You are on page 1of 96

Your Favorite MAA Books.

Now Digital.

Save 10% on your order!

coupon code: 353030655

Volume 119, No. 2 February 2012

Scott T. Chapman
Sam Houston State University


Sergei Tabachnikov Jeffrey Nunemacher
Pennsylvania State University Ohio Wesleyan University


Douglas B. West Gerald Edgar Doug Hensley
University of Illinois Ohio State University Texas A&M University

William Adkins Ulrich Krause
Louisiana State University Universitat 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

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
The MONTHLYs 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:
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

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,
theme, or a lively discussion of a single issue.
All at the address:
Beginning January 1, 2011, submission of articles and
notes is required via the MONTHLYs Editorial Man- The Mathematical Association of America
ager System. Initial submissions in pdf or LATEX form 1529 Eighteenth Street, N.W.
can be sent to the Editor Scott Chapman at Washington, DC 20036 Recent copies of the MONTHLY are available for pur-

chase through the MAA Service Center., 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 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 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-
// for informa- ciation of America (Incorporated), 2012, 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 provided that copies are not made or distributed for
profit or commercial advantage and that copies bear
The online MONTHLY archive at is a the following copyright notice: [Copyright the Math-
valuable resource for both authors and readers; it ematical Association of America 2012. All rights re-
may be searched online in a variety of ways for any served.] 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 MAAs 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,
Solving a Generalized Heron Problem
by Means of Convex Analysis
Boris S. Mordukhovich, Nguyen Mau Nam, and Juan Salinas Jr.

Abstract. The classical Heron problem states: on a given straight line in the plane, find a
point C such that the sum of the distances from C to the given points A and B is minimal. This
problem can be solved using standard geometry or differential calculus. In the light of modern
convex analysis, we are able to investigate more general versions of this problem. In this paper
we propose and solve the following problem: on a given nonempty closed convex subset of
Rs , find a point such that the sum of the distances from that point to n given nonempty closed
convex subsets of Rs is minimal.

1. PROBLEM FORMULATION. Heron from Alexandria (1075 AD) was a

Greek geometer and inventor whose writings preserved for posterity a knowledge
of the mathematics and engineering of Babylonia, ancient Egypt, and the Greco-
Roman world (from the Encyclopedia Britannica). One of the geometric problems he
proposed in his Catroptica was as follows: find a point on a straight line in the plane
such that the sum of the distances from it to two given points is minimal.
Recall that a subset  of Rs is called convex if x + (1 )y  whenever x
and y are in  and 0 1. Our idea now is to consider a much broader situation,
where the two given points in the classical Heron problem are replaced by finitely
many closed and convex subsets i , i = 1, . . . , n, and the given line is replaced by a
given closed and convex subset  of Rs . We are looking for a point in the set  such
that the sum of the distances from that point to i , i = 1, . . . , n, is minimal.
The distance from a point x to a nonempty set  is understood in the conventional

d(x; ) = inf ||x y|| y  , (1.1)

where || || is the Euclidean norm in Rs . The new generalized Heron problem is for-
mulated as follows:
minimize D(x) := d(x; i ) subject to x , (1.2)

where all the sets  and i , i = 1, . . . , n, are nonempty, closed, and convex; these are
our standing assumptions in this paper. Thus (1.2) is a constrained convex optimization
problem, and hence it is natural to use techniques of convex analysis and optimization
to solve it.

2. ELEMENTS OF CONVEX ANALYSIS. In this section we review some basic

concepts of convex analysis used in what follows. This material and much more can
be found, e.g., in the books [2, 3, 4, 7].
MSC: Primary 49J52, Secondary 49J53, 90C31


Let f : Rs R := (, ] be an extended-real-valued function, which may be
infinite at some points, and let

dom f := x Rs f (x) <

be its effective domain. The epigraph of f is the subset of Rs R defined by

epi f := (x, ) Rs+1 x dom f and f (x) .

The function f is closed if its epigraph is closed, and it is convex if its epigraph is a
convex subset of Rs+1 . It is easy to check that f is convex if and only if

f x + (1 )y f (x) + (1 ) f (y) for all x, y dom f and [0, 1].

Furthermore, a nonempty closed subset  of Rs is convex if and only if the corre-

sponding distance function f (x) = d(x; ) is a convex function. Note that the dis-
tance function f (x) = d(x; ) is Lipschitz continuous on Rs with modulus one, i.e.,

| f (x) f (y)| ||x y|| for all x, y Rs .

A typical example of an extended-real-valued function is the indicator function

0 if x ,
(x; ) := (2.3)

of the set . It follows immediately from the definitions that  Rs is closed (resp.
convex) if and only if the indicator function (2.3) is closed (resp. convex).
An element v Rs is called a subgradient of a convex function f : Rs R at
x dom f if it satisfies the inequality

f (x) f (x) + hv, x xi for all x Rs , (2.4)

where h, i stands for the usual scalar product in Rs . Intuitively, a vector v Rs is a

subgradient of f at x if and only if f is bounded below by an affine function that agrees
with f at x in which the coefficients are given by v. The set of all the subgradients v in
(2.4) is called the subdifferential of f at x and is denoted by f (x). If f is convex and
differentiable at x, then f (x) = { f (x)}. For example, the function f (x) = |x| is
not differentiable at x = 0. However, one can draw several subtangent lines that go
through (0, 0) with slopes belonging to [1, 1], and these lines stay below the graph

Figure 1. The absolute value function and subtangent lines at (0, 0).


of the function. As a result, we obtain the following subdifferential formula for the
absolute value function f (x) = |x|:

if x < 0,
f (x) = [1, 1] if x = 0,
if x > 0.


A well-recognized technique in optimization is to reduce a constrained optimization

problem to an unconstrained one using the indicator function of the constraint set. It is
obvious that x  is a minimizer of the general constrained optimization problem:

minimize f (x) subject to x  (2.5)

if and only if it solves the unconstrained problem

minimize f (x) + (x; ), x Rs . (2.6)

By the definitions, for any convex function : Rs R,

x is a minimizer of if and only if 0 (x), (2.7)

which is a nonsmooth convex counterpart of the classical Fermat stationary rule. Ap-
plying (2.7) to the constrained optimization problem (2.5) via its unconstrained de-
scription (2.6) requires the usage of subdifferential calculus. The most fundamental
calculus result of convex analysis is the following Moreau-Rockafellar theorem for
the subdifferential of sums; see, e.g., [4, p. 51].

Theorem 2.1. Let i : Rs R, i = 1, . . . , m, be closed convex functions. Assume

that there is a point x i=1 dom i at which all (except possibly one) of the functions
1 , . . . , m are continuous. Then we have the equality
X  m
X m
i (x) = i (x) := vi vi i (x) for i = 1, . . . , m

i=1 i=1 i=1

for all x i=1 dom i .

Given a convex set  Rs and a point x , the corresponding geometric coun-

terpart of (2.4) is the normal cone to  at x defined by

N (x; ) := v Rs hv, x xi 0 for all x  . (2.8)

Observe that a vector v belongs to N (x; ) if and only if it makes a right or obtuse
angle with the vector from x to x for any x . It easily follows from the definitions

(x; ) = N (x; ) for every x , (2.9)

which allows us, in particular, to characterize minimizers of the constrained problem

(2.5) in terms of the subdifferential (2.4) of f and the normal cone (2.8) to  by
applying Theorem 2.1 to the function f (x) + (x; ) in (2.7).


Figure 2. A set and its normal vectors at x = (0, 0).

Finally in this section, we present a useful formula for computing the subdifferential
of the distance function (1.1) via the Euclidean projection

5(x; ) := x  ||x x|| = d(x; ) (2.10)

of x Rs on the closed and convex set  Rs . It follows from the definition of the
Euclidean projection that 5(x; ) = {x} if x  and it is a singleton when x / .
In the sequel, we identify the set 5(x; ) with its unique element.

Proposition 2.2. Let  6 = be a closed and convex subset of Rs . Then

x 5(x; ) o
/ ,
if x
d(x; )

d(x; ) =

N (x; ) B if x ,

where B is the closed unit ball of Rs .

The proof of this formula can be found in [3, p. 181].


this section we derive efficient characterizations of optimal solutions to the generalized
Heron problem (1.2), which allow us to completely solve this problem in some impor-
tant settings.
First let us present general conditions that ensure the existence of optimal solutions
to (1.2).

Proposition 3.1. Assume that one of the sets  and i , i = 1, . . . , n, is bounded.

Then the generalized Heron problem (1.2) admits at least one optimal solution.

Proof. Consider the optimal value

:= inf D(x)

in (1.2) and take a minimizing sequence {xk }  with D(xk ) as k . If

the constraint set  is bounded, then by the classical Bolzano-Weierstrass theorem the
sequence {xk } contains a subsequence converging to a point x, which belongs to the


set  due to its closedness. Since the function D(x) in (1.2) is continuous, we have
D(x) = , and so x is an optimal solution to (1.2).
It remains to consider the case when one of sets i , say 1 , is bounded. In this case
we have for the above sequence {xk } when k is sufficiently large that

d(xk ; 1 ) D(xk ) < + 1,

and thus there exists wk 1 with ||xk wk || < + 1 for such indices k. Then

||xk || < ||wk || + + 1,

which shows that the sequence {xk } is bounded. The existence of optimal solutions
follows in this case from the arguments above.

To characterize optimal solutions to the generalized Heron problem (1.2) in what

follows, for any nonzero vectors u, v Rs define the cosine of the angle between u
and v by the quantity

hv, ui
cos(v, u) := . (3.11)
||v|| ||u||

We say that N (x; ) is representable by a subspace L = L(x) 6 = {0} if

N (x; ) = L := v Rs hv, ui = 0 whenever u L . (3.12)

The next theorem gives necessary and sufficient conditions for optimal solutions to
(1.2) via projections (2.10) on i incorporated into quantities (3.11).

Theorem 3.2. Consider problem (1.2) in which

i  = for i = 1, . . . , n. (3.13)

Given x , define the vectors

x 5(x; i )
ai (x) := 6 = 0, i = 1, . . . , n. (3.14)
d(x; i )

Then x  is an optimal solution to the generalized Heron problem (1.2) if and only
if we have the inclusion
ai (x) N (x; ). (3.15)

Suppose in addition that the normal cone to the constraint set N (x; ) is representable
by a subspace L. Then (3.15) is equivalent to
cos ai (x), u = 0 whenever u L \ {0}. (3.16)


Proof. Problem (1.2) is equivalent to the following unconstrained optimization



minimize D(x) + (x; ), x Rs . (3.17)

Applying the generalized Fermat rule (2.7), we see that x is a solution to (3.17) if and
only if
0 d(; i ) + (; ) (x). (3.18)

Since all of the functions d(; i ), i = 1, . . . , n, are convex and continuous, we em-
ploy the subdifferential sum rule of Theorem 2.1 to (3.18) and arrive at
0 D + (, ) (x) = d(x; i ) + N (x; )

X (3.19)
= ai (x) + N (x; ),

where the second representation in (3.19) is due to (2.9), assumption (3.13), and the
subdifferential description of Proposition 2.2 with ai (x) defined in (3.14). It is obvious
that (3.19) and (3.15) are equivalent.
Suppose in addition that the the normal cone N (x; ) to the constraint set is repre-
sentable by a subspace L. Then the inclusion (3.15) is equivalent to
0 ai (x) + L ,

which in turn can be written in the form

X n 
ai (x), u = 0 for all u L .

Taking into account that ||ai (x)|| = 1 for i = 1, . . . , n by (3.14) and assumption
(3.13), the latter equality is equivalent to
X hai (x), vi
= 0 for all u L \ {0},
||ai (x)|| ||u||

which gives (3.16) due to the notation (3.11) and thus completes the proof of the

To further specify the characterization in Theorem 3.2, recall that a subset A of Rs

is an affine subspace if there is a vector a A and a subspace L such that A = a + L.
In this case we say that A is parallel to L. Note that the subspace L parallel to A is
uniquely defined by L = A A = {x y | x A, y A} and that A = b + L for any
vector b A.

Corollary 3.3. Let  be an affine subspace parallel to a subspace L, and let assump-
tion (3.13) of Theorem 3.2 be satisfied. Then x  is a solution to the generalized
Heron problem (1.2) if and only if condition (3.16) holds.


Proof. To apply Theorem 3.2, it remains to check that N (x; ) is representable by the
subspace L in the setting of this corollary. Indeed, we have  = x + L, since  is an
affine subspace parallel to L. Fix any v N (x; ) and get by (2.8) that hv, x xi 0
whenever x  and hence hv, ui 0 for all u L. Since L is a subspace, the latter
implies that hv, ui = 0 for all u L, and thus N (x; ) L . The opposite inclusion
is trivial, which gives (3.12) and completes the proof of the corollary.

The underlying characterization (3.16) can be checked easily when the subspace L
in Theorem 3.2 is given as the span of fixed generating vectors.

Corollary 3.4. Let L = span{u 1 , . . . , u m } with u j 6 = 0, i = 1, . . . , m, in the setting

of Theorem 3.2. Then x  is an optimal solution to the generalized Heron problem
(1.2) if and only if
cos ai (x), u j = 0 for j = 1, . . . , m. (3.20)


Proof. We show that (3.16) is equivalent to (3.20) in the setting under consideration.
Since (3.16) obviously implies (3.20), it remains to justify the opposite implication.
a := ai (x)

and observe that (3.20) yields the condition

ha, u j i = 0 for j = 1, . . . m, (3.21)

since u j 6= 0 for j = 1, . . . , m and ||ai || = 1 for i = 1, . . . , n. Taking now any vector

u L \ {0}, we represent it in the form
u= j u j with j R

and get from (3.21) the equalities

ha, ui = j ha, u j i = 0.

This justifies (3.16) and completes the proof of the corollary.

Let us examine in more detail the case of two sets 1 and 2 in (1.2) with the
normal cone to the constraint set  being a straight line generated by a given vector.
This is a direct extension of the classical Heron problem to the setting when the two
points are replaced by closed and convex sets, and the constraint line is replaced by a
closed convex set  with the property above. The next theorem gives a complete and
verifiable solution to the new problem.

Theorem 3.5. Let 1 and 2 be subsets of Rs such that  i = for i = 1, 2 in

(1.2). Suppose also that there is a vector a 6 = 0 such that N (x; ) = span{a}. The


following assertions hold, where the quantities ai := ai (x) are defined in (3.14):
(i) If x  is an optimal solution to (1.2), then

either a1 + a2 = 0 or cos(a1 , a) = cos(a2 , a). (3.22)

(ii) Conversely, if s = 2 and

either a1 + a2 = 0 or a1 6 = a2 and cos(a1 , a) = cos(a2 , a) , (3.23)


then x  is an optimal solution to the generalized Heron problem (1.2).

Proof. It follows from Theorem 3.2 that x  is an optimal solution to (1.2) if and
only if a1 a2 N (x; ). By the assumed structure of the normal cone to  the
latter is equivalent to the alternative:

either a1 + a2 = 0 or a1 + a2 = a for some 6 = 0. (3.24)

To justify (i), let us show that the second equality in (3.24) implies the correspond-
ing one in (3.22). Indeed, we have ||a1 || = ||a2 || = 1, and thus

ha1 , ai = ha1 , a1 + a2 i
= ha1 , a1 i + ha1 , a2 i
= 1 + ha1 , a2 i
= ha2 , a2 i + ha2 , a1 i
= ha2 , a1 + a2 i
= ha2 , ai,

which ensures that ha1 , ai = ha2 , ai as 6 = 0. This gives us the equality cos(a1 , a) =
cos(a2 , a) due to ||a1 || = ||a2 || = 1 and a 6 = 0. Hence we arrive at (3.22).
To justify (ii), we need to prove that the relationships in (3.23) imply

a1 a2 N (x; ) = span{a}. (3.25)

If a1 + a2 = 0, then (3.25) is obviously satisfied. Consider the alternative in (3.23)

when a1 6= a2 and cos(a1 , a) = cos(a2 , a). Choose a vector b 6 = 0 orthogonal to a and
express a1 and a2 in terms of the basis {a, b} by:

a1 = x1 a + y1 b and a2 = x2 a + y2 b.

Since cos(a1 , a) = cos(a2 , a), we have x1 = x2 . Then y1 = y2 by ||a1 ||2 = ||a2 ||2 .
Due to a1 6= a2 this implies y1 = y2 and thus completes the proof.

Finally in this section, we present two examples illustrating the application of The-
orem 3.2 and Corollary 3.4, respectively, to solving the corresponding generalized and
classical Heron problems.

Example 3.6. Consider problem (1.2) where n = 2, the sets 1 and 2 are two points
A1 and A2 in the plane, and the constraint  is a disk that does not contain A1 or A2 .
Condition (3.15) from Theorem 3.2 characterizes a solution M  to this generalized


Heron problem as follows. If the line segment A1 A2 intersects the disk, then any point
in the intersection is an optimal solution. In this case the problem may actually have
infinitely many solutions. Otherwise, there is a unique point M on the circle such that
a normal vector nE to  at M is the angle bisector of angle A1 M A2 , and that is the only
optimal solution to the generalized Heron problem under consideration; see Figure 3.

10 10
8 8
A2 A2
6 6
4 M 4
2 2 M
0 0

2 2
4 4
6 6
8 8
10 10
10 8 6 4 2 0 2 4 6 8 10 10 8 6 4 2 0 2 4 6 8 10
x x
Figure 3. Generalized Heron problems for two points with disk constraints.

Example 3.7. Consider problem (1.2), where i = {Ai }, i = 1, . . . , n, are n points

in the plane, and where  = L R2 is a straight line that does not contain any of
these points. Then, by Corollary 3.4 of Theorem 3.2, a point M L is a solution to
this generalized Heron problem if and only if

cos(MA1 , aE ) + + cos(MAn , aE ) = 0,

where aE is a direction vector of L. Note that the latter equation completely character-
izes the solution of the classical Heron problem in the plane in both cases when A1
and A2 are on the same side and different sides of L; see Figure 4.

10 10
8 8
6 6
4 4
2 2
0 M 0 M

2 2
4 4
6 6 A1
8 8
10 10
10 8 6 4 2 0 2 4 6 8 10 10 8 6 4 2 0 2 4 6 8 10
x x
Figure 4. The classical Heron problem.


we present and justify an iterative algorithm to solve the generalized Heron problem
(1.2) numerically and illustrate its implementations using MATLAB in two important
settings with disk and ball constraints. Here is the main algorithm.

Theorem 4.1. Let  and i , i = 1, . . . , n, be nonempty closed convex subsets of Rs

such that at least one of them is bounded. Picking a sequence {k } of positive numbers
and a starting point x1 , consider the iterative algorithm:

xk+1 = 5 xk k vik ;  , k = 1, 2, . . . , (4.26)

where the vectors vik in (4.26) are constructed by

xk 5(xk ; i )

/ i ,
if xk
d(xk ; i )

vik :=

0 if xk i .

Assume that the given sequence {k } in (4.26) satisfies the conditions

k2 < .
k = and (4.27)
k=1 k=1

Then the iterative sequence {xk } in (4.26) converges to an optimal solution of the gen-
eralized Heron problem (1.2) and the value sequence

Vk := min D(x j ) j = 1, . . . , k (4.28)

converges to the optimal value V in this problem.

Proof. Observe that algorithm (4.26) is well posed, since the projection to a convex
set used in (4.26) is uniquely defined. Since one of the sets  and i , i = 1, . . . , n,
is bounded, the problem has an optimal solution by Proposition 3.1. This algorithm
and its convergence under conditions (4.27) are based on the subgradient method for
convex functions in the so-called square summable but not summable case (see,
e.g., [1, Proposition 8.2.8, p. 480]), the subdifferential sum rule of Theorem 2.1, and
the subdifferential formula for the distance function given in Proposition 2.2. The
reader can compare this algorithm and its justifications with the related developments
in [6] for the numerical solution of the (unconstrained) generalized Fermat-Torricelli

Example 4.2. Consider the generalized Heron problem (1.2) for (not necessarily dis-
joint) squares i , i = 1, . . . , n, of right position in R2 (i.e., such that the sides of each
square are parallel to the x-axis and the y-axis) subject to a given disk constraint . Let
ci = (ai , bi ) and ri , i = 1, . . . , n, be the centers and half the side lengths of the squares
under consideration. The vertices of the ith square are denoted by q1i = (ai + ri ,
bi + ri ), q2i = (ai ri , bi + ri ), q3i = (ai ri , bi ri ), q4i = (ai + ri , bi ri ). Let
r and p = (, ) be the radius and the center of the constraint.


The projection P(x, y) := 5((x, y); ) for (x, y)
/  is calculated by

P(x, y) = (wx + , w y + )

r (x )
wx = p
(x )2 + (y )2

r (y )
wy = p .
(x )2 + (y )2

For (x, y) , P(x, y) = (x, y).

Let xk = (x1k , x2k ) be the sequence defined by (4.26), in which the quantities vik
are computed by the following formula:

0 if |x1k ai | ri and |x2k bi | ri ,

xk q1i

if x1k ai > ri and x2k bi > ri ,

kxk q1i k

xk q2i

if x1k ai < ri and x2k bi > ri ,

kxk q2i k

xk q3i

if x1k ai < ri and x2k bi < ri ,

kxk q3i k

vik = xk q4i
if x1k ai > ri and x2k bi < ri ,
kxk q4i k

(0, 1) if |x1k ai | ri and x2k bi > ri ,

(0, 1) if |x1k ai | ri and x2k bi < ri ,

(1, 0) if x1k ai > ri and |x2k bi | ri ,

(1, 0) if x1k ai < ri and |x2k bi | ri

for i = 1, . . . , n and k = 1, 2, . . .

To implement this algorithm we developed a MATLAB program. Figure 5 and

the corresponding table show the result of applying this algorithm for the disk con-
straint  with center (3, 4) and radius 1.5, for the squares i with centers (7, 1),
(5, 8), (4, 7), and (5, 1) of the same half side length 1, for the starting point x1 =
(3, 5.5) , and for the sequence k = 1/k in (4.26) satisfying conditions (4.27).
The approximate optimal solution and optimal value are x (2.04012, 2.84734)
and V 26.13419.


4 k xk Vk
1 (3, 5.5) 30.99674
y 100 (2.02957, 2.85621) 26.13427
1000 (2.03873, 2.84850) 26.13419
100,000 (2.04010, 2.84735) 26.13419
200,000 (2.04011, 2.84735) 26.13419
10 400,000 (2.04012, 2.84734) 26.13419
10 8 6 4 2 0 2 4 6 8 10 600,000 (2.04012, 2.84734) 26.13419

Figure 5. Generalized Heron problem for squares with a disk constraint.

Example 4.3. Consider the generalized Heron problem (1.2) for (not necessarily dis-
joint) cubes of right position in R3 subject to a ball constraint. In this case the projec-
tion 5((x, y, z); ) and quantities vik are computed similarly to Example 4.2.
Once again, we implemented this algorithm with a MATLAB program. Figure 6 and
the corresponding table present the calculation results for the ball constraint  with
center (0, 2, 0) and radius 1, for the cubes i with centers (0, 4, 0), (4, 2, 3),
(3, 4, 2), (5, 4, 4), and (1, 8, 1) of the same half side length 1, for the starting
point x1 = (0, 2, 0), and for the sequence k = 1/k in (4.26) satisfying (4.27). The ap-
proximate optimal solution and optimal value are x (0.92531, 1.62907, 0.07883)
and V 22.23480.


6 k xk Vk
2 1 (0, 2, 0) 24.18180
0 10 (0.92583, 1.63052, 0.07947) 22.23480
2 100 (0.92531, 1.62908, 0.07884) 22.23480
6 4 8
1,000 (0.92531, 1.62907, 0.07883) 22.23480
2 0 4 6
2 4
6 8 6 4 2
0 2
10,000 (0.92531, 1.62907, 0.07883) 22.23480
10 8
20,000 (0.92531, 1.62907, 0.07883) 22.23480
30,000 (0.92531, 1.62907, 0.07883) 22.23480
Figure 6. Generalized Heron problem for cubes with a ball constraint.

ACKNOWLEDGMENTS. Research of the first author was partially supported by the US National Science
Foundation under grant DMS-1007132 and by the Australian Research Council under grant DP-12092508.


1. D. Bertsekas, A. Nedic, A. Ozdaglar, Convex Analysis and Optimization, Athena Scientific, Belmont, MA,
2. J. M. Borwein, A. S. Lewis, Convex Analysis and Nonlinear Optimization: Theory and Examples, second
edition, Springer, New York, 2006.
3. J.-B. Hiriart-Urruty, C. Lemarechal, Fundamentals of Convex Analysis. Springer-Verlag, Berlin, 2001.
4. G. G. Magaril-Ilyaev, M. V. Tikhomirov, Convex Analysis: Theory and Applications, American Mathe-
matical Society, Providence, RI, 2003.


5. H. Martini, K. J. Swanepoel, G. Weiss, The Fermat-Torricelli problem in normed planes and spaces,
J. Optim. Theory Appl. 115 (2002) 283314; available at
6. B. S. Mordukhovich, N. M. Nam, Applications of variational analysis to a generalized Fermat-Torricelli
problem, J. Optim. Theory Appl. 148 (2011) 431454; available at
7. R. T. Rockafellar, Convex Analysis, Princeton University Press, Princeton, NJ, 1970.

BORIS MORDUKHOVICH is Distinguished University Professor and President of the Academy of Schol-
ars at Wayne State University. He has more than 300 publications including monographs and patents. Among
his best known achievements are the introduction of powerful constructions of generalized differentiation
(bearing his name), their development, and applications to broad classes of problems in variational analy-
sis, optimization, equilibrium, control, economics, engineering, and other fields. Mordukhovich is a SIAM
Fellow and a recipient of many international awards and honors including Doctor Honoris Causa degrees from
four universities.
Department of Mathematics, Wayne State University, Detroit, MI 48202

NGUYEN MAU NAM received his B.S. from Hue University, Vietnam, in 1998 and his Ph.D. from Wayne
State University in 2007 under the direction of Boris Mordukhovich. He is currently an Assistant Professor of
Mathematics at the University of Texas-Pan American.
Department of Mathematics, University of Texas-Pan American, Edinburg, TX 78539

JUAN SALINAS JR. received his B.S. in Electrical Engineering from the University of Texas-Pan American
in 1999. He is currently a graduate student at the University of Texas-Pan American.
Department of Mathematics, University of Texas-Pan American, Edinburg, TX 78539


Jacobi Sum Matrices
Sam Vandervelde

Abstract. In this article we identify several beautiful properties of Jacobi sums that become
evident when these numbers are organized as a matrix and studied via the tools of linear alge-
bra. In the process we reconsider a convention employed in computing Jacobi sum values by
illustrating how these properties become less elegant or disappear entirely when the standard
definition for Jacobi sums is utilized. We conclude with a conjecture regarding polynomials
that factor in an unexpected manner.

1. JACOBI SUMS. Carl Jacobis formidable mathematical legacy includes such con-
tributions as the Jacobi triple product, the Jacobi symbol, the Jacobi elliptic functions
with associated Jacobi amplitudes, and the Jacobian in the change of variables theo-
rem, to but scratch the surface. Among his many discoveries, Jacobi sums stand out
as one of the most brilliant gems. Very informally, a Jacobi sum adds together certain
roots of unity in a manner prescribed by the arithmetic structure of the finite field on
which it is based. (We will supply a precise definition momentarily.) For a given finite
field a Jacobi sum depends on two parameters, so it is natural to assemble these values
into a matrix. We have done so below for the Jacobi sums arising from the field with
eight elements. We invite the reader to study this collection of numbers and identify as
many properties as are readily apparent.


1 1 1 1 1 1
1 1 + i 7 5 1 i 7 1 i 7 5 1 i 7 1 + i 7 1

2 2 2 2
1 5 1 i 7 1 + i 7 1 + i 7 5 1 i 7

2 2 2 2
1 1 i 7
5 1
5 1


1 1 i 7 1 + i 7 1 i 7 1 + 2i 7 2 + 2i 7

1 52 12 i 7 52 12 i 7 1 + i 7 1 i 7 1 + i 7


5 1 5 1
1 1 + i 7 7 7 7 7

1 2
+ 2
i 1 i 1 i 2
+ 2

1 1 1 i 7 2
+ 12 i 7 1 + i 7 5
+ 21 i 7 1 i 7

Before enumerating the standard properties of Jacobi sums we offer a modest back-
ground on their development and applications. According to [2] Jacobi first proposed
these sums as mathematical objects worthy of study in a letter mailed to Gauss in 1827.
Ten years later he published his findings, with extensions of his work provided soon
after by Cauchy, Gauss, and Eisenstein. It is interesting to note that while Gauss sums
will suffice for a proof of quadratic reciprocity, a demonstration of cubic reciprocity
along similar lines requires a foray into the realm of Jacobi sums; Eisenstein formu-
lated a generalization of Jacobi sums (see [3]) in order to prove biquadratic reciprocity.
As shown in [5], Jacobi sums may be used to estimate the number of integral solutions
to congruences such as x 3 + y 3 1 mod p. These estimates played an important role
in the development of the Weil conjectures [6]. Jacobi sums were also employed by
Adleman, Pomerance, and Rumely [1] for primality testing.
MSC: Primary 11T24


Although Jacobi sums have been around for a long time, several of the results pre-
sented below seem to have gone unnoticed. We suspect this has to do in part with the
fact that the usual definition of Jacobi sums differs slightly from the one we use. Con-
ventional wisdom would have us forego the 6 in the upper left corner of (1) in favor of
an 8 and replace each 1 along the top row and left column by a 0. However, some of
the most compelling features of Jacobi sum matrices evaporate when the standard def-
inition is used. Therefore one of our purposes in presenting these results is to suggest
that this alternative warrants serious consideration as the primary definition, at least
in the setting of finite fields. To be fair, the version of Jacobi sums we study does ap-
pear in the literature: e.g., in [2, Section 2.5], which discusses the relationship between
Jacobi sums and cyclotomic numbers.

2. PRELIMINARIES. Recall that there exists a finite field Fq with q elements if and
only if q = pr is a power of a prime, and such a field is unique up to isomorphism.
We shall not require any specialized knowledge of finite fields beyond the fact that the
multiplicative group Fq of nonzero elements forms a cyclic group of order q 1. The
quantity q 1 appears throughout our discussion, so we set m = q 1 from here on.
Thus Fq has m elements.
Fix a generator g of Fq and let = e2i/m . The function defined by (g k ) = k
for 1 k m is an example of a multiplicative character on Fq ; that is, a function
: Fq C satisfying

(1) = 1, (uv) = (u) (v), u, v Fq . (2)

We use an mth root of unity since ((g))m = (g m ) = (1) = 1. As the reader may
verify, there are precisely m multiplicative characters on Fq , namely , 2 , . . . , m ,
where a (g k ) = ((g k ))a = ak as one would expect. Note that m (g k ) = 1 for all k,
so we call m the trivial character. It follows that the value of the exponent a only
matters mod m. In particular, the inverse of a (which is also the complex conjugate)
may be written either as a or as ma . By the same token, we will usually write the
trivial character as 0 .
To define a Jacobi sum it is necessary to extend each character a to all of Fq
by defining a (0). The multiplicative condition forces a (0) = 0 whenever 1 a <
m. But for the trivial character a seemingly arbitrary choice1 must be made, since
taking either 0 (0) = 0 or 0 (0) = 1 satisfies (2). Convention dictates that we declare
0 (0) = 1 for the trivial character. However, we opt for setting a (0) = 0 for all a. As
the opportunity arises we will point out the ramifications of this choice. Properties of
roots of unity now imply that
a (u) = 0, 1 a < m, 0 (u) = q 1 = m. (3)
uFq uFq

(One rationale behind taking 0 (0) = 1 is presumably rooted in the fact that the latter
sum would come to q rather than q 1, giving a more pleasing value.)
A Jacobi sum takes as its arguments a pair of multiplicative characters on a given
finite field and returns a complex number:
1 Ireland and Rosen explain that Jacobi sums arise when counting solutions to equations over F . In this
context 0 (0) tallies solutions to x e = 0, which would seem to motivate the value 0 (0) = 1. However, one
might also argue that the zero solution should not be included since the equations are homogenous, leading to
0 (0) = 0 instead.

February 2012] JACOBI SUM MATRICES 101

Jq ( a , b ) = a (u) b (1 u) = a (u) b (v). (4)
uFq u,vFq

The middle expression is more utilitarian, while the final one highlights the symmetry
in the definition. When the field Fq is clear we will drop the subscript q. We will also
often omit and refer to a particular Jacobi sum simply as J (a, b). Because the terms
of the sum corresponding to u = 0 and u = 1 always vanish, we may write
J (a, b) = a (u) b (1 u), (5)
u6 =0,1

where it is understood that the sum is over u Fq . Thus a Jacobi sum adds together
q 2 not necessarily distinct mth roots of unity.
In a marvelous manner this sum plays the additive and multiplicative structures of
the field off one another, yielding a collection of numbers with extraordinary prop-
erties. To illustrate how these numbers are computed we return to matrix (1), which
catalogs the values J8 ( a , b ) for 0 a, b 6 for a particular generator g of F8 .
(For aesthetic reasons we begin numbering rows and columns of this matrix at 0.) The
generator g of F8 chosen satisfies

g 1 + g 3 = 1, g 2 + g 6 = 1, g 4 + g 5 = 1, g 7 + 0 = 1. (6)

Letting = e2i/7 we may now calculate, for instance,

J (1, 2) = (g) 2 (1 g) + (g 2 ) 2 (1 g 2 ) + + (g 6 ) 2 (1 g 6 ) (7)

= (g) 2 (g 3 ) + (g 2 ) 2 (g 6 ) + + (g 6 ) 2 (g 2 )
= (g 7 ) + (g 14 ) + (g 5 ) + (g 14 ) + (g 13 ) + (g 10 )
= 7 + 14 + 5 + 14 + 13 + 10
= 1 + 1 + 5 + 1 + 6 + 3

= 25 21 i 7,

which explains the entry in row 1, column 2 of (1). For 1 a 6 we find

J (a, 0) = a (g) 0 (1 g) + a (g 2 ) 0 (1 g 2 ) + + a (g 6 ) 0 (1 g 6 ) (8)

2 6
= (g) + (g ) + + (g )
a a a

= a (g 7 )
= 1,

where the penultimate step follows from (3). If we had employed the conventional
value for 0 (0) the term a (g 7 ) would also appear in the sum, giving a total of 0 in-
way of further orientation the reader is encouraged to confirm that J (5, 1) =
stead. By
1 + i 7 and that J (3, 4) = 1.
A cursory examination shows that matrix (1) is symmetric, that the top left entry
equals q 2, and that the remaining entries along the top row, the left column, and the
secondary diagonal are 1. Slightly less obvious is the fact that all other entries have
an absolute value of 8. The sum of the entries along the top row is 0; a quick check


reveals the same is true for every row and column. We summarize these properties
below without proof. (One may consult [5] for details.)

Proposition 1. Fix a generator g of Fq , let (g k ) = k with = e2i/m be the cor-

responding character, take a (0) = 0 for all a, and abbreviate Jq ( a , b ) to J (a, b).
Then for 0 a, b m 1 we have
(i) J (a, b) = J (b, a),
(ii) J (0, 0) = q 2,
(iii) J (a, 0) = J (0, b) = 1 when a, b 6 = 0,
(iv) J (a, m a) = a (1),
(v) |J (a, b)|2 = q when a, b 6 = 0 and a + b 6 = m,
k=0 J (a, k) = k=0 J (k, b) = 0.
Pm1 Pm1

Observe that |J (a, b) + 1|2 and |J (a, b) + 8|2 are either 0 or of the form 2r 7s with
r, s N for every entry of (1). In general the quantities Jq (a, b) + 1 and Jq (a, b) +
q satisfy interesting congruences. We also remark that all the results presented here
continue to be valid regardless of the generator g of Fq used to define . The value of
Jq ( as , bs ) obtained by using the generator g s is identical to that of Jq ( a , b ) using
the original generator g, so altering the generator only permutes the rows and columns
of a Jacobi sum matrix in a symmetric fashion.

3. EIGENVALUES. Thus far our discussion has focused on properties of Jacobi

sums taken individually. However, we are primarily interested in what can be said
about the set of all Jacobi sum values for a particular finite field, viewed collectively
as a matrix. It may have occurred to the curious individual to calculate the eigenval-
ues of matrix (1). We are rewarded for our efforts upon finding that its characteristic
polynomial factors as

p(x) = x(x 7)2 (x 7)2 (x 7)2 , (9)

where = e2i/3 . One might speculate that cube roots of unity make an appearance
since we used characters of F8 , and 8 = 23 . But in fact the same phenomenon occurs
for every value of q. This is explained by the fact that powers of these matrices (suit-
ably scaled) cycle with period three, a property that depends on using the nonstandard
value for 0 (0).

Theorem 1. Defining J (a, b) as in Proposition 1, let B be the m m matrix with

entries J (a, b) for 0 a, b m 1. Then the powers of B satisfy
(i) B 2 = m B,
(ii) B 3 = m 3 I m 2 U ,
(iii) B n = m 3 B n3 for n 4,
where I is the m m identity matrix and U is the matrix all of whose entries are 1.

Proof. The first claim is equivalent to the assertion that

J (a, k)J (k, b) = m J (a, b) (10)

February 2012] JACOBI SUM MATRICES 103

for all a and b. Using definition (5) for J (a, b) we expand the left-hand side as

Xm1 X X
a (u) k (1 u) k (v) b (1 v) . (11)
k=0 u6 =0,1 v6=0,1

It is a standard opening gambit in these sorts of proofs to move the summation over k
k=0 (u) = 0 unless u = 1, in which case
to the inside and then use the fact that m1 k
the sum equals m. (It is this feature of characters that make them useful for counting
arguments.) Employing this strategy leads to

X X m1
a (u) b (1 v) k ((1 u)v). (12)
u6 =0,1 v6 =0,1 k=0

The final sum vanishes unless (1 u)v = 1, or u = 1 v1 . Hence our expression

reduces to
1 v 1
X   X    
m 1
(1 v) = m
, (13)
v6 =0,1
v v6 =0,1
v1 1v

where we have used a (u) = a ( u1 ). Next observe that J (a, b) = J (a, b) since
a = a ; we introduced negative exponents in anticipation of this fact. And now in
v 1
a beautiful stroke we realize that v1 + 1v = 1, and as v runs through all elements of
Fq other than 0 and 1 so does v1 . Hence the right-hand side of (13) is precisely the
sum defining J (a, b), thus proving the first part.
With this result in hand the second part will follow once we show B B = m 2 I
mU . This is equivalent to demonstrating that
m2 m

X a=b
J (a, k)J (k, b) = (14)
m a 6 = b.

The same ingredients are needed as above (but without negative exponents at the end),
so we omit the proof in favor of permitting the reader to supply the steps. There are no
major surprises along the way, and the explanation is quite satisfying.
The final claim is an immediate consequence of the second part. For n 4 we

B n = B n4 B(m 3 I m 2 U ) = m 3 B n3 , (15)

where we have used the fact that BU is the zero matrix because the entries within each
row of B sum to 0. This completes the proof.

Corollary 1. If is an eigenvalue of a Jacobi sum matrix B then {0, m, m, m},

where = e2i/3 .

Proof. Suppose that Bv = v for some nonzero vector v. Then multiplying B 4 =

m 3 B on the right by v yields 4 v = m 3 v. Therefore 4 = m 3 since v 6 = 0, which
implies that {0, m, m, m}.

Corollary 2. Every Jacobi sum matrix B has an orthogonal basis of eigenvectors.


Proof. Since B is symmetric its conjugate transpose B is just B. But B is a scalar
multiple of B 2 , so we deduce that B and B commute, and hence B is normal. The
assertion now follows from well-known properties of normal matrices as furnished
by [4], for instance.

The fact that the eigenspaces for = 7, 7, and 7 have the same dimension has
probably not escaped notice. In general the eigenspaces are always as close in size as
possible, a fact that depends upon ascertaining the traces of Jacobi sum matrices.

Proposition 2. The trace tr(B) of a Jacobi sum matrix B is equal to 0, m, or 2m

according to whether m 1 0, 1, or 2 mod 3.

Proof. The values occurring along the main diagonal of B are J (a, a). Hence
m1 m1 X m1
a (u u 2 ).
tr(B) = J (a, a) = (u) (1 u) =
a a
a=0 a=0 uFq uFq a=0

But the inner sum vanishes unless u u 2 = 1, in which case its value is m. When
Fq has characteristic 3 we find that u = 1 is a double root of the equation, while
for other characteristics u = 1 is not a root. Since (u 2 u + 1)(u + 1) = u 3 + 1, in
these cases we seek values of u with u 3 = 1 other than u = 1. For m 0 mod 3
there will be two such values, while for m 1 mod 3 there are no such values, because
Fq is a cyclic group of order m. In summary, u u 2 = 1 has one, two, or zero distinct
roots when m 2, 0, 1 mod 3, as claimed.

Proposition 3. The characteristic polynomial of a Jacobi sum matrix B has the form

p B (x) = x(x m)r (x m)s (x m)s (17)

for nonnegative integers r s satisfying r + 2s = m 1 with r as close to s as


Proof. Clearly p B (x) is monic. Furthermore, the sign will be positive unless m is odd;
i.e., when q = 2r . We will also see below that rank(B) = m 1, giving the single fac-
tor of x. Now let us show that p B (x) has real coefficients, meaning that the eigenval-
ues m and m occur in pairs. This follows from the relationship J (m a, m b) =
J (a, b). In other words, swapping rows a and m a as well as columns b and m b
for all a and b with 1 a, b m2 does not alter p B (x) but does conjugate every entry
of B, and therefore p B (x) is real. The trace of B is the sum of the eigenvalues, and
each triple m, m, m will cancel. The result now follows from the fact that tr(B) is
equal to 0, m, or 2m.

4. RELATED MATRICES. Observe that the list of eigenvalues for a Jacobi sum
matrix constructed using the conventional definition is nearly identical to the list given
by Proposition 3, the difference being that the eigenvalue = 0 is replaced by = 1
and a single occurrence of = m changes to = m + 1 = q. This is a consequence of
the close relationship in each case between the characteristic polynomial of the entire
matrix and that of the lower right (m 1) (m 1) submatrix of values they share.

Lemma 1. Let M be an n n matrix whose first column contains the entries c 1,

1, . . . , 1 for some number c, as shown below, and such that the sum of the entries in

February 2012] JACOBI SUM MATRICES 105

each row of M is 0. Let M 0 be the (n 1) (n 1) submatrix obtained by deleting the
first row and column of M. Then the list of eigenvalues of M 0 (with multiplicity) is given
by removing the values = 0, c from the list of eigenvalues for M and including = 1.


M = .. (18)

. M0


Proof. Multiplying the first column of M x I by (1 x) before taking the determi-

nant yields
(c 1 x)(1 x)

x 1

(1 x) det(M x I ) = .. . (19)

. 0
M xI 0

x 1

We next add columns 2 through n to the first column. Since the sum of the entries
within each row of M is zero this operation cancels every term in the first column
below the top entry, which becomes

(c 1 x)(1 x) + (1 c) = x(x c). (20)

Therefore the value of the determinant may be rewritten as

x(x c)

(1 x) det(M x I ) = .. = x(x c) det(M 0 x I 0 ). (21)

. M0 x I 0


The assertion follows.

If Jq (a, b) were computed in the traditional manner the top row of our Jacobi sum
matrix would be q followed by a row of 0s, so the list of eigenvalues would consist of
those of the lower right submatrix, augmented by the value = q. Invoking the lemma
now leads to the statement made above comparing lists of eigenvalues.
Purely to satisfy our curiosity, we now propose permuting the rows and columns
of a Jacobi sum matrix B before computing the eigenvalues. For example, take B to
equal matrix (1), let P be any 7 7 permutation matrix, and consider the degree-seven
polynomial det(PB x I ). Compiling the roots to all 5040 polynomials that arise in
this manner produces a list with somewhat more than 3500 distinct complex numbers;
locating them in the complex plane yields the scatterplot on the left in Figure 1. The
roots, whose locations are marked by small solid discs, form a nearly unbroken chain
along the circle of radius 7 centered at the origin, with discernible gaps located only
near the real axis. By way of comparison, the related 7 7 matrix B of conventional
Jacobi sum values yields the right-hand plot in Figure 1. Put another way, matrix B
generates in excess of 3500 algebraic integers, each of degree 14 or less over Q and
each having absolute value 7. As one might hope, this property is shared by all Jacobi
sum matrices. The following result was conjectured by the author and proved by Ron
Evans (personal communication, Jan. 2011); we present this proof below.


6 6

4 4

2 2

6 4 2 2 4 6 6 4 2 2 4 6
2 2

4 4

6 6

Figure 1. A plot of the roots of det(P B x I ) on the left and the roots of det(P B x I ) on the right for all
7 7 permutation matrices P.

Proposition 4. Let B denote a Jacobi sum matrix for the finite field Fq and let P be
any m m permutation matrix, where m = q 1. Then every nonzero eigenvalue
of the matrix PB satisfies || = m.

Proof. Let be a nonzero eigenvalue of PB, so that PBv = v for some nonzero
vector v. Letting M denote the conjugate transpose of a matrix M, it follows that
(PBv) (PBv) = (v) (v). Expanding yields v B P PBv = ||2 v v, which implies

v m1 B 3 v = ||2 v v, (22)

since P P = I for any permutation matrix and B = B = m1 B 2 using Theorem 1

and the fact that B is symmetric. Appealling once more to Theorem 1, we find that
1 3
B = m 2 I mU , where every entry of U equals 1. Next observe that U v = 0, since
multiplying PBv = v on the left by U gives UPBv = U v, and UPB = UB = 0
while 6= 0. Therefore (22) becomes

v (m 2 I mU )v = m 2 v v mv U v = m 2 v v = ||2 v v. (23)

But v v > 0 since v is a nonzero vector, and hence || = m, as desired.

5. DETERMINANTS. One of the more striking properties of Jacobi sum matrices

emerges once we begin to examine submatrices and their determinants, in particular.
Thus the alert reader may have wondered about the determinant of (1). Since the sum
of the entries in each row is zero, it is clear that det(B) = 0 for any Jacobi sum matrix.
Not content, the truly enterprising individual next computes det(B 0 ) for the lower right
6 6 submatrix B 0 of matrix (1), obtaining the intriguing value det(B 0 ) = 16807 = 75 .
The obvious generalization is true, and the groundwork for a proof has largely been
laid. We need only one further observation, which is a nice result in its own right.

Proposition 5. Let B denote a Jacobi sum matrix with lower right (m 1) (m 1)

submatrix B 0 . Then (B 0 )1 = m12 (B 0 + (m + 1)U 0 ), where every entry of U 0 is 1.

Proof. The statement follows readily from the equality B B = m 2 I mU stated in

Theorem 1. We omit the details.

February 2012] JACOBI SUM MATRICES 107

Proposition 6. With B and B 0 as above, we have det(B 0 ) = m m2 .

Proof. According to Corollary 1 the eigenvalues of B belong to the set {0, m, m, m}.
We know det(B) = 0, so rank(B) < m. But by the previous lemma B 0 is nonsingular;
therefore rank(B) = m 1, implying that exactly one eigenvalue of B is 0. We next
apply Lemma 1 to conclude that the eigenvalues of B 0 are among {1, m, m, m},
with the value 1 occurring precisely once. Finally, the discussion within Proposition 3
indicates that the values m and m come in pairs. Hence the product of the m 1
eigenvalues, which is det(B 0 ), comes to m m2 .

Corollary 3. Let A be the submatrix of a Jacobi sum matrix B obtained by deleting

row i and column j. Then we have det(A) = (1)i+ j m m2 .

Proof. The case i = j = 0 is handled by Proposition 6. When j > 0 note that adding
all other columns of B 0 to column j effectively replaces that column with the negative
of column 0 of B, since the sum of the entries within every row is 0. Moving this
column back to the far left and negating it introduces a sign of (1) j to the value of
the determinant. The same reasoning applies to the rows; therefore B 0 is transformed
into A by operations that change the sign of det(B 0 ) by (1)i+ j .

But why stop there?

If A is the lower right 5 5 submatrix of (1), we discover that
det(A) = 343(7 i 7). The power of 7 is nice, but even more interesting is

7 i 7 = J (0, 0) J (0, 1) J (1, 0) + J (1, 1). (24)

In other words, the determinant of this submatrix appears to be related to the con-
jugates of the entries in the complementary upper left 2 2 submatrix. The same
phenomenon occurs elsewhere; for instance, if A is the upper left 5 5 submatrix
of (1) then we find that det(A) = 343(7 + 3i 7), and sure enough

7 + 3i 7 = J (5, 5) J (5, 6) J (6, 5) + J (6, 6). (25)

These computations hint at a beautiful extension to Corollary 3. We first formalize a

few of the above ideas.
A k k submatrix A is determined by a subset r1 , . . . , rk of the rows of B, where
0 r1 < < rk m 1, and a similar subset c1 , . . . , ck of k columns. Deleting
these rows and columns yields the complementary submatrix Ac , which contains ex-
actly those entries of B that are not in the same row or column as any element of A. The
sign of the submatrix, denoted by  A , is based on its position within B. It is given by

 A = (1)r1 ++rk +c1 ++ck . (26)

It is routine to verify that  A =  Ac . Finally, the diminished determinant ddet(A) of A

is an alternating sum of the determinants of all maximal submatrices of A. Letting Aij
represent the matrix obtained by deleting row i and column j of A we have
ddet(A) = (1)i+ j det(Aij ) =  A0 det(A0 ), (27)
i, j=1 A0 A

where A0 A signifies a (k 1) (k 1) submatrix of A. We have chosen the term

diminished since the degree of ddet(A) as a polynomial in the entries of A is one
less than the degree of det(A).


So that the upcoming result will apply to all possible submatrices of B, we adopt
the convention that ddet(A) = 1 for a 1 1 matrix A, while ddet(A) = 0, det(A) = 1,
and  A = 1 when A is the 0 0 empty matrix. With the foregoing definitions in
hand we are now prepared to state our main result.

Theorem 2. Given a Jacobi sum matrix B, let A be any k k submatrix of B, where

0 k m. Denote the complementary submatrix to A and its sign by Ac and  Ac ,
respectively. Then the following identity holds:

det(A) ddet(Ac )
=  A c . (28)
mk m mk

Observe that the power of m in each denominator corresponds to the size of the matrix
in the numerator. Also, the examples outlined above illustrate the case m = 7, k = 5;
in both examples the sign happened to be  Ac = 1. We provide a proof of this result in
the appendix. The reader is encouraged to peruse the argumentamong other things,
a number of steps would make excellent exercises for linear algebra students.
Before considering a collection of multivariable polynomials with unlikely factor-
izations, we pause to present a couple of elementary facts concerning the diminished
determinant, which arose naturally in the preceding discussion. Early in the proof of
Theorem 2 we will need an analogue to expansion by minors to handle the transition
between diminished determinants for matrices of different sizes. To clarify the analogy,
let M be an n n matrix with entries m i j and let M ij denote the submatrix obtained
by deleting row i and column j from M. Then expansion by minors implies that

1 X
det(M) = (1)i+ j m i j det(M ij ). (29)
n i, j=1

Lemma 2. With M and M ij as above we have

1 X
ddet(M) = (1)i+ j m i j ddet(M ij ). (30)
n 1 i, j=1

Proof. Applying (29) to the definition of ddet(M) yields

ddet(M) = (1)k+l det(Mlk )
X 1 X 0 0
= (1)k+l (1)i + j m i j det(M ik
jl ). (31)
n 1 i6=k

Here M ik jl is the submatrix of M obtained by deleting rows i, k and columns j, l. Note

that if row i is below row k then we must use i 1 in the exponent when applying (29)
to det(Mlk ); otherwise i is the correct value. Hence we set i 0 = i 1 when i > k and
i 0 = i when i < k, and similarly for j 0 relative to l.
0 0
The key to ensuring that the signs behave is to realize that (1)i +k = (1)i+k +1 ,
where k 0 = k 1 when k > i and k 0 = k otherwise. Defining l 0 in the same manner

February 2012] JACOBI SUM MATRICES 109

relative to j enables us to rewrite (31) as

1 X X 0 0
ddet(M) = (1)i+ j m i j (1)k +l det(M ik
jl ) (32)
n 1 i, j k6=i
l6= j
1 X
= (1)i+ j m i j ddet(M ij ).
n 1 i, j=1

This completes the proof.

Diminished determinants also resemble determinants with respect to row and col-
umn transpositions.

Lemma 3. Interchanging a pair of adjacent rows or columns in a matrix M negates

the value of ddet(M).

Proof. Every term in the sum (1)i+ j det(M ij ) is negated by such an operation, for
one of two reasons. If column i and row j stay put then a pair of rows or columns
within M ij trade places, negating det(M ij ) without affecting (1)i+ j . On the other
hand, if column i or row j is involved in the exchange then M ij still appears in the
sum with entries intact, but now with an attached sign of (1)i+ j1 .

6. FURTHER INQUIRY. To conclude we offer an observation regarding Jacobi sum

matrices that suggests there is still gold left to be mined. Define the three permutation
1000000 1000000 1000000

0 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 1 0 0 0 0
0 0 1 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0
P1 = 0 0 0 1 0 0 0 , P2 = 0 0 0 0 0 1 0 , P4 = 0 0 0 0 0 0 1 .

0 0 0 0 1 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0
0000010 0000001 0001000

0000001 0001000 0000010

In each case the 1s are situated along a line through the origin, where the origin is
the upper left entry and we reduce coordinates mod 7; the subscript indicates the slope
of the line. We have already observed that the characteristic polynomial of matrix (1),
which we shall denote as B once again, splits completely over the field Q():

p B (x) = det(B x P1 ) = x(x 7)2 (x 7)2 (x 7)2 . (34)

Remarkably, much more is true:

det(B x P1 y P2 z P4 ) = (x + y + z)(x + y + z 7)2 (35)

(x + y + z 7)(x + y + z 7)
(x + y + z 7)(x + y + z 7).

Further experimentation suggests that it is not a coincidence that the slopes used for P1 ,
P2 , and P4 are powers of 2. For instance, det(B w P1 x P2 y P4 z P8 ) splits into


linear and quadratic factors, where B is the Jacobi sum matrix for F16 and all matrices
are 15 15 in size. This phenomenon persists for finite fields of odd characteristic as
well. Thus when working over F9 we find that det(B x P1 y P3 ) splits completely
over Q(). We also point out the related beautiful factorization

det(B x P5 y P7 ) = (x + y)(x + y + 8)(x y + 8)(x y 8)2 (x + y 8)3 .

Based on these observations we surmise the following.

Conjecture 1. Let B be a Jacobi sum matrix for the finite field Fq , where q = pr and
m = q 1. For (k, m) = 1 denote by Pk the m m permutation matrix whose entry
in row s, column t is 1 for all 0 s, t < m with s kt mod m. Then the polynomial

det(B x0 P1 x1 Pp x2 Pp2 xr 1 Ppr 1 ) (37)

in the r variables x0 , x1 , . . . , xr 1 may be written as a product of factors each of which

has degree at most two in these variables.

Other evidence that we have not included here suggests that this conjecture can be
extended in scope.
In summary, we have examined an elegant tool from number theory via the lens
of linear algebra and uncovered several nice results in the process. At the very least
this approach demonstrates a tidy manner in which many of the elementary (though
perhaps not fully mapped out) facts concerning Jacobi sums may be packaged. On an
optimistic note, this avenue of inquiry may even lead to a more complete understanding
of Jacobi sums.

7. APPENDIX. Our main result relates the determinant of a submatrix of a Jacobi

sum matrix to the diminished determinant of the conjugate complementary submatrix.

Theorem 2. Given a Jacobi sum matrix B, let A be any k k submatrix of B, where

0 k m. Denote the complementary submatrix to A and its sign by Ac and  Ac ,
respectively. Then the following identity holds:

det(A) ddet(Ac )
=  Ac . (38)
m m mk
Proof. For k = 0 and k = m the statement to be proved reduces to

ddet(B) det(B)
1= , = 0. (39)
mm mm
The former is a consequence of Corollary 3, while the latter is clear. Furthermore,
the statement for k = m 1 is equivalent to Corollary 3. Hence we need only show
that case k follows from case k + 1 for 1 k m 2. In the interest of present-
ing a lucid argument, we will provide a sketch of the proof in the case k = m 3,
followed by a summary of the algebra for the general case, which is qualitatively no
Therefore suppose the result holds for k = m 2 and that A is an (m 3) (m
3) submatrix of B. For the sake of organization we permute the rows and columns of B
in order to situate the entries of Ac in the upper left corner, but otherwise maintain the

February 2012] JACOBI SUM MATRICES 111

original order of the rows and columns within A and Ac . Let us label the permuted
matrix as C, having entries i j for 0 i, j m 1.

00 01 02 03 04

10 11 12 13 14

21 22 23 24

C =
30 31 32

40 41 42



We claim that the result for k = m 2 continues to hold for matrix C, up to a sign
which we now determine. For exchanging a pair of adjacent rows or columns of B
will negate exactly one of det(A), ddet(Ac ) or  Ac , according as the pair of rows or
columns both intersect A, both intersect Ac (by Lemma 3), or intersect both. If the
entries of Ac reside in rows r1 , r2 , r3 and columns c1 , c2 , c3 then it requires

r1 + (r2 1) + (r3 2) + c1 + (c2 1) + (c3 2) (41)

swaps of adjacent rows or columns to transform B into C; hence we must include a

factor of  Ac when applying (28) to C. In other words, if D is an (m 2) (m 2)
submatrix of C then

det(D) ddet(D c )
 Ac =  D c . (42)
m m2 m2
The final observation to be made before embarking upon a grand calculation is
that the dot product of any row vector of C with the conjugate of another row vector
is m, while the dot product of a row vector with its own conjugate is m 2 m. This
relationship holds for B since B is symmetric and B B = m 2 I mU , as noted in the
proof of Theorem 1. Permuting rows and columns of B does not destroy this property,
which consequently holds for C as well. Now to begin.
We wish to relate ddet(Ac ) to det(A). By Lemma 2 we may begin

ddet(Ac ) = 00 ddet 11 12 01 ddet 10 12 +
2 21 22 20 22

12 12 12

= 00 det(C 12 ) + 01 det(C 02 ) + 02 det(C 01 ) + , (43)
2m m4

since the result holds for k = m 2 with a correction factor of  Ac . As before C ikjl
denotes the submatrix of C obtained by deleting rows i, k and columns j, l. We then
12 12 12
expand each of det(C12 ), det(C02 ), and det(C01 ) by minors along row 0, which gives
( 00 00 + 01 01 + 02 02 ) det(A), along with a fair number of other terms. We next
collect the remaining terms according to whether they involve 03 , 04 , 05 , and so on.
The reader may verify that the sum of the terms containing a factor of 03 is

03 03 det(A) + m det(A03 F3 ), (44)

where A03 F3 refers to matrix A with all entries in the left column replaced by 03 .
Combining the terms involving 04 , 05 , . . . in the same manner, we may rewrite the


first three terms of (43) as
(00 00 + 01 01 + 02 02 + 03 03 + 04 04 + 05 05 + ) det(A)
+ m det(A03 F3 ) + m det(A04 F4 ) + m det(A05 F5 ) + . (45)
The coefficient of det(A) is the dot product of row 0 of C with its own conjugate, so
this simplifies to
(m 2 m) det(A) + m det(A03 F3 ) + m det(A04 F4 ) + m det(A05 F5 ) + . (46)
Finally, this entire sequence of steps may be performed on the second trio and third
trio of terms in (43), yielding
3(m 2 m) det(A) + m(det(A03 F3 ) + det(A04 F4 ) + det(A05 F5 ) + )
+ m(det(A13 F3 ) + det(A14 F4 ) + det(A15 F5 ) + )
+ m(det(A23 F3 ) + det(A24 F4 ) + det(A25 F5 ) + ). (47)
Since the sum of the entries of any column of C is 0, we have

det(A03 F3 ) + det(A13 F3 ) + det(A23 F3 ) = det( A3 ), (48)

where A3 represents matrix A with each entry in its leftmost column replaced by the
sum of all the entries in that column. Defining A j similarly for j 3, (47) reduces to

3(m 2 m) det(A) m(det( A3 ) + det( A4 ) + det( A5 ) + ). (49)

It is a neat exercise in linear algebra to confirm that det( A j ) = (m 3) det(A).

Each term of det(A) appears in det( A j ) for every j and hence appears m 3 times in
the sum; all other terms cancel in pairs, as the reader may verify. Hence we are left with
3(m 2 m) det(A) m(m 3) det(A) = 2m 2 det(A). (50)
In summary, we have shown that
 Ac det(A)
ddet(Ac ) = (2m 2 det(A)) =  Ac m6 . (51)
2m m4 m
Dividing through by  Ac m 3 gives the desired equality.
The calculation proceeds in an identical fashion for other values of k. One arrives
at the expression
k(m 2 m) det(A) m(m k) det(A) = (k 1)m 2 det(A) (52)
in place of (50), yielding
 Ac det(A)
ddet(Ac ) = ((k 1)m 2 det(A)) =  Ac m2k . (53)
(k 1)m m2k+2 m
Rearranging gives the result.

ACKNOWLEDGMENTS. I would like to thank the referees for many helpful remarks and suggestions. In
particular, the idea of generating the right-hand scatterplot in the figure as well as the insightful remarks
contained in the footnote were both due to the referees. I am also grateful to Ron Evans for sharing the (quite
rapidly found) proof appearing in this article.

February 2012] JACOBI SUM MATRICES 113


1. L. Adleman, C. Pomerance, R. Rumely, On distinguishing prime numbers from composite numbers, Ann.
of Math. 117 (1983) 173206; available at
2. B. C. Berndt, R. J. Evans, K. S. Williams, Gauss and Jacobi Sums. Wiley, New York, 1998.
3. G. Eisenstein, Einfacher beweis und verallgemeinerung des fundamental theorems fur die biquadratischen
reste, in Mathematische Werke, Band I, 223245, Chelsea, New York, 1975.
4. R. A. Horn, C. R. Johnson, Matrix Analysis. Cambridge University Press, Cambridge, 1985.
5. K. Ireland, M. Rosen, A Classical Introduction to Modern Number Theory, second edition. Springer, New
York, 1990.
6. A. Weil, Number of solutions of equations in a finite field, Bull. Amer. Math. Soc. 55 (1949) 497508;
available at

SAMUEL K. VANDERVELDE is an assistant professor of mathematics at St. Lawrence University. His

mathematical interests include number theory, graph theory, and partitions. He is an enthusiastic promoter
of mathematicshe conducts math circles for students of all ages, helped to found the Stanford Math Circle
and the Teachers Circle, and composes problems for the USA Math Olympiad. He also writes and coordinates
the Mandelbrot Competition, a nationwide contest for high schools. He is an active member of his church and
enjoys singing, hiking, and teaching his boys to program.
Department of Mathematics, Computer Science, and Statistics, St. Lawrence University, Canton, NY 13617


Alcuins Sequence
Donald J. Bindner and Martin Erickson

Abstract. Alcuin of York (c. 740804) lived over four hundred years before Fibonacci. Like
Fibonacci, Alcuin has a sequence of integers named after him. Although not as well known
as the Fibonacci sequence, Alcuins sequence has several interesting properties. The purposes
of this note are to acquaint the reader with Alcuins sequence, to give the simplest available
proofs of various formulas for Alcuins sequence, and to showcase a new discovery about the
period of Alcuins sequence modulo a fixed integer.

1. INTRODUCTION. A famous problem posed by Alcuin of York gives rise to an

integer sequence not unlike the famous Fibonacci sequence, though Alcuins problem
predates Fibonacci by about 400 years. We will describe and prove some basic prop-
erties of Alcuins sequence and demonstrate a new property.
Little is known about the background and early life of Alcuin of York (c. 740804).
Apparently, he was born in Northumbria in what is now England. He was a scholar
and teacher who became instrumental in Charlemagnes court. He is noted for his
letters, poetry, and other writings, and his fostering of an educational system focused
on humanitarianism and religious teaching.
Alcuins collection of mathematical story problems, Propositiones ad acuendos ju-
venes (Problems to Sharpen the Young) contains 53 problems that can be classified
as puzzles or recreational mathematics. A notable example is the problem of the goat,
wolf, and cabbage, where a man must transport all three across a river but his boat
will only carry himself and one of the three at a time. The problem that interests us is
number 12 in Alcuins list.

A certain father died and left as an inheritance to his three sons 30 glass flasks,
of which 10 were full of oil, another 10 were half full, while another 10 were
empty. Divide the oil and flasks so that an equal share of the commodities should
equally come down to the three sons, both of oil and glass [4].

There are five solutions, of which Alcuin gives only the first:

Solution 1 Solution 2 Solution 3 Solution 4 Solution 5

Son 1 5 5 0 5 5 0 5 5 0 4 4 2 4 4 2
Son 2 5 5 0 4 4 2 3 3 4 4 4 2 3 3 4
Son 3 0 0 10 1 1 8 2 2 6 2 2 6 3 3 4

The numbers of full, empty, and half-full flasks are represented by the columns F,
E, and H, respectively. We dont regard solutions with the sons permuted as distinct.
Notice that in each solution, each son receives an equal number of full and empty
MSC: Primary 11B50

February 2012] ALCUINS SEQUENCE 115

flasks, and half-full flasks are used to make a total of ten flasks. Hence, the number of
full flasks that a son receives completely determines his share.
The problem may be generalized to any number 3n of flasks, n full, n empty, and
n half full, to be distributed among three sons. The numbers of full flasks for each
son form a triple of nonnegative integers (a, b, c), with a b c and a + b + c = n,
satisfying the weak triangle inequality: a + b c. The reason is that a, b, c n/2,
since the number of full flasks must be balanced by an equal number of empty flasks.
Because the integer triple (a, b, c) satisfies the weak triangle inequality if and only if
the triple (a + 1, b + 1, c + 1) satisfies the triangle inequality, the number of solutions
to the flask-sharing problem is the same as the number of incongruent triangles with
integer sides and perimeter n + 3. A triangle with integer sides is called an integer
triangle. For n 0, let t (n) be the number of incongruent integer triangles of perimeter
n. The sequence {t (n)} is called Alcuins sequence, after a tradition that seems to have
started in [10]. The first few terms are

0, 0, 0, 1, 0, 1, 1, 2, 1, 3, 2, 4, 3, 5, 4, 7, 5, 8, 7, 10, 8, 12, . . . .

2. FORMULAS FOR ALCUINS SEQUENCE. When we have an intriguing se-

quence, it is natural to want to find a direct formula for it as well as a recurrence
relation. Formulas for {t (n)}, the number of incongruent integer triangles of perimeter
n, have been derived many times (see, e.g., [1], [5], [6], [7], [8]). We give a streamlined
version of the derivation in [11]. The key idea is to relate integer triangles to partitions
of integers into three parts.

Theorem 1. Let kxk be the nearest integer to x. Then for all n 0,

48 if n is even,
t (n) =

if n is odd.

(There is no ambiguity because n 2 /48 is never a half-integer.)

Proof. Let pk (n) denote the number of partitions of the nonnegative integer n into
k positive integer parts (summands). The order of the parts is unimportant, and for
uniformity we list them in nonincreasing order. For example, p3 (5) = 2, and the two
relevant partitions are 3 + 1 + 1 and 2 + 2 + 1.
The number of integer triangles of even perimeter is equal to the number of parti-
tions of half the perimeter into three parts:

t (2n) = p3 (n), n 0.

This is seen via the bijection

(a, b, c) {n a, n b, n c},

where (a, b, c) represents the sides of an integer triangle of perimeter 2n and {n a,

n b, n c} represents the summands in a partition of n into three parts. The point is
that a, b, and c satisfy the triangle inequality if and only if each of these quantities is
less than n, and this occurs if and only if the three parts n a, n b, and n c are


The sequence { p3 (n)} satisfies the recurrence relation

p3 (n + 6) = p3 (n) + n + 3, n 0.

To prove this, consider the last summand in a partition of n + 6 into three parts. If this
summand is at least 3, then subtract 2 from each summand to obtain a partition of n
into three parts. If the last summand is a 2, then subtract 2 from each part to obtain a
partition of n into one or two parts. If the last summand is a 1, then subtract 1 from each
part to obtain a partition of n + 3 into one or two parts. Using the obvious formulas
p1 (n) = 1 and p2 (n) = bn/2c, where bxc is the greatest integer less than or equal to
x, we obtain
 j n k  

p3 (n + 6) = p3 (n) + 1 + + 1+ , n 0,
2 2

which simplifies to our desired recurrence relation by examining the cases n even and
n odd.
Initial values of the sequence { p3 (n)} can be obtained from the relation p3 (n) =
t (2n) or from scratch:

0, 0, 0, 1, 1, 2, . . . .

We will prove the formula

12 ,
p3 (n) = n 0,

where kxk is the nearest integer to x. (It is easy to show that n 2 /12 is never a half-
integer.) This formula for p3 (n) has the correct six initial values, and it satisfies the
recurrence relation:
(n + 6)2 n 2

12 = 12 + n + 3, n 0.

Hence, the formula holds for all n 0 by mathematical induction.

We have shown that for n even,
48 .
t (n) = p3 (n/2) =

Via the bijection

(a, b, c) (a + 1, b + 1, c + 1),

where (a, b, c) represents an integer triangle with odd perimeter n, we have

t (n) = t (n + 3), for n odd.

From the above formula, we notice a couple of interesting features of Alcuins se-
quence. The only fixed point is t (48) = 48. Furthermore, the sequence has the zig-zag
property that

t (2n) < t (2n + 1) > t (2n + 2), n 3.

February 2012] ALCUINS SEQUENCE 117

The ordinary generating function for a sequence {a(n)} is the power series

a(n)x n .

Like the Fibonacci sequence, Alcuins sequence has a rational generating function.

Theorem 2. Alcuins sequence {t (n)} has the generating function

X x3
t (n)x n = .
(1 x 2 )(1 x 3 )(1 x 4 )

Proof. We claim first that the triples (a, b, c) that represent integer triangles (i.e.,
triples of positive integers satisfying a b c < a + b) are precisely the triples of
the form

(a, b, c) = (1, 1, 1) + (0, 1, 1) + (1, 1, 1) + (1, 1, 2),

where , , and are nonnegative integers. Furthermore, the choice of , , and in

this representation is unique. It is easy to see that every triple of this form represents
an integer triangle. And given a triple (a, b, c) that represents an integer triangle, the
formulas = b a, = a + b c 1, = c b give the unique coefficients , ,
and .
Since 2 + 3 + 4 = a + b + c 3 = n 3, it follows that t (n) is equal to the
number of partitions of n 3 as a sum of 2s, 3s, and 4s. Our claimed generating
function may be written as

2 3
(x 4 ) =
(x ) (x ) x 2+3+4 +3 ,
=0 =0 =0 ,, 0

and each way of writing n 3 as a sum of 2s, 3s, and 4s leads to an x n term in the
sum on the right.

The denominator of the rational generating function yields an order-nine linear re-
currence relation for Alcuins sequence.

Theorem 3. Alcuins sequence is determined by the recurrence relation t (n) =

t (n 2) + t (n 3) + t (n 4) t (n 5) t (n 6) t (n 7) + t (n 9), for
n 9, together with the values of t (n) for 0 n 8.

Proof. From the generating function, we have

x 3 = (1 x 2 )(1 x 3 )(1 x 4 ) t (n)x n

= t (n)(x n x n+2 x n+3 x n+4 + x n+5 + x n+6 + x n+7 x n+9 ).

Now the order-nine recurrence relation can be read off by equating coefficients of x n
for n 9.


This recurrence relation is palindromic, so that if we use it to define {t (n)} for
n < 0, then the sequence is the same forwards and backwards: t (n) = t (n 3).

3. THE PERIOD OF ALCUINS SEQUENCE. An integer sequence {a(n)} is pe-

riodic if there exists a positive integer L such that a(n + L) = a(n) for all n. The least
such L is called the period of the sequence. Every integer recurrent sequence modulo a
fixed integer is periodic, where the period is the least common multiple of the periods
with respect to the prime powers which divide the modulus.
The period of the Fibonacci sequence with respect to prime power moduli is un-
known in general, except when the modulus is a power of 2 or 5 (see [12]). The period
of Alcuins sequence was first given in [2].

Theorem 4. For any integer m 2, the sequence {t (n) mod m} is periodic with period
12m. Moreover, the range of the sequence consists of all integers modulo m if and only
if m is one of the following:

7, 10, 19, 2 j , 3 j , 5 j , 11 j , 13 j , 41 j , 2 3 j , 5 3 j , for j 1.

This collection parallels a similar family of moduli for which the Fibonacci se-
quence contains all residues [3, p. 318].

Proof. For even values of the argument, we have

t (2n + 12m) = k(2n + 12m)2 /48k = k(2n)2 /48k + nm + 3m 2 t (2n) (mod m).

Since t (n) = t (n + 3) for n odd, the sequence {t (n) mod m} is periodic with period
L 12m.
Now we will show that L 12m. Let = L or L 1, so that is even. We then
have k2 /48k 0 (mod m) and k( + 2)2 /48k 0 (mod m), since t (1) = t (0) =
t (1) = t (2) = 0. Hence m divides k( + 2)2 /48k k2 /48k, which is nonzero be-
cause L > 12 (since {t (n)} is an aperiodic pattern of 0s and 1s for 4 n 8). This
difference is less than ( + 2)2 /48 2 /48 + 1 = ( + 13)/12, and it follows that
12m < + 13 L + 13. By definition of period, L is a divisor of 12m. If L < 12m,
then L 6m, but this contradicts the inequality 12m < L + 13, for m 3. The re-
sult is checked by inspection for m = 2. We conclude that L 12m, and therefore
L = 12m.
The question of when the range of {t (n) mod m} contains all integers modulo m re-
duces to a question of when a certain set of polynomials represents all integers modulo
m. Our proof will use two venerable tools of number theory, namely Hensels lemma
and quadratic nonresidues.
Letting n = 12k + r , where 0 r < 12, we find from Theorem 1 that the values of
t (n) are given by six quadratic polynomials:

3k 2 , 3k 2 + k, 3k 2 + 2k, 3k 2 + 3k + 1, 3k 2 + 4k + 1, 3k 2 + 5k + 2.

By the substitution k 7 (k + 1), the last two polynomials may be eliminated as

redundant. Now the problem is to show that the range of the first four polynomials,
where k is an integer modulo m, consists of all m residues if and only if m is as stated.
To show that these values of m have the required property, the solitary cases 7, 10,
and 19 are checked. For the remaining values, we will use a version of Hensels lemma
that is slightly more general than the version in [9]. If f (x) is a polynomial with in-
teger coefficients, p is a prime, f (a) 0 (mod kpi ), and f 0 (a) 6 0 (mod p), then

February 2012] ALCUINS SEQUENCE 119

there exists a unique integer t such that f (a + tkpi ) 0 (mod kpi+1 ). We check that
for m equal to each of the base primes 2, 3, 5, 11, 13, and 41, every integer mod-
ulo m is equal to f (x) for some x modulo m, where f is one of the polynomials and
the derivative f 0 (x) is nonzero modulo m. By induction, the solution to each congru-
ence can be lifted to a solution modulo 2 j , 3 j , 5 j , etc., for j > 1. A computer search
establishes that there are such solutions for each of the base primes.
Similarly, solutions modulo 6 and modulo 15 can be lifted to solutions modulo 2 3 j
and 5 3 j , for j > 1, respectively. This entails checking that every integer modulo 6 is
congruent to f (x) for some x with f 0 (x) 6 0 (mod 3), and every integer modulo 15
is congruent to f (x) for some x with f 0 (x) 6 0 (mod 3). This shows that the stated
values of m have the required property.
To show that the only valid moduli m are those on our list, we begin by showing
that if the prime p is sufficiently large, then the four polynomials do not represent
every integer modulo p; that is, some integer modulo p is not represented by any of
the polynomials. We complete the squares of the polynomials and observe that if the
integer s is represented, then 12s is of at least one of the following forms:

(6k)2 , (6k + 1)2 1, (6k + 2)2 4, (6k + 3)2 + 3.

Hence, at least one of

12s, 12s + 1, 12s + 4, 12s 3

must be a square modulo p. However, by a theorem of Andre Weil (see, e.g., [9]),
there exist arbitrarily many consecutive quadratic nonresidues modulo p, if p is large
enough. The number N of sequences of h consecutive quadratic nonresidues modulo
p satisfies
N 3h p.


For h = 8, if N > 0 then we are guaranteed a sequence of eight consecutive quadratic

nonresidues modulo p. Any such sequence includes four quadratic nonresidues which
yield a value of s not represented by the four polynomials. Thus, we check primes up
to 37,748,736 (in order to exclude all primes other than our base primes), and this is
done in a matter of seconds by computer.
Finally, we must eliminate all composite numbers except those in our list. If m is
a valid modulus, then m 0 is a valid modulus for any divisor m 0 of m. So it remains
to check that 72 and 192 do not work, nor does any modulus that is a product of two
primes in our list, except for the products 2 5, 2 3, and 3 5. We must also rule out
22 5, 2 52 , 3 52 , 22 3, and 2 3 5. All of this is done instantly by computer.

ACKNOWLEDGMENTS. The authors wish to thank the referees for valuable suggestions that improved the
motivation and clarity of this paper.


1. E. G. Andrews, A note on partitions and triangles with integer sides, Amer. Math. Monthly 86 (1979)
477478; available at
2. M. Erickson, Aha! Solutions, Mathematical Association of America, Washington, DC, 2009.
3. R. L. Graham, D. E. Knuth, O. Patashnik, Concrete Mathematics: A Foundation for Computer Science,
second edition, Addison-Wesley, Reading, MA, 1994.


4. J. Hadley, D. Singmaster, Problems to Sharpen the Young, An annotated translation of Propsitiones ad
acuendos juvenes, the oldest mathematical collection in Latin, attributed to Alcuin of York, Math. Gaz.
76 (1992) 102126; available at
5. D. M. Hirschorn, Triangles with integer sides, Math. Mag. 76 (2003) 306308; available at http://dx.
6. T. Jenkyns, E. Muller, Triangular triples from ceilings to floors, Amer. Math. Monthly 107 (2000) 634
639; available at
7. R. H. Jordan, R. Walch, R. J. Wisner, Triangles with integer sides, Amer. Math. Monthly 86 (1979) 686
689; available at
8. N. Krier, B. Manvel, Counting integer triangles, Math. Mag. 71 (1998) 291295; available at http:
9. I. Niven, H. Zuckerman, H. Montgomery, An Introduction to the Theory of Numbers, fifth edition, Wiley,
New York, 1991.
10. D. Olivastro, Ancient Puzzle: Classic Brainteasers and Other Timeless Mathematical Games of the Last
10 Centuries, Bantam, New York, 1993.
11. J. Tanton, Young students approach integer triangles, FOCUS 22 (2002) no 5, 46.
12. D. D. Wall, Fibonacci series modulo m, Amer. Math. Monthly 67 (1960) 525532; available at http:

DONALD J. BINDNER received his B.S. from Truman State University in 1992 and his Ph.D. from the
University of GeorgiaAthens in 2001. His interests include computer programming and free software. He and
Martin recently wrote the book A Students Guide to the Study, Practice, and Tools of Modern Mathematics
(CRC Press).
Department of Mathematics and Computer Science, Truman State University, Kirksville, MO 63501

MARTIN ERICKSON received his B.S. and M.S. from the University of Michigan in 1985 and his Ph.D.
from the University of Michigan in 1987. His mathematical interests are combinatorics, number theory, and
problem solving.
Department of Mathematics and Computer Science, Truman State University, Kirksville, MO 63501

February 2012] ALCUINS SEQUENCE 121

A Case of Continuous Hangover
Burkard Polster, Marty Ross, and David Treeby

Abstract. We consider a continuous analogue of the classic problem of stacking identical

bricks to construct a tower of maximal overhang.

1. INTRODUCTION. How much of an overhang can we produce by stacking iden-

tical rectangular blocks at the edge of a table? Most mathematicians know that the
overhang can be as large as desired: we arrange the blocks in the form of a staircase as
shown in Figure 1. This stack will (just) fail to topple over, and with n blocks of length
2 the overhang sums to

1 1 1 1
1+ + + + + .
2 3 4 n
Since the harmonic series diverges, it follows that the overhang can be arranged to be
as large as desired, simply by using a suitably large number of blocks.

table ... 1 1 1 1 1
65 4 3 2

1 1 1
Figure 1. The total overhang of this tower of twenty blocks is 1 + 2 + 3 + + 20 .

In practice, these special stacks are constructed from top to bottom: the top block
is placed so that its middle, balancing point is at the upper right corner of the second
block. Then the top two blocks are placed together so that their combined balancing
point is at the upper right corner of the third block, and so on.1
Okay, review is over; now for something new. Lets rescale our special stacks in
the vertical direction, so that each stack has height 1; the resulting stacks resemble
decks of playing cards, as indicated in Figure 2. Well call these stacks the harmonic
MSC: Primary 26A06
1 Recently, stacks have been investigated for which it is permitted to place two blocks upon any lower

block. Stacking in this way, one can use some blocks as counterweights and thus achieve significantly greater
overhangs than with the staircases. See [5] for these recent results and a comprehensive bibliography for the
stacking problem.



table x

y = 1 e x


Figure 2. The harmonic staircases converge to the harmonic stack.

staircases. Notice that weve arranged for the top right corner of the table to coincide
with the origin of our coordinate system.
We will prove (Theorem 2.1) that the sequence of harmonic staircases converges to
the harmonic stack, determined by the function 1 ex . And, just like the harmonic
staircases, the harmonic stack wont topple over. In fact, we will show (Theorem 3.2)
that the harmonic stack is stable in a correspondingly stricter sense.
Motivated by the limiting process above, in this article we shall consider general
stacks of width 2 and height 1. A general stack is not one solid piece, but rather consists
of infinitely many infinitely thin and unconnected horizontal blocks. Similar to a tower
of finite blocks, a general stack is capable of toppling at any level: to avoid toppling at
a given height, the center of mass of the stack above that height must lie directly above
the cross-section of the stack at that height. (By comparison, a solid stack is safe from
toppling just as long as its total center of mass lies above its base.)
As indicated above, the harmonic staircases are distinguished in the framework of
the original problem by each block being extended as far as possible. We will show
(Theorem 3.2) that the harmonic stack is similarly characterized among the stable
stacks: cutting the stack horizontally at any height into two pieces, the center of mass
of the top piece lies directly above the upper right corner of the lower piece.
The harmonic staircase consisting of n blocks has maximum overhang within a
natural class of stacks made up of the same blocks. Similarly, we will show (Theorem
6.1) that the harmonic stack is a fastest growing stable stack. What may be surprising is
that the harmonic stack is not the uniquely fastest growing stable stack (Theorem 6.2).
Other results include various methods of transforming stable stacks into new stable
stacks, and further characterizations of the harmonic stack amongst stable stacks. All
the arguments employed are elementary.

Getting ready to stack. The original stacking problem is posed in terms of three-
dimensional blocks. However, the harmonic staircases and all stacks that we are inter-
ested in are simply figures in the x y-plane, orthogonally extended in the z-direction.
Clearly, the two-dimensional stacks will be stable if and only if their extensions are. So,
we lose nothing by restricting ourselves to discussing and drawing two-dimensional
stacks. (In Section 7, we make a short excursion into the world of 3D blocks).


For a further simplification, note that scaling a stack horizontally or vertically can-
not affect its stability. It follows that we can normalize, making all stacks one unit
high and two units wide. There will be one further normalization in Section 3, once we
introduce the notion of the gravity curve of a stack: effectively, all stacks considered
will just balance at the table level.
We can now formally define stacks and the stability of stacks.

2. WHAT IS A STACK? The stack S given by the stack function f : [0, 1) R

is the region in the plane bordered by the graphs of x = f (y) (not y = f (x)!) and
x = f (y) 2:

S = {(x, y) : 0 y < 1, f (y) 2 x f (y)}.

Though more general functions can be considered, it is natural to restrict to stack

functions that are integrable,2 and piecewise continuous and right continuous: that
is, the one-sided limits of f exist at any y [0, 1), and lim f (t) = f (y). We shall
ty +
assume this throughout.3
As the simplest example, the constant stack function f (y) = 1 gives a vertical
stack. The harmonic stack HAR, pictured in Figure 2, has stack function har(y) given
by the inverse of 1 ex :

har(y) = log(1 y).

The n-block harmonic staircase HARn has piecewise constant stack function
X 1
harn (y) = ,

where the floor function bnyc is the largest integer m ny. Weve shaded the graph of
har6 (y) in the picture of HAR6 in Figure 3. Essentially, it is comprised of the vertical
right-hand borders of the rectangular blocks in the stack.

table 1 1 1 1 1 x
6 5 4 3 2 1

Figure 3. The 6-block harmonic staircase.

As indicated in the picture, all stacks are resting on the x-axis, and the table extends
from to 0, making the upper right corner of the table the origin. The weight of
part of a stack is simply its area. Since stacks are of height 1 and constant width 2, it
follows that all stacks have total weight 2.
2 The stack function being either Riemann or Lebesgue integrable suffices. It is also sufficient that the stack

have an improper Riemann integral as y 1 .

3 For some of what follows it is only required that f be Lebesgue integrable. However, such considerations

are a bit arcane, even for us.


Theorem 2.1 (The harmonic staircases converge to the harmonic stack). The har-
monic stack function har(y) is the pointwise limit as n of the harmonic staircase
functions harn (y).4

Proof. We use the estimate

X 1
= log(n + 1) + + n ,

where is the Euler-Mascheroni constant and n 0. Now let y [0, 1). Then

n n1bnyc
X 1 X 1
harn (y) =
m m=1

= (log(n + 1) + + n ) log (n bnyc) + + n1bnyc

= log 1 + log + n n1bnyc .
n n

Now, 0 ny bnyc < 1, from which it follows that bnyc

y. Also, since y < 1,

n 1 bnyc = ny bnyc 1 + n(1 y) .


lim harn (y) = log(1 y) = har(y).



stack in Figure 4 below. Thought of as a stack of six unconnected rectangles, it is clear
that the top two rectangles will fall down: the center of mass of the top rectangle is not
above the second rectangle, and the center of mass of the combined top two rectangles
is not above the stack consisting of the remaining four rectangles. However, considered
as one solid piece, the center of mass of this stack is above the table, and so it will not
topple over.

table x

Figure 4. As one solid piece, this stack has a huge overhang and will not topple over.
4 The proof actually shows that, for h < 1, har (y) converges uniformly to har(y) on [0, h]. Note also that
Theorem 2.1 can be related to the Maclaurin expansion of the harmonic stack function:

y2 y3 y4
har(y) = log(1 y) = y + + + + .
2 3 4
For the limiting value y = 1 this identity says that the overhang of the harmonic stack is equal to the limit of
the harmonic series.


We want to consider our stacks as unconnected, and thus capable of toppling, at any
height; in effect, we are thinking of stacks as consisting of infinitely many infinitely
thin blocks. It is then natural to define a stack to be stable if it balances at every
possible level.

The gravity function and the gravity curve of a stack. More formally, consider a
general stack S, and fix a height y. Consider the slab of S lying above y, and let g(y) be
the x-coordinate of the center of mass of this slab. Then we call (g(y), y) the gravity
point of S at height y. Notice that a gravity point always lies directly below the center
of mass of the slab defining it, as pictured in Figure 5. We also call the function g(y)
the gravity function of the stack S, and the graph of x = g(y) is the gravity curve of S.

center of mass of top slab

gravity point at height y


table x

Figure 5. The gravity point at height y is directly below the center of mass of the top slab.

Proposition 3.1 (Equation of the gravity function). Suppose S is a stack with stack
function f . Then the gravity function of S is continuous and is given by
( f (t) 1) dt
g(y) = .

Wherever f (y) is continuous, the gravity function g(y) is differentiable and

f (y) = 1 [g(y)(1 y)]0 .

At every point, g is differentiable from the right, and the above equation for f holds
everywhere with the derivative so interpreted. Consequently, the stack function
uniquely determines the gravity function and vice versa.5

Proof. If we consider the stack S to be made of infinitesimally thin blocks, then the
block at height t has mass 2dt, and the x-coordinate of its center of mass is f (t) 1.
It then follows that
( f (t) 1) 2 dt
g(y) =
mass of the slab above height y
( f (t) 1) dt
= .
(1 y)
5 In the more general setting of Lebesgue integrable functions, the gravity function g is absolutely continu-

ous, and determines f almost everywhere.


The rest of the proposition follows by multiplying by 1 y and applying the funda-
mental theorem of calculus, recalling that we are only considering stack functions that
are continuous from the right.

From here on, g 0 (y) shall always denote, if need be, the right derivative of the
gravity function g. The previous result then promises that this right derivative always

Normalizing the tables. We say that a stack S is stable if S contains its gravity curve.
If f is the stack function of S, this is the case exactly when

f (y) 2 g(y) f (y) for all y [0, 1).

Further, we say that a stack is balanced at 0 if g(0) = 0: that is, if the center of mass
of the whole stack is above the top right corner of the table.
It is easy to check that all harmonic staircases are stable stacks balanced at 0. In
fact, a harmonic staircase is constructed exactly so that its gravity curve will contain
the top right corner of the table, as well the top right corners of all but the topmost
block; through the top block, the gravity curve is simply a vertical line directly up the
As part of the next result, we prove that HAR is balanced at 0. As well, it is obvi-
ous that any stack can be translated to be balanced at 0. This justifies the following
From here on, we shall consider only those stacks that are balanced at 0.

y y

table x table x

Figure 6. The gravity curves of the vertical stack and a harmonic staircase.

The gravity curve of the harmonic stack. Intuitively, the stack functions of HARn
approximate the gravity functions of HARn , suggesting that the stack and gravity func-
tions of HAR should coincide. This is indeed the case.

Theorem 3.2 (The harmonic stack and gravity functions coincide). The harmonic
stack HAR is the unique stack whose gravity function and stack function coincide. In
particular, HAR is stable.

Proof. Suppose S is a stack with stack function f . By Proposition 3.1, the stack and
gravity functions of S will coincide if and only if f is continuous and

( f (t) 1) dt
= f (y).


It is easy to verify that the harmonic stack function satisfies this equation. Conversely,
suppose the equation holds. It follows that the integral is differentiable, and thus f
must be differentiable (in fact infinitely differentiable). Multiplying the equation by
1 y and differentiating,

1 f (y) = f 0 (y)(1 y) f (y),

and so
f 0 (y) = .

Antidifferentiating gives

f (y) = log(1 y) + C.

Since we have normalized to have all stacks balanced at 0, the only possibility is
C = 0, giving the harmonic stack.

For completeness, we also prove the following.

Proposition 3.3. The gravity functions of the stacks HARn converge pointwise to har.

Proof. Let gn be the gravity function of HARn and let g be the gravity function of
HAR. We prove that, as suggested above, gn (y) harn (y) 0 pointwise on [0, 1).
The proposition then follows immediately from Theorem 2.1 and  Theorem
Fix y [0, 1), suppose n N, and set m = bnyc. Then y mn , m+1
. If n is large
then n > m + 1, and therefore
m  m+1 m  1
gn gn (y) harn (y) = gn = gn + .
n n n nm
1 1
From the proof of Theorem 2.1, we know that nm
= nbnyc
0. It follows that
gn (y) harn (y) 0, as desired.

4. CUT AND PASTE. The following diagram shows two stacks S and T together
with their gravity curves. We now slice both stacks at height t, and then combine them
to make a new stack as pictured: the bottom slab of S stays fixed, and the top slab of
T is horizontally translated so that the ends of the gravity curves coincide. We will
denote this new stack by S \t T . From Proposition 3.1 we know that gravity functions
are continuous. It then follows immediately from the definition of the gravity curve
that the gravity curve of S \t T is exactly the union of the two part-curves.
We immediately conclude the following.

Proposition 4.1 (Properties of cut and pasted stacks). Let S and T be two stacks. If
both S and T are stable then so is S \t T .

Here is a nice application of this construction. Suppose that S and T are stable
stacks. Then St = S \t T, t [0, 1] is a continuous deformation of T into S with all
intermediate stacks being stable.



t S T

t S\tT

Figure 7. Combining parts of two stacks by aligning the gravity curves.

1 1
2 2 1

Figure 8. The stack har2 \ 1 har4 . The gray curve is the gravity curve.

Cutting and pasting is also a useful technique for transforming finite stacks of rect-
angular blocks. As an example, the stack har2 \ 1 har4 consists of the three blocks
pictured in Figure 8, with the lower two overhangs of length 12 .
Now replace the top block, by cutting and pasting with har8 at t = 34 , giving
(har2 \ 1 har4 ) \ 3 har8 . Continuing this process forever and taking the limit, we arrive
2 4
at the infinite-block stack

HALF = ((((har2 \ 1 har4 ) \ 3 har8 ) \ 7 har16 ) \ 15 har32 )

2 4 8 16

The stack HALF consists of blocks of heights 12 , 41 , 18 , 161 , . . . , with all overhangs
of length 12 . So, HALF has infinite overhang. Also, by applying Proposition 3.1 one

1 1 1 1 1
2 2 2 2 2

Figure 9. Finite stacks converging to HALF. The gray curve is the gravity curve.


can show that the gravity curves
n of HALF and the nth finite approximation to HALF
coincide on the interval 0, 2 21
n . It follows that HALF is stable and that its gravity
curve passes through the top right corners of the blocks, as indicated in Figure 9.

5. STRETCHING AND TRANSLATING. Another useful method of modifying a

stack S is to take the top slab of S, above some height t, and then dilate and translate
this slab to form a new stack; see Figure 10. We denote the resulting stack by t S.

t S discard bottom part

rescale down to make

t new stack

translate to balance at 0

t S

Figure 10. Stretching and translating the top part of a stack into a new stack.

The stack function for t S is easily determined from the construction, and then the
gravity function is easily determined from Proposition 3.1. We summarize this in the
following proposition.

Proposition 5.1 (The defining functions of stretched stacks). Let S be a stack with
stack function f and gravity function g, and let t [0, 1). Then t S has stack function

f (y(1 t) + t) g(t),

and gravity function

g(y(1 t) + t) g(t).

If S is stable then so is t S.

A direct calculation using Proposition 5.1 shows that HAR is self-similar:

t HAR = HAR for any t [0, 1).


However, HAR is not the only self-similar stack.

Theorem 5.2 (Classification of self-similar stacks). Let S be a stack such that t

S = S for all t [0, 1). Then the stack function f and gravity function g of S are
of the form

f (y) = a log(1 y) + a + 1


g(y) = a log(1 y),

for some a R.

Notice that a = 0 gives the vertical stack, and a = 1 gives HAR. Also, a = 1 gives
HAR reflected in the y-axis. It is clear that the stable stacks are given by a [1, 1],
and so a = 1 give the extreme stacks.

Proof of Theorem 5.2. Since t S = S, the gravity curves of the two stacks have to be
identical. So, by Proposition 5.1,

g(y(1 t) + t) g(t) = g(y) for all y, t [0, 1).

Rearranging and dividing by t (1 y), we have

g(y + t (1 y)) g(y) g(t)

= .
t (1 y) t (1 y)

Letting t 0+ and applying g(0) = 0 gives

g 0 (0)
g 0 (y) = ,

where g 0 (y) denotes the right derivative of g. Since g(0) = 0, we can conclude that

g(y) = a log(1 y)

for some a R.6 Then, from Proposition 3.1,

f (y) = 1 (g(y)(1 y))0 = a log(1 y) + a + 1.

Notice that the HALF stack introduced in the previous section is self-similar for
infinitely many values of t: it is easy to check that

1 1  HALF = HALF for all n N.


6. WHICH STACKS GROW THE FASTEST? When stacking n identical blocks,

with each block supporting at most one block above, it can be shown that the har-
monic staircase HARn has maximal overhang: see [4, Section II]. However, HARn is
6 Even though g may only have a one-sided derivative, the continuity of g still permits us to antidifferentiate

g 0 , obtaining the desired expression for g. See, for example, [1, Theorem 7.1].


Figure 11. Converting HAR6 into another stable stack of equal, maximal overhang.

not unique in this regard, as an adaptation of HARn does just as well, as shown in
Figure 11.
In this section we will prove similar results for HAR. First, we prove (Theorem
6.1) that HAR has the fastest growing gravity function amongst stable stacks. We then
prove (Theorem 6.2) that HAR has one of the fastest growing stack functions, but that
it is not unique in this regard.
In what follows, we will consider stacks growing to the right of the table. Of course
stacks can also grow to the left, and there are obvious left versions of all our results
Theorem 6.1 (HAR has the fastest growing gravity function). Suppose S is a stable
stack with gravity function g. Then har g is a nondecreasing and nonnegative func-
Note that it is possible that g = har on an initial interval [0, t]. However, Theorem
6.1 implies that once har gets in front of g, it remains so.
Proof of Theorem 6.1. Since har(0) = g(0) = 0, we only need to prove that har g
is nondecreasing. To do this, assume by way of contradiction that har0 (t) g 0 (t) < 0
for some t [0, 1), where g 0 (t) refers to the right derivative of g. Let h be the gravity
function of t S. From Proposition 5.1 and the self-similarity of HAR, it follows that
har =t har < h on some interval (0, s].
Now consider the stack T = (t S) \s HAR. By Proposition 4.1 and Proposition 5.1,
T is stable. Further, if k is the gravity function and f is the stack function of T , then
the stability of T and the choice of s ensures that har < k f on (0, 1); see Figure 12.


Figure 12. The gravity function k of the stack T is to the right of har.


But this is a contradiction: since HAR is balanced at 0, har < f implies that the center
of mass of T will be to the right of the origin.

Having proved the gravity function har of HAR is dominant, what can we say about
har as a stack function relative to other stack functions? Certainly, HAR need not al-
ways be in front of other stacks. For example, the vertical stack begins in front of HAR,
before being overtaken.
On the other hand, to have a stable stack S strictly in front of HAR from a certain
height on is impossible: if this were the case, then above that height the gravity function
of S would also be in front of har, contradicting our previous theorem. However, there
do exist stable stacks that effectively compete for the lead.

Theorem 6.2 (The fastest growing stack functions). There are no stable stacks that
stay ahead of HAR for all y near 1. However, there do exist stacks that grow as fast as
HAR. That is, there is a stable stack S, with stack function f , such that f (y) har(y)
changes sign infinitely often as y approaches 1.

Proof of Theorem 6.2. We have already argued that a stable stack cannot stay ahead of
HAR for all y near 1. We will now construct a stack S that repeatedly alternates with
HAR for the lead: S has stack function f satisfying

har(y) < f (y) for values of y arbitrarily close to 1.

To see how to construct such a stack S, we first assume that S has the desired lead-
changing property, and we use this to derive explicit sufficient conditions for the stack
function f . We then show that f can indeed be chosen so that these conditions are
We begin by considering the gravity function g of S. Since S is assumed stable, it
follows that for any y for which har(y) < f (y), we must also have

har(y) g(y) < 2.

But, by Theorem 6.1, har g is nondecreasing. It follows that if S is forever changing

the lead with HAR then har(y) g(y) < 2 for all y near 1, and thus that har g
converges to some m [0, 2] as y goes to 1. We can therefore write

g(y) = har(y) m + (y),

where  0 is a nonincreasing function, and (y) 0 as y 1.

Any m (0, 2) will suffice for what follows, but for definiteness we take m = 1.
Then, since S is to be balanced at 0, we must have

(0) = 1.

We now assume that (y) is differentiable. Then, using Proposition 3.1, we can

f (y) = 1 [g(y)(1 y)]0

= 1 [(har(y) 1 + (y))(1 y)]0
= [har(y)(1 y)]0  0 (y)(1 y) + (y).


By Theorem 3.2, we also know that 1 [har(y)(1 y)]0 = har(y), and so

f (y) = har(y) 1 + (y)  0 (y)(1 y).

We want S to be stable, for which we require f 2 g f . Since g = har 1 + ,

this is equivalent to

0  0 (y)(1 y) 2 for all y.

Finally, S being ahead of HAR at height y amounts to f (y) > har(y). So, for the
lead-changing property, it suffices to have

1 <  0 (y)(1 y) for values of y arbitrarily close to 1.

In summary, it suffices for us to find a nonincreasing and differentiable function

 : [0, 1) R such that

(0) = 1

lim (y) = 0


 0 (y)(1 y) 2 for all y

 0 (y)(1 y) > 1 for values of y arbitrarily close to 1.

These conditions will guarantee that the corresponding stack S is stable and balanced
at 0, and will repeatedly overtake HAR.
To construct an explicit function  satisfying these conditions, we shall take  to
have constant segments that are connected by small S-bends of just the right size and
slope. To do this, we first define a suitable prototype S-bend; see Figure 13:
5 1
B(y) = y y5, y [1, 1].
4 4
Note that B(1) = 1 and B 0 (1) = 0. Also, B 0 0 on [1, 1], with a maximum
of 54 . We also take B(y) = 1 for y > 1 and B(y) = 1 for y < 1.
We now define  by subtracting a sum of suitable linear transformations of B.
Specifically, define

(y) = 1 Bn (y),

where Bn is the S-bend B transformed to rise from 0 to 21n on the interval

2 1 1 2n 1

2n , ;
2n 2 2n
see Figure 14.
Clearly  is nonincreasing, with (0) = 1 and (y) 0 as y 1. Also, Bn has a
maximum slope of 5 2n2 . And, on the interval where Bn is bending, we have 21n
1 y 2n+1 . It follows that at the point of maximum slope of Bn , we have

1 5
 0 (y)(1 y) 5 2n2 = > 1.
2n 4


1 x = (y)
x = B (y)

1 1

0 1/8 1/4 1/2 1

Figure 13. The prototype S-bend B. Figure 14. The function .

Furthermore, at every point on the bending interval of Bn , we have

3 15
 0 (y)(1 y) 5 2n2 n+1 = 2.
2 8
It follows that  has exactly the properties desired. The resulting stack S is shown
shaded in Figure 15, with har and the gravity curve of S superimposed.

Figure 15. A stable stack that grows as fast as HAR.

7. AFTERTHOUGHT 1: FLIPPING. Figure 16 illustrates another natural opera-

tion for transforming a stack S into a new stack: take the slab above a certain height
t [0, 1), and reflect that slab about the vertical line through its gravity point. We
denote the new stack by t S. It is clear that if S is stable then so is t S.
We now apply this flipping procedure to construct a stable stack that has infinite
overhang to both the left and the right. (The following was inspired by similar con-
structions involving finitely many blocks, described in [2, Chapters 12.5 and 12.7].)

t S

t t S

Figure 16. Reflecting the top slab of a stack about an axis through its gravity point.



Figure 17. Constructing a stable stack with infinite overhang to both the left and right.

Recall the stack HALF constructed at the end of Section 4, consisting of blocks of
heights 21n , each placed with overhang 12 . We now flip this stack above the 1st, 3rd,
6th, and in general the (n+1)n
th block; each pair of flips results in the stack extending
further in both directions. Using Proposition 3.1, it is then easy to show that the
limiting result of these flips is a stable stack S, with stack function having unbounded
oscillation in both directions as y 1.
Note also that this flipping procedure can be used to construct a stack that continu-
ally overtakes the harmonic stack, similar to that constructed at the end of the previous
section. For this we modify the harmonic stack by flipping out infinitely many small
horizontal slivers that get arbitrarily close to the top of the stack. It is then possible to
arrange for these slivers to jut beyond the harmonic stack.
We now momentarily venture into the world of 3-dimensional blocks. Well create a
stack that casts a shadow over the whole x z-plane. (Well continue to label the vertical
direction as y.)
Begin with the oscillating stack S just constructed. Notice that there are infinitely
many blocks of S such that a top corner of the block lies above the origin and the grav-
ity curve of S also passes through that corner. Now, thicken the blocks in S to have
a thickness d in the z direction, giving a 3D stack b S. Next, take any fixed irrational
number a, and consider the angle = a. Finally, at the height of each of the distin-
guished corners of S above the origin, successively rotate the top slab of b S the angle
around the y-axis.
Any angle is closely approximated by arbitrarily large integer multiples of . It
follows that, no matter how small the thickness d, any point in the direction will
eventually lie under some block of b S. It follows that b
S casts a shadow over the whole
x z-plane.


this section, we derive an interesting balancing property of the exponential function.
We use this property to reprove the result from Theorem 3.2, that the stack function
and gravity function of the harmonic stack coincide.

Proposition 8.1 (Balancing the exponential function). The region under the graph
of the function e x to the left of the point x = a balances over a fulcrum at x = a 1.

Proof. Using the standard formula, we find that the x-coordinate of the center of mass
of the tail region is
xe x dx

= a 1.
e x dx


y = ex

Figure 18. The tail of the region under y = e x always balances 1 unit to the left of the cut.

Figure 19 shows the region between the graph of the exponential function e x and its
horizontal translate e x2 , truncated at a certain height. If the height is less than 1, then
this is exactly a top slab of the harmonic stack, rotated 180 degrees about the point
(0, 1/2). What we want to prove is that no matter where we cut, the shaded region
balances with the fulcrum at a: this establishes again that the gravity function and the
stack function of HAR are identical.


y = ex

Figure 19. The truncated region between y = e x and y = e x2 balances at x = a.

Consider the region lying between y = e x and the horizontal translate y = e xd ,

and to the left of x = a; see Figure 20. By the previous proposition this sliver is the
difference of two regions that balance over a 1, and hence the sliver also balances
over a 1.

y = ex
y = e x d

Figure 20. The sliver trapped by y = e x and its translate balances over a pivot at a 1.

Now set the horizontal difference to be d = n2 . Then n + 1 copies of the sliver fit
together seamlessly into the shaded region in Figure 19, with a few curvy triangles
missing at the top and an extra sliver sticking out on the right; see Figure 21.


a1 a
Figure 21. Translates of the sliver combine to approximate the shaded region in Figure 19.

The fulcrums of these n + 1 slivers are equally spaced from a 1 to a + 1 meaning

the fulcrum of the entire region is at the middle point a. Letting n , the triangular
gaps at the top and the extra sliver on the right vanish, proving again that the gravity
function and stack function of HAR coincide.
It can also be shown that, within a natural class of functions, the exponential func-
tion is characterized by the balancing property established in Theorem 8.1. The proof
is a straightforward exercise.


vide an interesting situation, where the natural candidate for optimality is not optimal.
Suppose, ignoring what we have learned from Theorem 6.2, that we try to prove HAR
is the stack with eventual greatest overhang: that is, for every stable stack function f it
is eventually the case that har f . We now argue that if there is a stack with eventual
greatest overhang, then it is HAR. This is reminiscent of Jakob Steiners original and
famously flawed approach to the isoperimetric problem, that the circle maximizes area
amongst closed curves of a given perimeter; see, for example, [3].
Consider any other stable stack S with stack function f and gravity function g.
Then, by Theorem 3.2, there exists an interval [s, t] on which g < f . Figure 22 shows
the stack S and its gravity curve subdivided into a top, middle, and bottom slab, with
the middle slab corresponding to the interval [s, t].


Figure 22. A stable non-harmonic stack together with its gravity curve.

Now, if we slide the top slab slightly to the right, then its gravity point a will stay
within the top of the middle slab, ensuring that the top slab will not topple. However,
sliding the top slab will move the gravity point of the whole stack to the right, and the
stack will not be balanced at 0.
However, we can avoid this by simultaneously sliding the top slab to the right and
the middle slab to the left. Clearly, we can do this in such a way that the gravity point
b of the combined top and middle slabs stays fixed, and so leaving the gravity curve
inside the bottom slab unchanged. This means that the adjusted stack still balances,


with the top slab being further to the right than for the original stack: the adjusted
stack eventually has greater overhang.
This argument can be applied to any stack other than HAR, and thus establishes that
if there is a stack of eventual maximum overhang, then it must be HAR. Note that this
does imply that HAR can be regarded as the fastest growing stable stack in a certain
sense: HAR is the only stable stack that cannot be completely overtaken by any other
stable stack.


1. D. Bressoud, A Radical Approach to Lebesgues Theory of Integration, MAA Textbooks, Cambridge Uni-
versity Press, Cambridge, 2008.
2. J. Bryant, C. Sangwin, How Round is Your Circle? Princeton University Press, Princeton, 2008.
3. R. Courant, H. Robbins, I. Stewart, What is Mathematics? Oxford University Press, New York, 1996.
4. J. F. Hall, Fun with stacking blocks, Amer. J. Phys. 73 (2005) 11071116; available at http://dx.doi.
5. M. Paterson, Y. Peres, M. Thorup, P. Winkler, U. Zwick, Maximum overhang, Amer. Math. Monthly 116
(2009) 763787; available at

BURKARD POLSTER received his Ph.D. in 1993 from the University of Erlangen-Nurnberg in Germany.
He currently teaches at Monash University in Melbourne, Australia. Readers may be familiar with some of his
books dealing with fun and beautiful mathematics such as The Mathematics of Juggling, Q.E.D.: Beauty in
Mathematical Proof, or the Shoelace Book.
School of Mathematical Sciences, Monash University, Victoria 3800, Australia

MARTY ROSS is a mathematical nomad. He received his Ph.D. in 1991 from Stanford University. Burkard,
Marty, and their mascot the QED cat are Australias tag team of mathematics. They have a weekly column
in Melbournes AGE newspaper and are heavily involved in the popularization of mathematics. Their various
activities can be checked out at When he is not partnering Burkard, Marty enjoys
smashing calculators with a hammer.
PO Box 83, Fairfield, Victoria 3078, Australia

DAVID TREEBY studied mathematics at Monash University in Australia where he graduated in 2005. He
currently teaches mathematics to high school students at Presbyterian Ladies College in Melbourne, Australia.
He delights in exploring beautiful mathematics with students at his school.
Presbyterian Ladies College, 141 Burwood Hwy, Burwood, Victoria 3125, Australia


Edited by Sergei Tabachnikov

Riemann Maps and Diameter Distance

David A. Herron

Abstract. We use the intrinsic diameter distance to describe when a Riemann map has a con-
tinuous extension to the closed unit disk.

1. INTRODUCTION. Many basic complex analysis courses include the cele-

brated Riemann mapping theorem; a few discuss the Caratheodory-Osgood extension
theorem which asserts that a Riemann map admits an extension to the closed unit disk
that is a homeomorphism if and only if the target is a Jordan domain (i.e., bounded
by a plane Jordan curve). When the target domain is not Jordan, one considers prime
ends, a theory beyond the scope of beginning complex analysis courses and texts.
The purpose of this note is to provide a written record of the following intermediate
result; of special interest is the interplay between complex analysis, plane topology,
and metric geometry. This by-product of a more general theory is folklore among the
experts (cf. [3, 6.1.(4)]), but perhaps not widely known.
Theorem. Let D  be a holomorphic homeomorphismfrom the open unit disk D
in the complex plane C onto a bounded domain  in C. The following are equivalent:
(a) There exists a continuous map D  with g|D = f .
(b) There exists a homeomorphism D  d with h|D = f .
(c) d  is a closed Jordan curve, i.e., a topological circle.
(d)  is a closed curve.
(e)  is locally connected.

Here  d and d  :=  d \  are the metric completion and metric boundary of the
metric space d := (, d) where d is the diameter distance on . Also, in (a) =
(b) we have g = i B h where i :  d  is the extension of the identity map d .
See 2.C for definitions.
Figure 1 illustrates a simple non-Jordan domain that satisfies conditions (a) through
(e); Figure 2 pictures two domains that do not.
It is straightforward to check that (b) implies (c) and that (d) implies (e). We refer
to [4, Theorem 2.1, p. 20] for a proof of the nontrivial fact that (e) implies (a). In 3
we verify that (a) implies (b) and that (c) implies (d).
Our ideas and proofs should be accessible to students possessing basic knowledge
of complex analysis and plane topology, and could possibly serve as a capstone experi-
ence for undergraduate mathematics majors. The reader is forewarned that our proofs
employ basic plane and metric topology arguments; the role of holomorphicity is ex-
plained in 2.B.
MSC: Primary 30C20, Secondary 30C35, 30J99


f slit disk

unit disk

A point in  that
has two preimages
in d  and in D

(z+1) 2 z 2 +1
f (z) =
(z+1)+ 2 z 2 +1

Figure 1. A Riemann map onto a slit disk.


2.A. Basic notation and terminology. Throughout this article  denotes a simply
connected bounded domain in the complex plane C. We write D(a; r ) := {z C |
|z a| < r } for the open disk centered at a C with radius r > 0. Then D := D(0; 1)
is the unit disk with boundary T := D = {eit | t [0, 2]}, the unit circle.
A path is a continuous map of a compact interval, and unless explicitly indicated
otherwise, we assume that the parameter interval is [0, 1]. We use the phrase path in
 with a terminal endpoint in  to describe a path

[0, 1]  { } with ([0, 1))  and (1) = .

We write | | := ([0, 1]) for the image of the path . However, we write [a, b]
both for the Euclidean line segment joining a and b as well as the affine path [0, 1] 3
t 7 a + t (b a); the reader can distinguish these two meanings by context.
A path joins (0) to (1). When (0) = (1), we call a closed path. By a
curve we mean the image of a path, a closed curve is the image of a closed path, and
an arc is the image of an injective path. A crosscut of  is an arc with endpoints in 
and all other points in . An endcut of  is an arc having one endpoint in  and all
other points in .
We note that every path contains an injective subpath that joins its endpoints; see [5].
A (closed) Jordan curve is a topological circle, that is, the homeomorphicimage
of T. We call 0 a plane Jordan curve if 0 is a Jordan curve in C; in this setting,
the Jordan curve theorem asserts that C \ 0 has exactly two components: the bounded
component int(0) called the interior of 0 and the unbounded component ext(0) called
the exterior of 0.

2.B. Riemann maps. A Riemann map D  is a holomorphic homeomorphism;
complex analysts call such an f a conformal map. We require two properties of Rie-
mann maps. First, according to [4, Prop. 2.14, p. 29], each path in  with a terminal
endpoint in  will have a preimage that is a path in D with a terminal endpoint in
T. This fact, whose proof is based on a length-area estimate known as Wolfs lemma
(see [4, Prop. 2.2, p. 20]), does not require that f have a continuous extension to the
closed disk.

February 2012] NOTES 141

Second, suppose a Riemann map f has a continuous extension to a map g : D .

According to Lemma 2.2, each nondegenerate subarc of T has a nondegenerate image;
that is,

for all subarcs A T, diam(A) > 0 diam(g(A)) > 0. (2.1)

2.2 Lemma. Let f be holomorphic and bounded in D. Suppose there is a nondegen-

erate subarc A T such that

for all A, lim f (z) = 0.


Then f = 0.

Proof. We assume that A {eit | 0 t 2/n} for some positive integer n. Define

F(z) := f 1 (z) f 2 (z) f n (z) where f k (z) := f (e2ik/n z).

Then F is holomorphic and bounded in D with the property that for each T,

lim F(z) = 0.

An appeal to the maximum principle reveals that F = 0, so f = 0.

2.C. Diameter distance. The diameter distance1 d on  is defined by

d(a, b) := inf{diam( ) | a path in  joining a and b }.

It is easy to check that d is a distance function on  and that for all a, b ,

|a b| d(a, b). We write d := (, d), and then  d and d  := 
d \  are the
metric completion and metric boundary of the metric space d .
In fact, the identity map d  is a 1-Lipschitz homeomorphism. Therefore, it
follows that id extends naturally to a 1-Lipschitz map i :  d .
In general, i need
not be surjective nor injective, but using continuity of id we see that i(d ) .

For the slit disk, pictured in Figure 1, the map i is surjective but not injective. For the
two domains pictured in Figure 2, i is neither surjective nor injective; here the left-

1 7i
1 7i 8
+ 8
+ 8
1 3i
+ 4  

1 i 1 i
+ 2 +
... ... 2 2

1 i
+ 4

Figure 2. Domains  with i noninjective and nonsurjective.

1 This is sometimes called the inner or internal or intrinsic diameter distance.


hand-side  has d  homeomorphic to (0, 1] whereas the right-hand-side  has d 
homeomorphic to (0, 1).
In our setting we can visualize d  as the set pa  of path accessible boundary
points of . More precisely, pa  = i(d ) is the set of all points  with the
property that there exists a path in  with terminal endpoint ; we recall that this

[0, 1]  { } with ([0, 1))  and (1) = .

There is also a natural connection between d  and endcuts of . To see this, let
be a path in  with terminal endpoint (1) = in . By choosing a sequence of
points along that is d-Cauchy but nonconvergent in , we obtain a point d 
with i( ) = . The continuity of at t = 1 ensures that limt1 d( (t), ) = 0, so
we can define a continuous path d in d by
(t) if t [0, 1),
d (t) :=
if t = 1.

Thus each such path corresponds to a path d in d with a unique terminal endpoint
d (1) d  and with the property that = i B d , so (1) = i(d (1)).
Let and be two paths in  with terminal endpoints in . We declare and
to be d-equivalent provided limt1 d((t), (t)) = 0. There is a natural one-to-
one correspondence between d  and the equivalence classes of such paths. A general
discussion, with detailed proofs, can be found in [2].
In 3 we use the following elementary facts.

2.3 Lemma. If 
d is compact, then i is surjective; in particular, i(d ) = .

Proof. Assume  d is compact. Let . Choose a sequence (z n ) n=1 in  with

|z n | 0 as n . There exists a subsequence (z nk ) k=1 and a point  d such
that d(z n k , ) 0 as k . As id is continuous, 6 , so d . Since

| i( )| | z n k | + |z nk i( )| | z n k | + d(z n k , ) 0 as k ,

it follows that = i( ).

2.4 Lemma. The space 

d is compact if and only if d  is compact.

Proof. Since d  is a closed subset of  d , compactness of the latter implies compact-

ness of the former. Assume that d  is compact. Let U be a collection of d-open sets
that cover 
d . There is a finite subcollection V U such that
U := V d .

Since U is d-open, A :=  d \ U = d \ U =  \ U is d-closed, so also closed. As

A is closed and bounded, it is compact, and hence d-compact too. So, there is another
finite subcollection W U that covers A, and V W is a finite subset of U that forms
a d-open cover of  d.

February 2012] NOTES 143

3. PROOFS. Let D  be a holomorphic homeomorphismwith  C bounded.
We continue with the notation introduced in 2. Here we demonstrate that (a) implies
(b) and that (c) implies (d).
3.A. (a) = (b). Assume there is a continuous map D  with g|D = f . For each
point a T, consider the path := g B [0, a] in  {g(a)} defined by (t) := g(t a).
As described in 2.C, this determines a unique point d (1) d  with the property
that i(d (1)) = (1) = g(a). Now define h : D  d by
g(z) if z D,
h(z) :=
(g B [0, z])d (1) if z T.

Note that g = i B h. We claim that h is a homeomorphism. Since D is compact and  d

is Hausdorff (being a metric space), it suffices to show that h is a continuous bijection.

Proof that h is continuous. Using the fact that d  is a homeomorphism we see
that h|D = id1 B g|D is continuous. Thus it suffices to check that h is continuous at
each point of T. Let a T and > 0 be given. Select > 0 so that g(D D(a; ))
D(g(a); /4). Let z D D(a; /2).
The continuity of g, in conjunction with the definition of h, guarantees that

lim d(h(ta), h(a)) = 0 = lim d(h(t z), h(z)).

t1 t1

In particular, we may choose t0 (0, 1) so that a0 := t0 a and z 0 := t0 z satisfy

a0 , z 0 D D(a; ) and d(h(a0 ), h(a)) < /4, d(h(z 0 ), h(z)) < /4.

As D D(a; ) is convex, the line segment path := [a0 , z 0 ] lies in D D(a; ), so

g B lies in D(g(a); /4). Thus

d(h(a0 ), h(z 0 )) = d(g(a0 ), g(z 0 )) d(g B ) < /2

and therefore

d(h(z), h(a)) d(h(z), h(z 0 )) + d(h(z 0 ), h(a0 )) + d(h(a0 ), h(a)) < .

Proof that h is surjective. Let d . Pick a path in  with terminal endpoints

d (1) = and (1) = := i( ). As described in 2.B, the preimage of is a path
(so, g B = ) in D with terminal endpoint a := (1) T = D. Since h is continu-
ous, h B is a path in 
d . For each t [0, 1), h((t)) = g((t)) = (t), so it follows
that h B = d . Thus h(a) = h((1)) = d (1) = .

Proof that h is injective. Let a and b be distinct points in T = D and let I and J be
the components of T \ {a, b} (so I and J are open subarcs of T). We demonstrate that
if h(a) = h(b), then min{diam[g(I )], diam[g(J )]} = 0; since this would contradict
(2.1), it follows that h(a) 6 = h(b).
As there is no harm in doing so, we assume that g(0) = 0. Then the paths := g B
[0, a] and := g B [0, b], given by (t) := g(t a) and (t) := g(t b), define endcuts of
 both having initial endpoint (0) = g(0) = 0 = (0) and with terminal endpoints
(1) = g(a), (1) = g(b) in . See Figures 3 and 4.


Figure 3. The paths and .

Now suppose that := h(a) = h(b). Then := i( ) = g(a) = g(b), and are
d-equivalent paths that join 0 to , and C := || || is a plane Jordan curve
in  { }. Note that C  = { }; however, in general, D := int(C) need not be
contained in . See Figure 3.
Let > 0 be given. We show that either g(I ) or g(J ) has diameter smaller than .
We assume that < | |/10. It is helpful to examine topology in the D disk picture
in Figure 4, but the reader must remember that g is only a homeomorphism in D.

g 1 (10 ) ra


g 1 (11 )

Figure 4. The disk picture.

The continuity of D , in conjunction with the d-equivalence of and , guar-
antees the existence of a (0, 1) such that

diam([, 1]) < /10, diam([, 1]) < /10,

and t [, 1), d((t), (t)) < /10.

In particular, there is an injective path in  that joins ( ) = g( a) to ( ) =

g( b) with diam( ) < /10. Note that | |  D( ; /5), so 0 / | | and thus
6 C (because the subarc of C that joins ( ) and ( ) in  passes through the
origin). See Figure 5.
Next we select an appropriate subpath of . We exhibit R, S satisfying 0 R <
S 1 and such that (R) is the last point of in [, 1] and (S) is the first point
of |[R,1] in [, 1], and such that := |[R,S] is injective and meets C only at the
endpoints of . To this end, let R := sup{t [0, 1] | (t) [, 1]}. Then 0 R < 1
and there exists an r [, 1) so that (R) = (r ) = g(r a). Next, let S := inf{t
[R, 1] | (t) [, 1]}. Then R < S 1 and there exists an s [, 1) so that (S) =
(s) = g(s b). We examine the paths

:= |[R,S] and := g 1 B .

February 2012] NOTES 145


( )


( )

Figure 5. The path and its subpath .

By its definition, is an injective path in  that joins (r ) to (s) and is such that
|| C = {(r ), (s)}. Thus || is a crosscut either of D = int(C) or of E := ext(C)
(and || separates 0 and , either in D or in E). To see that the former holds (i.e., that
|| is a crosscut of D), note that there are (closed) Jordan curves

00 := g([0, ra] || [sb, 0]) = [0, r ] || [s, 0],

01 := g([a, ra] || [sb, b]) = [1, r ] || [s, 1]


00 , 01 ( { }) D( ; /5),


10 := int(00 )  , 11 := int(01 ) D( ; /5).

In particular, / 10 and 0 6 11 . If || were a crosscut of E, then either || would

separate 0 from in E or || would separate from in E: the first case would
imply that 0 11 while the second case would mean that 10 . Since neither of
these holds, || is not a crosscut of E and so is a crosscut of D.
Note that 10 and 11 are the components of D \ ||. See Figure 4.
Evidently, || is a crosscut of one of the components, call it W , of D \ ([0, a]
[0, b]). The boundary of W consists of [0, a], [0, b], and one of the subarcs I or J ; we
assume that W = [0, a] I [0, b]. Again, see Figure 4. We claim that g(I ) 11 .
Since 11 D( ; /5), this means that diam[g(I )] < as required.
To verify this claim, we first check that g(I ) D { }. Let c I . Then K :=
g([0, c]) is an endcut of  and K C {0, }, so either K D {0, } or K E
{0, }. However, as [0, c] is a crosscut of W that joins 0 to c with 6 = [0, c] || W ,
we see that 6 = K || D. Thus, K D {0, }, so g(c) D { }.
Finally, recall that 0 and are separated by || in D. As g(I ) is connected and
g(I ), g(I ) lies in the -component of D \ ||, which is 11 \ ||. Thus g(I ) 11 .

3.B. (c) = (d). We assume that d  is a closed Jordan curve. In particular then,
d  is compact. To see that  is a closed curve, it suffices to verify that i(d ) = .
This latter requirement follows from Lemmas 2.3 and 2.4.


In closing we mention that in the literature one also encounters the (inner) length
distance ` which is defined similarly to d but using the length `( ) (of a joining path
) instead of diam( ). However, there are bounded Jordan plane domains  that have
finite length distance and  locally connected but  ` noncompact, so nonhomeo-
morphic to D. Such an example can be constructed by starting with an open square
and taking a convergent sequence (n ) of boundary points; at the points n we attach
disjoint spiraling tentacles that have length approximately one and diameters tending
to zero. See [1, Example 2.10] for more details.

ACKNOWLEDGMENTS. The author thanks the referees for their helpful comments. He was partially sup-
ported by the Charles Phelps Taft Research Center.


1. D. Freeman, D. Herron, Bilipschitz homogeneity and inner diameter distance, J. Anal. Math. 111 (2010)
146; available at
2. D. Herron, Geometry and topology of intrinsic distances, J. Anal. 18 (2010) 197231.
3. R. Nakki, J. Vaisala, Jon disks, Expo. Math. 9 (1991) 343.
4. C. Pommerenke, Boundary Behavior of Conformal Maps, Grunlehren der mathematischen Wis-
senschaften, no. 299, Springer-Verlag, Berlin, 1992.
5. J. Vaisala, Exhaustions of John domains, Ann. Acad. Sci. Fenn. Math. 19 (1994) 4757.

Department of Mathematics, University of Cincinnati, OH 45221

A Power Series Approach to

Some Inequalities
Cristinel Mortici

Abstract. The aim of this note is to introduce a new technique for proving and discovering
some inequalities.

1. INTRODUCTION. We give here a method for proving a class of inequalities us-

ing infinite series. This method is useful, because many difficult problems can be easily
solved and often they can be extended. To begin, let us consider the well-known Nes-
bitt inequality [4]

a b c 3
+ + , a, b, c > 0. (1)
b+c c+a a+b 2

To introduce our method, we assume, without loss of generality, that a + b + c = 1,

so we have to prove that
MSC: Primary 26D15

February 2012] NOTES 147

a b c 3
+ + , a, b, c (0, 1) .
1a 1b 1c 2
If we look carefully, then we discover that the fractions from the left-hand side are
the sums of some convergent geometric series. By also using the generalized means
inequality, we deduce that

a b c
an + bn + cn
+ + =
1a 1b 1c n=1 n=1 n=1

a+b+c n

a n + bn + cn
X X 
=3 3
3 n=1
 n 1
X 1 3 3
=3 =3 1
= ,
3 1 3

which justifies (1).

Next we discuss and refine some inequalities to illustrate our technique.

2. DISCOVERING NEW INEQUALITIES. Vasile Crtoaje [2] states that for all
nonnegative real numbers a1 , a2 , . . . , ak < 1 satisfying
a= a12 + a22 + + ak2 /k 3/3,

we have
a1 a2 ak ka
+ + + . (2)
1 a12 1 a22 1 ak2 1 a2

Here we give a sort of extension. Note that, by using (2) and the inequality between
the quadratic mean and the arithmetic mean, we obtain
a1 a2 ak

+ 1a22
+ + 1ak2
+ + +
1 a12 1 a22 1 ak2 k
ka 2
(1 a 2 )2

Hence, the inequality

2 2 2
a1 a2 ak ka 2
+ + + , (3)
1 a12 1 a22 1 ak2 (1 a 2 )2

a consequence of (2), holds for all a1 , a2 , . . . , ak (0, 1), with the condition a
We use our method to prove that inequality (3) holds without the condition
a 3/3. In this case, using

nx n =
, x (0, 1) , (4)
(1 x)2


we have:
2 2 2
a1 a2 ak
na12n + na22n + + nak2n
+ + + =
1 a12 1 a22 1 ak2 n=1 n=1 n=1

X a12n + a22n + + ak2n
=k n
 2 n
X a1 + a22 + + ak2
k n

a 2n

=k .
(1 a 2 )2

New inequalities can be discovered via this method, as we will see next. Consider
the well-known inequality:

a 2 + b2 + c2 ab + bc + ca, (5)

which is true for all a, b, c, but we take a, b, c (1, 1). For every n N,

a 2n + b2n + c2n (ab)n + (bc)n + (ca)n , (6)

and by addition,

a 2n + b2n + c2n (ab)n + (bc)n + (ca)n .

n=0 n=0 n=0 n=0 n=0 n=0

We obtain the following nice inequality:

1 1 1 1 1 1
+ + + + , a, b, c (1, 1) .
1a 2 1b 2 1c 2 1 ab 1 bc 1 ca

Further, if we multiply (6) by n and sum, then by (4), we obtain

2 2 2
a b c ab bc ca
+ + + + .
1 a2 1 b2 1 c2 (1 ab)2 (1 bc)2 (1 ca)2

By applying inequality (5) twice, we get

a 4n + b4n + c4n (a 2 bc)n + (ab2 c)n + (abc2 )n ,

and then using the summation method with respect to n again, we obtain

1 1 1 1 1 1
+ + + + .
1a 4 1b 4 1c 4 1 a bc 1 ab c 1 abc2
2 2

February 2012] NOTES 149

If we multiply by n before summing, we obtain
2 2 2
a2 b2 c2
+ +
1 a4 1 b4 1 c4
a b c
abc + + .
(1 a 2 bc)2 (1 ab2 c)2 (1 abc2 )2

Vasile Crtoaje [1] proved that for all x, y, z R,

17(x 3 + y 3 + z 3 ) + 45x yz 32(x 2 y + y 2 z + z 2 x).

If we consider x, y, z (1, 1) and n N, then by adding the inequalities

n n n 
17(x 3n + y 3n + z 3n ) + 45(x yz)n 32 x 2 y + y 2 z + z 2 x ,

we obtain the following interesting inequality:

17 17 17 45 32 32 32
+ + + + + .
1 x3 1 y3 1 z3 1 x yz 1 x2y 1 y2z 1 z2 x

Our method is also suitable for obtaining new results related to convexity. Walter
Janous [3] showed

(z n x n ) f (y) (z n y n ) f (x) + (y n x n ) f (z), (7)

where f : [0, ) R is any increasing and concave function, 0 < x y z, and n

a positive integer.
By adding inequalities (7) for n = 0, 1, 2, . . . , with 0 < x y z < 1, we get

1 1 1 1 1 1
f (y) f (x) + f (z).
1z 1x 1z 1y 1y 1x

By multiplying by (1 x)(1 y)(1 z), we obtain the interesting inequality:

(z x)(1 y) f (y) (z y)(1 x) f (x) + (y x)(1 z) f (z),

which shows that the function g : (0, 1) R given by g(x) = (1 x) f (x) is also
Finally, we use Holders inequality to establish two inequalities that are new, as far
we know.

Theorem 1. For a, b (0, 1) and p, q > 0 with p 1 + q 1 = 1, we have

q p pq
1 ap 1 bq 1 ab

ap bp ab
+ .
p(1 a p )2 q(1 bq )2 (1 ab)2


ap bq
Proof. We use Holders inequality: p
+ q
ab. We have

1 1 1 1 1 X p n 1 X q n
+ = (a ) + (b )
p 1 ap q 1 bq p n=0 q n=0

(a n ) p (bn )q 1

(ab)n =
= + .
p q n=0
1 ab

For the second inequality, we have

1 ap 1 bq 1 X p n 1 X
+ = n(a ) + n(bq )n
p (1 a p )2 q (1 bq )2 p n=1 q n=1
 n p
(a ) (bn )q ab
n(ab)n =
= n + .
p q n=1
(1 ab)2

ACKNOWLEDGMENTS. This work was supported by a grant of the Romanian National Authority for
Scientific Research, CNCS-UEFISCDI, project number PN-II-ID-PCE-2011-3-0087.


1. V. Crtoaje, Problem 2972, Crux Mathematicorum 6 (2004) 372.

2. , Problem 2983, Crux Mathematicorum 7 (2004) 430.
3. W. Janous, Problem 1861, Crux Mathematicorum 7 (1993) 203.
4. A. M. Nesbitt, Problem 15114, Educational Times 2 (1903) 3738.

Department of Mathematics, Valahia University of Targoviste, Romania

Chebyshev Mappings of Finite Fields

Julian Rosen, Zachary Scherr, Benjamin Weiss,
and Michael E. Zieve

Abstract. For a fixed prime p, we consider the set of maps Z/ pZ Z/ pZ of the form
a 7 Tn (a), where Tn (x) is the degree-n Chebyshev polynomial of the first kind. We observe
that these maps form a semigroup, and we determine its size and structure.

1. INTRODUCTION. Some of the worlds most interesting polynomials [2] are

the Chebyshev polynomials [4], which are defined for any positive integer n to be
n n k n2k1 n2k
Tn (x) = (1) 2 x .
nk k

The following table lists the first few Chebyshev polynomials:
MSC: Primary 37P25, Secondary 11T06, 33C45

February 2012] NOTES 151

n Tn (x)
1 x
2 2x 2 1
3 4x 3 3x
4 8x 4 8x 2 + 1
5 16x 5 20x 3 + 5x
6 32x 6 48x 4 + 18x 2 1
7 64x 7 112x 5 + 56x 3 7x

Chebyshev polynomials have integer coefficients and satisfy Tn (cos ) = cos n

for any R. The induced mappings a 7 Tn (a) are of particular interest, in part
because the identity Tn Tm = Tnm = Tm Tn implies that any two such mappings
commute. The Chebyshev polynomials induce especially remarkable mappings on
the rings Z/ pZ for prime p: for instance, if f (x) (Z/ pZ)[x] has degree at most
p 1/4 , and the map a 7 f (a) describes a bijection Z/ pZ Z/ pZ, then f is a com-
position of Chebyshev polynomials, cyclic polynomials x d , and linear polynomials1
[1, 5]. The purpose of this note is to analyze the collection of maps Z/ pZ Z/ pZ
which are induced by Chebyshev polynomials. Since there are only finitely many maps
Z/ pZ Z/ pZ of any sort, there must be infinitely many pairs (n, m) of distinct pos-
itive integers such that Tn and Tm induce the same map Z/ pZ Z/ pZ. This leads to
the following questions.
(1) When do Tn and Tm induce the same map Z/ pZ Z/ pZ?
(2) For fixed p, how many distinct maps Z/ pZ Z/ pZ are induced by Chebyshev
We can say more about the structure of the collection of maps Z/ pZ Z/ pZ
induced by Chebyshev polynomials, which we will call Chebyshev maps. For, the
identity Tn Tm = Tnm implies that (for a fixed prime p) the set of Chebyshev maps
Z/ pZ Z/ pZ is closed under composition, and hence forms a semigroup. This fact
already distinguishes Chebyshev polynomials from most other classes of polynomials,
and raises the following question.
(3) What is the structure of the semigroup of Chebyshev maps Z/ pZ Z/ pZ?
As often happens, the prime p = 2 behaves differently from other primes. The
answers to our questions for p = 2 are as follows.

Theorem 1. The polynomials Tn and Tm induce the same map Z/2Z Z/2Z if and
only if n m (mod 2). There are a total of two Chebyshev maps Z/2Z Z/2Z,
namely the identity and the constant map 1. These form a semigroup isomorphic to
Z/2Z under the operation of multiplication.

For odd primes p, the answers to our questions are as follows.

Theorem 2. Let p be an odd prime. The polynomials Tn and Tm induce the same map
Z/ pZ Z/ pZ if and only if n is congruent to either m or pm modulo ( p 2 1)/2.
1 The coefficients of these linear polynomials are only required to lie in the algebraic closure of Z/ pZ.


The number of distinct Chebyshev maps Z/ pZ Z/ pZ is ( p + 1)( p + 3)/8. The
semigroup of Chebyshev maps Z/ pZ Z/ pZ is isomorphic to the quotient of the
multiplicative semigroup Z/(( p 2 1)/2)Z by the subgroup {1, 1, p, p}.

Before proving these results, we illustrate Theorem 2 by writing it out in the two
smallest cases.
When p = 3, there are three Chebyshev maps on Z/ pZ, induced by T1 , T2 , and
T4 . Here T1 is the identity, T4 is the constant map 1, and T2 T2 = T4 . These three
maps comprise the quotient of the semigroup Z/4Z (under multiplication) by the
subgroup {1, 3}; the cosets of this subgroup are {1, 3}, {0}, and {2}.
When p = 5, there are six Chebyshev maps on Z/ pZ. These maps correspond to
the cosets of the subgroup {1, 5, 7, 11} of the semigroup Z/12Z (under multiplica-
tion), namely,

{0}, {1, 5, 7, 11}, {2, 10}, {3, 9}, {4, 8}, {6}.

Here a prescribed coset corresponds to the map a 7 Tn (a), where n is any positive
integer whose image in Z/12Z lies in the prescribed coset. The coset containing 1 is
the identity element, and in this case it is the only invertible element in the quotient
semigroup. Note that the cosets have sizes 1, 2, and 4. This also holds for larger
primes, and will be made explicit in the proof of Theorem 2.

2. EVEN CHARACTERISTIC. In this section we prove the following result, which

implies Theorem 1.

Proposition 3. If n is even then Tn (x) 1 (mod 2); if n is odd then Tn (x) x

(mod 2).

We begin with an alternate development of Chebyshev polynomials. For any pos-

itive integer n, the fundamental theorem of symmetric polynomials [6, p. 99] im-
plies that there is a unique f Z[x, y] such that f (u + v, uv) = u n + v n . More-
over, f (t x, t 2 y) is homogeneous in t of degree n, so f (x, y) = i=0 f i x n2i y i for

some integers f i . Now put g(x) := f (x, 1) = i=0 f i x n2i , so that g(u + u 1 ) =
u n + u n . Then h(x) := g(2x)/2 satisfies h((z + z 1 )/2) = (z n + z n )/2, which for
z = ei implies that h(cos ) = cos n . Hence h Tn vanishes at cos , and since is
arbitrary it follows that h = Tn .
We now determine the lowest-degree term of h, and use it to compute the reduction
of h mod 2. If n is even then substituting u = v yields

2v n = (v)n + v n = f (0, v 2 ) = f n/2 (v 2 )n/2 ,

so that f n/2 = 2 (1)n/2 . Since h = i=0 f i x n2i 2n2i1 and each f i is an integer,
it follows for even n that h 1 (mod 2). If n is odd then
u n + vn
f i (u + v)n12i (uv)i ;
u+v i=0

substituting u = v on the right yields f (n1)/2 (v 2 )(n1)/2 , and evaluating the left
side at u = v (for instance, via lHopitals rule) yields nv n1 . Thus we find that

February 2012] NOTES 153

f (n1)/2 = n (1)(n1)/2 is odd, so that h x (mod 2). This proves the proposition
(and more).

3. ODD CHARACTERISTIC. In this section we prove Theorem 2. Let p be an odd

prime, and write F p and F p for the field Z/ pZ and its algebraic closure. As noted in
the previous section, Tn ((z + z 1 )/2) = (z n + z n )/2.

Lemma 4. For any F p , the number of elements F p such that + 1 = is
either one or two, and if it is two then the elements are reciprocals of one another.

Proof. For F p , the equality + 1 = holds precisely when is a root of

x 2 x + 1. But this polynomial has either one or two roots in F p , and if it has two
then they are reciprocals.

For any F p , write = + 1 with as in the lemma; then, since pth power-
ing is an automorphism of F p which fixes , we have

p + p = p = = + 1 ,

so the lemma implies that p {, 1 }, whence p1 = 1. Conversely, if F p sat-

isfies p1 = 1, then + 1 is fixed by pth powering, and hence lies in F p . Thus the
elements of F p are precisely the elements ( + 1 )/2 where F p and p1 = 1.
Now, if p1 = 1 then

+ 1 + 1
Tn = Tm n + n = m + m
2 2
either n = m or n = m
either nm = 1 or n+m = 1.

Letting n denote the map F p F p defined by a 7 Tn (a), it follows that n = m

if and only if every ( p 1)th root of unity in F p is either an (n m)th root of unity
or an (n + m)th root of unity. Since F p contains both primitive ( p + 1)th roots of
unity and primitive ( p 1)th roots of unity, this says that n = m if and only if
n m (mod p + 1) and n m (mod p 1), or equivalently, n m or pm
(mod ( p 2 1)/2).
We have shown that the number of maps F p F p induced by Chebyshev poly-
nomials equals the number of orbits of the action of multiplication by {1, 1, p, p}
on residue classes mod ( p 2 1)/2. There are precisely two orbits of size 1, namely
{0} and {( p 2 1)/4}. The orbits of size 2 are {k( p 1)/2} for k = 1, 2, . . . ,
( p 1)/2 and {`( p + 1)/2} for ` = 1, 2, . . . , ( p 3)/2. The remaining
( p 2 4 p + 3)/2 residue classes split into orbits of size 4. Hence the number of
distinct orbits, which equals the number of distinct maps n , is ( p 2 + 4 p + 3)/8.
Finally, since Tn Tm = Tnm , the map n 7 n is a semigroup homomorphism from
the multiplicative semigroup of positive integers to the semigroup of maps F p F p
induced by Chebyshev polynomials. Since we showed above that n = m precisely
when n m or pm (mod ( p 2 1)/2), it follows that the semigroup of Cheby-
shev maps F p F p is isomorphic to the quotient of the multiplicative semigroup
Z/(( p 2 1)/2)Z by the subgroup {1, p}.


4. FINAL REMARKS. It would be interesting to consider similar questions over
more general fields or rings. Proposition 3 shows that if K is any commutative ring
of characteristic 2 then the identity map and the constant map 1 are the only maps
K K induced by Chebyhsev polynomials. If K is a finite field whose order q is
odd, then the proof of Theorem 2 shows that the number of Chebyshev maps K K
is (q + 1)(q + 3)/8, and the semigroup of Chebyshev maps is the quotient of the
multiplicative semigroup Z/((q 2 1)/2)Z by the subgroup {1, 1, q, q}.
Theorem 2 implies that, for any odd prime p, the group of permutations of Z/ pZ
induced by Chebyshev polynomials, or equivalently the group of invertible elements
in our semigroup, is the quotient group (Z/(( p 2 1)/2)Z) /h1, pi. This recovers
the main result of [3].
Finally, we note that when examining Chebyshev-like mappings of arbitrary fields
K , it is often convenient to treat the related class of Dickson polynomials. These are
defined for any positive integer n and any K by
n nk
()k x n2k
Dn (x, ) =
nk k

n nk
(it turns out that nk is an integer). If 2 6 = 0 then the Dickson polynomial is

related to the Chebyshev polynomial via the change of variables

Dn (x, ) = 2 Tn .

ACKNOWLEDGMENTS. We thank Florian Block, Kevin Carde, Jeff Lagarias, and the referees for valuable
suggestions which improved the exposition in this paper. The first, third and fourth authors were partially sup-
ported by the NSF under grants DMS-0502170, DMS-0801029, and DMS-0903420, respectively. The second
author was supported by an NSF Graduate Research Fellowship.


1. M. Fried, On a conjecture of Schure, Michigan Math. J. 17 (1970) 4155; available at http://dx.doi.

2. R. Lidl, G. L. Mullen, The worlds most interesting class of integral polynomials, J. Combin. Math.
Comnin. Comput. 37 (2001) 87100.
3. W. Nobauer, Uber eine Klasse von Permutationspolynomen und die dadurch dargestellten Gruppen, J.
Reine Angew. Math. 231 (1968) 216219.
4. T. J. Rivlin, Chebyshev Polynomials: From Approximation Theory to Algebra and Number Theory, second
edition. John Wiley, New York, 1990.
5. G. Turnwald, On Schurs conjecture, J. Austral. Math. Soc. Ser. A 58 (1995) 312357; available at http:
6. B. L. van der Waerden, Algebra, Vol. I. Springer-Verlag, New York, 1991.

Department of Mathematics, University of Michigan, Ann Arbor, MI 48109

{rosenjh, zscherr, blweiss, zieve}

February 2012] NOTES 155

Collapsing Walls Theorem
Igor Pak and Rom Pinchasi

Abstract. Let P R3 be a pyramid with the base a convex polygon Q. We show that when
other faces are collapsed (rotated around the edges onto the plane spanned by Q), they cover
the whole base Q.

1. INTRODUCTION. Let P be a convex pyramid in R 3 over the base Q, which is a

convex polygon in a horizontal plane. Think of the other faces F of P as the walls of
a wooden box, and that each wall F is hinged to the base Q along the edge. Suppose
now that the walls are collapsed, i.e., rotated around the edges towards the base onto
the horizontal plane. The question is: do they cover the whole base Q?
At first, this may seem obvious, but in fact the problem is already nontrivial even in
the case of four-sided pyramids, which can possibly have some obtuse dihedral angles
(see Figure 1). Formally, we have the following result:

Figure 1. An impossible configuration of four collapsing walls of a pyramid leaving a hole in the base.

Collapsing Walls Theorem. Let P R3 be a pyramid over a convex polygon Q. For

a face F of P, denote by e F the edge between F and the base: e F = F Q, and let
A F denote the result of rotation of F around e F in the direction of P, onto the plane
which contains Q. Then

Q F A F ,

where the union is over all faces F of P, different from Q.

For example, suppose pyramid P in the theorem has a very large height, and all
walls are nearly vertical. The theorem then implies that every point O Q has an
orthogonal projection into the interior of some edge e of Q. This is a classical result
with a number of far-reaching generalizations (see [4, 9]). Thus, the collapsing walls
theorem can be viewed as yet another generalization of this result (see Section 3).

2. PROOF OF THE THEOREM. Consider R3 endowed with the standard Carte-

sian coordinates (x, y, z). Without loss of generality assume that the plane H spanned
by Q is horizontal, i.e., given by z = 0, and that P is contained in the half-space z 0.
MSC: Primary 52B10


Denote by F1 , . . . , Fm the faces of P different from Q. For 1 i m, denote by Hi
the plane spanned by Fi and by ei = Fi Q the edge of Q adjacent to Fi .
Denote by 8i the rotation about ei of Hi onto H (the rotation is performed in the
direction dictated by P, so that throughout the rotation Hi intersects the interior of P).
Similarly, let Ai = 8i (Fi ) be the rotation of the face F of P onto Q, 1 i m. We
need to show that every point in Q lies in i=1 Ai . Without loss of generality and in
order to simplify the presentation, we can take this point to be the origin O.
Further, denote by L i = Hi H the line through ei . Let ri be the distance from
the origin to L i , and let i be the dihedral angle of P at ei , i.e., the angle between H
and Hi which contains P.
Suppose now F1 is a face such that
i = ri tan is minimized at 1 .
We will show that the origin O is contained in A1 . In other words, we prove that if
O / A1 , then i < 1 for some i > 1.
Let B H1 be such that the rotation of B onto Q is the origin: 81 (B) = O. It
suffices to show that B F1 . Let v = (a, b, 0) be the unit vector that is normal to L 1
in the horizontal plane and pointing outwards from Q. It is easy to see that

O B = r1 (1 cos 1 )a, r1 (1 cos 1 )b, r1 sin 1 .

To prove the theorem, assume to the contrary that B / F1 . Then there exists a face
of P, say F2 , such that H2 separates B from the origin. Denote by C the closest point
to B on L 2 , and by 0 the angle between the line BC and the horizontal plane H , where
the angle is taken with the half-plane of H which contains Q (and thus the origin). In
this notation, the above condition implies that 0 > 2 .
Without loss of generality we may assume that line L 2 is given by the equations
y = r2 and z = 0. Then

C = r1 (1 cos 1 )a, r2 , 0 ,

r2 r1 (1 cos 1 )b
cos 0 = cos OCB
d =q .
r12 sin2 1 + (r2 r1 (1 cos 1 )b)2

Note that the quantity t/ a 2 + t 2 is monotone increasing as a function of t, and that
b 1. We get

r2 r1 (1 cos 1 )
cos 0 q .
r12 sin2 1 + (r2 r1 (1 cos 1 ))2

Applying cos 0 < cos 2 , we conclude:

r2 r1 (1 cos 1 )
q < cos 2 . (1)
r12 sin2 1 + (r2 r1 (1 cos 1 ))2

Recall the assumption that 1 2 . This gives r1 tan 21 r2 tan 22 , or

February 2012] NOTES 157

r2 tan 21
. (2)
r1 tan 22

The rest of this section is dedicated to showing that (1) and (2) cannot both be true.
This gives a contradiction with our assumptions and proves the claim. We split the
proof into two cases depending on whether the dihedral angle 2 is acute or obtuse. In
each case we repeatedly rewrite (1) and (2), eventually leading to a contradiction.

Case 1 (obtuse angles). Suppose 2
< 2 < . In this case cos 2 < 0 and r2 r1 (1
cos 1 ) < 0. Now (1) implies

r12 sin2 1 1
1+ < , (3)
(r2 r1 (1 cos 1 ))2 cos2 2
r1 sin 1
> tan 2 . (4)
r2 r1 (1 cos 1 )
This can be further rewritten as:
r2 sin 1
< 1 cos 1 + . (5)
r1 tan 2
Now (5) and (2) together imply

tan 21 sin 1
2 < 1 cos 1 + ,
tan 2 tan 2

which is impossible. Indeed, suppose for some and satisfying 0 < , < we
tan 2 sin

< 1 cos + . (6)
tan 2

Dividing both sides by tan 2 , after some easy manipulations, we conclude that (6) is
equivalent to

1 1 + cos

< sin + , (7)
tan 2

which in turn is equivalent to

1 1
sin < cos( ). (8)
tan 2 tan

Since the left-hand side of (8) is equal to 1, we get a contradiction and complete the
proof in Case 1.

Case 2 (right and acute angles). Suppose now that 0 < 2 2 . Then cos 2 0, and
0 < tan 22 1. Let us first show that the numerator of (1) is nonnegative, i.e., that r2
r1 (1 cos 1 ). From the contrary assumption we have r2 /r1 < (1 cos 1 ). Together


with (2), this implies:

r2 tan 21 1
1 cos 1 > tan ,
r1 tan 22 2

which is impossible whenever 0 < 1 < .

From the above, we can exclude the right angle case 2 = 2 , for otherwise the
left-hand side of (1) is nonnegative, while the right-hand side is equal to zero. Thus,
cos 2 > 0.
Therefore, the inequality (1) in this case can be rewritten as

r12 sin2 1 1
1+ > , (9)
(r2 r1 (1 cos 1 ))2 cos2 2
r1 sin 1
> tan 2 . (10)
r2 r1 (1 cos 1 )
Note now that (10) coincides with (4). Since (6) does not hold for any and satisfy-
ing 0 < , < , we obtain the contradiction verbatim as in the proof of Case 1. This
completes the analysis of Case 2 and finishes the proof of the theorem.


3.1. The collapsing walls theorem extends verbatim to higher dimensions. Moreover,
it also extends to every polytope P Rd , as follows. For each facet F of P, let HF
denote the hyperplane containing F. Fix one facet Q of P. If all other facets F of P
are rotated around the affine subspace HF HQ onto HQ (or if HF is parallel to HQ we
just consider the orthogonal projection of F onto HQ ), then they cover the whole facet
Q. We refer to [5], where this result is proved in full generality. We should mention
that after we advertised the result in this paper, other people (we should mention here
personal communications with Arseniy Akopyan and independently with Gunter Rote)
found alternative elementary and beautiful proofs that are not more complicated and
perhaps technically even easier for the simple three-dimensional case presented in this
paper. However, we have not yet seen another proof (simple or not) that generalizes to
higher dimensions or to the case of a general convex polytope, as does the argument
in this paper.

3.2. Let us note that when the walls of a pyramid are collapsed outside, rather than
onto the base, they are pairwise nonintersecting (see Figure 2). We leave this easy
exercise to the reader.

Figure 2. Walls of a pyramid collapsing outside the base do not intersect.

February 2012] NOTES 159

3.3. Continuing with the example of vertical walls as given in the introduction right
after the theorem, recall that for the center of mass O = cm(Q), there are at least two
edges onto which orthogonal projection of O lies in the interior (see, e.g., [4, 9]).1 It
would be interesting to see if this result extends to the setting of the theorem (of course,
the notion of the center of mass would have to be modified appropriately). Let us note
here that the center of mass result is closely related to the four vertex theorem [6], and
fails in higher dimension [2].

3.4. The proof of the theorem is based on an implicit subdivision of Q given by the
smallest of the linear functions i at every point O Q. Recall that i is a weighted
distance to the edge ei . Thus this subdivision is in fact a weighted analogue of the
dual Voronoi subdivision in the plane (see [1, 3]). As a consequence, computing this
subdivision can be done efficiently, both theoretically and practically.

ACKNOWLEDGMENTS. The authors are thankful to Yuri Rabinovich for his interest in the problem. The
first author was partially supported by the National Security Agency and the National Science Foundation. The
second author was supported by the Israeli Science Foundation (grant No. 938/06).


1. F. Aurenhammer, Voronoi diagramsA survey of a fundamental geometric data structure, ACM Comput.
Surv. 23 (1991) 345405; available at
2. J. H. Conway, M. Goldberg, R. K. Guy, Problem 66-12, SIAM Review 11 (1969) 7882; available at
3. S. Fortune, Voronoi Diagrams and Delaunay Triangulations, Computing in Euclidean Geometry, second
edition. Edited by F. Hwang and D.-Z. Du, Lecture Notes Ser. Comput. 4, 225265, World Scientific,
Singapore, 1995.
4. I. Pak, Lectures on Discrete and Polyhedral Geometry, monograph (to appear). Available at http://www.
5. I. Pak, R. Pinchasi, How to cut out a convex polyhedron, (to appear).
6. S. Tabachnikov, Around four vertices, Russian Math. Surveys 45 (1990) 229230; available at http:

Department of Mathematics, UCLA, Los Angeles, CA

Mathematics Department, TechnionIsrael Institute of Technology, Haifa 32000, Israel

1 One can give a construction in which there is only one such edge, if the center of mass is replaced by a

general point in Q (see [2] and [4, 9]).


Edited by Gerald A. Edgar, Doug Hensley, Douglas B. West
with the collaboration of Mike Bennett, Itshak Borosh, Paul Bracken, Ezra A. Brown,
Randall Dougherty, Tamas Erdelyi, Zachary Franco, Christian Friesen, Ira M. Ges-
sel, Laszlo Liptak, Frederick W. Luttmann, Vania Mascioni, Frank B. Miles, Bog-
dan Petrenko, Richard Pfiefer, Cecil C. Rousseau, Leonard Smiley, Kenneth Stolarsky,
Richard Stong, Walter Stromquist, Daniel Ullman, Charles Vanden Eynden, Sam Van-
dervelde, 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 April 30, 2012. Additional information, such as general-
izations and references, is welcome. The problem number and the solvers 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.

11621. Proposed by Z. K. Silagadze, Budker Institute of Nuclear Physics and Novosi-
birsk State University, Novosibirsk, Russia. Find
Z Z s1 Z s2 Z s3
cos(s12 s22 ) cos(s32 s42 ) ds4 ds3 ds2 ds1 .
s1 = s2 = s3 = s4 =

11622. Proposed by Oleh Faynshteyn, Leipzig, Germany. In triangle ABC, let A1 ,

B1 , C1 be the points opposite A, B, C at which symmedians of the triangle meet the
opposite sides. Prove that
m a (c cos 1 b cos 2 ) + m b (a cos 1 c cos 2 ) + m c (b cos 1 a cos 2 ) = 0,
m a (sin 1 sin 2 ) + m b (sin 1 sin 2 ) + m c (sin 1 sin 2 ) = 0, and
m a (cos 1 + cos 2 ) + m b (cos 1 + cos 2 ) + m c (cos 1 + cos 2 ) = 3s,
where a, b, c are the lengths of the sides, m a , m b , m c are the lengths of the medians, s
is the semiperimeter, 1 = C A A1 , 2 = A1 AB, and similarly with the j and j .
11623. Proposed by Aruna Gabhe, Pendharkars College, Dombivali, India, and M.
N. Deshpande, Nagpur, India. A fair coin is tossed n times and the results recorded
as a bit string. A run is a maximal subsequence of (possibly just one) identical tosses.
Let the random variable X n be the number of runs in the bit string not immediately
followed by a longer run. (For instance, with bit string 1001101110, there are six runs,
of lengths 1, 2, 2, 1, 3, and 1. Of these, the 2nd, 3rd, 5th, and 6th are not followed by
a longer run, so X 10 = 4.) Find E(X n ).
11624. Proposed by David Callan, University of Wisconsin, Madison, WI, and Emeric
Deutsch, Polytechnic Institute of NYU, Brooklyn, NY. A Dyck n-path is a lattice path of


n upsteps U (changing by (1, 1)) and n downsteps D (changing by (1, 1)) that starts
at the origin and never goes below the x-axis. A peak is an occurrence of U D and the
peak height is the y-coordinate of the vertex between its U and D.
The peak heights multiset of a Dyck path is the set of peak heights for that Dyck
path, with multiplicity. For instance, the peak heights multiset of the Dyck 3-path
UU DU D D is {2, 2}. In terms of n, how many different multisets occur as the peak
heights multiset of a Dyck n-path?
11625. Proposed by Lane Bloome, Peter Johnson, and Nathan Saritzky (students)
Auburn University Research Experience for Undergraduates in Algebra and Discrete
Mathematics 2011. Let V (G), E(G), and (G) denote respectively the vertex set,
edge set, and chromatic number of a simple graph G. For each positive integer n, let
g(n) and h(n) respectively denote the maximum and the minimum of (G) + (H )
(G H ) over all pairs of simple graphs G and H with |V (G) V (H )| n and
E(G) E(H ) = . Find g(n) and limn h(n)n
11626. Proposed by Cezar Lupu, University of Pittsburgh, Pittsburgh, PA. Let x1 , x2 ,
and x3 be positive numbers such that x1 + x2 + x3 = x1 x2 x3 . Treating indices modulo
3, prove that
3 3 3
X 1 X 1 X 1 3
q 2
+ q .
1 xk2 + 1 1
xk + 1 1 (x 2 + 1)(x 2 + 1) 2
k k+1

11627. Proposed by Samuel Alexander, The Ohio State University, Columbus, Ohio.
Let N be the set of nonnegative integers. Let M be the set of all functions from N
to N. For a function f 0 from an interval [0, m] in N to N, say that f extends f 0 if
f (n) = f 0 (n) for 0 k m. Let F( f 0 ) be the set of all extensions in M of f 0 , and
equip M with the topology in which the open sets of M are unions of sets of the form
F( f 0 ). Thus, { f M : f (0) = 7 and f (1) = 11} is an open set. S T
Let S be a proper subset of M that can be expressed both as iN jN X i, j and as
iN jN Yi, j , where each set X i, j or Yi, j is a subset of M that is both closed
and open
(clopen). Show that there is a family Z i, j of clopen sets such that S = iN jN Z i, j
and S = iN jN Z i, j .


Eigenvalues of Sums and Differences of Idempotent Matrices

11466 [2009, 845]. Proposed by Tian Yongge, Central University of Finance and
Economics, Beijing, China. For a real symmetric n n matrix A, let r (A), i + (A),
and i (A) denote the rank, the number of positive eigenvalues, and the number of
negative eigenvalues of A, respectively. Let s(A) = i + (A) i (A). Show that if P
and Q are symmetric n n matrices, P 2 = P, and Q 2 = Q, then i + (P Q) =
r (P + Q) r (Q), i (P Q) = r (P + Q) r (P), and s(P Q) = r (P) r (Q).
Solution by Oskar Maria Baksalary, Adam Mickiewicz University, Poznan, Poland, and
Gotz Trenkler, Dortmund University of Technology, Dortmund, Germany. We solve
the more general problem in which idempotent P and Q are Hermitian with complex
entries. We view them as n n complex orthogonal projectors.
The solution is based on a joint decomposition of the projectors P and Q. Let P
have rank , where 0 < n. By the Spectral Theorem, there is an n n unitary


matrix U such that
I 0
P =U U ,
0 0

where I is the identity matrix of order and U denotes the conjugate transpose of
U . We use this expression for P to partition the projector Q. Using the same matrix
U , we write
Q=U U ,

where A and D are Hermitian matrices of orders and n , respectively. Let A =

I A. Since Q 2 = Q, we have A = A2 + B B = A A + B B . Since A A and B B
are both nonnegative definite, A is also nonnegative definite. With R() denoting the
column space of a matrix argument, we obtain

R( A) = R( A A + B B ) = R( A A + R(B B ) = R( A) + R(B),
and hence R(B) R( A). Other relationships among A, B, and D are found in Lem-
mas 15 of [1]; we use two of these. The first expresses the orthogonal projector PD
onto the column space of D as PD = D + B A B, where A is the MoorePenrose
inverse of A. The second expresses the rank of A as r ( A) = r (A) + r (B). Fur-
thermore, Theorem 1 of [1] gives r (Q) = r (A) r (B) + r (D), and Lemma 6 of [1]
gives r (P + Q) = + r (D). Taking differences of these expressions yields r (P +
Q) r (Q) = r (A) + r (B) and r (P + Q) r (P) = r (D). Since the third of the
desired equations is just the difference of the first two, it suffices to show that P Q
has r ( A) positive eigenvalues and r (D) negative eigenvalues.
Theorem 5 in [1] expresses P Q as
PQ=U U ,

which can be rewritten as

I 0 A 0 I A B
PQ=U U .
B A In 0 PD 0 In

The matrices before and after the central matrix in the product on the right are nonsin-
gular and are conjugate transposes of each other. By Sylvesters Law of Inertia (see [2,
Section 1.3]), the numbers of positive and negative eigenvalues are unchanged by con-
jugation. Since A and PD are nonnegative definite (the eigenvalues of an idempotent
matrix lie in the interval [0, 1]), we conclude that P Q has r ( A) positive eigenvalues
and r (D) negative eigenvalues, as desired.
[1] O. M. Baksalary and G. Trenkler, Eigenvalues of functions of orthogonal projec-
tors, Linear Alg. Appl. 431 (2009) 21722186.
[2] R. A. Horn, F. Zhang, Basic properties of the Schur complement, in The Schur
Complement and its Applications, edited by F. Zhang, Springer Verlag, New York,
2005, 1746.
Also solved by R. Chapman (U. K.), E. A. Herman, O. Kouba (Syria), J. H. Lindsey II, K. Schilling, J. Simons
(U. K.), R. Stong, Z. Voros (Hungary), S. Xiao (Canada), GCHQ Problem Solving Group (U. K.), and the


A Hankel Determinant Limit
11471 [2009, 941]. Proposed by Finbarr Holland, University College Cork, Cork,
Ireland. Let A be an r r matrix with distinct eigenvalues 1 , . . . , r . For n 0,
let a(n) be the trace of An . Let H (n) be the r r Hankel matrix with (i, j) entry
a(i + j + n 2). Show that
lim |det H (n)| 1/n
= |k |.

Solution by Jim Simons, Cheltenham, U. K. The eigenvalues of An are n1 , . . . , rn ,

so a(n) = rk=1 nk . Therefore, H (n)i, j = rk=1 k
P P n+i+ j2
. It is well known that the
matrix V , given by Vi, j = i for i, j {1, . . . , n}, has determinant
j<i (i j ). Multiplying row i of V by i yields a matrix V (n) in which V (n)i, j =

i , having determinant ( k=1 k ) j<i (i j ). With V denoting the transpose

n+ j1 Qr n
Q 0

of V ,
X r
(V 0 V (n))i, j = i1
k k k = H (n)i, j .
n+ j1 n+i+ j2
k=1 k=1

Y n Y
det H (n) = det(V 0 V (n)) = k (i j )2 .
k=1 j<i

The second factor is constant, so its nth root tends to 1.

Editorial comment. Simons observed that the computation of det H (n) is valid over
any field.
Also solved by R. Chapman (U. K.), M. Goldenberg & M. Kaplan, J.-P. Grivaux (France), E. A. Herman,
O. Kouba (Syria), J. H. Lindsey II, O. P. Lossers (Netherlands), R. Stong, E. I. Verriest, GCHQ Problem
Solving Group (U. K.), Microsoft Research Problems Group, and the proposer.

Pretty Boxes All in a Row

11477 [2010, 86]. Proposed by Antonio Gonzalez, Universidad de Sevilla, Seville,
Spain, and Jose Heber Nieto, Univesidad del Zulia, Maracaibo, Venezuela. Several
boxes sit in a row, numbered from 0 on the left to n on the right. A frog hops from
box to box, starting at time 0 in box 0. If at time t, the frog is in box k, it hops one
box to the left with probability k/n and one box to the right with probability 1 k/n.
Let pt (k) be the probability that the frog launches its (t + 1)th hop from box k. Find
limi p2i (k) and limi p2i+1 (k).
Solution by Robin Chapman, University of Exeter, Exeter, U. K. We show that
limi p2i (k) is nk /2n1 when k is even and 0 when k is odd. Also, limi p2i+1 is
when k is odd and 0 when k is even.
In standard language, we have a Markov chain with states 0, . . . , n and transition
probabilities pk,k1 = k/n and pk,k+1 = 1 k/n (all others equal 0). This Markov
chain is periodic with period 2, since the state switches parity on each move. Thus
p j (k) = 0 when j and k have opposite parity.
Taking two hops at once converts the Markov chain into two others, one on the odd
states and one on the even states. Each is ergodic and thus has a unique stationary


distribution. (For a chain to be ergodic it suffices that one can reach any state from any
other and that it is possible to remain in the current state at any step.) The stationary
distributions by definition are the limits to be computed, so the limits exist.
Let ak = limi p2i (k) for even k and ak = limi P p2i1 (k) for odd k. In order for
these to form stationary distributions, we must have nk=0 ak = 2 and
nk+1 k+1
ak = ak1 + ak+1
n n
for 0 k n (with a1 = an+1 = 0). These linear equations determine the n + 1 val-
ues {ak }nk=0 . Therefore, it suffices to check that setting ak = nk 2n1 for all k satisfies

the equations.
Editorial comment. Stephen J. Herschkorn wrote The problem begs the question as to
why, from a probabilistic point of view, the binomial should be the stationary distribu-
tion for this simple random walk. Herschkorn communicated the following intuition
from Sheldon Ross: Flip n fair coins; the number of heads has a binomial distribution.
Pick a random coin and turn it over. The new number of heads arises from the old by
the same transition probability as in the random-walk model, but the new number of
heads still has the binomial distribution, because each coin still has probability 1/2 of
being heads.
Daniel M. Rosenblum noted a similarity to Problem 11032 (2003, 637), in which
the frogs probabilities of jumping to the right and left are reversed, yielding the same
Markov chain as the Ehrenfest urn model (see, for example, Sections 4 and 5 of M.
Kac, Random Walk and the Theory of Brownian Motion, Amer. Math. Monthly 54
(1947) 269391.)
Some solvers used generating functions. It is also possible to avoid mentioning the
theorem on stationary distributions and instead prove that the limits exist by direct
methods particular to the problem.
Also solved by A. Agnew, M. Andreoli, D. Beckwith, K. David & P. Fricano, D. Fleischman, O. Geupel
(Germany), C. Gonzalez-alcon & A. Plaza (Spain), S. J. Herschkorn, O. Kouba (Syria), J. H. Lindsey II,
O. P. Lossers (Netherlands), D. M. Rosenblum, R. K. Schwartz, J. Simons (U. K.), N. C. Singer, R. Stong,
R. Tauraso (Italy), M. Tetiva (Romania), GCHQ Problem Solving Group (U. K.), NSA Problems Group, Skid-
more College Problem Group, and the proposer.

Separating the Degrees of Polynomials

11478 [2010, 87]. Proposed by Marius Cavachi, Ovidius University of Constanta,
Constanta, Romania. Let K be a field of characteristic 0, and let f and g be relatively
prime polynomials in K [x] with deg(g) < deg( f ). Suppose that for infinitely many
in K there is a sublist of the roots of f + g (counting multiplicity) that sums to 0.
Show that deg(g) < deg( f ) 1 and that the sum of all the roots of f (again counting
multiplicity) is 0.
Solution by Richard Stong, Center for Communications Research, San Diego, CA. For
a monic polynomial p of degree n with roots 1 , . . . , n (taken with multiplicity) the
product Q k defined by
Qk = (i1 + + ik )
1i 1 <i 2 <<i k n

is a symmetric function in the roots of p. Hence Q k is given by a universal polynomial

in the coefficients of p. When p is a constant multiple of f + g (choosing the constant
to make p monic), Q k is a polynomial in . By hypothesis, there are infinitely many


values of such that nk=1 Q k vanishes. Hence one of these polynomials, say Q j , is
the 0 polynomial. Thus Q j vanishes for all , and the desired sublist exists for all .
The same conclusion holds even when we replace K by a larger field, specifically
the field K (t) of rational functions in a new indeterminate t. By Gausss Lemma,
if the polynomial f (x) + tg(x) is reducible over K (t), then it is reducible over the
polynomial ring K [t]. However, since it is linear in t, one of the factors would be in-
dependent of t and would give a common factor of f and g. Thus f (x) + tg(x) is ir-
reducible over K (t). Hence its Galois group G acts transitively on the roots 1 , . . . , n
of f (x) + tg(x). Suppose without loss of generality that 1 + + k = 0. Now
X |G|k
0= (1 + + k ) = (1 + + n ).

Thus 1 + + n = 0, and hence the coefficient of x n1 in f + tg vanishes. Now

deg(g) n 2, and the sum of the roots of f vanishes as desired.
Also solved by R. Chapman (U. K.), O. P. Lossers (Netherlands), and the proposer.

Orthogonality of Matrices under Additivity of Traces of Powers

11483 [2010, 182]. Proposed by Eric Pite, Paris, France. Let A and B be real n n
symmetric matrices such that tr (A + B)k = tr Ak + tr B k for every nonzero integer k.
Show that AB = 0.
Composite solution by the editors. We prove the stronger statement that if A and B
are n n Hermitian matrices such that tr (A + B)k = tr Ak + tr B k for every integer k
such that 1 k 3n, then AB = 0.
We show first that if the sums of the kth powers of two lists of complex numbers,
of length l and m respectively, are equal for 1 k l + m, then the lists are the same
(up to order of the entries). To see this, let the first list have distinct entries 1 , . . . , r
with multiplicities a1 , . . . , ar , and let the second list have P 1 , . . . , s
P distinct entries
with multiplicities b1 , . . . , bs . The hypothesis is now that ri=1 ai ik sj=1 b j kj = 0
for 1 k ai + b j = l + m. Since the Vandermonde matrix is invertible, the
hypothesis requires the lists to have the same entries.
This immediately yields the following: If S1 , S2 , and S3 are three lists of complex
numbers, and the sum of the kth powers of the entries in S1 and S2 equals the sum of
the kth powers of the entries in S3 whenever k is at most the sum of the lengths of the
three lists, then the entries of the concatenation of S1 and S2 are the same as the entries
in S3 .
Now let A and B be Hermitian matrices, and let C = A + B. Let the lists of
s t
nonzero eigenvalues of these matrices be {i }ri=1 , {i }i=1 , andP{ i }i=1 , respectively.
tr (A + B) = tr A + tr B is the same as i=1 i =
i=1 i +
k k k k
Pr k
Ps condition
i=1 i , imposed for 1 k 3n. Hence, the nonzero eigenvalues of C are ex-

actly the nonzero eigenvalues of A and B, including multiplicities. Consequently,

rank (C) = rank (A) + rank (B). On the other hand, the images satisfy Im(A + B)
Im(A) + Im(B). Thus, Im(A + B) = Im(A) + Im(B). Let V = Im(A + B). Viewed
as a linear transformation on V , C is invertible.
Finally, we argue that AB = 0. By spectral factorization, since A and B are Her-
mitian, there are orthonormal vectors {u i }ri=1 for A and {vi }i=1 for B such that A =
v v
Pr Ps
i=1 i u i u
i and B = j=1 j j

j . Moreover, the space V is spanned by {u i }ri=1 and
s r s
{vi }i=1 . Since r + s = dim V , it follows that {u i }i=1 {vi }i=1 is a linearly independent


set and forms a basis for V . It follows that
(A + B)u i = i u i + j v j (v j u i ) for 1 i r

(A + B)v j = i u i (u i v j ) + j v j for 1 j s.

Under the basis {u 1 , . . . , u r , v1 , . . . , vs }, the matrix representation of A + B is

D D E D 0 I E
= , (1)
D E D 0 D E I
where D = diag(1 , . . . , r ), D = diag(1 , . . . , s ), and E i, j = u i v j .
Since the nonzero eigenvalues of A + B are {i }ri=1 and {i }i=1 , the determinants of
both sides of (1) equal i=1 i i=1 i . This yields det E I = 1. Also, EI EI is just
Qr Qs I E

the Gram matrix of u 1 , . . . , u r , v1 , . . . , vs . By the Hadamard determinant inequality,
E = 0; that is, u i v j = 0 for all i and j. It follows that
X s
X  r X
X s
AB = i u i u i j v j v j = i j u i (u i v j )v j = 0.
i=1 j=1 i=1 j=1

Editorial comment. It would be nice to extend the result to normal matrices. The prob-
lem is that C = A + B is not normal when A and B are normal. Thus the rank of C is
not necessarily the same as the number of nonzero eigenvalues of C. Other than this,
everything works for normal matrices.
One may wonder whether the condition k 3n be replaced with k n. This
fails at least when n = 1, since tr (A + B) = tr A + tr B for all numbers A and B, but
AB 6= 0.
Also solved by J. Simons (U. K.), R. Stong, and the proposer.

Friendly Paths
11484 [2010, 182]. Proposed by Giedrius Alkauskas, Vilnius University, Vilnius,
Lithuania. An uphill lattice path is the union of a (doubly infinite) sequence of di-
rected line segments in R2 , each connecting an integer pair (a, b) to an adjacent pair,
either (a, b + 1) or (a + 1, b). A downhill lattice path is defined similarly, but with
b 1 in place of b + 1, and a monotone lattice is an uphill or downhill lattice path.
Given a finite set P of points in Z2 , a friendly path is a monotone lattice path for which
there are as many points in P on one side of the path as on the other. (Points that lie
on the path do not count.)
(a) Show that if N = a 2 + b2 + a + b for some positive integer pair (a, b) satisfying
a b a + 2a, then for some set of N points there is no friendly path.
(b)* Is it true that for every odd-sized set of points there is a friendly path?
Solution to (a) by the proposer. Let P be the centrally symmetric configuration con-
sisting of triangles of points in four quadrants as in the figure (where a = 4 and
b = 7). The first and third quadrants contain triangles meeting a diagonals, compris-
ing a(a + 1)/2 points. The second and fourth quadrants contain triangles meeting b




diagonals, comprising b(b + 1)/2 points. In total, |P| = N . Let A, B, C, D denote the
subsets in the four quadrants.
We prove that there is no friendly path for P. If the first and last points of P on
a monotone path Q lie in neighboring quadrants, then at least N /2 points lie on one
side, and Q is not friendly. If the first and last points are in C and A, then Q hits one
point in each of 2a + 1 diagonals. Since N is even, this leaves an odd number of points
of P outside Q, and they cannot be split equally.
It remains to consider a downhill lattice path Q whose first and last points are in
B and D. If Q hits a point of P at every step between these extremes, then Q hits
2b + 1 points of P, and again the remainder cannot be split equally. Hence, we may
assume that an odd number of lattice points along Q between its ends are not in P. By
symmetry, we may assume these points are in A. We claim that every such path has
more points of P below it than above it.
Consider the point x just above the leftmost column of the triangle in A. The down-
hill path Q containing x that has the most points of P above and to its right goes
directly rightward to x and then down. There are ba1 2
points of P above Q in B,
a b
points of P to the right of Q in A, and 2 points of P to the right of Q in D.
Meanwhile, on the other side of Q are a+1 b+1 ba+1
+ 2
points. An equal split
a b ba1 a+1 b+1 ba+2
+ + + ,
2 2 2 2 2 2
which simplifies to a + b (b a)2 . The left side is at least 2a, and the right side is
at most 2a, so b = a is necessary, but then 2a 0.
As we move from x to any other point in the first quadrant outside A as a point of Q
outside P between points of Q in P, the number of points above Q decreases, while
the number of points below Q increases. Hence the two sides can never have equal size.
Editorial comment. We do not know the answer to part (b)*. Parity considerations
made part (a) easy using a centrally symmetric configuration. However, a centrally
symmetric configuration of odd size has a central point. Any symmetric path through
that point is a friendly path. This makes it difficult to construct a counterexample.
No other solutions were received.


Edited by Jeffrey Nunemacher
Mathematics and Computer Science, Ohio Wesleyan University, Delaware, OH 43015

Roads to Infinity. The mathematics of truth and proof. By John Stillwell. A K Peters, Natick,
MA, 2010. xi + 203 pp., ISBN 978-1-56881-466-7. $39.95.

Reviewed by Jose Ferreiros

Mathematics appears to be an ever-unfolding dialectic between the finite and the in-
finite, between discrete structures and continuous forms, but also between symbolic
content and idealisation. In a well-known paper On the infinite, Hilbert proposed a
distinction between the contentual in mathematicswhich he tentatively identified
with the study of strictly finite structuresand the ideal elements that are constantly
being introduced to help explore the realm of mathematical truths. Hilbert also claimed
that the infinite is not to be found in Nature, yet it may still be the case that the infinite
occupies a well-justified place in our thinking, that it plays the role of an indispensable
concept [6]. Broadly construed, that is the topic of the book under review. More than
dealing with roads to infinity, or with infinity studied for its own sake (the topic of
higher set theory), the books core focus is roads from infinity, in the sense explained
In that same paper of 1925, Hilbert offered new clarifications of his celebrated pro-
gram to vindicate infinitarian mathematics by methods employing only finitary, con-
tentual mathematicsby means of a proof theory studying the structure of proofs
inside any given (axiomatized) theory, which would show that mathematics is consis-
tent, i.e., free from contradiction. 1 This became the springboard for Godels surprising
and celebrated Incompleteness Theorems, but also for new developments in proof the-
ory initiated in the 1930s by Gentzen, who was able to establish the consistency of the
axiomatic system called PA (Peano Arithmetic) by an extension of Hilbertian meth-
ods. This is the mathematics of truth and proof to which Stillwells book is devoted,
and in which he chooses to emphasize the contributions of Emil Post and Gerhard
Stillwell is a master expositor and does a very good job explaining and weaving
together many core issues in mathematical logic and foundational studies. Less than
halfway through the book the reader has reached the limitation results that affect for-
mal systems capable of codifying a certain amount of arithmetic: incompleteness, un-
provability of consistency, the halting problem in computation, the decision problem.
But, although Stillwell sets himself the task of dispelling the myth that incomplete-
ness is a difficult concept, I doubt that he has managed to do so. Diagonalization
is a simple and clear technique, and the author does very well presenting it in several
versionsfrom the original set theoretic one (Chapter 1) to the relevant proof-theoretic
1 He also had the courage to propose a way of solving the Continuum Problemwhich turned out to be
seriously flawed.

February 2012] REVIEWS 169

versions. But the idea of incompleteness is itself a sophisticated concept (as is even the
notion of a formal system), so less mature readers will have some trouble adequately
grasping the contents of Chapter 3, where these ideas are discussed. I believe many
readers will not be quite satisfied with the proof sketch provided on p. 75, and par-
ticularly not with the (vague) considerations of soundness and consistency offered at
this point in the book. Even more difficultsince vagueis the problem of obtaining
an adequate understanding of the links and differences between formal systems and
mathematical practice, which however is central to fully digesting the implications of
Godels and other limitation results.2
In this part of the book, Stillwell chooses to underscore the contributions of the
American (Polish-born) mathematician Emil Leon Post (18971954). In 1921, having
completed his Ph.D., Post developed ideas about unsolvable problems that brought
him close to the path-breaking results of Godel and Turing, but he also could not find
the direct link to arithmetic and Principia Mathematica later established by Godel.
And, feeling the need for full generality and a complete analysis, he refrained
from publishing. (Stillwell cites a 1938 postcard to Godel: As for any claims I might
make perhaps the best I can say is that I would have proved Godels Theorem in
1921had I been Godel.) Nonetheless, Post had found a notion of canonical or
normal system that he would publish many years later, and which turned out to be
equivalent to Turing machines (Churchs thesis says that all computation can be done
by normal systems).3
The core of the book deals with results that (to the best of my knowledge) have
not yet made it into the popular, or semi-popular, literature: Gentzens 1936 proof
of the consistency of Peano Arithmetic (PA) and later work on natural unprov-
able arithmetic sentences. Both are crucial post-Godel developments in proof theory.
Gerhard Gentzen (19091945) employed an extension of Hilbertian methods to in-
clude induction up to 0 as the basis to establish the consistency of PA (this is a
miniature version of transfinite induction; 0 is a countable transfinite ordinal, i.e.,
0 < 0 < 1 , to be defined below). And, years after Godels incompleteness result,
mathematicians started looking for unprovable sentences that arise more naturally,
unlike Godels sentence which is concocted metatheoretically. The examples that
Stillwell discusses are Goodsteins theorem, the finite and infinite Ramsey theorems,
Kruskals theorem in graph theory, and Friedmans finite form of it. 4
One of the most beautiful examples is Goodsteins theorem, which is about se-
quences of natural numbers arising from the following process. Take any natural
number and express it in base 2 normal form (with all digits at most 2), then replace
each base 2 by a 3; subtract 1 and write the result in base 3 normal form (with digits at
most 3) and then repeat the procedure, replacing each 3 by a 4, subtracting 1, and writ-
ing the result in base 4 normal form; and continue repeating. The numbers obtained
grow bigger and bigger, since the procedure goes on for very long, and the bases keep
rising, but only up to a point. Goodsteins theorem states that the sequence of numbers
thus obtained is finite and the process terminates at 0.

2 This has given rise to all kinds of misunderstandings, even among expert mathematicians. To suggest

some readings, I find very interesting Celluccis distinction between closed and open symbolic systems
(see, e.g., his [2]); a masterful exposition of mistaken readings of Godels theorems is provided in [4].
3 Stillwell does not mention that Post suffered all his adult life from crippling manic-depressive disease at

a time when no drug therapy was available for this malady (M. Davis). His case, together with those of Cantor
and Godel, has reinforced the popular, romantic notion that there is some link between logic and madness.
4 Kruskals theorem has to do with trees, finite graphs that are connected and contain no closed paths: For

any infinite sequence of trees T1 , T2 , T3 , . . . there are indices i < j such that Ti embeds T j (i.e., the infinite
sequence of trees contains an infinite increasing subsequence of trees embedding each other).


If we start with 3 = 2 + 1, the Goodstein sequence of numbers is 2 + 1, 3, 3, 2, 1, 0.
Very simple. Starting with number 4 the process already becomes quite intractable;
the sequence begins

22 ,
32 2 + 3 2 + 2,
42 2 + 4 2 + 1,
52 2 + 5 2,
62 2 + 6 + 5, . . .

and, according to Stillwell (following Kirby & Paris, see pages 4950), it reaches 0 at
base 3 2402653211 1. The Goodstein process looks very natural if you have studied
some set theory, to the point of becoming acquainted with countable ordinals and with
the fact that any descending sequence of countable ordinals will be finite. For the
Goodstein process is a finitary, arithmetic translation of the finiteness of descending
sequences of ordinals.
The interesting fact is that Goodsteins theorem, albeit a truth of arithmetic which
can be stated in PA, and whose statement involves only the basic operations of ad-
dition, product and exponentiation, cannot be proved in PA. Indeed, it can be proved
in PA that the Goodstein theorem (expressed in this theory by means of an axiom
schema) implies the consistency of PA. But then, by Godels second incompleteness
theorem, a.k.a. the unprovability of consistency, it follows that Goodsteins theorem is
a formally unprovable sentence of PA (better: schema of sentences).
Of course, Goodsteins theorem can be proved in set theory, for instance in ZFC.
The process by which the numbers in any Goodstein sequence (like the two we showed
above) change and eventually come down to 0 can be understood by considering de-
scending sequences of countable ordinals. Countable ordinals are objects like , 2 , or

2 2 + + 5, or even + + +3 + 3, written in what is called the Cantor normal
form; obviously there are infinitely many such regimented polynomials in . And
they are bounded above by 0 , which can be defined as the limit of the sequence

, , , . . . , or alternatively as the first ordinal such that = ;5 Cantor knew
this number already and established that it is merely countable. That is, our friend 0 is
the first countable ordinal that cannot be written as one of the above polynomials in .
For instance, the sequence for number 4 above corresponds to this (one may simply
interpret as a variable that takes value n + 1 in row n):

2 2 + 2 + 2,
2 2 + 2 + 1,
2 2 + 2,
2 2 + + 5 . . .

Given that such descending sequences of ordinals must be finite, Goodsteins theo-
rem is true. Actually the result can be proved in a theory much simpler than set theory:
it is sufficient to employ Gentzens expanded arithmetic, which adds 0 -Induction to
the system of Peano Arithmetic.
5 Of course there are uncountably many such numbers, even among the countable ordinals!

February 2012] REVIEWS 171

Here we observe the delicate interplays between truth and proof that Stillwell
wants to present to his readers. We see a beautiful example of infinitary mathematics
translated into the finitary, which is actually the main topic of Roads to Infinity; for
the book is not so much devoted to the higher infinite as to proof-theoretic results,
e.g., concrete arithmetical statements that are obtained by mirroring certain methods
thattruewere introduced in the set-theoretic study of infinity. That is what I mean
by roads from infinity. Stillwell himself quotes Terence Tao talking about principles
that allow one to tap the power of the infinitary world in order to establish results in
the finitary world, or at least to take the intuition gained in the infinitary world and
transfer it to a finitary setting (p. 163).
Now, what are the roads to infinity? And how do they turn into roads from infinity?
Stillwell is explicit about this. He considers two main roads to infinity, explored in
the first two chapters: 1. The road to transfinite cardinals through diagonalization,
which led mathematicians from the simple infinity of N to the infinity of real numbers
R, namely, to 20 (here Ill have something to qualify, see below). And 2. The road
of transfinite ordinals through principles of ordinal generation, which led Cantor to
(= 0 ) and + 1 and all the ordinals mentioned above, and which in a bold step
he brought to completion by considering the set of all countable ordinals, 1 , and
proceeding beyond.
The attractive idea of Stillwells presentation is to emphasize the parallel between
those two upward roads, and the crucial methods employed in proof theory by Godel
and Post, and by Gentzen. For diagonalization in proof-theoretic dressing is crucial
to the proof of incompleteness for systems that codify primitive recursive arithmetic.
And the set of all countable ordinals that can be constructed from , hence a minia-
turization of Cantors boundless generation of ordinals, is crucial to Gentzens proof
of the consistency of PA. (In order to prove that the set of all countable ordinals is
uncountable, Cantor had to introduce into his considerations a new bigger ordinal, 1 ;
similarly, Gentzen has to introduce a new ordinal bigger than those constructible from
, namely 0 .)
Despite all of the admirable coverage of this book (and I recommend it highly), at
some points Stillwell presents rather conventional views that deserve closer scrutiny
and revision. An unimportant example is his comments on Principia Mathematica (p.
68ff), which historically was very important, but has some rather serious shortcomings
in content and presentation; for instance he makes no distinction between the complex
formal system of Principia and the simple type theory which was actually studied
by Godel in his famous paper. By doing so Stillwell echoes and amplifies the myth
of Principia, which looms large in the secondary literature. Also, in connection with
Freges predicate logic (p. 90), the author does not remark on the difference between
the first-order part and the full Fregean system. You may think that the difference
between first-order and second-order logic is sophisticated and not appropriate for a
popular book, but this one is not a usual popular book. It is rather a semi-popular expo-
sition, which often enters into technical material, and the above-mentioned distinction
plays a role in several passages of the exposition.
More important are some aspects of the exposition of set theory, in particular
the (important) side-comments of a conceptual nature. The following subsection is
devoted to a number of comments concerning this issue, pertaining to the roads to
infinity and the status of the Axiom of Choice.

It is not that Stillwells technical exposition, sketchy as it may be, has shortcomings;
the author is very proficient and very clever in finding ways to present his material.
But in general the book offers more on the side of proof theory and computability, than


on set theory proper6 (an indication of this is the fact that Chapter 7, entitled Axioms
of infinity and devoted to higher set theory, is presented as a kind of epilogue on
p. 165). In Chapter 2, while presenting Cantors transfinite ordinals, nothing is said
about the details of Zermelos proof of the well-ordering theorem, but above all what
the author says regarding the axiom of choice (AC) is conventional wisdom that I find
misguided. The reader will excuse me for being opinionated on this topic, since it is
one to which I have contributed and in which Id like to help change the received view.
The received view is that AC is a fishy axiom, which is necessary for some technical
purposes but which also has very inconvenient consequences in some other contexts, a
principle that is not quite clear and that we rather want to avoid. AC is ritually blamed
for the appearance of such inconvenient facts as the existence of subsets of R that are
not Lebesgue-measurable, and the fact that a ball of unit volume can be decomposed
into two balls of unit volume (Banach-Tarski paradox). The received view is that there
is not much difference between the evidence we have for AC and the evidence for
Cantors continuum hypothesis, that 20 = 1 . I disagree strongly [3]. And of course,
I am not alone.
The view I insist upon has been urged by great experts in set theory and foundations
such as Godel and Bernays, who, starting in 1935, emphasized that AC is a natural
principle of set theory. One of the key ingredients of set theory is the viewpoint called
quasi-combinatorialism, according to which arbitrary sets and arbitrary functions are
just as much mathematical objects as any set or function which may be explicitly
defined through some condition. That is to say, set theory (and classical mathematics,
in particular classical analysis) accepts arbitrary or random sets of natural numbers
on a par with definable sets (e.g., the set of prime numbers or the set of multiples of
2011). Classical analysis reasons about the totality of arbitrary sets of real numbers,
and set theory generalizes this to reason about the totality of arbitrary sets of numbers,
points, functions, or sets (in the cumulative hierarchy V = V which has levels V
corresponding to all transfinite ordinals ). Classical mathematics follows the lead of
Dedekind, Cantor, Hilbert and Zermelo in not making any requirement of definability,
even when we are thinking about infinite sets, however big in cardinality. It treats
infinite sets in analogy with the combinatorial nature of finite sets, which we take for
granted, and this analogy explains the label quasi-combinatorialism.
The Axiom of Choice AC is the one axiom in Zermelos set theory that imple-
ments on a technical level this crucial idea.7 But many mathematicians around 1900
had second thoughts about admitting arbitrary sets and functions, and they were not
fully aware of the extent to which such sets were involved in the traditional results of
analysis. This is how the practice of treating AC as different and dangerous started.
However, it is not difficult to show that the simple idea of the real number system as
the totality of all possible infinite decimals leads directly to quasi-combinatorialism
(hence to choice, AC).
The real numbers in any interval of unit length correspond to all possible decimal
expansions after the comma (but for the well-known identification in the case of se-
quences ending in 9s). This much was clear, well understood, and accepted as early as
1800. Now, these expansions are infinitary objects, and as such they can be understood
in several different ways: most simply perhaps as infinite sequences of ciphers; but

6 Readers interested in this topic will find a lot in the philosophically oriented [1].
7 Bernays in 1935: The axiom of choice is an immediate application of the quasi- combinatorial concepts
in question. Godel, assuming the extensional, combinatorial notion of class or set, says in 1944: nothing
can express better the meaning of the term class than the axiom of [separation] and the axiom of choice
(Collected Papers, vol. II, 139; compare p. 131).

February 2012] REVIEWS 173

also as functions f : N {0, 1}, if we assume we are using binary notation. Quasi-
combinatorialism is nothing but the denial of any definability requirement, and this is a
common ingredient to both of the reconstructions just mentioned. Let me spell this out.
There is a notion of sequence which is arguably more elementary than sets or
functions: the concept of lawlike sequence that is a generalization of the most basic
mathematical structure, the sequence of natural numbers. (The notion of lawlike se-
quence is usual in constructive mathematics, referring to sequences determined by
an explicitly given rule. Examples are the sequence of Fibonacci numbers, and any
explicitly given sequence you haveever encountered; think, e.g., of a determinate
sequence of rationals converging to 2, like

1/1, 7/5, 41/29, 239/169, . . . , 8

or even the algebraic numbers ordered in a sequence according to their correspond-

ing minimal polynomials.9 ) This simple notion is however not the one involved in
understanding the reals as decimals: to get the real numbers by means of sequences
of ciphers, one has to move beyond lawlike sequences and consider also arbitrary
Thus the classical real numbers force consideration of arbitrary infinitary objects,
they force acceptance of quasi-combinatorialism. And AC is nothing but an expres-
sion of this general idea. Hence, contrary to the received view, the axiom AC is clearly
required in a system that aims to establish foundations for classical analysis. Its conse-
quences (complicating as they may be: non-measurable sets of reals, non-determinacy
of sets of reals, etc.) are important discoveries regarding the subject matter of classical
mathematics. And there is much difference between the evidence we have for AC and
the evidence for Cantors continuum hypothesis.
Let me now come to the methodological point of the role of AC in implying strange
consequences. What I want to argue is just that AC is never alonein particular,
the axiom asserting the existence of power sets (APow) is always required too. For
example, APow is no less present in the proof of Zermelos well-ordering theorem
than is AC. To show that a well-ordering of S exists for any given set S, one employs
a choice function f : (S) S, and this step is essential. (Then what is required
is to define a substructure of (S) that will play the role of the family of rests of
(well-ordered) set S, in the sense that the rest of an element s S is the set of all
elements x > s in the well-ordering. Once a family of sets having the structure of the
set of rests is defined, the choice function f mentioned above specifies the ordering
of elements, beginning with f (S).)
This brings me to the promised criticism of Stillwells identification of the cardi-
nal road to infinity with diagonalization. I must disagree. The first road to infinity
consists in heavy reliance on the axiom of powersets, understood in the sense of quasi-
combinatorialism explained above. It is the assumption that all possible subsets of
N forms a new well-determined totality, a set (which is just as much a mathematical
object as the number 1), which brings us boldly into the realm of uncountable infinite
sets. In a paper published this year I wrote: The importance of Cantors diagonal pro-
cedure is, precisely, that it constitutes a method for transcending any given sequence of

8 Where numerators and denominators obey the rule: tn+2 = 6 tn+1 = tn .

9 Thisis what Cantor (and Dedekind) used to prove that the set of algebraic numbers is denumerable.
10 From a didactic point of view, it is important to reflect on the fact that the essential difference between

sequences and functions applies only if we restrict to lawlike sequences. Once we are considering arbitrary
sequences, this is essentially the same as functions f : N (0, 1, . . . , 9).


definable subsets of N (analogously for other sets). In and by itself, however, Cantors
diagonal method does not lead to arbitrary sets [nor, I add here, to higher infinities]. In
fact, if a countable sequence of definable sets of natural numbers is given explicitly, so
that we can compute whether n belongs to the nth set, the diagonal procedure yields
a computation of the truth value of n B (where B refers to the new set defined by
Cantors method). [3]11
Diagonalization establishes that a certain domain (e.g., the set of real numbers) is
not exhaustible by a given countable sequencewhich is transcended by employing
the diagonal procedure. However, to establish that the domain is an uncountable set,
a higher infinity, obviously one needs to postulate that it is a set, and this is based on
the axioms of ZFC, very especially on APow.
Actually, the axiom of powersets is also behind the second, ordinal road to infinity.
The formation of countable ordinals , + 1, . . . , 2 , . . . goes on and on, leading
up to 0 and much, much farther. But to follow Cantor in his path towards the higher
infinite, one needs to claim that there is a set of all countable ordinals, denoted by
1 . Now, the existence of 1 is again a claim that depends crucially on APow: in the
system ZFC without APow one cannot prove that 1 exists. And of course APow is
crucially behind the kind of infinitary structure by which the continuum is modelled
in classical mathematics, hence behind results such as the Banach-Tarski paradox.
The question now is, why do we put all the blame on AC, and disregard APow
completely? The answer is, merely by tradition and conventional wisdom. AC has
been regarded as intuitively more dubious than APow; but this may be a superficial
impression and invites revision insofar as powersets are interpreted in the standard,
quasi-combinatorial way.
Notice that the classical idea of the real numbers involves both (1) the assumption
of particular arbitrary (non-definable) objects, and (2) the assumption of a totality
of such objects. The first corresponds, roughly speaking, to a randomly chosen real,
the second to the set of all reals. In the system of axiomatic set theory, these two
assumptions are paralleled by AC, corresponding to (1), and the axiom of powersets,
corresponding to (2). From a conceptual point of view, both are equally dubious, and
any qualm we may have concerning AC would apply equally to APow. If anything,
such doubts should be multiplied, for the powerset (N ) contains uncountably many
arbitrary sets of naturals! Thus, your doubts concerning a single application of AC
should be multiplied times 20 in the case of the simplest powerset of an infinite set.

And the powerset (R) has 22 0 -many arbitrary sets of reals!

The above may help emphasize, once again, that there is something inconcrete in the
subject matter of classical set theoryin 1 and 20 , hence on both sides of Cantors
equation 20 = 1 . We have pinned it down to the heavily idealizing tendencyand
the associated vaguenessof quasi-combinatorialism, which is expressed directly
in the contentious axiom AC and indirectly in the standard reading of the axiom of
powersets. Coming back to Stillwells discussion of the finitary counterparts of those
ideas, we find that the material on arithmetic and proof theory is rather concrete,
contentual: there is something tangible about 0 and about Gentzens measuring the
proof-theoretic strength of the proposition that PA is consistent, while there is some-
thing intangible about the real number system and about the problem of measuring
its uncountability. Yet mathematicians have always found a need for ideal horizons
11 Compare the case with the algebraic numbers: one can define in the strict sense a transcendental number

by using Cantors diagonal argument applied to an explicit enumeration of the set of algebraic numbers. See
R. Gray, Georg Cantor and Transcendental Numbers, The American Mathematical Monthly 101 (1994), pp.
819832; available at

February 2012] REVIEWS 175

against which to determine our more concrete procedures. Prominent among these
are the step to infinity, and the step to the continuum, which are always moves that
allow us to determine in new ways (top-down) significant procedures which are more
constructive (bottom-up). We find again Hilberts topic of the contentual vs. the ideal,
symbols and processes vs. (possibly vague) ideas!12
To sum up, Stillwells book is highly commendable, very informative and well
organized. It is very carefully produced. I was able to find only minor errors: an unfor-
tunate one on p. 8 in Euclids proof, on p. 26 a misplaced sentence, and a few others.
I should mention also a historical mistake on p. 62 (Cantor did know about sets larger
than R before 1891), and an opinion that I do not share on 135 (Dirichlets geometric
proof of ab = ba probably had a didactic intention). There is a place where the reader
may easily be misled: on pp. 6869 it is said that Godel proved the logical complete-
ness (!) of Principia Mathematica before proving its mathematical incompleteness.
This must be understood in the sense that the logical completeness of PM is that of
its first-order subsystemfor the type theory of PM is incomplete.13
As already said, I believe the book is not really appropriate for high school students,
since it presupposes some measure of mathematical maturity (e.g., comfort with func-
tions). So it seems to presuppose the level of a mature first-year university student. A
feature that deserves special mention is that Stillwell has included rich and interesting
historical notes, which he often employs to mention more advanced material. Thus,
in the early chapters, some commentary on Godels constructible hierarchy (the basis
for proving the consistency of AC and the Continuum Hypothesis relative to the ax-
iomatic system ZF) and Cohens method of forcing; later on, a discussion of Chaitins
incompleteness theorem and of links between logic and computation (Cooks theorem,
the problem N = N P). In this same spirit, the last chapter discusses large cardinals
and projective determinacy.


1. T. Arrigoni, What is Meant by V? Reflections on the Universe of All Sets, Mentis-Verlag, Paderborn, 2007.
2. C. Cellucci, Why Proof? What is Proof? in Deduction, Computation, Experiment. Exploring the
Effectiveness of Proof. Edited by G. Corsi and R. Lupacchini. Springer, Berlin, 2008.
3. J. Ferreiros, On arbitrary sets and ZFC, Bull. of Symbolic. Logic 17 (2011) 361393.
4. T. Franzen, Godels theorem: An incomplete guide to its use and abuse, A K Peters, Natick, MA, 2005.
5. L. Henkin, Completeness in the theory of types, J. of Symbolic Logic 15 (1950) 8191; available at
6. D. Hilbert, On the Infinite, in From Frege to Godel, a Source Book in Mathematical Logic. Edited by J.
van Heijenoort. Harvard University Press, Cambridge, MA, 2002.

Institute of Philosophy, CCHS

CSIC (Spanish Higher Council for scientific research), Madrid

12 Two more examples given by Stillwell: we have both the finite and infinite Ramsey theorems, but few

Ramsey numbers are actually known! (p. 152153); also what is said about the Green-Tao theorem (p. 158).
13 Unless one does not employ the full set-theoretic semantics, but so-called Henkin semantics, which is

actually quite natural as a semantics for higher-order logics [5].


New from the MAA
The Hungarian Problem Book IV
Edited and Translated by
Robert Barrington Leigh and Andy Liu
The Etvs Mathematics Competition
is the oldest high school mathematics
competition in the world, dating back
to 1894. This book is a continuation of
Hungarian Problem Book III and takes
the contest through 1963. Forty-eight
problems in all are presented in this
volume. Problems are classified under
combinatorics, graph theory, number
theory, divisibility, sums and differ-
ences, algebra, geometry, tangent lines
and circles, geometric inequalities,
combinatorial geometry, trigonometry
and solid geometry. Multiple solutions
to the problems are presented along with background materi-
al. There is a substantial chapter entitled "Looking Back," which
provides additional insights into the problems.
Hungarian Problem Book IV is intended for beginners, al-
though the experienced student will find much here. Beginners
are encouraged to work the problems in each section, and
then to compare their results against the solutions presented
in the book. They will find ample material in each section to
help them improve their problem-solving techniques.
114 pp., Paperbound, 2011
ISBN 978-0-88385-831-8
Catalog Code: HP4
List: $40.95
Member: $33.95

To order visit us online at or call us at 1-800-

MATHEMATICAL ASSOCIATION OF AMERICA 1529 Eighteenth St., NW Washington, DC 20036

New title from the MAA!

Rediscovering Mathematics:
You Do the Math

Shai Simonson
Rediscovering Mathematics is an eclectic collection
of mathematical topics and puzzles aimed at
talented youngsters and inquisitive adults who
want to expand their view of mathematics. By
focusing on problem solving, and discouraging
rote memorization, the book shows how to learn
and teach mathematics through investigation,
experimentation, and discovery. Rediscovering
Mathematics is also an excellent text for training
math teachers at all levels.

Topics range in difficulty and cover a wide range of historical periods,

with some examples demonstrating how to uncover mathematics in
everyday life, including:

number theory and its application to secure communication

over the Internet,
the algebraic and combinatorial work of a medieval
mathematician Rabbi, and
applications of probability to sports, casinos, and everyday life.

Rediscovering Mathematics provides a fresh view of mathematics for

those who already like the subject, and offers a second chance for those
who think they dont.

To order call 1-800-331-1622 or visit us online at!