Professional Documents
Culture Documents
NOTES
An Unnoticed Consequence of Szegö’s Distribution 917
Theorem
William F. Trench
Coset Intersection Graphs for Groups 922
Jack Button, Maurice Chiodo, and Mariano Zeron-Medina Laris
Complex Descartes Circle Theorem 927
Sam Northshield
The Length of an Arithmetic Progression Represented 932
by a Binary Quadratic Form
Pallab Kanti Dey and R. Thangadurai
Cosines and Cayley, Triangles and Tetrahedra 937
Marshall Hampton
A Note on the Spectral Theorem in the Finite-Dimensional 942
Real Case
Felipe Acker
PROBLEMS AND SOLUTIONS 946
BOOK REVIEW
Mostly Surfaces 954
By Richard Evan Schwartz
Genevieve S. Walsh
EDITOR
Scott T. Chapman
Sam Houston State University
ASSOCIATE EDITORS
William Adkins Jeffrey Lawson
Louisiana State University Western Carolina University
David Aldous C. Dwight Lahr
University of California, Berkeley Dartmouth College
Elizabeth Allman Susan Loepp
University of Alaska, Fairbanks Williams College
Jonathan M. Borwein Irina Mitrea
University of Newcastle Temple University
Jason Boynton Bruce P. Palka
North Dakota State University National Science Foundation
Edward B. Burger Vadim Ponomarenko
Southwestern University San Diego State University
Minerva Cordero-Epperson Catherine A. Roberts
University of Texas, Arlington College of the Holy Cross
Allan Donsig Rachel Roberts
University of Nebraska, Lincoln Washington University, St. Louis
Michael Dorff Ivelisse M. Rubio
Brigham Young University Universidad de Puerto Rico, Rio Piedras
Daniela Ferrero Adriana Salerno
Texas State University Bates College
Luis David Garcia-Puente Edward Scheinerman
Sam Houston State University Johns Hopkins University
Sidney Graham Anne Shepler
Central Michigan University University of North Texas
Tara Holm Frank Sottile
Cornell University Texas A&M University
Roger A. Horn Susan G. Staples
University of Utah Texas Christian University
Lea Jenkins Daniel Ullman
Clemson University George Washington University
Daniel Krashen Daniel Velleman
University of Georgia Amherst College
Ulrich Krause
Universität Bremen
Abstract. The envelope of straight lines normal to a plane curve C is its evolute; the envelope
of lines tangent to C is the original curve, together with the entire tangent line at each inflexion
of C. We introduce some standard techniques of singularity theory and use them to explain
how the first of these envelopes turns into the second, as the (constant) angle between the set of
lines forming the envelope and the set of tangents to C changes from 12 π to 0. In particular, we
explain how cusps disappear and what happens at inflexions, where the evolute goes to infinity.
We also study the family of “wavefronts” or “parallels” associated with these envelopes.
Figure 1. Left: We consider the envelope τα of lines L such that the counterclockwise angle α between the
tangent T to σ and L is constant. Center and right: An ellipse σ (t) = (2 cos t, sin t) and the envelope with
α = 14 π = 0.785 . . . , clearly showing four cusps, and α = 0.5, where the cusps have almost disappeared and
the envelope is more closely approximating the ellipse itself. For clarity, the lines L are drawn only in the
“forward” direction at each point. See also Example 3.2.
http://dx.doi.org/10.4169/amer.math.monthly.121.10.871
MSC: Primary 53A04, Secondary 53A05; 57R45; 58K05
Figure 2. A curve with two inflexions—the curve appears as the kidney-shaped line on the left. Left and
middle: The envelope of normals, which has six cusps and two asymptotes, shown in full on the left. The
center diagram shows the normals themselves; for clarity they are drawn only in the direction in which they
“focus” on the envelope. Right: The envelope of tangents, this time drawn in both directions. Here the envelope
includes the original curve and the whole of the tangent lines at the inflexions.
The key ingredients of singularity theory on which we call are the theory of un-
foldings and discriminants and the theory of functions on discriminants. We cannot
(alas) present all the details of these theories here, but we hope that enough is said to
show how powerful abstract techniques yield highly concrete geometrical results. (For
details of most of the techniques, and other geometrical applications, see [6].)
The article is organized as follows. In Section 2 we firm up the definitions and
give an explicit formula for the envelope τα , recalling some basic facts about plane
parametrized curves. In Section 3 we study the cusps of τα , introducing the first ideas
from singularity theory and the classification of functions. In Section 4 we set the study
of envelopes in the general context of discriminants and functions on discriminants. In
Section 5 we state the necessary results from singularity theory and apply them to the
example of evolutoids, giving the main results as Corollary 5.7 and Theorem 5.9. In a
nutshell, the results say that as α varies the local appearance of the evolutoids changes
in one of only two ways, a “swallowtail transition” as in Figures 5 and 6, or a “beaks”
transition as in Figures 4 and 9.
In Section 6 we study the wavefronts associated to a given value of α; the cusps
on these wavefronts trace out the envelope τα , but in general the wavefronts are not
closed curves. Finally, in Sections 7 and 8 we draw things together and offer some
more details of proofs.
872
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
2. ENVELOPES. We shall abuse notation mildly by using the same symbol σ to
denote a plane curve and a parametrization. For us, the most interesting examples occur
when σ is a closed curve, such as an ellipse (σ (t) = (a cos t, b sin t) for a > 0, b > 0,
and 0 ≤ t ≤ 2π), and we shall give such examples below. Since the line L is not itself
oriented, the angle α in Figure 1 can be considered modulo π. Parametrizing σ by t
automatically gives it an orientation (increasing t), but if we reverse this orientation,
then α gives the same line L.
We assume throughout that σ (t) = (X (t), Y (t)) is a regular parametrized curve;
that is, X and Y are smooth functions of t and (using for d/dt), X (t) = Y (t) =
0 never happens, so that the speed ||σ (t)|| is always nonzero. We use the standard
notation T , or T (t), for the unit tangent σ (t)/||σ (t)|| and N , or N (t), for the unit
normal, obtained by rotating T counterclockwise through 12 π. The curvature κ(t) is
given by T (t) = κ(t)||σ (t)||N (t), and in terms of X and Y has the rather unattractive
formula (omitting the variable t) κ = (X Y − X Y )/(X 2 + Y 2 )3/2 .
Finally, we make the following assumption.
Assumption 2.1. The curve σ is generic in the precise sense that all inflexions (zeros
of curvature) and vertices (extrema of curvature) are ordinary. This means that, for σ ,
inflexions are simple zeros and vertices are simple extrema: κ = 0, κ = 0 at inflexions,
and κ = 0, κ = 0, κ = 0 at vertices.1
We are interested in the line L obtained by rotating the tangent a fixed angle α; the
direction of L is therefore T (t) cos α + N (t) sin α. Hence, T (t) sin α − N (t) cos α is
perpendicular to L and a vector equation of the line L is F(x, t) = 0, where
Here · is the usual scalar product of vectors and x = (x, y) ∈ R2 . Thus, we regard α as
fixed; then F(x, t) = 0 represents a family of lines: Each t gives a line, and as t varies
the line moves in the (x, y)-plane.
There is a simple method for finding the envelope τα of these lines. Visually, the
envelope is a curve tangent to all the lines, or the “limit of intersection points of con-
secutive lines”; see Figure 3. In fact, both these latter definitions, when made precise,
give curves which are always subsets of τα , as defined below. This is shown in [6, pp.
107–109].
Figure 3. Envelopes of families of lines are intuitively formed by (left) intersections of “consecutive” lines or
(right) a curve tangent to all the lines. Generally these definitions coincide with Definition 2.2 used here.
Definition 2.2. The envelope τα of the family of lines given by (1), for a fixed α, is
the set of points x = (x, y) in the plane for which there exists t with
∂F
F(x, t) = (x, t) = 0.
∂t
1 It can be shown that, in a precise sense, “almost all” closed curves are generic in the sense used here. See
which relate the derivatives of T and N to T and N themselves, the curvature κ and
the speed ||σ || of the curve σ . (See any book on differential geometry, or alternatively
[6, Chapter 2].)
Using the fact that α is constant gives
∂F
= − sin α||σ (t)|| + (x − σ (t)) · κ(t) sin α N (t)||σ (t)||
∂t
+ κ(t) cos αT (t)||σ (t)|| .
Now any vector is a linear combination of the form λT (t) + μN (t). Applying this to
the vector x − σ (t) and substituting in F = ∂∂tF = 0, we obtain two equations for λ
and μ:
λ sin α − μ cos α = 0,
λκ(t) cos α + μκ(t) sin α = sin α. (3)
provided κ(t) = 0. Thus, if α = ± 12 π, the lines F = 0 are the normals to σ and the
envelope is the usual evolute, namely the set of points σ (t) + κ(t)
1
N (t); these points are
also called the centers of curvature of σ . If α = 0 (or π), then the lines F = 0 are the
tangents to σ and the envelope, away from inflexions where κ(t) = 0, is the original
curve σ .
Note that when α = 0 and κ(t) = 0, then using (3) directly we have μ = 0, λ arbi-
trary so, as mentioned in the Introduction, we have the following.
Proposition 2.3. The envelope of tangent lines consists of the original curve σ to-
gether with the whole tangent line at inflexion points.
Our aim in what follows is to describe exactly how, as α moves from 12 π to 0, the
evolute of σ turns into the curve σ itself, and in particular to explain what happens at
inflexions, where κ(t) = 0.
Remark 2.4. There are numerous attractive properties of the envelope τα . Here are
two, the first pointed out to us by the referees, and the second proved in [10, Proposi-
tion 5].
(i) The line joining the center of curvature σ (t) + (1/κ(t))N (t) of the curve σ
at the point with parameter t to the envelope point as in (4) has the direction
sin αT (t) − cos α N (t), which is perpendicular to the line L. This has the fol-
lowing interpretation, using “angle in a semicircle”: the envelope point (4) is
the intersection of the line L with the circle tangent to σ at σ (t) and passing
874
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
through the center of curvature. When α = 12 π, this point is the center of cur-
vature. When α = 0 the result holds—the envelope point is σ (t) itself—except
of course at inflexions, where κ(t) = 0.
(ii) This relates the evolutoid τα with the center symmetry set (CSS) or Minkowski
set of σ (it is given yet a third name in [10], namely the midenvelope of σ ). This
is the envelope of all lines joining pairs of distinct points of σ at which the tan-
gent lines are parallel. The CSS is the subject of several investigations, both in
the plane, and through generalizations (using tangent planes rather than tangent
lines) in higher dimensions, employing more advanced techniques of singular-
ity theory. See [8, Section 5] for an introduction and [9] for a more technical
discussion. Now, consider the points P, Q of the evolutoid τα corresponding
with two distinct points p, q of σ at which the tangent lines are parallel. Let R
be the point where the line PQ meets the line pq. Then R is the CSS point on
pq. It is not clear to us whether there are sensible generalizations of evolutoids
to higher dimensions, and hence a generalization of this result. Note that the
CSS, unlike the evolutoid, is invariant under nonsingular linear transformations
of the plane, since such a transformation preserves parallel lines.
3. CUSPS ON THE ENVELOPE. Consider the envelope curve given by (4). When
will this curve not be regular? The condition is that the speed is zero, that is, the deriva-
tive of x with respect to t is the zero vector. Again, using the Serret–Frenet formulae
(2), this derivative is, assuming κ(t) = 0 and omitting now the variable name t,
κ κ
x = ||σ || − 2 sin α cos α − ||σ || sin α T + ||σ || sin α cos α − 2 sin α N
2 2
κ κ
κ
= ||σ || cos α − 2 sin α (cos αT + sin α N ).
κ
This is zero if and only if κ 2 ||σ || cos α − κ sin α = 0. Writing s for the arclength
parameter on σ , dκ
dt
= dκ ds
ds dt
= dκ ds
||σ (t)||; hence, we have the following.
Proposition 3.1. The envelope (4) (still assuming κ = 0) is not regular if and only
if κ 2 cos α − κs sin α = 0, where κs is the derivative of curvature with respect to ar-
clength s on σ . This condition can also be written in terms of the radius of curvature
ρ = 1/κ : ρs sin α + cos α = 0.
Note that κ 2 and κs are independent of the direction of orientation of σ , and that
α > 0, say, always means a counterclockwise rotation of the oriented tangent to σ .
This gives the same line, whichever orientation of σ is used. That is why the equation
in the Proposition is unaltered when the orientation of σ is reversed.
For α = ± 21 π, that is, for the envelope of normals (the evolute of σ ), the Proposi-
tion gives the familiar condition κ = 0 (this can be the derivative with respect to any
regular parameter), which says that σ has an extremum of curvature, that is, a vertex.
For α = 0, the envelope of tangents, it says that, away from inflexions, the envelope
has no cusps—but the envelope is σ itself, so this is nothing new.
Example 3.2. In the special case where σ (t) = (r cos t, r sin t), a circle of radius
r > 0, then κ is r1 , a nonzero constant. The Proposition shows that there are no cusps
at all, the only exception being cos α = 0, when all the lines are radii and pass through
the center of the circle so the envelope degenerates to a point. In fact, for other val-
ues of α we have ||x(t)||2 = r 2 cos2 α, so that the envelope is a concentric circle of
When a plane curve such as this envelope curve is not regular, then in general we
expect it to have an “ordinary cusp,” that is, a singular point which is “like” the cusp
at the origin on the curve (t 2 , t 3 ). More formally, a local diffeomorphism of the plane
should take the given curve to this standard cusp.2 The condition for an ordinary cusp
is in fact that the second and third derivatives of σ (with respect to any regular param-
eter), evaluated at the cusp point, should be independent vectors.3 Note that for (t 2 , t 3 )
at t = 0 these vectors are (2, 0) and (0, 6), hence certainly independent. Some modest
calculation (see Section 8.1) reveals the following.
Proposition 3.3. Assume as before that κ is nonzero, and also that sin α = 0; that
is, the lines forming the envelope are not the tangent lines to σ . Then the cusp as in
Proposition 3.1 is an ordinary cusp if and only if 2κs2 − κκss = 0, where the deriva-
tives, with respect to arclength s, are evaluated at the cusp point. This can also be
written as ρss = 0 where ρ = 1/κ is the radius of curvature of σ .
Our object in this article is not to study a single value of α, but to study what
happens to τα as α varies. For this we put the investigation in a wider context.
in the plane to a neighborhood of the origin, taking the one curve to the other. Similar definitions apply to
higher dimensions.
3 Starting with (x, y) = (at 2 + bt 3 + · · · , ct 2 + dt 3 + · · · ) this means that ad − bc = 0. In fact, an in-
876
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
Thus, combining our notation yields
Example 4.1.
(i) Cusp. Let G(X, t) = G(x1 , x2 , t) = t 3 + x1 t + x2 . For a fixed X, this is a (re-
duced) polynomial of degree 3 in t, and DG consists exactly of those polyno-
mials with a repeated root. It is the curve X = (−3t 2 , 2t 3 ) parametrized by t
with an ordinary cusp at the origin.
(ii) Cuspidal edge surface. A bit more bizarrely, let G(X, t) = G(x1 , x2 , x3 , t) =
t 3 + x1 t + x2 , where x3 plays no role on the right side. The discriminant is
called a (standard) cuspidal edge surface. It is the product of an ordinary cusp
with a line, as in Figure 4, left, and is parametrized by x3 and t, namely
The product of the cusp point itself with this line (the x3 -axis here) is called
the line of cusps. This surface has an important property.
Consider a point p of the line of cusps, and the tangent T to the line of
cusps at p. Then the tangent planes at points away from the line of cusps
but with limit p, have a limit that contains T.
For the surface as in (7), a normal is in direction (t, 1, 0) at points away from
the line of cusps, and this has limit (0, 1, 0) as t → 0, so the limiting tangent
plane is x2 = 0. Of course, in general when we encounter a cuspidal edge
surface it will not be so “straight up and down” (see Figures 7 and 8)—it will
be locally diffeomorphic to the standard surface in a neighborhood of a point
p as above—but the property stated will still be true locally.
Figure 4. Left: The level sets x3 = c of the function h 1 (x1 , x2 , x3 ) = x3 in the cuspidal edge surface are
all curves with a cusp. Center and right: The level sets defined by functions h 2 (x1 , x2 , x3 ) = x1 + x32 and
h 3 (x1 , x2 , x3 ) = x1 − x32 . Below each figure is drawn a schematic diagram of the transitions undergone by
these level sets as c changes. The transition for h 2 is called a “beaks” or “bec-à-bec,” and that for h 3 a “lips,”
where for c > 0 the level set is empty and for c = 0 it is a single point.
Figure 5. Left: A swallowtail surface with the curved lines of cusps and self-intersection curve marked. These
all pass through the origin O. Right: A planar section x1 = constant c < 0 of a swallowtail surface, and the
“swallowtail transition” which these sections—level sets of the function h 0 (x1 , x2 , x3 ) = x1 —undergo as the
constant c moves through 0.
The origin is then called the swallowtail point and there are two lines of
cusps through the origin, given by G = G t = G tt = 0 and parametrized by
(−6t 2 , 8t 3 , −3t 4 ), t > 0, and t < 0. These represent polynomials with a triple
root. There is also a curve of self-intersection, parametrized by (−2t 2 , 0, t 4 )
(a half-parabola), representing polynomials with two double roots. The origin
represents the polynomial t 4 with a fourfold root. One important property of
this surface is the following.
The tangents to the lines of cusps, and the tangent to the self-intersection
curve, all have the same limit at the origin (here the limit is the x1 -axis).
Furthermore, the tangent planes to the swallowtail surface at points
away from these curves all have the same limit (here the x1 x2 -plane),
which contains the above limit of tangent lines.
878
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
This is the union of all the envelopes, for all α: they are spread out in the
α-direction. (Figures 6 and 7 below illustrate this discriminant.)
Figure 6. See Example 4.2. Two views of the discriminant surface DF , as in (9), for σ an ellipse. The α-axis
is vertical in the right-hand figure and α = 0 is at the bottom (the envelope is the original ellipse) and α = 12 π
is at the top (the envelope of normals has four cusps). The surface appears to have (curved) cuspidal edges,
with (curved) lines of cusps C marked, and the horizontal sections—the level sets of α close to the marked
swallowtail point, where α = α0 , say—appear to undergo a swallowtail transition.
Figure 7. Two views of the cuspidal edge discriminant DF close to an ordinary inflexion of a curve and close
to α = 0, looking from below the plane α = 0 (left), and from above (right). The lines of cusps C are labelled
as are the visible parts of the curve σ in the plane α = 0; also the x-axis, which is the inflexional tangent to σ
at the origin. The plane sections α = constant of DF evolve through α = 0 by a “beaks” transition, as shown
in Figure 9.
for the cuspidal edge surface in (ii) and the function h 0 (X) = x1 for the swallowtail
surface in (iii). These are illustrated in Figures 4 and 5 by means of their level sets, that
is, the sets of points of the cuspidal edge or swallowtail surface for which h i = c, for
values of c passing through 0. We are interested in the answers to two questions here.
Thus, the type measures how many partial derivatives of G with the xi parameters
held fixed vanish at t0 . In the special case of the family F (x, y, α, t) in (5), defin-
ing the envelopes τα , X is replaced by (x, y, α). The following Proposition gives the
conditions for A2 and A3 singularities, expressed in terms of the curvature κ and its
derivatives. Part (i) is a routine and not very interesting calculation using the Serret–
Frenet formulas (2). Part (ii) deals with the case of an inflexion and we shall verify
this since it is slightly more surprising. Recall from Proposition 2.3 that the envelope
of tangents to a curve σ , that is τα with α = 0, when σ has inflexions, consists of σ
and the whole tangent line at inflexion points. We shall be interested in the discrimi-
nant DF close to a point (x0 , 0) where x0 = (x0 , y0 ) is the inflexion point itself. It will
turn out that this discriminant is locally diffeomorphic to a cuspidal edge surface, as in
Figure 7, and we need the result of (ii) to show this.
5 Perestroika in Russian means approximately the same as “restructuring” in English, and the Western World
heard a great deal about it in the 1980s and 1990s during the Gorbachev era in the Soviet Union and then the
Russian Federation. Its use in a mathematical context was popularized by the great Russian mathematician
Vladimir Igorevich Arnol’d (1937–2010).
880
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
Proposition 5.2. Let the point (x0 , α0 , t0 ) = (x0 , y0 , α0 , t0 ) satisfy F = Ft = 0, so
that (x0 , α0 ) ∈ DF .
(i) Let s denote the arclength function on σ , as in Section 3. Suppose that κ(t0 ) = 0
so that x0 is given by (4). Then f (t) = F (x0 , α0 , t) has, at t = t0 ,
(a) type A2 provided, at t0 ,
Note that the A2 condition in (i)(a) is the same as that for an ordinary cusp given in
Proposition 3.3.
Proof of (ii). The calculations are marginally simplified by taking the parameter t to
be itself arclength, so that Serret–Frenet formulas become T = κ N and N = −κ T .
(But we shall still write t for the parameter.) Let us take t0 = 0. We know that, locally,
F = Ft = 0 at (x, 0, t) implies that either x = σ (t) or x is on the tangent line at the
inflexion σ (0). Using the formula (5), we differentiate twice and put α = t = 0; this
gives Ftt (x, 0, 0) = (x − σ (0)) · (κ (0)T (0)), which can only be zero, for x close to
the inflexion point, when x coincides with that point. (This actually implies that at
other points of the tangent line at the inflexion, DF is a nonsingular surface.) Further-
more, Fttt (σ (0), 0, 0) = −2κ (0) = 0, so F (σ (0), 0, t) has exactly an A2 singularity
at t = 0.
Example 5.5.
(i) For the G, and X0 = (0, 0, 0), t0 = 0 in Example 5.3, where r = 2, the deriva-
tives with respect to xi are t 2 , t, 0, respectively (independently of X0 , in fact).
The Taylor polynomials up to degree 1, expanded about 0, are 0, t, 0, respec-
tively, so the criterion of Definition 5.4 does not hold. It is also clear that there is
“something missing” from this family for, whatever the values of x1 , x2 , x3 , the
function g(t) = G(x1 , x2 , x3 , t) will always have a singularity at t = 0, in fact
of type A1 unless x1 = 0. We cannot turn the graph of the function g(t) = t 3
into a graph without turning points by adjusting the xi . On the other hand, for
Example 4.1(ii) the condition for a critical point is 3t 2 + x1 = 0 and there are
no critical points if x1 > 0.
(ii) For the G of Examples 4.1(i), (ii), (iii), the criterion is easily shown to hold.
For example, with 4.1(iii), where r = 3, the Ti are t 2 , t, 1, respectively.
Theorem 5.6. Suppose that G satisfies the criterion of Definition 5.4. Then, in a neigh-
borhood of X0 ∈ DG , the discriminant is locally diffeomorphic to a standard cuspidal
edge surface when r = 2 and a standard swallowtail surface when r = 3 (as in Ex-
amples 4.1(ii), (iii)).
The most important example for us is of course DF , and there we must work a little
harder to verify the criterion of Definition 5.4 when G = F . We give some details in
Section 8.2 and the result is then as follows.
Corollary 5.7. The family F , as in (5), satisfies the conditions of the above the-
orem. Thus, when f (t) = F (x0 , t) has an Ar singularity at t0 , r = 2 or 3, in the
cases covered by Proposition 5.2, the discriminant DF , which is the union of the en-
velopes τα spread out in the α direction, is always locally diffeomorphic to a standard
cuspidal edge (r = 2) or a standard swallowtail surface (r = 3) in a neighborhood
of x0 .
5.2. How to recognize level sets of a function. There are two cases to consider,
namely level sets of a function on a cuspidal edge surface, and on a swallowtail surface
in 3-space R3 . Fortunately, in both cases the conditions to realize the transitions of
Figures 4 and 5 are intuitively very reasonable. We shall only state the answers here;
there is more discussion of functions on discriminants in [2, 5]. For a smooth function
h : R3 → R, defined in a neighborhood of X0 ∈ DG , there will be a kernel plane K
through X0 . This is the plane tangent to the level set of h through X0 and has equation
(X − X0 ) · (h x1 , h x2 , h x3 ) = 0, where the partial derivatives are evaluated at X0 . We
refer back to Examples 4.1(ii), (iii) for properties of tangent planes to a cusped edge
and swallowtail surface.
882
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
Proposition 5.8.
(i) For a cuspidal edge surface, with X0 on the line of cusps, the level sets of h on
DG will all be cusped curves, as in Figure 4, left, provided the plane K does
not contain the tangent to the line of cusps through X0 . (“K is transverse to the
line of cusps.”)
On the other hand, the levels sets undergo a “beaks” or “lips” transition, as
in Figure 4, center and right, provided K does contain this tangent but does not
coincide with the limiting tangent plane to the cuspidal edge surface at points
approaching X0 . (“K is transverse to this limiting tangent plane.”)
(ii) For a swallowtail surface, with X0 at the swallowtail point, the level sets on DG
undergo a swallowtail transition, as in Figure 5, with two cusps merging and
disappearing, provided K does not contain the limiting tangent to the lines of
cusps on DG at X0 . (“K is transverse to this limiting tangent line.”)
See Figure 8 for examples where the conditions of (i) hold and also where they do
not hold. Similarly, level sets of a function on a swallowtail surface failing to satisfy
the condition of (ii) will not resemble a swallowtail transition.
Figure 8. (a) An example of a function h(X) = x2 + x32 on the cuspidal edge illustrated, which does not satisfy
the conditions of Proposition 5.8(i), since the level set for c = 0, appearing as the lighter colored surface on the
left, has tangent plane at the origin equal to the x1 x3 -plane, which coincides with the limiting tangent planes
to the cuspidal edge. The level sets where the surface x2 + x32 = c meets the cuspidal edge evolve in the way
illustrated. Clearly this is not anything like the transitions of Figure 4. In (b) and (c), the function on the curved
cuspidal edge M is assumed to be height in the direction of the vertical arrow. For (b), the level set at any level
is a horizontal plane, which is transverse to the line of cusps C and all the horizontal sections are cusps. For
(c), the level set (horizontal plane) through P is tangent to C but does not coincide with the (vertical) limiting
tangent planes to M, hence producing a “beaks” transition as in Proposition 5.8(i).
Remarkably, the conditions of Proposition 5.8 are always satisfied for the discrim-
inant DF , provided only that the original curve σ satisfies the Assumptions 2.1. We
sketch the proof of this in Section 8.3 below. In particular, the transition on the enve-
lope τα through α = 0, at an inflexion point of σ , is a “beaks” transition in which two
cusps collide, momentarily giving the envelope the entire tangent line at the inflex-
ion point, and then separate into two smooth branches. This is illustrated in Figure 9
for a closed curve with two inflexions, in fact the curve σ (t) = ((cos t + 2 cos(2t)
+10) cos t, (12 sin t − 8) sin t). (The transition cannot be of “lips” type as in Figure 4,
right, since the envelope cannot become empty.) Thus we have the following, which
completely describes the local changes in the envelopes τα for a curve σ satisfying the
genericity Assumption 2.1.
Theorem 5.9. The evolutoids τα evolve locally according to a stable cusp (Figures 4,
left, 6 and 8(b)) at A2 points where the curvature κ is nonzero; according to a swal-
lowtail transition (Figure 6) at A3 points where κ = 0; and according to a “beaks”
For example, in Figure 1, the envelope is undergoing swallowtail transitions that are
identical, up to a local diffeomorphism, with those of the standard swallowtail sections
in Figure 5.
Figure 9. Center: A curve σ with two inflexions, and (left) the envelope τα for α small and negative; (right)
for α small and positive. Two cusps approach and merge close to each inflexion, leaving smooth branches, in
a “beaks” transition as in Figure 4, center. The envelope for α = 0 is the original curve and the tangents at
the inflexion points, which are drawn dashed in the center figure (see Proposition 2.3 and compare Figure 2,
right). The figure also shows two swallowtail configurations on each envelope, which collapse in a swallowtail
transition as α → 0. The envelope τα goes “to infinity” at points corresponding to the inflexions themselves,
since the denominator in (4) vanishes.
Figure 10. Left: The “ordinary” wavefronts of an ellipse σ , which are obtained by displacing σ a constant
distance along its normals. The cusps on the wavefronts all lie on the envelope of normals, the 4-cusped curve,
which is also drawn. Right: Nonclosed wavefronts, given by (10), corresponding to the envelope of lines shown
in Figure 1, center, for which α = 14 π . Note that as the wavefronts (in either example) approach a cusp on the
envelope, two cusps on the wavefronts collapse together—this is, in fact, another example of a swallowtail
transition.
σ . The wavefronts have the parametrization σ (t) + wN (t), where w takes a constant
value along each wavefront. Thus σ is displaced a constant distance w along the nor-
mals to σ . The general prescription for wavefronts from a family of lines (or curves)
884
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
F(x, t) is as follows. We “integrate” F, that is, we look for a family G satisfying
G t = F. Then the wavefronts are given by F = 0 and G = constant.
In the case of the family of normals, which is given by F(x, t) = (x − σ (t)) · T (t),
it is easy to write down a suitable G. Let us take t to be the arclength parameter;
then we can take G(x, t) = − 12 ||x − σ (t)||2 = − 21 (x − σ (t)) · (x − σ (t)), since by
(2) σ (t) = T (t), so that G t = F. The solutions of F = 0, G = w1 = constant
√ <0 are
the points of the normal (F = 0) for which the distance to the curve is ± −2w1 , that
is, the ordinary parallels (offsets) of σ .
For the family F in (1), with α fixed, we have to work a little harder, but writing
F = F1 sin α − F2 cos α, we can use the above solution for G 1 with (G 1 )t = F1 and,
for F2 we need
G 2 = (x − (X (t), Y (t))) · (−Y (t), X (t)) dt,
where σ (t) = (X (t), Y (t)). Note that the integral is independent of the parametrization
of σ and in fact it represents, up to an added constant, twice the area swept out by
a line from x to σ (t), as t travels from some arbitrary starting value t0 . When t is
arclength, then G 2 = (x − σ (t)) · N (t) dt. In the special case of an ellipse σ (t) =
(a cos t, b sin t), we can write down the integral explicitly:
Note that this contains t on its own, so that unless cos α = 0, in which case G 2 is not
needed, the solution is not periodic. The wavefronts will not be closed curves, even
though the ellipse σ is closed.
Returning to F for a fixed α and any unit speed curve σ , and using G = G 1 sin α −
G 2 cos α, we have a prescription for finding the wavefronts. It is not hard to check
that the following formula satisfies F(σw (t), t) = 0, G(σw , t) = − 21 w2 sin α, which is
constant for a given w and α,
σw (t) = σ (t) + (w − t cos α)(T (t) cos α + N (t) sin α). (10)
This gives an explicit formula for the family of wavefronts corresponding to a given
σ and α. Fixing w and letting t vary, we get an individual wavefront, parametrized
by t, corresponding to that w. It can be checked that the singular points (cusps) on
the wavefronts are given by the additional condition G tt = 0, that is Ft = 0, which
says that the cusps lie on the envelope τα of lines given by F = Ft = 0. We say that
the cusps of the wavefronts sweep out the envelope τα . The exception is sin α = 0, for
which τα is the envelope of tangent lines to σ . In that case, there are no cusps to sweep
out anything.
If α = 12 π (the envelope of normals to σ ), then σw (t) = σ (t) + wN (t), which is the
usual parallel or offset of σ , obtained by moving down the normals a distance w. Note
that w and −w give different parallels, but the same value of G. Fixing the value of G
gives two parallels, corresponding to values of w of opposite sign. If we use α = − 21 π
instead, then w and −w are interchanged.
For a general α = ± 21 π, the factor w − t cos α in (10) tells us how far along the line
given by t we must go to reach the wavefront point. For a closed curve σ , parametrized
by 0 ≤ t < T , we can add any multiple of T on to t and obtain the same point of σ
and the same line of the family. Thus, the wavefront meets the line corresponding to
the value t infinitely often.
An example, using the ellipse, is shown in Figure 10, right.
8.1. Proof of Proposition 3.3. Here is an indication of how to prove Proposition 3.3
without getting tangled in too much algebra. We will assume for this that ||σ (t)|| = 1
for all t, which says that σ is unit speed, or that t = s is the arclength parameter.
This is no loss of generality (see any book on differential geometry of curves, or
alternatively [6, pp. 27–28]) and saves a little writing. In fact, write λ = − cos α +
(κ sin α/κ 2 ), so that the condition for a cusp is λ = 0. Then the envelope in (4) has
x = −λ cos αT − λ sin α N . Differentiating this twice, to give x and x , and putting
λ = 0 (after differentiating!) quickly gives the condition for these vectors to be linearly
dependent as 2κλ2 = 0, that is, λ = 0, and this gives the required formula, provided
sin α = 0.
8.2. Sketch of a proof of Corollary 5.7. Let us write s for sin α and c for cos α,
σ (t) = (X (t), Y (t)) and assume that σ is unit speed, that is, ||σ (t)|| = 1 for all t, and
that κ = 0 (see below for the case κ = 0). Recall (5) that
To show that F is “versal,” we need to consider the Taylor series in t, including con-
stant term, up to degree 2, of the derivatives Fx , F y , Fα , and check that they are in-
dependent when evaluated at an A3 (swallowtail) point, that is, one at which the con-
ditions of Proposition 5.2(i)(b) hold. So we need to take these three derivatives and
differentiate them with respect to t twice to get the necessary terms of the Taylor se-
ries. We are, of course, allowed to assume (4). For example,
but using (4), this boils down to simply − cos α. Differentiating again and substituting
from (4) shows Fαtt = 0. Let us write u = (s, c); then, for example, Fx = T · u and
F y = −N · u. Altogether we find that the 2-jets are the columns of the matrix J1 below
886
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
1
(where the binomial coefficient 2
has been omitted):
⎛ ⎞ ⎛ ⎞
Fx Fy Fα T ·u −N · u s
κ
J1 = ⎝ Fxt F yt Fαt ⎠ = ⎝ κN · u κT · u −c ⎠ .
Fxtt F ytt Fαtt 2 2
(−κ T + κ N ) · u (κ N + κ T ) · u 0
The determinant of this matrix is κ 2 sin α + κ cos α. If this is zero, then, using
Proposition 5.2(i) (where κs is our κ ), we have κ 4 + κ 2 = 0 so that κ = 0, contrary to
assumption. This proves the required independence. See [6, Chapter 6] for full details
of this method.
In the case of an A2 singularity where κ = 0, the versal unfolding condition
is simply that the top left 2 × 2 minor of J1 is nonsingular, and this amounts
to κ = 0.
Finally, when considering an (ordinary) inflexion, so that κ = 0, and working at
α = 0, it is easy to check that F has exactly an A2 singularity. Let us take the unit
tangent at the inflexion to be (1, 0) so that the unit normal is (0, 1). The vector u is
(0, 1) here, and the Taylor expansions to degree 1 of Fx , F y , Fα come to 0, −1, and
−t, respectively. These span polynomials of degree 1 in t.
8.3. Verifying the conditions needed to apply Proposition 5.8 to the discriminant
DF . First, consider the case where the curve σ does not have an inflexion at σ (t0 ), but
f (t) = F (x0 , α0 , t) has an A2 or A3 singularity at t = t0 . In the A2 case, by Corol-
lary 5.7, DF is locally diffeomorphic to a cuspidal edge surface close to (x0 , α0 ). This
cuspidal edge is given by F = Ft = Ftt = 0, that is, three equations in the four vari-
ables x, y, α, t, and the solutions are then projected to (x, y, α)-space, where DF lies.
A standard technique for calculating tangent vectors (the implicit function theorem,
which is covered in books of advanced calculus, or see [6, Chapter 4]) says that we look
for nonzero kernel vectors of the 3 × 4 matrix J2 of partial derivatives of F , Ft , Ftt
with respect to the four variables (x, y, α, t), evaluated at (x0 , α0 , t0 ). Note that the
first three columns of J2 are the same as the three columns of J1 in the previous sec-
tion, while the fourth column is (0, 0, Fttt = 0) at an A2 point and (0, 0, 0) at an
A3 point. Thus for an A2 point, we can always find a kernel vector, (x, y, α, t) say,
whose first three components are not all zero, using the first two rows of J2 , and then
determine t using the third row of J2 , since Fttt = 0. Then (x, y, α) is a nonzero tan-
gent vector to the line of cusps C on DF in (x, y, α)-space. However, this cannot be
done with α = 0 in view of the nonsingularity of the top left 2 × 2 submatrix of J2 (or
J1 ). So a tangent vector to C will never be horizontal and changing α to nearby values
gives a stable cusp as in Figure 8(b) and not (c).
There is clearly a problem with this argument at an A3 point (x0 , α0 ), where Fttt =
0, since in view of the nonsingularity of J1 , all kernel vectors of J2 have the form
(0, 0, 0, t). This simply says that, in (x, y, α)-space, the curve C on DF is singular at
(x0 , α0 )—not surprising, since DF is a swallowtail surface and the space curve C itself
has a cusp at (x0 , α0 ). However, the above argument still applies, by taking, say, a unit
tangent vector (x, y, α) and moving toward (x0 , α0 ) along C . The last component can-
not tend to 0 without the other two tending to 0 as well, which is a contradiction.
√ In the
present case, we can be more explicit: A tangent vector (of length 1 + κ 2 > 1) to C ,
obtained from the first two rows of J2 , is ((sin(2α), cos(2α)) · N , (sin(2α), cos(2α)) ·
T, κ). It is plain that this cannot have a limit in which the third component is 0. This
is certainly visible in Figure 6, where the limiting tangent to the lines of cusps is far
from horizontal.
It is clear that the limiting tangent to C is in the direction (1, 0, 0), which is in the plane
α = 0. Since we know that DF is a cuspidal edge surface, we can find the limiting
tangent plane to DF at points away from C by taking any path on DF that avoids
C (apart from at (x, α) = (0, 0, 0)), such as the path given by t = 0. The normal to
DF then comes to (0, −6ax + · · · , 6ax 2 + · · · ), which has limit (0, 1, 0) so that the
limiting tangent plane is the plane y = 0 and hence does not coincide with the plane
α = 0, as required.
ACKNOWLEDGMENTS. The second author is grateful for financial support from the Engineering and
Physical Sciences Research Council in the UK during his Ph.D. programme; the present article derives from
part of Chapter 3 of [14]. Both authors thank Joel Haddley for suggesting that they consider wavefronts corre-
sponding to the envelopes studied, Joel and Victor Goryunov for insisting that the inflexion case is also a beaks
transition, and the referees who encouraged us to improve our exposition and gave us earlier references to the
literature on evolutoids.
REFERENCES
1. T. M. Apostol, M. A. Mnatsakanian, Tanvolutes: Generalized involutes, Amer. Math. Monthly 117 (2010)
701–713, http://dx.doi.org/10.4169/000298910x515767.
2. V. I. Arnol’d, Wave front evolution and equivariant Morse lemma, Comm. Pure Appl. Math. 29 (1976)
557–582, http://dx.doi.org/10.1002/cpa.3160290603.
3. , Catastrophe Theory. Third edition. Springer-Verlag, Berlin, 1992, http://dx.doi.org/
10.1007/978-3-642-58124-3.
4. Th. Bröcker, L. Lander, Differentiable Germs and Catastrophes. London Math. Soc. Lecture Notes. Vol.
17. Cambridge Univ. Press, Cambridge, 1975, http://dx.doi.org/10.1017/cbo9781107325418.
5. J. W. Bruce, Functions on discriminants, J. London Math. Soc. 30 (1984) 551–567, http://dx.doi.
org/10.1112/jlms/s2-30.3.551.
6. J. W. Bruce, P. J. Giblin, Curves and Singularities. Cambridge Univ. Press, Cambridge, 1992.
7. J. W. Bruce, N. P. Kirk, A. du Plessis, Complete transversals and the classification of singularities, Non-
linearity 10 (1997) 253–275, http://dx.doi.org/10.1088/0951-7715/10/1/017.
8. P. Giblin, Affinely invariant symmetry sets, Caustics ’06, Banach Center Publications, Warsaw, Vol. 82
(2008) 71–84, http://dx.doi.org/10.4064/bc82-0-5.
9. P. J. Giblin, V. M. Zakalyukin, Singularities of centre symmetry sets, Proc. London Math. Soc. 90 (2005)
132–166, http://dx.doi.org/10.1112/s0024611504014923.
10. M. Hamman, A note on ovals and their evolutoids, Beiträge Algebra Geom. 50 (2009) 433–441.
11. G. Katz, How tangents solve algebraic equations, or a remarkable geometry of discriminant varieties,
Expo. Math. 21 (2003) 219–261, http://dx.doi.org/10.1016/s0723-0869(03)80002-6.
12. Medial Representations. Edited by S. M. Pizer and K. Siddiqi. Springer, New York, 2008, http://dx.
doi.org/10.1007/978-1-4020-8658-8.
13. R.-A. F. de Réaumur, Methode générale pour déterminer le point d’intersection de deux lignes droites
infiment proches, qui rencontrent une courbe quelconques vers le même côté sous des angles égaux
moindres, ou plus grandes qu’un droit, et pour connoı̂tre la nature de la courbe décrite par une infinité de
tels points d’intersection, Histoire de l’Académie Royale des Sciences (1709) 149–162, http://www.
academie-sciences.fr/activite/archive/ressource.htm.
888
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
14. J. P. Warder, Symmetries of Curves and Surfaces. Ph.D. thesis, University of Liverpool, 2009, http://
www.liv.ac.uk/~pjgiblin/papers/pw-thesis.pdf.
PETER J. GIBLIN is Professor of mathematics (Emeritus) at the University of Liverpool, where he was
Head of the Department of Mathematical Sciences 2000–2004. He has been a Visiting Professor at UNC
Chapel Hill, the University of Massachusetts at Amherst, and Brown University. His research interests are in
singularity theory and its applications to differential geometry and computer vision. He also enjoys working
with High School students on challenging projects. Now that he is retired, he has more time, equal enthusiasm,
but less energy for doing mathematics himself, thereby preserving the status quo.
Department of Mathematical Sciences, The University of Liverpool, Liverpool L69 7ZL, UK
pjgiblin@liv.ac.uk
J. PAUL WARDER is a career mathematician of some 25 years experience, having been engaged in varied
mathematical modelling and applied statistics roles in the nuclear, finance, and public service sectors. After
completing his undergraduate studies at the University of Liverpool in 1988, he returned to the mathematics
department in 2004 to complete successive M.Sc. and Ph.D. studies under the supervision of Peter Giblin.
paul.warder@gmail.com
Abstract. We give two proofs of a folklore result relating numerical semigroups of embedding
dimension two and binary cyclotomic polynomials and explore some consequences. In partic-
ular, we give a more conceptual reproof of a result of Hong et al. (2012) on gaps between the
exponents of nonzero monomials in a binary cyclotomic polynomial.
The intent of this paper is to better unify the various results within the cyclotomic polyno-
mial and numerical semigroup communities.
Then S is a semigroup (that is, it is closed under addition). The semigroup S is said
to be numerical if its complement Z≥0 \S is finite. It is not difficult to prove that
S(a1 , . . . , am ) is numerical if and only if a1 , . . . , am are relatively prime (see, e.g.,
[15, p. 2]). If S is numerical, then max{Z≥0 \S} = F(S) is the Frobenius number of S.
Alternatively, by setting d(k, a1 , . . . , am ) equal to the number of nonnegative integer
representations of k by a1 , . . . , am , one can characterize F(S) as the largest k such
that d(k, a1 , . . . , am ) = 0. The value d(k, a1 , . . . , am ) is called the denumerant of k.
That F(S(4, 6, 9, 20)) = 11 is well known to fans of Chicken McNuggets, as 11 is the
largest number of McNuggets that cannot be exactly purchased; hence, the notion of
of the Frobenius number is less abstract than it might appear at first glance. A set of
generators of a numerical semigroup is a minimal system of generators if none of its
proper subsets generates the numerical semigroup. It is known that every numerical
semigroup S has a unique minimal system of generators and also that this minimal
system of generators is finite (see, e.g., [18, Theorem 2.7]). The cardinality of the min-
imal set of generators is called the embedding dimension of the numerical semigroup
S and is denoted by e(S). The smallest member in the minimal system of generators
is called the multiplicity of the numerical semigroup S and is denoted by m(S). The
Hilbert series of the numerical semigroup S is the formal power series
HS (x) = x s ∈ Z[[x]].
s∈S
http://dx.doi.org/10.4169/amer.math.monthly.121.10.890
MSC: Primary 20M14, Secondary 11C08; 11B68
890
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
From PS one immediately reads off the Frobenius number:
with ζn an nth primitive root of unity (one can take ζn = e2πi/n ). It has degree ϕ(n),
with ϕ Euler’s totient function. The polynomial n (x) is irreducible over the rationals
(see, e.g., Weintraub [22]) and has integer coefficients. The polynomial x n − 1 factors
as
xn − 1 = d (x) (3)
d|n
where μ(n) denotes the Möbius function. From (4) one deduces that if p|n is a prime,
then
pn (x) = n (x p ). (5)
In case p and q are distinct primes it follows from (4) and Theorem 1 that
Already Carlitz [5] in 1966 implicitly mentioned this result without proof.
The Bernoulli numbers Bn can be defined by
z zn ∞
= Bn , |z| < 2π. (7)
ez − 1 n=0
n!
m
Bm = (1 + 2m−1 + 3m−1 + 5m−1 + 6m−1 + 9m−1 + 10m−1 + 13m−1 + 17m−1 )
4m − 1
m−1 r
7m m 4
+ (1 + 2m−r + 3m−r )Br .
4(1 − 4 ) r =0 r
m 7
The natural numbers 1, 2, 3, 5, 6, 9, 10, 13, and 17 are precisely those that are not in
the numerical semigroup S(4, 7).
Let f = c1 x e1 + · · · + cs x es , where the coefficients ci are nonzero and e1 < e2
< · · · < es . Then the maximum gap of f , written as g( f ), is defined by
Hong et al. [9] studied g(n ) (inspired by a cryptographic application [10]). They re-
duce the study of these gaps to the case where n is square-free and odd and established
the following result for the simplest nontrivial case.
Theorem 2. [9]. If p and q are arbitrary primes with 2 < p < q, then
g( pq ) = p − 1.
(x pq − 1)(x − 1)
Q { p,q} (x) = . (11)
(x p − 1)(x q − 1)
892
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
It can be shown that Q ρ (x) defines apolynomial of degree d := i (ri − 1). We define
its coefficients aρ (k) by Q ρ (x) = k≥0 aρ (k)x k . Furthermore, Q ρ (x) is selfrecipro-
cal; that is aρ (k) = aρ (d − k) or, what amounts to the same thing,
1
Q ρ (x) = x Q ρ d
. (12)
x
If all elements of ρ are prime, then comparison of (10) with (4) shows that
If n is an arbitrary integer and γ (n) = p1 · · · ps its square-free kernel, then by (5) and
(13) we have Q { p1 ,..., ps } (x n/γ (n) ) = n (x) and hence inclusion-exclusion polynomials
generalize cyclotomic polynomials. They can be expressed as products of cyclotomic
polynomials.
then Q ρ (x) = d∈Dρ d (x).
ρ−1
σ −1
q−1
p−1
Q { p,q} (x) = xip x jq − x − pq xip x jq , (14)
i=0 j=0 i=ρ j=σ
Lemma 1. Let p, q > 1 be coprime integers. Let ρ and σ be the (unique) nonnegative
integers for which 1 + pq = ρp + σ q. Let 0 ≤ m < pq. Then either m = αp + βq or
m = αp + βq − pq with 0 ≤ α ≤ q − 1 the unique integer such that αp ≡ m(mod q)
and 0 ≤ β ≤ p − 1 the unique integer such that βq ≡ m(mod p). The inclusion-
exclusion coefficient a{ p,q} (m) equals
⎧
⎪
⎨1 if m = αp + βq with 0 ≤ α ≤ ρ − 1, 0 ≤ β ≤ σ − 1;
−1 if m = αp + βq − pq with ρ ≤ α ≤ q − 1, σ ≤ β ≤ p − 1;
⎪
⎩0 otherwise.
This corollary (in case p and q are distinct primes) is due to Carlitz [5].
Lemma 1 can be nicely illustrated with an LLL-diagram (for Lenstra, Lam and
Leung). Here is one such diagram for p = 5 and q = 7:
28 33 3 8 13 18 23
21 26 31 1 6 11 16
14 19 24 29 34 4 9
7 12 17 22 27 32 2
0 5 10 15 20 25 30.
We start with 0 in the lower left and add p for every move to the right and q for every
move upward. Reduce modulo pq. Every integer 0, . . . , pq − 1 is obtained precisely
once in this way (by the Chinese remainder theorem).
Lemma 1 can be reformulated in the following way.
Lemma 2. Let p, q > 1 be coprime integers. The numbers in the lower left corner of
the LLL-diagram are the exponents of the terms in Q { p,q} with coefficient 1. The num-
bers in the upper right corner are the exponents of the terms in Q { p,q} with coefficient
−1. All other coefficients equal 0.
Theorem 4. If p, q > 1 are coprime integers, then PS( p,q) (x) = Q { p,q} (x).
Our first proof will make use of “what is probably the most versatile tool in numer-
ical semigroup theory” [18, p. 8], namely Apéry sets.
First proof of Theorem 4. The Apéry set of S with respect to a nonzero m ∈ S is de-
fined as
Ap(S; m) = {s ∈ S : s − m ∈ S}.
Note that
S = Ap(S; m) + mZ≥0
and that Ap(S; m) consists of a complete set of residues modulo m. Thus, we have
∞
1
HS (x) = xw x mi = x w. (15)
w∈Ap(S;m) i=0
1 − xm w∈Ap(S;m)
1 + x q + · · · + x ( p−1)q 1 − x pq
HS( p,q) (x) = = .
1 − xp (1 − x q )(1 − x p )
Using this identity and (11) easily completes the proof.
894
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
Remark. This proof is an adaptation of the arguments given in [16]. Indeed, once
we know the Apéry set of a numerical semigroup S, by using [16, (4)], we obtain an
expression for HS (x) and consequently for PS (x). Theorem 4 is a particular case of
[16, Proposition 2], with { p, q} = {a, a + d} and k = 1.
Our second proof uses the denumerant (see [15, Chapter 4] for a survey) and the start-
ing point is the observation that
1
= r ( j)x j , (16)
(1 − x p )(1 − xq) j≥0
Proof. Put α ≡ kp −1 (mod q), 0 ≤ α < q and β ≡ kq −1 (mod p), 0 ≤ β < p and k0 =
αp + βq. Note that k0 < 2 pq. We have k ≡ k0 (mod pq). Now if k ∈ S, then k < k0
and k + pq = k0 ∈ S (since k0 < 2 pq). It follows that if r (k) = 0, then r (k + pq)
= 1. If k ∈ S, then k = k0 + t pq for some t ≥ 0 and we have r (k) = 1 + t, where we
use that
Remark. It is not difficult to derive an explicit formula for r (n) (see, e.g., [2, Section
1.3] or [13, pp. 213–214]). Let p−1 , q −1 denote inverses of p modulo q, respectively,
q modulo p. Then we have
−1
n p−1 n q n
r (n) = − − + 1,
pq q p
where {x} denote the fractional-part function. Note that Lemma 3 is a corollary of this
formula.
pq−1
∞
(1 − x pq ) r ( j)x j = r ( j)x j + (r ( j) − r ( j − pq))x j
j≥0 j=0 j= pq
pq−1
= r ( j)x j + xj = x j,
j=0 j≥ pq j∈S( p,q)
where we used that r ( j) ≤ 1 for j < pq and r ( j) ≥ 1 for j ≥ pq. Using this identity
and (16) easily completes the proof.
S ∪ (F(S) − S) = Z,
where F(S) − S = {F(S) − s|s ∈ S}. Symmetric semigroups occur in the study of
monomial curves that are complete intersections, Gorenstein rings, and the classifica-
tion of plane algebraic curves; see, e.g., [15, p. 142]. For example, Herzog and Kunz
showed that a Noetherian local ring of dimension one and analytically irreducible is a
Gorenstein ring if and only if its associate value semigroup is symmetric.
We will now show that the self-reciprocity of Q { p,q} (x) implies that S( p, q) is sym-
metric (a well-known result; see, e.g., [18, Corollary 4.7]).
Proof. If s ∈ S ∩ (F(S) − S), then s = F(S) − s1 for some s1 ∈ S. This implies that
F(S) ∈ S, a contradiction. Thus, S and F(S) − S are disjoint sets. Since every inte-
ger n ≥ F(S) + 1 is in S and every integer n ≤ −1 is in F(S) − S, the assertion is
equivalent to showing that
xj + x F(S)− j = 1 + x + · · · + x F(S) , (17)
0≤ j≤F(S) 0≤ j≤F(S)
j∈S j∈S
we see that x F(S)+1 PS (1/x) = PS (x) if and only if (17) holds. Clearly, (17) holds if
and only if S is symmetric.
Using the latter result and Theorem 4 we infer the following classical fact.
Theorem 4 together with Theorem 3 shows that if e(S) = 2, then PS (x) can be written
as a product of cyclotomic polynomials. This leads to the following problem.
Problem 1. Characterize the numerical semigroups S for which PS (x) can be written
as a product of cyclotomic polynomials.
896
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
5. GAP DISTRIBUTION. The nonnegative integers not in S are called the gaps of
S. For example, the gaps in S(4, 7) are 1, 2, 3, 5, 6, 9, 10, 13, and 17. The number of
gaps of S is called the genus of S and denoted by N (S). The set of gaps is denoted by
G(S). The following well-known result holds; see [15, Lemma 7.2.3] or [18, Corollary
4.7].
From (2) and Theorem 1 we infer the following well-known result due to Sylvester:
For four different proofs of (18) and more background, see [15, pp. 31–34]; the shortest
proof of (18) and (19) the author knows of is in the book by Wilf [23, p. 88].
Additional information on the gaps is given by the so-called Sylvester sum
σk ( p, q) := sk .
s∈ S( p,q)
∞
zk e pqz − 1 1
σk ( p, q) = pz − z . (21)
k=0
k! (e − 1)(e − 1) e − 1
qz
We obtain from (21), on multiplying by z and using the Taylor series expansion (7),
that
∞
zm ∞
zi
∞
z j ( pqz)k
∞
∞
zm
mσm−1 ( p, q) = Bi pi Bjq j − Bm .
m=1
m! i=0
i! j=0 j! k=0 (k + 1)! m=0 m!
1
σ1 ( p, q) = ( p − 1)(q − 1)(2 pq − p − q − 1)
12
(this result is due to Brown and Shiue [3]) and
1
σ2 ( p, q) = ( p − 1)(q − 1) pq( pq − p − q).
12
The proof we have given here of Theorem 8 is due to Rødseth [17], with the difference
that we gave a different proof of the identity (21).
By using the formula (9) for power sums, we obtain from Theorem 8 the identity
m
m
mσm−1 ( p, q) = pm−r −1 Bm−r q r Sr ( p) − Bm ,
r =0
r
p−1
( f (n + p) − f (n)) = ( f (nq) − f (n)),
n∈ S n=1
898
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
The following result gives some information on gap blocks and element blocks in
a numerical semigroup of embedding dimension 2. Recall that the smallest positive
integer of S is called the multiplicity and denoted by m(S).
Lemma 4.
1) The longest gap block, g(G(S)), has length m(S) − 1.
2) The longest element block, g(S), has length not exceeding m(S) − 1.
3) If S is symmetric, then g(S) = m(S) − 1.
Remark. The second observation was made by my intern Alexandru Ciolan. It allows
one to prove Theorem 10.
Theorem 9. If p, q > 1 are coprime integers, then g(Q { p,q} (x)) = min{ p, q} − 1.
Proof. Note that g(Q { p,q} (x)) equals the maximum of the longest gap block length and
the longest element block length, and hence, by Lemma 4 equals m(S( p, q)) − 1 =
min{ p, q} − 1.
Proof. Using that PS (x) = (1 − x)HS (x) and Lemma 4 we infer that g(PS (x)) =
max{g(S), g(G(S))} = m(S) − 1.
The next result gives an example where an existing result on cyclotomic coefficients
yields information about numerical semigroups.
Proof. In view of Theorem 11 we have a{ p,q} (k) = 1 if and only if k is at the start of an
element block (including the infinite block [F(S) + 1, ∞) ∩ Z). Moreover, a{ p,q} (k) =
−1 if and only if k is at the end of a gap block. The proof is now completed by invoking
Corollary 1.
Using Lemma 2 and Theorem 11 our folklore result can now be reformulated in
terms of the LLL-diagram.
900
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
The distribution of the quantity ρσ that appears at various places in this article has
been recently studied using deep results from analytic number theory by Bzdega [4]
and Fouvry [8]. In particular, they are interested in counting the number of integers
m = pq ≤ x with p, q distinct primes such that θ(m), the number of nonzero coeffi-
cients of m , satisfies θ(m) ≤ m 1/2+γ , with γ > 0 fixed. (Note that by Corollary 1 we
have θ(m) = 2ρσ − 1.)
ACKNOWLEDGMENT. I would like to thank Matthias Beck, Scott Chapman, Alexandru Ciolan, Pedro A.
Garcı́a-Sánchez, Nathan Kaplan, Bernd Kellner, Jorge Ramı́rez Alfonsı́n, Ali Sinan Sertoz, Paul Tegelaar, and
the three referees for helpful comments. Alexandru Ciolan pointed out to me that g(S) ≤ m(S) − 1, which
allows one to prove Theorem 10.
REFERENCES
PIETER MOREE is since 2004 Researcher/Scientific Coordinator (half/half) at the Max Planck Insitute for
Mathematics. His main scientific interest is in problems that require both algebraic and analytic number the-
ory for their solution, e.g., problems related to the Artin primitive root conjecture. He enjoys playing tennis,
reading, and making new friends.
Max-Planck-Institut für Mathematik, Vivatsgasse 7, D-53111 Bonn, Germany
moree@mpim-bonn.mpg.de
902
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
A Perron-type Theorem on the
Principal Eigenvalue of Nonsymmetric
Elliptic Operators
Lei Ni
Abstract. We provide a proof for a Perron-type theorem on the principal eigenvalue of non-
symmetric elliptic operators based on the strong maximum principle. This proof is modeled
after a variational proof of Perron’s theorem for matrices with positive entries that does not
appeal to Perron–Frobenius theory.
1. INTRODUCTION. Perron’s theorem (cf. [3, Theorem 1, Ch. 13]) asserts that a
square matrix A = (αi j ) with positive entries αi j > 0 must possess a positive eigen-
value with multiplicity one. Moreover, for this positive eigenvalue, there exists an
eigenvector whose entries are all positive. The purpose of this note is to prove an
analogous result for second order elliptic operators, which we will now describe.
Let be a smooth bounded domain in Rn and let
n
∂2 n
∂
L=− a (x)ij
+ bk (x) + c(x)
i, j=1
∂ xi ∂ x j k=1
∂ xk
n
a i j (x)ξi ξ j ≥ θ|ξ |2 (1)
i, j=1
for any ξ ∈ Rn . For the purpose of our discussion on the spectrum of L (with zero
boundary data), we may assume, by adding a constant if necessary, that c(x) ≥ 0.
Since L is not necessarily self-adjoint, its eigenvalues are in general complex numbers.
However, there exists the following analog of Perron’s theorem for positive matrices.
Theorem 1.1.
(i) There exists a real eigenvalue λ1 > 0 for the operator L with zero boundary
condition.
(ii) This eigenvalue is of multiplicity one, in the sense that there exists an eigen-
function w1 (x) > 0 in with w1 |∂ = 0, and if u is any other function not
http://dx.doi.org/10.4169/amer.math.monthly.121.10.903
MSC: Primary 35P15, Secondary 58J05; 47A75
Lemma 2.1. For any vector y ≥ 0 with y = 0, Ay > 0. In particular, for any real
vector z, there exists real number ε > 0 such that Ay > εz.
δ
ε=
max1≤i≤m (|z i | + 1)
904
3. THE PDE CASE. We proceed to give a proof of Theorem 1.1 along the same line
of arguments as above. Let L be the uniformly elliptic operator of Section 1. First,
recall the following strong maximum principle (see Corollary 2.8, 2.9 of [4], as well
as Theorem 4 and Lemma of Section 6.4 in [2]).
Theorem 3.1. Assume that u ∈ C 2 () ∩ C() satisfies Lu ≥ 0 and u|∂ = 0. Then
u > 0 in unless u ≡ 0. If u is not identically 0, then ∂u
∂ν
< 0 on ∂, where ν is the
exterior unit normal of ∂.
L −1 ( f )2 ≤ C f L 2 . (2)
Here · k denotes the Sobolev norm of H k (). Moreover, by [2, Theorem 5 of Ch.
6.3] there exists Ck ≥ 0 such that L −1 ( f )k+2 ≤ Ck f k for any k ≥ 0. Let k0 be
an integer so large, but fixed, that C 2 () ⊂ H k0 −1 (), and let X be the Hilbert space
H k0 () ∩ H01 (). Elliptic regularity ensures that L −1 maps X into X . In our discus-
sions below, L −1 : X → X is the infinite dimensional analogue of the linear transfor-
mation defined by a positive square matrix A : Rm → Rm in the previous section.
For the proof of Theorem 1.1, we need an infinite dimensional analogue of Lemma
2.1.
Lemma 3.1. For any nonzero u ∈ X , u ≥ 0, let w = L −1 (u). Then w > 0 in and
w|∂ = 0. Furthermore, for any v ∈ C 2 () with v|∂ = 0, there exists an ε > 0 such
that w ≥ εv.
and
∂v
v(x , xn ) = v(x , 0) + (x , 0)xn + o(xn ).
∂ xn
Now we present the proof of Theorem 1.1, the infinite dimensional analogue of
Perron’s theorem.
Proof. Let = {ε > 0|L −1 ( f ) > ε f in , for some f ∈ X, f > 0 in }. In view of
(2), it is easy to see that is bounded. Lemma 3.1 implies that = ∅. Let μ1 = sup .
Pick μ( j) ∈ with μ( j) → μ1 . By the definition of , there exist u j ∈ X , with
u j (x) > 0 for x ∈ and u j |∂ = 0 such that L −1 (u j ) > μ( j) u j . Without loss of gen-
erality, we assume that u j L 2 = 1. Since we can only infer that a subsequence of
{u j } is weakly convergent, we shall employ a finite iteration to get a better convergent
subsequence {z j } with each term z j satisfying the same properties as the correspond-
ing u j .
First, for any fixed j and l with 1 ≤ l ≤ k0 , let z j,l = (L −1 )l (u j ). We prove induc-
tively using the maximum principle that z j,l > μ( j) z j,l−1 in . Indeed, for l = 1, this
follows from the choice of z j,1 = L −1 (u j ) and z j,0 = u j . Assume that the claimed
inequality holds for l, namely z j,l > μ( j) z j,l−1 . Since z j,l+1 − μ( j) z j,l = L −1 (z j,l −
μ( j) z j,l−1 ), Theorem 3.1 and the inductive hypothesis imply that z j,l+1 > μ( j) z j,l .
On the other hand, the standard elliptic estimate ([2, Theorem 2, Section 6.3]) as-
serts that there exists a positive constant C depending on L and such that z j,l 2l
≤ C l u j L 2 . Let z j = z j,k0 . In particular, there exists a constant C independent of
j such that z j 2k0 ≤ C . By the Rellich–Kondrachov compactness theorem, after
passing to a subsequence if necessary, lim j→∞ z j = w1 in H k0 ()-norm, for some
w1 ∈ X . Note that L −1 (z j ) = z j,k0 +1 ≥ μ( j) z j,k0 = μ( j) z j . From this, we deduce that
L −1 (w1 ) ≥ μ1 w1 , w1 |∂ = 0, and w1 ≥ 0. To see that w1 = 0, we observe that z j
≥ (μ( j) )k0 u j ≥ 0; hence, z j L 2 ≥ (μ( j) )k0 . Now the argument in the linear algebra
proof, replacing Lemma 2.1 by Lemma 3.1, implies that L −1 w1 = μ1 w1 . In particular,
w1 > 0 in and w1 |∂ = 0. Taking λ1 = μ11 , this proves the existence of the principal
eigenvalue and a corresponding positive eigenfunction.
To show that λ1 is of multiplicity one, assume that Lu = λ1 u for a real-valued func-
tion u. Equivalently, we have L −1 u = μ1 u with μ1 = λ11 . Let = {η > 0|w1 − ηu
≥ 0}. Clearly is nonempty by Lemma 3.1 and is bounded. Then let η1 = sup .
906
n
∂v ∂ v̄ n
∂ log w1 ∂|v|2
L |v| = 2(e(λ) − λ1 )|v| − 2
2 2
a +2
ij
ai j .
i, j=1
∂ xi ∂ x j i, j=1
∂ xi ∂x j
As in [7], the regularity of |v|2 (which implies the finiteness of L |v|2 ) and the sin-
gularity of ∂ log w1 2 2
∂ xi
along ∂ imply that ∂|v|
∂N
= 0 on ∂. Here ∂|v|
∂N
is defined as
n i j ∂|v| 2 ∂
i, j=1 a ∂ xi ν j with ν being the exterior normal and ν j = ν, ∂ x j . Suppose that
e(λ) ≤ λ1 . Since v is nonconstant, the strong maximum principle implies that |v|2
can only attain its maximum (on ) at some boundary point x0 ∈ ∂. But Hopf’s
2
lemma (see Theorem 2.5 of [4]) asserts that ∂|v|
∂N
> 0 at x0 , which contradicts the just-
∂|v|2
established conclusion that ∂N
= 0 on ∂. The contradiction proves e(λ) > λ1 .
Note that the same argument as the one above proves that if μ is any Neumann
eigenvalue of the operator L with nonconstant eigenfunction, then e(μ) > 0. In fact,
the slightly better result e(μ) > min c(x) holds.
It seems interesting to estimate the gap e(λ) − λ1 from below in terms of the
geometry of the coefficients a i j (x), bk (x), and c(x), as well as that of . Another in-
teresting question is whether there is a generalization of this result to hypo-elliptic
operators. Similarly, we can ask for an effective positive lower estimate for the non-
trivial Neumann eigenvalue μ (i.e., the eigenvalue with nonconstant eigenfunctions)
for the operator L with c(x) = 0. For this last problem, consult [5] for some recent
progress for the case that the domain is convex.
ACKNOWLEDGMENT. The work is partially supported by NSF grant DMS-1105549. We thank Professor
Guershon Harel for the encouragement on writing this article up and Professor Jim Isenberg for editing help.
We particularly thank Professor Peter Li, Professor Nolan Wallach, and Professor Hung-Hsi Wu for their help,
which greatly improved the article’s exposition. Finally, we thank one of the referees who pointed out an
oversight in the first version and suggested various improvements.
REFERENCES
1. R. Bellman, Introduction to Matrix Analysis. Second edition. McGraw-Hill, New York, 1970.
2. L. Evans, Partial Differential Equations. Second edition. American Mathematical Society, Providence, RI,
2010.
3. F. Gantmacher, The Theory of Matrices. Chelsea, New York, 1977.
4. Q. Han, F.-H. Lin, Elliptic Partial Differential Equations. Second edition. American Mathematical Society
and Courant Institute of Mathematical Sciences, Providence, RI, 2011.
5. L. Ni, Estimates on the modulus of expansion for vector fields solving nonlinear equations, J. Math. Pures
Appl. 99 (2013) 1–16, http://dx.doi.org/10.1016/j.matpur.2012.05.009.
LEI NI received his Ph.D from the University of California, Irvine, in 1998. He was a research assistant pro-
fessor at Purdue University from 1998–2000 and a Szegő assistant professor at Stanford from 2000–2002
before he joined the faculty at the University of California, San Diego, where he is now a professor of mathe-
matics. His research interests include differential geometry and partial differential equations.
Department of Mathematics, University of California at San Diego, La Jolla, CA 92093
lni@math.ucsd.edu
908
Abstract. We explore the distinction between convergence and absolute convergence of se-
ries in both Archimedean and non-Archimedean ordered fields and find that the relationship
between them is closely connected to sequential (Cauchy) completeness.
1. INTRODUCTION.∞ A real series ∞ n=1 an is absolutely convergent if the “ab-
solute series” n=1 |an | converges. This is a strange and rather sneaky terminology;
many calculus students have heard “the series is absolutely convergent” as “absolutely,
the series is convergent.” The language strongly (and somewhat subliminally) hints
that
∞ an absolutely convergent series should converge. Fortunately this holds. Since
n=1 |an | converges, the partial sums form a Cauchy sequence—for all ε > 0, there is
N ∈ Z+ such that for all n ≥ N and m ≥ 0, we have n+m k=n |ak | < ε. Thus
m+n m+n
ak ≤ |ak | < ε,
k=n k=n
and ∞ n=1 an converges by the Cauchy criterion.
Above we used the convergence of Cauchy sequences—i.e., that R is sequentially
complete. In the spirit of “real analysis in reverse”—cf. [5] and [3]—it is natural to
ask about convergence versus absolute convergence in an arbitrary ordered field. In
a recent note, Kantrowitz and Schramm ask whether in an ordered field F, every ab-
solutely convergent series is convergent if and only if F is sequentially complete [2,
Question 3]. We will answer this question and also determine the ordered fields in
which every convergent series is absolutely convergent.
Let us begin with a taxonomic refresher on ordered fields. An ordered field F is
Dedekind complete if every nonempty subset S of F that is bounded above admits
a least upper bound. Dedekind complete ordered fields exist [1, §5], and if F1 and F2
are Dedekind complete ordered fields, there is a unique field homomorphism ι : F1 →
F2 , which is moreover an isomorphism of ordered fields [1, Cor. 3.6, Cor. 3.8]. This
essentially unique Dedekind complete ordered field is, of course, denoted by R and
called the field of real numbers.
For x, y ∈ F we write x y if n|x| < |y| for all n ∈ Z+ . An ordered field F
is non-Archimedean if there is x ∈ F with 1 x—equivalently, if Z+ is bounded
above in F—otherwise F is Archimedean. Notice that for x, y ∈ F with x = 0, x
y holds if and only if 1 xy . Thus F is non-Archimedean if and only if x y for
some x, y ∈ F with x = 0.
http://dx.doi.org/10.4169/amer.math.monthly.121.10.909
MSC: Primary 40J05
∞
a= bm t m , b M = 0, (1)
m=M
and then a has the same sign as b M . Put v(a) = M and also put v(0) = ∞. Then a
sequence {an } in R((t)) converges to 0 if and only if v(an ) converges to ∞.
Let {an } be a Cauchy sequence in R((t)). Then {an } is bounded; there is M ∈ Z
such that for all n ∈ Z+ we may write
∞
an = bm,n t m .
m=M
The Cauchy condition also implies that for every m ≥ M, the real sequence
∞ {bm,n }∞
n=1
is eventually constant, say with value bm , and thus {a
n } converges to
m
m=M bm t . Thus
R((t)) is sequentially complete. Similarly, a series ∞ n=1 an is convergent if and only
if an → 0, so absolute convergence is equivalent to convergence in R((t)).
An ordered field F is Archimedean if and only if there is a homomorphism
ι : (F, <) → (R, <) [1, Thm. 3.5], i.e., if and only if F is isomorphic to a sub-
field of R with the induced ordering. An ordered field is Dedekind complete if and
only if it is Archimedean and sequentially complete [1, Lemma 3.10, Thm. 3.11]. It
follows that the Archimedean ordered fields that are not (sequentially = Dedekind)
complete are, up to isomorphism, precisely the proper subfields of R.
Now we can state our main result.
Part a) of the main theorem is familiar from calculus. Part (i) has already been recalled,
and for part (ii) we need only exhibit the alternating harmonic series ∞ n=1
(−1)n+1
n
. We
have included part a) to facilitate comparison with the other cases.
A natural next step is to finish off the case of Archimedean ordered fields by estab-
lishing part c) for proper subfields of R. We do so in §2, using arguments that could be
presented in an undergraduate honors calculus/real analysis course. (This turns out to
910
2. SUBFIELDS OF R.
∞
Theorem 1. Let F R be a proper subfield. There is a series n=1 an with terms in
F that is absolutely convergent but not convergent.
and we have made an irrevocable mistake! The series ∞ sn
n=1 2n has terms in F and is
not convergent in F, but the associated absolute series is ∞n=1 2n = 1 ∈ F.
1
Proof. We may assume L ∈ [0, ∞). Let N1 be the least positive integer such that
N1
n=1 an > L, and put s1 = · · · = s N1 = 1. Let N2 be the least integer greater than
N2
N1
N1 such that n=1 sn an − n=N a < L, and put s N1 +1 = · · · = s N2 = −1. We
1 +1 n
continue in this manner, taking just enough terms of constant sign to place the partial
sum on the opposite side of L as the previous partial sum. The condition ∞ n=1 an = ∞
ensuresthis is well-defined, and the condition an → 0 guarantees that the resulting
series ∞ n=1 sn an converges to L.
The constructions of this section are intended to complement those of [2]. In particular,
the proof of Theorem 1 answers their Question 1 for b = 2.
Lemma 4. Let (X, <) be a totally ordered set, and let {xn }∞
n=1 be a sequence in X .
Then at least one of the following holds:
Proof. If the image of the sequence is finite, we may extract a constant subse-
quence. So assume the image is infinite. By passing to a subsequence we may assume
n
→ xn is injective. We say m ∈ Z+ is a peak of the sequence if for all n > m we have
xn < xm . If there are infinitely many peaks, the sequence of peaks forms a strictly de-
creasing subsequence. So suppose there are only finitely many peaks and thus there
is N ∈ Z+ such that no n ≥ N is a peak. Let n 1 = N . Since n 1 is not a peak, there
is n 2 > n 1 with xn2 > xn1 . Since n 2 is not a peak, there is n 3 > n 2 with xn3 > xn2 .
Continuing in this way we build a strictly increasing subsequence.
Now we introduce an invariant of a linearly ordered set that plays an important role
in the theory of ordered fields. A subset S of a linearly ordered set X is cofinal if for
all x ∈ X there is y ∈ S with x ≤ y. The cofinality of X is the least cardinality of a
cofinal subset. An ordered field is Archimedean if and only if Z+ is a cofinal subset, so
Archimedean ordered fields have countable cofinality. The subset {t −n }∞ n=1 of R((t)) is
countable and cofinal, so R((t)) is non-Archimedean of countable cofinality. For any
infinite cardinal κ there is an ordered field of cofinality κ [4, Cor. 2.7].
A Z-sequence in F is a sequence {an }∞ ∞an > 0 for all n and an → 0. A
n=1 with
ZC-sequence is a Z-sequence {an }∞ n=1 such that n=1 an converges.
912
Proof. (0) =⇒ (i): Let {Un }∞ n=1 be a countable neighborhood base at 0. For n ∈ Z ,
+
+
choose εn > 0 such that (−εn , εn ) ⊂ Un . Then εn | n ∈ Z is cofinal.
1
(i) =⇒ (o): Let {sn }∞ n=1 be a cofinal sequence of positive elements. Let εn = sn and
1
∞
Un = (−εn , εn ). Then {Un }n=1 is a countable base at zero, and thus for all x ∈ F, the
collection {Un + x}∞ n=1 is a countable base at x. Thus F is first countable.
(i) =⇒ (ii): Let S = {sn }∞ n=1 be a cofinal sequence. Put a1 = max(1, s1 ). Having de-
fined an , put an+1 = max(an + 1, sn+1 ). Then {an }∞ n=1 is a strictly increasing sequence
whose set of terms is a cofinal subset. The sequence {an−1 }∞ n=1 converges to 0 and is not
eventually constant.
(ii) =⇒ (iii): Let {xn }∞ n=1 be a sequence that is convergent and is not eventually con-
stant; its set of terms must then be infinite. By Lemma 4, {xn }∞ n=1 has a subsequence
that is either strictly increasing or strictly decreasing; by changing the signs of all
terms if necessary and adding a constant we get a strictly increasing convergent se-
quence 0 < S1 < S2 < · · · . Put S0 = 0, and for n ∈ Z+ , put an = Sn − Sn−1 . Then
{an }∞
n=1 is a ZC-sequence.
(iii) =⇒ (iv) =⇒ (v) is immediate.
(v) =⇒ (i): If {an }∞ n=1 is a Cauchy sequence and not eventually constant, then
1
m, n ∈ Z+ , am = an
|am − an |
is countable and cofinal. Let α > 0 in F. There is an N ∈ Z+ such that for all
m, n ≥ N , we have |am − an | ≤ α1 . Since the sequence is not eventually constant,
there are m, n ≥ N with am = an , and then α < |am −a
1
n|
.
Theorem 6 is a key step towards establishing the main theorem. For starters, it
disposes of the case of ordered fields of uncountable cofinality; by (i) ⇐⇒ (v),
must be sequentially complete. Moreover, by (i) ⇐⇒ (ii) an infinite
such fields
series ∞ n=1 an converges if and only if an = 0 for all sufficiently large n if and
only if an → 0. And in the case of countable cofinality it gives us some useful se-
quences.
Consider the sequence {t n }∞ +
n=1 in R((t)). It converges to 0, and for all n ∈ Z we
n ∞
have t n+1
t . One can use the sequence {t }n=1 to “test for convergence”; a sequence
n
{an }∞
n=1 in R((t)) converges if and only if for each N ∈ Z+ we have |an | ≤ t N for all
sufficiently large n.
Here is a useful generalization to arbitrary ordered fields. A test sequence is a Z-
sequence {εn }∞ +
n=1 such that εn+1 εn for all n ∈ Z .
4. PROOF OF THE MAIN THEOREM. We will now prove part b) of the main
theorem.
Let F be a sequentially complete non-Archimedean field. First we must show that
a series ∞ n=1 an in F is convergent if and only if an → 0. That a convergent series
has an → 0 follows (as usual) from the fact that a convergent sequence is a Cauchy se-
quence. Suppose an → 0. If F has uncountable cofinality, then by the remarks follow-
ing Theorem 6, a series ∞ n=1 an converges if and only if an = 0 for all large enough n,
hence if and only if an → 0. Now suppose F has countable cofinality. By Proposition
7, there is a test sequence {εn }∞ +
n=1 . For k ∈ Z , choose Nk such that for all n ≥ Nk , we
have |an | ≤ εk+1 . Then for all n ≥ Nk and ≥ 0, we have
Thus the sequence is a Cauchy sequence, and hence convergent because F is sequen-
tially complete. The fact that a series in F is convergent if and only if it is absolutely
convergent follows immediately, since an → 0 if and only if |an | → 0.
Before proving part c) of the main theorem, we want to build one more type of
∞
sequence. A ZD-sequence is a Z-sequence {an }∞ n=1 such that n=1 an diverges.
Proof. (i) =⇒ (ii): Let {an }∞ n=1 be a ZD-sequence in F. By part b) of the main theo-
rem, F cannot be non-Archimedean and sequentially
∞ complete.
(ii) =⇒ (i): If F is Archimedean, n1 n=1 is a ZD-sequence. Suppose F is non-
Archimedean and not sequentially complete, and let {an }∞ n=1 be a divergent Cauchy
sequence. By Lemmas 4 and 5, after passing to a subsequence and possibly changing
the sign, we get a strictly increasing, divergent Cauchy sequence {Sn }∞n=1 . Put S0 = 0,
and for n ∈ Z+ , put an = Sn − Sn−1 . Then {an }∞n=1 is a ZD-sequence.
Finally we can prove both assertions of part c) of the main theorem.
914
ank+1 − ank + ck
d2k−1 = , and
2
an − ank − ck
d2k = k+1 .
2
Then
k
k
This is a divergent
∞ subsequence of the sequence of partial sums associated to {dk }∞
k=1 ,
and hence k=1 dk diverges. Since −ck < ank+1 − ank < ck , we have
∞ ∞
Hence k=1 |dk | = k ck is convergent, i.e., k=1 dk is absolutely convergent.
Theorem 10. Let F be an ordered field that is not sequentially complete. Then F
admits a convergent series that is not absolutely convergent.
dn −dn
a2n−1 = , a2n = .
2 2
ACKNOWLEDGMENTS. We thank Paul Pollack and the referees for several comments that led to im-
provements of the exposition. The authors were partially supported by National Science Foundation grant
DMS-0701771.
REFERENCES
PETE L. CLARK is an associate professor at the University of Georgia. His primary research interests lie in
arithmetic geometry and number theory.
University of Georgia, Athens, GA 30606
pete@math.uga.edu
NIELS J. DIEPEVEEN is a private citizen of the Netherlands. He studies topology and related topics.
niels@dv1.demon.nl
916
Abstract. Suppose f = k=1 w f , where w1 , w2 , . . . , wk are positive constants,
k
=1 w = 1, and f 1 , f 2 , . . . , f k are real-valued and Riemann integrable functions on [−π, π ].
Let λ() () ()
1n ≤ λ2n ≤ · · · ≤ λnn be the eigenvalues of the Hermitian Toeplitz matrices Tn
()
() n () π −ir x
= [tr −s ]r,s=1 , where
tr = 2π −π e
1
f (x) d x , and let λ1n ≤ λ2n ≤ · · · ≤ λnn be the eigen-
values of Tn = n=1 w Tn() . We give conditions implying that
1
n
lim |λr n − w1 λr(1) (2) (k)
n − w2 λr n − · · · − wk λr n | = 0.
n→∞ n r =1
and
μr n = w1 λr(1) (2) (k)
n + w2 λr n + · · · + wk λr n , (5)
http://dx.doi.org/10.4169/amer.math.monthly.121.10.917
MSC: Primary 15A18, Secondary 15A51; 15B05
1
n
lim |λr n − μr n | = 0 (6)
n→∞ n
r =1
and
have positive Lebesgue measure for all > 0; thus, a and b are the essential lower
and upper bounds of h. We say that x0 is an essential minimum point of h if N (x0 ) ∩
[−π, π] ∩ L has positive measure for every neighborhood N (x0 ) of x0 and every
> 0, or an essential maximum point if N (x0 ) ∩ [−π, π] ∩ U has positive measure
for every N (x0 ) and > 0.
Let
and
π
1
n
1
lim G(γr n ) = G(h(x)) d x for all G ∈ C[a, b].
n→∞ n 2π −π
r =1
Given Szegö’s result and f as in (1), we assume that f 1 , f 2 , . . . , f k have the same
essential minimum and maximum a and b, so (1) and (4) imply that a ≤ f (x) ≤ b for
918
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
all x in [−π, π]. The essential minimum and maximum of f are a and b if and only
f 1 , f 2 , . . . , f k have at least one essential minimum point and one essential maximum
point in common, which we also assume.
Our final assumption stems from the following lemma, first stated in more general
form in [2]. See [3] and [4] for expository discussions of this special case.
Lemma 1. If
and
1
∞
lim (G(αr n ) − G(βr n )) = 0 for all G ∈ C[a, b];
n→∞ n
n=1
1
n
lim |αr n − βr n | = 0;
n→∞ n
r =1
1
∞
lim |G(αr n ) − G(βr n )| = 0 for all G ∈ C[a, b].
n→∞ n
n=1
For reasons that will soon become clear, Lemma 1 leads us to assume that
and 1 ≤ ≤ k, and
π
1
n
1
lim G(λr n ) = G( f (x)) d x for all G ∈ C[a, b]. (9)
n→∞ n 2π −π
r =1
Let
2(r − 1)
xr n = −1 π for 1 ≤ r ≤ n.
n
From (7), for each n ≥ 2 there is a single permutation σn of {1, 2, . . . , n} such that
Rearranging the terms on the left sides of (10) and (11) yields
1 1 π
n
lim G f xσn (r ),n = G( f (x)) d x for all G ∈ C[a, b], (14)
n→∞ n
r =1
π −π
where 1 ≤ ≤ k, and
1 1 π
n
lim G f xσn (r ),n = G( f (x)) d x for all G ∈ C[a, b]. (15)
n→∞ n
r =1
π −π
1 ()
n
lim G λr n − G f xσn (r ),n = 0 for all G ∈ C[a, b],
n→∞ n
r =1
1
()
n
lim λr n − f xσn (r ),n
= 0 for 1 ≤ ≤ k. (16)
n→∞ n
r =1
1
n
lim (G(λr n ) − G( f (xσn (r ),n ))) = 0 for all G ∈ C[a, b],
n→∞ n
r =1
1
m
lim |λr n − f (xσn (r ),n )| = 0. (17)
n→∞ n
r =1
k
μr n − f (xσn (r ),n )) = w (λr()n − f (xσn (r ),n )),
=1
920
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
so (16) implies that
1
n
lim |μr n − f xσn (r ),n )
= 0.
n→∞ n
r =1
1
n
lim |G(λr n − G(μr n )| = 0 for all G ∈ C[a, b].
n→∞ n
r =1
It is important to note that this paper is about a very special case of Szegö’s dis-
tribution theorem, which—in its original formulation—is about families of Hermitian
Toeplitz matrices Tn = [tr −s ]r,s=1
n
, where {tr }r∞=−∞ are the Fourier coefficients of a real-
valued (not necessarily bounded) Lebesgue integrable function on [−π, π]. This theo-
rem and its numerous extensions comprise part of the core of modern operator theory.
Although the arguments in this paper use the properties of the Riemann integral in an
essential way, it seems reasonable to hope that more knowledgeable investigators will
be motivated to extend our result to a more general setting.
REFERENCES
1. U. Grenander, G. Szegö, Toeplitz Forms and Their Applications, Univ. of California Press, Berkeley and
Los Angeles, 1958.
2. W. F. Trench, Absolute equal distribution of families of finite sets, Linear Algebra Appl. 367 (2003) 131–
146, http://dx.doi.org/10.1016/s0024-3795(02)00597-9.
3. , Simplification and strengthening of Weyl’s definition of equal distribution of two families of
finite sets, Cubo. A Mathematical Journal 06 no. 3 (2004) 47–54.
4. , An elementary view of Weyl’s theory of equal distribution, Amer. Math. Monthly 119 no. 10
(2012) 852–861, http://dx.doi.org/10.4169/amer.math.monthly.119.10.852.
Abstract. Let H, K be subgroups of G. We show that the intersection properties of left cosets
of H and right cosets of K exhibit considerable symmetry, allowing us to prove a generaliza-
tion of Hall’s theorem for transversals.
If H and K are subgroups of G, then G can be partitioned as the disjoint union of all
left cosets of H , as well as the disjoint union of all right cosets of K . But how do these
two partitions of G intersect each other?
A useful tool for studying the way left and right cosets interact, and obtaining
transversals, is the coset intersection graph that we introduce as follows.
Observing that left (respectively, right) cosets do not intersect, we see that H,K
G
is
a bipartite graph, split between {li H }i∈I and {K r j } j∈J .
For H a subgroup of a finite group G, the existence of a left-right transversal is well
known, sometimes presented as the following application of Hall’s marriage theorem
[4]. A set of k left cosets of H contains a total of k|H | elements, and these cannot fit
into fewer than k right cosets of H . It follows that every set of left cosets of H meets
at least as many right cosets of H . Hence, by Hall’s theorem there is a matching on the
bipartite graph H,H
G
and thus a left-right transversal by taking one element from each
edge in this matching.
This paper shows that in fact a much stronger result is true; we can completely
describe the way that left and right cosets of H intersect, without any need for Hall’s
theorem but instead by studying and applying the properties of the coset intersection
graph. We begin this now.
Theorem 3. H,K
G
is always a disjoint union of complete bipartite graphs.
922
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
K such that ah 1 = k1 b, k2 b = ch 2 , ch 3 = k3 d. Rearranging gives c = k3 dh −1 3 , so
b = k2−1 k3 dh −1
3 h 2 , so a = k k
1 2
−1
k 3 dh −1
3 h h
2 1
−1
and thus ah h
1 2
−1
h 3 = k k
1 2
−1
k 3 d. Hence,
a H − K d as required.
Take any li H and some K r j in the connected component of li H in H,K G
; there must
be at least one such K r j . We show li H and K r j are connected by an edge. For if not,
then there must be at least one finite path of length > 1 connecting them; take a min-
imal such path γ from li H to K r j . Then γ begins with li H − K a − bH − K c − . . .,
where K a = K r j . But by the previous remark, li H and K c must be joined by an edge,
contradicting the minimality of γ . So li H and K r j are joined by an edge for every K r j
in the connected component of li H .
Theorem 4. Let H, K < G. Suppose that |H | = m and |K | = n. Then the graph H,K G
Note that with this we have proved the existence of a left-right transversal for H a
finite subgroup of G without the use of Hall’s theorem.
Under the hypothesis of Theorem 4, we see that sets of si left cosets of H com-
pletely intersect sets of ti right cosets of K , with si /ti constant over i. With this in mind,
another way of visualizing H,K G
is by the following simultaneous double-partitioning
of G; draw left cosets of H as columns and right cosets of K as rows, partitioning G
into irregular “chessboards” denoted Ci , each with edge ratio n : m. Each chessboard
Ci corresponds to the connected component Ksi ,ti of H,K G
, and individual tiles in Ci
correspond to the nonempty intersection of a left coset of H and a right coset of K
(i.e., edges in Ksi ,ti ). Corollary 5 would then follow by choosing one element from
each tile on the leading diagonals of the Ci ’s. We give an example of this here.
left cosets: {eH, (1 2)H, (1 4)H, (2 4)H, (3 4)H, (1 2 4)H, (1 3 4)H, (1 4 2)H },
right cosets: {H e, H (1 2), H (1 4), H (2 4), H (3 4), H (1 4 2), H (1 4 3), H (1 2 4)}.
Note that, for convenience, the right transversal for H was chosen by taking the inverse
of the elements in the left transversal for H ; this can always be done.
The coset intersection graph H,H
G
, along with its corresponding chessboards, can
be seen in Figure 1 below. From these, we can easily read off a left-right transversal
for H in G, for example:
{e, (1 2), (1 4), (2 4), (3 4), (1 3)(2 4), (1 2)(3 4), (1 4)(2 3)}.
Proposition 6. Let H, K < G and g ∈ G, with H and K both finite. Then the number
of right cosets of K intersecting g H , denoted Mg , is given by
|H |
Mg = .
|g H g −1 ∩ K |
924
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
K = g H g −1 k ∩ K , which is the same as the number in g H g −1 ∩ K k −1 = g H g −1 ∩
K . Observe that this number will be |g H g −1 ∩ K : N |. This immediately gives
Proof. Consider the finite quotient G/ core(H ∩ K ) and set H and K to be H/ core
(H ∩ K ) and K / core(H ∩ K ), respectively. Apply Corollary 5 to H , K < G/ core
(H ∩ K ), forming a left transversal T of H that extends to a right transversal of K .
Now choose one preimage of each element of T from G and call this set T , which has
all the desired properties.
Proposition 9. Let H, K < G and g ∈ G, with H and K both of finite index. Then
the number of right cosets of K intersecting g H , denoted Mg , is given by
|G : g H g −1 ∩ K |
Mg = .
|G : H |
|H :N |
Proof. Follow the proof of Proposition 6 again to obtain Mg = |g H g −1 ∩K :N |
(where
N := core(H ∩ K )). The proposition now follows immediately.
All of our results can be derived from the work of Ore [7], who makes use of double
cosets; partitions of G into sets of the form K g H (where H, K < G). Observe that
every right coset of K in the double coset K g H meets every left coset of H in K g H .
To see this, note that a right coset of K in K g H has the form K gh, and a left coset
of H in K g H has form kh H . Thus K gh and kg H have common element kgh. So by
ACKNOWLEDGMENTS. We would like to thank Imre Leader for suggesting that we prepare this note
for the MONTHLY, the Department of Pure Mathematics and Mathematical Statistics at the University of
Cambridge for hosting us while we carried out this research, and Warren Dicks for showing such a keen
interest in the historical context of this material. The second author was partially supported by the Italian FIRB
“Futuro in Ricerca” project RBFR10DGUA 002 at the University of Milan. The main results in this paper
were first announced by the authors in the extended abstract [1].
REFERENCES
1. J. Button, M. Chiodo, M. Zeron-Medina Laris, Coset intersection graphs, and transversals as generat-
ing sets for finitely generated groups, Edited by J. González-Meneses et al., Extended Abstracts Fall
2012, Trends in Mathematics, Vol. 1, 29–34, Birkhäuser, Basel, 2014, http://dx.doi.org/10.1007/
978-3-319-05488-9_5.
2. H. Chapman, A note on the elementary theory of groups of finite order, Messenger of Math. 42 (1913)
132–134.
3. , On a note on the elementary theory of groups of finite order, Messenger of Math. 43 (1914) 85.
4. P. Hall, On representatives of subsets, J. Lond. Math. Soc. 10 (1935) 26–30.
5. W. Lederman, Introduction to Group Theory. Oliver and Boyd, London, 1973.
6. G. Miller, On a method due to Galois, Quart. J. Math. Oxford Ser. 41 (1910) 382–384.
7. O. Ore, On coset representatives in groups, Proc. Amer. Math. Soc. 9 no. 4 (1958) 665–670.
8. G. Scorza, A proposito di un teorema del Chapman, Boll. Unione Mat. Ital. 6 (1927) 1–6.
9. S. Shü, On the common representative system of residue classes of infinite groups, J. Lond. Math. Soc.
16 (1941) 101–104.
10. H. Zassenhaus, Lehrbuch der Gruppentheorie (German). Teubner, Berlin, 1937.
Mathematics Department, University of Neuchâtel, Rue Emile-Argand 11, Neuchâtel, 2000, Switzerland
maurice.chiodo@unine.ch
926
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
Complex Descartes Circle Theorem
Sam Northshield
Abstract. We present a short proof of the Descartes circle theorem on the “curvature-centers”
of four mutually tangent circles. Key to the proof is associating an octahedral configuration of
spheres to four mutually tangent circles. We also prove an analog for spheres.
It can be traced back to at least Descartes that four mutually tangent circles have cur-
vatures (reciprocals of radii) satisfying the relation
This M ONTHLY has published several papers concerning this fascinating topic: [1, 3,
4, 5, 6]. It was only in 2001 [3] that is was noticed, and proved, that the “curvature-
centers” (curvature times center where the center is considered a complex number)
satisfy the same relation. We present a short proof of this result (Theorem 1) and an
analogous version for spheres (Corollary 1).
For the purposes of this paper, a sphere will always be contained in the half-space
C × [0, ∞) and be tangent to the complex plane. Let S(z, r ) denote the sphere with
radius r tangent to C at z. It is obvious that S(z, r ) and S(w, s) are tangent to each
other if and only if
|z − w|2 = 4r s.
It is also immediate that given any three points z 1 , z 2 , z 3 ∈ C, there are unique numbers
r1 , r2 , r3 such that the spheres S(z i , ri ) are mutually tangent. In particular, if {i, j, k}
= {1, 2, 3} then
|z i − z j | · |z i − z k |
ri = . (2)
2|z j − z k |
We say that two circles are orthogonal if they intersect at right angles; see Figure 1.
4
(a) k1 k2 = = c12 + c22
|w1 − w2 |2
and
4w1 w2
(b) k1 k2 w1 w2 = = c12 z 12 + c22 z 22 .
|w1 − w2 |2
http://dx.doi.org/10.4169/amer.math.monthly.121.10.927
MSC: Primary 52C26
2|w2 − w| 2|w1 − w| 4
k1 k2 = = .
|w2 − w1 ||w − w1 | |w2 − w1 ||w − w2 | |w1 − w2 |2
r12 + r22 |z 1 − z 2 |2 4
c12 + c22 = = =
(r1r2 ) 2 |z 1 − z 2 |2 |w1 − w2 |2 /4 |w1 − w2 |2
W1
r1 r2
a
Z1 Z2
Z
W2
928
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
and thus
(a) ki j = ci + c j , and
(b) c2 z 2 = c1 z 1 c2 z 2 + c1 z 1 c3 z 3 + c2 z 2 c3 z 3 .
√
Proof. By Coxeter [2, p. 15], c = c1 c2 + c1 c3 + c2 c3 . By Lemma 1,
ki j z i j = ci z i + c j z j .
By Lemma 1b,
and therefore
c2 z 2 = c1 z 1 c2 z 2 + c1 z 1 c3 z 3 + c2 z 2 c3 z 3
Given four mutually tangent circles, there are six points of tangency between them
(see Figure 2). At each such point z, assign a sphere tangent to the plane at z with
curvature equal to the sum of the curvatures of the two circles that meet at z. By
Lemma 2(a), the six spheres have disjoint interiors and any two are tangent to each
other if and only if their points of tangency to the plane are on the same circle. We will
refer to this collection of spheres as the octahedral arrangement of spheres associated
with the four circles since the adjacency graph of these six circles forms an octahedron.
A key idea for the proof of the main theorem is that the respective octahedral ar-
rangements of spheres associated to {Ci } and to {Ci } are the same. We now show the
complex Descartes circle theorem.
C'1
C4 C3
Figure 2. Four circles and their duals; three circles and their duals are labeled.
Theorem 1. Given four mutually tangent circles with curvatures ci and respective
centers z i ,
2
ci z i =2 ci2 z i2 .
2(w1 + w2 + w3 − w4 ) = K 12 + K 34 + K 12 + K 13 + K 23 − K 14 − K 24 − K 34
= K 34 + K 12 + K 34 + K 24 + K 14 − K 23 − K 13 − K 12
√
= 4w4 = 4σ w1 w2 + w1 w3 + w2 w3 .
It follows that
√
w4 = w1 + w2 + w3 + 2σ w1 w2 + w1 w3 + w2 w3
and the result follows from the fact that equation (1) is equivalent to
√
d = a + b + c ± 2 ab + ac + bc.
930
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
The complex Descartes circle theorem is a true generalization of the Descartes circle
theorem since replacing z i by (z i + z)/z in the formula and taking the limit as z goes
to infinity gives the old version.
The spheres in an octahedral arrangement associated with four mutually tan-
gent circles obeys a similar formula (the proof of which follows immediately from
Theorem 1).
⎛ ⎞2
(b) 2⎝ ki j z i j ⎠ = 9 ki2j z i2j .
i< j i< j
REFERENCES
1. H. S. M. Coxeter, The problem of Apollonius, Amer. Math. Monthly 75 (1968) 5–15, http://dx.doi.
org/10.2307/2315097.
2. H. S. M. Coxeter, Introduction to Geometry. Second edition, John Wiley and Sons, New York, 1969.
3. J. Lagarias, C. Mallows, A. Wilks, Beyond the Descartes circle theorem, Amer. Math. Monthly 109 (2002)
338–361, http://dx.doi.org/10.2307/2695498.
4. D. Pedoe, On a theorem in geometry, Amer. Math. Monthly 74 (1967) 627–640.
5. P. Sarnak, Integral Apollonian packings, Amer. Math. Monthly 118 (2011) 291–306, http://dx.doi.
org/10.4169/amer.math.monthly.118.04.291.
6. J. B. Wilker, Four proofs of a generalization of the Descartes circle theorem, Amer. Math. Monthly 76
(1969) 278–282, http://dx.doi.org/10.2307/2316373.
see for example [2, p. 298]. Clearly (a, |N |) = 1 so, by Linnik’s theorem [5], there
are absolute constants C > 0 and L > 0 such that the least prime p in the arithmetic
progression
{|N |k + a : k = 0, 1, 2, . . .}
satisfies
p ≤ C|N | L .
http://dx.doi.org/10.4169/amer.math.monthly.121.10.932
MSC: Primary 11E25, Secondary 11E12
932
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
Since p belongs to this arithmetic progression, we have p ≡ a (mod |N |) and, by [2,
Lemma 2.3, p. 291] we deduce
N N
= = −1.
p a
N
Finally, as N ≡ 0 (mod 4) and = −1, we see that p =
2.
p
Remark. By a deep result of Xylouris [8], one has L ≤ 5.2. Xylouris’s work is a
refinement of that of Heath-Brown [4], who showed that L satisfies L ≤ 5.5.
If p|ac, then
b2 − 4ac b2
−1 = = =0 or 1,
p p
kt = 1 + up 2 , 1 ≤ t < p2 , 0 ≤ u < k.
We now construct an integer n with 1 ≤ n < C 3 |N |3L such that p|(kn + ) and
p2 (kn + ).
then p|(2ax + by) so p|2ax. But p = 2 and p a hence p|x. Therefore, p2 divides
ax 2 + bx y + cy 2 = kn + , contradicting p2 (kn + ). This completes the proof that
the integer kn + is not represented by Q.
934
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
The least length of a nontrivial arithmetic progression is 3. Does every nonzero in-
tegral binary quadratic form represent an arithmetic progression of length 3? Two deep
results of Weber [7] and Green [3] positively answer the above question in the partic-
ular case when the integral binary quadratic form is positive-definite. In 1882, Weber
[7] proved that if Q is a primitive integral binary quadratic form which is positive-
definite, then the set of primes that are represented by Q has positive relative density.
In 2006 Green [3] proved that any subset of primes having positive relative density has
a 3-term arithmetic progression. Thus, by putting these two deep results together, we
see that every primitive, positive-definite, integral binary quadratic form represents a
3-term arithmetic progression. In this paper we shall prove in an elementary way that
any nonzero integral, binary quadratic form represents a nontrivial arithmetic progres-
sion of length 3 infinitely often.
and
ACKNOWLEDGMENTS. We are grateful to the referees for going through the paper very carefully and
modifying it to a much nicer form.
1. A. Alaca, Ş. Alaca, K. S. Williams, Arithmetic progressions and binary quadratic forms, Amer. Math.
Monthly 115 (2008) 252–254.
2. R. Ayoub, An Introduction to the Analytic Theory of Numbers. Mathematical Surveys, Number 10, Ameri-
can Mathematical Society, Providence, Rhode Island, 1963, http://dx.doi.org/10.1090/surv/010.
3. B. Green, Roth’s theorem in the primes, Ann. Math. 161 (2005) 1609–1636,
http://dx.doi.org/10.4007/annals.2005.161.1609.
4. D. R. Heath-Brown, Zero-free regions for Dirichlet L-functions and the least prime in an
arithmetic progression, Proc. London Math. Soc. (3) 62 (1992) 265–338,
http://dx.doi.org/10.1112/plms/s3-64.2.265.
5. Y. V. Linnik, On the least prime in an arithmetic progression I . The basic theorem, Rec. Math. (Mat.
Sbornik) N. S. 15 no. 57 (1944) 139–178.
6. L. J. Mordell, Diophantine Equations. Pure and Applied Mathematics, Vol. 30, Academic Press, London,
1969.
7. H. Weber, Beweis des Satzes, daB jede eigentlich primitive quadratische Form un endliche viele
Primzahlen darzustellen fa̋hig ist, Math. Ann. 20 (1882) 301–329.
8. T. Xylouris, Über die Linniksche Konstante, Diplomarbeit, Universität Bonn, 2009. (arXiv:0906.2749v1
[math.NT] 15 Jun 2009).
936
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
Cosines and Cayley, Triangles and Tetrahedra
Marshall Hampton
Abstract. This article surveys some of the more aesthetically appealing and useful formulas
relating distances, areas, and angles in triangles and tetrahedra. For example, a somewhat ne-
glected trigonometric identity involving only the cosines of a triangle is an instance of the
famous Cayley cubic surface. While most of these formulas are well known, some novel iden-
tities also make an appearance.
in which ri j is the distance between vertex i and vertex j, and s is the semiperimeter
rAB + rAC + rBC
s= .
2
Almost two thousand years later, a form of Heron’s formula was found that gen-
eralizes to simplices of any dimension. This is the Cayley–Menger determinant; for a
triangle:
⎛ ⎞
0 1 1 1
⎜ 1 0 r2 r2 ⎟
⎜ AC ⎟
−162ABC = det ⎜ AB
2 ⎟
⎝ 1 rAB2
0 rBC ⎠
2 2
1 rAC rBC 0
= −(−rAB + rAC + rBC )(rAB − rAC + rBC )(rAB + rAC − rBC )
× (rAB + rAC + rBC ).
The matrix in the determinant is called the Cayley–Menger matrix. Cayley found
the determinantal form [7]—the polynomial itself was known earlier, by Lagrange.
Menger discovered a number of further properties of the matrix and closely related
variants of it [14]. For an n − 1-dimensional simplex of n vertices A1 , . . . An , the
volume formula generalizes to
⎛ ⎞
0 1 1 ... 1
⎜ 1 0 r A2 1 A2 ... r A2 1 An ⎟
(−1)n+1 ⎜ ⎟
⎜ ⎟
2A1 ...An = n det ⎜ 1 r A2 1 A2 0 ... r A2 2 An ⎟. (2)
2 (n!)2 ⎜ ⎟
⎝ ... ... ... ... ... ⎠
1 r A2 1 An r A2 2 An ... 0
The definitive article about the Cayley–Menger matrix is [5], and its properties
are also nicely summarized in [4]. The use of distances as coordinates is masterfully
http://dx.doi.org/10.4169/amer.math.monthly.121.10.937
MSC: Primary 51K05, Secondary 52-02
Laws of cosines. For brevity in some of the more complex formulae, we write cABC
for the cosine of the angle ABC.
The classic law of cosines is
2
rAB − rAC
2
− rBC
2
+ 2rACrBC cACB = 0.
One derivation of the law of cosines makes use of the relations of the form
which can be used for many purposes because of their linearity in the distances
and cosines. These relations generalize to the n-dimensional simplex through the
polyhedral version of the divergence theorem. If we choose a vector field to be the
outward-pointing normal ni to the ith facet of a bounded convex polyhedron, then
the divergence of the field is 0. This is equal to the surface integral
0= ni · nj j = (i) − cij j (4)
j
where ci j is the cosine of the angle between facets i and j, and i is the area of the
ith facet. Multiplying this equation by i and subtracting j times the corresponding
equation for face j gives
i2 = 2j − 2 cjk j k . (5)
j=i j,k=i
While equation (5) has the same form as the law of cosines for triangles, it seems much
less useful since more than one cosine is involved in each equation. The lower degree
conditions in (4) are usually a better starting point for other identities.
which are complementary to the law of cosines in that they express a single distance in
terms of the cosines. Of course, these formulae break down if the triangle is collinear.
938
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
In order to have a nontrivial kernel, the determinant must vanish, which gives us the
beautiful relation:
⎛ ⎞
−1 cBAC cABC
det ⎝ cBAC −1 cACB ⎠ = cACB
2
+ cABC
2
+ cBAC
2
+ 2cACB cABC cBAC − 1 = 0. (6)
cABC cACB −1
This trigonometric identity has been known for a long enough time that it is difficult
to determine its first appearance. But it does not seem well known, and it is not usually
cited as an example of the famous Cayley cubic, a surface with the maximal number
(four) of isolated singular points (where both the surface and the gradient of its defining
function vanish). Perhaps we should call it the Cayley cosine cubic.
The four singular points of the Cayley cosine cubic are
(c123 , c132 , c213 ) ∈ {(1, 1, −1), (1, −1, 1), (−1, 1, 1), (−1, −1, −1)}.
The first three of these points correspond to collinear triangles, while the last one is
unrealizable as a triangle. However, it is consistent algebraically with a triangle whose
edge lengths add up to zero (rAB + rAC + rBC = 0). Intriguingly, this is the fourth factor
in the Cayley–Menger determinant as well.
The Cayley cosine cubic contains six lines that intersect the planes ci jk = ±1. The
three lines with ci jk = 1 are the boundary of the portion of the surface corresponding
to triangles with positive distances.
This formula may relate to a rigidity result [15] on tetrahedra, which states that if
all of the dihedral angles of a tetrahedron are less than or equal to the corresponding
angles of another tetrahedron, then the two tetrahedra must be similar.
The dihedral angles can be computed in terms of facial cosines at a vertex. For each
dihedral angle, there are two choices for the vertex. For instance, the cosine of the
dihedral angle between faces A and B can be expressed as
from which we can eliminate the sines by squaring and substituting the Pythagorean
identity to obtain a relation between the facial cosines:
(1 − cACD
2
)(1 − cBCD
2
)(cADB − cADC cBDC )2
= (1 − cADC
2
)(1 − cBDC
2
)(cACB − cBCD cACD )2 . (7)
For each of the six edges of the tetrahedron, we have such an equation. These cannot
be independent since the space of similarity classes of tetrahedra is five-dimensional.
Somewhat similarly, the law of sines can be used once for each triangular face to
obtain the identity
which could be converted into a cosine identity by squaring both sides. This is some-
what unsatisfactory, however, since it involves eight angles.
If a geometric problem can be cast into polynomial form, the computation of a
Gröbner basis (or bases) provides an automated path for eliminating variables and
obtaining new or simpler relations [6] (caveat emptor: in practice, many Gröbner bases
require excessive memory and computational time to compute). It is beyond the scope
of this article to describe Gröbner bases in full. They are analogous to the reduction of
a linear system to echelon form (Gaussian elimination), but for polynomial (nonlinear)
systems. For more background on Gröbner bases, see [9].
By computing a Gröbner basis for the system of equations (4) for a tetrahedron
(using Singular [10]), we found a fairly simple condition on the six angles bordering
one face. The six angles are on three faces but do not include any of the three angles
meeting at the common vertex. For a common vertex A, this condition is
2
cACB 2
cADC 2
cABD − cADB
2 2
cABC 2
cACD + cADB
2 2
cABC − cACB
2 2
cADC
− cACB
2 2
cABD − cADC
2 2
cABD + cADB
2 2
cACD + cABC
2 2
cACD + cACB
2
− cADB
2
− cABC
2
+ cADC
2
+ cABD
2
− cACD
2
= 0.
940
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
If we use the squares as variables, i.e., let qi jk = ci2jk , then there are nine singular
planes corresponding to collinear configurations, along with the origin. This singular
point at the origin could be interpreted as a projective closure having point A at infinity.
This can be written in a nicer form:
⎛ ⎞ ⎛ ⎞
1 1 1 1 1 1 1 1
⎜ 1 2 2 ⎟ ⎜ 1 2 2 ⎟
⎜ 1 cABC cABC ⎟ ⎜ 1 cACB cACB ⎟
det ⎜ ⎟ = det ⎜ ⎟. (8)
⎝ 1 cACD2
1 2
cACD ⎠ ⎝ 1 cADC 2
1 2
cADC ⎠
2 2 2 2
1 cADB cADB 1 1 cABD cABD 1
ACKNOWLEDGMENT. The author thanks the reviewers of this article as well as Gareth Roberts and
Richard Moeckel for their comments and suggestions, which have improved it in many ways.
REFERENCES
1. A. Albouy, A. Chenciner, Le problème des n corps et les distances mutuelles, Invent. Math. 131 no. 1
(1997) 151–184.
2. H. Alexandrinus, Metrika. ca. 60.
3. D. Audet, Déterminants sphérique et hyperbolique de Cayley-Menger, Bulletin AMQ 51 (2011) 45–52.
4. M. Berger, Geometry I. Translated from the French by M. Cole and S. Levy, Springer, New York, 1987.
5. L. M. Blumenthal, B. E. Gillam, Distribution of points in n-space, Amer. Math. Monthly 50 (1943) 181–
185, http://dx.doi.org/10.2307/2302400.
6. B. Buchberger, F. Winkler, Gröbner Bases and Applications. Lecture note series. Cambridge Univ. Press,
Cambridge, 1998, http://dx.doi.org/10.1017/cbo9780511565847.
7. A. Cayley, On a theorem in the geometry of position, Camb. Math. J. 2 (1841) 267–271, http://dx.
doi.org/10.1017/cbo9780511703676.002.
8. A. Chenciner, The “form” of a triangle, Rend. Mat. Serie VII 27 (2007) 1–16.
9. D.A. Cox, J. Little, D. O’Shea, Ideals, Varieties, and Algorithms: An Introduction to Computational
Algebraic Geometry and Commutative Algebra. Springer, New York, 2007.
10. W. Decker, G.-M. Greuel, G. Pfister, H. Schönemann, SINGULAR 3-1-5—A computer algebra system
for polynomial computations, (2012), http://www.singular.uni-kl.de.
11. A. W. M. Dress, T. F. Havel, Distance geometry and geometric algebra, Found. Phys. 23 (1993), 1357–
1374, http://dx.doi.org/10.1007/bf01883783.
12. T. F. Havel, Some examples of the use of distances as coordinates for euclidean geometry, J. Symbolic
Comput. 11 (1991) 579–593, http://dx.doi.org/10.1016/s0747-7171(08)80120-4.
13. D. A. Klain, An intuitive derivation of Heron’s formula, Amer. Math. Monthly 111 (2004) 709–712,
http://dx.doi.org/10.2307/4145045.
14. K. Menger, Untersuchungen über allgemeine Metrik, Math. Ann. 100 (1928) 75–163.
15. I. Rivin, J. H. Lindsey II, A similarity criterion: 10462, Amer. Math. Monthly 105 no. 7 (1998) 671.
Abstract. We give a proof of the spectral theorem for self-adjoint operators in the finite-
dimensional real case that involves no complexification of the space, no determinants, and no
Lagrange multipliers. The tools in our proof are the fundamental theorem of algebra and the
intermediate value theorem.
1. THE THEOREM.
Definition. Let V be a real inner product space. We denote the inner product of u and
v by u, v. A linear transformation A : V → V is self-adjoint if
Av = λv,
2. ABOUT THE PROOF. The proof is carried out by induction on the dimension of
V. The clue is in the following lemma.
Lemma. Let V be an inner product space and let the linear transformation A : V →
V be self-adjoint. Suppose that the subspace E of V is invariant under A (that means
A(E) ⊂ E). Then the orthogonal complement of E,
E ⊥ = {u ∈ V| u, v = 0 ∀v ∈ E} ,
Proof. Let us suppose A(E) ⊂ E and let u ∈ E ⊥ . Since Av ∈ E, we have, for all
v ∈ E,
http://dx.doi.org/10.4169/amer.math.monthly.121.10.942
MSC: Primary 15A18
942
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
Au, v = u, Av = 0.
This proves Au ∈ E ⊥ .
Suppose now the theorem holds in the n-dimensional case. It follows from the
lemma that we can prove the theorem in the (n + 1)-dimensional case if we can prove
the existence of an eigenvector v. This follows since the subspace E spanned by v is
invariant under A and the theorem applies to the n-dimensional space E ⊥ . The proof
of the existence of the eigenvector v can use the complexification of V or Lagrange
multipliers, searching for critical points of
1
f (u) = Au, u
2
on the unit sphere of V [1, pp 106 and 113].
Our alternative strategy requires that we begin with the one- and two-
dimensional cases (dimension 2 requires the intermediate value theorem). Then we
prove that, in the general case, there always exists an invariant subspace of dimension
1 or 2. We will avoid, under all circumstances, the use of determinants and, a fortiori,
the characteristic polynomial of A will not even be mentioned (despite the algebraic
flavor of the final part of the proof).
Theorem (dimension 2). Let V be a two-dimensional real inner product space and
let A : V → V be a self-adjoint linear map. Then there is a set of eigenvectors of A
that form an orthonormal basis of V.
p(A)v = ak Ak v + · · · + a1 Av + a0 v.
Proof. Let I be the set of monic polynomials q such that q(A) = 0. We take p in I of
minimal degree. Since V is nontrivial, the degree of p must be at least one. It follows
from the fundamental theorem of algebra that p factors as a product of polynomials of
degree at most two. So, we can decompose p as a product of two monic polynomials,
p(x) = a(x)g(x),
in which the degree of a equals one or two. We study these two possibilities.
(i) If the degree of a is one, we may suppose that a(A) = A − λI . Since g(A) = 0
(otherwise, p would not have minimal degree), there is a nonzero vector v in the
range of g(A). Consequently, a(A)v = (A − λI )v = 0, so v is an eigenvector of
A. Let E be the space spanned by v.
(ii) If the degree of a is two, we may suppose that a(A) = A2 + b A + cI . As in
the previous case, the range of g(A) must contain a nonzero vector v. If v is an
eigenvector of A, we are done. If not, then v and Av span a two-dimensional
subspace E. But, since a(A)v = 0, we get A(Av) = −b Av − cv, and this shows
that E is invariant under A.
Theorem. Let V be a finite-dimensional real inner product space and let the linear
transformation A : V → V be self-adjoint. Then there is a set of eigenvectors of A that
form an orthonormal basis of V.
The real Schur form [2, p. 341] also follows from Proposition 2.
944
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
Proposition 3 (The real Schur form). Every real square matrix M can be factored
as
M = QQ T ,
We must prove the existence of an orthonormal basis γ of IR n such that the matrix of A
relative to γ is a block upper triangular matrix in which each diagonal block is 1 × 1 or
2 × 2. We argue inductively on the order n of M. By Proposition 2, there exists a one-
or a two-dimensional subspace E of IR n such that A(E) ⊂ E. Let α be an orthonormal
basis of E. Define B : E ⊥ → E ⊥ by Bx = P Ax, where P is the orthogonal projec-
tion onto E ⊥ . By the induction hypothesis, there exists an orthonormal basis β of E ⊥
such that the matrix of B relative to β is a block upper triangular matrix in which
each diagonal block is 1 × 1 or 2 × 2. The matrix of A relative to the (ordered) basis
γ = α ∪ β (starting with the elements of α) is a block upper triangular matrix in which
each diagonal block is 1 × 1 or 2 × 2
REFERENCES
1. P. D. Lax, Linear Algebra and its Applications. Second edition. Wiley-Interscience, Hoboken, NJ, 2007.
2. G. H. Golub, C. F. van Loan, Matrix Computations. Third edition. Johns Hopkins Univ. Press, Baltimore,
MD, 1996.
PROBLEMS
11803. Proposed by Sam Speed, Germantown, PA. Let a1 (k, n) = (9k (24n + 5) −
5)/8, a2 (k, n) = (9k (24n + 13) − 5)/8, a3 (k, n) = (3 · 9k (24n + 7) − 5)/8, and
a4 (k, n) = (3 · 9k (24n + 23) − 5)/8. Show that for each nonnegative integer m
there is a unique integer triple ( j, k, n) with j ∈ {1, 2, 3, 4} and k, n ≥ 0 such that
m = a j (k, n).
11804. Proposed by George Stoica, University of New Brunswick, Saint John, Canada.
Prove that 10|x 3 + y 3 + z 3 − 1| ≤ 9|x 5 + y 5 + z 5 − 1| for real numbers x, y, and z
with x + y + z = 1. When does equality hold?
11805. Proposed by Gleb Glebov, Simon Fraser University, Burnaby, Canada.
(a) Show that
√
∞
(−1)k
∞
(−1)k 5π 3 3
+ =
k=0
(3k + 1)3 k=0 (3k + 2)3 243
and
∞
(−1)k
∞
(−1)k 13
− = ζ (3).
k=0
(3k + 1)3 k=0
(3k + 2) 3 18
946
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
11806. Proposed by István Mező, Nanjing University of Information Science and Tech-
nology, Nanjing, China. Prove that
2π x ∞
log n
log ecos x sin(x + sin x) d x = (e − 1)(log(2π) + γ ) + .
0 2π n=2
n!
Here denotes the gamma function and γ denotes the Euler–Mascheroni constant.
11807. Proposed by Robin Oakapple, Albany, OR. Given a quadrilateral ABCD in-
scribed in a circle K , and a point Z inside K , the rays AZ, BZ, CZ, and DZ meet K
again at points E, F, G, and H , respectively, to yield another quadrilateral also in-
scribed in K . Develop a construction that takes as input A, B, C, and D and returns
a point Z such that this second quadrilateral has (at least) three of its sides of equal
length.
11808. Proposed by D. M. Bătineţu-Giurgiu, “Matei Basarab” National College,
Bucharest, Romania, and Neculai Stanciu, “George Emil Palade” School, Buzău,
Romania. Let be the gamma function. Compute
(n!)−1/n
lim n 2
(nx) d x.
n→∞ ((n+1)!)−1/(n+1)
11809. Proposed by Omran Kouba, Higher Institute for Applied Science and Technol-
ogy, Damascus, Syria. Let an
be a sequence of real numbers.
(a)
∞Suppose √ that an
consists of nonnegative√ numbers and is nonincreasing, and
∞ n
n=1 an / n converges. Prove that n=1 (−1) an converges.
(b) Find√ a nonincreasingsequence a√ n
of positive numbers such that
limn→∞ nan = 0 and ∞ n=1 (−1) n
a n diverges.
11795. Proposed by Mircea Merca, University of Craiova, Craiova, Romania. Let p be
the partition counting function on the set Z+ of positive integers, and let g be the func-
tion on N given by g(n) = 12
n/2 ,
(3n + 1)/2. Let A(n) be the set of nonnegative
integer triples (i, j, k) such that g(i) + j + k = n. Prove for n ≥ 1 that
1
p(n) = (−1)
i/2−1 g(i) p( j) p(k).
n (i, j,k)∈A(n)
SOLUTIONS
948
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
money wins, the path reaches a new high and the money is again split k + 1 to k, but
the two gamblers interchange roles.
In the classical problem starting with $a and $b, it is well known that the expected
number of steps until one gambler is ruined is ab, in this case k(k + 1). We have
shown that this also is the expected number of steps until the path is first k steps below
its all-time high.
Also solved by M. Andreoli, W. Barta, R. Chapman (U. K.), C. Delorme (France), S. J. Herschkorn, G. Lau
(U. K.), O. P. Lossers (Netherlands), H. M. Mahmoud, R. Martin (Germany), I. Pinelis, M. A. Prasad (India),
R. Pratt, R. Tauraso (Italy), M. Wildon (U. K.), and the proposer.
Equality can hold if and only if f (x) = a a.e. on [0, ξ ] and f (x) = b a.e. on [ξ, 1].
This can happen only if f is discontinuous at ξ , so the inequality is strict for all f ∈
S. On the other hand, this upper bound can be approached as closely as we like by
choosing ε small and positive and taking f (x) = a for 0 ≤ x ≤ ξ − ε, f (x) = b for
[ξ + ε, 1], and f linear on the interval [ξ − ε, ξ + ε].
Also solved by. K. F. Andersen (Canada), R. Bagby, P. Bracken, R. Chapman (U. K.), S. J. Herschkorn, B.
Karaivanov, O. Kouba (Syria), J. C. Linders (Netherlands), J. H. Lindsey II, O. P. Lossers (Netherlands), I.
Pinelis, Á. Plaza (Spain), A. Stenger, R. Stong, E. I. Verriest, S. V. Witt, GCHQ Problem Solving Group
(U. K.), TCDmath Problem Group (Ireland), and the proposer.
Solution by Roberto Tauraso, Università di Roma “Tor Vergata,” Rome, Italy. Let P(x)
be the generating function for integer partitions,
∞
∞
1
P(x) = p(n)x = n
.
n=0 n=1
1 − xn
∞
1 ∞
1
(x 2 − x 3 ) + (1 + x + x 2
) .
n=3
1−x n
n=4
1 − xn
950
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
A Powered Gamma Limit
11676 [2012, 801]. Proposed by D. M. Bătineţu-Giurgiu, “Matei Basarab” National
College, Bucharest, Romania, and Neculai Stanciu, “George Emil Palade” Secondary
School, Buzau, Romania. For real t, find
2
2 2
lim x sin t (x + 2)(cos t)/(x+1) − (x + 1)cos t/x .
x→∞
L(a, b) = lim x sin t (x + a)cos t/(x+a−1) −(x + b)cos t/(x+b−1) = (a −b) f (t).
x→∞
Assume that the limit exists (applying L’Hopital’s rule at the end of the calculation
verifies that it does). Apply Stirling’s formula to both (x + a) and (x + b) to obtain
cos2 t
cos2 t
2 x + a − 1 x + b − 1
L(a, b) = lim x sin t −
x→∞ e e
2 2
2 2
= lim e− cos t x 1−cos t (x + a − 1)cos t − (x + b − 1)cos t
x→∞
2
2
x + a − 1 cos t x + b − 1 cos t
−
− cos2 t e e
=e lim .
x→∞ 1/x
This is an indeterminate limit that can be evaluated using l’Hopital’s rule:
2
2
a − 1 − sin t b − 1 − sin t
L(a, b) = f (t) lim (a − 1) 1 + − (b − 1) 1 +
x→∞ x x
2
= f (t) (a − 1) − (b − 1) = (a − b) cos2 t e− cos t .
2
For the case in the problem as stated, a = 2 and b = 1 so L = e− cos t cos2 t.
Also solved by K. F. Andersen (Canada), R. Boukharfane (Canada), P. Bracken, R. Chapman (U. K.), H. Chen,
P. P. Dályay (Hungary), A. Ercan (Turkey), D. Fleischman, C. Georghiou (Greece), O. Geupel (Germany),
M. L. Glasser, J.-P. Grivaux (France), O. Kouba (Syria), K.-W. Lau (China), J. Li, O. P. Lossers (Netherlands),
H. M. Mahmoud, G. Martin (Canada)R. Nandan, M. Omarjee (France), P. Perfetti (Italy), I. Pinelis, R. Stong,
D. B. Tyler, GCHQ Problem Solving Group (U. K.), and the proposers.
Solution by Radouan
√ √ Boukharfane, Polytechnique Montréal, Montreal, Canada. The
π 3/18 4
answer is e / 3. We use the Dedekind η function defined for a complex number
t with positive imaginary part by
πit
∞
η(t) = e 12 1 − e2πint .
n=1
n=1
∞
∞
n=1 1 − e
6πint πit
e− 4 η(3t)
= 1+e 2πint
+e 4πint
= ∞ = πit
n=1 1 − e e− 12 η(t)
2πint
n=1
√
π 3
η(−1/t) π √ e 18
− πit √
=e 6 = e 6 3 −it = √ .
η(t) 4
3
Editorial comment. Unfortunately the problem appeared with typos, making the prod-
uct divergent. Solutions showing divergence were also accepted.
Also solved by G. Apostolopoulos (Greece), R. Chapman (U. K.), D. Fleischman, O. Geupel (Germany),
M. Omarjee (France), R. Stong, R. Tauraso (Italy), and the proposer.
yi, j = xi. j − xi−1, j = (xi−1. j + xi, j−1 ) − (xi−2, j + xi−1, j−1 ) = yi−1, j + yi, j−1 ,
so Y is an NW-matrix. The matrix Z obtained from Y by deleting the first row and
column is a unit NW-matrix; by the induction hypothesis, det Z = 1. Expanding the
determinant of Y along the first column yields det(X ) = det(Y ) = det(Z ) = 1.
Clearly det A1 = −1, so choose n ≥ 2. Obtain B from An by leaving the first two
rows unchanged and subtracting from each subsequent row the two rows immediately
above it; note that det B = det An . For i ≥ 3 and j ≥ 1,
bi, j − bi, j−1 = (ai, j − ai−1, j − ai−2, j ) − (ai, j−1 − ai−1, j−1 − ai−2, j−1 )
= ai−1, j − ai−2, j − ai−3, j = bi−1, j .
952
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
The matrix B takes the form
⎡ ⎤
0 1 0 ···
⎢1 2 3 · · ·⎥
⎢ ⎥
⎢0 0 2 · · ·⎥
⎢0 · · ·⎥
⎢ 0 2 ⎥.
⎢. .. .. .. ⎥
⎣ .. . . .⎦
0 0 2 ···
Expanding det B along the first two columns yields det B = − det(2W ), where W is a
unit NW-matrix of order n − 1. By the claim, det An = −2n−1 det(W ) = −2n−1 .
Editorial comment. Sergio Falcón and Ángel Plaza observed that the problem and its
solution appear as an example in A. R. Moghaddamfar, S. M. H. Pooya, General-
ized Pascal triangles and Toeplitzmatrices,
Electron. J. Lin. Alg. 18 (2009), 564–588.
Ionin’s matrix W , with i, j-entry i+i j , is shown to have determinant 1 via four proofs
in A. Edelman and G. Strang, Pascal Matrices, this MONTHLY 111 (2004), 189–197;
an earlier such proof appears in C. A. Rupp, Problem 3468, this MONTHLY 37 (1930),
552 (solution by H.T.R. Aude, 38 (1931), 355).
Also solved by D. Beckwith, R. Chapman (U. K.), P. P. Dályay (Hungary), C. Delorme (France), S. Falcón
& Á. Plaza (Spain), O. Geupel (Germany), J. P. Grivaux (France), E. A. Herman, B. Karaivanov, O. Kouba
(Syria), O. P. Lossers (Netherlands), M. Omarjee (France), R. E. Prather, C. P. Rupert, R. Stong, R. Tauraso
(Italy), J. van Hamme (Belgium), Armstrong Problem Solvers, GCHQ Problem Solving Group (U. K.), and
the proposer.
Solution by Traian Viteam, Punta Arenas, Chile. For n = 2 the inequality reduces to
4 > 2/e, which is trivial. For 2 < k ≤ n, the AM–GM inequality implies that
1 1 ak ak k kk ak2
(1 + ak )k = + ··· + + + ≥ .
k−2 k−2 2 2 (k − 2)k−2 4
Multiplying (1 + a2 )2 > a22 together with these inequalities for k ∈ {3, ..., n} yields
1 n−1 n 2n−1
n
(n − 1)n−1 n n a22 · · · an2
(1 + ak ) >
k
=2 1− ,
k=2
22 4n−2 n 2
n
since i=2 ai = 1 is assumed. Using e x ≥ 1 + x with x = n−1 1
, it follows that
n−1
1 1 −(n−1)
e ≥ 1+ = 1−
n−1 n
and hence (1 − n1 )n−1 ≥ e−1 . Substituting this inequality into the product inequality
above yields the stated result.
Also solved by G. Apostolopoulos (Greece), R. Boukharfane (Canada), E. Eyeson, D. Fleischman, N. Grivaux
(France), S. Kaczkowski, O. Kouba (Syria), O. P. Lossers (Netherlands), R. E. Prather, D. B. Tyler, GCHQ
Problem Solving Group (U. K.), and the proposer.
954
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
universal covers for “good” metric spaces, referring the reader to Hatcher for a more
formal treatment. Allen Hatcher’s book [3] is, in my mind, the definitive text for this
material. Anyone (strong undergraduate or graduate student) with plans to use the
fundamental group later in their work should be pointed here.
At this point I should emphasize that this book is very different from other under-
graduate books which concern surfaces. Armstrong [1], for example, is for a much
different kind of course. The two books cover some of the same material, namely the
fundamental group and classification of surfaces. However, Armstrong [1] is much
more traditional, and less concerned about reaching research material. The same goes
for most books (or course notes) on Riemann surfaces. In fact, one of the things that
makes Mostly Surfaces so worthwhile is that the material (in totality) and point of view
are not presented elsewhere at an (even allegedly) undergraduate level.
The text that most reminds me of Mostly Surfaces is a book by Francis Bona-
hon, Low-Dimensional Geometry: From Euclidean Surfaces To Hyperbolic Knots, [2],
which is also published under the AMS Student Mathematical Library. There is some
material overlap with Mostly Surfaces: they both cover hyperbolic, Euclidean and
spherical surfaces, and gluing constructions. More significantly, Bonahon [2] is very
similar in spirit: it starts with surfaces, with a decidedly geometric point of view, and
then uses this to launch into research level material. Bonahon [2] goes in the direc-
tion of hyperbolic 3-manifolds, whereas Mostly Surfaces delves into different struc-
tures on surfaces, such as flat cone surfaces and hyperbolic structures. Both books
have been used by my colleagues in successful topics courses. I can easily imagine
teaching courses at a variety of levels based on either, or both, of these books. For
example, a little bit of Chapter 1 along with Part 1: “Surfaces and Topology” and Part
2: “Surfaces and Geometry” of Mostly Surfaces would make a great (although some-
what unusual) introductory geometry/topology course. One could even just do Part 2
on its own, lingering on the more beautiful parts, and filling in the classification of
surfaces. It could be nice to supplement this with some explicit examples from Chap-
ter 5 of Bonahon [2] and of geometric universal covers in Chapters 6 and 7 of that
same book. For a more advanced audience, Parts 3, 4, and 5 of Mostly Surfaces would
provide an overview/review of complex analysis and an entry into the fields of cone
surfaces, Veech surfaces, and Teichmüller theory. That same audience would appreci-
ate the three-dimensional hyperbolic geometry, Kleinian groups, and the example of
the figure-8 knot complement in [2].
Both Mostly Surfaces and Bonahon [2] have too much material for a beginning
mathematician to digest at one meal. But that itself is part of an introduction to re-
search: there is usually too much to do. And there is a lot of beautiful math in this
book.
The material discussed above is essentially Section 1, “Surfaces and Topology.”
Section 2, “Surfaces and Geometry,” is not really standard in the undergraduate cur-
riculum, but it should be. There are sections on the three geometries in dimension 2:
Euclidean, spherical, and hyperbolic. The section on Euclidean geometry contains par-
ticularly nice and unusual tidbits, like Pick’s theorem. Actually I like all of it: the area
of a hyperbolic surface in terms of Euler characteristic, the Hairy Ball theorem, the
isometries of S 2 , etc. Although I think some of the other material in this book might
be difficult for beginning undergraduates to digest, Section 2 is accessible, interesting,
important, and often not taught in an undergraduate curriculum. Spherical and hyper-
bolic geometry ideally would be taught widely, not just to develop geometric intuition
but because they are applicable to so many areas of science. For people interested in
learning more deeply about these topics, as well as many others, I would highly advise
the now classic Thurston [5].
REFERENCES
1. M. A. Armstrong, Basic Topology. Undergraduate Texts in Mathematics. Springer, New York, 1997,
http://dx.doi.org/10.1007/978-1-4757-1793-8.
2. F. Bonahon, Low-Dimensional Geometry: From Euclidean Surfaces To Hyperbolic Knots. Student Mathe-
matical Library: IAS/Park City Mathematical Subseries. American Mathematical Society, Providence, RI,
2009, http://dx.doi.org/10.5860/choice.47-5697.
3. A. Hatcher, Algebraic Topology. Cambridge Univ. Press, Cambridge, 2001.
4. T. Needham, Visual Complex Analysis. Oxford Univ. Press, Oxford, 1999, http://dx.doi.org/10.
2307/3618747.
5. W. P. Thurston, Three-Dimensional Geometry and Topology. Princeton Univ. Press, Princeton, NJ, 1997.
6. A. Zorich, Flat surfaces, in Frontiers in Number Theory, Physics, and Geometry. Vol. I. Springer, Berlin,
2006, http://dx.doi.org/10.1007/3-540-31347-8_13.
956
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
EDITOR’S ENDNOTES
q
ψ j (q) = e j+1 ,
which implies
∞ qj −1 log(1−q) 1
ψ j (q j ) = e j=1 j+1 = e−1−q = e−1 (1 − q)− q .
j≥1
http://dx.doi.org/10.4169/amer.math.monthly.121.10.957
√
3= 1 + 2 1 + 3 1 + 4 1 + ....
Then he says that in 1935 Herschfeld proved that (a) the limit
√
lim 1 + 2 1 + 3 1 + ...(n − 1) 1 + n 1
n→∞
exists and equals 3, and (b) for a nonnegative sequence an , the infinite nested radical
√
a1 + a2 + a3 + ...
−n
converges if and only if the sequence (an )2 is bounded.
While all that is true, Herschfeld was not the first. Eight years earlier, in a note on
page 348 of “The Collected Papers of Srinivasa Ramanujan” (G.H. Hardy, P.V. Seshu
Aiyar and B.M. Wilson, eds., Cambridge University Press, 1927; reprinted by Chelsea,
1962), T. Vijayaraghavan proved statements (a) and (b).
958
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
Jerry Koliha sends the following comments about his MONTHLY paper “Evaluating
Lebesgue integrals efficiently with the FTC” (this MONTHLY 121 (2014), 361–364).
There is a misprint in my paper, but it occurs four times. In the theorems Lebesgue’s
FTC, Theorem 1, Theorem 2, and Theorem 3, the first occurrence of ∞ in each of these
theorems should be −∞.
The MONTHLY expresses its appreciation to the following people for their help in
refereeing during the past year (July 2013 to July 2014).We could not function suc-
cessfully without such people and their hard work.
http://dx.doi.org/10.4169/amer.math.monthly.121.10.960
960
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
Giblin, Peter Kaftal, Victor Look, Daniel
Glasser, Lawrence Kain, Ben Loper, Alan
Glaz, Sarah Kantrowitz, Robert Lorch, John
Goldman, William Kaplan, Samuel Lovett, Shachar
Goodman-Strauss, Chaim Katok, Svetlana Lovett, Stephen
Gora, Pavel Katz, Eric Luca, Florian
Gordon, Cameron Katz, Victor Lucas, Tom
Gorkin, Pamela Kenyon, Richard Lucchini, Andrea
Gornet, Ruth Khavinson, Dmitry Ludwig, Lew
Graham, Ron Khovanova, Tanya Luecke, John
Granville, Andrew Kim, Junhyong Luijten, Erik
Gray, Jeremy Kim, Peter Lutwak, Erwin
Gremaud, Pierre Klain, Dan Maclagan, Diane
Gressman, Philip Kleitman, Daniel Malandro, Martin
Griffiths, Phillip Klingler, Lee Manlove, David
Grinberg, Eric Knill, Oliver Mann, Stephen
Grinstead, Charles Knopfmacher, Arnold Manning, Anthony
Gross Mark Kobayashi, Mitsuo Mansour, Toufik
Grujic, Zoran Koch, Sarah Margetis, Dionisios
Grundman, Helen Koliha, Jerry Margolius, Barbara
Grynkiewicz, David Kontorovich, Alex Markarian, Roberto
Hajja, Mowaffaq Koralov, Leonid Matousek, Jirı́
Hales, Tom Koshlukov, Plamen Matthews, Gretchen
Halter-Koch, Franz Kossak, Roman Maurer, Stephen
Hampton, Marshall Kowalski, Travis McCleary, John
Harbater, David Kozek, Mark McCuan, John
Hart, Joan Kra, Irwin McLean, K. Robin
Hartshorne, Robin Krantz, Steven Mellor, Blake
Hasenauer, Richard Krivelevich, Michael Melman, Aaron
Hasfura-Buenaga, Roberto Kuczmarski, Fred Mendivil, Franklin
Hasselblatt, Boris Kufner, Alois Mercer, Peter
Hell, Pavol Kutsia, Temur Merino, Dennis
Henk, Martin Lambers, James Miller, Steven
Henle, James Lamoureux, Mike Minda, David
Hinkkanen, Aimo Langer, Joel Moll, Victor
Hodges, Wilfrid Larson, Lee Montgomery, Hugh
Hoffman, Michael Lawlor, Gary Moree, Pieter
Hofmann, Karl Lazebnik, Felix Mortini, Raymond
Holmer, Justin Leader, Imre Mossinghoff, Michael
Holmsen, Andreas Leamer, Micah Muir, Jerry
Howard, Fredric Leckband, Mark Murty, Ram
Howards, Hugh Lehnen, Al Nathanson, Melvyn
Howe, Roger Leise, Tanya Nelsen, Roger
Hrbacek, Karel Lemmermeyer, Franz Nica, Bogdan
Hubbard, John Levesque, Claude Niederhausen, Heinrich
Ipsen, Ilse Levi, Mark Nitecki, Zbigniew
Irving, Rob Levy, Doron Oh, Byeong-Kweon
Isaacs, Martin Lewis, Barry Olberding, Bruce
Isaksen, Dan Lewis, Mark O’Leary, Robbin
Ivanov, Sergei Lewis, Tim Olofsson, Peter
Jarai, Antal Li, Hanfung Ono, Ken
Jerónimo-Castro, Jesús Li, Zhongshan Osborn, Judy-Anne
Johnson, Warren Lindstrom, Tom Osgood, Brad
Jones, Nathan Little, John Ott, Katharine
Jones, Rafe Littlejohn, Lance Pak, Igor
962
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
From the Monthly Over 100 Years Ago. . .
Halsted on Gauss
“It has been said the heroic age of non-euclidean geometry is passed, since long gone are
the days when Lobachevski flinched into calling his system ‘imaginary geometry,’ and, I might
add, Gauss kept a more cowardly silence because, as he confesses, he ‘fears the outcry of the
Boeotians,’ (proverbial for stupidity).
“. . . the designation ‘Absolute Geometry’ was first used by myself as a rendering for John
Bolyai’s phrase scientiam spatii absolute veram, in which I have always gloried as showing
the magnificent verve of the young Magyar hero, victim of the meanness of Gauss, as was also
his own son who passed his life an exile here in Colorado.”
Reaction 1 . . .
“I am sure that you know by reputation, if not personally, Mr. George Bruce Halsted, who
has given much attention to the study of Non-Euclidean Geometry, a branch of mathematics
in which you have taken great interest. If I am correct in this, you are aware that he has
been a great champion of John Bolyai. For some reason this has developed in him a spirit of
antagonism to my grandfather and led him into a very unjust attack upon him. In the January
number for this year of ‘The American Mathematical Monthly’ he had an article in which he
declared that John Bolyai was a victim of ‘the meanness of Gauss’. To this he added that one
of Gauss’ own sons also was a victim of this ‘meanness’, and that he had spent his life an exile
in the State of Colorado.
“Professor Florian Cajori of Colorado Springs in this State, wrote me after reading Profes-
sor Halsted’s article. He knew that the reference could only be to Eugene Gauss, my father.
. . . I promptly replied to Professor Cajori that my father was in no sense an exile from his
home in Göttingen, and furthermore, that he had never been in the State of Colorado.
“. . . The enclosed letters also show the circumstances under which my father left Göttingen,
and I think they are a complete vindication of my grandfather against the charge made by
Halsted that he had ‘meanly’ treated my father.”
Reaction 2 . . .
“American mathematicians are indebted to Halsted for making the writings of the creators
of non-Euclidean geometry accessible to them in the English language. His commentaries
were always spicy and valuable, even though, as a historian, Halsted was not always able to
maintain the attitude of an impartial judge. At his hands Gauss, for instance, received scant
justice.”
http://dx.doi.org/10.4169/amer.math.monthly.121.10.963
MSC: Primary 01A70; Secondary 53-03
Proof. Clearly,
√ 12 < 2 < 22 . Since 2 lies between
√ two consecutive
√ integral
squares, 2 cannot be an integer. Suppose 2 is rational, i.e., 2 = mn , 1 <
n<m< √2n; m, n ∈ N. Since m > n, we may write: m = n + r, n > r, r ∈ N.
Hence, 2 = n+r . Let g = (n, r ). So, n = gs and r = gt, s > t; (s, t) =
n √
1; g, s, t ∈ N. Then, 2 = s+t
s
. In other words, 2s 2 = s 2 + 2st + t 2 . So,
http://dx.doi.org/10.4169/amer.math.monthly.121.10.964
MSC: Primary 11J72
964
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 121
A MERICAN M ATHEMATICAL S OCIETY
New R
elea ses fr
ARNO om th
SWIMM
LD
ING
AGAIN
e AMS
Boris A ST THE
.
Ontari Khesin, Un TIDE
o, Can iv MATH
ICERM ada, an ersity of Toron
, Brown d Serg
e L. Ta to,
CIRCL
Pennsylv Univer bachnik ELEM ES FO
Editors ania State Un sity, Providen o ENTAR R
This b iv ersity, S ce, RI, v, STUDE Y SCH
Vladim
ook re
counts ta te Colle and NTS OOL
ir Arno the wo g e, PA, N
ld. rk an d life atasha
2014; 22
4 of emin Univer Rozhkovska
US$23.2 pages; Softcov ent mat sity, Ma y
0; Order er ; ISBN hematic nhatta a, Kansas Sta
code M : 978-1-
4704-1 ian gram, w An ove n, KS te
BK/86 699-7; Li ith pro rview
blems, o
st US$29
; AMS m Titles in activiti f the popular
es, and
embers this serie
s are co teachin elementary sc
MATH MSRI -publishe g guides hool pro
EMAT Mathem d w ith the
Mat
. -
UNDE ICAL ISBN: 97
8-1-4704
atical
Circles
hematic
al Scienc
es Rese
RSTAN -1695-9;
List US$
Library
, Volum arch In
stitute (M
ESSAY DING 25; All in e 13; 20 SRI).
S ON A OF NA dividual 14;
s US$20 166 pages; So
PHY MAZ TURE ; Order ftc
code M over ;
SIC ING PEARL
THEIR AL PHENOM S FRO CL/13
UNDER ENA A
N THE LV MAL
MATH
EMATIC STANDING D OV SC OST C
IANS BY Roman HOOL ITY
V. I. A
rn Duda, OF MA
A collec old Univer THEM
Arnold tion of short Translat sity of W ATICS
. stories edby Dan rocław,
by Russian This ch iel Dav Poland
2014; 16
7 mathem ronic ies
US$23.2 pages; Softcov atician key asp one seek le of th
e Lvov
0; Order er ; ISBN Vladim ec ing a
: 978-1- ir anywher ts of twenti school
eth-cen cultural and
code M 4704-1 wil
instituti l appeal to an
BK/85 701-7; Li e else in
st US$29 th e extan tu ry Poli
; AMS m
embers
History
of Mat t English
-l
sh mathemat onal overview y-
4704-1 an guage li ics of
A DEC 076-6; Li hematics, Vo terature not described
ADE O st US$39
; AMS
lume 40
; 2014 .
MATH F THE membe ; 23
rs US$31 1 pages; Har
CIRCL BERK .20; Ord dcover
THE E ELEY er code ; ISBN
HMATH : 978-1-
AM THE W /40
VOLUM ERICAN EX AR OF
GUNS
E II PERIEN MATH AND
Zvezdel C E, EMAT
ICS
CA, an ina Stankova E
MATH
d ,
Editors Tom Rike, O Mills College, COMM MATICAL PR
U A
This se
co
akland O
High S akland, WESTE NITIES IN F CTICES AND
mathem nd volume R R
atical to contain
s a va
chool, C
A, WAR I N ALLIES A ANCE AND
pic s, some riety o David
A R OUND ITS
Marie ubin, Sorbon
Titles in new an f
this serie d some enticing and WORL
s are co continu st Curie, ne Un D
ing from imulating
-publishe iversité
MSRI d with th Gauche, Institu s, U
Mathem td
mathém France, and C e mathématiq niversité Pierr
e Mathe Volume
pages; So atical matical
I.
ftcover Circles Sciences
atiques atherin ues de Ju e
ssieu-P et
Order ; ISBN: L ib rary, Vo Re se ar
code M 978-0-82 lume 14
ch Institu
te (MSR d e Ju ssieu-Pa e G o ld
This bo
ris Riv stein, CNRS aris Rive
CL/14 18-491 ; 2014; I).
2-5; List o
US$25;
All indi
approx
imately War I o k suggests a n e Gauch , Instit
364 n math e, Fran u
ce, Edit t de
viduals ew visi
US$18.7 ematics on
5; History and mat of the long- ors
o f Mat h ematicia term influen
4704-1
469-6; Li hematics, Vo ns. ce of W
st US$12 lume 42 orld
6; AMS ; 2014;
membe 39
rs US$10 1 pages; Hardc
0.80; O over ; IS
rder co BN
de HMAT : 978-1-
H/42
www.a Orde Or
ms.org r Online: (8 der by Phon
/books 00)321 e:
tore
(401)4 -4267 (U.S. &
55-400
0 (Worl Canada),
dwide
)
facebo
ok.com
/amerm
athsoc
plus.goo @amermaths
gle.com oc
/+AmsO
rg
1529 Eighteenth St., NW ● Washington, DC 20036