Noiembrie 2013

THE AMERICAN MATHEMATICAL
MONTHLY
VOLUME 120, NO. 9 NOVEMBER 2013
Christiane’s Hair 771

Jacques Lévy Véhel and Franklin Mendivil
What to Expect in a Game of Memory 787

Daniel J. Velleman and Gregory S. Warrington
Mixing Problems with Many Tanks 806

Antonı́n Slavı́k
Reciprocal Sums as a Knowledge Metric: Theory, 822

Computation, and Perfect Numbers
Jonathan Bayless and Dominic Klyve
NOTES
Certain Inequalities Associated with the Divisor Function 832

Jorge Luis Cimadevilla Villacorta
On the Nonexistence of Certain Limits for the 837

Complex Exponential
Charles S. Kahane
A Generalization of Routh’s Triangle Theorem 841

Árpád Bényi and Branko Ćurgus
A Maximum Principle for High-Order Derivatives 846

David Pan
Nothing New about Equiangular Polygons 849

Gerhard J. Woeginger
Quotients of Gaussian Primes 851

Stephan Ramon Garcia
PROBLEMS AND SOLUTIONS 854
REVIEWS
Calculus: Modeling and Application 862

By David A. Smith and Lawrence C. Moore
Betty Mayfield
An Official Publication of the Mathematical Association of America

MAA’s New
Membership
Structure
Member Plus
$169/Year
$249/Year
Individual Member Benefits
$35/Year
Member
Student
Electronic Subscriptions to:
• The American Mathematical Monthly (10 per year) X X X
• The College Mathematics Journal (5 per year) X X X
• Mathematics Magazine (5 per year) X X X
• Math Horizons (4 per year) X X X
• MAA FOCUS (6 per year) X X X
Print Subscription to MAA FOCUS (6 per year) X X
Regional Section Membership (USA/CAN Members) X X X
Discounted Registration for MAA MathFest X X X
Discounted Registration for Joint Mathematics Meetings X X X
MAA Math Alert—the e-newsletter for MAA members X X X
Member Discounts on MAA books, videos & merchandise X X X
Member Discounts on Print Subscriptions to:

• The American Mathematical Monthly ($50/year) X X X
• The College Mathematics Journal ($25/year) X X X
• Mathematics Magazine ($25/year) X X X
• Math Horizons ($20/year) X X X
Access to MAA Affinity Program Discounts X X X
Eligibility to Join MAA SIGMAAs ($12 each) X X X
Access to MAA columns, blogs, and online resources X X X
Membership in up to 3 SIGMAAs of your choice X
Print Subscription to Math Horizons X
MAA Member Plus Book Selection X
MAA Annual Design Contest T-shirt X
Be sure to visit www.maa.org for all the details!

THE AMERICAN MATHEMATICAL
MONTHLY
Volume 120, No. 9 November 2013
EDITOR
Scott T. Chapman
Sam Houston State University
NOTES EDITOR BOOK REVIEW EDITOR

Sergei Tabachnikov Jeffrey Nunemacher
Pennsylvania State University Ohio Wesleyan University
PROBLEM SECTION EDITORS

Douglas B. West Gerald Edgar Doug Hensley
University of Illinois Ohio State University Texas A&M University
ASSOCIATE EDITORS
William Adkins Ulrich Krause
Louisiana State University Universität Bremen
David Aldous Jeffrey Lawson
University of California, Berkeley Western Carolina University
Elizabeth Allman C. Dwight Lahr
University of Alaska, Fairbanks Dartmouth College
Jonathan M. Borwein Susan Loepp
University of Newcastle Williams College
Jason Boynton Irina Mitrea
North Dakota State University Temple University
Edward B. Burger Bruce P. Palka
Williams College National Science Foundation
Minerva Cordero-Epperson Vadim Ponomarenko
University of Texas, Arlington San Diego State University
Beverly Diamond Catherine A. Roberts
College of Charleston College of the Holy Cross
Allan Donsig Rachel Roberts
University of Nebraska, Lincoln Washington University, St. Louis
Michael Dorff Ivelisse M. Rubio
Brigham Young University Universidad de Puerto Rico, Rio Piedras
Daniela Ferrero Adriana Salerno
Texas State University Bates College
Luis David Garcia-Puente Edward Scheinerman
Sam Houston State University Johns Hopkins University
Sidney Graham Susan G. Staples
Central Michigan University Texas Christian University
Tara Holm Dennis Stowe
Cornell University Idaho State University
Roger A. Horn Daniel Ullman
University of Utah George Washington University
Lea Jenkins Daniel Velleman
Clemson University Amherst College
Daniel Krashen
University of Georgia
EDITORIAL ASSISTANT
Bonnie K. Ponce
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:
Submission of articles, notes, and filler pieces is re-
quired via the MONTHLY’s Editorial Manager System. The Mathematical Association of America
Initial submissions in pdf or LATEX form can be sent 1529 Eighteenth Street, N.W.
to the Editor Scott Chapman at Washington, DC 20036
http://www.editorialmanager.com/monthly Recent copies of the MONTHLY are available for pur-

chase through the MAA Service Center.
maahq@maa.org, 1-800-331-1622
The Editorial Manager System will cue the author
for all required information concerning the paper. Microfilm Editions: University Microfilms Interna-
Questions concerning submission of papers can tional, Serial Bid coordinator, 300 North Zeeb Road,
be addressed to the Editor at monthly@shsu.edu. Ann Arbor, MI 48106.
Authors who use LATEX are urged to use arti-
cle.sty, or a similar generic style, and its stan- The AMERICAN MATHEMATICAL MONTHLY (ISSN
dard environments with no custom formatting. 0002-9890) is published monthly except bimonthly
A formatting document for MONTHLY references June-July and August-September by the Mathe-
can be found at http://www.shsu.edu/~bks006/ matical Association of America at 1529 Eighteenth
FormattingReferences.pdf. Follow the link to Elec- Street, N.W., Washington, DC 20036 and Lancaster,
tronic Publications Information for authors at http: PA, and copyrighted by the Mathematical Asso-
//www.maa.org/pubs/monthly.html for informa- ciation of America (Incorporated), 2013, including
tion about figures and files, as well as general edi- rights to this journal issue as a whole and, except
torial guidelines. where otherwise noted, rights to each individual
Letters to the Editor on any topic are invited. contribution. Permission to make copies of individ-
Comments, criticisms, and suggestions for mak- ual articles, in paper or electronic form, including
ing the MONTHLY more lively, entertaining, and posting on personal and class web pages, for ed-
informative can be forwarded to the Editor at ucational and scientific use is granted without fee
monthly@shsu.edu. provided that copies are not made or distributed
for profit or commercial advantage and that copies
The online MONTHLY archive at www.jstor.org is a bear the following copyright notice: [Copyright the
valuable resource for both authors and readers; it Mathematical Association of America 2013. All rights
may be searched online in a variety of ways for any reserved.] Abstracting, with credit, is permitted. To
specified keyword(s). MAA members whose institu- copy otherwise, or to republish, requires specific
tions do not provide JSTOR access may obtain indi- permission of the MAA’s Director of Publications and
vidual access for a modest annual fee; call 800-331- possibly a fee. Periodicals postage paid at Washing-
1622. ton, DC, and additional mailing offices. Postmaster:
See the MONTHLY section of MAA Online for current Send address changes to the American Mathemati-
information such as contents of issues and descrip- cal Monthly, Membership/Subscription Department,
tive summaries of forthcoming articles: MAA, 1529 Eighteenth Street, N.W., Washington, DC,
20036-1385.
http://www.maa.org/
Christiane’s Hair
Jacques Lévy Véhel and Franklin Mendivil
Abstract. We explore the geometric and measure-theoretic properties of a set built by stacking
central Cantor sets with continuously varying scaling factors. By using self-similarity, we
are able to describe its main features in a fairly complete way. We show that it is made of
an uncountable number of analytic curves, compute the exact areas of the gaps of all sizes,
and show that its Hausdorff and box-counting dimensions are both equal to 2. It provides a
particularly good example to introduce and showcase these notions because of the beauty and
simplicity of the arguments. Our derivation of explicit formulas for the areas of all of the gaps
is elementary enough to be explained to first-year calculus students.
1. INTRODUCTION. Consider the beautiful and striking set illustrated below (let
us call the set CH).
Clearly, this set has an intricate recursive (fractal) structure. In fact, it is constructed
by “stacking” Cantor sets (in a way that we will describe momentarily). We can also
see that CH is composed of uncountably many “strands.” Our name for the set (and this
paper) was inspired by the resemblance of CH to the braided hair of the first author’s
wife, Christiane.
Here we explore some geometric and measure-theoretic properties of this set. We
hope to convince you that the elegance and simplicity of the geometrical arguments
are just as striking as the set itself.
2. CONSTRUCTION OF CH. First we must describe the construction of CH. The

basic construction is very simple. Above, we said that CH is composed of a “stack” of
Cantor sets, so first we meditate on so-called “central” Cantor sets, with the classical
Cantor set as the most famous example (also called Smith–Cantor sets, as Henry Smith
had described similar sets independently of Cantor [6, p. 45]). The classic Cantor set is
constructed by starting with the unit interval I = [0, 1] and then removing the middle
1/3 (as an open interval) to obtain the two closed intervals I0 = [0, 1/3] and I1 =
http://dx.doi.org/10.4169/amer.math.monthly.120.09.771
MSC: Primary 28A80, Secondary 28A78
November 2013] CHRISTIANE’S HAIR 771

[2/3, 1]. Repeating for each of the closed intervals, we remove the open middle 1/3
to obtain four closed intervals I00 = [0, 1/9], I01 = [2/9, 1/3], I10 = [2/3, 7/9], and
I11 = [8/9, 1]. After the nth iteration, we have 2n closed intervals of length 3−n . The
union of these 2n intervals together comprise a set Cn . Noticing that {Cn } is a nested
collection of non-empty compact sets, we then see that C = ∩n Cn is non-empty and
compact; this is the classical Cantor set, also called the “middle 1/3-Cantor set.” This
construction is illustrated in Figure 1.
Figure 1. Iterative construction for the classical Cantor set C
In this figure, we can see the binary structure of C clearly. In fact, C is homeomor-
phic to the countably infinite product of the two-point discrete space {0, 1}. We can
think of a point x ∈ C as resulting from an infinite sequence of choices of left or right.
Each such sequence of choices selects some nested sequence of closed intervals, one
from each stage of the construction. The intersection of the resulting nested intervals
is always a single point. As an example, choosing the left interval at each stage will
result in the point x = 0.
It is simple to modify the construction of C where we remove some other (fixed)
ratio of the length at each stage. We shall denote by C y the set obtained by removing
the length 1 − 2y in the middle of [0, 1] at the first stage, where y ranges in [0, 1/2]
(the reason for our funny choice of 1 − 2y for the gap length will become evident
below). At the nth stage, we have 2n−1 remaining closed intervals, each of length y n−1 ,
and we remove a length of (1 − 2y)y n−1 from the middle of each interval. We build in
this way a continuum of Cantor sets, where the size of the gaps at each stage decrease.
When y = 1/2, we do not remove anything, so the resulting set is just the interval
[0, 1], while for y = 0, the resulting set is reduced to the two-point set {0, 1}. The
vertical stacking of the C y for y from 0 to 1/2 is our set CH.
More formally,
CH = (x, y) : x ∈ C y , y ∈ [0, 1/2] ⊂ [0, 1] × [0, 1/2].

(1)
Another, more useful way of constructing the classical Cantor C set is by using an
Iterated Function System (IFS) [1, 7]. Consider the two functions
w0 (x) = x/3 and w1 (x) = x/3 + 2/3.
Notice that w0 (C ) = C ∩ [0, 1/3] and w1 (C ) = C ∩ [2/3, 1], so that
C = w0 (C ) ∪ w1 (C ), (2)
with the union being disjoint. This self-tiling or self-similarity property uniquely de-
fines C in that if A ⊂ R is any non-empty compact set with A = w0 (A) ∪ w1 (A),
then it must be the case that A = C . Moreover, the set-valued mapping Ŵ given by
Ŵ (A) = w0 (A) ∪ w1 (A) is contractive in the Hausdorff metric, and thus we have that
Ŵ n (A) converges to C for any non-empty compact A ⊂ R. Note that Figure 1 illus-
trates this convergence with the initial set A = [0, 1]. The first line in the figure shows
772 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 120

A = [0, 1], the second line shows Ŵ (A) = [0, 1/3] ∪ [2/3, 1], the third line shows
Ŵ 2 (A) = [0, 1/9] ∪ [2/9, 1/3] ∪ [2/3, 7/9] ∪ [8/9, 1], and so on.
The self-similarity and uniqueness properties of C also allow us to give another
useful description of it. Consider the set
(∞ )
X
A= tn 3 : tn ∈ {0, 2} .
−n
(3)
n=1
Then it is not so hard to see that

( )
X
w0 (A) = tn 3−n : t1 = 0, tn ∈ {0, 2} for n ≥ 2
n≥1
and
( )
X
w1 (A) = tn 3−n : t1 = 2, tn ∈ {0, 2} for n ≥ 2 ,
n≥1
so A = w0 (A) ∪ w1 (A). However, this means that A = C and so (3) gives an explicit
description of C . For the more general situation below it will be useful to notice that
(∞ ) ( )
X X
tn 3−n : tn ∈ {0, 2} = 2 bn 3−n : bn ∈ {0, 1} .
n=1 n≥1
The sets C y introduced above are also obviously associated to iterated functions
systems. For y ∈ [0, 1/2], we define the two maps
w0 (x, y) = yx and w1 (x, y) = yx + (1 − y) . (4)
Thus, y = 1/3 yields the classical Cantor set, while y = 1/4 yields a middle-1/2
version of it.
Each C y is the unique invariant set under the two maps given in (4). The entire set
CH is invariant under these two maps as well, where we now think of these maps as
w0 , w1 : [0, 1] × [0, 1/2] → [0, 1] × [0, 1/2], abusing notation slightly. The reason
we label the sets C y is that y is the contraction factor for the IFS (4). Labeling each
Cantor set in the “stack” by its central gap length would result in messier formulas for
the IFS (4).
Comments. The set CH is a nice illustration of how the interval [0, 1] has been
“ripped apart” to form the classical Cantor set. We see at all the points with two binary
representations, that is, all the numbers of the form i/2 j , that the interval [0, 1] has
been cut and a gap has been inserted. More specifically, a gap of length 3−n has been
inserted at each point of the form i/2n for 1 ≤ i ≤ 2n − 1. It is best to think of this
construction in stages, as with the usual construction. First, we cut at x = 1/2, scale
each half by a factor of 2/3, and insert a gap of length 1/3. The scaling preserves the
total length. Then, we cut at the points that originally had coordinates 1/4 and 3/4
(now their coordinates have changed), scale each part by 2/3, and insert gaps of length
1/9. This is illustrated in CH where a gap originates at each dyadic point on the top
line, when they are inserted.

The IFS (4) for CH is not contractive and has many compact invariant sets on the
space [0, 1] × [0, 1/2]. In fact, it has uncountably many. Simply take any closed subset
S ⊂ [0, 1/2] and obtain an invariant set of the form
(x, y) : x ∈ C y , y ∈ S .

The set CH is the maximal compact invariant subset of [0, 1] × [0, 1/2] in the sense
that it contains any other invariant compact subset.
We make no claim to having discovered CH. For instance, Mandelbrot’s book [10]
contains a version of it on page 81.
3. AREAS OF THE “GAPS” OF CH. The starting point for this paper was a sur-
prising (to us) observation about the areas of the “gaps” of CH. In this short section we
explain this pretty little geometric fact, which is simple enough that it can be explained
to calculus students.
Consider the set CH as enclosed in the box [0, 1] × [0, 1/2]. We notice that the
central “gap” (or void) is a triangle with base length equal to one and height equal to
1/2. (It is worthwhile spending a little bit of time understanding why it is actually a
triangle.)
Remarkably, it is possible to exactly compute the areas of all the other gaps, even
though their sides are complicated curves. For instance, each of the two “stage-two”
gaps (the images of the central gap) have area
Z y=1/2
1
(1 − 2y)y dy = .
y=0 24
How does this formula arise? We see that for any given y, the set C y has a central
gap of length 1 − 2y. Because of the self-similarity of C y under the two maps given
in (4), the common length of the next largest gaps (one on either side of the central
gap) is equal to y(1 − 2y). Thus, the area of either one of these “stage-two” gaps is
the integral of this length over y.
In a similar way, the “nth-stage” gaps all have area equal to
y=1/2
2−n
Z
(1 − 2y)y n−1 dy = .
y=0 n(n + 1)
We can check that the sum of these areas is, in fact, equal to 1/2, the entire area of the
rectangle. Computing this sum is simple, as it is a telescoping sum:
X 2−n 1X1 1 1
2n−1 = − = .
n≥1
n(n + 1) 2 n≥1 n n + 1 2
Thus CH has Lebesgue measure zero.
4. EACH STRAND OF “HAIR” IS A SMOOTH CURVE. CH is comprised of

an uncountable collection of continuous curves, each of which goes from some point
on the top horizontal line down to one of the two points (0, 0) or (1, 0). All of the
curves (the “hairs” or “strands”) that you can see in the image are actually smooth,
in fact polynomials (the horizontal, x, coordinate is a polynomial function of y). To
understand this, just notice that each curve that bounds a “gap” is the image of one of

the two diagonal edges of the central triangle under some finite composition of the two
IFS maps. We say more about this in this section.
We will prove that each strand is a C ∞ curve, and also that each strand is an ana-
lytic curve. Clearly, knowing that they are analytic curves imply that they are C ∞ , so
why present both proofs? We are interested in more than just arriving at the strongest
results; we are mainly interested in presenting engaging mathematics, and we believe
both proofs are appealing. The proof that strands are C ∞ is a good illustration of
a standard technique in fractals. The argument defines an IFS on functions and their
derivatives and shows that this “vector IFS” process converges uniformly, thus yielding
the result. The proof that the strands are analytic uses an explicit power series repre-
sentation of the strand functions. This argument also uses IFS, but in a different way.
The collection of strands is not, in fact, indexed by the points on the top line (the
points in [0, 1]), but is indexed by infinite binary sequences. One clue to this is that the
two sides of the central triangle both originate at x = 1/2. What is special about x =
1/2 is that it has two binary representations, so we need to consider them separately;
each representation leads to its own strand. As the Cantor set has a binary structure,
this is not surprising.
To describe the strands, we need to establish some notation. For any n ∈ N, we
set 6 n = {0, 1}n as the space of length n binary sequences. Further, 6 = {0, 1}N , and
for σ ∈ 6 we define σ n ∈ 6 n as σ n = (σ1 , σ2 , . . . , σn ), the truncation of σ to the
first n places. Given two functions φ0 and φ1 and σ ∈ 6 n , we define the nth order
composition φσ by
φσ := φσ1 ◦ φσ2 ◦ · · · ◦ φσn . (5)
The order of composition in (5) is very important. Notice the difference between
φσ n = φσ1 ◦ φσ2 ◦ · · · ◦ φσn
and
φσ n+1 = φσ1 ◦ φσ2 ◦ · · · ◦ φσn ◦ φσn+1
for a given σ ∈ 6. The “newest” map is applied on the “inside.”

We consider everything as a function of y (since we are trying to show that all
the “hairs” are smooth functions of y). Acting on a function f : [0, 1/2] → [0, 1],
x = f (y), we have that the two IFS maps from (4) are given as
φ0 ( f )(y) = y f (y) and φ1 ( f )(y) = y f (y) + 1 − y, for y ∈ [0, 1/2].
Notice that, by the definition of φi , we have φi ( f ) : [0, 1/2] → [0, 1] whenever f :

[0, 1/2] → [0, 1]. Since y ∈ [0, 1/2], both φ1 and φ2 are contractions on C[0, 1/2]
in the uniform norm with contractivity 1/2. Furthermore, for any σ ∈ 6 n , the con-
tractivity of φσ is 2−n . In addition, for σ ∈ 6 and m > n ≥ 1, we have |φσ m ( f )(y) −
φσ n ( f )(y)| ≤ 2−n . Thus, for any fixed σ ∈ 6, we have that the limit
Tσ (y) := φσ ( f )(y) = lim φσ n ( f )(y) (6)

n
exists and is uniform in y (since the contraction factor is uniformly bounded in y). For
the “hairs,” our starting functions are f 0 (y) = 0 for all y (the left edge) and f 1 (y) = 1
for all y. The two edges of the central triangle are φ0 ( f 1 )(y) = y and φ1 ( f 0 )(y) =
1 − y. Then the boundaries of the next gap (at “stage” 2) are the four curves:

φ0 (φ0 ( f 1 ))(y) = y 2 ,
φ0 (φ1 ( f 0 ))(y) = y − y 2 ,
φ1 (φ0 ( f 1 ))(y) = y 2 + 1 − y,
φ1 (φ1 ( f 0 ))(y) = 1 − y 2 .
From these considerations, it is clear that all the boundary curves of any of the nth
stage “gaps” are polynomials of degree n. The strand associated with σ ∈ 6 is the
uniform limit in (6) and is thus a continuous function. We show that all of the deriva-
tives converge uniformly as well. To do this, just notice that
(φ0 ( f ))0 (y) = y f 0 (y) + f (y) and (φ1 ( f ))0 (y) = y f 0 (y) + f (y) − 1. (7)
Continuing on to the second derivative, we see that
(φ0 ( f ))00 (y) = y f 00 (y) + 2 f 0 (y) and (φ1 ( f ))00 (y) = y f 00 (y) + 2 f 0 (y), (8)
and the nth derivatives mappings for n > 1 are
(φ0 ( f ))(n) (y) = y f (n) (y) + n f (n−1) (y) = (φ1 ( f ))(n) (y). (9)
This means that we have two linear mappings on the function value and its first n
derivatives, which are given by
f y 0 0 ··· 0 0 f
    
 f0  1 y 0 ··· 0 0 f  0
f 00  0 2 y ··· 0 0   f 00 
    
80 

.. =.
 . .. .. . . .. ..   .. 
  (10)

 .  . . . . . .  . 

(n−1)  0 (n−1) 
 f 0 0 ··· y 0   f
f (n) 0 0 0 ··· n y f (n)
and
f y 0 0 ··· 0 0 f 1−y
      
 f0  1 y 0 ··· 0 0   f 0   −1 
f 00  0 2 y ··· 0 0   f 00   0 
      
81  .

.. =.
 . .. .. . . .. ..   ..   .. 
   +  (11)

 .  . . . . . .  .   . 
 f (n−1)   0 0 0 ··· y 0   f (n−1)   0 
f (n) 0 0 0 ··· n y f (n) 0
It is clear that both 80 and 81 are contractive (recall y ≤ 1/2), and thus the first n
derivatives converge uniformly as well. Since this is true for any n, Tσ (y) = φσ ( f )(y)
is a C ∞ function of y for any σ ∈ 6.
We note that it is easy to show from (7) that |Tσ0 (y)| ≤ 2 for all y and σ ∈ 6. This
fact will be important in section 5.
Thread functions are analytic. A completely different approach to the thread func-
tions shows that these functions are, in fact, real analytic.

The idea is based on the representation given in (3) for the special case of the clas-
sical Cantor set with y = 1/3. For y ∈ (0, 1/2], we claim that
( )
X
C y = (1 − y)y −1
σn y : σ ∈ 6 .
n
(12)
n≥1
We use the self-similarity of C y to show this.

Let the set defined on the right-hand side P of (12) be denoted by Sy . As a set of
subsums of the convergent geometric series n y n , Sy is a non-empty and compact set
(see [5, 8, 11]). Furthermore,
( )
X
w0 (Sy ) = (1 − y)y −1
σn y n+1
:σ ∈6
n≥1
( )
X
= (1 − y)y −1
αn y : α ∈ 6, α1 = 0 ,
n
n≥1
and
( )
X
w1 (Sy ) = (1 − y) + (1 − y)y −1
σn y n+1
:σ ∈6
n≥1
( )
X
= (1 − y)yy −1
+ (1 − y)y −1
σn y n+1
:σ ∈6
n≥1
( )
X
= (1 − y)y −1 αn y n : α ∈ 6, α1 = 1 .
n≥1
Thus Sy = w0 (Sy ) ∪ w1 (Sy ), and so Sy = C y . This means that the “thread function”
T : 6 × [0, 1/2] → [0, 1] is given by
X
Tσ (y) = (1 − y)y −1 σn y n . (13)
n≥1
For a fixed σ ∈ 6, this is a real-analytic function for y ∈ (0, 1/2].
5. GEOMETRIC MEASURE-THEORETIC PROPERTIES OF CH. We now

turn to deeper geometric properties of CH, in particular the Hausdorff measure and
dimension and the box-counting dimension. We discuss some very basic background
on these two topics. For a more complete discussion, we suggest that the reader con-
sult the books [3, 4, 9, 12, 13]. In particular, chapter 3 of Falconer’s book [3] has a
very nice general discussion about dimensions. In our brief overview, we will provide
some of the simpler arguments to give the reader a feel for these dimensions, but we
will omit the more complicated ones to keep our discussion to a reasonable length. A
full understanding of these dimensions is not necessary to appreciate our discussion,
so the reader should feel safe in skipping some background details.
We start with the Hausdorff dimension, even though the box-counting dimension is
more elementary. In our case, the Hausdorff dimension is easier to compute.

5.1. The Hausdorff dimension of CH. First, we remind the reader about the defini-
tion and one or two simple properties of the Hausdorff s-dimensional measures, which
we denote by h s . We use |A| for the diameter of A ⊂ Rd . For δ > 0, a δ-cover of A
is a countable cover {Ui } of A with |Ui | ≤ δ for each i. Given a Borel set A ⊂ Rd ,
we define
( )
X
h δ (A) = inf
s
|Ui | : {Ui } δ-cover of A
s
and h s (A) = lim h sδ (A).
δ→0
i
From standard results in measure theory, h s is a Borel measure. Furthermore, it is

possible to show that for integer values of n, the Hausdorff measure h n is a constant
multiple of n-dimensional Lebesgue measure. The definition of h s is guided by the
same intuition as for Lebesgue measure, but with the “size” of a “basic” set U given
by |U |s (rather than its n-dimensional volume).
We see that if s < t, then h sδ (A) ≥ δ s−t h tδ (A). This implies that if h t (A) > 0, then
h (A) = +∞. Therefore, there is a special value, denoted by dim H (A) and called the
s
Hausdorff dimension of A, such that for 0 ≤ s < dim H (A) we have h s (A) = +∞, and
for s > dim H (A) we have h s (A) = 0. Any A ⊂ Rd with non-empty interior has Haus-
dorff dimension d, so this notion agrees with our intuitive idea of dimension for nice
sets. However, unlike our intuitive notion of dimension, it is certainly possible for sets
to have fractional Hausdorff dimension. As an example, dim H (C y ) = − ln(2)/ ln(y).
The Hausdorff dimension is monotone (A ⊆ B implies dim H (A) ≤ dim H (B)) and
countably stable (dim H (∪i Ai ) = supi dim H (Ai )). Sets with 0 < h s (A) < ∞ are
called s-sets and have been extensively studied. In fact, any generalized Cantor set is
an s-set for an appropriate dimension function [2].
The measure h s has a nice behaviour under Lipschitz mappings, which will be very
important for us. Let f : Rd → Rd be Lipschitz (that is, k f (x) − f (y)k ≤ K kx − yk
for some constant K ), then | f (U )|s ≤ K s |U |s and thus h s ( f (A)) ≤ K s h s (A). In par-
ticular, dim H ( f (A)) ≤ dim H (A).
dim H (CH) = 2. We now show that the Hausdorff dimension of CH is equal to two.
One interesting feature is that “locally” the Hausdorff dimension of CH is strictly less
than two everywhere except in neighborhoods of the top line. Locally, CH is close to
being a product of a Cantor set (in the horizontal direction) with an interval (in the
vertical direction). The local geometry of CH varies considerably from top to bottom.
Clearly, the Hausdorff dimension of CH is at least one, since it contains many
smooth curves any of which have dimension equal to one. We show that it is actually
equal to two. For [a, b] ⊂ (0, 1/2) let K ab = ([0, 1] × [a, b]) ∩ CH. We show that
dim H (K ab ) ≥ 1 − ln(2)/ ln(a). By 7.2 in [3] we know that
dim H (Ca × [a, b]) ≥ dim H (Ca ) + dim H ([a, b]) = − ln(2)/ ln(a) + 1,
an intuitively plausible but nontrivial result to prove. Thus, if we can construct a

Lipschitz surjection 8 : K ab → Ca × [a, b], we then know that dim H (K ab ) ≥ 1 −
ln(2)/ ln(a) as well.
Our first step is to define a family of Lipschitz surjections φαβ : Cβ → Cα when-
β
ever α < β. We use the representation from (12) P to define φα . Given x ∈ Cβ , there
is a unique σ ∈ 6 for which x = (1 − β)β −1 n σn β n . Using this binary encoding,
we define
φαβ (x) = (1 − α)α −1

X
σn α n . (14)
n

Notice that x and φαβ (x) are on the same strand. Geometrically, we can think that
φαβ “slides” the point x down the strand from Cβ to Cα . We say “down”, since Cβ
is higher on CH than is Cα . We claim that φαβ is Lipschitz with Lipschitz constant
(1 − 2α)/(1 − 2β). To see this, notice that for x, y ∈ Cβ with x < y, we know that
|y − x| is equal to the sum of the gap lengths contained in the interval [x, y]. Since
φαβ (x) is on the same strand as x and φαβ (y) is on the same strand as y, for any stage-
n gap g ⊂ [x, y] of length |g| = (1 − 2β)β n−1 , there is a corresponding stage-n gap
g 0 ⊂ [φαβ (x), φαβ (y)] of length |g 0 | = (1 − 2α)α n−1 . Thus
|φαβ (y) − φαβ (y)| = {|g 0 | : g 0 ⊂ [φαβ (x), φαβ (y)] is a stage-n gap in Cα }
XX
X (1 − 2α)α n−1 X
= {|g| : g ⊂ [x, y] is a stage-n gap in Cβ }
n
(1 − 2β)β n−1
1 − 2α
≤ |y − x|.
1 − 2β
In words, φαβ acts by mapping each gap in Cβ to its corresponding gap in Cα ; the
mapping of the gaps completely defines the action of φαβ on Cβ by “squeezing” each
point of Cβ between its corresponding gaps. Because of the decay rates for the gap
lengths in Cβ and Cα , the most stretching is done in mapping the largest gap in Cβ to
the largest gap in Cα .
We define our desired Lipschitz surjection 8 : K ab → Ca × [a, b] by
8(x, y) = (φay (x), y).
Let (x1 , y1 ), (x2 , y2 ) ∈ K ab with y1 < y2 . We note that, by construction, φay2 = φay1 ◦
φ yy12 . This suggests that we decompose the action of 8 by first moving (x2 , y2 ) to
(φ yy12 (x2 ), y2 ), and then mapping to (φay1 (φ yy12 (x2 )), y2 ). For the point (x1 , y1 ), we need
only do the one step to (φay1 (x1 ), x1 ). By the uniform bound |Tσ0 (y)| ≤ 2 (from above),
we see that
|x1 − φ yy12 (x2 )| ≤ |x1 − x2 | + 2|y1 − y2 | ≤ 2(|x1 − x2 | + |y1 − y2 |).
We know that the mapping φay1 : C y1 → Ca is Lipschitz, so altogether 8 is also Lip-
schitz. This proves that dim H (K ab ) ≥ 1 − ln(2)/ ln(a). Taking an = (1/2)(1 − 1/n)
and bn = (1/2 + an )/2, we see that
ln(2)
2 ≥ dim H (CH) ≥ dim H (K abnn ) ≥ 1 + n−1
, for all n > 1,
ln(2) − ln n
and thus dim H (CH) = 2.

By a slight modification of the above procedure, it is possible to construct a Lip-
schitz surjection 9 : K ab → K cd whenever c ≤ a and d ≤ b. By this method, we can
prove that dim H (K ab ) = 1 − ln(2)/ ln(b), so the “local” Hausdorff dimension of CH
varies from 1 at y = 0 up to 2 at y = 1/2.
We can also compute the Hausdorff measure of any K ab . For b = 1/2, we know
that h 2 (CH) = 0 and so h 2 (K a1/2 ) = 0 for any a < 1/2. Continuing on to b < 1/2, let
0 ≤ a < b < 1/2 be fixed, s = dim H (K ab ) = − ln(2)/ ln(b), and yn = b − (b − a)/n
for n ≥ 1. Then
[
K ab = K yynn+1 ∪ (Cb × {b}) .
n≥1

y y
Since dim H (K ynn+1 ) < s for all n, we know that h s (K ynn+1 ) = 0, and so
X
h s (K ab ) ≤ h s (K yynn+1 ) + h s (Cb × {b}) = 0.
n≥1
Thus the Hausdorff measures of all the K ab are equal to zero, in their dimension.
5.2. The box-counting dimension of CH. The Hausdorff measures have very nice
properties but are somewhat difficult to work with. This makes computing the Haus-
dorff dimension difficult as well. For these, as well as other reasons, many different di-
mensions and corresponding measures of the “size” of a set have been defined. Among
the simplest of these is the box-counting dimension (also called the box dimension).
Given a bounded subset A ⊂ Rd , let Nδ (A) be the smallest number of sets of diam-
eter δ > 0 that will cover A. The box-counting dimension measures the asymptotic
growth rate of Nδ (A) as δ decreases to zero, specifically by fitting a model of the form
Nδ (A) ∼ Cδ −s , and so
ln(Nδ (A))
dim B (A) = lim . (15)
δ→0 − ln(δ)
Of course the limit doesn’t have to exist, so in general we have the upper box dimension
and lower box dimension given by
ln(Nδ (A)) ln(Nδ (A))

dim B (A) = lim sup and dim B (A) = lim inf .
δ→0 − ln(δ) δ→0 − ln(δ)
Clearly, for a bounded A ⊂ Rd we have Nδ (A) = O(δ −d ), and so dim B (A) ≤ d. Fur-
thermore, since Nδ (cl(A)) = Nδ (A), we have that the box dimensions of A and cl(A)
agree, where cl(A) is the closure of the set A.
Unfortunately, the box dimensions are only finitely stable with dim B (∪i=1 k
Ai ) =
maxi dim B (Ai ). The set A = {0} ∪ {1/n : n ≥ 1} ⊂ R provides a nice counter-
example to countable stability, since dim B (A) = 1/2, as is easy to show.
We also have a simple relation between the Hausdorff and box-counting dimen-
sions. It is easy to see that h sδ (A) ≤ Nδ (A)δ s . Thus, if h s (A) > 1, then for sufficiently
small δ > 0 we have 0 < ln(Nδ (A)) + s ln(δ), and so s ≤ dim B (A). In particular, for
any s ≤ dim H (A), we have h s (A) = +∞, and so s ≤ dim B (A). This means that
dim H (A) ≤ dim B (A) ≤ dim B (A)
for any compact A.

Another way to compute the box dimensions will be more useful for us. Given a set
A ⊂ Rd and > 0, we define the -dilation of A as
A = {x : d(x, a) < , for some a ∈ A}.
Since Ld (Aδ ) ≤ cNδ (A)δ d for some constant c > 0 depending only on d, it’s not sur-
prising that there is a relationship between the box dimensions and the decay rate of
Ld (Aδ ), as δ tends to zero. In fact, for A ⊂ Rd we have
ln Ld (Aδ )

dim B (A) = d − lim sup (16)
δ ln(δ)

and
ln Ld (Aδ )

dim B (A) = d − lim inf . (17)
δ ln(δ)
An especially interesting class of sets is the one composed of the so-called Min-
kowski measurable sets. These are sets A ⊂ Rd such that limδ→0 Ld (Aδ )/δ d−s exists,
where s = dim B (A). The limit is then called the Minkowski content of A. Minkowski
measurable sets are analogues of s-sets for box-counting dimension.
Computing the box dimension of K ab . The box dimension of CH is equal to two,

since 2 = dim H (CH) ≤ dim B (CH) ≤ dim B (CH) ≤ 2. Furthermore, because CHδ de-
creases to CH and L2 (CH) = 0, we know that limδ→0 L2 (CHδ )δ 2−2 = 0, so that CH
is also Minkowski measurable with content equal to zero.
Getting the box dimension of K ab is more difficult, and to do this we explicitly
estimate the exponential rate of decay in the “tube formula” (which gives the area of
(K ab ) ). Amazing as it sounds, this is simpler than estimating Nδ (K ab ). For the rest of
this section we fix [a, b] ⊂ [0, 1/2) and set V () = L2 ((K ab ) ). We know that
1 − ln(2)/ ln(b) = dim H K ab ≤ dim B K ab ,

and thus we just need to get an upper bound on the box dimension.
Our idea is to use an explicit expression for the length, L y (), of the -dilation of C y
for each y ∈ [a, b], and then integrate this over y to get an area. Now, this will give the
area of a “horizontal” dilation of K ab , and not a true -dilation of K ab . However, if V ()
is the area of the -dilation and Vh () is the area of our “horizontal” dilation, then
√
Vh () ≤ V () ≤ Vh 2 . (18)
So if Vh () → 0 as → 0, then the same is true for V (). For (18), it is important to
know that the derivative of any of the functions which parameterize the strands is uni-
formly bounded by 2. In particular, none of these curves are close to being horizontal.
This is also why it is reasonable to integrate the length of the dilation of C y and obtain
the area of the “horizontal” dilation of K ab . That is, for a random stacking of Cantor
sets (even with these same interval lengths), there is no reason to expect a geometric
relationship between an -dilation at height y and an -dilation at a height y 0 . Since
the hairs are analytic functions, there is a smooth change from one C y to the nearby
C y 0 , and thus the integration is a reasonable thing to do.
For a given y, C y is a central Cantor set with scaling factor y, so from equation (1.9)
in [9] we see that the length of the -dilation of C y is
X
L y () = 2#{gaps ≥ 2} + {all gaps < 2} . (19)
To understand (19), think about the gap endpoints. If x is an endpoint for a gap g and
|g| < 2, then the g ⊂ (C y ) .
As there are 2n gaps of length (1 − 2y)y n , then for (1 − 2y)y n ≥ 2 we must have
2
$ %
ln 1−2y
n ≤ N (y, ) := . (20)
ln(y)

Thus,
L y () = 2(2 N (y,)+1 − 1) + (1 − 2y)

X
(2y)k
k>N (y,)+1
N (y,)+1 N (y,)+2
= 2(2 − 1) + (2y) . (21)
We will integrate this expression in y over the range [a, b]. The main difficulty is
that y influences the value of an integer in both terms. So, we simply break the in-
tegral up into parts where this integer value is constant. For each k between N (a, )
and N (b, ), let yk ∈ [a, b] satisfy (1 − 2yk )ykk = 2, so that the intervals [yk , yk+1 ]
form a partition of [a, b]. Let γ = 1/(1 − 2b). Since (2)1/k solves x k = 2, we
know that
(2)1/k ≤ yk ≤ (2)1/k (1 − 2b)−1/k = (2)1/k γ 1/k .
Thus we get
Z b k=N
X (b,) Z yk+1 Z yk+1
L y () dy ≤ 2 (2 k+1
− 1) dy + (2y) k+2
dy
a k=N (a,) yk yk
(b,)
( )
k=N
X Z (2γ )1/(k+1) Z (2γ )1/(k+1)
≤ 2 (2k+1
− 1) dy + (2y) dy . (22)
k
k=N (a,) (2)1/k (2)1/k
First, estimate the second integral in (22) and its contribution to the sum. We see
that
(2γ )1/(k+1)
2k 2k
Z
(2y)k dy = 2 γ − (2)1/k ≤ 2γ .

(2)1/k k+1 k+1
Thus we have that the integral of the second term goes as
k=N (b,) N (b,)

2k 2x
X Z
2γ ∼ 2γ dx
k=N (a,)
k+1 N (a,) x + 1

N (b,)
1
∼ 2γ 2 + H.O.T.
ln(2)N (b, )
C
∼ (2)1+ln(2)/ ln(b) . (23)
|ln(2) − ln(1 − 2b)|
Estimating the first term in (22) is similar. Using the Mean Value Theorem for the
function (x, y) 7 → x 1/y , we have
Z (2γ )1/(k+1)
2k+1 − 1 dy = 2k+1 − 1 (2γ )1/(k+1) − (2)1/k

(2)1/k
γ − 1 ln(2)

≤2 k+1
(2γ ) 1/(k+1)
− .
k k2

Thus,
k=N
X (b,) Z (2γ )1/(k+1)
2k+1 − 1 dy

2
k=N (a,) (2)1/k
(b,) (b,) k
" k=N k=N
#
2k X X 2
≤ 4γ (γ − 1) − ln(2) . (24)
k=N (a,)
k k=N (a,)
k2
The first sum has the same estimate as (23). For the second sum, we see that
N (b,)
2x
Z
1
d x ∼ 2x .
N (a,) x2 ln(2)N (b, )2
Putting this information together, we obtain
k=N (b,) Z (2γ )1/(k+1)

X C
2 − 1 dy = O (2)1+ln(2)/ ln(b)
k
.

2
k=N (a,) (2)1/k |ln(2) − ln(1 − 2b)|
(25)
The final result of (23) and (25) is that

1
V ()/ 1+ln(2)/ ln(b)
=O .
ln(2)
So dim B (K ab ) = 2 − [1 + ln(2)/ ln(b)] = 1 − ln(2)/ ln(b). Furthermore, K ab is Min-

kowski measurable with content equal to zero, because of the ln(2) in the denomina-
tor of both (23) and (25). Note, however, that the decay to zero is very slow.
Comments. We notice that dim H (CH) = 2, but that L2 (CH) = 0 = h 2 (CH), and so
CH is a rather simple example of a set with maximal dimension and zero measure. It
is a more interesting and simpler example of this than the set

#{1 ≤ i ≤ n : ith binary digit of x is one}
S = x ∈ [0, 1] : lim does not exist ,
n n
which is a 1-dimensional set with zero Lebesgue measure. However, S is not compact
and cl(S) = [0, 1], so the Minkowski content doesn’t agree with the Lebesgue mea-
sure, unlike in the case of CH. The behaviour of CH and K ab is the same with respect
to both the Hausdorff and box-counting dimensions and with respect to the Hausdorff
measure and Minkowski content.
6. GENERALIZATIONS. It is pretty simple to change the construction of CH to

obtain variations, such as the sets illustrated in Figure 2.
The idea is to vary the “stack” of Cantor sets. This is accomplished by changing the
contraction factor for the maps in (4) from y to some function g(y) of y. That is, we
use the two maps
w0 (x, y) = g(y) x and w1 (x, y) = g(y) x + (1 − g(y)), (26)

Figure 2. Variations on CH
for some appropriate (and interesting) choice for g(y). Figure 2 illustrates p (clock-
wise from upper left) the sets associated
√ with the functions g(y) = 1/2 − 1/4 − y 2 ,
2
g(y) = sin(2π y)/2, and g(y) = y/2, g(y) = 2y . The only real restriction on g is
that 0 ≤ g(y) ≤ 1/2 for y ∈ [0, 1/2]. For polynomial g, it is easy to do explicit com-
putations. We see that, just as in the standard case of g(y) = y, all the visible strands
in the image are polynomial functions of y.
We will use CHg to denote the version of CH associated with the function g(y).
Area of “gaps.” For particular choices of g(y) it is also possible to obtain explicit
values for the pareas of the “gaps” in CHg . One non-polynomial case, where we set
g(y) = 1/2 − 1/4 − y 2 , is particularly interesting. In this case, the central gap is a
semi-circle. The calculations in this case are more complicated than in the standard
case, but it is still relatively simple to obtain explicit formulas for the areas of the
“gaps.” The formula for the area of an nth stage “gap” is
Z 1/2 p p n−1
2−n−1 1 − 4y 2 1 − 1 − 4y 2 dy.
0
This is a perfect integral for a trigonometric substitution of the form y = sin(θ )/2,
giving the integral
Z π/2
An = 2−n−2 cos2 (θ )(1 − cos(θ ))n−1 dθ.
0
Explicitly computing these and checking to see if they sum to 1/2 is rather tedious. A
nice trick, however, is to see that the total area is equal to
1 π/2
X Z X
2n An = cos2 (θ ) (1 − cos(θ ))n dθ
n
2 0 n≥0
Z π/2
1 1
= cos2 (θ ) dθ
2 0 1 − 1 + cos(θ )
Z π/2
1 1
= cos(θ ) dθ = .
2 0 2

We leave it up to the reader to justify the exchange of the infinite sum and the integral.
This calculation is part of a more general fact. For a general function g(y), we have
that the Cantor set at “level” y with scaling ratio g(y) has Lebesgue measure zero (as
long as g(y) < 1/2), since the sum of the gaps is
X 1 − 2g(y)
(1 − 2g(y))(g(y))n−1 = = 1.
n≥1
1 − 2g(y)
The area of one of the 2n−1 nth “stage” gaps is

Z 1/2
(1 − 2g(y))(g(y))n−1 dy,
0
and so the sum of all of them is

X Z 1/2 Z 1/2 X
(1 − 2g(y))(g(y)) dy =
n−1
(1 − 2g(y))(g(y))n−1 dy
n≥1 0 0 n≥1
Z 1/2
1
= 1 dy = .
0 2
Thread functions are real-analytic. If the function g(y) is real-analytic, then so are
all the thread functions. This is rather simple to show, since in this case the thread
function for σ ∈ 6 is given by
X
Tσg (y) = (1 − g(y))g(y)−1 σn g(y)n , (27)
n≥1
for all y such that g(y) 6 = 0, 1. In fact, we see that Tσg = Tσ ◦ g, and so Tσg is real-
analytic, as it is the composition of two real-analytic functions. From this viewpoint,
the standard version of CH has a “universality” property, since the thread functions for
any variation are constructed in a simple way from the thread functions of the standard
version.
Geometric measure-theoretic properties. Suppose that both g and g −1 are C 1 . Then

the following function 8 : CH → CHg is easily seen to be bi-Lipschitz:
8(x, y) = (x, g −1 (y)).
Thus dim H (CHg ) = 2 and L2 (CHg ) = h 2 (CHg ) = 0 under these conditions. Notice
that this condition is satisfied for three of the examples in Figure 2.
Furthermore, under this same condition the integral estimates for the “horizontal”
tube formula for CHg are a straightfoward change-of-variable from that for CH, so we
obtain the same decay rate, up to a constant multiplier.
Since we obtained the Hausdorff dimension, Hausdorff measure, box dimension,
and Minkowski content for K ab for any 0 ≤ a < b ≤ 1/2, we can use this information
to analyze these same local properties of any CHg .
ACKNOWLEDGMENTS. The second author was partially supported by a Discovery Grant from the Natural
Sciences and Engineering Research Council of Canada.

REFERENCES
1. M. F. Barnsley, Fractals Everywhere, Academic Press, New York, 1988.

2. C. Cabrelli, F. Mendivil, U. Molter, R. Shonkwiler, On the Hausdorff h-measure of Cantor sets, Pac. J.
Math. 217 (2004) 45–59, available at http://dx.doi.org/10.2140/pjm.2004.217.45.
3. K. J. Falconer, Fractal Geometry–Mathematical Foundations and Applications, second edition, Wiley,
New York, 2003.
4. , Techniques in Fractal Geometry, Wiley, New York, 1997.
5. J. A. Guthrie, J. E. Nymann, The topological structure of the set of subsums of an infinite series, Colloq.
Math. 55 (1988) 323–327.
6. K. Hannabus, Mathematics in Victorian Oxford: A Tale of Three Professors, in Mathematics in Victorian
Britain, Edited by R. Flood, A. Rice, R. Wilson, Oxford University Press, Oxford, 2011. 35–52.
7. J. E. Hutchinson, Fractals and self-similarity, Indiana Univ. J. Math. 30 (1981) 713–747, available at
http://dx.doi.org/10.1512/iumj.1981.30.30055.
8. S. Kakeya, On the partial sums of an infinite series, Tohoku Sci. Rep. 3 no. 4 (1914) 159–164.
9. M. L. Lapidus, M. Van Frankenhuijsen, Fractal Geometry, Complex Dimensions and Zeta Functions:
Geometry and Spectra of Fractal Strings, Springer, New York, 2006.
10. B. Mandelbrot, The Fractal Geometry of Nature, W. H. Freeman, New York, 1983.
11. P. K. Menon, On a class of perfect sets, Bull. Amer. Math. Soc. 54 (1948) 706–711, available at http:
//dx.doi.org/10.1090/S0002-9904-1948-09060-7.
12. C. A. Rogers, Hausdorff Measures, reprint of the 1970 original, Cambridge University Press, Cambridge,
1998.
13. S. G. Krantz, H. R. Parks, The Geometry of Domains in Space, Birkhäuser, Boston, 1999.
JACQUES LÉVY VÉHEL is a research director at Inria in France. He is interested in understanding irregu-
larity in natural phenomena, using tools in harmonic analysis, probability theory, and fractal analysis.
Regularity team, INRIA Saclay and MAS Laboratory, Ecole Centrale Paris, Grande Voie des Vignes,
92295 Chatenay-Malabry Cedex, France
jacques.levy-vehel@inria.fr
FRANKLIN MENDIVIL is a professor of mathematics at Acadia University in Nova Scotia. His research is
a blend of fractal geometry and analysis, image processing, and optimization. He considers himself extremely
lucky to be in a profession that allows him to explore many different topics.
Department of Mathematics and Statistics, Acadia University, 12 University Avenue,
Wolfville, NS Canada B4P 2R6
franklin.mendivil@acadiau.ca
Growth Rate of Calculus Textbooks

When students express surprise that Year Author Pages
there is anything new in mathe- 1716 L’Hôpital 202
matics, I point them to this table 1836 Davies 286
of calculus book page counts. The 1937 Courant 661
growth rate is 0.7% per year and 1993 Thomas/Finney 2nd 1094
increasing (not to mention that pages 2011 Stewart 7th 1368
are getting larger). Based on this,
we can safely predict a 2000 page
calculus book by the year 2090. —Submitted by Vadim Ponomarenko
MSC: Primary 26-01, Secondary 26A06

What to Expect in a Game of Memory
Daniel J. Velleman and Gregory S. Warrington
Abstract. The game of memory is played with a deck of n pairs of cards. The cards in each
pair are identical. The deck is shuffled and the cards laid face down. A move consists of
flipping over first one card and then another. The cards are removed from play if they match.
Otherwise, they are flipped back over and the next move commences. A game ends when all
pairs have been matched. We determine that, when the game is played optimally, as n → ∞:
• The expected number of moves is (3 − 2 ln 2)n + 7/8 − 2 ln 2 ≈ 1.61n.
• The expected number of times two matching cards are unwittingly flipped over is ln 2.
√
• The expected number of flips until two matching cards have been seen is 22n / 2nn ∼ πn.
1. INTRODUCTION. The game memory (also commonly known as concentration

or Hūsker Dū?) is played with a deck of 2n cards. The cards are numbered from 1
up to n, with each number appearing twice. The deck is shuffled and then the cards
are laid face down in a tableau. A move consists of flipping first one card and then a
second. If the cards match, both are removed from play and the current player moves
again. Otherwise, they are flipped back over and play passes to the next player. Play
ends when all pairs have been removed; the player who has removed the most pairs is
the winner. The length of a game is the total number of moves.
Perhaps surprisingly, strategy plays a role in the two-player version of the game [1].
In some situations, it may not be to a player’s advantage to acquire new information by
flipping over a card that has not been flipped before. The reason is that her opponent
also acquires the same information.
In this exploration we consider the solitaire version of memory, in which strategy
plays no part. Many games, such as poker and go fish, do not fare well as one-person
games. Memory does, as long as the goal shifts from collecting the largest number of
pairs to collecting all pairs in the fewest moves. As an additional simplification, we
assume that the player has perfect memory.
You may find it helpful to play a few games of solitaire memory on your own. You
can use some subset of an ordinary deck of cards; play online at a number of websites,
such as [2] or [3]; or, as of this writing, find a computer version on the wall at several
airports, including Houston, Minneapolis, JFK, and Dulles.
In the solitaire version of memory, acquiring information by flipping over an un-
known card is always helpful. It is therefore not hard to show that an optimal strategy
is to proceed as described in Algorithm 1. Throughout this paper, we assume that the
player is using this optimal strategy. We consider three questions regarding this one-
person game of memory.
1. What is the expected length of a game (i.e., how many moves are required)?
Each card must be flipped over at least once. If it is not matched the first time it
is flipped over, then it is flipped again, and removed, once the location of its mate is
known. Thus, each card is flipped either once or twice, and therefore the total number
of card flips is between 2n and 4n. Since two cards are flipped over on each move, this
MSC: Primary 60C05, Secondary 05A16
November 2013] WHAT TO EXPECT IN A GAME OF MEMORY 787

Algorithm 1 Optimal strategy
1: if the positions of both cards of a pair are known then

2: flip them over and remove them
3: else
4: flip over an unknown card
5: if the location of its mate is known then
6: flip the mate and remove both
7: else
8: flip over another unknown card
9: (and remove the pair if you are lucky enough to find the mate)
10: end if
11: end if
means that the number of moves in any game will be between n and 2n. Analysis of
the possibilities for the end of the game shows that the last card to be flipped over will
only be flipped once, and therefore the length of the game cannot be 2n. Therefore,
the length lies between n and 2n − 1. Example 2 (in Section 2) illustrates that both of
these lengths are possible. (It is not hard to show that if the player has no memory at
all, and plays by simply flipping over cards at random, then the expected length of a
game is n 2 .)
Computer investigations indicate that the expected length of a game is roughly 1.6n.
In Sections 4 and 5 we will determine the exact length, and show that it is approxi-
mately (3 − 2 ln 2)n + 7/8 − 2 ln 2, with the error in this approximation approaching 0
as n → ∞. Figure 1 shows the distribution of game lengths for n = 100. The expected
length in this case is about 160.8589, and our approximation gives about 160.8593.
1 0.20
10 −50 0.15
10 −100 0.10
0.05
10 −150
120 140 160 180 200 159 161 163 165

Figure 1. On the left is the distribution of game lengths for n = 100, with logarithmic y-scale. On the right
we use a linear y-scale and a restricted domain to highlight the fact that over 98% of games for n = 100 will
have a length lying between 157 and 165.
A lucky move is one in which the player flips two cards that have not been flipped
before and they happen to match. Note that a game of length n is one in which every
move is lucky. The second question, to be addressed in Section 5.2, is as follows.
2. What is the expected number of lucky moves in a game?

Early in the game, the player has little information about the locations of the various
cards. In all likelihood, each card flipped over will be the first of its pair to be seen. But
certainly by the (n + 1)th flip, the player will begin encountering the mates of cards

she has already seen. This brings us to our third and final question, to be addressed in
Section 3.
3. How many flips are required before the player should expect to have seen both
cards of a pair?
2. COMBINATORIAL SETUP. A game of memory is not affected by how the cards

are physically laid out. In light of this, it will be convenient to assume that the cards are
arranged in a single row. And when our strategy calls for the player to flip an unknown
card, there is no advantage to flipping one particular unknown card in preference to
another, since the cards are shuffled before they are laid out. We may therefore assume
that unknown cards are simply flipped from left to right. With this convention, game
play depends only on the order in which the cards are laid out.
Example 1. Suppose that n = 6 and that the cards are laid out in the order
1 2 1 6 2 3 3 5 4 4 5 6.
Play will proceed as follows:

1. The player flips over the first two cards, which are a 1 and a 2. Since they don’t
match, they are flipped back over.
2. The player flips over the next card, which is another 1. Remembering that the
first card was also a 1, the player flips the first card to get the match and removes
the two 1s.
3. The player flips the next two cards, which are a 6 and a 2. They don’t match, so
they are flipped back.
4. The player now knows the locations of both 2s. She therefore flips them and
removes them.
5. The player flips the next two cards, which are both 3s. They match, so she re-
moves them. Since these cards had not been flipped before, this is an example of
a lucky move.
6. The player flips the next two cards: a 5 and a 4. They don’t match, so they are
flipped back.
7. The player flips the next card, which is the second 4. Remembering that the
previous card was also a 4, she flips it and removes the two 4s.
8. The player flips the second 5, flips the first 5 again, and removes the two 5s.
9. The player flips the second 6, flips the first 6 again, and removes the two 6s.
Example 2. We can now give examples of the shortest and longest possible games. If
the cards are laid out in the order
1 1 2 2 3 3 · · · n n,
then the game will consist of n lucky moves; this is the shortest possible game. It is
only slightly harder to illustrate the longest possible game. We will let the reader verify
that if the cards are dealt in the order
1 2 3 1 4 2 · · · k k − 2 · · · n − 1 n − 3 n n − 2 n − 1 n,
then it will take 2n − 1 moves to complete the game.

Although we have presented all of our examples by specifying the order in which
the cards are laid out, in fact game play depends on only the relative positions of
the pairs. These relative positions can be visualized using the notion of an intercon-
nection network. An interconnection network [4, 6] consists of 2n points along with
n endpoint-disjoint arcs that connect pairs of points. (Equivalently, such a network
encodes an involution without fixed points.) Figure 2 illustrates two deals for n = 6
along with the associated interconnection network, in which points labeled with the
same number are connected by an arc. (The first of these deals is the one in Example 1
above.)
1 2 1 6 2 3 3 5 4 4 5 6
4 5 4 6 5 1 1 2 3 3 2 6
Figure 2. Two deals for n = 6 with the same interconnection network.
The two deals in Figure 2 differ only in the labels we have assigned to each pair
of cards, and therefore the first deal leads to essentially the same game as the second.
We wish to identify those games with the same underlying interconnection network.
In general, there is an n!-to-1 map from size-2n deals to size-2n interconnection net-
works. This map for n = 2 is illustrated in Figure 3. We prefer to remain in the realm
of card games, so we designate a representative for each class of deals. We call a deal
standard if, when the game is played, the pairs are removed in order from 1 to n. Put
another way, a deal is standard if the second occurrence of i occurs to the left of the
second occurrence of i + 1 for each i between 1 and n − 1. The deals in Examples 1
and 2 are all standard.
{1 1 2 2, 2 2 1 1}
{1 2 1 2, 2 1 2 1}
{2 1 1 2, 1 2 2 1}
Figure 3. Memory deals for n = 2 paired according to interconnection network.
We denote the set of standard permutations of the multiset {1, 1, 2, 2, . . . , n, n} by

Mn . At this point, we have reduced a game of memory to an element of some Mn . We
write these elements in one-line notation (reminiscent of an actual row of cards). For
example, define σ ∈ M2 by σ (1) = σ (4) = 2 and σ (2) = σ (3) = 1. The element σ
represents the game 2 1 1 2, in which the first and fourth cards are 2s and the second
and third are 1s.
For completeness, we include a proof of the following formula for the number of
inequivalent games of memory—that is, the cardinality of Mn .
Lemma 3. For n ≥ 1, the number of standard games of memory on a deck of 2n

cards is
(2n)!
(2n − 1)!! = 1 · 3 · 5 · · · (2n − 1) = .
2n n!

Proof. The first equality involving the double factorial function is true by definition.
The second equality can be checked by simple algebra. We prove by induction that the
number of games is given by the above product of odd integers.
For n = 1, the formula is trivially true. Now suppose that n ≥ 2. Any standard deal
of a deck of size 2n can be constructed by adding two copies of n to a standard deal
of a deck of size 2(n − 1). In order for the augmented deal to be standard, the second
occurrence of n must be at the end of the row. Thus, there are 2(n − 1) + 1 = 2n − 1
possible locations for the first occurrence of n. The lemma therefore follows by the
principle of mathematical induction.
3. EXPECTED POSITION OF THE FIRST MATCH. Fix n ≥ 1. For any σ ∈

Mn , let f (σ ) denote the smallest index j such that σ ( j) = σ (i) for some i < j. In
other words, f (σ ) is the position of the first card in σ that matches a previous card.
Since we are assuming that the deal is standard, this first matching card will be a 1. Let
a(n, j) count the number of permutations σ ∈ Mn for which f (σ ) = j. Notice that
f (σ ) is always between 2 and n + 1, so a(n, j) = 0 if j < 2 or j > n + 1. Then for
a randomly chosen σ ∈ Mn , the probability that f (σ ) = j is
a(n, j)
,
(2n)!/(2n n!)
and therefore the expected value of f (σ ) is
n+1 Pn+1
X a(n, j) j=2 j · a(n, j)
j· = . (1)
j=2
(2n!)/(2n n!) (2n)!/(2n n!)
It turns out that there is a simple formula for the numerator of the last fraction.
Lemma 4. For every positive integer n,
n+1
j · a(n, j) = 2n n!.
X
j=2
Proof. We proceed by induction on n. The equation is easily seen to be correct when

n = 1. Now suppose that n ≥ 2, and that the equation holds for n − 1.
A permutation σ ∈ Mn with f (σ ) = j can be constructed from a permutation in
Mn−1 in one of two ways. The first way is to start with a permutation σ 0 ∈ Mn−1 with
f (σ 0 ) = j, insert an n into σ 0 somewhere after position j, and then add a second n
at the end. The number of permutations that can be constructed in this way is (2n −
1 − j) · a(n − 1, j). The second way is to start with a permutation σ 0 ∈ Mn−1 with
f (σ 0 ) = j − 1, insert an n before position j − 1, and then add another n at the end.
This can be done in ( j − 1) · a(n − 1, j − 1) ways. Thus, we have the recurrence
a(n, j) = (2n − 1 − j) · a(n − 1, j) + ( j − 1) · a(n − 1, j − 1).

We can now compute:
n+1
X n+1
X
j · a(n, j) = j · [(2n −1− j) · a(n −1, j) + ( j −1) · a(n −1, j −1)]
j=2 j=2
n+1
X n+1
X
= j · (2n −1− j) · a(n −1, j) + j · ( j −1) · a(n −1, j −1)
j=2 j=2
n+1
X n
X
= j · (2n −1− j) · a(n −1, j) + ( j +1) · j · a(n −1, j)
j=2 j=1
n
X n
X
= 2n j · a(n −1, j) = 2n · j · a(n −1, j)
j=2 j=2
= 2n · 2n−1 (n −1)! (by the inductive hypothesis)

= 2n n!.
Theorem 5. For n ≥ 1, the expected position of the first match is
22n
2n
.
n
Proof. Using equation (1) and Lemma 4, we find that the expected position of the first
match is
Pn+1
j=2 j · a(n, j) 2n n! 22n 22n
= = = .
(2n)!/(2n n!) (2n)!/(2n n!) (2n)!/n!2 2n
n
√
Corollary 6. The expected position of the first match grows as π n.
√ n n
Proof. Straightforward applications of Stirling’s approximation n! ∼ 2πn (see,

e
for example, [5]) yield
√ n
22n 2n
( 2π n ne )2 √
2n
∼2 √ = πn.
2n 2n
n 2π(2n) e
4. EXPECTED LENGTH OF A GAME. Our main goal in this section is to prove

the following theorem.
Theorem 7. The expected length of a game with 2n cards is (3 − 2 ln 2)n + 7/8 −

2 ln 2 + n , where limn→∞ n = 0.
Consider any standard deal σ ∈ Mn . In the game based on σ , the player will turn
over unknown cards two at a time until she comes to the first matching card at position
f (σ ). What happens at that point depends on whether f (σ ) is even or odd. If f (σ )
is odd, then the card at position f (σ ), which is the second 1, will be flipped over at
the beginning of a move. Since we are assuming that the player has perfect memory,
she will then flip the first 1 and remove both 1s. The number of moves required to

flip all the cards up to position f (σ ) and then remove the matching 1s is therefore
( f (σ ) + 1)/2 = f (σ )/2 + 1/2. On the other hand, if f (σ ) is even, then the card at
position f (σ ) will be the second card flipped in a move. If the first card flipped in that
move was also a 1, then the move is lucky and the two 1s are removed right away. If
not, then it will take an additional move to flip the two 1s and remove them. Thus, if
f (σ ) is even, then the number of moves to flip the first f (σ ) cards and remove the two
1s is either f (σ )/2, if the first 1 is at position f (σ ) − 1, or f (σ )/2 + 1, if not.
Once the two 1s have been removed, the player continues by flipping the unknown
card at position f (σ ) + 1, and she again flips unknown cards two at a time until she
reaches the next matching card, the second 2. If the second 2 is at position f (σ ) + k,
then as above, the number of moves required to flip all of the cards from position
f (σ ) + 1 to f (σ ) + k and remove the matching 2s is k/2 + 1/2 if k is odd, k/2 if k
is even and the first 2 is at position f (σ ) + k − 1, and k/2 + 1 otherwise. Once the
2s have been removed, the player flips unknown cards two at a time until she finds
the second 3, and she continues in this way until all the pairs have been matched and
removed.
Thus, we see that the second occurrences of each of the numbers from 1 to n can be
thought of as breaking the deal σ into n blocks, with each block ending with a second
occurrence. The function f (σ ) that we studied in the last section is the length of the
first block, but we will have to consider the lengths of all of the blocks to determine the
number of moves in the game. Let the lengths of the blocks be b1 (σ ) = f (σ ), b2 (σ ),
. . . , bn (σ ), and notice that since the entire deck contains 2n cards, b1 (σ ) + b2 (σ ) +
· · · + bn (σ ) = 2n.
We will call a block lucky if the block has length at least 2 and the last two cards in
the block match. The analysis above shows that the number of moves required to flip all
of the cards in the ith block and then remove the two cards labeled i is bi (σ )/2 + ai (σ ),
where ai (σ ) is 1/2 if bi (σ ) is odd, 0 if bi (σ ) is even and the ith block is lucky, and 1
if bi (σ ) is even and the ith block is not lucky. Thus, if e(σ ) is the number of blocks in
σ of even length, and `(σ ) is the number of lucky even-length blocks, then the length
of the game is
n n
bi (σ ) 1X 1
X
+ ai (σ ) = bi (σ ) + (n − e(σ )) · + (e(σ ) − `(σ )) · 1
i=1
2 2 i=1 2
3n e(σ )
= + − `(σ ).
2 2
For example, if n = 6 and σ is the deal given in Example 1, then the lengths of the
blocks are 3, 2, 2, 3, 1, 1, and the third and fourth blocks are lucky. Therefore e(σ ) = 2
and `(σ ) = 1, and the length of the game is
3·6 2
+ − 1 = 9,
2 2
which is in agreement with our analysis of this game in Example 1.

To prove Theorem 7, we will need to find the expected values of e(σ ) and `(σ ):
the expected number of even blocks and the expected number of lucky even blocks.
To compute these, we will keep track of the number of blocks of each kind and each

length among all of the (2n)!/(2n n!) standard deals. For any positive integer j, let
b(n, j) be the total number of blocks of length j in all standard deals of a deck of 2n
cards, and for j ≥ 2, let `(n, j) be the total number of lucky blocks of length j. The
greatest possible length for a block is n + 1, so b(n, j) = `(n, j) = 0 for j ≥ n + 2.
The expected number of blocks of length j is then
b(n, j)
b(n, j) = ,
(2n)!/(2n n!)
and similarly, the expected number of lucky blocks of length j is
`(n, j)
`(n, j) = .
(2n)!/(2n n!)
Using this notation, we can write the expected value of e(σ ) as i=1
P∞
b(n, 2i), and the
expected value of `(σ ) as i=1 `(n, 2i). Notice that, although we have written these
P∞
expected values as infinite sums, in fact the sums are finite, since all but finitely many
terms in each sum are 0. Thus, the expected length of a game is
∞ ∞
3n 1 X X
+ b(n, 2i) − `(n, 2i). (2)
2 2 i=1 i=1
Several steps are required to pass from equation (2) to Theorem 7. The first is to
replace the b(n, j) and `(n, j) in (2) with (what turn out to be) their asymptotic limits.
We will show that, for large n,
2(2n + 1) 1
b(n, j) ≈ and for j ≥ 2, `(n, j) ≈ . (3)
j ( j + 1)( j + 2) j ( j − 1)
We therefore define
2(2n +1) ˚ 1
b̊(n, j) = b(n, j)− and for j ≥ 2, `(n, j) = `(n, j)− .
j ( j +1)( j +2) j ( j −1)
(4)
Equivalently,
2(2n +1) 1 ˚
b(n, j) = + b̊(n, j) and for j ≥ 2, `(n, j) = + `(n, j).
j ( j +1)( j +2) j ( j −1)
We wish to rewrite (2) using the asymptotic limits of the b(n, j) and `(n, j). To this
end, we first compute
∞ ∞
2(2n + 1)
X X
b(n, 2i) = + b̊(n, 2i)
i=1 i=1
2i(2i + 1)(2i + 2)
∞ ∞
X 2(2n + 1) X
= + b̊(n, 2i).
i=1
2i(2i + 1)(2i + 2) i=1

Although the sum on the left-hand side is finite, the two sums on the right-hand side
are infinite. However, it is not hard to see that these two infinite sums converge. Indeed,
we can evaluate the first one by using the partial fractions decomposition
2 1 2 1
= − + . (5)
j ( j + 1)( j + 2) j j +1 j +2
This gives us
∞ N
2(2n +1) 1 2 1
X X
= (2n +1) lim − +
i=1
2i(2i +1)(2i +2) N →∞
i=1
2i 2i +1 2i +2
1 2 1 1 2 1 1 2 1

= (2n +1) lim − + + − + +· · ·+ − +
N →∞ 2 3 4 4 5 6 2N 2N +1 2N +2
1 2 2 2 2 2 1

= (2n +1) lim − + − +· · ·+ − +
N →∞ 2 3 4 5 2N 2N +1 2N +2
3 1 1 1 1 1

= (2n +1) lim + −2 1− + − +· · ·+
N →∞ 2 2N +2 2 3 4 2N +1
3

= (2n +1) −2 ln 2 ,
2
since the alternating harmonic series converges to ln 2. Therefore,

∞ ∞
3
X X
b(n, 2i) = (2n + 1) − 2 ln 2 + b̊(n, 2i). (6)
i=1
2 i=1
Next, we compute
∞ ∞ ∞ ∞
1 1
X
˚ 2i) = ˚ 2i).
X X X
`(n, 2i) = + `(n, + `(n,
i=1 i=1
2i(2i − 1) i=1
2i(2i − 1) i=1
For the first part, we again use a partial fractions decomposition:
1 1 1
= − .
j ( j − 1) j −1 j
This gives us
∞ N
1 1 1
X X
= lim −
i=1
2i(2i − 1) N →∞ i=1 2i − 1 2i
1 1 1 1 1

= lim 1 − + − + · · · + − = ln 2,
N →∞ 2 3 4 2N − 1 2N
so
∞ ∞
˚ 2i).
X X
`(n, 2i) = ln 2 + `(n, (7)
i=1 i=1

Combining (2), (6), and (7), we see that the expected length of the game is
∞ ∞
3n 1 X X
+ b(n, 2i) − `(n, 2i)
2 2 i=1 i=1
∞ ∞
3 1X X
˚ 2i).
= (3 − 2 ln 2)n + − 2 ln 2 + b̊(n, 2i) − `(n,
4 2 i=1 i=1
In the next section, we will show that
∞
X 1
lim b̊(n, 2i) =
n→∞
i=1
4
and
∞
˚ 2i) = 0.
X
lim `(n,
n→∞
i=1
Substituting in these values will complete the proof of Theorem 7.
5. EVALUATION OF LIMITS. In order to determine the limits required to com-

plete the proof of Theorem 7, we need to establish some facts about the b̊(n, j) and
˚
`(n, j). To establish these facts, we begin by finding effective recurrences that can be
˚
used to compute b̊(n, j) and `(n, j).
5.1. Sum of b̊(n, 2i). First we determine recurrences for the b(n, j). If n = 1, then
the only standard deal is 1 1, which has a single block of length 2. So b(1, 2) = 1 and
b(1, j) = 0 for j 6 = 2. For n ≥ 2, as we saw before, any standard deal σ ∈ Mn can
be constructed from a standard deal σ 0 ∈ Mn−1 by inserting an n in one of the 2n − 1
possible positions in σ 0 , and then adding a second n at the end. Now, what happens to
the blocks of σ 0 when this is done? A block of length j in σ 0 either becomes a block
of length j + 1 (if the first n is inserted into that block) or length j (if the first n is
inserted somewhere else). Also, there is a new block at the end of length either 2 (if
both copies of n are at the end) or 1 (if not). So, for n ≥ 2:
(2n − 2)!
b(n, 1) = (2n − 2) · b(n − 1, 1) + (2n − 2) · ,
2n−1 (n − 1)!
(2n − 2)!
b(n, 2) = b(n − 1, 1) + (2n − 3) · b(n − 1, 2) + ,
2n−1 (n
− 1)!
and
b(n, j) = ( j − 1) · b(n − 1, j − 1) + (2n − 1 − j) · b(n − 1, j) (for j ≥ 3).
Dividing by (2n)!/(2n n!), we get corresponding recurrences for b: b(1, 2) = 1,

b(1, j) = 0 for j 6 = 2, and for n ≥ 2,

(2n − 2) · (b(n − 1, 1) + 1)
b(n, 1) = ,
2n − 1
b(n − 1, 1) + (2n − 3) · b(n − 1, 2) + 1
b(n, 2) = ,
2n − 1
and
( j − 1) · b(n − 1, j − 1) + (2n − 1 − j) · b(n − 1, j)
b(n, j) = (for j ≥ 3).
2n − 1
Finally, we use the definition of b̊(n, j) (equation (4)) to convert these recurrences into
recurrences for b̊.
Lemma 8. The numbers b̊(n, j) satisfy the following equations:

3
b̊(1, 2) = ,
4
6
b̊(1, j) = − ( for j 6 = 2),
j ( j + 1)( j + 2)
and for n ≥ 2,
(2n − 2) · b̊(n − 1, 1) − 1
b̊(n, 1) = ,
2n − 1
b̊(n − 1, 1) + (2n − 3) · b̊(n − 1, 2) + 1
b̊(n, 2) = ,
2n − 1
and
( j − 1) · b̊(n − 1, j − 1) + (2n − 1 − j) · b̊(n − 1, j)

b̊(n, j) = ( for j ≥ 3).
2n − 1
Proof. The derivation of these equations involves only routine algebra. We illustrate
by deriving the formula for b̊(n, 2):
2(2n + 1)
b̊(n, 2) = b(n, 2) −
2·3·4
b(n − 1, 1) + (2n − 3) · b(n − 1, 2) + 1 2n + 1
= −
2n − 1 12
1 2(2n − 1)

= + b̊(n − 1, 1)
2n − 1 1·2·3
2(2n − 1) 2n + 1

+ (2n − 3) · + b̊(n − 1, 2) + 1 −
2·3·4 12
1 b̊(n − 1, 1) 2n − 3 (2n − 3) · b̊(n − 1, 2) 1 2n + 1
= + + + + −
3 2n − 1 12 2n − 1 2n − 1 12
b̊(n − 1, 1) + (2n − 3) · b̊(n − 1, 2) + 1
= .
2n − 1

Using these recurrences, we can now compute the values of b̊(n, j). Table 1 shows
these values for n up to 8 and j up to 7. We begin our study of these numbers by
determining formulas for b̊(n, j) for small values of j.
Table 1. The values of b̊(n, j) for 1 ≤ n ≤ 8 and 1 ≤ j ≤ 7.
j
1 2 3 4 5 6 7
1 −1 3/4 −1/10 −1/20 −1/35 −1/56 −1/84
2 −1 1/4 1/2 −1/12 −1/21 −5/168 −5/252
3 −1 3/20 3/10 17/60 −1/15 −1/24 −1/36
n 4 −1 3/28 3/14 1/4 1/7 −3/56 −1/28
5 −1 1/12 1/6 53/252 11/63 31/504 −11/252
6 −1 3/44 3/22 71/396 17/99 85/792 7/396
7 −1 3/52 3/26 89/572 23/143 425/3432 3/52
8 −1 1/20 1/10 107/780 29/195 439/3432 53/660
Lemma 9.
1. For n ≥ 1,
b̊(n, 1) = −1.
2. For n ≥ 1,
3
b̊(n, 2) = .
4(2n − 1)
3. For n ≥ 2,
3
b̊(n, 3) = 2b̊(n, 2) = .
2(2n − 1)
Proof. All three statements are proven by induction using Lemma 8, with the proof of
each statement after the first also making use of the previous statement.
Lemma 10. For all n and all j ≥ 2,
b̊(n, j) ≤ 3 · 2 · 4 · 6 · · · (2n − 2) . (8)

4 3 · 5 · 7 · · · (2n − 1)
Proof. The proof is by induction on n. Let
3 2 · 4 · 6 · · · (2n − 2)
Cn = · .
4 3 · 5 · 7 · · · (2n − 1)
For n = 1, we interpret the products in the numerator and denominator of the last
fraction in the formula for
Cn as empty products, which are equal to 1. So the inequality
to be proven is b̊(1, j) ≤ 3/4, which follows easily from the formulas in Lemma 8.

Now suppose that n ≥ 2, and the lemma holds for n − 1. Then the inductive hy-
pothesis is that for all j ≥ 2, b̊(n − 1, j) ≤ Cn−1 . We now consider three cases.

Case 1: j = 2. Then by Lemma 9,
b̊(n, 2) =
3 3 2
≤ ·
4(2n − 1) 4 2n − 1
3 2 4 · 6 · · · (2n − 2)
≤ · · = Cn .
4 2n − 1 3 · 5 · · · (2n − 3)
Case 2: 3 ≤ j ≤ 2n − 1. Then

( j − 1) · b̊(n − 1, j − 1) + (2n − 1 − j) · b̊(n − 1, j)
b̊(n, j) =
2n − 1

j −1 2n − 1 − j
· b̊(n − 1, j − 1) + · b̊(n − 1, j)

≤
2n − 1 2n − 1
j −1 2n − 1 − j 2n − 2
≤ · Cn−1 + · Cn−1 = · Cn−1
2n − 1 2n − 1 2n − 1
2n − 2 3 2 · 4 · · · (2n − 4)
= · · = Cn .
2n − 1 4 3 · 5 · · · (2n − 3)
Case 3: j ≥ 2n. Since n ≥ 2, this implies that j ≥ n + 2. Therefore b(n, j) = 0, so
2(2n + 1) 2(2n + 1)
b̊(n, j) = b(n, j) − =− .
j ( j + 1)( j + 2) j ( j + 1)( j + 2)
Thus

b̊(n, j) = 2(2n + 1) 2(2n + 1) 1 1
≤ = ≤
j ( j + 1)( j + 2) 2n(2n + 1)(2n + 2) n(2n + 2) 2n − 1
3 2 3 2 4 · 6 · · · (2n − 2)
≤ · ≤ · · = Cn .
4 2n − 1 4 2n − 1 3 · 5 · · · (2n − 3)
To get a better idea of the size of the bound in the last lemma, note that
2
2 · 4 · · · (2n − 2) 2 2 4 4 2n − 2 2n − 2 1

= · · · ··· · · .
3 · 5 · · · (2n − 1) 1 3 3 5 2n − 3 2n − 1 2n − 1
Now, it is not hard to see that the product
2 2 4 4 2n − 2 2n − 2
· · · ··· ·
1 3 3 5 2n − 3 2n − 1
increases as n increases, and it is well known that as n → ∞, it converges to π/2 (see,
for example, [7]). Therefore, for j ≥ 2 we have
b̊(n, j) ≤ 3 · π
r
. (9)

4 2(2n − 1)
In particular, it follows that b̊(n, j) → 0 as n → ∞, which justifies our claim that (3)
gives the asymptotic limits for the b(n, j).

To prove Theorem 7, we need to find, for fixed n, the sum of the numbers b̊(n, j)
for j even. To start, we compute the sum of all of the b̊(n, j):
∞ ∞ ∞
X X X 2(2n + 1)
b̊(n, j) = b(n, j) − . (10)
j=1 j=1 j=1
j ( j + 1)( j + 2)
The first sum on the right-hand side of equation (10) is the expected number of blocks
(of all lengths). But for every standard deal, the number of blocks is n, so this expected
number is n. We can evaluate the second sum on the right-hand side of (10) by using
the partial fractions decomposition (5):
∞
X 2(2n + 1)
j=1
j ( j + 1)( j + 2)
N
1 2 1
X
= (2n + 1) lim − +
N →∞
j=1
j j +1 j +2
2 1 1 2 1 1 2 1

= (2n + 1) lim 1 − + + − + + · · · + − +
N →∞ 2 3 2 3 4 N N +1 N +2
1 1 1 1

= (2n + 1) lim − + =n+ .
N →∞ 2 N +1 N +2 2
Thus,
∞ ∞ ∞
2(2n + 1) 1 1
X X X
b̊(n, j) = b(n, j) − =n− n+ =− .
j=1 j=1 j=1
j ( j + 1)( j + 2) 2 2
According to Lemma 9, b̊(n, 1) = −1, so

∞
X 1
b̊(n, j) = . (11)
j=2
2
To separate out the contributions of the even- and odd-numbered terms in (11), we
will use the following fact, which is illustrated in the first graph in Figure 4.
0.12
0.025
0.10
0.020
0.08
0.015
0.06
0.010
0.04
0.005
0.02
0.000 j
0.00 j 5 10 15 20
5 10 15 20 − 0.005
− 0.02
˚
Figure 4. Plots of the values of b̊(n, j) (left) and `(n, j) (right) for j from 2 to 21 for n equal to 10 (circles),
15 (squares), and 20 (diamonds).

Lemma 11. For n ≥ 1, there exist numbers kn and m n such that 2 ≤ kn < m n ≤ n + 2
and
b̊(n, 2) ≤ b̊(n, 3) ≤ · · · ≤ b̊(n, kn ),

b̊(n, kn ) ≥ b̊(n, kn + 1) ≥ · · · ≥ b̊(n, m n ),
and
b̊(n, m n ) ≤ b̊(n, m n + 1) ≤ · · · .
Proof. We use induction on n. It is easy to verify that the lemma holds for n = 1 and
n = 2 (with k1 = 2, m 1 = 3, k2 = 3, and m 2 = 4). Now suppose that n ≥ 3, and the
lemma holds for n − 1. To prove that the lemma holds for n, we will show that
b̊(n, 2) ≤ b̊(n, 3) ≤ · · · ≤ b̊(n, kn−1 ), (12)

b̊(n, kn−1 + 1) ≥ b̊(n, kn−1 + 2) ≥ · · · ≥ b̊(n, m n−1 ), (13)
b̊(n, m n−1 + 1) ≤ b̊(n, m n−1 + 2) ≤ · · · . (14)
We can therefore set kn to be one of kn−1 or kn−1 + 1, and set m n to be one of m n−1 or
m n−1 + 1.
To prove (12), suppose that 2 ≤ j < kn−1 ; we must verify that b̊(n, j) ≤ b̊(n, j +
1). If j = 2, this follows immediately from Lemma 9. Now suppose that j ≥ 3, and
note that j < m n−1 ≤ n + 1 ≤ 2n − 2. Then by the inductive hypothesis, we have
b̊(n − 1, j − 1) ≤ b̊(n − 1, j) ≤ b̊(n − 1, j + 1),
and therefore
( j − 1) · b̊(n − 1, j − 1) + (2n − 1 − j) · b̊(n − 1, j)
b̊(n, j) =
2n − 1
(2n − 2) · b̊(n − 1, j)
≤ (15)
2n − 1
j · b̊(n − 1, j) + (2n − 2 − j) · b̊(n − 1, j + 1)
≤ = b̊(n, j + 1).
2n − 1
Similar reasoning can be used to prove (13). To prove (14), suppose that j ≥
m n−1 + 1. If j ≤ 2n − 2, then we can repeat the reasoning in (15) to show that
b̊(n, j) ≤ b̊(n, j + 1). If j ≥ 2n − 1, then j ≥ n + 2, so b(n, j) = b(n, j + 1) = 0,
and therefore
2(2n + 1) 2(2n + 1)
b̊(n, j) = − ≤− = b̊(n, j + 1).
j ( j + 1)( j + 2) ( j + 1)( j + 2)( j + 3)
Lemma 11 will be useful to us because of the following fact.
Lemma 12. Let a1 , a2 , . . . , ak be a monotonic sequence with the absolute value of

each term bounded by some constant δ. Write A for the sum of the odd-indexed terms,
i.e.,
Pk a1 + a3 + · · · + ak if k is odd, and a1 + a3 + · · · + ak−1 if k is even. Write B =
i=1 ai − A, so that B is the sum of the even-indexed terms. Then
|B − A| ≤ 2δ. (16)

Proof. Suppose that the sequence is monotonically increasing. If k is even, then
A = a1 + a3 + · · · + ak−1 ≤ a2 + a4 + · · · + ak = B
and
A − a1 = a3 + a5 + · · · + ak−1 ≥ a2 + a4 + · · · + ak−2 = B − ak .
This implies that 0 ≤ B − A ≤ ak − a1 , so |B − A| ≤ |ak | + |a1 | ≤ 2δ. If k is odd,

then we obtain the inequality a1 ≤ A − B ≤ ak , which leads to the stronger bound of
|B − A| ≤ δ. The argument is similar if the sequence is monotonically decreasing.
We are finally ready to evaluate the first limit needed to complete our proof of
Theorem 7.
Lemma 13.
∞
X 1
lim b̊(n, 2i) = .
n→∞
i=1
4
Proof. By Lemma 11, for N > 0 the sequence b̊(n, 2), b̊(n, 3), . . . , b̊(n, 2N + 1) con-
sists of at most three monotonic subsequences.
√ We can therefore apply Lemma 12 to
each subsequence with the bound δ = (3/4) π/(2(2n − 1)) from (9) to show that
N
X N
X N
X
b̊(n, 2i + 1) − 6δ ≤ b̊(n, 2i) ≤ b̊(n, 2i + 1) + 6δ.
i=1 i=1 i=1
PN
Adding i=1 b̊(n, 2i), we can rewrite this as
2N
X +1 N
X 2N
X +1
b̊(n, j) − 6δ ≤ 2 b̊(n, 2i) ≤ b̊(n, j) + 6δ.
j=2 i=1 j=2
Letting N → ∞ and applying equation (11), we conclude that

∞
1 X 1
− 6δ ≤ 2 b̊(n, 2i) ≤ + 6δ,
2 i=1
2
or in other words,
∞
1 9 π 1 9 π
r X r
− ≤ b̊(n, 2i) ≤ + .
4 4 2(2n − 1) i=1
4 4 2(2n − 1)
The lemma follows by letting n → ∞.
5.2. Sum of `(n,

˚ 2i). We now do similar calculations for the lucky blocks. Since the
steps are similar to those in the last subsection, we skip many of the details.
If n = 1, then the only block in the standard deal 1 1 is lucky, so `(1, 2) = 1 and
`(1, j) = 0 for j ≥ 3. For n ≥ 2, the numbers `(n, j) satisfy the recurrences
(2n − 2)!
`(n, 2) = (2n − 3) · `(n − 1, 2) +
2n−1 (n − 1)!

and
`(n, j) = ( j − 2) · `(n − 1, j − 1) + (2n − 1 − j) · `(n − 1, j) (for j ≥ 3).
Therefore, `(1, 2) = 1, `(1, j) = 0 for j ≥ 3, and for n ≥ 2,
(2n − 3) · `(n − 1, 2) + 1
`(n, 2) =
2n − 1
and
( j − 2) · `(n − 1, j − 1) + (2n − 1 − j) · `(n − 1, j)

`(n, j) = (for j ≥ 3).
2n − 1
˚
Finally, applying the definition of `(n, j) we get
˚ 2) = 1 ,
`(1,
2
˚ 1
`(1, j) = − ( j ≥ 3),
j ( j − 1)
and for n ≥ 2,
˚
˚ 2) = (2n − 3) · `(n − 1, 2)
`(n,
2n − 1
and
˚ − 1, j − 1) + (2n − 1 − j) · `(n
( j − 2) · `(n ˚ − 1, j)
˚
`(n, j) = (for j ≥ 3).
2n − 1
Lemma 14. For all n and all j ≥ 2,
˚
`(n, 1
.

j) ≤
2(2n − 1)
Also, for n ≥ 2,
˚ 2) = `(n,
˚ 3) = 1
`(n, .
2(2n − 1)
Proof. The proof is by induction on n.
Lemma 15. For all n,

∞ ∞
˚
X X
`(n, j) = 1 and `(n, j) = 0.
j=2 j=2
Note that the first equation in Lemma 15 indicates that the expected number of
adjacent matching pairs in a random deal is 1.

Proof. The first equation can be proven by induction on n. For the second, we begin
˚
with the definition of `(n, j):
∞ ∞ ∞ ∞
X
˚
X X 1 X 1
`(n, j) = `(n, j) − =1− .
j=2 j=2 j=2
j ( j − 1) j=2
j ( j − 1)
Now we use partial fractions to evaluate the last sum:

∞ N
1 1 1
X X
= lim −
j=2
j ( j − 1) N →∞ j=2 j − 1 j
1 1 1 1 1

= lim 1 − + − + · · · + −
N →∞ 2 2 3 N −1 N
1

= lim 1 − = 1.
N →∞ N
Thus ∞
˚
X
`(n, j) = 1 − 1 = 0.
j=2
The second graph in Figure 4 suggests our next lemma.
Lemma 16. For n ≥ 1, there exists a number pn such that 2 ≤ pn ≤ n + 2, as well as

˚ 2) ≥ `(n,
`(n, ˚ 3) ≥ · · · ≥ `(n,
˚ pn ) and ˚ pn ) ≤ `(n,
`(n, ˚ pn + 1) ≤ · · · .
Proof. The proof is by induction on n. It is easy to check that p1 = 3 and p2 = 4. To

establish the lemma for n ≥ 3, we show that
˚ 2) ≥ `(n,
`(n, ˚ 3) ≥ · · · ≥ `(n,
˚ pn−1 ) ˚ pn−1 + 1) ≤ `(n,
and `(n, ˚ pn−1 + 2) ≤ · · · .
The proofs are similar to those in the proof of Lemma 11.

Lemma 17.
∞
˚ 2i) = 0.
X
lim `(n,
n→∞
i=1
˚ 2), `(n,
Proof. By Lemma 16, for N > 0 the sequence `(n, ˚ 3), . . . , `(n,
˚ 2N + 1) con-
sists of at most two monotonic subsequences. As in the proof of Lemma 13, we apply
Lemma 12 with δ = 1/(2(2n − 1)) to show that
N N N
˚ 2i + 1) − 4δ ≤ ˚ 2i) ≤ ˚ 2i + 1) + 4δ.
X X X
`(n, `(n, `(n,
i=1 i=1 i=1
˚ 2i), letting N → ∞, and dividing by 2, we conclude that

PN
Adding i=1 `(n,
∞
1 X
˚ 2i) ≤ 1
− ≤ `(n, ,
2n − 1 i=1
2n −1
and the lemma follows.

This completes the proof of Theorem 7. Observe that the proofs of Lemmas 13
and 17 indicate that n in Theorem 7 is
1 1 1

O √ + =O √ .
n n n
In fact, numerical data indicate that −1/(25n) may be a reasonable approximation to

n . The resulting O n1 bound could possibly be proved by a more intricate analysis in
the proof of Lemma 13.
Lemma 17 also resolves Question 2.
Corollary 18. The expected number of lucky moves in a game tends to ln 2 as n → ∞.
Proof. There is a lucky move for P

each lucky block whose length is even, so the ex-
pected number of lucky moves is i=1∞
`(n, 2i). The lemma now follows from equa-
tion (7) and Lemma 17.
ACKNOWLEDGMENTS. The authors would like to thank Dan Archdeacon for useful conversations leading
to this paper, and the anonymous referees for several helpful suggestions.
REFERENCES
1. E. Alfthan, Optimal strategy in the childrens [sic] game memory (2007), available at http://www.math.
kth.se/xComb/x1.pdf.
2. Classic memory game (2012), available at http://classicmemorygame.com/.
3. Concentration memory game (2012), available at http://www.mathsisfun.com/games/memory/
index.html.
4. P. Flajolet, R. Sedgewick, Analytic Combinatorics, Cambridge University Press, Cambridge, 2009.
5. R. L. Graham, D. E. Knuth, O. Patashnik, Concrete Mathematics, second edition. Addison-Wesley, Read-
ing, MA, 1994.
6. J. C. Lagarias, A. M. Odlyzko, D. B. Zagier, On the capacity of disjointly shared networks, Com-
put. Networks ISDN Systems 10 (1985) 275–285, available at http://dx.doi.org/10.1016/0169-
7552(85)90070-4.
7. J. Wästlund, An elementary proof of the Wallis product formula for pi, Amer. Math. Monthly 114 (2007)
914–917.
DANIEL J. VELLEMAN received his B.A. from Dartmouth College in 1976 and his Ph.D. from the Univer-
sity of Wisconsin–Madison in 1980. He taught at the University of Texas before joining the faculty of Amherst
College in 1983. He was the editor of the American Mathematical Monthly from 2007 to 2011. In his spare
time he enjoys singing, bicycling, and playing volleyball.
Department of Mathematics, Amherst College, Amherst, MA 01002
djvelleman@amherst.edu
GREGORY S. WARRINGTON received his B.A. from Princeton University in 1995 and his Ph.D. from
Harvard University in 2001. He is now an algebraic combinatorialist at the University of Vermont. He and his
family enjoy playing games together; the inspiration for this paper came while he was playing the game of
memory with his daughter. This work was partially supported by a grant from the Simons Foundation (grant
number 197419) and National Science Foundation grant DMS-1201312.
Department of Mathematics and Statistics, University of Vermont, Burlington, VT 05401
gregory.warrington@uvm.edu

Mixing Problems with Many Tanks
Antonı́n Slavı́k
Abstract. We revisit the classical calculus problem of describing the flow of brine in a sys-
tem of tanks connected by pipes. For various configurations involving an arbitrary number of
tanks, we show that the corresponding linear system of differential equations can be solved
analytically. Finally, we analyze the asymptotic behavior of solutions for a general closed sys-
tem of tanks. It turns out that the problem is closely related to the study of Laplacian matrices
for directed graphs.
1. INTRODUCTION. Consider two tanks filled with brine, which are connected by
a pair of pipes. One pipe brings brine from the first tank to the second tank at a given
rate, while the second pipe carries brine in the opposite direction at the same rate. This
situation suggests the following problem. Assuming that the initial concentrations in
both tanks are known and that we have a perfect mixing in both tanks, find the concen-
trations in both tanks after a given period of time. Does the problem sound familiar?
Similar mixing problems appear in many differential equations textbooks (see, e.g.,
[3], [10], and especially [5], which has an impressive collection of mixing problems).
Most authors restrict themselves to mixing problems involving two or three tanks ar-
ranged in various configurations (a cascade with brine flowing in a single direction
only, a linear arrangement of tanks connected by pairs of pipes, a cyclic arrangement
of tanks, etc.). The problem then leads to a linear system of differential equations for
the unknown concentrations, which is solved by calculating the eigenvalues and eigen-
vectors of the corresponding matrix.
An exercise in [5, p. 380] asks the reader to consider a cascade of n tanks and to
solve the corresponding differential equations numerically in the case when n = 10.
In this article, we discuss a variety of mixing problems with n tanks (including the one
from [5]) and show that they can be solved exactly. Not only is it satisfying to obtain
analytic solutions of these problems, but we will also have the opportunity to review
a number of other interesting topics such as the Gershgorin circle theorem, recurrence
relations, the heat equation, circulant matrices, and Laplacian matrices.
2. STAR ARRANGEMENT OF TANKS. We start with the case where n tanks

(with n > 1) of the same volume V are arranged in the shape of a star. The central
tank T1 is connected by a pair of pipes to each of the remaining tanks T2 , . . . , Tn as
shown in Figure 1.
Assume that the flow through each pipe is f gallons per unit of time. Consequently,
the volume V in each tank remains constant. Let xi (t) be the amount of salt in tank Ti
at time t. This yields the differential equations
n
x1 (t) X xi (t)
x10 (t) = −(n − 1) f + f ,
V i=2
V
x1 (t) xi (t)
xi0 (t) = f − f , for 2 ≤ i ≤ n.
V V
MSC: Primary 34A30, Secondary 05C50, 34A05

T2 T8
T3 T7
T1
T4 T6
T5
Figure 1. Tanks in a star arrangement
Without loss of generality (by taking suitable time units), we may assume that f = V .
The system of differential equations can be written in the form x 0 (t) = Ax(t), where
x(t) = (x1 (t), . . . , xn (t))T and
−(n − 1) 1 · · · 1
 
1 −1 · · · 0 
.

A= .. .. .. .. 
 . . . . 
1 0 ··· −1
The problem of obtaining a general solution of this system is equivalent to finding

the eigenvalues and eigenvectors of A. The first task is to calculate the characteristic
polynomial of A, i.e., the determinant of the n × n matrix
−(n − 1) − λ 1 ··· 1
 
1 −1 − λ · · · 0
B = A − λI =  .
 
.. .. .. ..
 . . . . 
1 0 ··· −1 − λ
By definition,
X
det B = sgn(π )b1π(1) · · · bnπ(n) ,
π∈Sn
where the summation runs over all permutations π of {1, . . . , n}. Clearly, the only
permutations that contribute nonzero terms are the following.
• The identity permutation, which contributes (−(n − 1) − λ)(−1 − λ)n−1 .
• Permutations where π(1) = i for a certain i ∈ {2, . . . , n}, π(i) = 1, and π( j) = j
for j ∈ {2, . . . , n}\{i}. Each of these permutations is a transposition, and con-
tributes −(−1 − λ)n−2 to the final sum.
It follows that
det(A − λI ) = (−(n − 1) − λ)(−1 − λ)n−1 − (n − 1)(−1 − λ)n−2

= (−1 − λ)n−2 λ(n + λ),
November 2013] MIXING PROBLEMS WITH MANY TANKS 807

i.e., A has two simple eigenvalues λ = 0 and λ = −n, as well as the eigenvalue λ =
−1 of multiplicity n − 2.
The next step is to calculate the eigenvectors. For λ = 0, all eigenvectors v =
(v1 , . . . , vn ) have to satisfy v1 = v2 , v1 = v3 , . . . , v1 = vn ; the simplest choice is v =
(1, . . . , 1). This corresponds to the physical situation where every tank contains the
same amount of salt. For λ = −n, we obtain
1 1 ··· 1
 
1 −1 + n ··· 0
A − λI =  .
 
.
 .. .
.. .. ..
. . 
1 0 ··· −1 + n
The eigenvector components must satisfy vi = v1 /(1 − n), for 2 ≤ i ≤ n, and we can
take v = (1 − n, 1, . . . , 1). Finally, if λ = −1, then
−n + 2 1 · · · 1
 
 1 0 ··· 0 
A − λI =  .
.. .. . . ..  .
 . . . 
1 0 ··· 0
All eigenvectors must be such that v1 = 0 and v2 + · · · + vn = 0. For example, we

obtain a linearly independent set of n − 2 eigenvectors v 3 , . . . , v n ∈ Rn by taking v2k =
1, vkk = −1, and vlk = 0 for other values of l.
The general solution is a linear combination of functions of the form eλt v, where v
is an eigenvector corresponding to λ. Since eλt → 0 for λ = −n and λ = −1, every
solution approaches the state where all tanks contain the same amount of salt. In other
words, this state corresponds to a globally asymptotically stable equilibrium.
3. TANKS IN A ROW. In our next example, we have a row of n tanks T1 , . . . , Tn ,

with neighboring tanks connected by a pair of pipes (see Figure 2).
T1 T2 T3 T4 T5 T6 T7 T8
Figure 2. A linear arrangement of tanks
Again, the flow through each pipe is f gallons per unit of time. Using the same
notation as before, we obtain the following system of differential equations:
x1 (t) x2 (t)
x10 (t) = − f + f ,
V V
xi−1 (t) xi (t) xi+1 (t)
xi0 (t) = f −2f + f , for 2 ≤ i ≤ n − 1,
V V V
and
xn−1 (t) xn (t)
xn0 (t) = − f − f .
V V

Without loss of generality, we assume that f = V , and switch to the vector form
x 0 (t) = Ax(t), where
1 0 ··· 0 0 0
 
−1
 1 −2 1 ··· 0 0 0
 0 1 −2 · · · 0 0 0
 
 . .. .. . . .. .. .. 
A= .
 . . . . . . ..

 0 0 0 · · · −2 1 0
 
 0 0 0 ··· 1 −2 1
0 0 0 ··· 0 1 −1
The matrix A is an example of a tridiagonal and “almost-Toeplitz” matrix (a Toeplitz

matrix has constant values along lines parallel to the main diagonal). Before calculat-
ing its characteristic polynomial, it is useful to obtain some preliminary information
about the location of the eigenvalues of A. Since A is a symmetric matrix, all eigen-
values must be real. According to the Gershgorin circle theorem, the eigenvalues of A
are contained in the union of the intervals [aii − ri , aii + ri ], where
X
ri = |ai j |, for 1 ≤ i ≤ n.
j∈{1,...,n}\{i}
In our case, we have a11 = ann = −1 and r1 = rn = 1, while aii = −2 and ri = 2 for
i ∈ {2, . . . , n − 1}. Consequently, all eigenvalues of A are contained in the interval
[−4, 0].
Let us now return to the calculation of det(A − λI ). We start by expanding the
determinant with respect to the first column:
−1 − λ 1 0 ··· 0 0
 
 1 −2 − λ 1 ··· 0 0 
 0 1 −2 − λ · · · 0 0
 
det(A − λI ) = det 

.. .. .. .. .. .. 

 . . . . . .


 0 0 0 ··· −2 − λ 1 
0 0 0 ··· 1 −1 − λ
−2 − λ 1 ··· 0 0
 
 1 −2 − λ · · · 0 0 
= (−1 − λ) det 
 .. .. .. .. .. 
 . . . . .


 0 0 ··· −2 − λ 1 
0 0 ··· 1 −1 − λ
1 0 ··· 0 0
 
 1 −2 − λ · · · 0 0 
. .. .. .. ..
− det  . .

. . . . . 
0 0 · · · −2 − λ 1 
0 0 ··· 1 −1 − λ
The second determinant remains unchanged if we omit the first row, together with the
first column. Consequently,
det(A − λI ) = (−1 − λ)Dn−1 − Dn−2 ,

where Dk (for every k ∈ N) stands for the determinant of the k × k matrix
−2 − λ 1 0 ··· 0 0
 
 1 −2 − λ 1 ··· 0 0 
 0 1 −2 − λ ··· 0 0
 
.

 .. .. .. .. .. ..

 . . . . . .


 0 0 0 ··· −2 − λ 1 
0 0 0 ··· 1 −1 − λ
The value of Dk can be calculated by the same method that works for tridiagonal
Toeplitz matrices (see [8, p. 133]). Expanding this determinant with respect to the first
column, we obtain the recurrence relation
Dk = (−2 − λ)Dk−1 − Dk−2 , for k ≥ 2.
The initial values are D1 = −1 − λ and D0 = 1. We apply the standard procedure for
solving second-order linear homogeneous recurrence relations (see [6]). The auxiliary
equation
x 2 + (2 + λ)x + 1 = 0
has discriminant (2 + λ)2 − 4 = λ(λ + 4), which is nonpositive for λ ∈ [−4, 0]. (We
restrict our attention to λ ∈ [−4, 0] because we are trying to find the roots of det(A −
λ), and we already know they are contained in [−4, 0].)
√
−λ − 2 −λ(λ + 4)
x1,2 = ±i .
2 2
Performing the substitution λ + 2 = u, we obtain
√
−u 4 − u2
x1,2 = ±i = e±iγ ,
2 2
√
4−u 2
where γ ∈ [0, π] satisfies cos γ = −u/2, sin γ = 2 . Thus Dk is a linear combi-
nation of x1k = eikγ and x2k = e−ikγ , and Dk = α cos(kγ ) + β sin(kγ ) for some con-
stants α, β and every k ≥ 0. Using the initial conditions D0 = 1 and D1 = −1 − λ =
1 − u = 1 + 2 cos γ , we find α = 1 and β = (1 + cos γ )/ sin γ = cot(γ /2), provided
that γ 6= 0 (this assumption will be justified shortly). Therefore, Dk = cos(kγ ) +
cot(γ /2) sin(kγ ) and
det(A−λI ) = (−1−λ)Dn−1 − Dn−2 = Dn + Dn−1

γ
= cos (nγ )+cos ((n −1) γ )+cot (sin (nγ )+sin ((n −1) γ ))
2
γ γ γ γ γ
= 2 cos cos (2n −1) +2 cot cos sin (2n − 1)
2 2 2 2 2
γ γ γ γ γ
× 2 cot sin cos (2n −1) +cos sin (2n −1)
2 2 2 2 2
γ
= 2 cot sin (nγ ) .
2

For γ ∈ (0, π], the last expression vanishes when γ = kπ/n, for k ∈ {1, . . . , n}. Re-
turning back from γ to λ, we conclude that
kπ
λk = −2 cos − 2, for k ∈ {1, . . . , n},
n
are the eigenvalues of A; since we found n distinct eigenvalues, there is no need to
investigate the case γ = 0, which we excluded earlier.
The eigenvectors corresponding to λk are solutions of the linear homogeneous sys-
tem with the matrix
1 + 2 cos kπ 1 0 0 0
 
n
···
 1 2 cos kπ
n
1 ··· 0 0 
kπ
 
 0 1 2 cos n · · · 0 0 
.. .. .. . . . .
 

 . . . . . .
. .
. 
kπ
0 0 0 · · · 2 cos n 1
 
 
0 0 0 ··· 1 1 + 2 cos kπ
n
It follows that the components of an eigenvector v = (v1 , . . . , vn ) corresponding to the

eigenvalue λk are given by the relations
kπ

v2 = − 1 + 2 cos v1 , (1)
n
and
kπ

vi = −2 cos vi−1 − vi−2 , for i ∈ {3, . . . , n}, (2)
n
while v1 can be an arbitrary nonzero number. In particular, λn = 0, and the corre-

sponding eigenvectors have all components identical. Since the remaining eigenvalues
are negative, we conclude again that a general solution always approaches the state
with all tanks containing the same amount of salt.
Note that (2) is a second-order linear homogeneous recurrence relation with con-
stant coefficients; by considering (1) as the initial condition, we can apply the standard
algorithm (details are left to the reader) to conclude that
kπi kπ kπi

i i
vi = (−1) cos − (−1) cot sin , for i ∈ {1, . . . , n},
n 2n n
satisfies both (1) and (2); hence, v = (v1 , . . . , vn ) is an eigenvector corresponding

to λk .
The mixing problem we have just solved is closely related to the one-dimensional
heat equation, i.e., the partial differential equation of the form
∂f ∂2 f
(t, x) = k 2 (t, x),
∂t ∂x
which models the conduction of heat in a one-dimensional rod (see [3], [10]); in par-
ticular, f (t, x) is the temperature at the point x on the rod at time t. The same equation
also describes the one-dimensional diffusion in a homogeneous medium; in this case,
f (t, x) corresponds to the concentration of a chemical at the point x at time t. Consider

a line segment, represented by the interval [a, b], whose endpoints are insulated—
no heat transfer or diffusion takes place across the endpoints. Mathematically, this
means that
∂f ∂f
(t, a) = (t, b) = 0
∂x ∂x
for every t. In view of these boundary conditions, we extend the spatial domain of f by
letting f (t, x) = f (t, a) for x < a and f (t, x) = f (t, b) for x > b. We now discretize
the spatial domain of the heat equation by considering the grid of equally spaced
points xi = a + i1x, where 1x = (b − a)/n and n is an arbitrary positive integer.
Let yi (t) = f (t, xi ) for every i ∈ {−1, . . . , n + 1}. We approximate the second-order
2
partial derivatives ∂∂ x 2f at the grid points xi by the second-order central differences:
∂2 f . 1 f (t, x1 ) − f (t, x0 ) f (t, x0 ) − f (t, x−1 )

(t, x0 ) = −
∂x2 1x 1x 1x
− f (t, x0 ) + f (t, x1 )
= ,
(1x)2
∂2 f . 1 f (t, xi+1 ) − f (t, xi ) f (t, xi ) − f (t, xi−1 )

(t, x i ) = −
∂x2 1x 1x 1x
f (t, xi−1 ) − 2 f (t, xi ) + f (t, xi+1 )
= , for i ∈ {1, . . . , n − 1},
(1x)2
∂2 f . 1 f (t, xn+1 ) − f (t, xn ) f (t, xn ) − f (t, xn−1 )

(t, x n ) = −
∂x2 1x 1x 1x
f (t, xn−1 ) − f (t, xn )
= .
(1x)2
.
(The symbol = means “is approximately equal to”.) Since
∂f ∂2 f
yi0 (t) = (t, xi ) = k 2 (t, xi ),
∂t ∂x
it follows that
. k
y00 (t) = (−y0 (t) + y1 (t)),
(1x)2
. k
yi0 (t) = (yi−1 (t) − 2yi (t) + yi+1 (t)), for i ∈ {1, . . . , n − 1},
(1x)2
and
. k
yn0 (t) = (yn−1 (t) − yn (t)).
(1x)2
The right-hand side of this system is the same as in our mixing problem for n + 1 tanks
in a row. Thus, we have discovered that this particular mixing problem is equivalent
to the spatial discretization of the one-dimensional heat equation. Conversely, we can
start with the mixing problem as a toy model for heat conduction or diffusion, and
derive the one-dimensional heat equation by letting n → ∞ and 1x → 0 (see also [6,
p. 167] for a related model corresponding to both spatial and temporal discretization
of the heat equation).

It should be mentioned that our matrix A appears also in situations where we dis-
cretize a second-order ordinary differential equation subject to suitable boundary con-
ditions; many examples can be found in [12].
4. TANKS IN A CIRCLE. Another quite natural mixing problem is obtained by

placing n tanks along a circle and connecting neighboring tanks by pairs of pipes
(see Figure 3); note that this type of mixing problem can be interpreted as the spatial
discretization of the one-dimensional heat equation on a circle.
T2 T3
T1 T4
T8 T5
T7 T6
Figure 3. A cyclic arrangement of tanks
This configuration is quite similar to the linear arrangement and we leave it to the
reader to verify that the matrix of the resulting linear system of differential equations is
1 0 ··· 0 0 1
 
−2
 1 −2 1 ··· 0 0 0
0 1 −2 · · · 0 0 0
 

 . .. .. . . .. .. .. 
A=  .. . . . . . ..

 0 0 0 · · · −2 1 0
 
 0 0 0 ··· 1 −2 1
1 0 0 ··· 0 1 −2
This is an example of a circulant matrix, i.e., a matrix whose rows are obtained by
cyclically shifting the first row. Since the problem of calculating the eigenvalues and
eigenvectors of a circulant matrix is well known, we state only the results and refer
the reader to appropriate sources (for the general theory of circulant matrices, see [9]
and the references given there; for a direct calculation of the eigenvalues of A without
circulant matrix theory, see Lemma 4.9 in [2]).
1. All circulant matrices of order n share the same set of eigenvectors, namely
v j = (1, ω j , ω2j , . . . , ωn−1

j ), j ∈ {0, . . . , n − 1},
where ω j = exp(2πi j/n) = cos(2π j/n) + i sin(2π j/n).

2. A circulant matrix whose first row is (c0 , c1 , . . . , cn−1 ) has eigenvalues
λ j = c0 + c1 ω j + c2 ω2j + · · · + cn−1 ωn−1

j , j ∈ {0, . . . , n − 1}.

In our case, we have c0 = −2, c1 = cn−1 = 1, and ci = 0 otherwise. Consequently,
λ j = −2 + exp(2πi j/n) + exp(2πi(n − 1) j/n)

= −2 + exp(2πi j/n) + exp(−2πi j/n) = −2 + 2 cos(2π j/n).
In particular, λ0 = 0 and the corresponding eigenvector is v 0 = (1, . . . , 1).

Note that
λ j = −2 + 2 cos(2π j/n) = −2 + 2 cos(2π(n − j)/n) = λn− j
for j ∈ {1, . . . , n − 1}. Thus, for odd values of n, the spectrum of A consists of the
simple eigenvalue λ0 and the eigenvalues λ1 , . . . , λ(n−1)/2 having multiplicity 2. For
even values of n, we have the simple eigenvalues λ0 , λn/2 (with (−1, 1, . . . , −1, 1) be-
ing the eigenvector corresponding to λn/2 ) and the eigenvalues λ1 , . . . , λ(n/2)−1 having
multiplicity 2. In any case, all eigenvalues except λ0 are negative, and thus all solu-
tions of our system approach the equilibrium state with all tanks containing the same
amount of salt.
Again, we remark that the matrix A occurs quite frequently in situations where we
discretize second-order ordinary differential equations subject to suitable boundary
conditions; see [12] for a detailed discussion.
5. ANALYSIS OF A GENERAL MIXING PROBLEM. Suppose we are not able

to find the general solution of a certain mixing problem, i.e., we are unable to calcu-
late the eigenvalues and eigenvectors of the corresponding matrix. Is it still possible
to obtain some information about the general qualitative behavior of the solutions?
According to physical intuition, it seems reasonable to expect that every solution ap-
proaches the state with all tanks containing the same amount of salt. Can we prove this
fact rigorously?
We provide an affirmative answer for mixing problems satisfying the following
conditions.
(1) All tanks hold the same constant volume V of brine; consequently, the total
amount of brine flowing into a particular tank equals the total amount of brine
flowing out of the tank.
(2) Each pipe connecting a pair of tanks transports the same volume f of brine per
unit of time. By a suitable choice of time units, we can assume that f /V = 1.
(3) The mixing problem is irreducible in the following sense: The set of tanks
{T1 , . . . , Tn } cannot be partitioned into two disjoint nonempty groups such that
every pipe connects tanks from the same group.
Such a mixing problem again leads to a linear system of differential equations. What
does the corresponding matrix A look like? Note that A need not be symmetric, i.e.,
the fact that there is a pipe transporting brine from tank Ti to tank T j need not imply
the existence of a pipe transporting brine in the opposite direction. For example, the
simple cyclic arrangement displayed in Figure 4 satisfies our assumptions.
By the first condition, the sum of each row of A is zero. By the second condition,
ai j = 1 if there is a pipe transporting brine from the ith tank to the jth tank, aii equals
minus the number of pipes originating in the ith tank, and all remaining entries of A
are zero.
Matrices of a similar form are well known in algebraic graph theory. Given a di-
rected graph G = (V, E) with V = {v1 , . . . , vn }, the matrix L = {li j }i,n j=1 given by

T2 T3
T1 T4
T8 T5
T7 T6
Figure 4. A simple cycle with neighbors connected by a single pipe only

deg (vi ) if i = j,
 +
li j = −1 if i 6 = j and (vi , v j ) ∈ E,
0

otherwise
(where deg+ (v) stands for the outdegree of the vertex v) is known as the Laplacian
matrix of the directed graph G.
What is the relation between mixing problems and Laplacian matrices? A mixing
problem can be represented by a directed graph G with vertices T1 , . . . , Tn correspond-
ing to tanks and edges representing the pipes. It follows from our previous discussion
that such a mixing problem leads to the linear system x 0 (t) = Ax(t), where L = −A
is the Laplacian matrix of G.
Clearly, λ is an eigenvalue of A if and only if −λ is an eigenvalue of L (in this case,
both eigenvalues share the same eigenvectors). For an arbitrary directed graph, the sum
of each row in L is zero, and we see that (1, . . . , 1) is an eigenvector corresponding to
the zero eigenvalue. To prove that all solutions of our general mixing problem tend to
the equilibrium state with all tanks containing the same amount of salt, it is sufficient
to show that the zero eigenvalue is a simple one, and that all remaining eigenvalues
of L have positive real parts. The second assertion is an immediate consequence of
the Gershgorin circle theorem, which says that P all eigenvaluesP are contained in the
union of the n disks with centers lii and radii j6=i |li j | = − j6=i li j = lii , where
i ∈ {1, . . . , n}. To prove that the zero eigenvalue is a simple one, note that the graphs
arising from our mixing problems have two important properties: They are connected
and balanced, i.e., the indegree deg− (v) of an arbitrary vertex v equals its outdegree
deg+ (v). However, a balanced connected graph is also strongly connected, i.e., there
is a directed path from each vertex in the graph to every other vertex. (Assume that the
graph has more than one strongly connected component. The number of edges leaving
a certain component equals the number of edges entering that component, and we can
find an oriented cycle passing through different components—a contradiction.)
We now proceed to show that every balanced connected directed graph has a simple
zero eigenvalue. For undirected graphs, this result is well known (see [2], [7]); the
usual proof is based on the observation that L = M M T , where L is the Laplacian
matrix and M is the incidence matrix. Such a decomposition is possible only when
L is symmetric, and thus we have to use a different approach. Unfortunately, the lit-

erature on Laplacian matrices of directed graphs is rather scarce. The authors of [11]
consider a connected balanced directed graph and its Laplacian matrix L and prove
that the null space of L has dimension 1 (i.e., the geometric multiplicity of the zero
eigenvalue is 1). Then they conclude that zero must be a simple eigenvalue (i.e., the
algebraic multiplicity of the zero eigenvalue is 1). This result is correct, but a further
explanation is necessary to show that the geometric multiplicity of the zero eigenvalue
is equal to its algebraic multiplicity. Another paper dealing with Laplacian matrices of
directed graphs is [1]. The authors consider weighted directed graphs (not necessarily
balanced), and the results presented in their paper are far too general for our purposes.
Here we present a short proof for balanced unweighted graphs based on the following
lemma.
Lemma 1. If L is the Laplacian matrix of a balanced directed graph G = (V, E) on

n vertices and x ∈ Rn is an arbitrary vector, then
1 X
xT Lx = (xi − x j )2 .
2 (v ,v )∈E
i j
Proof. We have
n n n
T
lii xi2 l j j x 2j
X X X X
2x L x = 2 li j x i x j = +2 li j x i x j +
i, j=1 i=1 i, j∈{1,...,n}, j=1
i6= j
n n
deg+ (vi )xi2 − 2 deg+ (v j )x 2j .
X X X
= xi x j +
i=1 (vi ,v j )∈E j=1
Observing that
n
deg+ (vi )xi2 = xi2 ,
X X
i=1 (vi ,v j )∈E
n n
deg+ (v j )x 2j = deg− (v j )x 2j
X X
j=1 j=1
x 2j ,
X
=
(vi ,v j )∈E
we obtain
2x T L x = (xi2 − 2xi x j + x 2j ) = (xi − x j )2 ,

X X
(vi ,v j )∈E (vi ,v j )∈E
which proves the statement.
Theorem 2. The null space of the Laplacian matrix of a connected balanced directed
graph G = (V, E) has dimension 1.

Proof. Let L be the Laplacian matrix and consider an arbitrary vector x ∈ Rn such
that L x = 0. By Lemma 1,
1 X
0 = xT Lx = (xi − x j )2 .
2 (v ,v )∈E
i j
Consequently, all terms in the last sum must vanish and xi = x j whenever (vi , v j ) ∈ E.
It follows that xi = x j if there is a directed path between vi and v j . Since our graph is
strongly connected, we conclude that x1 = · · · = xn , i.e., the null space of L is spanned
by the single vector (1, . . . , 1).
Theorem 2 was already proved in [11], but our proof is different and shorter. We
now proceed to show that the algebraic multiplicity of the zero eigenvalue equals its
geometric multiplicity.
Theorem 3. For the Laplacian matrix of a connected balanced directed graph, the
zero eigenvalue is a simple one.
Proof. Let L be the Laplacian matrix. We have
det(L − λI ) = (−1)n λn + an−1 λn−1 + · · · + a1 λ + a0

= (−1)n (λ − λ1 ) · · · (λ − λn ),
where a0 , . . . , an−1 are certain constants and λ1 , . . . , λn are the eigenvalues of L, in-
cluding multiplicity. Without loss of generality, assume that λ1 is the zero eigenvalue.
Then
 
n
a1 = (−1)n
X  Y
(−λ ) = −λ2 · · · λn ,

i

 
k=1 i∈{1,...,n},
i6 =k
and the proof will be complete once we show that a1 6 = 0.

By Theorem 2, the dimension of the null space of L is 1, and it is thus possible to
obtain an invertible matrix L 0 by deleting the ith row and jth column of L for suitable
values of i and j. The value of the determinant
l11 − λ · · · l1n
 
 .. .. ..
det(L − λI ) = det  . . .


ln1 ··· lnn − λ
remains unchanged if we first add all rows except the ith one to the ith row, and then
add all columns except the jth one to the jth column. Using the fact that both the rows
and columns of L have zero sums, we see that

l11 − λ · · · −λ ··· l1n
 
 .. .. ..
 . . .


det(L − λI ) = det  −λ · · · −nλ ··· −λ 
 
 . .. .. 
 .. . . 
ln1 · · · −λ · · · lnn − λ
l11 − λ · · · −λ ··· l1n
 
 .. .. .. 
 . . . 
= λ det  −1 · · · −n ··· −1  .
 
 . .. .. 
 .. . . 
ln1 · · · −λ · · · lnn − λ
Consequently, the value a1 , which is the coefficient of the linear term in the charac-
teristic polynomial of L, equals the coefficient of the absolute term in the last determi-
nant. This absolute term can be obtained by substituting λ = 0, which gives
l11 · · · 0 · · · l1n
 
 .. .. .. 
 . . . 
a1 = det  −1 · · · −n · · · −1  = −n(−1)i+ j det L 0 6 = 0
 
 . .. .. 
 .. . . 
ln1 · · · 0 · · · lnn
(we have expanded the determinant with respect to the jth column).
Corollary 4. Consider a general mixing problem with n tanks such that conditions
(1)–(3) are satisfied. Then
x1 (0) + · · · + xn (0)
lim xi (t) = , for i ∈ {1, . . . , n}.
t→∞ n
Proof. The mixing problem leads to a linear system x 0 (t) = Ax(t), where A has a
simple zero eigenvalue and all remaining eigenvalues have negative real parts. The
eigenvectors corresponding to the zero eigenvalue have all components identical. Con-
sequently, for t → ∞, any solution of the system tends toward the state where all tanks
contain the same amount of salt. Since the total amount of salt is preserved, all com-
ponents of the solution must approach the arithmetic mean of the initial conditions.
Note that for disconnected balanced graphs, we have the following result.
Theorem 5. For the Laplacian matrix of a balanced directed graph, the algebraic
multiplicity of the zero eigenvalue equals the number of connected components.
Proof. After a suitable permutation of the vertices, the Laplacian matrix of the given
graph is the block diagonal matrix
L1 0 · · · 0
 
 0 L2 · · · 0 
L=  ... .. . . . ,
. . .. 
0 0 ··· Lk

where k is the number of connected components and L i is the Laplacian matrix of the
ith component. We have
k
Y
det(L − λI ) = det(L i − λI ),
i=1
and the conclusion follows from the fact that each L i has a simple zero eigenvalue.
Returning to mixing problems, we see that the assumption of irreducibility cannot

be omitted; for example, if we take two disjoint circular tank arrangements, it is true
that each cycle approaches the equilibrium state with all tanks having the same amount
of salt, but the equilibrium value can be different for each cycle.
6. CASCADE OF TANKS. So far, we have been interested in tank configurations

that are closed in the sense that the total amount of salt in the system remains constant.
The exercise from [5] mentioned in the introduction is concerned with a configuration
of n tanks T1 , . . . , Tn that is not closed, namely a cascade of n tanks (see Figure 5). For
every i ∈ {1, . . . , n − 1}, there is a pipe transporting brine from Ti to Ti+1 at the rate
of f gallons per unit of time. At the same rate, fresh water flows into tank T1 and brine
flows out of tank Tn , ensuring that the amount of liquid in each tank remains constant.
This type of mixing problem can serve as a model describing the spread of pollution
in a cascade of lakes.
T1
T2
T3
T4
T5
T6
T7
T8
Figure 5. A cascade of tanks
Assuming that the tanks have volumes V1 , . . . , Vn , we obtain the system
x10 (t) = −k1 x1 (t),

xi0 (t) = ki−1 xi−1 (t) − ki xi (t), for 2 ≤ i ≤ n,
where ki = f /Vi . We have

−k1 − λ 0 0 ··· 0 0
 
 k1 −k2 − λ 0 ··· 0 0 
0 k2 −k3 − λ ··· 0 0
 
A − λI =  ,
 
.. .. .. .. .. ..

 . . . . . .


 0 0 0 ··· −kn−1 − λ 0 
0 0 0 ··· kn−1 −kn − λ
and therefore
n
Y
det(A − λI ) = (−ki − λ).
i=1

We focus on the case when k1 = · · · = kn , which corresponds to identical volumes
V1 = · · · = Vn (the reader can verify that, somewhat surprisingly, the case when all
volumes are different is easier to solve). In this case, the number λ = −k1 = · · · = −kn
is an eigenvalue of multiplicity n, and
0 0 0 ··· 0 0
 
λ 0 0 ··· 0 0 
0 λ 0 ··· 0 0 
 
A − λI =  .. .. .. . . . . 
. .. .. 

. . . 
0 0 0 ··· 0 0 
0 0 0 ··· λ 0
has the form of a nilpotent Jordan block. The nullspace of A − λI has dimension 1
and is spanned by the vector v 1 = (0, . . . , 0, 1). Since we need n linearly independent
solutions, we have to look at the nullspaces of higher powers of A − λI . With each
successive power, the only nonzero band containing the values λ shifts one position
lower. It follows that the vector v k = (0, . . . , 0, 1, . . . , 1), whose first n − k compo-
nents are zero and the remaining equal to one, satisfies (A − λI )k v k = 0. This gives a
system of n linearly independent solutions of the form
e At v k = e(A−λI )t eλt v k
2 k−1
t 2t k−1 t

= I + (A − λI ) + (A − λI ) + · · · + (A − λI ) eλt v k ,
1! 2! (k − 1)!
where k ∈ {1, . . . , n}. Since λ < 0, all solutions approach the state with no salt in the
tanks.
7. CONCLUSION. The aim of this article was to present a topic that is accessible
to undergraduate students and displays a nice interplay between differential equations,
linear algebra, and graph theory. We remark that similar mathematical problems occur
in a completely different setting, namely in the study of coordination of multiagent
systems (see [11], [4] and the references there).
A possible project for the interested reader is to find other configurations (closed
or not) of n tanks leading to systems of differential equations that can be solved ana-
lytically. A slightly more ambitious project is to consider mixing problems where we
take into account the time necessary for the transport of brine between two tanks. The
corresponding mathematical model then becomes a linear system of delay differential
equations, which is naturally more difficult to analyze.
ACKNOWLEDGMENTS. I am grateful to Stan Wagon, who brought my attention to mixing problems, and
to the anonymous referees, whose suggestions helped to improve the paper.
REFERENCES
1. R. Agaev, P. Chebotarev, On the spectra of nonsymmetric Laplacian matrices, Linear Algebra Appl. 399
(2005) 157–168, available at http://dx.doi.org/10.1016/j.laa.2004.09.003.
2. R. B. Bapat, Graphs and Matrices, Universitext, Springer, London, 2010.
3. W. E. Boyce, R. C. DiPrima, Elementary Differential Equations and Boundary Value Problems, ninth
edition, Wiley, New York, 2009.
4. P. Yu. Chebotarev, R. P. Agaev, Coordination in multiagent systems and Laplacian spectra of di-
graphs, Autom. Remote Control 70 no. 3 (2009) 469–483, available at http://dx.doi.org/10.1134/
S0005117909030126.

5. C. H. Edwards, D. E. Penney, Elementary Differential Equations, sixth edition, Pearson, Upper Saddle
River, NJ, 2008.
6. S. Elaydi, An Introduction to Difference Equations, third edition, Springer, New York, 2005.
7. C. Godsil, G. Royle, Algebraic Graph Theory, Springer, New York, 2001.
8. A. Jeffrey, Matrix Operations for Engineers and Scientists. An Essential Guide in Linear Algebra,
Springer, Dordrecht, 2010.
9. I. Kra, S. R. Simanca, On Circulant Matrices, Notices Amer. Math. Soc. 59 (2012) 368–377, available at
http://dx.doi.org/10.1090/noti804.
10. S. G. Krantz, Differential Equations Demystified, McGraw-Hill, New York, 2005.
11. R. Olfati-Saber, R. M. Murray, Consensus problems in networks of agents with switching topology and
time-delays, IEEE Trans. Automat. Control 49 no. 9 (2004) 1520–1533.
12. G. Strang, Computational Science and Engineering, Wellesley-Cambridge Press, Wellesley, MA, 2007.
ANTONÍN SLAVÍK is an assistant professor at the Charles University in Prague, where he received his
Ph.D. in 2005. His professional interests include differential equations, integration theory, history of math-
ematics, and computer science.
Charles University, Faculty of Mathematics and Physics, Sokolovská 83, 186 75 Praha 8, Czech Republic
slavik@karlin.mff.cuni.cz
Yet Another Generalization of a Celebrated Inequality of the

Gamma Function
Let n be a positive integer and let p = ( p1 , . . . , pn ) ∈ Rn . If a ∈ R, we let a stand for (a, . . . , a) ∈ Rn .
By p q we mean that pi ≥ qi for i = 1, 2, . . . , n. For a fixed p 0, the generalized unit p -ball is
defined by B p = {x = (x1 , . . . , xn ) ∈ Rn : |x1 | p1 + · · · + |xn | pn ≤ 1}, and its volume is given by (see
[2, 3])

0 1 + p11 0 1 + p12 · · · 0 1 + p1n
n
V p = V (B p ) = 2 . (1)
0 1 + p11 + p12 + · · · + p1n
If 1 p q, then B1 ⊆ B p ⊆ Bq ⊆ B∞ = {x ∈ Rn : max1≤i≤n |xi | ≤ 1}, as is easy to check (in fact,

the proof goes as in the case of the classical unit p-balls). Therefore
V1 ≤ V p ≤ Vq ≤ V∞ = 2n . (2)
Substituting (1) into (2), and denoting αi = 1/ pi and βi = 1/qi , we obtain
Qn Qn
1 i=1 0(1 + αi ) i=1 0(1 + βi )
≤ Pn ≤ Pn ≤ 1, 0 < βi ≤ αi ≤ 1.
n! 0 1 + i=1 αi 0 1 + i=1 βi
n
The case αi = x ∈ [0, 1] gives 1/n! ≤ 0(1 + x) / 0(1 + nx) ≤ 1, which is an inequality that has

sparked much interest in recent years (see [1]).
REFERENCES
1. C. Alsina, M. S. Tomás, A geometrical proof of a new inequality for the Gamma function, J. Ineq.
Pure Appl. Math. 6 (2005) Art. 48.
2. F. Gao, Volumes of generalized balls, Amer. Math. Monthly 120 (2013) 130.
3. X. Wang, Volumes of generalized unit balls, Math. Mag. 78 (2005) 390–395.
—Submitted by Esther M. Garcı́a–Caballero and Samuel G. Moreno,
Departamento de Matemáticas, Universidad de Jaén, 23071 Jaén, Spain, and
Michael P. Prophet, Department of Mathematics, University of Northern Iowa,
Cedar Falls, IA, USA
MSC: Primary: 33B15, Secondary: 26D07

Reciprocal Sums as a Knowledge Metric:
Theory, Computation, and Perfect Numbers
Jonathan Bayless and Dominic Klyve
Abstract. We first provide a short survey of reciprocal sums. We discuss some of the history
of their computation and application, show how they are applied in various modern contexts,
and discuss some ways that their values are computed. We give an example of computing a
reciprocal sum by providing (we believe) the first computation of the sum of the reciprocals
of perfect numbers. Second, we introduce a new use for reciprocal sums; that is, they can be
used as a knowledge metric to classify the current state of number theorists’ understanding of
a given class of integers.
1. INTRODUCTION. Most math majors learn early in their education that there are
at least two “sizes” of infinity: countable and uncountable. Since all countably infinite
sets can be placed in one-to-one correspondence, such sets are said to have the same
cardinality. This beautiful fact, however, can obscure important distinctions among
these sets.
For example, the set of perfect (integer) squares, which we will call S, is an infinite
subset of the integers. Similarly, the set of primes P is also an infinite subset of the in-
tegers, and hence both sets are countable. However, in a very real sense, these two sets
have different “sizes”. Specifically, say we fix a positive real number x and consider
the integers of size at most x. Given a set A of integers, its counting function A(x) is
the number of elements of A up to x; thus S(x) represents the number of squares less
than or equal to x, and P(x) represents the number of primes less than or equal to x.
We can describe the growth of S(x) approximately by writing
√
S(x) = O( x),
which indicates
√ that for sufficiently large x, the value of S(x) is less than a constant
multiple of x. (Of course it is possible to get more precise bounds on S(x), but we
do not need those at this time.)
Similarly, it is known that

x
P(x) = O ,
log x
where log x denotes the natural logarithm of x (see, for example, the discussion and
proof in Chapter 3 of [24]). While both underlying sets are countably infinite, these
bounds let us see that there are more primes than squares up to x (at least for suffi-
ciently large x).
Comparing counting functions is a useful exercise. However, the comparison does
not take into account computational work on finding and enumerating the small ele-
ments of a set. For this, we need a different metric. One interesting measure comes
MSC: Primary 11A99

from something number theorists do often: summing reciprocals. For instance, if we
consider the reciprocal sum of the squares,
X1
,
n∈S
n
not only do we find that this sum converges, but we also know that its precise value is
π2
6
. In contrast, the reciprocal sum of the primes,
X1
n∈P
n
diverges. In this case, the best bounds that apply for x ≥ 3 [13] are
X1
log log x < < log log x + 1.
n∈P
n
n≤x
2. A BIT OF HISTORY. Reciprocal sums have a long and intimate connection to

number theory. They have hinted at great conjectures, helped to prove landmark theo-
rems, and inspired the formation of whole new fields. In a sense, this history stretches
to ancient Egypt, where fractions were represented as the sum of distinct unit fractions.
In the “modern” world, the earliest example of reciprocal sums playing a role in the
development of mathematics is probably due to Nicole d’Oresme, who showed in the
14th century that the harmonic series diverges. His result was lost for some centuries,
and was rediscovered independently three times about 300 years later (see [9, pp. 8–9]
and [17, ch. 2] for details).
Then, in 1744, Euler published a paper in which he used reciprocal sums to compare
the set of primes to the set of integers. The precise way he stated his result looks
strange to modern eyes, but well demonstrates the relationship between the two sets.
In particular, he stated the following [14].
Theorem 1 (Euler). The sum of the reciprocals of the prime numbers,
1 1 1 1 1 1
+ + + + + + ···
2 3 5 7 11 13
is infinitely great, but is infinitely smaller than the sum of the harmonic series
1 1 1 1
1+ + + + + ··· .
2 3 4 5
Also, the sum of the former is the logarithm of the latter.
Neither the proof, nor even the statement of the theorem, are rigorous by modern
standards, but it is not too difficult to see a way to modernize Euler’s claim to get a
deep (and correct!) result, namely that
1
P
log n≤x n
lim P 1
= 1.
x→∞
p≤x p
November 2013] RECIPROCAL SUMS AS A KNOWLEDGE METRIC 823

This formulation of Euler’s claim is correct, and Euler’s result does hint at a logarith-
mic relationship between the integers and the primes. In fact, Sandifer [28] points out
that reciprocal sums presaged the conjecture of the Prime Number Theorem.
A century after Euler’s inspired deductions, reciprocal sums were applied in a rig-
orous proof by Gustav Lejeune Dirichlet. Dirichlet used reciprocal sums to prove that
for any coprime a, m, the sum
X 1
p≤x
p
p≡a (mod m)
diverges [10]. This allowed him to answer one of the most important open questions
of his time; namely, that there are infinitely many prime numbers in any arithmetic
progression.
Interest in reciprocal sums continued throughout the 20th century. Viggo Brun fa-
mously proved [5] that the sum of the reciprocals of the twin primes converges, giving
birth to sieve theory—a major tool in the modern arsenal of analytic number theorists.
His sieve, now called Brun’s Sieve, has inspired many other sieve methods with fa-
miliar names, including the Large Sieve and Selberg’s Sieve [7, Chs. 7 and 8]. Even
more recently, reciprocal sums have motivated research on amicable numbers [1, 25]
and on prime-indexed-primes [2]. Thus, reciprocal sums have continued to influence
mathematics for at least 250 years.
3. A USEFUL MEASURING STICK. Our primary purpose here is to propose a

new application of reciprocal sums. In particular, we claim that looking at the best
bounds currently available on the reciprocal sum of a particular class of integers
provides a good measure of current mathematical knowledge about these integers.
Consider the following: If there were a good measure of our knowledge of a class of
numbers, what properties would it have? We suggest three.
1. Every time we show another integer to be a member of this class, the measure
should (at least slightly) increase.
2. Similarly, every time we show another integer not to be a member of this class,
the measure should increase.
3. If we find better explicit upper or lower bounds on this class of numbers, the
measure should increase.
For a nonempty set, each of these three goals can be met by the following metric.
Definition 2. Given a (nonempty) set of integers with a convergent reciprocal sum, its
Reciprocal Sum Metric is the quotient of the best-known upper bound and the best-
known lower bound on this reciprocal sum.
It is easiest to show how such a metric might be calculated using an example; we

shall give one in the next section, in which we use standard techniques on reciprocal
sums to show a result on the sum of the reciprocals of the perfect numbers.
4. FINDING A RECIPROCAL SUM: A “PERFECT” EXAMPLE. How then do

we estimate the sum of the reciprocals of a set of numbers? Typically, this is done
in three or four steps. First, we explicitly find or compute all elements of a set up
to some value x0 . Second, we calculate explicit upper and lower bounds on the set’s
counting function, which hold for all x > x1 for some x1 . Third, we use these bounds

in a technique based on Abel’s theorem called partial summation (see [24]). If x1 > x0 ,
we employ the fourth step, and find a way to bound the sum over “medium” values in
the range [x0 , x1 ].
Partial summation is a powerful and versatile method that lets us move easily from
estimates of a function to estimates of reciprocal sums, by writing a (discrete) sum
using a (continuous) integral. In the particular case in which we just want to estimate
the sum of the reciprocals of the integers n ≤ x from a set A, we can write
X1 Z x
A(x) A(t)
= + dt,
n∈A
n x n0 t2
n≤x
where A(t) is again the counting function of A up to t and n 0 is the smallest member
of A. Occasionally, the bounds available only apply for very large values of x or have
extremely large constants involved, and thus are only useful for very large x. When
this happens, it is usually a good idea to use some other method to bound the sum over
“medium” values.
We demonstrate this process by computing the reciprocal sum of the perfect num-
bers. To our knowledge, this has never appeared in print. It is known that all even
perfect numbers have the form
2 p−1 (2 p − 1)
where both p and 2 p − 1 are prime. This was shown by Euclid and Euler, and in their
honor we denote the set of perfect numbers by E. As sets go, E is relatively small.
However, the set of odd perfect numbers is probably even smaller. There are none
known, though we do know that the smallest odd perfect number must be greater than
10300 . Furthermore, we can show
√ that the number of odd perfect numbers up to some
(large) value of x is at most x. Since the proof of this is fairly straightforward, we
recall it here. Though this was first proven by Hornfeck [18], we shall follow the clear
exposition in [24, p. 251].
√
Theorem 3. The number of odd perfect numbers up to x is at most x.
Proof. We start by recalling a result of Euler, from his posthumous Tractatus [15], that
an odd perfect number n can be written as n = p k m 2 with p prime and gcd( p, m) = 1.
(Euler’s result itself is proven as Theorem 8.2 in [24].) √ We seek the number of such
n less than or equal to x. Certainly we must have m ≤ x, so we fix such an m and
consider the prime powers p k for which gcd( p, m) = 1. Since n is perfect and σ (n) is
a multiplicative function, we have σ ( p k m 2 ) = σ ( p k )σ (m 2 ) = 2 p k m 2 , and thus
σ ( pk ) 2m 2
= . (1)
pk σ (m 2 )
Since the fraction σ (ppk ) is already in lowest terms, we have that the values of σ (ppk ) are
k k
all distinct as p k ranges over prime powers. Thus,√there is at most one prime power p k
(with p - m) satisfying (1). Therefore,
√ each m ≤ x generates at most one odd perfect
n ≤ x, and there are at most x odd perfect numbers up to x.
To sum the reciprocals of the perfect numbers, we first calculate the sum over known
small values. Thanks to centuries of searching, many even perfect numbers have been

enumerated, and we now know that there are precisely twelve less than 10300 . Through
the simple expedient of taking the reciprocal of these 12 perfect numbers and summing
them, using software permitting arbitrary precision arithmetic, we can accurately cal-
culate the reciprocal sum of all perfect numbers up to 10300 . We denote this value by
S, so that S :=
X 1
≈ 0.20452014283892643017813442909845557667731148935076
n∈E
n 33970064248248986227440451319854070768557639532877
n≤10300
0871819162718674223226559418497366349720745705892.
In this case, the “≈” symbol indicates that this sum is accurate to 150 decimal
places. It now remains for us to put bounds on the reciprocal sum of perfect numbers
greater than 10300 . We do this by splitting them into two cases, considering even and
odd perfect numbers separately.
We begin with the sum of reciprocals of the odd perfect numbers.
Theorem 4. The sum of the reciprocals of all odd perfect numbers greater than 10300
is less than 2 × 10−150 .
Proof. We let E o be the set of odd perfect numbers,

√ and E o (x) its associated counting
function. Recall from above that E o (x) < x for all positive values of x. If no odd
perfect number exists, then E o (x) = 0 for all x, and the sum in question is zero. Other-
wise, we know that the smallest odd perfect number n 0 , if it exists, satisfies n 0 > 10300 .
We now bound our sum using partial summation:
Z x√
E o (x) E 0 (t)
X 1 X1
= lim = lim + 2
dt .
n∈E
n x→∞
n∈E
n x→∞ x n 0
t
o o
n>10300
√
Since n 0 > 10300 and E o (t) < t for all positive t, this is at most
√ Z x √ Z ∞ √
x t t
lim + 2
dt = 2
dt < 2 × 10−150 ,
x→∞ x 10 300 t 10 300 t
giving the result.
In order to sum the reciprocals of even perfect numbers greater than 10300 , we recall
the Euclid–Euler criterion that all such numbers are of the form 2 p−1 (2 p − 1) for prime
p. We use this result in the following.
Theorem 5. The sum of the reciprocals of all even perfect numbers greater than 10300
is less than 2 × 10−300 .
Proof. If an even perfect number m is larger than 10300 , then m = 2n−1 (2n − 1) for
some n ≥ 499. We derive an upper bound on the sum of the even perfect reciprocals
by assuming that in fact, the expression 2n−1 (2n − 1) gives a perfect number for all
n ≥ 499. We sum all such values, finding
X 1 X 1 X 1
= < < 2 × 10−300 ,
n≥499
(2 n−1 )(2n − 1)
n≥499
22n−1 − 2n−1
n≥499
2 2n−2
from which the theorem follows.

Of course, much better results are possible. Given a more careful sum, and a better
use of existing perfect numbers, we could determine this sum to millions of decimal
places. Even the bound above, however, is more than we need for our purposes.
Combining these results, we have the following.
Theorem 6. The sum of the reciprocals of the perfect numbers greater than 10300 is
less than 3 × 10−150 .
This gives us the pleasing conclusion that the explicitly calculated value of S above
gives the reciprocal sum of all perfect numbers with an accuracy of at least 149 decimal
places.
Theorem 7. The sum of the reciprocals of the perfect numbers lies in the range
[S, S + 3 × 10−150 ].
5. SOME KNOWN RESULTS. We have seen how we might compute a reciprocal

sum, and we have found explicit bounds on the reciprocal sum of the perfect num-
bers. Let us now compare these bounds for reciprocal sums on the perfect numbers to
reciprocal sums over some other classes of numbers. For the sequences above whose
reciprocal sums converge, we give the current value of their reciprocal sum metric in
Table 1.
Table 1. The Reciprocal Sum Metric (RSM) on the current knowledge of some arithmetic sequences
Sequence Lower Bound (L) Upper Bound (U) RSM (U/L)
Perfect numbers S S + 3 × 10−150 1 + 10−149

Prime-indexed primes 1.04299 1.04365 1.00063
Twin primes 1.83408 2.346765 1.2795
Amicable numbers .01198 6.56 × 108 5.476 × 1010
5.1. Prime-indexed primes. The prime-indexed primes are primes whose index in
the list of primes is itself prime. This sequence was first considered in 1975, and since
then mathematicians have improved their knowledge of the behavior of these integers
[2, 4, 12]. Using extensive computations and the best-available explicit bounds on
primes [13, 27], the best bounds on this reciprocal sum have been strengthened [2] to
X1
1.04299 < < 1.04365,
q
q
where q is taken over the set of prime-indexed primes.
5.2. Twin primes. When Brun proved that the reciprocal sum of the twin primes
converges, his result was notable for the fact that in order to show this, he gave, for the
first time, an effective upper bound on the density of twin primes (here we can think
of the “density” of a set as its counting function). The prime numbers have asymptotic
density x/ log(x), and heuristically we expect that twin primes have density x/ log2 x.
Although Brun was not able to show this precisely, he was able to show that π2 (x), the
number of twin primes up to x, satisfies
x(log log x)2

π2 (x) ≤ c
log2 x

for sufficiently large x, where c is a constant that could “in principle” be computed.
Brun himself made no attempt to find the value of the reciprocal sum of twin primes,
later named Brun’s Constant, denoted B2 . It wasn’t until 1975 that Brent put an actual
upper bound on B2 [3]. Since then, several authors have further refined his bounds (see
[8, 19, 29]). The best-known bounds are
1.83 < B2 < 2.347.
5.3. Amicable numbers. Amicable numbers have been studied at least since the
time of the Pythagoreans. (The reader who has not thought lately about elementary
number theory or ancient Greek mathematics may wish to be reminded that two pos-
itive integers m, n are amicable if s(m) = n and s(n) = m, where s(n) is the sum
of the divisors of n other than n itself.) Amicable numbers remain fairly mysterious
objects. It was only in 1981 that Pomerance showed that the reciprocal sum of the
amicable numbers converges [25]. This sum is known as the Pomerance constant, and
is denoted P. No explicit bound was known for this sum until 2009, when the authors
showed [1] that
0.01198 < P < 6.56 × 108 .
Note that the large gap between the lower and upper bounds on P reflects a deep
lack of understanding about the behavior of amicable numbers, particularly for small
values, which may have been difficult to notice without considering reciprocal sums.
All of the insights above about the accuracy of bounding these sums can be effec-
tively captured by the reciprocal sum metric from Definition 2.
6. A MEASURE FOR DIVERGENT SUMS. A bit of thought suggests that our

definition of a reciprocal sum metric can be easily modified for divergent sums. In this
case, we shall compare the best-known upper and lower bounds on the partial sums of
the series.
Definition 8. Given a sequence A whose reciprocal sum is known to be divergent, let

U (x) and L(x) be the best-known bounds such that there exists an x0 with
X1
L(x) ≤ ≤ U (x)
n∈A
n
n≤x
for all x ≥ x0 . Then the Reciprocal Sum Metric is
U (x)
lim .
x→∞ L(x)
To illustrate this definition, we note that the reciprocal sum metric can tell us some-
thing interesting about the history of our knowledge of primes. Using the methods
described in Section 4, it is straightforward to show that if we know constants c1 , c2
such that c1 logx x < π(x) < c2 logx x , then the reciprocal sum metric is c2 /c1 . Since the
Prime Number Theorem was proven in 1896 by de la Valeé Poussin [26] and Hadamard
[16], the Reciprocal Sum Metric for the primes has been precisely 1. Before the proof,

however, other work had been done relating to the prime number theorem. Most inter-
estingly for our purposes, Chebychev proved in 1852 [6] that
π(x) π(x)
.92129 ≤ lim inf ≤ 1 ≤ lim sup ≤ 1.10555.
x→∞ x/ log x x→∞ x/ log x
This work can be neatly encapsulated by the Reciprocal Sum Metric.
Table 2. The historical progress of the Recipro-

cal Sum Metric (RSM) of the set of primes
Year RSM for primes
< 1850 ∞
1850 1.200
1896 1
7. ONGOING WORK, UNANSWERED QUESTIONS, AND INFINITELY BAD

BOUNDS. Interestingly, there are some well-known sequences of numbers for which
the reciprocal sum metric can’t be shown to be less than any given value. In this case,
we shall say that the RSM of the sequence is infinity. For an example, consider the
Wieferich primes.
Fermat showed that for all odd prime numbers p, 2 p−1 ≡ 1 (mod p). If additionally
we have that 2 p−1 ≡ 1 (mod p 2 ), then p is said to be a Wieferich prime. There are only
two known Wieferich primes, 1093 and 3511, and it is known that there are no others
less than 6.7 × 1015 [11]. Even so, there is a great deal about Wieferich primes that we
don’t know. It is unknown whether there are infinitely many Wieferich primes. It is also
unknown whether there are infinitely many primes which are not Wieferich primes!1
The reciprocal sum of the Wieferich primes could be precisely 1/1093 + 1/3511, but
the sum may diverge. Thus the RSM of the Wieferich primes is infinity.
8. CONCLUSION. Humans love to break records. Being able to break a record nec-
essarily implies that something relating to the record is measurable. In sport, this is
often easy—running 100 meters faster than anyone else neatly proves that the runner
is the fastest. Even in computational mathematics, records are easy to measure and
break. There is a record for the largest prime yet discovered, a record for the best
bound on the density of abundant numbers [20], and a record for the highest level to
which the Goldbach Conjecture has been verified [22]. These records seem to be good
for us; their very existence pushes the mathematical community to greater efforts. It is
our hope that the reciprocal sum metric may help motivate somebody to continue the
centuries’ long tradition of exploring the values and applications of reciprocal sums.
9. ADDENDUM: KNOWLEDGE PROGRESSES. When this paper was in page

proof, we discovered a 2012 result [21] showing that the smallest odd perfect number
is at least 101500 . The reader is invited to verify that with this new information, the
RSM of the perfect numbers is at most 1 + 10−750 .
1 We note that the existence of infinitely many non-Wieferich primes does, however, follow from the ABC
conjecture. If Shinichi Mochizuki’s proof of the ABC conjecture [23] is verified, one of the many consequences
will be to render Wieferich primes a bit less mysterious.

ACKNOWLEDGMENTS. Paul Pollack suggested to us years ago that it would be easy to bound the recipro-
cals of the perfect numbers. Rhea Caldwell, Jonathan Sands, and Steve Kokoska unwittingly worked together
to provide a conducive work environment for this paper. We also thank two referees, whose suggestions mea-
surably improved the final product.
REFERENCES
1. J. Bayless, D. Klyve, On the sum of reciprocals of amicable numbers, Integers 11A (2011) Article 5.
2. J. Bayless, D. Klyve, T. Oliviera e Silva, New bounds and computations on prime-indexed primes, to
appear in Integers.
3. R. Brent, Irregularities in the distribution of primes and twin primes, Math. Comp. 29 (1975) 43–56,
available at http://dx.doi.org/10.1090/S0025-5718-1975-0369287-1.
4. K. A. Broughan, A. R. Barnett, On the subsequence of primes having prime subscripts, J. Integer Seq. 12
(2) (2009).
5. V. Brun, La serie 1/5 + 1/7 + 1/11 + 1/13 + 1/17 + 1/19 + 1/29 + 1/31 + 1/41 + 1/43 + 1/59 +
1/61 + . . ., les dénominateurs sont nombres premiers jumeaux est convergente oú finie, Bull. Sci. Math.
43 (1919) 124–128.
6. P. Chebyshev, Mémoire sur les nombres premiers. Académie Impériale des Sciences, 1850.
7. A. Cojocaru, M. R. Murty, An Introduction to Sieve Methods and Their Applications, London Mathemat-
ical Society Texts, Vol. 66. Cambridge University Press, Cambridge, 2006.
8. R. Crandall, C. Pomerance, Prime Numbers: A Computational Perspective. second edition. Springer,
New York, 2005.
9. J. Derbyshire, Prime Obsession: Bernhard Riemann and the Greatest Unsolved Problem in Mathematics,
Plume, New York, 2004.
10. P. G. L. Dirichlet, Über eine neue Anwendung bestimmter Integrale auf die Summation endlicher oder
unendlicher Reihen. Abh. Königl. Pr. Wiss. Berlin (1835) 391.
11. F. Dorais, D. Klyve, A Wieferich Prime Search up to 6.7 × 1015 , J. Integer Seq. 14 (2011) Article 11.9.2.
12. R. E. Dressler, S. T. Parker, Primes with a prime subscript, J. Assoc. Comput. Mach. 22 (1975) 380–381,
available at http://dx.doi.org/10.1145/321892.321900.
13. P. Dusart, Autour de la fonction qui compte le nombre de nombres premiers. Ph.D. thesis, Université de
Limoges, Limoges, France, 1998.
14. L. Euler, Variae observationes circa series infinitas. (E72) Commentarii academiae scientiarum Petropoli-
tanae 9, 1744, 160–188. Republished in Opera Omnia: Series 1, Vol. 14, 217–244. A scan of the orig-
inal paper and an English translation by P. Viader Sr., L. Bibiloni, and P. Viader Jr. are available at
www.eulerarchive.org.
15. , Tractatus de numerorum doctrina capita sedecim quae supersunt, Commentationes arithmeticae
2, 1849, 503–575. Republished in Opera Omnia: Series I, Vol. 5, 182–283. Also available online at
eulerarchive.maa.org.
16. J. Hadamard, Sur la distribution des zéros de la fonction ζ (s) et ses conséquences arithmétiques, Bull.
Soc. Math. France 24 (1896) 199–220.
17. J. Havil, The Harmonic Series, in Gamma: Exploring Euler’s Constant. Princeton University Press,
Princeton, NJ, 2003, 21–25.
18. B. Hornfeck, Zur Dichte der Menge der vollkommenen Zahlen, Arch. Math. (Basel) 6 (1955) 442–443,
available at http://dx.doi.org/10.1007/BF01901120.
19. D. Klyve, Explicit Bounds on Twin Primes and Brun’s Constant. Ph.D. Thesis, Dartmouth College, 2007.
20. M. Kobayashi, On the Density of Abundant Numbers. Ph.D. Thesis, Dartmouth College, 2010.
21. P. Ochem, M. Rao, Odd perfect numbers are greater than 101500 , Math. Comp., 81 (2012) 1869–1877,
available at http://dx.doi.org/10.1090/S0025-5718-2012-02563-4.
22. T. Oliveira e Silva, S. Herzog, S. Pardi, Empirical Verification of the Even Goldbach Conjecture up to
4 · 1018 , to appear in Math Comp.
23. S. Mochizuki. Inter-universal teichmuller theory I-IV, Available on the author’s webpage, http://www.
kurims.kyoto-u.ac.jp/~motizuki/top-english.html.
24. P. Pollack, Not Always Buried Deep, American Mathematical Society, Providence, RI, 2009.
25. C. Pomerance, On the distribution of amicable numbers. II, J. Reine Angew. Math. 325 (1981) 183–188.
26. C. J. de la Vallée Poussin, Recherches analytiques sur la théorie des nombres (3 parts). Ann. Soc. Sci.
Bruxelles 20, Part II (1896) 183–256, 281–397.
27. J. B. Rosser, L. Schoenfeld, Sharper bounds for the Chebyshev functions θ(x) and ψ(x), Math. Comp.
29 (1975) 243–269.

28. E. Sandifer, Infinitely Many Primes, Ch. 33, in How Euler Did It, Mathematical Association of America,
Washington, DC, 2007.
29. D. Shanks, J. W. Wrench, Jr., Brun’s Constant, Math. Comp. 28 (1974) 293–299; Corrigendum, ibid, 28
(1974) 1183.
JONATHAN BAYLESS is an assistant professor and mathematics coordinator at Husson University in Ban-
gor, Maine. In addition to number theory, his interests include philosophy and inquiry-based learning. He
enjoys logic puzzles, playing basketball, and spending time with his family.
Husson University, 1 College Circle, Bangor, ME 04401
baylessj@husson.edu
DOMINIC KLYVE is an associate professor at Central Washington University, where he also serves as the
Director of the Math Honors Program. He works in number theory and the history of mathematics, and is
always pleased to have the chance to combine these fields in a single work. He also enjoys learning about
new (non-mathematical) fields, and looking for opportunities to use mathematics to help him understand them
better.
Central Washington University, 400 E University Way, Ellensburg WA 98926
klyved@cwu.edu
Another Simple Proof that the Sum of the Reciprocals

of the Primes Diverges
Let pi denote the ith prime number, and suppose the sum converges. Then there is an index k
such that
∞
X 1
< 1.
p
i=k+1 i
Take A to be the set of positive integers whose prime factors are all ≤ pk , and let B be the
positive integers whose prime factors are all ≥ pk+1 . Note that 1 is vacuously a member of
both A and B since it has no prime factors. By the fundamental theorem of arithmetic, every
positive integer can be uniquely expressed as a product ab for some a ∈ A and b ∈ B. We
exploit the convergence of the geometric series to get
   
X1 ∞ ∞ ∞ ∞
X X 1 X 1 X 1  < ∞.
= ··· n n =
 n
··· n
a∈A
a n =0
p 1 · · · pk k
n =0 1
p1
n =0 1 n =0 k
pk
1 k 1 k
Letting Bm denote the members of B with exactly m prime factors (not necessarily distinct),
we also get
∞ X ∞ ∞
!m
X1 X 1 X X 1
= ≤ < ∞.
b∈B
b m=0 b∈B
b m=0 i=k+1 pi
m
Combining these sums then produces our contradiction:

∞
! !
X 1 XX 1 X1 X1
= = < ∞.
n=1
n a∈A b∈B
ab a∈A
a b∈B
b
—Submitted by Dustin G. Mixon
MSC: Primary 11A41

NOTES
Edited by Sergei Tabachnikov
Certain Inequalities Associated with the

Divisor Function
Jorge Luis Cimadevilla Villacorta
Dedicado a mi hijo Jorge Cimadevilla Pereira
Abstract. In this paper, we prove certain inequalities of the type

n
X n
X
d(si + t) ≥ d(ui + v),
i=1 i=1
where d(n) is the divisor function and s, t, u, v ∈ Z.
1. INTRODUCTION. The divisor function, denoted d(n), counts the number of

positive divisors of n and can be written as
X
d(n) = 1.
d|n
Over the centuries, this function and its various generalizations have turned up in many
central problems in number theory.
One place where the function has appeared is in countingPthe number Pof lattice
points below a hyperbola. It is a simple matter to show that x y≤c 1 = n≤c d(n),
where the former sum can be interpreted as the number of integer lattice points (x, y)
lying on or below the hyperbola x y = c, which are in the open first quadrant region
x > 0 and y > 0.
In 1849, Dirichlet [1] proved that
X
d(n) = x log x + (2γ − 1)x + q(x), (1.1)
n≤x
where | q(x) |≤ 4x 1/2 and γ is the Euler–Mascheroni constant. Since then, there have
been many papers devoted to exploring q(x) (we refer the reader to [4]–[10]).
Another noteworthy result is the paper written by J. B. Friedlander and H. Iwaniec
[3], in which asymptotic formulas for the sum of the divisor function applied to arith-
metic progressions are established.
InP this brief note, we establish some unanticipated inequalities among sums of the
n
type i=1 d(kq + a), using elementary q-series identities. The proofs reduce to show-
ing that the coefficients of certain infinite series are nonnegative. Such inequalities
seem to have no precedents in the mathematical literature.
Our main results are given in the following theorem.
MSC: Primary 11A25, Secondary 26D15

Theorem 1.1. The following inequalities hold:
n
X n
X
d(24i + 1) ≥ d(6i + 1), (1.2)
i=1 i=1
n
X n
X
d(24i + 17) ≥ d(6i + 5), (1.3)
i=1 i=1
n
X n
X
d(3(8i + 3)) ≥ d(3(2i + 1)), (1.4)
i=1 i=1
n
X n
X n
X
d(8i + 1) ≥ d(4i + 1) ≥ d(2i + 1), (1.5)
i=1 i=1 i=1
n
X n
X
d(8i + 5) ≥ d(4i + 3), (1.6)
i=1 i=1
and
n
X n
X
d(72i + 1) ≥ d(18i + 1). (1.7)
i=1 i=1
2. THREE SERIES. To prove the inequalities in Theorem 1.1, we need to establish

three identities, which are given in the following lemma. These identities are proved
using the identity
∞ ∞ ∞
2 1 + qn X qn
qn d(n)q n .
X X
= = (2.1)
n=1
1 − qn n=1
1 − q n
n=1
According to N. J. Fine [3, p. 16, eq. (14.5)], (2.1) is due to T. Clausen.
Lemma 2.1. Let n be a positive integer and q be a complex number such that |q| < 1.
It follows that
∞ ∞
1 + q 2n+1
d(2n + 1)q n = q 2n(n+1)
X X
, (2.2)
n=0 n=0
1 − q 2n+1
∞ ∞
n 1 + q 2n+1
q n(n+1)
X X
d(4n + 1)q = , (2.3)
n=0 n=0
1 − q 2n+1
and
∞ ∞
X
n
X n(n+1) 1 + q 2n+1
d(8n + 1)q = q 2 . (2.4)
n=0 n=0
1 − q 2n+1
November 2013] NOTES 833

Proof. We first give the proof of (2.2). We rewrite (2.1) by splitting the sums in both
sides into sums over odd and even integers. This yields
∞ ∞ ∞ ∞
2 1 + q 2n X (2n+1)2 1 + q 2n+1
d(2n)q 2n + d(2n + 1)q 2n+1 =
X X X
q (2n) + q .
n=1 n=0 n=1
1 − q 2n n=0
1 − q 2n+1
(2.5)
Replacing q by −q in (2.5), we find that
∞ ∞ ∞ ∞
2 1 + q 2n X (2n+1)2 1 − q 2n+1
d(2n)q 2n − d(2n + 1)q 2n+1 =
X X X
q (2n) − q .
n=1 n=0 n=1
1 − q 2n n=0
1 + q 2n+1
(2.6)
Subtracting (2.5) from (2.6), we deduce that
∞ ∞
1 + q 2n+1 1 − q 2n+1

2
d(2n + 1)q 2n =
X X
2q q (2n+1) +
n=0 n=0
1 − q 2n+1 1 + q 2n+1
∞
2 +n) 1 + q 2(2n+1)
q 4(n
X
= 2q . (2.7)
n=0
1 − q 2(2n+1)
Next, we divide both sides of (2.7) by q, and replace q 2 by q to obtain (2.2). The iden-
tity (2.2) is obtained by first replacing q by −q in (2.2) and then adding the resulting
identity to (2.2). The final result is
∞ ∞ ∞
1 + q 2n+1
d(2n + 1)q n + (−1)n d(2n + 1)q n = q 2n(n+1)
X X X
n=0 n=0 n=0

1 − q 2n+1
∞
1 − q 2n+1
q 2n(n+1)
X
+ , (2.8)
n=0
1 + q 2n+1
from which (2.2) follows. The proof of (2.3) is identical to that of (2.2), and so we
leave this to the reader.
3. AN ELEMENTARY METHOD. In this section, we complete the proof of

Theorem 1.1. First, subtract (2.2) from (2.4) to conclude that
∞ ∞
1 + q 2n+1
n
q n(n+1)/2 (1 − q 3n(n+1)/2 )
X X
(d(8n + 1) − d(2n + 1))q = . (3.1)
n=1 n=1
1 − q 2n+1
Since
∞
1
q 3n =
X
,
n=0
1 − q3
we find that
∞
! ∞ ! ∞
3n n 1−q 3n(n+1)/2 1+q 2n+1
q n(n+1)/2
X X X
q (d(8n +1) − d(2n +1))q = .
n=0 n=1 n=1
1−q 3 1−q 2n+1
(3.2)

Observe that the coefficients of q 3m , q 3m+1 , and q 3m+2 on the left-hand side of (3.2) are
m
X
(d(8(3i) + 1) − d(2(3i) + 1)),
i=1
m
X
(d(8(3i + 1) + 1) − d(2(3i + 1) + 1)),
i=1
and
m
X
(d(8(3i + 2) + 1) − d(2(3i + 2) + 1)),
i=1
and that the terms

1 − q 3k 1 + qk
and
1 − q3 1 − qk
on the right-hand side of (3.2) have Taylor series expansions with nonnegative coeffi-
cients. We complete the proofs of (1.2), (1.3), and (1.4) using these observations.
Following the above method, we subtract (2.2) from (2.3) to deduce that
∞ ∞
n 1 + q 2n+1
q n(n+1) (1 − q n(n+1) )
X X
(d(4n + 1) − d(2n + 1)) q = . (3.3)
n=1 n=1
1 − q 2n+1
Next, we multiply both sides of (3.3) by

∞
1
q 2n =
X
n=0
1 − q2
and find that

∞
! ∞ ! ∞
2n n 1 − q n(n+1) 1 + q 2n+1
q n(n+1)
X X X
q (d(4n + 1) − d(2n + 1)) q = .
n=0 n=1 n=1
1 − q 2 1 − q 2n+1
(3.4)
By arguing as previously, the proofs of

m
X
(d(4(2i) + 1) − d(2(2i) + 1)) ≥ 0
i=1
and
m
X
(d(4(2i + 1) + 1) − d(2(2i + 1) + 1)) ≥ 0
i=1
1
now follow. Another possibility is to multiply both sides of (3.3) by 1−q
to obtain the
right side of (1.5).

4. A MORE DEMANDING RESULT. The inequality (1.7) is more difficult to
prove. We split (3.1) into three cases, according to whether m ≡ 0, 1, or 2 (mod 3),
∞ ∞ 1 + q 6m+1
3m(3m+1) 9m(3m+1)

(d(8m + 1) − d(2m + 1))q m =
X X
q 2 1−q 2
m=1 m=1
1 − q 6m+1
∞ 1 + q 3(2m+1)
9m(m+1) 3(3m+2)(3m+1)
X
+ q 2 +1
1−q 2
m=1
1 − q 3(2m+1)
∞ 1 + q 6m+5
3(3m+2)(m+1) 9(3m+2)(m+1)
X
+ q 2 1−q 2 . (4.1)
m=1
1 − q 6m+5
In the second series on the right side, all the exponents of q are of the form 3n + 1.
Therefore, only the first and second sum on the right-hand side of (4.1) will contribute
to the coefficient of q 9k .
Thus, if
∞
1
q 9i =
X
i=0
1 − q9
is multiplied by (4.1), then

∞
! ∞
!
(d(8m + 1) − d(2m + 1))q m q 9i
X X
m=1 i=0
∞ 9m(3m+1)
X 3m(3m+1) 1−q 2 1 + q 6m+1
= q 2 · ·
m=1
1 − q9 1 − q 6m+1
∞ 3(3m+2)(3m+1)
X 9m(m+1) 1−q 2 1 + q 6m+3
+ q 2+1 · ·
m=1
1 − q9 1 − q 6m+3
9(3m+2)(m+1)
∞
!
X 3(3m+2)(m+1) 1−q 2 1 + q 6m+5
+ q 2 · · . (4.2)
m=1
1−q 9 1 − q 6m+5
is found.
In the second series on the right side, the Taylor series expansion of
3(3m+2)(3m+1)
1−q 2
1 − q9
does not consist solely of positive coefficients. However, the coefficient of q 9m in (4.2)
is always positive or zero. Therefore,
m
X
(d(72n + 1) − d(18n + 1)) ≥ 0
n=1
now follows.

ACKNOWLEDGMENTS. The author would like to thank Fernando Chamizo, Jonathan Sondow, and espe-
cially my friends Chan Heng Huat and Jesus Guillera Goyanes for their advice and support in the writing of
this paper.
REFERENCES
1. P. G. L. Dirichlet, Über die Bestimmung der mittleren Werthe in der Zahlentheorie, Abhandlungen der
Königlich Preussischen Akademie der Wissenchaften 2 (1849) 69–83.
2. N. J. Fine, Basic Hypergeometric Series and Applications. Mathematical Surveys and Monographs, 27,
American Mathematical Society, Providence, RI, 1988.
3. J. B. Friedlander, H. Iwaniec, The Divisor Problem for Arithmetic Progressions, Acta Arithmetica 45 (3)
(1985) 273–277.
4. M. N. Huxley, Exponential Sums and Lattice Point, Proc. London Math. Soc. 60 (1990) 471–502.
5. , Exponential Sums and Lattice Points II, Proc. London Math. Soc. 66 (1993) 279–301.
6. , Exponential Sums and Lattice Points III, Proc. London Math. Soc. 87 (2003) 591–609, available
at http://dx.doi.org/10.1112/S0024611503014485.
7. G. A. Kolesnik, An Improvement of the Remainder Term in the Divisor Problem, Mat. Zametki 6 (1969)
545–554.
8. J.G van der Corput, Zum Teilerproblem, Math. Ann. 98 (1928) 697–716, available at http://dx.doi.
org/10.1007/BF01451619.
9. I. M. Vinogradov, Anzahl der Gitterpunkte in der Kugel, Traveaux Inst. Phys. Math. Stekloff 9 (1935)
17–38.
10. G. Voronoı̈, Sur un probleme du calcul des fonctions asymptotiques, Journal für die reine und ange-
wandte Mathematik 126 (1903) 241–282.
Gijon, Asturias, Spain

villacorta1968@hotmail.com
On the Nonexistence of Certain Limits

for the Complex Exponential
Charles S. Kahane
Abstract. With {n j } an arbitrary subsequence of natural numbers, it is shown that
lim ein j α
j→∞
does not exist for almost all α ∈ R. The result is then applied to provide an alternative proof
that L 1 (R) is not weakly sequentially compact.
1. INTRODUCTION. It is well known that the sequence {einα }, for α a real number,
is divergent when α is an irrational multiple of 2π. More generally, consider the se-
quences {ein j α } with {n j } a subsequence of natural numbers. It is the purpose of this
note to explore the divergence properties of such sequences.
As a first example, consider the sequence {ei(an+b)α } with a and b positive integers.
For α an irrational multiple of 2π , this sequence is divergent. To see this, assume that
lim ei(an+b)α = L
n→∞

exists with L 6 = 0, since |ei x | = 1 for x real. It follows that the same exponential with
n replaced by n + 1 also converges to L. Hence,
eiaα = lim ei([n+1]a+b)α e−i(na+b)α = L/L = 1.
n→∞
But this is impossible since α is an irrational multiple of 2π.

Next, consider subsequences generated by polynomials
k
ajn j
X
p(n) =
j=0
with positive integer coefficients. We claim that if α is an irrational multiple of 2π , then

{ei p(n)α } is also a divergent sequence. This is established inductively on the degree k of
such polynomials, with the result for k = 1 having been established in the preceding
paragraph.
Suppose now that the assertion has been proven for polynomials of degree k − 1.
Then to prove it for polynomials p(n) of degree k, assume that
lim ei p(n)α = L 6 = 0
n→∞
exists. The same then holds with n replaced by n + 1; consequently,

lim ei[ p(n+1)− p(n)]α = L/L = 1.
n→∞
But p(n + 1) − p(n) is a polynomial with positive integer coefficients of degree k − 1.

Hence, by the induction hypothesis, the last limit cannot exist; this is a contradiction
that establishes the divergence of {ei p(n)α } for α an irrational multiple of 2π.
From the examples above, it would appear that the set of α’s for which divergence
occurs is rather substantial. This then suggests the following question: With {n j } denot-
ing an arbitrary subsequence, can we expect the same for the sequence {ein j α }? More
precisely, how extensive is the set of α’s for which {ein j α } is divergent? An answer to
this question is furnished by the following.
Theorem. If {n j } is any subsequence of natural numbers, then the lim j→∞ ein j α fails
to exist for almost all α ∈ R.
2. PROOF OF THE THEOREM. Consider α ∈ [0, 2π ). Assume, on the contrary,

that lim j→∞ ein j α existed on a set E ⊂ [0, 2π ) of positive measure. Set f (α) =
lim j→∞ ein j α on E and note that by the Lebesgue dominated convergence theorem,
Z Z
f (α) dα = lim ein j α dα.
E j→∞ E
According to the Riemann–Lebesgue lemma,

Z 2π
lim einα g(α) dα = 0
n→∞ 0
for g integrable on [0, 2π ] (see [4, p. 2]). Applying it with g = χ E (α), the character-
istic function of the set E, we find that
Z Z 2π
inα
lim e dα = lim einα χ E (α) dα = 0;
n→∞ E n→∞ 0

and thus for the subsequence
Z
lim ein j α dα = 0
j→∞ E
as well. Consequently,
Z
f (α) dα = 0.
E
The same argument can be applied to any measurable subset F of E, and so

Z
f (α) dα = 0
F
holds for any such subset. From this it follows that the subsets of E on which
Re f (α) and Im f (α) are positive or negative have measure zero, and therefore
we conclude that f (α) = 0 almost everywhere on E. But this is impossible, since
| f (α)| = lim j→∞ |ein j α | = 1 on E; hence lim j→∞ ein j α does not exist on any set of
positive measure in [0, 2π).
A similar result holds for the sequences {sin(n j α)} and {cos(n j α)}. We sketch
the proof for the sine sequence. Assuming that lim j→∞ sin(n j α) exists on a set E ⊂
[0, 2π ) of positive measure, the same reasoning as used above shows that this is only
possible if lim j→∞ sin(n j α) = 0 almost everywhere on E. But then
lim sin(2n j α) = lim 2 sin(n j α) cos(n j α) = 0

j→∞ j→∞
and
lim cos(2n j α) = lim [1 − 2 sin2 (n j α)] = 1

j→∞ j→∞
almost everywhere on E. Hence,
lim ei2n j α = 1
j→∞
almost everywhere on the set E of positive measure, in contradiction to the theorem.
3. AN APPLICATION. The theorem may be used to provide another proof that

L 1 (R) is not weakly sequentially compact. That is, not every norm bounded sequence
of functions in L 1 (R) contains a subsequence that converges weakly to another func-
tion in L 1 (R). For this purpose, we take as the terms of our sequence the translates
f (x + n), for n = 1, 2, . . . , of an appropriate f (x) ∈ L 1 (R), to be determined in
a moment. Clearly, as k f (x + n)k L 1 = k f (x)k L 1 for n = 1, 2, . . . , our sequence is
norm bounded. Consider now any subsequence { f (x + n j )}; since e−iαx is an L ∞ (R)
function for any choice of α and hence generates a bounded linear functional on L 1 (R),
it will suffice to show that there exists an α so that
Z
e −iαx
f (x + n j ) d x
R

does not approach a limit as j → ∞. Noting that
Z Z
e −iαx
f (x + n j ) d x = e−iα(x−n j ) f (x) d x
R R
Z
= ein j α e−iαx f (x) d x,
R
we can assure the divergence of

Z
e −iαx
f (x + n j ) d x
R
by choosing an α for which lim j→∞ ein j α does not exist, provided that the last integral
on the right, the Fourier integral of f , never vanishes for α ∈ R. The latter can be
arranged for by (for example) taking f (x) = e−x for x ≥ 0, and f (x) = 0 for x < 0,
in which case
1
Z
e−iαx f (x) d x = 6= 0
R 1 + iα
for α ∈ R.
This takes care of the complex case. The real case is taken care of by observing that
at least one of
Z Z
cos(−αx) f (x + n j ) d x or sin(−αx) f (x + n j ) d x
R R
must diverge as j → ∞.
Another direct proof of the failure of weak sequential compactness for L 1 (R) can
be found in [1, p. 173]. An indirect proof of this failure can be based on the equivalence
of weak sequential compactness with reflexivity (see [2, p. 119 and p. 251] as well as
[5, p. 141]). Since L 1 (R) is not reflexive [3, p. 23], it follows immediately that it is not
weakly sequentially compact.
ACKNOWLEDGMENTS. I wish to thank the reviewers for suggestions that greatly improved the exposition.
REFERENCES
1. P. M. Fitzpatrick, H. L. Royden, Real Analysis, fourth edition, Prentice Hall, Boston, 2010.
2. R. E. Megginson, An Introduction to Banach Space Theory, Springer-Verlag, New York, 1998.
3. R. Shakarchi, E. M. Stein, Functional Analysis, Princeton University Press, Princeton, NJ, 2011.
4. E. M. Stein, G. Weiss, Fourier Analysis on Euclidean Spaces, Princeton University Press, Princeton, NJ,
1971.
5. K. Yosida, Functional Analysis, sixth edition, Springer-Verlag, Berlin, 1980.
Department of Mathematics, Vanderbilt University, Nashville, TN 37240

c.kahane@att.net
Editor’s Note: Dr. Charles S. Kahane and his wife, Claire, were killed in an automobile accident in
May of this year. Charles received his Ph.D. from New York University in 1962, under the direction
of Professor Louis Nirenberg. Before coming to Vanderbilt University in 1969, he was an Assistant
Professor at the University of Minnesota. He retired from Vanderbilt in 1998 as Professor of Math-
ematics. His seminal work was on nonlinear parabolic partial differential equations. We extend our
deepest condolences to Dr. and Mrs. Kahane’s family.

A Generalization of
Routh’s Triangle Theorem
Árpád Bényi and Branko Ćurgus
Abstract. We prove a generalization of the well-known Routh’s triangle theorem. As a con-

sequence, we get a unification of the theorems of Ceva and Menelaus. A connection to Feyn-
man’s triangle is also given.
1. INTRODUCTION. In [7, p. 82], just after his now famous triangle theorem,
Routh writes:
The author has not met with these expressions for the area of two triangles which
often occur. He has therefore placed them here in order that the argument in the
text may be more easily understood.
In this note, we revisit “these expressions” of Routh and show that they can be thought
of as special cases of only one expression. To our surprise, this unifying expression
relating the areas of two triangles seems to be missing from the literature. Thus, para-
phrasing Routh, we deemed it useful to place it here. We hope that the reader will find
our generalization of Routh’s triangle theorem not only natural but also intrinsically
beautiful in its symmetry.
Let ABC be a triangle determined by three non-collinear points A, B, C in the Eu-
clidean plane. A line joining a vertex to a point on the line containing the opposite side
is called a cevian. Given a point D on the line BC, the line AD is a cevian through
the vertex A. If D 6 = C, then this cevian is uniquely determined by x ∈ R \ {−1} such
−→ −→ −
→ −→
that BD = x DC. Conversely, for x ∈ R \ {−1}, the equality BD = x DC determines
uniquely a point D 6 = C on the line BC. In this note, for such a point D, we write
D = A x . It follows directly from the definition that A0 = B. We adopt the notation
A∞ = C. The line through A, which is parallel to BC, will also be considered as a
cevian through the vertex A. We denote it by AA−1 , thinking of A−1 as the “point at
infinity” on the line BC. In this way, we establish a bijection between the points of the
line BC and the set R ∪ {∞}. This bijection associates positive numbers to the points
between B and C; it associates numbers in (−1, 0) to the points between the point at
infinity and B; and it associates numbers less than −1 to the points between the point
at infinity and C. We use an analogous notation for the cevians through vertices B
−−→ −−→
and C. We will denote by B y the unique point on the line CA such that CB y = y B y A,
−→ −−→
and by C z the unique point on the line AB such that ACz = z C z B. By definition,
B0 = C, B∞ = A, C0 = A, and C∞ = B.
2. TWO EXPRESSIONS OF ROUTH. The first of the two expressions to which

Routh refers in the displayed quote gives the ratio between the area 10 of the triangle
A x B y C z , and the area 1 of the triangle ABC (see [7, p. 82]):
10 x yz + 1
= . (1)
1 (1 + x)(1 + y)(1 + z)
MSC: Primary 51M04

The natural domain of (1) is (R \ {−1})3 . In fact, it is an exercise in multivariable
limits that (1) can be continuously extended to ((R ∪ {∞}) \ {−1})3 . In this way, (1)
holds whenever A x , B y , C z , with x, y, z ∈ R ∪ {∞}, are “finite points”.
In (1) and all subsequent formulas involving areas, we consider signed areas of
triangles. A positive area corresponds to a positively-oriented triangle, that is, a tri-
angle in which the increasing alphabetic order of vertices proceeds counterclockwise.
A negative area corresponds to a triangle of opposite orientation.
If x = y = z = 1, then the triangle A x B y C z is commonly known as the medial
triangle of ABC. Analogous with this terminology, we will refer to the general triangle
A x B y C z as a cevial triangle of ABC. It follows from (1) that the points A x , B y , C z are
collinear if and only if x yz = −1. This is the well-known theorem of Menelaus; see
[2, p. 220].
The second of the two expressions to which Routh refers in the quote is another
equality, nowadays known as Routh’s theorem. Whenever the cevians AAx and BB y
intersect at a single point, we denote by P their intersection point. Similarly, we denote
by Q the intersection point of BB y and CCz , and by R the intersection point of CCz
and AAx . Then, Routh’s theorem gives the ratio between the area 100 of the triangle
PQR and the area 1 of the triangle ABC; see again [7, p. 82]:
100 (x yz − 1)2
= . (2)
1 (1 + x + x y)(1 + y + yz)(1 + z + zx)
The natural domain of (2) is R3 \ S, where S is the set of all triples (x, y, z) ∈ R3 such
that (1 + x + x y)(1 + y + yz)(1 + z + zx) = 0. As before, (2) can be continuously
extended to (R ∪ {∞})3 \ S, where S is the closure of S in (R ∪ {∞})3 . To understand
the set S, notice that the only solutions of 1 + x + x y = 0 involving ∞ are (x, y) =
(∞, −1) and (x, y) = (0, ∞). This extended domain of (2) coincides with the set
of triples (x, y, z) ∈ (R ∪ {∞})3 , for which each of the pairs of cevians (AAx , BB y ),
(BB y , CCz ), and (CCz , AAx ) intersects at exactly one point. More specifically, with
x, y ∈ R ∪ {∞}, the cevians AAx and BB y are parallel if and only if 1 + x + x y = 0
and (x, y) 6= (0, ∞); this follows from the calculations preceding (6) below. Note also
that the cevians AA0 and BB∞ coincide.
The triangle PQR will be called a Routh’s triangle of ABC. Notice that the points
P, Q, and R are collinear if and only if they coincide. Therefore, (2) implies that
the lines AAx , BB y , and CCz are concurrent if and only if x yz = 1. This statement is
known as Ceva’s theorem [2, p. 220].
3. A UNIFICATION. An extension of the construction by Nakamura and Oguiso [5,

§3] allows for formulas (1) and (2) to be unified. For u, v, w, x, y, z ∈ R ∪ {∞}, con-
sider the following six cevians, two from each vertex: AAu , AAx , BBv , BB y , CCw , CCz .
Assuming that each of the pairs of cevians (AAx , BBv ), (BB y , CCw ), and (CCz , AAu )
intersects at exactly one point, we define the generalized Routh’s triangle PQR of the
given triangle ABC (see Figure 1), by setting
{P} = AAx ∩ BBv , {Q} = BB y ∩ CCw , {R} = CCz ∩ AAu . (3)
Note that the discussion in the paragraph following (2) implies that the points P, Q,
and R in (3) are well defined if and only if x, y, z, u, v, w ∈ R ∪ {∞} satisfy
(1 + x + xv)(1 + y + yw)(1 + z + zu) 6 = 0. (4)
Choosing u = x, v = y, and w = z, the triangle PQR defined in (3) becomes a Routh’s
triangle; see Figure 2. On the other hand, choosing u = v = w = 0, we have AA0 =

C
Bv Ax
P
By
Q
Au
R
A Cw Cz B
Figure 1. A generalized Routh’s triangle
AB, BB0 = BC, CC0 = CA, and (3) yields P = A x , Q = B y , R = C z . In this case, the
triangle PQR is the cevial triangle A x B y C z ; see Figure 3.
C C
Bv
Au Ax
Ax P
R
By By Q
Bv
P
Q
R Au
A B A B
Cz Cw Cw Cz
Figure 2. Almost Routh’s PQR Figure 3. Almost a cevial PQR
Theorem. With the points P, Q, R defined in (3), the ratio between the area 11 of the
triangle PQR and the area 1 of the triangle ABC is given by
11 1 − x yw − xvz − uyz + x yz + x yzuvw

= . (5)
1 (1 + x + xv)(1 + y + yw)(1 + z + zu)
The natural domain of (5) is the set of all (x, y, z, u, v, w) ∈ R6 that satisfy (4).
However, it is again an exercise in multivariable limits, this time with six variables,
to check that formula (5) extends by continuity to all (x, y, z, u, v, w) ∈ (R ∪ {∞})6
that satisfy (4). Beautifully, with u = x, v = y, w = z, (5) simplifies to (2), and with
u = v = w = 0, it simplifies to (1). Also, as x, y, z → ∞, formula (5) becomes (1),
with x in (1) substituted by u, y by v, and z by w.
It is quite possible that any of the many proofs of Routh’s theorem (see, for ex-
ample, [2, Section 13.7], [4], [6], to mention a few) can be modified to prove this
generalization. We will prove it by using two standard undergraduate tools: linear
algebra and analytic geometry. First, we observe that the ratio of the areas remains

unchanged under affine transformations of R2 . Therefore, instead of an arbitrary
triangle ABC, we can consider the triangle with the vertices
A = (0, 0), B = (1, 0), and C = (0, 1).
Next, we find the coordinates of the point P for these special vertices A, B, C.
Let ξ denote its first coordinate and calculate A x = (1/(1 + x), x/(1 + x)) and Bv =
(0, 1/(1 + v)). Now, intersecting the lines AAx and BBv , we get ξ x = (1 − ξ )/(1 + v).
Solving for ξ gives the first coordinate of P, while the second coordinate is ξ x. Sim-
ilarly, we find Q and R; we have
1 x

P= , ,
1 + x + xv 1 + x + xv
yw 1

Q= , ,
1 + y + yw 1 + y + yw
and
z zu

R= , . (6)
1 + z + zu 1 + z + zu
Let 2 stand for the left-hand side of (4). Since we assume that each pair of cevians
(AAx , BBv ), (BB y , CCw ) and (CCz , AAu ) intersects at exactly one point, we have
2 6= 0. As the area of the triangle ABC is 1/2, the ratio in the theorem is given by the
determinant
1 yw z

1 yw z

1+x+xv 1+y+yw 1+z+zu
x 1 zu
1
x 1 zu

=
1+x+xv 1+y+yw 1+z+zu 2

1 1 1 1 + x + xv 1 + y + yw 1 + z + zu

1 yw z 1 yw z 1 yw z

1 1 1
= x 1 zu + x 1 zu + x 1 zu

2 2 2
1 1 1 x y z xv yw zu
1
= (1 + ywzu + zx − z − zu − x yw)
2
1
+ (z + ywzux + zx y − zx − zuy − zx yw)
2
1
+ (zu + ywzuxv + zx yw − zxv − zuyw − zux yw)
2
1
= (1 − x yw − xvz − uyz + x yz + uvwx yz).
2
This proves (5). As a consequence of our theorem, we also obtain the following
unification of the theorems of Ceva and Menelaus.
Corollary. The points P, Q, and R defined in (3) are collinear if and only if
1 − x yw − xvz − uyz + x yz + x yzuvw = 0.

For an animation that ties together these two classical theorems of Euclidean geom-
etry, and which is inspired by this corollary, see [3].
4. A CONNECTION TO FEYNMAN’S TRIANGLE. The Routh’s triangle with

coefficients x = y = z = 2 attracted more attention than any other; see, for example,
[8, p. 9] and [1]. The area of this special Routh’s triangle equals 1/7 of the area of
its host triangle. The proof of this fact in [8, p. 9] is a tiling type argument almost
without words, while [1] gives three different proofs using standard tools of Euclidean
geometry. According to [1], this result kept Richard Feynman busy during a dinner.
This story gave the triangle the name Feynman’s triangle; another name for it is the
“one-seventh area triangle”.
Further on, we assume that the host triangle has area 1. Feynman’s triangle is unique
in the sense that no other Routh’s triangle with equal integer coefficients x, y, and z
has an area that is a reciprocal of a nonzero digit. For example, if x = y = z = 4, the
area is 3/7. For all other equal integers as coefficients, the areas are more complicated
fractions. In fact, even if we allow unequal positive digit coefficients x, y, and z, we
can only get the area 1/2 for x = 7, y = 6, z = 3, and 1/4 for x = 7, y = 4, z = 1.
No other rational number in (0, 1) \ {1/7, 1/4, 3/7, 1/2} with digits as its numerator
and denominator is attainable as an area of Routh’s triangle with positive digits as
coefficients.
Does this change if we consider generalized Routh’s triangles with u = v = w,
x = y = z, and x 6 = u being digits? The answer is yes. We can get the area 1/9 with
u = 1, x = 4 (see Figure 4), the area 1/4 with u = 4, x = 1 (see Figure 5), and the
area 4/9 with u = 7, x = 1 (see Figure 6.) However, none of these triangles are very
interesting, since they are simply scaled images of the host triangle with the same
center of gravity.
C C C
Au
Ax Au
R R
Bu P Au Bx Ax Bx Ax
R
Bx Bu P Q
Bu
Q P Q
A B A B A B
Cu Cx Cx Cu Cx Cu
Figure 4. Figure 5. Figure 6.
C C
Bx Au Ax
Bu
P
Q
Au
R
Bu P Bx R
Ax Q
A B A B
Cx Cu Cu Cx
Figure 7. 1/4 area; u = 4, x = 1/9 Figure 8. 1/7 area; u = 1/2, x = 4
Nevertheless, if we allow one of the coefficients u or x to be a reciprocal of a

nonzero digit, then, besides the previous three configurations, we can also get the area

1/4 with u = 9, x = 1/4 or u = 4, x = 1/9, and our old acquaintance, the area 1/7,
with u = 4, x = 1/2, or u = 2, x = 1/4, or u = 1/2, x = 4; see Figures 7 and 8.
We conclude this note with an invitation for the reader to find a purely geometric
argument along the lines of [8, p. 9] for the claims indicated in the last two figures.
REFERENCES
1. R. J. Cook, G. V. Wood, Feynman’s triangle, Mathematical Gazette 88 (2004) 299–302.

2. H. S. M. Coxeter, Introduction to Geometry, second edition. Wiley, New York, 1969.
3. B. Ćurgus, The theorems of Ceva and Menelaus: an animation, available at http://faculty.wwu.edu/
curgus/Papers/Monthly2012.html.
4. J. S. Kline, D. Velleman, Yet another proof of Routh’s theorem, Crux Mathematicorum 21 (1995) 37–40.
5. H. Nakamura, K. Oguiso, Elementary moduli space of triangles and iterative processes, The University of
Tokyo Journal of Mathematical Sciences 10 (2004) 209–224.
6. I. Niven, A new proof of Routh’s theorem, Math. Mag. 49 (1976) 25–27, available at http://dx.doi.
org/10.2307/2689876.
7. E. J. Routh, A Treatise on Analytical Statics with Numerous Examples, Vol. 1, second edition. Cambridge
University Press, London, 1909, available at http://www.archive.org/details/texts.
8. H. Steinhaus, Mathematical Snapshots, third English edition. Translated from the Polish. With a preface
by Morris Kline. Dover Publications, Mineola, NY, 1999.
Department of Mathematics, Western Washington University, Bellingham, WA 98225, USA,

Arpad.Benyi@wwu.edu
Department of Mathematics, Western Washington University, Bellingham, WA 98225, USA,

Branko.Curgus@wwu.edu
A Maximum Principle for

High-Order Derivatives
David Pan
Abstract. We prove a maximum principle for high-order derivatives under initial conditions.
1. INTRODUCTION We begin with a well-known definition.
Definition 1.1. The maximum principle holds for f (x) on [a, b] if f (x) ≤ max{ f (a),
f (b)} on [a, b].
It is well known that f 0 (x) ≥ 0 implies the maximum principle because of nonde-
creasingness, and that f 00 (x) ≥ 0 implies the maximum principle because of noncon-
cavity. It is also well known that f (n) (x) ≥ 0 where n ≥ 3 does not necessarily imply
the maximum principle. For example, consider f (x) = ±x 2 and its third derivative.
We present conditions under which f (n) (x) ≥ 0 where n ≥ 3 does imply the maxi-
mum principle. Let I = [a, b] be a closed interval of the real line, and let C n (I ) be the
MSC: Primary 26A06

set of real-valued functions on I that are n-times continuously differentiable. Recall
that
n
f (i) (a)
(x − a)i
X
Pn (x) :=
i=0
i!
is the nth-degree Taylor polynomial of f (x) at a, where f (x) ∈ C n (I ) and a ∈ I .
Theorem 1.1. Let f (x) ∈ C n ([a, b]) for some n ≥ 2. If f (n) (x) ≥ 0 on [a, b], then
f (x) ≤ vn (x) on [a, b], where
f (b) − Pn−2 (b)

vn (x) := Pn−2 (x) + (x − a)n−1 .
(b − a)n−1
The following theorem is a corollary of Theorem 1.1. It shows that the maximum
principle holds under initial conditions.
Theorem 1.2. Let f (x) ∈ C n ([a, b]) for some n ≥ 3. If

(i) f (n) (x) ≥ 0 on [a, b], and
(ii) f (b) ≥ Pi (b) for i = 1, . . . , n − 2,
then f (x) ≤ max{ f (a), f (b)} on [a, b] (i.e., the maximum principle holds for f (x)
on [a, b]).
Example 1.1. Below is a graph of f (x) = (x − 2)(x − 1)(x + 1)(x + 2) = x 4 −

5x 2 + 4.
40
30
20
10
−3 −2 −1 1 2 3
It is false that f 0 (x), f 00 (x) ≥ 0 on [0, 3], but it is true that f 000 (x) ≥ 0 on [0, 3].
Also, f (3) = 40 ≥ P1 (3) = 4. Therefore, by Theorem 1.2, the maximum principle
holds for f (x) on [0, 3].
Proof of Theorem 1.1. Suppose that f (x) ∈ C n ([a, b]) for some n ≥ 2 and f (n) (x) ≥
0 on [a, b]. By Taylor’s theorem,
1
f (x) − Pn−2 (x) 1
Z
g(x) := = (1 − t)n−2 f (n−1) (a + t (x − a)) dt.
(x − a) n−1 (n − 2)! 0

It follows that
1
1
Z
g (x) =
0
t (1 − t)n−2 f (n) (a + t (x − a)) dt.
(n − 2)! 0
Since f (n) (x) ≥ 0 on [a, b], g 0 (x) ≥ 0 on [a, b], so g(x) ≤ g(b) on [a, b], which
implies
f (b) − Pn−2 (b)

f (x) ≤ Pn−2 (x) + (x − a)n−1 = vn (x)
(b − a)n−1
on [a, b].
Proof of Theorem 1.2. We prove Theorem 1.2 by mathematical induction.

First, we prove the initial case n = 3. Suppose that for some f (x) ∈ C 3 ([a, b]), we
have f 000 (x) ≥ 0 on [a, b] and f (b) ≥ P1 (b). By Theorem 1.1, we have
f (b) − P1 (b)
f (x) ≤ v3 (x) = P1 (x) + (x − a)2
(b − a)2
on [a, b]. Because P100 (x) = 0 and f (b) − P1 (b) ≥ 0, we have v300 (x) ≥ 0, which, by
convexity, implies that v3 (x) ≤ max{v(a), v(b)} on [a, b]. It is obvious that v3 (a) =
f (a) and v3 (b) = f (b), so f (x) ≤ v3 (x) ≤ max{ f (a), f (b)} on [a, b]. Therefore, the
theorem holds in the initial case n = 3.
Now, we assume that the theorem holds in the case n = k where k ≥ 3, and we
want to show that the theorem holds in the case n = k + 1. To do this, suppose that
for some f (x) ∈ C k+1 ([a, b]), we have f (k+1) (x) ≥ 0 on [a, b] and f (b) ≥ Pi (b) for
i = 1, . . . , k − 1. By Theorem 1.1, we have
f (b) − Pk−1 (b)

f (x) ≤ vk+1 (x) = Pk−1 (x) + (x − a)k
(b − a)k
(k) (k)
on [a, b]. Because Pk−1 (x) = 0 and f (b) − Pk−1 (b) ≥ 0, we have vk+1 (x) ≥ 0. It is
(i)
obvious that vk+1 (a) = f (a) for i = 0, . . . , k − 1 and vk+1 (b) = f (b), so
(i)
j j (i)
f (i) (a) vk+1 (a)
(b − a)i = (b − a)i
X X
vk+1 (b) = f (b) ≥ P j (b) =
i=0
i! i=0
i!
for j = 1, . . . , k − 2. By the induction assumption,
vk+1 (x) ≤ max{vk+1 (a), vk+1 (b)} = max{ f (a), f (b)}
on [a, b]. Therefore, f (x) ≤ vk+1 (x) ≤ max{ f (a), f (b)} on [a, b]. This concludes
the mathematical induction, and the proof is complete.
Canterbury High School, 3210 Smith Road, Fort Wayne, IN 46804

david.l.pan@gmail.com

Nothing New about Equiangular Polygons
Gerhard J. Woeginger
Abstract. We survey results on equiangular n-vertex polygons with edge lengths in arithmetic
progression. Such a polygon exists if and only if n has at least two distinct prime factors.
We tell a tale of mathematics, old and aged, and rooted deeply in the last century.
Indeed, let us go back, all the way to the year 1983, when the proud nation of Sweden
proposed the following problem for the 24th International Mathematical Olympiad in
Paris; see page 167 in the problem collection [3].
Let n be a positive integer having at least two different prime factors. Show that
there exists a permutation a1 , a2 , . . . , an of the integers 1, 2, . . . , n such that
n
X 2πak
k · cos = 0.
k=1
n
Alas, the Swedish proposal was only one out of 25 strong candidates on the problem
short-list, and the international jury decided not to select it for the competition. The
problem fell into oblivion until 1986, when Murray Klamkin published his collection
[4] of IMO problems together with 40 supplementary problems. The first one of the
40 supplementary problems was the Swedish proposal from 1983.
Klamkin’s solution [4, p. 61]) writes n = pq as the product of two relative prime
integers p, q > 1, and considers the unit vectors v k = e2kπi/n for k = 1, . . . , n. If o
denotes the zero vector, then Klamkin observes that
vr + vr + p + vr +2 p + · · · + vr +(q−1) p = o (1)
holds for r = 1, 2, . . . , p, as the vectors in each of these sums form a (closed!) regular
q-gon. Analogously, he observes that
v s + v s+q + v s+2q + · · · + v s+( p−1)q = o (2)
for s = 1, 2, . . . , q. Klamkin multiplies the r th equation (1) by r , multiplies the sth

equation (2) by (s − 1) p, and finally sums these p + q equations. In the resulting sum,
the coefficients of the unit vectors v 1 , . . . , v n form a permutation π of 1, . . . , n, as
every integer k = 1, . . . , pq has a unique representation of the form k = r + (s − 1) p
with 1 ≤ r ≤ p and 1 ≤ s ≤ q. As the sum is the zero-vector, its x-coordinate and
y-coordinate are zero, so that permutation π simultaneously satisfies
n n
X 2kπ X 2kπ
π(k) · cos =0 and π(k) · sin = 0. (3)
k=1
n k=1
n
This does not only settle the Swedish problem, but also yields the following strength-
ening.
MSC: Primary 12D10, Secondary 52B12

Theorem 1. [4] If the positive integer n has at least two different prime factors,
then there exists an equiangular n-gon whose edge lengths form a permutation of
1, 2, 3, . . . , n.
The plot thickened again in fall 2005, after two decades of endless silence, when
Brendan McKay posted the following problem to the SEQFAN mailing list.
Suppose you have n objects with weights 1, 2, 3, . . . , n. How many ways are
there to place these objects evenly spaced around the circumference of a disk so
that the disk will exactly balance on the center point?
In algebraic terms, how many permutations π ∈ Sn are there such that the
polynomial π(1) + π(2)x + · · · + π(n)x n−1 has e2πi/n as a zero?
McKay [5] said that his question was based on a recent puzzle in the New Scientist,
and he also wrote that he didn’t even know which values n have a solution. Within a
couple of days, the positive result of Murray Klamkin was rediscovered, and William
Edwin Clark from the University of South Florida showed that all remaining cases are
unsolvable.
Clark’s argument [1] is as elegant as it is simple: Let n = p k be a prime power. If
a polynomial P(x) = π(1) + π(2)x + · · · + π(n)x n−1 has the nth primitive root of
unity e2πi/n as a zero, then it is divisible by the nth cyclotomic polynomial 8n (x) =
P p−1 i pk−1
i=0 x . In other words. P(x) = 8n (x)Q(x) for some polynomial Q ∈ Z[x] of
degree p k−1 − 1. But then the coefficients of P will just be the p k−1 coefficients of Q
repeated p times, and can never form a permutation of the integers 1, . . . , p k .
Theorem 2. [1] If n is a prime power, then there is no equiangular n-gon whose edge
lengths form a permutation of 1, . . . , n.
Although Theorems 1 and 2 provide the full picture on such equiangular n-gons,
there are three further chapters to this fascinating tale. The first chapter added se-
quence A118887 to the On-Line Encyclopedia of Integer Sequences [6], the sequence
described in McKay’s posting. The second chapter was written by Robert Dawson [2],
who rediscovered the above results for even n and for n = 3, 5, 7. The third and last
chapter is the current article, summarizing the history of this problem. And that’s the
end of this story.
REFERENCES
1. E. Clark, post to SEQFAN mailing list (13 Sep 2005).

2. R. Dawson, Arithmetic polygons, Amer. Math. Monthly 119 (2012) 695–698, available at http://dx.
doi.org/10.4169/amer.math.monthly.119.08.695.
3. D. Djukić, V. Janković, I. Matić, N. Petrović, The IMO Compendium: A Collection of Problems Suggested
for the International Mathematical Olympiads: 1959–2004, Springer, New York, 2006.
4. M. S. Klamkin, International Mathematical Olympiads 1978–1985 and Forty Supplementary Problems,
Mathematical Association of America, Washington, DC, 1986.
5. B. McKay, post to SEQFAN mailing list (12 Sep 2005).
6. The On-Line Encyclopedia of Integer Sequences, Sequence A118887, available at http://oeis.org/
A118888.
Department of Mathematics, TU Eindhoven, The Netherlands

gwoegi@win.tue.nl

Quotients of Gaussian Primes
Stephan Ramon Garcia
Abstract. It has been observed many times, both in the M ONTHLY and elsewhere, that the
set of all quotients of prime numbers is dense in the positive real numbers. In this short note
we answer the related question: “Is the set of all quotients of Gaussian primes dense in the
complex plane?”
Quotient sets {s/t : s, t ∈ S} corresponding to subsets S of the natural numbers have

been intensely studied in the M ONTHLY over the years [1, 4, 7, 8, 10, 13]. Moreover, it
has been observed many times in the M ONTHLY and elsewhere that the set of all quo-
tients of prime numbers is dense in the positive reals (e.g., [2, Ex. 218], [3, Ex. 4.19],
[4, Cor. 5], [8, Thm. 4],[11, Ex. 7, p. 107], [12, Thm. 4], [13, Cor. 2]).
In this short note we answer the related question: “Is the set of all quotients of
Gaussian primes dense in the complex plane?” The author became convinced of the
nontriviality of this problem after consulting several respected number theorists who
each admitted not seeing a simple solution.
In the following, we refer to the traditional primes 2, 3, 5, 7, . . . as rational primes,
remarking that a rational prime p is a Gaussian prime (i.e., a prime in the ring Z[i] :=
{a + bi : a, b ∈ Z} of Gaussian integers), if and only if p ≡ 3 (mod 4). In general, a
nonzero Gaussian integer is prime if and only if it is of the form ± p or ± pi where p
is a rational prime congruent to 3 (mod 4) or if it is of the form a + bi where a 2 + b2
is a rational prime (see Figure 1). We refer the reader to [5] for complete details.
Figure 1. Gaussian primes a + bi satisfying |a|, |b| ≤ 50 and |a|, |b| ≤ 100, respectively

Theorem. The set of quotients of Gaussian primes is dense in the complex plane.
Proof. It suffices to show that each region of the form
{z ∈ C : α < arg z < β, r < |z| < R}, (1)
contains a quotient of Gaussian primes.

We first claim that if 0 < a < b, then for sufficiently large real x, the open interval
(xa, xb) contains a rational prime congruent to 3 (mod 4). Let π3 (x) denote the num-
ber of rational primes congruent to 3 (mod 4) which are ≤ x. By the prime number
theorem for arithmetic progressions [5, Thm. 4.7.4],
π3 (x) 1
lim = ,
x→∞ x/ log x 2
whence

π3 (xa)
lim [π3 (xb) − π3 (xa)] = lim π3 (xb) 1 −
x→∞ x→∞ π3 (xb)
xa log xb

= lim π3 (xb) 1 −
x→∞ xb log xa
a
= 1− lim π3 (xb)
b x→∞
= ∞,
which establishes the claim.

Next observe that the sector α < arg z < β contains Gaussian primes of arbitrarily
large magnitude. This follows from an old result of I. Kubilyus (illustrated in Figure
2) which states that the number of Gaussian primes γ satisfying 0 ≤ α ≤ arg γ ≤ β ≤
2π and |γ |2 ≤ u is
u
2 dx
Z p
(β − α) + O u exp(−b log u) (2)
π 2 log x
where b > 0 is an absolute constant [9] (see also [6, Thms. 2,3]).
ρ N K ρ N K
100 50 53 1,000 0 5
500 946 940 5,000 0 100
1,000 3,327 3,346 10,000 369 367
5,000 66,712 66,651 50,000 7,823 7,732
10,000 245,085 245,200 100,000 28,964 28,971
25,000 1,384,746 1,385,602 250,000 167,197 167,099
50,000 5,168,740 5,167,941 500,000 632,781 631,552
2π 2π
(a) π
24
≤ arg z ≤ 47
(b) π
31415
≤ arg z ≤ 31415
Figure 2. The number N of Gaussian primes in the specified sector with |z| < ρ, along with the corresponding
estimate K (rounded to the nearest whole number) provided by (2)

Putting this all together, we conclude that there exists a Gaussian prime γ in the
sector α < arg z < β whose magnitude is large enough to ensure that

|γ | |γ |
π3 − π3 ≥ 2.
r R
This yields a rational prime q ≡ 3 (mod 4) such that
|γ | |γ |
<q< .
R r
Since q is real and positive, it follows that r < | γq | < R and α < arg γq < β so that
γ /q is a quotient of Gaussian primes which belongs to the desired region (1).
ACKNOWLEDGMENTS. We thank the anonymous referees for several helpful suggestions. This work was
partially supported by National Science Foundation Grant DMS-1001614.
REFERENCES
1. J. Bukor, J. T. Tóth, On accumulation points of ratio sets of positive integers, Amer. Math. Monthly 103
(1996) 502–504, available at http://dx.doi.org/10.2307/2974720.
2. J.-M. DeKonick, A. Mercier, 1001 Problems in Classical Number Theory, American Mathematical So-
ciety, Providence, RI, 2007.
3. B. Fine, G. Rosenberger, Number Theory: An Introduction via the Distribution of Primes, Birkhäuser,
Boston, 2007.
4. S. R. Garcia, V. Selhorst-Jones, D. E. Poore, N. Simon, Quotient sets and diophantine equations, Amer.
Math. Monthly 118 (2011) 704–711.
5. G. H. Hardy, E. M. Wright, An Introduction to the Theory of Numbers, sixth edition, Oxford University
Press, Oxford, 2008. Revised by D. R. Heath-Brown and J. H. Silverman, with a foreword by Andrew
Wiles.
6. G. Harman, P. Lewis, Gaussian primes in narrow sectors, Mathematika 48 (2001) 119–135, available at
http://dx.doi.org/10.1112/S0025579300014388.
7. S. Hedman, D. Rose, Light subets of N with dense quotient sets, Amer. Math. Monthly 116 (2009) 635–
641, available at http://dx.doi.org/10.4169/193009709X458618.
8. D. Hobby, D. M. Silberger, Quotients of primes, Amer. Math. Monthly 100 (1993) 50–52, available at
http://dx.doi.org/10.2307/2324814.
9. I. Kubilyus, The distribution of Gaussian primes in sectors and contours, Leningrad. Gos. Univ. Uč. Zap.
Ser. Nauk 137(19) (1950) 40–52.
10. A. Nowicki, Editor’s endnotes, Amer. Math. Monthly 117 (2010) 755–756.
11. P. Pollack, Not Always Buried Deep: A Second Course in Elementary Number Theory, American Math-
ematical Society, Providence, RI, 2009.
12. P. Ribenboim, The Book of Prime Number Records, second edition, Springer-Verlag, New York, 1989.
13. P. Starni, Answers to two questions concerning quotients of primes, Amer. Math. Monthly 102 (1995)
347–349, available at http://dx.doi.org/10.2307/2974957.
Department of mathematics, Pomona College, Claremont, CA 91711

Stephen.Garcia@pomona.edu

PROBLEMS AND SOLUTIONS
Edited by Gerald A. Edgar, Doug Hensley, Douglas B. West
with the collaboration of Itshak Borosh, Paul Bracken, Ezra A. Brown, Randall
Dougherty, Tamás Erdélyi, Zachary Franco, Christian Friesen, Ira M. Gessel, László
Lipták, Frederick W. Luttmann, Vania Mascioni, Frank B. Miles, Richard Pfiefer,
Dave Renfro, Cecil C. Rousseau, Leonard Smiley, Kenneth Stolarsky, Richard Stong,
Walter Stromquist, Daniel Ullman, Charles Vanden Eynden, Sam Vandervelde, and
Fuzhen Zhang.
Proposed problems and solutions should be sent in duplicate to the MONTHLY

problems address on the back of the title page. Proposed problems should never
be under submission concurrently to more than one journal. Submitted solutions
should arrive before March 31, 2014. Additional information, such as general-
izations and references, is welcome. The problem number and the solver’s name
and address should appear on each solution. An asterisk (*) after the number of
a problem or a part of a problem indicates that no solution is currently available.
PROBLEMS
11733. Proposed by Donald Knuth, Stanford University,

Stanford, CA.
Let V = {0, 1, 2, 3, 4}2 . Say that nonnegative integers
a and b are adjacent when their base-5 expansions
· · · a2 a1 a0 and · · · b2 b1 b0 satisfy the condition that if i >
j ≥ 0 and (ai , a j ) 6 = (bi , b j ), then (ai , a j ) and (bi , b j ) are
consecutive in the path through V shown at right (horizon-
tal coordinate listed first). Thus, for example, 0 is adjacent
to 1. Similarly, 48 (expansion 1435 ) is adjacent to 47 (ex-
pansion 1425 ) and 73 (expansion 2435 ).
(a) Prove that every positive integer is adjacent to exactly two nonnegative integers.
(b) Prove that with this definition of adjacency, the nonnegative integers form a path
hx0 , x1 , x2 , . . . i starting with x0 = 0.
(c) Explain how to compute efficiently from n the number xn that comes n steps after
0, and determine x1,000,000 .
11734. Proposed by Vahagn Aslanyan, Yerevan State University, Yerevan, Armenia.
Find all lists (a, k, m, n) of positive integers such that
a m+n + a n − a m − 1 = 15k .
11735. Proposed by Cosmin Pohoata, Princeton University, Princeton, NJ. Let P be a

point inside triangle ABC. Let d A , d B , and dC be the distances from m to A, B, and C,
respectively. Let r A , r B , and rC be the radii of the circumcircles of P BC, PC A, and
P AB, respectively. Prove that
1 1 1 1 1 1
+ + ≥ + + .
dA dB dC rA rB rC

11736. Proposed by Mircea Merca, University of Craiova, Craiova, Romania. For
n ≥ 1, let f be the symmetric polynomial in variables x1 , . . . , xn given by
n−1
X
f (x1 , . . . , xn ) = (−1)k+1 ek (x1 + x12 , x2 + x22 , . . . , xn + xn2 ),
k=0
where ek is the kth elementary polynomial in n variables. (For example, when n = 6,

e2 has 15 terms, each a product of two distinct variables.) Also, let ξ be a primitive nth
root of unity. Prove that
f (1, ξ, ξ 2 , . . . , ξ n−1 ) = L n − L 0 ,
where L k is the kth Lucas number (that is, L 0 = 2, L 1 = 1, and L k = L k−1 + L k−2 for
k ≥ 2).
11737. Proposed by Nguyen Thanh Binh, Hanoi, Vietnam. Given an acute triangle
ABC, let O be its circumcenter, let M be the intersection of lines AO and BC, and
let D be the other intersection of AO with the circumcircle of ABC. Let E be that
point on AD such that M is the midpoint of E D. Let F be the point at which the
perpendicular to AD at M meets AC. Prove that E F is perpendicular to AB.
11738. Proposed by Stefano Siboni, University of Trento, Trento, Italy. Three point par-
ticles are constrained to move without friction along a unit circle. Three ideal massless
springs of stiffness k1 , k2 , and k3 connect the particles pairwise. Show that an equilib-
rium in which the particles occupy three distinct positions exists if and only if 1/k1 ,
1/k2 , and 1/k3 can be the lengths of the sides of a triangle. Show also that if this
happens, the equilibrium length L of the spring with stiffness k1 is given by
s
1 2

p 1 1
L = k2 k3 + − 2.
k2 k3 k1
11739. Proposed by Fred Adams, Anthony

Bloch,
and Jeffrey Lagarias, University of
1 x
Michigan, Ann Arbor, MI. Let B(x) = . Consider the infinite matrix product
x 1
Y
M(t) = B(2−t )B(3−t )B(5−t ) · · · = B( p −t ),
p
where the product runs over all primes, taken in increasing order. Evaluate M(2).
SOLUTIONS
A Singular Matrix
11593 [2011, 747]. Proposed by Peter McGrath, Brown University, Providence, RI.
For positive integers k and n, we let T (n, k) be the n × n matrix with (i, j)-entry
((i − 1)n + j)k . Prove that for n > k + 1, det(T (n, k)) = 0.
Solution by Moubinool Omarjee, Paris, France. We prove a more general result. Let
K be a commutative field with characteristic 0. Let Q be a polynomial of degree k
with k > 0, and let a1 , . . . , an be distinct scalars in K . For n > k + 1, we show that
the determinant of the matrix M whose (i, j)-entry is Q(ai + j − 1) is 0.
November 2013] PROBLEMS AND SOLUTIONS 855

Let W (x) be the determinant of the matrix that agrees with M in all columns after
the first, but has entry Q(ai + x) instead of Q(ai ) in row i of column 1. Due to repeti-
tion of columns, W (1) = · · · = W (n − 1) = 0. Also, W is a polynomial of degree k.
With n > k + 1, necessarily W is identically zero. Hence, det M = W (0) = 0.
Also solved by U. Abel (Germany), G. Apostolopoulos (Greece), D. Beckwith, N. Caro (Brazil), R. Chap-
man (U. K.), D. Constales (Belgium), P. P. Dályay (Hungary), D. Fleischman, S. M. Gagola Jr., O. Geupel
(Germany), M. Goldenberg & M. Kaplan, J.-P. Grivaux (France), E. A. Herman, Y. J. Ionin, D. A. Jackson,
W. P. Johnson, B. Karaivanov, M. E. Kidwell & M. D. Meyerson, J. C. Kieffer, M. J. Knight, O. Kouba
(Syria), Y.-J. Kuo, E. Lalov, L. Lipták, O. P. Lossers (Netherlands), Á. Plaza (Spain), C. R. Pranesachar (In-
dia), R. E. Prather, F. G. Schmitt, U. Schneider (Switzerland), N. C. Singer, J. H. Steelman, R. Stong, J. Stuart,
D. B. Tyler, J. Vinuesa (Spain), Z. Vörös (Hungary), H. Widmer (Switzerland), M. Wildon (U. K.), A. L. Yandl,
Z. Zhang, BSI Problems Group (Germany), GCHQ Problem Solving Group (U. K.), University of Louisiana
at Lafayette Math Club, Missouri State University Problem Solving Group, Texas State University Problem
Solvers, and the proposer.
An Integral Product
11594 [2011, 747]. Proposed by Harm Derksen and Jeffrey Lagarias, University of
Michigan, Ann Arbor, MI. Let
 
n k−1
Y Y j ,
Gn = 
k=1 j=1
k
and let G n = 1/G n .

(a) Show that if n is an integer greater than 1, then G n is an integer.
(b) Show that for each prime p, there are infinitely many n greater than 1 such that p
does not divide G n .
Solution by John H. Smith, Needham, MA.
(a) The inner product is empty when k = 1 and then equals 1, so we may start the
outer product with k = 2:
Qn Qk−1 Qn−1 Qn Qn−1 n! n−1
j=1 k k= j+1 k j=1 j! n
k=2 j=1
Y
G n = Qn Qk−1 = Qn = Qn−1 = ∈ Z.
k=2 j=1 j k=2 (k − 1)! j=1 (n − j)! j=1
j
(b) If n = p m − 1 with p prime, then nj = ( p −1)(1×2×···

m p m −2)···( p m − j)
j
. For 1 ≤ i ≤ j, the
m
power of p dividing p − i is the same as that dividing i, and p does not divide G n .
Editorial comment. It is possible to determine exactly which G n are divisible by p
by using the theorem of Kummer that the power of p dividing nj is the number of
“borrows” in subtracting j from n when they are written in base p. If n = ar pr − 1 =
i=0 ( p − 1) p , then there are no “borrows” and p does not divide G n ; otherwise there
Pr −1 i
is a borrow for some j, and p divides G n (see E. Kummer, Jour. für Math. 44, 1852,
115–116). Theorems by Lucas and N.J. Fine also can be used.
Also solved by M. Bataille (France), A. Bostan (France), P. Budney, B. S. Burdick, N. Caro (Brazil), R. Chap-
man (U. K.), W. ChengYuan (Singapore), P. P. Dályay (Hungary), A.-M. Ernvall-Hytönen (Finland), D. Fleis-
chman, J. Freeman, O. Geupel (Germany), M. Goldenberg and M. Kaplan, J.-P. Grivaux (France), D. Hender-
son, E. Hysnelaj & E. Bojaxhiu (Australia & Germany), E. J. Ionascu, Y. J. Ionin, B. Karaivanov, J. C. Kieffer,
O. Kouba (Syria), H. Kwong, J. H. Lindsey II, L. Lipták, O. P. Lossers (Netherlands), R. Martin (Germany),
Á. Plaza (Spain), P. Pongsriiam & T. Pongsriiam (U. S. A. & Thailand), C. R. Pranesachar (India), R. E. Prather,
B. Schmuland (Canada), A. Stenger, R. Stong, R. Tauraso (Italy), M. Tetiva (Romania), D. B. Tyler, Z. Vörös

(Hungary), M. Vowe (Switzerland), H. Widmer (Switzerland), Y. Zhang, Z. Zhang, BSI Problems Group (Ger-
many), GCHQ Problem Solving Group (U. K.), University of Louisiana at Lafayette Math Club, NSA Prob-
lems Group, ONU–SOLVE Problem Group, and the proposers.
Another Version of the Axiom of Choice

11599 [2011, 748]. Proposed by Fred Galvin, University of Kansas, Lawrence, KS, and
Péter Komjáth, Eötvös Loránd University, Budapest, Hungary. Prove that the follow-
ing statement is equivalent to the Axiom of Choice: For any finite family A1 , . . . , An
of sets, there is a finite set F such that |Ai ∩ F| < |A j ∩ F| whenever |Ai | < |A j |.
Here, equivalence is to be judged in the context of Zermelo–Fraenkel set theory not
assuming the Axiom of Choice, and to say that |C| < |D| is to say that there is an
injection from C to D, but none from D to C.
Solution by the proposers. First, assume the Axiom of Choice, and consider sets
A1 , . . . , An . Without loss of generality, we may assume that |A1 | ≤ · · · ≤ |An |. Let
F0 = S ∅. If A j is finite, then let F j = F j−1 ∪ A j . If A j is infinite, then let B j =
A j − {Ai : |Ai | < |A j |}. The set B j is infinite; choose a finite subset X j of B j with
|X j | > |F j−1 |, and let F j = F j−1 ∪ X j . Now Fn is finite, and |Ai ∪ Fn | < |A j ∩ Fn |
whenever |Ai | < |A j |, so we may let F = Fn .
For the converse, if the Axiom of Choice is false, then there is a set S that is not
well-orderable. Let κ = |S × ω|, so that κ + κ = κ. By Hartog’s Theorem, κ is in-
comparable with some ℵα . Construct disjoint sets X, Y, U, V with |X | = |U | = κ and
|Y | = |V | = ℵα . Let A1 = X ∪ U , A2 = Y ∪ U , A3 = Y ∪ V , and A4 = X ∪ V . Now
|A1 | = κ + κ = κ < ℵα + κ = |A2 | and |A3 | = ℵα + ℵα = ℵα , and ℵα < κ + ℵα =
|A4 |. If there is a finite set F such that |A1 ∩ F| < |A2 ∩ F| and |A3 ∩ F| < |A4 ∩ F|,
then the finite cardinals x, y, u, v defined by z = |Z ∩ F| for Z ∈ {X, Y, U, V } satisfy
the contradictory inequalities x + u < y + u and y + v < x + v.
Editorial comment. The proposers note that the solution shows that the assertion for
n = 4 is already equivalent to the Axiom of Choice. Indeed, the Axiom of Choice is
equivalent to the assertion that for any sets A1 , A2 , A3 , A4 with |A1 | < |A2 | and |A3 | <
|A4 | there is a finite set F such that |A1 ∩ F| < |A2 ∩ F| and |A3 ∩ F| < |A4 ∩ F|.
On the other hand, the statement for n = 3 is easily proved in ZF.
No other correct solution was received.
The Product of the Farey Series

11601 [2011, 846]. 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
Let F n = 1/Fn . Show that F n is an integer for only finitely many n.

Solution by Richard Stong, Center for Communications Research, San Diego, CA. It
follows from the prime number theorem that for all but finitely many n there is an
odd prime p with n3 < p ≤ 3n8 . For such a prime p, the only denominators in Fn that
are multiples of p are when k = p and k = 2 p (both with p − 1 numerators, since
ϕ( p) = ϕ(2 p) = p − 1). Thus the denominator of Fn has 2( p − 1) factors of p. The

numerators that are multiples of p are p itself for p + 1 ≤ k ≤ 2 p − 1 and 2 p + 1 ≤
k ≤ n (a total of n − p − 1 terms), plus 2 p for the odd values of k such that 2 p + 1 ≤
k ≤ 2 b(n − 1)/2c + 1 (a total of b(n + 1)/2c − p terms). For the total number of
times p divides Fp , we compute

n+1 3n − 8 p + 3
(n − p − 1) + − p − 2( p − 1) = ≥ 1 > 0.
2 2
Thus 1/Fn has a factor of p in its denominator and is not an integer.
Editorial comment. Stong remarked that from the explicit bounds of Dusart (1998),
a prime p as in the solution exists when n ≥ 9825. Checking cases reduces this to
n ≥ 142. Computing Fn for n < 142, we find that the last value of n where F n is an
integer is F 58 .
Also solved by N. Caro (Brazil), D. Fleischman, S. M. Gagola Jr., Y. J. Ionin, B. Karaivanov, O. Kouba (Syria),
O. P. Lossers (Netherlands), R. E. Prather, R. Tauraso (Italy), M. Tetiva (Romania), Ellington Management
Problem Solving Group, and the proposers.
A Fibonacci Congruence
11602 [2011, 846]. Proposed by Roberto Tauraso, Università di Roma “Tor Vergata,”
Rome, Italy. Let p be a prime. Let Fn denote the nth Fibonacci number. Show that
X Fi
≡0 (mod p).
0<i< j<k< p
i jk
(A rational number is deemed congruent to 0 mod p if, when put in reduced form, the
numerator is a multiple of p.)
Solution by Yury J. Ionin, Champaign, IL. The statement holds vacuously for p = 2
and p = 3, and it can be easily verified for p = 5 (the left side simplifies to 5/12),
so assume that p > 5. Consider the field F p of the residue classes modulo p. The
discriminant of the quadratic equation x 2 − x − 1 = 0 is 5, so the equation has distinct
roots α and β. These roots are in F p if 5 is a quadratic residue modulo p; otherwise,
they are in a quadratic extension of F p . In either case, the Fibonacci numbers are given
by Fn = (α n − β n )/(α − β), and they lie in F p . The congruence to be shown is then
equivalent to the following identity in a field K containing α, β, and all of F p :
p−1 k−1 j−1 i
X X X α − βi
= 0. (1)
k=3 j=2 i=1
i jk
Define a polynomial S in F p [x] by

p−1 k−1 j−1
X X X xi
S(x) = ,
k=3 j=2 i=1
i jk
and let T (x) = S(1 − x).P Since α + β = 1, identity (1) holds if S = T . Clearly S(0) =
0, and T (0) = S(1) = 16 1
i jk
, where the summation extends over all ordered triples
(i, j, k) of distinct nonzero elements of F p . Choose an element l ∈ F p such that l 3 6 = 1
and l 6= 0. Since Tl(0) = 16 1
= T (0), we obtain T (0) = 0. Since S and T
P
3 (li)(l j)(lk)
are polynomials of degree at most p − 3, identity (1) holds if S 0 = T 0 . We perform

computations in the field F p (x) of rational functions over F p . Using the sum of a
geometric series,
p−1 k−1 j−1 i−1 p−1 k−1
X XX x X X x j−1 − 1
S 0 (x) = =
k=3 j=2 i=1
jk k=3 j=2
jk(x − 1)
and
p−1 k−1
X X (1 − x) j−1 − 1
T 0 (x) = −S 0 (1 − x) = .
k=3 j=2
jkx
Define the polynomial U in F p [x] by

p−1 k−1
X X x j + (1 − x) j − 1
U (x) = x(x − 1) S 0 (x) − T 0 (x) = .
k=3 j=2
jk
Since S 0 and T 0 are polynomials, U (0) = 0. The degree of U is at most p − 2, so (1)

holds if U 0 = 0. We have
p−1 k−1 p−1 k
X X x j−1 − (1 − x) j−1 1 X x − (1 − x)k − 2x + 1
U (x) =
0
= .
k=3 j=2
k x(x − 1) k=3 k
Let V (x) = x(x − 1)U 0 (x). Since U 0 is a polynomial, V (0) = 0. The degree of V is
at most p − 1, so (1) holds if V 0 (a) = 0 for all a ∈ F p . Since k=1
P p−1 1 P p−1
k
= k=1 k = 0
P p−1 2
in F p , we have k=3 k = −3. Thus
p−1
X x p−1 − x 2 (1 − x) p−1 − (1 − x)2
V 0 (x) = 3 + x k−1 + (1 − x)k−1 = 3 + .

−
k=3
x −1 x
In particular, V 0 (0) = V 0 (1) = 3 + ( p − 3) = 0. We have a p = a and (1 − a) p =

1 − a for all a ∈ F p , so for a ∈ F p \ {0, 1} we have
3a(a − 1) + (a p−1 − a 2 )a − (1 − a) p−1 − (1 − a)2 (a − 1)

V (a) =
0
= 0,
a(a − 1)
and (1) is proved.
Editorial comment. O. Geupel referred to a paper by Sandro Mattarei and Roberto
Tauraso (Congruences of multiple sums involving sequences invariant under the bino-
mial transform, Journal of Integer Sequences, 13 (2010)), in which a more general re-
sult is shown for sequences that belong to the eigenspaces corresponding to the eigen-
values 1 or −1 of the binomial transform T , given by (T hai)n = nk=0 nk (−1)k ak .
P
Also solved by O. Geupel (Germany), O. P. Lossers (Netherlands), Ellington Management Problem Solving
Group, and the proposer.
Resolvent Cubics for Quartics with Small Galois Groups

11606 [2011, 847]. Proposed by Kent Holing, Trondheim, Norway. Let a, b, c, d be
integers, the first two even and the other two odd. Let Q be the polynomial x 4 + ax 3 +
bx 2 + cx + d, and assume that the Galois group of Q has order less than 24.

(a) Show that the resolvent cubic
P(x) = x 3 − bx 2 + (ac − 4d)x + (4bd − c2 − a 2 d)
of Q has exactly one integer root; call it m.
(b) Show that a 2 + 4(m − b) cannot be a nonzero square.
(c) Show that if a = 0, then the Galois group of Q is cyclic if and only if (m − b)(m +
b)2 − 4c2 is a square.
Solution by Richard Stong, Center for Communications Research, San Diego, CA. The
statement has been changed in part (a) to use the resolvent cubic rather than the La-
grange resolvent cubic; this is needed to make (b) and (c) true. An alternate fix keeps
the Lagrange resolvent cubic and modifies (b) and (c).
(a) Let p, q, r, s be the roots of Q, let K be the splitting field Q[ p, q, r, s] of Q,
and let G be the Galois group of Q viewed as a subgroup of the symmetric group S4
on { p, q, r, s}. Reducing Q modulo 2 yields x 4 + x + 1, which is irreducible over the
two-element field F2 . This implies that Q is irreducible and that G contains a 4-cycle.
Since G 6= S4 by assumption, G is the cyclic group of order 4 or the dihedral group
of order 8. Hence 3 does not divide |G|, which equals [K : Q]. Hence P, whose roots
pq + r s, pr + qs, and ps + qr are all in K , cannot be irreducible. If all three roots of
P are rational (hence integral), then
a 2 + 4( pr + qs − b) = ( p − q + r − s)2 = n 2 ∈ Z
and similarly for the other two roots. Thus
√ √ √
a + n1 + n2 + n3
p= ;
4
√ √ √
q, r , and s can be similarly written. This implies that K = Q[ n 1 , n 2 , n 3 ]. Such
a biquadratic extension has the Klein 4-group Z2 as its Galois group, but Z22 has no
2
4-cycle. Independence of the three square roots is ruled out, since p only has degree
4 over the rationals. Thus, P has exactly one rational (and hence integral) root, which
we may assume is pr + qs. Henceforth call it m.
(b) We have a 2 + 4(m − b) = ( p − q + r − s)2 ; suppose that the value is the
nonzero square k 2 . Note that p + r = (a + k)/2 and q + s = (a − k)/2, so both are
rational, and hence they are preserved by any σ ∈ G. Thus, σ ( p) + σ (r ) = p + r 6 =
q + s, and so {σ ( p), σ (r )} and { p, r } cannot be disjoint. Since these two-element sets
overlap and have the same sums, they must agree. Hence, G preserves both { p, r } and
{q, s}. This again forces G to be a subgroup of Z22 , again a contradiction.
(c) Note that G is a subgroup of the group h( p q r s), ( p r )i of permutations. We
can verify that
2( p − r )(q − s)( p − q + r − s)
is fixed by ( p q r s) and is negated by ( p r ). Therefore, G = h( p q r s)i is cyclic
if and only if X is a square, where X = 4( p − r )2 (q − s)2 ( p − q + r − s)2 . But X =
(4m − 4b + a 2 )(2m + 2b − a 2 )2 = (a 3 − 4ab + 8c)2 , so X is a square. In the special
case where a = 0, we obtain
X/16 = (m − b)(m + b)2 − 4c2 ,
so the characterization is as claimed.
Editorial comment. The Ellington Management Problem Solving Group also changed
the problem statement in the same way, and for the same reason.

Also solved by G. Apostolopoulos (Greece) (part (a) only), B. Karaivanov, the Ellington Management Problem
Solving Group, and the proposer.
A Harmonic Number Sum

11609 [2011, 936]. Proposed by M. N. Deshpande, Nagpur, India. Let n be an integer
greater than 1, and let Sk (n) be the family of all subsets of {2, . . . , n} with k elements.
Let H (k) = kj=1 1j . Show that
P
n−1
X X Y1
(2n + 1 − 2k) = (n + 1)((n + 2) − H (n + 1)).
k=0 A∈S (n) j∈A
j
k
Solution by C. R. Pranesachar, Indian Institute of Science, Bangalore, India. Let

n
Y 1
g(x) = x 3 x2 + .
j=2
j
Expanding the product yields

X Y1 X n−1 X Y1
g(x) = x 3
x 2(n−1−|A|)
= x 2n+1−2k .
A⊆{2,...,n} j∈A
j k=0 A∈S (n) j∈A
j
k
Thus the desired sum is g 0 (1). Since

n
3 X 2x
g (x) = g(x)
0
+ ,
x j=2
x 2 + 1/j
we have
n
Y n
j +1 X j
g (1) =
0
3+2
j=2
j j=2
j +1
n
n+1 X 1
= 3+2 1−
2 j=2
j +1

3 1 1 1
= (n + 1) + n − 1 − − − ··· −
2 3 4 n+1
= (n + 1)(n + 2 − H (n + 1)),
as desired.
Also solved by T. Amdeberhan & A. Straub, D. Beckwith, N. Caro (Brazil), R. Chapman (U.K.), P. P. Dályay
(Hungary), E. S. Eyeson, O. Geupel (Germany), Y. J. Ionin, B. Karaivanov, O. Kouba (Syria), R. Leroy,
J. H. Lindsey II, O. P. Lossers (Netherlands), G. Martin (Canada), M. A. Prasad (India), R. Pratt, J. Schlosberg,
E. Schmeichel, B. Schmuland (Canada), J. H. Steelman, A. Stenger, R. Stong, R. Tauraso (Italy), M. Tetiva
(Romania), D. B. Tyler, Z. Zhang, GCHQ Problem Solving Group (U.K.), NSA Problems Group, and the
proposer.

REVIEWS
Edited by Jeffrey Nunemacher
Mathematics and Computer Science, Ohio Wesleyan University, Delaware, OH 43015
Calculus: Modeling and Application, second edition. By David A. Smith and Lawrence
C. Moore. Mathematical Association of America, Washington, D.C., 2010. http://maa.
pinnaclecart.com. Price: $35.00. ISBN 978-1-6144-610-1.
Reviewed by Betty Mayfield

Sit back. Relax. Close your eyes. Now imagine that you are picking up a calculus
textbook. Feel its mighty heft. Open it to see the formulas inside the front cover, the
table of integrals at the back of the book, the solutions to odd-numbered exercises.
Scan the table of contents for the familiar topics, in the familiar order. Flip open to a
page at random and admire the colorful graphics, the theorems helpfully boxed and
shaded, the examples carefully worked out. Note the many exercises you may assign
to your students. Resist the urge to call the university bookstore and ask how much
they are charging this semester for this impressive tome.
Now imagine a completely different kind of text—different both in content and
delivery. In fact, you are reading it on your iPad. That book might be David Smith
and Lang Moore’s Calculus: Modeling and Application (CMA), available through the
MAA Bookstore as an electronic text. It is a direct descendent of The Calculus Reader
[2], one of the more popular texts to grow out of the calculus reform movement of the
1990s [1], and it retains much of the flavor of that earlier book. Its emphasis is on the
use of authentic real-world problems to motivate learning new mathematics; on the use
of technology for graphing, computation and exploration; and on collaboration among
students as they discover patterns and construct their own mathematical knowledge.
As an example of how this book is different from a standard calculus text, we ex-
amine its treatment of several standard calculus topics. In most texts, the section on the
Product Rule begins with a statement of the answer, dtd (uv) = u · dvdt
+ v · du
dt
, set off in
a color-coded box. Then the student is shown several examples to convince him or her
that the rule seems to work. This discussion is followed by a formal proof, and maybe
a picture of why it makes sense. The text quickly goes on to state the Quotient Rule
and a Power Rule for Negative Exponents for differentiation. There are homework
problems of each of these types at the end of the section.
In CMA, the section on the Product Rule begins with a motivating real-world prob-
lem: In the 1980s, the U.S. population was growing (at a roughly exponential rate,
though admittedly with a very small exponent), and per capita energy consumption
was also growing (at a roughly linear rate), and so the total energy consumption, rep-
resented by the product of those two functions, was certainly growing as well—but at
what rate? Students realize that they know how to differentiate each factor separately,
but they are not sure if they know how to differentiate their product.
Instead of giving them the answer, the text launches into an Activity (a familiar
sight to students of this text) in which they look at several examples of functions u, v,
their product uv, and their derivatives. Not only do students quickly notice that the

derivative of a product is apparently not the product of the derivatives, they are led to
conjecture a formula on their own. In my experience, after experimenting with four
or five examples of simple functions, someone in the class always says, “Hey, what
if we added those two? No wait, multiply those first. . . ” and we really do “discover”
the Product Rule together. (There will undoubtedly be students in your class who have
taken calculus before, and some of them might actually remember the Product Rule
correctly; you may have to ask them to keep quiet for a few minutes.) Then we can go
back and answer the question about energy consumption.
In fact, the introduction to the derivative itself is unusual in CMA. There is no
preliminary discussion of limit or continuity, and the word “secant” does not make an
appearance until we reach trigonometric functions at the end of the semester. Instead,
students are reminded of what they already know about slope and average rate of
change and are introduced to the concept of a difference quotient, all in the context
of the motion of a falling body. Then the text shifts to a discussion of local linear-
ity: Zoom in really close on the graph of a parabola; what does it look like? How
can we use that information to define the instantaneous speed of a falling body at
a certain time? We agree that we can get a pretty good idea by computing the dif-
ference quotient using a tiny value of 1t— say, 0.001. After students compute this
nearly-instantaneous rate for several values of t, and are convinced that we can do so
for any value, they begin to understand that we have defined a new function of t, an
instantaneous speed function. And we can graph it. Using a computer algebra system,
students graph the function f (t) = k · t 2 , along with its almost-instantaneous speed
function, for several values of k, and then try to write down a formula for the speed in
each case. By the end of the activity, whether they realize it or not, students have just
discovered the derivative of the quadratic function f (t) = k · t 2 . Eventually students
do verify these results by evaluating the difference quotient algebraically and taking
the “limiting value” (the word “limit” is not used until we reach definite integrals and
their limits of integration many chapters later) of the result, and a formal definition
of the derivative appears—in the requisite box. But we return over and over to this
graphical/numerical approach of approximating a derivative with a difference quotient
and trying to guess what the resulting function is, whenever we are confronted with a
function whose derivative we do not know.
In a later chapter, after working with left- and right-hand sums to approximate the
area under a curve, and noting their inherent limitations, students are led to consider
two improvements on them: using the midpoint of each interval instead of either end-
point, or taking the average of the two sums. Thus they are introduced to the Midpoint
Rule and the Trapezoidal Rule in a natural way. Simpson’s Rule is presented as a
weighted average of those two new approximations—something I had never seen
before.
Those easily shocked will have many provocations: The definition of a continuous
function does not appear until Chapter 10; the Quotient Rule is hidden in a homework
exercise and may be skipped entirely; the name Riemann does not appear in the index.
Instructors will search in vain for the standard related rates problems (the conical pile
of sand, the sliding ladder, the shadow of the man who is doomed to spend eternity
walking away from that light pole). Instead, there is an Air Traffic Control project
involving velocities and three-dimensional distances which students are expected to
solve in groups, working with minimal guidance and using their knowledge of deriva-
tives. There are no solids of revolution whose volumes must be found by shells or
discs. But students compute the balance point of a pool cue and, later, the centers of
mass of several regions in the plane. (In our classes, students cut those regions out
of cardboard, and we make a class mobile using their computed balance points.) Stu-
November 2013] REVIEWS 863

dents become adept at fitting models to data, often using semilog and log-log plots to
determine what kind of function may work best. We should take the title of this book
seriously. To help them decide which topics to include—and which to omit—from this
text, the authors consulted with faculty from our “partner disciplines” and asked them
how and when they really used calculus in their fields. The answers they got drove the
content of this book.
Another unique feature of this text is the fact that it is only available in electronic
form, in either a computer or tablet version. This is not a static pdf file which happens
to be available on a computer; it is truly electronic, truly interactive. As students read
the text, they can click on live buttons that lead them to pictures, graphs, a calculator,
and to computer activities where they can explore and experiment. The computer ver-
sion requires the use of the Firefox browser and one of the computer algebra systems
Maple, Mathematica, or Mathcad for most of the activities. It can be downloaded
and then used without internet access. The tablet version runs on either the Safari or
Firefox browser and uses interactive Sage worksheets; it does require internet access.
The price is right for this text: students pay $35 for one-year electronic access from
the time of purchase.
There are ample homework exercises; in fact, for each section there are two levels of
assignments—Exercises (the more customary mechanical/rote problems, reinforcing
the material in the text) and Problems (deeper, richer, requiring more exploration and
thought). There are no answers to odd-numbered problems at the end of the text, but
almost all of the Exercises are available through the CMA Library in WeBWorK, the
open-source online homework system now managed by the MAA (http://webwork.
maa.org/). In that system, students are given immediate feedback and, if the instruc-
tor allows, may go back and try the problem again; faculty have no homework to grade
but are given detailed information about students’ performance. Our experience here at
Hood College has been that students will exhibit a remarkable level of perseverance in
trying to solve a problem correctly and to receive the computer’s affirmation—a trait
we have never particularly observed when assigning pencil-and-paper assignments.
Although this text is designed to be used in a class—and a rather unconventional
one at that—its features make it appropriate to be used in many settings (with or
without group work, with or without a computer algebra system, with or without the
WeBWorK exercises), and even for independent study. The activities and checkpoints
(all answers are in the text), the exercises and problems, and even the projects could
be completed by a student working alone.
Lots of information is available about this text on the MAA website, including
sample chapters. If you are ready to make a change in the way you and your students
experience calculus, this just might be the book for you.
You can open your eyes now.
REFERENCES
1. B. Darken, R. Wynegar, S. Kuhn, Evaluating calculus reform: a review and a longitudinal study, CBMS
Issues in Mathematics Education, Vol. 8. Edited by E. Dubinsky, A. Schoenfeld, and J. Kaput. American
Mathematical Society, Providence, RI, 2000, 16–41.
2. D. Smith, L. Moore, The Calculus Reader. D.C. Heath, Lexington, MA, 1994.
Hood College, Frederick, MD 21701

mayfield@hood.edu

Jump into Abstract Mathematics
with
Bridge to Abstract Mathematics
By Ralph Oberste-Vorth, Aristides Mouzakitis,
and Bonita A. Lawrence
MAA Textbooks
A Bridge to Abstract Mathematics will prepare the math-

ematical novice to explore the universe of abstract math-
ematics. Mathematics is a science that concerns theorems
that must be proved within the constraints of a logical
system of axioms and definitions, rather than theories
that must be tested, revised, and retested. Readers will
learn how to read mathematics beyond popular computa-
tional calculus courses. Moreover, readers will learn how
to construct their own proofs.
The book is intended as the primary text for an introductory course in proving
theorems, as well as for self-study or as a reference. Throughout the text, some pieces
(usually proofs) are left as exercises; Part V gives hints to help students find good
approaches to the exercises. Part I introduces the language of mathematics and the
methods of proof. The mathematical content of Parts II through IV were chosen so as
not to seriously overlap the standard mathematics major. In Part II, students study
sets, functions, equivalence and order relations, and cardinality. Part III concerns
algebra. The goal is to prove that the real numbers form the unique, up to isomor-
phism, ordered field with the least upper bound; in the process, we construct the real
numbers starting with the natural numbers. Students will be prepared for an abstract
linear algebra or modern algebra course. Part IV studies analysis. Continuity and dif-
ferentiation are considered in the context of time scales (nonempty closed subsets of
the real numbers). Students will be prepared for advanced calculus and general topol-
ogy courses. There is a lot of room for instructors to skip and choose topics from among
those that are presented.
2012, 252 pp., Catalog Code: BTAM

ISBN: 978-0-88385-779-3
List: $60.00 MAA Member: $50.00
To order visit us online at www.maa.org or call 1-800-331-1622.

MATHEMATICAL ASSOCIATION OF AMERICA
1529 Eighteenth St., NW • Washington, DC 20036
Fantastic Titles from Cambridge!

Manifold Mirrors A Course in Beauty
The Crossing Paths Mathematical Analysis Edited by Lauren Arrington,
of the Arts and Mathematics D. J. H. Garling Zoe Leinhardt, and Philip Dawid
Felipe Cucker Volume 1: Foundations and Darwin College Lectures
$90.00: Hb: 978-0-521-42963-4: 424 pp. Elementary Real Analysis $19.99: Pb: 978-1-107-69343-2: 210 pp.
$29.99: Pb: 978-0-521-72876-8 $125.00: Hb: 978-1-107-03202-6: 314 pp. Formal Languages in Logic
$50.00: Pb: 978-1-107-61418-5
A Philosophical and Cognitive Analysis
Volume 2: Metric and Topological Spaces, Catarina Dutilh Novaes
Functions of a Vector Variable $95.00: Hb: 978-1-107-02091-7: 282 pp.
$125.00: Hb: 978-1-107-03203-3: 328 pp.
$50.00: Pb: 978-1-107-67532-2 How Humans Learn to
Volume 3: Complex Analysis,
Think Mathematically
Measure and Integration Exploring the Three Worlds of Mathematics
$125.00: Hb: 978-1-107-03204-0: 310 pp. David Tall
$50.00: Pb: 978-1-107-66330-5 Learning in Doing: Social, Cognitive and
Computational Perspectives
Who’s Bigger?
Where Historical Figures Really Rank
An Introduction to 2nd
$99.00: Hb: 978-1-107-03570-6: 488 pp.
Gödel’s Theorems Edition! $39.99: Pb: 978-1-107-66854-6
Steven Skiena and
Charles Ward Peter Smith Programming with
$27.99: Hb: 978-1-107-04137-0: 408 pp.
Cambridge Introductions to Philosophy Mathematica®: An Introduction
$99.00: Hb: 978-1-107-02284-3: 402 pp. Paul Wellin
Prices subject to change. $34.99: Pb: 978-1-107-60675-3 $95.00: Hb: 978-1-107-00946-2: 728 pp.
www.cambridge.org/mathematics
800.872.7423
@cambUP_maths

Noiembrie 2013

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Noiembrie 2013

Uploaded by

Copyright:

Available Formats

THE AMERICAN MATHEMATICAL

Christiane’s Hair 771

What to Expect in a Game of Memory 787

Mixing Problems with Many Tanks 806

Reciprocal Sums as a Knowledge Metric: Theory, 822

Certain Inequalities Associated with the Divisor Function 832

On the Nonexistence of Certain Limits for the 837

A Generalization of Routh’s Triangle Theorem 841

A Maximum Principle for High-Order Derivatives 846

Nothing New about Equiangular Polygons 849

Quotients of Gaussian Primes 851

PROBLEMS AND SOLUTIONS 854

Calculus: Modeling and Application 862

An Official Publication of the Mathematical Association of America

Print Subscription to MAA FOCUS (6 per year) X X

Regional Section Membership (USA/CAN Members) X X X

Discounted Registration for MAA MathFest X X X

Discounted Registration for Joint Mathematics Meetings X X X

MAA Math Alert—the e-newsletter for MAA members X X X

Member Discounts on MAA books, videos & merchandise X X X

Member Discounts on Print Subscriptions to:

Access to MAA Affinity Program Discounts X X X

Eligibility to Join MAA SIGMAAs ($12 each) X X X

Access to MAA columns, blogs, and online resources X X X

Membership in up to 3 SIGMAAs of your choice X

Print Subscription to Math Horizons X

MAA Member Plus Book Selection X

MAA Annual Design Contest T-shirt X

Be sure to visit www.maa.org for all the details!

NOTES EDITOR BOOK REVIEW EDITOR

PROBLEM SECTION EDITORS

http://www.editorialmanager.com/monthly Recent copies of the MONTHLY are available for pur-

2. CONSTRUCTION OF CH. First we must describe the construction of CH. The

November 2013] CHRISTIANE’S HAIR 771

Figure 1. Iterative construction for the classical Cantor set C

CH = (x, y) : x ∈ C y , y ∈ [0, 1/2] ⊂ [0, 1] × [0, 1/2].

w0 (x) = x/3 and w1 (x) = x/3 + 2/3.

Notice that w0 (C ) = C ∩ [0, 1/3] and w1 (C ) = C ∩ [2/3, 1], so that

772 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 120

Then it is not so hard to see that

w0 (x, y) = yx and w1 (x, y) = yx + (1 − y) . (4)

November 2013] CHRISTIANE’S HAIR 773

Thus CH has Lebesgue measure zero.

4. EACH STRAND OF “HAIR” IS A SMOOTH CURVE. CH is comprised of

774 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 120

φσ := φσ1 ◦ φσ2 ◦ · · · ◦ φσn . (5)

φσ n = φσ1 ◦ φσ2 ◦ · · · ◦ φσn

φσ n+1 = φσ1 ◦ φσ2 ◦ · · · ◦ φσn ◦ φσn+1

for a given σ ∈ 6. The “newest” map is applied on the “inside.”

φ0 ( f )(y) = y f (y) and φ1 ( f )(y) = y f (y) + 1 − y, for y ∈ [0, 1/2].

Notice that, by the definition of φi , we have φi ( f ) : [0, 1/2] → [0, 1] whenever f :

Tσ (y) := φσ ( f )(y) = lim φσ n ( f )(y) (6)

November 2013] CHRISTIANE’S HAIR 775

Continuing on to the second derivative, we see that

and the nth derivatives mappings for n > 1 are

776 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 120

We use the self-similarity of C y to show this.

For a fixed σ ∈ 6, this is a real-analytic function for y ∈ (0, 1/2].

5. GEOMETRIC MEASURE-THEORETIC PROPERTIES OF CH. We now

November 2013] CHRISTIANE’S HAIR 777

From standard results in measure theory, h s is a Borel measure. Furthermore, it is

an intuitively plausible but nontrivial result to prove. Thus, if we can construct a

φαβ (x) = (1 − α)α −1

A = {x : d(x, a) < , for some a ∈ A}.

L y () = 2(2 N (y,)+1 − 1) + (1 − 2y)

(2)1/k ≤ yk ≤ (2)1/k (1 − 2b)−1/k = (2)1/k γ 1/k .

k=N (a,) (2)1/k (2)1/k

k=N (b,) N (b,)

k=N (b,) Z (2γ )1/(k+1)