0 ; : : : ; An1 ; B1 ; : : : ; Bn1 D
mex 0; : : : ; n 1 D n, since Bn1 > > B1 D f .1/ > An .
Theorem 7.27. Let f W Z0 ! Z0 be a polynomial of the form .7:5/ with t 2 and
a0 > 0. Then the set of P -positions of Wyt.f / is given by W WD [1
iD0 .Ai ; Bi /,
where Ai , Bi are defined in .7:7/.
Proof. As in the proof of Theorems 7.10 and 7.17, it suffices to show that
(1) every move from any position in W lands in a position outside W ;
(2) for every position outside W , there is a move to some position in W:
(1) The proof is the same as in part 1 of Theorem 7.17, using g instead of f and
using g.x/ f .x/.
(2) The proof for this statement is as in part 2 of Theorem 7.17 up to and including
Case (iii) x D An y < f .An / D Bn . There are two additional cases:
Case (iv) x D An , f .An / y < Bn and there exists m for which y D Bm < Bn .
Then move to the position .Am ; Bm / by subtracting An Am tokens from the first
pile, using a move of the first type.
Case (v) x D An , f .An / y < Bn and y D Bm for no m. Since Bn D g.An / >
f .An / it follows from the definition of g and Remark 7.24 that:
Bm D g.Am / D f .1/Am for all 0 m n: (7.8)
Further, by Lemma 7.25,
Am D m < f .1/ for all 0 m n: (7.9)
Write y D f .1/s C r; 0 r < f .1/; s; r 2 Z0 ; and u WD An As : We claim

that the move .An ; y/ ! .As ; Bs / is a valid move of the second type. For proving this
claim, it suffices to show:
(a) 0 s < n.
(b) y > Bs .
(c) Either 0 < u r < f .u/ or 0 < r u < f .r/. Thus, condition (7.1) of the
move of the second type is satisfied by subtracting u tokens from the first pile and
r tokens from the second pile.
(a) We have y D f .1/s C r < Bn D f .1/n by (7.9). Thus s < n, since r < f .1/.
Clearly s 0.
(b) By (7.8) and (a), r D y f .1/s D y Bs . Therefore r > 0 since y Bs .
(c) By (7.9) and (a), u D An As D n s > 0 and n s n < f .1/. Thus if
u r, then 0 < u r < f .1/ f .u/. Otherwise, if u > r, then 0 < r < u <
f .1/ f .r/.
Corollary 7.28. Let f be a polynomial of the form .7:5/ with t 2 and a0 > 0. Then
there is a tractable strategy for winning Wyt.f /.
Proof. By Lemma 7.23, g is a strictly increasing superadditive function, and clearly

polynomial-time computable. Using the algorithm in Theorem 7.21 and the instruc-
tions how to move from any N -position to a P -position in the second part of the proof
of Theorem 7.27, it is clear how to construct a tractable strategy for winning Wyt.f /
for any given N -position.
7.5 Further Work

In this chapter we studied the P -positions and winning strategies of the game Wyt.f /
for three classes of functions: constant, superadditive, and polynomial with nonnega-
tive coefficients. There are two main direction in extending this work. One is to exam-
ine other classes of functions and possibly to generalize the results presented here to
other classes of functions. For example, Wyt.f / where f .k/ D sk C t , for s; t 2 R>0
(the case s; t 2 Z>0 was studied in [11]), or for an arbitrary polynomial function f .
Another direction is to study deeper properties in each class. One of the important
properties of impartial games is the SpragueGrundy function, which enables playing
sums of games. Some results on the SpragueGrundy function for Wythoffs game
can be found in [4, 17, 21]. An additional direction is to study the sets of restrictions
and extensions of the game, in the sense of a subset or superset of the set of possible
moves, that preserve its P -positions. These directions were studied for Wythoff and
other Wythoff-like games in [6, 14, 19].
References
[1] E. R. Berlekamp, J. H. Conway, and R. K. Guy, Winning Ways for your Mathematical
Plays, vols. 14, A .K. Peters, Wellesley, MA, 20012004; 2nd edition: vol. 1 (2001), vols.
2, 3 (2003), vol. 4 (2004); translation by G. Seiffert of 1st edition (1982) into German:
Gewinnen, Strategien fr Mathematische Spiele, Foreword by K. Jacobs, M. Remnyi and
Seiffert, four volumes, Friedr. Vieweg & Sohn, Braunschweig, 1985.
[2] A. M. Bruckner and E. Ostrow, Some function classes related to the class of convex func-
tions, Pacific Journal of Mathematics 12(4) (1962), 12031215.
[3] I. G. Connell, A generalization of Wythoffs game, Canad. Math. Bull 2 (1959), 181190.
[4] A. Dress, A. Flammenkamp, and N. Pink, Additive periodicity of the sprague-grundy
function of certain nim games, Adv. in Appl. Math. 22 (1999), 249270.
[5] E. Duchne, A. Fraenkel, S. Gravier, R. J. Nowakowski, et al., Another bridge between
Nim and Wythoff, Australasian Journal of Combinatorics 44 (2009), 4356.
[6] E. Duchne, A. S. Fraenkel, R. J. Nowakowski, and M. Rigo.,Extensions and restrictions
of Wythoffs game preserving its P positions, Journal of Combinatorial Theory, Series
A 117(5) (2010), 545567.
[7] E. Duchne and S. Gravier, Geometrical extensions of Wythoffs game, Discrete Mathe-
matics 309(11) (2009), 35953608.
[8] A. S. Fraenkel, How to beat your Wythoff games opponent on three fronts, Amer. Math.
Monthly 89 (1982), 353361.
[9] A. S. Fraenkel, Systems of numeration, Amer. Math. Monthly 92 (1985), 105114.
[10] A. S. Fraenkel, Wythoff games, continued fractions, cedar trees and Fibonacci searches,
Theoretical Computer Science 29(12) (1984), 4973.
[11] A. S. Fraenkel, Heap games, numeration systems and sequences, Annals of Combina-
torics 2(3) (1998), 197210.
[12] A. S. Fraenkel, Complexity, appeal and challenges of combinatorial games, Expanded
version of a keynote address at Dagstuhl Seminar Feb. 1722, 2002, Theoretical Com-
puter Science 313 (2004), 393-415, special issue on Algorithmic Combinatorial Game
Theory.
[13] A. S. Fraenkel, New games related to old and new sequences, Integers, Electr. J of Com-
binat. Number Theory 4 (2004), #G6, 18 pp., Comb. Games Sect.
[14] A. S. Fraenkel and M. Lorberbom, Nimhoff games. Journal of Combinatorial Theory,
Series A 58(1) (1991), 125.
[15] V. E. Hoggatt Jr, M. Bicknell-Johnson, and R. Sarsfield, A generalization of Wythoffs
game, Fibonacci Quart. 17 (1979), 198211.
[16] L. Hrmander. Notions of convexity, vol. 127. Birkhauser, 1994.
[17] H. Landman, A simple FSM-based proof of the additive periodicity of the Sprague
Grundy function of Wythoffs game, in: R. J. Nowakowski (ed.), More Games of No
Chance, Proc. MSRI Workshop on Combinatorial Games, July, 2000, Berkeley, CA,
MSRI Publ., vol. 42, pp. 383386, Cambridge University Press, Cambridge, 2002.
[18] U. Larsson, Blocking Wythoff Nim, Electronic J. of Combinatorics 18(P120) (2011), 1.
[19] U. Larsson, A generalized diagonal Wythoff Nim, Integers 12 (2012), #G2, 24 pp., Comb.
Games Sect.
[20] W. A. Liu, H. Li, and B. Li, A restricted version of Wythoffs game, Electronic J. of
Combinatorics 18(P207) (2011), 1.
[21] G. Nivasch, More on the SpragueGrundy function for Wythoffs game, Games of No
Chance III 56 (2009), 377410.
[22] D. Singmaster, Almost all games are first person games, Eureka 41 (1981), 3337.
[23] D. Singmaster, Almost all partizan games are first person and almost all impartial games
are maximal, J. Combin. Inform. System Sci. 7 (1982), 270274.
[24] W. A. Wythoff, A modification of the game of Nim, Nieuw Arch. Wisk 7 (1907), 199202.
Author information
Aviezri S. Fraenkel, Department of Applied Mathematics and Computer Science, Weizmann
Institute of Science, Rehovot, Israel.
Email: fraenkel@wisdom.weizmann.ac.il
Yuval Tanny, Department of Applied Mathematics and Computer Science, Weizmann Insti-
tute of Science, Rehovot, Israel.
Email: yuval.tanny@gmail.com
Chapter 8
On the Multiplicative Order of FnC1=Fn

Modulo Fm
Takao Komatsu, Florian Luca, and Yohei Tachiya
Abstract. Here, we show that if s 62 1; 2; 4 is a fixed positive integer and m and n are coprime
positive integers such that the multiplicative order of FnC1 =Fn modulo Fm is s, where Fk is
the k-th Fibonacci number, then m < 500s 2 .
Keywords. Fibonacci Sequence, Multiplicative Order.
Mathematics Subject Classification 2010. 11A07, 11B39.
8.1 Introduction
Let Fk k0 be the Fibonacci sequence given by F0 D 0; F1 D 1 and
FkC2 D FkC1 C Fk for all k 0:
Let m 3 and n be positive integers such that Fm and Fn are coprime. Since
gcd.Fm ; Fn / D Fgcd.m;n/ , this last property holds when gcd.m; n/ 2 1; 2. Then
Fn is invertible modulo Fm . Assuming also that FnC1 is coprime to Fm , we can think
of the rational number FnC1 =Fn as an invertible element modulo Fm . Here we look
at its order denoted by s. Formally, s depends on both m and n, but we shall omit this
dependence in what follows.
It is quite possible that this order is s D 1. Indeed, this happens precisely when
FnC1 Fn .mod Fm /, so Fm j FnC1 Fn D Fn1 , and this holds when m j n 1.
Hence, when n 1 .mod m/.
2
It is also possible that s D 2. In this case, FnC1 Fn2 .mod Fm /, so
2
Fm j FnC1 Fn2 D .FnC1 Fn /.FnC1 C Fn / D Fn1 FnC2 :
Assume that m > 12. Then, by Carmichaels primitive divisor theorem (see [1]), Fm
has a primitive prime factor p. This primitive prime has the property that p j Fm but
p F` for any positive integer 1 ` < m. Furthermore, p j F` if and only if m j `.
From the above divisibilities, we see that either p j Fn1 , case in which m j n 1, or
p j FnC2 , case in which m j n C 2. The situation when m j n 1 leads to s D 1 and
this is not convenient, so we must have m j n C 2. Thus, n 2 .mod m/.
4
It is also possible that s D 4. In this case, FnC1 Fn4 .mod Fm /, so
4
Fm j FnC1 Fn4 D .FnC1 Fn /.FnC1 C Fn /.FnC1
2
C Fn2 / D Fn1 FnC2 F2nC1 :
If m > 12, then Fm has a primitive prime factor p. Since p divides the right-hand side
of the above divisibility relation, we get that m divides one of n 1; n C 2 or 2n C 1.
The first two cases lead to s 2 1; 2. The third case is possible only when m is odd
and n .m 1/=2 .mod m/.
From the above discussion, we see that for each of s 2 1; 2; 4 there exist infinitely
many positive integers m such that the set of invertible residue classes modulo Fm
contains a class representable as FnC1 =Fn for some appropriate positive integer n
whose multiplicative order is s. We asked ourselves if this property holds for some
other positive integers s. Maybe quite surprisingly, the answer is no.
Our main result is the following:
Theorem 8.1. If s 62 1; 2; 4 is a positive integer and m is such that there exists

an invertible class modulo Fm of the form FnC1 =Fn of multiplicative order s, then
m < 500s 2 .
For an algebraic number field K we put OK for the ring of algebraic integers in K.
8.2 Preliminary Results

We need the following four lemmas.
Lemma 8.2. Let X 3 be a real number. Let a and b be positive integers p with
maxa; b X . Then there p exist integers u; v not both zero with maxjuj; jvj X
such that jau C bvj 3 X.
p
Proof.pConsider the nonnegative numbers as Cbt p for s; t 2 0; 1; : : : ; b Xc. There
are .b Xc C 1/2 > X such numbers all in 0; 2X X. By the pigeon hole principle,
there exist .s1 ; t1 / .s2 ; t2 / such that
p
2X X p
ja.s1 s2 / C b.t1 t2 /j D j.as1 C bt1 / .as2 C bt2 /j 3 X:
X 1
Putting u D s1 s2 and v D t1 t2 , we get the desired conclusion.
p
We put D .1 C 5/=2 and D 1 .
Lemma 8.3. Let D e 2 iu=v with coprime positive integers u and v be a primitive
root of unity of order v. If v 62 1; 2; 4, then the two numbers

and
C
Chapter 8 On the Multiplicative Order of FnC1 =Fn Modulo Fm 111
are multiplicatively independent.
Proof. Assume on the contrary that there exist integers m and n not both zero such
that
m
D n : (8.1)
C
If m D 0, then n D 1, therefore n D 0, which is impossible. So, we assume that
m 0. Up to replacing the pair .m; n/ by .m; n/, we may assume that m > 0.
Assume first that v is coprime to 5. Then 2 K D Q.e 2 i=5 / and 2 L D Q.e 2 i=v /
and K and L are both Galois extensions of Q whose intersection is trivial (i.e., equal
to Q). Thus, every Galois automorphism of G D Gal.L=Q/ can be extended to
a Galois automorphism of the compositum M D KL D Q.e 2 i=5v / of K and L in
such a way that ./ D . Applying an arbitrary such 2 G to (8.1), we deduce that
Equation (8.1) holds when we replace by any conjugate of it. In particular, given
u1 ; u2 2 1; : : : ; v both coprime to v, we have
!m !m
e 2 iu1 =v 2e 2 iu2 =v
D n D : (8.2)
C e 2 iu1 =v C e 2 iu2 =v
Taking absolute values in (8.2) and then extracting mth roots, we get
2 2 C 2 2 2 cos.2u1 =v/ C 1
1 C D
2 C 2 cos.2u1 =v/ C 1 2 C 2 cos.2u1 =v/ C 1

e 2 iu1 =v 2 e 2 iu2 =v 2

D D
C e 2 iu1 =v C e 2 iu2 =v
2 2 cos.2u2 =v/ C 1
D
2 C 2 cos.2u2 =v/ C 1
2 2 C 2
D 1 C ;
2 C 2 cos.2u2 =v/ C 1
giving
cos.2u1 =v/ D cos.2u2 =v/:
This gives
q q
sin.2u1 =v/ D 1 cos.2u1 =v/2 D 1 cos.2u2 =v/2
D sin.2u2 =v/:
This argument shows that there exist at most 2 primitive roots of unity of order v,
therefore .v/ 2, and since v 62 1; 2; 4, we get that v 2 3; 6. Let us look at these
p p
cases. In this instance, M D Q. 5; i 3/ is of degree 4 over Q. We compute
8 9
p p !1
< 2 C 5 C "i 3 =
2 p p W " 2 1 :
C : 5 C "i 3 ;
p ispa unit, Equation (8.1)

Since p that the principal ideals in OM given by
p tells us
. 5 C "i 3/m OM and .2 C 5 C "i 3/m OM are equal for some " 2 1. By
unique factorization of ideals in OM , we get that
p p p p
. 5 C "i 3/OM D .2 C 5 C "i 3/OM :
p p
In particular, we deduce that 5 C "i 3 j 2. Taking norms in this last divisibility
relation, we get that
p p
64 D jNM=Q . 5 C "i 3/j j jNM=Q .2/j D 16;
which is false.
A similar argument applies when 5 j v. In this case K D Q.e 2 i=5 / L, and so
M D L and G D Gal.M=Q/ is isomorphic with the group of invertible elements
modulo v which has order .v/. Further, by Galois theory, there are exactly .v/=2
Galois automorphisms such that ./ D . We deduce that there exists a subset
U
1; 2; : : : ; v of positive integers coprime to v having exactly .v/=2 elements,
such that Equation (8.1) holds for all D e 2 iu=v with all u 2 U. The preceding
argument shows that
cos.2u1 =v/ D cos.2u2 =v/ holds for all u1 ; u2 2 U;
therefore
sin.2u1 =v/ D sin.2u2 =v/ holds for all u1 ; u2 2 U:
This shows that the number of elements in U is at most 2, so .v/ 4. Since we

already have that 5 j v, we get that v 2 5; 10: We calculated that all numbers of the
form

;
C
when is a primitive root of unity of order v 2 5; 10, are algebraic numbers in
M D Q.e 2 i=5 / of norm 111 , and therefore Equation (8.1) does not hold in this
instance either for any pair of integers m; n with not both zero.
Lemma 8.4. Let D e 2 iu=v , where v 4 is a positive integer and u 2 1; 2; : : : ; v

is coprime to v. Then the divisibility relation 1 C 2 j 2v holds in OK , where K is any
number field containing Q./.
Proof. We distinguish four cases. For a positive integer m we put m .X / for the mth
cyclotomic polynomial.
If v is odd, then 2 is also a primitive root of order v of unity, so

1 C 2 j v .1/ j X v 1 D 2:
XD1
If 2 j v and v=2 is odd, then 2 is a primitive root of unity of order v=2 and

1 C 2 j v=2 .1/ j X v=2 1 D 2:
XD1
If 4 j v and v=4 is odd, then, since v=4 > 1, it follows that .X v=4 1/ and .X C 1/
are proper divisors of X v=2 1 and they do not have any common roots. Thus,
X v=2 1 X v=4 C 1

1 C 2 j v=2 .1/ j D D
.X v=4 1/.X C 1/ XD1 X C 1 XD1

X v=41 X v=42 C C 1 D v=4:
XD1
If 8 j v, then
X v=2 1

1 C 2 j v=2 .1/ j D X v=4
C 1 D 2:
X v=4 1 XD1 XD1
For a prime number p and a nonzero integer m, we put p .m/ for the exponent of
the prime p in the factorization of m. For a finite set of primes S and a positive integer
m, we put Y
mS D p p .m/
p2S
for the largest divisor of m whose prime factors are in S.
Lemma 8.5. If S is any finite set of primes and m is a positive integer, then
Y
.Fm /S 2m FpC1 :
p2S
Proof. For a prime p, let fp be its order of appearance in the Fibonacci sequence,
which is the minimal positive integer k such that p j Fk . It is well-known that
8

0 if m 6 0 .mod fp /I

< .F / C .m=f / if m 0 .mod f /; p is oddI
p fp p p p
p .Fm / D

1 if m 3 .mod 6/; p D 2I

: 2 C 2 .m/ if m 0 .mod 6/; p D 2:
In particular, the inequality
p .Fm / p .Ffp / C p .m/ C p;2
always holds with p;2 being 0 if p is odd and 1 if p D 2. Since fp p C 1 holds for
all primes p, we get that
! !
Y Y
p .Ffp / p .m/
.Fm /S p p 22 .m/C1
p2S pjm
p>2
Y Y
2m Ffp 2m FpC1 ;
p2S p2S
which is what we wanted to prove.
8.3 Proof of Theorem 8.1

We use the Binet formula
n n
Fn D valid for all n 0: (8.3)

We also use the inequalities
n2 Fn n1 valid for all n 1: (8.4)
We also use the fact that if m 1, then the sequence Fk k0 is periodic modulo Fm
with period 4m. Assume now that s 1; 2; 4 is a positive integer and that m > 1000
is such that there exist n with Fn coprime to Fm and FnC1 =Fn is invertible modulo
Fm of multiplicative order exactly s. From the periodicity of Fk k0 modulo Fm , we
may assume that n 4m, and since Fn FnC1 and Fm are coprime, we may assume
that n 4m 2. We shall exploit the relation
Y
s
Fm j FnC1 Fns D .FnC1 Fn /: (8.5)
W s D1
We split Fm into various factors.

Step 1. We put
A D gcd.Fm ; FnC1 Fn /;
B D gcd.Fm ; FnC1 C Fn /;
2
C D gcd.Fm ; FnC1 C Fn2 /;
and we bound ABC .

Then,
A D gcd.Fm ; Fn1 / D Fd1 ; where d1 D gcd.m; n 1/I

B D gcd.Fm ; FnC2 / D Fd2 ; where d2 D gcd.m; n C 2/I
C D gcd.Fm ; F2nC1 / D Fd3 ; where d3 D gcd.m; 2n C 1/:
The numbers d1 ; d2 ; d3 are divisors of m and they are proper, since if di D m for
some i 2 1; 2; 3, then, from what we have seen in the Introduction, we would get that
s 2 1; 2; 4, which is not the case. Observe that any two of d1 ; d2 ; d3 are coprime,
or the greatest common divisors of any two of them is exactly 3. The second condition
holds precisely when m 0 .mod 3/ and n 1 .mod 3/. Indeed, this holds because
gcd.d1 ; d2 / D gcd.m; n 1; n C 2/ D gcd.m; n 1; 3/I

gcd.d1 ; d3 / D gcd.m; n 1; 2n C 1/ D gcd.m; n 1; 3/I
gcd.d2 ; d3 / D gcd.m; n C 2; 2n C 1/ D gcd.m; n C 2; 3/ D gcd.m; n 1; 3/:
Let i 2 1; 2; 3 be such that di D maxd1 ; d2 ; d3 and let j; k be indices such that

i; j; k D 1; 2; 3. Noting that since di is a proper divisor of m, we have di m=2.
When any two of d1 ; d2 ; d3 are coprime, we then have that
d1 d2 d3 m; therefore dj dk m2=3 : (8.6)
When the greatest common divisor of any two of the numbers d1 ; d2 ; d3 is exactly 3,
we get

d1 d2 d3 m dj dk m 2=3
; therefore ;
3 3 3 3 3 3 3
leading to the slightly worse bound than Equation (8.6), namely
dj dk 34=3 m2=3 : (8.7)
Thus, using Equation (8.4), we get that

4=3 m2=3 2
ABC D Fd1 Fd2 Fd3 d1 Cd2 Cd3 3 m=2Cdj Cdk 3 m=2C3 ;
(8.8)
where we used also the fact that the inequality a C b ab C 1 is valid for all positive
integers a and b with a D dj and b D dk .
Step 2. We put S D 2 [ p W p j s and D D .Fm /S , and bound D.
By Lemma 8.5 and inequalities (8.4), we have that
Y P
D 2mF3 FpC1 < 4m pjs p < sClog.4m/= log ; (8.9)
pjs
P
where we used the fact that pjs p s, which is easily proved by induction on the
number of distinct prime factors of s.
Step 3. We put
Fm
ED ;
gcd.ABCD; Fm /
and bound E.
We shall estimate the number E by using the fact that E is coprime to 2s, as well
as divisibility (8.5), which in particular tell us that
Y
Fm j ABC .FnC1 Fn /;
W s D1
621;i
which shows that Y

Ej .FnC1 Fn /: (8.10)
W s D1
621;i
p
Let K D Q.e 2 i=s ; 5/, which is a number field of degree d equal to .s/ or to
2.s/, according to whether s is a multiple of 5 or not. Assume that there are ` roots
of unity participating in the product appearing in the righthand side of (8.10) and
label them 1 ; : : : ; ` . Clearly, ` 2 s 4; s 1. Write
Ei D gcd.E; FnC1 i Fn / for all i D 1; : : : ; `; (8.11)
where Ei are ideals in OK . Then relations (8.10) and (8.11) tell us that
Y
`
EOK j Ei : (8.12)
iD1
Our next goal is to bound the norm NK=Q .Ei / of Ei for i D 1; : : : ; `: First of all,
Fm 2 Ei . Thus, with Equation (8.3) and the fact that D 1 , we get
m .1/m m .mod Ei /:
Multiplying the above congruence by m , we get
2m .1/m .mod Ei /: (8.13)
We next use Equations (8.3) and (8.11) to deduce that
. nC1 .1/nC1 n1 / . n .1/n n / 0 .mod Ei /; . D i /:
Multiplying both sides above by n , we get
2n . / .1/nC1 . 1 C / 0 .mod Ei /: (8.14)

Let us show that and Ei are coprime. Assume this is not so and let be some
prime ideal of OK dividing both and Ei . Then we get .mod / and so
1 .mod / by (8.14). Multiplying these two congruences we get 1 2
.mod /. Hence, j 1 C 2 , and so by Lemma 8.4 we get that j 2s. However, this
contradicts the fact that j Ei j E, with E an integer coprime to 2s. Thus, indeed
and Ei are coprime, so is invertible modulo Ei . Now congruence (8.14)
shows that
1 C
2n .1/nC1 .mod Ei /;

therefore !
2nC1 nC1 C
.1/ .mod Ei /: (8.15)

We now apply Lemma 8.2 to a D 2m and b D 2n C 1 2.4m 2/ C 1 < 8m
with the choice XpD 8m to deduce that there exist integersp u, v not both zero with
maxjuj; jvj X such that j2mu C .2n C 1/vj 3 X : We raise congruence
(8.13) to u and congruence (8.15) to v and multiply the resulting congruences getting
!v
2muC.2nC1/v muC.nC1/v v C
D .1/ .mod Ei /:

We record this as
!b
a C
.mod Ei / (8.16)

for suitable roots of unity and of order dividing 2s with not of order 1; 2 or 4,
where a D 2mu C .2n C 1/v and b D v. We may assume that a 0, for if not, we
replace the pair .u; v/ by the pair .u; v/, thus replacing .a; b/ by .a; b/ and
by 1 and leaving unaffected. We may additionally assume that b 0, for if not,
we replace b by b and D by D , again a root of unity of order dividing 2s
but not of order 1; 2 or 4, and leave a and unaffected. Thus, Ei divides the algebraic
integer
Ei D a . i /b i . C i /b ; (8.17)
where i 2 i ; i and i is some suitable root of unity of order dividing 2s. Let us
show that Ei 0. If Ei D 0; we then get
!b
a C i
D i ;
i
and after raising both sides of the above equality to the power 2s, we get, since 2s
i D 1,
that
!2bs
2sa C i
D :
i
By Lemma 8.3, we have that as D bs D 0, and so a D b D 0. Since b D 0, we get

that v D 0, and later since 2mu C .2n C 1/v D a D 0 and v D 0, we get mu D 0,
so u D 0, therefore u D v D 0, but this is not allowed. We now bound the absolute
values of the conjugates of Ei . We find it more convenient to work with the associate
of Ei given by
Gi D ba=2c Ei D aba=2c . i /b ba=2c i . C i /b :
Note that
p p p p
a j2m C .2n C 1/vj 3 X D 6 2m; and b D jvj X D 2 2m:
Let be an arbitrary element of G D Gal.K=Q/. We then have that .i / D 0i ,
.i / D i0 , where 0i and i0 are roots of unity of order dividing 2s. Furthermore,
./ 2 ; . If ./ D , we then get
j .Gi /j D j aba=2c . i0 /b 0i ba=2c . C i0 /b j
.aC1/=2 . C 1/b C . C 1/b
p p p
2 .aC1/=2 . C 1/b .2 / 3 2m . 2 /2 2m
p p
D .2 / 7 2m ; (8.18)
while if ./ D , we also get
j .Gi /j D j aba=2c . i0 /b ba=2c 0i . C i0 /b j
. 1 C 1/b C a=2 . 1 C 1/b
p p
D b C a=2Cb 2 3 2m 2 2m

p
5 2m
D 2 :
.1/ .d /
In conclusion, inequality (8.18) holds for all 2 G. Thus, if we write Gi ; : : : ; Gi
for the d conjugates of Gi in K, we then get that
p p
jNK=Q .Ei /j jNK=Q .Ei /j D jNK=Q .Gi /j .2 /d 7d 2m ;
where the first inequality above follows because Ei divides Ei ; hence Gi , and Ei 0.
Multiplying the above inequalities for i D 1; : : : ; ` we get, also using Equation (8.12),
that !
Y`
E d D NK=Q .E/ D NK=Q .EOK / N Ei
iD1
Y
`
p p
NK=Q .Gi / .2 /`d 7d ` 2m ;
iD1
and therefore
p p p p
E .2 /` 7` 2m D 7` 2mC` log.2 /= log : (8.19)
Thus, we have bounded E.
Step 4. The final inequality.

We now use Equation (8.4) to bound Fm from below as Fm > m2 , and the fact
that Fm ABCDE and the estimates (8.8), (8.9), and (8.19), to bound Fm from
above as
p p
4=3 m2=3 2CsClog.4m/= log C7`
Fm m=2C3 2mC` log.2 /= log
;
to conclude that
p
m 4=3 2=3 log.4m/ p ` log.2 /
m2< C3 m 2CsC C 7` 2m C ; (8.20)
2 log log
where ` s 1. We look at
p
m 4=3 2=3 log.4m/ p .s 1/ log.2 /
f .m; s/ D 3 m s 7.s 1/ 2m :
2 log log
Computing the partial derivative with respect to m, we get
@f 1 2 31=3 1 7.s 1/
g.m; s/ D .m; s/ D 1=3
p : (8.21)
@m 2 m m log 2m
The function g.m; s/ is positive when m 500s 2 and s 3, because in this range
1 2 31=3 1 7
g.m; s/ p > 0:103:
2 .4500/1=3 4500 log 1000
Thus, in order to prove that m < 500s 2 , it suffices to prove f .500s 2 ; s/ > 0. We
checked with Mathematica that this inequality holds for s 17. For the remaining
values s 2 3; 16, we checked individually by noticing that for each one of these
values of s a slightly better inequality than (8.20) holds. For example, in the case
when s 2 3; 5; 7; 9; 11; 13; 15, there is no need for d2 and d3 because s is odd. Thus,
the analogue of inequality (8.20) for such values of s is simply
p
m log.4m/ p .s 1/ log.2 /
m2< 1CsC C 7.s 1/ 2m C : (8.22)
2 log log
Plugging in s D 3; 5; 7; 9; 11; 13, and 15 into (8.22), we got m bounded by 2000,
7000, 15000, 26000, 40000, 57000, and 77000, respectively, and so the inequality
m < 500s 2 definitely holds for these values of s as well. When s D 6; 10; 14, we keep
only two divisors in Case 1, namely d1 and d2 since there is no need for d3 . Putting
i 2 1; 2 such that di D maxd1 ; d2 and letting j be such that i; j D 1; 2, the
analog of inequality (8.7) is p
dj 3m:
Since ` s 2, when s D 6; 10; 14, the analog of inequality (8.20) in this case is
p
m p log.4m/ p .s 2/ log.2 /
m 2 < C 3m 2 C s C C 7.s 2/ 2m C ; (8.23)
2 log log
giving for s D 6; 10, and 14 that m is bounded by 8000; 27000, and 60000; respec-
tively. Thus, the inequality m < 500s 2 holds also for s D 6; 10; 14.
Finally, for s D 8; 12; 16, we use the analog of inequality (8.20) with the value
` s 4, yielding
p
m 4=3 2=3 log.4m/ p .s 4/ log.2 /
m2 < C3 m 2CsC C7.s4/ 2mC ; (8.24)
2 log log
which at s D 8; 12, and 16 gives that m is bounded by 16000; 45000, and 88000,
respectively, so the inequality m < 500s 2 holds in these last three cases as well.
This completes the proof of the theorem.
8.4 Comments and Numerical Results

Numerical results are few because the bounds of Theorem 8.1 are very weak. However,
from what we have said above at Step 4 of the proof of Theorem 8.1, we have that
m < 2000 when s D 3, and m < 7000 for s D 5. We ran a Mathematica code run for
about a day and searched for all such m and for all n 2 1; 4m for which FnC1 =Fn is
indeed an element of order s modulo Fm . No example was found with s D 3, and the
example F7 =F6 modulo F10 is the only example with s D 5.
In [3], it was shown that the Diophantine equation
Fnx C FnC1
x
D Fm
has no positive integer solutions with n 2 and x 3. The method there was based
on linear forms in logarithms. Since for any potential solution of the above equation
it is easy to check that FnC1 =Fn is an invertible element modulo Fm which is not of
order 1; 2; or 4 but whose order divides 2x, we get right away from Theorem 8.1 that
m < 2000x 2 . Next, by Equation (8.4), we get
2 1
nx FnC1
x
< Fnx C FnC1
x
D Fm < m1 < 2000x ;
and we derive that n < 2000x. However, we could not find an elementary upper
bound on x out of this equation (without appealing to linear forms in logarithms). We
conclude by mentioning that all the solutions of the more general Diophantine equation
y
Fnx C FnC1
x D Fm in positive integers .n; m; x; y/ were found in [2].
Acknowledgments. T. K. was supported in part by the Grant-in-Aid for Scientific

research (C) (No. 22540005), the Japan Society for the Promotion of Science. F. L.
worked on this project during a visit to Hirosaki in January and February of 2012 with
a JSPS Fellowship (No. S-11021).
We thank the referee for comments which improved the quality of this paper.
References
[1] R. D. Carmichael, On the numerical factors of the arithmetic forms n n , Ann. Math.
(2) 15 (1913), 3070.
[2] N. Hirata-Kohno and F. Luca, On the Diophantine equation Fnx C FnC1x
D Fmy , preprint,
2012.
[3] F. Luca and R. Oyono, On the sum of powers of two consecutive Fibonacci numbers, Proc.
Japan Acad. Ser. A. 87 (2011), 4550.
Author information
Takao Komatsu, Graduate School of Science and Technology, Hirosaki University, Hirosaki,
Japan.
Email: komatsu@cc.hirosaki-u.ac.jp
Florian Luca, Centro de Ciencias Matemticas UNAM, Morelia, Michoacn, Mexico.

Email: fluca@matmor.unam.mx
Yohei Tachiya, Graduate School of Science and Technology, Hirosaki University, Hirosaki,
Japan.
Email: tachiya@cc.hirosaki-u.ac.jp
Chapter 9
Outcomes of Partizan Euclid

Neil A. McKay and Richard J. Nowakowski
Abstract. PARTIZAN EUCLID is a game based on the Euclidean algorithm. The outcome of any
position .p; q/ is determined by a single path of the game tree; this path has connections to the
furthest integer continued fraction of p=q. We convert the question of Who wins? to a word
problem, then give a list of reductions that reduces the word/position to one of 9 positions.
Keywords. Combinatorial Games, Partizan Euclid.
9.1 Introduction
Suggested by Euclid [3] and Richard K. Guy, the game of PARTIZAN EUCLID is
played by two players, Left and Right, and starts with a pair of positive integers .p; q/
with p > q. Let p D kq C t where 0 t < q. If q j p (i.e., t D 0), then the game
is over; otherwise, Left moves to .q; t / and Right moves to .q; q t /. The game may
seem trivial as there is only one move available for each player. However, as we shall
show, answering the question Who wins? reveals some of the interesting structure
of the game. We would like to answer the question of who wins in the disjunctive sum
of this game, but this appears to be difficult. See the last section for a discussion of
that problem.
In the (impartial) game EUCLID, which is also played with .p; q/, a pair of positive
integers, a player is allowed to remove any multiple of the smaller from the larger
provided the remainder is positive. Lengyel [7] reports that Schwartz first found that
EUCLID is the sequential sum [10] of nim-heaps: given .p; q/, suppose the normal
continued fraction of pq is a1 ; a2 ; : : : ; an (an > 1, except if Fibonacci numbers are
involved); then the EUCLID position .p; q/ corresponds to playing the sequential sum
of NIM with nim-heaps a1 ; a2 ; : : : ; an . EUCLID has attracted much attention and has
been generalized (see, e.g., [4, 5]). PARTIZAN EUCLID is related to nearest and farthest
integer continued fractions (NICF and FICF) (see [8]).
In the case of both NICFs and FICFs we write rational numbers as a sum or differ-
ence of an integer and a rational less than 1. For example, the FICF for 11 8 is obtained
by rewriting, noting that
11 1 11

8 D2 8=5
, since 2 is further away from 8 than 1;
8 1

5 D1C 5=3
, since 1 is further away than 2;
5 1

3 D1C 3=2
, since 1 is further away than 2;
3 1 1

2 D2 2=1
D1C 2=1
, since 1 and 2 are equally distant.
We are not interested in the continued fraction itself but in noting that during the cal-
culation (i) integer subtract fraction corresponds to a move by Right and (ii) fraction
subtract integer corresponds to a Left move. We will use the word rl le to represent
this where e is the common move to .2; 1/. Section 9.2 reports on the structure of the
game tree and shows there is one path, the path obtained from the FICF algorithm, that
determines the whole game tree.
For example, in Figure 9.1 the path formed by the moves (edges)
Right Left Left
.11; 8/ ! .8; 5/ ! .5; 3/ ! .3; 2/
is the important path. Why is it important? Nontrivial parts of the tree that are not on
that path are isomorphic to parts rooted on the path; .8; 3/0 is isomorphic to .5; 3/0 ,
.5; 2/0 is isomorphic to .3; 2/, and .3; 1/ is isomorphic to .2; 1/. All the infor-
mation needed to determine the outcome and value of .11; 8/ is found on this path.
A game tree (position) is represented as a word from the alphabet r; l ending in e.
In Lemma 9.13, we find reduction rules that preserve the outcome class of the word,
moreover, any word reduces to one of just 9 words each with length at most 4. This can
be accomplished in time linear in the length of the corresponding FICF. Unfortunately,
these reductions most of the time do not preserve the value.
Figure 9.1. Some of the game tree of .11; 8/:

Chapter 9 Outcomes of Partizan Euclid 125
We will denote a game position as E.p; q/. Also, we will use p%q for p mod q.
We try to present a sufficient amount of game theory to make the paper self-con-
tained (with the exception of the last two sections). For terms not defined in the pa-
per we follow [1]. A position, say h, is defined in terms of its options as follows:
h D hL jhR . For example, where t D p%q, E.p; q/ D E.q; t /jE.q; q t /. The
outcome of a position is Left, Right, Next or Previous depending on (under perfect
play), respectively, whether Left can win going first and second, Right can win going
first and second, the next player to move wins regardless if this Left or Right, the next
player cannot win regardless if this Left or Right. We phrase this more formally.
Lemma 9.1. Let h be a position. The outcome of h is determined by the outcomes of

its options. Specifically,
o.h/ D L iff 9hL , o.hL / 2 L; P and 8hR , o.hR / 2 L; N ;
o.h/ D P iff 8hL , o.hL / 2 N ; R and 8hR , o.hR / 2 L; N ;
o.h/ D N iff 9hL , o.hL / 2 L; P and 9hR , o.hr / 2 P ; R;
o.h/ D R iff 8hL , o.hL / 2 N ; R and 9hR , o.hr / 2 P ; R.
Let g D E.p; q/. Since there is at most one option for each player we will abuse
notation and write o.g/ D o.g L jg R / as o.gL /jo.gR /. For example, N jP D R.
9.2 Game Tree Structure

Lemmas 9.2 and 9.3 each show that for every position there are infinitely many po-
sitions with the same game tree. We call positions equivalent if they have the same
game tree.
Lemma 9.2. For all k, E.kp; kq/ D E.p; q/.
Proof. Recall that q j p if and only if kq j kp. Thus, E.p; q/ D j if and only if
E.kp; kq/ D j. Let t D p%q, then by induction E.p; q/ D E.q; t /jE.q; q
t / D E.kq; k t /jE.kq; kq k t / D E.kp; kq/.
Note that if h D E.n; m/ is a follower of a position g D E.p; q/ with gcd.p; q/ D 1,
then gcd.n; m/ D 1. In the rest of the chapter we will assume that every position has
gcd.p; q/ D 1 and thus p%q D 0 if and only if q D 1.
Lemma 9.3. If p > 2q, then E.p; q/ D E.p q; q/.
Proof. Let p D kq C t , 0 t < q and k 2. Then p q D .k 1/q C t , 0 t < q
and k 1 1. Consider the options of both positions:
E.p; q/ D E.q; t /jE.q; q t /
E.p q; q/ D E.q; t /jE.q; q t /:
Since they have identical options the two positions are equivalent.
A position E.p; q/ will be called standard if q < p < 2q. All positions E.p; q/
with q > 2 are equivalent to some standard position (which is reachable with repeated
applications of Lemma 9.3). Notably E.2; 1/ is not standard, and positions of the form
E.k; 1/ are neither standard nor equivalent to some standard position. A follower of
a standard position may not be standard. For example, E.3; 2/ has only one proper
follower, E.2; 1/, which is not standard.
Lemma 9.4. Let g D E.p; q/ and t D p%q. If t 6D 0 then g has exactly one standard
option except when q D 2t (i.e. E.3; 2/). Moreover;
if 2t > q, then exactly g L is standard;
if 0 < 2t < q; then exactly gR is standard.
Proof. As t > 0, g L D E.q; t / and g R D E.q; q t /. Recall by the definition

of standard; that t > q2 if and only if g L is standard and t < q2 if and only if g R
is standard. Otherwise, 2t D q, implying that p D 3t and subsequently that g D
E.3; 2/; from E.3; 2/ both players have the option E.2; 1/, which is not standard.
There is a unique standard position with a given left or right option.
Lemma 9.5. Let g D E.p; q/ and 0 < t < q. If g is standard, then

if g L D E.q; t /; then g D E.q C t; q/;
if g R D E.q; q t /; then g D E.q C t; q/.
Proof. If g L D E.q; t / then p D kq C t ; as g is standard, k D 1. If gR D E.q; q t /

then p D kq C .t q/ D .k 1/q C t ; as g is standard, k D 2.
The essential game tree structure is given by Theorem 9.6.
Theorem 9.6. Let g D E.p; q/ and t D p%q:

if t D 0 then g D j;
if 2t D q then g L D g R ;
if 2t > q then gLL D gR ;
if 0 < 2t < q then g L D g RL .
Proof. If t D 0 then the game is over and g D j. Thus we suppose t > 0 and hence
g L D E.q; t / and g R D E.q; q t /.
Suppose 2t D q; then g L D E.q; t / D E.q; q 2t C t / D E.q; q t / D gR .
Moreover, q t j q, so E.q; t / D E.q; q t / D 0. (For the purists, g D 0j0 D .)
Suppose 2t > q; then q > 2.q t / and t > q t so gR D E.q; q t / D
E.q .q t /; q t / D E.t; q t / by Lemma 9.3, and g LL D E.q; t /L D E.t; q t /,
giving
g R D E.t; q t / D g LL :
Suppose 2t < q; then g L D E.q; t / D E.q t; t / by Lemma 9.3. Since gR D

E.q; q t / we have that g RL D E.q t; t /, giving g RL D g L .
Corollary 9.7. Let p > q with p D kq C t , 0 < t < q and let g D E.p; q/:
if 2t > q; then g D g L jgLL ;
if 2t < q; then g D g RL jgR .
Proof. This is a simplification of Theorem 9.6.
We note the similarity of Lemma 9.4 and Corollary 9.7. In Corollary 9.7, the two
options are the standard option and its left option. The case 2t > q is when the lower
integer is the farthest integer when calculating the FICF and 2t < q is when the
higher integer is the farthest.
This motivates our next definition, the signature of a position, in which we high-
light the important option at each stage. Recall that when we refer to E.p; q/ we are
assuming that gcd.p; q/ D 1.
Definition 9.8. Let g D E.p; q/. The signature of g, denoted Sg , is defined as fol-
lows. If q D 1 then Sg D , the empty word. If q D 2 then Sg D e. Otherwise, let h
be the standard option from g. If gL D h then Sg D lSh . If gR D h then Sg D rSh .
The position g and the standard positions that are successively the standard option
(as per Lemma 9.4) starting from g are the spine of g.
For example, if g D E.12; 7/ then the signature of g is lrle and the spine of g
is E.12; 7/; E.7; 5/; E.5; 3/; E.3; 2/. Often, we will write the signature with super-
scripts; for example, S D l l lrrl l lre is the same as S D l 3 r 2 l 3 e.
If two positions have the same signature then they have the same game tree. We
use signatures liberally to represent positions. Furthermore, we use f to denote the
position g where Sg D Sf .
The position E.3; 2/ is the unique standard position with signature e. This position
is at the bottom of every spine for every position other than E.k; 1/ and E.2k C 1; 2/
for k 2.
Theorem 9.9. Let g be a PARTIZAN EUCLID position. Every follower of g not of the
form E.k; 1/ is equivalent to some position on the spine of g.
Proof. A position is on its spine, so we only need consider proper followers. If the
length of the signature is 0 then there are no proper followers. If the length of the sig-
nature is 1; then Sg D e and g D E.3; 2/; where the only proper follower is E.2; 1/.
We proceed by induction on the length of signature. If the length of the signature is
at least 2, then the standard option is on the spine; the nonstandard option is (by The-
orem 9.6) either gL D g RL or g R D gLL , which is the left option of the standard
option and is by induction on the spine of the standard option or of the form E.k; 1/. As
the spine of the standard option is part of the spine of g, this completes the proof.
Corollary 9.10. Consider the position g, and let k represent an unfixed nonnegative
integer.
We can write Sg as either r k le or r k e; Lefts move has signature e or , respec-
tively.
We can also write Sg as either re, lr k le, or lr k e; Rights move has signature
e, e, or , respectively.
Proof. If Sg is r k le or r k e, then Sg L is e or , respectively, because gL D gRL D

k
g RRL D g RRRL D D g R L .
If Sg D re then Sg R D e. Otherwise, gR D g LL so Sg R is the signature of the
position resulting from two Left moves namely e or , as seen by the first part.
In the examples below, we repeatedly use Theorem 9.6, but using Corollary 9.10
one can easily jump from the leftmost term in a line of equalities to the rightmost. Let
g D l lrle then
e D e L je R D j;
le D le L jle R D ejle LL D eje L D ej;
rle D rle L jrle R D rle RL jle D le L jle D ejle;
lrle D lrle L jlrle R D rlejlrle LL D rlejrle L D rlejrle RL
D le L je D rleje;
l lrle D l lrle L jl lrle R D lrlejl lrle LL D lrlejlrle L D lrlejrle:
9.3 Reducing the Signature
The paired outcome of a position g (or signature Sg ) is the pair .o.g L /; o.g//, denoted
by po.g/ or po.Sg /. For example, if Sg D e, then po.g/ D po.e/ D .P ; N /. Note
that po./ is not defined.
The paired outcome of a position depends upon which option is standard and the
paired outcome of that option.
Lemma 9.11. If Sg D lSh then po.g/ D .o.h/; o.h/jo.hL //. If Sg D rSh then
po.g/ D .o.hL /; o.hL /jo.h//.
Proof. Follows immediately from Theorem 9.6.

That is, the paired outcome of a position with a standard option is determined by the
paired outcome of the standard option. We use l and r to denote functions on paired
outcomes; we write l po.Sg / to mean po.l Sg / and r po.Sg / to mean po.r Sg /.
There are 4 4 D 16 ordered pairs of outcome classes. However, as stated in
Lemma 9.1, there are relationships between the outcome of a position and the outcome
of its options; there are only eight ordered pairs that are paired outcomes of positions.
Figure 9.2 has a vertex for each of the eight paired outcomes. The directed edges
labelled l and r from a paired outcome, say x, lead to l x and r x, respectively.
The outcome of a position is given by the paired outcome of the signature. Provided
we know the paired outcome of some suffix of the signature, we can find the paired
outcome of the next larger suffix using Figure 9.2 and eventually the desired paired
outcome. As e is the suffix of every nonempty signature, we only need to know that
po.e/ D po.E.3; 2// D .P ; N / is where we start and to read the signature from the
right starting after e.
We have now described a relatively efficient method to determine the outcome of a
position given its signature, but we give a better way to determine the outcome than a
walk through the graph for each letter in the signature.
Just as l and r are functions on paired outcomes, we use words (denoted by Greek
letters) from l; r such as D lrr as functions on paired outcomes in the natural
way where x D l r r x. For any such , po.Sh / D po.Sh /.
We give reduction rules by which we can simplify a word (signature) that preserves
outcome, in the sense that if two positions have the same reduced signature then they
have the same outcome. The main goal of this section is to prove Theorem 9.15, in
which we give a short list of words to which any signature will reduce.
Lemma 9.12. If 2 l; r then .L; L/ D .L; L/ and .R; R/ D .R; R/.
Proof. Immediate from Figure 9.2.
Figure 9.2. Paired outcome of signatures.

Lemma 9.13. Let x be a paired outcome:

1. l 3 x D x;
2. r 2 x D r x;
3. rlr x D rlr x;
4. .rl 2 /2 r x D r x;
5. rl lr .P ; N / D l .P ; N /;
6. rl l .P ; N / D rl l .P ; N /;
7. rl .P ; N / D l .P ; N /.
Proof. In this proof we make extensive use of Figure 9.2. The most common use is to
find a vertex with the desired paired outcome then look up the paired outcomes reached
by the directed edges. For the first four rules, we need to show that these equation holds
for all paired outcomes x.
Rule 1: we apply l to each x three times to observe that we return to the x with
which we started.
Rule 2: following two edges marked r and, regardless, the second edge is a loop.
Rule 3: following three edges marked r, l and r results in .L; L/ or .R; R/ so the
first part of the signature it is irrelevant.
Rule 4: following the walk with edges marked rl lrl lr either (i) goes once around
the 6-cycle in Figure 9.2 ending at the starting vertex; (ii) or if the starting vertex is
.L; L/ or .R; R/ then remains at .L; L/ or .R; R/ respectively; (iii) starts at .L; N /
or .R; P / then the first r-edge goes to .L; L/, .R; R/ respectively. In all cases the
effect of following the last r edge in the walk is the same as following the first.
Rule 5: obvious from the figure.
Rule 6: from .P ; N / the walk l, l, r ends at .R; R/ and any further edges do not
change the paired outcome.
Rule 7: obvious from the figure.
Lemma 9.14. If x is a paired outcome and x D x, then po.e/ D

po.e/.
Proof. po.e/ D po.e/ D po.e/ D po.e/.
In applying Lemma 9.14 to reduce a word, we make reference to particular rules

from Lemma 9.13.
We call a word irreducible if none of the reduction rules are applicable. Reduction
rules may be applied in any order to the signature of a position, say g, to derive an
irreducible word; the irreducible word corresponds to some other position, say h. The
outcome of g is the same as the outcome of h, as they have the same paired outcome,
Table 9.1. Irreducible signatures with corresponding positions and outcomes.
Sg o.g/ g
P E.2; 1/
e N E.3; 2/
re L E.4; 3/
le R E.5; 3/
l le P E.8; 5/
lre N E.7; 4/
rlre L E.10; 7/
rl le R E.11; 8/
l lre P E.11; 7/
which is a stronger condition. As there are a finite number of irreducible words, we

will be able to compute and store the outcomes of those positions.
Theorem 9.15. There are nine irreducible words: , e, re, le, l le, lre, rlre, l lre,
and rl le.
For the proof of Theorem 9.15 we need the following Lemma:
Lemma 9.16. A word containing 4 rs is reducible.
Proof. Suppose is an irreducible word containing 4 rs. By Rule 2, each pair of

consecutive rs is separated by at least one l. By Rule 3, each pair of rs except possibly
the leftmost, is separated by more than one l. By Rule 1, each pair of rs is separated
by at most 2 ls. That is, the rightmost 3 rs form the pattern rl lrl lr, which contradicts
the assumption that is irreducible by Rule 4.
Proof of Theorem 9.15. The irreducible words that do not end in e are easy to list and
count with the help of Lemma 9.16; such words have at most 3 rs. In what follows,
and are one of either , l, or l l.
Words with 3 rs are of the form rlrl lr, of which there are 3.
Words with 2 rs are of the form rl lr or rlr, of which there are 12.
Words with 1 r are of the form r, of which there are 9.
Words with no r are of the form , of which there are 3.
There are a total of 27 such words. The only irreducible signatures are among the
set containing these strings but with a trailing e appended, and the empty word.
We show that 19 of the 27 strings reduce to the remaining 8: e, le, l le, re, lre,
l lre, rl le and rlre.
The 7 strings of the form rl le where is nonempty reduce by Rule 6 to rl le.

The 4 words of the form rl lrle reduce to rl l le and then to re by Rules 7 and 1,
respectively, leaving re, lre, l lre, and rlre.
The 4 words of the form rl lre reduce to le by Rule 5, leaving the 3 strings with
no r (note l l le D e) and rl le.
The 4 words of the form rle reduce to le by Rule 7, leaving the 3 strings with
no r (note l l le D e) and rl le.
We note that of the none of the reductions apply to the 9 claimed irreducible words
(8 from above and ).
9.3.1 Algorithm
We present an algorithm that efficiently determines the outcome of a PARTIZAN EUCLID
position.
Step 0. Let S be the signature of E.p; q/. Let S 0 be the empty string.
Step 1. If S is nonempty, remove the first letter of S and add it to the end of S 0 ; go to
Step 2. Otherwise, go to Step 3.
Step 2. If you added l to S 0 , use Rule 1 on the suffix of S 0 if applicable. Go to
Step 1.
If you added r to S 0 , use Rule 2, 3, or 4 on the suffix of S 0 if applicable;
at most one will apply. Go to Step 1.
If you added e to S 0 , use Rule 5, 6, or 7 on the suffix of S 0 if applicable,
at most one will apply. If you applied Rule 5 or 7, then use Rule 1 if
applicable. Go to Step 3.
Step 3. The outcome of E.p; q/ is the outcome of S 0 given in Table 9.1.
Reductions occur at the end of the word and the application of a reduction does
not cause another reduction, except possibly in Step 2: part 3, with an l l l reduction.
As such, Step 2 finishes in constant time (as do Steps 1 and 3). Step 1 takes about as
long as the Euclidean algorithm. Steps 1 and 2 have to be performed at most p times;
Steps 0 and 3 are each performed once.
By Lemma 9.16, S 0 reaches a length of at most 8, as demonstrated by l lrl lrl l.
That is, if at any point the length of S 0 is 9, then it will be irreducible in the next step.
In the algorithm as given above, S is computed in full at the beginning, for ease of
description. However, we can easily modify our algorithm to be an on-line algorithm
by computing the next letter of S as we need it to add to S 0 . In that case, to run the
algorithm we store at most 3 integers no larger than p and a string of length at most 9.
9.4 Outcome Observations

There are several interesting observations that can be made about the outcomes which
may be useful in actual play.
Observation 9.17. If Sg D re then o.g/ 2 L; R.
All signatures in Table 9.1 starting with r are in L or R. The reductions (from
Lemma 9.13) change signatures starting with r to shorter signatures starting with r or
to le, which is in R.
Observation 9.18. Let g D E.p; q/ be a standard position. If o.g/ 2 N ; P , then

2p p
3 q > 2.
If o.g/ 2 N ; P , then Sg is (in which case g is not standard), e (in which case
2p
3 D q), or starts with l. As g is standard, if t D p%q, then t D p q and q > p2 .
For the signature to start with l, we need 2t > q, which is 2.p q/ > q or 2p 3 > q;
combining this with q > p2 gives the result.
Lemma 9.13 can be restated in terms of positions in the game.
Observation 9.19. If a and b are integers with a > b > 0; then

(1) o.E.a C b; a// D o.E.5a C 3b; 3a C 2b//;
(2) if o.E.2a C b; a C b// D P and a > 2b, then E.2a C 3b; a C 2b/ D P .
For (1). Let E.a Cb; a/ D e and consider E.5a C3b; 3a C2b/. First suppose a >
b; the consecutive left options E.3aC2b; 2aCb/, E.2aCb; aCb/, and E.aCb; a/ are
standard and are on the spine so E.5a C 3b; 3a C 2b/ D l l le. By Rule 1, o.E.5a C
3b; 3a C 2b// D o.E.a C b; a//. Now suppose a < b; from E.5a C 3b; 3a C 2b/,
both E.3a C 2b; 2a C b/, E.2a C b; a C b/ are still on the spine, and now E.a C b; b/
is, too. Since E.a C b; a/ D P , then by Theorem 9.6 o.E.a C b; b/L / D P , and thus
o.E.2aCb; aCb// D L. Now o.E.3aC2b; 2aCb/L / D o.E.2aCb; aCb// D L,
and again by Theorem 9.6 o.E.3a C 2b; 2a C b/R / D o.E.a C b; a// D P , thus
o.E.3aC2b; 2aCb// D N . Finally, o.E.5aC3b; 3aC2b/L / D o.E.3aC2b; 2aC
b// D N , and again by Theorem 9.6, o.E.5a C 3b; 3a C 2b/R / D o.E.2a C b; a C
b// D L, and thus o.E.5a C 3b; 3a C 2b// D P .
For (2): Since a > 2b then E.2a C b; a C b/ D lre, where the left option is
E.a C b; a/ and its right option is E.a; a b/, they are both standard and on the
spine. The left option of E.2a C 3b; a C 2b/ is E.a C 2b; a C b/ and its right option is
E.a C b; a/; again both on the spine. Thus E.2a C 3b; a C 2b/ D lrre. Therefore,
from Lemma 9.13, o.E.2a C 3b; a C 2b// D o.E.2a C b; a C b//.
9.5 Open Questions

Our main work is describing the structure of postitions of PARTIZAN EUCLID and giving
an efficient algorithm for determining the outcome. Thus we arrive at one main open
question.
Question 9.20. Is there an efficient method to play disjunctive sums of PARTIZAN

EUCLID positions?
For some families of positions (signatures) we can give the value easily. Obser-
vations 9.21 and 9.22 happen to correspond to the extreme cases of the Euclidean
algorithm.
Observation 9.21. Positions of the form E.k C 1; k/ have value C .k 2/" for
k 2.
The signature r k e corresponds to E.k C 1; k/. When k 2 the Left option is to

E.k; 1/ which is equal to 0, and the Right option is to E.k; k 1/.
Observation 9.22. Let fn be the n-th Fibonacci number, where f0 D 0 and f1 D 1.

The position E.fk ; fk1 / has the signature l k e, and the value is periodic in k;
E.f3k ; f3k1 / D", E.f3kC1 ; f3k / D , and E.f3kC2 ; f3kC1 / D 0. Starting with
E.2; 1/ D 0, E.3; 2/ D , and E.5; 3/ D", an easy induction gives the result.
It seems unlikely that we would find a heuristic method to play a sum; we expect
a solution would require first computing the values of the summands and a method to
play on a sum of such values (see [1] for more on values). Values are harder to calculate
and there are few easy reductions of the signature that allow short cuts. However, we
present a general rule that we have found.
Note from Corollary 9.10 that a Left move from a position whose signature has
at least one l in it removes exactly one l, and a Right move from a position whose
signature has at least two ls in it removes exactly one r or exactly two ls.
Observation 9.23. Let two positions g and h have signatures lr a le and lr b le;
respectively. If r a le D r b le and e D e; then g D h.
To see this, if D , then gL D r a le D r b le D hL and g R D g LL D

e D e D hLL D hR . If D l, then g L D lr a le D lr b le D hL and
g R D e D e D hR . If D r, then gL D r a le D r b le D hL and g R D
lr a le D lr b le D hR .
As there are many nontrivial P positions in PARTIZAN EUCLID, and all P positions
have value 0, we think it is reasonable to expect many other values to occur repeatedly,
perhaps with similar patterns to that of the P positions.
Table 9.2. E.q; p/, q D 11; : : : ; 20, p D 10; : : : ; 15.
q D 11 12 13 14 15 16 17 18 19 20
p D 10 6j2 3=2 2 0 1 0 2 0 0 0
11 7j2 0 3 0 4 3 0 4 0
12 8j2 2j2 1 0 3 1 1 0
13 9j2 0 2 3 0 5 4
14 10j2 3j2 3 0 0 2
15 11j2 0 3=2 0 0
Question 9.24. Which signature reductions preserve value in which instances?
Canonical forms become messy with even small values of p and q. Atomic weights
(see [1]) are an approximation to the value of a position. The atomic weights of Ta-
ble 9.2 were generated by CGSuite [9]. (In version 1.0 of CGSuite, PARTIZAN EUCLID
is used as an example, and tables of both canonical forms and atomic weights can be
easily generated.)
Figure 9.3. Graph showing mean atomic weights of positions against ratio of p to q.
The mean values of atomic weights show some regularity on a large scale; see Fig-
ure 9.3, where the figure is cut off above 25, removing the points corresponding to
E.k C 1; k/ for k > 27.
In addition to the variation of the means of atomic weights of positions there is great
complexity among the atomic weights, which include positions such as 6j9 12 , z3 ,
8*, and 7jj6j5.
Question 9.25. Is there an efficient way to determine the atomic weight of a position
from its signature?
Acknowledgments. The authors wish to thank the NSERC for financial support.
References
[1] M. H. Albert, R. J. Nowakowski, and D. Wolfe, Lessons in Play, A. K. Peters, Ltd.,
Nattick, MA., 2007.
[2] E. R. Berlekamp, J. H. Conway, and R. K. Guy, Winning ways for your mathematical
plays, vol. 1, A. K. Peters, Ltd., 2001.
[3] A. J. Cole and A. J. T. Davie, A game based on the Euclidean algorithm and a winning
strategy for it, Math. Gaz. 53 (1969), 354357.
[4] D. Collins, Variations on a theme of Euclid, Integers, 5 (2005), #G3.
[5] D. Collins and T. Lengyel, The game of 3-Euclid, Discrete Mathematics 308 (2008),
11301136.
[6] J. H. Conway, On Numbers and Games, 2nd edn., A. K. Peters, Ltd., 2001.
[7] T. Lengyel, A Nim-type game and continued fractions, Fibonacci Quart. 41 (2003), 310
320.
[8] O. Perron, Die Lehre von den Kettenbrchen, B.G. Teubner, Leipzig and Berlin, 1913.
[9] A. N. Siegel, Combinatorial Game Suite, http://cgsuite.sourceforge.net/, 2000, A soft-
ware tool for investigating games, last accessed 6/2013.
[10] W. Stromquist and D. Ullman, Sequential compounds of combinatorial games, Theoret.
Comput. Sci. 119 (1993), 311321.
Author information
Neil A. McKay, Department of Mathematics and Statistics, Dalhousie University, Halifax,
Nova Scotia, Canada.
Email: nmckay@mathstat.dal.ca
Richard J. Nowakowski, Department of Mathematics and Statistics, Dalhousie University,

Halifax, Nova Scotia, Canada.
Email: rjn@mathstat.dal.ca
Chapter 10
Lecture Hall Partitions and the Wreath Products

Ck o Sn
Thomas W. Pensyl and Carla D. Savage
Abstract. It is shown that statistics on the wreath product groups, Ck oSn , can be interpreted in
terms of natural statistics on lecture hall partitions. Lecture hall theory is applied to prove dis-
tribution results for statistics on Ck o Sn . Finally, some new statistics on Ck o Sn are introduced,
inspired by lecture hall theory, and their distributions are derived.
Keywords. Partitions, Lecture Hall, Wreath Product Group.
Mathematics Subject Classification 2010. 05A17, 52B11.
10.1 Introduction
The purpose of this note is to show that statistics on the wreath product Ck o Sn of a
cyclic group Ck , of order k, and the symmetric group Sn , can be interpreted in terms
of natural statistics on lecture hall partitions. We demonstrate that lecture hall theory
can be used to prove results about the distribution of statistics on Ck oSn . We introduce
some new statistics on Ck oSn , inspired by lecture hall partitions, including a quadratic
version of flag-major index, and prove distribution results for these statistics.
This chapter is organized as follows. In Section 10.2, we define the s-lecture hall par-
titions and state a few useful results. Section 10.3 is devoted to statistics of interest on
the wreath product groups and a very brief discussion of what is known. Section 10.4
introduces s-inversion sequences, which will be used to relate statistics on Ck o Sn to
statistics on lecture hall partitions.
Section 10.5 describes a bijection between .k; 2k; : : : ; nk/-inversion sequences and
Ck o Sn that allows statistics to be translated from one domain to another.
Section 10.6 reviews recent work of Savage and Schuster [13] relating inversion
sequences to lecture hall partitions. This work was developed with the intention of
extending work on permutation statistics to a more general setting.
Section 10.7 is the heart of the chapter. We prove there a theorem which allows us
to apply the tools of Section 10.6 to Ck o Sn . This contains our main results relating
statistics such as descent, flag-major index and flag-inversion number to statistics on
lecture hall partitions, also proving an EulerMahonian distribution result.
In Section 10.8 we define a new statistic lhall on Ck oSn and derive its surprisingly
nice distribution.
In Section 10.9, we are led to define a distorted version of the descent statistic on
Ck o Sn , that reveals an even closer connection to lecture hall partitions.
A few words about notation: Z is the set of integers, R the set of real numbers, Sn
the set of permutations of n elements; j D 1; 2; : : : ; j , where 0 D ;; n q D
.1 q n /=.1 q/; and for x D .x1 ; x2 ; : : : ; xn /, jxj D x1 C x2 C C xn .
10.2 Lecture Hall Partitions
For a sequence s D si i1 of positive integers, the s-lecture hall partitions are the
elements of the set

.s/ n 1 2 n
Ln D 2 Z 0 :
s1 s2 sn
.1;2;3;4/
For example, .0; 1; 3; 4/ 2 Ln , but .0; 1; 3; 4/ 62 L.1;3;5;7/
n , since 3=5 > 4=7.
.1;2;:::;n/
The original lecture hall partitions Ln D Ln were introduced by Bousquet-
Mlou and Eriksson in [3], where they showed that
X Y
n
1
y jj D : (10.1)
1 y 2i1
2Ln iD1
In [4] they proved the following refinement, which will be useful in the present work.
Theorem 10.1 (Refined lecture hall theorem [4]). For any nonnegative integer n,
X Y
n
1 C qy i
jdej jj
q y D ; (10.2)
1 q 2 y nCi
2Ln iD1
where de D .d1 =1e ; d2 =2e ; : : : ; dn =ne/.
If the largest part in a lecture hall partition in Ln is constrained, we have the follow-
ing:
Theorem 10.2 ([8, 13]). For integers n 1 and t 0,

X
q jdej D t C 1 nq : (10.3)
2Ln I n tn
Chapter 10 Lecture Hall Partitions and the Wreath Products Ck o Sn 139
For example, when n D 3 and t D 1, the set 2 L3 j 3 3 has the eight

elements

.0; 0; 0/; .0; 0; 1/; .0; 0; 2/; .0; 0; 3/; .0; 1; 2/; .0; 1; 3/; .0; 2; 3/; .1; 2; 3/
and X
q d1 =1eCd2 =2eCd3 =3e D 1 C 3q C 3q 2 C q 3 D 2 3q :
2L3 I 3 3
10.3 Statistics on Ck o Sn
An element 2 Sn is a bijection W n ! n , and we write D .1 ; : : : ; n /,
to mean that .i / D i . A descent in 2 Sn is a position i 2 n 1 such that
i > iC1 . The set of all descents of is Des and des D jDes j. The inversion
number of is

inv D .i; j / j 1 i < j n and i > j :
For example, if D .5; 4; 1; 3; 2/, then Des D 1; 2; 4, des D 3 and inv D 8.

For positive integers k and n, we view Ck o Sn combinatorially as a set of pairs
.; /:
Ck o Sn D .; / j 2 Sn ; 2 0; 1; : : : ; k 1n :
We use the notation to denote .; / and write
D .1 1 ; 2 2 ; : : : ; n n / D ..1 ; : : : ; n /; .1 ; : : : ; n // D .; /:
Statistics on Ck o Sn (or k-colored permutations or k-indexed permutations) have

been studied by many, starting with Reiners work on signed permutations [12], fol-
lowed by independent work of Brenti [5] and Steingrmsson [14] on the more general
wreath products. Pairs of (descent, major index) statistics have been found, satisfy-
ing relations like Carlitzs q-Eulerian polynomials, starting with work of Adin, Brenti,
and Roichman [1]. There have very recently been many new and exciting discoveries,
including [2, 7, 9, 10]. It is remarkable how many variations in the definitions of the
statistics there are, even when they give the same distribution.
We start with a fairly standard definition of descent. The descent set of 2 Ck oSn
is

Des D i 2 0; 1; : : : ; n 1 j i < iC1 ; or i D iC1 and i > iC1 ;
(10.4)
with the convention that 0 D 0 D 0.

We will consider the following statistics defined on Ck o Sn .
des D jDes j
X
comaj D .n i /
i2Des
X
n
fmaj D k comaj i
iD1
X
n
finv D inv C i i :
iD1
As an example, for D .51 ; 41 ; 10 ; 30 ; 22 / 2 C3 o S5 , we have Des D 0; 1; 4;

des D 3; comaj D 10; fmaj D 26; and finv D 21. Note that this
definition of fmaj differs a bit from those appearing elsewhere, even among those
who define the descent set as in Equation(10.4) [1, 7].
Using lecture hall theory, we will show, among other things,
X X
q fmaj D q finv ; (10.5)
2Ck oSn 2Ck oSn
P fmaj x des
X 2Ck oSn q
kt C 1 nq x t D Qn ki
; (10.6)
t0 iD0 .1 xq /
P fmaj x des
X 2Ck oSn q
jdej dn =.k n/e
q x D Qn ki
: (10.7)
2Ln iD1 .1 xq /
Relations of the form (10.6), for general k, have been found only recently, starting with
Chow and Mansour [7] and Hyatt [10], sometimes with slightly different definitions
of Des or fmaj . Our intention here is to highlight our methods, which are quite novel,
and which allow us to prove new results like Equation (10.7).
10.4 Statistics on s-Inversion Sequences

The connection between statistics on Ck o Sn and statistics on lecture hall partitions
will be made via statistics on inversion sequences.
Given a sequence s D si i1 of positive integers, and positive integer n, the set
.s/
In of s-inversion sequences is defined by

I.s/
n D .e1 ; : : : ; en / 2 Zn j 0 ei < si for 1 i n :
The familiar inversion sequences associated with permutations are the elements of
.s/
In for s D .1; 2; : : : ; n/.
The ascent set of an inversion sequence e 2 I.s/

n is the set
e
i eiC1
Asc e D i 2 0; 1; : : : ; n 1 < ;
si siC1
with the convention that e0 D 0 and s0 D 1. For example, as an element of

.3;6;9;12;15/
I5 , the inversion sequence e D .1; 3; 2; 2; 13/ has the ascent set Asc e D
0; 1; 4.
.s/
The following statistics on In were defined in [13]:

asc e D Asc e ;
X
amaj e D .n i /;
i2Asc e
Xn
jej D ei ;
iD1
X
lhp e D jej C .siC1 C ::: C sn / :
i2Asc e
.3;6;9;12;15/
For e D .1; 3; 2; 2; 13/ 2 I5 , we have asc e D 3, amaj e D 10, jej D 21,
and lhp e D 81.
In this chapter, our focus is the sequence s D .k; 2k; : : : ; nk/, where k is a positive
.k;2k;:::;nk/
integer. Let In;k D In . We will require two new statistics on In;k :
n
X
ej
N.e/ D ;
j
j D1
Ifmaj e D k amaj e N.e/:
.3;6;9;12;15/
For e D .1; 3; 2; 2; 13/ 2 I5 , N.e/ D 4 and Ifmaj e D 26.
10.5 From Statistics on Ck o Sn to Statistics on In;k

We will make use of the following bijection between Sn and In;1 which was proved
in [13] to have the required properties.
Lemma 10.3. For positive integer n, the mapping W Sn ! In;1 defined by ./ D
t D .t1 ; t2 ; : : : ; tn /, where

ti D j 2 i 1 j j > i
is a bijection satisfying both Des D Asc t and inv D jt j.

For example, if D .5; 4; 1; 3; 2/, then t D ./ D .0; 1; 2; 2; 3/ 2 I5;1 . Checking

the statistics, Des D 1; 2; 4 D Asc t and inv D 8 D jt j.
Noting that, as sets, In;k and Ck o Sn have the same cardinality, we set up a bijection
which translates statistics from one domain to the other in a useful way.
Theorem 10.4. For each pair of integers .n; k/ with n 1, k 1, there is a bijection
W Ck o Sn ! In;k
with the following properties. If . / D e D .e1 ; : : : ; en / then

Asc e D Des ; (10.8)
X
n
N.e/ D i ; (10.9)
iD1
Ifmaj e D fmaj ; (10.10)
en D n.n C 1/ n ; (10.11)
X
n
jej D inv C i i D finv : (10.12)
iD1
Proof. Define by
e D .1 1 ; 2 2 ; : : : ; n n / D .1 C t1 ; 22 C t2 ; : : : ; nn C tn /;
where .t1 ; t2 ; : : : ; tn / D ./, as in Lemma 10.3.

For example, for D .51 ; 41 ; 10 ; 30 ; 22 / 2 C3 o S5 , t D .5; 4; 1; 3; 2/ D
.0; 1; 2; 2; 3/, so we get e D . / D .1; 3; 2; 2; 13/. Note that properties (10.8)
through (10.12) hold for this example:
Asc e D 0; 1; 4 D Des ;
N.e/ D 4 D 1 C 1 C 0 C 0 C 4 D jj;
Ifmaj e D 26 D fmaj ;
e5 D 13 D 5.5 C 1/ 5 ;
jej D 21 D finv :
Clearly, . / 2 In;k . Since Ck o Sn and In;k have the same cardinality, to show that
is a bijection, it suffices to show that is onto. Let e D .e1 ; : : : ; en / 2 In;k . Define
D .1 ; : : : n / by i D bei =i c. Then 2 0; 1; : : : ; k 1n . Define t D .t1 ; : : : tn /
by ti D ei i i . Then t 2 In;1 . Finally, let D 1 .t / 2 Sn . Then 2 Ck o Sn
and 1 .e/ D .
To prove properties (10.8) through (10.12), observe first that tn D n n , so prop-
erty (10.11) holds. It is clear from the definition of that Equation (10.12) is true.
Also, note that bei =i c D i since 0 ti < i and property (10.9) holds. So property
(10.10) will follow once we prove Equation (10.8). By Lemma 10.3, since t D ./,
we know that Asc t D Des , so it remains to show Asc e D Des .
Note first that e1 D 1 C t1 D 1 , since t1 D 0. Thus,
0 2 Des 1 > 0 e1 > 0 0 2 Asc e:
For 1 i n, i 2 Asc e if and only if

eiC1 ei .i C 1/iC1 C tiC1 i i C ti
0< D
k.i C 1/ ki k.i C 1/ ki
i.i C 1/.iC1 i / C i tiC1 .i C 1/ti
D
ki.i C 1/
i
D ;
ki.i C 1/
where
i D i.i C 1/.iC1 i / C i tiC1 .i C 1/ti :
Therefore, i 2 Asc e if and only if i > 0.
If i D iC1 , then
i > 0 i tiC1 .i C 1/ti > 0 i 2 Asc t
i 2 Des i 2 Des :
For the remaining cases, note that since 0 tiC1 i and 0 ti i 1,
i.i C 1/.iC1 i / i 2 C 1 i i.i C 1/.iC1 i / C i 2 :
If i 6D iC1 , then i 2 Des if and only if i < iC1 . But if i < iC1 , then
i i.i C 1/ i 2 C 1 D i C 1 > 0;
and so i 2 Asc e. And, if i > iC1 , then
i i.i C 1/ C i 2 D i 0
and i 62 Asc e. This completes the proof.
10.6 Lecture Hall Polytopes and s-Inversion Sequences

The s-lecture hall polytope was introduced in [13], for an arbitrary sequence s D
si i1 of positive integers, as

.s/ n 1 2 n
Pn D 2 R 0 1 :
s1 s2 sn
P.s/
n is a convex, simplicial polytope with the n C 1 vertices:
.0; 0; : : : ; 0/; .s1 ; s2 ; : : : ; sn /; .0; s2 ; : : : ; sn /; .0; 0; s3 ; : : : ; sn /; : : : ; .0; 0; : : : ; 0; sn /;

.s/
all with integer coordinates. The t -th dilation of Pn is the polytope

n 1 2 n
t P.s/
n D 2 R 0 t :
s1 s2 sn
A multivariate function, fn.s/ .t I q; y; z/, was used in [13] to enumerate lattice points in
.s/
t Pn according to statistics significant in the theory of lecture hall partitions:
X C
fn.s/ .t I q; y; z/ D q jdes j y jj z j ./ j ;
2tP.s/
n \Z
n
where

1 2 n
des D ; ;:::; ; (10.13)
s1 s2 sn

1 2 n
C ./ D s1 1 ; s2 2 ; : : : ; sn n : (10.14)
s1 s2 sn
The following theorems show the connection between statistics on s-inversion se-
quences and statistics on s-lecture hall partitions.
Theorem 10.5 ([13]). For any sequence s of positive integers, and any positive inte-
ger n,
P
X e2I.s/
x asc e q amaj e y lhp e z j e j
.s/ t
fn .t I q; y; z/ x D Qn n
ni y si C1 CCsn /
: (10.15)
t0 iD0 .1 xq
Theorem 10.6 ([13]). For any sequence s of positive integers, and any positive inte-
ger n,
P
X e2I.s/
x asc e q amaj e y lhp e z j e j
jdejs jj j C ./ j dn =sn e
q y z x D Qn1n : (10.16)
ni y si C1 CCsn /
.s/
2Ln iD0 .1 xq
10.7 Lecture Hall Partitions and the Inversion

Sequences In;k
In order to apply the results of the previous section to the problem of interest, we need
an analog of Ifmaj on In;k for lecture hall partitions.
First, observe that the following sets of lecture hall partitions are all the same:
Ln D L.1;2;:::;n/
n D L.2;4;:::;2n/
n D L.3;6;:::;3n/
n D :
However, the lecture hall polytopes Pn;k defined by

1 2 n
Pn;k D 2 Rn 0 1
k 2k nk
are different for different k. On the other hand, the following dilations are the same:
t Pn;k D k t Pn;1 ; (10.17)
a fact which we will exploit. Furthermore,
k t Pn;1 \ Zn D 2 Ln j n k t n:
Since the definitions (10.13) and (10.14) depend on the sequence s D .k; 2k;
: : : ; nk/, we will make the dependence explicit in the notation. For 2 Ln and k 1,
let

1 2 n
dek D ; ;:::; ; (10.18)
k 2k nk

1 2 n
kC ./ D k 1 ; 2k 2 ; : : : ; nk n ; (10.19)
k 2k nk
k ./ D k dek de : (10.20)
Note that for 2 Ln ,
de1 D de ;
where de was defined in Theorem 10.1.
We now show that the new statistic k on Ln corresponds to the statistic N on In;k .
Theorem 10.7. For positive integers n; k, let

X C
fn;k .t I q; y; z; w/ D q jdek j y jj z j k ./j w j k ./ j : (10.21)
2tPn;k \ Zn
Then,
P
X
t e2In;k x asc e q amaj e y lhp e z j e j w N.e/
fn;k .t I q; y; z; w/ x D Qn : (10.22)
t0 iD0 .1 xq ni y k.n.nC1/i.iC1//=2 /
Proof. If w D 1, this is just the case s D .k; 2k; : : : ; nk/ of Theorem 10.5. To include
w, we appeal to the combinatorial proof of (10.15) in Theorem 10.5 which was pre-
sented in [13]. In that proof, 2 .t Pn;k \Zn / is associated with the inversion sequence
kC ./, which, by definition, is in In;k . It suffices to check that jk ./j D N.kC .//:
Xn
C i k di =.i k/e i
N.k .// D
i
iD1
Xn
D bk di =.i k/e i =i c
iD1
Xn
D .k di =.i k/e di =i e/
iD1
D jk dek de1 j D jk ./j:
The Ifmaj statistic is obtained by setting q D q k and w D q 1 in Theorem 10.7.
Corollary 10.8. For positive integers n; k,

X X C
q jdej y jj z j k ./j x t D
t0 2ktPn;1 \ Zn P
e2In;k x asc e q Ifmaj e y lhp e z j e j
Qn : (10.23)
iD0 .1 xq k.ni/ y k.n.nC1/i.iC1//=2 /
Proof. With q D q k and w D q 1 , the numerator in the right-hand side of Equation

(10.22) becomes
x asc e q k amaj eN.e/ y lhp e z j e j D x asc e q Ifmaj e y lhp e z j e j :
From (10.21), the left-hand side summand of Equation (10.22) becomes
X C
fn;k .t I q k ; y; z; q 1 / D q kjdek jj k ./ j y jj z j k ./j
2tPn;k \ Zn
X C
D q jdej y jj z j k ./j
2ktPn;1 \ Zn
by (10.18)(10.20) and by (10.17).

P
X x asc e q Ifmaj e y lhp e z j e j
jdej jj j kC ./j dn =.nk/e e2In;k
q y z x D Qn1 :
2Ln iD0 .1 xq k.ni/ y k.n.nC1/i.iC1//=2 /
(10.24)
P C
jde1 j y jj z j k ./j from Equation
Proof. For t > 0, let H.t / D 2ktPn;1 \ Zn q
(10.23), with H.0/ D 1. Then for t > 0, since

n n n
2 Ln I D t D 2 Ln I t 2 Ln I t 1
nk nk nk

D k t Pn;1 \ Zn k.t 1/Pn;1 \ Zn ;
we have
X C X X C
q jdej y jj z j k ./j x dn =.nk/e D xt q jdej y jj z j k ./j
l m
2Ln t0 n
2Ln I nk Dt
X
D1C .H.t / H.t 1//x t
t1
X X
D1C H.t /x t H.t 1/x t
t1 t1
X X
t
D H.t /x x H.t /x t
t0 t0
X
D .1 x/ H.t /x t :
t0
P
But t0 H.t /x t is the left-hand side of Equation (10.23), so we simply multiply the
right-hand side of Equation (10.23) by .1 x/ to complete the proof.
We can now apply these results to the wreath product groups. First, we have the
expected result that the pair .des ; fmaj / is EulerMahonian.
Theorem 10.10. For positive integers n; k,

P fmaj x des
X 2Ck oSn q
n t
k t C 1 q x D Qn ki
:
t0 iD0 .1 xq /
Proof. Set y D z D 1 in Equation (10.23). On the left-hand side, in the summand, we

get X
q jdej :
2ktPn;1 \ Zn
Since k t Pn;1 \ Zn D 2 Ln j n k t n, by Theorem 10.2,

X
q jdej D k t C 1 nq :
2ktPn;1 \ Zn
For the right-hand side, we get

P
e2In;k x asc e q Ifmaj e
Qn :
iD0 .1 xq k.ni/ /
Reindex the product in the denominator and for the numerator, using the fact that by
Theorem 10.4, the distribution of .des ; fmaj / on Ck oSn is the same as the distribution
of .asc ; Ifmaj / on In;k .
Now, to interpret the distribution .des ; fmaj ; finv / on Ck o Sn in terms of lecture

hall partitions, set y D 1 in Equation (10.24) and use Theorem 10.4.

P fmaj x des z finv
X 2Ck oSn q
jdej j kC ./j dn =.nk/e
q z x D Qn ki
:
2Ln iD1 .1 xq /
The implication of Theorem 10.11 for z D 1 is quite interesting. We have

P fmaj x des
X 2Ck oSn q
jdej dn =.nk/e
q x D Qn ki
: (10.25)
2Ln iD1 .1 xq /
In the left-hand side of equaqtion (10.25), the only dependence on k is in the exponent
of x, in a statistic involving only the last part of . We take this further in Section 10.9.
10.8 A Lecture Hall Statistic on Ck o Sn

From the point of view of partition theory, the most important statistic for a lecture
hall partition is the number jj D 1 C C n being partitioned. So, what does
jj correspond to on Ck o Sn ?
In [6], a quadratic version of the major index was defined on Sn by bin D
P iC1
i2Des 2 : In that spirit, we define cobin on Ck o Sn by
! !!
X nC1 i C1

cobin D :

2 2
i2Des
Now define the statistic lhall on Ck o Sn by

lhall D k cobin finv :
Observe that under the bijection of Theorem 10.4, if e D . / then lhall D
lhp e. This can be seen as follows, since jej D finv and Asc e D Des e:
X
lhp e D jej C .k.i C 1/ C C k n/
i2Asc e
! !!
X nC1 i C1
D jej C k
2 2
i2Asc e
! !!
X nC1 i C1

D finv C k
2 2
i2Des e
D finv C kcobin

D lhall :
The joint distribution of .lhall ; fmaj / on Ck o Sn has the following form.

X Y
n
.1 C qy i /.1 q k.nC1i/ y k.iCCn/
lhall fmaj
y q D
1 q 2 y nCi
2Ck oSn iD1
Y
dn=2e Y
bn=2c
D k.2i 1/ qy nC1i . 2 qy i ki q 2 y 2.ni /C1 /
iD1 iD1
Proof. Under the bijection of Theorem 10.4, if e D . /, then lhall D lhp e

and fmaj D Ifmaj e. Therefore,
X X
y lhall q fmaj D y lhp e q Ifmaj e :
2Ck oSn e2In;k
So, by Corollary 10.9 with x D z D 1,

P lhall q fmaj
P lhp e q Ifmaj e
2Ck oSn y e2In;k y
Qn1 k.ni/ y k.n.nC1/i.iC1//=2 /
D Qn1
k.ni/ y k.n.nC1/i.iC1//=2 /
iD0 .1 q iD0 .1 q
X
D y jj q jdej :
2Ln
Now apply Theorem 10.1 to get

P lhall q fmaj
2Ck oSn y
Yn
1 C qy i
Qn1 D :
iD0 .1 q
k.ni/ y k.n.nC1/i.iC1//=2 / 1 q 2 y nCi
iD1
So,
X Y
n Y
n
1 C qy i
lhall fmaj k.niC1/ k.n.nC1/i.iC1//=2
y q D .1q y / ;
1 q 2 y nCi
2Ck oSn iD1 iD1
which, after simplification, gives the theorem.
Setting y D 1 in Theorem 10.12 and simplifying, we get

X Y
n
fmaj
q D ki q ;
2Ck oSn iD1
the same distribution as finv , Ifmaj , and jej, as expected. But the statistic lhall itself
also has a surprisingly simple distribution:
X Y
n
lhall
q D ki q 2.ni /C1 :
2Ck oSn iD1
Proof. Set q D 1 and y D q in the proof of the Theorem 10.12, but apply Equation
(10.1) instead of (10.2) to get
X Y
n
1 q k.iCCn/
q lhp e D
.1;2;:::;n/
1 q 2i1
e2In iD1
Y
bn=2c
1 q k.2i1/.niC1/ Y 1 q ki.2.ni/C1/
dn=2e
D
1 q 2i1 1 q 2.ni/C1
iD1 iD1
Y
n
D ki q 2.ni /C1 :
iD1
10.9 Inflated Eulerian Polynomials for Ck o Sn

We showed in [11] how to obtain more refined information about the s-lecture hall
.s/
partitions by considering the rational lecture hall polytope Rn :

.s/ n 1 2 n
Rn D 2 R 0 and n 1 :
s1 s2 sn
R.s/
n is a convex simplicial polytope, whose vertices are

s1 s2 sn s2 sn
.0; 0; : : : ; 0/ ; ; ;:::; ; 0; ; : : : ; ;
sn sn sn sn sn

s3 sn sn
0; 0; ; : : : ; ; : : : ; 0; 0; : : : ; 0; ;
sn sn sn
with rational (but not necessarily integer) coordinates. Let
X C
gn.s/ .t I q; y; z/ D q jdejk y jj z j k ./ j : (10.26)
2tR.s/
n \Z
n
The following theorems were proved in [11]. These are analogs of Theorems 10.5
and 10.6.
Theorem 10.14 ([11]). For any sequence s of positive integers, and positive integer n,
P
X e2I.s/
q amaj e y lhp e z j e j x sn asc een
.s/ t
gn .t I q; y; z/x D n
Q :
t0 .1 x/ n1 iD0 .1 x q
sn ni y si C1 CCsn /
Theorem 10.15 ([11]). For any sequence s of positive integers, and positive integer n,
P
X e2I.s/
q amaj e y lhp e z j e j x sn asc een
jdejs jj j sC ./ j n n
q y z x D Qn1 :
sn ni y si C1 CCsn /
.s/
2Ln iD0 .1 x q
We can specialize Theorems 10.14 and 10.15 to s D .k; 2k; : : : ; nk/ and modify
to track Ifmaj as in Theorem 10.7 and its corollaries. We should expect something
interesting because
R.1;2;:::;n/
n D R.2;4;:::;2n/
n D R.3;6;:::;3n/
n :
We get the following theorem, which is an analog of Theorem 10.7. The proof, which
is analog to that of Theorem 10.6, is omitted.
Theorem 10.16. For positive integers n; k, let

X C
gn;k .t I q; y; z; w/ D q jdek j y jj z j k ./j w j k ./ j : (10.27)
2tRn \ Zn
Then
P k nasc een q amaj e y lhp e z j e j w N.e/
X e2In;k x
t
gn;k .t I q; y; z; w/ x D Qn1 :
t0 .1 x/ iD0 .1 x k n q ni y k.n.nC1/i.iC1//=2 /
(10.28)
The following corollaries of Theorem 10.16 are analogs of Corollaries 10.8 and 10.9
with y D z D 1. Note that in the left-hand sides of the equations there is no depen-
dence on k.

P
X X e2In;k x
k nasc een q Ifmaj e
jdej t
q x D Q :
t0 2tR \ Zn .1 x/ n1iD0 .1 x q
k n k.ni/ /
n

P
X e2In;k x
k nasc een q Ifmaj e
jdej n
q x D Qn1 k n k.ni/ /
:
2Ln iD0 .1 x q
Making use of Theorem 10.4 giving the correspondence between statistics on In;k
and on Ck o Sn , we have the following analogs of Theorems 10.10 and 10.11. First,
We need a result from [8].
Lemma 10.19 ([8]). For integers t 0 and n > 0, let j and i be the unique integers
satisfying t D j n C i , where j 0 and 0 i < n. Then,
X
q jdej D j C 1 ni
q j C 2 iq :
2tRn \Zn

P fmaj x n.k des 1 n /Cn
X
X n1 2Ck oSn q
ni i nj Ci
j C 1 q j C 2 q x D Q :
j 0 iD0
.1 x/ niD1 .1 x k n q ki /
Proof. By Lemma 10.19,
X n1
X X
X n1 X
j C 1 ni
q j C 2 iq x nj Ci D q jdej x j nCi :
j 0 iD0 j 0 iD0 2.j nCi/Rn \ Zn
Since every t 0 can be written uniquely as t D j n C i for nonnegative integers j

and i with i < n, the last expression can be rewritten as
X X
q jdej x t ;
t0 2tRn \ Zn
which, by Corollary 10.17, is equal to

P k nasc een q Ifmaj e
e2In;k x
Qn k n k.ni/ /
:
iD0 .1 x q
Under the bijection of Theorem 10.4, if e D . /, then Ifmaj e D fmaj ,

asc e D des , and en D n.n C 1/ n . The result follows then, since
k n asc e en D k n des n.n C 1/ C n :

Theorem 10.21. For any positive integers n; k,

P fmaj x n.k des 1 n /Cn
X 2Ck oSn q
jdej n
q x D Qn k n ki
:
2Ln iD1 .1 x q /
Proof. Start from Corollary 10.18 and apply Theorem 10.4.
(Note: there is no dependence on k in the left-hand side).

Let Qn;k .x/ be the q D 1 specialization:
X 1
Qn;k .x/ D x n.k des n /Cn
:
2C k oSn
The Qn;k .x/ are referred to as inflated Eulerian polynomials in [11]. To contrast the
usual, Eulerian polynomials for Ck o Sn are
X
En;k .x/ D x des :
2Ck oSn
It is interesting that Qn;k .x/ is self-reciprocal, but in general En;k .x/ is not when
k > 2.
10.10 Concluding Remarks

It is interesting from the results in Sections 10.7 10.9 that for fixed n, statistics on
Ck o Sn such as descent, flag-major index, and flag-inversion number appear naturally
in the geometry of the same simplicial cone, Rn , independent of k.
It would be interesting to see to what extent other statistics on Ck o Sn can be inter-
preted in terms of lecture hall partitions. Different orderings on Ck o Sn and different
bijections Ck o Sn ! In;k would give different results.
Lecture hall partitions were discovered in the setting of affine Coxeter groups, and
Theorem 10.1 was inspired by Botts formula. It should be possible to trace through
backwards to discover the algebraic significance of the statistic lhall , at least in the
Coxeter groups An D C1 o Sn or Bn D C2 o Sn but we have not seen how to do this.
Acknowledgments. The second author would like to thank Matthias Beck, Benjamin
Braun, and Mattias Koeppe for discussions on Ehrhart theory and signed permutations.
Thanks also to the American Institute of Mathematics, where some of those discus-
sions were hosted. We are grateful to the referee for a careful reading of the manuscript
and for many helpful suggestions to improve the presentation.
References
[1] Ron M. Adin, F. Brenti, and Y. Roichman, Descent numbers and major indices for the
hyperoctahedral group, Adv. in Appl. Math. 27(23) (2001) 210224, special issue in
honor of Dominique Foatas 65th birthday (Philadelphia, PA, 2000).
[2] R. Biagioli and J. Zeng, Enumerating wreath products via GarsiaGessel bijections, Eu-
ropean J. Combin. 32(4) (2011) 538553.
[3] M. Bousquet-Mlou and K. Eriksson, Lecture hall partitions. Ramanujan J. 1(1) (1997),
101111.
[4] M. Bousquet-Mlou and K. Eriksson, A refinement of the lecture hall theorem, J. Com-
bin. Theory Ser. A 86(1) (1999), 6384.
[5] F. Brenti. q-Eulerian polynomials arising from Coxeter groups, Eur. J. Comb. 15 (1994),
417441.
[6] K. L. Bright and C. D. Savage, The geometry of lecture hall partitions and quadratic
permutation statistics, in: DMTCS Proceedings, 22nd International Conference on For-
mal Power Series and Algebraic Combinatorics (FPSAC 2010), San Francisco, (AN),
pp. 569, 580, 2010.
[7] C.-O. Chow and T.Mansour, A Carlitz identity for the wreath product Cr o Sn , Adv. in
Appl. Math. 47(2) (2011), 199215.
[8] S. Corteel, S. Lee, and C. D. Savage, Enumeration of sequences constrained by the ratio
of consecutive parts, Sm. Lothar. Combin. 54A (2005/07), Art. B54Aa, (electronic).
[9] H. L. M. Faliharimalala and A. Randrianarivony, Flag-major index and flag-inversion
number on colored words and wreath product, Sm. Lothar. Combin. 62 (2009/10), Art.
B62c, 10.
[10] M. Hyatt, Quasisymmetric functions and permutation statistics for Coxeter groups and
wreath product groups, . Ph.D. Thesis, University of Miami, 2011.
[11] T. W. Pensyl and C. D. Savage, Rational lecture hall polytopes and inflated Eulerian poly-
nomials, Ramanujan J. (2012), DOI 10.1007/s11139-012-9393-7, to appear.
[12] V. Reiner, Signed permutation statistics, European J. Combin. 14(6) (1993), 553567.
[13] C. D. Savage and M. J. Schuster, Ehrhart series of lecture hall polytopes and Eulerian
polynomials for inversion sequences, J. Combin. Theory Ser. A 119 (2012), 850870.
[14] E. Steingrmsson, Permutation statistics of indexed permutations, European J. Combin.
15(2) (1994), 187205.
Author information
Thomas W. Pensyl, Department of Computer Science, North Carolina State University,
Raleigh, North Carolina, USA.
Email: twpensyl@ncsu.edu
Carla D. Savage, Department of Computer Science, North Carolina State University,

Raleigh, North Carolina, USA.
Email: savage@ncsu.edu
Index
admissible set, 39 polynomial time computable, 101

affine Coxeter Groups, 153 Sprague-Grundy, 106, 107
alternating game, 1 superadditive, 91, 92, 99, 100, 101,
AP -set, 19 102, 103, 106
atomic weight, 135 function fields, 7374
Bhargava, M. 5862, 6770

Botts formula, 153 g; , 15
bounded Wythoff game, 93 game tree, 124, 125
Bush, Albert, 52 geometry of numbers, 6770
Green, B., 53,
C -set, 18
C -set, 18
Central Sets Theorem, 17 h , 19
central, 16 Heegner points, 7073
central*, 16 Hermites theorem, 5657
circle group, 25 Hirzebruch surfaces, 7374
compositions, 79 Hough, B., 7073
Conjecture H, 38, 39, 42
conjugate, 3 ideal class groups, 5960, 7073
cubic forms, 6061 ideals, 15
preserved, 33
-set, 16
-set, 16 if and only if, 28
Diananda, P. H., 53 impartial games, 92, 106, 107
difference set, 48 inclusion-exclusion, 37, 39, 40
dilation, 144 indistinguishable, 3
distinguishable, 3 inflated Eulerian polynomials, 150, 153
intractable, 92
Erdos, P., 53 IP , 15
Euler-Mahonian, 147 irreducible
Eulerian polynomials, 153 signature, 131
word, 130
f , 25
farthest continued fraction, 123
Fibonacci numbers, 80, 109, 123 J -set, 18
function, preserved, 27
constant, 91, 93, 98, 106 J -set, 18
convex, 103, 104 J.S /, 18
mex, 92, 96 ideal, 28
polynomial 91, 100, 103, 106 jumping champions, 38
156 Index
k-colored permutations, 139 prime k-tuple conjecture, 37, 38, 39

k-indexed permutations, 139 prime number theorem, 37, 39
Kedlaya, K. S., 53 product-free, 45
lattice points, 144 Pryby, Chris 52
lecture hall
polytopes, 143 rational lecture hall polytope, 150
statistics on wreath products (cobin, really complicated proof for simple facts,
lhall), 148 58
left-handed game, 2, 5 reduced signature, 128
right-handed game, 2, 5
MacMahon 79, 86 Ruzsa, I. Z., 52, 53
Maroni invariant, 7374
Mertens theorem, 52 s-inversion sequences, 140
Minkowskis theorem, 5657 s-lecture hall partitions, 138
misre S C -set, 20
equivalence, 3 Schinzel, A., 53
inequality, 3 self-reciprocal, 153
play, 91 Shankar, A., 6770
multiplicative Shintani zeta functions, 6367
independence, 111 signature, 127, 128
order, 109 singular series, 38
average, 41
n-colored compositions, 79 Solymosi, J., 53
N -position, 92, 98, 99, 103, 106 spectrum, 15
nearest continued fraction, 123 spine, 127
nonhomogeneous, 15 spotted tilings, 80
normal play, 91, 92 statistics
number fields, counting, 5659 on lecture hall partitions, 145
numeration system, 93, 94, 95, 97, 98, on permutations (Des, des, inv), 139
101, 107 on s-inversion sequences (Asc, asc,
amaj, lhp, N, fmaj), 141
one-handed game, 2, 6, 7 on wreath products (Des), 139
outcome classes, 2, 129 on wreath products (des, comaj, fmaj,
outcomes, 10 finv), 140
Stirlings formula, 42
paired outcomes, 129 strongly central, 26
P -position, 92, 93, 95, 96, 98, 99, 100, succinct, 101
101, 103, 105, 106 sum-free, 45
PENNY NIM, 2, 5 sum-product problem, 45
periodic, periodicity, 93, 99 syndetic, 18, 23
additive, 99, 106, 107 Szemerdi, E., 53
arithmetic, 99
piecewise syndetic, 18 take-away games, 91, 92
Poisson Taniguchi, T., 6367
density, 38 Taylors theorem, 41
distribution, 37 Terquems problem, 80
prehomogeneous vector space, 5859 thick, 20, 23
preserved, 27, 32 time complexity, 101
Index 157
tractable, 92, 93, 98, 101, 103, 105, 106 Wythoffs game, 91, 92, 93, 101, 106,
Tsimerman, J., 6770 107
two-handed game, 2
Yap, H. P., 53
upper asymptotic density, 46
Vandehey, Joseph, 52 Z , 25
Zhao, Y., 7374
wreath product Ck o Sn , 137 zig-zag graph, 86

Bruce Landman - Combinatorial Number Theory

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bruce Landman - Combinatorial Number Theory

Uploaded by

Copyright:

Available Formats

De Gruyter Proceedings in Mathematics

Prof. Dr. Jaroslav Neetril Prof. Dr. Richard J. Nowakowski

Prof. Dr. Carl Pomerance Prof. Dr. Aaron Robertson

Library of Congress Cataloging-in-Publication Data

Bibliographic information published by the Deutsche Nationalbibliothek

2013 Walter de Gruyter GmbH, Berlin/Boston

Typesetting: PTP-Berlin Protago-TEX-Production GmbH, www.ptp-berlin.eu

Carrollton, Georgia, USA, March 2013 The Editors

1 Rebecca Milley, Richard J. Nowakowski, and Paul Ottaway

2 Neil Hindman and John H. Johnson

3 Daniel A. Goldston and Andrew H. Ledoan

4 Pr Kurlberg, Jeffrey C. Lagarias, and Carl Pomerance

4.3 An Upper Bound for the Density in Z=nZ . . . . . . . . . . . . . . . . . . . . . 50

7.4 Polynomial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

The Misre Monoid of One-Handed Alternating

Definition 1.1. A game G is said to be alternating if G LL D ; and G RR D ; for all

G  H (mod X) if o .G C X / D o .H C X / for all games X 2 X;

while misre inequality (modulo X) is defined by

G = H (mod X) if o .G C X / o .H C X / for all games X 2 X:

We use the words equivalent and indistinguishable interchangeably, and if G 6 H

In Section 1.2, we determine the equivalence classes of one-handed alternating

along with their corresponding right-handed conjugates3 .

Lemma 1.3. If G is a single right-handed game in P C then G = 0 (mod A), and if

Theorem 1.4. If G is any one-handed alternating game in N \ P C , then G  0

Proof. Assume G is left-handed (the other case is symmetric). By Lemma 1.3 we

We want to show that every one-handed alternating game not in N \ P C is also

Theorem 1.5. If G 2 N \ RC and H 2 L \ P C are one-handed alternating

Theorem 1.6. If G is a one-handed alternating game, then the following statements

(i) If G 2 R \ LC then G  A. (ii) If G 2 L \ RC then G  A.

( C , by induction) or has moves to both L \ RC ( A by (i)) and N \ P C ( 0).

In the first case we have H  B immediately. It remains to show that C j  B

Theorem 1.7. If G 2 N \ RC ; H 2 L \ P C ; and K 2 L \ RC are one-handed

Proof. By Theorem 1.6 we need only show B 5 A 5 C . As in the proof of Claim 2

Theorem 1.8. If G is a one-handed alternating game, then G C G  0 (mod A).

Theorem 1.9. Let G D aA C bB C cC . Then

.a1;b;c/ .a;b1;c/; .a;b1;c1/ .aC1;b;c/ .a;bC1;c/; .a;bC1;cC1/

if a>0; if b>0; if c>0; if a<0; if b<0; if c<0;

A7!0 B7!0;A C 7!B A7!0 B7!0;A C 7!B

Figure 1.4. Left and Right options from G D aA C bB C cC .

Corollary 1.10. If O is the closure of one-handed alternating games, then B  0

Corollary 1.11. If G is a sum of one-handed games then G  aA (mod O) for some

1.4 The Misre Monoid

As noted earlier, this monoid is actually a group; the mapping aA 7! a for a 2 Z

Richard J. Nowakowski, Department of Mathematics and Statistics, Dalhousie University,

Paul Ottaway, Department of Mathematics, Thompson Rivers University, Kamloops, British

Images of C-Sets and Related Large Sets under

finite nonempty subsets of X . It is trivial that for each n 2 N, nN is an IP -set

Proof. [7, Theorem 2.2].

Corollary 2.2. Let l 2 N and for each i 2 1; 2; : : : ; l, let hyi;n i1

Definition 2.3. Let .S; C/ be a commutative semigroup. A set A  S is a C -set if

(2) whenever m 2 N, G1 ; G2 ; : : : ; Gm 2 Pf .NS /, G1 G2 : : : Gm , and for

There is an elementary characterization of C -sets [14] which is very similar in form

Definition 2.4. Let .S; C/ be a commutative semigroup. A set A  S is a J -set if

Definition 2.5. Let S be a semigroup. Then J.S / D p 2 S W .8A 2 p/.A is a

Definition 2.6. AP D p 2 N W .8A 2 p/.A is an AP -set/.

2.2 The Various Notions of Size

Theorem 2.7. Let A  N.

Property Set with no extra properties

Property Set with no extra properties

is a C -set. And, by Lemma 2.22 below, 22n C m2n C 1 W n; m 2 N and m < n is

Theorem 2.8. Let A  N, let > 0, and let 0 < < 1.

Proof. (1) Pick b 2 N such that for all x 2 !, x C 1; x C 2; : : : ; x C b \ A ;

Proof. Let r 2 N n 1; 2. We need to show that g; A contains a length r arith-

2.3 The Functions f and h

G H (mod X) if o .G C X / D o .H C X / for all games X 2 X;

We use the words equivalent and indistinguishable interchangeably, and if G 6 H

Theorem 1.4. If G is any one-handed alternating game in N \ P C , then G 0

(i) If G 2 R \ LC then G A. (ii) If G 2 L \ RC then G A.

( C , by induction) or has moves to both L \ RC ( A by (i)) and N \ P C ( 0).

In the first case we have H B immediately. It remains to show that C j B

Theorem 1.8. If G is a one-handed alternating game, then G C G 0 (mod A).

Corollary 1.10. If O is the closure of one-handed alternating games, then B 0

Corollary 1.11. If G is a sum of one-handed games then G aA (mod O) for some

Definition 2.3. Let .S; C/ be a commutative semigroup. A set A S is a C -set if

Definition 2.4. Let .S; C/ be a commutative semigroup. A set A S is a J -set if

Theorem 2.7. Let A N.

Theorem 2.8. Let A N, let > 0, and let 0 < < 1.

Proof. Let A D x 2 N W f .x/ 2 . 2 ; 2 /. Since f f is a homomorphism, for any

d < a C d 22nC1 so a C 2d < 22nC2 . Since 22n C .n 1/2n C 1 < a C 2d <

Proof. Let 1, 0 < < 1, and A N. The sufficiency holds by assumption.

Corollary 2.34. Let 0 < 1, let 0 < < 1, and let A N. If A is a PS -

r 2 N such that x C mk D g; .r/. Then x C mk m n r C < x C mk C 1 so

./ if l C > x , then l C x C 1 C :

To see this, let l 2 Z. If l t , then l C t C x . Thus, assume that

and the required estimate now follows since, if d log x, dN as x ! 1. Hence,

Note that x 0 for jA.x/j 12 x: If a is an integer, we write a C A for a C A and

Proof. Suppose that g 2 A, so that there exist a1 ; a2 2 A such that a1 a2 D g. If

Proof. We first show that A is a subgroup of Z. Since A is closed under multipli-

Adding these two inequalities and dividing by 2 gives that jA.x/j . 12 1

2j1 b1 2j2 b2 2j3 b3 .mod N /:

Let ai 2 A be such that bi ai .mod n/ for i D 1; 2; 3. We have that a1 ; a2 ; a3

Dx0 D d j`x W d odd; k < .d / < 2k ;