Combinatorial Number Theory

Combinatorial Number
Theory: Proceedings of the

'Integers Conference 2007'
Edited by
B. Landman et al.
Walter de Gruyter
Combinatorial Number Theory
Combinatorial Number Theory
Proceedings of the
Integers Conference 2007
Carrollton, Georgia, USA
October 2427, 2007
Editors
B. Landman
M. B. Nathanson
J. Nesetril
R. J. Nowakowski
C. Pomerance
A. Robertson
Walter de Gruyter Berlin New York

Editors
Bruce Landman Richard J. Nowakowski
Department of Mathematics Department of Mathematics and Statistics
University of West Georgia Dalhousie University
1601 Maple Street Halifax, Nova Scotia, Canada B3H 3J5
Carrollton, GA 30118, USA e-mail: rjn@mathstat.dal.ca
e-mail: landman@westga.edu
Carl Pomerance
Melvyn B. Nathanson Department of Mathematics
Department of Mathematics Dartmouth College
Lehman College (CUNY) Hanover, NH 03755-3551, USA
250 Bedford Park Boulevard West e-mail: carl.b.pomerance@dartmouth.edu
Bronx, NY 10468, USA
Aaron Robertson
e-mail: melvyn.nathanson@lehman.cuny.edu
Department of Mathematics
Jaroslav Nesetril Colgate University
Department of Applied Mathematics 13 Oak Drive
Charles University Hamilton, NY 13346, USA
Malostranske nam. 25 e-mail: aaron@math.colgate.edu
118 00 Praha 1, Czech Republic
e-mail: nesetril@kam.mff.cuni.cz
Keywords: Combinatorics, Number Theory, Primes, Euler Product, Euler Function, Pseudosquares,
Pseudonumbers, Perfect Numbers, Theory of Partitions, Ramsey Theory
Mathematics Subject Classification 2000: 11-06
Printed on acid-free paper which falls within the guidelines of the
ANSI to ensure permanence and durability.
ISBN 978-3-11-020221-2
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.d-nb.de.
Copyright 2009 by Walter de Gruyter GmbH & Co. KG, 10785 Berlin, Germany.
All rights reserved, including those of translation into foreign languages. No part of this book
may be reproduced or transmitted in any form or by any means, electronic or mechanical, includ-
ing photocopy, recording or any information storage and retrieval system, without permission in
writing from the publisher.
Printed in Germany.
Cover design: Thomas Bonnie, Hamburg.
Typeset using the authors T
E
X files: Kay Dimler, Mncheberg.
Printing and binding: Hubert & Co. GmbH & Co. KG, Gttingen.
Preface
The Integers Conference 2007 was held October 2427, 2007, at the University
of West Georgia in Carrollton, Georgia. This was the third Integers Conference,
held bi-annually since 2003.
It featured sixty-four invited talks including six plenary lectures presented
by George Andrews, Vitaly Bergelson, Bryna Kra, Florian Luca, Ken Ono, and
Van Vu.
This volume consists of sixteen refereed articles, which are expanded and re-
vised versions of talks presented at the conference. These sixteen articles will
appear as a special volume of the journal Integers: Electronic Journal of Com-
binatorial Number Theory. They represent a broad range of topics in the ar-
eas of number theory and combinatorics including multiplicative number theory,
additive number theory, Ramsey theory, enumerative combinatorics, elementary
number theory, the theory of partitions, algebraic number theory, and integer se-
quences.
The conference was made possible with the generous support of the National
Science Foundation and the University of West Georgia. The Integers conferences
are organized by the Editors of Integers, which publishes articles in the eld of
combinatorial number theory. The conferences are held in order to further support
and strengthen this growing eld.
December, 2008 The Editors
Table of contents
Preface : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : v
GEORGE E. ANDREWS
The Finite Heine Transformation : : : : : : : : : : : : : : : : : : : : : : : 1
TSZ HO CHAN
Finding Almost Squares III : : : : : : : : : : : : : : : : : : : : : : : : : : 7
DENNIS EICHHORN, MIZAN R. KHAN, ALAN H. STEIN,
CHRISTIAN L. YANKOV
Sums and Differences of the Coordinates of Points on Modular Hyperbolas : 17
DAVID GARTH, JOSEPH PALMER, HA TA
Self Generating Sets and Numeration Systems : : : : : : : : : : : : : : : : 41
NEIL HINDMAN
Small Sets Satisfying the Central Sets Theorem : : : : : : : : : : : : : : : 57
BRIAN HOPKINS
Column-to-Row Operations on Partitions: The Envelopes : : : : : : : : : : 65
XIAN-JIN LI
On the Euler Product of Some Zeta Functions : : : : : : : : : : : : : : : : 77
FLORIAN LUCA, CARL POMERANCE
On the Range of the Iterated Euler Function: : : : : : : : : : : : : : : : : 101
GRETCHEN L. MATTHEWS
Frobenius Numbers of Generalized Fibonacci Semigroups : : : : : : : : : 117
JAMES MCLAUGHLIN, ANDREW V. SILLS
Combinatorics of RamanujanSlater Type Identities: : : : : : : : : : : : : 125
KEN ONO
A Mock Theta Function for the Delta-function : : : : : : : : : : : : : : : 141
RAM KRISHNA PANDEY, AMITABHA TRIPATHI
On the Density of Integral Sets with Missing Differences : : : : : : : : : : 157
viii Table of contents
CARL POMERANCE, IGOR E. SHPARLINSKI
On Pseudosquares and Pseudopowers : : : : : : : : : : : : : : : : : : : : 171
FRANK THORNE
Maier Matrices Beyond Z: : : : : : : : : : : : : : : : : : : : : : : : : : 185
TOMOHIRO YAMADA
Linear Equations Involving Iterates of .N / : : : : : : : : : : : : : : : : 193
PAUL YIU, K. R. S. SASTRY, SHANZHEN GAO
Heron Sequences and Their Modications : : : : : : : : : : : : : : : : : 199
List of participants : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 205
Combinatorial Number Theory de Gruyter 2009
The Finite Heine Transformation
George E. Andrews
Abstract. We shall present nite summations that converge to the Heine
2
1
transfor-
mations in the limit as n !1. We shall investigate their partition-theoretic implications.
Keywords. q-series, partitions, basic hypergeometric series, Heines transformation.
AMS classication. 11P81, 11P83, 05A17, 05A19.
1 Introduction
In an expository article describing Eulers pioneering work on partitions, I was
particularly drawn to Eulers assertion [6, p. 566, eq. (5.2) corrected]
1
Y
nD0
q
3
n
C1 Cq
3
n
D
1
X
nD1
q
n
. (1.1)
an identity valid only in a formal sense in that neither the series nor the product
converges for any value of q.
This led to my comparisons of the two innite series identities ([6, p. 567,
eq. (5.5)] and [6, p. 567, eq. (5.6)] respectively):
1
X
nD0
q
n
2
(1 q)
2
(1 q
2
)
2
(1 q
n
)
2
D
1
Y
nD1
1
1 q
n
. (1.2)
and
1
X
nD0
q
n
(1 q)
2
(1 q
2
)
2
(1 q
n
)
2
D
1
Y
nD1
1
(1 q
n
)
2
1
X
jD0
(1)
j
q
j.jC1/=2
. (1.3)
Each of the left-hand series is analytic inside jqj < 1 with jqj D 1 as a nat-
ural boundary, and the second series is formally transformable into the rst by
the mapping q ! 1,q. The fact that jqj D 1 is a natural boundary means we
should not be surprised when the same transformation applied to the right-hand
side produces only nonsense.
Partially supported by National Science Foundation Grant DMS 0200097.
2 George E. Andrews
However, it was observed in [4] that it is sometimes possible to nd polynomial
or rational function identities that converge to innite q-series in the limit. This
observation in [7] was the secret to dealing with Regime II of Baxters generalized
hard-hexagon model (cf. [5, Ch. 8]).
So this led to the question: Are there nite identities that would both (A)
simplify to (1.2) and (1.3) in the limit, and (B) allow the mapping q ! 1,q
prior to taking limits?
The answer to this question is yes. In Section 2 we provide q-analogs of the
Heine transformations of the
2
1
. In Section 3, we shall derive generalizations of
the following corollaries:
N
X
nD0
q
n
2
(1 q)
2
(1 q
2
)
2
(1 q
n
)
2
D
N
Y
nD1
1
(1 q
n
)
N
X
jD0
q
.NC1/j
(1 q)(1 q
2
) (1 q
j
)
.
(1.4)
and
N
X
nD0
q
n
(1 q)
2
(1 q
2
)
2
(1 q
n
)
2
D
N
Y
nD1
1
(1 q
n
)
N
X
jD0
(1)
j
q
j.jC1/=2
(1 q)(1 q
2
) (1 q
Nj
)
.
(1.5)
Clearly (1.4) and (1.5) converge to (1.2) and (1.3) as N !1, and by reversing
the sum on the right-hand side it is a simple matter to see that (1.4) becomes (1.5)
under the now legitimate mapping q !1,q.
In Section 4, we shall note quite transparent combinatorial proofs of (1.4) and
(1.5).
2 Finite Heine Transformations
We shall employ the following standard notation
(a)
n
D (aI q)
n
D
n1
Y
jD0
(1 aq
j
). (2.1)
(a
1
. . . . . a
r
I q)
n
D (a
1
I q)
n
(a
2
I q)
n
(a
r
I q)
n
. (2.2)
and
rC1
a
0
. a
1
. . . . . a
r
I q. t
b
1
. . . . . b
r
!
D
1
X
nD0
(a
0
. a
1
. . . . . a
r
I q)
n
t
n
(q. b
1
. . . . . b
r
I q)
n
. (2.3)
The Finite Heine Transformation 3
Lemma 1. For non-negative integers n,
3
q
n
. . I q. q
;. q
1n
,t
!
D
(tI q)
n
(tI q)
n
3
q
n
. ;,. I q. tq
n
;. t
!
. (2.4)
Proof. In (III.13) of [8, p. 242], set b D ;,, c D , J D ;, e D t. The result
after simplication is (2.4).
We note in passing that Lemma 1 is, in fact, a nite version of Jacksons sum-
mation [9] (cf. [8, p. 11, eq. (1.54)], [2, p. 527, Lemma]).
Theorem 2. For non-negative integers n,
3
q
n
. . I q. q
;. q
1n
,t
!
D
(. tI q)
n
(;. tI q)
n
3
q
n
. ;,. tI q. q
t. q
1n
,
!
. (2.5)
Remark. When n ! 1, this is Heines classic
2
1
transformation [8, p. 9,
eq. (1.4.1)], [3, p. 28, Cor. 2.3].
Proof. If in Lemma 1, we replace , , ;, and t by ;,, t, t and respectively,
we nd that
3
q
n
. ;,. I q. tq
n
;. t
!
D
(I q)
n
(;I q)
n
3
q
n
. ;,. tI q. q
t. q
1n
,
!
. (2.6)
Now substituting the left-hand side of (2.6) into the right-hand side of (2.4) we
deduce (2.5).
Corollary 3. For non-negative integers n,
3
q
n
. . I q. q
;. q
1n
,t
!
D
(;,. tI q)
n
(;. tI q)
n
3
q
n
. t,;. I q. q
t. q
1n
,;
!
. (2.7)
Proof. Apply Theorem 2 (with , , ; and t replaced by t, ;,, t and
respectively) to transform the
3
2
on the right-hand side of (2.5).
Corollary 4. For non-negative integers n,
3
q
n
. . I q. q
;. q
1n
,t
!
D
(
I q)
n
(tI q)
n
3
q
n
.

I q. q
;. ;q
1n
,(t)
!
. (2.8)
Proof. Apply Theorem 2 (with , , ; and t replaced by , t,;, t, ;,
respectively) to transform the
3
2
on the right-hand side of (2.7).
Corollaries 3 and 4 reduce to the second and third Heine transformations [8, p.
10] when n !1.
4 George E. Andrews
3 Identities (1.4) and (1.5)
n
X
jD0
q
j
(q. ;I q)
j
D
1
(;)
n
n
X
jD0
(1)
j
;
j
q
j.j1/=2
(q)
nj
. (3.1)
Proof. Set D 0, t D q and let !0 in Theorem 2. The desired result follows
after algebraic simplication.
n
X
jD0
q
j
2
;
j
(q. ;qI q)
j
D
1
(;q)
n
n
X
jD0
;
j
q
j.nC1/
(q)
j
. (3.2)
Proof. Replace q by 1,q and ; by 1,q; in (3.1), then reverse the sum on the
right-hand side and simplify.
Identity (1.5) is Theorem 5 with ; D q, and (1.4) is Theorem 6 with ; D 1.
4 Combinatorial Proofs
Replacing q by q
2
in Theorem 5 and then replacing ; with ;q, we see that
Theorem 5 is equivalent to the following assertion:
n
X
jD0
q
2j
;q
2jC1
I q
2
nj
(q
2
I q
2
)
j
D
n
X
jD0
;
j
q
j
2
(q
2
I q
2
)
nj
. (4.1)
Proof of (4.1). The left-hand side of (4.1) is the generating function for partitions
in which (1) all parts are 5 2n, (2) odd parts are distinct, and (3) each odd is >
each even. The general two-modular Ferrers graph [3, p. 13] for such partitions is
thus
2 2 2 2 2 1
2 2 2 2 1
2 2 2 2 1
. . . . . . . . . . . . . . . . . . . . . . . . .
2 2 2 2
2 2 2
.
.
.
2
The Finite Heine Transformation 5
Now remove the columns that have a 1 at the bottom. In light of the fact that the
odds were distinct, we see that if there were originally odd parts, then we have
removed 1C3C5C C(2 1) (D
2
). The remaining parts are all even and the
largest is at most 2n 2 . Thus this transformation (which is clearly reversible)
provides the partitions generated by the right-hand side of (4.1) and thus we have
a bijective proof of Theorem 5.
Proof of (3.2). Classical arguments immediately reveal that the left-hand side of
(3.2) is the generating function for partitions with Durfee square of side at most
n. ; keeps track of the number of parts.
On the other hand, the side of the Durfee square is the largest such that the
th
part is = . So we may replicate the partitions generated by the left-hand side
of (3.2) by exhibiting the generating function for partitions in which the parts > n
are at most n in number. If there are parts greater than n, the generating function
is
;
j
q
j.nC1/
(;q)
n
(q)
j
.
Hence summing on from 0 to n we obtain a new expression for the generating
function for partitions with Durfee square at most n, and this proves (3.2).
5 Conclusion
There are many other corollaries obtainable from the nite Heine transforma-
tions. The q-Pfaff-Saalschtz summation is merely [8, p. 13, eq. (1.7.2)] with
t D ;,. One can also obtain a nite version of the q-analog of Kummers
theorem [2], however, the result does not reduce to the hoped for sum equals
product identity. Also it should be possible to provide a fully combinatorial
proof of Theorem 2 along the lines given in [1] for the n !1case.
References
[1] G. E. Andrews, Enumerative proofs of certain q-identities, Glasgow Math. J. 8
(1967), 3340.
[2] G. E. Andrews, On the q-analog of Kummers theorem and applications, Duke Math.
J. 40 (1973), 525528.
[3] G. E. Andrews, The Theory of Partitions, Encycl. of Math. and Its Appl., Vol. 2,
Addison-Wesley, Reading, 1976 (Reissued: Cambridge University Press, 1998).
[4] G. E. Andrews, The hard-hexagon model and RogersRamanujan type identities,
Proc. Nat. Acad. Sci. (USA) 78 (1981), 52905292.
6 George E. Andrews
[5] G. E. Andrews, q-Series: Their Development and Application in Analysis, Number
Theory, Combinatorics, Physics and Computer Algebra, CBMS Regional Conference
Lecture Series 66, Amer. Math. Soc., Providence, 1986.
[6] G. E. Andrews, Eulers De Partitio(ne) Numerorum, Bull. Amer. Math. Soc. 44
(2007), 561573.
[7] G. E. Andrews, R. J. Baxter and P. J. Forrester, Eight-vertex SOS model and general-
ized RogersRamanujan-type identities, J. Stat. Phys. 35 (1984), 193266.
[8] G. Gasper and M. Rahman, Basic Hypergeometric Series, Encycl. of Math. and Its
Appl., Vol. 35, 1st ed., Cambridge University Press, Cambridge, 1990.
[9] F. H. Jackson, Tranformations of q-series, Messenger of Mathematics 39 (1910), 145
153.
Author information
George E. Andrews, Department of Mathematics, The Pennsylvania State University,
University Park, PA 16802, USA.
E-mail: andrews@math.psu.edu
Finding Almost Squares III
Tsz Ho Chan
Abstract. An almost square of type 2 is an integer n that can be factored in two different
ways as n = a
1
b
1
= a
2
b
2
with a
1
, a
2
, b
1
, b
2
~
_
n. In this paper, we shall improve
upon our previous results on short intervals containing such an almost square. This leads
to another question of independent interest: given some 0 < c < 1, nd a short interval
around . which contains an integer divisible by some integer in .
c
. 2.
c
|.
Keywords. Almost square, Erd osTurn inequality, exponent pairs, almost divisible.
AMS classication. 11B75, 11L07, 11N25.
1 Introduction and Main Results
In [1] and [2], the author started studying almost squares which are integers n that
can be factored as n = ab with a. b close to
_
n. For example, n = 9999 =
99 101 is an almost square. We say that an integer n is an almost square of
type 2 if it has two different representations, n = a
1
b
1
= a
2
b
2
, with a
1
. b
1
. a
2
. b
2
close to
_
n. For example n = 99990000 = 9999 10000 = 9900 10100 is an
almost square of type 2.
More precisely, for 0 _ 0 _ 1,2 and C > 0,
Denition 1. An integer n is a (0, C)-almost square of type 1 if n = ab for some
integers a. b in the interval n
1{2
Cn
0
. n
1{2
Cn
0
|.
Denition 2. An integer n is a (0, C)-almost square of type 2 if n = a
1
b
1
= a
2
b
2
for some integers a
1
< a
2
_ b
2
< b
1
in the interval n
1{2
Cn
0
. n
1{2
Cn
0
|.
Let . be a large positive real number. Following [1] and [2], we are interested
in nding almost squares of type 1 or 2 near to .. In particular, given 0 _ 0 _ 1,2,
we want to nd admissible
i
_ 0 (as small as possible) such that, for some
constants C
0,i
. D
0,i
> 0, the interval . D
0,i
.
i
. . D
0,i
.
i
| contains a (0,
C
0,i
)-almost square of type i (i = 1. 2) for all large ..
Denition 3. (0) := inf
1
and g(0) := inf
2
where the inma are taken over
all the admissible
i
(i = 1. 2) respectively.
8 Tsz Ho Chan
Clearly and g are non-increasing functions of 0. Summarizing the results in
[1] and [2], we have:
Theorem 1. For 0 _ 0 _ 1,2,
(0)
8
<
:
= 1,2, if 0 _ 0 < 1,4,
= 1,4, if 0 = 1,4,
= 1,2 0, if 1,4 _ 0 _ 3,10 and a conjectural upper bound on
certain average of twisted incomplete Salie sum is true,
_ 1,2 0, if 1,4 _ 0 _ 1,2.
Theorem 2. For 0 _ 0 _ 1,2,
g(0)
8
<
:
does not exist, if 0 _ 0 < 1,4,
_ 1 20, if 1,4 _ 0 _ 1,2,
_ 1 0, if 1,4 _ 0 _ 1,3.
And we gave the following conjecture.
Conjecture 1. For 0 _ 0 _ 1,2,
(0) =
1,2, if 0 _ 0 < 1,4,

1,2 0, if 1,4 _ 0 _ 1,2;
and
g(0) =
does not exist, if 0 _ 0 < 1,4,

1 20, if 1,4 _ 0 _ 1,2.
In this paper, we improve Theorem 2:
Theorem 3. For 1,4 _ 0 _ 1,2,
(i) g(1,4) _ 5,8,
(ii) g(0) _ 9,16, if 5,16 _ 0 _ 1,2,
(iii) g(0) _ 17,32, if 5,16 _ 0 _ 1,2,
(iv) g(0) _ 1,2, if 1,3 < 0 _ 1,2,
(v) g(0) _ 1,2, if 743,2306 < 0 _ 1,2.
Clearly (iii) is better than (ii). The reason we keep (ii) is that (ii) and (iii) use
different approaches. Also (v) includes (iv). We keep (iv) because it is a prototype
of (v).
Finding Almost Squares III 9
0
1
2
1
g(0)
1
4
1
2
0
5
16
743
2306
5
8
17
32
?
?
?
H
H
H
H
H
H
H
H
H
H
6
X
X
X
X
?
?
The above picture summarizes Theorems 2 and 3. The thin line segments are
the upper and lower bounds from Theorem 2. The thick horizontal line segments
are the upper bounds from Theorem 3. The next challenge is to beat the
1
2
upper
bound for g(0).
Some Notations: Throughout the paper, c denotes a small positive number. Both
(.) = O(g(.)) and (.) g(.) mean that [(.)[ _ Cg(.) for some constant
C > 0. Moreover (.) = O
2
(g(.)) and (.)
2
g(.) mean that the implicit
constant C = C
2
may depend on the parameter z. Finally (.) g(.) means
that (.) g(.) and g(.) (.).
2 Proof of Theorem 3 (i)
Let 1,4 _ 0 _ 1,2. From [2], we recall that a (0, C)-almost square of type 2 is
of the form
n = (J
1
e
1
)(J
2
e
2
) = (J
1
e
2
)(J
2
e
1
).
where a
1
= J
1
e
1
, b
1
= J
2
e
2
, a
2
= J
1
e
2
, b
2
= J
2
e
1
; n
1{2
Cn
0
_ a
1
< a
2
_
b
2
< b
1
_ n
1{2
Cn
0
;
1
2C
n
1
2
0
1
2
_ J
1
. J
2
. e
1
. e
2
_ 2Cn
0
. e
2
e
1
_ 2C
n
0
J
2
. J
2
J
1
_ 2C
n
0
e
2
.
Let 1 _ k 1 be any integer. By the 0 = 1,4 case in Theorem 1, for some
constant C > 0, we can nd integers J. e .
1{4
C.
1{8
. .
1{4
C.
1{8
| such
that
Je = .
1{2
2k.
1{4
O(.
1{8
).
Then (J 2k)(e 2k) = Je 2k(J e) k
2
= .
1{2
2k.
1{4
O(.
1{8
).
and
Je(J 2k)(e 2k) = . 4k
2
.
1{2
O(.
5{8
) = . O(.
5{8
).
This gives g(1,4) _ 5,8.
10 Tsz Ho Chan
3 Proof of Theorem 3 (ii)
The key idea is the identity
ab =
a b
2
a b
2
2
as used in [1]. Using this identity,
J
1
e
1
J
2
e
2
=
h
J
2
J
1
2
J
2
J
1
2
2
ih
e
2
e
1
2
e
2
e
1
2
2
i
=
J
2
J
1
2
e
2
e
1
2
J
2
J
1
2
e
2
e
1
2
e
2
e
1
2
J
2
J
1
2
J
2
J
1
2
e
2
e
1
2
2
=: G
2
H
2
g
2
H
2
h
2
G
2
g
2
h
2
where G =
d
2
Cd
1
2
, H =
e
2
Ce
1
2
, g =
d
2
d
1
2
and h =
e
2
e
1
2
. Now we want
. ~ J
1
e
1
J
2
e
2
= G
2
H
2
g
2
H
2
h
2
G
2
g
2
h
2
.
G
2
H
2
. ~ g
2
H
2
h
2
G
2
g
2
h
2
.
(GH
_
.)(GH
_
.) ~ g
2
H
2
h
2
G
2
g
2
h
2
. (1)
By the 0 = 1,4 case in Theorem 1, for some constant C > 0, there exist integers
G. H .
1{4
C.
1{16
. .
1{4
C.
1{16
| such that 0 < GH
_
. .
1{8
. Then
the left hand side of (1) is .
1{2C1{8
. As for the right hand side of (1), observe
that, for xed h (say h = 1), the increment
(i 1)
2
H
2
h
2
G
2
(i 1)
2
h
2
| i
2
H
2
h
2
G
2
i
2
h
2
|
= (2i 1)H
2
(2i 1)h
2
.
1{2
i.
Now observe that
g
2
H
2
h
2
G
2
g
2
h
2
= h
2
G
2
X
0i~
(i 1)
2
H
2
h
2
G
2
(i 1)
2
h
2
| i
2
H
2
h
2
G
2
i
2
h
2
|
.
1{2
X
1i~
i g
2
.
1{2
.
Therefore, for some integer 1 _ g .
1{16
,
[Right hand side of (1) Left hand side of (1)[ .
1{2
g .
1{2C1{16
.
This gives
[. (G
2
g
2
)(H
2
h
2
)[ .
1{2C1{16
or
[. J
1
J
2
e
1
e
2
[ = [. (G g)(G g)(H h)(H h)[ .
1{2C1{16
.
Consequently, with
a
1
= J
1
e
1
= (G g)(H h) = .
1{2
O(.
1{4C1{16
).
b
1
= J
2
e
2
= (G g)(H h) = .
1{2
O(.
1{4C1{16
).
a
2
= J
1
e
2
= (G g)(H h) = .
1{2
O(.
1{4C1{16
).
b
2
= J
2
e
1
= (G g)(H h) = .
1{2
O(.
1{4C1{16
).
we have a (
1
4

1
16
. C
0
)-almost square n = a
1
b
1
= a
2
b
2
of type 2 in the interval
. C
00
.
1{2C1{16
. . C
00
.
1{2C1{16
| for some C
0
. C
00
> 0. This proves that
g(0) _
9
16
for 0 _
5
16
.
4 Proof of Theorem 3 (iii)
This time we try to approximate the left hand side of (1) by the quadratic form
g
2
H
2
h
2
G
2
directly. As in the proof of Theorem 3 (ii), for some C > 0,
there exist integers .
1{4
C.
1{16
_ G. H _ .
1{4
C.
1{16
such that 0 <
GH
_
. .
1{8
. The left hand side of (1) is .
1{2C1{8
. Without loss of
generality, G _ H. Then g
2
H
2
h
2
G
2
= G
2
(g
2
h
2
) (H
2
G
2
)g
2
.
Observe that 0 _ H
2
G
2
= (H G)(H G) .
1{4C1{16
. By an elementary
argument, for any real number X > 0, we can nd a sum of two squares g
2
h
2
such that [X (g
2
h
2
)[ X
1{4
. In particular, we can nd 1 _ g. h .
1{16
such that
(GH
_
.)(GH
_
.)
G
2
(g
2
h
2
)
.
1{32
.
This implies
[(GH
_
.)(GH
_
.) (g
2
H
2
h
2
G
2
g
2
h
2
)[
_ [(GH
_
.)(GH
_
.) G
2
(g
2
h
2
)[ [(H
2
G
2
)g
2
[ [g
2
h
2
[
.
1{2C1{32
.
Hence
[. J
1
J
2
e
1
e
2
[ = [. (G g)(G g)(H h)(H h)[ .
1{2C1{32
.
12 Tsz Ho Chan
Consequently, with
a
1
= J
1
e
1
= (G g)(H h) = .
1{2
O(.
1{4C1{16
).
b
1
= J
2
e
2
= (G g)(H h) = .
1{2
O(.
1{4C1{16
).
a
2
= J
1
e
2
= (G g)(H h) = .
1{2
O(.
1{4C1{16
).
b
2
= J
2
e
1
= (G g)(H h) = .
1{2
O(.
1{4C1{16
).
there is a (
1
4

1
16
. C
0
1
b
1
= a
2
b
2
. C
00
.
1{2C1{32
. . C
00
.
1{2C1{32
| for some C
0
. C
00
> 0. This proves that
g(0) _
17
32
for 0 _
5
16
.
5 Proof of Theorem 3 (iv)
Let 1,2 _ _ 1. Observe that, for large ., the interval . .
1
. . 2.
1
|
contains an integer n which is divisible by an integer a .
1
,2. .
1
|. In
particular n = ab with integer b .
. 3.
|.
Again we use (1). Instead of having G. H close to .
1{4
as in the proof of
Theorem 3 (iii), we want
G ~ .
(1){2
and H ~ .
{2
for some
1
2
< <
2
3
.
By the observation at the beginning of this section, we can nd H .
{2
. 3.
{2
|
and G .
(1){2
,2. .
(1){2
| such that 0 < GH
_
. .
(1){2
. Then the
left hand side of (1) is 1 := (GH
_
.)(GH
_
.) .
1{2
.
Firstly, we approximate 1 by g
2
H
2
. For some choice of g .
1{23{4
, we
have 0 < 1 g
2
H
2
gH
2
.
1{2C{4
. Note that
1
2

3
4
> 0 as <
2
3
.
Secondly, we approximate 1 g
2
H
2
by h
2
G
2
. For some choice of h
.
5{81{4
, we have [1g
2
H
2
h
2
G
2
[ hG
2
.
3{43{8
. Note that
5
8

1
4
> 0
as >
1
2
.
Thirdly, observe that g
2
h
2
.
1{2{4
.
3{43{8
as < 2. Therefore,
[1 g
2
H
2
h
2
G
2
g
2
h
2
[ .
3{43{8
which gives
[. J
1
J
2
e
1
e
2
[ = [. (G g)(G g)(H h)(H h)[ .
3{43{8
.
Consequently, as
1
2
< <
2
3
, with
a
1
= J
1
e
1
= (G g)(H h) = .
1{2
O(.
1{2{4
).
b
1
= J
2
e
2
= (G g)(H h) = .
1{2
O(.
1{2{4
).
a
2
= J
1
e
2
= (G g)(H h) = .
1{2
O(.
1{2{4
).
b
2
= J
2
e
1
= (G g)(H h) = .
1{2
O(.
1{2{4
).
there is a (
1
2

4
. C
0
1
b
1
= a
2
b
2
. C
00
.
3{43{8
. . C
00
.
3{43{8
| for some C
0
. C
00
> 0. By picking close
to
2
3
, we have g(0) _
1
2
for 0 >
1
3
.
6 Integers Almost Divisible by Some Integer
in an Interval
Again let 1,2 _ _ 1. In the previous section, we found an interval of length
.
1
around . containing an integer divisible by some integer in the interval
.
1
,2. .
1
|. This is obviously true. Our goal in this section is to nd a
shorter interval still containing an integer divisible by some integer in the interval
.
1
,2. .
1
|. We hope that this will give some improvements to Theorem 3
(iv). Let us reformulate the question as follows:
Question 1. Let 0 < _ 1,2 and X > 0 be a large integer. Given 0 < c
1
<
c
2
_ 1, nd 1, as small as possible, such that the interval X 1. X| contains an
integer that is divisible by some integer in the interval c
1
X
. c
2
X
|.
One may interpret the above as nding an integer in the interval c
1
X
. c
2
X
|
that almost divides X (with a remainder less than or equal to 1). We suspect that
Conjecture 2, below, is true, but are only able to prove Proposition 1, below.
Conjecture 2. For any c > 0, one can take 1 = X
e
in the above question as long
as X is sufciently large (in terms of c).
Proposition 1. Suppose (. q) with 0 _ _
1
2
_ q _ 1 is an exponent pair for
exponential sums. Then one can take 1 = X
.qp/
1Cp
C
p
1Cp
Ce
in the above question
for any c > 0 as long as X is sufciently large (in terms of c).
Our method of proof uses Erd osTurn inequality in the following form (see
H. L. Montgomery [4, Corollary 1.2] for example):
Lemma 1. Suppose M is a positive integer chosen so that
X
ID1
J
X
}D1
e(l.
}
)
_
J
10
.
Then every arc J = . | _ 0. 1| of length _
4
C1
contains at least
1
2
J( ) of the points .
}
, 1 _ _ J. Here [[.[[ = min
n2Z
[. n[, the
distance from . to the nearest integer, and e(.) = e
2tix
.
14 Tsz Ho Chan
Proof of Proposition 1. Our sequence {.
}
}
J
}D1
should be {
A
o
: a Z and a
c
1
X
. c
2
X
|}. We want to nd some a such that the fractional part of

A
o
is small.
For if {
A
o
} 0.
4
|, then
A
o
= k
0
for some integer k and 0 _ 0 _ 4. This

gives X = ka
0
a and X
0
a = ka. Hence, with 1 =

4c
2
A
, the interval
X 1. X| contains an integer that is divisible by some integer in c
1
X
. c
2
X
|.
Thus, in view of Lemma 1, it sufces to show
S :=
21
X
ID1
X
c
1
A
oc
2
A
lX
a
_ X
e
for any 21 _ M and c > 0 as long as X is sufciently large in terms of c. Keep
in mind that we want M as large as possible.
By the theory of exponent pairs on exponential sums (for example, see Chap-
ter 3, Section 4 of [4] for an overview),
X
c
1
A
oc
2
A
lX
a
(lX(X
)
2
)
]
(X
)
q
1
]
X
]2]Cq
(2)
if (. q) with 0 _ _
1
2
_ q _ 1 is an exponent pair. Using (2), we have
21
X
ID1
X
c
1
A
oc
2
A
lX
a
1
1C]
X
]2]Cq
.
Thus, S _ X
e
provided that, for X large enough,
1
1C]
X
]2]Cq
_ X
e
or 1 _ X
.1qC2p/
1Cp

p
1Cp
e
.
Therefore, we can pick M = X
.1qC2p/
1Cp

p
1Cp
e
, which gives the identity 1 =
4c
2
X
.qp/
1Cp
C
p
1Cp
Ce
. This proves Proposition 1 as c is arbitrary.
7 Proof of Theorem 3 (v)
We follow closely the proof of Theorem 3 (iv). Applying Proposition 1 with =
1 and X = .3.
.1/.qp/
1Cp
C
p
1Cp
Ce
, the interval ..
.1/.qp/
1Cp
C
p
1Cp
Ce
. .
3.
.1/.qp/
1Cp
C
p
1Cp
Ce
| contains an integer n = ab with integers a .
1
,2.
.
1
| and b .
. 3.
|. Thus we can nd
H .
{2
. 3.
{2
| and G .
(1){2
,2. .
(1){2
|
such that
0 < GH
_
. .
.1/.qp/
2.1Cp/
C
p
2.1Cp/
C
2
.
Then for the left hand side of (1) we have 1 := (GH
_
.)(GH
_
.)
.
1CpCq
2.1Cp/
qp
2.1Cp/
C
2
.
Firstly, approximate 1by g
2
H
2
. For some choice of g .
1CpCq
4.1Cp/
2CpCq
4.1Cp/
C
4
,
we have
0 < 1 g
2
H
2
gH
2
.
1CpCq
4.1Cp/
C
2C3pq
4.1Cp/
C
4
.
Note that we need
1C]Cq
4(1C])

2C]Cq
4(1C])
_ 0 which means _
1C]Cq
2C]Cq
.
Secondly, we approximate 1 g
2
H
2
by h
2
G
2
. For some choice of h
.
6C7pq
8.1Cp/

3C3pq
8.1Cp/
C
8
, we have
[1 g
2
H
2
h
2
G
2
[ hG
2
.
5C5pCq
8.1Cp/

2CpCq
8.1Cp/
C
8
.
Note that
6C7]q
8(1C])

3C3]q
8(1C])
_ 0 as _ 1,2 and . q _ 0.
Thirdly, observe that g
2
h
2
.
3qp1
4.1Cp/

3q5p2
4.1Cp/
C
3
4
.
5C5pCq
8.1Cp/

2CpCq
8.1Cp/
C
8
provided <
7C7]5q
6C11]5q
and c is small enough. One can easily check that
7C7]5q
6C11]5q
>
1C]Cq
2C]Cq
. Therefore we have [1 g
2
H
2
h
2
G
2
g
2
h
2
[
e
.
5C5pCq
8.1Cp/

2CpCq
8.1Cp/
C
8
which gives
[. J
1
J
2
e
1
e
2
[ = [. (G g)(G g)(H h)(H h)[
e
.
5C5pCq
8.1Cp/

2CpCq
8.1Cp/
C
8
provided
1
2
_ _
1C]Cq
2C]Cq
. Choose =
1C]Cq
2C]Cq
, we have, after some simple
algebra,
[. J
1
J
2
e
1
e
2
[ = [. (G g)(G g)(H h)(H h)[
e
.
1
2
C
8
.
Now observe that with =
1C]Cq
2C]Cq
, after some algebra,
GH .
1{2
.
q
2.1Cp/
qp
2.1Cp/
C
2
= .
pCq
2.2CpCq/
C
2
.
gH .
1CpCq
4.1Cp/
2CpCq
4.1Cp/
C
2
C
4
= .
1CpCq
2.2CpCq/
C
4
.
hG .
6C7pq
8.1Cp/

3C3pq
8.1Cp/
C
1
2
C
8
= .
1CpCq
2.2CpCq/
C
8
.
and
gh .
3qp1
8.1Cp/

3q5p2
8.1Cp/
C
3
8
= .
pCq
2.2CpCq/
C
3
8
.
Therefore a
1
= J
1
e
1
= (G g)(H h), b
1
= J
2
e
2
= (G g)(H h),
a
2
= J
1
e
2
= (G g)(H h) and b
2
= J
2
e
1
= (G g)(H h) are all
= .
1
2
O
e
(.
1CpCq
2.2CpCq/
C
2
). Therefore, there is a (
1C]Cq
2(2C]Cq)
e
2
. C
e
)-almost square
16 Tsz Ho Chan
of type 2 in the interval . .
1
2
C
8
. . .
1
2
C
8
|. This shows that g(0) _
1
2
for
0 >
1C]Cq
2(2C]Cq)
. Since
1Cu
2Cu
is an increasing function of u, we try to nd exponent
pairs that make q as small as possible.
For example, recently, Huxley [3] proved that (. q) = (
32
205
c.
1
2
32
205
c)
is an exponent pair for any c > 0. This gives
1C]Cq
2(2C]Cq)
_
743
2306
c for any c > 0
and hence Theorem 3 (v).
Note that
743
2306
= 0.3222029488 . . . <
1
3
. However, we still cannot beat the
1
2
bound for g(0). Assuming the exponent pair conjecture that (c.
1
2
c) is an
exponent pair, we can push the range for 0 to 0 > 0.3 with g(0) _
1
2
but this
is still shy of the range 0 _
1
4
. Furthermore, if one assumes Conjecture 2 in the
previous section and imitates the proof of part (iv) or (v) of Theorem 3, one can
get g(0) _
1
2
for 0 >
1
4
. This comes close to the conjecture g(
1
4
) =
1
2
.
Acknowledgments. The author would like to thank the American Institute of
Mathematics where the study of almost squares began during a visit from 2004
to 2005. He also thanks Central Michigan University where the main idea of this
paper was worked out during a one-year visiting position (20052006). Finally,
he thanks the University of Hong Kong where the Erd osTurn and exponent pair
part was worked out during a visit there in the summer of 2007.
References
[1] T. H. Chan, Finding almost squares, Acta Arith. 121 (2006), no. 3, 221232.
[2] T. H. Chan, Finding almost squares II, Integers 5 (2005), no. 1, A23, 4 pp. (elec-
tronic).
[3] M. N. Huxley, Exponential sums and the Riemann zeta function. V., Proc. London
Math. Soc. (3) 90 (2005), no. 1, 141.
[4] H. L. Montgomery, Ten Lectures on the Interface Between Analytic Number The-
ory and Harmonic Analysis, CBMS Regional Conference Series in Mathematics 84,
published for the Conference Board of the Mathematical Sciences, Washington, DC,
American Mathematical Society, Providence, RI, 1994.
Author information
Tsz Ho Chan, Department of Mathematical Sciences, University of Memphis,
Memphis, TN 38152, USA.
E-mail: tchan@memphis.edu
Sums and Differences of the Coordinates
of Points on Modular Hyperbolas
Dennis Eichhorn, Mizan R. Khan, Alan H. Stein and
Christian L. Yankov
Abstract. The modular hyperbola H
n
is {(.. ,) : ., = 1(mod n). 1 _ .. , _ n 1}.
This simply dened set of points has connections to a variety of other mathematical topics
including Kloosterman sums, quasirandomness, and consecutive Farey fractions. These
connections have inspired a closer look at the distribution of the points of H
n
, and many
questions remain open. In this paper, we examine the propensity of these points to collect
on lines of slope 1.
Keywords. Modular hyperbola, arithmetical function, average order.
AMS classication. 11A07, 11A25, 11N37.
1 Introduction
Let H
n
denote the modular hyperbola H
n
= {(.. ,) : ., = 1(mod n). 1 _
.. , _ n 1}. An important property of these sets is that the sequence {n
1
H
n
}
is uniformly distributed in the unit square. More precisely, if C _ 0. 1|
2
has
piecewise smooth boundary then
lim
n!1
#(C n
1
H
n
)
(n)
= area(C). (1)
(We note that the cardinality of H
n
is (n), where denotes the Euler phi func-
tion.) To prove (1) it sufces to only consider rectangles 1 _ 0. 1|
2
. We can
express #(1n
1
H
n
) as an exponential sum and then invoke bounds for Kloost-
erman sums to obtain that
#(1 n
1
H
n
) = area(1)(n) O(t
2
(n) log
2
(n)
_
n).
where t(n) is the number of positive divisors of n. The limit (1) is an immedi-
ate consequence of this asymptotic formula. The details of this calculation are
elegantly presented in [2, Lemma 1.7].
18 Dennis Eichhorn, Mizan R. Khan, Alan H. Stein and Christian L. Yankov
Using a computer algebra package, such as MAPLE, we can easily generate
graphs of H
n
. A typical example is shown in Figure 1. Such pictures provide
convincing visual evidence of the validity of (1) and we encourage the reader to
generate other examples. The relevant MAPLE code is given in the appendix.
5000
4000
4000
3000
3000
1000
2000
2000 0 1000 5000
0
Figure 1. The graph H
5001
In recent years quantitative forms of (1) have been given in a number of papers,
see [3, 5, 15, 17, 18] and references therein. For example, it follows from general
results of [5] that for primes ,
area(C)
#(C
1
H
p
)
1
= O
_
1=4
log
_
. (2)
where the implied constant depends only on C.
On a whimsical note we observe that from a visual perspective the graphs of
H
n
are particularly interesting when n is small. For such integers, the cardinality
of H
n
, (n), is small and so one can try to identify patterns in the graphs in the
same vein as one looks at clouds in the sky and identies fanciful shapes! (Once
(n) takes on values in the thousands you simply see a mass of points illustrating
the uniform distribution of {n
1
H
n
}.) We give a few of our favorite examples
below.
Sums and Differences of the Coordinates of Points on Modular Hyperbolas 19
40
40
30
10
30 10 20
20
47
80
80
0
60 20
60
40
20
40
0
88
249
In displaying these images, we are delighted to reveal that a buttery, a dragony,
and a scholar all lie hidden in the arithmetic structure of the integers!
Returning to matters mathematical, let D(n). S(n).

D(n) and

S(n) be the fol-
lowing sets:
D(n) = {. , : (.. ,) H
n
}.
S(n) = {. , : (.. ,) H
n
}.
D(n) = {(. ,) mod n : (.. ,) H

n
}.
S(n) = {(. ,) mod n : (.. ,) H

n
}.
The quantities #D(n) and #S(n) count the number of lines, of slope 1 and 1
respectively, that have nonempty intersection with H
n
. The central results of this
paper are precise formulas for #

D(n) and #
S(n).
Since {n
1
H
n
} is uniformly distributed in the unit square, it is natural to be-
lieve that the ratio #D(n),#S(n) should be close to 1 when n is large. Further-
more, it is easy to show that for primes ,
#D()
#S()
= 1
1 (1,)
1
.
where (a,) is the Legendre symbol. (We prove this assertion at the end of this
section.) However whilst looking at some graphs of H
n
(typically with n having
several factors of 2), we were quite surprised to see that there seemed to be many
more lines of slope 1 intersecting the graph than lines of slope 1. Two such
unusual examples are H
1024
and H
1728
.
0 800
600
800
0
200
200 400 1000
400
600
1000
1024
1600
1600
0
1200
1200 400 800
0
400
800
1728
We then used MAPLE to generate some data. In particular we noticed that for
powers of 2, the ratio #D(2
k
),#S(2
k
), with k _ 10, seemed to lie between 4 and
5 (see Table 1). Our numerical work at this juncture suggested the asymptotic
#D(n)
#S(n)
1:
but we eventually proved that
liminf
n!1
#D(n)
#S(n)
= 0 and limsup
n!1
#D(n)
#S(n)
= o.
a result completely contrary to our initial intuition and belief!
m 4 8 16 32 64 128 256 512 1024 2048 4096 8192 16384
#D.m/ 1 1 3 5 13 21 53 97 205 393 797 1549 3089
#S.m/ 2 4 4 8 8 12 16 28 44 84 162 328 652
Table 1. Values of #D(m) and #S(m) with m = 2
t
, t = 2. . . . . 14
In the course of our work we realized that we could apply the Chinese Remain-
der Theorem to the sets

D(n) and

S(n) and consequently determine formulas for
#

D(n) and #
S(n). This is not the case for the sets D(n) and S(n); but our formu-
lae for #

D(n) and #
S(n), in conjunction with the inequalities

#

D(n) _ #D(n) _ 2#

D(n) and #
S(n) _ #S(n) _ 2#
S(n). (3)
give us upper and lower bounds for #D(n). #S(n) and related ratios. For ex-
ample they allow us to prove that 3 _ #D(2
t
),#S(2
t
) _ 12, for t large. Two
interesting consequences of the formulas are that the mean-value of c(n), where
c(n) = #
S(n),#

D(n), is approximately 1.3; and for more than 80% of all inte-
gers c(n) > 1.
We end this section by proving our earlier assertion that for primes
#D()
#S()
= 1
1 (1,)
1
.
Proposition 1. For primes > 2,
#S() =
1
2
(4)
and
#D() =
(1,)
2
. (5)
where (a,) denotes the Legendre symbol.
Proof. Let k S() and let (a. b) l
k
H
p
, where l
k
denotes the line . , =
k. It is easy to check that a is a root of .
2
k. 1 = 0 (mod ). Since any
quadratic congruence modulo a prime has at most two roots, we conclude that
1 _ #(l
k
H
p
) _ 2. Now . = , is a line of symmetry of H
p
and therefore
(b. a) l
k
H
p
. If a = b then l
k
H
p
= {(a. a)}, and if a = b then
l
k
H
p
= {(a. b). (b. a)}.
There are two of the former, {(1. 1)} and {( 1. 1)}, and ( 3),2 of
the latter, so #S() = ( 1),2.
The proof of (5) is similar. We look at lines . , = k, and since .
, = is a line of symmetry of H
p
the points again come in pairs, (a. b) and
( b. a). If 1 is a quadratic residue, the counting is exactly the same and
#D() = ( 1),2. If 1 is not a quadratic residue, there are no singleton sets
and #D() = ( 1),2.
The set {[. ,[ (.. ,) H
p
} has been studied in [12] and each result in [12]
has an analogous result for D() and S(). In particular, the above result and
proof is essentially [12, Theorem 1]. As mentioned in the abstract, there are many
interesting questions that one can ask about modular hyperbolas. For a discussion
of recent results and open problems on modular hyperbolas we refer the reader to
the survey article [11].
2 General Strategy
From this point on, will always denote a prime. In this section, we will apply
the Chinese Remainder Theorem to prove that the quantities #

D(n) and #
S(n)
are multiplicative. We will then translate the problem of counting #

D(
k
) and
#
S(
k
) to one of counting squares.
Proposition 2. Let n =
m
iD1
e
i
i
be the canonical factorization of n. Then
#

D(n) =
m
iD1
#

D(
e
i
i
) (6)
and
#
S(n) =
m
iD1
#
S(
e
i
i
). (7)
Proof. Since the proofs (6) and (7) are identical we only prove the former. The
Chinese Remainder Theorem states that the map
: Z
n

m
iD1
Z
p
e
i
i
via
(.) = (. mod
e
1
1
. . . . . . mod
e
m
m
)
is an isomorphism of rings. Consequently, when we restrict to

D(n) we obtain
a map g :

D(n)
m
iD1
D(
e
i
i
). We now show that g is a bijection.
The injectivity of g is clear, so we need to only worry about the surjectivity. Let
(k
1
. . . . . k
m
)
m
iD1
D(
e
i
i
). So there exist (a
i
. b
i
) H
p
e
i
i
, with i = 1. . . . . m,
such that (a
i
b
i
) mod
e
i
i
= k
i
. By the Chinese Remainder Theorem, the two
systems of congruences
. a
i
(mod
e
i
i
). , b
i
(mod
e
i
i
). i = 1. . . . . m.
have a unique solution . = a. , = b modulo n. Since a
i
b
i
1 (mod
e
i
i
)
for i = 1. . . . . m, we have that ab 1 (mod n). Clearly g((a b) mod n) =
(k
1
. . . . . k
m
).
We now translate the problem of counting #

D(
k
) and #
S(
k
) to one of
counting squares. We start with a minor observation.
Lemma 3. Let (a. b) H
p
t . Then a b 2k
1
(mod
t
) and a b 2k
2
(mod
t
) for some k
1
. k
2
Z.
Proof. If = 2 then a. b are both odd. If ,= 2, then 2 is invertible modulo
t
.
Theorem 4. Let (a. b) H
p
t . Then
(i) (2k mod
t
)

D(
t
) (k
2
1) is a square modulo
t
.
Furthermore, the map J
p
t (k) = 2k mod
t
denes a bijection
J
p
t : {k : k
2
1 is a square modulo
t
. 0 _ k <
t
}

D(
t
).
when ,= 2.
For the special case = 2, the map J
2
t (k) = 2k mod
t
denes a bijec-
tion
J
2
t : {k : k
2
1 is a square modulo 2
t
. 0 _ k < 2
t1
}

D(2
t
).
(that is, we restrict the elements of the domain to lie between 0 and 2
t1
1).
(ii) (2k mod
t
)

S(
t
) (k
2
1) is a square modulo
t
.
Furthermore, the map s
p
t (k) = 2k mod
t
denes a bijection
s
p
t : {k : k
2
t
. 0 _ k <
t
}

S(
t
).
when ,= 2.
For the special case = 2, the map s
2
t (k) = 2k mod
t
denes a bijection
s
2
t : {k : k
2
t
. 0 _ k < 2
t1
}

S(2
t
).
(that is, we restrict the elements of the domain to lie between 0 and 2
t1
1).
Proof. Since the proofs of the two parts are identical, we only prove the result for
D(
t
).
Let (a. b) H
p
t . By Lemma 3, a b 2k (mod
t
) for some k Z. Upon
completing the square, we obtain k
2
1 (a k)
2
(mod
t
). Conversely, if
k
2
1 is a square, then there exists c Z such that c
2
k
2
1 (mod
t
). It
follows that
(a. b) = ((c k) mod
t
. (c k) mod
t
) H
p
t .
and a b 2k (mod
t
).
If ,= 2, then 2 is invertible modulo
t
, and consequently
J
1
p
t
(.) = 2
1
. mod
t
.
The case when = 2 is slightly more involved. Let k be an integer, with
0 _ k < 2
t
, such that k
2
t
. It follows immediately
that for the integer k
1
= (k 2
t1
) mod 2
t
, k
2
1
1 is also a square modulo 2
t
.
The congruence 2. 2k (mod 2
t
) has precisely two distinct solutions, which
must be k and k
1
. Since either k or k
1
is less than 2
t1
, we conclude that J
2
t is a
bijection.
From this we see that counting #

D(
t
) or #
S(
t
) is equivalent to counting
the ks such that k
2
1 and k
2
1 are squares. In this context, we will on two
separate occasions invoke the following formulas of Stangl [13].
Theorem 5 (Stangl). Let be an odd prime. Then
#{k
2
mod
t
} =
tC1
2( 1)
(1)
t1
1
4( 1)
3
4
. (8)
For the special case = 2 we have that
#{k
2
mod 2
t
} =
2
t1
3
(1)
t1
6
3
2
. t _ 2. (9)
Finally, we will need the following criteria concerning the solvability of qua-
dratic congruences. (See [8, Propositions 4.2.3, 4.2.4, page 46].)
Proposition 6. For the congruence
.
2
a (mod
t
)
where is prime and a is an integer such that ,[ a, we have the following:
(i) ,= 2 : If the congruence .
2
a (mod ) is solvable, then for every
t _ 2 the congruence .
2
a (mod
t
) is solvable with precisely 2 distinct
solutions.
(ii) = 2 : If the congruence .
2
a (mod 2
3
) is solvable, then for every
t _ 3 the congruence .
2
a (mod 2
t
) is solvable with precisely 4 distinct
solutions.
3 The Formulas for #
N
S.p
t
/ and #
N
D.p
t
/
3.1 Case n D 2
t
In this section we determine the cardinality of

D(2
t
) and

S(2
t
).
Theorem 7. The cardinality of the set

D(2
t
) is
#

D(2
t
) =
_
1. 1 _ t _ 3.
2
t3
. t _ 4.
(10)
Proof. Direct computations show that the result is true for t _ 4. So we assume
that t _ 5. By Theorem 4,
#

D(2
t
) = #{k : k
2
t
. 0 _ k < 2
t1
}.
We claim that
k
2
t
=k = 4l for some l Z.
We obtain the (=) direction by reducing modulo 8 and observing that k
2
1 is
a square modulo 8 if and only if k 0 (mod 8) or k 4 (mod 8). To obtain
the (=) direction we note that .
2
16l
2
1 (mod 8) is solvable for any l, and
therefore by the second part of Proposition 6, (4l)
2
1 = 16l
2
1 is a square
modulo 2
t
for all l. Hence,
{k : k
2
t
. 0 _ k < 2
t1
} = {4l : 0 _ l < 2
t3
}.
and therefore #

D(2
t
) = 2
t3
.
Theorem 8. The cardinality of the set

S(2
t
) is
#
S(2
t
) =
_
_
1. t = 1. 2.
2. t = 3. 4.
2
t4
3

.1/
t1
3
3. t _ 5.
(11)
Proof. Direct computations show that the result is true for t _ 6 and so we may
assume that t _ 7. We will prove that
#{k : k
2
t
. 0 _ k < 2
t1
} = 2 #{k
2
mod 2
t4
}
and conclude by applying (9).
Let {k : k
2
t
. 0 _ k < 2
t1
}. Since
2
1 is a
square modulo 2
t
, we have that = 2l 1 for some l. 0 _ l < 2
t2
. (Otherwise
we would have that 1 is a square modulo 4.) It follows that (l
2
l) is a square
modulo 2
t2
and so we can conclude that
#{k : k
2
t
. 0 _ k < 2
t1
}
= #{l : l
2
l is a square modulo 2
t2
. 0 _ l < 2
t2
}.
The set {l : l
2
t2
. 0 _ l < 2
t2
} is the union of
{l : l
2
t2
. 0 _ l < 2
t2
. l odd}
and
{l : l
2
t2
. 0 _ l < 2
t2
. l even}.
Since l
2
l (2
t2
1 l)
2
(2
t2
1 l) (mod 2
t2
),
#{l : l
2
t2
. 0 _ l < 2
t2
. l odd}
= #{l : l
2
t2
. 0 _ l < 2
t2
. l even}.
Therefore,
#{l : l
2
t2
. 0 _ l < 2
t2
}
= 2#{l : l
2
t2
. 0 _ l < 2
t2
. l odd}.
Now
#{l : l
2
t2
. 0 _ l < 2
t2
. l odd}
= #{l
1
1 : l
1
t2
. 0 _ l < 2
t2
. l odd}.
If l
1
1 is square modulo 2
t2
, then it is a multiple of 4 and consequently
l
1
1 4m
2
(mod 2
t2
)
for some m. 0 _ m < 2
t4
. Consequently,
#{l
1
1 : l
1
t2
. 0 _ l < 2
t2
}
= #{k
2
mod 2
t4
}.
which ends the proof.
3.2 Case n D p
t
, p an Odd Prime
Proposition 9. If 1 (mod 4), then for all t ,
#

D(
t
) = #
S(
t
). (12)
Proof. If 1 (mod 4), then for any value of t the congruence .
2
1
(mod
t
) has a solution, i
p
t . The map 1
p
t : Z
p
t Z
p
t via 1
p
t (.) = i
p
t . is a
bijection.
Let k {k mod
t
: k
2
t
}. Then k
2
1 m
2
(mod
t
) for some m, and consequently
i
2
p
t
k
2
1 (k
2
1) m
2
(i
p
t m)
2
(mod
t
).
The above calculation shows that
1
p
t
_
{k mod
t
: k
2
t
}
_
= {k mod
t
: k
2
t
}.
and therefore by Theorem 4, #

D(
t
) = #
S(
t
).
For the rest of this section we will use the following notation: Let
S
0
(
t
) = {k mod
t
: k
2
t
. ,[ (k
2
1)}.
S
00
(
t
) = {k mod
t
: k
2
t
. [ (k
2
1)}.
D
0
(
t
) = {k mod
t
: k
2
t
. ,[ (k
2
1)}.
and
D
00
(
t
) = {k mod
t
: k
2
t
. [ (k
2
1)}.
We note that the bijections from Theorem 4 imply
#
S(
t
) = #S
0
(
t
) #S
00
(
t
)
and
#

D(
t
) = #D
0
(
t
) #D
00
(
t
).
which explains our notation. In our next theorem we determine #D
0
(
t
) and
#S
0
(
t
) by calculating a sum of Legendre symbols. We then determine #D
00
(
t
)
and #S
00
(
t
) by applying Stangls formula (8).
Theorem 10. Let p be an odd prime, and let (a,) denote the Legendre symbol.
Then
#S
0
(
t
) =
( 3)
t1
2
(13)
and
#D
0
(
t
) =
_
( 3)
t1
,2. 1 (mod 4).
( 1)
t1
,2. 3 (mod 4).
(14)
Proof. The proofs of (13) and (14) are identical and so we will only do the second
one. If l D
0
(
t
), then the Legendre symbol
_
(l
2
1),
_
= 1. Therefore,
#D
0
(
t
) =
1
2
p
t
1
lD0; gcd.l
2
C1;p/D1
__
l
2
1
_
1
_
=
1
2
p
t1
1
kD0
p1
lD0; l
2
6D1 .mod p/
__
(l k)
2
1
_
1
_
=
_
p1
lD0; l
2
6D1 .mod p/
__
l
2
1
_
1
__
t1
2
=
_
1
p1
lD0; l
2
6D1 .mod p/
1
_
t1
2
.
where the 1 term in the last expression arises by invoking
p1
aD0
_
a
2
1
_
= 1.
(see [1, Theorem 2.1.2, page 58]). We complete our proof by noting that
p1
lD0; l
2
6D1 .mod p/
1 =
_
2. 1 (mod 4).
. 3 (mod 4).
Lemma 11. If 3 (mod 4), then
#D
00
(
t
) = 0.
and consequently
#

D(
t
) = #D
0
(
t
) =
(
t
)
2
. (15)
Proof. If 3 (mod 4) then the congruence .
2
1 = 0 (mod ) has no
solutions and consequently
#D
00
(
t
) = 0.
Proposition 12. The cardinality of the set S
00
(
t
) is
#S
00
(
t
) =
_
2. t _ 2.
p
t1
pC1

3
2
(1)
t3
p1
2.pC1/
. t _ 3.
(16)
Proof. Let k S
00
(
t
). Then k
2
1
2
m
2
(mod
t
) for some m. 0 _ m <
t2
. We have two cases to consider.
(i) t _ 2: In this case we have m = 0, and since the congruence .
2
1
(mod
t
) has exactly 2 solutions, we conclude that #S
00
(
t
) = 2.
(ii) t _ 3: In this case we dene a map
S
00
(
t
) {m
2
mod
t2
}
via k m
2
. By Proposition 6, for each m
2
{m
2
mod
t2
}, the con-
gruence .
2

2
m
2
1 (mod
t
) is solvable with precisely two solu-
tions. Hence, the map is surjective with each element in the image hav-
ing exactly two elements in its preimage. Consequently, we can infer that
#S
00
(
t
) = 2#{m
2
mod
t2
}, and we conclude by invoking (8).
On combining all of the pieces we obtain the following formulas for #

D(
t
)
and #
S(
t
).
Theorem 13. The cardinalities of the sets

D(
t
) and

S(
t
) are
#

D(
t
) =
_
_
.p3/p
t1
2
2. 1 (mod 4). t _ 2.
.p3/p
t1
2

p
t1
pC1

3
2
(1)
t3
p1
2.pC1/
. 1 (mod 4). t _ 3.
.p1/p
t1
2
=
'.p
t
/
2
. 3 (mod 4).
(17)
and
#
S(
t
) =
_
.p3/p
t1
2
2. t _ 2.
.p3/p
t1
2

p
t1
pC1

3
2
(1)
t3
p1
2.pC1/
. t _ 3.
(18)
4 Some Properties of c.n/ D #
N
S.n/=#
N
D.n/
To simplify notational matters, let c(n) = #
S(n),#

D(n). We use our formulas
for #
S(n) and #

D(n) to determine some properties of c(n). We begin with some
elementary, but useful remarks about c(
k
).
Lemma 14. c(
k
) has the following properties.
(i) For 1 (mod 4),
c(
k
) = 1. (19)
(ii) For 3 (mod 4).
c() =
_
1
2
1
_
. (20)
(iii) For 3 (mod 4) and k _ 2.
c(
k
) c(
k1
) =
_
4
p
k1
. k even.
2
p
k1
. k odd.
(21)
(iv) For k _ 6,
c(2
k
) c(2
k1
) =
_
1
2
k5
. k even.
1
2
k4
. k odd.
(22)
(v) For k _ 1,
1,6 _ c(2
k
) _ 2. (23)
(vi) For k _ 2,
1
4
_ c(3
k
) _
2
3
. (24)
(vii) For _ 5. 3 (mod 4) and k _ 2,
_
1
2
1
_
_ c(
k
) _
_
1
2
1
2

_
< 1. (25)
Proof. These are immediate consequences of formulas (17) and (18). We make a
few remarks: the maximum value of c(2
k
) is c(8) = 2; in (25) the upper bound
is achieved when k = 2; if 3 (mod 4) then for any k _ 2, c(
k
) < c().
Finally, the reason we treated the case of 3
k
separately is that the LHS of (25)
equals 0 when = 3. We could have merged (24) and (25) by replacing the LHS
of (25) with
_
1
2
1
2
1
_
.
We now prove that the maximal and minimal order of c(n) is log log n and
1, log log n respectively.
Theorem 15. Let N
k
=
k
iD1
i
, where
i
is the i -th prime that is congruent to
3 modulo 4.
(1) We have
c(N
k
) log log N
k
: (26)
and for any t _ 2,
c(N
t
k
) (log log N
k
)
1
. (27)
Consequently,
limsup
n!1
c(n) = o. (28)
and
liminf
n!1
c(n) = 0. (29)
(2) Furthermore,
c(n) log log n (30)
and
c(n) ;
1
log log n
. (31)
Proof. For (1), we only prove (26) as the proof of (27) is similar. Since
c(N
k
) =
k
iD1
_
1
2
i
1
_
.
we have by the Taylor series of log(1 .) that
log c(N
k
) = 2
k
iD1
1
i
1
iD1
1
mD2
(1)
m1
m
_
2
i
1
_
m
. (32)
For _ 5,
4
( 1)
2
1
mD2
(1)
m1
m
_
2
1
_
m2
_
4
( 1)
2
1
mD0
2
m
=
8
( 1)
2
:
and therefore the double series
C =
1
iD1
1
mD2
(1)
m1
m
_
2
i
1
_
m
converges since the i = 1 termis a convergent alternating series and the remaining
terms are bounded in magnitude by
1
iD2
8i
2
.
Since C < 0, we obtain from (32) the inequality
2
k
iD1
1
i
C _ log c(N
k
) _ 1 2
k
iD1
1
i
. (33)
We now apply to (33) Mertenss formula (see [4, Section 2.2]) for primes 3
(mod 4),
lim
k!1
_
_
k
iD1
1
1
2
log log
k
_
_
= 0.048239 . . . .
and obtain
log(c(N
k
)) log log(
k
) = O(1). (34)
The prime number theorem for arithmetic progressions for primes 3
(mod 4) implies
lim
k!1
log N
k
k
=
1
2
.
Consequently log log log N
k
log log
k
= o(1) and so we can replace log log
k
with log log log N
k
in (34) and obtain the desired conclusion that log(c(N
k
))
log log log N
k
= O(1).
The proof of (27) is nearly identical. The main difference is that we start with
the inequality
1
4
k
iD2
_
1
2
i
1
_
_ c(N
t
k
) _
k
iD1
_
1
2
i
1
2

_
.
that we obtain by applying the inequalities (24) and (25).
We next prove item (2). If n has no prime factors congruent to 3 modulo 4,
then we have the inequality 1,6 _ c(n) _ 2. So without loss of generality we
may assume that q
1
. . . . . q
k
are the distinct prime factors of n that are congruent
to 3 modulo 4. Clearly n _ q
1
. . . q
k
_ N
k
. From (20) and (26) we get that
c(n) _ 2c(N
k
) log log N
k
_ log log n.
We now prove (31). An immediate consequence of the asymptotic
2
k
iD1
1
i
log log log N
k
= O(1)
is that
1
4
k
iD2
_
1
2
i
1
_
1
log log N
k
.
We combine this with the inequality
c(n) _
1
24
k
iD2
_
1
2
q
i
1
_
_
1
24
k
iD2
_
1
2
i
1
_
to obtain (31).
Corollary 16. For the sequence #S(n),#D(n) we have
liminf
#S(n)
#D(n)
= 0 and limsup
#S(n)
#D(n)
= o. (35)
Proof. Since #

D(n) _ #D(n) _ 2#

D(n) and #
S(n) _ #S(n) _ 2#
S(n), we
obtain the inequality
0.5c(n) _
#S(n)
#D(n)
_ 2c(n).
We now apply (29), (28) to obtain (35).
Corollary 17. The Dirichlet series
1
nD1
c(n)n
s
converges absolutely in the
half-plane m(s) > 1.
Proof. This is an immediate consequence of (30).
We preface our calculation of the mean value of c(n) with the following ob-
servation about mean values of arithmetical functions. Let
M(g) = lim
x!1
1
.
nx
g(n)
denote the mean value of g. If M(g) exists, then, via partial summation, we have
that
M(g) = lim
s!1
C
1
nD1
g(n)n
s
(s)
.
Thus if g is multiplicative then we can represent M(g) by the innite product
M(g) =
p prime
_
1
1
_
_
1
kD0
g(
k
)
k
_
=
p prime
_
1
1
kD1
g(
k
) g(
k1
)
k
_
.
Since we have explicit formulas for c(
k
) we can easily compute M(c) provided
we rst show that this mean value exists. Arguably the simplest proof of the
existence of M(c) is to invoke Wintners mean-value theorem for multiplicative
functions.
Theorem 18 (Wintner). If g is a multiplicative function satisfying the conditions
p prime
[g() 1[
< oand
p prime
1
kD2
[g(
k
) g(
k1
)[
k
< o.
then M(g) exists.
This was rst proved in [16] a monograph that is very difcult to nd. For a
more recent and accessible reference see [10, II.2, Corollary 2.3]. The proof is a
straightforward convolution argument.
Theorem 19. The mean-value of c(n), M(c), is given by the innite product
M(c) = lim
x!1
1
.
nx
c(n) =
337
320
p3 .mod 4/
_
1
2(
2
1)
4
1
_
~ 1.32.
(36)
Proof. From the properties of c(
k
) listed in Lemma 14 the series
p prime
[c() 1[
and
p prime
1
kD2
[c(
k
) c(
k1
)[
k
are both convergent and therefore by Wintners theorem we conclude that M(c)
exists. We now use (10), (11), (19), (20), (21) and (22) to obtain M(c).
Our nal result shows that c(n) > 1 for over 80% of all integers. We prove it
by applying Wirsings mean-value theorem for multiplicative functions [14, III.4,
Theorem 5].
Theorem 20 (Wirsing). If g is a real multiplicative function with [g(n)[ _ 1 for
all n Z
C
, then M(g) exists.
Wirsings theorem is a deep theorem. For example, it contains the Prime Num-
ber Theorem in its equivalent form M(j) = 0, where j is the Mbius function,
see [7, Section 3]. We would have preferred to have used a simpler result such
as Wintners theorem; however, the condition that
p prime
[g() 1[
1
is con-
vergent is not satised in one part of our argument. We will need the following
lemma to justify the use of Wirsings theorem.
Lemma 21. For each prime , let
p
(n) denote the exponent of the prime in the
canonical factorization of n and let
p
denote a non-empty subset of Z
C
L {0}.
Then the characteristic function of the set {n :
p
(n)
p
}L{1} is multiplica-
tive.
Proof. We construct a function : Z
C
{0. 1} in the following way. Let
(1) = 1 and, for k _ 1, let
(
k
) =
_
1. k
p
.
0. k ,
p
:
and
(n) =
l
(
e
l
l
).
where
l

e
l
l
is the canonical factorization of n. Clearly, is both multiplicative
and also the characteristic function of {n :
p
(n)
p
}L{1}.
Theorem22. Let C = {n : c(n) > 1} and let denote the characteristic function
of C. Then the lower density of C, liminf .
1
nx
(n), satises the inequality
liminf
x!1
1
.
nx
(n) _
63
64
p3 .mod 4/
_
1
1
2
_
~ 0.84.
Furthermore, for any positive constant 1, the set {n : c(n) _ 1} has positive
lower density.
Proof. Let C
1
. C
2
. C
3
be the sets
C
1
= {n :
2
(n) _ 5. and
p
(n) _ 1 if 3 (mod 4)}.
C
2
= {n : n C
1
.
2
(n) ,= 3. and
p
(n) = 0 if 3 (mod 4)}.
and C
3
= C
1
\ C
2
; and let
1
and
2
be the characteristic functions of C
1
and C
2
respectively. The conditions on C
1
ensure that for any n C
1
and for any prime
, c(
v
p
.n/
) _ 1. Therefore for any n C
1
we have c(n) _ 1 with equality
precisely when n C
2
, showing that C
3
C.
By Lemma 21,
1
and
2
are multiplicative functions and so we can apply
Wirsings theorem to obtain that
density(C
3
) = M(
1

2
) = M(
1
) M(
2
)
=
p prime
_
1
1
_
_
1
1
iD1
1
(
i
)
i
_
p prime
_
1
1
_
_
1
1
iD1
2
(
i
)
i
_
=
63
64
p3 .mod 4/
_
1
1
2
_
0 =
63
64
p3 .mod 4/
_
1
1
2
_
.
(We cannot invoke Wintners theorem here as the series
p
[
2
() 1[
1
is
divergent.) By our formulas for #

D(n) and #
S(n) we have the inclusion C

3
_ C,
and so we can conclude that the lower density of C is greater or equal to the
density of C
3
.
A slight variation of the above proof gives the second assertion. Recalling the
notation in Theorem 15, let
i
denote the i -th prime congruent to 3 modulo 4 and
let N
k
=
k
iD1
i
. Since c(N
k
) log log N
k
(see asymptotic (26)), we can nd
an integer l such that c(N
k
) _ 1 for k _ l. Let
L
1
= {n :
2
(n) = 0 and
p
i
(n) _ 1 for i = 1. 2. . . . }.
L
2
= {n : n L
1
and
p
i
(n) = 0 for i = 1. . . . . l}.
and L
3
= L
1
\L
2
. A slight modication of our earlier calculation of density(C
3
)
gives that
density(L
3
) =
1
2
_
_
1
l
iD1
1
1
1
p
i
_
_
p3 .mod 4/
_
1
1
2
_
.
We conclude by observing that for any n L
3
, c(n) _ c(N
l
) _ 1.
4.1 Unanswered Questions and Ongoing Work
We have not resolved the following 3 questions.
(i) Does the density of A = {n Z
C
: c(n) = 1} equal 0? This would follow
if we could prove that for any n A, the odd prime factors of n are all
congruent to 1 mod 4.
(ii) What is the density of {n Z
C
: c(n) < 1}? Is it non-zero?
(iii) What is the normal order of c(n)?
It is easy to generalize Proposition 2 to arbitrary polynomials in Z.. ,|. Spe-
cically, if Z.. ,| and we dene the map
n
: H
n
Z
n
via
n
((.. ,)) =
(.. ,) mod n, then the quantity #Image(
n
) is a multiplicative function of n.
One possible extension of our work is to determine formulas for #Image(
p
e ) for
some other polynomials Z.. ,|, especially for cases where one can apply the
formulas in [13] and this paper. S. Hanrahan, under the supervision of M. Khan,
is currently writing an undergraduate honors thesis on this topic for the quadratic
forms .
2
,
2
and .
2
,
2
.
5 Appendix
This is the MAPLE code that generated the graph of H
5001
.
n:=5001:
a:=array(1..numtheory[phi](n)):
b:=array(1..numtheory[phi](n)):
count:=1:
for i from 1 to n-1 do;
if gcd(i,n)=1 then
a[count]:=i: b[count]:=(i^(-1)mod n):
count := count+1;
end if;
end do;
printf("n=\%d, no. of points on graph=\%d \n",n,count-1):
points := zip((x,y) -> [x,y],a,b):
p1:=plot(points,style=POINT,symbol=CROSS):
plots[display](p1);
References
[1] B. C. Berndt, R. J. Evans and K. S. Williams, Gauss and Jacobi Sums, Wiley, 1997.
[2] F. Boca, C. Cobeli and A. Zaharescu, Distribution of Lattice Points Visible from the
Origin, Commun. Math. Phys. 213 (2000), 433470.
[3] C. Cobeli and A. Zaharescu, On the Distribution of the F
p
-points on an Afne Curve
in r Dimensions, Acta Arithmetica 99 (2001), 321329.
[4] S. R. Finch, Mathematical Constants, Encylopedia of Mathematics and its Applica-
tions 94, Cambridge, 2003.
[5] A. Granville, I. E. Shparlinski and A. Zaharescu, On the Distribution of Rational
Functions Along a Curve over F
p
and Residue Races, J. Number Theory 112 (2005),
216237.
[6] G. H. Hardy and E. M. Wright, An Introduction to the Theory of Numbers, 5th ed.,
Oxford, 1985.
[7] A. Hildebrand, Some New Applications of the Large Sieve, in: Number Theory
(New York 19851988), pp. 7688, Lecture Notes in Mathematics 1383, Springer-
Verlag, 1989.
[8] K. Ireland and M. Rosen, A Classical Introduction to Modern Number Theory,
Springer-Verlag, 1982.
[9] H. L. Montgomery and R. C. Vaughan, Multiplicative Number Theory I. Classical
Theory, Cambridge, 2007.
[10] W. Schwarz and J. Spilker, Arithmetical Functions, Cambridge, 1994.
[11] I. E. Shparlinski, Distribution of Points on Modular Hyperbolas, Sailing on the Sea
of Number Theory: Proc. 4th China-Japan Seminar on Number Theory, Weihai,
2006, World Scientic, 2007, 155189.
[12] I. E. Shparlinski and A. Winterhof, On the Number of Distances Between the Co-
ordinates of Points on Modular Hyperbolas, J. Number Theory 128 (2008), 1224
1230.
[13] W. Stangl, Counting squares in Z
n
, Math. Mag. 69 (1996), 285289.
[14] G. Tenenbaum, Introduction to Analytic and Probabilistic Number Theory, Cam-
bridge, 1995.
[15] M. Vajaitu and A. Zaharescu, Distribution of Values of Rational Maps on the F
p
-
points on an Afne Curve, Monathsh. Math. 136 (2002), 8186.
[16] A. Wintner, Eratosthenian Averages, Baltimore, 1943.
[17] W. Zhang, On the Distribution of Inverses Modulo n, J. Number Theory 61 (1996),
301310.
[18] Z. Zheng, The Distribution of Zeros of an Irreducible Curve over a Finite Field, J.
Number Theory 59 (1996), 106118.
Author information
Dennis Eichhorn, Department of Mathematics, University of California,
Irvine, CA 92697, USA.
E-mail: deichhor@math.uci.edu
Mizan R. Khan, Department of Mathematics and Computer Science, Eastern Connecticut
State University, Willimantic, CT 06226, USA.
E-mail: khanm@easternct.edu
Alan H. Stein, Department of Mathematics, University of Connecticut,
Waterbury, CT 06702, USA.
E-mail: stein@math.uconn.edu
Christian L. Yankov, Department of Mathematics and Computer Science, Eastern Con-
necticut State University, Willimantic, CT 06226, USA.
E-mail: yankovc@easternct.edu
Self Generating Sets and
Numeration Systems
David Garth, Joseph Palmer and Ha Ta
Abstract. Kimberling has studied a variety of sets generated in the following way. Let
J be a countable set of functions, and let S = S
F
be the smallest set containing 0 that
is closed under any function in J. When J = {2.. 4. 1}, the resulting set S is pre-
cisely the set of nonnegative integers whose binary expansion does not contain the block
11. The set of all such binary expansions corresponds to the set of greedy representa-
tions of the natural numbers with respect to the Fibonacci sequence. Other examples of
a similar nature can be found in the literature. In this paper we explore the following
question; which self generating sets consist of integers whose digit expansions in base
two correspond to the digit expansions of the natural numbers with respect to a linearly
recurrent base sequence? We study this problem in the framework of abstract numeration
systems. That is, we consider an abstract numeration system as an innite language over
a nite alphabet, ordered under a genealogical ordering. We then dene a self generating
numeration system as one that can be realized as the set of base two expansions of the
integers in some self generating set. Our rst result is to prove a necessary and sufcient
condition for an abstract numeration system to have a base. This result is then used to
prove that certain families of generating functions give rise to self generating numeration
systems that have a base sequence. Finally, we prove that the base sequence in any based
self generating numeration system satises a linear recurrence. Many of our results make
use of a natural tree structure that can be put on an abstract numeration system.
Keywords. Numeration systems, greedy expansions, lazy expansions, self generating
sets.
AMS classication. 11B13, (11B85).
1 Introduction
The purpose of this paper is to explore some of the connections between self
generating sets and abstract numeration systems. Kimberling ([9], [10], [11]) has
studied a variety of sets generated in the following way. Let J be a countable set
The work of the rst and second authors was supported in part by NSF Grant No. 0431664.
42 David Garth, Joseph Palmer and Ha Ta
of functions, and let S = S
F
be the smallest set containing 0 that is closed under
all the functions in J. In other words, 0 S, and if . S and J then
(.) S. Moreover, no other elements are in S.
As an example, the set S generated by J = {2.. 4. 1} has been consid-
ered in the literature ([1], [10], [8]). Since 2. simply adds a zero to the binary
expansion of ., and since 4. 1 adds a 01, the ordered sequence
{0. 1. 2. 4. 5. 8. 9. 10. 16. . . . }
of elements of S consists precisely of those integers whose binary expansions have
no adjacent ones. Allouche, Shallit, and Skordev [2] studied the self generating
set S arising from J = {2. 1. 4. 2} and showed that in this case
S = {0. 1. 2. 3. 5. 6. 7. 10. 11. 13. 14. 15. . . . }.
the set of natural numbers whose binary expansions do not contain the block 00.
The connection with numeration systems becomes apparent when considering
the so-called greedy representations of the natural numbers with respect to some
sequence. More precisely, let {b
i
}
i0
be a strictly increasing sequence of natural
numbers, with b
0
= 1. The sequence {b
i
} serves as a base for a positional numer-
ation system for the natural numbers as follows. For n _ 1, let k _ 0 be such that
b
k
_ n < b
kC1
. Using the division algorithm we write
n = q
k
b
k
r
k
. where 0 _ r
k
< b
k
.
Thus, q
k
= ]
n
b
k
where ] is the greatest integer function. For i = k 1. . . . . 0
let q
i
= ]
r
iC1
b
i
and r
i
= r
iC1
b
i
q
i
. It follows that n = q
0
b
0
q
k
b
k
,
and we say that the string q
k
q
k1
q
1
q
0
is the greedy representation of n with
respect to the sequence {b
i
}. We dene the greedy representation of 0 to be 0,
although most authors dene it to be the empty word.
Suppose {b
i
} is the Fibonacci sequence, indexed as b
0
= 1. b
1
= 2, and
b
i
= b
i1
b
i2
for i _ 2. Using the above algorithm it is not hard to show that
the set of greedy representations of the natural numbers in this case consists of 0
and all words over {0. 1} that begin with 1 and do not contain the block 11. This
is precisely the set of binary expansions of the elements of the set S generated by
J = {2.. 4. 1}.
The lazy representation of a natural number n with respect to the Fibonacci
sequence is obtained by successively replacing all occurrences of the string 100
in the greedy Fibonacci representation of n with the string 011 until the resulting
string contains no occurrences of the block 00. Any leading zeros that arise in
this process are disregarded. Thus, for J = {2. 1. 4. 2}, the set of base
Self Generating Sets and Numeration Systems 43
2 expansions of elements of S is the set of lazy Fibonacci representations of the
natural numbers.
These examples are easy to generalize, and some examples are considered in
[8]. For example, if J = {2.. 4. 1. 8. 3}, then
S
F
= {0. 1. 2. 4. 8. 9. 10. 16. 17. 18. 32. 33. 34. 36. . . . }
consists of the natural numbers whose binary expansions do not contain the block
111. These binary expansions correspond to the greedy expansions of the natural
numbers with respect to the base {b
i
} of Tribonacci numbers, where b
0
= 1, b
1
=
2, b
2
= 4, and b
i
= b
i1
b
i2
b
i3
for i _ 3. If J = {2.1. 4.2. 8.4},
then the elements of the resulting self generating set S are precisely those natural
numbers whose binary expansions do not contain the block 000. The set of all
such expansions corresponds to the set of lazy Tribonacci representations of N
that are dened in a manner analogous to the lazy Fibonacci representations of N.
In this paper we will place these results in a more general context. In Section
2 we dene the notion of an abstract numeration system. Under this denition,
the set of base two expansions of the elements of any self generating set can be
considered as the set of representations of the natural numbers in such a system.
We will then prove a result that will be useful for determining whether such an
abstract numeration system has a base sequence. In Section 3 we apply this result
to give conditions on the set of generating functions J which guarantee that the
binary expansions of the elements of S
F
correspond to the digit expansions of the
natural numbers with respect to some base sequence {b
i
}. Of natural interest is the
question of whether the base sequence satises a linear recurrence. In Section 4
we show that for any set of generating functions J, if the binary expansions of the
elements of S
F
correspond to the expansions of the natural numbers with respect
to some base sequence, then the base sequence must satisfy a linear recurrence
relation.
2 Numeration Systems
Non-standard numeration systems and more generally the so-called abstract nu-
meration systems have been considered in a variety of settings in the literature
([4], [3], [5], [7], [12], [13], [14], [15]). Generally such a numeration system is
regarded as a set S of words over some alphabet along with a bijection between
that set of words and the natural numbers. The alphabet is the set of allowable
digits, the words in S are the valid representations of the natural numbers, and the
bijection gives a means for obtaining the numerical value of a given representa-
tion. In this paper we take our digit set to be
2
= {0. 1}. To dene our bijection
we recall a few denitions. Let
2
denote the set of all nite words over
2
, and
let
C
2
be the set of nonempty words over
2
. For n
2
, let [n[ denote the
length of n. The radix order [13] on
2
is the ordering where, for n.
2
,
n < if [n[ < [[ or if [n[ = [[ and n = uan
0
and = ub
0
with a. b
2
and a < b in the natural order on
2
. Enumerating the elements of S under the
radix order induces a natural order preserving bijection N : S N. (We will use
the convention that 0 N). We are now ready for our denition of a numeration
system.
Denition 1. A numeration system is an ordered pair (S. N), where S is an in-
nite subset of (
C
2
\ 0
2
) L {0} and N is the natural order preserving bijection
from S to N that maps the (n 1)
st
word of S to n. The map N is referred to as
the evaluation map, and N
1
is the representation map. If N() = n, then is
the S-representation of n.
A few remarks about this denition are in order. First, while the denition
generalizes naturally to allow for more general digit sets, we will consider only
numeration systems with digit set
2
. Also, since the evaluation map will always
be obtained from the radix order, we therefore write S instead of (S. N). Sec-
ond, notice that our denition requires that 0 S. While this restriction is not
necessary in general, it is natural for our purposes. Finally, our condition that
S _ (
C
2
\ 0
2
) L {0} ensures that for n _ 1 the S-representation of n begins
with 1. More general numeration systems that allow for leading zeros have been
considered elsewhere in the literature (e.g. [12], [14]), but will not be considered
here.
Suppose S is a numeration system, and let T = {b
i
}
1
iD0
be an arbitrary
sequence of natural numbers. Let
B
: S N be the function that assigns
n = n
k
n
k1
n
0
S to
B
(n) =
k
X
iD0
b
i
n
i
. (1)
A numeration systemS is based if there exists a strictly increasing sequence T =
{b
i
}
1
iD0
of natural numbers, with b
0
= 1, such that N(n) =
B
(n) for all n S.
In this case, the sequence {b
i
} is called the base sequence of S. The following
lemma will be important in Sections 3 and 4.
Lemma 1. If S is a based numeration system, then 1 S.
Proof. Let T = {b
i
} be the base sequence. By denition b
0
= 1. Since S is a
numeration system, there exists a n S for which N(n) = 1. Since S is based,
B
(n) = N(n) = 1. Since {b
i
} is an increasing sequence, if [n[ > 1 then
B
(n) > 1. It follows then that [n[ = 1, and therefore n = 1. Thus, 1 S.
We say a based numeration system S is greedy if whenever n S {0} and

2
{0} with
N(n) =
B
(n) =
B
().
it follows that n _ under the radix order. In other words, for every n _ 1 the S-
representation of n is the largest possible representation under the radix order. The
numeration systemS is said to be lazy if for every n _ 1 the S-representation of n
is the smallest possible representation. The greedy and lazy Fibonacci representa-
tions mentioned in Section 1 provide examples of these denitions. In general, if
{b
i
} is the base sequence in a greedy numeration systemS and if sup
i0
b
iC1
b
i
< 2
then the digits in the S-representation for n N can be obtained by the algorithm
given in Section 1.
It is natural to consider numeration systems with the property that n0 S\{0}
whenever n S \ {0}. Such a numeration system is said to be right extendable
(see [13], Section 7.3.2). Similarly, S is a Bertrand numeration system if n
S \ {0}= n0 S \ {0} [4]. The set of greedy Fibonacci representations of
the natural numbers mentioned in Section 1 is a Bertrand numeration system,
while the set of lazy Fibonacci representations is not even right extendable. The
following lemma gives a necessary condition for a numeration system that is also
right extendable to be based.
Lemma 2. Let S be a right extendable based numeration system. If n. n1 S
then n01 S.
Proof. Let T = {b
i
} be the base sequence for S. Since S is right extendable, it
follows that n00 S and n10 S. It must be true then that
b
1
=
B
(10)
B
(0) =
B
(n10)
B
(n00) = N(n10) N(n00).
If n01 S, then N(n10) N(n00) = 1, which implies that b
1
= 1. However,
since b
1
> 1 we have a contradiction, and so n01 S.
To any numeration system S we associate a graph T(S) with vertex set S and
edge set
{(. j) : j {0. 1} and . j S}.
In other words, we draw an edge between and j whenever both are members
of S. This graph structure clearly partitions S into a collection of trees. To every
tree t in T(S) we dene the root of t to be the vertex in t of minimal length
under the radix order. Notice that by denition every vertex in T(S) has at most
two children. We say that for n S, if n0 S, then n0 is the left child of n,
and if n1 S, then n1 is the right child of n. Whenever S is right extendable,
every vertex in T(S) has at least one child. Numeration systems in which T(S)
is a single rooted tree are of particular interest.
Denition 2. Let S be a Bertrand numeration system, with 1 S. If T(S) is such
that n S whenever n1 S, then S is treelike.
The conditions of the denition guarantee that T(S) is a single rooted tree.
These conditions also imply that if n S then every prex of n is in S. Treelike
numeration systems were introduced in [5]. As an example, Figure 1 shows the
rst few levels of the tree for the greedy Fibonacci numeration system mentioned
in the introduction. The vertices in the tree of Figure 1 are the greedy Fibonacci
representations of the natural numbers. The numbers in parentheses are the stan-
dard base 10 values of these representations.
0, (0)
1, (1)
10, (2)
100, (3)
1000, (5)
10000, (8) 10001, (9)
1001, (6)
10010, (10)
101, (4)
1010, (7)
10100, (11) 10101, (12)
Figure 1. The tree T(S) for the greedy Fibonacci numeration system.
In a given numeration system S, for k _ 0 we dene M
k
to be the word in
S having k digits that is maximal under the radix order. We take M
0
to the the
empty word. Similarly, for k _ 1 let m
k
be the minimal word having k digits in
S. If 1 S it will be convenient to dene m
1
to be 1. If S has the property that
M
k
is a prex of M
kC1
for every k, we let
M = lim
k!1
M
k
. (2)
This M is then referred to as the maximal word associated with S. We dene the
minimal word m of S similarly. It is clear that in a treelike numeration system,
m
k
= 10
k1
.
Our rst theorem establishes a necessary and sufcient condition for a nu-
meration system to be based. We point out that the theorem actually resembles
Theorem 5.2 of [5]. That theorem is restricted to treelike numeration systems, and
so the following theorem is more general. We also comment that the hypotheses
of the theorem guarantee that any two words of length at least 2 share a common
nonempty prex.
Theorem 1. Let S be a right extendable numeration system, and assume 1 S.
Then S is based if and only if for any two consecutive words . n S of length
l _ 2 having maximal common prex it follows that
= 0M
ljpj1
. (3)
n = m
ljpj
.
Proof. Suppose rst that S is based, with base T = {b
i
}. Consider two arbitrary
adjacent words and n in S of the same length l _ 2, with < n. The property
of the theorem clearly holds if l = 2. Suppose that l _ 3, and let be the
maximal common prex of and n. Then we can write
= 0.
k2
.
0
and n = 1,
k2
,
0
.
where k = l [[. Note that k _ 1. If k = 1, then = 0 = 0M
0
and
n = 1 = m
1
, and the property of the theorem is satised. Assume then that
k _ 2. Since and n are adjacent in S, it is clear that N(n)N() = 1. Since the
numeration system is based, it follows that N() =
B
() and N(n) =
B
(n),
where
B
is as dened in (1). Thus
1 =
B
(n)
B
() =
B
(1,
k2
,
0
)
B
(0.
k2
.
0
)
=
B
(1,
k2
,
0
)
B
(.
k2
.
0
).
Since the word .
k2
.
0
has k 1 digits and the word 1,
k2
,
0
has k digits,
it must be true that
.
k2
.
0
= M
k1
= M
ljpj1
and 1,
k2
,
0
= m
k
= m
ljpj
.
Now suppose that S has the property mentioned in the statement of the theo-
rem. For i _ 0 let b
i
= N(10
i
). We will show that {b
i
} is a base for the numera-
tion system. We need to show that N(n) =
B
(n) for every n S. We use in-
duction on the length of n. Clearly N(0) = 0 =
B
(0), and N(1) = 1 =
B
(1).
Let l _ 2, and assume that N(n) =
B
(n) whenever [n[ < l. Let n
1
. . . . . n
m
be
the words in S of length l in increasing lexicographic order under <. We need to
showthat N(n
j
) =
B
(n
j
) for 1 _ _ m. Since S is right extendable it follows
immediately that n
1
= 10 0 = 10
l1
. Therefore N(n
1
) = b
l1
=
B
(n
1
).
Assume 1 _ < m. and that
N(n
1
) =
B
(n
1
). N(n
2
) =
B
(n
2
). . . . . N(n
j
) =
B
(n
j
).
We need to show that N(n
jC1
) =
B
(n
jC1
). Let be the maximal common
prex for n
j
and n
jC1
. If n
j
= 0 and n
jC1
= 1, then
B
(n
jC1
) =
B
(n
j
) 1 = N(n
j
) 1 = N(n
jC1
). Assume then that
n
j
= 0. and n
jC1
= 1,.
where .. , {0. 1}
and 1 _ [.[. [,[ _ l [[ 1. Since n

j
and n
jC1
are adja-
cent and since S is right extendable it follows from the property in the statement
of the theorem that
. = M
ljpj1
and , = 0
ljpj1
.
Thus, by the induction hypothesis,
B
(n
jC1
)
B
(n
j
) =
B
(1,)
B
(0.) =
B
(1,)
B
(.)
= N(M
ljpj1
) N(m
ljpj
) = 1.
Since N(n
jC1
) N(n
j
) = 1, and N(n
j
) =
B
(n
j
) it follows that,
N(n
jC1
) =
B
(n
jC1
).
So for all n of length l, N(n) =
B
(n). And by induction the result follows.
3 Self Generating Numeration Systems
We now return to self generating sets. For a family of functions J, a self generat-
ing set S = S
F
is the smallest set containing 0 that is closed under the functions
in J. A numeration system S is self generating if there exists a collection of
functions J for which
S = {m|
2
: m S
F
}.
where m|
2
is the base 2 expansion of the integer m S
F
. As mentioned in
Section 1, the sets of greedy and lazy representations of the natural numbers with
respect to both the Fibonacci and Tribonacci sequences give rise to numeration
systems that are self generating and based.
In light of these examples it is natural to consider the question of which sets
of afne functions produce self generating numeration systems that are based. In
this section we will show how to construct a large class of such functions. These
generating functions are perhaps the most natural generalization of the sets of
functions generating the greedy Fibonacci and Tribonacci numeration systems.
First, we require that 2. J. Since multiplication of . by 2 adds a 0 to the
binary expansion of ., the resulting numeration system will be right extendable.
Furthermore, since multiplication of . by 2
n
adds n zeros to the binary expansion
of ., it is natural to restrict our attention to those families of functions in which the
coefcient on . is a power of 2. We also require that if (.) = 2
n
.c J, then
0 _ c < 2
n
. Thus, the binary expansion of (.) ends in the binary expansion of
c. Our next lemma gives us a further restriction on J.
Lemma 3. Let J = {
0
.
1
. . . . .
n
} be a family of functions dened as follows.
Let
0
= 2. and for 1 _ i _ n let (.) = 2
k
i
. c
i
, where k
i
_ 1 and
0 _ c
i
< 2
k
i
. Let S
F
be the set generated by J, and let S be the numeration
system {s|
2
: s S
F
}. If S is based then c
i
= 1 for some i .
Proof. If S is based, then by Lemma 1 S must contain 1. By denition of S it
follows that 1 S
F
. Thus, 1 =
i
(0) for some
i
J with 1 _ i _ n.
In light of Lemma 3 we add the assumption that if 2
k
is the smallest coefcient
on . for all functions in J that have an odd constant term, then 2
k
. 1 J.
We will prove that a nite number of iterations of this minimal function on the
function 2. produces a family of functions J that generates a based numeration
systemS. Our method will be to show that the self generating numeration system
S satises the conditions of Theorem 1. Our rst step in this process is to show
that S is treelike.
Lemma 4. Let n _ 1, k _ 2, and let J be the set of functions {
0
. . . . .
n
} where
0
(.) = 2., and
i
(.) = 2
k1
(
i1
(.)) 1 for 1 _ i _ n. (4)
Let S
F
be the set generated by J, and let S be the numeration system {s|
2
: s
S
F
}. Then S is treelike.
Proof. Clearly 1 S. Since 2. J, S is right extendable. Suppose 0 S.
Then there is a function J and an s S
F
such that (s)|
2
= 0. Since (s)
is even, it follows that =
0
. Then = s|
2
S, and therefore S is a Bertrand
numeration system. Now let {0. 1}
be such that 1 S. We need to show

that S. By denition, there exists an s S
F
and an
i
J, where 1 _ i _ n,
such that
i
(s)|
2
= 1. Thus, 2
k1
i1
(s)1|
2
= 1, and so 2
k1
i1
(s)|
2
=
0, and 2
k2
i1
(s)|
2
= . Notice that =
k2
0
(
i1
(s))|
2
. It follows from
the denition of S that S. This completes the proof that S is treelike.
The next lemma enables us to establish that the property of Theorem 1 is sat-
ised. If S and if there exists an J and an s S
F
such that = (s)|
2
then we say that is produced by . Recall also that for _ 0, M
j
is the maximal
word of length in S.
Lemma 5. Let n _ 1, k _ 2, and let J be the set of functions {
0
. . . . .
n
} where
0
(.) = 2., and
i
(.) = 2
k1
(
i1
(.)) 1 for 1 _ i _ n.
Let S be the self generating numeration system generated by J. Let S, and
assume 1 S. If 1 is produced by
1
, then for _ 1, M
j
is the maximal
sequence in S of length [[ having as a prex.
Proof. Since S is treelike, a maximal sequence in S having as a prex corre-
sponds to a path in T(S) beginning at and following the rightmost branches at
each level. Let s S be such that 1 =
1
(s)|
2
= 2
k
s 1|
2
= s|
2
0
k1
1. Let
0
= s|
2
, and dene
i
= 2
k1
i1
(s) 1|
2
=
i
(s)|
2
for 1 _ i _ n.
Also, let
nC1
= 2
k1
n
(s)|
2
. Notice that
i
=
i
(s)|
2
= 2
k1
i1
(s) 1|
2
=
i1
0
k2
1 for 1 _ i _ n.
and
nC1
=
n
0
k1
. It follows from the denition of T(S) then that for 1 _
i _ n,
i
is a right child of 2
k2
i1
(s)|
2
in S. Also, for any integer l _ 1,
2
l
i1
(s)|
2
is a left child of 2
l1
i1
(s)|
2
in S. Thus, for 1 _ i _ n it follows
that the
i
are connected by a path from to
nC1
of length (k 1)(n 1) k.
For 1 _ i _ n, the
i
are right children in this path, and all the other vertices in
this path are left children. Let .
1
.
2
.
j
be the
th
vertex in this path. We show
that this is the maximal word in S of length [[ having as a prex.
Assume to the contrary that ,
1
,
2
,
j
S is lexicographically larger than
.
1
.
2
.
j
. We will consider two cases. First, suppose that .
1
.
j
lies on the
path from to
n
. Let l be the smallest integer such that ,
l
= .
l
. This gives that
.
1
= ,
1
. .
2
= ,
2
. . . . . .
l1
= ,
l1
. .
l
= 0. and ,
l
= 1. (5)
By Lemma 4, S is treelike, so every prex of a word in S is also in S. Thus there
exists a , S and an integer m, with 1 _ m _ n, such that ,
1
,
l
=
m
(,)|
2
.
Now,
m
(,) = 2
k1
m1
(,) 1, and so by (5) we have that 2
k1
m1
(,)|
2
=
,
1
,
2
,
l1
0 = .
1
.
2
.
l
. However, since .
1
.
l
is a left vertex in the
aforementioned path from to
n
, there exist s S
F
and integers i and t with 1 _
i < n and 1 _ t _ k2 such that .
1
.
l
= 2
t
i
(s)|
2
. Thus, 2
k1
m1
(,) =
2
t
i
(s). Therefore 2
kt1
m1
(,) =
i
(s). Since k t 1 _ 1, this is a
contradiction, since
i
(s) is odd for i _ 1. Thus, .
1
.
l
is the maximal word
in the path from to
n
.
Now, assume that .
1
.
l
lies on the path from
n
to
nC1
. Then the argu-
ment of the previous paragraph holds, except that i = n, and 1 _ t _ k 1.
We have shown then that for _ (k 1)(n 1) k, .
1
.
j
is the maximal
word in S having as a prex. Notice that our argument reveals that the .
j
s
depend on the functions in J, and not on . Thus, if we take = 0 we see that
.
1
.
j
= M
j
for 1 _ _ (k 1)(n 1) k.
So far we have restricted our attention to the case that is less than the length
of
1
(
n
(0))|
2
. Since
nC1
1 is produced by
1
, the above argument can be ex-
tended to arbitrary .
Theorem 2. Let n _ 1, k _ 2, and let J be the set of functions {
0
. . . . .
n
}
where
0
(.) = 2., and
i
(.) = 2
k1
(
i1
(.)) 1 for 1 _ i _ n.
Let S
F
be the set generated by J, and let S be the numeration system {s|
2
: s
S
F
}. Then S is based.
Proof. Let and n be adjacent words in S of the same length. Assume that
< n. Let be the maximal common prex of and n in S. Then = 0.
and n = 1, for some .. , {0. 1}
. Since S is treelike, every prex of a word

in S is also in S. Thus, 1 S and since S is self generating, there exists an
s S and an
i
J such that
i
(s)|
2
= 1. By denition we have that
1 =
i
(s)|
2
= 2
k1
i1
(s) 1|
2
.
Therefore,
2
k1
i1
(s)|
2
= 0.
and so
1
(
i1
(s))|
2
= 2
k
(
i1
(s)) 1|
2
= 2(2
k1
i1
(s)) 1|
2
= 01.
Therefore 01 S and is produced by
1
. Thus, by Lemma 5, it follows that
= 0M
jvjjpj1
. Since S is right extendable, it is clear that n = m
jwjjpj
.
Thus, by Theorem 1, it follows that S is based.
We conclude this section with a few examples. Note that in each case the
elements in the base sequence are obtained from the indices of the powers of 2 in
the sequence of ordered elements of S
F
.
Example 1. Let J = {2.. 8. 1}. Then
S = {0. 1. 2. 4. 8. 9. 10. 16. 17. 18. 32. 33. 34. 36. . . . }.
The numeration system S is dened as the set of base two representations of
the elements of S
F
. The elements of S correspond to the greedy expansions
of the natural numbers with respect to the base b
0
= 1, b
1
= 2, b
2
= 3, and
b
n
= b
n1
b
n3
for n _ 3.
Example 2. In general, if J = {2.. 2
k
. 1}, then S
F
is the set of nonnegative
integers whose base two expansions correspond to the greedy expansions of the
natural numbers with respect to the base sequence b
0
= 1, b
1
= 2. . . . . b
k1
=
k 1, and b
n
= b
n1
b
nk
for n _ k.
Example 3. Let J = {2.. 8. 1. 32. 5}. Then
S
F
= {0. 1. 2. 4. 5. 8. 9. 10. 16. 17. 18. 20. 32. 33. 34. 36. 37. 40. 41. . . . }.
The numeration system S consists of the greedy expansions of the natural num-
bers with respect to the base b
0
= 1, b
1
= 2, b
2
= 3, b
3
= 5, b
4
= 8, and
b
n
= b
n1
b
n3
b
n5
for n _ 5.
The generating functions used in Theorem 2 generate a numeration system that
is treelike. It is possible to construct examples of based self generating numeration
systems for which this is not the case. The following example is perhaps the
simplest such example.
Example 4. Let J = {2.. 4. 1. 4. 2}. Then
S
F
= {0. 1. 2. 4. 5. 6. 8. 9. 10. 12. 16. 17. 18. 20. 21. 22. 24. 25. 26. . . . }.
The numeration system S is dened as the set of base two representations of the
elements of S
F
. The rst 18 elements of S are shown in Figure 2, where we leave
out the root vertex at 0 for the sake of brevity.
1
10
100
1000
10000 10001
1001
10010
101
1010
10100 10101 10110
110
1100
11000 11001 11010
Figure 2. Tree structure for S generated by J = {2.. 4. 1. 4. 2}.
We have listed the elements of S in rows, where the elements of a given row
k are all the elements of S of length k. The tree connections between vertices are
shown. The elements in the last row that are not connected to anything are roots
of new trees in T(S). It is interesting to note that the rooted trees in Figure 2 are
all isomorphic. It is a routine exercise to prove a result similar to Lemma 5 for
this S, and therefore the proof that S is based is similar to the proof of Theorem 2.
We can also compute the base and see that the elements of S correspond to the
greedy expansions of the natural numbers with respect to the base b
0
= 1, b
1
= 2,
b
2
= 3, and b
n
= b
n1
2b
n2
b
n3
for n _ 3.
4 Linearly Recurrent Base Sequences
We now show that the base sequence in any based self generating numeration
system satises a linear recurrence. In [5] it was shown that the base sequence
in any based treelike numeration system satises a linear recurrence if and only if
the maximal sequence M dened in (2) is periodic. As example 4 shows, a self
generating numeration system need not be treelike. Shallit [15] also established
some rather general conditions which guarantee that the base sequence in a based
numeration system satises a linear recurrence. It can be shown that the self
generating numeration systems we are considering satisfy these conditions. In
this section we give a different proof which gives us a relatively simple means for
constructing the linear recurrence. Note also that in the following theorem, we are
no longer restricting our attention to the family J given by (4).
Theorem 3. Let J = {
0
.
1
. . . . .
n
} be a family of functions dened as follows.
Let
0
(.) = 2., and for 1 _ i _ m, let
i
(.) = 2
k
i
. c
i
, where k
i
_ 1 and
0 _ c
i
< 2
k
i
. Let S be the numeration system generated by J. If S is based,
then the base sequence {b
i
}
1
iD0
satises a linear recurrence.
Proof. Let {s
i
}
1
iD0
be the sequence of elements of S listed in order. Since S is
based, by Lemma 3 it follows that c
i
= 1 for some i with 1 _ i _ n. By
denition it follows that for n _ 0, b
n
is the index of 2
n
in {s
i
}. We need to show
that {b
i
} satises a linear recurrence.
Let {u
i
}
1
iD0
be the characteristic sequence of S
F
. That is, for i _ 0, u
i
= 1
if i S
F
and u
i
= 0 otherwise. It was shown in [8] that the characteristic
sequence is 2-automatic (for a denition of automatic sequences see [1]). By
Cobhams Theorem [6], (see also Theorem 6.3.2 of [1]), {u
i
} is the image under
a coding of the xed point of a morphism o of constant length 2 over a nite
alphabet {a
1
. . . . . a
k
}. Let o : {a
1
. a
2
. . . . . a
k
}
{a
1
. a
2
. . . . . a
k
}
be this mor-
phism, and suppose the xed point is generated by iteration on a
1
. Recall that
the incidence matrix of o is dened as the n by n matrix =
m
i;j
, where
m
ij
= [o(a
j
)[
a
i
, the number of occurrences of a
i
in o(a
j
) (see [1], Chapter 8).
Let be the incidence matrix of o and let
1(.) = .
k
c
1
.
k1
c
2
.
k2
c
k1
. c
k
be the characteristic polynomial of . Finally, let {r
n
} be the sequence dened by
the recurrence relation
r
n
= c
1
r
n1
c
2
r
n2
c
k
r
nk
. (6)
We show that {b
i
} satises this recurrence.
Let e
1
= 1 0 0|
T
. We rst show that for 1 _ i _ k the i
th
coordinate of
n
e
1
satises (6). By the CayleyHamilton Theorem 1() = 0, and so
k
c
1
k1
c
2
k2
c
k1
c
k
1 = 0.
Therefore 1() e
1
= 0. and so
k
e
1
c
1
k1
e
1
c
2
k2
e
1
c
k1
e
1
c
k
1e
1
= 0.
For n _ k, multiplying both sides of this last equation by
nk
gives
n
e
1
c
1
n1
e
1
c
2
n2
e
1
c
k1
nkC1
e
1
c
k
nk
e
1
= 0.
Now, by denition of , [o
n
(a
1
)[
a
i
, the number of occurrences of a
i
in o
n
(a
1
),
is the i
th
coordinate of
n
e
1
. Thus, for n > 0, [o
n
(a
1
)[
a
i
satises the recurrence
(6) with initial conditions [o(a
1
)[
a
i
. [o
2
(a
1
)[
a
i
. . . . . [o
k
(a
1
)[
a
i
.
Now, for n _ 0 we have that b
n
is the index of 2
n
in {s
i
}. Equivalently, since
{b
i
} is the base system for the abstract numeration system S, b
n
is the number of
ones in the sequence {u
n
} between u
0
and u
2
n
1
. On the other hand, since o has
constant length 2, it follows that
k
X
iD1
[o
n
(a
1
)[
a
i
= [o
n
(a
1
)[ = 2
n
.
Now let a
i
1
. a
i
2
. . . . . a
i
l
, where 1 _ l _ k, be the characters that map to 1 under
the coding that maps the xed point of o to {u
n
}. We therefore have that
b
n
= [o
n
(a
1
)[
a
i
1
[o
n
(a
1
)[
a
i
l
.
Since each of [o
n
(a
1
)[
a
i
j
satises (6), it follows that b
n
satises (6) as well.
5 Further Considerations
We close with some suggestions for further study. There are many examples of
sets of generating functions that do not satisfy the conditions of Theorem 2 that
seem to generate based numeration systems. For example, numerical evidence
seems to suggest that if J = {2.. 4. 1. 16. 3. 32. 11} then the result-
ing numeration system is based. In particular, the base two expansions of the
elements of S
F
seem to correspond to the set of greedy representations of N
with respect to the sequence b
0
= 1, b
1
= 2, b
2
= 4, b
3
= 7, b
4
= 13, and
b
n
= b
n1
b
n2
b
n4
b
n5
. On the other hand, not every self generating
numeration system is based. For example, it follows from Lemma 2 that the nu-
meration system generated by J = {2.. 8. 3} is not based. In light of these
examples we see that the characterization of all self generating based numeration
systems remains an open problem.
It is also natural to try to extend the results of this paper to numeration systems
with digit sets
k
= {0. 1. 2. . . . . k 1}. Some examples were considered in [8].
For example, if J = {3. 1. 3. 2. 9. 3. 9. 6} then
S
F
= {0. 1. 2. 3. 4. 5. 6. 7. 8. 10. 11. 12. 13. 14. 15. 16. 17. 19. 20. 21. . . . }.
This is the set of all nonnegative integers whose base 3 expansion does not contain
the block 00. If S = {s|
3
: s S
F
}, where s|
3
denotes the base 3 expansion
of s, then S is a numeration system with digit set
3
. It was noted in [8] that S
corresponds to the set of lazy representations of the natural numbers with respect
to the base sequence b
0
= 1, b
2
= 3, and b
n
= 2b
n1
2b
n2
for n _ 2.
Finally, the reader may have noticed that Theorem2 does not guarantee that the
numeration systems are greedy with respect to their bases. It is not hard to verify
this for specic examples. However, the question of which based numeration
systems are greedy and which are lazy also remains open.
Acknowledgments. We would like to thank the referee for a thorough reading
of the manuscript and for many helpful suggestions.
References
[1] J.-P. Allouche and J. Shallit, Automatic Sequences: Theory, Applications, General-
izations, Cambridge University Press, Cambridge, 2003.
[2] J.-P. Allouche, J. Shallit and G. Skordev, Self generating sets, integers with missing
blocks, and substitutions, Discrete Math. 292 (2005), no. 3, 115.
[3] A. Bertrand-Mathis, Comment crire les numbers entiers dans une base qui nest
pas entire, Acta Math. Acad. Sci. Hungar. 54 (1989), 237241.
[4] V. Bruyre and G. Hansel, Bertrand numeration systems and recognizability, Theo-
ret. Comput. Sci. 181 (1997), no. 1, 1743.
[5] P. J. Cameron and D. G. Fon-Der-Flaass, Fibonacci Notes, unpublished notes, 1996.
[6] A. Cobham, Uniform tag sequences, Math. Systems Theory 6 (1972), 164192.
[7] A. S. Fraenkel, Systems of numeration, Amer. Math. Monthly 92 (1985), 105114.
[8] D. Garth and A. Gouge, Afnely self generating sets and morphisms, J. Int. Seq. 10
(2007), Article 07.1.5.
[9] C. Kimberling, A self generating set and the golden mean, J. Int. Seq. 3 (2000)
Article 00.2.8.
[10] C. Kimberling, Afnely recursive sets and orderings of languages, Discrete Math.
274 (2004), 147159.
[11] C. Kimberling, Ordering of words and sets of numbers: the Fibonacci case, in:
Applications of Fibonacci Numbers, vol. 9, pp. 137144, Kluwer Acacamy Publ.,
Dordrecht, 2004
[12] P. B. A. Lecomte and M. Rigo, Numeration systems on a regular language, Theory
of Comput. Syst. 34 (2001), 2744.
[13] M. Lothaire, Algebraic Combinatorics on Words, Cambridge University Press, Cam-
bridge, 2002.
[14] M. Rigo, Numeration systems on a regular language: Arithmetic operations, rec-
ognizability, and formal power series, Theoret. Comput. Sci. 269 (2001), no. 12,
469498.
[15] J. Shallit, Numeration systems, linear recurrences, and regular sets, Inform. and
Comput. 113 (1994), 331347.
Author information
David Garth, Department of Mathematics and Computer Science,
Truman State University, Kirksville, MO 63501, USA.
E-mail: dgarth@truman.edu
Joseph Palmer, Department of Mathematics and Computer Science,
E-mail: jap875@truman.edu
Ha Ta, Department of Mathematics and Computer Science,
E-mail: tdh186@truman.edu
Small Sets Satisfying
the Central Sets Theorem
Neil Hindman
Abstract. The Central Sets Theorem is a powerful theorem, one of whose consequences
is that any central set in N contains solutions to any partition regular system of homoge-
neous linear equations. Since at least one set in any nite partition of N must be central,
any of the consequences of the Central Sets Theorem must be valid for any partition of N.
It is a result of Beiglbck, Bergelson, Downarowicz, and Fish that if is an idempotent
in (N. ) with the property that any member of has positive Banach density, then any
member of satises the conclusion of the Central Sets Theorem. Since all central sets
are members of such idempotents, the question naturally arises whether any set satisfy-
ing the conclusion of the Central Sets Theorem must have positive Banach density. We
answer this question here in the negative.
Keywords. Central, density.
AMS classication. 05D10.
1 Introduction
In [6] H. Furstenberg introduced the notion of central subsets of N in terms of
notions from topological dynamics. He showed that one cell of any nite partition
of a N must contain a central set and proved the original Central Sets Theorem.
(Given a set X, we denote by P
f
(X) the set of nite nonempty subsets of X.)
Theorem 1.1. Let C be a central subset of N. Let l N and for each i {1. 2.
. . . . l}, let
i
be a sequence in Z. Then there exist sequences (a
n
)
1
nD1
in N and
(H
n
)
1
nD1
in P
f
(N) such that
(1) for all n, max H
n
< min H
nC1
and
(2) for all J P
f
(N) and all i {1. 2. . . . . l},
n2F
_
a
n

t2H
n
i
(t )
_

C.
The author acknowledges support received from the National Science Foundation via Grant
DMS-0554803.
58 Neil Hindman
Proof. [6, Proposition 8.21].
Furstenberg used central sets to prove Rados Theorem [10] by showing that
any central subset of N contains solutions to all partition regular systems of ho-
mogeneous linear equations.
Based on an idea of V. Bergelson, central sets in N were characterized quite
simply [4] as members of minimal idempotents of (N. ), and this characteri-
zation extended naturally to dene central subsets of an arbitrary discrete semi-
group S.
What is currently the most general version of the Central Sets Theorem (for
commutative semigroups) is the following.
Theorem 1.2. Let (S. ) be a commutative semigroup and let T =
N
S, the set
of sequences in S. Let C be a central subset of S. There exist functions :
P
f
(T ) S and H : P
f
(T ) P
f
(N) such that
(1) if J. G P
f
(T ) and J _
6
G, then max H(J) < min H(G) and
(2) whenever m N, G
1
. G
2
. . . . . G
m
P
f
(T ), G
1
_
6
G
2
_
6
. . . _
6
G
m
, and for
each i {1. 2. . . . . m},
i
G
i
, one has
m
iD1
_
(G
i
)
t2H.G
i
/
i
(t )
_
C.
Proof. [5, Theorem 2.2].
To derive Theorem 1.1 from Theorem 1.2, note that one may assume that the
sequences
1
.
2
. . . . .
l
in the statement of Theorem 1.1 are distinct. Choose
additionally distinct sequences
k
for k > l and let for each n N, G
n
=
{
1
.
2
. . . . .
n
}. For n N, let a
n
= (G
n
) and let H
n
= H(G
n
).
For some of the motivating results that we will present, it is necessary to de-
scribe briey the algebraic structure of the Stone
Cech compactication. If the

reader is willing to accept that the question of whether every subset of N which
satises the conclusion of Theorem 1.2 must have positive Banach density is in-
teresting, she may proceed directly to Section 2 where that question is answered.
Given a discrete semigroup (S. ), the Stone
Cech compactication S of
S is the set of ultralters on S, the principal ultralters being identied with the
points of S. Given _ S, c = = { S : }. The family
{ : _ S} is a basis for the open sets (and a basis for the closed sets) of S.
The operation extends to S so that (S. ) is a right topological semigroup
(meaning that for each S the function j
p
: S S dened by j
p
(q) =
q is continuous) with S contained in its topological center (meaning that
for each . S the function z
x
: S S dened by z
x
(q) = . q is
continuous). Given . q S and _ S, one has that q if and only if
{. S : . q} , where . = {, S : . , }.
Small Sets Satisfying the Central Sets Theorem 59
As is true of any compact Hausdorff right topological semigroup, S has a
smallest two sided ideal 1(S) and there are idempotents in 1(S). Such idem-
potents are said to be minimal, and a subset C of S is central if and only if it is a
member of a minimal idempotent. The reader is referred to [8] for an elementary
introduction to the algebra of S.
The following notion was originally introduced by Polya in [9], but it is com-
monly referred to as Banach density.
Denition 1.3. Let _ N. Then
J
() = sup{ R : (Vk N)(Jn _ k)(Ja N)

([ {a 1. a 2. . . . . a n}[ _ n)}.
^
= { N : (V )(J
() > 0)}.
Since ^
is a two sided ideal of N, one has that 1(N) _ ^
, and in par-
ticular, if C is a central subset of N, then J
(C) > 0. The following result of

Beiglbck, Bergelson, Downarowicz, and Fish establishes that a weaker assump-
tion than central yields the conclusion of the original Central Sets Theorem.
Theorem 1.4. Let C _ N and assume that C is a member of an idempotent in
^
. Let l N and for each i {1. 2. . . . . l}, let

i
be a sequence in Z. Then
there exist sequences (a
n
)
1
nD1
in N and (H
n
)
1
nD1
in P
f
(N) such that
(1) for all n, max H
n
< min H
nC1
and
(2) for all J P
f
(N) and all i {1. 2. . . . . l},
n2F
_
a
n

t2H
n
i
(t )
_

C.
Proof. [2, Theorem 10].
In fact, the proof of [2, Theorem 10] is easily modied to show that any mem-
ber of an idempotent in ^
satises the conclusion of Theorem 1.2. It is a result

of C. Adams [1, Theorem 2.21] that there is a set C which is a member of an
idempotent in ^
but C misses the closure of the smallest ideal of N and in

particular, C is not central.
One is naturally led by the above results to ask whether any subset of N which
satises the conclusion of Theorem 1.2 must in fact have positive Banach density.
We show in Section 2 that this is not the case.
We close this introduction with an interesting contrast between members of
idempotents in ^
and central sets, that is members of idempotents in 1(N).

Those sets _ N such that 1(N) = 0 are exactly the piecewise syndetic
subsets of N by [8, Theorem 4.40] while a set _ N has ^
= 0 if and
60 Neil Hindman
only if J
() > 0 by [8, Theorem 3.11]. If is piecewise syndetic, then by [8,

Theorem 4.43] there is some . N such that . is central. On the other
hand, it is a result of Ernst Straus that there exist sets _ N with asymptotic
density arbitrarily close to 1 (and thus J
() arbitrarily close to 1) such that no

translate of is a member of any idempotent. (See [3, Theorem 2.20].)
2 A Small Subset of N Satisfying the Conclusion of the
Central Sets Theorem
We produce in this section a subset of N with zero Banach density which satises
the conclusion of Theorem 1.2 applied to the group (Z. ). The construction is
based on that of [7, Lemma 5.2]. For . N we denote by supp(.) the subset of
o = N L {0} such that . =
t2supp.x/
2
t
.
Theorem 2.1. Let T =
N
Z, the set of sequences in Z. There is a subset of
N such that J
() = 0 and there exist functions : P

f
(T ) N and H :
P
f
(T ) P
f
(N) such that
(1) if J. G P
f
(T ) and J _
6
G, then max H(J) < min H(G) and
(2) whenever m N, G
1
. G
2
. . . . . G
m
P
f
(T ), G
1
_
6
G
2
_
6
. . . _
6
G
m
, and for
each i {1. 2. . . . . m},
i
G
i
, one has
m
iD1
_
(G
i
)
t2H.G
i
/
i
(t )
_

.
Proof. For n N, let a
n
= min{t N : (
2
n
1
2
n
)
t
_
1
2
} and let s
n
=
n
iD1
a
i
.
(So s
1
= 1 and s
2
= 4.) Let b
0
= 0, let b
1
= 1, and for n N and t
{s
n
. s
n
1. s
n
2. . . . . s
nC1
1}, let b
tC1
= b
t
n 1. For k o, let
T
k
= {b
k
. b
k
1. b
k
2. . . . . b
kC1
1}. Let
= {. N : (Vk o)(T
k
\ supp(.) = 0)}
and let
0
= {. o : (Vk o)(T
k
\ supp(.) = 0)} (so
0
= L {0}).
We show rst that J
() = 0. Notice that for any . and m in N,

[ {.. . 1. . 2. . . . . . 2
m
1}[ _ [
0
{0. 1. 2. . . . . 2
m
1}[ .
Indeed, given any , {0. 1. 2. . . . . 2
m
1}\
0
, there is some k with b
kC1
_ m
such that T
k
_ supp(,) and there is a unique :(,) {.. .1. .2. . . . . .2
m
1} such that the rightmost m bits in the binary representation of :(,) are equal to
those of , and so T
k
_ supp
_
:(,)
_
. Further, if , = ,
0
, then :(,) = :(,
0
).
Let .. m N, let k = s
mC1
and let l _ 2
b
k
. We shall show that
[ {.. . 1. . 2. . . . . . l 1}[
l
<
_
1
2
_
m
.
Pick r N such that 2
r1
_ l < 2
r
. Then
[ {.. . 1. . . . . . l 1}[ _ [ {.. . 1. . . . . . 2
r
1}[
_ [
0
{0. 1. . . . . 2
r
1}[
so
[ {.. . 1. . 2. . . . . . l}[
l
_
[
0
{0. 1. 2. . . . . 2
r
1}[
2
r1
.
Now
[
0
{0. 1. 2. . . . . 2
r
1}[
=
2
rb
k
1
tD0
[
0
{t 2
b
k
. t 2
b
k
1. . . . . (t 1)2
b
k
1}[
_
2
rb
k
1
tD0
[
0
{0. 1. . . . . 2
b
k
1}[
= 2
rb
k
[
0
{0. 1. . . . . 2
b
k
1}[
so
[
0
{0. 1. 2. . . . . 2
r
1}[
2
r1
_
2
rb
k
[
0
{0. 1. . . . . 2
b
k
1}[
2
r1
=
[
0
{0. 1. . . . . 2
b
k
1}[
2
b
k
1
.
We have that [
0
{0. 1. . . . . 2
b
k
1}[ =
k1
tD0
(2
b
tC1
b
t
1) and 2
b
k
1
=
1
2
k1
tD0
2
b
tC1
b
t
so
[
0
{0. 1. . . . . 2
b
k
1}[
2
b
k
1
= 2
k1
tD0
_
2
b
tC1
b
t
1
2
b
tC1
b
t
_
= 2
2
1
1
2
1

m
nD1
s
nC1
1
tDs
n
_
2
b
tC1
b
t
1
2
b
tC1
b
t
_
=
m
nD1
_
2
nC1
1
2
nC1
_
a
nC1
_
_
1
2
_
m
.
Now we show that satises the conclusion of Theorem 1.2. First note that
if n. k N and and b
kC1
b
k
> n, then whenever :
1
. :
2
. . . . . :
n
N, there
must exist r T
k
such that for all t {1. 2. . . . . n}, T
k
\ supp(2
r
:
t
) = 0.
62 Neil Hindman
Indeed, if r T
k
, : N, and T
k
_ supp(2
r
:) then supp(:) T
k
= T
k
\ {r}.
Consequently
[{r T
k
: there is some i {1. 2. . . . . n} with T
k
_ supp(2
r
:
i
)}[ _ n.
Now we claim that
(+) for each n. m N and each J P
f
(T ), there exist J N and H P
f
(N)
such that min H > m and for all J, J
t2H
(t ) N2
n
.
To see this, let r = [J[ and pick k such that b
kC1
b
k
> r and b
k
> n. Pick
H P
f
(N) such that min H > m and for all J,
t2H
(t ) Z2
b
k
.
(Choose an innite subset C of N such that for all s. t C and all J,
(s) (t ) (mod 2
b
k
). Then pick H _ C such that min H > m and [H[ =
2
b
k
.) Pick c N2
b
k
such that for all J, c
t2H
(t ) > 0.
Let l = max
__
supp
_
c
t2H
(t )
_
: J
_
and pick such that
l < b
j
. Pick r
0
T
k
such that T
k
\ supp
_
2
r
0
c
t2H
(t )
_
= 0 for
each J. Inductively for i {1. 2. . . . . k}, pick r
i
T
kCi
such that
T
kCi
\ supp
_
2
r
i
i1
tD0
2
r
t
c
t2H
(t )
_
= 0 for each J. Let
J = c
jk
iD0
2
r
i
. Then (+) is established.
Now we dene (J) N and H(J) P
f
(N) for J P
f
(T ) inductively
on [J[. If J = { }, pick (J) N and H(J) P
f
(N) by (+) such that
(J)
t2H.F/
(t ) . Now let J P
f
(T ) with [J[ > 1 and assume that
we have dened (G) and H(G) for all G such that 0 = G _
6
J so that
(1) (G)
t2H.G/
(t ) for each G and
(2) if 1 _
6
G, then
(a) max H(1) < min H(G) and
(b) there exists k N such that for all 1 and all g G,
max supp
_
(1)
t2H.K/
(t )
_
< b
k
< min supp
_
(G)
t2H.G/
g(t )
_
.
Let m = max
_
{H(G) : 0 = G _
6
J} and pick k N such that for all G
P
f
(T ) with G _
6
J and all G, max supp
_
(G)
t2H.G/
(t )
_
< b
k
.
Pick by (+) some H(J) P
f
(N) and (J) N such that min H(J) > m and
for all J, (J)
t2H.F/
(t ) N2
b
k
C1
.
To verify that and H are as required for Theorem 1, let m N, let
G
1
. G
2
. . . . . G
m
P
f
(T ) .
and assume that G
1
_
6
G
2
_
6
. . . _
6
G
m
, and for each i {1. 2. . . . . m},
i
G
i
.
We claim that
m
iD1
_
(G
i
)
t2H.G
i
/

i
(t )
_
. Suppose instead one has
some k N such that T
k
_ supp
m
iD1
_
(G
i
)
t2H.G
i
/

i
(t )
_
. Then
there is some i such that T
k
_ supp
_
(G
i
)
t2H.G
i
/

i
(t )
_
, contradicting
hypothesis (1) of the construction.
References
[1] C. Adams, Large nite sums sets with closure missing the smallest ideal of N,
Topology Proceedings, to appear.
[2] M. Beiglbck, V. Bergelson, T. Downarowicz and A. Fish, Solvability of Rado sys-
tems in D-sets, Topology and its Applications, to appear.
[3] V. Bergelson, M. Beiglbck, N. Hindman and D. Strauss, Multiplicative structures
in additively large sets, J. Comb. Theory (Series A) 113 (2006), 12191242.
[4] V. Bergelson and N. Hindman, Nonmetrizable topological dynamics and Ramsey
Theory, Trans. Amer. Math. Soc. 320 (1990), 293320.
[5] D. De, N. Hindman and D. Strauss, A new and stronger Central Sets
Theorem, to appear in Fundamenta Mathematicae (currently available at
http://members.aol.com/nhindman/).
[6] H. Furstenberg, Recurrence in Ergodic Theory and Combinatorial Number Theory,
Princeton University Press, Princeton, 1981.
[7] N. Hindman, A. Maleki and D. Strauss, Central sets and their combinatorial charac-
terization, J. Comb. Theory (Series A) 74 (1996), 188208.
[8] N. Hindman and D. Strauss, Algebra in the Stone

Cech Compactication: Theory
and Applications, de Gruyter, Berlin, 1998.
[9] G. Polya, Untersuchungen ber Lcken und Singularitten von Potenzreihen, Math.
Zeit. 29 (1929), 549640.
[10] R. Rado, Studien zur Kombinatorik, Math. Zeit. 36 (1933), 242280.
Author information
Neil Hindman, Department of Mathematics, Howard University,
Washington, DC 20059, USA.
E-mail: nhindman@aol.com
Column-to-Row Operations on Partitions:
The Envelopes
Brian Hopkins
Abstract. Conjugation and the Bulgarian solitaire move are considered as extreme cases
of several column-to-row operations on integer partitions. Each operation generates a
state diagram on the partitions of n, which leads to the questions: How many Garden of
Eden states are there? How many cycle states? How many connected components? All of
these questions are answered for partitions of n when at least
n1
2
columns are switched
to rows.
Keywords. Partitions, Bulgarian solitaire.
AMS classication. 05A17, 37E15.
1 Introduction
Conjugation is the fundamental operation on integer partitions. Write a partition
z as (z
1
. . . . . z
`./
) where (z) denotes the partitions length, its number of parts.
The conjugate partition z
0
is dened as z
0
= (z
0
1
. . . . . z
0
s
) where z
0
i
is the number
of parts {z
i
} greater than or equal to i . This is more easily understood in terms
of the Ferrers diagram: the dots are reected along the diagonal, so that columns
and rows are swapped; see Figure 3.
We write 1(n) for the set of partitions of n. Since conjugation is an involution,
the state diagram of 1(n) determined by conjugation consists of singletons and
pairs, i.e., self-conjugate partitions and conjugate pairs. See Figure 1 for an ex-
ample, which also introduces the superscript notation for partitions, e.g., writing
21
3
for (2. 1. 1. 1).
Consider the effect of conjugation on 1(n) as a state diagram. Notice in Figure
1 that all seven partitions of 5 are in cycles and the diagram has four connected
components.
Bulgarian solitaire is an operation on partitions introduced by Brandt in 1982
[3]. We dene it as D
1
(z) = (z
0
1
. z
1
1. . . . . z
`./
1) where any zeros are
removed and the parts may not be in the standard non-increasing order. In terms
of the Ferrers diagram, the operation takes the rst (leftmost) column and makes
it a row; see Figure 3. Figure 2 shows the effect of D
1
on partitions of 5.
66 Brian Hopkins
1
5
T
c
5
41
21
3
T
c
311
c
32
T
c
221
Figure 1. Conjugation on 1(5); all 7 partitions are in cycles and there are 4
components.
1
5
c
5
E
21
3
c
41
E
32
E
311
T
221
Figure 2. Bulgarian solitaire on 1(5); there are 2 Garden of Eden partitions, 3
cycle partitions, and one component.
Like conjugation, the D
1
operation also produces a state diagram on 1(n).
Notice in Figure 2 that three partitions of 5 are in cycles and the diagram con-
sists of a single component. There are also two partitions that have no pre-image
under the operation (1
5
and 21
3
); these are called Garden of Eden partitions (sub-
sequently abbreviated GE-partitions).
In this article, we introduce a sequence of column-to-row operations; con-
jugation and the Bulgarian solitaire operation are the extreme cases. Bulgarian
solitaire has been the subject of several articles; HopkinsJones [4] includes a
fairly complete bibliography. Many of the questions concern state diagram con-
cepts: partitions in cycles, partitions with no preimages, and number of connected
components. In this article, we consider these same questions for all general-
ized column-to-row operations. We determine the number of GE-partitions, the
number of cycle partitions, and the number of connected components for approx-
imately half of all possible cases.
2 General Row-to-Column Operations
Conjugation can be thought of as moving all columns to rows; Bulgarian solitaire
moves one column to a row. We connect these ideas by introducing the sequence
of operations
D
k
(z) = (z
0
1
. . . . . z
0
k
. z
1
k. . . . . z
`./
k)
Column-to-Row Operations on Partitions: The Envelopes 67
where any nonpositive numbers are removed and the parts may not be in the stan-
dard non-increasing order. In terms of the Ferrers diagram, the operation takes the
rst k columns and makes them rows. Figure 3 shows a partition and its images
under various D
k
.
Figure 3. Ferrers diagrams for z = (4. 1), D
1
(z) = (3. 2), D
2
(z) = (2. 2. 1),
and D
3
(z) = D
4
(z) = z
0
= (2. 1. 1. 1), with shaded dots showing which rows
came from columns of z.
These operations all generate state diagrams on 1(n). Figures 4 and 5 show
1(5) under D
2
and D
3
, respectively.
1
5
c
5
E
c
311
21
3
41
c
E
32
T
c
221
Figure 4. The D
2
operation on 1(5); there are 2 GE-partitions, 3 total partitions
in cycles, and 2 components.
1
5
c
5
E
41
21
3
T
c
311
c
32
T
c
221
Figure 5. The D
3
operation on 1(5); there is 1 GE-partition, 5 cycle partitions,
and 3 components.
Notice that for 1(5), conjugation (Figure 1) is equivalent to the operation D
4
.
This is an example of a general fact.
Lemma 1. For a partition z with z
1
_ k or with z
1
= k 1 and z
2
_ k, the
operation D
k
is equivalent to conjugation. In particular, D
n1
is equivalent to
conjugation on 1(n).
68 Brian Hopkins
Proof. For z with k or fewer columns, the claim is evident. Assume z
1
= k 1
and z
2
_ k, i.e., that z has k 1 columns with (k 1)-st having height 1.
Moving k columns to rows leaves a single row of length 1, so that the effect of
D
k
is equivalent to moving all columns to rows. Every z 1(n) has z
1
_ n 1
except the single-part partition (n), which satises the other condition since the
second part of (n) is 0. Therefore, for all partitions of n, D
n1
(z) = z
0
.
3 Results on Partitions with Many Parts
This section consists of results about partitions of n with at least (n 1),2 or
n,2 parts. One result is well known and others are particular to the purposes
of this article. First, we introduce some notation. Capital letters signify sets,
corresponding lower-case letters the number of elements in the set, e.g., (5) = 7.
Recall the convention that (0) = 1. We will use part-wise addition on partitions,
e.g., (3. 1. 1) (2. 2) = (5. 3. 1).
Let 1(n. ) denote the set of partitions of n with exactly parts; from the
examples, we see (5. 3) = 2. Table 1 shows the (n. ) values for 1 _ n. _
12. Notice that, reading right to left, roughly half of each row are initial values of
(n), i.e., 1. 1. 2. 3. 5. 7. . . . . We call that portion of the triangle the envelope.
n\k 1 2 3 4 5 6 7 8 9 10 11 12
1 1
2 1 1
3 1 1 1
4 1 2 1 1
5 1 2 2 1 1
6 1 3 3 2 1 1
7 1 3 4 3 2 1 1
8 1 4 5 5 3 2 1 1
9 1 4 7 6 5 3 2 1 1
10 1 5 8 9 7 5 3 2 1 1
11 1 5 10 11 10 7 5 3 2 1 1
12 1 6 12 15 13 11 7 5 3 2 1 1
Table 1. (n. k), the number of partitions of n with k parts.
Lemma 2. Let a positive integer n be given. For each integer _
n
2
, (n. ) =
(n ).
Proof. We demonstrate a bijection between 1(n. ) and 1(n ). Any z
1(n. ) can be written as z = 1
j
j where j 1(n ); let z j. Any
j 1(n ) has at most parts, since the restriction on implies n _ ,
so that z = 1
j
j 1(n. ); let j z. Clearly these are inverse maps.
This result shows that, in some sense, the difculty of studying partitions lies
in the partitions with fewer than n,2 parts, which correspond to roughly the left-
hand half of each row in Table 1. There are direct formulas for (n. ) with
_ 5 (see, e.g., [2] and [6]), but they quickly become complicated. Notice that
the z j relation determined by z = 1
j
j is equivalent to removing the
rst column of the Ferrers diagram of z, i.e., the Bulgarian solitaire D
1
operation
without including the part z
0
1
.
Lemma 3. Let a positive integer n be given. For each integer _
n1
2
, the
following hold.
(a) All z 1(n. k) with k _ 2 have z
j
= 1.
(b) All j 1(n) with (j) _ j
0
1
have j
1
_ n (j) 1.
(c) All v 1(n) with v
1
= 1 have v
2
_ . That is, all v 1(n) with
v
1
_ 1 satisfy one of the conditions of Lemma 1.
Proof. (a) Assume that z 1(n. 2). Since z has 2 parts, z
j
= 0. Suppose
that z
j
_ 2. Then the sum of the parts of z would be at least 2 1 1 _ n1,
a contradiction. For k > 2, the sum of the parts has a higher lower bound, so
the result follows.
(b) The rst row and the rst column share a dot in the Ferrers diagram, so
their sum is at most n 1.
(c) If v
1
= v
2
= 1, then v
1
v
2
> n, contradicting j 1(n).
Note that (a) is sharp in the sense that, for example, every z 1(9. 6) has
z
4
= 1 but 1(9. 5) includes 32211 and 2
4
1 whose third parts are 2, not 1.
4 Garden of Eden Partitions
Let GE(n. k) denote the GE-partitions of n under the operation D
k
. From the
examples of the previous section, we know ge(5. 1) = ge(5. 2) = 2, ge(5. 3) = 1,
and ge(5. 4) = 0. Those data correspond to the n = 5 row of Table 2.
The diagonal of zeros corresponds to the fact that every partition has a pre-
image under conjugation. Notice that other diagonals seem to eventually stabilize
at some value; these limiting values comprise the envelope.
70 Brian Hopkins
1 2 3 4 5 6 7 8 9 10
2 0
3 1 0
4 1 1 0
5 2 2 1 0
6 3 3 2 1 0
7 5 5 4 2 1 0
8 7 8 6 4 2 1 0
9 10 12 10 7 4 2 1 0
10 14 18 15 11 7 4 2 1 0
11 20 25 23 17 12 7 4 2 1 0
12 27 35 33 26 18 12 7 4 2 1
13 37 48 47 38 28 19 12 7 4 2
14 49 66 65 55 41 29 19 12 7 4
15 66 88 89 77 60 43 30 19 12 7
16 86 118 120 107 85 63 44 30 19 12
17 113 155 161 145 119 90 65 45 30 19
18 147 203 213 196 163 127 93 66 45 30
19 190 263 280 260 222 175 132 95 67 45
20 243 340 364 344 297 240 183 135 96 67
21 311 435 471 449 394 323 252 188 137 97
Table 2. ge(n. k), the number of GE-partitions in 1(n) under D
k
.
What is the sequence of values 0. 1. 2. 4. 7. 12. 19. 30. 34. 67. 97. . . . in the en-
velope? One possibility is the partial sum of partition numbers (A000070 in [8]).
Let
s(n) =
n
X
iD0
(i )
and s(1) = 0. Before we can verify that this sequence describes the envelope,
we need to characterize GE-partitions.
Lemma 4. A partition z = (z
1
. . . . . z
`./
) 1(n) is in GE(n. k) precisely when
z
k
(z) _ 1 k.
Proof. In terms of the Ferrers diagram, z has a pre-image under D
k
for every
set of k rows each greater than or equal to (z) k, i.e., long enough to be
moved to become columns to the left side of the remaining dots. This fails when
z
k
< (z) k.
This generalizes the initial lemma and corollary of HopkinsJones [4] for Bul-
garian solitaire (D
1
).
We now show that, in the envelope, the GE-partitions are precisely the parti-
tions with many parts.
Theorem 1. Let a positive integer n be given. For each integer
n1
2
_ _ n 1,
we have ge(n. ) = s(n 2).
Proof. If = n 1, then the operation is equivalent to conjugation and there
are no GE-partitions, matching ge(n. n 1) = s(1) = 0. So assume that
n1
2
_ _ n 2. By the preceding lemma, GE(n. ) consists of all z 1(n)
with z
j
(z) _ 1 . This means that any GE-partition z must have (z) _
z
j
1, so z
j
= 0 and in fact (z) _ 2.
But by Lemma 3a, all j 1(n) with 2 or more parts have j
j
= 1, so
that j
j
(j) = 1 (j) _ 1 ( 2) = 1 . That is, GE(n. ) is exactly
1(n. 2) L L 1(n. n). By Lemma 2, (n. 2) = (n 2), . . . ,
(n. n) = (0). We conclude that ge(n. ) =
P
(i ) = s(n 2).
As with the (n. ) values of Table 1, one would like to have formulas for
the columns of Table 2. HopkinsSellers [5] provides two proofs of the following
result.
Theorem 2.
ge(n. 1) = (n 3) (n 9) (n 18)
=
X
j1
(1)
jC1
n
3
2
3
2
.
We make the following conjectures about the next few columns.
Conjectures
ge(n. 2) = (n 4) (n 5) (n 11) (n 12) (n 13)
(n 21) (n 22) (n 23) (n 24)
=
X
j1
j
X
kD0
(1)
jC1
n
3
2
3
2
k
.
ge(n. 3) = (n 5) (n 6) (n 7) (n 13) (n 14)
2(n 15) (n 16) (n 17) (n 24) .
72 Brian Hopkins
ge(n. 4) = (n 6) (n 7) (n 8) (n 9) (n 15)
(n 16) 2(n 17) 2(n 18) 2(n 19) (n 20)
(n 21) (n 27) (n 28) 2(n 29)
where complete expressions for ge(n. 3) and ge(n. 4) involve q-binomial coef-
cients. These are consistent with Theorem 1, since only the initial positive terms
arise in the envelope. These conjectures will be considered in future work with
Louis Kolitsch.
5 Cycle Partitions
Let CP(n. k) denote the partitions of n in cycles under the operation D
k
. Fromthe
examples of the previous section, we know cp(5. 1) = cp(5. 2) = 3, cp(5. 3) = 5,
and cp(5. 4) = 7. Those data correspond to the n = 5 row on the left-hand side
of Table 3.
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
2 2 0
3 1 3 2 0
4 3 3 5 2 2 0
5 3 3 5 7 4 4 2 0
6 1 5 7 9 11 10 6 4 2 0
7 4 6 7 11 13 15 11 9 8 4 2 0
8 6 8 10 14 18 20 22 16 14 12 8 4 2 0
9 4 6 11 16 22 26 28 30 26 24 19 14 8 4 2 0
10 1 5 15 20 28 34 38 40 41 37 27 22 14 8 4 2
11 5 5 15 23 32 42 48 52 51 51 41 33 24 14 8 4
12 10 10 20 28 41 53 63 69 67 67 57 49 36 24 14 8
13 10 16 22 32 46 63 77 87 91 85 79 69 55 38 24 14
14 5 23 29 37 56 77 97 111 130 112 106 98 79 58 38 24
15 1 28 35 42 63 91 116 138 175 148 141 134 113 85 60 38
16 6 33 41 49 75 108 143 171 225 198 190 182 156 123 88 60
17 15 35 45 57 83 124 168 207 282 262 252 240 214 173 129 90
18 20 42 48 68 98 145 202 253 365 343 337 317 287 240 183 132
19 15 39 45 79 107 166 233 301 475 451 445 411 383 324 257 189
20 6 41 43 93 126 190 275 360 621 586 584 534 501 437 352 267
21 1 46 42 108 142 215 314 423 791 746 750 684 650 577 478 369
Table 3. On the left, cp(n. k), the number of cycle partitions. On the right,
(n) cp(n. k), the number of partitions not in cycles.
Under conjugation, we know every partition is in a cycle, either self-conjugate
or half of a conjugate pair. By Lemma 1, then, cp(n. n 1) = (n). The right-
hand side of Table 3 shows (n) cp(n. k). The envelope of this triangle of
differences appears to be 2s(n) for s(n) dened in Section 4.
n1
2
_ _ n 1,
cp(n. ) = (n) 2s(n 2).
Proof. Given z GE(n. ), we claim that its iterated images under D
j
have the
form
z z
0
j j
0
where indicates that the operation D
j
coincides with conjugation and we allow
the possibility j = j
0
.
First, we know from the proof of Theorem 1 that z GE(n. ) has at least
2 parts, and then z
1
_ by Lemma 3b. Therefore, by Lemma 1, D
j
(z) = z
0
.
Let D
j
(z
0
) = j. Since z is a Garden of Eden partition, j = z and applying D
j
to z
0
is not equivalent to conjugation. By the denition of D
j
,
j = D
j
(D
j
(z)) = D
j
(z
0
) = (z
1
. . . . . z
j
. z
0
1
)
with z
0
2
and subsequent terms removed since z
0
2
_ by Lemma 3c. Therefore
j
0
1
= 1 and, by Lemma 3b, j
1
_ 1. By Lemmas 3c and 1 we conclude
that D
j
(j) = j
0
. Likewise, D
j
(j
0
) = j. Since conjugation is an involution,
the sets {z. z
0
} and {j. j
0
} are disjoint, completing the claim.
We can now complete the proof of the theorem. If some partition z 1(n)
is not in a cycle, it is either a GE-partition or between a GE-partition and a cycle
partition. By the claim above, we know that a GE-partition maps to its conjugate,
which maps to a cycle partition. From Theorem 1, we know that there are s(n
2) GE-partitions, which are not in cycles. Their conjugates are the other
s(n 2) partitions not in cycles.
It is important to realize that the structural results of the proof do not imply
that every component of 1(n) under D
j
in the envelope contains at most four
partitions. It is true that z GE(n. ) has iterates z z
0
jfor j CP(n. ),
but multiple GE-partitions can lead to the same j, e.g.,
D
3
(D
3
((21
5
)) = D
3
(61) = 3211. D
3
(D
3
((31
4
)) = D
3
(511) = 3211.
Also, the structural results of the proof do not hold in general outside the enve-
lope. For instance, D
2
(D
2
(51
3
)) = 3221 which does not have k 1 = 3 parts.
Also, Figure 1 shows that D
1
(D
1
(1
5
)) = 41 CP(5. 1), so two steps from a
74 Brian Hopkins
GE-partition is not always a cycle partition. Lengths from GE-partitions to cycle
partitions for D
1
are among the data tabulated in [4].
A formula for the column cp(n. 1) is proven in [3].
Theorem 4. Write n =
mC1
2
a where 0 _ a _ m 1. Then cp(n. 1) =
m
a
.
Formulas for other columns would seem to require generalizing the charac-
terization of cycle partition for D
1
found in [3]. The proof in the next section
describes cycle partitions in the envelope, but does not apply for smaller .
6 Connected Components
Let cc(n. k) denote the number of connected components in the state diagram of
1(n) under the operation D
k
. From the examples of the previous section, we
know cc(5. 1) = 1, cc(5. 2) = 2, cc(5. 3) = 3, and cc(5. 4) = 4. Those data
correspond to the n = 5 row on the left-hand side of Table 4.
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
2 1 0
3 1 2 1 0
4 1 2 3 2 1 0
5 1 2 3 4 3 2 1 0
6 1 2 4 5 6 5 4 2 1 0
7 1 2 4 6 7 8 7 6 4 2 1 0
8 2 3 5 8 10 11 12 10 9 7 4 2 1 0
9 1 2 5 9 12 14 15 16 15 14 11 7 4 2 1 0
10 1 2 6 10 15 18 20 21 21 20 16 12 7 4 2 1
11 1 2 6 11 17 22 25 27 28 27 23 18 12 7 4 2
12 2 3 7 13 21 28 33 36 38 37 33 27 19 12 7 4
13 2 4 7 14 23 33 40 45 50 48 45 38 29 19 12 7
14 1 5 9 15 27 39 50 57 68 64 60 54 42 30 19 12
15 1 6 11 17 30 46 60 71 89 84 79 73 60 44 30 19
16 1 7 12 18 34 54 73 88 117 111 106 100 84 64 45 30
17 3 8 13 20 37 61 85 106 148 143 138 131 114 90 66 45
18 4 9 14 23 41 69 101 128 191 186 181 172 154 126 94 67
19 3 9 13 25 44 78 116 152 245 239 235 223 204 170 132 96
20 1 8 12 29 49 87 135 181 316 309 305 288 268 230 182 136
21 1 7 12 33 53 97 153 212 399 393 388 367 347 303 247 188
Table 4. On the left, cc(n. k), the number of connected components. On the right,
cc(n) cc(n. k).
The numbers of self-conjugate partitions and conjugate pairs were studied by
Osima [7]. It follows that the total number of components of 1(n) under conju-
gation (equivalently D
n1
) is given by
cc(n) = (n) (n 2) (n 8) (n 18) (n 32)
=
X
k0
(1)
k
(n 2k
2
)
with initial terms 1. 1. 2. 3. 4. 6. 8. 12. 16. 22. 29 . . . (A046682 in [8]). The right-
hand side of Table 4 shows cc(n) cc(n. k). The envelope of this triangle of
differences appears once again to be s(n) dened in Section 4.
n1
2
_ _ n 1,
we have cc(n. ) = cc(n) s(n 2).
Proof. Recall from the proof of Theorem 3 that a partition of 1(n) not in a cycle
under D
j
is either a GE-partition or the conjugate of a GE-partition. For each
z GE(n. ), the conjugate pair {z. z
0
} counted in cc(n) is part of another com-
ponent. The discussion of j = D
j
(D
j
(z)) in the proof of Theorem 3 established
that the conjugate pair {j. j
0
} is still a 2-cycle under D
j
or the self-conjugate j
is still self-conjugate under D
j
.
We show that no partitions in CP(n. ) are part of larger cycles. Recall from
the proof of Theorem1 that GE(n. ) = 1(n. 2)L L1(n. n). It follows that
{conjugates of GE-partitions} is the set of z 1(n) with z
1
_ 2. Therefore
CP(n. ) is the remainder of 1(n), namely, the 1(n) with
1
_ 1 and
0
1
_ 1, i.e., the partitions of n that t inside a ( 1) ( 1) square.
By Lemmas 3c and 1, the operation D
j
is equivalent to conjugation for CP(n. ),
which means that it consists of conjugate pairs and self-conjugate partitions.
Of the singletons and pairs counted by cc(n), exactly ge(n. ) pairs are no
longer components in 1(n) under D
j
. By Theorem 1, we conclude cc(n. ) =
cc(n) s(n 2).
Viewing 1(n) dynamically under D
n1
, then D
n2
, . . . , some conjugate pairs
are opened into z z
0
fragments that attach to preserved conjugate pairs
or self-conjugate partitions. To the left of the envelope, larger cycles develop and
there are longer paths from GE-partitions to cycle partitions, as in Figure 1.
A formula for the column cc(n. 1) is proven in [3].
Theorem 6. Write n =
mC1
2
a where 0 _ a _ m 1. Then
cc(n. 1) =
1
m
X
dj.m;a/
(J)
m,J
a,J
!
76 Brian Hopkins
where the summation is over all divisors of the greatest common divisor of m and
a, and is the Euler phi function.
It would be very interesting to determine formulas for other columns, as they
would transition between the number-theoretic formula for cc(n. 1) and the for-
mulas involving (n) for cc(n. ) in the envelope.
Acknowledgments. This article was developed from a talk given at the 2007
Integers conference at the University of West Georgia. Thanks to Bruce Landman
for coordinating the conference, where I enjoyed fruitful discussions on this mate-
rial with James Sellers and Louis Kolitsch. Thanks also to the anonymous referee
for close reading and helpful suggestions. Antonio Pane (SPC 08) prepared some
of the gures. Helpful data were derived using Mathematica.
References
[1] G. Andrews, The Theory of Partitions, Cambridge University Press, 1984.
[2] G. Andrews and K. Eriksson, Integer Partitions, Cambridge University Press, 2004.
[3] J. Brandt, Cycles of partitions, Proc. Amer. Math. Soc. 85 (1982), 483486.
[4] B. Hopkins and M. A. Jones, Shift-induced dynamical systems on partitions and com-
positions, Electron. J. Combin. 13 (2006), R80.
[5] B. Hopkins and J. A. Sellers, Exact enumeration of Garden of Eden partitions, Inte-
gers 7 (2007), no. 2, A19.
[6] A. Munagi, Computation of q-partial fractions, Integers 7 (2007), no. 2, A25.
[7] M. Osima, On the irreducible representations of the symmetric group, Canad. J.
Math. 4 (1952), 381384.
[8] N. Sloane, The On-line Encyclopedia of Integer Sequences, published electronically
at www.research.att.com/~njas/sequences/.
Author information
Brian Hopkins, Department of Mathematics, Saint Peters College,
Jersey City, NJ 07306, USA.
E-mail: bhopkins@spc.edu
On the Euler Product of
Some Zeta Functions
Xian-Jin Li
Abstract. It is well known that the Euler product for the Riemann zeta function (s) is
still valid for m(s) = 1 and s = 1. In this paper, we extend this result to zeta functions of
number elds. In particular, the Dedekind zeta function
k
(s) for any algebraic number
eld k and the Hecke zeta function (s. ) for the rational number eld are shown to have
the Euler product on the line m(s) = 1 except at s = 1. A functional equation is obtained
for the nite Euler product, which is the product of rst nite number of factors in the
Euler product of (s. ).
Keywords. Dirichlet series, Euler product, zeta-fuction.
AMS classication. 11M41, 30B40.
1 Introduction
The gamma function I(s,2) can be regarded as a factor for the innity place of
the rational number eld in the Euler product of the Riemann zeta function (s) as
shown in [11], and is required for the functional equation of (s). A spectral the-
ory of the gamma function was discovered by Sonine [10]. Remarkable examples
of functions were presented for which the analogue of the Riemann hypothesis
is true (cf [6]). Sonines theory was further developed by de Branges [1] whose
theory can be regarded as a generalization of the part of Fourier analysis involving
Fourier transforms and the Plancherel formula.
Fourier analysis over number elds is a fundamental tool for studying zeta-
and 1-functions. It started in Tates thesis [11], which had a deep inuence on
number theory as a piece of clandestine literature. Therefore, it is natural for us
to generalize de Branges theory to number elds. In trying to do so, we need
the Euler product for the Dedekind zeta function
k
(s) and for the Hecke zeta
function (s. ) to be still valid for m(s) = 1 and s = 1.
The way, in which the convergence of the Euler product of zeta functions on
the line ms = 1 is needed in my current work in progress, goes briey as the
Research supported by National Security Agency H98230-06-1-0061.
78 Xian-Jin Li
following: For the Hecke zeta function we expect a certain relation between (1
i:. ) and (1 i:. ) to be true for all complex :. This relation is similar to
the statement of Theorem 1.3 with S being the set of all places of the rational
number eld. Since the Euler product for (1 i t. ) is valid for all nonzero
real t as shown in Theorem 1.2, we can prove that the expected relation between
(1 i t. ) and (1 i t. ) holds for all nonzero real t . Because one side of the
expected relation is analytic in the upper half-plane and is continuous in the closed
upper half-plane, and because the other side of the expected relation is analytic in
the lower half-plane and is continuous in the closed lower half-plane, by analytic
continuation we see that the expected relation holds for all complex :. The results
obtained may be important for studying the analyticity of Artin 1-functions in the
whole complex plane.
For the Fourier transformation, we have a similar construction. But, the cor-
responding relation turns out to be between (1,2 i:. ) and (1,2 i:. ).
In other words, the line where both sides of the relation are extended across by
analytic continuation is the line 1,2 i t , t R instead of the line 1 i t , t R.
Since the Hecke zeta function (s. ) does not have Euler product on the line
ms = 1,2, we dont have a similar result. The result obtained is only for nite
Euler products. It will be described after the statement of Theorem 1.2.
Let k be any algebraic number eld. The Dedekind zeta function
k
(s) of k is
dened by
k
(s) =
p
1
1 Np
x
for o > 1 with s = o i t , where the product is over all prime ideals p of k.
For the convergence of the Euler product of the Dedekind zeta function
k
(s)
on the line ms = 1, we prove the following theorem in Section 2.
Theorem 1.1. We have the Euler product
k
(s) =
p
1
1 Np
x
for o = 1 and s = 1.
A Hecke character of the rational number eld Q is a character of the mul-
tiplicative group generated by primes of Q not in some set 1 and has value 0 on
nite primes in 1, where 1 is a nite set of primes including the innite prime of
Qand is called the exceptional set of .
The Hecke zeta-function (s. ) is dened by
(s. ) =
]621
1
1 ()
x
for o > 1, s = o i t .
On the Euler Product of Some Zeta Functions 79
For the convergence of the Euler product of the Hecke zeta function (s. ) on
the line ms = 1, we obtain the following theorem which is proved in Section 3.
Theorem 1.2. Let be a Hecke character of Q. Then
(s. ) =
]621
1
1 ()
x
for o = 1 and s = 1, where the product is over all rational primes , 1.
For every place , we denote by Q
, O
, and 1
the completion of Q at
, the maximal compact subring of Q
, and the unique maximal ideal of O
,
respectively. We denote by [ [
the valuation of Q normalized so that [ [
is
the ordinary absolute value if is real, and [
= 1, if O
,1
contains
elements where 1
.
The idele group J of Q is the restricted direct product of the multiplicative
groups Q
relative to subgroups O
of units in Q
. Let J
1
be the set of ideles
= (
) such that
= 1.
Let S be a nite set of places of Q containing the innite place. We dene
A
S
=
2S
Q
. Let
[
: . e
2ti2
v
(x)
be the character on the additive group Q
given in Section 2.2 of [11]. It is trivial

on O
, and is nontrivial on 1
1
for nite places . We have order([
) = 0 and
d
1
= O
for all nite places of Q. Let

[() =
2S
[
)
for A
S
. Then [ is a character on A
S
.
For any place of Q, we select a xed Haar measure J
on the additive
group Q
as follows: J
:= the ordinary Lebesgue measure on the real line if

is real, and J
:= that measure for which O
gets measure 1 if is nite.

Then J =
2S
J
is the unique Haar measure on A

S
such that the inversion
formula
() =
_
A
S
()[() J
holds if is continuous and

1
1
(A
S
), where
() =
_
A
S
()[() J.
for 1
1
(A
S
); see Section 3.3 in [11].
80 Xian-Jin Li
The Schwartz space S(R) is the space of all smooth functions , all of whose
derivatives are of rapid decay; that is
d
k
d.
k
(.) = O((1 [.[)
1
)
for all integers k _ 0 and N > 0. Let S(A
S
) be the SchwartzBruhat space on
A
S
, whose functions are nite linear combinations of functions of the form
() =
2S
)
where
(1)
is in the Schwartz space S(R) if is the innite place of Q, and

(2)
belongs to S(Q
), the space of locally constant and compactly supported

functions on Q
, if is nite.
Let
c() =
all places of Q
c
) =
_
_
]21
c
)[
[
it
v
_
_
(
1
()) (1.1)
be a character of J
1
given in Section 4.5 of [11] with t
1
= 0, where is a Hecke
character of Qwith exceptional set 1 and
1
() =
]621
ord
v
v
.
In particular, c
is unramied for every , 1, and c() = 1 for all Q
. We
denote 1
0
= 1 {o}.
The ramication degree of c
is denoted by e
for 1
0
. We always assume
that e
_ 1. Otherwise, if e
= 0 we do not include this prime in the set 1. By

Corollary 2.4.1 in [11] and the remark after Theorem 4 in [9],
c
(
e
v
)
e
v
{2
_
jxjD1
[
(
e
v
.) c
(.)J. =
i0
v
(1.2)
for some real number 0
when 1
0
.
S is always chosen so that it contains 1. We dene a nite Euler product
S
by
S
(s. ) =
2S,1
1
1 (
)
x
.
A functional equation is obtained in the following for the nite Euler product
S
(s. ).
Theorem 1.3. Assume that c, and 1 are given as in (1.1) with c
1
(1) = 1.
Let be a function in S(A
S
) such that () and

() vanish for [[ _ for
a positive number and such that (
0
) = () for all A
S
, where
0
is
obtained from by replacing
1
by
1
. Then
_

]21
0
i0
v
Ce
v
xe
v
{2
_
s
2
I(
s
2
)
S
(s. )
_
A
S
() c()[[
x
J
=
1s
2
I(
1 s
2
)
S
(1 s. )
_
A
S
()c()[[
x1
J (1.3)
for all complex s, where the left side for ms < 1,2 and the right side for ms > 1,2
are dened by analytic continuation.
2 Proofs of Theorem 1.1
Let k be an algebraic number eld with r
1
real places and r
2
imaginary places.
Put G
1
(s) =
x{2
I(s,2) and G
2
(s) = (2)
1x
I(s). Let
k
(s) = s(s 1)[d[
s
2
G
1
(s)
i
1
G
2
(s)
i
2
k
(s). (2.1)
where d is the discriminant of k. By Theorem 3 in Chapter VII of Weil [14],
k
(s)
is an entire function.
Dene c
to be one when is a real place of k and to be two when is an

imaginary place of k. Let . =
be the variable in the half space R

i
1
Ci
2
C
.
Denote by [.[ the product
.
e
v
taken over all innite places of k. Let N =

r
1
2r
2
. The Hecke theta function
k
(.) is dened by
k
(.) =
b
exp
_
[d[
1
N
(Nb)
2
N
_
(2.2)
where the sum on b runs over all nonzero integral ideals of k and the sum on
is over all innite places of k. Put J. =
J.
. It follows from Theorem 3 in

82 Xian-Jin Li
Chapter XIII of Lang [5] that
k
(s) = 2
i
1
(2)
i
2
h1,e s(s 1)
_
jxj>1
k
(.)([.[
s
2
[.[
1s
2
)
J.
.
(2.3)
for all complex s, where h, 1 and e are respectively the number of ideal classes
of k, the regulator of k and the number of roots of unity in k.
Lemma 2.1 (Satz 184 in Landau [4]).
k
(s) = 0 on the line o = 1.
Lemma 2.2 (Satz 186 in Landau [4]). There are positive constants c
0
. t
0
depend-
ing on k with t
0
> 1 such that
0
k
(s)
k
(s)
= O(ln
3
[t [)
for o _ 1
1
c
0
ln jtj
and [t [ > t
0
.
Lemma 2.3 (Lemma 3.12 in Titchmarsh [13]). Let
(s) =
1
nD1
a
n
n
x
for o > 1, where a
n
= O([(n)) with [ being non-decreasing. Assume that
1
nD1
[a
n
[
n
c
= O
_
1
(o 1)
_
as o 1. If c > 0 and o c > 1, . is not an integer, and N is the integer
nearest to ., then
n~x
a
n
n
x
=
1
2i
_
cCiT
ciT
(s n)
.
u
n
Jn O
_
.
c
T(o c 1)
_
O
_
[(2.).
1c
ln .
T
_
O
_
[(N).
1c
T [. N[
_
for T ; 0.
Lemma 2.4. Let
0
k
k
(s) =
1
nD1
n
n
x
for o > 1. Then
n
= O
_
(ln n)
1Ce
_
as n o, where c is a small positive number.
Proof. For o > 1,
0
k
k
(s) =
p
1
nD1
ln Np
(Np)
nx
where the sum on p is over all prime ideals of k. For each prime ideal p of k there
is exactly one rational prime which is divisible by p. Thus Np =
(
for some
positive integer less than or equal to the degree of k over Q; see Theorem 108
in [3].
Let (Np)
n
=
. We can write
p
1
nD1
ln Np
(Np)
nx
=
]
1
D1
_

p,(1p)
m
D]
ln Np
_
x
.
where the sum on is over all rational primes. We have
p,(1p)
m
D]
ln Np _ o
1
(v) ln .
where
o
1
(v) =
dj
J _ J(v)v v
1Ce
with J(v) being the number of positive divisors of v. Thus
p,(1p)
m
D]
ln Np v
1Ce
ln .
If we denote
= n, then
n
=
p,(1p)
m
D]
ln Np.
It follows that
n
(ln n)
1Ce
as n o.
This completes the proof of the lemma.
Lemma 2.5. Let
n
be given as in Lemma 2.4. Then
1
nD1
[
n
[
n
c
= O
_
1
o 1
_
as o 1.
84 Xian-Jin Li
Proof. We write
1
nD1
[
n
[
n
c
=
p
1
nD1
ln Np
(Np)
nc
=
0
k
k
(o)
for o > 1. By (2.1) and (2.2)
0
k
k
(o) =
1
o
1
o 1
ln [d[
2
(r
2
r
1
2
) ln
r
1
2
I
0
I
(
o
2
) r
2
I
0
I
(o)
0
k
k
(o).
Since
k
(1) = c
k
and since poles of I(s) are at s = 1. 2. . . . , multiplying
both sides of the above identity by o 1 and letting o 1we nd that
lim
c!1C
(o 1)
0
k
k
(o) = 1.
It follows that
1
nD1
[
n
[
n
c
= O
_
1
o 1
_
as o 1.
Lemma 2.6. Let
n
be given as Lemma 2.4, and let s = 1 i t for any xed
nonzero real number t . Then the partial sums
n~x
n
n
x
are bounded as . o.
Proof. By Lemma 2.3, Lemma 2.4, and Lemma 2.5
n~x
n
n
x
=
1
2i
_
cCiT
ciT
0
k
k
(s n)
.
u
n
Jn O
_
.
c
Tc
_
O
_
(ln .)
2Ce
T
_
(2.4)
for c > 0 and T ; 0.
Let
=
1
c
0
ln([t [ T )
.
By Lemma 2.2, Lemma 2.1, and (2.3), we can choose T to be sufciently large
so that
k
(s n) has no zeros for m(n) _ and [`(s n)[ _ [t [ T . By
Cauchys residue theorem,
1
2i
_
cCiT
ciT
0
k
k
(s n)
.
u
n
Jn
=
0
k
k
(s)
.
1x
1 s
1
2i
_
_
cCiT
CiT
_
CiT
iT
_
iT
ciT
_
0
k
k
(s n)
.
u
n
Jn.
(2.5)
By Lemma 2.2 and the choice of ,
1
2i
_
cCiT
CiT
0
k
k
(s n)
.
u
n
Jn
ln
3
T
T
.
c
. (2.6)
Similarly, we have
1
2i
_
iT
ciT
0
k
k
(s n)
.
u
n
Jn
ln
3
T
T
.
c
.
In this paragraph we assume that [t `(n)[ _ t
0
, m(n) = , and s = 1i t .
Then

1
s n
1
s n 1
ln [d[
2
(r
2
r
1
2
) ln
.
By using the identity
I
0
(:)
I(:)
=
1
:
:
1
nD1
1
n(n :)
;
where ; is Eulers constant, we get that
r
1
2
I
0
I
(
s n
2
) r
2
I
0
I
(s n)
1.
By (2.3),
0
k
k
(s n)
1.
Thus it follows from the identity
0
k
k
(s n) =
1
s n
1
s n 1
ln [d[
2
(r
2
r
1
2
) ln
r
1
2
I
0
I
(
s n
2
) r
2
I
0
I
(s n)
0
k
k
(s n)
86 Xian-Jin Li
that
0
k
k
(s n)
(2.7)
when [t `(n)[ _ t
0
.
When [t `(n)[ > t
0
, as m(s n) = 1 by Lemma 2.2 we nd that
0
k
k
(s n)
ln
3
T.
Note that t is a xed real number. Thus by (2.7) we obtain that
1
2i
_
CiT
iT
0
k
k
(s n)
.
u
n
Jn .
max
_
1
. ln
3
T
__
T
0
1
_
2
u
2
Ju.
Since
_
T
0
1
_
2
u
2
Ju = ln
_
T,
_
(T,)
2
1
_
and
=
1
c
0
ln([t [ T )
.
we have
1
2i
_
CiT
iT
0
k
k
(s n)
.
u
n
Jn .
ln
4
T. (2.8)
We can take c = 1,ln . and T = exp
__
ln .
_
with . being sufciently large.
Then .
c
= e and
.
ln
4
T =
ln
2
.
e
ln x{c
0
ln(jtjCT)

ln
2
.
e
ln x{2c
0
ln T
=
ln
2
.
e
p
ln x{2c
0
= o(1)
as . o. It follows from (2.5)(2.8) that
1
2i
_
cCiT
ciT
0
k
k
(s n)
.
u
n
Jn =
0
k
k
(s)
.
1x
1 s
o(1) (2.9)
as . o.
Since
.
c
Tc
=
e ln .
e
p
ln x
= o(1)
and
(ln .)
2Ce
T
=
(ln .)
2Ce
e
p
ln x
= o(1).
by (2.4) and (2.9)
n~x
n
n
x
=
0
k
k
(s)
.
1x
1 s
o(1)
as . o. By Lemma 2.1, the stated result then follows.
Proof of Theorem 1.1. For o > 1,
0
k
k
(s) =
p
1
nD1
ln Np
(Np)
nx
and
ln
k
(s) =
p
1
nD1
1
m(Np)
nx
where the sum on p is over all prime ideals of k. We write
ln
k
(s) =
p
1
nD1
ln Np
(Np)
nx

1
ln {(Np)
n
}
.
Let
n
s be given as in Lemma 2.4. Then
ln
k
(s) =
1
nD2
n
n
x

1
ln n
(2.10)
for o > 1, because the series is absolutely convergent.
For any xed non-zero real number t , by Lemma 2.6 the partial sums
1
nD2
n
n
1Cit
are bounded for all sufciently large integers N. Since 1,ln n tends steadily to 0
as n o, by Dirichlets test the series
1
nD2
n
n
x

1
ln n
converges for o = 1 and s = 1.
Since ln
k
(s) is continuous for o _ 1 and s = 1 by Lemma 2.1, and since the
right side of (2.10) represents an analytic function of s in the half-plane o > 1
88 Xian-Jin Li
and is convergent for o = 1 and s = 1, by the continuity theorem for Dirichlet
series (see Section 9.12 of Titchmarsh [12])
ln
k
(s) =
1
nD2
n
n
x

1
ln n
By the proof of Lemma 2.4, the series
p
1
nD2
1
m(Np)
nx
is absolutely convergent for o = 1. This implies that
ln
k
(s) =
p
1
nD1
1
m(Np)
nx
=
p
ln
1
1 Np
x
(2.11)
for o = 1 and s = 1. By taking exponentials of both sides of (2.11), we nd that
k
(s) =
p
(1 Np
x
)
1
This completes the proof of Theorem 1.1.
3 Proof of Theorem 1.2
The proof of Theorem 1.2 is a minor modication of that for Theorem 1.1. We
present it for the convenience of readers.
Lemma 3.1 (Theorem 7.15 in [7]). (s. ) = 0 on the line o = 1.
Lemma 3.2 (Theorem 7.20 and its proof in [7]). There are positive constants t
0
>
e and c
0
, depending on , such that
0
(s. )
(s. )
ln
[t [
for o _ 1
1
c
0
ln jtj
and [t [ > t
0
, where M is some constant greater than 1.
Lemma 3.3. Let
(s. ) =
1
nD1
T
n
n
x
for o > 1. Then [T
n
[ _ ln n for all n.
Proof. For o > 1,
(s. ) =
]621
1
nD1
(
n
) ln
nx
.
The stated assertion then follows.
Lemma 3.4. Let T
n
be given as in Lemma 3.3. Then
1
nD1
[T
n
[
n
c
= O
_
1
o 1
_
as o 1.
Proof. We have
1
nD1
[T
n
[
n
c
_
]
1
nD1
ln
nc
=
(o)
for o > 1, where is the Riemann zeta-function.
By (2.12.7) in [13],
0
(o)
(o)
=
1
o 1
1
;
2
ln 2
1
2
I
0
I
(1
o
2
)
p
o
j(j o)
where the sum is over all complex zeros j of (s). This implies that
0
(o)
(o)
1
o 1
as o 1. It follows that
1
nD1
[T
n
[
n
c
= O
_
1
o 1
_
as o 1.
90 Xian-Jin Li
In the next two paragraphs, we review some analytic properties for (s. ); see
(3.3) and (3.5). They will be needed for the proof of Lemma 3.5. Let
=
all places of Q
where
) =
_
_
1
O
v
(
) if , 1
[
)1
1
e
v
v
(
) if 1
0
e
t
2
v
if = o. c
(1) = 1
e
t
2
v
if = o. c
(1) = 1.
(3.1)
We have
) =
_
_
1
O
v
(
) if , 1
e
v
1
1C1
e
v
v
(
) if 1
0
e
t
2
v
if = o. c
(1) = 1
i
e
t
2
v
if = o. c
(1) = 1.
(3.2)
We dene t
1
= 0 if 1 contains at least one nite prime of k and t
1
= 1 if 1
contains only the innite place of k. By the proof of Theorem 4.4.1 in [11], Tates
zeta-function (. c[[
x
) equals
(. c[[
x
) =
_
1
1
t
(. c[[
x
)
Jt
t
_
1
1
t
(
. c[[
1x
)
Jt
t
t
1
_
(0)
s 1
(0)
s
_
.
(3.3)
where
t
(. c[[
x
) =
_
J
1
(t b)c(b)t
x
Jb
with Jb being given as in [11].
The two integrals on the right side of (3.3) are entire functions of s. By the
argument in Section 4.5 and Section 2.5 of [11], we can write
(. c[[
x
) =
_

]21
]
(
. c
[[
x
)
_
(s. ). (3.4)
where
]
(
. c
[[
x
) =
e
v
x
_
jj
v
D]
e
v
[
()c
()J
with J
being the standard Haar measure on Q
given as in Section 2.3 of

Tate [11] for 1
0
, and where
]
(
. c
[[
x
) =
_
x{2
I(s,2) for = o. c
1
(1) = 1
sC1
2
I(
xC1
2
) for = o. c
1
(1) = 1.
Let (s. ) = s(s 1)|
r
P
(. c[[
x
). By (3.3), (s. ) is an entire function of s.
We have
(s. ) =
t
1
s
t
1
s 1
1
2
I
0
I
_
2s 1 c
1
(1)
4
_
ln
2
]21
0
e
ln
(s. ).
(3.5)
Note that (3.5) also follows from Theorem 8.5 in Chapter VII of Neukirch [8],
which is proved using the classical language. The author wishes to thank the
referee for pointing out this to him.
Lemma 3.5. Let T
n
be given as Lemma 3.3, and let s = 1 i t for any xed
nonzero real number t . Then the partial sums
n~x
T
n
n
x
are bounded as . o.
Proof. By Lemma 2.3, Lemma 3.3, and Lemma 3.4
n~x
T
n
n
x
=
1
2i
_
cCiT
ciT
(s n. )
.
u
n
Jn O
_
.
c
Tc
_
O
_
ln
2
.
T
_
(3.6)
for c > 0 and T ; 0.
Let
=
1
c
0
ln([t [ T )
.
By Lemma 3.2 and Lemma 3.1, we can choose T to be sufciently large so that
(s n. ) has no zeros for m(n) _ and [`(s n)[ _ [t [ T . By Cauchys
residue theorem and (3.5), we nd that
1
2i
_
cCiT
ciT
(s n. )
.
u
n
Jn =
(s. ) t
1
.
1x
1 s
1
2i
_
_
cCiT
CiT
_
CiT
iT
_
iT
ciT
_
(s n. )
.
u
n
Jn. (3.7)
By Lemma 3.2 and the choice of ,
1
2i
_
cCiT
CiT
(s n. )
.
u
n
Jn
ln
T
T
.
c
. (3.8)
92 Xian-Jin Li
Similarly, we also nd that
1
2i
_
iT
ciT
(s n. )
.
u
n
Jn
ln
T
T
.
c
. (3.9)
By Lemma 3.1, (s. ) = for ms = 1. Since (s. ) is an entire function, we
can choose T large enough so that [(o i t. )[ _ c
0
for a positive number c
0
depending only on and t
0
when [t [ _ t
0
and 1 _ o _ 1. Thus, if we use
(3.5) when [t `(n)[ _ t
0
and use Lemma 3.2 when [t `(n)[ > t
0
, we nd
that
1
2i
_
CiT
iT
(s n. )
.
u
n
Jn .
max{
1
. ln
T }
_
T
0
1
_
2
u
2
Ju.
Since
_
T
0
1
_
2
u
2
Ju = ln
_
T,
_
(T,)
2
1
_
and
=
1
c
0
ln([t [ T )
.
we have
1
2i
_
CiT
iT
(s n. )
.
u
n
Jn .
ln
C1
T. (3.10)
We can take c = 1,ln . and T = exp
__
ln .
_
with . being sufciently large.
Then .
c
= e and
.
ln
C1
T =
ln
MC1
2
.
e
ln x{c
0
ln(jtjCT)

(
_
ln .)
C1
e
ln x{2c
0
ln T
=
(
_
ln .)
C1
e
p
ln x{2c
0
= o(1)
as . o. It follows from (3.7)(3.10) that
1
2i
_
cCiT
ciT
(s n. )
.
u
n
Jn =
(s. ) t
1
.
1x
1 s
o(1) (3.11)
as . o.
Since
.
c
Tc
=
e ln .
e
p
ln x
= o(1)
and
ln
2
.
T
=
ln
2
.
e
p
ln x
= o(1).
by (3.6) and (3.11)
n~x
T
n
n
x
=
(s. ) t
1
.
1x
1 s
o(1)
as . o. By Lemma 3.1, the stated result then follows.
Proof of Theorem 1.2. For o > 1,
(s. ) =
]621
1
nD1
(
n
) ln
nx
and
ln (s. ) =
]621
1
nD1
(
n
)
m
nx
.
We write
ln (s. ) =
]621
1
nD1
(
n
) ln
nx

1
ln(
n
)
.
Let T
n
s be given as in Lemma 3.3. Then
ln (s. ) =
1
nD2
T
n
n
x

1
ln n
(3.12)
for o > 1, because the series is absolutely convergent.
For any xed non-zero real number t , by Lemma 3.5 the partial sums
1
nD2
T
n
n
1Cit
are bounded for all sufciently large integers N. Since 1,ln n tends steadily to 0
as n o, by Dirichlets test the series
1
nD2
T
n
n
x

1
ln n
converges for o = 1 and s = 1.
Since ln (s. ) is continuous for o _ 1 and s = 1 by Lemma 3.1, and since
the right side of (3.12) represents an analytic function of s in the half-plane o > 1
94 Xian-Jin Li
and is convergent for o = 1 and s = 1, by the continuity theorem for Dirichlet
series (see Section 9.12 of Titchmarsh [12])
ln (s. ) =
1
nD2
T
n
n
x

1
ln n
The series
]621
1
nD2
(
n
)
m
nx
is absolutely convergent for o = 1. It follows that
ln (s. ) =
]621
1
nD1
(
n
)
m
nx
=
]621
ln
1
1 ()
x
(3.13)
for o = 1 and s = 1. By taking exponentials of both sides of (3.13), we nd that
(s. ) =
p
(1 ()
x
)
1
4 Proofs of Theorem 1.3
Let be an even function in S(R) such that and

vanish outside a nite
interval (a. a). For existence of such functions , see Sonine [10]. We dene
J(:) =
1=2iz
2
I
_
1,2 i:
2
_
_
1
0
(t )t
1{2Ciz
Jt
and
G(:) =
1=2iz
2
I
_
1,2 i:
2
_
_
1
0
(t )t
1{2Ciz
Jt
for `: _ 0.
Lemma 4.1. J and G have analytic extensions to the lower half-plane and satisfy
the identity
G(:) = J(:) (4.1)
for all complex :. In particular,
1=2ix
2
I
_
1,2 i.
2
_
_
1
0
(t )t
1{2Cix
Jt
=
1=2Cix
2
I
_
1,2 i.
2
_
_
1
0
(t )t
1{2ix
Jt
(4.2)
for all real ..
Proof. Let , be any xed positive number. If (t ) = e
t,
2
t
2
, then
(t ) = ,
1
e
t,
2
t
2
.
By Plancherels formula
_
1
0
(t )(t ) Jt =
_
1
0
(t )
(t ) Jt :
that is,
_
1
0
(t )e
t,
2
t
2
Jt =
_
1
0
(t ),
1
e
t,
2
t
2
Jt
for all , > 0. It follows that
_
1
0
,
1{2Ciz
J,
_
1
0
(t )e
t(,t)
2
Jt
=
_
1
0
,
1{2Ciz
J,
_
1
0
(t ),
1
e
t(,
1
t)
2
Jt
(4.3)
for real :. The left side of (4.3) is equal to
_
1
0
(t )t
1{2iz
Jt
_
1
0
,
1{2Ciz
e
t,
2
J,
=
1
2
1=2Ciz
2
I
_
1,2 i:
2
_
_
1
0
(t )t
1{2iz
Jt =
1
2
J(:)
for real :, and the right side of (4.3) equals
_
1
0
(t )t
1{2Ciz
Jt
_
1
0
,
1{2iz
e
t,
2
J,
=
1
2
1=2iz
2
I
_
1,2 i:
2
_
_
1
0
(t )t
1{2Ciz
Jt =
1
2
G(:)
for real :. Therefore
G(:) = J(:)
96 Xian-Jin Li
for all real :. Since J and G are analytic and continuous in the closed upper half-
plane, J and G have analytic extensions to the lower half-plane. The extended J
and G satisfy the identity
G(:) = J(:)
for all complex :.
Lemma 4.2. Assume that c
is an unramied character of Q
. Let be a function
in S(Q
). Then
1
1 c
)
1{2Ciz
_
Q
v
()[[
1{2Ciz
c
()J
=
1
1 c
)
1{2iz
_
Q
v
()[[
1{2iz
c
()J
for all complex :.
Proof. We have
_
Q
v
()[[
1{2Ciz
c
()J
=
_
Q
v
()J
_
Q
v
[
()[[
1{2Ciz
c
()J
=
_
Q
v
()[[
1{2iz
c
()J
_
Q
v
[
(.) c
(.)[.[
1{2Ciz
J.
for all complex :.
By Theorem 1 in [9],
_
Q
v
[
(.) c
(.)[.[
1{2Ciz
J. =
1 c
)
1{2Ciz
1 c
)
1{2iz
(4.4)
for all complex :. It follows that
_
Q
v
()[[
1{2Ciz
c
()J
=
1 c
)
1{2Ciz
1 c
)
1{2iz
_
Q
v
()[[
1{2iz
c
()J
for all complex :. The stated identity then follows.
Lemma 4.3. Assume that c
is a unitary character on Q
and has ramication

degree e
> 0. Let
be a function in S(Q
). Then
i0
v
ie
v
z
_
Q
v
()[[
1{2Ciz
c
()J =
_
Q
v
()[[
1{2iz
c
()J
for all complex :, where 0
is given in (1.2).
Proof. As in the proof of Lemma 4.2 we have
_
Q
v
()[[
1{2Ciz
c
()J
=
_
Q
v
()[[
1{2iz
c
()J
_
Q
v
[
(.) c
(.)[.[
1{2Ciz
J.
for all complex :.
By Lemma 1 in Sally and Taibleson [9],
_
Q
v
[
(.) c
(.)[.[
1{2Ciz
J.
=
e
v
(1{2Ciz)
c
(
e
v
)
_
jxjD1
[
(
e
v
.) c
(.)J..
(4.5)
By (1.2), we can write
_
Q
v
[
(.) c
(.)[.[
1{2Ciz
J. =
ie
v
z
i0
v
.
It follows from (4.5) that
i0
v
ie
v
z
_
Q
v
()[[
1{2Ciz
c
()J =
_
Q
v
()[[
1{2iz
c
()J
for all complex :.
Let c, and 1 be given as in (1.1). Then c
) = () for all , 1 and

() = 0 for 1
0
.
Proof of Theorem 1.3. Since elements in S(A
S
) are nite linear combinations of
functions of the form
2S
).
without loss of generality we can assume that is of this form.
98 Xian-Jin Li
Because () belongs to S(A
S
) and vanishes for [[ _ , the function
J(:) =
1=2iz
2
I
_
1,2 i:
2
_
S
(1,2 i:. )
_
A
S
()c()[[
1{2Ciz
J
(4.6)
is analytic in the upper half-plane `(:) _ 0 and is continuous in the closed upper
half-plane.
Similarly, we nd that the expression
G(:) =
1=2iz
2
I
_
1,2 i:
2
_
_

]21
0
i0
v
e
v
iz
_

S
(1,2 i:. )
_
A
S
() c()[[
1{2Ciz
J
represents an analytic function of : in the upper half-plane. It is continuous in the
closed upper half-plane.
By Lemma 4.1, Lemma 4.2 and Lemma 4.3,
1=2it
2
I
_
1,2 i t
2
_
_

]21
0
i0
v
e
v
it
_

S
(1,2 i t. )
_
A
S
() c()[[
1{2Cit
J
=
1=2Cit
2
I
_
1,2 i t
2
_
S
(1,2 i t. )
_
A
S
()c()[[
1{2it
J
for all real t ; that is, G(t ) = J(t ) for all real t . By analytic continuation, we see
that J and G can be extended to become analytic functions in the whole complex
plane and satisfy the identity
G(:) = J(:) (4.7)
for all complex :. If we let s = 1,2 i:, then the stated identity follows from
(4.7).
Remark. By Lemma 1 in Sally and Taibleson [9] and Sonine [10], there exist
functions in S(A
S
) such that () and

() vanish for [[ _ for a positive
number and such that (
0
) = () for all A
S
, where
0
is obtained from
by replacing
1
by
1
.
Acknowledgments. The author wishes to thank the referee for his helpful sug-
gestions of improving the presentation of the original manuscript.
References
[1] L. de Branges, Hilbert Spaces of Entire Functions, Prentice-Hall, N.J., 1968.
[2] H. Davenport, Multiplicative Number Theory, Third Edition, Revised by Hugh
L. Montgomery, Springer-Verlag, New York, 2000.
[3] E. Hecke, Lectures on the Theory of Algebraic Numbers, Springer-Verlag, New
York, 1981.
[4] E. Landau, Einfhrung in die elementare und analytische Theorie der algebraischen
Zahlen und der Ideale, 2. Auage, Chelsea Publishing Company, New York, 1949.
[5] S. Lang, Algebraic Number Theory, Second Edition, Springer-Verlag, New York,
1994.
[6] Xian-Jin Li, On zeros of dening functions for some Hilbert spaces of polynomi-
als, in: Operator Theory and Interpolation, edited by C. Foias and H. Bercovici,
pp. 235243, Birkhuser-Verlag, Basel, 2000.
[7] W. Narkiewicz, Elementary and Analytic Theory of Algebraic Numbers, Third Edi-
tion, Springer-Verlag, Berlin, 2004.
[8] J. Neukirch, Algebraic Number Theory, Springer-Verlag, Heidelberg, 1999.
[9] P. J. Sally, Jr. and M. H. Taibleson, Special functions on locally compact elds, Acta
Math. 116 (1966), 279309.
[10] N. Sonine Recherches sur les fonctions cylindriques et le dveloppement des fonc-
tions continues en sries, Math. Ann. 16 (1880), 180.
[11] J. T. Tate, Fourier analysis in number elds and Heckes zeta-functions, in: Alge-
braic Number Theory, edited by J. W. S. Cassels and A. Frhlich, pp. 305347,
Academic Press, New York, 1967.
[12] E. C. Titchmarsh, The Theory of Functions, Second Edition, Oxford University
Press, 1958.
[13] E. C. Titchmarsh, The Theory of the Riemann Zeta-Function, Second Edition, edited
by D. R. Heath-Brown, Oxford University Press, New York, 1986.
[14] A. Weil, Basic Number Theory, Springer-Verlag, Heidelberg, 1967.
Author information
Xian-Jin Li, Department of Mathematics, Brigham Young University,
Provo, Utah 84602, USA.
E-mail: xianjin@math.byu.edu
On the Range of the
Iterated Euler Function
Florian Luca and Carl Pomerance
Abstract. For a positive integer k let
k
be the k-fold composition of the Euler function
. In this paper, we study the size of the set {
k
(n) _ .} as . tends to innity.
Keywords. Iterations of Eulers function, applications of sieve methods.
AMS classication. 11N36, 11N56.
1 Introduction
Let be Eulers function. For a positive integer k, let
k
be the k-fold composi-
tion of . In this paper, we study the range V
k
of
k
. For a positive real number
. we put
V
k
(.) = {
k
(n) _ .}.
In 1935, Erd os [7] showed that #V
1
(.) = .,(log .)
1o(1)
. (Stronger estimates
are known for #V
1
(.), see [10], [17].) In 1977, Erd os and Hall [8] considered the
more general problem of estimating #V
k
(.), suggesting that it is .,(log .)
ko(1)
for each xed integer k _ 1. They were able to prove that
#V
2
(.) _
.
(log .)
2o(1)
.
and in fact, they were able to establish a somewhat more explicit form for this
inequality. Our rst result is the following general upper bound on #V
k
(.) which
is uniform in k.
Theorem 1. The estimate
#V
k
(.) _
.
(log .)
k
exp
_
13k
3{2
(log log . log log log .)
1{2
_
(1)
holds uniformly in k _ 1 once . is sufciently large.
Work by the rst author was done in Spring of 2006 while he visited Williams College. The
second author was supported in part by NSF grants DMS-0401422 and DMS-0703850.
102 Florian Luca and Carl Pomerance
As a corollary we have, when . o,
#V
k
(.) _
.
(log .)
ko(1)
(2)
when k = o((log log ., log log log .)
1{3
), and
#V
k
(.) _
.
(log .)
(1o(1))k
when k = o(log log ., log log log .). Note that (1) is somewhat stronger than the
explicit upper bound in [8] for the case k = 2.
Let k _ 1 be xed. Let m > 2 be such that m. 2m1. . . . . 2
k-1
m2
k-1
1
are all prime numbers. Then
k
(2
k-1
m2
k-1
1) = m 1. The quantitative
version of the Prime k-tuples Conjecture of Bateman and Horn [2] implies that the
number of such values m _ . should be _ c
k
.,(log .)
k
for . sufciently large,
where c
k
> 0 is a constant depending on k. Thus, we see that up to the factor of
size (log .)
o(1)
appearing on the right hand side of estimate (2), it is likely that
#V
k
(.) = .,(log .)
ko(1)
holds when k is xed as . o, thus verifying the
surmise of Erd os and Hall.
Next, we prove a lower bound on #V
2
(.) comparable to the one predicted by
the above heuristic construction.
Theorem 2. There exists an absolute constant c
2
> 0 such that the inequality
#V
2
(.) _ c
2
.
(log .)
2
holds for all . _ 2.
In [8], Erd os and Hall assert that they were able to prove such a lower bound with
the exponent 2 replaced by any larger real number.
In the last section we study the integers that are in every V
k
and we also discuss
analogous problems for Carmichaels universal exponent function z(n).
In what follows, we use the Vinogradov symbols ; and and the Landau
symbols O and o with their usual meaning. The constants and convergence im-
plied by them might depend on some other parameters such as k. 1. c, etc. We
use and q with or without subscripts for prime numbers. We use o(n) for the
number of distinct prime factors of n, C(n) for the number of prime power divi-
sors (> 1) of n, (n) and 1(n) for the smallest and largest prime divisors of n,
respectively, and
2
(n) for the exponent of 2 in the factorization of n. We write
log
1
. = max{1. log .}, and for k _ 2 we put log
k
. for the k-fold iterate of the
function log
1
evaluated at .. For a subset A of positive integers and a positive
real number . we write A(.) for the set A 1. .|.
On the Range of the Iterated Euler Function 103
2 The Proof of Theorem 1
Let . be large. By a result of Pillai [18], we may assume that k _ log ., log 2,
since otherwise V
k
(.) = {1}. Furthermore, we may in fact assume that k _
10
-2
log
2
., log
3
., since otherwise the upper bound on #V
k
(.) appearing in es-
timate (1) exceeds .. We may also assume that n _ .,(log .)
k
, since other-
wise there are at most .,(log .)
k
possibilities for n, and, in particular, at most
.,(log .)
k
possibilities for
k
(n) also.
By the minimal order of the Euler function, there exists a constant c
0
> 0 such
that the inequality (m),m _ c
0
m, log log m holds for all m _ 3. From this it is
easy to prove by induction on k that if . is sufciently large and
k
(n) _ ., then
n _ .(2c
0
log
2
.)
k
for all k in our stated range. Let X := .(log
2
.)
2k
, so that
for large ., we may assume that n _ X.
Let , = .
1{(log
2
x)
2
and write n = m, where = 1(n). By familiar esti-
mates (see, for example, [3]), the number of n _ X such that _ , is at most,
for large .,
X
(log .)
log
2
x
=
.(log
2
.)
2k
(log .)
log
2
x
_
.
(log .)
k
.
so we need only deal with the case > ,. Assume that C(
k
(n)) _ 2.9k log
2
..
Lemma 13 in [15] shows that the number of such possibilities for
k
(n) _ . is
k. log . log
2
.
2
2.9k log
2
x
_
.(log
2
.)
2
(log .)
2.9k log 2-1

.
(log .)
k
for all k in our range. It follows that we may assume that
C(
k
(n)) _ 2.9k log
2
..
It is easy to see that C((a)) _ C(a) 1 for every natural number a. Thus, since
k
(m) [
k
(n), we have
C((m)) _ 2.9k log
2
. k 1 _ 3k log
2
. (3)
for all . sufciently large.
Since also
k
() [
k
(n), we may assume that
C(
k
()) _ 2.9k log
2
..
Since > ,, we have log
2
> log
2
. 2 log
3
., so that C(
k
()) _ 3k log
2
for . large. Since _ X,m, we thus have, in the notation of Lemma 4 below,
that A
k,3k
(X,m), and that result shows that the number of such possibilities
is at most
#A
k,3k
(X,m) _
X
m(log(X,m))
k
exp
_
3k(6k log
2
X log
3
X)
1{2
3k
2
log
3
X
_
.
Observe further that with our bound on k,
3k(6k log
2
X log
3
X)
1{2
3k
2
log
3
X
= k
3{2
(log
3
X)
_
3(6 log
2
X, log
3
X)
1{2
3k
1{2
_
_ k
3{2
(log
2
X log
3
X)
1{2
(3
_
6 3,10).
Since 3
_
6 3,10 < 7.7, it thus follows that if we put
U(.) = exp(7.7k
3{2
(log
2
. log
3
.)
1{2
).
then for large .,
#A
k,3k
(X,m) _
.U(.)(log
2
.)
2k
m(log ,)
k
_
.U(.)(log
2
.)
4k
m(log .)
k
uniformly in m and k. Thus, the number of such possibilities for n _ X is
_
.U(.)(log
2
.)
4k
(log .)
k
nM
1
m
.
where M is the set of all possible values of m. Such m satisfy, in particular, the
inequality (3). Lemma 3 below shows that if . is sufciently large then
nM
1
m
_ exp
_
2.9(3k log
2
X log
3
X)
1{2
_
.
which together with the fact that 2.9
_
3 < 5.1 and the previous estimate shows
that the count on the set of our n _ X is
_
.
(log .)
k
exp
_
13k
3{2
(log
2
. log
3
.)
1{2
_
for large values of .. We thus nish the proof of Theorem 1 and it remains to
prove Lemmas 3 and 4.
Lemma 3. Let . be large, 1 be any positive integer and let N(1. .) denote the
set of natural numbers n _ . with C((n)) _ 1 log
2
.. Then
nN(1,x)
1
n
_ exp(2.9(1 log
2
. log
3
.)
1{2
)
holds for large values of . uniformly in 1.
Proof. We assume that 1 _ log
2
., log
3
. since otherwise the right hand side
above exceeds (log .)
2.9
, while the left hand side is at most log . O(1), so the
desired inequality holds anyway.
Let : be a parameter that we will choose shortly. For each integer n _ . write
n = n
0
n
1
, where each prime q [ n
0
has C(q 1) < log : and each prime q [ n
1
has C(q 1) _ log :. For n N(1. .) we have that C(n
1
) _ 1 log
2
., log :.
Let N
0
(.) denote the set of numbers n
0
_ . divisible only by primes q with
C(q 1) < log : and let N
1
(.) denote the set of numbers n
1
_ . with C(n
1
) _
1 log
2
., log :. We thus have
nN(1,x)
1
n
_
_
n
0
N
0
(x)
1
n
0
__
n
1
N
1
(x)
1
n
1
_
. (4)
Note that
n
0
N
0
(x)
1
n
0
_
o
}=0
1

_
q_x
D(q-1)~log z
1
q
1
q
2

_
}
= exp
_
q_x
D(q-1)~log z
1
q 1
_
.
It follows from Erd os [7] that there is some c > 0 such that the number of primes
q _ t with o(q1) _
1
2
log
2
q is O(t ,(log t )
1c
). Since o(q1) _ C(q1), the
same O-estimate holds for the distribution of primes q with C(q 1) _
1
2
log
2
q.
In particular the sum of their reciprocals is convergent, so that
e
z
2
~q_x
D(q-1)~log z
1
q 1
_
e
z
2
~q
D(q-1)~
1
2
log
2
q
1
q 1
1.
Thus,
q_x
D(q-1)~log z
1
q 1
_
q_e
z
2
1
q 1
e
z
2
~q_x
D(q-1)~log z
1
q 1
_ 2 log : O(1).
and so
n
0
N
0
(x)
1
n
0
:
2
. (5)
For the sum over N
1
(.), we have
n
1
N
1
(x)
1
n
1
_
}_1log
2
x{ log z
1

_
q_x
1
q 1
_
}
_
}_1log
2
x{ log z
1

(log
2
. O(1))
}
.
We choose : = exp((
1
2
1 log
2
. log
3
.)
1{2
). Observe that the inequalities
1 log
2
., log : = (21 log
2
., log
3
.)
1{2
< 2
1{2
log
2
., log
3
. < log
2
.
hold for large values of .. Thus,
n
1
N
1
(x)
1
n
1
(2 log
2
.)
1log
2
x{ log z
. (6)
Putting (5) and (6) into (4) and using the fact that 2
_
2 < 2.9, we have
nN(1,x)
1
n
_ exp(2.9(1 log
2
. log
3
.)
1{2
))
for all sufciently large .. This proves the lemma.
Remark 1. The above proof uses ideas from Erd os [7] and is also similar to
Lemma 4 in Luca [14].
Lemma 4. Let k, 1 be positive integers not exceeding
1
2
log
2
.. Put
A
k,1
= { : C(
k
()) _ 1 log
2
}.
We have
#A
k,1
(.) _
.
(log .)
k
exp
_
3k(21 log
2
. log
3
.)
1{2
3k
2
log
3
.
_
for all sufciently large values of ., independent of the choices of k. 1.
Proof. When k = 1, this trivially follows from the Prime Number Theorem. We
assume that k > 1. We let A
k,1
(.) and assume that _ .,(log .)
k
because
there are only (.,(log .)
k
) _ .,(log .)
k
primes failing this condition. Let
0
= and write
0
1 =
1
m
1
.
1
1 =
2
m
2
.
.
.
.
k-2
1 =
k-1
m
k-1
.
where
i
= 1(
i-1
1) for all i = 1. . . . . k 1. Since C((n)) _ C(n) 1,
we have that
C(
i-1
1) _ C(
i
()) _ C(
k
()) k _ 21 log
2
.
for all i = 1. 2. . . . . k 1 if . is sufciently large. In particular
i
_
1{(21log
2
x)
i-1
_
1{(log
2
x)
2
i-1
.
so that for . sufciently large we have
i
_
1{(log
2
x)
2i
0
_ ,
i
:=
1
2
.
1{(log
2
x)
2i
for i = 1. 2. . . . . k 1.
Consider the k linear functions 1
}
(.) =
}
. T
}
for = k. k 1. . . . . 1
given by 1
k
(.) = . and
1
k-1
(.) = m
k-1
. 1.
1
k-2
(.) = m
k-2
m
k-1
. m
k-2
1.
.
.
.
1
1
(.) = m
1
m
k-1
. (m
1
m
k-2
m
1
m
k-3
m
1
1).
Note that
k-1
_ .,(m
1
m
k-1
) is such that 1
}
(
k-1
) is a prime for all =
1. . . . . k. If some (
i
. T
i
) > 1, then there is at most one prime
k-1
for which
all of 1
}
(
k-1
) are prime. Further, since 0 = T
k
< T
k-1
< < T
1
, it follows
that if some
}
T
i
=
i
T
}
for some 0 _ < i _ k 1, then 1 <
i
,
}
[ T
i
so that (
i
. T
i
) > 1. Thus, we may assume that each
}
T
i

i
T
}
= 0. The
following result allows us to use something like a traditional sieve upper bound
for prime k-tuples, where it is not assumed that k is bounded. Note that a stronger
form of this lemma will appear in [11].
Lemma 5. Let 1
i
(n) =
i
n T
i
be linear functions for i = 1. . . . . k with
integer coefcients such that each
i
> 0, each (
i
. T
i
) = 1, and
1 :=
1

k
1_}~i_k
(
}
T
i

i
T
}
)
is nonzero. Put J(n) =
k
i=1
1
i
(n) and for each let j() be the number of
congruence classes n mod such that J(n) 0 (mod ). Assume that for
each , we have j() < . If N _ 2 and k _ log N,(10 log
2
N)
2
, then the
number of n _ N such that each 1
i
(n) is prime is at most
(ck log
1
k)
k
_
^
(^)
_
k
N(log
2
N)
k
(log N)
k
.
where c is an absolute constant and ^ is the product of the distinct primes [ 1
with > k.
Proof. We may assume that N is large since the constant c may be adjusted for
smaller values. Let 7 denote the number of n _ N with each 1
i
(n) prime. We
rst show
7 _ N
k~]_1
1=.100k log
2
N/
_
1
j()
_
O
_
N
(log N)
10k
_
. (7)
For the proof, let j(m) be the number of solutions n modulo m of the congru-
ence J(n) 0 (mod m). Clearly, j is a multiplicative function. Put N
1
=
N
1{(100k log
2
1)
. Noting that j() _ k, it follows that j(J) _ k
o(d)
holds for all
squarefree positive integers J. Taking M to be the rst even integer exceeding
10k log
2
N, we get, by the Principle of Inclusion and Exclusion and the Bonfer-
roni upper-bound inequality, that
7 _ N
1{2
k~](d)_1(d)_1
1
o(d)_
_
Nj(J)j(J)
J
O(k
o(d)
)
_
_ N
k~]_1
1
_
1
j()
_
O
_
N
1{2
d : 1(d)_1
1
o(d)_
k
o(d)
N
d : (d)y0,
1(d)_1
1
o(d)>
k
o(d)
J
_
.
It remains to look at the O-terms. For the rst sum, we have that
k
o(d)
_ k
10k log
2
12
= exp((10k log
2
N 2) log k) < N
1{9
for all large values of N uniformly in our range for k. The number of possibilities
for J is _ N
1
_ N
(10k log
2
12){(100k log
2
1)
< N
1{9
for large values of N.
Hence, the rst sum is < N
2{9
. The second one is
_
}>
N

_
_
]_1
1
k
_
_
}
_
}>
N

(k log
2
N O(k))
}
_ N
}>
_
ek log
2
N O(k)
_
}
_ N
}>
_
e
9
_
}
_
N
e
_
N
(log N)
10k
for large values of N. Note that in our range for k, this last error estimate domi-
nates the other two. Thus, we have (7).
To nish the proof of the lemma, we estimate the main term in (7). We have
log
_
k~]_1
1
_
1
j()
__
_
k~]_1
1
j()
k~]_1
1
k
][Z
k
= k log
2
N
1
k log
2
k k
][Z
log(1 1,) O(k).
Since the last sum above is log(^,(^)) and log
2
N
1
= log
2
N log
3
N
log
1
k O(1), the main term in (7) is at most
(ck log
1
k)
k
_
^
(^)
_
k
N(log
2
N)
k
(log N)
k
for some absolute constant c. Thus, by adjusting the constant c if necessary, we
have the lemma.
We apply Lemma 5 to our system of linear functions with
N = .,(m
1
. . . m
k-1
) _ ,
k-1
.
Thus, the number of choices for
k-1
_ N with each 1
i
(
k-1
) prime is at most
.(log log .)
k
m
1
. . . m
k-1
(log ,
k-1
)
k
_
c
^
(^)
k log k
_
k
.
We need an estimate for ^,(^). For this, note that each
}
T
i
in our setting is
at most .
2
, so that ^ _ .
O(k
2
)
, therefore by the minimal order of , we have
^,(^) log
1
k log
2
. log
2
.. (8)
With our choice for ,
k-1
, our upper bound for k in the lemma, and the estimate
(8), our count for the number of choices for
k-1
is now at most
.
m
1
. . . m
k-1
(log .)
k
exp(3k
2
log
3
.).
for . sufciently large.
Observe that C(
k-}
(m
}
)) _ 1 log log . holds for all = 1. . . . . k 1,
so that C((m
}
)) _ 21 log log . for each = 1. . . . . k 1 if . is sufciently
large. It then follows, by Lemma 3, that summing up over all possibilities for
m
1
. . . . . m
k-1
(positive integers m _ . such that C((m)) _ 21 log
2
.), we
have
#A
k,1
(.) _
. exp(3k
2
log
3
.)
(log .)
k
_
1_n_x
D((n))_21log log x
1
m
_
k-1
_
.
(log .)
k
exp
_
3k(21 log
2
. log
3
.)
1{2
3k
2
log
3
.
_
once . is large. This completes the proof of Lemma 4.
3 The Proof of Theorem 2
Here, we use the following theorem essentially due to Chen [5, 6].
Lemma 6. There exists .
0
such that if . > .
0
the interval .,2. .| contains ;
.,(log .)
2
primes such that (1),2 is either prime or a product of two primes
each of them exceeding .
1{10
.
Let
C
1
(.) = { .,2. .| : ( 1),2 is prime}
and let
C
2
(.) = { .,2. .| : ( 1),2 = q
1
q
2
. q
i
> .
1{10
is prime for i = 1. 2}.
We distinguish two cases.
Case 1. #C
1
(.) _ #C
2
(.).
In this case, for large .,
2
() = ( 3),2 is injective when restricted to
C
1
(.). Hence,
#V
2
(.) _ #C
1
(.) ;
.
(log .)
2
.
where the last inequality follows from Lemma 6.
Case 2. #C
1
(.) < #C
2
(.).
Let C
2
(.) and write 1 = 2q
1
q
2
, where .
1{10
< q
1
_ q
2
. Put
, = exp((log .)
4{5
). Let C
3
(.) be the subset of C
2
(.) such that q
1
> .
1{2
,,.
Since q
1
q
2
< ., we get that q
2
< .,q
1
< .
1{2
,. We nd an upper bound on
#C
3
(.). Let q
1
.
1{2
,,. .
1{2
| be a xed prime. By Bruns sieve, the number of
primes q
2
_ .,q
1
such that 2q
1
q
2
1 is a prime is
.
(q
1
)(log(.,q
1
))
2

.
q
1
(log .)
2
.
Summing the above bound for all q
1
.
1{2
,,. .
1{2
|, we get that
#C
3
(.)
.
(log .)
2
x
1=2
{,_q
1
_x
1=2
1
q
1
.
(log .)
2

log ,
log .
=
.
(log .)
11{5
= o (#C
2
(.))
as . o, where the last estimate follows again from Lemma 6.
We now look at primes C
2
(.)\C
3
(.) and we let C
4
(.) be the set of
such primes with the property that
2
() =
2
(
t
) for some
t
= also in
C
2
(.)\C
3
(.). Writing 1 = 2q
1
q
2
and
t
1 = 2q
t
1
q
t
2
, we have (q
1
1)(q
2
1) = (q
t
1
1)(q
t
2
1). Fix q
1
and q
t
1
. If q
1
= q
t
1
, we then get that q
2
= q
t
2
,
therefore =
t
, which is false. So, q
1
= q
t
1
and they are both < .
1{2
,,. Let
D = gcd(q
1
1. q
t
1
1). Then the equation
(q
1
1)(q
2
1) = (q
t
1
1)(q
t
2
1)
can be rewritten as
q
2
_
q
1
1
D
_
q
t
1
q
1
D
= q
t
2
_
q
t
1
1
D
_
.
Let = (q
1
1),D. T = (q
t
1
q
1
),D. C = (q
t
1
1),D. Then q
2
T = Cq
t
2
and and C are coprime. This puts q
2
into a xed class modulo C, namely the
congruence class of T
-1
modulo C. Let this class be C
0
, where 1 _ C
0
_
C 1. Then q
2
= C C
0
for some _ 0. We have q
2
_ .,q
1
, therefore
_ .,(q
1
C). To count such s for a given choice of q
1
. q
t
1
, note that
C C
0
= q
2
. 2q
1
C 2q
1
C
0
1 = 2q
1
q
2
1 = .

C
0
T
C
= q
t
2
. 2q
t
1
2q
t
1
_
C
0
T
C
_
1 = 2q
t
1
q
t
2
1 =
t
are all four prime numbers. By the Brun sieve (it is easy to see that since T = 0,
the four forms above satisfy the hypothesis from the Brun sieve for large .), it
follows that if we put
^ = 2q
1
q
t
1
C
0
(2q
1
C
0
1)(C
0
T)(2q
t
1
(C
0
T),C 1).
then the number of _ .,(q
1
C) with the above property is bounded by
.
(q
1
C)(log(.,q
1
C))
4
_
^
(^)
_
4
.D
q
1
q
t
1
(log log .)
4
(log ,)
4
=
.D(log log .)
4
q
1
q
t
1
(log .)
16{5
.
by the minimal order of the Euler function. Keeping now D xed and summing
the above inequality over all pairs of primes q
1
. q
t
1
_ .
1{2
which are congruent
to 1 modulo D we get, by the BrunTitchmarsh theorem, that the number of such
primes once D is xed is
.D(log log .)
4
(log .)
16{5
_
1_q_x
1=2
q=1 (mod T)
1
q
_
2
.D(log log .)
6
(D)
2
(log .)
16{5

.(log log .)
8
D(log .)
16{5
.
where we again used the minimal order of the Euler function. Summing up over
all the values for D, we nally get that
#C
4
(.)
.(log log .)
8
(log .)
16{5
T_x
1=2
1
D
.(log log .)
8
(log .)
11{5
= o(#C
2
(.))
as . o. Thus, putting C
5
(.) = C
2
(.)\(C
3
(.) L C
4
(.)), we have, by the
above calculations and Lemma 6, that #C
5
(.) ; .,(log .)
2
. Certainly,
2
is
injective when restricted to C
5
(.). This takes care of the desired lower bound.
4 Further Problems
Observe that V
k
_ V
k-1
for all k _ 2. Put V
o
=
_
k_1
V
k
. The following
result, which was conjectured by A. Chakrabarti [4], characterizes V
o
.
Theorem 7. The set V
o
is equal to the set of positive integers n whose largest
squarefree divisor is 1. 2, or 6.
Proof. It is clear that such numbers n are in V
o
, since if the largest squarefree
divisor of n is 1 or 2, then
k
(2
k
n) = n for every k, while if the largest squarefree
divisor of n is 6, then
k
(3
k
n) = n.
Suppose that n V
o
. There is thus a sequence n = n
0
. n
1
. n
2
. . . . such that
(n
i
) = n
i-1
for each i _ 1. Note that
2
((m)) _
2
(m) for m not a power
of 2. In addition, if we have equality, then m = 2
c
b
where b. c are positive
and is a prime that is 3 (mod 4). Assume that n
0
is not a power of 2, so that
2
(n
0
) _
2
(n
1
) _ . Thus, starting at some point, say n
k
, we have equality;
that is,
2
(n
k
) =
2
(n
k1
) = . Thus, for i _ 1 we have
n
ki
= 2
c
b
i
i
.
i
3 (mod 4).
We may assume that all
i
> 3 for otherwise the theorem holds. If some b
i
> 1,
then n
ki-1
= (n
ki
) is divisible by two different odd primes, namely
i
and
an odd prime factor of
i
1. Thus, we may assume that each b
i
= 1 for i _ 2.
We have
n
ki
= 2
c
i
. i _ 2.
i
= 2
i-1
1. i _ 2.
We can solve this last recurrence, getting
i
= 2
i-1
(
1
1) 1. i _ 2. But note
then since 2
]
1
-1
1 (mod
1
), we have
]
1
(
1
1) 1 0 (mod
1
).
Thus,
]
1
cannot be prime, a contradiction which proves the theorem.
Remark 2. Note that the numbers n with largest squarefree divisor 1, 2, or 6 are
precisely those n with (n) [ n. Note too that from the counting function up to .
of the integers whose largest squarefree factor is 1, 2, or 6, we have
#V
o
(.) =
1
log 3 log 4
(log .)
2
O(log .). (9)
It is possible to use the proof of Theorem 7 to show that there is a number k =
k(n) such that if n V
k
, then the largest squarefree divisor of n is 1, 2, or 6.
That is, if n is not of this form, not only does there not exist an innite reverse
Euler chain" starting at n, there also cannot exist arbitrarily long nite reverse
Euler chains starting at n. It is an interesting question to estimate k(n); in [11] it
is shown on the generalized Riemann hypothesis that k(n) log n for n > 1.
Let z(n) be the Carmichael function of n; that is, the universal exponent mod-
ulo n. This is the largest possible multiplicative order of invertible elements mod-
ulo n. For k _ 1 let z
k
(n) be the k-fold iterate of z evaluated at n. It would
be interesting to study L
k
= {z
(k)
(n)}. For k = 1, an upper bound of the
shape #L
1
(.) .,(log .)
c
1
with an inexplicit positive constant c
1
was outlined
in [9], and an actual numerical value for c
1
was established in [12]. Trivially,
#L
1
(.) ; ., log .. A slightly stronger lower bound appears in [1]. Stronger
upper and lower bounds on #L
1
(.) will appear in [16]. While #L
k
(.) seems dif-
cult to study for larger values of k, it is easy to see that the method of the present
paper shows that uniformly for . large,
#{z
k
(n) : n _ .}_
.
(log .)
k
exp
_
16k
3{2
(log
2
. log
3
.)
1{2
_
. (10)
Indeed, to see this, assume in the notation of the proof of Theorem 1, that n =
m _ ., and that > ,. Further, we may assume that z
k
(n) _ .,(log .)
k
, since
there are at most .,(log .)
k
positive integers failing this condition. We assume
that C(z
k
(n)) _ 2.9k log
2
., since otherwise Lemma 13 in [15] tells us again
that there are at most O(.,(log .)
k
) possibilities for the number of such positive
integers z
k
(n). We now note that z
k
(n) [
k
(n) and that
k
(n) _ ., therefore
k
(n),z
k
(n) _ (log .)
k
. Hence,
C(
k
(n)) = C(z
k
(n)) C(
k
(n),z
k
(n))
_ 2.9k log log .
_
k
log 2
_
log log . < 4.5k log log ..
In particular, both C(
k
()) and C(
k
(m)) are at most 4.5k log log .. The ar-
gument from the end of the proof of Theorem 1 combined with the fact that
3
_
9 3,10 2.9
_
4.5 < 16 shows that the number of possibilities for such
n _ . is at most what is shown in the right hand side of inequality (10). The
conditional argument from the introduction suggests that c
k
.,(log .)
k
should be
a lower bound on the cardinality of the above set.
Finally we remark that if n has the property that z(n) [ n, then n is in every set
L
k
, as is easy to see. It is not clear if the converse holds; for example, is n = 10
in every L
k
? It is not so easy to nd values of z that are not values of z
2
, but
in fact, one can use Bruns method to show most shifted primes 1 have this
property. By using the basic argument at the end of [7] plus the latest results on
the distribution of primes with 1( 1) small, one can prove that for large .
there are at least .
0.7067
numbers n _ . with z(n) [ n. Thus, there are at least
this many numbers n _ . which are in every L
k
, a result which stands in stark
contrast to (9).
Acknowledgments. The rst author would like to thank Williams College for
their hospitality and Professor Igor Shparlinski for enlightening conversations.
The second author would like to thank Bob Vaughan for helpful correspondence.
References
[1] W. Banks, J. Friedlander, F. Luca, F. Pappalardi and I. E. Shparlinksi, Coincidences
in the values of the Euler and Carmichael functions, Acta Arith. 122 (2006), 207
234.
[2] P. T. Bateman and R. A. Horn, A heuristic asymptotic formula concerning the distri-
bution of prime numbers, Math. Comp. 16 (1962), 363367.
[3] N. G. de Bruijn, On the number of positive integers _ . and free of prime factors
> ,, Nederl. Acad. Wetensch. Proc. Ser. A 54 (1951), 5060.
[4] A. Chakrabarti, private communication.
[5] J. R. Chen, On the representation of a large even integer as a sum of a prime and a
product of at most two primes, Kexue Tongbao 17 (1966), 385386.
[6] J. R. Chen, On the representation of a large even integer as a sum of a prime and a
product of at most two primes, Sci. Sinica 16 (1973), 157176.
[7] P. Erd os, On the normal number of prime factors of 1 and some related problems
concerning Eulers -function, Quart. J. Math. (Oxford Series) 6 (1935), 205213.
[8] P. Erd os and R. R. Hall, Eulers -function and its iterates, Mathematika 24 (1977),
173177.
[9] P. Erd os, C. Pomerance and E. Schmutz, Carmichaels lambda function, Acta Arith.,
58 (1991), 363385.
[10] K. Ford, The distribution of totients, Ramanujan J. 2 (1998), 67151.
[11] K. Ford, S. Konyagin and F. Luca, Prime chains and Pratt trees, in preparation.
[12] J. Friedlander and F. Luca, On the value set of the Carmichael z-function, J. Austral.
Math. Soc. 82 (2007), 123131.
[13] H. Halberstam and H.-E. Richert, Sieve Methods, Academic Press, London, 1974.
[14] F. Luca, On the distribution of perfect totients, J. Integer Sequences 9 (2006),
A. 06.4.4.
[15] F. Luca and C. Pomerance, Irreducible radical extensions and Euler function
chains, in: Combinatorial Number Theory, edited by Landman, Nathanson, Neetril,
Nowakowski, Pomerance, pp. 351362, Proceedings of the INTEGERS Conference
in honor of R. Grahams 70th birthday, de Gruyter, 2007.
[16] F. Luca and C. Pomerance, On the range of the Carmichael z-function, in prepara-
tion.
[17] H. Maier and C. Pomerance, On the number of distinct values of Eulers -function,
Acta Arith. 49 (1988), 263275.
[18] S. S. Pillai, On a function connected with (n), Bull. Amer. Math. Soc. 35 (1929),
837841.
Author information
Florian Luca, Instituto de Matemticas, Universidad Nacional Autonoma de Mxico,
C.P. 58089, Morelia, Michoacn, Mxico.
E-mail: fluca@matmor.unam.mx
Carl Pomerance, Department of Mathematics, Dartmouth College,
Hanover, NH 037553551, USA.
E-mail: carl.pomerance@dartmouth.edu
Frobenius Numbers of Generalized
Fibonacci Semigroups
Gretchen L. Matthews
Abstract. The numerical semigroup generated by relatively prime positive integers
a
1
; : : : ; a
n
is the set S of all linear combinations of a
1
; : : : ; a
n
with nonnegative inte-
gral coefcients. The largest integer which is not an element of S is called the Frobenius
number of S. Recently, J. M. Marn, J. L. Ramrez Alfonsn, and M. P. Revuelta deter-
mined the Frobenius number of a Fibonacci semigroup, that is, a numerical semigroup
generated by a certain set of Fibonacci numbers. In this paper, we consider numerical
semigroups generated by certain generalized Fibonacci numbers. Using a technique of
S. M. Johnson, we nd the Frobenius numbers of such semigroups obtaining the result of
Marn et al. as a special case. In addition, we determine the duals of such semigroups and
relate them to the associated Lipman semigroups.
Keywords. Fibonacci semigroup, Frobenius number, Frobenius problem, numerical
semigroup.
AMS classication. 20M99, 20M14.
1 Introduction
Given a set of relatively prime positive integers a
1
; : : : ; a
n
, let S denote the set
of linear combinations of a
1
; : : : ; a
n
with nonnegative integral coefcients. Since
a
1
; : : : ; a
n
are relatively prime, every sufciently large integer N is an element of
S. The largest integer which is not an element of S is called the Frobenius number
of S and is denoted by g.S/. The Frobenius problem is to determine g.S/. An
excellent general reference on the Frobenius problem is [11].
In discussing the Frobenius problem, it is convenient to use the terminology of
numerical semigroups. The set S dened above is called the numerical semigroup
generated by a
1
; : : : ; a
n
and is denoted by S = (a
1
; : : : ; a
n
); that is,
(a
1
; : : : ; a
n
) :=
_
n
iD1
c
i
a
i
: c
i
N
_
This work was supported in part by NSA H-98230-06-1-0008.
118 Gretchen L. Matthews
where N denotes the set of nonnegative integers. Typically, we assume that
a
i
(a
1
; : : : ; a
i1
; a
iC1
; : : : ; a
n
)
for all i , 1 _ i _ n. Then we say that S is an n-generated semigroup. General
references on numerical semigroups include [1, 7, 6, 8].
The Frobenius problem takes its name from the fact that Frobenius is said to
have mentioned it repeatedly in his lectures [3]. However, the rst published work
on this problem appears to be due to Sylvester [12] where he determined the num-
ber of elements of N \ (a; b) where a and b are relatively prime. Though not
stated explicitly in [12] (or in the often cited [13]), it is suspected that Sylvester
knew that g .(a; b)/ = abab and this fact is typically attributed to him. Given
such a simple formula for the Frobenius number of a two-generated semigroup,
it is natural to try to nd the Frobenius number of an n-generated semigroup for
other small values of n. However, Curtis proved that such a closed-form expres-
sion cannot be given for the Frobenius number of a general n-generated semigroup
for n > 2 [4]. For this reason, Frobenius problem enthusiasts often consider semi-
groups whose generators are of a particular form.
In [10], the authors determine the Frobenius numbers of so-called Fibonacci
semigroups which are numerical semigroups of the form (F
i
; F
iC2
; F
iCk
) where
F
j
denotes the j
th
Fibonacci number. In studying these semigroups, it is useful
to recall the convolution property of Fibonacci numbers:
F
n
= F
m
F
nmC1
F
m1
F
nm
for all m; n Z
C
(where Z
C
denotes the set of positive integers). A Fibonacci
semigroup is three-generated if and only if 3 _ k < i (equivalently, F
k
< F
i
).
To see this, we consider two cases depending on the value of k. If k = i , then
F
iCk
= F
2i
= L
i
F
i
(F
i
; F
iC2
)
where L
i
denotes the i
th
Lucas number. If k > i , then
F
iCk
_ F
2iC1
= F
i
F
iC2
F
i1
F
iC1
> g .(F
i
; F
iC2
)/
and so F
iCk
(F
i
; F
iC2
).
In this paper, we consider semigroups of the form
S = (a; a b; aF
k1
bF
k
)
where a > F
k
and gcd.a; b/ = 1. Such a semigroup will be called a generalized
Fibonacci semigroup. Notice that if a = F
i
and b = F
iC1
, then a b = F
iC2
and
aF
k1
bF
k
= F
i
F
k1
F
iC1
F
k
= F
iCk
Frobenius Numbers of Generalized Fibonacci Semigroups 119
and so every Fibonacci semigroup is a generalized Fibonacci semigroup. Using
a method of S. M. Johnson [9], we nd the dual of a generalized Fibonacci semi-
group. Recall that the dual of a numerical semigroup S is dened to be
B.S/ := {x N : x .S \ {0}/ _ S} :
It is immediate that g.S/ B.S/ for any numerical semigroup S = N as g.S/
s > g.S/ for all s Z
C
. Moreover,
g.S/ = max {x B.S/ : x S}
provided S = N. Hence, we determine the Frobenius number of a generalized
Fibonacci semigroup obtaining the result of [10] as a corollary.
This paper is organized as follows. Section 2 outlines Johnsons method and
applies it to nd the dual of a generalized Fibonacci semigroup. Section 3 con-
tains results relating the dual and Lipman semigroups. The paper concludes with
Section 4 where several open problems are posed.
2 Johnsons Method
We begin this section with a review of S. M. Johnsons method [9] for determining
the dual of a semigroup generated by three relatively prime positive integers.
Let S := (a
1
; a
2
; a
3
) where a
1
, a
2
, and a
3
are pairwise relatively prime. Sup-
pose N B.S/ \ S. Then N a
i
S for i = 1; 2; 3; that is,
N = y
ij
a
j
y
ik
a
k
a
i
for some y
ij
; y
ik
N. Since the a
i
are relatively prime, the semigroup generated
by any two of them has a Frobenius number. Hence, any sufciently large integer
will be contained in such a semigroup. Let
L
i
:= min
_
c : ca
i

a
j
; a
k
_
:
Then there exist x
ij
; x
ik
N such that
L
i
a
i
:= x
ij
a
j
x
ik
a
k
:
According to [9, Theorem 3], x
ij
and x
ik
are positive integers and are unique.
Recall that
N = y
21
a
1
y
23
a
3
a
2
= y
31
a
1
y
32
a
2
a
3
:
Then
N =
_
.L
2
1/ a
2
.x
13
1/ a
3
a
1
if y
21
< y
31
.x
21
1/ a
2
.L
3
1/ a
3
a
1
if y
31
< y
21
and so
B.S/ \ S =
_
.L
2
1/ a
2
.x
13
1/ a
3
a
1
;
.x
21
1/ a
2
.L
3
1/ a
3
a
1
_
:
Next, we apply this method to a generalized Fibonacci semigroup. Consider
S := (a; a b; aF
k1
bF
k
) ;
where a > F
k
and the generators of S are pairwise relatively prime. Note that
F
k
.a b/ = F
k2
a .F
k1
a F
k
b/ :
From this and the argument [9, p. 395396], it follows that
L
2
= F
k
; x
21
= F
k2
; and x
23
= 1:
In addition, we nd that
L
1
= .a b/ F
k2
_
a
F
k
_
; x
12
= a F
k
_
a
F
k
_
; x
13
=
_
a
F
k
_
;
L
2
= F
k
; x
21
= F
k2
; x
23
= 1;
L
3
=
_
a
F
k
_
1:
As a consequence,
B.S/ \ S =
__
a b F
k2
_
a
F
k
_
2
_
a b;
.F
k2
2/ a
_
a
F
k
_
.aF
k1
bF
k
/ b
_
:
This proves the next two results.
Proposition 1. Assume a > F
k
. If S = (a; a b; aF
k1
bF
k
) is generated by
three pairwise relatively prime integers, then the dual of S is
B.S/ = S L
__
a b F
k2
_
a
F
k
_
2
_
a b;
.F
k2
2/ a
_
a
F
k
_
.aF
k1
bF
k
/ b
_
:
Theorem 2. The Frobenius number of S = (a; a b; aF
k1
bF
k
) where a >
F
k
and the generators of S are pairwise relatively prime is
g.S/ = max
__
a b F
k2
_
a
F
k
_
2
_
a b;
.F
k2
2/ a
_
a
F
k
_
.aF
k1
bF
k
/ b
_
:
This theorem gives a formula for g .F
i
; F
iC2
; F
iCk
/, due to Marin et al. [10],
when a = F
i
and b = F
iC1
. However, we should point out that the technique
used in [10], different from ours, allowed the authors not only to state explicitly
when the maximum is obtained in each case but also to give a formula for the
genus (meaning [N \ S[) of such Fibonacci semigroups.
3 Duals and Lipman Semigroups
In this section, we compare two chains of semigroups. One of the chains is based
on the dual construction. The other chain arises by taking Lipman semigroups. We
rst describe the Lipman semigroup. Then we relate it to the dual of S. Finally,
we consider these for Fibonacci semigroups.
Suppose S = (a
1
; a
2
; : : : ; a
n
) is a numerical semigroup with a
1
< < a
n
.
The Lipman semigroup of S is dened as
L.S/ := (a
1
; a
2
a
1
; : : : ; a
n
a
1
) :
Clearly, S _ L.S/. Moreover, B.S/ _ L.S/ since x B.S/ implies x a
1
=
n
iD1
c
i
a
i
for some c
i
N.
Given a numerical semigroup S, both its dual B.S/ and its Lipman semigroup
L.S/ are numerical semigroups. Hence, one may iterate the B and L construc-
tions to obtain two ascending chains of numerical semigroups
B
0
.S/ := S _ B
1
.S/ := B.B
0
.S// _ _ B
hC1
.S/ := B.B
h
.S// _
and
L
0
.S/ := S _ L
1
.S/ := L.L
0
.S// _ _ L
hC1
.S/ := L.L
h
.S// _
as in [1]. We will refer to these as the B- and L-chains. Notice that for S = N,
S B.S/ and S L.S/. This together with the fact that N \ S is nite
implies that there exist smallest non-negative integers .S/ and .S/ such that
B
.S/
.S/ = N
0
= L
.S/
.S/. Since
B
0
.S/ = S = L
0
.S/;
B
1
.S/ _ L
1
.S/;
and
B
.S/
.S/ = N
0
= L
.S/
.S/;
it is natural to compare the two chains. In [1] the authors suggest that B
j
.S/ _
L
j
.S/ for all 0 _ j _ .S/. While true for two-generated semigroups, this
containment may fail in general; there are examples of four-generated semigroups
T for which B
2
.T / L
2
.T /; see [5]. This prompts the question of whether or
not B
j
.S/ _ L
j
.S/ for all j _ 0 for a three-generated semigroup S. Here, we
consider this question for Fibonacci semigroups.
Suppose that S = (a; a b; aF
k1
bF
k
) where a = F
i
and b = F
iC1
;
that is, suppose
S = (F
i
; F
iC2
; F
iCk
) :
Then the Lipman semigroup of S is L
1
.S/ = (a; b). Since a < b, L
2
.S/ =
(a; b a). Continuing this process yields the following result.
Proposition 3. Given a Fibonacci semigroup S = (F
i
; F
iC2
; F
iCk
),
L
j
.S/ =
F
ijC1
; F
ijC2
for all j , 1 _ j _ i 3. In particular, .S/ = i 3.

Of course, one may obtain similar results for generalized Fibonacci semi-
groups. Because the description depends on sizes of a and b, we omit this here
and leave the details to the reader. Since L
j
.S/ is a two-generated semigroup,
L
j
.S/ is symmetric for all 1 _ j < .S/. Therefore, L
j
.S/ is maximal in the
set of all numerical semigroups with Frobenius number g
_
L
j
.S/
_
. In particular,
we notice that if
F
i
= min
_
x B
F
i
_
F
i
1
F
k
_
F
k2
2
.S/ : x = 0
_
; (1)
then
g
_
B
F
i
_
F
i
1
F
k
_
F
k2
1
.S/
_
= g.S/
_
F
i

_
F
i
1
F
k
_
F
k2
2
_
F
i
= g .L
1
.S//
[1, Proposition I.1.11]. Hence, if (1) holds and
F
iC1
B
F
i
_
F
i
1
F
k
_
F
k2
1
.S/; (2)
then
B
F
i
_
F
i
1
F
k
_
F
k2
1
.S/ = L
1
.S/
and B
j
.S/ _ L
j
.S/ for all nonnegative integers j would follow from [5, Theo-
rem 2.6]. Unfortunately, because B
1
.S/ is not three-generated, B
2
.S/ may not be
computed using the method in Section 2. Instead, one may compute this directly
from the denition and obtain the next result.
Proposition 4. Given a Fibonacci semigroup S = (F
i
; F
iC2
; F
iCk
)
B
2
.S/ =
_
F
i
; F
iC2
; F
iCk
; l F
iC2
; h F
iCk
if x
32
= 1;
l F
i
if k > 4 or x
12
= 1; l F
iCk
if x
13
_ 2 or x
13
= x
31
= 1;
h F
i
if x
31
_ 2 or x
31
= x
13
= 1;
h F
iC2
if x
12
_ 2 or x
12
= x
21
= 1
_
where
h = .F
k
1/ F
iC2

_
F
iC2
F
k2
__
F
i
F
k
_
1
_
1
_
F
i
F
iCk
and
l = .F
k
1/ F
iC2

___
F
i
F
k
1
__
F
k
F
i
1
_
F
iCk
F
i
:
Consequently,
B
2
.S/ _ L
1
.S/ _ L
2
.S/:
In light of Proposition 4, determining B
j
.S/ for j > 3 will not be an imme-
diate consequence of Johnsons method. We leave this (and settling when (1) and
(2) hold) as a problem for further study.
4 Conclusion
In this paper, we determined the Frobenius number of generalized Fibonacci semi-
groups. In addition, we also obtained the dual of such a semigroup. We leave as
open problems to
(i) determine if B
j
.S/ _ L
j
.S/ for each j _ 0 for Fibonacci semigroups S
(or, more generally, three-generated semigroups), and
(ii) nd the Frobenius number and dual of other semigroups generated by gener-
alized Fibonacci numbers.
We conclude by mentioning another interesting problem relating numerical semi-
groups and Fibonacci numbers. M. Bras-Amoros conjectures that the number of
semigroups with a particular genus g behaves asymptotically as the Fibonacci
sequence [2].
References
[1] V. Barucci, D. E. Dobbs and M. Fontana, Maximality properties in numerical semi-
groups and applications to one-dimensional analytically irreducible local domains,
Memoirs Amer. Math. Soc. 125/598 (1997).
[2] M. Bras-Amoros, Fibonacci-like behavior of the number of numerical semigroups
of a given genus, to appear in Semigroup Forum.
[3] A. Brauer, On a problem of partitions, Amer. J. Math. 64 (1942), 299312.
[4] F. Curtis, On formulas for the Frobenius numer of a numerical semigroup, Math
Scand. 67 (1990), 190192.
[5] D. E. Dobbs and G. L. Matthews, On comparing two chains of numerical semigroups
and detecting Arf semigroups, Semigroup Forum 63 (2001), 237246.
[6] R. Frberg, C. Gottlieb and R. Hggkvist, Semigroups, semigroup rings and analyt-
ically irreducible rings, Reports Dept. Math. Univ. Stockholm no. 1 (1986).
[7] R. Frberg, C. Gottlieb and R. Hggkvist, On numerical semigroups, Semigroup
Forum 35 (1987), no. 1, 6383.
[8] R. Gilmer, Commutative Semigroup Rings, Chicago Lectures in Mathematics, Uni-
versity of Chicago Press, Chicago, IL, 1984.
[9] S. M. Johnson, A linear Diophantine problem, Canad. J. Math. 12 (1960), 390398.
[10] J. M. Marn, J. L. Ramrez Alfonsn and M. P. Revuelta, On the Frobenius number
of Fibonacci numerical semigroups, Integers 7 (2007), A14.
[11] J. L. Ramrez Alfonsn, The Diophantine Frobenius Problem, Oxford Lecture Series
in Mathematics and its Applications 30, Oxford University Press, 2005.
[12] J. J. Sylvester, On Subvariants, i.e. semi-invariants to binary quantics of an unlimited
order, Amer. J. Math. 5 (1882), no. 14, 79136.
[13] J. J. Sylvester, Mathematical questions with their solutions, Educational Times 41
(1884), 21.
Author information
Gretchen L. Matthews, Department of Mathematical Sciences, Clemson University,
Clemson, SC 29634-0975, USA.
E-mail: gmatthe@clemson.edu
Combinatorics of
RamanujanSlater Type Identities
James McLaughlin and Andrew V. Sills
Abstract. We provide the missing member of a family of four q-series identities related
to the modulus 36, the other members having been found by Ramanujan and Slater. We
examine combinatorial implications of the identities in this family, and of some of the
identities we considered in Identities of the Ramanujan-Slater type related to the moduli
18 and 24.
Keywords. RogersRamanujan type identities, partitions, signed partitions.
AMS classication. 11B65, 05A17, 11P55, 05A19.
1 Introduction
The RogersRamanujan identities,
1
X
jD0
q
j
2
(q: q)
j
=
Y
k=1
k1.mod 5/
1
1 q
k
(1.1)
and
1
X
jD0
q
j
2
Cj
(q: q)
j
=
Y
k=1
k2.mod 5/
1
1 q
k
. (1.2)
where
(a: q)
j
:=
j1
Y
kD0
(1 aq
k
)
were rst proved by L. J. Rogers [13] in 1894 and later independently redis-
covered (without proof) by S. Ramanujan [12, Vol II, p. 33]. Many additional
q-series = innite product identities were found by Ramanujan and recorded in
his lost notebook [5], [6]. A large collection of such identities was produced by
L. J. Slater [17].
126 James McLaughlin and Andrew V. Sills
Just as the RogersRamanujan identities (1.1), (1.2) are a family of two similar
identities where the innite products are related to the modulus 5, most Rogers
Ramanujan type identities exist in a family of several similar identities where the
sum sides are similar and the product sides involve some common modulus.
In most cases Ramanujan and Slater found all members a given family, but in a
few cases they found just one or two members of a family of four or ve identities.
In [11], we found some missing members of families of identities related to the
moduli 18 and 24 where Ramanujan and/or Slater had found one or two of the
family members, as well as two new complete families.
In this paper, we nd the missing member in a family of four identities related
to the modulus 36. We examine combinatorial implications of the identities in this
family, and of some of the identities we considered in [11].
2 Combinatorial Denitions
Informally, a partition of an integer n is a representation of n as a sum of positive
integers where the order of the summands is considered irrelevant. Thus the ve
partitions of 4 are 4 itself, 3 1, 2 2, 2 1 1, and 1 1 1 1. The
summands are called the parts of the partition, and since the order of the parts
is irrelevant, 2 1 1, 1 2 1, and 1 1 2 are all considered to be the
same partition of 4. It is often convenient to impose a canonical ordering for the
parts and to separate parts with commas instead of plus signs, and so we make the
following denitions:
A partition z of an integer n into parts is an -tuple of positive integers
(z
1
. z
2
. . . . . z
`
) where
z
i
= z
iC1
for 1 5 i 5 1, and
`
X
iD1
z
i
= n.
The number of parts = (z) of z is also called the length of z. The sum of the
parts of z is called the weight of z and is denoted [z[.
Thus in this notation, the ve partitions of 4 are (4), (3. 1), (2. 1. 1), and
(1. 1. 1. 1).
In [4], G. Andrews considers some of the implications of generalizing the no-
tion of partition to include the possibility of some negative integers as parts. We
may formalize this idea with the following denitions:
A signed partition o of an integer n is a partition pair (. v) where n = [[
[v[. We may call (resp., v) the positive (resp., negative) subpartition of o and
1
.
2
. . . . .
`./
(resp. v
1
. v
2
. . . . . v
`./
) the positive (resp. negative) parts of o.
Combinatorics of RamanujanSlater Type Identities 127
Thus ((6. 3. 3. 1). (4. 2. 1. 1)), which represents 6 3 3 1 1 1 2 4,
is an example of a signed partition of 5. Of course, there are innitely many
unrestricted signed partitions of any integer, but when we place restrictions on
how parts may appear, signed partitions arise naturally in the study of certain q-
series.
Remark 2.1. Notice that the way we have dened signed partitions, the nega-
tive parts are positive numbers (which count negatively toward the weight of the
signed partition), much as the imaginary part of a complex number is real.
3 Partitions and q-series Identities of Ramanujan
and Slater
Using ideas that originated with Euler, MacMahon [12, vol. II, Ch. III] and
Schur [14] independently realized that (1.1) and (1.2) imply the following par-
tition identities:
Theorem 3.1 (First RogersRamanujan identity combinatorial version). For all
integers n, the number of partitions z of n where
z
i
z
iC1
= 2 for 1 5 i 5 (z) 1 (3.1)
equals the number of partitions of n into parts congruent to 1 (mod 5).
Theorem 3.2 (Second RogersRamanujan identity combinatorial version). For
all integers n, the number of partitions z of n where
z
i
z
iC1
= 2 for 1 5 i 5 (z) 1 (3.2)
and
z
`./
> 1. (3.3)
equals the number of partitions of n into parts congruent to 2 (mod 5).
When studying sets of partitions where the appearance or exclusion of parts is
governed by difference conditions such as (3.1), it is often useful to introduce a
second parameter a. The exponent on a indicates the length of a partition being
enumerated, while the exponent on q indicates the weight of the partition.
For example, it is standard to generalize (1.2) and (1.1) as follows:
J
1
(a. q) :=
1
X
jD0
a
j
q
j
2
Cj
(q: q)
j
=
1
(aq: q)
1
1
X
jD0
(1)
j
a
2j
q
j.5jC3/=2
(a: q)
j
(1 aq
2jC1
)
(q: q)
j
(3.4)
J
2
(a. q) :=
1
X
jD0
a
j
q
j
2
(q: q)
j
=
1
(aq: q)
1
1
X
jD0
(1)
j
a
2j
q
j.5j1/=2
(a: q)
j
(1 aq
2j
)
(1 a)(q: q)
j
.
(3.5)
where
(a: q)
1
:=
1
Y
kD0
(1 aq
k
).
It is then easily seen that J
1
(a. q) and J
2
(a. q) satisfy the following system of
q-difference equations:
J
1
(a. q) = J
2
(aq. q). (3.6)
J
2
(a. q) = J
1
(a. q) aqJ
1
(aq. q). (3.7)
Notice that there are straightforward combinatorial interpretations to (3.6) and
(3.7). Equation (3.6) states that if we start with the collection of partitions satis-
fying (3.1) and add 1 to each part (i.e., replace a by aq), then we obtain the set of
partitions that satsify (3.2) and (3.3); the difference condition is maintained, but
the new partitions will have no ones. The left-hand side of (3.7) generates par-
titions that satisfy (3.1) while the right-hand side segregates these partitions into
two classes: those where no ones appear (generated by J
1
(a. q)) and those where
a unique one appears (generated by aqJ
1
(aq. q)).
Remark 3.3. It may seem awkward to have the a-generalization of the rst (resp.
second) RogersRamanujan identity labeled J
2
(a. q) (resp., J
1
(a. q)), but this is
actually standard practice (see, e.g., Andrews [2, Ch. 7]). Here and in certain
generalizations, the subscript on J corresponds to one more than the maximum
number of ones which can appear in the partitions enumerated by the function.
Remark 3.4. While the a-generalizations are useful for studying the relevant par-
titions, the price paid for generalizing (1.1) to (3.5) and (1.2) to (3.4) is that the a-
generalizations no longer have innite product representations; only in the a = 1
cases will Jacobis triple product identity [10, p. 15, Eq. (1.6.1)] allow the right-
hand sides of (1.2) and (1.1) to be transformed into innite products.
An exception to Remark 3.4 may be found in one of the identities in Ramanu-
jans lost notebook [6, Entry 5.3.9]; cf. [11, Eq. (1.16)]:
1
X
jD0
q
j
2
(q
3
: q
6
)
j
(q: q
2
)
2
j
(q
4
: q
4
)
j
=
Y
j=1
j1.mod 2/ or j2.mod 12/
1
1 q
j
. (3.8)
Equation (3.8) admits an a-generalization with an innite product:
1
X
jD0
a
j
q
j
2
(q
3
: q
6
)
j
(q: q
2
)
j
(aq: q
2
)
j
(q
4
: q
4
)
j
=
Y
j=1
1 aq
4j2
a
2
q
8j4
1 aq
2j1
. (3.9)
Notice that the right-hand side of (3.9) is easily seen to be equal to
X
n;`=0
s(. n)a
`
q
n
.
where s(. n) denotes the number of partitions of n into exactly parts where
no even part appears more than twice nor is divisible by 4. Note also that the
right-hand side of (3.8) generates partitions where parts may appear as in Schurs
1926 partition theorem [15] (i.e. partitions into parts congruent to 1 (mod 6)),
dilated by a factor of 2, along with unrestricted appearances of odd parts. It is
a fairly common phenomenon for a RogersRamanujan type identity to generate
partitions whose parts are restricted according to a well-known partition theorem,
dilated by a factor of m, and where nonmultiples of m may appear without restric-
tion. See, e.g., Connor [8] and Sills [16].
Apartner to (3.8) was found by Slater [17, p. 164, Eq. (110), corrected], cf. [11,
Eq. (1.19)]:
1
X
jD0
q
j
2
C2j
(q
3
: q
6
)
j
(q: q
2
)
j
(q: q
2
)
jC1
(q
4
: q
4
)
j
=
Y
j=1
j1.mod 2/ or j4.mod 12/
1
1 q
j
. (3.10)
An a-generalization of (3.10) is
1
X
jD0
a
j
q
j
2
C2j
(q
3
: q
6
)
j
(q: q
2
)
j
(aq: q
2
)
jC1
(q
4
: q
4
)
j
=
Y
j=1
1 aq
4j
a
2
q
8j
1 aq
2j1
=
X
n;`=0
t (. n)a
`
q
n
.
(3.11)
where t (. n) denotes the number of partitions of n into parts where even parts
appear at most twice and are divisible by 4.
Remark 3.5. An explanation as to why (3.8) and (3.10) admit a-generalizations
which include innite products and (1.1) and (1.2) do not, may be found in the
theory of basic hypergeometric series. The RogersRamanujan identities (1.1)
and (1.2) arise as limiting cases of Watsons q-analog of Whipples theorem [18],
[10, p. 43, Eq. (2.5.1)]; see [10, pp. 4445, 2.7]. In contrast, (3.9) and (3.11) are
special cases of Andrewss q-analog of Baileys
2
J
1
(
1
2
) sum [1, p. 526, Eq. (1.9)],
[10, p. 354, Eq. (II.10)].
Remark 3.6. S. Corteel and J. Lovejoy interpreted (3.8) and (3.10) combinatori-
ally using overpartitions in [9].
4 A Family of Ramanujan and Slater
4.1 A Long-lost Relative
Let us dene Q(n. .) := (n.
1
. .. n: n)
1
(n.
2
. n.
2
: n
2
)
1
. where
(a
1
. a
2
. . . . . a
r
: n)
1
:=
r
Y
kD1
(a
k
: n)
1
.
Then it is clear that an identity is missing from the family
1
X
jD0
q
2j.jC2/
(q
3
: q
6
)
j
(q
2
: q
2
)
2jC1
(q: q
2
)
j
=
Q(q
18
. q
7
)
(q
2
: q
2
)
1
(Slater [17, Eq. (125)]). (4.1)
1
X
jD0
q
2j.jC1/
(q
3
: q
6
)
j
(q
2
: q
2
)
2jC1
(q: q
2
)
j
=
Q(q
18
. q
5
)
(q
2
: q
2
)
1
(Slater [17, Eq. (124)]). (4.2)
1
X
jD0
q
2j
2
(q
3
: q
6
)
j
(q
2
: q
2
)
2j
(q: q
2
)
j
=
Q(q
18
. q
3
)
(q
2
: q
2
)
1
(Ramanujan [6, Entry 5.3.4]). (4.3)
The following identity completes the above family:
1
X
jD0
q
2j.jC1/
(q
3
: q
6
)
j
(q
2
: q
2
)
2j
(q: q
2
)
jC1
=
Q(q
18
. q)
(q
2
: q
2
)
1
. (4.4)
Theorem 4.1. Identity (4.4) is valid.
Proof. We show that (4.2)q(4.1) =(4.4). For the series side,
1
X
jD0
q
2j.jC1/
(q
3
: q
6
)
j
(q
2
: q
2
)
2jC1
(q: q
2
)
j
q
1
X
jD0
q
2j.jC2/
(q
3
: q
6
)
j
(q
2
: q
2
)
2jC1
(q: q
2
)
j
=
1
X
jD0
q
2j.jC1/
(q
3
: q
6
)
j
(1 q
2jC1
)
(q
2
: q
2
)
2jC1
(q: q
2
)
j
=
1
X
jD0
q
2j.jC1/
(q
3
: q
6
)
j
(q
2
: q
2
)
2j
(q: q
2
)
jC1
.
For the product side, we make use of the quintuple product identity:
Q(n. .) = (n.
3
. n
2
.
3
. n
3
: n
3
)
1
.(n.
3
. n
2
.
3
. n
3
: n
3
)
1
.
Hence
Q(q
18
. q
5
) qQ(q
18
. q
7
)
= (q
33
. q
21
. q
54
: q
54
)
1
q
5
(q
3
. q
51
. q
54
: q
54
)
1
q((q
39
. q
15
. q
54
: q
54
)
1
q
7
(q
3
. q
57
. q
54
: q
54
)
1
)
= (q
33
. q
21
. q
54
: q
54
)
1
q
5
(q
3
. q
51
. q
54
: q
54
)
1
q((q
39
. q
15
. q
54
: q
54
)
1
q
4
(q
51
. q
3
. q
54
: q
54
)
1
)
= (q
21
. q
33
. q
54
: q
54
)
1
q(q
15
. q
39
. q
54
: q
54
)
1
= Q(q
18
. q).
The result now follows.
4.2 Combinatorial Interpretations
We interpret (4.3) combinatorially.
Theorem 4.2. The number of signed partitions o = (. v) of n, where
() is even, and each positive part is even and = (), and
the negative parts are odd, less than (), and may appear at most twice
equals the number of (ordinary) partitions of n into parts congruent to 2, 3,
4, 8 (mod 18).
Proof. Starting with the left-hand side of (4.3), we nd
1
X
jD0
q
2j
2
(q
3
: q
6
)
j
(q
2
: q
2
)
2j
(q: q
2
)
j
=
1
X
jD0
q
2j
2
Q
j
kD1
(1 q
2k1
q
4k2
)
(q
2
: q
2
)
2j
=
1
X
jD0
q
2j
2
Q
j
kD1
q
4k2
(1 q
.2k1/
q
.4k2/
)
(q
2
: q
2
)
2j
=
1
X
jD0
q
2j
2
C4.1C2CCj/2j
Q
j
kD1
(1 q
.2k1/
q
.4k2/
)
(q
2
: q
2
)
2j
=
1
X
jD0
q
4j
2
(q
2
: q
2
)
2j
j
Y
kD1
(1 q
.2k1/
q
.4k2/
).
Notice that
1
(q: q)
2j
is the generating function for partitions into at most 2 parts, thus
q
2j
2 1
(q: q)
2j
=
q
2j terms

(q: q)
2j
is the generating function for partitions into exactly 2 parts, where each part is
at least . Thus
q
4j
2 1
(q
2
: q
2
)
2j
is the generating function for partitions into exactly 2 parts, each of which is
even and at least 2 . Also,
Q
j
kD1
(1 q
.2k1/
q
.4k2/
) is the generating
function for signed partitions into odd negative parts < 2 and appearing at most
twice each. Summing over all , we nd that the left-hand side of (4.3) is the
generating function for signed partitions o = (. v) of n, where () is even, and
each positive part is even and = (), and the negative parts are odd, less than
(), and may appear at most twice.
Now the right-hand side of (4.3) is
Q(q
18
. q
3
)
(q
2
: q
2
)
1
=
(q
3
. q
15
: q
18
: q
18
)
1
(q
12
. q
24
: q
36
)
1
(q
2
: q
2
)
1
=
Y
i=1
i2;3;4;8.mod 18/
1
1 q
i
.
which is clearly the generating function for partitions into parts congruent to
2. 3. 4. 8 (mod 18).
Remark 4.3. Andrews provided a different combinatorial interpretation of (4.3)
in [3, p. 175, Theorem 2].
Following the ideas in the proof of Theorem 4.2, the analogous combinatorial
interpretation of Identity (4.2) is as follows.
() is odd, and each positive part is even and = () 1, and
5, 6 (mod 18).
We next interpret (4.1) combinatorially. Note that the theorem equates the
number in a certain class of signed partitions of n1 with the number in a certain
class of regular partitions of n.
Theorem 4.5. The number of signed partitions o = (. v) of n 1, where
() is odd, and each positive part is odd and = (), and
7, 8 (mod 18).
Proof. The proof is similar to that of Theorem 4.2, except that
1
X
jD0
q
2j.jC2/
(q
3
: q
6
)
j
(q
2
: q
2
)
2jC1
(q: q
2
)
j
=
1
q
1
X
jD0
q
4j
2
C4jC1
(q
2
: q
2
)
2jC1
j
Y
kD1
(1q
.2k1/
q
.4k2/
).
and since
4
2
4 1 = (2 1) (2 1) (2 1)

2jC1 terms
.
it follows that
q
4j
2
C4jC1
(q
2
: q
2
)
2jC1
is the generating function for partitions into exactly 2 1 parts, each of which
is odd and at least 2 1.
Lastly, we give a combinatorial interpretation of (4.4).
Theorem 4.6. The number of signed partitions o = (. v) of n 1, where
contains an odd positive part m (which may be repeated), exactly m 1

positive even parts, all = m 1, and
negative parts are all odd, < m, and appear at most twice,
6, 8 (mod 18).
Proof. This time
1
X
jD0
q
2j.jC1/
(q
3
: q
6
)
j
(q
2
: q
2
)
2j
(q: q
2
)
jC1
=
1
q
1
X
jD0
q
4j
2
(q
2
: q
2
)
2j
q
2jC1
1 q
2jC1
j
Y
kD1
(1 q
.2k1/
q
.4k2/
).
so that, as before,
q
4j
2
(q
2
: q
2
)
2j
is the generating function for partitions into exactly 2 parts, each of which is
even and at least 2 , and
q
2jC1
1 q
2jC1
generates partitions consisting of the part 2 1 and containing at least one such
part.
5 Combinatorial Interpretations of a Family of
Mod 18 Identities
In [11], we presented the following family of RogersRamanujanSlater type
identities related to the modulus 18:
1
X
jD0
q
j.jC1/
(1: q
3
)
j
(1: q)
j
(q: q)
2j
=
(q. q
8
. q
9
: q
9
)
1
(q
7
. q
11
: q
18
)
1
(q: q)
1
. (5.1)
1
X
jD0
q
j
2
(1: q
3
)
j
(1: q)
j
(q: q)
2j
=
(q
2
. q
7
. q
9
: q
9
)
1
(q
5
. q
13
: q
18
)
1
(q: q)
1
. (5.2)
1
X
jD0
q
j.jC1/
(q
3
: q
3
)
j
(q: q)
j
(q: q)
2jC1
=
(q
3
. q
6
. q
9
: q
9
)
1
(q
3
. q
15
: q
18
)
1
(q: q)
1
. (5.3)
1
X
jD0
q
j.jC2/
(q
3
: q
3
)
j
(q
2
: q
2
)
j
(q
jC2
: q)
jC1
=
(q
4
. q
5
. q
9
: q
9
)
1
(q. q
17
: q
18
)
1
(q: q)
1
. (5.4)
We give a combinatorial interpretation of (5.4).
Theorem 5.1. The number of signed partitions o = (. v) of n 2, wherein
1
, the largest positive part, is even,
the integers 1. 2. . . . .

1
2
1 all appear an even number of times and at least
twice,
the integer

1
2
does not appear,
the integers

1
2
1.

1
2
2. . . . .
1
all appear at least once, and
there are exactly

1
2
1 negative parts, each 1 (mod 3) and 5
3
1
2
2,
with the parts greater than 1 occurring at most once
6, 7, 8 (mod 18).
Proof. We consider the general term on the left side of (5.4),
q
j.jC2/
(q
3
: q
3
)
j
(q
2
: q
2
)
j
(q
jC2
: q)
jC1
=
q
j
2
C2j
(q
2
: q
2
)
j
q
.3j
2
C3j/=2
Q
j
kD0
(1 q
jC2Ck
)
j
Y
kD1
(1 q
3k
)
=
1
q
2
q
j
2
Cj
(q
2
: q
2
)
j
q
.3j
2
C7jC4/=2
Q
j
kD0
(1 q
jC2Ck
)
q
j
j
Y
kD1
(1 q
3k
).
The factors
q
j
2
Cj
(q
2
: q
2
)
j
=
q
2C4C6CC2j
(q
2
: q
2
)
j
generates parts in {2. 4. 6. . . . . 2 } where each part appears at least once. Then
by mapping each even part 2r to r r, we have parts {1. 2. 3. . . . . } where each
part appears an even number of times and at least twice.
The factors
q
.3j
2
C7jC4/=2
Q
j
kD0
(1 q
jC2Ck
)
=
q
.jC2/C.jC3/CC.2j C2/
Q
j
kD0
(1 q
jC2Ck
)
generates partitions from the parts { 2. 3. . . . . 2 1. 2 2} and where
each part appears at least once. Lastly,
q
j
j
Y
kD1
(1 q
3k
)
is the generating function for signed partitions with negative parts that are congru-
ent to 1 modulo 3, _ 3 1, the parts greater than 1 occur at most once, and the
total number of parts is (the number of 1s being minus the number of other
parts).
Upon summing over _ 0, we get that
1
X
jD0
q
j
2
Cj
(q
2
: q
2
)
j
q
.3j
2
C7jC4/=2
Q
j
kD0
(1 q
jC2Ck
)
q
j
j
Y
kD1
(1 q
3k
)
is the generating function for signed partitions with the properties itemized in the
statement of the theorem.
The right side of (5.4) is
(q
4
. q
5
. q
9
: q
9
)
1
(q. q
17
: q
18
)
1
(q: q)
1
=
Y
i=1
i2;3;6;7;8.mod 18/
1
1 q
i
.
which is the generating function for partitions into parts congruent to 2, 3,
6, 7, 8 (mod 18).
The corresponding combinatorial interpretation of (5.2) is given by the follow-
ing theorem.
1

1
2
twice,
the integers

1
2
.

1
2
1. . . . .
1
there are exactly

1
2
3
1
2
1,
with the parts greater than 2 occurring at most once,
4, 6, 8 (mod 18).
Proof. The proof is similar to that of Theorem 5.1, except we rewrite the general
term on the left side of (5.2) as follows
q
j
2
(1: q
3
)
j
(1: q)
j
(q: q)
2j
=
q
j
2
(q
2
: q
2
)
j1
q
.3j
2
3j/=2
Q
j
kD0
(1 q
jCk
)
j1
Y
kD1
(1 q
3k
)
=
q
j
2
j
(q
2
: q
2
)
j1
q
.3j
2
C3j/=2
Q
j
kD0
(1 q
jCk
)
q
2j
j1
Y
kD1
(1 q
3k
).
The identity at (5.1) may be interpreted combinatorially as follows.
1

1
2
twice,
the integers

1
2
.

1
2
1. . . . .
1
there are exactly

1
2
3
1
2
2,
with the parts greater than 1 occurring at most once,
4, 5, 6 (mod 18).
Proof. The general term on the left side of (5.1) may be written as
q
j
2
Cj
(1: q
3
)
j
(1: q)
j
(q: q)
2j
=
q
j
2
Cj
(q
2
: q
2
)
j1
q
.3j
2
3j/=2
Q
j
kD0
(1 q
jCk
)
j1
Y
kD1
(1 q
3k
)
=
q
j
2
j
(q
2
: q
2
)
j1
q
.3j
2
C3j/=2
Q
j
kD0
(1 q
jCk
)
q
j
j1
Y
kD1
(1 q
3k
).
Finally, we provide a combinatorial interpretation of (5.3).
Theorem 5.4. The number of signed partitions o = (. v) of n 1, wherein
1
, the largest positive part, is odd,

1
1
2
all appear an even number of times and at least
twice,
the integers

1
1
2
1.

1
1
2
2. . . . .
1
there are exactly

1
1
2
negative parts, each 1 (mod 3) and 5
3
1
2
2, with
the parts greater than 1 occurring at most once,
equals the number of (ordinary) partitions of n into parts congruent to 1, 2, 4, 5,
7 or 8 modulo 9, such that for any nonnegative integer , 9 1 and 9 2 do
not both appear, and for any nonnegative integer k, 9k 7 and 9k 8 do not
both appear.
Proof. The general term on the left side of (5.3) may be written as
q
j.jC1/
(q
3
: q
3
)
j
(q: q)
j
(q: q)
2jC1
=
q
j
2
Cj
(q
2
: q
2
)
j
q
.3j
2
C3j/=2
(q
jC1
: q)
jC1
j
Y
kD1
(1 q
3k
)
=
1
q
q
j
2
Cj
(q
2
: q
2
)
j
q
.3j
2
C5jC2/=2
(q
jC1
: q)
jC1
q
j
j
Y
kD1
(1 q
3k
):
thus, the interpretation of the left side is similar to that of the previous identities.
The right side of (5.3) provides a challenge because of the double occurrence
of the factors (q
3
: q
18
)
1
and (q
15
: q
18
)
1
in the numerator. Accordingly, we turn
to a partition enumeration technique introduced by Andrews and Lewis. In [7,
p. 79, Eq. (2.2) with k = 9], they show that
(q
aCb
: q
18
)
1
(q
a
. q
b
: q
9
)
1
(5.5)
is the generating function for partitions of n into parts congruent to a or b modulo
9 such that for any k, 9k a and 9k b do not both appear as parts, where
0 < a < b < 9.
With this in mind, we immediately see that
(q
3
. q
6
. q
9
: q
9
)
1
(q
3
. q
15
: q
18
)
1
(q: q)
1
=
1
(q
4
. q
5
: q
9
)
1
(q
3
: q
18
)
1
(q. q
2
: q
9
)
1
(q
15
: q
18
)
1
(q
7
. q
8
: q
9
)
1
generates the partitions stated in our theorem.
Acknowledgments. We thank the referee for carefully reading the manuscript
and supplying helpful comments, and the conference organizers for the opportu-
nity to present at Integers 2007.
References
[1] G. E. Andrews, On the q-analog of Kummers theorem and applications, Duke Math.
J. 40 (1973), 525528.
[2] G. E. Andrews, The Theory of Partitions, Encyclopedia of Mathematics and its Ap-
plications, vol. 2, Addison-Wesely, 1976; reissued Cambridge, 1998.
[3] G. E. Andrews, Ramanujans lost notebook II: 0-function expansions, Adv. in
Math. 41 (1981), 173185.
[4] G. E. Andrews, Eulers De Partitio Numerorum, Bull. Amer. Math. Soc. 44 (2007),
561574.
[5] G. E. Andrews and B. C. Berndt, Ramanujans Lost Notebook. Part I, Springer, 2005
[6] G. E. Andrews and B. C. Berndt, Ramanujans Lost Notebook. Part II, Springer,
2009.
[7] G. E. Andrews and R. Lewis, An algebraic identity of F. H. Jackson and its implica-
tions for partitions, Discrete Math. 232 (2001), 7783.
[8] W. G. Connor, Partition theorems related to some identities of Rogers and Watson,
Trans. Amer. Math. Soc. 214 (1975), 95111.
[9] S. Corteel and J. Lovejoy, Overpartitions and the q-Bailey identity, to appear in
Proc. Edinburgh Math. Soc.
[10] G. Gasper and M. Rahman, Basic Hypergeometric Series, 2nd ed., Cambridge,
2004.
[11] J. Mc Laughlin and A. V. Sills, Identities of the RamanujanSlater type related to
the moduli 18 and 24, J. Math. Anal. Appl. 344 (2008), no. 2, 765777.
[12] P. A. MacMahon, Combinatory Analysis, two volumes, Cambridge, 19151916.
Reprinted: Chelsea, 1960.
[13] L. J. Rogers, Second memoir on the expansion of certain innite products, Proc.
London Math. Soc. 25 (1894), 318343.
[14] I. Schur, Ein Beitrag zur additiven Zahlentheorie und zur Theorie der Kettenbrche,
Sitz. Preuss. Akad. Wiss. Phys.-Math. Kl. (1917), 302321.
[15] I. Schur, Zur additiven Zahlentheorie, Sitz. Preuss. Akad. Wiss. Phys.-Math. Kl.
(1926), 488495.
[16] A. V. Sills, Identities of the RogersRamanujanSlater type, Int. J. Number Theory
3 (2007), 293323.
[17] L. J. Slater, Further identities of the RogersRamanujan type, Proc. London Math.
Soc. (2) 54 (1952), 147167.
[18] G. N. Watson, A new proof of the RogersRamanujan identities, J. London Math.
Soc. 4 (1929), 49.
Author information
James McLaughlin, Department of Mathematics, West Chester University,
West Chester, PA 19383, USA.
E-mail: jmclaughl@wcupa.edu
Andrew V. Sills, Department of Mathematical Sciences, Georgia Southern University,
Statesboro, GA 30460, USA.
E-mail: ASills@GeorgiaSouthern.edu
A Mock Theta Function for the
Delta-function
Ken Ono
Abstract. We dene a mock theta function
M
Z
(:) =
1
nD1
a
Z
(n)q
n
= 11 q
1
2615348736000
691

for Ramanujans Delta-function. Although it is not a modular form, we are able to deter-
mine its images under the Hecke operators. These images are in terms of Ramanujans
tau-function and the modular polynomials
n
(:) which encode the denominator formula
for the innite dimensional Monster Lie algebra. Using these results, we obtain a criterion
for Lehmers Conjecture on the nonvanishing of t(n). We show that t() = 0 if and only
if M
Z
(:) [ T
10
() is a modular form. We also relate Lehmers Conjecture to integrality
properties of the a
Z
(n).
Keywords. Mock theta function, Maass form.
AMS classication. 11F03, 11F11.
For George E. Andrews in celebration of his 70th birthday.
1 Introduction and Statement of Results
As usual, let
^(:) =
1
nD1
t(n)q
n
:= q
1
nD1
(1 q
n
)
24
= q 24q
2
252q
3
(1.1)
(note q := e
2tiz
throughout) be the unique normalized weight 12 cusp form
on SL
2
(Z). Its coefcients, the so-called values of Ramanujans tau-function,
provide examples of some of the deepest phenomena in the theory of holomorphic
modular forms. Here we show how its values also arise in the theory of non-
holomorphic modular forms, the so-called Maass forms.
142 Ken Ono
To make this precise, we rst recall some classical modular forms. As usual,
let
1
4
(:) := 1 240
1
nD1
o
3
(n)q
n
.
1
6
(:) := 1 504
1
nD1
o
5
(n)q
n
denote the normalized weight 4 and 6 Eisenstein series for the modular group
SL
2
(Z). Throughout, we let o
(n) :=
djn
J
. Let (:) be the classical modular

function
(:) :=
1
4
(:)
3
^(:)
=
1
nD1
c(n)q
n
= q
1
744 196884q . (1.2)
We now recall an important sequence of modular functions related to the Mon-
ster. Let
0
(:) := 1, and for every positive integer m, let
n
(:) be the unique mod-
ular function which is holomorphic on H, the upper-half of the complex plane,
whose q-expansion is of the form
n
(:) = q
n
nD1
c
n
(n)q
n
. (1.3)
It is a standard fact that each
n
(:) is a monic degree m polynomial in (:) with
integer coefcients. The rst few
n
are:
0
(:) = 1.
1
(:) = (:) 744 = q
1
196884q .
2
(:) = (:)
2
1488(:) 159768 = q
2
42987520q .
3
(:) = (:)
3
2232(:)
2
1069956(:) 36866976
= q
3
2592899910q .
These polynomials are easily described using Hecke operators. For primes
and integral weights k, the Hecke operator T
k
() is dened by
_
a(n)q
n
_
[ T
k
() :=
_
a(n)
k1
a(n,)
_
q
n
. (1.4)
These operators generate all of the Hecke operators, and it turns out that
n
(:) = m(
1
(:) [ T
0
(m)).
A Mock Theta Function for the Delta-function 143
Remark. The
n
(:) have an elegant generating function (see [2, 11, 12]). If
polynomials J
n
(.) are dened by
1
nD0
J
n
(.)q
n
:=
1
4
(:)
2
1
6
(:)
^(:)
1
(:) .
= 1 (. 744)q . (1.5)
then we have that
n
(:) = J
n
((:)).
The
n
(:) encode the identity
(t) (:) =
1
exp
_
nD1
n
(:)
n
m
_
.
where = e
2tir
. It is equivalent, by a straightforward calculation, to the famous
denominator formula for the Monster Lie algebra
(t) (:) =
1
n>0 and n2Z

(1
n
q
n
)
c(nn)
.
We show that these polynomials also arise in the Hecke theory of certain har-
monic weak Maass forms (For more on such Maass forms and their applications,
see [3, 4, 5, 6, 7, 8, 9]). We shall recall the denition of these forms in Sec-
tion 2. The mock theta functions of Ramanujan provide non-trivial examples of
such Maass forms (for example, see [3, 5, 17]). As an example, it turns out that
q
1
(q
24
) 2i
_
3 N
(
(:) (1.6)
is a weight 1/2 harmonic weak Maass form, where
N
(
(:) :=
_
i1
24z
1
nD1
_
n
1
6
_
e
3ti(nC
1
6
)
2
r
_
i(t 24:)
Jt
is a period integral of a theta function, and (q) is Ramanujans mock theta func-
tion
(q) := 1
1
nD1
q
n
2
(1 q)
2
(1 q
2
)
2
(1 q
n
)
2
.
All of Ramanujans mock theta functions are parts of such Maass forms in this
way.
Since such period integrals of modular forms are non-holomorphic, we shall
refer to the holomorphic projection of a harmonic weak Maass form as a mock
144 Ken Ono
theta function. Using Poincar series, in Section 2.1 we dene such a mock theta
function M
Z
(:)
M
Z
(:) =
1
nD1
a
Z
(n)q
n
= 39916800q
1
2615348736000
691

= 11 q
1
24 11
T
12
73562460235.68364 . . . q
929026615019.11308 . . . q
2
8982427958440.32917 . . . q
3
71877619168847.70781 . . . q
4
.
(1.7)
Here T
12
= 691,2730 is the 12th Bernoulli number. This function enjoys the
property that
M
Z
(:) N
Z
(:)
is a weight 10 harmonic weak Maass form on SL
2
(Z), where N
Z
(:) is the
period integral of ^(:)
N
Z
(:) = (2)
11
11i
Z
_
i1
z
^(t)
(i(t :))
10
Jt. (1.8)
We shall see that
Z
~ 2.840287 . . . is the rst coefcient of a certain Poincar
series.
For positive integers n, the coefcients a
Z
(n) appear to be real numbers with-
out nice arithmetic properties. Indeed, we do not have a simple description for
any of these coefcients. However, we shall show that these coefcients satisfy
deep arithmetic properties thanks to the rationality of a
Z
(1) and a
Z
(0). In par-
ticular, we shall describe the behavior of the Hecke operators on M
Z
(:) in terms
of power series with rational coefcients, and we shall also be able to speak of
congruences.
Although M
Z
(:) is not a modular form, for every prime we shall show that
11
M
Z
(:) [ T
10
() t() M
Z
(:)
=
1
nD]
_
11
a
Z
(n) t()a
Z
(n) a
Z
(n,)
_
q
n
(1.9)
is a weight 10 weakly holomorphic modular form. It will turn out to be a mod-
ular form with integer coefcients. We explicitly describe these forms in terms of
Ramanujans tau-function and the modular polynomials
n
(:). For convenience,
if is prime, then dene the modular functions
]
(:) and T
]
(:) by
]
(:) :=
24
T
12
(1
11
)
]
(:) 264
]
nD1
o
9
(m)
]n
(:). (1.10)
T
]
(:) := t()
_
264
24
T
12
1
(:)
_
. (1.11)
Theorem 1.1. If is prime, then
1
nD]
_
11
a
Z
(n) t()a
Z
(n) a
Z
(n,)
_
q
n
=
11
1
4
(:)1
6
(:)
]
(:) T
]
(:)
_
.
Theorem 1.1 gives many congruences for the mock theta function M
Z
(:) such
as those given by the following corollary.
Corollary 1.2. The following congruences are true:
(1) If is prime, then
1
nD]
_
11
a
Z
(n) t()a
Z
(n) a
Z
(n,)
_
q
n
0 (mod 11).
(2) If is prime, then
1
11
1
nD]
_
11
a
Z
(n) t()a
Z
(n) a
Z
(n,)
_
q
n

]
(:) t()
1
(:) (mod 24).
Remark. Since
24
B
12
=
65520
691
and since the coefcients of each
n
(:) are inte-
gers, it follows that the coefcients in Corollary 1.2 (1) are in
11
691
Z. The absence
of 691 in the denominators follows from the famous Ramanujan congruence
t(n) o
11
(n) (mod 691).
It is interesting to note that this congruence and the appearance of 691 in the
formula for a
Z
(0) both arise from the fact that the numerator of T
12
is 691.
146 Ken Ono
Remark. One may simplify the right hand side of Corollary 1.2 (2) for primes
_ 5. For such primes , classical congruences of Ramanujan imply that
t() 1 (mod 24).
A famous conjecture of Lehmer asserts that t(n) = 0 for every positive in-
teger n. It is well known that the truth of this conjecture would follow from the
nonvanishing of t() for all primes . With this in mind, we obtain the following
corollary which reinterpretes this conjecture in terms of the mock theta function
M
Z
(:).
Corollary 1.3. The following are true for all primes .
(1) We have that t() = 0 if and only if
1
nD]
_
11
a
Z
(n) a
Z
(n,)
_
q
n
is a weight 10 modular form on SL
2
(Z).
(2) We have that t() = 0 if and only if
1
nD]
_
11
a
Z
(n) a
Z
(n,)
_
q
n
=
11
1
4
(:)1
6
(:)

]
(:).
(3) If t() = 0, then for every positive integer n coprime to we have that
11
a
Z
(n) is an integer for which
11
a
Z
(n) 0 (mod 11).
Based on numerical experiments, we make the following conjecture which
implies Lehmers Conjecture.
Conjecture. The coefcients a
Z
(n) are irrational for every positive integer n.
Remark. After this paper was written, Bruinier, Rhoades and the author [10]
proved general theorems relating the algebraicity of coefcients of certain inte-
ger weight harmonic weak Maass forms to the vanishing Hecke eigenvalues.
Remark. This conjecture is closely related to results obtained by the author and
Bruinier on Heegner divisors and derivatives of 1-functions (see [9]). In that work
results were obtained on the transcendence on the coefcients of harmonic weak
Maass forms. These results, combined with a well-known conjecture of Goldfeld
on modular 1-functions, implies that almost all coefcients of certain weight
1,2 harmonic weak Maass forms are transcendental, and hence irrational. This
work provides further evidence supporting the plausibility of our conjecture.
In Section 2.1 we construct the mock theta function M
Z
(:) using results in
earlier joint work with Bringmann [4]. In particular, we give complicated exact
formulas for the coefcients of M
Z
(:), and we give a description of the real
number
Z
in (1.8). In Section 3 we prove Theorem 1.1 and Corollaries 1.2 and
1.3. In the last section we conclude with a detailed discussion of Theorem 1.1 and
Corollary 1.2 when = 2.
2 Harmonic Weak Maass Forms and M
.z/
We recall the notion of a harmonic weak Maass form of integer weight k on
SL
2
(Z) due to Bruinier and Funke [8]. If : = . i, H with .. , R,
then the weight k hyperbolic Laplacian is given by
^
k
:= ,
2
_
d
2
d.
2

d
2
d,
2
_
i k,
_
d
d.
i
d
d,
_
. (2.1)
A weight k harmonic weak Maass form on SL
2
(Z) is any smooth function M :
H C satisfying the following:
(1) For all =
_
o b
c d
_
SL
2
(Z) and all : H, we have
M(:) = (c: J)
k
M(:).
(2) We have that ^
k
M = 0.
(3) The function M(:) has at most linear exponential growth at the cusp innity.
2.1 Poincar Series and the Denition of M
.z/
Here we recall two relevant Poincar series. The Poincar series H(:) will turn
out to equal
Z
^(:), and the second Poincar series 1(:) shall be of the form
1(:) = M
Z
(:) N
Z
(:).
Here we recall the constructions of these Poincar series following [4]. We
rely on classical special functions whose properties and denitions may be found
in [1]. Suppose that k Z. For =
_
o b
c d
_
SL
2
(Z), dene (. :) by
(. :) := (c: J). (2.2)
As usual, for such and functions : H C, we let
( [
k
)(:) := (. :)
k
(:). (2.3)
148 Ken Ono
Let m be an integer, and let
n
: R
C
C be a function which satises
n
(,) = O(,
), as , 0, for some R. If e() := e

2ti
as usual, then let
n
(:) :=
n
(,)e(m.). (2.4)
Such functions are xed by the translations I
1
:= {
_
1 n
0 1
_
: n Z}. Given
this data, we dene the Poincar series
1(m. k.
n
: :) :=
2I
1
nSL
2
(Z)
(
n
[
k
)(:). (2.5)
Now dene the function H(:) by
H(:) := 1(1. 12. e(i,): :) =
1
nD1
b(n)q
n
. (2.6)
It is well known that H(:) is a weight 12 cusp form on SL
2
(Z), and so it follows
that
H(:) =
Z
^(:).
where
Z
= b(1) ~ 2.840287 . . . . It is well known that (for example, see Chap-
ter 3 of [15]) that
Z
may be described in terms of the Petersson norm of H(:).
Now we construct the relevant weight 10 harmonic weak Maass form. Let
M
,
(:) be the usual M-Whittaker function. For complex s, let
M
x
(,) := [,[
k
2
Mk
2
sgn(,), x
1
2
([,[).
and for positive m let
n
(:) := M
1
k
2
(4m,). We now let
1(:) := 1(1. 10.
1
: :). (2.7)
Now recall the period integral N
Z
(:) from (1.8), and dene M
Z
(:) by
M
Z
(:) := 1(:) N
Z
(:). (2.8)
We shall give an exact formula for the Fourier expansion of M
Z
(:) in terms of
the classical Kloosterman sums
1(m. n. c) :=
(c)
e
_
m n
c
_
. (2.9)
In the sums above, runs through the primitive residue classes modulo c, and
denotes the multiplicative inverse of modulo c. Theorem 1.1 of [4] (also
see [13, 14] for related earlier works) then implies the following result relating
H(:). 1(:) and M
Z
(:).
Theorem 2.1. The following are true:
(1) The function 1(:) is a weight 10 harmonic weak Maass form on SL
2
(Z).
(2) The function M
Z
(:) is holomorphic on H, and it has a Fourier expansion of
the form
M
Z
(:) =
1
nD1
a
Z
(n)q
n
= I(12)q
1
2
12
12
(12)
.
where for positive integers n we have
a
Z
(n) = 2I(12)n
11
2
cD1
1(1. n. c)
c
1
11
_
4
_
n
c
_
.
Here 1
11
(.) is the usual 1
11
-Bessel function.
Remark. We note that M
Z
(:) begins with the terms
M
Z
(:) = 39916800q
1
2615348736000
691
.
Obviously, we have that a
Z
(1) = I(12) = 11. Using classical formulas for
Riemanns zeta function at positive even integers, we nd that a
Z
(0) =
2411
B
12
.
Strictly speaking, the constant term for M
Z
(:) in [4] is given in terms of the
innite sum
1
cD1
1(1. 0. c)
c
12
.
A straightforward calculation relates these Kloosterman sums to j(c), the classi-
cal Mbius function, and consequently provides the connection to the reciprocal
of (12).
Remark. It would be very interesting to have a simpler description of the coef-
cients a
Z
(n) for positive n.
Remark. It turns out that there are innitely many choices for the mock theta
function M
Z
(:). This ambiguity arises for two reasons. First of all, one may
choose other descriptions of ^(:) in terms of Poincar series to obtain other mock
theta functions. We simply selected the simplest choice. Secondly, we point out
that if J(:) is a weakly holomorphic modular form of weight 10 on SL
2
(Z)
with algebraic coefcients, then the sum
M
Z
(:) N
Z
(:) J(:)
is also a weight 10 harmonic weak Maass form, and so M
Z
(:) J(:) is also a
mock theta function for ^(:). In both situations one can easily modify the argu-
ments here to obtain the corresponding versions of Theorem1.1 and Corollary 1.2.
150 Ken Ono
3 Proofs
Here we prove Theorem 1.1 and Corollary 1.2. The proof of Theorem 1.1 re-
lies on facts about harmonic weak Maass forms and their behavior under Hecke
operators. The proofs of Corollaries 1.2 and 1.3 are elementary.
3.1 Proof of Theorem 1.1
The Hecke operators T
10
() act independently on the holomorphic and non-
holomorphic parts of 1(:). Since ^(:) is a Hecke eigenform, it follows that
N
Z
(:) is an eigenform of T
10
() (for example, see Section 7 of [9]). In particu-
lar, since
^(:) [ T
12
() = t()^(:).
it then follows that
1
]
(:) := 1(:) [ T
10
() t()
11
1(:)
is holomorphic on H, and so it is a weight 10 weakly holomorphic modular form
on SL
2
(Z). Recall that a weakly holomorphic modular form is a meromorphic
modular form whose poles (if any) are supported at cusps. Obviously, we then
have that
1
]
(:) = M
Z
(:) [ T
10
() t()
11
M
Z
(:). (3.1)
Consequently, it follows that 1
4
(:)1
6
(:)1
]
(:) is a weakly holomorphic modular
form of weight 0. All such forms are polynomials in (:).
Now we turn to the problem of determining the exact formula for the modular
form 1
4
(:)1
6
(:)1
]
(:). We rst note that
M
Z
(:) [ T
10
() = 11
11
q
]
a
Z
(0)(1
11
) O(q).
where a
Z
(0) =
2411
B
12
as before. Therefore, since
1
4
(:)1
6
(:) = 1
10
(:) = 1 264
1
nD1
o
9
(n)q
n
.
it follows that
1
4
(:)1
6
(:) (M
Z
(:) [ T
10
())
= 11
11
q
]
264 11
11
]1
nD1
o
9
(m)q
]Cn
264 11
11
(1
9
) a
Z
(0)(1
11
) O(q).
(3.2)
Similarly, we have that
t()
11
1
4
(:)1
6
(:)M
Z
(:)
= t()
11
_
11 q
1
264 11 a
Z
(0)
_
O(q).
Combining these observations, we nd that
1
4
(:)1
6
(:)1
]
(:) = 11
11
q
]
264 11
11
]1
nD1
o
9
(m)q
]Cn
264 11
11
(1
9
) a
Z
(0)(1
11
)
t()
11
_
11 q
1
264 11 a
Z
(0)
_
O(q)
= 11
11
_
q
]
264
]1
nD1
o
9
(m)q
]Cn
264(1
9
)
a
Z
(0)
11
11
(1
11
)
_
t() 11
11
_
q
1
264
a
Z
(0)
11
_
O(q).
Since every polynomial in (:) is uniquely determined by its principal part (i.e.,
the coefcients corresponding to non-positive exponents), it follows that
1
4
(:)1
6
(:)1
]
(:) =
11
11

_
]
(:) T
]
(:)
_
.
Here we use the fact that each
n
(:) satises
n
(:) = q
n
O(q).
This completes the proof.
3.2 Proof of Corollary 1.2
Here we prove the two claimed congruences.
(1) With exception of the constant terms, by (1.10) and (1.11) we have that all of
the coefcients of
]
(:) and T
]
(:) are integers. Since
24
B
12
=
65520
691
, and since
1
4
(:)1
6
(:) has integer coefcients, the rst claimfollows immediately fromThe-
orem 1.1 and the Ramanujan congruence
t(n) o
11
(n) (mod 691).
152 Ken Ono
(2) By Theorem 1.1, we have that
1
11
1
nD]
_
11
a
Z
(n) t()a
Z
(n) a
Z
(n,)
_
q
n
=
1
1
4
(:)1
6
(:)
(
]
(:) T
]
(:)).
Moreover, its coefcients are in Zby part (1). Therefore, we may reduce this mod-
ulo 24. Since 1
4
(:)1
6
(:) 1 (mod 24), by (1.10) and (1.11) we then nd that
1
11
1
nD]
(a
Z
(n) t()a
Z
(n) a
Z
(n,)) q
n

]
(:) T
]
(:) (mod 24)

]
(:) t()
1
(:) (mod 24).
This gives the second claim.
3.3 Proof of Corollary 1.3
Theorem 1.1, combined with the fact that M
Z
(:) is not modular, gives the rst
claim. Theorem 1.1 and (1.11) immediately implies the second claim. The third
claim now follows from Corollary 1.2 (1).
4 Examples
Here we give a detailed discussion of Theorem 1.1 and Corollary 1.2 when = 2.
In particular, we use the exact formulas for the coefcients of M
Z
(:) to nu-
merically approximate the theoretical exact formulas for M
Z
(:) [ T
10
(2) and
compare them with the exact formulas we derive using just the rst two coef-
cients of M
Z
(:).
We begin by giving the numerical approximations for the rst few terms of
M
Z
(:). Using Theorem 2.1, we nd that
M
Z
(:) =
1
nD1
a
Z
(n)q
n
= 39916800q
1
2615348736000
691
73562460235.68364 . . . q
929026615019.11308 . . . q
2
8982427958440.32917 . . . q
3
71877619168847.70781 . . . q
4
497966668914961.54321 . . . q
5
3074946857439412.02739 . . . q
6
.
Again we stress that a
Z
(1) and a
Z
(0) are exact.
4.1 Theorem 1.1 when p D 2
Using Theorem 2.1, one numerically nds that
M
Z
(:) [ T
10
(2) t(2)2
11
M
Z
(:)
=
155925
8
q
2
467775q
1
3831077250
929888675100.00000 . . . q
71888542118662.49999 . . . q
2
3075052120267049.99999 . . . q
3
.
(4.1)
The rst three terms are exact.
The modular functions
2
(:) and T
2
(:) are:
2
(:) =
2
(:) 264
1
(:)
227833992
691
.
T
2
(:) = 24
1
(:)
5950656
691
.
Therefore, Theorem 1.1 implies that
M
Z
(:) [ T
10
(2) t(2)2
11
M
Z
(:)
=
11
2
11
1
4
(:)1
6
(:)
(
2
(:) T
2
(:))
=
155925(
2
(:) 240
1
(:) 338328)
81
4
(:)1
6
(:)
=
155925
8
q
2
467775q
1
3831077250 929888675100q
143777084237325
2
q
2
3075052120267050q
3
.
This agrees with the numerics in (4.1).
4.2 Corollary 1.2 when p D 2
Using the calculations from the previous subsection, we have that
M
Z
(:) [ T
10
(2) t(2)2
11
M
Z
(:) =
155925(
2
(:) 240
1
(:) 338328)
81
4
(:)1
6
(:)
.
The coefcients are easily seen to lie in
1
8
Z. This illustrates Corollary 1.2 (1).
154 Ken Ono
To illustrate Corollary 1.2 (2), simply multiply both sides of this expression by
2
11
,11 to obtain
2
11
_
M
Z
(:) [ t(2)2
11
M
Z
(:)
_
11
=
2
(:) 240
1
(:) 338328
1
4
(:)1
6
(:)

2
(:) t(2)
1
(:)

2
(:) (mod 24).
Here we used t(2) = 24 and 1
4
(:)1
6
(:) 1 (mod 24).
Acknowledgments. The author thanks Matt Boylan, Kathrin Bringmann,
Amanda Folsom, and Rob Rhoades for their comments on early drafts of this
paper. The author also thanks the National Science Foundation for its generous
support, he thanks the support of the Manasse family, and the support of the Hill-
dale Foundation.
References
[1] G. E. Andrews, R. Askey and R. Roy, Special Functions, Cambridge Univ. Press,
Cambridge, 1999.
[2] T. Asai, M. Kaneko and H. Ninomiya, Zeros of certain modular functions and an
application, Comm. Math. Univ. Sancti Pauli 46 (1997), 93101.
[3] K. Bringmann and K. Ono, The (q) mock theta function conjecture and partition
ranks, Invent. Math. 165 (2006), 243266.
[4] K. Bringmann and K. Ono, Lifting cusp forms to Maass forms with an application
to partitions, Proc. Natl. Acad. Sci. USA 104 (2007), no. 10, 37253731.
[5] K. Bringmann and K. Ono, Dysons ranks and Maass forms, to appear in Ann. of
Math.
[6] K. Bringmann, K. Ono and R. Rhoades, Eulerian series as modular forms, J. Amer.
Math. Soc. 21 (2008), 10851104.
[7] J. H. Bruinier, Borcherds products on O(2. l) and Chern classes of Heegner divisors,
Springer Lecture Notes in Mathematics 1780, Springer-Verlag, 2002.
[8] J. H. Bruinier and J. Funke, On two geometric theta lifts, Duke Math. J. 125 (2004),
4590.
[9] J. H. Bruinier and K. Ono, Heegner divisors, 1-functions, and harmonic weak Maass
forms, to appear in Ann. Math.
[10] J. H. Bruinier, K. Ono and R. Rhoades, Differential operators for harmonic weak
Maass forms and the vanishing of Hecke eigenvalues, Math. Ann. 342 (2008), 673
693.
[11] G. Faber, ber polynomische Entwicklungen, Math. Ann. 57 (1903), 389408.
[12] G. Faber, ber polynomische Entwicklungen. II, Math. Ann. 64 (1907), 116135.
[13] J. D. Fay, Fourier coefcients of the resolvent for a Fuchsian group, J. Reine Angew.
Math. 293/294 (1977), 143203.
[14] D. A. Hejhal, The Selberg Trace Formula for PSL(2. 1). Vol. 2, Springer Lect.
Notes in Math. 1001, Springer-Verlag, Berlin, 1983.
[15] H. Iwaniec, Topics in the Classical Theory of Automorphic Forms, Grad. Studies in
Math. 17, Amer. Math. Soc., Providence, R.I., 1997.
[16] K. Ono, The Web of Modularity: Arithmetic of the Coefcients of Modular Forms
and q-Series, Conference Board of the Mathematical Sciences 102, Amer. Math.
Soc., 2004.
[17] S. P. Zwegers, Mock Theta Functions, Ph.D. Thesis, Universiteit Utrecht, 2002.
Author information
Ken Ono, Department of Mathematics, University of Wisconsin,
Madison, Wisconsin 53706, USA.
E-mail: ono@math.wisc.edu
On the Density of Integral Sets with
Missing Differences
Ram Krishna Pandey and Amitabha Tripathi
Abstract. For a given set M of positive integers, a well-known problem of Motzkin
asks for determining the maximal density j(M) among sets of nonnegative integers in
which no two elements differ by M. The problem is completely settled when [M[ _ 2,
and some partial results are known for several families of M for [M[ _ 3. In this paper,
we consider the case M = {a. b. c}, with c a multiple of a or b. In most cases, we obtain
lower bounds for j(M), which are conjecturally the exact values of j(M), while in some
we obtain the exact value of j(M).
Keywords. Density, M-set.
AMS classication. 11B05.
1 Introduction
For . R and a set S of nonnegative integers, let S(.) denote the number of
elements n S such that n _ .. The upper and lower densities of S, denoted by
(S) and (S) respectively, are given by
(S) := limsup
x!1
S(.)
.
. (S) := liminf
x!1
S(.)
.
.
If (S) = (S), we denote the common value by (S), and say that S has density
(S). Given a set of positive integers M, S is said to be an M-set if a S, b S
imply a b M. Motzkin in [3] asked to determine j(M) given by
j(M) := sup
S
(S)
where S varies over the class of all M-sets. Cantor & Gordon in [1] showed the
existence of j(M) for any M, determined j(M) when [M[ _ 2:
j({m
1
}) =
1
2
.
j({m
1
. m
2
}) =
](m
1
m
2
),2
m
1
m
2
for gcd(m
1
. m
2
) = 1.
158 Ram Krishna Pandey and Amitabha Tripathi
and gave the following lower bound for j(M):
j(M) _ sup
gcd.k;m/D1
1
m
min
i
[km
i
[
m
.
where m
i
are the elements of M and [.[
m
denotes the absolute value of the abso-
lutely least remainder of . mod m. Haralambis in [2] gave the equivalent expres-
sions for the right-hand side expression of the above inequality:
J
1
(M) = sup
x2.0;1/
min
i
[.m
i
[.
J
2
(M) = sup
gcd.k;m/D1
1
m
min
i
[km
i
[
m
.
J
3
(M) = max
mDm
j
Cm
`
1k
m
2
1
m
min
i
[km
i
[
m
where [.[ denotes the distance fromthe nearest integer. Thus J
1
(M) =J
2
(M) =
J
3
(M), and we denote this common value by J(M). Hence J(M) serves as a
lower bound for j(M). A useful upper bound for j(M) is due to Haralambis
in [2]:
j(M) _ provided there exists a positive integer k such that S(k) _ (k 1)
for every M-set S with 0 S.
In fact, Haralambis in [2] conjectured that j(M) = J(M) for [M[ = 3, so that,
conjecturally, determining J(M) = J
3
(M) gives the value of j(M).
We consider the problem for the families M = {a. b. c}, where c is a multiple
of a or b. By a result of Cantor & Gordon in [1], we know that j(kM) = j(M).
Thus, it is no loss of generality to assume that gcd(a. b) = 1, and that a < b. In
most cases, we determine the value of J(M), which is the lower bound for j(M)
and conjecturally equal to it, and in some cases we determine the value of j(M).
2 Exact Results
We begin by dealing with one special case where we determine j(M). We use
the upper and lower bounds for j(M) to achieve this.
Theorem 1. Let M = {a. b. c}, where a b is odd, gcd(a. b) = 1, c {na. nb},
and n 1(mod a b). Then
j(M) =
a b 1
2(a b)
.
On the Density of Integral Sets with Missing Differences 159
Proof. By Cantor & Gordons result we have j(M) _ j({a. b}) =
aCb1
2.aCb/
.
For the reverse inequality, choose . such that a.
aCb1
2
(mod a b). Since
c. a. b.(mod a b), j(M) _
aCb1
2.aCb/
. Hence the result.
3 The Case M D a; b; nb
In this section, we deal with the family M = {a. b. nb}, with a < b, gcd(a. b) =
1, and n _ 2. We compute J
3
(M) by comparing the rational numbers in the three
cases, as mentioned in Section 1.
Theorem 2. Let M = {a. b. nb}, where a < b, gcd(a. b) = 1, and a, b are odd
integers. Then
j(M) =
_
_
_
n
2.nC1/
if n is even:
1
2
if n is odd.
Proof. By Cantor and Gordons result we have j(M) _ j({a. b}) =
1
2
. If n is
odd, then {1. 3. 5. . . . } is an M-set. Hence j(M) =
1
2
in this case. If n is even,
then j(M) _ j({b. nb}) = j({1. n}) =
n
2.nC1/
. To show the reverse inequality,
let m = (n 1)b. Observe that m is odd. Choose .
m1
2
(mod m). Then
a.
ma
2
(mod m). nb. b.
mb
2
(mod m).
Since
1
2
(m a) >
1
2
(m b) =
1
2
nb, we have j(M) _ J(M) _
nb
2m
=
n
2.nC1/
.
Lemma 1. For r. s _ 0, let
r
:= {2r(a b) 2t 1 : 1 _ t _ b}.
T
s
:= {2s(a b) 2b 2t 1 : 1 _ t _ a}.
Then {
0
.
1
. . . . . T
0
. T
1
. . . . } partitions the set of positive odd integers 2N 1.
Proof. Observe that [
r
[ = b and [T
s
[ = a for each r. s _ 0, and that
rC1
=
r
2(a b) and T
sC1
= T
s
2(a b). The lemma now follows from
the observation that {
0
. T
0
} partitions the odd integers in the interval 1. 2(a
b) 1|.
Theorem 3. Let M = {a. b. nb}, where a < b, gcd(a. b) = 1, a b and n
are odd integers. Let the family of sets {
r
}
r0
and {T
s
}
s0
be dened as in
Lemma 1. If n , 1(mod a b), then
J(M) =
_
_
_
m.2rbC2t1/
2m
if n
r
and where m = a nb:
m2.sC1/b
2m
if n T
s
and where m = (n 1)b.
Proof. We compute J(M) by using the expression for J
3
(M) in Section 1. Thus
there are three choices for m, and we determine J
3
(M) by comparing the three
rational numbers corresponding to these case. By Lemma 1, n belongs to a unique
set among the two families {
r
}
r0
and {T
s
}
s0
.
CASE I: (m = a nb) Observe that m is odd, and that gcd(b. m) = 1.
Subcase (i): (n
r
) Choose . such that
b.
m.2rbC1/
2
(mod m).
Thus we have 2a. 2nb. n(2rb 1) = 2r(m a) n n 2ra =
2rb 2t 1(mod m), and
a.
m.2rbC2t1/
2
(mod m).
Since nb. a.(mod m),
min
_
[a.[
m
. [b.[
m
. [nb.[
m
_
=
m.2rbC2t1/
2
. (1)
We now show that
min
_
[a,[
m
. [b,[
m
. [nb,[
m
_
_
m.2rbC2t1/
2
=
m1
2
.
for each ,, 1 _ , _
1
2
(m1), where = rbt 1. Let I :=
_
m1
2
.
mC1
2

_
and J :=
_
m1
2
.
mC1
2

_
. We show that, for 1 _ , _
1
2
(m 1), if
b, (mod m) I, then a, (mod m) J. Accordingly, write
b,
m1
2
i (mod m).
Then b, (mod m) I if and only if 0 _ i _ 2 1, and 2a, 2nb,
n
_
1 2( i )
_
(mod m). Since n
_
1 2( i )
_
is odd, we get
a,
m1
2

nC1
2
n( i ) (mod m).
To show that a, (mod m) J, we consider the two cases 0 _ i _ and 1 _
i _ 2 1.
First consider the case 0 _ i _ . For each k, 0 _ k _ r, dene
1
k
:=
_

1
n
_
(k 1)m
nC1
2
_
.
1
n
_
km
nC1
2
_
1
_
.
J := { kb : 1 _ k _ r} L{}.
Then it can be shown that 1
0
L 1
1
L . . . L1
r
L J contains the set {0. 1. 2. . . . . }.
A simple computation shows that
a,
_
_
_
mC1
2
. m
_
if i = :
_
0.
m1
2

_
if i J. i = :
_
km
mC1
2
. (k 1)m
m1
2

_
if i 1
k
with 0 _ k _ r.
For the cases 1 _ i _ 2 1, dene for 0 _ k _ r,
1
0
k
:=
_

1
n
_
km
n1
2
_
1.
1
n
_
(k 1)m
n1
2
__
.
J
0
:= { 1 kb : 1 _ k _ r}.
Then it can be shown that 1
0
0
L 1
0
1
L . . . L 1
0
r
L J
0
contains the set { 1.
2. . . . . 2 1}, and that a, (mod m) J. This completes the subcase when
n
r
.
Subcase (ii): (n T
s
b.
m
_
2.sC1/bC1
_
2
(mod m).
The computation in subcase (i) can be employed to show that
a.
m
_
2.sC1/b2.at/1
_
2
(mod m).
so that
min
_
[a.[
m
. [b.[
m
. [nb.[
m
_
=
m
_
2.sC1/bC1
_
2
. (2)
To show that
min
_
[a,[
m
. [b,[
m
. [nb,[
m
_
_
m
_
2.sC1/bC1
_
2
=
m1
2
.
for each ,, 1 _ , _
1
2
(m1), where = (s 1)b, we mimic the proof in subcase
(i), the only change being in the value of . All notations and congruences carry
through, so the derivations are omitted. This completes the subcase when n T
s
,
and Case I.
CASE II: (m = (n 1)b) Observe that m is even. As in Case I, we consider two
subcases.
Subcase (i): (n
r
) For each ., 1 _ . _
1
2
m, we show that
min{[a.[
m
. [b.[
m
. [nb.[
m
}_
m
2
(rb t ). (3)
Suppose [b.[
m
>
m
2
(rb t ). Then
zm
m
2
rb t < b. < zm
m
2
rb t
for some integer z. Thus .
_
z(n 1)
nC1
2
r. z(n 1)
nC1
2
r
_
. Write
.
nC1
2
r i (mod n 1), with 0 _ i _ 2r. Since
nC1
2
= r(a b) t , it is
easy to verify that
a. ar ai
nC1
2
br t ai (mod n 1) when b is odd
and
a. br t ai (mod n 1) when b is even.
Now 2ar = (n 1) 2(br t ). Hence 0 _ ai _ (n 1) 2(rb t ) for
0 _ i _ 2r, and [a.[
m
_
m
2
(br t ), as desired. This completes the subcase
when n
r
.
Subcase (ii): (n T
s
) As in subcase (i), we can choose an integer z such that
with . = z(n 1)
nC1
2
(s 1), we have
b.
m
2
(s 1)b (mod m). a.
m
2
(s 1)b (t a) (mod m).
Hence
min{[a.[
m
. [b.[
m
. [nb.[
m
}=
m
2
(s 1)b. (4)
Again, an argument similar to the one in subcase (i) shows that
min{[a,[
m
. [b,[
m
. [nb,[
m
}_
m
2
(s 1)b
for each ,, 1 _ , _
1
2
m. This completes the argument in Case II.
CASE III: (m = a b) Observe that m is odd, and that gcd(a. m) = 1 =
gcd(b. m). Choose . such that a. b.
aCb1
2
(mod m). Since n is odd,
it is easy to see that
nb.
mn
2

m1
2
(mod m)
if and only if n 1(mod 2m). Since
m1
2
is the maximum absolute remainder
mod m and since n 1(mod 2(a b)) is excluded by assumption,
min{[a.[
m
. [b.[
m
. [nb.[
m
}_
aCb3
2
for each ., 1 _ . _
1
2
(m 1).
To determine J(M), we consider the two cases n
r
and n T
s
separately,
and compare the values given by the three cases. For n
r
, n , 1(mod 2(a
b)), observe that
.nC1/b2.rbCt/
2.nC1/b
=
1
2

rbCt
.nC1/b
<
1
2

2.rbCt/1
2.aCnb/
=
.aCnb/.2rbC2t1/
2.aCnb/
.
and
aCb3
2.aCb/
=
1
2

3
2.aCb/
<
1
2

2.rbCt/1
2.aCnb/
=
.aCnb/.2rbC2t1/
2.aCnb/
.
Thus the upper bounds for J(M) in Cases II and III are each less than the value
of J(M) in Case I. Hence, in this case
J(M) =
1
2

2.rbCt/1
2.aCnb/
=
.aCnb/.2rbC2t1/
2.aCnb/
.
For n T
s
, n , 1(mod 2(a b)), we have
.aCnb/.2.sC1/bC1/
2.aCnb/
=
1
2

.sC1/bC
1
2
aCnb
<
1
2

sC1
nC1
=
.nC1/2.sC1/
2.nC1/
.
and
aCb3
2.aCb/
=
1
2

3
2.aCb/
<
1
2

sC1
nC1
=
.nC1/2.sC1/
2.nC1/
.
Thus the upper bounds for J(M) in Cases I and III are each less than the value of
J(M) in Case II, and
J(m) =
1
2

sC1
nC1
=
.nC1/2.sC1/
2.nC1/
in this case. This completes the comparison, and the proof of the theorem.
0
r
:= {(2r 1)(a b) 2t 1 : 1 _ t _ b}.
T
0
s
:= {(2s 1)(a b) 2b 2t 1 : 1 _ t _ a}.
Then {
0
0
.
0
1
. . . . . T
0
0
. T
0
1
. . . . } partitions the set b a (2N 1).
Proof. Observe that
0
r
=
r
(ab) and T
0
s
= T
s
(ab) for each r. s _ 0.
The proof is similar to that of Lemma 1. We have [
0
r
[ = b, [T
0
s
[ = a for each
r. s _ 0,
0
rC1
=
0
r
2(a b) and T
0
sC1
= T
0
s
2(a b). The lemma
now follows from the observation that {
0
0
. T
0
0
} partitions the even integers in the
interval b a 1. 3b a 1|.
Theorem 4. Let M = {a. b. nb} where a < b, gcd(a. b) = 1, a b is odd,
n _ b a 1 and even. Let the family of sets {
0
r
}
r0
and {T
0
s
}
s0
be as dened
in Lemma 2. If n , 1(mod 2(a b)), then
J(M) =
_
_
_
m.2rC1/bC2t1
2m
where m = a nb and n
0
r
:
m.2sC1/b
2m
where m = (n 1)b and n T
0
s
.
Proof. We use the method of proof given in Theorem 3, and place every even
integer n _ b a 1 in a unique set among the two families {
0
r
}
r0
and
{T
0
s
}
s0
.
CASE I: (m = a nb) Observe that gcd(b. m) = 1.
Subcase (i): (n
0
r
b.
m.2rC1/bC1
2
(mod m).
This is an analogue of the corresponding subcase in Theorem 3 with 2r 1 replac-
ing 2r. The argument of this subcase carries through if we make this replacement
throughout this subcase. We omit the details. This completes the subcase when
n
0
r
.
Subcase (ii): (n T
0
s
b.
m.2sC1/bC1
2
(mod m).
This is an analogue of the corresponding subcase in Theorem 3 with 2s 1 re-
placing 2(s 1). The argument of this subcase carries through if we make this
replacement throughout this subcase. We omit the details. This completes the
subcase when n T
0
s
, and Case I.
CASE II: (m = a nb)
Subcase (i): (n
0
r
) This is an analogue of the corresponding subcase in The-
orem 3 with 2r 1 replacing 2r, obtaining only an upper bound. We omit the
details.
Subcase (ii): (n T
0
s
) This is an analogue of the corresponding subcase in Theo-
rem 3 with 2s 1 replacing 2(s 1). We again omit the details.
CASE III: (m = a b)
We may use the exact same computation of Case III in Theorem 3 for this case
as well. The rest of the proof is the same as that given in Theorem 3, and is
omitted.
Theorem 5. Let M = {a. b. nb}. where a < b, gcd(a. b) = 1, a b is odd,
n _ b a 1, and n is even. Then j(M) =
n
2.nC1/
.
Proof. By Cantor and Gordons result we have j(M) _ j({b. nb}) = j({1. n})
=
n
2.nC1/
. For the reverse inequality, let m = (n 1)b and choose . such that
.
n
2
(mod n 1). Then . = z(n 1)
n
2
for some integer z, and a simple
calculation shows that
a.
ma.nC1/
2
(mod m) (2z 1)a 1 (mod b).
If b is even, then a must be odd, and so any solution of a, 1(mod b) is
necessarily odd. If b is odd, then a must be even, and we may choose , to be odd
by replacing , by b ,, if necessary. In either case, we may choose z such that
a(2z 1) 1(mod b), and hence satisfy a.
ma.nC1/
2
(mod m). Since
b.
nb
2
=
mb
2
(mod m) and b _ an1, it follows that j(M) _
nb
2.nC1/b
=
n
2.nC1/
. This completes the proof.
4 The Case M D a; b; na
We deal with the family M = {a. b. nb}, with a < b, gcd(a. b) = 1, and n _ 2.
The results are analogous to those obtained in Section 3, and proofs similar. We
begin by considering the case where a. b are both odd. However, unlike the anal-
ogous case in Theorem 2, we are able to determine J(M) only for all sufciently
large n.
Theorem 6. Let M = {a. b. na}, where a < b, gcd(a. b) = 1, a, b are odd
integers, and n _
b.aCb2/
2a
and even. Then J(M) =
na
2.naCb/
.
Proof. We compute J(M) by using the expression for J
3
(M) in Section 1.
CASE I: (m = na b) Observe that m is odd. Choose . such that .
m1
2
(mod m). Then
a.
ma
2
(mod m) b.
mb
2
(mod m).
Thus
min{[a.[
m
. [b.[
m
. [na.[
m
}=
mb
2
=
na
2
. (5)
We now show that
min{[a,[
m
. [b,[
m
. [na,[
m
}_
mb
2
for each ,, 1 _ , _
1
2
(m 1), by an argument similar to the one in Theorem 2.
Let I :=
_
mb
2
.
mCb
2
_
. We show that, for 1 _ , _
1
2
(m 1), b, (mod m) I
and a, (mod m) I only when ,
m1
2
(mod m). With ,
m1
2
i (mod m),
a simple calculation shows that
b, (mod m) I i
_
k
m
b
. k
m
b
1
_
for some integer k, with 0 _ k _ b 1.
If k = 0, i = 0 gives ,
m1
2
.(mod m) while i = 1 gives ,
m1
2

.(mod m). For 1 _ k _ b 1, let ka = qb r, where 0 _ r _ b 1. In
fact, r = 0 since b [ ka otherwise, and this is impossible since gcd(a. b) = 1
and 1 _ k _ b 1. A routine calculation shows that a, (mod m) lies between
ma
2

mr
b
(mod m) and
mCa
2

mr
b
(mod m), and another shows
mCb
2
_
ma
2

mr
b
<
mCa
2

mr
b
_ m
mb
2
.
the rst and third inequalities being valid since 2na _ b(a b 2). This proves
our claim that a, (mod m) I for 1 _ , <
m1
2
, and completes the proof in this
case.
CASE II: (m = (n 1)a) As in Case I, m is odd. The same choice of . gives
min{[a.[
m
. [b.[
m
. [na.[
m
}=
mb
2
=
.nC1/ab
2
. (6)
The proof of
min{[a,[
m
. [b,[
m
. [na,[
m
}_
mb
2
for each ,, 1 _ , _
1
2
(m 1) is similar to the one in Case I, and omitted.
CASE III: (m = a b) In this case m is even, and gcd(a. m) = gcd(b. m) = 1.
Observe that
a. b.
m
2
(mod m)
implies na. 0(mod m) since n is even. Therefore
min{[a.[
m
. [b.[
m
. [na.[
m
}_
m
2
1
for each ., 1 _ . _
1
2
m.
It is easy to check that
na
2.bCna/
>
.nC1/ab
2.nC1/a
, and that
na
2.bCna/
_
aCb2
2.aCb/
if and
only if n _
b.aCb2/
2a
. Hence the result.
C
r
:= {2r(a b) 2t 1 : 1 _ t _ a}.
D
s
:= {2s(a b) 2a 2t 1 : 1 _ t _ b}.
Then {C
1
. C
2
. . . . . D
1
. D
2
. . . . } partitions the set 2(a b) (2N 1).
Proof. Observe that the families {C
r
}
r0
and {D
s
}
s0
are obtained from the fam-
ilies {
r
}
r0
and {T
s
}
s0
by interchanging a and b. Observe also that [C
r
[ = a
and [D
s
[ = b for each r. s _ 0, and that C
rC1
= C
r
2(a b) and D
sC1
=
D
s
2(a b). The lemma now follows from the observation that {C
1
. D
1
}
partitions the odd integers in the interval 2(a b) 1. 4(a b) 1|.
Theorem 7. Let M = {a. b. na} where a < b, gcd(a. b) = 1, a b is odd,
n _ 2(a b) 1 and odd. Let the family of sets {C
r
}
r1
and {D
s
}
s1
be dened
as in Lemma 3. If n , 1(mod a b), then
J(M) =
_
_
m.2raC2t1/
2m
if n C
r
and where m = na b:
m2.sC1/a
2m
if n T
s
, b _ 2(s 1)a t ,
and where m = (n 1)a.
Proof. The argument in Theorem 3 carries over with the roles of a and b inter-
changed. We omit the proof.
Remark 1. We remark that just as the families {C
r
}
r1
and {D
s
}
s1
are obtained
from the families {
r
}
r0
and {T
s
}
s0
by interchanging a and b, so are the cor-
responding results from Theorem 3. However, these formulae do not hold for
n C
0
L D
0
.
C
0
r
:= {(2r 1)(a b) 2t 1 : 1 _ t _ a}.
D
0
s
:= {(2s 1)(a b) 2a 2t 1 : 1 _ t _ b}.
Then {C
0
0
. C
0
1
. . . . . D
0
0
. D
0
1
. . . . } partitions the set (a b) (2N 1).
Proof. Observe that C
0
r
= C
r
(ab) and D
0
s
= D
s
(ab) for each r. s _ 0.
The lemma now follows from Lemma 3.
Theorem 8. Let M = {a. b. na} where a < b, gcd(a. b) = 1, a b is odd,
n _ a b 1 and even. Let the family of sets {C
0
r
}
r0
and {D
0
s
}
s0
be dened
as in Lemma 4. If n , 1(mod a b), then
J(M) =
_
_
m
_
.2rC1/aC2t1
_
2m
if n C
0
r
and where m = na b:
m.2sC3/a
2m
if n D
0
s
, b _ (2s 3)a t ,
and where m = (n 1)a.
Proof. The argument in Theorem 4 carries over with the roles of a and b inter-
changed, and in addition, replacing s by s 1 in the case n T
0
s
. We omit the
proof.
5 Concluding Remarks
In the previous sections, we have been able to obtain exact values of J(M) in
many cases, and in some special cases, even the value of j(M). We expect that
the values of J(M) are in fact equal to j(M), although we have not been able to
show this. There are, however, a few cases where the exact value of J(M) has
eluded us. We state this as the following concluding remark.
Remark 2. Let M = {a. b. na}, with a < b and gcd(a. b) = 1. We have been
unable to determine J(M) in the following cases:
(i) a b is odd, n _ a b 1 and even, and satises b _ (2s 3)a t ;
(ii) a b is odd, n _ 2(a b) 1 and odd, and satises b _ 2(s 1)a t ;
(iii) a, b odd, n <
b.aCb2/
2a
and even.
Acknowledgments. The authors are grateful to the referee for his or her careful
reading and comments.
References
[1] D. G. Cantor and B. Gordon, Sequences of integers with missing differences, J. Com-
bin. Theory Ser. A 14 (1972), 281287.
[2] N. M. Haralambis, Sets of Integers with Missing Differences, J. Combin. Theory Ser.
A 23 (1977), 2233.
[3] T. S. Motzkin, Unpublished problem collection.
Author information
Ram Krishna Pandey, Department of Mathematics, Indian Institute of Technology,
Hauz Khas, New Delhi 110 016, India.
E-mail: pandey_mathiitd@yahoo.co.in
Amitabha Tripathi, Department of Mathematics, Indian Institute of Technology,
Hauz Khas, New Delhi 110 016, India.
E-mail: atripath@maths.iitd.ac.in
On Pseudosquares and Pseudopowers
Carl Pomerance and Igor E. Shparlinski
Abstract. Introduced by Kraitchik and Lehmer, an .-pseudosquare is a positive integer
n 1 (mod 8) that is a quadratic residue for each odd prime _ ., yet is not a square.
We use bounds of character sums to prove that pseudosquares are equidistributed in fairly
short intervals. An .-pseudopower to base g is a positive integer which is not a power of
g yet is so modulo for all primes _ .. It is conjectured by Bach, Lukes, Shallit, and
Williams that the least such number is at most exp(a
., log .) for a suitable constant a
.
A bound of exp(a
. log log ., log .) is proved conditionally on the Riemann Hypothesis

for Dedekind zeta functions, thus improving on a recent conditional exponential bound
of Konyagin and the present authors. We also give a GRH-conditional equidistribution
result for pseudopowers that is analogous to our unconditional result for pseudosquares.
Keywords. Pseudosquare, pseudopower, character sums.
AMS classication. 11L40.
1 Introduction
1.1 Pseudosquares
An .-pseudosquare is a nonsquare positive integer n such that n 1 (mod 8)
and (n,) = 1 for each odd prime _ .. The subject of pseudosquares was
initiated by Kraitchik and more formally by Lehmer in [14]. It was later shown by
Weinberger (see [19]) that if the Generalized Riemann Hypothesis (GRH) holds,
then the least .-pseudosquare, call it N
x
, satises N
x
_ exp(c.
1{2
) for a positive
constant c. The interest in this inequality is that there is a primality test, due
to Selfridge and Weinberger, for integers n < N
x
that requires the verication of
some simple Fermat-type congruences for prime bases _ .. Thus, a large lower
bound for N
x
leads to a fast primality test, and in particular this result gives an
alternate and somewhat simpler form of Millers GRH-conditional polynomial-
time deterministic primality test. See [19] for details.
The rst author was supported in part by NSF grant DMS-0703850. The second author was
supported in part by ARC grant DP0556431.
172 Carl Pomerance and Igor E. Shparlinski
By the GRH, we mean the Riemann Hypothesis for Dedekind zeta functions,
that is, for algebraic number elds. Note that this conjecture subsumes the Ex-
tended Riemann Hypothesis (ERH), which is the Riemann Hypothesis for rational
Dirichlet 1-functions. The Weinberger lower bound for N
x
in fact only requires
the ERH.
As the concept of an .-pseudosquare is a natural one, it is also of interest
to nd a reasonable upper bound for N
x
and also to study the distribution of .-
pseudosquares. Let
M(.) =
]_x
denote the product of the primes up to .. For nonsquare integers n coprime to

M(.), the probability that n satisfy n 1 (mod 8) and (n,) = 1 for all odd
primes _ . is 2
-t(x)-1
. Thus, it is reasonable perhaps to conjecture that
N
x
_ 2
(1o(1))t(x)
.
for example, see Bach and Huelsbergen [1]. In [18], Schinzel proves conditionally
on the GRH that
N
x
_ 4
(1o(1))t(x)
. (1)
and in particular, he conditionally shows that this inequality holds as well for the
smallest prime .-pseudosquare. Unconditionally, he uses the Burgess bound [5]
(see also [11, Theorem 12.6]) to show that
N
x
_ exp((1,4 o(1)).). (2)
We start with an observation, communicated to us by K. Soundararajan, that
the pigeonhole principle (used in the same fashion as in [7, Lemma 10.1]) gives
an unconditional proof of (1), though not for prime pseudosquares. Indeed, let us
put X = 2
t(x)
. and consider the
(1 o(1))
X
log X
= (1 o(1))
2
t(x)
log .
log 2
vectors of Legendre symbols
__
3
_
.
_
5
_
. . . . .
_
__
for all primes (.. X|, where is the largest prime with _ .. Clearly
there are at most 2
t(x)-1
possibilities for such vectors, so for large . there are
ve distinct primes
1
. . . . .
5
(.. X| for which they coincide. Thus, at least
two of them, say
1
.
2
, have the property that
1

2
(mod 8). Then
1
2
is an
On Pseudosquares and Pseudopowers 173
.-pseudosquare and we have N
x
_
1
2
_ X
2
, implying (1). (Note that it is not
necessary that the numbers be prime in this proof, just coprime to M(.).)
Our contribution to the subject of pseudosquares is on their equidistribution.
For this we follow Schinzels proof of (2), but use a character sum estimate given
in [11, Corollary 12.14], which dates back to work of Graham and Ringrose [6],
to prove the following result. Let S
x
be the set of .-pseudosquares.
Theorem 1. Uniformly for > 0 and N _ exp(3., log log .), we have
# (S
x
(. N|) = (1 o(1))
N
M(.)
# (S
x
(0. M(.)|) . . o.
We also show
# (S
x
(0. M(.)|) = (1 o(1))
M(.)
2
t(x)1
e
;
log .
. . o. (3)
so that one can rewrite Theorem 1 in a more explicit form.
We remark that, by the prime number theorem, M(.) = exp ((1 o(1)).)
thus our result is nontrivial starting with very small values of N compared to
M(.).
Note that Granville and Soundararajan [7] also discuss the equidistribution
of .-pseudosquares via the GrahamRingrose result on character sums, but their
context is different and it is not clear that Theorem 1 follows directly from their
paper.
1.2 Pseudopowers
Let g be a xed integer with [g[ _ 2. Following Bach, Lukes, Shallit, and
Williams [2], we say that an integer n > 0 is an .-pseudopower to base g if
n is not a power of g over the integers but is a power of g modulo all primes
_ .. Denote by q
(.) the least .-pseudopower to base g.

In [2] it is conjectured that for each xed g, there is a number a
such that for

. _ 2,
q
(.) _ exp(a
., log .). (4)

In addition, a heuristic argument is given for (4), with numerical evidence pre-
sented in the case of g = 2. For any g, we have (see [12]) the trivial bound
q
(.) _ 2M(.) 1, where M(.) is the product of the primes _ .. Thus,

q
(.) _ exp((1 o(1)).).

Using an estimate for exponential sums due to Heath-Brown and Konyagin [9] and
results of Baker and Harman [3, 4] on the BrunTitchmarsh theorem on average,
Konyagin, Pomerance, and Shparlinski [12] proved that
q
(.) _ exp(0.88715.)
for all sufciently large . and all integers g with 2 _ [g[ _ .. Further, it was
noted in [12] that the method implied that for xed g,
q
(.) _ exp((1,2 o(1)).).

assuming the GRH. In this paper we make further progress towards (4), again
assuming the GRH. Our proof makes use of the approach in Schinzel [18] for
pseudosquares.
Theorem 2. Assume the GRH. Then for each xed integer g with [g[ _ 2 there is
a number a
such that for . _ 3,

q
(.) _ exp(a
. log log ., log .).

We are also able to prove an equidistribution result conditional on the GRH
that is similar in strength to Theorem 1. Let P
x
be the set of .-pseudopowers
base g.
Theorem 3. Assume the GRH. Let g be a xed integer with [g[ > 1. There is a
positive number b
such that for any interval (. N| _ (0. o), uniformly

over N _ exp(b
., log log .), we have

# (P
x
(. N|) = (1 o(1))
N
M(.)
# (P
x
(0. M(.)|) . . o.
We derive an asymptotic formula for # (P
x
(0. M(.)|) in Section 3.2, see
(22), so that one can get a more explicit form of Theorem 3.
We note that in [12] an unconditional version of Theorem 3 is given which
however holds only for N _ exp(0.88715.). Under the GRH, the method of [12]
gives a somewhat stronger result but still requires N to be rather large, namely it
applies only to N _ exp ((0.5 c).) for an arbitrary c > 0.
As for lower bounds for q
(.), it follows from Schinzel [16, 17] that

q
(.) o. . o.
In [2] it is shown that assuming the GRH there is a number c
> 0 such that

q
(.) _ exp(c
_
.(log log .)
3
,(log .)
2
).
1.3 Notation
We recall that the notation U = O(V ) and U V are equivalent to the assertion
that the inequality [U[ _ c V holds for some constant c > 0.
2 Distribution of Pseudosquares
In this section we prove Theorem 1 by making use of the following character sum
estimate, which is [11, Corollary 12.14].
Lemma 4. Let be a primitive character to the squarefree modulus q > 1. Sup-
pose all prime factors of q are at most N
1{9
and let r be an integer with N
i
_ q
3
.
Then for any number ,
~n_1
(n)
_ 4Nt(q)
i{2
r
q
-1{i2
r
.
where t(q) is the number of positive divisors of q.
Recall that M(.) is the product of the primes in 1. .|. Let . be a large number
and let S
x
denote the set of positive integers n 1 (mod 8) with (n,) = 1 for
each odd prime _ .. That is, S
x
consists of the .-pseudosquares and actual
squares coprime to M(.). In particular, S
x
_ S
x
. We let M
2
(.) = M(.),2, the
product of the odd primes up to ..
Theorem 1 is routine once N is large compared with M(.), so we assume that
N _ M(.)
2
. Note that for a positive integer n with (8n 1. M
2
(.)) = 1, we
have that
][
2
(x)
_
1
_
8n 1
__
=
_
2
t(x)-1
. if 8n 1 S
x
;
0. else.
(5)
Thus, if . N are positive numbers, then the sum
S
,1
:=
~8n1_1
(8n1,
2
(x))=1
][
2
(x)
_
1
_
8n 1
__
satises
S
,1
= 2
t(x)-1
#(S
x
(. N|). (6)
The product in (5) can be expanded, so that we have
S
,1
=
~8n1_1
(8n1,
2
(x))=1
( [
2
(x)
_
8n 1
_
=
( [
2
(x)
~8n1_1
(8n1,
2
(x))=1
_
8n 1
_
.
(7)
The contribution to S
,1
from = 1 is
~8n1_1
(8n1,
2
(x))=1
1 ~
N
4e
;
log .
(8)
uniformly for . N with N _ exp(.
1{2
). This estimate follows immediately from
the fundamental lemma of the sieve; for example, see [8, Theorem 2.5].
Our goal now is to show that the contributions to S
,1
coming from values of
> 1 is small in comparison to (8). Suppose that [ M
2
(.), > 1 is xed. We
can rewrite the contribution in (7) corresponding to as
1
(
=
d[
2
(x){(
j(J)
~8n1_1
d[8n1
_
8n 1
_
. (9)
where j(J) is the Mbius function.
The PlyaVinogradov inequality (see [11, Theorem 12.5]) immediately im-
plies that
[1
(
[ _ 3 2
t(x)-2
_
log < 2
t(x)
_
log (10)
for any choice of > 1. We use (10) when is not much larger than N, namely
we use it when
_ N
i2
r
{(i2
r1
1)
.
where r shall be chosen later. In this case it gives
[1
(
[ _ 2
t(x)
N
1-2{(i2
r
2)
log(N
2
) _ 2
(1o(1))t(x)
N
1-2{(i2
r
2)
. (11)
For large values of , that is, when
> N
i2
r
{(i2
r1
1)
. (12)
we use a different approach which relies on Lemma 4.
Let r
(
= (1
2
),8, so that r
(
is an integer and 8r
(
1 (mod ). Then
1
(
=
d[
2
(x){(
j(J)
_
J
_

~dk_1
k=d (mod 8)
_
k
_
=
d[
2
(x){(
j(J)
_
8J
_

~d(8Id)_1
_
l Jr
(
_
=
d[
2
(x){(
j(J)
_
8J
_

nI
d;f
_
m
_
.
where I
d,(
= T
d,(
1. T
d,(
N
d,(
|, an interval of length
N
d,(
=
N
8J
O(1).
Thus,
[1
(
[ _
d[
2
(x){(
nI
d;f
_
m
. (13)
The character sums in (13) where 8J > N
0.1
are trivially bounded by N
d,(
=
O(N
0.9
) in absolute value, so their total contribution to 1
(
is
d[
2
(x){(
8d>1
0:1
nI
d;f
_
m
2
t(x)
N
0.9
. (14)
We now assume that 8J _ N
0.1
. Note that the conductors of the characters
which appear in (13) are squarefree. We choose r as the largest integer with
r2
i
2 _
log .
log log .
and apply Lemma 4 to the inner sum in (13) with this value of r. To do this we
need
(N,(8J))
i
_
3
and . _ (N,(8J))
1{9
.
These inequalities hold since
r =
_
1
log 2
o(1)
_
log log . and N _ exp(3., log log .). (15)
so that
_
N
8J
_
i
_ N
0.9i
_ exp(2.7r., log log .) _ M(.)
3
_
3
(16)
and
_
N
8J
_
1{9
_ N
0.1
_ exp(0.3., log log .) _ . (17)
for all large ..
Thus, by Lemma 4,
d[
2
(x){(
8d_1
0:1
nI
d;f
_
m
_ 4 2
t(x)-1
N2
(t(x)-1)i{2
r
-1{i2
r
. (18)
We now derive from (14), (18), and (12) that
[1
(
[ _ 2
(1o(1))t(x)
N
0.9
2
(1o(1))t(x)
N
-1{i2
r
_ 2
(1o(1))t(x)
N
1-2{(i2
r
2)
which is also the bound in (11) for those > 1 not satisfying (12).
Now summing over we see that the contribution to S
,1
from values of
> 1 is at most
( [
2
(x)
( >1
[1
(
[ _ 4
(1o(1))t(x)
N
1-2{(i2
r
2)
= N
1-2{(i2
r
2)
exp((log 4 o(1))., log .).
Note that by our choice of r we have
N
2{(i2
r
2)
_ N
2 log log x{ log x
_ exp(6., log .).
Since 6 > log 4, we have that the contribution to (7) from terms with > 1 is
small compared to the main term given by (8), so that
S
,1
= (1 o(1))
N
4e
;
log .
.
Together with (6), we now have
#(S
x
(. N|) = (1 o(1))
N
2
t(x)1
e
;
log .
.
Since the number of squares in the interval (. N| is at most N
1{2
, we obtain
# (S
x
(. N|) = (1 o(1))
N
2
t(x)1
e
;
log .
.
Taking = 0 and N = M(.) we derive (3) and conclude the proof of Theorem1.
We remark that by being a little more careful with the estimates, we can prove
the theorem with 3 replaced with any xed number larger than log 8.
3 Distribution of Pseudopowers
3.1 Proof of Theorem 2
Let g be a given integer with [g[ _ 2 which we assume to be xed. Let
(.)
be the least positive integer which is not a power of g yet is a power of g modulo
every prime _ . with g. It is easy to see that
q
(.) _ g
(.). (19)
Indeed, the integer g
(.) is not a power of g, it is a power of g modulo every

prime _ . with g, and it is zero modulo for every prime [ g, and so is
a power of g modulo these primes too.
For every prime g let l
() be the multiplicative order of g modulo ,

and let i
() = ( 1),l
(), the index of the subgroup of powers of g in the

multiplicative group modulo . Let
M
(.) =
]_x
]
. 1
(.) =
][
g
(x)
i
().
It follows from [12, Theorem 1] that 1
(.) _ exp(0.42.) for all sufciently large

.. We conditionally improve this result.
Lemma 5. Assume the GRH. There is a number c
such that for . _ 3,

1
(.) _ exp(c
. log log ., log .).

Proof. In Kurlberg and Pomerance [13, Theorem 23], following ideas of Hoo-
ley [10] and Pappalardi [15], it is shown conditionally on the GRH that
][
g
(x)
I
g
(])_]{,
1
(.)
,
. log log .
(log .)
2
for 1 _ , _ log .. Applying this result with , = log ., we have
][
g
(x)
i
g
(])_log x
1
. log log .
(log .)
2
.
Indeed, i
() _ , implies that l
() _ ( 1),, < ,,. Since we trivially

have i
() _ . for each prime [ M
(.) with g, we thus have

1
(.) =
][
g
(x)
i
g
(])~log x
i
()
][
g
(x)
i
g
(])_log x
i
() _ (log .)
t(x)
.
O
g
(x log log x{(log x)
2
)
.
The lemma follows. |L
For each prime g, let
]
be a character modulo of order i
(). Then
i
g
(])
}=1
}
]
(n) =
_
i
(). if n is a power of g modulo .

0. else.
Thus,
][
g
(x)
i
g
(])
}=1
}
]
(n)
=
_
1
(.). if n is a power of g modulo every [ M
(.).
0. else.
(20)
Let (n) denote the von Mangoldt function. From the denition of
(.) we
deduce that
S
:=
n~]
g
(x)
(n)
][
g
(x)
i
g
(])
}=1
}
]
(n) = 1
(.)
n~]
g
(x)
nis a power of
(n).
The last sum is 0 if g is not a prime or prime power, and in any event is always at
most log
(.) .. Thus,
S
(.).. (21)
We now multiply out the product in (20); it is seen as a sum of 1
(.) characters
modulo M
(.). The contribution to S
from the principal character
]

i
g
(])
]
is [(
(.)) O(.). We may assume that
(.) _ e
t(x)
since otherwise the
theorem follows immediately from (19), so that the contribution to S
from the
principal character is (1 o(1))
(.), by the prime number theorem.

The contribution to S
from each nonprincipal character is
n~]
g
(x)
(n)(n)
(.)
1{2
_
(log M
(.))
2
(log
(.))
2
_
(.)
1{2
.
2
assuming the GRH. Hence, the contribution to S
from nonprincipal characters is

O(1
(.)
(.)
1{2
.
2
). Thus,
S
= (1 o(1))
(.) O(1
(.)
(.)
1{2
.
2
).
so that from (21) we deduce that
(.) 1
(.)
2
.
4
.
Theorem 2 now follows immediately from (19) and Lemma 5.
We remark that an alternate way to handle primes dividing g is to eschew (19)
and instead multiply the product in (20) by
_ mod
(n). This sum is (g) when
n 1 (mod g) and is 0 otherwise. Note that a number that is 1 mod is always a
power of g modulo . Although this is somewhat more complicated, it does lead
to a proof that there is a prime number below the bound exp(a
. log log ., log .)

that is an .-pseudopower base g.
3.2 Proof of Theorem 3
We use the method of proof of Theorem 1. Accordingly we only outline some
new elements and suppress the details.
For a nonzero integer n, let rad(n), the radical of n, be the largest squarefree
divisor of n. That is, rad(n) is the product of the distinct prime factors of n. Also,
let o(n) denote the number of distinct prime factors of n. Let P
x
denote the set
of positive integers which are either an .-pseudopower base g or a true power of
g. Then a positive integer n P
x
if and only if both
(i) n is in the subgroup (g) of (Z,Z)
+
when _ . and g;
(ii) n 0 or 1 (mod ) when _ . and [ g.
Assuming then that . _ [g[, the cardinality of P
x
(0. M(.)| is
2
o()

][
g
(x)
l
() = 2
o()

][
g
(x)
1
i
()
=
2
o()
(M
(.))
1
(.)
~
2
o()
M(.)
e
;
(rad(g))1
(.) log .
.
by the formula of Mertens.
It is easy to see that this expression is exponentially large, either from the
observation that l
() _ 2 whenever
_
]
_
= 1, so that #(P
x
(0. M(.)|)) _
2
(1{2o(1))t(x)
, or using 1
(.) _ e
0.42x
from [12]. Further, the number of true
powers of g in (0. M(.)| is small; it is O(.). Thus,
# (P
x
(0. M(.)|) = (1 o(1))
2
o()
M(.)
e
;
(rad(g))1
(.) log .
. . o. (22)
To prove Theorem 3 we again use Lemma 5, which is GRH-conditional. But
the framework of the proof follows the argument of Theorem 1, and in particular
it uses the unconditional Lemma 4. Notice that the proof of Theorem 2 used
Riemann Hypotheses a second time, namely in the estimation of the weighted
character sums. Now we use unweighted character sums and so are able to use
Lemma 4. The set-up is as follows. Let
1
,1
=
ob=rad()
~n_1
n=0(mod o)
n=1(mod b)
][
g
(x)
i
g
(])
}=1
}
]
(n). (23)
This expression counts integers n (. N| that are 0 or 1 (mod ) for each
prime [ g and in the subgroup (g) of (Z,Z)
+
for each prime _ . with
g. Namely, it counts members of P
x
, and does so with the weight 1
(.).
To prove Theorem 3 one then expands the product in (23). As usual, the contri-
bution from the principal character is easily estimated: it is the number of integers
n (. N| which are 0 or 1 (mod ) for each prime [ g and are coprime
to M
(.). Thus, the principal character gives the contribution

(1 o(1))
2
o()
rad(g)
(M
(.))
M
(.)
N = (1 o(1))
2
o()
N
e
;
(rad(g)) log .
.
which when divided by the weight 1
(.) gives the main term for our count.

The nonprincipal characters have conductors corresponding to those divisors
of M
(.) with > 1 and i
() > 1 for each prime [ . For such integers

, the characters that occur with conductor are induced by characters in the set
X
(
=
_
][(
}
p
]
: 1 _
]
_ i
() 1 for [
_
.
Thus, the contribution of the nonprincipal characters to 1
,1
is
1
+
,1
:=
ob=rad()
( [
g
(x)
( >1
_A
f
~n_1
n=0 (mod o)
n=1 (mod b)
(n,
g
(x){( )=1
(n).
where X
(
is empty if i
() = 1 for some prime [ . The inner sum is
d[
g
(x){(
j(J)
~n_1
n=0 (mod od)
n=1 (mod b)
(n).
As before we estimate the character sum here trivially if J is large, we use the
PlyaVinogradov inequality if is small, and we use Lemma 4 in the remaining
cases. However, we modify slightly the choice of r in the proof of Theorem 1 and
take it now as the largest integer with
r2
i
2 _
log .
(log log .)
2
.
Then we still have (15) and thus the conditions (16) and (17) are still satised so
that Lemma 4 may be used. Since each [X
(
[ < 1
(.), we obtain
[1
+
,1
[ _ 4
(1o(1))t(x)
1
(.)N
1-2{(i2
r
2)
.
This leads us to the asymptotic formula
1
,1
= (1 o(1))
2
o()
N
e
;
(rad(g)) log .
O
_
1
(.)4
(1o(1))t(x)
N
1-2{(i2
r
2)
_
.
For N _ exp(b
., log log .) we have

N
2{(i2
r
2)
_ N
2(log log x)
2
{ log x
> exp
_
2b
. log log ., log .

_
.
Using Lemma 5 and taking b
= c
, we conclude the proof of Theorem 3.

Acknowledgments. The authors are grateful to K. Soundararajan for commu-
nicating the proof we presented in Section 1.1 and for his permission to give it
here.
References
[1] E. Bach and L. Huelsbergen, Statistical evidence for small generating sets, Math.
Comp. 61 (1993), 6982.
[2] E. Bach, R. Lukes, J. Shallit and H. C. Williams, Results and estimates on pseu-
dopowers, Math. Comp. 65 (1996), 17371747.
[3] R. C. Baker and G. Harman, The BrunTitchmarsh theorem on average, in: Proc.
Conf. in Honor of Heini Halberstam (Allerton Park, IL, 1995), pp. 39103, Progress
in Mathematics, vol. 138, Birkhuser, Boston, 1996.
[4] R. C. Baker and G. Harman, Shifted primes without large prime factors, Acta Arith.
83 (1998), 331361.
[5] D. A. Burgess, On character sums and L-series. II, Proc. London Math. Soc. 13
(1963), 524536.
[6] S. W. Graham and C. J. Ringrose, Lower bounds for least quadratic nonresidues, in:
Analytic Number Theory, Allerton Park 1989, pp. 269309, Progress in Mathemat-
ics, vol. 85, Birkhuser, Basel, 1990.
[7] A. Granville and K. Soundararajan, The distribution of values of 1(1.
d
), Geom.
and Func. Anal. 13 (2003), 9921028.
[8] H. Halberstam and H.-E. Richert, Sieve Methods, Academic Press, London, 1974.
[9] D. R. Heath-Brown and S. V. Konyagin, New bounds for Gauss sums derived from
kth powers, and for Heilbronns exponential sum, Quart. J. Math. 51 (2000), 221
235.
[10] C. Hooley, On Artins conjecture, J. Reine Angew. Math. 225 (1967), 209220.
[11] H. Iwaniec and E. Kowalski, Analytic Number Theory, Colloquium Pubs., vol. 53,
Amer. Math. Soc., Providence, RI, 2004.
[12] S. Konyagin, C. Pomerance and I. E. Shparlinski, On the distribution of pseudopow-
ers, Canad. J. Math., to appear.
[13] P. Kurlberg and C. Pomerance, On the period of the linear congruential and power
generators, Acta Arith. 119 (2005), 149169.
[14] D. H. Lehmer, A sieve problem on pseudo-squares, Math. Tables and Other Aids
to Computation 8 (1954), 241242.
[15] F. Pappalardi, On the order of nitely generated subgroups of Q
+
(mod ) and
divisors of 1, J. Number Theory 57 (1996), 207222.
[16] A. Schinzel, On the congruence a
x
b (mod ), Bull. Acad. Polon. Sci., Sr. Sci.
Math. Astronom. Phys. 8 (1960), 307309.
[17] A. Schinzel, A renement of a theorem of Gerst on power residues, Acta Arith. 17
(1970), 161168.
[18] A. Schinzel, On pseudosquares, in: New Trends in Probability and Statistics, Pa-
longa, 1996, vol. 4, pp. 213220, VSP, Utrecht, 1997.
[19] H. C. Williams, Primality testing on a computer, Ars Combinatoria 5 (1978), 127
185.
Author information
Carl Pomerance, Department of Mathematics, Dartmouth College,
Hanover, NH 03755-3551, USA.
E-mail: carl.pomerance@dartmouth.edu
Igor E. Shparlinski, Department of Computing, Macquarie University,
Sydney, NSW 2109, Australia.
E-mail: igor@ics.mq.edu.au
Maier Matrices Beyond Z
Frank Thorne
Abstract. This paper is an expanded version of a talk given at the 2007 Integers Con-
ference, giving an overview of the Maier matrix method and surveying the authors work
in extending it beyond the integers.
Keywords. Maier matrix method, distribution of primes, Shius theorem, function elds.
AMS classication. 11N05, 11N13, 11T55.
1 Maier Matrices
Loosely speaking, a Maier matrix is a combinatorial device used to prove the ex-
istence of irregular or interesting patterns in the distribution of primes or related
sequences. We will illustrate the technique with two particularly interesting ex-
amples. The rst is Maiers 1985 proof [6] that unexpected irregularities exist
in the distribution of primes in short intervals. In particular, Maier proved that for
any > 0 there exists a constant
A
> 0 such that
limsup
n!1
(n log
A
n) (n)
log
A1
n
_ 1
A
.
liminf
n!1
(n log
A
n) (n)
log
A1
n
_ 1
A
.
(1.1)
These irregularities are unexpected in the sense that they contradict probabilistic
heuristics for > 2.
The proof is as follows. For a variable ,, let Q =
Q
p<y
, let .
1
= Q
D
for
some xed large D, and let C be a parameter to be determined later. Consider the
following matrix of integers:
2
6
6
6
4
Q.
1
1 Q.
1
2 . . . Q.
1
,
C
Q(.
1
1) 1 Q(.
1
1) 2 . . . Q(.
1
1) ,
C
.
.
.
.
.
.
.
.
.
.
.
.
Q(2.
1
) 1 Q(2.
1
) 2 . . . Q(2.
1
) ,
C
3
7
7
7
5
.
The author is grateful for nancial support from an NSF VIGRE fellowship.
186 Frank Thorne
The columns form arithmetic progressions modulo Q, and so the prime number
theorem for arithmetic progressions predicts
1
that for each i 1. ,
C
| which is
coprime to Q, the corresponding column should contain ~
Q
.Q/
x
1
log.Qx
1
/
primes.
Therefore, the number of primes in the matrix, and thus in an average row, can
be asymptotically determined by counting the number of such i . In fact, the latter
quantity is
(,
C
. ,) ~ ,
C
(Q)
Q
e
o(C). (1.2)
for a function o(C) which converges to e
, but oscillates above and below e
as C o.
2
The short intervals occurring in the Maier matrix are of the sort
considered in (1.1), and Maiers theorem soon follows.
In 1997, Shiu [8] similarly proved the remarkable result that if a. q, and k are
arbitrary integers with (a. q) = 1, there exists a string of k consecutive primes
nC1

nC2

nCk
a (mod q).
(Here
n
denotes the nth prime.) Furthermore, for k sufciently large in terms of
q, these primes can be chosen to satisfy the bound
1
(q)
log log
nC1
log log log log
nC1
(log log log
nC1
)
2
1=.q/
k. (1.3)
To prove (1.3) Shiu constructed a similar Maier matrix; the primary difference is
in the choice of Q. For example, if a = 1, primes , 1 (mod m) are excluded
from the product. This forces most primes in the matrix to be 1 (mod m), and
Shius result easily follows.
The method has been similarly adapted to prove a host of interesting results
about the distribution of the primes and related integer sequences. For more
on this, we recommend the outstanding survey articles of Granville [4] and of
Soundararajan [9]. In this article we will consider the problem of adapting the
Maier matrix method to different settings. In particular we will describe exten-
sions of the method to the polynomial ring F
q
t | and to imaginary quadratic elds,
where we obtained analogous results.
2 Maier Matrices in F
q
t
The polynomial ring F
q
t | (here F
q
is a nite eld) has long been studied as an
analogue to the integers. Like the integers, F
q
t | enjoys unique factorization, and
1
This is not known to be true for all Q, except under GRH. However, a theorem of Gallagher [3]
implies the correct asymptotic for an innite set of such Q, where the error term in the asymptotic
depends on D.
2
In general, (.. ,) denotes the number of n _ ., all of whose prime factors are at least ,.
Maier Matrices Beyond Z 187
has the additional property that the residue class rings are all nite, so that one
may naturally talk about the size of elements.
Classical methods of analytic number theory have been extremely successful
in analyzing the distribution of primes (e.g., monic irreducible polynomials) in
F
q
t |. For example, one denes the zeta function as
F
q
t
(s) :=
X
x2F
q
t
1
[.[
s
. (2.1)
where the sum is over all monic polynomials ., and [.[ := #[F
q
t |,(.)[ = q
deg x
.
As there are exactly q
n
monics of degree n, one easily obtains the formula
F
q
t
(s) =
1
1 q
1s
.
The Riemann hypothesis is then a triviality, and in fact one has an exact prime
number theorem
(n) =
1
n
X
djn
j(J)q
n=d
.
Thinking of n as log
q
(q
n
), this closely mirrors the classical case.
It is therefore natural to ask whether the Maier matrix method can be adapted
to F
q
t |, and we have answered this question in the afrmative. To start with, we
have the following
Theorem 2.1 ([11, Theorem 1.1]). For any xed > 0, there exists a constant
A
> 0 (depending also on q) such that
limsup
k!1
sup
deg f Dk
(. {log k)
q
dAlog keC1
,k
_ 1
A
.
liminf
k!1
inf
deg f Dk
(. {log k)
q
dAlog keC1
,k
_ 1
A
.
Here (. i ) denotes the number of irreducible monic polynomials with deg(
) _ i .
The proof follows similar lines. We write Q =
Q
deg pn
, and consider the
following matrix:
2
6
6
6
4
Qg
1
h
1
Qg
1
h
2
. . . Qg
1
h
j
Qg
2
h
1
Qg
2
h
2
. . . Qg
2
h
j
.
.
.
.
.
.
.
.
.
.
.
.
Qg
i
h
1
Qg
i
h
2
. . . Qg
i
h
j
3
7
7
7
5
.
188 Frank Thorne
Here g
1
through g
i
range through all monic polynomials of degree 2 deg Q (or of
degree deg Q for any xed > 1), and h
1
through h
j
run through all polynomi-
als of degree s (of arbitrary leading coefcient), for a parameter s to be determined
later. The number of primes in the whole matrix is ~
ij
3 deg Q
e
o(s,n), and by
appropriately choosing s in terms of n, the matrix and thus some row can be made
to contain more or fewer primes than expected.
We also proved the following function eld analogue of Shius theorem:
Theorem 2.2 ([11, Theorem 1.2]). Suppose that k is a positive integer, and a and
m are polynomials with m monic and (a. m) = 1. Then there exists a string of
consecutive primes
rC1

rC2

rCk
a (mod m).
Furthermore, for sufciently large k, these primes may be chosen so that their
common degree D satises
1
(m)
log D
(log log D)
2
1=.m/
k. (2.2)
The implied constant depends only on q.
The observant reader will notice that it is not obvious what consecutive
means; the elements of F
q
t | are not naturally ordered in the same way as the
integers. We may in fact order our primes with respect to any ordering compatible
with our Maier matrix construction, and in particular our theorem applies with
respect to lexicographic order.
This theorem was extended in an interesting way by Tanner [10], who proved
the following:
Theorem (Tanner). Under the same hypotheses there exists an integer D
0
(de-
pending on q, k, and m) such that for each D _ D
0
there exists a string of
consecutive primes
rC1

rC2

rCk
a (mod m)
of degree D. Furthermore, for sufciently large k, D
0
satises (2.2).
Tanners proof is an extension of the authors proof; the point is that since
there are many polynomials of the same degree in F
q
t |, it is possible to construct
appropriate Maier matrices where all the polynomials in the matrix are of a given
degree.
3 Prime Bubbles in Imaginary Quadratic Fields
We will now consider the problem of adapting Maiers matrix method to number
elds. Let 1 be an imaginary quadratic eld. In this setting a positive propor-
tion of ideals correspond to elements (although the unit group interferes), and the
prime elements of O
K
can be naturally visualized as lattice points in Z. Adapt-
ing the proof of Shius theorem, we proved that there are clumps of primes, all of
which lie in an arbitrary xed arithmetic progression, up to multiplication by units:
Theorem 3.1 ([12, Theorem 1.1]). Suppose 1 is an imaginary quadratic eld, k
is a positive integer, and a and q are elements of O
K
with q = 2 and (a. q) = 1.
Then there exists a bubble
T(r. .
0
) := {. C : [. .
0
[ < r} (3.1)
with at least k primes, all of which are congruent to ua modulo q for units u
O
K
. Furthermore, for k sufciently large in terms of q (and 1), .
0
can be chosen
to satisfy
1
K
(q)
log log [.
0
[ log log log log [.
0
[
(log log log [.
0
[)
2
!
K
=h
K
K
.q/
k. (3.2)
The implied constant is absolute.
Here o
K
denotes the number of units in O
K
, h
K
is the class number of 1, and
K
(q) := [(O
K
,(q))
[. As an example of such a prime bubble in Zi | (which

we found by computer search), the ball of radius
_
23.5 centered at 59 779i
contains six primes, all congruent to 1 or i modulo 5 i .
To prove our result we construct the following Maier matrix:
2
6
6
6
4
Qi
1
b
1
Qi
1
b
2
. . . Qi
1
b
J
Qi
2
b
1
Qi
2
b
2
. . . Qi
2
b
J
.
.
.
.
.
.
.
.
.
.
.
.
Qi
I
b
1
Qi
I
b
2
. . . Qi
I
b
J
3
7
7
7
5
. (3.3)
Q is dened to be any generator of the ideal Q, given by
Q:= q
Y
p2P;pp
0
p. (3.4)
where P ranges over primes of norm _ , with restrictions on the residue classes
modulo q (which depend on a). The need to exclude one prime p
0
will be ex-
plained shortly.
The i range over all elements of O
K
with norm in (NQ
D
. 2NQ
D
), and the
b range over all elements of norm less than either ,: or 9,: (where : will be
190 Frank Thorne
chosen later in terms of ,). In effect we are constructing two Maier matrices, a
good matrix (the smaller one, where Nb < ,:) and a bad one. We then prove
that nearly all of the primes in the matrix are a (mod q), where good primes
a (mod q) are counted only in the good matrix, and bad primes , a (mod q)
are counted in the larger bad matrix.
There are two more important ingredients in the proof. The rst is an appropri-
ate version of the prime number theorem for arithmetic progressions, valid when
the relative size of Q and i is as in (3.3). We cannot expect to prove such a re-
sult for all Q, but we can for a large class of moduli Q, as dened in (3.4). The
starting point is a zero-density estimate for Hecke 1-functions proved by Fogels
[2], and we then follow techniques of Gallagher [3] and Shiu to obtain our result.
(We remove the prime p
0
to ensure that the associated 1-functions do not have
any Siegel zeroes.)
The last ingredient in the proof is a bit of combinatorial geometry. Using the
above techniques, we can prove that that some row of our Maier matrix is a pair
of concentric balls in the complex plane, such that the inner ball contains many
more good primes than the outer ball has bad. We must now prove that this bubble
contains a sub-bubble containing many good primes and no bad ones.
To do this we rely on the existence of a Delaunay triangulation. The Delaunay
triangulation of a set of points has the property that no point in the triangulation
is inside the circumcircle of any triangle. We take our set of points to be the set
all bad primes within the outer ball, as well as a regular 7-gon (of a certain radius)
outside the inner ball but inside the outer one. The circumcircles associated to
the Delaunay triangulation contain all of the good points and none of the bad,
and the number of such circumcircles is easily bounded from above. Moreover,
any circumcircle intersecting the inner ball can be proved to lie entirely within
the outer ball. The circumcircle containing the most good primes is therefore our
bubble of congruent primes.
4 Concluding Remarks
In the rst place, we would like to discuss some additional results which we do not
have the space to fully describe here. In particular, Granville and Soundararajan
[5] recently generalized Maiers theorem and proved that similar irregularities
occur in any arithmetic sequence. Here an arithmetic sequence is any sequence
A of integers, such that for all integers J coprime to some bad modulus S, the
proportion of elements of A divisible by J is asymptotic to h(J),J, where h(J)
is a multiplicative function h(J) taking values in 0. 1|. It is also assumed that a
suitable weighted average of h() is sufciently smaller than 1.
Examples of such sequences include the primes and arbitrary subsets thereof,
almost primes, sums of two squares, norms of algebraic integers from extensions
of Q, and many other interesting sequences. Granville and Soundararajans main
result is then that any such sequence cannot be uniformly distributed in both short
intervals and arithmetic progressions to somewhat large moduli. To prove their
result they combine a generalized Maier matrix construction with a detailed anal-
ysis of oscillation in arithmetic functions (such as the function o(C) occurring
in (1.2)).
In [13], the present author translated their mechanism to F
q
t |. In brief, the
method works. In particular we obtained several results on general arithmetic se-
quences in F
q
t |, exactly along the lines suggested by Granville and Soundarara-
jans work. Furthermore, in some cases we were be able to be quite precise about
where irregularities occur, proving (for example) that they occur among the poly-
nomials of every sufciently large degree.
We conclude with a few remarks about some related work and some questions
that remain. Recently, Pollack [7] has proved an F
q
t | version of the quantita-
tive BatemanHorn conjecture (which implies the HardyLittlewood prime tuple
conjecture as a special case), valid when q is coprime to 2n and large in relation
to n. Conversely, Conrad, Conrad, and Gross [1] have found a global obstruc-
tion to a somewhat different version of this conjecture. This obstruction is related
to a certain average of the Mbius function, and these authors propose a revised
conjecture based on geometric considerations as well as numerical calculations.
Finite extensions of F
q
t | are naturally associated to algebraic curves, and one
wonders whether the geometry of these curves may have additional consequences
for the distribution of primes.
In the number eld case we have only scratched the surface, and one could
hope to prove all sorts of additional results. For example, one might ask whether
one could prove a result similar to Theorem 2.2 for any number eld. The state-
ment of such a result might be somewhat more involved, but certainly we believe
that the proof should generalize.
One might also ask whether irregularities of the form (1.1) can be proved to
exist in number elds. The norms of primes are already known to be irregularly
distributed, as these form an arithmetic sequence in the sense of [5]. But nothing
has yet been proved about these sequences themselves. We are optimistic that
techniques similar to those discussed in this article should yield interesting results.
Acknowledgments. This work was completed at the University of Wisconsin,
where the author was a Ph.D. student. I would like to thank the very many people
who read my papers and who listened to my talk at Integers Conference 2007
and elsewhere, and who made many useful suggestions. I would also like to again
192 Frank Thorne
point out the excellent survey articles of Granville [4] and Soundararajan [9], from
which I learned much.
References
[1] B. Conrad, K. Conrad and R. Gross, Prime specialization in genus 0, Trans. Amer.
Math. Soc. 360 (2008), 28672908.
[2] E. Fogels, On the zeros of 1-functions, Acta Arith. 11 (1965), 6796.
[3] P. X. Gallagher, A large sieve density estimate near o = 1, Invent. Math. 11 (1970),
329339.
[4] A. Granville, Unexpected irregularities in the distribution of prime numbers, in: Pro-
ceedings of the International Congress of Mathematicians (Zrich, 1994), pp. 388
399, Birkhuser, Basel, 1995.
[5] A. Granville and K. Soundararajan, An uncertainty principle for arithmetic se-
quences, Ann. of Math. 165 (2007), no. 2, 593635.
[6] H. Maier, Primes in short intervals, Michigan Math. J. 32 (1985), 221225.
[7] P. Pollack, Simultaneous prime specializations of polynomials over nite elds, ac-
cepted for publication in Proc. London Math. Soc.
[8] D. K. L. Shiu, Strings of congruent primes, J. London Math. Soc. 61 (2000), 359
373.
[9] K. Soundararajan, The distribution of prime numbers, in: Equidistribution in Num-
ber Theory. An Introduction, pp. 5983, NATO Sci. Ser. II Math. Phys. Chem. 237,
Springer, Dordrecht, 2007.
[10] N. Tanner, Strings of consecutive primes in function elds, accepted for publication
in Int. J. Number Theory.
[11] F. Thorne, Irregularities in the distribution of primes in function elds, J. Number
Theory 128 (2008), 17841794.
[12] F. Thorne, Bubbles of congruent primes, submitted.
[13] F. Thorne, An uncertainty principle for function elds, submitted.
Author information
Frank Thorne, Department of Mathematics, University of South Carolina,
1523 Greene Street, Columbia, SC 29208, USA.
Current address:
Department of Mathematics, Stanford University,
450 Serra Mall, Stanford, CA 94305, USA.
E-mail: fthorne@math.stanford.edu
Linear Equations Involving Iterates
of .N/
Tomohiro Yamada
Abstract. We study integers N satisfying the equation o(o(N)) = o(N) TN.
Keywords. Perfect numbers, superperfect numbers, sum of divisors.
AMS classication. 11A05, 11A25.
1 Introduction
We denote by o(N) the sum of divisors of N. N is said to be perfect if o(N) =
2N and multiperfect if o(N) = kN for some integer k. It is not known whether or
not an odd perfect/multiperfect number exists. There are many known conditions
which must be satised by such a number. But these results are far fromanswering
whether or not an odd perfect/multiperfect number exists.
Suryanarayana [7] called N superperfect if o(o(N)) = 2N and Pomerance
[5] called N super multiply perfect if o(o(N)) = kN for some integer k. More
generally, Cohen and te Riele [2] dened N to be (m. k)-perfect if o
.m/
(N) =
kN, where o
.m/
(N) denotes o(o
.m1/
(N)) with o
.0/
(N) = N.
We introduce another analogous notion of perfect/multiperfect numbers. We
say that N is (n: a
0
: . . . : a
n
)-perfect if
P
n
iD0
a
ni
o
.i/
(N) = 0 and (n: a
1
. . . . .
a
n
)-perfect if N is (n: 1 : a
1
: . . . : a
n
)-perfect. In particular, N is (2: . T)-
perfect if o(o(N)) = o(N) TN.
We begin by noting that almost all integers are (2: . T)-perfect for some in-
tegers . T. This fact follows from Katai and Subbarao [3, Theorem 1]. They
proved that for any xed m _ 1, we have (N. o
.m/
(N)) =
Q
p;p
a
jjN

a
, where
runs over all primes below .
m
2
, for all integers N _ . with o(.) exceptions (Here
we use the notation .
0
= .. .
iC1
= max{1. log .
i
} for all i _ 0, where log .
is the natural logarithm of ., introduced in [4]). Hence, for almost all integers
N, (N. o(N)) divides o
.2/
(N), which implies that there exist some integers . T
such that o(o(N)) = o(N) TN.
On the other hand, we can show that for any xed positive integers . T, the
set of (2: . T)-perfect numbers has density zero.
194 Tomohiro Yamada
Theorem 1.1. Let . T. C be integers not all zero and satisfying C _ 0. Then
the number of (2: : T: C)-perfect numbers below . is at most . exp((2
1=3
o(1))(log .)
1=3
(log log .)
2=3
).
Our argument is similar to the argument of Pomerance [6] to study the distri-
bution of integers n satisfying o(n) a (mod n). His argument rests on the fact
that almost all integers n can be written as m, where m is prime, large and
uniquely determined by m except some special cases. Our argument adopts a fac-
torization of 1 into lq with q large and (l. q) = 1, to show that q is uniquely
determined by m and l under some condition.
We note that this result does not seem to apply to the case C > 0. If . 21
are both prime, then n = 21 satises 2o(o(n))3o(n)6n = 0 and therefore
n is (2: 2 : 3 : 6)-perfect. More generally, we can easily conrm the following
result.
Theorem 1.2. Let 1 be an arbitrary positive integer. If and N = 1 1 are
both prime, then N is (2: 1 : (1 1)o(1) : 1o(1))-perfect.
Corollary 1.3. If 1 is a k-multiperfect number and . N = 1 1 are both
prime, then N is (2: k(1 1) : k1)-perfect.
A well-known conjecture states that, for any even integer 1, the number of
primes < . with 1 1 also prime is asymptotically equal to c.,(log .)
2
for some constant c > 0 depending on 1, contrary to the above given estimate
O(. exp((2
1=3
o(1))(log .)
1=3
(log log .)
2=3
)).
2 Notations and Preliminary Lemmas
We denote by 1(n). (n) the largest and smallest prime factor of n respectively.
For the positive real number ., let us denote .
0
= .. .
iC1
= max{1. log .
i
}
as mentioned in the previous section. We denote by c some positive constant
not necessarily same at every occurrence. Furthermore, we denote by .. ,. : real
numbers and we put u = log ., log , and = log ,, log :.
Lemma 2.1. Denote by (.. ,) the number of integers n _ . divisible by no
prime > ,. If , > .
2
1
, then we have (.. ,) < . exp((1o(1))ulog u) as .. u
tend to innity.
Proof. This follows from a well-known theorem of de Bruijn [1]. For details on
the distribution of integers free from large prime factors, we refer the readers to
[8, Chapter III. 5], where a simple proof of the lemma is also given.
Linear Equations Involving Iterates of o(N) 195
Lemma 2.2. Let
s(.. k) =
X
px;p1 .mod k/
1
. (1)
Uniformly in k and . _ e
2
, we have
s(.. k)
.
2
(k)
. (2)
Proof. This inequality can be immediately obtained using partial summation and
the BrunTitchmarsh inequality. The complete proof is given in [4, Lemma 2].
Lemma 2.3. Let S(.) = {n [ n _ ..
a
[ n for some . a with
a
> ,. a _ 2}.
Then we have #S(.) .,
1=2
.
Proof. Let (t ) be the number of perfect powers below t . It is clear that (t ) <
t
1=2
t
1=3
t
1=k
< t
1=2
ct
1=3
log t t
1=2
, where k = ](log .),(log 2).
Let us denote by ;
p
the smallest integer ; for which
> , and ; > 1.

Clearly we have #S(.) _ .
P
px

p
. Since
p
> ,, we have by partial
summation
X
px
1
p
_
(.)
.
(,)
,
Z
x
y
(t )
t
2
Jt ,
1=2
. (3)
This proves the lemma.
Lemma 2.4. If , > : > 2.
2
1
and > .
2
, then the number of integers _ .
divisible by some prime _ , with 1( 1) < : is at most . exp((1
o(1)) log ).
Proof. The number of such integers is at most
X
ypx;P.pC1/<z
.
_ .
X
ymx;P.mC1/<z
1
m
. (4)
where and m respectively run over primes and integers satisfying the described
conditions. By partial summation, we nd that the last sum is at most
(,. :)
,
Z
2x
y
(t. :)Jt
t
2
. (5)
Since we have (t. :),t < exp((1 o(1)) log ) uniformly for t ,. 2.|
by Lemma 2.1, the last sum in (4) can be bounded from above by
exp((1 o(1)) log )
1
Z
2x
y
Jt
t
!
. (6)
196 Tomohiro Yamada
This integral is O(.
1
) = O(exp ) since, by assumption, .
2
< . Thus we obtain
.
X
y<mx;P.mC1/<z
1
m
= . exp((1 o(1)) log ). (7)
3 Proof of Theorem 1.1
Let ,. : be real numbers and put u = log ., log , and = log ,, log :. We
choose ,. : later so that . > , > : > .
2
1
. > .
2
and u. . ,. : tend to innity as
. does so.
Let
S
1
= {n [ n _ .. 1(n) _ ,} (8)
and
S
2
= {n [ n _ ..
a
[ n for some . a with
a
_ :. a _ 2}. (9)
We immediately obtain #S
1
= O(. exp((1 o(1))ulog u)) by Lemma 2.1
and #S
2
= O(.,:
1=2
) by Lemma 2.3.
Denote by S
3
the set of integers n _ . not in S
1
L S
2
which can be written
in the form m, where is a prime _ ,, m is an integer not divisible by and
(o(m). 1) is divisible by some prime q _ :.
Let n be an integer in S
3
and write n = m in the above way. Then q [ o(r
a
)
for some prime r dividing m and some integer a with r
a
[[ m. Since q _ :, we
have a = 1 by the assumption n , S
2
. Hence, m is divisible by some prime r
congruent to 1 (mod q).
Since m _ .,, the number of integers n satisfying n = m. q [ (o(m). 1)
for some q _ : is at most
X
qz
X
p1 .mod q/
X
r1 .mod q/
.
r
_
X
qz
c..
2
2
q
2
_
c..
2
2
:
. (10)
by Lemma 2.2. Hence, we have #S
3
= O(..
2
2
,:).
Denote by S
4
the set of integers n _ . divisible by some prime _ , with
1( 1) < : or q
2
[ ( 1) for some q _ :. By Lemma 2.4, the num-
ber of integers n _ . divisible by some prime _ , with 1( 1) < : is
O(. exp((1 o(1)) log )). The number of integers n _ . divisible by some
prime with q
2
[ ( 1) for some q _ : is at most
.
X
qz
X
p1 .mod q
2
/
1
_ c..
2
X
qz
1
q
2
_
c..
2
:.
1=3
1
.
:
. (11)
by the assumption that : > .
2
1
.
Linear Equations Involving Iterates of o(N) 197
Combining these estimates yield #S
4
= O(.(1,:exp((1o(1)) log ))).
We may assume that at least one of and C is nonzero since the equation
o(o(N)) To(N) CN = 0 does not hold if exactly one of , T and C is
nonzero. Now let us denote by T the set of (2: : T : C)-perfect numbers n _ .
belonging to none of S
i
(i = 1. 2. 3. 4). We assume that n T . Since n , S
1
LS
2
,
we have 1(n) > , and 1(n)
2
n. Thus n can be expressed as n = m, > ,
and m. Now it follows from n , S
3
that (o(m). 1) has no prime factor
_ :. Let T
m
denote the set of such integers. Moreover, we write 1 = N
1
N
2
in the way 1(N
1
) < : _ (N
2
) and divide each of T
m
into sets T
m;N
1
according
to the value of N
1
. Since n , S
4
, we have 1 is divisible by some prime Q _ :
exactly once. By the denition of N
1
. N
2
, we have N
2
= N
3
Q and T
m;N
1
can be
covered by sets T
m;N
1
;N
3
according to the value of N
3
.
We shall show each T
m;N
1
;N
3
consists at most one element. We have o(n) =
o(m)(1) and o(o(n)) = o(M
1
N
1
)o(M
2
)o(N
2
), where M
1
. M
2
are uniquely
determined by o(m) = M
1
M
2
, 1(M
1
) < : _ (M
2
). Furthermore, noting
that Q does not divide ( 1),Q by the assumption that n , S
4
, we obtain
o(N
2
) = o(N
3
)(Q 1). Since o(o(n)) To(m)( 1) Cm = 0, we
have
o(M
1
N
1
)o(M
2
)o(N
3
)(Q1) (To(m) Cm)N
1
N
3
Q Cm = 0. (12)
Denoting
C
1
= o(M
1
N
1
)o(M
2
)o(N
3
). C
2
= (To(m) Cm)N
1
N
3
. C
3
= Cm. (13)
this equation can be written
(C
1
C
2
)Q = (C
3
C
1
). (14)
Since C _ 0, we have C
1
= C
3
and therefore Q can be uniquely determined as
Q =
C
3
C
1
C
1
C
2
. (15)
This is the crucial point where we use the assumption C _ 0. The uniqueness
of Q yields that #T
m;N
1
;N
3
_ 1 for any m. N
1
. N
3
.
By the denition of T
m;N
1
;N
3
, each element of T must belong to T
m;N
1
;N
3
for
at least one triple (m. N
1
. N
3
). Since N
1
N
3
= ( 1),Q _ (.,m 1),: _
2.,(m:), we have
#T _
X
mx=y
X
N
1
X
N
3
1 _
X
mx=y
2.
m:
_
c..
1
:
_
c.
:
1=2
. (16)
198 Tomohiro Yamada
Now we conclude that the number of (2: . T)-perfect numbers _ . is at most
#S
1
#S
2
#S
3
#S
4
#T , which is
O(.(:
1=2
exp((1 o(1))ulog u) exp((1 o(1)) log ))) (17)
by those estimates for #S
i
s and #T given above.
In order to search the optimal estimate, we put log : = c
1
.
1=3
1
.
2=3
2
. log , =
c
2
.
2=3
1
.
1=3
2
and we have ulog u = (1 o(1))c
1
2
.
1=3
1
.
2=3
2
and log = (1
o(1))(c
2
,c
1
).
1=3
1
.
2=3
2
. We see that max{c
1
,2. c
2
,c
1
. 1,c
2
}_ 2
1=3
with the equal-
ity attained when we choose c
1
= 2
2=3
. c
2
= 2
1=3
. This choice gives the desired
estimate. This completes the proof of Theorem 1.1.
We remark that (2: 1 : (1 1)o(1) : 1o(1))-perfect numbers given in
Theorem 1.2 correspond to the case m = N
1
= 1. C = o(1) = (T C)1.
Acknowledgments. The author thanks Prof. Florian Luca for his advice on
exceptional triples (. T. C).
References
[1] N. G. de Bruijn, On the number of positive integers _ . and free of prime factors
> ,, Nederl. Akad. Wetensch. Proc. Ser. A 54 (1951), 5060.
[2] G. L. Cohen and H. J. J. te Riele, Iterating the Sum-of-Divisors Function, Experiment.
Math. 5 (1996), 91100.
[3] I. Ktai and M. V. Subbarao, Some further remarks on the and o functions, Annales
Univ. Sci. Budapest. Sect. Comput. 26 (2006), 5163.
[4] I. Ktai and M. Wijsmuller, On the iterates of the sum of unitary divisors, Acta Math.
Hungar. 79 (1998), 149167.
[5] C. Pomerance, On multiply perfect numbers with a special property, Pacic J. Math.
57 (1975), 511517.
[6] C. Pomerance, On the congruences o(n) a (mod n) and n a (mod (n)), Acta
Arith. 26 (1975), 265272.
[7] D. Suryanarayana, Super perfect numbers, Elem. Math. 24 (1969), 1617.
[8] G. Tenenbaum, Introduction to Analytic and Probabilistic Number Theory, Cam-
bridge University Press, 1995.
Author information
Tomohiro Yamada, Department of Mathematics, Kyoto University,
Kyoto, 606-8502, Japan.
E-mail: tyamada@math.kyoto-u.ac.jp
Heron Sequences and Their Modications
Paul Yiu, K. R. S. Sastry and Shanzhen Gao

Abstract. We study sequences whose consecutive terms determine Heron triangles.
With a few exceptions, we show that given two positive integers u < v, there is an
innite sequence every three consecutive terms of which determine a Heron triangle. We
also construct arbitrarily long sequences every three consecutive terms of which are the
sides of a Heron triangle.
Keywords. Heron triangles, Pell equations, Fibonacci numbers.
AMS classication. 11D09, 51M04, 11B99.
1 Heron Sequences
Heron triangles are triangles with integer sides and integer areas. They have been
studied since ancient times. For some recent studies, see [1, 2, 3, 5, 6, 7]. It is clear
that the angles of a Heron triangle all have rational sines and cosines. Such are
called Heron angles, see for example [4]. Conversely, if the angles of a triangle
are all Heron angles, then an appropriate magnication yields a Heron triangle.
In this paper we study natural number sequences associated with Heron triangles.
One most natural and interesting question is whether there is an innite increasing
sequence in which every three consecutive terms are the sidelengths of a Heron
triangle. While we do not know the answer to this question, we can prove that
such a sequence can be arbitrarily long. The proof makes use of the Fibonacci
numbers F
n
dened recursively by F
nC1
D F
n
CF
n1
, F
1
D 1, F
2
D 1.
Theorem 1. Given an integer n 3, there is a sequence
a
1
; a
2
; : : : ; a
n
every three consecutive terms of which are the sides of a Heron triangle.
Proof. Let be a Heron angle such that F
nC1
<

2
. Consider the points
P
k
D .cos 2F
kC2
; sin 2F
kC2
/;
Sequences from The Encyclopedia of Integer Sequences discussed in this paper: A003500,
A103977, A103974, A104009, A103772, A104008.
200 Paul Yiu, K. R. S. Sastry and Shanzhen Gao
for k D 1; : : : ; n C 1. These are rational points on the unit circle. For k D
1; : : : ; n, the length of the chord P
k
P
kC1
is
a
k
D 2 sin
2F
kC3
2F
kC2
2
D 2 sin F
kC1
2 Q:
Since F
nC1
<

2
, the sequence of rational numbers a
1
; a
2
; : : : ; a
n
is strictly
increasing. We claim that every three consecutive terms are the sides of a triangle
with rational area. To see this, note that for k D 3; : : : ; n, two of the sides of
triangle P
k2
P
k1
P
k
have lengths a
k2
and a
k1
. The length of the third side
P
k2
P
k
is
2 sin
2F
kC2
2F
k
2
D 2 sin F
kC1
D a
k
:
Since the vertices of the triangle P
k2
P
k1
P
k
are on the unit circle, its area
is
a
k2
a
k1
a
k
4
, a rational number. By clearing denominators, we obtain an n-
term sequence of integers every three consecutive terms of which form a Heron
triangle.
The numbers realizing the Heron triangles in the above construction increase
very rapidly in n. By a routine computer search beginning with two positive
integers < 1000, we have found that the longest Heron sequence contains 9 terms.
There are three such sequences. One of these is
sides 60 275 325 500 525 697 746 1345 1797
area 4950 41250 78750 130872 175644 175644 452844
This sequence has been recorded in the Encyclopedia of Integer Sequences as
sequence A134587.
The other two 9-term sequences are obtained by multiplying the sides in this
sequence by 2 and 3 respectively. Again, beginning with two integers < 1000,
there are two 8-term modied Hereon sequences:
sides 445 485 850 1095 1435 2516 2691 3505
area 80100 197100 464100 165648 1782270 3369960
sides 825 975 1500 1575 2091 2238 4035 5391
area 371250 708750 1177848 1580796 1580796 4075596
Heron Sequences and Their Modications 201
2 Modied Heron Sequences
A common way to construct Heron triangles is to choose three integers u, v, w
such that
uvw.u Cv Cw/ D 4
2
(1)
for an integer 4. The triangle with sides
a D u Cv; b D u Cw; c D v Cw
is a Heron triangle with area 4. By a modied Heron sequence we mean an
increasing sequence .u
n
/ of positive integers such that every three consecutive
terms u
n2
, u
n1
and u
n
determine a Heron triangle with sides
a
n
D u
n2
Cu
n1
; b
n
D u
n2
Cu
n
; c
n
D u
n1
Cu
n
;
and integer area 4
n
. While it is not a priori obvious that a Heron sequence may
be innite, we show that, apart from a few exceptions, an innite modied Heron
sequence always results if one begins with two distinct positive integers.
Theorem 2. Let u < v be positive integers such that
.u; v/ .1; 4/; .1; 9/; .2; 8/; .2; 18/; .4; 16/: (2)
There is an innite modied Heron sequence .u
n
/ satisfying
(i) u
1
D u, u
0
D v;
(ii) u
n
is the smallest positive integer > u
n1
such that the triangle determined
by u
n2
, u
n1
, and u
n
is a Heron triangle.
Proof. Given two positive integers u < v, to nd an integer w which together
with u and v gives a Heron triangle we solve (1) by rewriting it in the form
x
2
uv y
2
D uv.u Cv/
2
(3)
by putting x D 24and y D 2w Cu Cv.
Case I. uv not equal to a square. It is readily seen that .x; y/ D .0; u C v/
is a solution of (3). If uv is not the square of an integer, then there is an in-
nite sequence of integer solutions generated by the fundamental solution of the
FermatPell equation
x
2
uv y
2
D 1: (4)
Namely, if .a; b/ is the smallest positive solution of (4), then there is a sequence
.x
n
; y
n
/ of solutions of (3) given by
x
nC1
y
nC1
!
D
a uvb
b a
!
x
n
y
n
!
;
x
0
y
0
!
D
0
u Cv
!
: (5)
Note that a 3 except when uv D 3. In particular, y
1
D a.uCv/ 3.uCv/.
We can determine an integer w fromy
1
D 2wCuCv which satises w uCv >
v. For .u; v/ D .1; 3/, the solution y
2
would give w > v.
Case II. uv equal to a square. If uv is a square, equation (3) reduces to
.2w Cu Cv/
2
D .u Cv/
2
Cz
2
(6)
for some integer z. This requires u C v to be a shorter side of a Pythagorean
triangle whose hypotenuse is 2w C u C v for some integer w > v. Now, it
is well known (and easy to verify) that every integer 3 is a shorter side of a
Pythagorean triangle. However, to ensure a hypotenuse of sufcient length, we
examine the details. Note that we require the hypotenuse to be 2wCuCv, which
has the same parity as u Cv.
(i) If u Cv D 4k for some integer k, consider the Pythagorean triangle
.4k; 2.k
2
1/; 2.k
2
C1//:
If we set the hypotenuse to be 2w C 4k, then w D .k 1/
2
. If k 6, w D
.k 1/
2
> 4k > v. The only pairs .u; v/ which do not satisfy this condition are
.u; v/ D .2; 18/ and .4; 16/.
(ii) If u Cv D 4k C2 for some integer k, consider the Pythagorean triangle
.2.2k C1/; .2k C1/
2
1; .2k C1/
2
C1/:
If we set 2w C 4k C 2 D .2k C 1/
2
C 1, then w D 2k
2
. If k 3, then
w D 2k
2
> 4k C2 > v. The only pairs .u; v/ which do not satisfy this condition
are .1; 9/ and .2; 8/.
(iii) If u Cv D 4k C1 for some integer k, consider the Pythagorean triangle
.4k C1; 4k.2k C1/; .2k C1/
2
C.2k/
2
/:
If we set 2wC4k C1 D .2k C1/
2
C.2k/
2
, then w D 4k
2
. If k 2, w D 4k
2
>
4k C1 > v. The only pair .u; v/ which does not satisfy this condition is .1; 4/.
(iv) If u Cv D 4k C3 for some integer k, consider the Pythagorean triangle
.4k C3; 4.k C1/.2k C1/; .2k C1/
2
C.2k C2/
2
/:
Heron Sequences and Their Modications 203
If we set 2w C 4k C 3 D .2k C 1/
2
C .2k C 2/
2
, then w D .2k C 1/
2
, which
always exceeds 4k C3 for k 1.
It follows that if uv is a square, then apart from the pairs in the list (2), there
exists an integer w > v such that u, v, w determine a Heron triangle.
For .u; v/ given in the list (2), it is routine to check that Pythagorean triangle
in (6) is an integer multiple of .5; 12; 13/. In each case, the corresponding value
of w is smaller than v.
Therefore, beginning with .u
1
; u
0
/ which is not one of the pairs in the list
(2) we can determine inductively for n 1, u
n
as the smallest integer > u
n1
such that u
n2
, u
n1
, u
n
determine a Heron triangle. The sequence so obtained
is innite.
Remark. The above constructions may or not give the smallest possible w in
question. However, the existence of the smallest w > v is guaranteed.
We conclude with some examples of modied Heron sequences each deter-
mined by two initial entries. The rst three have been recorded in the Encyclope-
dia of Integer Sequences
EIS number Modied Heron sequence
A134588 1; 2; 3; 10; 27; 98; 120; 327; : : :
A134589 1; 3; 12; 49; 108; 243; 624; : : :
A134590 2; 5; 63; 112; 140; 315; 364; : : :
6; 7; 8; 27; 70; 750; 972; : : :
The Heron triangles determined by the last sequence are
.13; 14; 15I 84/; .15; 34; 35I 252/; .35; 78; 97I 1260/;
.97; 777; 820I 34650/; .820; 1042; 1722I 302400/; : : : :
Acknowledgments. The authors thank the referee for instructive comments
leading to improvements on the presentation of this paper.
References
[1] A. Bremner, On Heron triangles, Ann. Math. Inform. 33 (2006), 1521.
[2] R. H. Buchholz and R. L. Rathbun, An innite set of Heron triangles with two rational
medians, Amer. Math. Monthly 104 (1997), 107115.
[3] R. H. Buchholz and R. L. Rathbun, Heron triangles and elliptic curves, Bull. Austral.
Math. Soc. 58 (1998), 411421.
[4] K. R. S. Sastry, Heron angles, Math. Comput. Ed. 35 (2001), 5160.
[5] K. R. S. Sastry, Two Brahmagupta problems, Forum Geom. 6 (2006), 301310.
[6] P. Yiu, Construction of indecomposable Heronian triangles, Rocky Mountain J. Math.
28 (1998), 11891202.
[7] P. Yiu, Heronian triangles are lattice triangles, Amer. Math. Monthly 108 (2001), 261
263.
Author information
Paul Yiu, Department of Mathematical Sciences, Florida Atlantic University,
Boca Raton, Florida 33431-0991, USA.
E-mail: yiu@fau.edu
K. R. S. Sastry, Jeevan Sandhya, Dodda Kalsandra Post,
Raghuvana Halli, Bangalore, 560 062, India.
E-mail: yiu@fau.edu
Shanzhen Gao, Department of Mathematical Sciences, Florida Atlantic University,
Boca Raton, Florida 33431-0991, USA.
E-mail: sgao2@fau.edu
List of participants
Abebe, Fisseha
Anderson, Derrick
Andrews, George
Bachraoui, Mohamed El
Bagdasaryan, Armen
Banks, William
Barrientos, Christian
Bayless, Jonathan
Benevides, Fabricio
Siqueira
Bergelson, Vitaly
Boncek, John
Boumenir, Amin
Boylan, Matthew
Caldwell, Chris
Caneld, Rod
Chan, Tsz Ho
Chen, Fang
Corcino, Roberto B.
Dominy, Morgan
Dos Santos, Carlos
Pereira
Eichhorn, Dennis
Finch, Carrie
Flowers, Tim
Fraenkel, Aviezri
Franze, Craig
Freeman, J. M.
Gao, Shanzhen
Garth, David
Goldston, Dan
Gong, Ke
Grekos, Georges
Hadad, Udi
Hasbun, Javier
Hegarty, Peter
Hindman, Neil
Hoffman, Fred
Hooshmand, M. H.
Hopkins, Brian
Ionescu, Eugen
James, Kevin
Jungic, Veselin
Kang, Jeong-Hyun
Keith, William J.
Khodkar, Abdollah
Kolitsch, Louis
Komatsu, Takao
Kra, Bryna
Landman, Bruce
Larsson, Urban
Le, Thai Hoang
Leach, David
Lee, Jaewoo
Lee, Joon Yop
Li, Xian-Jin
Ljujic, Zeljka
Luca, Florian
Lyall, Neil
Ma, Jianmin
Matthews, Gretchen
Mawi, Henok
McKay, Neil
Myers, Kellen
Nagle, Brendan
Nathanson, Melvyn
Nguyen, Van Minh
Nowakowski, Richard
OBryant Kevin
Ono, Ken
Orosz, Brooke
Palmer, Joseph
Pandey, Ram Krishna
Pollack, Paul
Pomerance, Carl
Riet, Ago-Erik
Robertson, Aaron
Robinson, David
Ruskey, Frank
Rutherfoord, Vermont
Schmid, Wolfgang
Savage, Carla
Sellers, James
Siegel, Aaron
Sills, Andrew
Silva, Jorge Nuno
Silva, Manual
Smith, Matthew
Stanica, Pantelimon
Stewart, Fraser
Swisher, Holly
Szalay, Laszlo
Ta, Ha
Thorne, Frank
Tuan, Vu
Vasilyeva, Vira
Ventullo, Kevin
Vinh, Le Anh
Vu, Van
Walsh, Gary
Wheeler, Jeffery
Williamson, Kevin
Xu, Rui
Yamada, Tomohiro
Ye, Jing
Zamberlan, Anthony

Combinatorial Number Theory

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Combinatorial Number Theory

Загружено:

Авторское право:

Доступные форматы

Combinatorial Number

Theory: Proceedings of the

Walter de Gruyter Berlin New York

1,2, if 0 _ 0 < 1,4,

does not exist, if 0 _ 0 < 1,4,

|}. We want to nd some a such that the fractional part of

for some integer k and 0 _ 0 _ 4. This

a = ka. Hence, with 1 =

D(n) = {(. ,) mod n : (.. ,) H

S(n) = {(. ,) mod n : (.. ,) H

S(n), in conjunction with the inequalities

S(n) we have the inclusion C

and 1 _ [.[. [,[ _ l [[ 1. Since n

be such that 1 S. We need to show

. Since S is treelike, every prex of a word

Cech compactication. If the

() = sup{ R : (Vk N)(Jn _ k)(Ja N)

is a two sided ideal of N, one has that 1(N) _ ^

(C) > 0. The following result of

. Let l N and for each i {1. 2. . . . . l}, let

satises the conclusion of Theorem 1.2. It is a result

but C misses the closure of the smallest ideal of N and in

and central sets, that is members of idempotents in 1(N).

() > 0 by [8, Theorem 3.11]. If is piecewise syndetic, then by [8,

() arbitrarily close to 1) such that no

() = 0 and there exist functions : P

() = 0. Notice that for any . and m in N,

a where 0 _ a _ m 1. Then cp(n. 1) =

, and the unique maximal ideal of O

the valuation of Q normalized so that [ [

given in Section 2.2 of [11]. It is trivial

for nite places . We have order([

for all nite places of Q. Let

:= the ordinary Lebesgue measure on the real line if

:= that measure for which O

gets measure 1 if is nite.

is the unique Haar measure on A

is in the Schwartz space S(R) if is the innite place of Q, and

), the space of locally constant and compactly supported

is unramied for every , 1, and c() = 1 for all Q

= 0 we do not include this prime in the set 1. By

to be one when is a real place of k and to be two when is an

be the variable in the half space R

taken over all innite places of k. Let N =

. It follows from Theorem 3 in

being the standard Haar measure on Q

given as in Section 2.3 of

and has ramication

) = () for all , 1 and

for all j , 1 _ j _ i 3. In particular, .S/ = i 3.

() is even, and each positive part is even and = (), and

() is odd, and each positive part is even and = () 1, and

() is odd, and each positive part is odd and = (), and

contains an odd positive part m (which may be repeated), exactly m 1

there are exactly

there are exactly

there are exactly

there are exactly

. Let (:) be the classical modular

n>0 and n2Z

), as , 0, for some R. If e() := e

., log .) for a suitable constant a

. log log ., log .) is proved conditionally on the Riemann Hypothesis

denote the product of the primes up to .. For nonsquare integers n coprime to