Professional Documents
Culture Documents
MONTHLY
VOLUME 118, NO. 9 NOVEMBER 2011
NOTES
REVIEWS
Shai Simonson
Rediscovering Mathematics is an eclectic collection
of mathematical topics and puzzles aimed at
talented youngsters and inquisitive adults who
want to expand their view of mathematics. By
focusing on problem solving, and discouraging
rote memorization, the book shows how to learn
and teach mathematics through investigation,
experimentation, and discovery. Rediscovering
Mathematics is also an excellent text for training
math teachers at all levels.
7RSLFVUDQJHLQGLIoFXOW\DQGFRYHUDZLGHUDQJHRIKLVWRULFDOSHULRGV
with some examples demonstrating how to uncover mathematics in
everyday life, including:
v QXPEHUWKHRU\DQGLWVDSSOLFDWLRQWRVHFXUHFRPPXQLFDWLRQ
over the Internet,
v WKHDOJHEUDLFDQGFRPELQDWRULDOZRUNRIDPHGLHYDO
mathematician Rabbi, and
v DSSOLFDWLRQVRISUREDELOLW\WRVSRUWVFDVLQRVDQGHYHU\GD\OLIH
MONTHLY
VOLUME 118, NO. 9 NOVEMBER 2011
EDITOR
Daniel J. Velleman
Amherst College
ASSOCIATE EDITORS
William Adkins Jeffrey Nunemacher
Louisiana State University Ohio Wesleyan University
David Aldous Bruce P. Palka
University of California, Berkeley National Science Foundation
Roger Alperin Joel W. Robbin
San Jose State University University of Wisconsin, Madison
Anne Brown Rachel Roberts
Indiana University South Bend Washington University, St. Louis
Edward B. Burger Judith Roitman
Williams College University of Kansas, Lawrence
Scott Chapman Edward Scheinerman
Sam Houston State University Johns Hopkins University
Ricardo Cortez Abe Shenitzer
Tulane University York University
Joseph W. Dauben Karen E. Smith
City University of New York University of Michigan, Ann Arbor
Beverly Diamond Susan G. Staples
College of Charleston Texas Christian University
Gerald A. Edgar John Stillwell
The Ohio State University University of San Francisco
Gerald B. Folland Dennis Stowe
University of Washington, Seattle Idaho State University, Pocatello
Sidney Graham Francis Edward Su
Central Michigan University Harvey Mudd College
Doug Hensley Serge Tabachnikov
Texas A&M University Pennsylvania State University
Roger A. Horn Daniel Ullman
University of Utah George Washington University
Steven Krantz Gerard Venema
Washington University, St. Louis Calvin College
C. Dwight Lahr Douglas B. West
Dartmouth College University of Illinois, Urbana-Champaign
Bo Li
Purdue University
EDITORIAL ASSISTANT
Nancy R. Board
NOTICE TO AUTHORS Proposed problems or solutions should be sent to:
The MONTHLY publishes articles, as well as notes and DOUG HENSLEY, MONTHLY Problems
other features, about mathematics and the profes- Department of Mathematics
sion. Its readers span a broad spectrum of math- Texas A&M University
ematical interests, and include professional mathe- 3368 TAMU
maticians as well as students of mathematics at all College Station, TX 77843-3368
collegiate levels. Authors are invited to submit arti-
cles and notes that bring interesting mathematical
In lieu of duplicate hardcopy, authors may submit
ideas to a wide audience of MONTHLY readers.
pdfs to monthlyproblems@math.tamu.edu.
The MONTHLY’s readers expect a high standard of ex-
position; they expect articles to inform, stimulate,
challenge, enlighten, and even entertain. MONTHLY Advertising Correspondence:
articles are meant to be read, enjoyed, and dis- MAA Advertising
cussed, rather than just archived. Articles may be 1529 Eighteenth St. NW
expositions of old or new results, historical or bio- Washington DC 20036
graphical essays, speculations or definitive treat-
ments, broad developments, or explorations of a Phone: (877) 622-2373
single application. Novelty and generality are far E-mail: tmarmor@maa.org
less important than clarity of exposition and broad
appeal. Appropriate figures, diagrams, and photo- Further advertising information can be found online
graphs are encouraged. at www.maa.org
Notes are short, sharply focused, and possibly infor- Change of address, missing issue inquiries, and
mal. They are often gems that provide a new proof other subscription correspondence:
of an old theorem, a novel presentation of a familiar MAA Service Center, maahq@maa.org
theme, or a lively discussion of a single issue.
All at the address:
Beginning January 1, 2011, submission of articles and
notes is required via the MONTHLY’s Editorial Man- The Mathematical Association of America
ager System. Initial submissions in pdf or LATEX form 1529 Eighteenth Street, N.W.
can be sent to the Editor-Elect Scott Chapman at Washington, DC 20036
Abstract. S. Narayana Aiyar was chief accountant at the Madras Port Trust office, where Ra-
manujan worked as a clerk in 1912–1914. In this article, a short biography of S. Narayana
Aiyar is given, his mathematical contributions are discussed, and his personal and mathemat-
ical relationships with Ramanujan are examined and emphasized.
1. INTRODUCING S. NARYANA AIYAR. In June 2009, the author had the priv-
ilege of participating in a conference commemorating K. Venkatachaliengar on the
centenary of his birth. For many years, more than any other Indian mathematician,
Venkatachaliengar had kept the spirit of Ramanujan’s mathematics alive in his home-
land, and so it was fitting that a conference in his memory be held on the beautiful
campus of Infosys in Bangalore. At this five-day meeting, a film entitled “God, Zero
and Infinity” about the life of Ramanujan was screened for the delegates. At one partic-
ular juncture in the film, I was saddened to hear one of the narrators condescendingly
proclaim that S. Narayana Aiyar, chief accountant at the Madras Port Trust office, was
an “amateur mathematician,” who was unable to appreciate the work of his employee,
S. Ramanujan. Since S. Narayana Aiyar was not able to defend himself from this un-
charitable remark, this listener decided to construct a defense for him.
Readers familiar with the life of Ramanujan might recall that after receiving a
monthly stipend from R. Ramachandra Rao for more than one year, in February, 1912,
Ramanujan wrote the Madras Port Trust office seeking employment as a clerk. Having
obtained a position at the Port Trust office, Ramanujan was supervised by S. Narayana
Aiyar. More than any other person in India during the years 1910–1914 when Ramanu-
jan lived in Madras prior to his departure for England in March, 1914, S. Narayana
Aiyar was able to appreciate Ramanujan’s work. It has been related that during the
evenings and into the wee hours of the mornings, S. Narayana Aiyar and Ramanujan
worked together on mathematics. Of course, we are not claiming that S. Narayana Ai-
yar was anywhere near the equal of Ramanujan in mathematical ability, but clearly
Ramanujan had respect for him, both as chief accountant at the Madras Port Trust
office and as a mathematician. Since many readers will have some knowledge and
interest in Ramanujan and his work, but will have little familiarity with S. Narayana
Aiyar and his influence on Ramanujan, in this brief article we provide a few details
about the life of S. Narayana Aiyar. Some of our narration can be found in the author’s
article in [9, pp. 97–98]. However, in contrast to our account in [9], we also discuss his
mathematical contributions.
Indian names when translated into English often bear different spellings. Another
rendition of Aiyar is Iyer, which is the preferred family spelling. Since Aiyar was used
in Narayana Aiyar’s mathematical publications, we adhere to this spelling.
traffic manager and then became the first Indian to officiate as chairman of the Madras
Port Trust. Upon assuming the post, he gathered together all of the files connected with
Ramanujan and transferred them to the National Archives in New Delhi. Copies of
these papers were secured for the present author by Nobel laureate S. Chandrasekhar,
and many are reproduced in the author’s book Ramanujan: Letters and Commen-
tary, coauthored with R. A. Rankin [7]. Narayana Aiyar’s son, Subbanarayanan, also
joined the Madras Port Trust, serving as assistant section master and retiring as office
manager.
One incident in Narayana Aiyar’s life is often related with amusement by his de-
scendants. Narayana Aiyar’s wife had arranged a prospective alliance for her younger
sister and requested that Narayana Aiyar visit the young man and give his valued opin-
ion. Accordingly, Narayana Aiyar called on the prospective bridegroom and asked him
but one question. Satisfied with the reply, he gave his approval for the marriage. The
question he asked was, “What is the value of π ?”
Narayana Aiyar was known for his simplicity bordering on austerity. Although he
could have afforded a more comfortable mode of transportation, he traveled to and
from work by tramcar, the ordinary transport for common people. He did not even
own a radio, although in the evenings he often visited his daughter Kamakshi to listen
to the English news on her radio. He insisted on absolute silence while taking food, as
prescribed in the scriptures. He was religious and scrupulously honest, had an indepen-
dent mind, and was a (silent) champion of Indian independence. Although he suffered
from a mild case of diabetes, he took good care of his health and lived a normal life
until shortly before he died on January 17, 1937.
As indicated above, after working at the Madras Port Trust office during the day,
Narayana Aiyar and Ramanujan often worked on mathematics together at night. It is
interesting that Narayana Aiyar communicated to the Journal of the Indian Mathe-
matical Society two brief papers summarizing some of Ramanujan’s discoveries in the
theory of prime numbers, and some applications of his “master theorem” for evaluat-
ing integrals [3, 4]. By the time these two papers were communicated, Ramanujan had
already published three papers and several problems in the Journal of the Indian Math-
ematical Society, and so one might naturally ask why Ramanujan did not communicate
these results himself to the Journal of the Indian Mathematical Society. The answer
5. EPILOGUE. Except for a few family pictures, only one photograph of S. Na-
rayana Aiyar is extant. Narayana Aiyar was one of the delegates to the meeting of the
Indian Mathematical Society in Bombay in 1919, and so can be seen in this group
photo. This photograph is extremely valuable, because many important people in the
life of Ramanujan are present. These include (numbering from the left) S. Narayana
Aiyar (3, first row), P. V. Seshu Aiyar (4, first row), V. Ramaswami Aiyar (5, sec-
ond row), R. Ramachandra Rao (10, second row), and S. R. Ranganathan (2, third
row). Ranganathan wrote the first book biography of Ramanujan in English [15]. He
was the librarian at the University of Madras and is internationally recognized today
as one of the founders of modern library science. Also mentioned in this article are
M. T. Naraniengar (7, second row) and S. Narayanan (9, third row). For a recent ac-
count of the history of the Indian Mathematical Society and of mathematics in India
during the past century, see the article by J. W. Dauben and R. Parikh [11].
We are grateful to V. Viswanathan for providing detailed information about his
grandfather, S. Narayana Aiyar. Viswanathan’s mother was Kamakshi, who entered
the world in 1907 and possibly inspired her father to resume intense mathematical ac-
tivity. Narayana Aiyar’s son, N. Subbanarayanan [16], has written about his father and
his father’s relationship with Ramanujan. In particular, he writes that his father and
Ramanujan exchanged slates prior to the latter’s departure for England. In those days
paper was expensive, and so, in particular, school children performed all of their work
on slates. Before departing for England, Ramanujan did almost all of his mathematics
on his slate. The bestowing of his slate to Narayana Aiyar clearly indicates the respect
that Ramanujan had for him. Subbanarayanan tells that after Ramanujan’s death he
often went to the library to fetch books on elliptic functions for his father. Obviously,
his love for the subject was not abated by the demise of Ramanujan.
Jim Tattersall supplied copies of all the problems and solutions that Narayana Aiyar
submitted to The Educational Times, and also notified the author about the Progress
Reports. Moreover, he informed us of [17], from which we took much of our infor-
mation about The Educational Times. We are enormously beholden to Paul Yiu for
providing solutions to several of Narayana Aiyar’s problems in geometry, and to Mike
Hirschhorn for solutions to some of Narayana Aiyar’s difficult problems on integrals.
We are also grateful to Antreas Hatzipolakis, Clark Kimberling, and John Wetzel for
their comments. The referees also offered several helpful suggestions.
1. S. Narayana Aiyar, Note on expansion in series, Math. Quest. Sols. from the Educational Times 13 (1907)
49–50.
2. , On the integration of sin mx
sin x , J. Indian Math. Club 1 (1909) 9–10.
3. , The distribution of primes, J. Indian Math. Soc. 5 (1913) 60–61.
4. , Some theorems in summation, J. Indian Math. Soc. 5 (1913) 183–186.
5. , Some infinite determinants, J. Indian Math. Soc. 7 (1915) 51–55.
6. , A small theorem, J. Indian Math. Soc. 8 (1916) 181.
7. B. C. Berndt and R. A. Rankin, Ramanujan: Letters and Commentary, American Mathematical Society,
Providence, RI, 1995; London Mathematical Society, London, 1995.
8. , The books studied by Ramanujan in India, Amer. Math. Monthly 107 (2000) 595–601. http:
//dx.doi.org/10.2307/2589114
9. , Ramanujan: Essays and Surveys, American Mathematical Society, Providence, RI, 2001; Lon-
don Mathematical Society, London, 2001.
10. A. Cayley, An Elementary Treatise on Elliptic Functions, 2nd ed., 1895; reprinted by Dover, New York,
1961.
11. J. W. Dauben and R. Parikh, Beginnings of modern mathematics in India, Current Science 99 (2010)
300–322.
12. S. Ramanujan, On Question 330 of Professor Sanjana, J. Indian Math. Soc. 4 (1912) 59–61.
13. , Collected Papers, Cambridge University Press, Cambridge, 1927; reprinted by Chelsea, New
York, 1962; reprinted by the American Mathematical Society, Providence, RI, 2000.
14. , Notebooks (2 volumes), Tata Institute of Fundamental Research, Bombay, 1957.
15. S. R. Ranganathan, Ramanujan: The Man and the Mathematician, Asia Publishing, Bombay, 1967.
16. S. Subbanarayanan, My father and Ramanujan, in Ramanujan: Letters and Reminiscences, Vol. 1, P. K.
Srinivasan, ed., Muthialpet High School, Madras, 1968, 112–115.
17. J. J. Tattersall, S. L. McMurran, and F. Coughlin, Women and the ‘Educational Times’, Proc. Canad.
Soc. Hist. Philos. Math. 16 (2003) 250–260.
BRUCE C. BERNDT received an A.B. from Albion College, Albion, Michigan, in 1961, and a Ph.D. from
the University of Wisconsin, Madison, in 1966. He joined the faculty at the University of Illinois in 1967 after
one year at the University of Glasgow. Twenty-eight students have received their doctorates under his direction,
with six further students currently writing dissertations under him. Since 1977, he has devoted almost all of his
research efforts toward proving the claims left behind by Ramanujan in his earlier notebooks and lost notebook.
With George Andrews, he is currently preparing a projected five volumes on Ramanujan’s lost notebook, two
of which have been published to date.
Department of Mathematics, University of Illinois, 1409 West Green Street, Urbana, IL 61801, USA
berndt@illinois.edu
Abstract. This paper surveys a collection of results on finding special sets of vertices in graphs
with vertex partitions, all of which can viewed as models for “committee-choosing” problems.
ū y z̄ w̄ x ū
u
x̄
y ȳ
z̄
w
ȳ w z
Figure 1. The graph corresponding to the boolean formula A = (u ∨ y ∨ w) ∧ (ū ∨ y ∨ z̄) ∧ (w̄ ∨ x ∨ ū) ∧
(x̄ ∨ ȳ ∨ z̄) ∧ ( ȳ ∨ w ∨ z). The square vertices form an IT corresponding to the truth assignment u = x = T ,
y = z = F, and w can be either T or F.
What help can we then offer to the dean? The best we can do is to provide some
sufficient conditions for an IT to exist in a given vertex-partitioned graph that are not
too hard to check. In the next section we show that if the partition classes are big
enough compared to the maximum degree of the graph, then an IT always exists. In
Section 4 we describe a somewhat more complicated condition for the existence of an
IT, which nevertheless can be applied in certain circumstances. Here we will appeal to
Sperner’s lemma, a result from combinatorial topology.
A very optimistic dean might even want to form many disjoint committees, perhaps
even imagining a perfect world in which the entire faculty could be partitioned into
disjoint committees. This “happy dean” problem is captured by the notion of strong
colouring, which we discuss in Section 3. We end the paper with some remarks and
pointers to other related topics.
Theorem 2.1. Let G be a graph with a vertex partition. Suppose each partition class
has size at least 21(G). Then G has an IT.
Theorem 2.1 first appeared explicitly in [17], and has been applied in a number
of other settings, for example [2, 8, 11, 25]. It is best possible, as Szabó and Tardos
[29] gave constructions of graphs G with partition classes of size 21(G) − 1 that do
not have independent transversals (see also Bollobás, Erdős, and Szemerédi [10], Jin
[23], and Yuster [30] for earlier constructions for certain values of 1). A more precise
To see that Theorem 2.2 implies Theorem 2.1, let G be as in Theorem 2.1 and
suppose on the contrary that it has no IT. Remove edges one by one from G until the
resulting graph satisfies the assumptions of Theorem 2.2, and let S be the subset of
classes given by Theorem 2.2. The number of vertices that can be dominated by v(Z ),
a set of size at most 2|S| − 2, is at most (2|S| − 2)1, but G S contains 21|S| vertices.
Thus conclusions (2) and (3) cannot hold, so this contradiction shows that G must have
had an IT.
In fact it follows from the above proof that the set Z of edges is a matching, that is,
no two edges in Z share a vertex.
Figure 2. A partition into 3 committees (the white, grey, and black committees) for a faculty with 3 depart-
ments.
Let r and n be positive integers such that r divides n, and let G be a graph with n
vertices. We call G strongly r -colourable if for every partition of the vertex set V (G)
into parts Vk of size r , there exists a proper colouring of G with r colours with the
additional property that each Vk contains exactly one vertex of each colour. If r does
not divide n then we say G is strongly r -colourable if the graph obtained by adding
r ⌈n/r ⌉ − n isolated vertices to G is strongly r -colourable. It can be shown that if G is
strongly r -colourable then it is also strongly (r + 1)-colourable. The strong chromatic
number sχ(G) of a graph G is defined to be the minimum r such that G is strongly
r -colourable. This notion was introduced independently by Alon [6] and Fellows [13].
Thus the dean could be guaranteed a partition of the entire faculty G into com-
mittees provided the departments have size at least sχ(G) (and, of course, they all
have the same size). If the unlucky dean finds that the faculty is exactly the graph
given in [29] (see Section 2) with departments of size 21(G) − 1 but no independent
transversal, then not even one committee can be found, let alone a partition into com-
mittees. Thus for some graphs sχ(G) ≥ 21(G). It is conjectured that 21(G) is in
fact also an upper bound for every G, but the best known bounds for strong colouring
Proof. Let V1 , . . . , Vm denote the vertex classes of G (again we may remove any edge
that is inside a class). Assume without loss of generality that x ∈ V1 . Let G ′ be the
graph, with vertex classes Vi′ ⊆ Vi for 2 ≤ i ≤ m, obtained by removing V1 and all
neighbours of x from G. Note that each Vi′ is nonempty since |Vi | ≥ 21. If G ′ has an
IT T then T ∪ {x} is an IT of G as required. If G ′ has no IT then remove edges from G ′
until every remaining edge prevents an IT. Let e be an arbitrary edge of the resulting
graph H (which must exist since each class of H is nonempty). We apply Theorem 2.2
to H and e to obtain a subset S of classes and a set of edges Z of size at most |S| − 1
such that v(Z ) dominates HS . But v(Z ) can dominate at most 21|Z | ≤ 21(|S| − 1)
vertices, and HS contains at least 21|S| − 1 vertices. This contradiction shows that
H has an IT, and hence G has an IT containing x.
Finding the correct function of 1(G) that bounds sχ(G) from above seems to
be difficult (see Problem 4.14 in Jensen and Toft [22]). For example, it is not even
known whether sχ(G) ≤ 4 for every graph of maximum degree two. By [18] we know
sχ(G) ≤ 5 for such graphs.
Theorem 4.1. (Hall’s Theorem) A bipartite graph H with vertex classes A and X has
a matching of size |A| if and only if for every subset S of A we have |Ŵ(S)| ≥ |S|.
We may now state the main theorem of this section, which is from [4]. Note that
in contrast to Hall’s theorem this is only a one-way implication; for an if-and-only-if
version see Section 5 (where we also show that this theorem implies Hall’s theorem).
Theorem 4.2. Let G be a graph with a vertex partition into independent classes. Sup-
pose that for every subset S of classes, the graph G S contains an independent set I S
that is special for S. Then G has an IT.
The dean may be justifiably dubious that this theorem could possibly be useful,
as checking whether a partitioned graph satisfies the condition (of having a special
independent set for every subset of classes) looks hopelessly complicated. Thus to
convince the dean to read on, we first give a quick application of Theorem 4.2 to a well-
known combinatorial problem. Then in the following subsection we will describe the
proof, which uses Sperner’s lemma. We mention that another application of Theorem
4.2 (and in fact its original motivation) is an extension of Hall’s theorem to hypergraphs
(see [4]).
4.1. Application of Theorem 4.2. Our application is related to the famous “cycle
plus triangles” problem, popularised by Erdős in the 1980s, and finally solved by Fleis-
chner and Stiebitz [14] in 1992 and with a different proof by Sachs [28] in 1993. It asks
whether every union of a cycle C3k of length divisible by three, together with a set of k
disjoint triangles on the same vertex set, has a proper colouring with three colours (see
Figure 3). In the language of Section 3, this is the same as asking whether every C3k
is strongly 3-colourable. This question of Erdős was motivated by an earlier question
of Du, Hsu, and Hwang [12], who asked whether every C3k with an arbitrary vertex
partition into classes of size three has an IT. This problem was also unsolved until the
proofs of Fleischner and Stiebitz, and Sachs, of the stronger statement. Both of these
proofs are ingenious but quite difficult. Here we can give a solution to the question of
Du, Hsu, and Hwang that is almost immediate from Theorem 4.2, and in fact is more
general.
Theorem 4.3. Let G be a graph of maximum degree two, in which each cycle is of
length divisible by three. Then G has an IT with respect to any vertex partition into
classes of size at least three.
Proof. Fix a vertex partition, and remove any edge that is inside a class. Let S be any
subset of classes. Then the graph G S has at least 3|S| vertices, and its components are
cycles of length divisible by three and paths. Let I S be a maximum independent set in
G S that is spaced three apart. For example, I S could be formed by taking, from each
path component, an end vertex and then every third vertex starting from that end, and,
from each cycle component C3k , an independent set of size k containing every third
vertex. We claim that I S is special for S. To see this, let J be an arbitrary independent
set in G S of size |S| − 1. Since I S is spaced three apart, each vertex of J is adjacent
to at most one vertex of I S . Thus since |I S | ≥ |S|, there is a vertex v ∈ I S such that
J ∪ {v} is independent. Thus I S is special, and so by Theorem 4.2 we know that G has
an IT.
4.2. Proof of Theorem 4.2. First we recall Sperner’s lemma, a basic result from com-
binatorial topology. Suppose F is a triangulation of the (m − 1)-dimensional simplex
6m−1 . For each point x of F we denote by f (x) the face of 6m−1 containing x in its
interior. A labelling of the points of F with elements of {1, . . . , m} is called a Sperner
labelling if
• each vertex of 6m−1 receives a different label, and
• each point x of F receives the same label as some vertex of f (x).
An example of a Sperner labelling is shown in Figure 4. We call a simplex of the
triangulation fully-labelled if it receives all m labels on its vertices.
As is usually the case, the implication of Sperner’s lemma that is useful for us is that the
number of fully-labelled simplices is nonzero. Note that there are three fully-labelled
simplices in Figure 4.
For the proof of Theorem 4.2 we also need to know that certain special triangula-
tions of 6m−1 exist (see Figure 4). The 1-skeleton of a triangulation F is the graph
whose vertices are the points of F, and whose edges are the 1-dimensional simplices
of F.
1 2
1 2
1 2
3
2
1
2 1
2
2
3 2
2 3
Figure 4. A Sperner labelling of a triangulation F of 62 with the properties in Lemma 4.5.
Lemma 4.5. There exists a triangulation F of 6m−1 with the following properties.
(i) If x and y are adjacent in the 1-skeleton of F then one of f (x) and f (y)
contains the other,
(ii) if x has neighbours in the 1-skeleton of F on the boundary of f (x), then these
neighbours are the vertices of a simplex of F.
A proof of Lemma 4.5 can be found in [4]. A different construction was given in [3].
a A
b B
c C
d D
V1 α V2
δ
V3
Proof. We may now give the proof of Theorem 4.2. Let the graph G with vertex parti-
tion V1 ∪ · · · ∪ Vm be given (see Figure 5 for an example). Let F be the triangulation of
6m−1 given by Lemma 4.5, and let z 1 , . . . , z m denote the vertices of 6m−1 . We define a
function g that assigns to each point of F a vertex of G with the following properties.
b
D C
c
c A
A
d
d
A D
A
A
D
Figure 6. The triangulation F with the vertex g(z) of G assigned to each point z.
Our motivation here is as follows. Once we have found g, by (3), the labelling ℓ
is a Sperner labelling. (The labelling ℓ for the function g in Figure 6 is the labelling
in Figure 4.) Therefore by Theorem 4.4 there exists a simplex W of F that gets all m
labels on its vertices (e.g., the shaded triangle in Figure 6). By (2), the set of vertices
I = {g(x) : x ∈ W } is independent in G. By definition of ℓ, each v ∈ I lies in a distinct
class Vi . Therefore I is an IT of G as required. (I = {c, A, γ } in our example.)
Thus to finish the proof, we define g on the points x of F, in increasing order of
dim( f (x)). For the points in zero-dimensional faces (i.e., the vertices z 1 , . . . , z m of
6m−1 ) we choose an arbitrary vertex vi ∈ I{Vi } (which is nonempty by the assumption
of the theorem) and set g(z i ) = vi . This satisfies (1), satisfies (2) by Lemma 4.5(i), and
is consistent with (3).
Now suppose j ≥ 1 and we have defined g on all points in faces of dimension
smaller than j, and possibly some in faces of dimension j, such that (1)–(3) are satis-
fied (see Figure 7). Let x be in a face f of dimension j (e.g., the grey point in Figure 7),
and let S = {Vi : z i is a vertex of f }. (S = {V1 , V2 , V3 } in Figure 7.) By Lemma 4.5(i)
and (ii), if x has any neighbours in the 1-skeleton of F in faces of dimension smaller
than j, then the set U of such neighbours lies in the boundary of f and forms a simplex
of F. Thus |U | ≤ dim( f ) = |S| − 1 and by (2) J = {g(x) : x ∈ U } is independent in
G. (J = {b, γ } in Figure 7.) Thus J is a “small” independent set, and so the special in-
dependent set I S contains a vertex v such that J ∪ {v} is independent. (In our example
suppose I S = {c, d, A, D}. Then v = c is a suitable choice.) Then we set g(x) = v, so
(1) and (3) are satisfied. To check (2), if y is a neighbour of x in a smaller dimensional
face then (2) is satisfied by choice of v. If y is in a face of dimension j (e.g., the white
point in Figure 7) then by Lemma 4.5(i) it must be in f as well. Thus if g(y) has
b
D C
c
A
D
Figure 7. Defining g on the points of F.
already been defined then g(y) ∈ I S by (1) (g(y) = c in our example) and therefore is
not adjacent to g(x) in G. This completes the definition of g, and hence the proof.
REFERENCES
1. R. Aharoni and E. Berger, The intersection of a matroid and a simplicial complex, Trans. Amer. Math.
Soc. 358 (2006) 4895–4917. http://dx.doi.org/10.1090/S0002-9947-06-03833-5
2. R. Aharoni, E. Berger, and R. Ziv, Independent systems of representatives in weighted graphs, Combina-
torica 27 (2007) 253–267. http://dx.doi.org/10.1007/s00493-007-2086-y
3. R. Aharoni, M. Chudnovsky, and A. Kotlov, Triangulated spheres and colored cliques, Discrete Comput.
Geom. 28 (2002) 223–229.
4. R. Aharoni and P. Haxell, Hall’s theorem for hypergraphs, J. Graph Theory 35 (2000) 83–88. http:
//dx.doi.org/10.1002/1097-0118(200010)35:2<83::AID-JGT2>3.0.CO;2-V
5. N. Alon, The linear arboricity of graphs, Israel J. Math. 62 (1988) 311–325. http://dx.doi.org/10.
1007/BF02783300
6. , The strong chromatic number of a graph, Random Structures Algorithms 3 (1992) 1–7. http:
//dx.doi.org/10.1002/rsa.3240030102
7. , Problems and results in extremal combinatorics, Part I, Discrete Math. 273 (2003) 31–53. http:
//dx.doi.org/10.1016/S0012-365X(03)00227-9
8. N. Alon, G. Ding, B. Oporowski, and D. Vertigan, Partitioning into graphs with only small com-
ponents, J. Combin. Theory Ser. B 87 (2003) 231–243. http://dx.doi.org/10.1016/S0095-
8956(02)00006-0
9. R. Berke, P. Haxell, and T. Szabó, Bounded transversals in multipartite graphs (to appear).
10. B. Bollobás, P. Erdős, and E. Szemerédi, On complete subgraphs of r -chromatic graphs, Discrete Math.
13 (1975) 97–107. http://dx.doi.org/10.1016/0012-365X(75)90011-4
11. J. Britnell, A. Evseev, R. Guralnick, P. Holmes, and A. Maróti, Sets of elements that pairwise generate a
linear group, J. Combin. Theory Ser. A 115 (2008) 442–465. http://dx.doi.org/10.1016/j.jcta.
2007.07.002
12. D. Z. Du, D. F. Hsu, and F. K. Hwang, The Hamiltonian property of consecutive-d digraphs, Math. Com-
put. Modelling 17, no. 11 (1993) 61–63. http://dx.doi.org/10.1016/0895-7177(93)90253-U
13. M. Fellows, Transversals of vertex partitions in graphs, SIAM J. Discrete Math. 3 (1990) 206–215.
http://dx.doi.org/10.1137/0403018
14. H. Fleischner and M. Stiebitz, A solution to a colouring problem of P. Erdős, Discrete Math. 101 (1992)
39–48. http://dx.doi.org/10.1016/0012-365X(92)90588-7
15. M. Garey and D. Johnson, Computers and Intractability: A Guide to the Theory of NP-Completeness,
W. H. Freeman, New York, 1979.
16. P. Haxell, A condition for matchability in hypergraphs, Graphs Combin. 11 (1995) 245–248.
http://dx.doi.org/10.1007/BF01793010
17. , A note on vertex list colouring, Combin. Probab. Comput. 10 (2001) 345–348.
http://dx.doi.org/10.1017/S0963548301004758
18. , On the strong chromatic number, Combin. Probab. Comput. 13 (2004) 857–865.
http://dx.doi.org/10.1017/S0963548304006157
19. , A new bound on the strong chromatic number, J. Graph Theory 58 (2008) 148–158.
http://dx.doi.org/10.1002/jgt.20300
20. P. Haxell and T. Szabó, Odd independent transversals are odd, Combin. Probab. Comput. 15 (2006)
193–211. http://dx.doi.org/10.1017/S0963548305007157
21. P. Haxell, T. Szabó, and G. Tardos, Bounded size components—partitions and transversals, J. Combin.
Theory Ser. B 88 (2003) 281–297. http://dx.doi.org/10.1016/S0095-8956(03)00031-5
22. T. Jensen and B. Toft, Graph Coloring Problems, Wiley-Interscience, New York, 1994.
23. G. Jin, Complete subgraphs of r -partite graphs, Combin. Probab. Comput. 1 (1992) 241–250.
http://dx.doi.org/10.1017/S0963548300000274
24. , The number of complete subgraphs of equi-partite graphs, Discrete Math. 186 (1998) 157–165.
http://dx.doi.org/10.1016/S0012-365X(97)00190-8
25. M. Krivelevich, Almost perfect matchings in random uniform hypergraphs, Discrete Math. 170 (1997)
259–263. http://dx.doi.org/10.1016/S0012-365X(96)00310-X
PENNY HAXELL received her B.Math from the University of Waterloo in 1988 and her Ph.D. from the
University of Cambridge in 1993. She has been a faculty member in the Combinatorics and Optimization
department at the University of Waterloo since 1993. This article is based on an MAA invited address at the
AMS-MAA Joint Meetings in New Orleans in 2007. The author has no good excuse for taking this long to
write it.
Department of Combinatorics and Optimization, University of Waterloo, Waterloo Ont. Canada N2L 3G1.
pehaxell@math.uwaterloo.ca
Abstract. Classical dissections convert any planar polygonal region onto any other polygonal
region having the same area. If two convex polygonal regions are isoparametric, that is, have
equal areas and equal perimeters, our main result states that there is always a dissection, called
a complete dissection, that converts not only the regions but also their boundaries onto one
another. The proof is constructive and provides a general method for complete dissection using
frames of constant width. This leads to a new object of study: isoparametric polygonal frames,
for which we show that a complete dissection of one convex polygonal frame onto any other
always exists. We also show that every complete dissection can be done without flipping any
of the pieces.
(a) (b)
Figure 1. Dissection (a) of a triangle onto a rectangle; (b) of a rectangle onto another rectangle of prescribed
altitude. Dark lines show how the boundaries are transformed.
In Figure 2, the triangle and rectangle are isoparametric, but the dissection shown
is not complete because one boundary is not converted entirely onto the other. In fact,
there is no reason to expect that a complete dissection exists even though the figures are
isoparametric. Nevertheless, our first theorem reveals a surprising and profound result:
for any two isoparametric convex polygonal regions, a complete dissection always
exists. Moreover, the proof shows how to construct such a dissection.
http://dx.doi.org/10.4169/amer.math.monthly.118.09.789
2 2
6 6
Figure 2. Standard dissection converting a triangle onto an isoparametric rectangle.
Before proceeding further, the reader might try to find a complete dissection that
converts the triangle in Figure 3 onto the isoparametric isosceles trapezoid shown.
This is a simpler task than for the pair in Figure 2.
(a)
angle bisector
a A A
C A
B b B
(b) a b
B b C a C
Figure 4. Complete dissection converting (a) isosceles triangle onto isosceles trapezoid, and (b) any triangle
onto a trapezoid. In (a), the unshaded piece has been flipped. In (b) one piece is flipped and divided into two
right triangles, the smaller of which is flipped again.
Theorem 1. Any two isoparametric convex polygonal regions can be converted onto
one another by complete dissection.
Proof. Consider two isoparametric convex polygonal regions A and B. Figure 5 shows
an example with A triangular and B quadrilateral. This example displays all the essen-
tial features required in treating general convex isoparametric polygonal regions. The
method of proof is suggested by an oversimplified intuitive idea: Remove each bound-
ary and perform a standard dissection of the interior of A to produce the interior of B.
Then restore the two boundaries to obtain the complete dissection.
To make this intuitive idea rigorous, refer to Figure 6a, which shows a “frame” of
some constant width w protruding into each region along its boundary. Choose w small
enough so that the inner boundary of each frame will be a simple closed polygon with
the same number of sides. Each region now consists of two parts: the frame plus the
interior region surrounded by the frame. Keep in mind that:
The sum of areas, frame plus interior, is the same for both regions A and B.
Now unfold each frame at the outer vertices (thought of as hinges) and lay it out
horizontally, as shown in Figure 6b. To be specific, use angle bisectors to cut each
frame into trapezoidal pieces with isosceles triangular gaps between adjacent pieces.
(a)
w w
(b) w
Figure 6. Frames unfolded to form adjacent trapezoidal pieces with triangular gaps.
Next, use standard dissections (as in Figure 1) of the regions interior to the frames
in Figure 6a to convert them onto two rectangles with common altitude w (the frame
width), as indicated in Figure 7a. Each rectangle has area equal to that of the interior
that produced it, but, of course, the two rectangular areas are not equal to each other.
(a)
(b)
(c)
Figure 7. (a) Dissection of interior regions produces two rectangles of equal altitude, but not of equal area.
Removing shaded triangular pieces in (b) to fill the gaps as shown in (c) leaves two unshaded congruent
rectangles in (b).
From each of these rectangles, remove smaller triangles as needed (see Figure 7b)
to fill the triangular gaps in Figure 6b. This transforms the unfolded frames into two
congruent rectangles shown in Figure 7c, whose lower bases are the unfolded bound-
aries of A and B. By overlapping these rectangles we obtain their common dissection
(dissection 1) which also converts the full boundary of A onto the full boundary of B.
Obviously, these dissections can be done in many different ways. For example, we
can choose a different frame width w.
15
9 5
3
w= 2
4
12
8 4
w= 2
common dissection
Lemma 1. For any convex polygonal frame, its width w, area A, and total perimeter
P are related by the equation
A = 12 Pw. (1)
The proof follows easily by applying the area formula for adjacent trapezoids form-
ing the frame.
As an immediate consequence of Lemma 1 we have the following crucial result:
Corollary 1. All isoparametric convex polygonal frames have the same width.
Now return to Figure 8, which shows an unfolding of each frame into trapezoids of
constant altitude w = 2, followed by a dissection onto a rectangle of the same altitude
with two horizontal bases, the sum of whose lengths is the total perimeter of the frame.
Because the two frames are isoparametric, so are the two rectangles. At the bottom of
Figure 8, the dissected rectangles are superimposed to obtain a common dissection of
the two frames that converts their total boundaries, indicated by the heavy lines.
The same type of argument can be applied to any pair of isoparametric convex
polygonal frames to prove the following theorem.
Theorem 2. Any two convex isoparametric polygonal frames can be converted onto
one another by complete dissection.
Flipping a piece turns it into its mirror image, so the proof of Theorem 3 reduces
to showing that a complete dissection of a general polygonal piece onto its mirror
image can be done without flipping and in such a way that each edge of the polygon
is converted to the corresponding edge of the mirror image. Such a dissection requires
more than completeness, and we call it strongly complete. Now we will prove that
every polygon can be converted onto its mirror image by using strongly complete
dissection.
First dissect the polygonal piece into triangles, as illustrated in Figure 9. Then per-
form a strongly complete dissection on each triangle.
To do this, we can dissect each triangle into six right triangles as shown in Figure
10a, where now each right triangle has only one leg as part of the boundary of the
original triangle. This reduces the problem to that of dissecting a right triangle onto its
mirror image, without flipping, so that one leg gets converted onto its mirror image.
Figure 10b shows how a right triangle can be cut into two isosceles triangles. Rotate
one of them to convert the given leg as shown. This completes the proof of Theorem 3.
By combining Theorems 1 and 3 we conclude that any two isoparametric convex
polygons can be converted onto one another by complete dissection without flipping.
Examples. Figure 11 (taken from [1]) shows how easy it is to produce an endless
supply of incongruent isoparametric pairs in which boundaries are transformed onto
one another. In each case, a chord bisects the first region and one piece is flipped, thus
giving a simple but complete dissection of one onto the other. It is reassuring to realize
that every example of such a dissection can also be done without flipping.
isosceles triangle and parallelogram parallelogram and kite regular pentagon and hexagon
Figure 11. Isoparametric regions with boundaries converted onto one another.
Figure 13. Complete dissection of rectangular frame and isoparametric square frame.
w
w
w w
w w
w
w
w
outer edge length
r (r + 1) / 4 + (r + 6) / 2
w=2 w=2
Further examples. Figure 16a (borrowed from [1]) shows two isoparametric circular
sectors that are also isoparametric to the same rectangle. Figure 16b shows a circular
frame and an isoparametric square frame of the same width.
2 1
2 1 1
2 3 −1
4
1 3 +1
(a) (b) 4
Figure 16. Partially circular regions isoparametric to polygonal regions.
1 2 1
1 2
Figure 17. Radial slicing of two special circular sectors with their boundaries, rearranged differently to ap-
proximate the same isoparametric rectangle.
Figure 17 shows how the circular sectors in Figure 16a can be dissected into an
even number of radial slices that can be rearranged to form a figure approximating
the same rectangle. In one case the slices are arranged in horizontal layers, and in
the other case in vertical layers. As the number of slices increases without bound,
both approximations have the same rectangle as a limit, with the curved boundaries
becoming the rectangular boundaries, in the style of Archimedes. Figure 18 shows a
dissection of the circular frame in Figure 16b that approximates the square frame. The
Figure 18. Radial slices of a circular frame and its boundary, rearranged to approximate the isoparametric
square frame.
Figure 19. Regular polygonal frames and a circular frame, all isoparametric.
two frames in Figure 16b are members of an infinite family of isoparametric regular
frames of constant width w = 1, examples of which are shown in Figure 19.
Each inner regular n-gon in this family has an incircle of diameter dn , where
3π
dn = − 1,
n tan πn
In view of Max Dehn’s counterexample that settled Hilbert’s third problem, exten-
sions of our ideas to dissections of solids in 3-space are not always possible. Never-
theless, circumsolid shells, introduced in [3], have properties analogous to polygonal
frames. For example, all circumsolid shells have constant thickness (see Theorem 8a
in [3]). The analog of Lemma 1 is given by Theorem 10 in [3].
Frederickson’s book [4] gives an admirable introduction to the field of geometric
dissections, and contains a valuable bibliography of known results. Although there is a
vast literature on standard polygonal dissection, we were not able to find any references
relating to complete dissections of the type discussed in this paper.
ACKNOWLEDGMENTS. The authors wish to thank the referees for valuable suggestions that improved
this paper. They also wish to thank Dan Velleman for suggesting a way to considerably shorten our proof of
Theorem 3, and for suggesting the simple example of isoparametric frames in Figure 13.
REFERENCES
1. T. M. Apostol and M. A. Mnatsakanian, Isoperimetric and isoparametric problems, Amer. Math. Monthly
111 (2004) 118–136. http://dx.doi.org/10.2307/4145213
2. , Figures circumscribing circles, Amer. Math. Monthly 111 (2004) 853–863. http://dx.doi.
org/10.2307/4145094
3. , Solids circumscribing spheres, Amer. Math. Monthly 113 (2006) 521–540. http://dx.doi.
org/10.2307/27641977
4. G. N. Frederickson, Dissections: Plane and Fancy, Cambridge University Press, Cambridge, 1997.
TOM M. APOSTOL joined the Caltech faculty in 1950 and became professor emeritus in 1992. He is director
of Project MATHEMATICS! (http://www.projectmathematics.com), an award-winning series of videos
he initiated in 1987. His long career in mathematics is described in the September 1997 issue of The College
Mathematics Journal. He is currently working with colleague Mamikon Mnatsakanian to produce materials
demonstrating Mamikon’s innovative and exciting approach to mathematics.
California Institute of Technology, 253-37 Caltech, Pasadena, CA 91125
apostol@caltech.edu
MAMIKON A. MNATSAKANIAN received a Ph.D. in physics in 1969 from Yerevan University, where he
became professor of astrophysics. As an undergraduate he began developing innovative geometric methods
for solving many calculus problems by a dynamic and visual approach that makes no use of formulas. He is
currently working with Tom Apostol under the auspices of Project MATHEMATICS! to present his methods in
a multimedia format.
California Institute of Technology, 253-37 Caltech, Pasadena, CA 91125
mamikon@caltech.edu
Abstract. We present a selection of a few discontinuous functions and we discuss some peda-
gogical advantages of using such functions in order to illustrate some basic concepts of math-
ematical analysis to beginners.
This level of generality is commonly attributed to Dirichlet, who in 1829 proposed his
celebrated function D defined on (0, 1) as follows:
(
1, if x ∈ (0, 1) ∩ Q,
D(x) = (1)
0, if x ∈ (0, 1) \ Q.
In fact, this example opened a door to a new world: functions are not just formulas,
or analytic expressions, as was commonly assumed in the 18th century. Functions can
be defined by very general laws.
From a pedagogical point of view, deciding the level of generality of functions to
use with young students is not straightforward. According to Kleiner [9, pp. 187–188],
it is possible to teach an elementary model of analysis by placing emphasis solely on
curves and the equations that represent them, without necessarily talking about func-
tions. Kleiner argues that students would find curves more natural than functions and
http://dx.doi.org/10.4169/amer.math.monthly.118.09.799
In [14] another approach has been developed. The introduction of [14] reads: “after
all, analysis has few bases: the concepts of function, infinity, limit, continuity, and
differentiability. To visit these locations in a comprehensive way, one should allow
these concepts the faculty of expressing all of their potential and all their fantastic
and incredible situations. To achieve this high level it is necessary to stick closely to
the definition of function as a correspondence of free sets.” The modern definition
of function clearly allows a deeper comprehension of other concepts, such as those
of limit, continuity, and differentiability. It is important that students understand the
far-reaching features of these notions, because only in this way will they be able to
avoid mistakes due to a limited concept of function. Students should be aware from
the very beginning that formulas provide a very small class of functions: analyticity is
a property enjoyed by very special functions. Being discontinuous is not synonymous
with being pathological because most functions are discontinuous. In the same way,
being continuous is not synonymous with being smooth since continuous functions
can be very rough, as in the case of the celebrated Weierstrass functions.
Following this plan, in [14] the authors propose a collection of problems the solu-
tions of which require the modern definition of function: in some cases, using highly
discontinuous functions is not essential but it simplifies the argument and makes the
situation clear.
We note that there is a vast literature devoted to so-called pathological functions. We
mention the classic book by Gelbaum and Olmsted [6] and the recent extensive mono-
graph by Kharazishvili [7]. We also mention Thim [12], which is basically a treatise
concerning continuous nowhere differentiable functions. However, at a pedagogical
level, not much material seems to be available.
In this paper, we adopt the approach of [14] and we discuss some peculiar examples,
some of which are taken from [14]. The focus is on special discontinuous functions,
the definitions of which are algebraically and technically simple in most cases. We
believe that this approach could be used not only with good students but also with weak
–3 –2 –1 1 2 3
–1
–2
–3
Figure 2. Graph of function F1 .
–2 – 2
–1
–2
Figure 3. Graph of function G 1 .
The function G 1 is continuous only at a countable set of points, precisely at the points
x = kπ with k ∈ Z.
Lemma 2. Let f, g, and h be functions from R to itself such that f (x) ≤ h(x) ≤ g(x)
for all x ∈ R. Let x0 ∈ R be fixed. If f and g are differentiable at x0 , f ′ (x0 ) = g ′ (x0 ),
and f (x0 ) = g(x0 ), then h is differentiable at x0 and h ′ (x0 ) = f ′ (x0 ) = g ′ (x0 ).
3 3
2 2
1 1
–2 –1 1 2 –2 –1 1 2
–1 –1
–2 –2
–3 –3
Figure 4. Graph of function F2 . Figure 5. A randomly generated function differ-
entiable at x = 0.
3. A function with positive derivative at one point, which is not monotone in any
neighborhood of that point. Let F3 be the function from R to itself defined by
(
x 2, if x ∈ Q,
F3 (x) =
2x − 1, if x ∈ R \ Q.
By Lemma 2, the function F3 is differentiable at the point x = 1 and F3′ (1) = 2, but
F3 is not monotone in any neighborhood of 1. Those students who are used to identify-
ing increasing functions with functions with positive derivative may find this example
rather bizarre. However, this example points out a well-known subtle distinction con-
cerning increasing functions. In fact, if a function f from R to itself is differentiable
at a point x0 and f ′ (x0 ) > 0 then f is increasing at the point x0 in the sense that
for all x in a convenient neighborhood of x0 we have: f (x) > f (x0 ) if x > x0 and
f (x) < f (x0 ) if x < x0 . Monotonicity is a stronger notion and, in the case of differ-
entiable functions, it occurs when the derivative does not change sign in the whole of
an interval.
2.5
2.0
1.5
1.0
0.5
–0.5
Theorem 3. Let f be a function from R to itself. Let A be the set of those points x0 ∈ R
such that f is discontinuous at x0 and at least one of the two limits limx→x + f (x) and
0
limx→x − f (x) exists and is finite. Then A is at most countable.
0
0.8
0.6
0.4
0.2
0.0
0.0 0.2 0.4 0.6 0.8 1.0
Figure 8. Graph of function F4 (arrested at the fourth iteration).
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
1.0 1.0
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Figure 9. The first four iterations.
np o
n
B= : n, p ∈ N, p is odd, p < 2 (5)
2n
p p+1
F4 n = (6)
2 2n
for all n, p ∈ N, where p is odd and p < 2n . This alternative definition allows an
easy proof of (4) since the summand 1/2n in the right-hand side of (6) vanishes in the
limiting procedure (see also (8) and (9)).
A well-known function which enjoys a similar property can be found in several
textbooks. It is defined as follows. Let G 4 be the function from (0, 1) ∩ Q to itself
defined by
p 1
G4 = (7)
q q
1
3
1
4
1
5
1 1 1 2 3
4 3 2 3 4
1
for all p, q ∈ N, 0 < p/q < 1, p, q coprime. It is not difficult to prove that
for all x0 ∈ (0, 1) ∩ Q. The function G 4 can be extended to the whole of (0, 1) by
setting G 4 (x) = 0 for all x ∈ (0, 1) \ Q and turns out to be continuous at any point
x ∈ (0, 1) \ Q. According to Appell [1, p. 10] some students would call function G 4
“Dirichlet light” because its set of discontinuities is “lighter” than the set of disconti-
nuities of the Dirichlet function.
We note that by (6) and (7) it immediately follows that
for all x ∈ B. Equality (9) could be used to define F4 via G 4 and to give a further
proof of (4). In this way, the function F4 , whose original definition is based on a geo-
metric construction, can eventually be represented via a function whose definition has
a completely different origin.
5. A function with jump discontinuities at all points of its domain. Monotone func-
tions with many jump discontinuities are well known in real analysis and probability
theory: these functions can be obtained as cumulative distribution functions associated
with discrete measures (see, e.g., Folland [5, p. 102]). Here we present two functions
defined on the set B of the dyadic rational numbers in (0, 1) with jump discontinuities
at every point of B. These functions are defined directly by using binary notation and
do not require any advanced tool.
We recall that the set B in (5) can be represented also in the form
∞ Xn
[ x j
B= j j
x ∈ {0, 1} for all j ∈ {1, . . . , n}, xn = 1 . (10)
n=1 j=1
2
0. x1 x2 · · · xn(x) .
G 5 = F4 ◦ F5 .
1.0
0.8
0.6
0.4
0.2
0.0
0.0 0.2 0.4 0.6 0.8 1.0
Figure 11. Graph of function G 5 (arrested at the third iteration).
6. A function approaching infinity at any point of its domain. The following ex-
ample is a simple variant of the function G 4 . Let F6 be the function from (0, 1) ∩ Q to
itself defined by
p
F6 =q
q
0.8
0.6
0.4
0.2
0.0
0.0 0.2 0.4 0.6 0.8 1.0
Figure 12. Construction of function G 5 .
50
40
30
20
10
0
1 1 1 2 3 1
4 3 2 3 4
lim F6 (x) = +∞
x→x0
ACKNOWLEDGMENTS. The authors are thankful to Prof. Maurizio Emaldi for pointing out reference [9]
and useful discussions, and also to Dr. Matteo Dalla Riva for pointing out reference [10]. The first two authors
are deeply indebted to the third author, P. Toni, who, in different periods, was their teacher of mathematics
and physics in secondary school, where he has developed and experienced the main pedagogical ideas that
are the basis of the present paper for about three decades. Those stimulating lessons led the second author,
P. D. Lamberti, to define the function F4 . The functions F5 and G 5 were defined by the first author, G. Drago,
in his dissertation [4], written under the supervision of the second author.
The authors are very thankful to a referee for suggesting formula (6), which simplifies a binary representa-
tion given in the first version of this paper, and for pointing out formula (9).
REFERENCES
GIACOMO DRAGO is 22 years old. He got his bachelor’s degree with honors from the University of Padova
in 2009. He distinguished himself in the national finals of three consecutive editions of the Kangourou Mathe-
matics Competitions in Italy (2004–2006). He is currently working on his master’s degree in mathematics, and
he dreams of becoming a teacher of mathematics.
Dipartimento di Matematica Pura ed Applicata, Università degli Studi di Padova, Via Trieste 63, 35121
Padova, Italy
giacomo.drago@gmail.com
PIER DOMENICO LAMBERTI received his Ph.D. in mathematics from the University of Padova in 2003.
He currently has a position as researcher in mathematical analysis at the same university. His research in-
PAOLO TONI received his master’s degree in mathematics from the University of Firenze in 1972. He has
taught mathematics and physics in high school for almost 40 years. He was a pioneer in the organization of
mathematics competitions in Italy. His interests are mainly in mathematical creativity and teacher education.
He retired in 2010. Now, he enjoys staying with his nephews and playing the organ. He believes in students’
creativity.
Liceo Scientifico Statale E. Fermi, Via Vittorio Emanuele II 50, 35100 Padova, Italy
paolotoni@inwind.it
n+ 12
1
A Ramanujan-Type Formula for 1+ e
n
n+ 12
Motivated by Sanjay Khattri’s proofs [1] that e < 1 + n1
!
, I offer the follow-
ing formula, which I think Ramanujan would have liked, for the ratio of the two
sides of Khattri’s inequality:
1
1 n+ 2
Z 1
u2
1+ e = exp 2 2
du
n 0 (2n + 1) − u
Z 1
u2
= exp 2
du (where m = n(n + 1)/2)
0 8m + 1 − u
Z 1
αu 2
= exp 2
du (where α = 1/(8m))
0 1 + α(1 − u )
Z 1
2 2 2 2 3 2 2 2 4 2 2 3
= exp αu − α u (1 − u ) + α u (1 − u ) − α u (1 − u ) + · · · du
0
1 1 2 2 1 4 2 3 1 6 4 2 4 1 8 6 4 2 5
= exp α− · α + · · α − · · · α + · · · · α − ···
3 5 3 7 5 3 9 7 5 3 11 9 7 5 3
α 2α 4α 6α 8α !
= exp 1− 1− 1− 1− 1 − ···
3 5 7 9 11
1 1 2 3 4 !
= exp 1− 1− 1− 1− 1 − ··· .
24m 5 · 4m 7 · 4m 9 · 4m 11 · 4m
REFERENCES
n+0.5
1. S. K. Khattri, Three proofs of the inequality e < 1 + n1 , Amer. Math. Monthly 117 (2010)
273–277. http://dx.doi.org/10.4169/000298910X480126
Abstract. Starting with Wilson’s theorem and its generalization by Gauss, we define a Gauss
factorial Nn ! to be the product of all positive integers up to N that are relatively prime to n. We
present results on the Gauss factorials ( n−1
M n
) !, and more generally on partial products obtained
when the product (n − 1)n ! is divided into M equal parts, for integers M ≥ 2. Finally, exten-
sions of the Gauss binomial coefficient theorem are presented in terms of Gauss factorials.
A proof of this result can be found in most introductory books on number theory, and
it depends on the fact that if p is prime then any integer a with 1 < a < p − 1 has its
inverse a −1 6 ≡ a (mod p).
For any odd prime p, if we write out the factorial ( p − 1)! and exploit symmetry
modulo p, we obtain
p−1 p+1
! p−1 p−1 !
p−1
1 · 2··· 2 2
· · · ( p − 1) ≡ 2
!(−1) 2
2
! (mod p), (1.2)
This was apparently first observed by Lagrange (see [13, p. 275]), and this congru-
ence can be used, along with a result of Mordell [24] that involves the class numbers
of imaginary
! p−1 quadratic fields, to completely characterize the multiplicative order of
2
! modulo p. We will leave this aside, and instead consider now the two halves
of the product on the left-hand side of (1.2). We denote these two partial products by
5(2) (2)
1 and 52 , respectively, where the upper index indicates the fact that we divide
the entire product into two equal parts. Using Wilson’s theorem (1.1) and symmetry
modulo p, we obtain 5(2) (2)
1 52 ≡ −1 (mod p) and
p−1
52(2) ≡ (−1) 2 5
(2)
1 (mod p). (1.4)
This is, of course, equivalent to (1.3), but writing it in this way gives rise to the follow-
ing question:
What can we say about the three partial products 5(3) (3) (3)
1 , 52 , 53 obtained by di-
viding the entire product ( p − 1)!, that is, the left-hand side of (1.1), into three equal
parts? For this we require p to be of the form p ≡ 1 (mod 3) or, in fact (since p is
prime), p ≡ 1 (mod 6); we then have
5(3)
1 = 1 · 2···
p−1
3
, 5(3)
2 =
p+2
3
· · · 2 p−2
3
, 5(3)
3 =
2 p+1
3
· · · ( p − 1).
http://dx.doi.org/10.4169/amer.math.monthly.118.09.812
5(3) (3)
3 ≡ 51 (mod p), (1.5)
Table 1 illustrates the congruences (1.5) and (1.7). We also see that, indeed, there
are no obvious relationships between 5(M) 1 and 52(M) for M = 3 and 4 and p < 100,
with the exceptions of p = 7 and p = 61, where 5(3) (3)
1 ≡ −52 (mod p). These could,
of course, be coincidences, but it turns out that this congruence holds also for p = 331,
p = 547, p = 1951, and for further relatively rare primes, as explained later. Here it
is interesting to mention p = 3571, the first case with 5(3) (3)
1 ≡ −52 ≡ 1 (mod p). In
contrast, there are no primes p for which 5(3) (3)
1 ≡ 52 (mod p), p ≡ 1 (mod 6), or
5(4) (4)
1 ≡ ±52 (mod p), p ≡ 1 (mod 4). All this is readily explained by appealing to
two deep theorems of Jacobi and of Gauss; we will return to this later.
Table 1. The partial products modulo p, for ( p − 1)! split into M = 3 and M = 4 equal parts.
p 5(3)
1 5(3)
2 5(3)
3 p 5(4)
1 5(4)
2 53(4) 54(4)
7 2 −2 2 5 1 2 −2 −1
13 −2 3 −2 13 6 3 −3 −6
19 −2 −5 −2 17 7 −3 −3 7
31 2 −8 2 29 −6 −2 2 6
37 7 3 7 37 −16 5 −5 16
43 −3 19 −3 41 13 7 7 13
61 −14 14 −14 53 26 7 −7 −26
67 −20 −33 −20 61 19 7 −7 −19
73 33 −12 33 73 18 −35 −35 18
79 −37 3 −37 89 22 42 42 22
97 21 −11 21 97 20 −28 −28 20
This naturally leads to the question of dividing the product ( p − 1)! into 5, 6, or
in general M ≥ 2 partial products of equal length, for primes p ≡ 1 (mod M). In
analogy to (1.6) we define, for such M and p, the products
p−1
Y
M
!
5(M)
j = ( j − 1) p−1
M
+i , ( j = 1, 2, . . . , M). (1.8)
i=1
5(M) (M)
M− j ≡ ±5 j (mod p), j = 1, 2, . . . , ⌊ M−1
2
⌋,
(M)
with the “central product” 5(M+1)/2 playing a somewhat special role when M is odd.
Extensive computations suggest that while (for fixed M) there are instances where two
of the partial products, with 1 ≤ j ≤ ⌊ M+1
2
⌋, are congruent, there are no cases where
all are simultaneously congruent.
2. COMPOSITE MODULI. This might well have been the end of the story were it
not for the possibility of considering composite moduli. Since our point of departure
has been Wilson’s theorem (1.1), let us first recall why Lagrange’s converse is true. If
n is composite, we can write it as n = n 1 n 2 , with 1 < n 1 < n. But then n 1 | (n − 1)!,
and therefore (n − 1)! 6 ≡ ±1 (mod n). However, if we suitably modify the factorial
on the left-hand side of (1.1), we obtain a composite analogue of Wilson’s theorem. It
was Gauss who first proved the following theorem.
The number of integers j in the product in (2.1), that is, those positive integers up
to n that are relatively prime to n,Qis given by Euler’s totient function φ(n), which has
the explicit evaluation φ(n) = n p|n (1 − 1p ), with the product taken over all prime
divisors p of n. We recall that the integers n in (2.1) for which the product is −1
(mod n) are exactly those that have a primitive root; this fact is important for the proof
of the result.
In spite of the fact that Theorem 1 was stated in the famous Disquisitiones Arith-
meticae [17, §78] and in the equally influential books [14, §38] and [20, p. 102], sur-
prisingly little can be found on this topic in the literature. The few published references
to this result include [21] and [27], where Theorem 1 was further extended, and [22]
and [1], where (2.1) was used to extend the classical Wilson quotient to composite
moduli. The theorem was rediscovered at least once; see [26].
In order to state this theorem and numerous other results more concisely, we intro-
duce the following notation: for positive integers N and n let Nn ! denote the product
of all integers up to N that are relatively prime to n, i.e.,
Y
Nn ! = j. (2.2)
1≤ j≤N
gcd( j,n)=1
This notation is a slight variation of the one used in [18], a useful reference on factorial
and binomial congruences. To be able to refer more easily to Nn !, we shall call it a
Gauss factorial, a terminology suggested by Theorem 1, which we call from here on
the Gauss-Wilson theorem.
We now turn to the composite analogue of Lagrange’s observation! n−1 in (1.2) and (1.3)
and begin with a general discussion of the Gauss factorial 2 n ! for odd integers
On the other hand, φ(n) is divisible by 4 if n has at least two distinct odd prime factors.
Hence by (2.3) we get
(
! n−1 2 −1 (mod n) if n = p α , p ≡ 1 (mod 4),
2 n
! ≡ (2.4)
1 (mod n) otherwise.
This is analogous to (1.3) for prime powers. In connection with this congruence we
remark that the multiplicative orders of ( n−1
2 n
) ! modulo n were completely determined
in [7].
As we did following (1.3), we will now turn to dividing the product (n − 1)n ! into
partial products. In complete analogy to (1.8) we define our partial products 5(M)
j , for
integers M ≥ 2 and n ≡ 1 (mod M), as
Y
5(M)
j := i, ( j = 1, 2, . . . , M), (2.5)
(M)
i∈I j
where, for j = 1, 2, . . . , M,
I j(M) := i | ( j − 1) n−1
M
+ 1 ≤ i ≤ j n−1
M
, gcd(i, n) = 1 . (2.6)
n 5(3)
1 5(3)
2 5(3)
3 n 5(4)
1 5(4)
2 5(4)
3 5(4)
4
70 29 1 29 61 19 7 −7 −19
73 33 −12 33 65 8 8 8 8
76 −29 −15 −29 69 31 −26 −26 31
79 −37 3 −37 73 18 −35 −35 18
82 −33 −25 −33 77 16 31 31 16
85 −28 9 −28 81 2 40 −40 2
88 5 −7 5 85 13 13 13 13
91 29 29 29 89 22 42 42 22
94 −23 43 −23 93 34 −10 −10 34
97 21 −11 21 97 20 −28 −28 20
First, when n = p is a prime, then clearly for a given M all I j(M) have the same number
of elements, namely φ M, j ( p) = p−1M
. Next, when M = 1, then φ1,1 (n) = φ(n). When
M = 2, then by symmetry of the sets I1(2) and I2(2) we have φ2,1 (n) = φ2,2 (n) = 12 φ(n).
However, already in the case M = 3 the situation is not as straightforward, as the
example n = 4 shows: in this case we have φ3,1 (n) = φ3,3 (n) = 1, but φ3,2 (n) = 0.
According to D. H. Lehmer [23] it was J. J. Sylvester who coined the term totatives
for those positive integers up to a given n that are relatively prime to n. We are therefore
dealing with the distribution of totatives in subintervals of the interval [1, n]. Lehmer
[23] was the first to study this distribution, and to give a sufficient condition for the
equal distribution of totatives. This area of study has attracted the attention of later
mathematicians; for instance, a conjecture of Erdős [16] was proven by Hall and Shiu
[19].
Table 3. The first ten moduli n for which all φ M, j (n) are equal, for each of M = 3, 4, 5.
28 22 · 7 4 25 52 5 66 2 · 3 · 11 4
49 72 14 45 2
3 ·5 6 121 112 22
2 4
52 2 · 13 8 65 5 · 13 12 176 2 · 11 16
70 2·5·7 8 85 5 · 17 16 186 2 · 3 · 31 12
76 22 · 19 12 105 3·5·7 12 231 3 · 7 · 11 24
91 7 · 13 24 117 32 · 13 18 246 2 · 3 · 41 16
112 24 · 7 16 125 53 25 286 2 · 11 · 13 24
2
124 2 · 31 20 145 5 · 29 28 341 11 · 31 60
130 2 · 5 · 13 16 153 32 · 17 24 366 2 · 3 · 61 24
133 7 · 19 36 165 3 · 5 · 11 20 396 22 · 32 · 11 24
Table 3 seems to indicate that whenever n ≡ 1 (mod M) has a prime factor p sat-
isfying p ≡ 1 (mod M), then the corresponding totatives are equally distributed, that
is, all φ M, j (n) are equal. This is in fact true, as was shown by D. H. Lehmer [23,
Theorem 4]:
Lemma 1 (Lehmer). Let M ≥ 2 and n ≡ 1 (mod M). If n has at least one prime
factor p with p ≡ 1 (mod M), then the totatives in the interval [1, n] are equally
distributed, that is, we have
1
φ M, j (n) = φ(n), ( j = 1, 2, . . . , M). (3.2)
M
Lehmer actually showed something slightly different, namely that under the given
conditions the intervals
n n
( j − 1) <i < j , j = 1, 2, . . . , M, (3.3)
M M
By definition the left-hand side of (4.1) is obviously identical with the right-hand side
for j = 1.
(M)
Table 4. The first ten moduli n for which all 5 j are congruent, for each of M = 3, 4, 5.
Q(5)
n factored 5(3)
j n factored 5(4)
j n factored j
91 7 · 13 29 65 5 · 13 8 341 11 · 31 −85
133 7 · 19 58 85 5 · 17 13 451 11 · 41 −105
217 7 · 31 67 145 5 · 29 1 671 11 · 61 −304
244 22 · 61 1 185 5 · 37 −68 781 11 · 71 −117
247 13 · 19 −88 205 5 · 41 1 1111 11 · 101 36
259 7 · 37 100 221 13 · 17 −1 1271 31 · 41 264
301 7 · 43 36 265 5 · 53 23 1441 11 · 131 89
364 22 · 7 · 13 113 305 5 · 61 −121 1661 11 · 151 −545
403 13 · 31 118 325 52 · 13 −57 1891 31 · 61 497
427 7 · 61 135 365 5 · 73 27 1991 11 · 181 125
We see from this table that all the moduli, except n = 244 = 22 · 61, have at least
two distinct prime factors that are congruent to 1 modulo M. In fact, we have:
Theorem 2. Let M ≥ 2 be an integer, and suppose that the positive integer n has at
least two distinct prime factors congruent to 1 (mod M). Then all the partial products
5(M)
j are congruent modulo n, that is, the congruences (4.1) hold.
Our starting point for the proof of Theorem 2 is the observation that each partial
product 5(M)
j can be written as a quotient of two Gauss factorials that are similar in
nature. In particular, we see immediately from the definitions (2.5) and (2.2) that
! n−1
(M) j M n!
5j = ! , j = 1, 2, . . . , M, (4.2)
( j − 1) n−1
M n
!
with the convention that 0n ! = 1. We therefore need to study the Gauss factorials on the
right-hand side of (4.2). Our main tool is the following explicit congruence, obtained
as a generalization of Proposition 2 in [7].
By the Chinese remainder theorem, applied to (4.4) and (4.5), the partial product 5(M)j
is uniquely determined modulo p α q β w = n, and most importantly, it is independent
of j. This completes the proof of Theorem 2.
To prove Lemma ! 2,
the main idea is to break the whole range of the product in the
Gauss factorial i n−1
M n
! into a number of products of approximately equal length and
a shorter “tail.” We then evaluate the products of the first type using the Gauss-Wilson
theorem with modulus e n := q β w. To do so we divide i n−1
M
by e
n with remainder:
n−1 e
n−1 pα − 1
i = ise
n+i , where s := . (4.6)
M M M
By hypothesis we know that s and (en − 1)/M are both integers. Based on (4.6) we
now decompose our Gauss factorial into is products of similar lengths and one shorter
product; that is, we write
is
Y
! n−1
i M n! = P j Q, (4.7)
j=1
where
e
n −1 ie
n −1
Y ! Y M
!
P j := ( j − 1)e
n+k , Q := ise
n+k . (4.8)
k=1 k=1
gcd(( j−1)e
n +k,n)=1 gcd(ise
n +k,n)=1
For a given j, if the set of residues {( j − 1)e n+k | 1 ≤ k ≤e n − 1}, subject to gcd(( j −
n + k, n) = 1, formed a reduced residue system modulo e
1)e n , then the product P j
would, by the Gauss-Wilson theorem, be congruent to −1 (mod e n ) if w = 1, and to
1 (mod e n ) if w > 1. However,
! n−1 this is not always the case because the residues that
appear in the product i M n ! have none divisible by p; these residues have been re-
!
moved from the normal factorial i n−1 M
! in forming the corresponding Gauss factorial.
where
i ! α−1 β
s′ = p q w−1 ,
M
which comes from the obvious division
n−1 p−1
i = s′ p + i , (4.10)
M M
where s ′ and p−1
M
are integers, by hypothesis. To count the number of elements in the
product (4.9), we do yet another obvious division, namely
! α−1 β ! !
p q w − 1 = p α−1 − 1 q β w + q β w − 1 ,
giving
i ! α−1 i
s′ = p −1 e n − 1).
n + (e (4.11)
M M
Counting the number of elements in the product (4.9) for each of the intervals of length
e
n is no problem; there are exactly φ(en ) elements in each of these intervals. The only
problem is to deal with the remainder term in (4.11), and for that we need Lemma 1.
Using (4.11) and (3.2) with e
n in place of n, we see that the number of elements in the
product (4.9) is
i ! α−1 i i α−1
p − 1 φ(e n) =
n ) + φ(e p φ(e
n ). (4.12)
M M M
But this expression is i A, with A as defined in (4.3). We therefore get from (4.7),
! n−1 P1 · · · Pis · Q
i M n! ≡ Qs ′ (mod e
n ). (4.13)
p iA j
j=1,gcd( j,e
n )=1
Here the bars over the P j and Q indicate that the products (4.8) are taken over all k
relatively prime to e
n , that is,
e
n −1 ie
n −1
Y ! Y !
M
P j := ( j − 1)e
n+k , Q := ise
n+k .
k=1 k=1
gcd(k,e
n )=1 gcd(k,e
n )=1
Q≡ n + k) ≡
(ise k (mod e
n ). (4.15)
k=1 k=1
gcd(k,e
n )=1 gcd(k,e
n )=1
α−1
The Gauss factorial in the denominator of (4.13) can be split up into i p M −1 products
that are congruent to the P j and a remainder that is congruent to Q (mod e n ); this
follows from (4.11). Hence (4.14) and (4.15) together with (4.13) give
! n−1 εB
i M n ! ≡ i A (mod e
n ), (4.16)
p
with A defined by (4.12) and
This implies:
Corollary 1. Let M ≥ 2 be an integer, and suppose that the positive integer n has
at least two distinct prime factors congruent to 1 (mod M). Then the multiplicative
order of the Gauss factorial ( n−1
M n
) ! modulo n is a divisor of M.
While in Lemma 2 the factor w plays only an auxiliary role, collecting all the irrel-
evant prime powers in n (if any), it turns out that we obtain some interesting results if
we consider the Gauss factorial in (4.3) modulo w. Indeed, using the multiplicativity
of φ(n) and the fact that M divides q − 1, we can rewrite A in (4.3) as
This holds for w = 1 because in that case the congruences are trivially true. As a
further consequence we obtain the following result, a special case of which was already
proven as Proposition 4 in [7]. We formulate it as a theorem since it supplements
Theorem 2.
Theorem 3. Let M ≥ 2 be an integer, and suppose that the positive integer n has at
least three distinct prime factors congruent to 1 (mod M). Then
p−1 p−1
52(4) 5(4)
1 52
(4)
2
! 2
Q 4 ( p) := = 2 = ! p−1 2 = p−1
. (6.1)
51(4) 5(4) ! 4
1 4
Table 5 lists the values of Q 4 ( p) (mod p) for all the appropriate primes p < 100,
where both the least positive and the least absolute residues are given.
While the least positive residues do not perhaps reveal much, we see a strong con-
nection between the least absolute residues and the representation of p as a sum of
two squares, the existence and uniqueness of which are guaranteed by the well-known
two-squares theorem of Fermat. Even the sign pattern is now quite obvious: the least
absolute residue is positive if and only if a ≡ 1 (mod 4).
5 2 2 1 2
13 7 −6 3 2
17 2 2 1 4
29 10 10 5 2
37 2 2 1 6
41 10 10 5 4
53 39 −14 7 2
61 10 10 5 6
73 67 −6 3 8
89 10 10 5 8
97 18 18 9 4
All this is in fact explained by the following celebrated theorem of Gauss. We fix
p, a, and b such that
Theorem 4 (Gauss). Let the prime p and the integer a be as in (6.2). Then
p−1
2
p−1
≡ 2a (mod p). (6.3)
4
5(4) (4)
2 6 ≡ ±51 (mod p) for all p ≡ 1 (mod 4).
Indeed, if this were not the case, then Q 4 ( p) ≡ ±1 (mod p). But by (6.1) and (6.3)
we have Q 4 ( p) ≡ 2a (mod p), so that 2a ≡ ±1 (mod p). The smallest possible solu-
√
tions of this congruence are a = ± p−1
2
. However, by (6.2) we have |a| < p, but also
√
p < p−12
for p > 5. This means that there are no solutions, which was to be shown.
(The case p = 5 is clear from Table 1.)
The analogous investigation for the case M = 3 is a bit more involved, which is
why we present it second. For primes p ≡ 1 (mod 6) we begin, in analogy to (6.1), by
considering
! p−1 p−1
5(3)
2 5(3)
1 52
(3)
2 3 ! 2 3
Q 3 ( p) := (3) = 2 = ! p−1 2 = p−1 . (6.4)
51 5(3) ! 3
1 3
In an attempt to find a congruence for Q 3 ( p) that is analogous to (6.3) one might want
to try another two-squares formula of Fermat, namely p = a 2 + 3b2 for primes p ≡ 1
(mod 6), which is unique up to signs. (For the early history of such representations,
see [15, pp. 14ff.].) However, as columns 2 and 3 of Table 6 show, there seems to be
no apparent relationship modulo p between Q 3 ( p) and a or b.
It was Jacobi who, in 1837, used instead the representation 4 p = x 2 + 3y 2 , which
always has three distinct solutions, namely (|2a|, |2b|), (|a + 3b|, |a − b|), and (|a −
p Q 3 ( p) (mod p) a, b x1 , y1 x2 , y2 x1 , y1 r, s
7 −1 2, 1 5, 1 4, 2 1, 3 1, 3
13 5 1, 2 7, 1 5, 3 2, 4 −5,3
19 −7 4, 1 8, 2 7, 3 1, 5 7, 3
31 −4 2, 3 11, 1 7, 5 4, 6 4, 6
37 11 5, 2 11, 3 10, 4 1, 7 −11, 3
43 8 4, 3 13, 1 8, 6 5, 7 −8, 6
61 −1 7, 2 14, 4 13, 5 1, 9 1, 9
67 5 8, 1 16, 2 11, 7 5, 9 −5, 9
73 −7 5, 4 17, 1 10, 8 7, 9 7, 9
79 17 2, 5 17, 3 13, 7 4, 10 −17, 3
97 −19 7, 4 19, 3 14, 8 5, 11 19, 3
3b|, |a + b|). These three solutions, for p < 100, are listed in columns 4–6 of Table 6.
One of these solutions always satisfies y ≡ 0 (mod 3); it is then the corresponding
x, with its sign appropriately chosen, that gives the desired congruence. To be exact,
suppose that the prime p and integers r , s are such that
The integer r is then uniquely determined, and we can now state Jacobi’s binomial
coefficient theorem, which is illustrated in the last column of Table 6.
Proofs of the theorems of Gauss and Jacobi are nonelementary and can be found
in the book [2] by Berndt, Evans, and Williams, which is the standard reference in
the field. For remarks and references, see [2, p. 291]. It is worth giving an explicit
connection between the a in p = a 2 + 3b2 , with its sign fixed by the condition a ≡ −1
(mod 3), and the r as fixed in (6.5):
2a if b ≡ 0 (mod 3),
r = −(a − 3b) if b ≡ 1 (mod 3),
−(a + 3b) if b ≡ 2 (mod 3).
5(3) (3)
2 6 ≡ 51 (mod p) for all p ≡ 1 (mod 6).
Indeed, if this were not the case, we would have r ≡ −1 (mod p) by (6.4) and (6.6).
Now r = −1 is impossible since r ≡ 1 (mod 3). The next smallest solution, r = p −
√ √
1, is also impossible since by (6.5) we have |r | < 2 p, but we already saw that 2 p <
p − 1 for p > 5.
The first primes generated by this formula are 7, 61 (see Table 1), 331, 547, and
1951. As is easily seen, negative x give rise to the same primes. It is, of course, not
known whether there are infinitely many primes of this form.
We continue this section with some remarks on congruences for the factorials p−1 4
!
p−1
and 3 !, all of which follow from the theorems of Gauss and Jacobi, respectively.
First, consider primes of the form p ≡ 1 (mod 4). Then
p−1
(a) 4
! ≡ 1 (mod p) only if p = 5.
! p−1 k
(b) 4
! 6≡ −1 (mod p) for k = 1, 2, 4.
! p−1 8
(c) 4
! ≡ −1 (mod p) does hold for p = 17, 241, 3361, 46817, 652081, . . .
For primes of the form p ≡ 1 (mod 6) we have
p−1
(d) 3
! ≡ 1 (mod p) holds for p = 3571, 4219, 13669, 25117, 55897, . . .
! p−1 3
(e) 3
! ≡ 1 (mod p) holds if and only if p = 27x 2 + 27x + 7 for some integer
x.
! p−1 9
(f) 3
! ≡ 1 (mod p) holds if and only if p = 3y 2 + 3y + 1 for some integer
y. Furthermore, the multiplicative order of p−1
3
! (mod p) is 9 if and only if p
2 2
is of the form p = 27x + 9x + 1 or p = 27x + 45x + 19.
! p−1 k
(g) 3
! 6≡ −1 (mod p) for k = 1, 3, 9.
! p−1 18
(h) 3
! ≡ −1 (mod p) holds if and only if p satisfies p 2 = 3y 2 + 3y + 1 for
some integer y. The first few such primes are 13, 181, 2251, 489061.
Statements (c) and (h) are actually connected in the following surprising way: The
identity for p 2 in (h) can be rewritten in the form of the Pell equation (2 p)2 − 3(2y +
1)2 = 1. The infinitely many solutions (An , Bn ) of the equation A2 − 3B 2 = 1 are
given by the recurrence relations (see, e.g., [25, p. 354])
An+2 = 4An+1 − An , A0 = 1, A1 = 2,
Bn+2 = 4Bn+1 − Bn , B0 = 0, B1 = 1.
Then, as is shown in [9], the primes p in (h) are given by primes 12 A2k−1 , while those
2
in (c) are given by prime values of Bn−1 + Bn2 .
While details concerning (h) can be found in [9], statements (a)–(g) are derived and
further discussed in a forthcoming paper [10]. However, some of them follow imme-
diately from results earlier in this section. For instance, if we square (6.1) and use the
fact that by (1.3) we have ( p−1
2
!)2 ≡ −1 (mod p), then we get ( p−1 4
!)4 ≡ −Q 4 ( p)−2
(mod p). Since we know that Q 4 ( p) 6≡ −1 (mod p), this proves statement (b) for
k = 4. The solutions in statement (c) are related to a certain Pell equation; see [4,
We have seen that Q 3 ( p) ≡ 1 (mod p) is impossible, which proves statement (g) for
k = 3 and also for k = 1. On the other hand, the case Q 3 ( p) ≡ −1 (mod p) is equiv-
alent to Corollary 3; hence (6.7) gives statement (e). A proof along these lines was
suggested by Andrew Granville (private communication with the first author, Decem-
ber, 2004); see also [6].
Statement (e) means that the multiplicative order of p−1 3
! modulo p is 1 or 3 when
2
p = 27x + 27x + 7 for some integer x. Order 1 does actually occur, as statement
(d) indicates. Computations that were kindly carried out for us by Yves Gallot show
that there are 364 such primes up to 109 , while for 762 primes up to 109 the order of
p−1
3
! is 3. It appears to be a difficult question to find a criterion for when the order is
1 and when it is 3. Also, the data suggest that the split between these two classes may
approach 1:2.
Returning to statement (e) and Corollary 3, recall that the polynomial expression
for p comes from
where we have put s = 2y + 1 since s has to be odd. Now Jacobi’s theorem required
3 | s, that is, y = 3x + 1, which led to p = 27x 2 + 27x + 7. In the other two cases,
namely y = 3x and y = 3x + 2, we get p = 27x 2 + 9x + 1 and p = 27x 2 + 45x +
19, respectively. In both these cases the order of p−1 3
! is 9, as is shown in [10]. This,
together with (6.8), gives statement (f).
To conclude this section we note that for Gauss factorials with composite moduli
the situation related to statements (a) and (d) is very different: As Theorem 3 shows,
for each M ≥ 2 we have ( n−1 ) ! ≡ 1 (mod n) for infinitely many n, namely all those
M n
with at least three distinct prime factors p ≡ 1 (mod M).
However, if these composite moduli are prime powers, then the situation remains
very interesting. In fact, in [9] we showed that for a given M ≥ 2 and p ≡ 1 (mod M)
the sequence of multiplicative orders mod p α of the Gauss factorials
p α −1
M
!, α = 1, 2, . . . ,
p
Obviously α = 1 gives the usual binomial coefficient. For α = 2 we were able to show
that the congruence (7.1) is equivalent to
p
B (2) ( p) ≡ 2a − (mod p 2 ). (7.3)
2a
Seeing the simplicity of (7.3) as compared with (7.1), one is led to search numerically
for congruences modulo higher powers of p. Indeed, one readily conjectures that
p p2
B (3) ( p) ≡ 2a − − 3 (mod p 3 ). (7.4)
2a 8a
Continuing with higher powers, we were then able to conjecture and ultimately prove:
1 2n
!
where Cn := n+1 n
is the nth Catalan number, which is always an integer.
The first few Catalan numbers C0 , C1 , . . . are 1, 1, 2, 5, 14, 42, 132, . . . The proof
of Theorem 7 uses methods similar to those in the proof of Theorem 6; see [2, The-
orem 9.4.3]. The Catalan numbers enter through certain combinatorial identities that
are related to αth powers of particular binomial expressions in complex numbers. For
details, see [8].
If the summation on the right of (7.5) is considered 0 for α = 1, then Gauss’s The-
orem 4 can be seen as a special case of (7.5). We already remarked that for α = 2
the congruences (7.5) and (7.1) are equivalent. This leads to the natural question of
whether one can derive a binomial coefficient analogue to (7.5) for α = 3. This is in
fact possible, and we obtain the following mod p 3 extension of Theorem 6.
For further details and proofs, see [8], where Jacobi’s Theorem 5 has also been ex-
tended in a similar fashion.
To summarize: Comparing the congruences (7.5) and (7.6), it is clear that for higher
congruences the Gauss factorials with prime power moduli are the more natural objects
to study than the usual factorials.
ACKNOWLEDGMENTS. We would like to thank Yves Gallot for generously carrying out the computations
reported in Section 6. We also thank Andrew Granville for suggesting the proof mentioned in Section 6.
REFERENCES
1. T. Agoh, K. Dilcher, and L. Skula, Wilson quotients for composite moduli, Math. Comp. 67 (1998) 843–
861. http://dx.doi.org/10.1090/S0025-5718-98-00951-X
2. B. C. Berndt, R. J. Evans, and K. S. Williams, Gauss and Jacobi Sums, Wiley, New York, 1998.
3. F. Beukers, Arithmetical properties of Picard-Fuchs equations, in Seminar on Number Theory—Paris,
1982–83, Progr. Math., vol. 51, M.-J. Bertin and C. Goldstein, eds., Birkhäuser, Boston, MA, 1984,
33–38.
4. J. M. Borwein and D. Bailey, Mathematics by Experiment. Plausible Reasoning in the 21st Century, 2nd
ed., A K Peters, Wellesley, MA, 2008. !
5. S. Chowla, B. Dwork, and R. Evans, On the mod p 2 determination of (( p−1)/2 p−1)/4 , J. Number Theory 24
(1986) 188–196. http://dx.doi.org/10.1016/0022-314X(86)90102-2
6. J. B. Cosgrave, Trinity College Dublin Mathematical Society Lecture, available at http://staff.spd.
dcu.ie/johnbcos/jacobi.htm.
7. J. B. Cosgrave and K. Dilcher, Extensions of the Gauss-Wilson theorem, Integers 8 (2008) A39; available
at http://www.integers-ejcnt.org/vol8.html.
8. , Mod p 3 analogues of theorems of Gauss and Jacobi on binomial coefficients, Acta Arith. 142
(2010) 103–118. http://dx.doi.org/10.4064/aa142-2-1
9. , The multiplicative orders of certain Gauss factorials, Int. J. Number Theory 7 (2011) 145–171.
http://dx.doi.org/10.1142/S179304211100396X
10. , The Gauss-Wilson theorem for one-third, one-quarter and one-sixth intervals (in preparation).
11. R. Crandall, Topics in Advanced Scientific Computation, Springer-Verlag, New York, 1996.
12. R. E. Crandall, K. Dilcher, and C. Pomerance, A search for Wieferich and Wilson primes, Math. Comp.
66 (1997) 433–449. http://dx.doi.org/10.1090/S0025-5718-97-00791-6
13. L. E. Dickson, History of the Theory of Numbers. Volume I: Divisibility and Primality, Chelsea, New
York, 1966.
14. P. G. L. Dirichlet, Vorlesungen über Zahlentheorie, 4th ed., ed. and supplemented by R. Dedekind,
Chelsea, New York, 1968; translated as Lectures on Number Theory by J. Stillwell, American Mathe-
matical Society, Providence, RI, 1999.
15. H. M. Edwards, Fermat’s Last Theorem. A Genetic Introduction to Algebraic Number Theory, Springer-
Verlag, New York, 1977.
16. P. Erdős, Some remarks on a paper of McCarthy, Canad. Math. Bull. 1 (1958) 71–75. http://dx.doi.
org/10.4153/CMB-1958-008-7
17. C. F. Gauss, Disquisitiones Arithmeticae (trans. and preface by A. A. Clarke), Yale University Press,
New Haven, 1966; rev. by W. C. Waterhouse, C. Greither, and A. W. Grootendorst with preface by
W. C. Waterhouse, Springer-Verlag, New York, 1986.
18. A. Granville, Arithmetic properties of binomial coefficients. I. Binomial coefficients modulo prime pow-
ers, in Organic Mathematics—Burnaby, BC, 1995, CMS Conf. Proc., vol. 20, American Mathematical
Society, Providence, RI, 1997, 253–276.
19. R. R. Hall and P. Shiu, The distribution of totatives, Canad. Math. Bull. 45 (2002) 109–114. http:
//dx.doi.org/10.4153/CMB-2002-012-1
JOHN B. COSGRAVE was born in Bailieboro, County Cavan, Ireland. He received his B.Sc. (1968) and
Ph.D. (1972) in Mathematics from Royal Holloway College (London University). He worked at RHC,
Manchester, Jos (Nigeria), Carysfort College (Dublin), and finally St. Patrick’s College, Drumcondra, Dublin,
where—shortly before retiring in 2007—he had the pleasure of having Doron Zeilberger as his department’s
international external assessor. Besides elementary number theory, his interests include reading (all kinds),
music, and cycling, and—together with his wife Mary, whom he met at RHC—he is a daily swimmer in
Dublin Bay.
79 Rowanbyrn, Blackrock, County Dublin, Ireland
jbcosgrave@gmail.com
http://staff.spd.dcu.ie/johnbcos/
KARL DILCHER received his undergraduate education at the Technische Universität Clausthal in Germany.
He then did his graduate studies at Queen’s University in Kingston, Ontario, and finished his Ph.D. there
in 1983 under the supervision of Paulo Ribenboim. He is currently a professor at Dalhousie University in
Halifax, Nova Scotia, Canada, where he first arrived in 1984 as a postdoctoral fellow. His research interests
include classical analysis, special functions, and elementary and computational number theory.
Department of Mathematics and Statistics, Dalhousie University, Halifax, Nova Scotia, B3H 4R2, Canada
dilcher@mathstat.dal.ca
Abstract. The rationals can be enumerated using Stern-Brocot sequences SBn , Calkin-Wilf
sequences CW n , or Farey sequences Fn . We show that the rationals in SBn , CW n , and Fn can
be computed using very similar second-order linear recurrence relations. (This, incidentally,
obviates the need to compute previous sequences SBi , CW i , Fi with i < n.)
a c p
b d q
a +c p p+q
b+d p+q q
A simple induction shows that |CW n | = 2n−1 . Another simple induction (see [2,
p. 360]) shows that the fractions in CW n have the form
b−1 b0 b N −2
CW n = , ,..., where N = 2n−1 ,
b0 b1 b N −1
and the denominator of a given fraction is the numerator of the succeeding fraction.
Indeed, this property obtains even when the sequences CW 1 , CW 2 , CW 3 , . . . are con-
catenated to form
z}|{ z }| { z }| {
1 1 2 1 3 2 3 1 4 c0 c1 c2
CW ∞ := , , , , , , , , . . . , , . . . = , , ,... .
1 2 1 3 2 3 1 4 1 c1 c2 c3
p
The Farey sequence of order n contains all the reduced fractions q
with 0 ≤ p ≤
q ≤ n, in their natural order. Thus
0 1
F1 = , ,
1 1
0 1 1
F2 = , , ,
1 2 1
0 1 1 2 1
F3 = , , , , ,
1 3 2 3 1
0 1 1 1 2 3 1
F4 = , , , , , , ,....
1 4 3 2 3 4 1
A standard way to compute Fn from Fn−1 is to insert mediants between consecutive
fractions of Fn−1 only when this gives a denominator of size n (see [3, p. 118]). Thus
Fn is a subsequence of SB n . The number of reduced fractions an with 1 ≤ a < n is
denoted ϕ(n). Counting the P nonzero reduced fractions with denominator j, together
with zero, gives |Fn | = 1 + nj=1 ϕ( j). The mediant rule above implies that consecu-
tive fractions in SB n and Fn are adjacent (see also [3, p. 119]).
S
2. RECURSIVE RESULTS. It is shown in [3] and [2] that SB ∞ = ∞ n=0 SB n and
CW ∞ contain every (reduced) positive rational precisely once. Although SB n , CW n ,
and Fn are defined “top-down” they can be computed from “left to right” via almost
identical recurrence relations. The 2-exponent ν2 (i) of a positive integer i is defined
by i = 2ν2 (i) j, where j is an odd integer.
Theorem 2. Write
a0 a1 ai−1
CW ∞ = , ,..., ,...
a1 a2 ai
and
b−1 b0 b1 b N −2
CW n = , , ,...,
b0 b1 b2 b N −1
where N = 2n−1 . Then the ai and bi can be computed via the recurrence relations
i −1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
ai 0 1 1 2 1 3 2 3 1 4 3 5 2 5 3 4 1
bi 1 4 3 5 2 5 3 4 1 3 2 3 1 2 1 1 0
ki 1 3 1 5 1 3 1 7 1 3 1 5 1 3 1
Certainly the formulas for a−1 and a0 in (1a) are correct, as SB n starts with 01 , n1 .
Suppose now that 1 ≤ i < N , and consider the cases when i is even and odd.
a ′j−2 a ′j−1 a ′j
SBn−1 b′j−2 b′j−1 b′j
Case 1. i = 2 j is even and j ≥ 1. The following shows that (1a) holds for even i:
Case 2. i = 2 j + 1 is odd and j ≥ 1. It follows from k2 j+1 = 1 and (4) and (5) that
Define a−1 = 0. With this definition (6.1) and (6.2) also hold when j = 0. We shall
now prove, by induction on i, that the formulas for ai in (2a) are correct. This is clear
for i = −1, 0. Now suppose i > 0, and assume the formulas in (2a) hold for all i ′ < i.
a0 1
a1
= 1
a j−1
aj
a1 1 a2 2
a2
= 2 a3
= 1
a j−1 a j−1 +a j
a j−1 +a j aj
a3 1 a4 3 a5 2 a6 3
a4
= 3 a5
= 2 a6
= 3 a7
= 1
Proof of Theorem 3. As the first two fractions of Fn are 01 and n1 , the recurrences
Ai−2 Ai−1
(3a, b) are correct for i = −1, 0. Now suppose 1 ≤ i < N . By definition, Bi−2 , Bi−1 ,
Ai
and Bi
are consecutive terms in Fn . We shall prove that Ai = a and Bi = b, where
(8.1) (8.2)
1 ≤ a Bi − b Ai and 1 = Ai Bi−1 − Bi Ai−1 . (8)
This is a contradiction since ABii ∈ Fn has Bi ≤ n. Hence ABii = ab . As both fractions are
reduced (and Ai , Bi , b > 0), we conclude that Ai = a and Bi = b, as desired.
ACKNOWLEDGMENTS. The author would like to thank the referees for their insightful comments.
REFERENCES
1. B. Bates, M. Bunder, and K. Tognetti, Locating terms in the Stern-Brocot tree, European J. Combin. 31
(2010) 1020–1033. http://dx.doi.org/10.1016/j.ejc.2007.10.005
2. N. Calkin and H. S. Wilf, Recounting the rationals, Amer. Math. Monthly 107 (2000) 360–363. http:
//dx.doi.org/10.2307/2589182
3. R. L. Graham, D. E. Knuth, and O. Patashnik, Concrete Mathematics, 2nd ed., Addison-Wesley, Reading,
MA, 1994.
4. M. Hockman, Continued fractions and the geometric decomposition of modular transformations, Quaest.
Math. 29 (2006) 427–446. http://dx.doi.org/10.2989/16073600609486174
5. M. Niqui, Exact arithmetic on the Stern-Brocot tree, J. Discrete Algorithms 5 (2007) 356–379. http:
//dx.doi.org/10.1016/j.jda.2005.03.007
6. On-Line Encyclopedia of Integer Sequences, available at http://oeis.org.
Abstract. We prove the Banach-Steinhaus theorem and some of its consequences under the
assumption that the underlying space is a topological group of second Baire category.
In [5, p. 44], Theorem 1 is proved under the assumptions that X and Y are topo-
logical vector spaces, and the set of all elements of X with bounded orbits is a set of
second Baire category in X . In the next section we will extend Theorem 1 to the case
where X is a topological group of second Baire category.
Let (X, ∗, τ ) be a topological group. Recall that if (X, τ ) is Hausdorff, then it is of
second Baire category in each of the following cases [5, p. 43]:
(a) X is complete and metrizable;
(b) X is locally compact.
Also, if (X, ∗, τ ) is a Hausdorff topological group, then it is metrizable if and only
if the topological space (X, τ ) has a countable base of neighborhoods of the identity
element [1].
The identity element of a topological group (X, ∗, τ ) will be denoted by e, the
inverse of x ∈ X by x −1 , while x m will stand for |x ∗ ·{z
· · ∗ x}. For basic properties and
examples of topological groups, see [1]. m times
Theorem 2. Let (X, ∗, τ ) be a topological group of second Baire category, and let
(Y, k · k) be a normed space. If aλ : X → Y is a continuous homomorphism for ev-
ery λ ∈ 3 such that supλ∈3 kaλ (x)k < ∞ for all x ∈ X , then the family {aλ }λ∈3 is
equicontinuous.
http://dx.doi.org/10.4169/amer.math.monthly.118.09.836
836
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 118
Proof. Fix ǫ > 0. For L ∈ N, let B L = {y ∈ Y : kyk ≤ L} and put
\
XL = aλ−1 (B L ).
λ∈3
Then X L is closed for every L. Also, one can easily check that
[
X= XL.
L
Using the assumption that X is of second Baire category, we infer that there exists
L 0 such that X L 0 contains a nonempty open set. Therefore, aλ (U ) ⊂ B L 0 , λ ∈ 3, for
some nonempty open U ⊂ X . This allows us to conclude that
Note that, by the triangle inequality and the fact that each aλ is a homomorphism, we
have
Now let m ∈ N be such that 2Lm0 < ǫ, and let O be an open neighborhood of e for
which O m ⊂ V . If x ∈ O, then x m ∈ V , and we obtain
m · â(x) = â(x m ) ≤ 2L 0 ,
which implies that â(x) < ǫ. Thus, for a given ǫ > 0, we were able to find an open
neighborhood O of e such that kaλ (x)k < ǫ for all λ ∈ 3, that is, the family {aλ }λ∈3
is equicontinuous.
Theorem 3. Let (X, ∗, τ ) be a topological group of second Baire category, and let
(Y, k · k) be a Banach space. For a sequence an : X → Y of continuous homomor-
phisms, consider the conditions:
(A) the sequence {an (x)} converges for every x ∈ X ;
(B) â(x) = supn kan (x)k < ∞ for every x ∈ X ;
(C) the operator â : X → R is continuous at e;
(D) the set C = {x ∈ X : {an (x)} converges} is closed in X .
Then the implications (A)⇒(B)⇒(C)⇒(D) hold. If, in addition, the sequence {an (x)}
converges for every x in a dense subset of X , then conditions (A)–(D) are equivalent.
kam (x̄) − an (x̄)k ≤ kam (x̄ ∗ x0 ) − an (x̄ ∗ x0 )k + kam (x0 )k + kan (x0 )k < ǫ,
Theorem 4. Let (Y, k · k) be a normed space, and let X and {an } be as above. If,
for every x ∈ X , there exists yx ∈ Y satisfying limn→∞ kan (x) − yx k = 0, then the
operator a : X → Y given by the formula a(x) = yx , x ∈ X , is a continuous homo-
morphism.
Proof. Because the sequence {an (x)} converges for every x ∈ X , condition (A) of
Theorem 3 is satisfied, which implies that the operator â is continuous at e. Therefore,
since a(x) = limn→∞ an (x) and, consequently,
ACKNOWLEDGMENTS. The author is grateful to Professor Vyacheslav Chistyakov of the State University
Higher School of Economics, Nizhny Novgorod, for posting a question that is answered in Theorem 4 of the
present article.
REFERENCES
1. E. Hewitt and K. A. Ross, Abstract Harmonic Analysis, vol. I, Springer-Verlag, Berlin, 1963.
2. U. Krengel, Ergodic Theorems, Walter de Gruyter, Berlin, 1985.
3. S. Litvinov, The Banach principle for topological groups, Atti Sem. Mat. Fis. Univ. Modena e Reggio
Emilita LIII (2005) 323–330.
4. H. L. Royden and P. M. Fitzpatrick, Real Analysis, 4th ed., Prentice Hall, Boston, 2010.
5. W. Rudin, Functional Analysis, 2nd ed., McGraw Hill, New York, 1991.
838
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 118
Ramanujan’s “Most Beautiful Identity”
Michael D. Hirschhorn
Abstract. We give a simple proof of the identity which for Hardy represented the best of
Ramanujan. On the way, we give a new proof of an important identity that Ramanujan stated
but did not prove.
Of all the 4000 or so identities Ramanujan presented, Hardy chose one which for
him represented the best of Ramanujan. I would like to show you this identity, and
prove it.
Following Euler, we define a partition of the positive integer n as a representation
of n as a sum of positive integers, in which order is unimportant. The partitions of 4
are 4 = 3 + 1 = 2 + 2 = 2 + 1 + 1 = 1 + 1 + 1 + 1. The number of partitions of n
is denoted by p(n); thus, p(4) = 5. For convenience, we define p(0) = 1.
Euler showed that the partition generating function
X
P(q) = p(n)q n = 1 + 1q + 2q 2 + 3q 3 + 5q 4 + · · ·
n≥0
satisfies
1
p(q) = ,
(q; q)∞
where
Y
(a; q)∞ = (1 − aq n ).
n≥0
(q; q)∞ = 1 − q − q 2 + q 5 + q 7 − q 12 − q 15 + + − − · · · .
In the series on the right, the terms occur in pairs, alternately with coefficients −1
and +1. The powers, 1, 2, 5, 7, 12, 15, . . ., are known as the pentagonal numbers, and
Euler’s expansion above as the pentagonal number theorem. The easiest way of gener-
ating the pentagonal numbers is to proceed as follows.
The triangular numbers 1, 1 + 2, 1 + 2 + 3 and so on are
Two of every three are divisible by 3. If we divide these by 3 we obtain the pentagonal
numbers!
A beautiful combinatorial proof of Euler’s pentagonal number theorem was given
by F. Franklin in 1881, and is reproduced in Hardy and Wright [3].
Euler’s pentagonal number theorem is the special case a = 1 of Jacobi’s triple prod-
uct identity when written
http://dx.doi.org/10.4169/amer.math.monthly.118.09.839
A proof of the triple product identity is given in the Appendix. Also, it is important
that I mention that Ramanujan found a marvellous and powerful extension of the triple
product identity.
Back to the main story: We have that the partition generating function is the recip-
rocal of the series 1 − q − q 2 + q 5 + q 7 − − + + · · · . This implies
p(0) = 1,
p(1) − p(0) = 0,
p(2) − p(1) − p(0) = 0,
p(3) − p(2) − p(1) = 0,
Here, you recognise the pattern of + and − signs, and the numbers 1, 2, 5, 7, . . . . For
each n, the sum on the left terminates, since all terms with negative argument are zero.
P. MacMahon, who was in Cambridge with Hardy and Ramanujan, used the above
recurrence to calculate p(n) for n ≤ 200, and serendipitously listed the values in
groups of five thus:
Ramanujan observed that the numbers at the bottom of each group are divisible by 5,
that is, 5| p(5n + 4). He also observed that 7| p(7n + 5), 11| p(11n + 6), and, on the
basis of the very small amount of evidence provided by MacMahon’s table, formulated
a very general conjecture, which was essentially correct; the proof was completed by
Oliver Atkin in 1967.
Ramanujan did much more than prove 5| p(5n + 4). We can write down the gener-
ating function of the p(5n + 4),
X
p(5n + 4)q n = 5 + 30q + 135q 2 + 490q 3 + 1575q 4 + 4565q 5 + · · · .
n≥0
Ramanujan [4, p. 213] makes the claim that this series can be written as a neat product,
X
n(q 5 ; q 5 )5∞
p(5n + 4)q = 5 . (1)
n≥0
(q; q)6∞
Hardy [3, p. xxxv] says of (1): “It would be difficult to find more beautiful formulae
than the ‘Rogers-Ramanujan’ identities, but here Ramanujan must take second place
to Rogers; and, if I had to select one formula from all Ramanujan’s work, I would
= (q 5 ; q 5 )6∞ /(q 25 ; q 25 )∞ ,
(q; q)∞ = 1 + q 5 − q 15 − q 35 − q 40 − q 70 · · ·
!
−q 1 − q 25 − q 50 + q 125 · · ·
!
−q 2 1 − q 5 + q 10 − q 20 − q 55 + q 75 · · · .
!
(q 10 ; q 25 )∞ (q 15 ; q 25 )∞ (q 25 ; q 25 )∞
(q; q)∞ = − q(q 25 ; q 25 )∞ (4)
(q 5 ; q 25 )∞ (q 20 ; q 25 )∞
(q 5 ; q 25 )∞ (q 20 ; q 25 )∞ (q 25 ; q 25 )∞
− q2 .
(q 10 ; q 25 )∞ (q 15 ; q 25 )∞
Hardy regarded this as a gap in Ramanujan’s proof of (1), and said [2, p. 90] “Ra-
manujan never gave a complete proof.”
Proof. Here is a simple proof of (4) that I found recently. The triple product identity
can be written
or
2
X
(1 − a −1 )(a −1 q; q)∞ (aq; q)∞ (q; q)∞ = (−1)n (a n − a −n−1 )q (n +n)/2 .
n≥0
2π
In particular, if θ = 5
,
Y
(1 + αq n + q 2n )(1 − q n ) = (q 10 ; q 25 )∞ (q 15 ; q 25 )∞ (q 25 ; q 25 )∞ (5)
n≥1
− βq(q 5 ; q 25 )∞ (q 20 ; q 25 )∞ (q 25 ; q 25 )∞ ,
while if θ = π5 ,
Y
(1 + βq n + q 2n )(1 − q n ) = (q 10 ; q 25 )∞ (q 15 ; q 25 )∞ (q 25 ; q 25 )∞ (6)
n≥1
− αq(q 5 ; q 25 )∞ (q 20 ; q 25 )∞ (q 25 ; q 25 )∞ ,
√ √
where α = 1+2 5 and β = 1−2 5 , and where we have used the triple product identity to
sum the series that arise.
By way of explanation, when we substitute θ = 2π5 , the terms of the sum in which
n ≡ 0 or −1 (mod 5) are
X
5m
sin (10m + 1) 2π5 2 +5m)/2
(−1) 2π
q (25m
m≥0
sin 5
(q 10 ; q 25 )∞ (q 15 ; q 25 )∞ (q 5 ; q 25 )∞ (q 20 ; q 25 )∞
a= and b= = a −1 ,
(q 5 ; q 25 )∞ (q 20 ; q 25 )∞ 10 25
(q ; q )∞ (q ; q )∞15 25
(q; q)∞ = (q 25 ; q 25 )∞ (a − q − q 2 b)
= (q 25 ; q 25 )4∞
4 5 3 5 2 2 2 5 3 3 5 4 4
× (a − 3q b) + q(a + 2q b ) + q (2a − q b ) + q (3a + q b ) + 5q .
So (3) becomes
X (q 25 ; q 25 )5∞
p(n)q n =
n≥0
(q 5 ; q 5 )6∞
× (a 4 − 3q 5 b) + q(a 3 + 2q 5 b2 ) + q 2 (2a 2 − q 5 b3 ) + q 3 (3a + q 5 b4 ) + 5q 4 .
X (q 25 ; q 25 )5∞
p(5n + 4)q 5n+4 = 5q 4 ,
n≥0
(q 5 ; q 5 )6∞
which is (1).
(−aq; q 2 )∞ = (1 + aq)(−aq 3 ; q 2 )∞ .
If we write
X
(−aq; q 2 )∞ = a k ck (q)
k≥0
2
qk
ck (q) = 2 2
(q ; q )k
and
X ak q k2
2
(−aq; q )∞ = 2; q 2)
.
k≥0
(q k
and this is the triple product identity. Note that the triple product can be written in the
equivalent forms
∞
2 +k)/2
X
3 2 3 3 3
−1
(a q; q )∞ (aq ; q )∞ (q ; q )∞ = (−1)k a k q (3k
k=−∞
and
∞
2 +k)/2
X
−1
(a ; q)∞ (aq; q)∞ (q; q)∞ = (−1)k a k q (k
k=−∞
under the substitutions (a, q) → (−aq (1/2) , q (3/2) ) and (a, q) → (−aq (1/2) , q (1/2) ) re-
spectively.
POSTSCRIPT. The “marvellous and powerful” extension of the triple product iden-
tity that I alluded to earlier is known technically as Ramanujan’s 1 ψ1 -summation.
I recommend that the reader interested in learning more about Ramanujan’s work
might begin by consulting Bruce Berndt’s book [1].
REFERENCES
1. B. C. Berndt, Number Theory in the Spirit of Ramanujan, Student Mathematical Library, vol. 34, American
Mathematical Society, Providence, RI, 2006.
2. G. H. Hardy, Ramanujan, Twelve Lectures on Subjects Suggested by His Life and Work, AMS Chelsea,
Providence, RI, 1999.
3. G. H. Hardy and E. M. Wright, An Introduction to the Theory of Numbers, Oxford University Press,
Oxford, 1960.
4. S. Ramanujan, Collected Papers, G. H. Hardy, P. V. Seshu Aiyar, B. M. Wilson, eds., AMS Chelsea,
Providence, RI, 2000.
School of Mathematics and Statistics, University of New South Wales, Sydney, Australia 2052
m.hirschhorn@unsw.edu.au
PROBLEMS
11600. Proposed by Michael D. Hirschhorn, University of New South Wales, Sydney,
Australia. Suppose a > 0, n ≥ 1, and 0 < r < a/n. For given θ, let
kr sin θ p
φk = arctan , ρk = a 2 − 2kra cos θ + k 2r 2 .
a − kr cos θ
Show that
Z ∞
cos(φ1 + · · · + φn ) − θ sin(φ1 + · · · + φn ) dθ π
2
= n.
0 ρ1 · · · ρn 1+θ 2a
11601. Proposed by Harm Derksen and Jeffrey Lagarias, University of Michigan, Ann
Arbor, MI. The Farey series of order n is the set of reduced rational fractions j/k in
the unit interval with denominator at most n. Let Fn be the product of these fractions,
excluding 0/1. That is,
n k−1
Y Y j
Fn = .
k=1 j=1
k
( j,k)=1
http://dx.doi.org/10.4169/amer.math.monthly.118.09.846
SOLUTIONS
P3 Q3
Using trigonometric formulas and the relation k=1 tan Ak = k=1 tan Ak , we obtain
Q3 ! 3 "! 3 "3 Q3
k=1 (tan A k + tan A k+1 ) Y sin(Ak + Ak+1 ) Y cos Ak cos Ak
P3 = = Q3k=1 2 .
( k=1 tan Ak )3 k=1
cos Ak cos Ak+1 k=1
sin Ak k=1 sin A k
P3 Q3 P3 (2)
2 1
Now use the relations sin A k = 4 cos(A k /2), sin A k = (3 −
P3 Q3k=1 k=1 k=1 2
k=1 cos(2A k )) = 2(1 + k=1 cos Ak ), and the sine law to obtain
If L denotes the left-hand side of inequality (1), then using (2) and (3), we obtain
!Q " !Q " 3
3 3
cos Ak k=1 (1 + cos A k )
Y (1 − cos Ak ) 1
L = Q3k=1 2 Q3 = Q3 .
k=1 sin Ak 1 + k=1 cos Ak k=1 cos Ak 1 + k=1 cos Ak
Thus inequality (1) holds if and only if 1/(1 + 3k=1 cos Ak ) ≥ 8/9, that is, if and
Q
only if 3k=1 cos Ak ≤ 1/8. This is a known inequality, but the proof is short: note that
Q
the function f given by f (x) = log(cos x) is concave on (0, π/2) because f ′′ (x) =
− cos−2 x < 0. Thus we have
3
! ! 3 ""
X 1X 1 1
log(cos Ak ) ≤ 3 log cos Ak = 3 log = log .
k=1
3 k=1
2 8
The required inequality follows. Equality holds when the triangle is equilateral.
A Symmetric Inequality
11492 [2010, 278]. Proposed by Tuan Le, student, Freemont High School, Anaheim,
CA. Show that for positive a, b, and c,
√ √ √
a 3 + b3 b3 + c3 c3 + a 3 6(ab + bc + ca)
+ + ≥ √ .
a 2 + b2 b2 + c2 c2 + a 2 (a + b + c) (a + b)(b + c)(c + a)
√ √
Solution by M. A. Prasad, India. First, note that ( a 3 + b3 )/(a 2 + b2 ) ≥ 1/ a + b.
Next, put
X p
T1 = (b + c)(c + a) (a + b + c), T2 = 6(ab + bc + ca),
Glaisher–Kinkelin
11494 [2010, 279]. Proposed by Ovidiu Furdui, Campia Turzii, Cluj, Romania. Let A
be the Glaisher-Kinkelin constant, given by
n
−n 2 /2−n/2−1/12 n 2 /4
Y
A = lim n e k k = 1.2824 . . . .
n→∞
k=1
Show that
∞ (−1)n−1
Y n! A3
√ = .
n=1
2π n(n/e)n 27/12 π 1/4
Rearranging gives
n
1 Y (2k)4k 22n(n+1) pn4
√ = √ ·
en (2n)! k=1 (2k − 1)2k−1 (2k)2k en (2n)! p2n
2 2
22n(n+1) A4 n 2n +2n+1/3 e−n A3
∼ √ · 2 2
=
(2n)n 4 4πn A(2n)2n +n+1/12 e−n π 1/4 27/12
as claimed.
Also solved by P. Bracken, B. S. Burdick, R. Chapman (U. K.), P. P. Dályay (Hungary), O. Geupel (Germany),
J. Grivaux (France), E. A. Herman, O. Kouba (Syria), V. Krasniqi (Kosovo), O. P. Lossers (Netherlands), B.
Mulansky (Germany), M. Omarjee (France), P. Perfetti (Italy), M. A. Prasad (India), P. .F. Refolio (Spain), J.
Schlosberg, J. Simons (U. K.), S. D. Smith, R. Stong, R. Tauraso (Italy), M. Tetiva (Romania), GCHQ Problem
Solving Group (U. K.), and the proposer.
Elementary Probability for Applications. By Rick Durrett. Cambridge University Press, Cam-
bridge, 2009, ix + 243 pp., ISBN 978-0-521-86756-6, $52.98.
Few areas of mathematics have more immediate connection to our life than proba-
bility. The past is gone, the present is ephemeral, and the future is just a guess. We
sharpen this guess through the lens of probability theory. Maybe it was easier back in
the days when we didn’t ask questions about the future, when fate was predetermined
by the gods.1 But eventually some French mathematicians got interested in the mathe-
matics of gambling. Over the next couple of centuries, probability became formalized,
axiomatized, and made profoundly useful. Unfortunately, none of this made it easier.
We have events and sample spaces; measurable (and nonmeasurable) sets; and condi-
tional, marginal, and Bayesian probabilities. We can formally define probability, but
we aren’t quite sure what it means. Is the probability associated with pulling balls out
of (Polya’s) urn the same as the probability that it will rain tomorrow? What about
those Bayesians who think of probability as belief? How can knowing the gender of
one of the twins affect the probability that the other is a girl? Even what we call “el-
ementary probability” confuses the mathematics intelligentsia. Consider, for example,
the Monte Hall problem and Marilyn vos Savant controversy [5]. No wonder students
struggle in their first probability course.
The challenge is to present probability to students in such a way that they under-
stand the fundamental concepts, appreciate the beauty and power of the subject, and—
most of all—learn to think probabilistically. There is no shortage of probability books,
ranging from texts aimed at everyone from elementary school students to professional
mathematicians, to hundreds of popular books professing to describe the uses (and
misuses) of probability theory to wider audiences (e.g., Taleb’s “Black Swan” [6]).
With all of these options available, is Durrett’s Elementary Probability for Applica-
tions really useful? Probably.
A clue to Durrett’s approach is in the title and his choice of the word “for” instead
of the usual “and.” Durrett is well acquainted with applied probability. His research
spans areas as diverse as random graphs, spatial models, and genetics, and his text
Probability: Theory and Examples [3] is standard reading for graduate students. Dur-
rett believes in the need to motivate probability through and for applications, partic-
ularly via applications that are (hopefully) relevant to the students. In this text under
review, along with the usual war horses (polling, urns, coin-flipping, and Bayes’ rule),
http://dx.doi.org/10.4169/amer.math.monthly.118.09.854
1 A nice description of this can be found in Mlodinow’s The Drunkard’s Walk: How Randomness Rules Our
Lives [4].
To interpret our result note that the probability . . . is roughly 1/22,000 so even if
there were 220 players with 0.325 batting averages, it would take 100 years for
this to occur again.
From this, one can infer that this DiMaggio guy must have been a pretty good
ballplayer.
Another worthwhile feature of this text is Durrett’s way of interspersing more ad-
vanced mathematical concepts into elementary discussions. For example, while still
flipping coins to explain World Series winners, he mentions the chi-squared statistic,
almost as an aside. This allows Durrett to presage the notion of a p-value and hy-
pothesis testing. More sophisticated ideas, such as Stirling numbers, also pop up here
and there. There is a potential downside to this approach; topics that aren’t absolutely
necessary to the logical development of an idea could be a distraction and cause some
students to conflate key ideas with “extra for experts” concepts. Durrett trusts that the
instructor is capable of making this distinction.
The book is organized in a somewhat alternative way. Chapter One, Basic Concepts,
introduces the notion of probability via coin flipping and sports analogies. In short or-
k
Let A1 , . . . , Ak be disjoint events whose union ∪i=1 Ai = . Suppose we per-
form the experiment a random number of times N , where N has Poisson dis-
tribution with mean λ, and let X i be the number of times Ai occurs. [Then]
X 1 , . . . , X k are independent Poissons with parameters λ P(Ai ).
This is made tangible with an example to which the students can relate.
Chapter Four, Markov Chains, is creative and artfully done, but could be challeng-
ing for students with only a calculus background. The chapter starts innocently enough
with an example-motivated introduction to Markov chains. Very quickly, however, the
level of mathematical sophistication (especially the notation) increases dramatically.
Durrett admirably works to discuss topics such as stationary distributions, limit behav-
ior of Markov chains, and absorbing chains. It will take a nimble instructor to dance
between some pretty serious theory and the practical applications of Markov chains.
For example, I question how many general education students will appreciate a theo-
rem whose short version is the sentence “the transition probability converges to equi-
librium exponentially fast.” Not having the notion of an eigenvector makes Durrett’s
discussion of stationary distributions awkward, but not unworkable. He depends heav-
ily on two- and three-state examples, supplemented by TI-83 calculator instructions.
The motivating examples are excellent: Ehrenfest chains from statistical mechanics,
the Wright-Fisher model from genetics, Monopoly (the author missed a chance to cite
the definitive work (!) on this subject [1]), inventories, and gambler’s ruin (required,
if the Blackjack section from Chapter Three is covered). Durrett deserves credit for
bringing a sophisticated approach to Markov chains to an audience with a relatively
limited mathematical background.
Next comes the aforementioned Chapter Five, Continuous Distributions. As adver-
tised, relative to what has come before, this is pretty boring and perfunctory—gone is
the spark that infuses the first four chapters. This material will be pretty intimidating
to the average calculus-only student, and completely unapproachable to students tak-
ing this course for their general education mathematics credits. No doubt, there’s a lot
of good content here: density and distribution functions, expected values, functions of
random variables, and joint/marginal/conditional distributions. The standard continu-
ous distributions, such as uniform and exponential, are discussed. The all-important
normal distribution is not mentioned in this chapter; it will appear later. Overall, the
level of discussion is extremely formal and more sophisticated than what has come
before. My advice: follow Durrett’s advice and skip this chapter.
Chapter Six, Limit Theorems, introduces the various laws of large numbers, the
normal distribution, and the central limit theorem. Arguably, all students of mathemat-
ics, whether they are taking probability as a general education requirement or on the
way to a mathematics major, deserve to encounter these fundamental ideas about ran-
domness. The level of sophistication here is, by necessity, higher than what has come
before. This chapter does not depend on knowing anything about continuous distribu-
tions and it is quite believable that a student who has been carefully and intuitively led
through the first four of Durrett’s chapters can really appreciate these advanced top-
ics. It helps that the earlier spark has returned; Durrett really wants to get these ideas
across. You can tell from the whimsical way he presents his examples: “Assuming that
dog bites are a rare event, . . . ” and “Is there a difference between baseball and flip-
ping coins?” Depending on the audience, the instructor will have to make some hard
choices in this chapter. For example, the proofs of Chebyshev’s inequality and the law
of large numbers, though short, are extremely sophisticated. I like the way the central
limit theorem is presented, but not proven, and then followed with a short section on
statistics. This chapter is worthwhile and accessible, especially if the instructor builds
up some good will beforehand.
REFERENCES
1. S. Abbott and M. Richey, Take a walk on the boardwalk, College Math J. 28 (1997) 162–171. http:
//dx.doi.org/10.2307/2687519
2. R. Durrett, The Essentials of Probability, Duxbury Press, Belmont, CA, 1994.
3. , Probability: Theory and Examples, 4th ed., Cambridge University Press, Cambridge, 2010.
4. L. Mlodinow, The Drunkward’s Walk: How Randomness Rules Our Lives, Pantheon, New York, 2008.
5. J. Rosenhouse, The Monty Hall Problem: The Remarkable Story of Math’s Most Contentious Brain Teaser,
Oxford University Press, New York, 2009.
6. N. Taleb, The Black Swan: The Impact of the Highly Improbable, Random House, New York, 2007.
www.cambridge.org/us/mathematics
800.872.7423
CALL FOR NOMINATIONS
NEMMERS PRIZE IN
MATHEMATICS
$200,000
Northwestern University invites
nominations for the Frederic Esser
Nemmers Prize in Mathematics to be
awarded during the 2011-12 academic
RECIPIENTS OF year. The award includes payment to the
RECIPIENTS OF recipient of $200,000. Made possible by a
THE FREDERIC ESSER
THE FREDERIC ESSER generous gift to Northwestern by the late
NEMMERS PRIZE Erwin Esser Nemmers and the late Frederic
NEMMERS PRIZE
IN MATHEMATICS Esser Nemmers, the award is given every
IN MATHEMATICS other year.
Candidacy for the Nemmers Prize in
Mathematics is open to those with careers
of outstanding achievement in mathematics
JOHN H. CONWAY as demonstrated by major contributions to
JOHN H. CONWAY
SIMON DONALDSON new knowledge or the development of
SIMON DONALDSON significant new modes of analysis.
MIKHAEL GROMOV Individuals of all nationalities and
MIKHAEL GROMOV institutional affiliations are eligible except
JOSEPH B. KELLER current or recent members of the
JOSEPH B. KELLER
ROBERT P. LANGLANDS Northwestern University faculty and
ROBERT P. LANGLANDS recipients of the Nobel Prize.
YURI I. MANIN The recipient of the 2012 Nemmers Prize
YURI I. MANIN in Mathematics will deliver a public lecture
YAKOV G. SINAI and participate in other scholarly activities
YAKOV G. SINAI
TERENCE TAO at Northwestern University for 10 weeks
TERENCE TAO during the 2012-13 academic year.
EDWARD WITTEN Nominations for the Frederic Esser
EDWARD WITTEN Nemmers Prize in Mathematics will be
accepted until December 1, 2011.
Nominating letters of no more than three
pages should describe the nominee’s
professional experience, accomplishments,
and qualifications for the award. A brief
curriculum vitae of the nominee is helpful
but not required. Nominations from
experts in the field are preferred to
institutional nominations; direct
applications will not be accepted.
www.northwestern.edu/provost/awards/nemmers
Northwestern University is an equal opportunity, affirmative action educator
and employer.
MATHEMATICAL ASSOCIATION OF AMERICA 1529 Eighteenth St., NW • Washington, DC 20036
f o r P a pers
Call
Publishing outstanding research in applied and
computational mathematics by undergraduates
Society for Industrial and Applied Mathematics • 3600 Market Street, 6th Floor
Philadelphia, PA 19104-2688 USA • +1-215-382-9800
Fax: +1-215-386-7999 • siuro@siam.org • www.siam.org/students/siuro