Professional Documents
Culture Documents
NOTES
On the Greatest Common Divisor of the Value of Two 446
Polynomials
Péter E. Frenkel and József Pelikán
BOOK REVIEW
Elements of Mathematics: From Euclid to Gödel 475
by John Stillwell
by Reuben Hersh
EDITOR
Susan Jane Colley
Oberlin College
ASSOCIATE EDITORS
David Aldous Daniel Krashen
University of California, Berkeley University of Georgia
Elizabeth S. Allman Jeffrey Lawson
University of Alaska Fairbanks Western Carolina University
David H. Bailey Susan Loepp
University of California, Davis Williams College
Scott T. Chapman Jeffrey Nunemacher
Sam Houston State University Ohio Wesleyan University
Allan Donsig Bruce P. Palka
University of Nebraska-Lincoln National Science Foundation
Michael Dorff Paul Pollack
Brigham Young University University of Georgia
John Ewing Adriana Salerno
Math for America Bates College
Stephan Ramon Garcia Edward Scheinerman
Pomona College Johns Hopkins University
Luis David Garcia Puente Anne V. Shepler
Sam Houston State University University of North Texas
Sidney Graham Frank Sottile
Central Michigan University Texas A&M University
J. Roberto Hasfura-Buenaga Susan G. Staples
Trinity University Texas Christian University
Michael Henle Sergei Tabachnikov
Oberlin College Pennsylvania State University
Tara Holm Daniel Velleman
Cornell University Amherst College
Lea Jenkins Cynthia Vinzant
Clemson University North Carolina State University
Gary Kennedy Steven H. Weintraub
Ohio State University, Mansfield Lehigh University
Chawne Kimber Kevin Woods
Lafayette College Oberlin College
ELECTRONIC PRODUCTION
MANAGING EDITOR AND PUBLISHING MANAGER
Bonnie K. Ponce Beverly Joy Ruedi
NOTICE TO AUTHORS Proposed problems and solutions may be submitted to Prob-
lem Editor Daniel Ullman online via https://american
The MONTHLY publishes articles, as well as notes and other fea- mathematicalmonthly.submittable.com/submit.
tures, about mathematics and the profession. Its readers span
a broad spectrum of mathematical interests, and include pro- Questions but not submissions may be addressed to
fessional mathematicians as well as students of mathematics monthlyproblems@maa.org.
at all collegiate levels. Authors are invited to submit articles
Advertising correspondence should be sent to:
and notes that bring interesting mathematical ideas to a wide
audience of MONTHLY readers. MAA Advertising
1529 Eighteenth St. NW
The MONTHLY’s readers expect a high standard of exposition;
Washington DC 20036
they expect articles to inform, stimulate, challenge, enlighten,
Phone: (202) 319-8461
and even entertain. MONTHLY articles are meant to be read, en-
E-mail: advertising@maa.org
joyed, and discussed, rather than just archived. Articles may
be expositions of old or new results, historical or biographical Further advertising information can be found online at www.
essays, speculations or definitive treatments, broad develop- maa.org.
ments, or explorations of a single application. Novelty and
Change of address, missing issue inquiries, and other sub-
generality are far less important than clarity of exposition
scription correspondence can be sent to:
and broad appeal. Appropriate figures, diagrams, and photo-
graphs are encouraged. maaservice@maa.org.
Notes are short, sharply focused, and possibly informal. They or
are often gems that provide a new proof of an old theorem, a The MAA Customer Service Center
novel presentation of a familiar theme, or a lively discussion P.O. Box 91112
of a single issue. Washington, DC 20090-1112
(800) 331-1622
Submission of articles, notes, and filler pieces is required via the
(301) 617-7800
MONTHLY’s Editorial Manager System. Initial submissions in pdf or
LATEX form can be sent to Editor Susan Jane Colley at Recent copies of the MONTHLY are available for purchase
www.editorialmanager.com/monthly. through the MAA Service Center at the address above.
The Editorial Manager System will cue the author for all re- Microfilm Editions are available at: University Microfilms In-
quired information concerning the paper. The MONTHLY has ternational, Serial Bid coordinator, 300 North Zeeb Road, Ann
instituted a double-blind refereeing policy. Manuscripts that Arbor, MI 48106.
contain the author’s names will be returned. Questions con-
cerning submission of papers can be addressed to the Editor- The AMERICAN MATHEMATICAL MONTHLY (ISSN 0002-9890) is
Elect at monthly@maa.org. Authors who use LATEX can find published monthly except bimonthly June-July and August-
our article/note template at www.maa.org/monthly.html. September by the Mathematical Association of America
This template requires the style file maa-monthly.sty, which at 1529 Eighteenth Street, NW, Washington, DC 20036 and
can also be downloaded from the same webpage. A format- Lancaster, PA, and copyrighted by the Mathematical Asso-
ting document for MONTHLY references can be found there too. ciation of America (Incorporated), 2017, including rights to
this journal issue as a whole and, except where otherwise
Letters to the Editor on any topic are invited. Comments, criti- noted, rights to each individual contribution. Permission to
cisms, and suggestions for making the MONTHLY more lively, make copies of individual articles, in paper or electronic
entertaining, and informative can be forwarded to the Editor form, including posting on personal and class web pages,
at monthly@maa.org. for educational and scientific use is granted without fee
The online MONTHLY archive at www.jstor.org is a valuable provided that copies are not made or distributed for profit
resource for both authors and readers; it may be searched or commercial advantage and that copies bear the follow-
online in a variety of ways for any specified keyword(s). MAA ing copyright notice: [Copyright 2017 Mathematical Asso-
members whose institutions do not provide JSTOR access ciation of America. All rights reserved.] Abstracting, with
may obtain individual access for a modest annual fee; call credit, is permitted. To copy otherwise, or to republish,
800-331-1622 for more information. requires specific permission of the MAA’s Director of Pub-
lications and possibly a fee. Periodicals postage paid at
See the MONTHLY section of MAA Online for current informa- Washington, DC, and additional mailing offices. Postmas-
tion such as contents of issues and descriptive summaries of ter: Send address changes to the American Mathemati-
forthcoming articles: cal Monthly, Membership/Subscription Department, MAA,
www.maa.org/monthly.html. 1529 Eighteenth Street, NW, Washington, DC 20036-1385.
Friendly Frogs, Stable Marriage,
and the Magic of Invariance
Maria Deijfen, Alexander E. Holroyd, and James B. Martin
Abstract. We introduce a two-player game involving two tokens located at points of a fixed
set. The players take turns moving a token to an unoccupied point in such a way that the dis-
tance between the two tokens is decreased. Optimal strategies for this game and its variants
are intimately tied to Gale–Shapley stable marriage. We focus particularly on the case of ran-
dom infinite sets, where we use invariance, ergodicity, mass transport, and deletion-tolerance
to determine game outcomes.
Proof. Let M be the set of all unordered pairs {x, y} in L such that the game started
with two frogs at x and y is a loss for the next player. The key ingredient is a simple
algorithm that identifies M. (We postpone consideration of the two opening moves, in
which the frogs are placed.) In fact M will form a partial matching on L. We construct
this matching iteratively as follows. The idea is to work backwards from positions
where the outcome is known. Order the set of all n2 pairs in L in increasing order of
distance between the pair. Then for each pair in turn (starting with the closest pair),
http://dx.doi.org/10.4169/amer.math.monthly.124.5.387
MSC: Primary 91A46, Secondary 60D05; 60G55
3 4 5
Figure 1. A game of friendly frogs on a set L of size 5. Alice starts. Alice’s moves are shown in amber
(lighter), Bob’s in blue (darker). After move 5, Bob has no legal move, so Alice wins.
match the two points to each other if and only if neither is already matched. The
algorithm ends with at most one point not matched. See Figure 2 for an example.
To show that this M has the claimed property for the game, we need to check that
from any position in M, it is impossible to move to another position in M, while from
a position not in M, it is possible to move to a position in M. The former is immediate
because M is a partial matching (and a move consists of moving only one frog). For
the latter, suppose the frogs are located at x and y, and that x and y are not matched
to each other. Since x and y were not matched by the algorithm, at least one of them
was matched to a closer point; without loss of generality, x is matched to w, where
|x − w| < |x − y|. (Here and subsequently, | · | is the Euclidean norm on Rd .) Hence
we can move a frog from y to w.
If n is odd then there is exactly one point that is not matched, so Alice wins by
placing the first frog there; wherever Bob places the second frog, the two frog locations
are not matched to each other. If n is even, then the matching M is perfect (that is, every
point is matched). Therefore, wherever Alice places the first frog, Bob wins by placing
the second on its partner in M.
We will consider various extensions of the friendly frogs game, including versions
where frogs and/or points are player-specific (available only to one player), where
388
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
certain moves are forbidden, and where different winning criteria apply. Notwithstand-
ing the humble beginning of Theorem 1, we will be led into some very intriguing
waters. For concreteness we will focus throughout on points in Rd , although many
arguments carry over to more general metric spaces (and, for instance, the proof of
Theorem 1 extends even to any injective symmetric distance function on L). We will
continue to assume that all interpoint distances are distinct. (Relaxing this assumption
is also quite natural, but we choose instead to pursue other directions.)
Matters become particularly interesting when we allow the set of points (lily pads)
L to be infinite, and especially a random countable set. The “losing” two-frog posi-
tions will still form a matching, and this matching is most naturally interpreted as a
version of the celebrated stable marriage of Gale and Shapley, the topic of the 2012
Nobel prize in economics (awarded to Roth and Shapley). We will make crucial use of
invariance of the probability distribution of L under symmetries of Rd . This powerful
tool permits remarkably simple and elegant proofs of facts apparently not amenable to
other arguments. In games involving points of several types, we will see an example
of a phase transition, as well as a situation in which existence of a phase transition
is an open question. We will also analyze play of simultaneous games by making a
connection to the remarkable theory of Sprague–Grundy values (or “nimbers”).
This article contains a mixture of original research and expository material. We use
the friendly frogs game partly as a vehicle to showcase some beautiful known ideas,
and we assume a minimum of technical background. The game and its analysis are
novel, so far as we know. Stable marriage [10] and its variants have been extensively
studied, but the connection to games appears to be new. Many of the results that we use
on matchings of random point sets are taken from [17]. We will review the necessary
background and give proofs where appropriate. The general theory of combinatorial
games is highly developed (see for instance [3]). We will explain the relevant parts
of the theory as they apply in our context. Other recent work on games in random
settings appears, for example, in [1, 14, 15] and the review [19]. In a different direction,
certain games in infinite and random settings have intimate connections with general
topology [22] and logic [20].
2. INFINITE POINT SETS. Theorem 1 shows that the outcome of friendly frogs
on a finite set L is determined solely by the parity of the number of points (lily pads).
What happens when L is infinite? Is ∞ odd or even? The answer now depends on
the choice of set; we will focus especially on the behavior of typical (that is, random)
infinite sets.
Let L be an infinite subset of Rd . As before, we assume that all distances between
pairs of points in L are distinct. We call a sequence of points x1 , x2 , . . . a descending
chain if the distances (|xi − xi+1 |)i≥1 form a strictly decreasing sequence. If there
exists an infinite descending chain x1 , x2 , · · · ∈ L, then it is possible for the game to
last forever. See Figure 3. Therefore we make the additional assumption that L has
no infinite descending chains. This implies in particular that L is discrete, that is, any
bounded set contains only finitely many points.
It is easy to construct examples of infinite sets L, with all distances between pairs
of point distinct and with no infinite descending chains, for which either player wins
friendly frogs; see Figure 3. First, in dimension 1, place exactly two points in each
of the intervals [3i, 3i + 1] for i ∈ Z. (A simple way to make all interpoint distances
distinct is to choose each point uniformly at random in the appropriate interval, inde-
pendently of all others.) Then Bob wins by placing a frog at the unique point in the
same interval as Alice’s initial frog. Second, suppose the points are as above except
that the interval [0, 1] now contains only one point. Then Alice wins by placing the
first frog on this point; whichever point Bob chooses for the second frog, Alice can
then move the first frog to the “partner” of that point in the appropriate unit interval.
As in the previous section, the key to analyzing the game for general L is to identify
those positions from which the game is a loss for the player whose turn it is to move.
Following standard conventions of combinatorial game theory (see for instance [3]),
such positions are called P-positions to indicate that the [P]revious player wins, while
all other positions are called N-positions, since the [N]ext player wins. Since terminal
positions are P-positions, the P- and N-positions satisfy the following.
(N) From every N-position, there is at least one possible move to a P-position.
(P) From every P-position, every possible move is to an N-position.
Since the game terminates in a finite number of moves, it follows by induction that
these properties are sufficient to characterize the P- and N-positions. That is, to check
that a claimed partition of the positions into P- and N-positions is correct, it suffices to
check that it satisfies (N) and (P).
In many games, characterizing the set of P-positions is a difficult problem requiring
experimentation and insight. In contrast, checking via (N) and (P) that such a charac-
terization is correct may be essentially mechanical.
In friendly frogs, the two-frog P-positions are given by a matching. Here is some
notation. Let L ⊆ Rd . A matching of L is a set M of unordered pairs of distinct points
in L such that each point of L is included in at most one pair. The matching is perfect
if each point is included in exactly one pair. For x ∈ L, we write M(x) for the partner
of x, that is, the unique point y such that {x, y} ∈ M, or, if there is no such y, we set
M(x) := ∞ and say that x is unmatched.
As in the case of finite L in the last section, we will construct the relevant matching
iteratively. Now, however, there may be no closest pair of points, so we need a local
version of the algorithm.
The following abstraction will prove very useful. Imagine that each point of L
“prefers” to be matched to a partner that is as close as possible. Given a matching
M of L, a pair of points x, y ∈ L is called unstable if they both strictly prefer each
other over their own partners, that is, if |x − M(x)| and |y − M(y)| are both strictly
greater than |x − y| (where |x − M(x)| := ∞ if M(x) = ∞, so that any partner is
preferable to being unmatched). A matching M is called stable if there are no unstable
pairs. Note that any stable matching of L has at most one unmatched point.
Stable matching can be applied to a wide variety of settings involving agents each
of which has preferences over the others. The concept was introduced in a celebrated
paper of Gale and Shapley [10], who considered the setting of n heterosexual mar-
riages between n women and n men, each of whom has an arbitrary preference order
over those of the opposite sex. Gale and Shapley gave a beautiful algorithm proving
390
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
the existence of a stable matching in this case. (They showed, however, that stable
matchings are not necessarily unique, and may not exist in the same-sex “roommates”
variant.) As mentioned earlier, the 2012 Nobel prize in economics was awarded on
the basis of this and ensuing work, to Roth for practical applications, and to Shapley
for theoretical advances. Our setting differs from the standard Gale–Shapley same-sex
matching problem in that the set L is infinite; on the other hand, our preferences are
very special, since they are based on distance. This case was studied in [17].
Proposition 2 ([17]). Suppose L ⊂ Rd has all pairwise distances distinct and has no
infinite descending chains. Then there exists a unique stable matching of L.
Proof. We will show that the following algorithm leads to a stable matching. First
match all mutually closest pairs of points. Then remove them and match all mutually
closest pairs in the remaining point set. Repeat indefinitely (that is, for a countably
infinite sequence of stages), and take as the final matching the set of all pairs that are
ever matched.
By induction over the stages in the algorithm, every pair that is matched by the
algorithm must be matched in any stable matching.
Furthermore, at most one point can be left unmatched by the algorithm. To see this,
assume that there are at least two unmatched points. Since there are no descending
chains, the set of unmatched points then contains at least one pair of points that are
mutually closest in this set and, since L is discrete, this pair must have been mutually
closest at some finite stage of the algorithm. However, then they should have been
matched to each other, which is a contradiction.
Finally, we need to confirm that the resulting matching is in fact stable. To this end,
assume that there exist x, y ∈ L with |x − M(x)| and |y − M(y)| both strictly greater
than |x − y|. By the previous argument, at least one of x and y is matched, so consider
the earliest stage at which one of them was matched by the algorithm. Since both x
and y were unmatched prior to this stage, we obtain a contradiction.
Proposition 3. Suppose L ⊂ Rd has all pairwise distances distinct and has no infinite
descending chains. Let M be the stable matching of L and consider friendly frogs on
L. The position with the two frogs at x and y is a P-position if and only if x is matched
to y in M.
Proof. Since L has no infinite descending chains, the game terminates. Therefore, it
suffices to check that properties (N) and (P) above hold for the claimed partition of the
positions. For (N), if {x, y} ∈ M, then x (or, respectively, y) must have a partner that
is closer than y (x), since otherwise x and y would constitute an unstable pair. Without
loss of generality, M(x) = w where |x − w| < |x − y|, and we can then move a frog
from y to w, confirming (N). The property (P) is immediate, since M is a matching.
As before, if M leaves one point unmatched, then Alice wins by placing the first
frog at that point. If the matching is perfect, then Bob wins by placing the second frog
at the partner of Alice’s initial move. As we have seen, both situations are possible for
suitable infinite sets L.
Random infinite sets. It is natural to ask what happens for a typical infinite set of
points. A natural and canonical way to formalize this notion is the Poisson point pro-
cess, which is defined as follows. Fix λ > 0. Let any Borel set of finite volume contain
a random number of points with a Poisson distribution of mean equal to λ times its
volume, and let disjoint sets contain independent numbers of points. These conditions
characterize the distribution of the set of points, and the resulting random set is called
a (homogeneous) Poisson (point) process with intensity λ on Rd . It is a countable
infinite set with probability 1. (The Poisson process has other equivalent definitions—
for instance, it may be constructed as a limit as n → ∞ of n uniformly random points
in a ball of volume n/λ around the origin, or as a limit as ε → 0 of a grid of cubes
of volume ε each of which contains a point with probability ελ independently.) If L
is a Poisson process of intensity 1, then {λ1/d x : x ∈ L} is a Poisson process of inten-
sity λ. The intensity parameter will be unimportant for us until we consider several
Poisson processes together. See, for instance, [6] for background. It is straightforward
to check that with probability 1, all pairs of points have distinct distances, and that
there are no descending chains. See for instance [13] or [5] for proofs. The process is
translation-invariant, which is to say, its distribution is invariant under the action of
any translation of Rd .
392
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
2
5 2 4 1
3 4 3
3
1 1 2
Figure 5. Three variant games: (a) colored friendly frogs, in which each player may only move their own
frog; (b) colored friendly frogs on colored points, where in addition a frog may only occupy a point of its own
color; (c) fussy frogs, in which the two frogs may not both occupy red (darker) points. (Here, blue elements
are shown somewhat darker than their amber counterparts.)
Despite the simplicity of the proof of Theorem 4, there is something subtle and
mysterious about the argument. What probability-one property of the Poisson process
does it use? In other words, is there some easily described set A of subsets of Rd
such that (a) the Poisson process lies in A with probability 1, and (b) Bob wins on
any L ∈ A? We do not know of such a set, except for unsatisfying choices such as
A = {L : L has a perfect stable matching} or A = {L : Bob wins}. As we have seen,
the set of L with distinct interpoint distances and no descending chains satisfies (a)
but not (b). The proof of Theorem 4 uses translation-invariance of the Poisson process
in a fundamental way that apparently cannot be easily reduced to such a probability-
one property. Many elegant arguments in probability theory involve an appeal to some
symmetry or invariance property of this kind. In the next section we will use stronger
probabilistic properties of Poisson processes: deletion-tolerance and ergodicity.
In fact, the algorithm in the proof of Proposition 2 leads to a perfect stable matching
for a large class of translation-invariant point processes on Rd ; see [17, Proposition 9].
The conclusion of Theorem 4 hence remains valid for this class of processes.
The article [17] is also concerned with the distribution of the distance from a point
to its partner in the stable matching. These distances are potentially relevant to issues
of computational complexity and length of the game. For instance, if Alice is required
to place her first frog within distance r of the origin, how difficult can she make it for
Bob to win? We leave these interesting questions for future investigation.
Colored frogs. First we introduce the colored friendly frogs game. Here, Alice starts
by placing an amber frog on some point of L, then Bob places a blue frog on a
different point. Subsequently, the game proceeds exactly as before, except that Alice
may only move the amber frog, and Bob may only move the blue frog. As before, a
player who cannot move loses.
A two-frog position can now be specified by an ordered pair (x, y), where x is the
location of the frog of the previous player to move, and y the location of the frog of
the next player.
Rather than requiring an entirely new analysis, it turns out that the P-positions can
again be described in terms of the stable matching M of L. If |x − y| ≤ |x − M(x)|,
then we say that x desires y. (This terminology is natural given the interpretation
of preferences described earlier.) Note the use of the weak inequality ≤, so that a
Proposition 5. Suppose L ⊂ Rd has all pairwise distances distinct and has no infi-
nite descending chains. Let M be the stable matching of L, and consider the colored
friendly frogs game on L. The position (x, y) is a P-position if and only if x desires y.
Proof. Again it suffices to check the conditions (N) and (P). For (N), if |x − y| >
|x − M(x)|, then the frog at y can be moved to M(x). On the other hand, for (P), if
|x − y| ≤ |x − M(x)|, then there cannot exist z ∈ L with |x − z| < |x − y| and |x −
z| ≤ |z − M(z)|, since in that case x and z would constitute an unstable pair. Hence
moving the frog at y must result in a position (z, x) with |x − z| > |z − M(z)|.
Note that an unmatched point in the stable matching is not desired by any other
point, since that pair would be unstable. Hence, if the stable matching of L has one
unmatched point, then Alice wins colored friendly frogs by placing her amber frog at
the unmatched point. If the matching is perfect, then Bob wins, for instance by placing
his blue frog at the partner of Alice’s initial point. The outcome is hence the same as
in the original friendly frogs game. In particular, we have the following.
Indeed, Bob may use the same strategy in the colored and uncolored games, always
moving to a matched pair. Does this mean that the games are essentially identical?
No. To highlight an interesting difference, let us modify the rules in a way that favors
Alice. In shy friendly frogs, we fix a constant c > 0, and stipulate that Bob, on his
opening move, cannot place the second frog within distance c of the first frog. (But we
place no such restriction on subsequent moves.) Shy colored friendly frogs is defined
analogously. Surprisingly, the outcome now differs between the two variants; the proof
will employ an interesting probabilistic argument.
Proof. For (i), Alice places the first frog on any point x whose partner M(x) is at most
distance c away. Such a point exists: for instance there are pairs of mutually closest
points within distance c of each other.
Turning to (ii), we claim that with probability 1, every point is desired by infinitely
many others. This implies in particular that whatever Alice’s opening move x is, there
exists a point y with |x − y| > c that desires x, so Bob wins by placing his frog there,
by Proposition 5. The claim follows from [8, Theorem 1.3 (i)]. Since the proof in our
case is short, we include it.
Let X be the (random) point of L closest to the origin. It suffices to show that
infinitely many points desire X . Let D be the set of points that desire X . Modify
the set L as follows. Whenever D is finite, delete all points of D and their partners,
except for X itself (which is the partner of a point in D). It is easy to check that the
stable matching of the modified set is simply the restriction of M to the points that
remain. In particular, if D was finite, then X is now unmatched. However, the Poisson
394
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
process is deletion-tolerant, which is to say: deleting any finite set of points, even
in a way that depends on the process, results in a point process whose distribution is
absolutely continuous with respect to the original distribution. (See for instance [17,
Lemma 18] or [18].) That is, the deletion cannot cause any event of zero probability to
have positive probability. (Intuitively, the picture after deletion is still plausible.) Since
the stable matching of the Poisson process is perfect with probability 1, we deduce that
D was infinite with probability 1.
Colored points. There is a further natural variant of colored friendly frogs in which
the two frogs are restricted to different point sets. Let L A and L B be two disjoint
subsets of Rd whose union satisfies the assumptions of Proposition 2. We refer to
points of L A and L B as amber and blue, respectively. We stipulate that Alice’s amber
frog can only occupy an amber point, and Bob’s blue frog can only occupy a blue
point. Otherwise the rules are as for colored friendly frogs. We call this game colored
friendly frogs on colored points. The P-positions in this case are given by a two-color
variant of stable matching.
A two-color matching of (L A , L B ) is a set M of pairs of points (x, y) ∈ L A × L B
such that each point is contained in at most one pair. As in the one-color case, the
matching is perfect if each point of L A ∪ L B is included in a pair. A two-color matching
M of (L A , L B ) is stable if and only if there do not exist x ∈ L A and y ∈ L B with
|x − M(x)| and |y − M(y)| both strictly greater than |x − y|.
Proposition 2 and Proposition 5 remain true for this game, with L replaced by
(L A , L B ), “stable matching” replaced by “stable two-color matching,” and a revised
definition of desire under which a point can only desire a point of the opposite color
(see [17] for more detail). The same proofs apply with only minor adjustments. Specif-
ically, in the algorithm described in the proof of Proposition 2, points of the same
color cannot be matched to each other. Therefore, instead of leaving at most one point
unmatched, it follows from the same arguments that all unmatched points must be of
the same color.
Note that an unmatched point desires all points of the other color, and an unmatched
point cannot be desired by any point of the other color, since they would be an unstable
pair. If the two-color stable matching has unmatched amber points, then Alice wins by
placing her frog at one of these points. If not, Bob wins by placing his frog on an
unmatched blue point (if one exists), or on the partner of Alice’s opening move.
Proof of Theorem 8. First let us consider the case α = β. The set of unmatched points
in the stable matching is either empty, or consists only of amber points or only of blue
points. Applying ergodicity, one of these three events must have probability 1, and the
others probability 0. But by symmetry the probabilities of unmatched amber points and
of unmatched blue points must be equal. Hence they are both 0, and with probability
1 the matching is perfect, giving a win for Bob.
When the two intensities are different, it is natural to expect that we cannot match
amber points to blue points in a translation-invariant way without leaving some
of the higher-intensity set unmatched. Making this intuition rigorous may at first
appear tricky. We might compare the numbers of points in a large ball, but perhaps
many points have their partners outside the ball. Furthermore, where should we use
translation-invariance? Since L A and L B are countable infinite sets, there certainly
exists some perfect matching between them.
In fact, there is a clean solution, using a simple but powerful tool, the mass transport
principle. (See [2, 12] for background.) Consider any function f : Zd × Zd → [0, ∞)
that is translation-invariant
that f (s, t) = f
in the sense (s + u, t + u) for all s, t, u ∈
Zd . Then note that t∈Zd f (0, t) = t∈Zd f (−t, 0) = s∈Zd f (s, 0). It is sometimes
helpful to think of f (s, t) as the mass sent from s to t.
Now suppose α < β. For s ∈ Zd , let Q s be the unit cube s + [0, 1)d in Rd . Define
f (s, t) to be the expected number of amber points in Q s that are matched to blue
points in Q t . This f is translation-invariant in the sense of the previous paragraph,
because of translation-invariance of the Poisson processes. Thus, s f (s, 0), which is
the expected number of matched blue points in Q 0 , is equal to t f (0, t), which is the
expected number of matched amber points in Q 0 . The latter is at most α, the expected
total number of amber points in Q 0 . But the expected number of blue points in Q 0 is
β, so the expected number of unmatched blue points in Q 0 is at least β − α. In par-
ticular, the probability that there exists an unmatched blue point is positive. Applying
ergodicity again shows that this probability is therefore 1. Thus Bob wins.
Similarly, if α > β then with probability 1 there are unmatched amber points, lead-
ing to a win for Alice.
Once again, the proof of Theorem 8 uses invariance and ergodicity in a subtle and
fundamental way that cannot easily be reduced to probability 1 properties of the point
process. What property of (L A , L B ) guarantees Bob wins when α = β? It is not that
L A and L B have equal asymptotic density. Modifying the example in Figure 3, that
holds if L A consists of one point in every interval [3i, 3i + 1] for i ∈ Z while L B has
one point in each such interval except [0, 1]. But here Alice wins.
Again, the conclusion of Theorem 8 remains valid for a large class of translation-
invariant point processes; see [17] for details of the corresponding results for stable
matchings.
Fussy Frogs. Despite the relatively complete analysis in the last two cases, we need
not go far to reach an unsolved problem. In fussy friendly frogs, the points again have
two colors, now green and red, denoted by sets L and L R , respectively. The rules
are as in the original friendly frogs game (in particular, the two frogs are once again
identical and can be moved by either player), except that it is not permitted that both
frogs simultaneously occupy red points.
396
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Open Problem. Let L and L R be independent Poisson processes on Rd with respective
intensities 1 and ρ. Do there exist d ≥ 1 and ρ > 0 for which Bob wins fussy friendly
frogs with positive probability?
Fussy friendly frogs again has an associated matching, the analog of stable matching
under the restriction that red points cannot be matched to each other. This matching
can be constructed iteratively as in the proof of Proposition 2, and Bob wins if and
only if it is perfect. Ergodicity shows that this has probability 0 or 1 for each ρ and d.
When ρ > 1 (and even when ρ > 1 − ε for some ε = ε(d) > 0), it is not difficult to
show that there are unmatched red points (so Alice wins); the question is whether this
holds for every positive ρ. This is not known for any dimension d, although in [16] it is
proved that for any fixed ρ > 0, there exists d0 = d0 (ρ) such that there are unmatched
red points for all d ≥ d0 .
Proof. With misère rules, all terminal positions are N-positions, and the characteriza-
tion of N-positions and P-positions is modified by replacing condition (N) with:
(N ) From every N-position that is not terminal, there is at least one move to a
P-position.
As before, to check that a claimed partition into N-positions and P-positions is correct,
it suffices to show that it satisfies (P) and (N ).
Assume that x and y are not mutually closest and are not matched in M (accord-
ing to the claim, they hence define an N-position that is not terminal). If both
This argument shows that the misére friendly frogs is a win for Alice if and only if
the restricted stable matching has exactly one unmatched point.
Corollary 10. Let L be a Poisson process on Rd . Misère friendly frogs is a win for
Bob with probability 1.
Proof. The argument in the proof of Theorem 4 shows that the matching M is perfect
with probability 1: it is impossible for the unmatched points to form a nonempty finite
translation-invariant random set.
Proposition 11. Let L ⊂ Rd have distinct distances and no infinite descending chains,
and assume that the stable m-matching of L is perfect. Consider k-stone colored
friendly frogs. Suppose that the frogs are at x and y, with y being the frog of the
next player. This position is a P-position if and only if, in the stable (k + 1)-matching,
x desires y, and all partners of x that are closer than y are blocked by stones.
398
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Figure 6. Left: the stable 3-multimatching of random points in a torus. Right: pairs having friendly frogs
Sprague–Grundy values 0 (black, thick), 1 (red, medium), and 2 (blue, thin), for the same points.
Proof. We check (N) and (P). For (P), suppose the given conditions hold. Since x
desires y we have D(x) ≥ |x − y|. Thus, if the next player moves their frog from y
to z, then x also desires z. But z is not a partner of x, because we assumed that all
possible such z are blocked. Therefore z does not desire x (otherwise they would be
unstable), so the new position is an N-position (regardless of where the player moves
the stones). We now check (N). If x desires y but some closer partner z of x is not
blocked, then the next player can move to z. On the other hand, if x does not desire
y, then all the partners of x are closer than y, and at least one of them, z say, is not
blocked, so the next player moves there. In either case, this player then blocks all k
partners of z other than y.
Proof. We claim that the stable (k + 1)-matching is perfect with probability 1. Indeed,
there are at most k + 1 incompletely matched points. But the invariance argument
of Theorem 4 shows that a translation-invariant random set of points cannot have a
positive finite number of points with positive probability.
By Proposition 11, Bob wins by placing his frog on an unblocked partner z of the
location y of Alice’s opening frog, and placing stones on the other k partners of z.
Multiple Ponds and Bitwise XOR. Finally, we address how to play several games
of friendly frogs simultaneously. Consider k sets L 1 , . . . , L k ⊂ Rd , each assumed to
have no infinite descending chains and all distances distinct. (We imagine k disjoint
ponds, each with its own set of lily pads.) In a position of k-pond friendly frogs, each
set L i has two frogs on two distinct points. (We discuss the opening moves, in which
the frogs are placed, below.) Alice and Bob take turns, and a move consists of jumping
one frog in one set L i to a different point in the same set L i according to the usual
In fact Bob has a unique good opening move that depends in an intricate way on
Alice’s 2k − 1 initial frogs. The key to the proof is the following result extending stable
matching to an integer-valued labeling of all pairs of points. Write N := {0, 1, 2 . . .}.
For S N, let mex S := min(N \ S) be the minimum excluded value. For a set
L ⊂ Rd and an unordered pair of distinct points x, y of L, let F(x, y) be the set of
positions to which one can legally move in friendly frogs, that is, pairs that are strictly
closer to each other than x, y and share exactly one point with x, y.
Proof. As before, we construct the map via an iterative algorithm. Start with G(x, y)
undefined for all x, y. We say that each point of x ∈ L looks at the closest other point
y for which G(x, y) is currently undefined. For every pair x, y that are looking at
each other, set G(x, y) equal to the smallest nonnegative integer that is not currently
assigned to any pair containing x or y. Now repeat indefinitely.
We first check that the resulting G assigns an integer to every pair of points. Indeed,
if G(x, y) is undefined, then x, y never looked at each other, and so one of them, say
y, must have a closer point z for which G(y, z) is undefined. Passing to the closest
such z and iterating gives an infinite descending chain, a contradiction.
We now check the claimed properties. For (i), it is immediate that no two pairs con-
taining x are assigned the same integer. It remains to check that some pair containing x
has the label k. Let Uk be the set of points x that are not contained in any pair with label
G(x, y) = k. By invariance, if Uk is nonempty, then it is infinite. Let W ⊆ Uk be any
set of size k + 2. By the pigeon-hole principle, there exist u, v ∈ W with G(u, v) >
k. But this is a contradiction: the algorithm should instead have assigned u, v a
value of at most k.
To check (ii), note that, during the stages of the algorithm, a given point looks at
other points of L in order of increasing distance (perhaps looking at the same point for
multiple consecutive stages). Therefore, when the algorithm assigns a value to the pair
x, y, all pairs in F(x, y) have been assigned values, while all other pairs that share a
point with x, y have not. Hence G(x, y) is assigned the mex as claimed.
400
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
It is easy to see that the set of pairs {x, y} with G(x, y) = 0 is precisely the sta-
ble matching. More generally, G(x, y) is the so-called Sprague–Grundy value of the
associated position. Note, however, that the set of pairs with G(x, y) ≤ m does not in
general coincide with the m-matching considered earlier. See Figure 6. It should also
be noted that the analog of property (i) in Proposition 14 does not hold in general for
finite sets L, since it is possible that for some x the set {G(x, y) : y ∈ L \ {x}} is not
the interval {0, . . . , L − 2}.
Proof
of Theorem 13.Let ⊕ denote bitwise XOR of binary expansions, so if a =
j∈N α j 2 and b = j∈N β j 2 with α j , β j ∈ {0, 1}, then a ⊕ b := j∈N σ j 2 where
j j i
ACKNOWLEDGMENT. We thank the referees for very careful reading and helpful comments.
REFERENCES
1. R. Basu, A. E. Holroyd, J. B. Martin, J. Wästlund, Trapping games on random boards, Ann. Appl. Prob.
(forthcoming).
2. I. Benjamini, R. Lyons, Y. Peres, O. Schramm, Group-invariant percolation on graphs, Geom. Funct.
Anal. 9 (1999) 29–66.
3. E. R. Berlekamp, J. H. Conway, R. K. Guy, Winning Ways for Your Mathematical Plays. Second ed.
Vol. 1. A K Peters, Ltd., Natick, MA, 2001.
4. C. L. Bouton, Nim, a game with a complete mathematical theory, Ann. Math. 3 no. 2 (1901/02) 35–39.
5. D. J. Daley, G. Last, Descending chains, the lilypond model, and mutual-nearest-neighbour matching,
Adv. Appl. Probab. 37 (2005) 604–628.
6. D. J. Daley, D. Vere-Jones, An Introduction to the Theory of Point Processes. Second ed. Vol. II, Proba-
bility and its Applications. Springer, New York, 2008.
7. M. Deijfen, O. Häggström, A. E. Holroyd, Percolation in invariant Poisson graphs with i.i.d. degrees,
Ark. Mat. 50 (2012) 41–58.
8. M. Deijfen, A. E. Holroyd, Y. Peres, Stable Poisson graphs in one dimension, Electron. J. Probab. 16
(2011) 1238–1253.
9. M. Deijfen, F. M. Lopes, Bipartite stable Poisson graphs on R, Markov Process. Related Fields 18 (2012)
583–594.
10. D. Gale, L. S. Shapley, College admissions and the stability of marriage, Amer. Math. Monthly 69 (1962)
9–15.
11. P. M. Grundy, Mathematics and games, Eureka 2 (1939) 6–8.
MARIA DEIJFEN received her Ph.D. in 2004 from Stockholm University, where she is currently a professor.
Her research area is discrete probability theory, with particular emphasis on spatial structures and random
graphs.
Department of Mathematics, Stockholm University, 106 91 Stockholm.
mia@math.su.se
ALEXANDER E. HOLROYD received his Ph.D. in 2000 from the University of Cambridge. Before joining
the Theory Group at Microsoft Research, he held positions at the University of California in Los Angeles and
Berkeley, and the University of British Columbia. He works on discrete probability theory with emphasis on
percolation, cellular automata, matching, and coupling.
Microsoft Research, 1 Microsoft Way, Redmond, WA 98052, USA.
holroyd@microsoft.com
JAMES B. MARTIN received his Ph.D. in 1999 from the University of Cambridge. After working in Paris
for INRIA and for the CNRS, in 2005 he moved to Oxford, where he is based in the Statistics Department and
at St Hugh’s College. He works in probability theory, with particular interests including interacting particle
systems, models of random growth and percolation, and models of coalescence and fragmentation.
Department of Statistics, University of Oxford, 24-29 St Giles’, Oxford OX1 3LB, United Kingdom.
martin@stats.ox.ac.uk
402
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Holditch’s Ellipse Unveiled
Juan Monterde and David Rochera
Abstract. In plane geometry, Holditch’s theorem states that if a chord of fixed length is
allowed to rotate inside a convex closed curve, then the locus of a point on the chord a distance
p from one end and a distance q from the other is a closed curve whose area is less than that
of the original curve by π pq. In this article we obtain, first, sufficient conditions to ensure
the existence of the Holditch curve and, second, a version of Holditch’s theorem for convex
polygons where the ellipse involved is explicitly shown.
q
C p
C1
πpq
Our interest in this theorem started from the inclusion of this result as one of Clif-
ford Pickover’s 250 milestones in the history of mathematics [5]. In the two pages
of the book devoted to the theorem, the author echoes the question raised by Mark
Cooker in [2]. Namely, the statement refers to an ellipse, but. . . where is it? The proof
of Holditch’s theorem, after adding some necessary hypotheses as can be found in [1],
uses usual techniques from calculus, but no reference appears to any ellipse other than
the fact that the area is π pq.
In this article we will try to address this missing ellipse. The first steps were already
presented in the cited paper [2]: If C1 is a rectangle and the stick is shorter than the
shortest side, then the Holditch curve consists of four similar quarter-ellipses. More-
over, in the same paper it is also said that if the angle between two consecutive sides
is not a right angle, then a part of an ellipse still appears although the semiaxes are no
longer p and q. Our investigation starts by noting that in this case (not a right angle)
http://dx.doi.org/10.4169/amer.math.monthly.124.5.403
MSC: Primary 52A10
404
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
p α(g(s)) + ( − p) α(h(s))
Hα (s) = , s ∈ S1. (1)
Holditch curves may have retrograde motion. This happens when one of the end-
points of the chord passes through the same point at least twice, as happens in an
equilateral triangle. If both maps g and h are homeomorphisms of the circle, then
there is no retrograde motion. When g is a homeomorphism, Holditch curves can be
reduced to
p α(s) + ( − p) α( f (s))
Hα (s) = , s ∈ S1, (2)
where f = h ◦ g −1 .
First, we give the conditions that ensure the existence of Holditch curves.
Proof. We will prove the existence of the Holditch curve as an application of the
implicit function theorem. Let us define F : S 1 × S 1 → R as
Condition i) ensures that there are s0 and t0 such that F(s0 , t0 ) = 2 . More-
over, condition ii) is equivalent to saying that if s, t ∈ S 1 are such that (α(s) −
α(t)) ⊥ α (t), then α(s) − α(t) = . Since α(s0 ) − α(t0 ) = , we have that
α(s0 ) − α(t0 )), α (t0 )
= 0. (Figure 3 illustrates the geometric interpretation of the
second condition.)
Therefore, ∂∂tF (s0 , t0 ) = 0 and we can apply the implicit function theorem. There
exist Us0 , a neighborhood of s0 , Vt0 , a neighborhood of t0 , and a C 1 -function f 0 :
Us0 → Vt0 such that f 0 (s0 ) = t0 and, for any s ∈ Us0 ,
F s, f 0 (s) = 2 .
This means that for any s ∈ Us0 , α(s) − α( f 0 (s)) = . We have shown that if there
is one possible position of a chord of length with endpoints on the trace of α, specif-
ically α(s0 ) and α(t0 ), then a piece of any -Holditch curve can be built in neighbor-
hoods of s0 and t0 .
The extension of f 0 , and therefore of the Holditch curve, to the whole S 1 follows
from a typical argument using connectedness properties of S 1 . Let A ⊂ S 1 be the
subset where f 0 can be extended; that is,
F(s, f 1 (s)) = 2 .
α(s) − α( f (s)) = ,
for all s ∈ S 1 , which is the map that defines the -Holditch curves in α, according to
expression (2). This means that g = Id (which is obviously a homeomorphism) and
h = f is well-defined if we are working under the hypothesis of Proposition 1. We
shall restrict ourselves to the case without retrograde motion. In such a case, the map
f : S 1 → S 1 can be defined as follows: if the chord has an endpoint at α(s), then the
other endpoint is α( f (s)) according to the orientation of the curve (see Figure 2).
In this case we will say that α is -Holditch admissible (or, simply, admissible)
and we will denote by C−ad 1
(resp., P−ad
1
) the space of simple closed C 1 (resp.,
piecewise C ) -Holditch admissible curves.
1
α ( f (s))
α (s)
406
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Remark 1. When the initial curve is a circle, the associated map f : S 1 → S 1 such
that α(s) − α( f (s)) = is simply a translation f (s) = s + s0 . The value of s0 can
be computed explicitly in terms of the radius r of the circle and the length :
s0 = 2 arcsin .
2r
r(s) α (f (s))
α (s)
Example 1. Computation
of the Holditch radius for an ellipse. Let α(s) =
a cos(s), b sin(s) be the usual parameterization of an ellipse with semiaxes a and b.
The computation of its Holditch radius is a typical minimization problem with restric-
tions. Let
We have to compute the minimum value of F(s, t) under the restriction α(s) −
α(t), α (t)
= 0. Since
= 2a 2 cos(s) − cos(t) sin(t) + 2b2 cos(t) − sin(s) + sin(t) ,
27a 4 b4
4a 2 , 4b2 , 0, 3 .
a 2 + b2
27a 4 b4
Excluding the 0, the minimum is the last value (a 2 +b2 )3
. For example, taking a = 2 and
b = 1, the Holditch radius is
27a 4 b4 432
3 = = 3.456.
a 2 + b2 125
1.0
0.5
−
−2
2 −1 1 2
− 0.5
−1.0
1.0
The next result shows a sufficient condition to have an admissible curve (i.e., no
retrograde motion).
Theorem 1. Let α be a convex simple closed curve, let R H > 0 be its Holditch radius,
and > 0. If < R H , then α is -Holditch admissible.
408
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Proof. Since < R H , the circle of radius centered at a point α(s0 ) will intersect the
trace of the curve at two points in a nontangential way. Therefore, conditions i) and
ii) in Proposition 1 hold and show the existence of -Holditch curves in α. Moreover,
the expressions of those curves are reduced to (2). Thus, to prove that α is -Holditch
admissible, we only have to see that the map f : S 1 → S 1 is injective. We define the
function F̄ : S 1 → R as
F̄(s) = F s, f (s) ,
Now we know that F̄ is constant equal to 2 by definition of f . Thus F̄ (s) = 0 for all
s ∈ S1.
Since < R H , we have ∂∂tF (s, f (s)) = 0 for all s ∈ S 1 . If we have ∂∂sF (s, f (s)) = 0,
then f (s) = 0 for all s ∈ S 1 , which would imply the injectivity of f . Let us show that
∂F
∂s
(s, f (s)) = 0. We will prove that ∂∂sF (s, f (s)) = 0 implies ≥ R H . Let s ∈ S 1 . The
fact that
means that α(s) − α( f (s)) is one of the distances considered in the minimum
r ( f (s)), f (s) ∈ S 1 . By definition of R H , we have
α(s) − α( f (s)) ≥ R H ,
Remark 2. In Theorem 1 we have seen how to obtain Holditch curves with the map
f : S 1 → S 1 injective depending on the length of the Holditch chord. We will now
show the geometric interpretation that is behind this result. For any point α(s) on a
closed curve α, let us consider the intersection between the curve and circle centered
at α(s) with radius . If, for some point α(s0 ) the intersection is empty, then α is not
-Holditch admissible. If the intersection is reduced to just one point α(t0 ), then the
circle is tangent to α at α(t0 ), and thus α(s0 ) − α(t0 ), α (t0 )
= 0. If the intersection
has 3 or more points, then the function F(s0 , t) = α(s0 ) − α(t)2 will have some
local minimum at some t0 with F(s0 , t0 ) ≤ . Therefore, the case we were looking for
is when, for any point α(s), the intersection is made up of only two different points
(see Figure 5), and this happens only when we have no retrograde motion. According
to the orientation of the curve, one of the intersection points, let us say α(t1 ), will be
the one next to α(s0 ). Hence, the map f : S 1 → S 1 is well-defined as f (s0 ) = t1 and
the curve is -Holditch admissible.
Example 2. The Holditch radius can be zero but a Holditch curve still exists, as hap-
pens in a square or in an equilateral triangle. In this latter case, an equilateral triangle,
retrograde motion appears as the segment traces the Holditch curve.
Now let us study the continuity of the map that sends any admissible curve (thus,
without retrograde motion) to its Holditch curve. First we will need a preliminary
result.
α(f(s ))
2
α(s ) 2
α(f(s ))1
α(f(s ))
1
α(s )
1
α(s )
2
α(s )
1
Figure 5. If is less than the Holditch radius, the intersection between the curve and circle centered at α(s)
and with radius is reduced to only two points (left). If at some point that intersection has three or more points,
then we will have retrograde motion (right).
Lemma 1. Let α and β be two simple closed and admissible curves. Let us denote by
f (resp., g) : S 1 → S 1 the map of the circle defining the Holditch curve of α (resp.,
β); that is,
p α(s) + ( − p) α( f (s)) p β(s) + ( − p) β(g(s))
Hα (s) = , Hβ (s) = .
Given η > 0, there is δ > 0 such that, if α(s) − β(s) < δ for all s ∈ S 1 , then
f (s) − g(s) < η for all s ∈ S 1 .
Since α is -Holditch admissible, then for all s0 ∈ S 1 there is t0 = f (s0 ) such that
α(s0 ) − α(t0 ) = . Equivalently, F(s0 , 0, t0 , 0) = (2 , 0). Now, it is easy to check
that the 2 × 2 matrix given by the partial derivatives
∂F
∂t (s0 ,0,t0 ,0)
= −2 α(s0 ) − α(t0 ), α (t0 )
, 0 ,
∂F
∂v (s ,0,t ,0)
= (∗, −1),
0 0
has a maximal rank. We can then apply the implicit function theorem. There are neigh-
borhoods Us0 of (s0 , 0) and Vt0 of (t0 , 0) and a continuous map f s0 : Us0 → Vt0 such
that f s0 (s0 , 0) = (t0 , 0) and, for all (s, u) ∈ Us0 ,
F s, u, f s0 (s, u) = (2 , 0). (4)
s s
If we write f s0 (s, u) = f 1 0 (s, u), f 2 0 (s, u) , then (4) is equivalent to
s
f 2 0 (s, u) = u,
s s
α(s) + u h(s) − α( f 1 0 (s, u)) + u h( f 1 0 (s, u)) = . (5)
410
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Since S 1 is compact, there is a finite subcover {Ui }i=1 n
where continuous maps
f i : Ui → Vi are defined. If an intersection Ui ∩ U j is not empty, then f i and f j
agree on (s, 0) ∈ Ui ∩ U j . Thus, we can suppose that there is a continuous map
f 1 : S 1 ×] − 2δ0 , 2δ0 [→ S 1 with δ0 > 0. Since f 1 is continuous, given η > 0, there is
some δ > 0, which can be supposed to be less than δ0 , such that if |u| < δ, then
f 1 (s, 0) − f 1 (s, u) < η.
β(s) − α(s)
= α(s) + α(s) − β(s)
α(s) − β(s)
≤ α( f (s)) − β( f (s)) + β( f (s)) − β(g(s))
ε ε ε
≤δ+ < + = ε.
2 2 2
1
Hα (s) − Hβ (s) ≤ p α(s) − β(s) + ( − p) α( f (s)) − β(g(s))
1
≤ ( p ε + ( − p) ε) = ε.
Remark 3. Notice that, once we have proven the continuity of the Holditch map, the
existence result of Holditch curves can be extended to the piecewise case. Indeed, any
piecewise C 1 closed curve is the limit of a sequence of C 1 closed curves.
If, in addition, a horizontal translation of vector x0 is applied, the resulting linear map
will be denoted by Sc,x0 : R2 → R2 and its expression is
There are two main properties of shear mappings that we are going to need. First,
since their Jacobians are equal to 1, shear transformations preserve areas. Second,
shears transform orthogonal ellipses into oblique ellipses.
The following result indicates what kind of curve the Holditch curve is in the case
we are focusing on and shows a way to find out where the area of the Holditch ellipse
comes from. Let us recall, as stated in [6], page 65, that the Holditch curve associated
to an angle is a particular case of the ellipse construction invented by Leonardo da
Vinci.
412
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
θ
Figure 6. Holditch’s curve is an arc of an oblique ellipse which in turn is the image by a shear of an orthogonal
ellipse with semiaxes p and − p. Its area is θ2 p ( − p), where θ is the external angle.
Proof. Let us denote by θ the external angle formed by the two semistraight lines and
let t be the angle between the Holditch chord and one of the lines. We will use the
quantity t to parameterize the Holditch curve. The situation is represented in Figure 7.
B
−p
C
sin t sin t
p sin θ
A t θ
cos t − sin t cot θ O sin t cot θ V
We set the origin of the coordinates O = (0, 0) at the intersection of our two semi-
straight lines. Let A = A(t) and B = B(t) be the points defined by the ends of the
chord in the semistraight lines. Then it is easy to notice that the only thing we need
to be able to parameterize the Holditch curve is to know the coordinates of A and B,
since such parameterization would be defined as
1
γ (t) = ( − p) A + p B , t ∈ [0, θ]. (6)
If we focus on the triangle OVB and call its hypotenuse v, then we have sin(t) =
v sin(θ), from which it follows that v = sin(θ)
sin(t)
. Thus, the side O V of the triangle has
length v cos(θ) = sin(t) cot(θ). From this we get
B = sin(t) cot(θ), sin(t) .
Now we focus on the triangle AVB. If the length of the side AO is called x, then
cos(t) = x + sin(t) cot(θ). From that we obtain x = cos(t) − sin(t) cot(θ).
γθ, p, (t) = −( − p) cos(t) + sin(t) cot(θ), p sin(t) , t ∈ [0, θ] (7)
Now that we know the explicit parameterization of the Holditch curve given in (7),
we can see what it is like in some easy examples. For instance, Figure 8 shows the inner
curve when the initial curve is a square as well as when it is an equilateral triangle.
In these examples, we can understand the usefulness of the shear map defined in the
previous proof at each vertex of the polygonal curve. In the square we have four equal
parts of an ellipse with semiaxes lengths 1 and 1 − p and in the triangle, three pieces
of an oblique ellipse (obtained by a shear transformation of an orthogonal one). From
the fact that shear maps preserve areas, we deduce that the area in both examples is the
same as the area of the Holditch ellipse, as stated in Proposition 2.
Figure 8. Holditch curve in a square and in an equilateral triangle with different choices of p.
We will use the previous basic result about Holditch curves in the following two
sections, in which we conduct a separate study depending on the length of the chord.
414
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
4. POLYGONS WITH LONG SIDES. According to the previous section, we can
now easily deal with polygons with all sides longer than the length of the moving
chord. The next result shows that if we are in this case, we are finished.
Proposition 3. The area between a convex closed polygonal curve with sides of length
larger than the length of the moving chord and its Holditch curve is equal to the area
of an ellipse with both semiaxes lengths being the same as those into which the point
divides the chord.
Proof. Suppose that the initial curve is an n-sided convex closed polygon with external
angles θi , i = 1, . . . , n, defined at each vertex. By Proposition 2, the area at each
vertex between the polygon and the Holditch curve is the area of a (0, θi )-sector of an
orthogonal ellipse with semiaxes lengths p and − p. That area is well known and
equal to
θi
p ( − p).
2
Therefore, if we have n vertices, the total Holditch area would be
n
θi 1 n
p ( − p) = p ( − p) θi .
i=1
2 2 i=1
n
Now i=1 θi = 2 π since it is the sum of all the external angles in a simple closed
polygon. Thus we conclude that the total Holditch area is π p( − p), the area of the
ellipse with semiaxes lengths p and − p that we were looking for.
5. CUTTING OFF A CORNER. In the previous section we have seen how to deal
with the case of a polygonal curve with all sides longer than the length of the moving
chord. Since our interest is to build a sequence of convex polygonal curves approxi-
mating a curve, the next step is to deal with a polygonal curve, some of the sides of
which are shorter than the chord.
The situation now is that the endpoints of the moving chord could not be on con-
secutive sides of the polygon. Then the Holditch curve will be piecewise defined, as
we show in Figure 9.
θ2
θ1 θ1+θ2
C1 C2
A2 A1 O2 O1
Lemma 2. Let C1 be the oblique ellipse defined as the image of an orthogonal ellipse
by the horizontal shear transformation Sc1 (x, y) = (x + c1 y, y), and let C2 be the
oblique ellipse defined as the image of the same orthogonal ellipse by a horizontal
shear transformation plus a horizontal translation Sc2 ,x0 (x, y) = (x + x0 + c2 y, y)
with x0 < 0 (see Figure 10). Suppose that c1 < c2 . Let Ai be the leftmost intersection
between the line joining the centers O1 and O2 of both ellipses and the ellipse Ci ,
i = 1, 2. Among the four possible intersection points between the two ellipses, let P
be the one located above that line and to the left. The area defined by the three points
A1 , A2 , and P is equal to the area of a triangle with a base of length −x0 and height
x0
c1 −c2
.
Proof. If we apply Sc1 −c2 to the second ellipse, then what we get is the same first ellipse
C1 but translated according to the vector (x0 , 0). The region (P, A2 , C) (see Figure 11)
is transformed into the region defined by points (Q, A2 , C). The area we are looking
for is the area of (P, A2 , C) minus the area of (P, A1 , C). Since shear transformations
preserve area, the area of (P, A2 , C) is the same as the area of (Q, A2 , C). Moreover,
the area of (P, A1 , C) is the same as the area of (Q, A2 , D). Therefore, the area of
(P, A2 , A1 ) is the same as the area of the triangle (Q, D, C). The length of its base is
the norm of the vector (x0 , 0) and its height is the height of the intersection point P. A
straightforward computation shows that the second coordinate of P is c1x−c 0
2
.
416
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Q P
A2 D A1 C
The following idea illustrates the purpose of the previous result. The case we are
focusing on now, as we said before, is the one shown in Figure 9, which is none other
than the basic case of two semistraight lines that form an external angle θ1 + θ2 , but
cutting off its corner with another straight line that forms angles θ1 and θ2 with each
of the first two lines. Now, in this case we want to study the new area between the
Holditch curve, which is piecewise defined, and the sides of the polygon. The idea is
to prove that the area of the triangle removed by the new line is equal to the area of
both the new bumps of area that appear on each side. This is what the following results
deal with. See Figure 12.
Figure 12. The area of the removed triangle is equal to the sum of the areas defined by the two new pieces of
the Holditch curve.
In what follows in this section, for the sake of simplicity we will assume that = 1.
Clearly this can be achieved with a scale change.
Lemma 3. The area of the left bump is equal to p times the area of the removed
triangle.
Proof. We will apply Lemma 2 to the situation shown in Figure 9. In this case, the
two shears are defined by the parameters (according to the notation in the statement of
Lemma 2):
where c1 < c2 and x0 < 0. Therefore, the area of the left bump is equal to
On the other hand, let us compute the area of the removed triangle. We can get its
height from the system of equations
h
−x0 +a
= tan(θ1 ),
h
a
= tan(θ1 + θ2 ),
Proposition 4. The area of the removed triangle is equal to the sum of the areas of the
two new bumps of the Holditch curve.
1-p
p 1-p
Proof. Let AT be the area of the removed triangle. As we have seen, the area of the
left bump is equal to p AT . Thanks to a flip with respect to the bisecting line of the
inner angle at the corner, we can interchange both bumps (see Figure 13). Notice that
the flip also interchanges the lengths p and q = 1 − p in the chord. Hence, the area of
the second bump is equal to (1 − p) AT . Thus, the sum of the areas of both bumps is
equal to AT .
418
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
A new bump
A removed
A small
Figure 14. Case when the straight line that cuts off the corner also cuts the Holditch curve.
We remark now that in the previous procedure we have supposed something implic-
itly. The fact is that, depending on the choice of p, there is another possibility that we
have not talked about. This appears when the straight line that cuts off the corner also
cuts the Holditch curve we had. In this case we do not really remove the full area
inside the triangle, only Aremoved (see Figure 14). It is easy to check that the sense of
the previous statement is still true in this case, i.e., the sum of the areas of both bumps
is equal to the removed area. Indeed, let us suppose, without loss of generality, that the
left bump is the one with problems. We have proved before that
Then, with the notation of Figure 14, since AT = Aremoved + Asmall and Aleft bump =
Anew bump + Asmall , we conclude that
as we wanted to see.
If the previous problem appears in both bumps, we can argue similarly with each of
them to reach the same conclusion, i.e., the sum of the areas of both bumps is equal to
the removed area.
Figure 16. The three first terms in the sequence toward the Holditch curve of a circle (above) and of a convex
planar curve (below).
Since the sequence of polygons {Pn }∞n=2 converges to the curve α and the map send-
ing any admissible curve to its Holditch curve is continuous (Theorem 2), then the
sequence of Holditch curves {Hn }∞n=2 converges to the Holditch curve of α. Intuitively,
this shows that the Holditch area is the result of an infinite number of “cutting off a
corner” steps which distribute the area of the Holditch ellipse by pasting small pieces
of it.
ACKNOWLEDGMENT. The authors wish to thank Mark Cooker for his interest and careful reading of the
preprint we sent him. We have taken note of his comments to improve some details in the writing. The first
author is partially supported by DGICYT grant MTM2012-33073.
REFERENCES
420
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Note added in proof. After the acceptance of this paper, an article appeared where
the the existence of Holditch curves is also studied: H. Proppe et al., On Holditch’s
theorem and Holditch curves, Journal of Convex Analysis 24 (February, 2017) 239–
259.
JUAN MONTERDE received his Ph.D. from the University of Valencia in 1988. His interests range from
classical differential geometry to computer-aided geometric design.
Dept. of Mathematics, University of Valencia, Avd. Vicent Andrés Estellés, 1, E-46100-Burjassot (València),
Spain
monterde@uv.es
DAVID ROCHERA earned his master’s degree from the University of Valencia in 2015. He is especially
interested in the areas of classical differential geometry and applied mathematics.
Dept. of Mathematics, University of Valencia, Avd. Vicent Andrés Estellés, 1, E-46100-Burjassot (València),
Spain
David.Rochera@uv.es
Abstract. It is shown in [28] that, using only tools of elementary geometry, the classical
Steiner–Routh theorem for triangles can be fully extended to tetrahedra. In this article, we first
give another proof of the Steiner–Routh theorem for tetrahedra, where methods of elemen-
tary geometry are combined with the inclusion–exclusion principle. Then we generalize this
approach to (n − 1)-dimensional simplices. A comparison with the formula obtained using
vector analysis yields an interesting algebraic identity.
P
R L
B K C
http://dx.doi.org/10.4169/amer.math.monthly.124.5.422
MSC: Primary 97G30
422
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
1 + x yz
,
(1 + x)(1 + y)(1 + z)
(1 − x yz)2
.
(1 + x + x y)(1 + y + yz)(1 + z + zx)
Theorem 2. In the notation of Theorem 1, the lines AK, BL, and CM intersect at one
point if and only if x yz = 1.
Steiner–Routh, Ceva’s, and Menelaus’ theorems, in their general forms, are stated
using signed lengths and are closely related. We have conducted an extensive search of
the literature on these theorems and their generalizations to higher dimensions within
the context of Euclidean geometry (there are generalizations in other geometries, but
we did not include them here) that resulted in the bibliography of the present paper. We
believe that this list of articles is interesting from the historical perspective (although
we cannot guarantee its completeness) and is valuable since it represents the wide
range of generalizations of these classical theorems.
We have been able to find only two papers, [24] and [45], where the Steiner–Routh
theorem is generalized to higher dimensions. Unfortunately, neither of these papers
is readily accessible to many readers since they are written in Slovak and Chinese,
respectively. Theorem 2 of [24] uses the notation from [7], and the formula for the
volume is presented in a form that is difficult to identify as a generalization of the
Steiner–Routh theorem. On the other hand, the formula in Theorem 2 of [45] is eas-
ily identifiable, except that it is missing an absolute value. The papers employ similar
techniques. In short, the authors compute the coordinates of the vertices of relevant
(n − 1)-dimensional simplices and then evaluate their volumes using determinants. In
computing these determinants, most of the work is done with the tools of linear alge-
bra. Besides, since any two nondegenerate (n − 1)-dimensional simplices are affine
isomorphic, in [45], without loss of generality, the vertices of the initial simplex are
placed at the origin and at the points (1, 0, . . . , 0), . . . , (0, . . . , 0, 1).
The first geometric proof of the Steiner–Routh theorem for tetrahedra was given in
[28]. The purpose of the present paper is threefold. First, we present a more versatile
geometric proof of the Steiner–Routh theorem that is based on an application of the
inclusion–exclusion principle. Then we generalize this approach to obtain a geomet-
ric proof of the Steiner–Routh theorem for higher-dimensional simplices. Finally, we
present a remarkable algebraic identity (Theorem 5) that relates the formulae obtained
using our geometric approach and the approach that involves analytic geometry and
linear algebra, as described above.
Keeping in mind that we aim to generalize the next theorem to simplices, we need
to adjust the notation in this special case of the Steiner–Routh theorem for tetrahedra.
|1 − x1 x2 x3 x4 |3
V R1 R2 R3 R4 = , (2)
x13 x23 x33 x43
A1
a x1 P4
R2 Q3
d x4
P1 Q4
a R3
A2 A4
R1
c
b x2 R4 Q2
Q1 P3
P2 c x3
A3
424
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Formula (2) will be proved using geometric considerations together with the
principle of inclusion–exclusion. The inclusion–exclusion principle for finite sets
A1 , . . . , An states
n
n
A i = |Ai | − |Ai ∩ A j | + |Ai ∩ A j ∩ Ak | − · · ·
(IEP)
i=1 i=1 1≤i< j≤n 1≤i< j<k≤n
+ (−1) n−1
|A1 ∩ · · · ∩ An |,
where |A| stands for the number of elements in a finite set A.
It is clear that the inclusion–exclusion principle remains valid if finite sets are
replaced by solids and the numbers of elements of these finite sets are replaced by
the volumes of the solids.
In Section 3 of the article, we will extend our considerations to the cycle (A1 . . . An )
corresponding to a general (n − 1)-dimensional simplex A1 . . . An . Comparison of the
result with the formula given in [45] yields the algebraic identity (5).
|AP|
= v(1 + u).
|PK|
In order to simplify the notation, for given natural numbers i and j, let us define
xi xi+1 . . . xi+ j−1
xi j = 1 + xi + xi xi+1 + · · · + xi xi+1 . . . xi+ j−1 and X i j =
xi j
(see Remark 1).
D
B
vf
M
P
B ue K e C
Figure 4. Ratios.
426
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Lemma 4. Consider the triangle in Figure 4. Then
|MP| vu |MP| vu
= and = .
|PC| 1+v |MC| 1 + v + vu
A1
P1
A4
A2
Q1
P2
A3
|P1 Q 1 |
V A1 P1 Q 1 A4 = V A1 P1 A3 A4 .
|Q 1 A3 |
|P1 Q 1 |
Since, by Lemma 4, |Q 1 A3 |
= X 12 , the result follows.
The volumes VT2 ∩T3 , VT3 ∩T4 , and VT4 ∩T1 are evaluated similarly.
A1
P1
A2
R1 A4
Q1 Q2 P3
P2
A3
|Q 1 R1 |
V A1 P1 Q 1 R1 = V A1 P1 Q 1 A4 ,
|Q 1 A4 |
|Q 1 R1 |
where V A1 P1 Q 1 A4 = VT1 ∩T2 = X 11 X 12 by Lemma 5. To find the remaining ratio |Q 1 A4 |
,
consider the triangle A1 P2 A4 as depicted in Figure 7. Here, we have v = x1 (1 + x2 )
by Lemma 3 applied to the triangle A1 A2 A3 and u = 1+x x2 x3
2
by Lemma 4 applied to the
triangle A2 A3 A4 . Therefore, Lemma 4 applied to the triangle A1 P2 A4 yields
|Q 1 R1 | vu
= = X 13 ,
|Q 1 A4 | 1 + v + vu
Expressions for the volumes VT2 ∩T3 ∩T4 , VT3 ∩T4 ∩T1 , and VT4 ∩T1 ∩T2 are analogous.
Proof of (2) in Theorem 3. Assume first that x1 x2 x3 x4 > 1. Using (IEP), we obtain
VR1 R2 R3 R4 =V A1 A2 A3 A4 − VT1 − VT2 − VT3 − VT4
+ VT1 ∩T3 + VT2 ∩T4 + VT1 ∩T2 + VT2 ∩T3 + VT3 ∩T4 + VT4 ∩T1
− VT1 ∩T2 ∩T3 − VT2 ∩T3 ∩T4 − VT3 ∩T4 ∩T1 − VT4 ∩T1 ∩T2 .
428
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
A1
Q1
R1
P2 Q2 A4
|Q 1 R1 |
Figure 7. Ratio |Q 1 A4 | .
Formula (2) now follows from the previous formulae for the above volumes together
with the identity
1 − X 11 − X 21 − X 31 − X 41 + X 11 X 12 + X 21 X 22 + X 31 X 32 + X 41 X 42
+ X 11 X 31 + X 21 X 41 − X 11 X 12 X 13 − X 21 X 22 X 23 − X 31 X 32 X 33
(3)
(x1 x2 x3 x4 − 1)3
− X 41 X 42 X 43 = .
x13 x23 x33 x43
Identity (3) can be verified either manually or by using a software like Mathematica or
Maple.
The case x1 x2 x3 x4 < 1 can be treated similarly to [28] by reversing the orientation
of the cycle (A1 A2 A3 A4 ) to (A1 A4 A3 A2 ) and using the substitution x1 → x14 , x2 → x11 ,
x3 → x12 , x4 → x13 that reduces it to the case x1 x2 x3 x4 > 1.
Ai0
Aij–1
j
Ai
A0i+j
j–2 j–1
Ai+1 Ai+1
Proof. Note that all points in Figure 8 belong to the same plane. For j = 2, it follows
from the choice of the points Ai1 and Ai+11
. For j > 2, it follows by induction on j
j−1 0 j−2 j−1
since Ai lies on the line given by Ai and Ai+1 , and Ai+1 lies on the line given by
0 j−2
Ai+ j and Ai+1 .
The statement of the lemma is true for j = 2 since both Ai0 and Ai+1
1
belong to σi+1 .
For j > 2, it follows by induction on j since the point Ai belongs to σi+ j−1 and the
0
j−1 j−3 0
point Ai+1 belongs to σi+ j−1 by induction applied to the triangle Ai+1
0
Ai+2 Ai+ j.
430
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
For the inductive step, apply Lemma 3 to the triangle in Figure 8 with v = vi, j−1 and
u = u i, j−1 to infer that vi, j = vi, j−1 (1 + u i, j−1 ) = xi, j−1 − 1. When we apply Lemma
4 to the triangle in Figure 8 with v = vi+1, j−1 and u = u i+1, j−1 , we obtain u i, j =
vi+1, j−1 u i+1, j−1
1+vi+1, j−1
= X i+1, j−2 xi+ j−1 . Finally, the formula for ti, j follows from the formulae
for vi, j and u i, j with the help of Lemma 4.
n
We will determine the volume of the simplex i=1 1 ,...,
Ti (with vertices An−1
An ) using the inclusion–exclusion
n−1
principle (IEP). For this, we shall compute the
volumes of all simplices i∈I Ti , where I {1, . . . , n}. An important property of
such simplices i∈I Ti is that they contain the original vertices A0j , where j ∈ / I.
For the remainder, we will assume that I {1, . . . , n}. We now proceed to deter-
mine the vertices of
the simplices i∈I Ti and compute their volumes. When calcu-
lating the volume of i∈I Ti , the crucial role is played by the distribution of elements
i ∈ I along the cycle C = (1 . . . n). Assume that the set I consists of blocks of con-
secutive elements along the cycle C, and keep in mind that a block containing n can
start before n and continue through n to 1 and further. Denote by B (I ) the set of all
blocks of I along the cycle C. To each block of I , say B = {k, k + 1, . . . , k + l}, we
assign the expression
l+1
V (B) = Xkj .
j=1
Vi∈B Ti = V (B).
l
Vk+l−1 Ti = Xkj .
i=k
j=1
Since Tk+l has the vertices A0k , . . . , A0k+l , A1k+l , A0k+l+2 , . . . , A0n+k−1 , and all vertices
A0k , A1k , . . . , Alk are included in the simplex given by vertices A0k , . . . , A0k+l , when
k+l−1
we cut the simplex i=k Ti by Hk+l , all of its edges remain the same except the
edge Alk A0k+l+1 . The edge Alk A0k+l+1 is replaced by the edge Alk Al+1 k as can be seen
from Lemma 7 and Figure 8, because Ak belongs to σk+l . Therefore, the vertices of
l+1
|Al+1 l
k Ak |
Since |A0k+l+1 Alk |
= tk,l+1 = X k,l+1 , Lemma 8 concludes the inductive step.
Proposition 2.
Vi∈I Ti = V (B) = Vi∈B Ti .
B∈B(I ) B∈B(I )
k k +n−1
Proof. We will use the list of simplices S I 1 , . . . , S I 1 defined above and will show
that the vertices of S Ik consist of the first k − k1 + 2 vertices from the list (4) and the
vertices A0k+2 , . . . A0k1 +n−1 . By Proposition 1, this statement is true for k = k1 , . . . , k1 +
l1 , which corresponds to the first block B1 = {k1 , . . . , k1 + l1 } of I . Since the values
k = k1 + l1 + 1, . . . , k2 − 1 correspond to the indices that do not belong to I , we
k +l k +l +1 k −1
conclude immediately that S I 1 1 = S I 1 1 = · · · = S I 2 and its vertices are listed
correctly.
k k −1
The simplex S I 2 is the intersection of S I 2 and Hk2 . Since the vertices A0k1 ,
l +1 k −1
A1k1 , . . . , Ak11 , A0k1 +l1 +2 , . . . , A0k2 of S I 2 belong to the convex hull of A0k1 , . . . , A0k2 ,
k
the only edge of S I 2 that is cut by the hyperplane σk2 is the edge A0k2 A0k2 +1 . This edge
is replaced in S k2 by the edge A0k2 A1k2 , which confirms that the vertices of S k2 are
listed correctly. Taking the values of k in the second block B2 = {k2 , . . . , k2 + l2 }, we
proceed as before and always replace only one edge, analogously to that of the proof
of Proposition 1, and determine the vertices of S Ik .
Proceeding like this, in each step corresponding to k ∈ I , we replace a single edge
of S Ik−1 to obtain S Ik , while each step corresponding to k ∈ / I yields S Ik−1 = S Ik .
k
Having determined the vertices of the simplices S I , the volumes VSk are calculated
I
x
easily. The volume VSk1 = 1+xk1k by Lemma 1. If k ∈ / I , then VSk−1 = VSk . If k ∈ I and
I 1 I I
k
= k1 , then k = k j + s, where 0 ≤ s ≤ l j for an appropriate j. Then
k j Ak j |
|As+1 s
VSk = VSk−1 tk j ,s+1 , where tk j ,s+1 = .
I I |A0k j +s+1 Ais |
V l+1
S
Since VSk is the product of VSk1 and the ratios I
V l
for l = k1 , . . . , k − 1, it is clear
I I S I
that VSk1 +l1 = V (B1 ), VSk2 +l2 = V (B1 )V (B2 ), . . . , and
I I
432
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
s
Vi∈I Ti = VSk1 +n−1 = V (Bi ) = Vi∈B Ti .
I
i=1 B∈B(I )
Formula (IEP) together with Propositions 1 and 2 yield the following generalization
of Routh’s theorem (when n = 3) and Theorem 3 (when n = 4).
Theorem 4.
Vi=1
n T = V n−1
i A ...An−1
n
=1+ (−1)|I | V (B).
1
∅
= I {1,...,n} B∈B(I )
Theorem 5.
n
(
i=1 x i − 1)
n−1
|I |
1+ (−1) V (B) = n . (5)
∅
= I {1,...,n} B∈B(I ) k=1 x k,n−1
Proof. Assume x1 · · · xn > 1 and consider V∩i=1n T . By Theorem 4, this volume is given
i
by the expression
n on the left-hand side of the above equation. On the other hand, the
volume of i=1 Ti can be determined using vector analysis and determinants and, by
[45], it is equal to
n
( i=1 x i − 1)
n−1
n ,
k=1 x k,n−1
For the amusement of the reader, we display the identity (5) for n = 5:
1 − X 11 − X 21 − X 31 − X 41 − X 51 + X 11 X 12 + X 21 X 22 + X 31 X 32 + X 41 X 42
+ X 51 X 52 + X 11 X 31 + X 11 X 41 + X 21 X 41 + X 21 X 51 + X 31 X 51 − X 11 X 12 X 13
− X 21 X 22 X 23 − X 31 X 32 X 33 − X 41 X 42 X 43 − X 51 X 52 X 53 − X 11 X 12 X 41
− X 11 X 31 X 32 − X 21 X 22 X 51 − X 21 X 41 X 42 − X 31 X 51 X 52 + X 11 X 12 X 13 X 14
+ X 21 X 22 X 23 X 24 + X 31 X 32 X 33 X 34 + X 41 X 42 X 43 X 44 + X 51 X 52 X 53 X 54
(x1 x2 x3 x4 x5 − 1)4
= .
x14 x24 x34 x44 x54
Finally, formula (1) in Theorem 3 was proved in [28] as a consequence of the identity
x1 x2 x3 x4 x1 x3 x2 x4
1− − − − − −
x11 x21 x31 x21 x31 x41 x31 x41 x11 x41 x11 x21 x11 x31 x21 x41
1 − x1 x2 x3 x4
= .
x11 x21 x31 x41
x1 x2 x3 x4
1− − − −
x11 x31 x41 x41 x11 x21 x41 x51 x11 x21 x31 x51 x11 x21 x31 x41
x5 x1 x3 x1 x4 x2 x4 x2 x5 x3 x5
− − − − − −
x21 x31 x41 x51 x11 x31 x11 x41 x21 x41 x21 x51 x31 x51
x1 x2 x4 x1 x3 x4 x1 x3 x5 x2 x3 x5 x2 x4 x5
+ + + + +
x11 x21 x41 x11 x31 x41 x13 x31 x51 x21 x31 x51 x21 x41 x51
1 + x1 x2 x3 x4 x5
= .
x11 x21 x31 x41 x51
We finish by stating the formulae for the volumes of the previously considered
simplices in the special case when x1 = x2 = · · · = xn = k. In this case, the volume
|k−1|
n T = n
V = V∩i=1 i k −1
. In particular, if n = 3 and k = 2, then V = 17 ; if n = 4 and k = 2,
then V = 151 . The case when n = 3 and k = 2 is known in the literature as the area of
Feynman’s triangle.
The volume of the simplex A11 . . . A1n in the special case x1 = x2 = · · · = xn = k
k n +1 k n −1
equals V A1 A1 A1 = (k+1) n = 3 if n = 3 and k = 2 and equals V A1 A1 A1 A1 = (k+1)n = 27
1 5
1 2 3 1 2 3 4
if n = 4 and k = 2.
REFERENCES
434
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
23. M. S. Klamkin, S. H. Kung, Ceva’s and Menelaus’s theorems and their converses via centroids, Math.
Mag. 69 no. 1 (1996) 49–51.
24. T. Klein, A certain generalization of the theorems of Menelaos and Ceva (in Slovak, German summary),
Časopis Pěst. Mat. 98 (1973) 22–25.
25. J. S. Kline, D. Velleman, Yet another proof of Routh’s theorem, Crux Mathematicorum 21 (1995) 37–40.
26. S. Landy, A generalization of Ceva’s theorem to higher dimensions, Amer. Math. Monthly, 95 no. 10
(1988) 936–939.
27. J. Lipman, A generalization of Ceva’s theorem, Amer. Math. Monthly 67 (1960) 162–163.
28. S. Litvinov, F. Marko, Routh’s Theorem for tetrahedra, Geom. Dedicata 147 (2015) 155–167.
29. Q. J. Mao, An extension of Ceva’s theorem to higher dimensions, Yangzhou Shiyuan Ziran Kexue Xuebao
1 (1985) 33-35.
30. L. A. Masal’tsev, Incidence theorem in spaces of constant curvature, J. Math. Sci., 72 no. 4 (1994) 3201–
3206.
31. D. Maxin, Proving that three lines are concurrent, College Math. J. 40 no. 2 (2009) 128–130.
32. Z. A. Meizak, Companion to Concrete Mathematics. Wiley, New York, 1973.
33. F. Molnár, Über einige Verallgemeinerungen der Sätze von Ceva und Menelaos (in Hungarian; Russian,
German summaries), Mat. Lapok 10 (1959) 231–248.
34. ———, Eine Verallgemeinerung des Satzes von Ceva (in German), Ann. Univ. Sci. Budapest. Eötvös
Sect. Math., 3–4 (1960/1961) 197–199.
35. Z. Nádenı́k, L’élargissement du théoréme de Ménélaüs et de Céva sur les figures n-dimensionnelles (in
Czech; Russian, French summaries), Časopis Pěst. Mat. 81 (1956) 1–25.
36. ———, Několik vlastnostı́ vrcholových nadrovin normálnı́ho mnohoúhelnı́ka, Časopis Pěst. Mat. 81
(1956) 287–291.
37. ———, O ortocentru normálnı́ho mnohoúhelnı́ka, Časopis Pěst. Mat. 81 (1956) 292–298.
38. I. Niven, A new proof of Routh’s theorem, Math. Mag. 49 no. 1 (1976) 25–27.
39. B. J. Routh, A Treatise on Analytical Statics with Numerous Examples. Vol. I. Second ed. Cambridge
Univ. Press, London, 1909, http://www.archive.org/details/texts.
40. J. Steiner, Bemerkungen zu der zweiten Aufgabe in der Abhandlung No. 17 in diesem Hefte (in German),
J. Reine Angew. Math. 3 (1828) 201; see also J. Steiner, Gesammelte Werke. Vol. I, 1881, 163–168.
41. P. Wernicke, The theorems of Ceva and Menelaus and their extension, Amer. Math. Monthly 34 no. 9
(1927) 468–472.
42. K. Witczyński, Ceva’s and Menelaus’ theorems for tetrahedra, Zeszyty Nauk. Geom. 21 (1995) 99–107.
43. ———, Ceva’s and Menelaus’ theorems for tetrahedra. II, Demonstratio Math. 29 no. 1 (1996) 233–235.
44. ———, On some generalization of the Menelaus’ theorem, Zeszyty Nauk. Geom., 21 (1995) 109–111.
45. S. G. Yang, J. B. Qi, Higher-dimensional Routh theorem (in Chinese; English, Chinese summaries), J.
Math. (Wuhan), 31 no. 1 (2011) 152–156.
FRANTIŠEK MARKO received his Ph.D. in number theory from Slovak Academy of Sciences in Bratislava
under the supervision of Štefan Porubský and his second Ph.D. in algebra from Carleton University in Ottawa
under the supervision of Vlastimil Dlab. He held brief positions at Syracuse University and the University of
Minnesota at Duluth before joining the faculty at The Pennsylvania State University, Hazleton. His research
interests are in the areas of number theory, algebra, and representation theory.
76 University Drive, Pennsylvania State University, Hazleton PA 18202
fxm13@psu.edu
SEMYON LITVINOV received a Ph.D. in operator algebras in 1987 from Romanovsky Institute of Mathe-
matics of the Academy of Sciences of Uzbekistan and a Ph.D. in noncommutative ergodic theory in 1999 from
North Dakota State University. Before joining the faculty at The Pennsylvania State University, Hazleton, he
taught mathematics at Tashkent State University, North Dakota State University, and Saint Cloud State Univer-
sity. His main research interests lie in the area of functional analysis with emphases in operator algebras and
ergodic theory.
76 University Drive, Pennsylvania State University, Hazleton PA 18202
snl2@psu.edu
Abstract. We give a short argument showing that the set of openness points of a continuous
function from a metric space X into a metric space Y is of type G δ . If X is locally connected
and Y := R, the set of nonopenness points (of type Fσ ) coincides with the set of points of
extrema of f . We discuss which Fσ sets can be equal to the set of points of extrema for a
continuous function f from R into R, and we present a short survey of the known results on
this topic.
where
http://dx.doi.org/10.4169/amer.math.monthly.124.5.436
MSC: Primary 26A15, Secondary 54C30; 54C10
436
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Proof. Let x0 ∈ Aεδ . By the continuity of f at x0 , pick ε ∈ (0, ε] such that f (x) ∈
B( f (x0 ), δ/2) for all x ∈ B(x0 , ε ). We will show that B(x0 , ε ) ⊂ A2ε δ , which ends
2
the proof. Let x ∈ B(x0 , ε ). By the definition of A2ε δ , it is enough to check two
2
inclusions
Theorem 1. If X and Y are metric spaces and f : X → Y is continuous, then the set
Op( f ) is of type G δ .
Observe that
int A2ε δ = int Aεδ .
2
ε∈Q+ δ>0 ε∈Q+ δ>0
So by Lemma 1, we have
Aεδ ⊂ int A2ε δ = int Aεδ ⊂ Aεδ .
2
ε∈Q+ δ>0 ε∈Q+ δ>0 ε∈Q+ δ>0 ε∈Q+ δ>0
Hence,
Op( f ) = int Aεδ ,
ε∈Q+ δ>0
Now, from Theorem 1 and Corollary 1 we can infer that the set of points of local
extrema of a continuous real function on a locally connected metric space is of type Fσ .
However, this can be done simply via a direct proof in a general case dealing with an
arbitrary metric space. This fact was also observed by S. Geschke [10]. Our argument
is slightly different.
For a continuous function f : X → R, we denote by Extr( f ), Min( f ) and Max( f ),
the subsets of X that consist of points of local extrema, minima, and maxima of f ,
respectively.
Fact 1. For a continuous function f from a metric space (X, d) into R, the sets
Min( f ), Max( f ), and Extr( f ) are of type Fσ .
438
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
that for any two disjoint countable sets A, B ⊂ (0, 1) there exists a differentiable func-
tion f : [0, 1] → R that attains proper local minima exactly at points of A and proper
local maxima exactly at points of B. (Note that such a function may have some extra
nonproper extrema.) A short proof of this result can be found in [12].
In [14], Posey and Vaughan gave an elementary example of a continuous real func-
tion that has a proper local maximum at each point of a preassigned countable dense
set. A further result was obtained by Cater [5], who constructed a continuous, nowhere
differentiable function f on (0, 1) that has a proper local minimum at each point of
A and a proper local maximum at each point of B, where A and B are preassigned
disjoint countable dense subsets of (0, 1). It was shown in [6] that the subset of the
Banach space C[0, 1] that consists of functions with a dense set of points of proper
local maxima is residual (i.e., it is the complement of a countable union of nowhere
dense sets). More general studies in this direction were conducted in [3].
Among points of nonproper local extrema of a continuous function f : R → R, we
distinguish two types. The first type consists of points of local extrema belonging to
nondegenerate intervals where f is constant. We call them points of local c-extrema,
and the set of such points will be denoted by Extrc ( f ). Consider a simple example.
For the function
we have Extrc ( f ) = (−∞, −2] ∪ [−1, 1] ∪ [2, ∞). See Figure 1. However, there are
continuous functions f such that connected components of Extrc ( f ) are not closed
intervals. Namely, let
Then Extrc ( f ) = (−∞, 0) and f has infinitely many proper extrema. See Figure 2.
Figure 1. f (x) = max{0, min{1, −|x| + 2}}. Figure 2. f (x) = x sin(1/x), x > 0.
has infinitely many proper extrema and one nonproper minimum at 0 that is not a
c-extremum. See Figure 3.
However, the set of nonproper extrema of the second type can be nonempty and
perfect. Indeed, given the classic ternary Cantor set C, take a continuous function
Proof. Suppose that Extr( f ) is open, nonempty, and different from R. Fix its con-
nected component (a, b). Then −∞ < a or b < ∞. Assume that −∞ < a. By the
result of [2], we have f (x) = c for all x ∈ (a, b) (for some constant c). By the con-
tinuity of f , we have f (a) = c. However, a ∈ / Extr( f ), so the following cases are
impossible:
• f (x) ≥ c for all x ∈ (a − 1, a);
• f (x) ≤ c for all x ∈ (a − 1, a).
Hence, we can pick points s, t ∈ (a − 1, a) such that f (s) < c < f (t). We may
assume that s < t. Pick y ∈ (s, t) with f (y) = c. Let z ∈ [y, a] be such that f (z) =
max{ f (x) : x ∈ [y, a]}. Then f (z) ≥ f (t) > c, so z ∈ (y, a) and z ∈ Extr( f ). Let I
440
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
be an open component of Extr( f ) containing z. Then the endpoints of I are points of
local maxima of f , which contradicts the openness of the component I .
We propose the following negative result. First, we shall prove a lemma (see [9,
Lemma 1]).
Proof. It suffices to show that the sets f [Min( f )] and f [Max( f )] are countable.
Consider f [Min( f )]. (The proof for Max( f ) is analogous.) With any y ∈ f [Min( f )],
associate a fixed x y ∈ Min( f ) such that y = f (x y ) and an open interval I y
x y , with
rational endpoints, such that f (t) ≥ y for all t ∈ I y . The map y → I y is one-to-one
since I y = I y implies y = y . Hence, f [Min( f )] is countable.
Finally, let us focus on local c-extrema. For simplicity we will consider continuous
functions defined on [0, 1]. The following lemma is known. Proposition 3 seems to
be a folklore-like result. It characterizes points of c-extrema for monotone continuous
functions.
Lemma 3. If B is a countable union n Jn of open subintervals of J := [α, β], and
inf{|x − y| : x ∈ Jm , y ∈ Jn } > 0 for any distinct m and n, then the set W := J \ B is
uncountable (of cardinality continuum). The same holds if the intervals Jn are closed
with Jn = [α, β], or one-sided closed.
Proof. If int W = ∅, the assertion is clear. If int W = ∅, then the set W is nowhere
dense and the family of all Jn ’s is infinite. Note that the endpoints of J and of all Jn ’s
are accumulation points of W . Deleting the longest interval Jn (or one of the longest
intervals Jn ) from J , we obtain a disjoint union of closed intervals K 0 and K 1 . Then we
delete the respective longest intervals from K 0 and K 1 . We continue this construction,
which resembles that of the classic ternary Cantor set. Thus, we infer that W is a
Cantor-type perfect set, so it is uncountable. The second assertion follows from the
previous part.
Proof. To show the first assertion, consider a monotone and continuous function
f : [0, 1] → R. Note that if x ∈ Extrc ( f ), then there exists a maximal closed nonde-
generate interval I with x ∈ I ⊂ Extrc ( f ). All these intervals constitute a countable
disjoint family.
To prove the second assertion, consider the closure F := cl([0, 1] \ A). Let F0 stand
for the interior of F in [0, 1]. Then F \ F0 is a closed, nowhere dense set that, by
the Cantor–Bendixson theorem (see [7]), can be partitioned into a a perfect part P
(possibly empty) and a countable part E. If P = ∅, set g(x) := 0 for x ∈ [0, 1]. If
P = ∅, then P is a Cantor-type set, so we can consider a Cantor-type continuous and
nondecreasing function g from [a, b] onto [0, 1], where a := min P and b := max P.
Then g(a) = 0, g(b) = 1, and g is constant on each connected component of [a, b] \
P. We extend g to the whole interval [0, 1] by putting g(x) := 0 for x ∈ [0, a] and
g(x) := 1 for x ∈ [b, 1]. Let h(x) := λ(F ∩ [0, x]) for x ∈ [0, 1], where λ is Lebesgue
measure on R. Then h is nondecreasing, continuous, and h is constant on each con-
nected component of [0, 1] \ F. Such a component is contained in some connected
component of [0, 1] \ P. Therefore, f := g + h is continuous, and f is constant on
the closure of each connected component of [0, 1] \ F. We have [0, 1] \ F = int A =
n int In . Hence, f is constant on every interval In .
Suppose there exists an open interval J such that f is constant on J ∩ [0, 1] and
J \ A = ∅. By Lemma 3, the set J ∩ [0, 1] \ A is uncountable, so either J ∩ F0 = ∅ or
J ∩ P = ∅. If J ∩ F0 = ∅, then h is increasing on J ∩ F0 and so is f , a contradiction.
If J ∩ P = ∅, there exist x1 , x2 ∈ J ∩ P such that x1 < x2 and g(x1 ) < g(x2 ). Hence
f (x1 ) < f (x2 ), a contradiction. Summing up, Extrc ( f ) = A, as desired.
Remark. Consider a union A = n In of a countable disjoint family of intervals that
are nonempty open subsets of [0, 1]. We can infer from Corollary 2 that, for no con-
tinuous real function f on [0, 1], the equality A = Extr( f ) holds.
ACKNOWLEDGMENTS. We are grateful to Tomasz Natkaniec for a fruitful discussion and information on
references concerning sets of extrema of continuous functions. We thank the referees for several suggestions
for improvements to the paper.
442
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
REFERENCES
1. A. V. Arhangel’skii, Some metrization theorems, Uspehi Mat. Nauk, 18 (1963) 139–145 (in Russian).
2. E. Behrends, S. Geschke, T. Natkaniec, Functions for which all points are local extrema, Real Anal.
Exchange 33 (2007/2008) 467–470.
3. A. Bella, J. J. Charatonik, A. Villani, Many continuous functions have many proper local extrema,
J. Math. Anal. Appl. 154 (1991) 558–571.
4. W. G. Bloch, Open discontinuous maps from Rn onto Rn , Amer. Math. Monthly 122 (2015) 268–271.
5. F. S. Cater, Functions with preassigned local maximum points, Rocky Mountain J. Math. 15 (1985)
215–217.
6. V. Drobot, M. Morayne, Continuous functions with a dense set of proper local maxima, Amer. Math.
Monthly 92 (1985) 209–211.
7. R. Engelking, General Topology, PWN, Warsaw, 1977.
8. A. Fedeli, A. Le Donne, On metric spaces and local extrema, Topology Appl. 156 (2009) 2196–2199.
9. M. Filipczak, G. Ivanova, J. Wódka, Comparison of some families of real functions in porosity terms,
Math. Slovaca (forthcoming).
10. S. Geschke, Functions with many local extrema, KURENAI (Kioto University Research Information
Repository) (2008), 1619: 43–47; URL: http://hdl.handle.net/2433/140207
11. L. Holá, A. K. Mirmostafaee, Z. Piotrowski, Points of openness and closedness of some mappings,
Banach J. Math. Anal. 9 (2015) 243–252.
12. V. Kelar, On strict local extrema of differentiable functions, Real Anal. Exchange 6 (1980–1981)
242–244.
13. K. Kuratowski, Topology, Vol. 1, Academic Press, New York, 1966.
14. E. E. Posey, J. E. Vaughan, Functions with a proper local maximum in each interval, Amer. Math. Monthly
90 (1983) 281–282.
15. A. Schoenflies, Die Entwickelung der Lehre von den Punktmannigflatigkeiten, Jahresbericht Deutschen
Mathematiker-Vereinigung 8, Leipzig, 1900.
16. Z. Zalcwasser, Sur le fonctions de Köpcke, Prace Mat. Fiz. 35 (1927–1928) 57–99.
MAREK BALCERZAK received his Ph.D. in mathematics from the Łódź University in 1983. Since 2000,
he has been a full professor at the Łódź University of Technology. His research interests are real analysis,
measure theory, and descriptive set theory.
Institute of Mathematics, Łódź University of Technology, Wólczańska 215, 93-005 Łódź, Poland
marek.balcerzak@p.lodz.pl
MICHAŁ POPŁAWSKI received his M.Sc. degree from the Łódź University of Technology in 2015. Then
he started Ph.D. studies in mathematics at this university.
Institute of Mathematics, Łódź University of Technology, Wólczańska 215, 93-005 Łódź, Poland
michal.poplawski.m@gmail.com
JULIA WÓDKA received her M.Sc. degree from the Łódź University of Technology in 2013. Now, she is a
Ph.D. student of the fourth course of mathematics at this university. She is preparing her Ph.D. thesis in real
analysis.
Institute of Mathematics, Łódź University of Technology, Wólczańska 215, 93-005 Łódź, Poland
JuliaWodka@gmail.com
http://dx.doi.org/10.4169/amer.math.monthly.124.5.444
444
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Mathematical Evolution
Jeremiah Farrell and William Johnston
1 2 3 4 5 6 7 8 9 10 11 12 13
14 15 16
17 18 19
20 21 22
23 24 25 26 27 28 29 30
31 32 33
34 35
36 37 38 39
40 41
42 43 44 45 46
47 48 49
50 51 52 53
54 55 56 57 58 59
60 61 62
63 64 65
The clues begin on the left, on page 444. The Solution is on page 479.
Extra copies of the puzzle can be found at the Monthly’s website, http://www.maa.org/amm_
supplements.
Abstract. We show that if two monic polynomials with integer coefficients have a square-free
resultant, then all positive divisors of the resultant arise as the greatest common divisor of the
values of the two polynomials at a suitable integer.
Throughout this paper, f, g ∈ Z[x] are monic polynomials with integer coefficients:
and
N = 8424432925592889329288197322308900672459420460792433.
This number N has 52 digits, and the gcd for n = N is the 52-digit prime
p = 8936582237915716659950962253358945635793453256935559.
Turning to the general case, let r = R( f, g) ∈ Z be the resultant of the two poly-
nomials. Recall that, by definition, r is the determinant of the Sylvester matrix
http://dx.doi.org/10.4169/amer.math.monthly.124.5.446
MSC: Primary 11C08, Secondary 13P15; 15A03
446
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
⎛ ⎞
a0 a1 ... ak
⎜ a0 a1 ... ak ⎟
⎜ ⎟
⎜ ... ... ... ... ⎟
⎜ ⎟
⎜ a0 a1 ... ak ⎟
M =⎜ ⎟ (3)
⎜b0 b1 ... bl ⎟
⎜ ... ⎟
⎜ b0 b1 bl ⎟
⎝ ... ... ... ... ⎠
b0 b1 ... bl
of the two polynomials. Note that M is an (l + k)-square matrix; the first l rows are
built from the coefficients of f , and the last k rows are built from the coefficients of g,
padded with zeros.
The most widely applied fact about the resultant is that it is zero if and only if
the two polynomials have a common complex root, or, equivalently, a nonconstant
common divisor in C[x]. This holds true even if the coefficients are arbitrary complex
numbers. In our case, however, the coefficients are integers. In this setting, the resultant
is zero if and only if the two polynomials have a nonconstant common divisor in Z[x].
We start with two easy observations relating the resultant r to the gcd of the poly-
nomial values.
Note that r can be zero. By definition, any function is periodic with period 0.
Proof. (a) Let d = gcd( f (n), g(n)). Each coordinate of the column vector
is divisible by either f (n) or g(n), and therefore by d. Thus, the last column of M is
congruent modulo d to a linear combination, with integer coefficients, of the previous
columns. It follows that r = det M ≡ 0 mod d, as claimed.
(b) We have f (n + r ) ≡ f (n) and g(n + r ) ≡ g(n) mod r . It follows that
In view of statement (a), the third argument can be omitted from the gcd on both sides,
proving statement (b).
This example also shows that when r = 0, |r | need not be the smallest positive
period of gcd( f (n), g(n)). In Example 4, we have r = 4, but the smallest positive
period is 2.
Our main result, Theorem 6 below, says that when r is square-free, Proposition 2(a)
is the only restriction on the values attained by the gcd, and the smallest positive period
of the gcd is |r |.
For this, we shall need a basic fact about integer matrices: they can be brought
to Smith normal form. For any matrix M with integer entries, there exist matrices U
and V , also with integer entries and invertible over Z, such that U M V is a diagonal
matrix with diagonal entries d1 , d2 , . . . , where the so-called invariant factors di satisfy
di |di+1 for all i. See Smith’s original paper [7], or see, e.g., [1, Section 5.3] for a
textbook presentation. Note that U and V , being invertible over Z, are necessarily
square matrices with determinant ±1. If M is also square, it follows that
In the proof of our main result, we shall have to leave the realm of polynomials with
integer coefficients and consider polynomials over the field F p of prime cardinality p.
Given two polynomials f and g over any field F, of degree k and l, respectively, with
coefficients as in (1) and (2), their Sylvester matrix M is defined by the formula (3).
We shall need
Theorem 5. [4, Theorem 1.19] The corank (or kernel dimension) k + l − rank M of
M over F equals the degree of the gcd of the two polynomials f and g as elements of
the polynomial ring F[x].
For two proofs of this well-known fact, the reader may consult [4]. As this is an
Internet reference, and we were unable to find a textbook or journal reference, we
include a third proof.
Proof. Let us identify the vector space F k+l with the vector space of polynomials of
degree less than k + l. Let any such polynomial correspond to the list of its coefficients,
starting with the coefficient of x k+l−1 and ending with the constant term.
Under this correspondence, the row space of the Sylvester matrix M is identified
with the set of polynomials of the form φ f + ψg, where φ, ψ ∈ F[x] have degree
less than l and k, respectively. Any polynomial of this form is divisible by gcd( f, g).
Conversely, any polynomial that is divisible by gcd( f, g) and has degree less than k + l
is in the row space. To see this, we first write such a polynomial as φ0 f + ψ0 g, where
we know nothing about the degree of φ0 , ψ0 ∈ F[x], but then we write φ0 = qg + φ
with φ of degree less than l, and we define ψ = q f + ψ0 . Then φ0 f + ψ0 g = φ f +
ψg; moreover, this polynomial and φ f both have degree less than k + l, whence so
does ψg, showing that ψ has degree less than k.
The rank of M is the dimension of the row space. The theorem follows.
Theorem 6. Let f and g be monic polynomials with integer coefficients. Assume that
their resultant r is square-free. Then all positive divisors of r arise as gcd( f (n), g(n))
448
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
for suitable integer n. Moreover, any d|r arises exactly ( p − 1) times in each period
of length |r |, where the product is taken over all (positive) prime divisors p of r/d. In
particular, |r | itself arises once.
r =± p.
p∈P
We shall prove that for all subsets S of P , the product d = p∈S p arises as
gcd( f (n), g(n))
for a suitable integer n; moreover, in each period of length |r |, it
arises exactly p∈P−S ( p − 1) times.
For each p ∈ P , the gcd( f (n), g(n), p) is periodic with period p. It suffices to
prove that in each period of length p, this gcd is p exactly once. Indeed, the Chinese
remainder theorem will then finish the proof: in each period of length |r |, the integers
n such that gcd( f (n), g(n)) = d can be found by specifying their value mod p for
each p ∈ P . For each p ∈ S , there is a unique possibility for n mod p, and for each
p ∈ P − S , there are p − 1 possibilities.
It suffices to prove that for any prime p ∈ P , the polynomials f and g, when viewed
mod p, have a unique common root in F p ; equivalently, the gcd of f and g as elements
of F p [x] has a unique root in F p . In fact, we shall prove that this gcd is a polynomial
of degree exactly 1.
It suffices to prove that the mod p corank of the Sylvester matrix M of f and g is
1. But the determinant of M over Z is r , which is divisible by p but not by p 2 . Now M
can be brought to Smith normal form, and from (4), we see that the last invariant factor
dk+l is divisible by p, but the previous one is not. The mod p corank of the diagonal
matrix U M V , and therefore also of M, is 1, as claimed.
Remark 7. When |r | is prime, the gcd is |r | for n in a unique residue class mod r
and is 1 for all other n. This sheds some light on the seemingly peculiar behavior in
Example 1, since r = 13 for (a) and r = p for (b).
When r is not square-free, we know very little about the range of the gcd. At least
we can give a sufficient condition for 1 to appear in the range. This condition, however,
is not necessary; see Example 4.
Proposition 8. Let f and g be monic polynomials with integer coefficients and resul-
tant r .
(a) Suppose that p is prime and r is not divisible by p p . Then there exists an integer
n such that gcd( f (n), g(n)) is not divisible by p.
(b) If r has no divisor of the form p p with p prime, then there exists an integer n
such that f (n) and g(n) are coprime.
Note that Proposition 8(a) is a special case of [5, Theorem 1], which is, in turn, a
consequence of [2, Theorem]. Nevertheless, we give an independent proof.
Proof. (a) Again we exploit the fact that r = ±d1 · · · dk+l , where the di are the invari-
ant factors of the Sylvester matrix M. Since di |di+1 for all i, and p p |r , it follows that
at most the last p − 1 invariant factors di can be divisible by p. In other words, the
mod p corank of M is less than p, so the degree of the gcd of f and g as elements
of F p [x] is less than p, and therefore this gcd cannot vanish as a function F p → F p .
Remark 9. Throughout this paper, we have studied two monic polynomials over the
ring Z of integers. However, Z can be replaced by an arbitrary principal ideal domain
A. Our results and their proofs remain valid, with trivial modifications.
For example, Proposition 2(b) should be interpreted as saying that ( f (n), g(n)) =
( f (n ), g(n )) whenever n, n ∈ A and r |(n − n ) in A. Note that this is an equality of
ideals of A.
In this general setting, the conclusion of Theorem 6 is replaced by the following.
There exist constants c P ∈ A, one for each prime ideal P containing r , such that for
any divisor d of r , and any n ∈ A, we have ( f (n), g(n)) = (d) if and only if n − c P ∈
P for each P containing d but n − c P ∈ P for each P that does not contain d. Such
elements n exist for any divisor d of r . When d = r , they form a coset c + (r ).
The p p in Proposition 8 should be interpreted as p|A/( p)| . This can be p∞ , which,
by definition, divides only 0.
ACKNOWLEDGMENTS. We are grateful to the Editorial Board of the M ONTHLY and to the two unnamed
referees of this paper for many useful comments. Thanks to Dmitry I. Khomovsky for calling our attention to
the references [2, 5].
Research of the first author is partially supported by ERC Consolidator Grant 648017, by MTA Rényi
Lendület Groups and Graphs research group, and by the Hungarian National Research, Development and
Innovation Office—NKFIH, OTKA grants no. K109684 and K104206.
REFERENCES
1. W. A. Adkins, S. H. Weintraub, Algebra: An Approach via Module Theory, Springer, Berlin, 1992.
2. D. Gomez, J. Gutierrez, Á. Ibeas, D. Sevilla, Common factors of resultants modulo p, Bull. Aust. Math.
Soc. 79 (2009) 299–302.
3. R. K. Guy, The strong law of small numbers, Amer. Math. Monthly 95 no. 8 (Oct 1988) 697–712.
4. S. Janson, Resultant and discriminant of polynomials,
http://www2.math.uu.se/~svante/papers/sjN5.pdf.
5. D. I. Khomovsky, On the relationship between the number of solutions of congruence systems and the
resultant of two polynomials, INTEGERS—Electronic Journal of Combinatorial Number Theory 16 A41.
6. The Prime Glossary, http://primes.utm.edu/glossary/page.php?sort=LawOfSmall.
7. H. J. S. Smith, On systems of linear indeterminate equations and congruences, Philos. Trans. R. Soc.
London 151 no. 1 293–326. Reprinted in The Collected Mathematical Papers of Henry John Stephen
Smith I. Ed. J. W. L. Glaisher. Clarendon Press, Oxford, 1894. 367–409.
Eötvös University, Department of Algebra and Number Theory, Pázmány Péter sétány 1/c, H-1117 Budapest,
Hungary
Rényi Institute of Mathematics, Hungarian Academy of Sciences, 13-15 Reáltanoda utca, H-1053 Budapest,
Hungary
frenkelp265@gmail.com
Eötvös University, Department of Algebra and Number Theory, Pázmány Péter sétány 1/c, H-1117 Budapest,
Hungary
pelikan@cs.elte.hu
450
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
The Stern Diatomic Sequence via
Generalized Chebyshev Polynomials
Valerio De Angelis
Abstract. Let a(n) be the Stern diatomic sequence, and let x1 , . . . , xr be the distances between
successive 1’s in the binary expansion of the (odd) positive integer n. We show that a(n) is
obtained by evaluating generalized Chebyshev polynomials when the variables are given the
values x1 + 1, . . . , xr + 1. We also derive a formula expressing the same polynomials in terms
of sets of increasing integers of alternating parity and derive a determinant representation
for a(n).
It has already appeared in the M ONTHLY (see [6], [1], and the nice survey article [7]).
Define polynomials qr (y1 , . . . , yr ) inductively by
q0 = 1, q1 (y1 ) = y1 ,
qr (y1 , . . . , yr ) = y1 qr −1 (y2 , . . . , yr ) − qr −2 (y3 , . . . , yr ) for r ≥ 2. (2)
The main result of this note is that a(n) coincides with these polynomials when the
variables are given the value of the gaps between successive 1’s in the binary expansion
of n, increased by 1 (Theorem 1), and a formula expressing the same polynomials in
terms of sets of increasing integers of alternating parity (Theorem 2).
The polynomials qr have appeared before in connection with the eigenvalue
problem for certain Jacobi matrices [9] and with cluster algebras, an algebraic-
combinatorial construction that was conceived by Fomin and Zelevinsky in 2000
[5]. Cluster algebras have been the focus of much research in many different areas
of mathematics in recent years. We briefly outline the latter connection and refer the
reader to the survey article [10] for the notions mentioned here.
In [4, Lemma 3.2], qr are introduced as generalized Chebyshev polynomials satis-
fying the relation
This identity corresponds to the exchange relation used to define a new cluster of
variables in the definition of a cluster algebra, and in the same paper, the author proves
that a cluster algebra of Dynkin type Ar is isomorphic to
Z[y1 , . . . , yr +1 ]/(qr +1 (y1 , . . . , yr +1 ) − 1).
The case r = 3 coincides with a recent expression obtained by Defant [3, equa-
tion (29)]. Using the polynomial representation, we give simple proofs for a number
of known identities for a(n), including a convolution identity found by Coons [2]
(Corollary 3), and we derive a result on the divisibility of a(n) (Corollary 4).
http://dx.doi.org/10.4169/amer.math.monthly.124.5.451
MSC: Primary 11B83
The expressions for a([c1 , . . . , cr )] for the first few values of r are recorded below
(the case r = 3 was recently derived by Defant in [3, equation (29)]):
a([c1 ]) = c1 + 1, a([c1 , c2 ]) = c1 + c1 c2 + c2 ,
a([c1 , c2 , c3 ]) = c1 c2 c3 + c1 c2 + c1 c3 + c2 c3 + c2 − 1,
a([c1 , c2 , c3 , c4 ]) = c1 c2 c3 c4 + c1 c2 c3 + c2 c3 c4 + c1 c3 c4
+ c1 c2 c4 + c2 c4 + c1 c3 + c2 c3 − c1 − c4 − 1.
The following corollary of the previous theorem corresponds to Corollary 3.3 of [4].
452
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
⎛ ⎞
y1 1 0 0 0
⎜1 y2 1 0 0⎟
⎜ .. .. .. ⎟
⎜ . . . ⎟
Mr (y1 , . . . , yr ) = ⎜ 0 0 ⎟ r ≥ 1.
⎜ .. .. ⎟
⎝0 0 . . 1⎠
0 0 0 1 yr
Let ci : 1 ≤ i ≤ r be positive integers, and let Ir be the r × r identity matrix. Then
a([c1 , . . . , cr ]) = det(Ir + Mr (c1 , . . . , cr )).
Proof. It is easy to check that det(Mr (y1 , . . . , yr )) satisfies the same recurrence rela-
tion as qr (y1 , . . . , yr ), with the same initial conditions. So det(Mr (y1 , . . . , yr )) =
qr (y1 , . . . , yr ).
Theorem 2. If r ≥ 2, then
r
qr (y1 , y2 , . . . , yr ) = ωr,s yu . (11)
s=0 u∈Ar,s
Proof. We will show that the right side of (11) satisfies the recurrence (2). If r ≥ 1
and 1 ≤ s ≤ r , then let
Br,s = {(i 1 , i 2 , . . . , i s ) ∈ Ar,s : i 1 = 1}, Cr,s = Ar,s \ Br,s .
There are bijections φ : Br,s → Ar −1,s−1 and ψ : Cr,s → Ar −2,s given by
φ ((1, i 2 , . . . , i s )) = (i 2 − 1, i 3 − 1, . . . , i s − 1) and ψ ((i 1 , i 2 , . . . , i s )) = (i 1 − 2, . . . ,
i s − 2). If z i = yi+1 , 1 ≤ i ≤ r − 1 and wi = yi+2 , 1 ≤ i ≤ r − 2, then yu = y1 z φ(u)
for u ∈ Br,s and yu = wψ(u) for u ∈ Cr,s . The result then easily follows by splitting
the sum over Ar,s as u∈Br,s + u∈Cr,s and using the fact that ωr,s+1 = ωr −1,s , and
ωr,s = −ωr −2,s .
Proof. It is enough to consider the case n odd. Note that if n = [c1 , . . . , cr ], then n =
[cr , . . . , c1 ]. So the result will follow if we show that qr (y1 , . . . , yr ) = qr (yr , . . . , y1 ).
This follows easily from either Corollary 1, by a permutation of the rows and columns
that reverses the main diagonal of the matrix Mr , or from Theorem 2. To see the
latter, notice that there is an involution βr,s on the sets Ar,s given by (i 1 , . . . , i s ) →
(i 1 , . . . , i s ), where i j = r − i s− j+1 + 1, because if r and s have the same parity, then
r − i s− j+1 + 1 ≡ r − s + j ≡ j (mod 2), while if r
≡ s (mod 2), then ωr,s = 0.
Proposition 1. If r ≥ 0, k ≥ 0, then
qk+r (t1 , . . . , tk , y1 , . . . , yr ) = qk (t1 , . . . , tk )qr (y1 , . . . , yr )
− qk−1 (t1 , . . . , tk−1 )qr −1 (y2 , . . . , qr ).
Proof. The proposition is proved by induction on k by writing qr +(k+1) = q(r +1)+k and
making use of Corollary 2.
Proof. The result holds for c = 0 trivially and for c = 1 or u = 0 by using the basic
identities for the Stern sequence. Decreasing e if necessary, we may assume that c
is odd and c ≥ 3. So there is some k ≥ 2 and integers c1 , . . . , ck−1 such that c =
[c1 , . . . , ck−1 ]. Since c ≤ 2e , we can define the positive integer ck = e − (c1 + · · · +
ck−1 ), and then [c1 , . . . , ck ] = c + 2e . Write u + 2 = [u 1 , . . . , u r ] for some positive
integers u 1 , . . . , u r . Then [c1 , . . . , ck , u 1 , . . . , u r ] = c + 2e (u + 2), and it is easily
checked that qr −1 (u 2 + 1, . . . , u r + 1) = a(u + 1). Proposition 1 (with yi = u i +
1, ti = ci + 1 ) gives us the identity
a(c + 2e (u + 2)) + a(c)a(u + 2) = a(c + 2e )a(u + 2) + a(c)a(u + 3).
Use a(2u + 3) = a(u + 2) + a(u + 1), a(2u + 5) = a(u + 2) + a(u + 3), the basic
identity a(2e + c) = a(2e − c) + a(c) (see[7]), and the identity a(2e − c)a(u + 1) +
a(c)a(u + 2) = a(2e (u + 1) + c) (easily proved by induction on e) to get the result.
The following result is an easy consequence of the recurrence (2) satisfied by the
polynomials qr . Recall that s(n) is the number of 1’s appearing in the binary expansion
of n.
Corollary 4. Suppose k is a positive integer that divides the exponent of each power
of 2 appearing in the binary expansion of n. Then:
⎧
⎪
⎨0 (mod k) if s(n) ≡ 0 or 3 (mod 6)
a(n) ≡ 1 (mod k) if s(n) ≡ 1 or 2 (mod 6)
⎪
⎩−1 (mod k) if s(n) ≡ 4 or 5 (mod 6).
454
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Proof. We may assume n is odd. Let m = s(n) − 1. By assumption, n = [kc1 , . . . , kcm ],
and by Theorem 1, a(n) = qm (kc1 + 1, . . . , kcm + 1) ≡ qm (1, . . . , 1) (mod k). If
bm = qm (1, . . . , 1), then bm satisfies the recurrence bm = bm−1 − bm−2 with ini-
√ b0 = b1 = 1. This recurrence is easily solved as bm = cos(mπ/3) +
tial conditions
sin(mπ/3)/ 3, and the result follows.
√ √
If λ1 = (1 + 5)/2 is the golden mean, and λ2 = (1 − 5)/2 is its algebraic
conjugate, then the classical Binet formula for the Fibonacci numbers Fn is Fn =
(λn1 − λn2 )/(λ1 − λ2 ), n ≥ 0.
We conclude with the same type of formula for some special values of a(n), easily
obtained by letting all variables of qr equal a single variable t.
Proof. Note that (2r t − 1)/(2t − 1) = [t, . . . , t], where there are r − 1 entries. So
br = a((2r t − 1)/(2t − 1)) = qr −1 (t + 1, . . . , t + 1) satisfies the recurrence br = (t +
1)br −1 − br −2 with initial conditions b1 = 1, b2 = t + 1. Solving this recurrence we
obtain the result.
ACKNOWLEDGMENTS. I thank Sam Northshield for pointing out the article [4] (that led to the current
title of this note) and for several helpful comments on an earlier version of the paper, Christophe Vignat for
pointing out the article [2], and the referee for useful suggestions.
REFERENCES
1. N. Calkin, H. S. Wilf, Recounting the rationals, Amer. Math. Monthly 107 (2000) 360–363, http://dx.
doi.org/10.2307/2589182.
2. M. Coons, A correlation identity for Stern’s sequence, Integers 12 (2012) 1–5.
3. C. Defant, Upper bounds for Stern’s diatomic sequence and related sequences,
http://arxiv.org/abs/1506.07824.
4. G. Dupont, Cluster multiplication in regular components via generalized Chebyshev polynomials, Algebr.
Represent. Theory 15 no. 3 (2012) 527–549, http://dx.doi.org/10.1007/s10468-010-9248-0.
5. S. Fomin, A. Zelevinsky, Cluster algebras I: Foundations, J. Amer. Math. Soc. 15 no. 2 (2002) 497–529,
http://dx.doi.org/10.1090/S0894-0347-01-00385-X.
6. D. H. Lehmer, On Stern’s diatomic series, Amer. Math. Monthly 36 (1929) 59–67,
http://dx.doi.org/10.2307/2299356.
7. S. Northshield, Stern’s diatomic sequence 0, 1, 1, 2, 1, 3, 2, 3, 1, 4, . . ., Amer. Math. Monthly 117 (2010)
581–598.
8. B. Reznick, Regularity properties of the Stern enumeration of the rationals, J. Integer Seq. 11 (2008).
9. F. Štampach, P. Šťtovı́ček, On the eigenvalue problem for a particular class of finite Jacobi matrices,
Linear Algebra Appl. 434 (2011) 1336–1353, http://dx.doi.org/10.1016/j.laa.2010.11.010.
10. A. Zelevinsky, What is a cluster algebra?, Notices Amer. Math. Soc. 54 no. 11 (2007) 1494–1495.
Mathematics Department, Xavier University of Louisiana, 1 Drexel Drive, New Orleans, LA 70125.
vdeangel@xula.edu
Abstract. Let R1 be a commutative ring, let R2 be a finitely generated extension ring of R1 , and
let S be a ring that is intermediate between R1 and R2 . For R1 = R[x] and R2 = R[x, y], there
are simple combinatorial constructions of intermediate rings S that are not finitely generated
over R[x].
R1 ⊆ S ⊆ R2 .
Artin and Tate [1] proved that if R1 is Noetherian and if R2 is finitely generated as a
module over S, then S is finitely generated as a ring over R1 . They used this to prove
Hilbert’s Nullstellensatz (cf. Zariski [3], Kunz [2, Lemma 3.3]).
It is natural to ask: If R2 is finitely generated as a ring over R1 and if the ring S is
intermediate between R1 and R2 , then is S finitely generated as a ring over R1 ? The
answer is “no,” and the purpose of this note is to give simple combinatorial construc-
tions of intermediate rings S that are not finitely generated over R1 .
Let N denote the set of positive integers and N0 the set of nonnegative integers.
and
b
< λ for all (a, b) ∈ . (2)
a
Consider the set of monomials
M() = x a y b : (a, b) ∈ .
http://dx.doi.org/10.4169/amer.math.monthly.124.5.456
MSC: Primary 13E15
456
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Let R be a commutative ring, and let R[M()] be the subring of R[x, y] generated by
M(). Then
For example, the set 1 = {(1, n) : n ∈ N0 } satisfies conditions (1) and (2) with
λ = ∞. The corresponding set of monomials is
M(1 ) = x, x y, x y 2 , x y 3 , . . . ,
R[M(1 )] = R[x, x y, x y 2 , x y 3 , . . .]
Proof. Because (1, 0) ∈ , we have x ∈ M() and R[x] ⊆ R[M()] ⊆ R[x, y].
Let F be a finite subset of R[M()]. For every polynomial f in F , there is a finite
set M ∗ ( f ) of monomials in M() such that f is a linear combination of products of
monomials in M ∗ ( f ). This set of monomials is not necessarily unique (for example,
(x y)(x y 4 ) = (x y 2 )(x y 3 ) in R[M(1 )]), but we choose, for each polynomial f in F ,
one set M ∗ ( f ) of monomials in M() that generates f . Because F is a finite set of
polynomials, the set
M ∗ (F ) = M ∗( f )
f ∈F
Condition (1) implies that the ring R[M()] contains monomials x A y B with
/ R[M ∗ (F )] and so R[M ∗ (F )] = R[M()].
β < B/A < λ. It follows that x A y B ∈
Therefore, R[F ] = R[M()], and the ring R[M()] is not finitely generated.
This completes the proof.
ACKNOWLEDGMENT. I thank Ryan Alweiss for very helpful discussions on this topic at CANT 2016.
REFERENCES
1. E. Artin, J. T. Tate, A note on finite ring extensions, J. Math. Soc. Japan 3 (1951) 74–77.
2. E. Kunz, Introduction to Commutative Algebra and Algebraic Geometry. Birkhäuser/Springer, New York,
2013.
3. O. Zariski, A new proof of Hilbert’s Nullstellensatz, Bull. Amer. Math. Soc. 53 (1947) 362–368.
458
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
A Geometric Proof of the Siebeck–Marden
Theorem
Beniamin Bogosel
Abstract. The Siebeck–Marden theorem relates the roots of a third degree polynomial and
the roots of its derivative in a geometrical way. A few geometric arguments imply that every
inellipse for a triangle is uniquely related to a certain logarithmic potential via its focal points.
This fact provides a new direct proof of a general form of the result of Siebeck and Marden.
x1 c1
t2 f2
t1 g2 f1
g1
f1 f2 b c
a1
x2
Figure 1. Left: Basic property of the tangents to an ellipse. Right: Construction of an inellipse starting from
two isogonal conjugate points.
We give precise answers to all these questions in the next section, dedicated to the
geometric properties of inellipses. Once these properties are established, we are able
to prove a more general version of the Siebeck–Marden theorem. The proof of the
original Siebeck–Marden result will follow immediately from the two main geometric
properties of the critical points f 1 , f 2 .
• The midpoint of f 1 f 2 is the centroid of abc.
• The points f 1 , f 2 are isogonal conjugates relative to triangle abc.
We recall that two points f 1 , f 2 are isogonal conjugates relative to triangle abc if
the pairs of lines (a f 1 , a f 2 ), (b f 1 , b f 2 ), (c f 1 , c f 2 ) are symmetric with respect to the
bisectors of the angles a, b, c, respectively.
460
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Proof. The proof of 1. was discussed above, so it only remains to prove 2. Consider
the points x1 , x2 , the reflections of f 1 , f 2 with respect to the lines ab and bc (see Figure
1 right). The construction implies that b f 1 = bx1 , b f 2 = bx2 and ∠x1 b f 2 = ∠ f 1 bx2 ,
which, in turn, implies that x1 f 2 = f 1 x2 . We denote their common value with m. We
denote a1 = f 1 x2 ∩ bc and c1 = x1 f 2 ∩ ab. The construction of x1 , x2 implies that
f 1 a1 + f 2 a1 = f 1 x2 = f 2 x1 = f 1 c1 + f 2 c1 = m. Heron’s problem cited above implies
that a1 is the point that minimizes x → f 1 x + f 2 x with x ∈ bc and c1 is the point that
minimizes x → f 1 x + f 2 x with x ∈ ab.
Thus, the ellipse characterized by f 1 x + f 2 x = m is tangent to bc and ab in a1 and,
respectively, c1 . A similar argument proves that this ellipse is, in fact, also tangent to
ac. The unicity of this ellipse comes from the fact that m is defined as the minimum
of f 1 x + f 2 x where x is on one of the sides of abc, and this minimum is unique and
independent of the chosen side.
We are left to answer questions 1, 2, and 4. The first two questions were answered
by Chakerian in [4] using an argument based on orthogonal projection. We provide
a slightly different argument, which, in addition, gives us information about the rela-
tion between the barycentric coordinates of the center of the inellipse and its tangency
points. In the proof of the following results we use the properties of real affine trans-
formations of the plane.
Proof. 1. We begin with the particular case where the inellipse E is the incircle with
center o. Suppose E is another inellipse, with center o, and denote by f 1 , f 2 its focal
points. We know that f 1 , f 2 are isogonal conjugates relative to abc and the midpoint
of f 1 f 2 is o, the center of the inellipse. Thus, if f 1 = f 2 , then ao is at the same time a
median and a bisector in triangle a f 1 f 2 . This implies that ao ⊥ f 1 f 2 . A similar argu-
ment proves that bo ⊥ f 1 f 2 and co ⊥ f 1 f 2 . Thus a, b, c all lie on a line perpendicular
to f 1 f 2 in o, which contradicts the fact that abc is nondegenerate. The assumption
f 1 = f 2 leads to a contradiction, and therefore we must have f 1 = f 2 , which means
that E is a circle and E = E .
Consider now the general case. Suppose that the inellipses E , E for abc have
the same center. Consider an affine mapping h that maps E to a circle. Since h maps
ellipses to ellipses and preserves midpoints, the image of our configuration by h is
a triangle where h(E ) is the incircle and h(E ) is an inscribed ellipse with the same
center. This case was treated in the previous paragraph and we must have h(E ) =
h(E ). Thus E = E .
2. To find the locus of the centers of inellipses for abc, it is enough to see
which barycentric coordinates are admissible for the incircle of a general triangle.
We recall that the barycentric coordinates of a point p are proportional to the areas of
the triangles pbc, pca, pab, and their sum is chosen to be 1. Thus, barycentric
coordinates are preserved under affine transformations. The barycentric coordinates
of the center of an inellipse with respect to a, b, c are the same as the barycentric
1 The medial triangle is the triangle formed by the midpoints of the edges of a triangle.
where u, v, w are the lengths of the sides of a b c . Thus, we can see that x + y + z
= 1 and x < y + z, y < z + x, z < x + y. One simple consequence of these relations
is the fact that x, y, z < 1/2. Furthermore, since
we can see that the previous relations for x, y, z are satisfied if and only if o is in the
interior of the medial triangle for a b c . Thus, the locus of the center of an inscribed
ellipse is the interior of the medial triangle.
3. If the center of the inellipse E is αa + βb + γ c with α + β + γ = 1, then con-
sider an affine map h that transforms E into a circle. Let a b c be the image of abc
by h. It is known that α, β, γ are proportional with the sidelengths of the triangle
a b c . Thus, the tangency points of h(E ) with respect to a b c divide its sides into
ratios
α+γ −β α+β −γ β +γ −α
, , .
α+β −γ β +γ −α α+γ −β
The affine map h does not modify the ratios of collinear segments, thus, E divides the
sides of abc into the same ratios.
α β γ
L (z) = + + ,
z−a z−b z−c
462
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
which means that f 1 , f 2 are roots of
Without loss of generality, we can suppose that a = 0 and that the imaginary axis is the
bisector of the angle ∠bac (equivalently bc < 0). In this case we have f 1 f 2 = αbc <
0, and thus the imaginary axis is the bisector of the angle ∠ f 1 a f 2 . Repeating the same
argument for b and c, we deduce that f 1 , f 2 are isogonal conjugates relative to abc.
Steiner’s result (Theorem 1) implies that f 1 , f 2 are the foci of an inellipse E for abc.
The center of this inellipse has barycentric coordinates
β +γ α+γ α+β
o= a+ b+ c,
2 2 2
which, according to Theorem 2, implies that E is the unique inellipse for abc, which
divides the sides of abc in ratios β/γ , γ /α, α/β.
Conversely, given an inellipse E for abc, its tangency points must be of the form
β/γ , γ /α, α/β for some α, β, γ > 0, α + β + γ = 1. We choose L(z) = α log(z −
a) + β log(z − b) + γ log(z − c) and, according to the first part of the proof, the criti-
cal points f 1 , f 2 of L (z) are the foci of an ellipse E that divides the sides of abc into
ratios β/γ , γ /α, α/β. This means that E = E and L(z) is the associated logarithmic
potential.
ACKNOWLEDGMENT. The author wishes to thank the anonymous reviewer for suggestions that helped
improve the quality of this paper.
REFERENCES
1. R. E. Allardice, Note on the dual of a focal property of the inscribed ellipse, Ann. of Math. 2 (1900)
148–150.
2. H. Aref, M. Brøns, On stagnation points and streamline topology in vortex flows, J. Fluid Mech. 370
(1998) 1–27.
3. E. Badertscher, A simple direct proof of Marden’s theorem, Amer. Math. Monthly 121 (2014) 547–548.
4. G. Chakerian, A distorted view of geometry, in Mathematical Plums, Ed. R. Honsberger. The Dolciani
Mathematical Expositions, Vol. 4. Mathematical Association of America, Washington, D.C., 1979.
5. V. Dragović, M. Radnović, Poncelet Porisms and Beyond: Integrable Billiards, Hyperelliptic Jacobians
and Pencils of Quadrics. Frontiers in Mathematics, Birkhäuser/Springer Basel AG, Basel, 2011.
6. M. J. Kaiser, The dynamics and internal geometry of the three-city noxious location problem, Math.
Comput. Modelling 17 (1003) 81–98.
7. D. Kalman, An elementary proof of Marden’s theorem, Amer. Math. Monthly 115 (2008) 330–338.
8. M. Marden, A note on the zeros of the sections of a partial fraction, Bull. Amer. Math. Soc. 51 (1945)
935–940.
9. D. Minda, S. Phelps, Triangles, ellipses, and cubic polynomials, Amer. Math. Monthly 115 (2008)
679–689.
10. S. Northshield, Geometry of cubic polynomials, Math. Mag. 86 (2013) 136–143.
11. J. L. Parish, On the derivative of a vertex polynomials, Forum Geom. 6 (2006) 285–288.
12. J. Siebeck, Ueber eine neue analytische behandlungweise der brennpunkte, J. Reine Angew. Math. 64
(1864) 175–182.
13. J. Steiner, Géométrie pure. Développment d’une séie de théorèmes relatifs aux sections coniques, Ann.
Math. Pures Appl. 19 (1828/1829) 37–64.
Let x0 := sup B. By (i), there exists δ > 0 such that Iδ := [x0 − δ, bδ ] ⊆ A z 0 , for
some z 0 , where bδ := x0 + δ if x0 < b; and bδ := b if x0 = b. By the choice of
x0 , there exists x1 ∈ (x0 − δ, x0 ] such that [a, x1 ) ⊆ A z 1 , for some z 1 . Thus, by
(ii), [a, bδ ] ⊆ A z 1 ∪ A z 0 = A z 2 , where z 2 ∈ {z 0 , z 1 } such that f (z 2 ) = max{ f (z 0 ),
f (z 1 )}, and so bδ ∈ B. This implies that bδ ≤ x0 , thus x0 = b, and so bδ = b. But
then z 2 ∈ [a, b] = A z 2 , contradicting the fact that z 2 ∈ / Az2 .
REFERENCE
1. R. G. Bartle, D. R. Sherbert, Introduction to Real Analysis. Third edition. John Wiley & Sons, New
York, 2000.
464
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
PROBLEMS AND SOLUTIONS
Edited by Gerald A. Edgar, Daniel H. Ullman, Douglas B. West
with the collaboration of Paul Bracken, Ezra A. Brown, Zachary Franco, Christian Friesen,
László Lipták, Rick Luttmann, Frank B. Miles, Lenhard Ng, Leonard Smiley, Kenneth
Stolarsky, Richard Stong, Walter Stromquist, Daniel Velleman, and Fuzhen Zhang.
PROBLEMS
11978. Proposed by Hideyuki Ohtsuka, Saitama, Japan. Let Fn be the nth Fibonacci num-
ber, with F0 = 0, F1 = 1, and Fn = Fn−1 + Fn−2 when n ≥ 2. Find
∞
(−1)n
.
n=0
cosh Fn cosh Fn+3
11979. Proposed by Zachary Franco, Houston, Texas. Let O and I denote the circumcenter
and incenter of a triangle ABC. Are there infinitely many nonsimilar scalene triangles ABC
for which the lengths AB, BC, CA, and OI are all integers?
11980. Proposed by George Stoica, Saint John, NB, Canada. Let a1 , . . . , an be a nonin-
creasing list of positive real numbers, and fix an integer k with 1 ≤ k ≤ n. Prove that there
exists a partition {B1 , . . . , Bk } of {1, . . . , n} such that
1 1 n
min ai ≥ min ai .
1≤ j≤k
i∈B j
2 1≤ j≤k k + 1 − j i= j
11981. Proposed by Cezar Lupu, University of Pittsburgh, Pittsburgh, PA. Suppose that
f : [0, 1] → R is a differentiable function with continuous derivative and with
1 1
f (x) d x = x f (x) d x = 1.
0 0
Prove
2
1 3 128
f (x) d x ≥ .
0 3π
11982. Proposed by Ovidiu Furdui, Mircea Ivan, and Alina Sı̂ntămărian, Technical Uni-
versity of Cluj-Napoca, Cluj-Napoca, Romania. Calculate
∞
1/x
x n
lim .
x→∞
n=1
n
http://dx.doi.org/10.4169/amer.math.monthly.124.5.465
SOLUTIONS
Orthogonal Functions
11850 [2015, 605]. Proposed by Zafar Ahmed, Bhabha Atomic Research Centre, Mumbai,
India. Let
n
2 1 2 n/2 d 1
An (x) = (1 + x ) .
π n! dxn 1 + x2
∞
Prove that −∞ Am (x)An (x) d x = δ(m, n) for nonnegative integers m and n. Here δ(m, n) =
1 if m = n, and otherwise δ(m, n) = 0.
Solution by Ramya Dutta, Chennai Mathematical Institute, Chennai, India. We have
dn 1 1 dn 1 1 n n! 1 1
= − = (−1) − .
dxn 1 + x2 2i d x n x − i x +i 2i (x − i)n+1 (x + i)n+1
466
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Editorial comment. The problem as originally printed asserted that An (x) is a polynomial.
The editors are responsible for this error.
Also solved by T. Amdeberhan & H. Kilete-Seleste, T. Amdeberhan & S. B. Ekhad, R. Bagby, D. Beckwith,
G. E. Bilodeau, R. Boukharfane (France), P. Bracken, H. Chen, P. P. Dályay (Hungary), P. J. Fitzsimmons,
N. Grivaux (France), F. Holland (Ireland), O. Kouba (Syria), G. Kuldeep (India), O. P. Lossers (Netherlands),
R. Stong, R. Tauraso (Italy), J. Van Hamme (Belgium), M. Vowe (Switzerland), H. Widmer (Switzerland),
GCHQ Problem Solving Group (U. K.), and the proposer.
468
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
A Hyperbolic Sine Series
11853 [2015, 700]. Proposed by Hideyuki Ohtsuka, Saitama, Japan. Find
∞
1
.
n=1
sinh 2n
= coth x − 1,
for all x > 0. The requested sum is the special case where x = 1.
Solution II by Rituraj Nandan, SunEdison, St. Peters, MO.
∞
∞ ∞ n
∞ ∞
1 2 e−2 x n
= 2n x − e−2n x
= 2 = 2 e−2 (2 j+1)x
sinh(2n x) e 1 − e −2n+1 x
n=1 n=1 n=1 n=1 j=0
∞
2
=2 e−2kx = ,
k=1
e2x −1
where in the penultimate step we have used the fact that every even, positive integer 2k can
be written uniquely as 2k = 2n (2 j + 1) for n ≥ 1 and j ≥ 0.
Editorial comment. This sum appears as 1.121.2 in Gradshteyn and Ryzhik, Table of Inte-
grals, Series, and Products.
Also solved by U. Abel (Germany), Z. Ahmed (India), A. Ali (India), K. Andersen (Canada), M. Arake-
lian (Armenia), H. I. Arshagi, M. Bataille (France), D. Beckwith, M. Bello & M. Benito & Ó. Ciaurri &
E Fernández & L. Roncal (Spain), S. C. Bhoria (India), R. Boukharfane (France), P. Bracken, B. Bradie,
N. Caro (Brazil), R. Chapman (U. K.), H. Chen, S. Choi (Korea), C. Curtis, N. Curwen (U. K.), P. P. Dályay
(Hungary), B. E. Davis, R. Dutta (India), E. Errthum, D. Fleischman, J. Gaisser, O. Geupel (Germany),
H. B. Ghaffari (Iran), M. L. Glasser, M. Goldenberg & M. Kaplan, N. Grivaux (France), J. A. Grzesik, M. Hoff-
man, F. Holland (Ireland), T. Horine, B. Karaivanov (U. S. A.) & T. S. Vassilev (Canada), O. Kouba (Syria),
H. Kwong, P. Lalonde (Canada), W. C. Lang, K.-W. Lau (China), L. Lipták, O. P. Lossers (Netherlands),
J. Magliano, L. Matejı́čka (Slovakia), V. Mikayelyan (Armenia), J. Mooney, M. Omarjee (France), S. Pathak
(Canada), F. Perdomo & Á. Plaza (Spain), C. M. Russell, M. Sawhney, V. Schindler (Germany), N. C. Singer,
J. Sorel (Romania), A. Stenger, R. Stong, H. Takeda (Japan), R. Tauraso (Italy), C. I. Vălean (Romania), G. Vid-
iani (France), J. Vinuesa (Spain), T. Viteam (Japan), Z. Vörös (Hungary), M. Vowe (Switzerland), T. Wiandt,
H. Widmer (Switzerland), M. Wildon (U. K.), J. Zacharias, L. Zhou, FAU Problem Solving Group, GCHQ
Problem Solving Group (U. K.), GWstat Problem Solving Group, NSA Problems Group, Northwestern Uni-
versity Math Problem Solving Group, PHP Solving Team, and the proposer.
A Momentous Inequality
11855 [2015, 700]. Proposed by Cezar Lupu, University of Pittsburgh, Pittsburgh, PA.
1
For a continuous and nonnegative function f on [0, 1], let μn = 0 x n f (x) d x. Show that
μn+1 μ0 ≥ μn μ1 for n ∈ N.
Solution I by Ross Dempsey, student, Thomas Jefferson High School, Alexandria, VA. If
μn = 0 for some n, then f is identically zero. So we may assume μn > 0 for all n.
For n ≥ 1, consider the integral
1
μn 2 n−1
x− x f (x) d x.
0 μn−1
The integrand is nonnegative, so
1
μn 2 n−1
0≤ x− x f (x) d x
0 μn−1
1 1 1
μn μ2
= x n+1 f (x) d x − 2 x n f (x) d x + 2 n x n−1 f (x) d x
0 μn−1 0 μn−1 0
= μn+1 − μ2n /μn−1 .
It follows that μn+1 /μn ≥ μn /μn−1 . By induction, μn+1 /μn ≥ μ1 /μ0 , which is equivalent
to the required inequality.
Solution II by Oliver Geupel, Brühl, NRW, Germany. Let g : [0, 1] → R be defined by
x x x x
g(x) = t n+1 f (t) dt · f (t) dt − t n f (t) dt · t f (t) dt.
0 0 0 0
Therefore, g(x) is increasing on [0, 1], and since g(0) = 0, we have g(x) ≥ 0. This implies
that g(1) ≥ 0, or μn+1 μ0 − μn μ1 ≥ 0.
470
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Solution III by Ulrich Abel, Technische Hochschule Mittelhessen, Friedberg, Germany. We
prove more generally that μm μn ≤ μm+n μ0 . The required inequality is the case m = 1.
We have
1 1 1 1 m n
μm μn = (μm μn + μn μm ) = (x y + x n y m ) f (x) f (y) d xd y.
2 2 0 0
For 0 ≤ x, y ≤ 1, we have 0 ≤ (x m − y m )(x n − y n ) = (x m+n + y m+n ) − (x m y n + x n y m ),
or x m y n + x n y m ≤ x m+n + y m+n , and this implies
1 1 1 m+n 1
μm μn ≤ (x + y m+n ) f (x) f (y) d xd y = (μm+n μ0 + μ0 μm+n ).
2 0 0 2
Hence μm μn ≤ μm+n μ0 , as claimed.
Also solved by R. A. Agnew, A. Ali (India), T. Amdeberhan & A. Straub, K. F. Andersen (Canada),
M. Andreoli, H. I. Arshagi, R. Bagby, M. Bataille (France), M. Bello & M. Benito & Ó. Ciaurri &
E. Fernández & L. Roncal (Spain), P. Bracken, M. A. Carlton, R. Chapman (U. K.), H. Chen, L. V. P. Cuong
(Vietnam), C. Curtis, N. Curwen (U. K.), P. P. Dályay (Hungary), B. E. Davis, J. Duemmel, R. Dutta (India),
D. L. Farnsworth, P. J. Fitzsimmons, D. Fleischman, L. Giugiuc (Romania), N. Grivaux (France), J. A. Grzesik,
L. Han, E. A. Herman, F. Holland (Ireland), T. Horine, E. J. Ionaşcu, B. Karaivanov (U. S. A.) & T. S. Vassilev
(Canada), O. Kouba (Syria), P. T. Krasopoulos (Greece), J. H. Lindsey II, P. W. Lindstrom, O. P. Lossers
(Netherlands), L. Matejı́čka (Slovenia), V. Mikayelyan (Armenia), M. Omarjee (France), E. Omey (Belgium),
D. Ritter, M. Sawhney, K. Schilling, N. C. Singer, A. Stenger, R. Stong, R. Tauraso (Italy), N. Thornber, R. van
der Veer (Netherlands), E. I. Verriest, J. Vinuesa (Spain), J. Wakem, T. Wiandt, J. Zacharias, Z. Zhang (China),
L. Zhou, GCHQ Problem Solving Group (U. K.), NSA Problems Group, and the proposer.
p−1 |G| 1 1 2
≤ s p (G) + 2
< |G| + 2
< |G|,
p∈A
6 p
p 6 p
p 3
2
where we have used the fact that p 1p ≈ 0.452224742 < 1/2.
The argument shows that equality occurs only when G is the group S3 of order six.
Also solved by the proposer.
Solution by Borislav Karaivanov, Sigma Space, Lanham, MD, and Tzvetalin S. Vassilev,
Nipissing University, North Bay, ON, Canada. Write s for the semiperimeter of ABC. Using
the formulas
(s − a)(s − b)(s − c) s(s − b)(s − c)
r= , ra =
s (s − a)
and similar formulas for rb and rc , we derive
rra = (s − b)(s − c), rrb = (s − c)(s − a), rrc = (s − a)(s − b). (1)
Let denote the area of A B C . Using Heron’s formula, we find
√ √ √ √ √ √ √ √ √ √ √ √
16( )2 = a+ b+ c − a+ b+ c a− b+ c a+ b− c
√ √ 2 √ √ 2
= b+ c −a a− b− c
√ √
= 2 bc + (b + c − a) 2 bc − (b + c − a)
472
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Editorial comment. Sin Hitotumatu showed that for all triangles ABC, the triangle A B C
is acute.
Also solved by Z. Ahmed (India), A. Ali (India), A. Alt, M. Bataille (France), B. S. Burdick, M. V. Chan-
nakeshava (India), R. Chapman (U. K.), C. Curtis, N. Curwen (U. K.), P. P. Dályay (Hungary), P. De (India),
A. Fanchini (Italy), D. Fleischman, O. Geupel (Germany), M. Goldenberg & M. Kaplan, J.-P. Grivaux (France),
J. A. Grzesik, J. G. Heuver (Canada), S. Hitotumatu (Japan), O. Hughes, Y. J. Ionin, L. R. King, O. Kouba
(Syria), W.-K. Lai & J. Risher & W. D. Ethridge, K.-W. Lau (China), J. M. Lewis, J. H. Lindsey II, G. Lord,
O. P. Lossers (Netherlands), V. Mikayelyan (Armenia), J. Minkus, D. J. Moore, R. Nandan, P. Nüesch (Switzer-
land), C. G. Petalas (Greece), M. Sawhney, V. Schindler (Germany), M. A. Shayib, I. Sofair, N. Stanciu &
T. Zvonaru (Romania), R. Stong, W. Szpunar-Lojasiewicz, H. Takeda (Japan), R. Tauraso (Italy), T. Viteam
(Japan), Z. Vörös (Hungary), M. Vowe (Switzerland), T. Wiandt, L. Wimmer, J. Zacharias, L. Zhou, GCHW
Problem Solving Group (U. K.), and the proposer.
Avoiding Voids
11862 [2015, 802]. Proposed by David A. Cox and Uyen Thieu, Amherst College, Amherst,
MA. For positive integers n and k, evaluate
k
i k kn − in
(−1) .
i=0
i k+1
Solution I by Borislav Karaivanov, Sigma Space, Lanham, MD, and Tzvetalin S. Vassilev,
Nipissing University, North Bay, Ontario, Canada. The value is kn k−1 n2 .
Consider a deck of kn cards, with n distinct cards in each of k suits. Both the summation
and the value count the ways to pick k + 1 cards with at least one card from each suit. For
the value, we pick one of the k suits to contribute two cards and pick one card from each
of the other suits.
k k
i k kn − in i k
(−1) = (−1) [z k+1 ](1 + z)n(k−i)
i=0
i k + 1 i=0
i
k
(1 + z)n − 1)
= [z ]((1 + z) − 1) = [z]
k+1 n k
z
k
n n
= [z] n + z + ··· = kn k−1 .
2 2
Editorial comment. Extending Solution I, the FAU Problem Solving Group noted that
choosing k + 2 cards yields a similar formula:
k 2
k kn − in k n k n k−1
(−1)i = n k−2 + n .
i=0
i k + 2 2 2 1 3
474
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
REVIEWS
Edited by Jason Rosenhouse
Department of Mathematics and Statistics, James Madison University, Harrisonburg,
VA 22807
x = 1766319049, y = 2261539880.
This result was found by Bhaskara in India around 1150, and by Fermat in France in
1657.
If we introduce the “algebraic integers”
√
Q(a, b) = a + b m,
where a, b ∈ Z, and if (a, b) is any solution of Pell’s equation, then for all n, positive
or negative, the pair of integer coefficients arising from [Q(a, b)]n is also a solution!
http://dx.doi.org/10.4169/amer.math.monthly.124.5.475
476
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
geometrical limit of the properly scaled binomial curve. The exposition attains both
clarity and correctness, which are so often thought to be incompatible.
So much for the book’s contents. What about its philosophy? Just what is “elemen-
tary” anyway?
This is an interesting methodological and philosophical question. Stillwell pursues
it in philosophical comments at the end of several chapters.
In chapter one, on elementary topics, he writes,
The reason is that representative rings and fields, namely Z and Q, have been
familiar since ancient times. The situation is quite different with the group con-
cept. Before the concept was identified by Galois (and perhaps glimpsed by
Lagrange a generation earlier) the only familiar groups were quite atypical, with
a commutative group operation, such as Z under addition. The most important
groups, such as those arising from general polynomial equations, are not com-
mutative. So the group axioms had to omit a statement of commutativity, and
mathematicians had to get used to noncommutative multiplication.
REFERENCES
478
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
Solution to Mathematical Evolution
S A G A S A L P O T R U E
C I R C A C O R N R O S E
A R I T H M E T I C E W E R
M O D U L O M E T E
P U L P D A D E O L D A S
S T Y B E M E A L I E N S
E R I C L U M E N S
L I N E A R A L G E B R A
P E N U L T G O O N
A N O M I E O G R E D O S
D O Z E N A N S I M E R E
R E S T S P E C I E
A R E A C O H O M O L O G Y
N U L L O N E A M E D I A
T E D S T E S T P E E N S
http://dx.doi.org/10.4169/amer.math.monthly.124.5.479
We received the following from Lyle Ramshaw, concerning his M ONTHLY article
“Stråhle’s equal-tempered triumph” 123 (2016) 871–883.
I apologize for failing to find and, hence, failing to cite in my November 2016
article on Stråhle’s guitar-fret construction, the recent work of Andrew M. Rockett
and Joseph P. Ruggerio [1]. They show that Stråhle’s tilt ratio of 24/17 minimizes the
worst-case absolute error in the locations of the frets if we ignore the errors in pitch.
They show that 24/17 also makes the fret at their x = 7/12 quite accurate. (Since their
fret numbers vary inversely with pitch, though, that fret sounds, above the open string,
by the interval that musicians call a fourth, not the fifth that they claim.)
While Stråhle’s construction gives a temperament that is close to equal, the extent to
which he was striving for equality is unclear. Rockett and Ruggerio add helpful context
by citing Unnerbäck, who discusses two of Stråhle’s primary mentors [2, p. 134]: The
organ builder Johan Niclas Cahman was at least in the well-tempered camp, wanting
to play “with tolerable satisfaction” in all keys, while the scientist Christopher Polhem
went further by suggesting an equal-beating temperament.
REFERENCES
1. A. M. Rockett, J. P. Ruggerio, On Stråhle’s guitar frets, Math. Gaz. 95 (2011) 300–303, https://doi.
org/10.1017/S0025557200003077.
2. A. Unnerbäck, The Cahman tradition and its German roots, in The Organ as a Mirror of Its Time. Ed.
by K. J. Snyder. Oxford Univ. Press, Oxford, 2002. 126–136, https://books.google.com/books?
id=joY3CiRqGgUC&pg=P126.
http://dx.doi.org/10.4169/amer.math.monthly.124.5.480
480
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 124
New in the eBooks Store
To order go to www.maa.org/ebooks/CAM33.
Coming soon in hardcover to the MAA Store.
Catalog Code: CAM33 List: $63.00
ISBN: 978-0-88385-142-5 MAA Member: $47.25
MATHEMATICAL ASSOCIATION OF AMERICA
1529 Eighteenth St., NW O Washington, DC 20036