A M M P R o B L e M S 2 0 0 8

PROBLEMS AND SOLUTIONS
Edited by Gerald A. Edgar, Doug Hensley, Douglas B. West

with the collaboration of Paul T. Bateman, Mario Benedicty, Itshak Borosh, Paul
Bracken, Ezra A. Brown, Randall Dougherty, Tamás Erdélyi, Zachary Franco, Chris-
tian Friesen, Ira M. Gessel, Jerrold Grossman, Frederick W. Luttmann, Vania Mas-
cioni, Frank B. Miles, Richard Pfiefer, Cecil C. Rousseau, Leonard Smiley, John
Henry Steelman, Kenneth Stolarsky, Richard Stong, Walter Stromquist, Daniel Ull-
man, Charles Vanden Eynden, and Fuzhen Zhang.
Proposed problems and solutions should be sent in duplicate to the MONTHLY

problems address on the inside front cover. Submitted solutions should arrive at
that address before May 31, 2008. Additional information, such as generaliza-
tions and references, is welcome. The problem number and the solver’s name
and address should appear on each solution. An asterisk (*) after the number of
a problem or a part of a problem indicates that no solution is currently available.
PROBLEMS
11334. Proposed by Jonathan Bober and Jeffrey Lagarias, University of Michigan,

Ann Arbor, MI. Let x and y be positive integers with x 2 − 3y 2 = 1.
(a) Show that if x cannot be written as 2n , 2n ± 1, 3 · 2n , or 3 · 2n ± 1 for any natural
number n, then x y has more than 3 distinct prime factors.
(b) Show that if x = 2n − 1 with n > 3, then y has at least three distinct prime factors.
11335. Proposed by Juan López González, Madrid, Spain. Let σ (n) denote the sum of
the divisors of n.
(a) Find, with proof, the least positive integer m such that σ (6m) < σ (6m + 1).
(b) Show that there are infinitely many m such that σ (6m) < σ (6m + 1).
(cf G. Martin, The Smallest Solution of φ(30n + 1) < φ(30n) Is . . . , this M ONTHLY
106 (1999) 449-451.)
11336. Proposed by Donald Knuth, Stanford University, Stanford, CA. A near-
deBruijn cycle of order d is a cyclic sequence of 2d − 1 zeros and ones in which
all 2d − 1 substrings of length d are distinct. For all d > 0, construct a near-deBruijn
cycle of order d + 1 such that the front and back substrings of length 2d − 1 are both
near-deBruijn cycles of order d. (Thus, for example, 1100010 is near-deBruijn of
order 3, while 110 and 010 are both near-deBruijn of order 2.)
11337. Proposed by Marian Tetiva, National College “Gheorghe Roşca Codreanu”,
Bı̂rlad, Romania. Suppose in triangle ABC we have opposite sides of lengths a, b, and
c, respectively, with a ≤ b ≤ c. Let wa and wb be the lengths of the bisectors of the
angles A and B respectively. Show that a + wa ≤ b + wb .
Furdui, Cluj, Romania. Let denote the classical gamma

11338. Proposed by Ovidiu
function, and let G(n) = nk=1 (1/k). Find

lim G(n + 1)1/(n+1) − G(n)1/n .
n→∞
January 2008] PROBLEMS AND SOLUTIONS 71

11339. Proposed by José Luis Dı́az-Barrero, Universitat Politècnica de Catalunya,
Barcelona, Spain. Let Fn and L n denote the nth Fibonacci and Lucas numbers, respec-
tively. Prove that for all n ≥ 1,
1 1/Fn Fn+1
Fn + L 1/L
n
n
≤2− .
2 F2n
(The Fibonacci and Lucas numbers are given by the recurrence an+1 = an + an−1 , with
F0 = 0, F1 = 1, L 0 = 2 and L 1 = 1.)
11340. Proposed by Óscar Ciaurri and Luz Roncal, Universidad de la Rioja, Logroño,
Spain. An umbrella of radius 1 meter is spun with angular velocity ρ in the x z-plane
about an axis (call it the y-axis) parallel to the ground. It is wet, and drops of water
crawl along the ribs and fly off as they reach their ends.
Each drop leaves the umbrella with a velocity vector equal to the velocity of the
tip of the rib at the point where it exited. It then follows a parabolic trajectory. If a
drop spins off while on the downspin, then the high point in its arc will be the point
of departure. Otherwise, the high point is the vertex of a parabolic arc in the x z-plane.
Determine a parameterized family Pρ of polynomials in two variables such that when-
ever ρ 2 > g, the various arc vertices reached by the water droplets all lie on the curve
Pρ (x, z) = 0. (Here g denotes the magnitude of the downward acceleration due to
gravity.)
The figure shows the case ρ 2 /g = 2,
with the umbrella spinning counterclock-
wise.
SOLUTIONS
A Condition on the Commutator

11196 [2006, 79]. Proposed by Mohammad Hossein Mehrabi, Iran University of Sci-
A and B be real n × n matrices. Show that if
ence and Technology, Tehran, Iran. Let √
AB − B A is invertible and A2 + B 2 = 3(AB − B A), then n is a multiple of 6.
Solution by Pál Péter Dályay, Deák Ferenc High School, Szeged, Hungary. From
√
(A + i B)(A − i B) = A2 + B 2 − i(AB − B A) = ( 3 − i)(AB − B A),
√
it follows that det[(A + i B)(A − i B)] = ( 3 − i)n det(AB − B A). Since AB − B A
is invertible, so is (A + i B)(A − i B). We conclude that
det[(A + i B)(A − i B)] = det(A + i B) det(A − i B) = det(A + i B)det(A + i B) > 0.
√
Since A and B are real, also det(AB − B A) is real; hence ( 3 − i)n is real. Since
√
3 − i = 2e−πi/6 , we conclude that n is a multiple of 6.
Also solved by S. Amghibech (Canada), R. Chapman (U. K.), Y. Dumont (France), O. Furdui, J.-P. Gri-
vaux (France), C. C. Heckman, E. A. Herman, O. P. Lossers (Netherlands), M. Omarjee (France), S. Pierce,
A. K. Shafie (Iran), R. Stong, M. Tetiva (Romania), J. Vinuesa (Spain), L. Zhou, NSA Problems Group, and
the proposer.
72
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 115
Can One Recover a (Spherical) Triangle from Its Medial Triangle?
11201 [2006, 179]. Proposed by Robert Russell, New York, NY. Given the midpoints
of the sides of a spherical triangle, provide a construction for the original triangle.
Composite solution by Christoph Soland (Switzerland), William Dickenson, and the
editors. There is one configuration of midpoints, namely, mutually perpendicular, for
which many choices of original triangle work. Otherwise, the underlying triangle may
be recovered by a construction.
We take it to be part of the definition of a spherical triangle that the three vertices
do not lie on a common great circle. We also assume (by definition) that the sides of
a spherical triangle are less than π radians. Thus, to construct a spherical triangle, it
suffices to construct the three vertices. Let O be the center of our sphere S . Without
loss of generality, we take O to be the origin of R3 and the radius of S to be 1. When
A and B are not antipodal, let m(A, B) be the midpoint of the spherical line segment
AB, defined as the shorter of the two portions of the great circle through A and B
delimited by A and B.
We first consider the exceptional case. If any side of the medial triangle M1 M2 M3
is π/2, then, we claim, all of them are, and there are infinitely many triangles with the
same three midpoints, so we cannot uniquely construct the “original triangle”. For the
proof of this claim, suppose that M1 ⊥ M2 . Without loss of generality, M1 = (1, 0, 0)
and M2 = (0, 1, 0). By assumption, there exist distinct S1 , S2 , and S3 , no two antipodal,
such that for any permutation (i, j, k) of (1, 2, 3), m(Si , S j ) = Mk . Note that the ends
of a line segment in spherical geometry are nearer the midpoint than its antipode.
Now if S2 = (x2 , y2 , z 2 ), then x2 > 0, S3 = (x2 , −y2 , −z 2 ) so that y2 < 0, and S1 =
(−x2 , −y2 , z 2 ). Now z 2 = 0 since S1 = −S2 , and it follows that M3 = (0, 0, sign(z 2 )).
Thus all three axes are perpendicular. Glancing over the foregoing algebra, we see that
any choice of S2 = (x2 , y2 , z 2 ) with x2 , z 2 > 0 > y2 and x22 + y22 + z 22 = 1 gives rise to
S2 , S3 , and S1 for which (1, 0, 0), (0, 1, 0), and (0, 0, 1) are the vertices of the medial
triangle.
We now consider the composition R = H3 ◦ H2 ◦ H1 , where H j is the bijection of
S obtained by reflecting it through the axis of ±M j . Either R is the identity map on S ,
or it is a nontrivial rotation that fixes two antipodal points, call them ±S. We claim that
the M j are mutually perpendicular if and only if R = I . In one direction this is trivial
by coordinate algebra after setting M1 = (1, 0, 0) and so on. In the other direction,
suppose R = I . Then for any permutation (i, j, k) of (1, 2, 3), Hi ◦ H j = Hk . Thus
the M j are distinct, and no two are antipodal. Now let P = H2 M1 and Q = H3 M1 . If
M1 ⊥ M2 , then M1 and M2 determine a line L on S , P, M1 , and M2 are distinct points
on L, and P is not antipodal to M1 .
Furthermore, H1 Q = P since H1 H3 H2 = I . Thus Q = H1 P. Since P lies on L
and H1 carries any line through M1 onto itself, Q is on L. From Q = H3 M1 , we
have also that Q is on the line L through M1 and M3 . Thus Q = M1 , Q = −M1 , or
L = L . But Q = M1 because M3 = ±M1 , and Q = −M1 because that would make
P = H1 Q = −M1 , a contradiction. Finally, if L = L , each H j transposes the poles
of S if we take L as the equator, and so R transposes those poles, contrary to the the
assumption that R = I . This proves that M1 ⊥ M2 , and by the same logic, all three
M j are mutually perpendicular.
If, on the other hand, no two of the M j are perpendicular, then the underlying S j
may be recovered from the M j by a construction.
Henceforth, we assume that no two of the M j are orthogonal. Thus, R = I . We
first show that the axis ±S of rotation for R is not perpendicular to M1 . To see this,
suppose to the contrary that H3 H2 H1 S = S, and S ⊥ M1 . Take S = H1 S and S =

H2 S . Then H3 S = H3 H2 H1 S = S. If S and S are antipodal, then S = H2 S =
−H2 S. Since H3 S = S, M3 S, and thus M3 ⊥ M1 , a contradiction. If S and S are
not antipodal, then, viewing these points as vectors in R3 , from H2 S = S it follows
that M2 (S + S ). If also S and S are not antipodal then M3 S + S = S − S .
But (S + S ) ⊥ (S − S ) so M2 ⊥ M3 . If instead, S and S are antipodal, then S =
S , But then H2 S = S so that M2 ±S. From M1 ⊥ S we have M1 ⊥ M2 . This
shows that S ⊥ M1 . Now take S2 as the nearer of ±S to M1 . Then take S3 = H1 S2 ,
and S1 = H2 S3 .
It remains to discuss how S2 may be located via a construction. This can depend
on what tools are permitted. Here, we assume that lines through points, intersections
of lines, midpoints of line segments, and perpendicular bisectors of line segments are
constructible, that we can pick points on lines that are distinct from given points or
their antipodes, that we can construct reflections of a point about a given axis, that
we can determine which of two points is nearer a given point if the points are not
equidistant, and we can pick points not on specified lines.
Consider, then, an arbitrary pair (A, B) of distinct, non-antipodal points. If either of
these is fixed by R, then we have ±S2 and we are done. Otherwise, S2 lies on the per-
pendicular bisector of (A, R(A)), as well as the perpendicular bisector of (B, R(B)).
If these are distinct, their intersection again gives S2 . If not, then the axis of rotation
for R lies on the intersection of the lines AB and R A R B, and that gives S2 .
Editorial comment. This problem would be trivial in Euclidean geometry: draw
through each midpoint a line parallel to the segment joining the other two midpoints;
these lines define the original triangle. In spherical geometry there are no parallels, so
this construction doesn’t make sense. There are other snags, and the somewhat narrow
definition used here for the term ‘spherical triangle’ was the only means known to the
editors to avoid these.
Also solved by R. Stong, A. Tissier (France), and the Microsoft Research Problems Group.
Differentiable and Discontinuous Densely

11221 [2006, 367]. Proposed by Paolo Perfetti, University “Tor Vergata”, Rome, Italy.
Give an example of a function g from R into R such that g is differentiable everywhere,
g is differentiable on one dense subset of R, and g is discontinuous on another dense
subset of R.
Solution by Richard Bagby, New Mexico State University, Las Cruces, NM. Define
f : R → R by f (x) = x 2 sin2 (π/x) for 0 < x < 1 and f (x) = 0 otherwise. Let {qk }
be an enumeration of the rationals. We will show that

∞

g(x) = 2−3k f 2k (x − qk )
k=1
has the required properties.

Note that f is differentiable everywhere: 0 ≤ f (x) ≤ x 2 shows f (0) = 0 and 0 ≤
f (x) ≤ x 2 sin2 (π − π/x) ≤ π 2 (x − 1)2 shows f (1) = 0. Also note that | f (x)| is
bounded and f is discontinuous at 0.
Since f is continuous and the series for g(x) converges uniformly, the sum g(x)
exists and defines a continuous function. The derived series

∞

2−2k f 2k (x − qk ) (1)
k=1
74
converges uniformly, so g is differentiable and g (x) is the sum of this derived series.
We claim that g is discontinuous at each rational q j . Write (1) as
−2k k
g (x) = 2−2 j f 2 j (x − q j ) + 2 f 2 (x − qk ) .
k = j
The last series defines a function continuous at q j since each term is continuous there
and the series converges uniformly. The discontinuity of f at 0 then shows that g is
discontinuous at q j .
For n a positive integer, let E n = ∞ k=n (qk − 2
1−k
, qk + 21−k ) and E = ∞ n=1 E n .
Thus E n has Lebesgue measure at most 23−n and E has measure zero, so the set of
irrationals in the complement of E is dense. We claim g is differentiable at every
irrational x0 ∈ E. Since x0 is neither qk nor qk + 2−k for any k, each term 2−3k f (2k (x −
qk )) is twice differentiable at x0 . Let n be such that x0 ∈ E n . Then the derived series
(1) reduces to a finite series, and its derivative at x0 is

n−1

2−k f 2k (x0 − qk ) . (2)
k=1
To prove this finite sum is g (x0 ), let

∞

α(x) = 2−2k f 2k (x − qk ) .
k=n

The term f 2k (x − qk ) vanishes when |x − x0 | ≤ 2−k , so

α(x) − α(x0 ) ≤ C 2−k |x − x0 | ≤ 2C|x − x0 |2 ,
2−k ≤|x−x 0 |
where C is a bound for | f |. Hence α (x0 ) = 0, so g (x0 ) is indeed given by the finite
series (2).
Also solved by J. H. Lindsey II, M. D. Meyerson, A. Stadler (Switzerland), R. Stong, Szeged Problem Solving
Group “Fejéntaláltuka” (Hungary), GCHQ Problem Solving Group (U. K.), and the proposer.
Snapshots for Velocities

11223 [2006, 459]. Proposed by Christopher Hillar, Texas A& M University, College
Station, TX, and Lionel Levine, University of California at Berkeley, Berkeley, CA.
Consider n unlabeled particles moving each at its own constant velocity along the real
line. An observer is promised some number P of snapshots of the particles, to be taken
at uniformly spaced intervals of time. When particles coincide, the snapshot will show
how many are at a given point.
(a) Show that if P = n + 1 then the observer can determine the velocity of each of the
particles.
(b∗ ) As a function of n, what is the minimum value of P that will suffice to ensure that
the observer can determine all n velocities?
Solution to (a) by S. C. Locke, Florida Atlantic University, Boca Raton, FL. Suppose
the kth snapshot is taken at time tk and the positions of the n particles at that time
are (xk, j )nj =1 , where xk, j is not necessarily the position of the jth particle. Let G =
{(tk , xk, j ) : 1 ≤ k ≤ n + 1, 1 ≤ j ≤ n}. If p j (t) is the position of the jth particle at
time t, then the points L j = {(t, p j (t)) : t ∈ R} would constitute a line meeting G in
exactly n + 1 points (not counted with multiplicity).

If a line L meets G in at least n + 1 points (and therefore exactly n + 1 points),
then at least two of these points are generated by the same particle, say particle j.
Thus L = L j , since two distinct points determine a line. The lines meeting G in n + 1
points are therefore exactly the lines representing the motion of the individual particles,
and the velocities of the particles are the slopes of these lines.
Editorial comment. No solutions to (b) were received. Several solvers noted that
for n ≤ 3, n + 1 snapshots are required. For n = 3, the three snapshots (−2, 0, 2),
(−1, 0, 1) and (−2, 0, 2) are ambiguous. The GCHQ Problem Solving Group verified
that for n = 4, four snapshots suffice. The best lower bound was due to Petr Skovron.
Let P(n) be the minimum number of snapshots required for n particles. Suppose the
n trajectories { f i (t)}i=1
n
are indistinguishable from {gi (t)}i=1n
with snapshots at times
1, . . . , P − 1. For any constant a, the 2n trajectories { f i (t) + a(t − P)}i=1
n
∪ {gi (t) −
a(t − P)}i=1 are indistinguishable from { f i (t) − a(t − P)}i=1 ∪ {gi (t) + a(t − P)}i=1
n n n
with snapshots at times 1, . . . , P. Hence P(2n) ≥ P(n) + 1. It follows that P(n) ≥

log2 n + 2.
Part (a) also solved by D. Beckwith, K. Bernstein, D. R. Bridges, P. Budney, R. Chapman (U. K.), J. H.
Lindsey II, L. Pebody, R. E. Prather, F. Yang, GCHQ Problem Solving Group (U. K.), Houghton College
Problem Solving Group, Szeged Problem Solving Group “Fejéntaláltuka” (Hungary), and the proposer.
A Unique Solution
11226 [2006, 460]. Proposed by Franck Beaucoup, Ottawa, Canada, and Tamás
Erdélyi, Texas A& M University, College Station, TX. Let a1 , . . . , an be real numbers,
each greater than 1. For n ≥ 2, show that there is exactly one solution in the interval
(0, 1) to

n
(1 − x a j ) = 1 − x.
j =1
Solution by Microsoft Research Problems Group, Redmond, WA. If a ≥ 1 and x ≥ 0,

then x a ≥ 1 + a(x − 1) bythe mean value theorem (MVT). Equality occurs only for
a = 1 or x = 1. If f (x) = nj =1 (1 − x a j ) for 0 ≤ x ≤ 1, then
0 ≤ f (x) ≤ (1 − x a1 )(1 − x a2 ) ≤ a1 a2 (1 − x)2 .

Hence f (x) < 1 − x as x approaches 1− . On the other hand, 1 ≥ f (x) ≥ 1 − (x a1 +
· · · + x an ) shows that f (x) > 1 − x as x approaches 0+ . By the intermediate value
theorem, f (x) = 1 − x has a root in (0, 1).
There is only one such root. Let g(x) = log(1 − x) − log f (x). Suppose g(0) =
g(x1 ) = g(x2 ) = 0 where 0 < x1 < x2 < 1. By MVT, h(x) = (1 − x)g (x) has
at least one zero in (0, x1 ) and another
n in (x1a, x−12 ). By MVT again, h has a root
in (0, x2 ). However, h(x) = −1 + j =1 a j x j (1 − x)(1 − x ) , so h (x) =
a j −1
n a j −2 a j a j −2
j =1 a j x (x − 1 − a j x + a j )(1 − x ) and is thus positive on (0, 1).
Also solved by K. F. Andersen (Canada), R. Bagby, P. P. Dályay (Hungary), J. H. Lindsey II, O. P. Lossers
(Netherlands), T. L. McCoy (Taiwan), R. Mortini (France), L. Pebody, J. Rooin and A. Mahmoodi (Iran), A.
Stadler (Switzerland), R. Stong, BSI Problems Group (Germany), GCHQ Problem Solving Group (U. K.), and
the proposers.
76
A Triangle Inequality
11228 [2006, 460]. Proposed by Marian Tetiva, Bı̂rlad, Romania. Prove that in an
acute triangle with angles A, B, and C radians,
(1 − cos A)(1 − cos B)(1 − cos C)
cos A cos B cos C
8(tan A + tan B + tan C)3
≥ .
27(tan A + tan B)(tan A + tan C)(tan B + tan C)
Solution by Minh Can, Irvine Valley College, Irvine, CA.

(tan A + tan B + tan C)3 cos A cos B cos C
(tan A + tan B)(tan B + tan C)(tan C + tan A)(1 − cos A)(1 − cos B)(1 − cos C)
(tan A + tan B + tan C)3 cos A cos B cos C
=
sin C sin A sin B
(1 − cos A)(1 − cos B)(1 − cos C)
cos A cos B cos B cos C cos C cos A
(sin A)2 (sin B)2 (sin C)2 (1 − cos2 A)(1 − cos2 B)(1 − cos2 C)
= =
(1 − cos A)(1 − cos B)(1 − cos C) (1 − cos A)(1 − cos B)(1 − cos C)

3 + cos A + cos B + cos C 3
= (1 + cos A)(1 + cos B)(1 + cos C) ≤
3
3
3 + 3 cos (A + B + C)/3 27
≤ = .
3 8
The last inequality follows from the concavity of cos x on the interval [0, π/2]. The
inequality before that follows from the arithmetic–geometric mean inequality. Equality
holds if and only if the triangle is equilateral.
Also solved by S. Amghibech (Canada), A. Arkady, A. R. Avidon, M. Battaille (France), D. Beckwith, A. Bun-
gale (India), R. Chapman (U. K.), G. H. Chung, P. P. Dályay (Hungary), P. De (Ireland), O. Faynshteyn
(Germany), D. Fleischman, M. Goldenberg, M. Hajja (Jordan), E. A. Herman, Y.-J. Kuo, J. Lee (Korea),
O. P. Lossers (Netherlands), M. Mabuchi (Japan), D. J. Moore, L. Pebody, D. Perkins, C. R. Pranasachar (In-
dia), J. Rooin (Iran), V. Schindler (Germany), S. Shaebani (Iran), A. Stadler (Switzerland), R. Stong, T. Tam,
S. Varosanec (Croatia), J. Vinuesa (Spain), M. Vowe (Switzerland), GCHQ Problem Solving Group (U. K.),
Microsoft Research Problems Group, Szeged Problem Solving Group “Fejéntaláltuka” (Hungary), and the
proposer.
A Function Inequality
11232 [2006, 567]. Proposed by Michael W. Botsko, Saint Vincent College, Latrobe,
PA. Let f be a continuous mapping from R into R that is bounded below. Show that
there exists a real number x0 such that f (x0 ) − f (x) < |x − x0 | holds for all x other
than x0 .
Solution by Albert Stadler, Dübendorf, Switzerland. Put g(x) := f (x) + |x|/2. Then
g is continuous, bounded below, and g(x) → +∞ as x → ±∞. Therefore there is a
real number x0 such that g(x0 ) = minx∈R g(x). Thus f (x) + |x|/2 ≥ f (x0 ) + |x0 |/2
or f (x0 ) − f (x) + |x0 |/2 − |x|/2 ≤ 0. We conclude that for x = x0 ,
f (x0 ) − f (x) − |x − x0 | < f (x0 ) − f (x) − |x − x0 |/2
≤ f (x0 ) − f (x) + |x0 |/2 − |x|/2 ≤ 0.

Also solved by 41 others, and the proposer.
A Derivative Formula
11233 [2006, 568]. Proposed by Robert Downes, Mountain Lakes High School, Moun-
tain Lakes, NJ. Show that for positive integer n, and for x = 0,

dn 1 (−1)n 1 nπ
x n−1
sin = n+1 sin + .
dxn x x x 2
Solution by Jean-Pierre Grivaux, Paris, France. We will show more generally that if
f is an n-times differentiable real-valued function, then

dn 1 (−1)n (n) 1
x n−1
f = n+1 f .
dxn x x x
The proof is by induction on n. The case n = 1 is obvious. Suppose it is true for
n − 1. Applying the inductive hypothesis, and taking g(x) = x n−2 f (1/x) in Leibniz’s
formula
dn dn g d n−1 g
n
(xg(x)) = x n + nx n−1 ,
dx dx dx
we compute

dn 1 dn 1 d n−1 1
n
x n−1
f =x n x n−2
f + nx n−1 x n−2
f
dx x dx x dx x

d (−1) n−1
1 (−1) n−1
1
=x n
f (n−1) + nx n
f (n−1)
dx x x x x

(−1) n (n−1) 1
n
(−1) (n) 1
n
(−1) n (n−1) 1
n−1
=x f + f + f
x n+1 x x n+2 x x n−1 x

(−1) n
1
= n+1 f (n) .
x x
Specializing to f (x) = sin(x) gives the desired result.

Editorial comment. The more or less equivalent result (obtained by setting f (x) = e x ,)
d n n−1 1/x
x e = (−1)n x −n−1 e1/x ,
dxn
appears as a problem in 1850 Exercices de Mathématiques Posés à l’Oral du CAPES de
Mathématiques et des Concours des Grande Écoles, by L. Moisotte. The most general
formula of this kind submitted was given by O. P. Lossers, who showed

dk 1 k
k! m − j m−k− j j ( j) 1
m
x f = x (−1) f ,
dx k x j =0
j! k − j x
from which the above formula follows by taking k = n and m = n − 1.

Also solved by 74 other readers and the proposer.
78

that address before June 30, 2008. Additional information, such as generaliza-
PROBLEMS
11341. Proposed by Cezar Lupu, University of Bucharest, Bucharest, Romania (stu-
dent), and Tudorel Lupu, Decebal High School, Constanza, Romania. Consider an
acute triangle with side-lengths a, b, and c, with inradius r and semiperimeter p. Show
that
√
3 3r
(1 − cos A)(1 − cos B)(1 − cos C) ≥ cos A cos B cos C 2 − .
p
11342. Proposed by Luis H. Gallardo, University of Brest, Brest, France. Let p be a

prime and let Fq be a finite field of characteristic p, where q is a power of p. Let n be
a divisor of q − 1. With the natural mapping of Z onto F p and embedding of F p in Fq ,
show that (−1)(n+2)(n−1)/2 n n is a square in Fq .
11343. Proposed by David Beckwith, Sag Harbor, NY. Show that when n is a positive
integer,
n 2k n 2k
= 3n−2k .
k≥0
k k k≥0
2k k
11344. Proposed by Albert Stadler, Meilen, Switzerland. Let μ be the Möbius μ func-
tion of number theory. Show that if n is a positive integer and n > 1, then

n
(n−1)/2
n/(2 j +1)
μ( j) = − j μ(k).
j =1 j =1 k=(n+1)/(2 j +3)
11345. Proposed by Roger Cuculière, France. Find all nondecreasing functions f from
R to R such that f (x + f (y)) = f ( f (x)) + f (y) for all real x and y.
11346. Proposed by Christopher Hillar, Texas A&M University, College Station, TX,
and Lionel Levine, University of California, Berkeley, CA. Let n be an integer greater
166

than 1 and let S = {2, . . . , n}. For each nonempty subset A of S, let π(A) = j ∈A j.
Prove that when k is a positive integer and k < n,
n
lcm({1, . . . , n/i}) = gcd ({π(A) : |A| = n − k}) .
i=k
n
(In particular, setting k = 1 yields i=1 lcm({1, . . . , n/i}) = n!.)
√
11347. Proposed by Mihály Bencze, Brasov, Romania. Let A = (x + y)/2, G = x y,
and

1 x x 1/(x−y)
I = .
e yy
Determine all ordered 4-tuples (α, β, γ , δ) of positive numbers with α > β and γ > δ
such that for all distinct positive x and y,
α A + βG
1/(γ +δ) √
I > > Aγ G δ > AG.
α+β
SOLUTIONS
Orthogonal Sudoku Squares

11214 [2006, 268]. Proposed by Solomon W. Golomb, University of Southern Califor-
nia, Los Angeles, CA. A Sudoku solution is a 9 × 9 square array with integer entries
such that each of the nine possible entries occurs exactly once in each row, once in
each column, and once in each of the nine 3 × 3 subsquares that together tile the
main array. Is it possible for two Sudoku solutions to form a pair of orthogonal Latin
squares? (That is, can the 81 pairs of corresponding cells contain all 81 possible pairs
of entries?)
Solution by Kyle Calderhead, Illinois College, Jacksonville, IL. Yes. In fact, there is a
family of six pairwise orthogonal Latin Sudoku squares:
It comes from the usual construction of complete families of orthogonal Latin
squares, limited to the ones that also meet the Sudoku restrictions. The standard con-
struction permutes the rows of the addition table for the finite field F9 , with the permu-
tations determined by the (nonzero) rows of the multiplication table. If we represent
F9 as the set of congruence classes of polynomials over F3 modulo a quadratic polyno-
mial such as x 2 + 1 that is irreducible over F3 , then the rows of the multiplication table
that result in Latin Sudoku squares are precisely those that correspond to nonconstant
polynomials. This produces the six Latin Sudoku squares shown below. More gener-
ally, for any prime p, the same technique will produce a family of p( p − 1) mutually
orthogonal p 2 × p 2 Latin Sudoku squares.
1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9
4 5 6 7 8 9 1 2 3 5 6 4 8 9 7 2 3 1
7 8 9 1 2 3 4 5 6 9 7 8 3 1 2 6 4 5
3 1 2 6 4 5 9 7 8 6 4 5 9 7 8 3 1 2
6 4 5 9 7 8 3 1 2 7 8 9 1 2 3 4 5 6
9 7 8 3 1 2 6 4 5 2 3 1 5 6 4 8 9 7
2 3 1 5 6 4 8 9 7 8 9 7 2 3 1 5 6 4
5 6 4 8 9 7 2 3 1 3 1 2 6 4 5 9 7 8
8 9 7 2 3 1 5 6 4 4 5 6 7 8 9 1 2 3
February 2008] PROBLEMS AND SOLUTIONS 167

1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9
6 4 5 9 7 8 3 1 2 7 8 9 1 2 3 4 5 6
8 9 7 2 3 1 5 6 4 4 5 6 7 8 9 1 2 3
9 7 8 3 1 2 6 4 5 2 3 1 5 6 4 8 9 7
2 3 1 5 6 4 8 9 7 8 9 7 2 3 1 5 6 4
4 5 6 7 8 9 1 2 3 5 6 4 8 9 7 2 3 1
5 6 4 8 9 7 2 3 1 3 1 2 6 4 5 9 7 8
7 8 9 1 2 3 4 5 6 9 7 8 3 1 2 6 4 5
3 1 2 6 4 5 9 7 8 6 4 5 9 7 8 3 1 2
1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9
8 9 7 2 3 1 5 6 4 9 7 8 3 1 2 6 4 5
6 4 5 9 7 8 3 1 2 5 6 4 8 9 7 2 3 1
5 6 4 8 9 7 2 3 1 8 9 7 2 3 1 5 6 4
3 1 2 6 4 5 9 7 8 4 5 6 7 8 9 1 2 3
7 8 9 1 2 3 4 5 6 3 1 2 6 4 5 9 7 8
9 7 8 3 1 2 6 4 5 6 4 5 9 7 8 3 1 2
4 5 6 7 8 9 1 2 3 2 3 1 5 6 4 8 9 7
2 3 1 5 6 4 8 9 7 7 8 9 1 2 3 4 5 6
Editorial comment. Lyle Ramshaw provided a proof that 6 is the maximum number of
pairwise orthogonal Latin squares of order 9. These results and others are dealt with
in “Sudoku, gerechte designs, resolutions, affine space, spreads, reguli, and Hamming
codes” by R. A. Bailey, Peter J. Cameron, and Robert Connelly, to appear shortly in
this M ONTHLY.
Also solved by R. Bagby, J. Brawner, R. Chapman (U. K.), K. Dale (Norway), D. Degiorgi (Switzerland),
G. T. Fala, R. T. Guy, D. E. Knuth, O. P. Lossers (Netherlands), M. D. Meyerson & T. S. Michael, R. M. Ped-
ersen, L. Ramshaw, T. Q. Sibley, J. H. Steelman, R. Stong, H. Stubbs, R. Tauraso (Italy), L. Wenstrom, the
ABC Student Problem Solving Group, the BSI Problems Group (Germany), Szeged Problem Solving Group
“Fejéntaláltuka”(Hungary), the GCHQ Problem Solving Group (U. K.), the Northwestern University Math
Problem Solving Group, the VMI Problem Solving Group, and the proposer.
Maximum Velocity on a Car Trip

11215 [2006, 366]. Proposed by Shmuel Rosset, Tel Aviv University, Tel Aviv, Israel. A
car moves along the real line from x = 0 at t = 0 to x = 1 at t = 1, with differentiable
position function x(t) and differentiable velocity function v(t) = x
(t). The car begins
and ends the trip at a standstill; that is, v = 0 at both the beginning and the end of the
trip. Let L be the maximum velocity attained during the trip. Prove that at some time
between the beginning and end of the trip, |v
| > L 2 /(L − 1).
Solution by José Heber Nieto, Universidad del Zulia, Moracaibo,
t Venezuela. Let M =
L 2 /(L − 1). If |v
| ≤ M for all t ∈ [0, 1], then v(t) = 0 v
(u) du ≤ Mt and v(t) =
1
− t v
(u) du ≤ M(1 − t) for all t ∈ [0, 1]. Since L = v(t0 ) for some t0 ∈ [0, 1], we
have L ≤ Mt0 and L ≤ M(1 − t0 ). Thus 1 − L1 ≤ t0 and t0 ≤ L1 , and therefore L ≤ 2.
Now consider the function f defined by
⎧
⎪
⎨ Mt, if 0 ≤ t < 1 − L1 ,
f (t) = L , if 1 − L1 ≤ t < L1 ,
⎪
⎩ M(1 − t), if 1 ≤ t ≤ 1.
L
168
Observe that
1
L 1 2 L 1
f (t) dt = 1− +L −1 + 1− = 1.
0 2 L L 2 L
We have v(t) ≤ f (t) for t ∈ [0, 1]. Equality cannot hold everywhere, since f is not
differentiable at 1/L. Hence v(t1 ) < f (t1 ) at some t1 ∈ [0, 1] and, by continuity,
v(t) < f (t) in some neighborhood of t1 . Our assumption that |v
| ≤ M on [0, 1] thus
leads to this contradiction:
1 1
1 = x(1) − x(0) = v(t) dt < f (t) dt = 1.
0 0
Also solved by T. Achenbach, M. W. Botsko, D. Chakerian, P. Corn, J.-P. Grivaux (France), E. A. Herman, J. H.
Lindsey II, O. P. Lossers (Netherlands), T. L. McCoy (Taiwan), K. McInturff, M. D. Meyerson, R. E. Prather,
K. Schilling, J. Silver, J. G. Simmonds, R. Stong, J. B. Zacharias, L. Zhou, Szeged Problem Solving Group “Fe-
jéntaláltuka” (Hungary), GCHQ Problem Solving Group (U. K.), Houghton College Problem Solving Group,
Microsoft Research Problems Group, NSA Problems Group, and the proposer.
Goes Back to Boole

11234 [2006, 568]. Proposed by Jim Brennan and Richard Ehrenborg, University of
Kentucky, Lexington, KY. Let a1 , . . . , an and b1 , . . . , bn−1 be real numbers, with a1 <
b1 < a2 . . . < an−1 < bn−1 < an . Let h be an integrable function from R to R. Show
that
∞ ∞
(x − a1 ) · · · (x − an )
h dx = h(x) d x.
−∞ (x − b1 ) · · · (x − bn−1 ) −∞
Solution by Eugene A. Herman, Grinnell College, Grinnell, IA. Let

(x − a1 ) · · · (x − an )
f (x) = .
(x − b1 ) · · · (x − bn−1 )
On each of the open intervals
(−∞, b1 ), (b1 , b2 ), · · · , (bn−1 , ∞) (1)
the function f is continuous and has limits −∞ and ∞ at the left and right endpoints,
repectively. Thus the range of f on each of these intervals is (−∞, ∞). Hence, for
every real number y, the equation y = f (x) has a solution in each of these n intervals.
In fact, there is exactly one solution in each interval, since on these intervals, f (x) −
y = 0 if and only if
(x − a1 ) · · · (x − an ) − y (x − b1 ) · · · (x − bn−1 ) = 0, (2)
which is a polynomial equation in x of degree n. Therefore, on each of these n inter-
vals, the function f is invertible. We denote the inverse of f on the ith interval by
gi . In particular, for any y ∈ R, the numbers g1 (y), · · · , gn (y) are the solutions of the
polynomial equation (2). Consequently, the sum of these solutions is the negative of
the coefficient of x n−1 in (2); that is,

m
n
gi (y) = y + ai . (3)
i=1 i=1

∞
We transform the integral −∞ h( f (x)) d x by making the substitution y = f (x) in
each of the n intervals (1); that is, in the ith interval, we have x = gi (y) and hence
d x = gi
(y) dy. Therefore
∞ b1 b2 ∞
h( f (x)) d x = h( f (x)) d x + h( f (x)) d x + · · · + h( f (x)) d x
−∞ −∞ b1 bn−1
n
∞ ∞
n ∞
= h(y)gi
(y) dy = h(y) gi
(y) dy = h(y) dy.
i=1 −∞ −∞ i=1 −∞
The last step follows from equation (3).

Editorial comment. Six solvers noted that this result is a consequence of the following
m
theorem (∗): If h(x) ∞ is integrable and∞g(x) = x − i=1 ci /(x − bi ) where all the ci
are positive, then −∞ h(g(x)) d x = −∞ h(x) d x. This reduction is obtained by writ-
(x−a1 )···(x−an ) m
ing (x−b 1 )···(x−bn−1 )
in its partial fraction form g(x) = x + C − i=1 ci /(x − bi ), using
the given conditions to show that all the ci are positive, then appealing to theorem (∗)
and to the translation invariance of Lebesgue measure. References provided for theo-
rem (∗) were: (1) G. Boole, “On a general theorem of definite integration,” Cambridge
& Dublin Mathematical Journal, 1849; G. Boole, “On the comparison of transcen-
dents . . . ” Philos. Trans. Roy. Soc., London 147 (1857) 745–803 (the proof perhaps
not meeting modern standards of rigor); (2) M. L. Glasser, “A remarkable property of
definite integrals,” Math. Comput. 40 (1983) 561–563; (3) G. Pólya & G. Szegő, Prob-
lems and Theorems in Analysis, vol. 1 (Springer, New York, 1972), Problem II-12.
Also solved by K. F. Andersen (Canada), R. R. Avidon, R. Bagby, R. Chapman (U. K.), N. Cohen (Brazil), P. P.
Dalyay (Hungary), B. Dunn, P. J. Fitzsimmons, M. L. Glasser, J.-P. Grivaux (France), J. Groah, J. A. Grzesik,
F. Holland (Ireland), G. Keselman, J. H. Lindsey II, O. P. Lossers (Netherlands), D. Lubell, M. Mabuchi
(Japan), J. Mack, W. Matysiak (Poland), T. L. McCoy (Taiwan), A. Nijenhuis, P. Perfetti (Italy), A. Quas
(Canada), K. Schilling N. C. Singer, A. Stadler (Switzerland), R. Stong, C. T. Stretch (Ireland), R. Tauraso
(Italy), S. Vagi, E. I. Verriest, L. Zhou, BSI Problems Group (Germany), GCHQ Problem Solving Group
(U. K.), Microsoft Research Problems Group, NSA Problems Group, and the proposers.
Coloring the Plane

11236 [2006, 655]. Proposed by Jim Owings, Riderwood, MD. Given a positive integer
n and a positive real number x, consider the proposition P(x, n): If we color each
point in the Euclidean plane with one of n colors, then there exist two points of the
same color that are either
√ distance 1√ or distance x apart.
√
(a) Prove that P((1 + 5)/2, 4), P( 3, 4), and P( 2, 4) all hold.
(b)∗ Does P(t, 4) hold for any t > 1 other than those specified in part (a)?
(c)∗ Does there exist t > 1 such that P(t, 5) holds?
Solution to (a) and (b) by Marian Tetiva, Bı̂rlad, Romania.
√ In a regular pentagon with
side 1, any two of the vertices have distance 1 or (1 + 5)/2. If the plane is colored
in four colors,
√ then two of these five points must have the same color. This proves
P((1 + 5)/2, 4). √
Next consider P( 3, 4). Assume there √ is a 4-coloring of the plane such that any
two points separated by a distance of 1 or 3 have different colors. Let A and B be
any two points with distance 2. Choose C so that ABC is an equilateral triangle, and
let D, E, F be the midpoints of the segments BC, C A, AB, respectively.
√ The distance
between any two of the resulting four points A, D, E, F is 1 or 3, so they have
170
different colors. Similarly B, D, E, F have different colors. Thus A and B have the
same color. Since B was an arbitrary point on the circle with center A and radius 2,
that circle is monochromatic. This gives a contradiction, because there are two points
on that circle with distance √1. √
Next we prove P((1 + 3)/ 2, 4); since this is not listed in (a), it will be an
example for (b). It is equivalent to prove: In every 4-coloring of the
√ Euclidean plane
√
there are two points having the same color that are distance 1 + 3 or distance 2
apart. Let A and B be any two points with distance√ 2. Let AB be the hypotenuse of
isosceles right
√ triangle ABC so that AC = BC = 2. On the circle centered at A
with radius 2, take D and E to be the two points such that D AB =√ E AB = 105
degrees. Now D B E is an equilateral triangle√with sides of length 1 + 3, and AC D

is an equilateral triangle with sides of length 2.
D
A B
E
√
√ BC = AC = AD = AE = C D = 2, while B D = D E =
In this configuration
E B = EC = 1 + 3. Assuming a 4-coloring of the plane contrary to the claim, the
four points A, C, D, and E have different colors, as do the four points B, C, D, and
E. Therefore A and B have the same color. Since A and B were arbitrary points with
distance 2, the entire circle centered at A with radius 2 has just one color. This gives
√ a
contradiction, because√there are two points on that circle separated by distance 2.
Next
√ consider P( 2, 4). Assume a 4-coloring of the plane inconsistent√ with
P( 2, 4), that is, one √ in which
√ any two points with distance 1 or 2 have differ-
√ colors. By P((1 + 3)/ 2, 4), just proved, since any two points with distance
ent √
2 have different colors, there must exist two points X and T with distance 1 + 3
having the same color; call it color 1. Let X Y Z T U V be a convex hexagon such that
points Y and V are symmetric with respect to line X T , points Z and U are symmetric
with respect to line X T , and X Y V and T U Z are equilateral triangles with side 1.
Now Y , Z, U , and V all have colors different from color 1, so some two of them have
the same color.
√ This gives a contradiction, because Y ZU V forms a square with side 1
and diagonal 2.
Solution to (c) by √Matthew Huddleston, Washington State University, Pullman, WA.
We prove P((1 + 5)/2, 5). Suppose √ there is a 5-coloring of the plane so that no two
points with distance 1 or (1 + 5)/2 have the same color. Let S be a set of five points
forming the vertices of a regular pentagon with side 1. Let Q be the set of ordered
quintuples chosen from S, so that Q has 55 elements. Color Q by assigning to each of
its 5-tuples the color, in the original coloring, of the sum of its five entries.
For any 4-tuple of points in S, note that adding the sum of its entries to each of the
five points of S gives a regular pentagon with side 1, so these five points have different
colors. Therefore, each of the five colors is assigned to 54 of the 55 elements of Q. On
the other hand, permuting an ordered 5-tuple in Q cannot change its color. The number

of permutations of a given quintuple is a multinomial coefficient of the form

5 5!
= ,
a1 , a2 , a3 , a4 , a5 a1 ! a2 ! a3 ! a4 ! a5 !
where a j is the number of occurrences in the quintuple of the jth element of S. This
multinomial coefficient is a multiple of 5 except for the cases in which one of the a j is
5 and the rest are 0. In order for the sizes of all the color classes in Q to be multiples
of 5, these 5 exceptional cases must all be assigned the same color. Equivalently, the
points of 5S all have the same color. This shows that in any regular pentagon with side
5 all vertices have the same color, so that any two vertices with distance 5 have the
same color. An isosceles triangle with sides of length 5, 5, 1 thus has vertices of the
same color, and that is a contradiction.
√ √
Editorial comment.All solvers for Part (b) found the same example: (1 + 3)/ 2 =
√ √ √
( 6 + 2)/2 = 2 + 3 = (1/2) csc(π/12) = 2 sin(5π/12) = 2 cos(π/12). Is
there another t > 1 such that P(t, 4) holds?
Parts (a) and (b) also solved by W. C. Calhoun, R. Stong, the GCHQ Problem Solving Group (U. K.), and
Microsoft Research Problems Group. Part (a) also solved by O. P. Lossers (Netherlands) and the proposer.
Ex-To-In-Radius Ratio
11240 [2006, 655]. Proposed by Pál Péter Dályay, Deák Ferenc High School, Szeged,
Hungary. Let a, b, and c be the lengths of the sides of a triangle, and let R and r be
the circumradius and inradius of that triangle, respectively. Show that

R (a − b)2 (b − c)2 (c − a)2
≥ exp + + .
2r 2c2 2a 2 2b2
Solution by A. K. Shafie, IASBS, Zanjan, Iran. Write A, B, C for the angles opposite
sides a, b, c, repectively, and s for the semiperimeter. We have
A A B C A B C
r = (s − a) tan = a sec sin sin = 4R sin sin sin ,
2 2 2 2 2 2 2
so sin(A/2) sin(B/2) sin(C/2) = r/(4R). Now e ≤ 1/(1 − x) for all x ∈ [0, 1).
x
Since 0 ≤ (a − b)2 /c2 < 1 and

(a − b)2 a 2 + b2 − 2ab cos C − a 2 − b2 + 2ab 4ab sin2 (C/2)
1− = = ,
c2 c2 c2
we have

(a − b)2 (a − c)2 (b − c)2
exp + +
c2 b2 a2
c2 b2 a2 R2
≤ = 2.
4ab sin (C/2) 4ac sin (B/2) 4bc sin (A/2)
2 2 2
4r
This is the square of the required inequality.
Also solved by A. Alt, S. Amghibech (Canada), M. R. Avidon, O. Bagdasar (Romania), E. Braune (Austria), P.
De (Ireland), Y. Dumont (France), O. Faynshteyn (Germany), S. Hitotumatu (Japan), A. Ilić & M. Novaković
(Serbia), K. W. Lau (China), O. P. Lossers (Netherlands), D. Lovit, M. Mabuchi (Japan), J. Minkus, J. Rooin &
F. Ghanimat (Iran), V. Schindler (Germany), A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), M. Tetiva
(Romania), M. Vowe (Switzerland), J. Zacharias, B. Zhao, BSI Problems Group (Germany), Szeged Problem
Solving Group “Fejéntaláltuka” (Hungary), Microsoft Research Problems Group, and the proposer.
172
Iterational Rate of Convergence
11244 [2006, 759]. Proposed by Jolene Harris and Bogdan Suceavă, California State
University, Fullerton, CA. Let f be a differentiable function from the positive reals to
the positive reals with the property that f (x) < x for all x. Suppose that x1 > 0, and
for n > 1 let xn = f (xn−1 ). Suppose further that limn→∞ xn = 0 and that there exist
positive numbers a and k such that
x a − ( f (x))a 1
lim = a.
x→0 x ( f (x))
a a k
(a) Prove that limn→∞ n 1/a xn = k.
(b) Suppose that 0 < x1 < 1, and specialize to the case where f is given
√ by f (x) =
√
sin x if x < π/2 and f (x) = 1 if x ≥ π/2. Show that limn→∞ xn n = 3.
(c) Finally, suppose instead that 0 < x1 < 1 and f (x) = 1 − e−x . Show that, in this
case, limn→∞ nxn = 2.
Solution by Knut Dale, Telemark University College, Bø, Norway. (a) We are given
f (x)−a − x −a = k −a + (x), where (x) → 0 as x → 0. So
n−1

1 1 1 1 n−1
− a = a − a = + (x1 ) + · · · + (xn−1 ) and
xna x1 i=1
x i+1 x i ka
1 1 n − 1 (x1 ) + · · · + (xn−1 ) 1
a
= a + a
+ → a
nxn nx1 nk n k
as n → ∞. Therefore n 1/a xn → k.
(b) Let a = 2. Compute
x 2 − sin2 x x 2 − sin2 x 1
lim 2
= lim 4
=
2
x→0 x sin x x→0 x 3
√
by L’Hopital’s rule (or Maclaurin series). Hence k = 3.
(c) Let a = 1. Compute
x − (1 − e−x ) 1
lim =
x→0 x(1 − e−x ) 2
by writing e−x = 1 − x + x 2 /2 − x 2 ω(x) where ω(x) → 0 as x → 0. Hence k = 2.
Editorial comment. Note that the differentiability of f is not needed.
Also solved by S. Amghibech (Canada), M. R. Avidon, O. Bagdasar (Romania), P. Bracken, D. R. Bridges, R.
Chapman (U. K.), P. P. Dályay (Hungary), J.-P. Grivaux (France), J. H. Lindsey II, O. López & N. Caro (Brazil),
M. McMullen, M. D. Meyerson, J. Rooin & A. Morassaei (Iran), K. Schillling, H.-J. Seiffert (Germany),
N. C. Singer, A. Stadler (Switzerland), A. Stenger, R. Stong, M. Tetiva (Romania), D. Văcaru (Romania),
J. Vinuesa (Spain), Z. Vörös (Hungary), BSI Problems Group (Germany), GCHQ Problem Solving Group
(U. K.), Microsoft Research Problems Group, NSA Problems Group, Northwestern University Math Problem
Solving Group, and the proposers.


that address before July 31, 2008. Additional information, such as generaliza-
PROBLEMS
11348. Proposed by Richard P. Stanley, Massachusetts Institute of Technology, Cam-
bridge, MA. A polynomial f over a field K is powerful if every irreducible factor of f
has multiplicity at least 2. When q is a prime or a power of a prime, let Pq (n) denote
the number of monic powerful polynomials of degree n over the finite field Fq . Show
that for n ≥ 2,
Pq (n) = q n/2 + q n/2−1 − q (n−1)/3 .
11349. Proposed by Cezar Lupu (student), University of Bucharest, Bucharest, Roma-

nia. In triangle ABC, let h a denote the altitude to the side BC and let ra be the exradius
relative to side BC, which is the radius of the circle that is tangent to BC and to the
extensions of AB beyond B and AC beyond C. Define h b , h c , rb , and rc similarly. Let
p, r , R, and S be the semiperimeter, inradius, circumradius, and area of ABC. Let ν
be a positive number. Show that
ν
ν ν ν ν ν ν ν ν ν ν ν ν ν 3p
2(h a ra + h b rb + h c rc ) ≤ ra rb + rb rc + rc ra + 3S .
4R + r
11350. Proposed by Bhavana Deshpande, Poona College of Arts, Science & Com-
merce Camp, Pune, India, and M. N. Deshpande, Institute of Science, Nagpur, In-
dia. Given a positive integer n and an integer k with 0 ≤ k ≤ n, form a permutation
a = (a1 , . . . , an ) of (1, . . . , n) by choosing the first k positions at random and filling
the remaining n − k positions in ascending order. Let E n,k be the expected number of
left-to-right maxima. (For example, E 3,1 = 2, E 3,2 = 11/6, and E 4,2 = 13/6.) Show
that E n+1,k − E n,k = 1/(k + 1). (A left-to-right maximum occurs at k when a j < ak
for all j < k.)
Bârlad, Romania. Given positive integers p and q, find the least positive integer m
such that among any m distinct integers in [− p, q] there are three that sum to zero.
262
11352. Proposed by Daniel Reem, The Technion-Israel Institute of Technology, Haifa,
Israel. Let I be an open interval containing the origin, and let f be a twice-
differentiable function from I into R with continuous second derivative. Let T2 be
the Taylor polynomial of order 2 for f at 0, and let R2 be the corresponding remain-
der. Show that
R2 (u) − R2 (v)
lim √ = 0.
(u,v)→(0,0) (u − v) u 2 + v 2
u =v
11353. Proposed by Ernst Schulte-Geers, √ BSI, Bonn, Germany. For s > 0, let f (s) =
∞ s −x
0
(1 + x/s) e d x and g(s) = f (s) − sπ/2. Show that g maps R+ onto (2/3, 1)
and is strictly decreasing on its domain.
11354. Proposed by Matthias Beck, San Francisco State University, San Francisco,
CA, and Alexander Berkovich, University of Florida, Gainesville, FL. Find a polyno-
mial f in two variables such that for all pairs (s, t) of relatively prime positive integers,

s−1
t−1
|mt − ns| = f (s, t).
m=1 n=1
SOLUTIONS
Unsolved in 1990
6576 [1986, 1036]. Proposed by Hans V. Gerber, University of Lausanne, Switzerland.
Suppose X 1 , X 2 , . . . are independent identically distributed real random variables with
E(X k ) = μ. Put Sk = X 1 + X 2 + · · · + X k for k = 1, 2, . . . .
(a) If ρ < μ < 1, where ρ = −0.278465 . . . is the real root of xe1−x = −1, show
that the series
∞
Skk e−Sk /k!
k=1
converges with probability one.

(b) If X 1 , X 2 , . . . are positive and if μ < 1, show that the expectation of

∞
Skk e−Sk /k!
k=1
is μ/(1 − μ).
∗
(c) In (b) is it possible to relax the condition that the random variables are positive?
For example, would it suffice to assume E(|X k |) < ∞ and ρ < μ < 1?
Solution to (c) by Daniel Neuenschwander, Université de Lausanne, Lausanne, and
Universität Bern, Bern, Switzerland. We prove the validity of (c) under the additional
assumption that supp (X 1 ) [the support of X 1 , i.e., the intersection of all closed subsets
A of the real line for which P(X 1 ∈ A) = 1] is contained in the open interval (ρ, −ρ).
Let λ = sup{|x| : x ∈ supp (X 1 )}. Note that λ < −ρ. By Stirling’s formula, one sees
that
(kλ)k (nλ)n
< ∞. (1)
k,n≥0
k! n!
March 2008] PROBLEMS AND SOLUTIONS 263

Thus by Lebesgue’s Dominated Convergence Theorem, the expectation of the series
displayed in (a) is given by the absolutely convergent sum
(−1)n
E(Skk+n ).
k≥1, n≥0
k!n!
By approximation, we may assume that X 1 has finite support:

P(X 1 = z j ) = p j ( j = 1, 2, . . . , h)

where p1 , . . . , ph are positive, hj=1 p j = 1, and z j ∈ (ρ, −ρ) for 1 ≤ j ≤ h. Now
let p1 , . . . , ph be fixed. The required equation
μ
E(. . . ) = (2)
1−μ

[where (. . . ) is the series of (a) and μ = hj=1 p j z j ] can be viewed as an equality of
two functions in h real variables z 1 , . . . , z h on the open cube (ρ, −ρ)h . By (1), the
left side of (2) extends as a complex analytic function in h variables to the domain D h ,
where D = {z ∈ C : |z| < −ρ}. The same holds for the right side of (2). By (b), (2)
holds on the subcube (0, −ρ)h of C, and by standard methods of complex analysis, it
thus holds also on D h . This proves (c) in the asserted case.
Editorial comment. Solutions for (a) and (b) were published in the December, 1990,
issue of this M ONTHLY (pages 930–932).
A Determinant Identity
11242 [2006, 656 & 848]. Proposed by Gerd Herzog and Roland Lemmert, Univer-
sität Karlsruhe, Karlsruhe, Germany. (corrected) Let f and g be entire holomorphic
functions of one complex variable, and let A and B be complex n × n matrices. If
the application of such a function to a matrix means applying the power series of this
function to the matrix, prove that
det ( f (A) f (B) + g(A)g(B)) = det ( f (B) f (A) + g(B)g(A)) .
Solution by Roger A. Horn, University of Utah. The crucial property of these matrix
functions is that f (Z) and g(Z) commute whenever both are defined. The following
result is key. (See D. Carlson et al., Linear Algebra Gems, MAA, 2002, p. 13.)
Lemma. Let C, D, X, Y be n × n complex matrices. If C commutes with X , then

C Y
det = det(C D − X Y ).
X D
Proof. If C is nonsingular, then

C Y I 0 C Y

C Y

det = det = det
X D −XC −1 I X D 0 D − XC −1 Y
= det C · det(D − XC −1 Y ) = det(C D − C XC −1 Y )
= det(C D − XCC −1 Y ) = det(C D − X Y ).

(If C is singular, we invoke the foregoing result for C + ε I and take limits as ε → 0.)
264
Let

f (A) g(B) f (B) −g(A) 0 1
M= , N= , and P= .
−g(A) f (B) g(B) f (A) 1 0
Since M = P N P, det M = (det P)2 det N = det N . Because f (A) commutes with
−g(A) and f (B) commutes with g(B), the lemma ensures that
det( f (A) f (B) + g(A)g(B)) = det M = det N = det( f (B) f (A) + g(B)g(A)).
Also solved by S. Amghibech (Canada), R. Chapman (U. K.), K. Dale (Norway), G. Dospinescu (France), H.
Flanders, J. Grivaux (France), E. A. Herman, J. H. Lindsey II, O. P. Lossers (Netherlands), K. Schilling, R.
Stong, BSI Problems Group (Germany), GCHQ Problem Solving Group (U. K.), Microsoft Research Problems
Group, and the proposers.
A Bessel Function Identity

11246 [2006, 760]. Proposed by Lee Goldstein, Wichita, KS. Let Jn be the nth Bessel
function of the first kind, and let K n be the nth modified Bessel function of the second
kind (also known as a Macdonald function), defined by

(|n| + 1/2)(2z)|n| ∞ cos t
K n (z) = √ dt.
π 0 (t + z 2 )|n|+1/2
2
Show that, for any positive b and any real λ,

√
π 2 ∞ √
√ eλ −2b = J2n (4λ b )K n+1/2 (2b).
2 b −∞
Solution by Mourad E. H. Ismail, University of Central Florida, Orlando, FL. First

we note formula (4.10.2) from M. E. H. Ismail, Classical and Quantum Orthogonal
Polynomials in One Variable, Cambridge University Press, 2005:
√ −x −1/2 n
(−n)k (n + 1)k
K n+1/2 (x) = πe x (−2x)−k , n = 0, 1, · · · ,
k=0
k!

where (a)0 := 1 and (a)n := n−1 k=0 (a + k) for n > 0. Since K ν (x) = K −ν (x) and
J−n (x) = (−1)n Jn (x), the series on the right side of the identity to be proved is
√
π ∞ √ n
(−n)k (n + 1)k
√ e−2b J2n (4λ b ) (4b)−k .
b n=0 k=0
k!
Formula (9.0.1) in the same reference is

∞
(zw)m ∞
(−z)n ∞
bn+r zr n
(−n)k (n + γ )s
am bm = as w s .
m=0
m! n=0
n!(γ + n) n r =0
r !(γ + 2n + 1)r s=0
s!
This, together with

∞
(−1)s x 2s+2n
J2n (2x) = ,
s=0
s!(2n + s)!
proves the identity claimed.

Also solved by R. Chapman (U. K.), J. Grivaux (France), F. Holland (Ireland), G. Lamb, A. R. Miller, V.
Schindler (Germany), A. Stadler (Switzerland), V. Stakhovsky, R. Stong, BSI Problems Group (Germany),
GCHQ Problem Solving Group (U. K.), and the proposer.
A Strip Problem
11247 [2006, 760]. Proposed by Jürgen Eckhoff, University College London, London,
U. K. Let A, B, C, and D be distinct points in the plane with the property that any
three of
√ them can be covered by some strip of width 1. Show that there is a strip of
width
√ 2 covering all four points, and demonstrate that if no strip of √
width less than
2 covers all four, then the points are the corners of a square of side 2. (A strip of
width w is the closed set of points bounded by two parallel lines separated by distance
w.)
Solution by Li Zhou, Polk Community College, Winter Haven, FL. Let S be the set
{A, B, C, D}. If the convex hull of S is a triangle, then that triangle is covered by
a strip of width 1, and so is S. It thus suffices to assume that ABC D is a convex
quadrilateral. Given an arbitrary triangle X Y Z, let h X denote the altitude of X Y Z at
vertex X . For X, Y, Z ∈ S, by the assumed property of ABC D, min{h X , h Y , h Z } ≤ 1.
√
Lemma. Consider a triangle X Y Z√with X, Y, Z ∈ S. If h X ≥ 2, then√ X ≤ 45◦ .
Equality holds if and only if h X = 2, h Y = 1, and Y = 90◦ , or h X = 2, h Z = 1,
and Z = 90◦ .
Proof. Either h Y ≤ 1 or h Z ≤ 1, say h√ Y ≤ 1. By the Law of Sines, sin X ≤

sin X/ sin Y = Y Z/ X Z = h Y / h X ≤ 1/ 2. Thus X ≤ 45◦ .
Now assume that ABC D is labeled clockwise. Without loss of generality, assume that
−→ −→ −→ −→
rays B A and C D do not intersect, and similarly C B and D A do not intersect. The
minimum width of all strips that cover ABC D is the minimum length of the perpen-
diculars √
from A to BC, A to C D, B to AD, and D to AB. Suppose this minimum
exceeds 2. Applying the lemma repeatedly, taking X Y Z to be ABC, AC D, B AD,
and D AB, we conclude that all of B AC, C AD, AB D, and AD B are less than
45◦ . This contradicts the fact that these angles sum to 180◦ as internal
√ angles of tri-
angle AB√ D. Thus ABC D may be covered by a strip of width 2. If the minimum
equals 2, then each of the four angles must equal 45◦ . By the conditions for equality
in the lemma, it follows that ABC D is a square.
Also solved by E. A. Herman, J. H. Lindsey II, B. Schmuland (Canada), R. Stong, M. Tetiva (Romania), GCHQ
Problem Solving Group (U. K.), and the proposer.
Double Sum Inequality

11250 [2006, 847]. Proposed by Sun Wen Cai, Pinggang Middle School, Shenzhen,
Guangdong Province, China. Show that if n is a positive integer and x1 , . . . , xn are
nonnegative real numbers that sum to 1, then
n
√
n
1 n2
xj √ ≤√ √ .
j =1 k=1 1 + 1 + 2xk n+ n+2
Solution
√ by Vitaly Stakhovsky,
√ Redwood City, CA. Let n > 1, 0 ≤ x ≤ 1, x0 = 1/n,
β = 1 + 2x0 , f (x) = nx, and
√ √
n+ n+2 1+β
g(x) = √ √ = √ .
n + n + 2nx 1 + 1 + 2x
266
We want to prove that

1 n
1 n
f (xi ) g(xi ) ≤ 1.
n i=1 n i=1
Expand f (x) and g(x) by Taylor’s theorem:

√ f (x) = 1 + f (x0 )(x − x0 ) − φ(x)/2,
where f (x0 ) = n/2 and φ(x) = ( nx − 1)2 ; and g(x) = 1 + g (x0 )(x − x0 ) +

ψ(x)/2, where
√
( 1 + 2x − β)2 2 1 √ 2
ψ(x) = √ + ≤ 1 + 2x − β
1+β 1 + 1 + 2x β
√ √ 2
√ 2 1 + 2x + 1 + 2x0 2 √ 2
≤ 1 + 2x − β √ √ = nx − 1 ≤ φ(x).
2x + 2x0 n
n
Using i=1 (xi − x0 ) = 0, we obtain:

1 n
1 n
1 n
1 n
f (xi ) g(xi ) = 1 − φ(xi ) 1+ ψ(xi )
n i=1 n i=1 2n i=1 2n i=1

1 n
1 n
≤ 1− φ(xi ) 1+ φ(xi ) ≤ 1.
2n i=1 2n i=1
Also solved by D. R. Bridges, G. Crandall, P. P. Dályay (Hungary), K.-W. Lau (China), J. H. Lindsey II, O. P.
Lossers (Netherlands), B. Schmuland (Canada), A. Stenger, R. Stong, J. Sun, L. Zhou, GCHQ Problem Solving
Group (U. K.), Microsoft Research Problems Group, and the proposer.
An Inequality Proved Without Computer Assistance

11251 [2006, 847]. Proposed by Marian Tetiva, National College “Gheorghe Roşca
Codreanu”, Bârlad, Romania. Suppose that a, b, and c are positive real numbers, two
of which are less than or equal to 1, and ab + ac + bc = 3. Show that
1 1 1 3 3(a − 1)(b − 1)(c − 1)
+ + − ≥ .
(a + b) 2 (a + c) 2 (b + c) 2 4 2(a + b)(a + c)(b + c)
Solution by Vitaly Stakhovsky, Redwood City, CA. The inequality follows from
1 1 1 3 (a − 1)(b − 1)(c − 1) 3
+ + ≥ + (1)
a+b b+c c+a 2 (a + b)(b + c)(c + a) 2
and x 2 + 1/4 ≥ x with x = 1/(a + b), 1/(b + c), and 1/(c + a). Multiplying by the
positive factor (a + b)(b + c)(c + a), we see that (1) is equivalent to
(b + c)(c + a) + (a + b)(c + a) + (a + b)(b + c)
3 3
≥ (a − 1)(b − 1)(c − 1) + (a + b)(b + c)(c + a). (2)
2 2
Using S1 , S2 , and S3 for the symmetric expressions a + b + c, ab + bc + ca, and abc,
inequality (2) becomes
3 3 3
S12 + S2 ≥ (S3 − S2 + S1 − 1) + (S1 S2 − S3 ) = (S1 − 1)(S2 + 1). (3)
2 2 2

Now by hypothesis S2 = 3, so finally (3) is equivalent to S12 + 3 ≥ 6(S1 − 1). This is
equivalent to (S1 − 3)2 ≥ 0, which is always true.
Also solved by D. Beckwith, P. Bracken, J.-P. Grivaux (France), J. H. Lindsey II, K. McInturff, T.-L. Rădulescu
& V. Rădulescu (Romania), V. Schindler (Germany), R. Stong, S. Wagon, T. R. Wilkerson, L. Zhou, GCHQ
Problem Solving Group (U. K.), Microsoft Research Problem Solving Group, Northwestern University Math
Problem Solving Group, and the proposer.
A Productive Inequality
11252 [2006, 847]. Proposed by Ovidiu Bagdasar, Babeş Bolyai University, Cluj-
Napoca, Romania.Let
n n be an integer greater than 2 and let a1 , . . . , an be positive
n
numbers. Let S = i=1 ai . Let bi = S − ai , and let S = i=1 bi . Show that
n n
ai bi
n i=1 ≤ n i=1 .
i=1 (S − ai ) i=1 (S − bi )
Solution by Marian Tetiva, National College “Gheorghe Roşca Codreanu”, Bârlad,

Romania. We begin with two lemmas.
Lemma 1. Let m be a positive integer, and let a, a1 , . . . , am be positive. Then
1/m
(a + a1 ) · · · (a + am ) (a1 · · · am )1/m
≥
ma + a1 + · · · + am a1 + · · · + am
Proof. Apply Jensen’s inequality to the concave function f (x) = ln(x/(a + x)) to get
(a1 + · · · + am )/m 1 m
ak
ln ≥ ln
a + (a1 + · · · + am )/m m k=1 a + ak
which is equivalent to the claim.
Lemma 2. Let a, a1 , . . . , am be positive and let A = a1 + · · · + am . Then

(a + A − a1 ) · · · (a + A − am ) ≥ a + (m − 1)a1 · · · a + (m − 1)am
Proof. By the AM-GM inequality we have

(a + A − a1 )m−1 ≥ a + (m − 1)a2 · · · a + (m − 1)am
and m − 1 similar inequalities. Multiply them together to get the claimed result.
Now for the problem proposed, using Lemma 2 and then Lemma 1 gives
1/(n−1) 1/(n−1)
(S − a1 ) · · · (S − an−1 ) (S − a1 ) · · · (S − an−1 )
=
S − bn (n − 1)an + (n − 2)a1 + · · · + (n − 2)an−1
1/(n−1)
(an + (n − 2)a1 ) · · · (an + (n − 2)an−1 )
≥
(n − 1)an + (n − 2)a1 + · · · + (n − 2)an−1
1/(n−1) 1/(n−1)
((n − 2)a1 ) · · · ((n − 2)an−1 ) a1 · · · an−1
≥ = .
(n − 2)a1 + · · · + (n − 2)an−1 a1 + · · · + an−1
Thus we have
1/(n−1) 1/(n−1)
(S − a1 ) · · · (S − an−1 ) a1 · · · an−1
≥
S − bn a1 + · · · + an−1
and n − 1 similar inequalities. Multiply these to get the required result.
268
Also solved by P. P. Dályay (Hungary), O. P. Lossers (Netherlands), B. Schmuland (Canada), R. Stong, and
the proposer.
A Myth About Infinite Products

11257 [2006, 939]. Proposed by Raimond Struble, Santa Monica, CA. Let z n be a
sequence of complex numbers, and let sn = nk=1 z k . Suppose that all sn are nonzero.
(a) Given that sn does not tend to zero, show that ∞ n=1 z n /sn converges if and only if
limn→∞ sn exists.
(b) Show that if sn tends to a limit s, and s − sn is never zero, then ∞ k=1 z n /(s − sn−1 )
diverges.
Solution by Jean-Pierre Grivaux, Paris, France. We reduce both parts to a commonly-
believed (but false) myth about infinite products. In fact, both parts are false, in general.
(a) Write
zn sn − sn−1 sn−1
λn = = =1− . (1)
sn sn sn
Since the sums sn are nonzero, λn = 1 and
s1
sn = . (2)
(1 − λ2 )(1 − λ3 ) · · · (1 − λn )
So
the question has been reduced to: n
λn converges if and only if lim√ n→∞ k=2 (1 − λk ) exists and is nonzero.
This
is not true. Let λ
n = (−1) n
/ n. By (2), this defines sn , and by (1) z n . The series
λn converges, but nk=2 (1 − λk ) → 0, since

(−1)k (−1)k 1 1
log 1 − √ =− √ − +O .
k k 2k k 3/2
√
Alternatively, if λn =√ i(−1)n /n, then λn converges,
but nk=2 (1 − λk ) → ∞. If
λn = (1 + i)(−1)n / n, then λn converges, but nk=2 (1 − λk ) diverges while re-
maining bounded.
(b) Let rn = s − sn−1 . If
zn rn − rn+1 rn+1
μn = = =1− , (3)
rn rn rn
then
rn+1 = r1 (1 − μ1 )(1 − μ2 ) · · · (1 − μn ). (4)
Since limn→∞sn = s, limn→∞ rn = 0. This
reduces the problem to:
If limn→∞ nk=1 (1 − μk ) = 0,√then μk diverges.
This is not true, as μn = (−1)n / n shows.
Also solved by D. Borwein (Canada), J. H. Lindsey II, O. P. Lossers (The Netherlands), P. Perfetti (Italy), A.
Stadler (Switzerland), R. Stong, BSI Problems Group (Germany), and GCHQ Problem Solving Group (U.K.).


that address before August 31, 2008. Additional information, such as general-
izations and references, is welcome. The problem number and the solver’s name
PROBLEMS
11355. Proposed by Jeffrey C. Lagarias, University of Michigan, Ann Arbor, MI. De-
termine for which integers a the Diophantine equation
1 1 1 a
+ + =
x y z x yz
has infinitely many integer solutions (x, y, z) such that gcd(a, x yz) = 1.
11356. Proposed by Michael Poghosyan, Yerevan State University, Yerevan, Armenia.
Prove that for any positive integer n,
n 2
n
24n (n!)4
k
2n = .
k=0 (2k + 1) 2k
(2n)!(2n + 1)!
11357. Proposed by Mehmet Şahin, Ankara, Turkey. Let Ia , Ib , Ic and ra , rb , rc be

respectively the excenters and exradii of the triangle ABC. If ρa , ρb , ρc are the inradii
of triangles Ia BC, Ib C A, and Ic AB, show that
ρa ρb ρc
+ + = 1.
ra rb rc

Bârlad, Romania. Let d be a square-free positive integer greater than 1. Show that
there are infinitely many positive integers n such that dn 2 + 1 divides n!.
11359. Proposed by J. Monterde, University of Valencia, Valencia, Spain. Find an ex-
interval (0, 1) of the fam-
plicit parametric formula for the geometric envelope on the
ily of Bernstein polynomials Bsn (t), defined by Bsn (t) = ns t s (1 − t)n−s for s ∈ [0, n],

where ns = (s+1)(n−s+1)
n!
.
11360. Proposed by Cezar Lupu, student, University of Bucharest, Bucharest, and Tu-
dorel Lupu, Decebal High School, Constanta, Romania. Let f and g be continuous
April 2008] PROBLEMS AND SOLUTIONS 365

1
real-valued functions on [0, 1] satisfying the condition 0 f (x)g(x) d x = 0. Show that
1 21 2 1 2 1 2 1 2 1 2 1 2 1 2
1 1
0
f 0
g ≥ 4 0
f 0
g and 0
f 0
g + 0
g 0
f ≥ 4 0
f 0
g .
11361. Proposed by Finbarr Holland, University College Cork, Cork, Ireland. The
Lemoine point of a triangle is the unique point L inside the triangle such that the
distances from L to the sides are proportional to the corresponding side lengths.
Given a circle G and distinct fixed points B, C on G, let K be the locus of the
Lemoine point of ABC as A traverses the circle. Show that K is an ellipse.
SOLUTIONS
A Vanishing Alternating Sum

11212/11220 [2006, 268/367]. Proposed by David Beckwith, Sag Harbor, NY. Show
that when n is a positive integer,

n
r n 2n − 2r
(−1) = 0.
r =0
r n−1

k+n−1 k
Solution I by Harris Kwong, SUNY, Fredonia, NY. From (1 − t)−n = ∞ t

n k=0 n−1
and (1 − t ) = r =0 (−1) r t , we obtain
2 n r n 2r
∞
(1 − t 2 )n n
r n k + n − 1 k+2r
(1 + t)n = = (−1) t .
(1 − t)n k=0 r =0
r n−1
Comparing coefficients of t n+1 yields the desired identity.

Solution II by Rob Pratt, Raleigh, NC. More generally, if k < n, then

n
n 2n − 2r
(−1)r = 0.
r =0
r k
Since k < n, no k-person committees can be formed from n couples using at least one
member of each couple. To count these committees using inclusion-exclusion, let Ai
be the set of k-person committees using neither member of couple i. Now i∈S Ai =
2n−2|S|
k
. Hence the sum on the left is the inclusion-exclusion sum to count the empty
set.

Solution III by Nicholas Singer, Annandale, VA. Indeed, rn=0 (−1)r nr P(r ) = 0 for
any polynomial P of degree less than
n. The nth forward
difference of the sequence
(P(0), P(1), . . . ) has first term (−1)n rn=0 (−1)r nr P(r ). Since P is a polynomial of
degree less than n, its nth forward difference is identically zero.
Sometimes an Integer
11213 [2006, 268]. Proposed by Stanley Rabinowitz, Chelmsford, MA. For positive

integers n and m with n odd and greater than 1, let S(n, m) = (n−1)/2
k=1 sec2m ( n+1
kπ
).
(a) Show that if n is one less than a power of 2, then S(n, m) is a positive integer.
(b∗ ) Show that if n does not have the form of part (a), then S(n, m) is not an integer.
366
Solution to (a) by NSA Problems Group, Fort Meade, MD. Let n + 1 = 2s with s ≥ 2.
We use induction on s. The base case s = 2 is easy: sec2m (π/4) = 2m . Note that
2
s −1
2s−1
kπ (2 j − 1)π
S(2 s+1
− 1, m) = 2m
sec = S(2 − 1, m) +
s
sec2m
.
k=1
2s+1 j =1
2s+1
By the induction hypothesis, it suffices to show that the last sum is integral.
The Chebyshev polynomial Tr is a polynomial of degree r defined by Tr (x) =
cos(r arccos(x)). In particular, T2s is a polynomial of degree 2s with T2s (x) = 0 when
x = cos((2 j − 1)π/2s+1 ) for 1 ≤ j ≤ 2s . Furthermore, since T2 (x) = 2x 2 − 1 and
T2s (x) = 2 cos2 (2s−1 arccos(x)) − 1 = 2T2s−1 (x)2 − 1,
it follows that T2s (x) ∈ Z[x] and, apart from the constant term, all of its coeffi-
cients are even. Since T2s (0) = cos(2s π/2) = 1 for s ≥ 2, the reciprocal polynomial
s
x 2 T2s (1/x) is a monic integral polynomial that has a zero at each of the 2 S numbers
xk := sec((2 j − 1)π/2s+1 ), 1 ≤ j ≤ 2s . Furthermore, every elementary symmetric
function in 2 S variables evaluates to an even integer at (x1 , . . . , x2S ). Now

2s−1
1 1
s S
(2 j − 1)π 2
(2 j − 1)π 2
sec 2m
= sec2m = x 2m ,
j =1
2s+1 2 j =1 2s+1 2 j =1 j
and from the foregoing observations, this is an integer, as desired.

Editorial comment. Unfortunately, due to an error by the editors, part (b∗ ) of this prob-
lem was incorrectly stated. It should have read, “Show that if n does not have the form
of part (a), then there exists a positive integer m such that S(n, m) is not an integer.”
Several solvers noticed that part (b∗ ) as stated is false, with a simple counterexample
being S(9, 1) = 16. The intended statement of part (b∗ ) remains untreated.
Also solved by D. R. Bridges, R. Chapman (U. K.), P. P. Dalyay (Hungary), Y. Dumont (France), J. W. From-
meyer, J. Grivaux (France), G. Keselman, O. P. Lossers (The Netherlands), M. A. Prasad (India), A. Stadler
(Switzerland), A. Stenger, R. Stong, M. Tetiva (Romania), A. Tissier (France), BSI Problems Group (Ger-
many), GCHQ Problem Solving Group (U. K.), and the proposer.
The Number Between 1 and n That is Least Prime

11218 [2006, 366]. Proposed by Gary Gordon, Lafayette College. Consider the follow-
ing algorithm, which takes as input a positive integer n and proceeds by rounds, listing
in each round certain positive integers between 1 and n inclusive, ultimately producing
as output a positive integer f (n), the last number to be listed. In the 0th round, list 1.
In the first round, list, in increasing order, all primes less than n. In the second round,
list in increasing order all numbers that have not yet been listed and are of the form
2 p, where p is prime. Continue in this fashion, listing numbers of the form 3 p, 4 p,
and so on until all numbers between 1 and n have been listed. Thus f (10) = 8 because
the list eventually reaches the state (1, 2, 3, 5, 7, 4, 6, 10, 9, 8), while f (20) = 16 and
f (30) = 27.
(a) Find f (2006).
(b) Describe the range of f .
(c) Find limn→∞ f (n)/n and limn→∞ f (n)/n.
Solution by Bruce S. Burdick, Roger Williams University, Bristol, RI. In describing the
first round of the process, “less than n” should be “at most n”.

(b) Let R be the range of f , and let S = {2i 3 j : i is a nonnegative integer and j = 0
or 3 · 2l−1 ≤ 3 j < 2l+1 for some l ∈ N}. We show that R = S.
Let g(1) = 0. For k ≥ 2, let g(k) = k/ p, where p is the largest prime divisor of k.
When k ≤ n, the number k appears in round g(k). Thus f (n) is the largest element of
[n] that maximize g.
Let T = {k ∈ N : g(k) ≥ g(i) for 1 ≤ i < k}. If k ∈ T , then f (k) = k; hence
T ⊆ R. Furthermore, f (n) = k requires k ∈ T , so R = T . To compare T and S, we
show first that a number in T has no prime divisor outside {2, 3}. If g(k) = k/ p with
p > 4, then there is a positive integer m such that 2k/ p < 2m < k. Now g(2m ) =
2m−1 > k/ p = g(k) and k ∈ / T.
Now consider the exponents. Note that T contains all powers of 2, since g(2m ) =
2m−1 > g(k) for k < 2m . If 2i 3 j ∈ T with j
= 0, then 3 j −1 ≥ 2l−1 whenever 2l < 3 j ;
otherwise 2i+l < 2i 3 j and g(2i 3 j ) < g(2i+l ). Letting 2l be the largest power of 2 less
than 3 j , we obtain 3 · 2l−1 ≤ 3 j < 2l+1 . Hence T ⊆ S. Finally, if 2i 3 j ∈ S, then 2i 3 j is
listed after all smaller numbers divisible by 3 and after all smaller powers of 2. Hence
S ⊆ R, and we have R = S.
The condition for 2i 3 j ∈ S when j
= 0 is that the second bit from the left in the
binary expansion of 3 j is 1. It is also equivalent to the fractional part of j log2 3 being
at least log2 3. The first few values of j with this property are 1, 3, 5, 8, 10, 13, 15, 18,
and 20.
(a) f (2006) = 1944. To find f (2006) using (b), we create a list of candidates by
taking powers of 2 and numbers of the form 2i 3 j where j ∈ {1, 3, 5} and 2i is the
largest power of 2 bounded by 2006/3 j . The candidates are 1024, 1536, 1728, and
1944. The largest, 1944, is f (2006).
(c) limn→∞ f (n)/n = 1. When n ∈ R, the value of f (n)/n is 1. There are infinitely
many such n, and f (n) ≤ n for all n, so limn→∞ f (n)/n = 1.
limn→∞ f (n)/n = 2/3. When n = 3 · 2i − 1, the value of f (n) is 2i+1 , so f (n)/n =
2 /(3 · 2i − 1) = 2/(3 − 2−i ). These numbers tend to 2/3. If 2i+1 ≤ n ≤ 3 · 2i − 1,
i+1
then f (n) = 2i+1 , and f (n)/n > 2/3. Similarly, if 3 · 2i ≤ n ≤ 2i+2 − 1, then f (n) ≥
3 · 2i , and f (n)/n > 3/4. Hence limn→∞ f (n)/n = 2/3.
Solved also by M. R. Avidon, D. Beckwith, P. Corn, D. Cranston, D. Fleischman, J. W. Frommeyer, J.-P. Gri-
vaux (France), C. C. Heckman, E. A. Herman, M. Hildebrand, M. Huibregtse, G. Keselman, J. H. Lindsey
II, O. P. Lossers (The Netherlands), D. Lovit, F. B. Miles, D. Opitz, W. Y. Pong, M. A. Prasad (India),
A. Stadler (Switzerland), R. Staum, J. H. Steelman, R. Stong, M. Tetiva (Romania), L. Zhou, BSI Problem
Solving Group (Germany), Szeged Problem Solving Group “Fejéntaláltuka” (Hungary), GCHQ Problem Solv-
ing Group (U. K.), Manchester Problem Solving Group (U. K.), Microsoft Research Problems Group, and
Northwestern University Problem Solving Group.
A Romanian Olympiad Problem Generalized
11248 [2006, 760]. Proposed by Pál Péter Dályay, Deák Ferenc High School, Szeged,
Hungary. Let n be a positive integer, and let f be a continuous real-valued func-
1
tion on [0, 1] with the property that 0 x k f (x) d x = 1 for 0 ≤ k ≤ n − 1. Prove that
1
0
( f (x))2 d x ≥ n 2 . (The case n = 2, due to Ioan Raşa, appeared on the 55th Roma-
nian Mathematical Olympiad, and the solution was published in Gazeta Matematică
(CIX) 5-6 (2004), page 227.)
Solution by Jaime Vinuesa, University of Cantabria, Santander, Spain. The clas-

sical Legendre polynomial Pk is a polynomial of degree k given by 2k k!Pk (x) =
368
(d k /d x k )[(x 2 − 1)k ]. These polynomials satisfy Pk (1) = 1 and
⎧
1 ⎪
⎨0 if j
= k,
P j (x)Pk (x) d x = 2
−1 ⎪
⎩ if j = k.
2k + 1
√
Let Q k (x) = 2k + 1 Pk (2x − 1). 1 Now {Q 0 , Q 1 , . . . , Q n−1 } is√an orthonormal set
in L 2 [0, 1]. Moreover, f, Q k = 0 f (x)Q k (x) d x = Q k (1) = 2k + 1, so Bessel’s
inequality yields
1
n−1
n−1
( f (x))2 d x = f 2 ≥ | f, Q k |2 = (2k + 1) = n 2 .
0 k=0 k=0
Editorial comment. As some contributors noted, the inequality is sharp, and it holds
for all f ∈ L 2 [0, 1]. Some solvers observed that it suffices to construct a polynomial p
1
of degree at most n − 1 satisfying 0 ( p(x))2 d x = 1 and p(1) = n. The desired result
then follows from the Schwarz inequality using p, f = p(1). Various devices—e.g.,
orthogonal polynomials, Lagrange interpolation, Hilbert matrices—were used in con-
structing such a polynomial.
Also solved by U. Abel (Germany), O. J. L. Alfonso (Colombia), S. Amghibech (Canada), K. F. Andersen
(Canada), A. Bandeira & E. Dias (Portugal), M. Benito & Ó. Benito & E. Fernández (Spain), M. W. Botsko,
P. Bracken, D. R. Bridges, M. A. Carlton, R. Chapman (U. K.), A. Chaudhuri & R. Selukar, P. R. Chernoff,
K. Dale (Norway), Y. Dumont (France), B. Dunn III, J. Fabrykowski & T. Smotzer, S. Foucart, J.-P. Grivaux
(France), J. Groah, E. A. Herman, F. Holland (Ireland), E. J. Ionascu (Romania), M. E. H. Ismail, D. Jespersen,
J. Kalis, G. Keselman, O. Kouba (Syria), A. V. Kumchev, K.-W. Lau (China), J. H. Lindsey II, O. P. Lossers
(The Netherlands), R. Martin (U. K.), W. Matysiak (Poland), T. Nowak (Austria), M. A. Prasad (India), H.
Ricardo, J. Rooin & M. Hassani (Iran), H.-J. Seiffert (Germany), J. G. Simmonds, N. C. Singer, A. Stadler
(Switzerland), V. Stakhovsky, A. Stenger, R. Stong, J. Sun, M. Tetiva (Romania), S. Vagi, Z. Vörös (Hun-
gary), R. Whitley, L. Zhou, BSI Problems Group (Germany), Szeged Problem Solving Group “Fejéntaláltuka”
(Hungary), GCHQ Problem Solving Group (U. K.), Microsoft Research Problems Group, and the proposer.
A Symmetric Inequality
11255 [2006, 848]. Proposed by Slavko Simic, Mathematical Institute SANU, Bel-
grade, Serbia. Let n be a positive integer, x1 , . . . , xn be real numbers,

n and k let
p1 , . . . , pn be real numbers summing to 1. For k ≥ 1, let Sk = i=1 pi xi −

n k
i=1 pi x i . Show that for m ≥ 1,

1
S2m S2m+2 ≥ 1 − S2 .
m(2m + 1) 2m+1
Solution by O. P. Lossers, Eindhoven University of Technology, Eindhoven, The

Netherlands. We assume that p1 , . . . , pn are nonnegative real numbers summing
to 1, because if p = (1, −3, 3) and x = (1, 2, 3), then S2 S4 < (2/3)S32 and the in-
equality fails.
Define f : R2 → R by
t 2 x 2m+2 2t x 2m+1 x 2m
f (x, t) = + + .
(2m + 2)(2m + 1) 2m(2m + 1) 2m(2m − 1)
∂2
Now ∂x2
f (x, t) = (t x m + x m−1 )2 ≥ 0, so f (x, t) is a convex function of x for all t.

From Jensen’s inequality it then follows that, for all t,

n
n
pi f (xi , t) − f pi xi , t ≥ 0,
i=1 i=1
or, equivalently,
t 2 S2m+2 2t S2m+1 S2m
+ + ≥ 0.
(2m + 2)(2m + 1) 2m(2m + 1) 2m(2m − 1)
The expression on the left, considered as a polynomial in t, must therefore have a
nonpositive discriminant:
2
2S2m+1 S2m+2 S2m
−4 · ≤ 0,
2m(2m + 1) (2m + 2)(2m + 1) 2m(2m − 1)
and this is equivalent to the desired inequality.
Editorial comment. The requirement that pi ≥ 0 was given by the proposer but lost
by the editors. A more subtle oversight is the unstated requirement that m should be
an integer. Byron Schmuland provided a counterexample with m = 3/2, x1 = x2 = 1,
x3 = −1, and p1 = p2 = p3 = 1/3, in which case S3 S5 < (5/6)S42 .
Also solved by B. Schmuland (Canada), Microsoft Research Problems Group, and the proposer.
Perimeter by Integral
11256 [2006, 848]. Proposed by Finbarr Holland, University College Cork, Cork, Ire-
land. For complex a, b, and c, let f (x) = max{Re(aei x ), Re(bei x ), Re(cei x )}. Find
2π
0
f (x) d x.
2π
Solution by Michel Bataille, Rouen, France. Let I = 0 f (x) d x. We show that I =
|a − b| + |b − c| + |c − a|; this is the perimeter of the triangle formed by the three
complex numbers.
Let g(x) = min{Re(aei x ), Re(bei x ), Re(cei x )}. Now
2π 2π
g(x) d x = − max{−Re(aei x ), −Re(bei x ), −Re(cei x )} d x
0 0
2π
=− max{Re(aei(x+π) ), Re(bei(x+π) ), Re(cei(x+π) )} d x
0
3π
=− max{Re(aeit ), Re(beit ), Re(ceit )} dt = −I,
π
since f is 2π-periodic. Now max{u, v} = 12 (u + v + |u − v|) with u, v ∈ R and

2π i x 2π
0
e d x = 0, so with the notation Ja,b = 0 max{Re(aei x ), Re(bei x )} d x it follows
that
2π
1 1 2π
Ja,b = Re (aei x + bei x ) d x + |Re(aei x ) − Re(bei x )| d x
2 0 2 0
2π
1
= |Re(aei x ) − Re(bei x )| d x.
2 0
370
Now let a = α + iα and b = β + iβ where α, α , β, and β are real numbers. Then
Re(aei x ) − Re(bei x ) = (α − β) cos x − (α − β ) sin x = A sin(x + φ),
where A = ((α − β)2 + (α − β )2 )1/2 = |a − b|, sin φ = (α − β)/A, and cos φ =
−(α − β )/A. Thus,
2π 2π
[Re(aei x ) − Re(bei x )| d x = |a − b| | sin(x + φ)| d x
0 0
2π
= |a − b| | sin x| d x = 4|a − b|,
0
2π
and Ja,b = 2|a − b|. Similarly, with K a,b = 0 min{Re(aei x ), Re(bei x )} d x, we have
K a,b = −2|a − b|. Finally, for real numbers u, v, and w, we have
1
max{u, v, w} − min{u, v, w} = (max{u, v} + max{v, w} + max{w, u}
2
− min{u, v} − min{v, w} − min{w, u}),
so
2π
1
2I = (Ja,b + Jb,c + Jc,a − K a,b − K b,c − K c,a )
( f (x) − g(x)) d x =
0 2
1
= 2|a − b| + 2|b − c| + 2|c − a| + 2|a − b| + 2|b − c| + 2|c − a|
2
= 2|a − b| + 2|b − c| + 2|c − a|.
This is the stated result.
Also solved by S. Amghibech (Canada), M. R. Avidon, D. R. Bridges, B. S. Burdick, R. Chapman (U. K.),
A. Demis (Greece), J.-P. Grivaux (France), E. A. Herman, F. Hjouj, G. Keselman, J. H. Lindsey II, M. D.
Meyerson, J. G. Simmonds, V. Stakhovsky, R. Stong, R. Tauraso (Italy), E. I. Verriest, J. Vinuesa (Spain),
H. Widmer (Switzerland), J. H. Zacharias, L. Zhou, Szeged Problem Solving Group “Fejéntaláltuka” (Hun-
gary), GCHQ Problem Solving Group (U. K.), Microsoft Research Problems Group, and the proposer.
A Choice Sum
11263 [2006, 940]. Proposed by Gregory Keselman, Oak Park, MI, and formerly of
Lvov Polytechnic Institute, Ukraine. Show that when n is a positive integer and a is
real,
⎧ √ √
⎪ (1+ 1−4a)n+1√ −(1− 1−4a)n+1

n/2 ⎨ , if a < 1/4;
n−k 2 n+1 1−4a
(−1)k a k = (n + 1)/2n , if a = 1/4;
k ⎪
⎩ n/2 √
k=0
a cos(nβ) + sin(nβ)/ 4a − 1 , if a > 1/4.
√
Here, β denotes arcsin 1 − 1/(4a).
Solution by E. Leonard, University of Alberta, Edmonton, AL, Canada. For each non-
negative integer n, define

n/2
n−k n
k k n−k
bn = (−1) a
k k
= (−1) a .
k=0
k k=0
k

m+1 m m
Using r
= r
+ m
r −1
and r
= 0 if r > m, we see that the sequence bn satisfies
b0 = 1, b1 = 1, bn+1 = bn − abn−1 , n ≥ 1.
This linear recurrence has characteristic equation λ − λ + a = 0, which has roots:
2
1 1√ 1
λ= ± 1 − 4a if a < ,
2 2 4
1 1
λ= if a = ,
2 4
1 1√ 1
λ= ±i 4a − 1 if a > .
2 2 4
(i) For a < 1/4 there are two distinct roots, so the general solution of the recurrence
is
√ n √ n
1 + 1 − 4a 1 − 1 − 4a
bn = A +B .
2 2
We now compute A and B from the initial conditions b0 = b1 = 1 to arrive at
√ n+1 √ n+1
1 + 1 − 4a − 1 − 1 − 4a
bn = √ .
2n+1 1 − 4a
(ii) For a = 1/4 there is a double root, so the general solution of the recurrence is
A Bn
n
+ n. bn =
2 2
Again using the initial conditions to determine A and B, we obtain
n+1
. bn =
2n
(iii) For a > 1/4, there are two distinct complex roots, which may be written in
polar form as
√ √ √ √
a cos β + i sin β = aeiβ , and a cos β − i sin β = ae−iβ ,
√
where β = arcsin 1 − 1/(4a). The general solution of the recurrence is
bn = Aa n/2 einβ + Ba n/2 e−inβ .
Applying the initial conditions to evaluate A and B, and using
√
4a − 1 1
sin β = √ , cos β = √ ,
2 a 2 a
we get
a n/2
bn = a n/2 cos nβ + √ sin nβ.
4a − 1
Also solved by U. Abel (Germany), S. Amghibech (Canada), M. R. Avidon, M. Bataille (France), D. Beckwith,
J. Borwein (Canada), P. Bracken, R. Chapman (U.K.), P. P. Dályay (Hungary), P. De (Ireland), G.C. Greubel,
J.-P. Grivaux (France), H. Kwong, G. Lamb, O. P. Lossers (The Netherlands), J. Minkus, C. R. Pranesachar
(India), T. L. & V. Râdulescu (Romania), H. Roelants (Belgium), J. N. Senadheera, A. Stadler (Switzerland),
R. Stong, M. Tetiva (Romania), H. Widmer (Switzerland), C. Zhou (Canada), BSI Problems Group (Germany),
GCHQ Problem Solving Group (U.K.), Microsoft Research Problems Group, and the proposer.
372

that address before September 30, 2008. Additional information, such as gen-
eralizations and references, is welcome. The problem number and the solver’s
name and address should appear on each solution. An asterisk (*) after the num-
ber of a problem or a part of a problem indicates that no solution is currently
available.
PROBLEMS
11362. Proposed by David Callan, University of Wisconsin, Madison, WI. A bit string
arc diagram is an undirected graph in which the vertices are the positions in a single
string of bits and the edges are called arcs due to the visual representation in which
they are drawn joining positions in the string. To be a good diagram, arcs must occur
only between unequal bits, and each bit may be the left endpoint of at most one arc.
Thus, the first diagram is good but, for two reasons, the second is not.
1 1 0 0 1 1 0 1 1 0 0 1 1 0
There are six good diagrams on two bits, four with no arc and two with a single arc.
How many good diagrams are there on n bits?
11363. Proposed by Oleh Faynshteyn, Leipzig, Germany. Let m a , m b , and m c be the
lengths of the medians of a triangle T . Similarly, let Ia , Ib , Ic , h a , h b , and h c be the
lengths of the angle bisectors and altitudes of T , and let R, r , and S be the circumra-
dius, inradius, and area of T . Show that
Ia Ib Ib Ic Ic Ia
+ + ≥ 3(2R − r ),
Ic Ia Ib
and
m a Ib m b Ic m c Ia √
+ + ≥ 35/4 S.
hc ha hb
11364. Proposed by Pál Péter Dályay, Szeged, Hungary. Let p be a prime greater than
3, and t the integer nearest p/6.
(a) Show that if p = 6t + 1, then

2t−1
1 1
( p − 1)! (−1) j
+ ≡ 0 (mod p).
j =0
3j + 1 3j + 2
May 2008] PROBLEMS AND SOLUTIONS 461

(b) Show that if p = 6t − 1, then

(−1) j
2t−1 (−1) j
2t−2
( p − 1)! + ≡0 (mod p).
j =0
3j + 1 j =0
3j + 2
11365. Proposed by Aviezri S. Fraenkel, Weizmann√ Institute of Science, Rehovot,

Israel. Let t be a positive integer. Let γ = t 2 + 4, α = 12 (2 + γ − t), and β =
1
2
(2 + γ + t). Show that for all positive integers n,
nβ = (nα + n(t − 1))α + 1 = (nα + n(t − 1) + 1)α − 1.
11366. Proposed by Nicolae Anghel, University of North Texas, Denton, TX. Let φ :
R → R be a continuously differentiable function such that φ(0) = 0 and φ is strictly
increasing. For a > 0, let Ca denote the space of all continuous functions from [0, a]
a
into R, and for f ∈ Ca , let I ( f ) = x=0 (φ(x) f (x) − xφ( f (x))) d x. Show that I has
a finite supremum on Ca and that there exists an f ∈ Ca at which that supremum is
attained.
√
11367.
Proposed by Andrew Cusumano, Great Neck, NY. Let x = 1 + 2, x2 =
√ 1
1 + 2 1 + 3, and in general, let xn+1 be the number obtained by replacing the
innermost expression
√ (1 + (n + 1)) in the nested square root formula for xn with
1 + (n + 1) 1 + (n + 2). Show that
xn − xn−1
lim = 2.
n→∞ xn+1 − xn
11368. Proposed by Wei-Dong Jiang, Weihai Vocational College, Weihai, ShanDong,

China. For a triangle of area 1, let a, b, and c be the lengths of its sides. Let s =
(a + b + c)/2. Show that the weighted average of (s − a)2 , (s − b)2 , and (s − c)2 ,
weighted√by the radian measure of the angles opposite a, b, and c respectively, is at
least π/ 3.
SOLUTIONS
Generating Functions and Hypergeometric Series and Continued Fractions!

11198 [2006, 79]. Proposed by P. R. Parthasarathy, India Institute of Technology
Madras, Chennai, India. Let f (k) = 1 + k if k is odd, f (k) = 1 + k/2 if k is even.
Show that

0
1+i 1
1+i 2
1+i n−1
n
m

f (i 1 ) f (i 2 ) f (i 3 ) · · · f (i n ) = (−1)m−k mk k n .
i 1 =0 i 2 =0 i 3 =0 i n =0 m=1 k=1
Composite solution by the proposer and the editors. We begin by forming a generating
function from each side. Multiply by (−1)n q n , sum over n ≥ 1, and use 1 for the
constant term. It suffices to show that the two resulting generating functions are the
same. We start with the one on the left.
462
We first convert the power series to a continued fraction. With a = (a0 , a1 , . . . ), let
pn (a) be the coefficient of q n in the expression
1 1 a0 q a1 q ∞
= ··· = (−1)n pn (a)q n . (1)
a0 q 1+ 1+ 1+
1+ n=0
1 + a1 q
..
1+ .

We prove by induction on n that pn (a) is the sum of nj =1 ai j over all n-tuples
i 1 , . . . , i n such that i 1 = 0 and 0 ≤ i j ≤ i j −1 + 1 for 2 ≤ j ≤ n. This is precisely our
generating function when ak = f (k) for k ≥ 0.
For n = 0, we have the empty product and coefficient 1. For n ≥ 1, let a =
(a1 , a2 , . . . ). By the induction hypothesis,

∞
∞
(−1)n pn (a)q n 1 + a0 q (−1)k pk (a )q k = 1,
n=0 k=0
n
By examining the coefficient of q on both sides, it follows that

n−1
pn (a) = a0 p (a) pn−1− (a ). (2)
=0
Group the terms in the claimed formula pn (a) according to the maximum index
j such that i j = 0. By the induction hypothesis, the sum of the terms for which
the last position with subscript 0 in the n-tuple is position l + 1 can be written as
p (a)a0 pn−1− (a ). Hence (2) completes the proof of the formula for pn (a).
The sequence from which we form the continued fraction can be expressed as
a2r −2 = r and a2r −1 = 2r for r ≥ 1. Let s = 1/q. Our next aim is
k
1 a0 /s a1 /s 1 ∞
s 1
··· = . (3)
1+ 1+ 1+ 2 k=0 s + k 2
On the right is the hypergeometric series 12 2 F1 (s, 1; 1 + s; 12 ); in standard notation,

a (k) b(k) z n (k)
2 F1 (a, b; c; z) = k≥0 c(k) n! , where x is the rising factorial. We seek a continued
fraction expansion converging to the same value as the left side.
It is convenient to multiply both sides of (3) by 1/s. The continued fraction on the
left converges when s has positive real part. A continued fraction converges to the same
value as its even part, which is the limit of its even approximants (see L. Lorentzen and
H. Waadeland, Continued Fractions with Applications, North-Holland, 1992, p. 84).
The even part is
1/s a0 a1 /s 2 a2 a3 /s 2
··· .
1 + a0 /s− 1 + a1 /s + a2 /s− 1 + a3 /s + a4 /s−
With a2r −2 a2r −1 = 2r 2 and a2r −1 + a2r = 3r + 1, this becomes
1/s 2 · 12 /s 2 2 · 22 /s 2 2 · r 2 /s 2
··· ··· .
1 + 1/s− 1 + 4/s− 1 + 7/s− 1 + (3r + 1)/s−
This continued fraction is equivalent, in the sense of having the same sequence of
approximants, to
1 2 · 12 2 · 22 2 · r2
··· ··· .
s + 1− s + 4− s + 7− s + 3r + 1−

To obtain a continued fraction expansion for 2s1 2 F1 (s, 1; s + 1; 12 ), we use the con-
tinued fraction expansion for the standard Euler fraction (see p. 308 of the book cited
above). With c = c − a and b = b − a, we obtain
2 F1 (a, b; c; z)
c· =
F
2 1 (a, b + 1; c + 1; z)
(c + 1)(b + 1)z (c + 2)(b + 2)z (c + 3)(b + 3)z
c + (b + 1)z − ···
c + 1 + (b + 2)z− c + 2 + (b + 3)z− c + 3 + (b + 4)z−
To obtain the desired formula, we set c = a = s, b = 0, and z = 1/2. Since
s (k) 0(k) (1/2)k
2 F1 (s, 0; s; 2 ) =
1
k≥0 s (k) k!
= 1 and c + r + (b + r + 1)z = (s + 3r + 1)/2,
the previous display reduces to
s (1)(1)/2 (2)(2)/2 (3)(3)/2
= s + (1 − s)/2 − ···
2 F1 (s, 1; s + 1; 2 )
1
(s + 4)/2− (s + 7)/2− (s + 10)/2−

1 2 · 12 2 · 22 2 · 32
= s+1− .
2 s + 4− s + 7− s + 10−
Taking the reciprocal of the continued fraction yields

1 1 1 2 · 12 2 · 22 2 · 32
F
2 1 s, 1; s + 1, = .
2s 2 s + 1− s + 4− s + 7− s + 10−
We have thus proved (3).
1
k
It remains only to convert 12 ∞ s
k=0 s+k 2 into the generating function we obtained
from the right side of the problem statement. Written using q instead of s, the steps
(explained below) are
k ∞ k
1 ∞
1 1 1 ∞ 1
= (−1) qn n
kn
2 k=0 1 + kq 2 2 n=0 k=0
2
∞ k
1 ∞ 1 n
=1+ (−1) qn n
S(n, m)k(m)
2 n=1 k=0
2 m=1
∞ n ∞ n m
m−k m
=1+ (−1) q
n n
m!S(n, m) = 1 + (−1) q
n n
(−1) kn.
n=1 m=1 n=1 m=1 k=1
k
In step 1 we expand (1 + kq)−1 by the geometric series and interchange the order
of summation. In step 2 we invoke k n = S(n, m)k(m) , where S(n, m) is the Stirling
number of the second kind (the number of partitions of n positions into m unordered
blocks) and k(m) is the falling factorial. Both sides count the k-ary n-tuples, on the right
∞ k
−k
grouped by the number of distinct entries. In step 3 we use the identity k=0 m 2 =
k
k
2; this is the instance of ∞ k=0 m x = x /(1 − x)
m m+1
at x = 1/2, where the series
converges. Finally, in step 4 we use the inclusion-exclusion formula for the Stirling
numbers.
No solutions were received.
Groups, Rings, Fields, and Power Series

11216 [2006, 366]. Proposed by Ted Chinburg, University of Pennsylvania, Philadel-
phia, PA, and Shahriar Shahriari, Pomona College, Claremont, CA. Let K be a field,
464
let γG be an ordered Abelian group. The support Supp(a) of a formal sum a =
and
γ aγ t with coefficients aγ in K and exponents γ in G is the set {γ ∈ G : aγ = 0}.
The generalized power series ring K ((G ≤0 )) is the set of formal sums a for which
Supp(a) is a well-ordered subset of the nonpositive elements of G. Addition and mul-
tiplication in K ((G ≤0 )) are defined in the same way they are for ordinary power se-
ries. Show that K ((G ≤0 )) is Noetherian if and only if either G = {0} or G is order-
isomorphic to Z with the usual ordering. (An ordered Abelian group is an Abelian
group G with a total order ≤ such that a ≤ b implies a + c ≤ b + c for all a, b, c
in G.)
Solution by Reid Huntsinger, Lansdale, PA. Let R = K ((G ≤0 )). Assume G = {0}.
Given a positive element γ0 in G, let I (γ0 ) = {a ∈ R : aγ = 0 if γ > −γ0 }. Always
I (γ0 ) is an additive
subgroup. In fact, I (γ0 ) is an ideal, since a ∈ R and b ∈ I (γ0 )
imply (ab)γ = δ≤0 aδ bγ −δ . The sum is zero unless γ − δ ≤ −γ0 for some δ ≤ 0,
which requires γ ≤ −γ0 . Thus ab ∈ I (γ0 ), making I (γ0 ) an ideal.
On the other hand, if a ∈ R with min Supp(A) = 0 then a is invertible in R. (The
inversion algorithm is essentially the customary algorithm for inverting power series
with real coefficients with nonzero constant coefficient.) As a result, all proper ideals
of R have the form I (γ ) for some positive γ .
If γ0 < γ1 , then I (γ1 ) ⊂ I (γ0 ), and vice-versa. Thus the Noetherian property (no in-
finite ascending chains of ideals) is equivalent to having no infinite descending chains
in G >0 . In other words, R is Noetherian if and only if the positive elements of G are
well-ordered. This latter condition means that the positive elements are all powers of
the smallest positive element. Thus G is order-isomorphic to Z exactly when R is
Noetherian.
Also solved by N. Caro (Brazil), R. Chapman (U. K.), D. Fleischman, Szeged Problem Solving Group “Fején-
találtuka” (Hungary), and the proposers.
An Infinite Product Based on a Base

11222 [2006, 459]. Proposed by Jonathan Sondow, New York, NY. Fix an integer B ≥
2, and let s(n) denote the sum of the base-B digits of n. Prove that
∞
n B + k (−1)s(n) 1
=√ .
n=0 k odd
n B + k + 1 B
0<k<B
Solution for odd B by the GCHQ Problem Solving Group Cheltenham, UK. For odd
B we have n ≡ s(n) mod 2. Therefore we can write (−1)n rather than (−1)s(n) in the
infinite product. Thus the corresponding infinite series

∞ nB + k ∞ 1
(−1)n log = (−1)n+1 log 1 +
n=0 k odd
nB + k + 1 n=0 k odd
nB + k
0<k<B 0<k<B
converges by the classical alternating series test. Hence the product converges to a
finite positive limit. By induction (for example), it follows for each positive integer M
that

2M−1 n B + k (−1)n (2M B)!22M (M!)2
= 2M B
n=0 k odd
nB + k + 1 2 (M B)!(M B)!(2M)!
0<k<B

Using Stirling’s Formula N ! = N N e−N (2π N )1/2 {1 + O(1/N )}, we obtain

2M−1 n B + k (−1)n
= B −1/2 {1 + O(1/M)}.
n=0 k odd
n B + k + 1
0<k<B
Also, for n > 0 we have

n B + k (−1)n 1
1 1
log = log 1 + < < ,
k odd
nB + k + 1 k odd
nB + k k odd
nB + k n
0<k<B 0<k<B 0<k<B
so
2M
n B + k (−1)n 2M−1
n B + k (−1)n
= 1 + O(1/M) .
n=0 k odd
nB + k + 1 n=0 k odd
nB + k + 1
0<k<B 0<k<B
Thus for any positive integer N ,

N
n B + k (−1)n
= B −1/2 (1 + O(1/N )) ,
n=0 k odd
n B + k + 1
0<k<B
which yields the desired result with a good error estimate.

Editorial comment. The case B = 2 was the subject of M ONTHLY Problem E2692 [85
(1978) 48 and 86 (1979) 394–395].
Other even values of B were treated in a paper by J. O. Shallit [J. Number Theory
21 (1985) 128–134]. The case of even B is somewhat more difficult than the case of
odd B, since for even B the exponent (−1)s(n) is a more complicated function of n than
(−1)n . In particular the familiar alternating series test for establishing convergence of
a series is no longer applicable. For a detailed treatment of even B we refer the reader
to Shallit’s paper.
Also solved (at least for odd values of B) by K. L. Bernstein, E. J. Ionascu, K. McInturff, M. A. Prasad (India),
A. E. Stadler (Switzerland), A. L. Stenger, R. Tauraso (Italy), and the BSI Problems Group (Germany).
The Square Root of a Certain Matrix

11224 [2006, 459]. Proposed by Dietrich Trenkler, University of Osnabrück, Os-
nabrück, Germany, and Götz Trenkler, University of Dortmund, Dortmund, Germany.
Let a and b be linearly independent column vectors in R3 . Find a formula in terms of a
and b for a square matrix A such that A2 = bat − abt . (Here, xt denotes the transpose
of x.)
Solution by Patrick Corn, University of Georgia, Athens, GA. Let M = bat − abt . Let
k = a × b (note that k > 0 by linear independence). We compute
M 3 = bat (a · b)2 − bbt (a · b)(a · a) − bat (a · a)(b · b)
+ bbt (a · a)(a · b) − aat (b · b)(a · b) + abt (a · a)(b · b)
+ aat (a · b)(b · b) − abt (a · b)2

= (bat − abt ) (a · b)2 − (a · a)(b · b) ,

so

M 3 + k 2 M = M (a · b)2 − (a · a)(b · b) + a × b2 = 0.
466
Letting A = (2k 3 )−1/2 (M 2 − k M), we have
A2 − M = (2k 3 )−1 (M 4 − 2k M 3 + k 2 M 2 − 2k 3 M)
= (2k 3 )−1 (M − 2k I )(M 3 + k 2 M) = 0,
as desired (here I is the identity matrix). Writing out A in terms of a and b, we obtain
1 t
A= √ (ba − abt )2 − a × b(bat − abt ) .

2a × b3/2
Editorial comment. The restriction to vectors in R3 is unnecessary.

Also solved by S. Amghibech (Canada), C.P. Anilkumar (India), M. Bataille (France), S.J., Bernau, D. R.
Bridges, R. Chapman, (U.K.) P. Corn, K. Dale (Norway), J.A. Grzesik, E.A. Herman, K.J. Heuvers, R.A.
Horn, G.L. Isaacs, G. Keselman, T. Kezlan, J.H. Lindsey II, O.P. Lossers (The Netherlands), P. Magli (Italy),
J. McHugh, K. McInturff, M.D. Meyerson, K. Onneweer, L. Pebody, N.C. Singer, A. Stadler (Switzerland),
R. Stong, J. Stuart, T. Tam, N. Thornber, E.I. Verriest, W.C. Waterhouse, M. Woltermann, J.B. Zacharias, BSI
Problems Group (Germany) GCHQ Problem Solving Group (U.K.), and the proposer.
A New Bound for r/R

11245 [2006, 760]. Proposed by Cezar Lupu, University of Bucharest, Bucharest, Ro-
mania, and Tudorel Lupu, Decebal High School, Constanza, Romania. Consider an
acute triangle with sides of lengths a, b, and c, and with an inradius of r and a circum-
radius of R. Show that

r 2(2a 2 − (b − c)2 )(2b2 − (c − a)2 )(2c2 − (a − b)2 )
≤ .
R (a + b)(b + c)(c + a)
Solution by Ronald A. Kopas, Clarion University, Clarion, PA. Side a is the sum of
the two segments on either side of the point of tangency with the incircle, so

B C cos(B/2) sin(C/2) + cos(C/2) sin(B/2)
a = r cot + cot =r
2 2 sin(B/2) sin(C/2)
r sin(B/2 + C/2) r sin(π/2 − A/2) r cos(A/2)
= = = ,
sin(B/2) sin(C/2) sin(B/2) sin(C/2) sin(B/2) sin(C/2)
and r = a sin(B/2) sin(C/2)/ cos(A/2).
From the law
of sines, a = 2R sin A =
4R sin(A/2) cos(A/2), so R = a/ 4 sin(A/2) cos(A/2) . Therefore
r A B C
= 4 sin sin sin . (1)
R 2 2 2
Now we compute a lower bound for E = 2a 2 − (b − c)2 . We have

E = 2 a 2 − (b − c)2 + (b − c)2

= 2 b2 + c2 − 2bc cos A − b2 − c2 + 2bc + (b − c)2

= 4bc (1 − cos A) + (b − c)2 = (b + c)2 − (b − c)2 (1 − cos A) + (b − c)2

= (b + c)2 (1 − cos A) + (b − c)2 cos A
A A
= (b + c)2 2 sin2 + (b − c)2 cos A ≥ (b + c)2 2 sin2 ,
2 2
noting that cos A ≥ 0 since the triangle is acute.

√ √
Therefore 2 sin(A/2) ≤ √ 2a − (b − c)/(b + c). Similarly 2 sin(B/2) ≤
2 2
2b2 − (c − a)2 /(c + a) and 2 sin(C/2)

√ ≤ 2c2 − (a − b)2 /(a + b). Multiply the
product of these three inequalities by 2 and apply (1) to obtain the required result.
Editorial comment. In fact the result holds for right and obtuse triangles as well. Four
solvers (V. Schindler, R. Stong, R. Tauraso, and the Microsoft Research Problems
Group) provided proofs for this stronger result, but their proofs were all computer
assisted and much longer than the proof given here.
Also solved by A. Alt, S. Amghibech (Canada), O. Bagdasar (Romania), P. P. Dályay (Hungary), P. De (Ire-
land), J. Fabrykowski & T. Smotzer, V. V. Garcı́a (Spain), N. Lakshmanan (India), O. P. Lossers (The Nether-
lands), J. Minkus, C. R. Pranesachar (India), T.-L. & V. Rădulescu (Romania), J. Rooin & A. Emami (Iran),
V. Schindler (Germany), R. Stong, R. Tauraso (Italy), D. Văcaru (Romania), L. Zhou, Microsoft Research
Problems Group, Northwestern University Math Problem Solving Group, and the proposer.
Evaluate the Series

11260 [2006, 939]. Proposed by Paolo Perfetti, Mathematics Department, University
“Tor Vergata,” Rome, Italy. Find those nonnegative values of α and β for which

∞
n
α + k log k
n=1 k=1
β + (k + 1) log(k + 1)
converges. For those values of α and β, evaluate the sum.

Solution by Richard Stong, Rice University, Houston, TX. If α ≥ β, then

∞
n
α + k log k ∞
α
≥ ,
n=1 k=1
β + (k + 1) log(k + 1) n=1
α + k log k
which has all positive terms and diverges. So assume β > α ≥ 0. Now

n
α + k log k β n
α + k log k n+1
α + k log k
= − ,
k=1
β + (k + 1) log(k + 1) β − α k=1 β + k log k k=1 β + k log k
so we have

N −1
n
α + k log k α β N
α + k log k
= − .
n=1 k=1
β + (k + 1) log(k + 1) β − α β − α k=1 β + k log k
∞
Since k=1 (β − α)/(β + k log k) has all positive terms and diverges, we have
∞ ∞
β −α α + k log k
1− = = 0.
k=1
β + k log k k=1
β + k log k
Hence for β > α ≥ 0, we have convergence:

∞
n
α + k log k α
= .
n=1 k=1
β + (k + 1) log(k + 1) β −α
Also solved by O. Furdui, J.-P. Grivaux (France), O. Kouba (Syria), J. H. Lindsey II, O. P. Lossers (The
Netherlands), A. Stadler (Switzerland), Szeged Problem Solving Group “Fejéntaláltuka” (Hungary), GCHQ
Problem Solving Group (U. K.), Microsoft Research Problems Group, and the proposer.
468

that address before October 31, 2008. Additional information, such as general-
PROBLEMS
11369. Proposed by Donald Knuth, Stanford University, Stanford, CA. Prove that for
all real t, and all α ≥ 2,
α
eαt + e−αt − 2 ≤ et + e−t − 2α .
11370. Proposed by Michael Goldenberg and Mark Kaplan, Baltimore Polytechnic In-
stitute, Baltimore, MD. Let A0 , A1 , and A2 be the vertices of a non-equilateral triangle
T . Let G and H be the centroid and orthocenter of T , respectively. Treating all indices
modulo 3, let Bk be the midpoint of Ak−1 Ak+1 , let Ck be the foot of the altitude from
Ak , and let Dk be the midpoint of Ak H .
The nine-point circle of T is the circle through all Bk , Ck , and Dk . We now introduce
nine more points, each obtained by intersecting a pair of lines. (The intersection is not
claimed to occur between the two points specifying a line.) Let Pk be the intersection
of Bk−1 Ck+1 and Bk+1 Ck−1 , Q k the intersection of Ck−1 Dk+1 and Ck+1 Dk−1 , and Rk
the intersection of Ck−1 Ck+1 and Dk−1 Dk+1 .
Let e be the line through {P0 , P1 , P2 }, and f be the line through {Q 0 , Q 1 , Q 2 }.
(By Pascal’s theorem, these triples of points are collinear.) Let g be the line through
{R0 , R1 , R2 }; by Desargues’ theorem, these points are also collinear.
(a) Show that the line e is the Euler line of T , that is, the line through the circumcenter
and centroid of T .
(b) Show that g coincides with f .
(c) Show that f is perpendicular to e.
(d) Show that the intersection S of e and f is the inverse of H with respect to the
nine-point circle.
11371. Proposed by Ovidiu Furdui, University of Toledo, Toledo, OH. Let A denote
the Glaisher-Kinkelin constant, given by
2 /2−n/2−1/12 n 2 /4

n
A = lim n −n e k k = 1.2824 · · · .
n→∞
k=1
June–July 2008] PROBLEMS AND SOLUTIONS 567

Evaluate in closed form

∞
−1 (−1)n
A6 e (1 + 1/n)n .
n=1
11372. Proposed by Jennifer Vandenbussche and Douglas B. West, University of Illi-

nois at Urbana-Champaign, Urbana, IL. In a family of finite sets, let a splitting ele-
ment be an element that belongs to at least two of the sets and is omitted by at least
two of the sets. Determine the maximum size of a family of subsets of {1, . . . , n} for
which there is no splitting element.
11373. Proposed by Emeric Deutsch, Polytechnic University, Brooklyn, NY. Let Sn
be the symmetric group on {1, . . . , n}. By the canonical cycle decomposition of an
element π of S, we mean the cycle decomposition of π in which the largest entry of
each cycle is at the beginning of that cycle, and the cycles are arranged in increasing
order of their first elements.
Let ψn : Sn → Sn be the mapping that associates to each π ∈ Sn the permuta-
tion whose one-line representation is obtained by removing the parentheses
from the
canonical cycle decomposition of π. (Thus, the permutation 12345
34521
has one-line repre-
sentation 34521 and canonical cycle representation (42)(513) and is mapped by ψ5 to
42513.) Describe the fixed points of ψn and find their number.
11374. Proposed by Harley Flanders and Hugh L. Montgomery, University of Michi-
gan, Ann Arbor, MI. Let a, b, c, and m be positive integers such that abcm = 1 + a 2 +
b2 + c2 . Show that m = 4.
11375. Proposed by Cezar Lupu, student, University of Bucharest, Bucharest, Roma-
nia. The first Brocard point of a triangle ABC is that interior point for which the
angles BC, C A, and AB have the same radian measure. Let ω be that measure.
Regarding the triangle as a figure in the Euclidean plane R2 , show that if the vertices
belong to Z × Z, then ω/π is irrational.
SOLUTIONS
Consecutive Squares with a Square Average

11227 [2006, 460]. Proposed by Syrous Marivani, Louisiana State University –
Alexandria, Alexandria, LA. Call an integer n blocky if n > 1 and there is a run
of n consecutive integer squares the average of which is a square. (Thus, 25 is blocky
2
because ( 24 0 k )/25 = 4900/25 = 14 , and 31 is the next blocky integer.)
2
(a) Determine the set B of blocky integers.

(b) Given a blocky integer n, give a procedure that determines all integers k that can
serve as starting points for required runs of squares.
(c) Give a formula in terms of n for the number of k that can serve as starting points
of required runs of squares.
Solution by Missouri State University Problem Solving Group, Springfield, MO.
(a) We prove that n is blocky if and only if n > 1 and n ≡ ±1 or ±7 (mod 24). We
want to find integers k, n, and m such that

1
k+n−1
1
k+n−1
k−1
m2 = i2 = i2 − i2 .
n i=k n i=1 i=1
568
N 2
Using the standard formula i=1 i = N (N + 1)(2N + 1)/6 and algebraically simpli-
fying, we obtain m 2 = k 2 + (n − 1)k + (n − 1)(2n − 1)/6. Thus (n − 1)(2n − 1)/6
must be an integer, which forces n ≡ ±1 (mod 6). Completing the square on the right
side yields m 2 = (k + (n − 1)/2)2 + (n 2 − 1)/12, so

n2 − 1 n−1 2 n−1 n−1

=m − k+
2
= m+k+ m−k− .
12 2 2 2
Since n = 6l ± 1, (n − 1)/2 = 3l or 3l − 1, while (n 2 − 1)/12 = 3l 2 ± l, which is

an even integer. Because m + k + (n − 1)/2 and m − k − (n − 1)/2 have the same
parity and their product is even, we have (n 2 − 1)/12 ≡ 0 mod 4. This forces n ≡ ±1
or ±7 (mod 24). Conversely, if n is of this form, then (n 2 − 1)/12 = (2α)(2β), where
α and β are integers. Taking m = α + β and k = α − β − (n − 1)/2 proves that the
conditions are sufficient.
(b) Given a blocky integer n, taking all possible factorizations of the form (n 2 −
1)/12 = (2α)(2β) with α, β > 0 and setting k = α − β − (n − 1)/2 yields all pos-
sible starting points k for the summation. (Notice that if α = −a and β = −b with
a, b > 0, then k also arises from the positive factorization (n 2 − 1)/12 = (2a)(2b),
because k = α − β − (n − 1)/2 = b − a − (n − 1)/2.)
(c) From the construction above, we conclude that the number of such k is the
number of positive divisors of (n 2 − 1)/48, which is easily computed from its prime
factorization.
Also solved by K. Alpar-Vajk (Italy), M. R. Avidon, D. Beckwith, R. Chapman (U. K.), P. Corn, P. P. Dályay
(Hungary), J. Ferdinands, J. W. Frommeyer, S. M. Gagola, Jr., E. A. Herman, O. P. Lossers (Netherlands),
P. Magli (Italy), L. Pebody, F. Sami, R. Stong, Z. Vörös (Hungary), Con Amore Problem Group (Denmark),
GCHQ Problem Solving Group (U. K.), Microsoft Research Problems Group, NSA Problems Group, and the
proposer.
Inscribed Triangle Area

11261 [2006, 940]. Proposed by Isaac Sofair, Fredericksburg, VA. A triangle of area
1 has vertices A1 , A2 , and A3 . The sides A2 A3 , A3 A1 , and A1 A2 subtend angles of
measure α1 , α2 , and α3 , respectively, at an internal point P. The triangle has angles at
A1 , A2 , and A3 of measure a1 , a2 , and a3 , respectively. The extensions of A1 P, A2 P,
and A3 P to their opposite sides meet those sides at B1 , B2 , and B3 , respectively.
For 1 ≤ k ≤ 3, let Tk = sin ak sin αk sin(αk − ak ). For an even permutation p of
(1, 2, 3), let S p = (sin a p1 sin α p2 sin α p3 + sin α p1 sin a p2 sin a p3 ), and let S be the
product of S p over all three such permutations. Prove that the area of triangle B1 B2 B3
is 2T1 T2 T3 /S.
Solution by C. R. Pranesachar, Indian Institute of Science, Bangalore, India. Notation:
We write, for example, A2 B1 for the length of segment A2 B1 , and [P A2 A3 ] for the
area of P A2 A3 . If t1 = A2 B1 /B1 A3 , t2 = A3 B2 /B2 A1 , and t3 = A1 B3 /B3 A2 , then
t1 t2 t3 = 1 by Ceva’s theorem (see http://www.cut-the-knot.org/Gener-
alization/ceva.shtml). Now
1 1 t3 1
[A1 B3 B2 ] = A1 B3 · A1 B2 sin a1 = A1 A2 · A1 A3 sin a1
2 2 1 + t3 1 + t2
t3
= [A1 A2 A3 ].
(1 + t3 )(1 + t2 )

From this and the two similar relations for [A2 B1 B3 ] and [A3 B2 B1 ] we get
[B1 B2 B3 ] = [A1 A2 A3 ] − [A1 B3 B2 ] − [A2 B1 B3 ] − [A3 B2 B1 ]
2
= [A1 A2 A3 ].
(1 + t1 )(1 + t2 )(1 + t3 )
Subtracting areas, we have
[A1 A2 B1 ] [P A2 B1 ] [P A1 A2 ]
= t1 , = t1 , so = t1 ,
[A1 B1 A3 ] [P B1 A3 ] [P A3 A1 ]
and similarly for t2 , t3 .
A1
Now write
a1 = x1 + x2
a2 = y1 + y2 x1
x2
B2
a3 = z 1 + z 2 α2
B3
P
dividing the angles us- α3
ing lines through P as y2
α1
shown in the diagram. z1
y1 z2
A2 A3
B1
We obtain
sin x1 (A2 P/A1 A2 ) sin α3 sin α3 sin a1
= = .
sin z 2 (A2 P/A2 A3 ) sin α1 sin α1 sin a3
Also x1 + z 2 =
A1 P A3 −
A1 A2 A3 = α2 − a2 . Observe that if sin θ/ sin φ = k and
θ + φ = β, then
k sin β sin β
tan θ = and tan φ = .
1 + k cos β k + cos β
Therefore
sin a3 sin α1 sin(α2 − a2 )
tan z 2 = .
sin a1 sin α3 + sin a3 sin α1 cos(α2 − a2 )
Similarly (interchanging suffixes 2 and 3)
sin a2 sin α1 sin(α3 − a3 )
tan y1 = .
sin a1 sin α2 + sin a2 sin α1 cos(α3 − a3 )
Hence
1 tan y1 tan z 2
[P A2 A3 ] = (A2 A3 )2
2 tan y2 + tan z 2
(A2 A3 )2 sin a2 sin a3 sin α1 sin(α2 − a2 ) sin(α3 − a3 )
=
2 sin a3 sin(α2 − a2 ) {sin a1 sin α2 + sin a2 sin α1 cos(α3 − a3 )}

+ sin a1 sin(α3 − a3 ) {sin a1 sin α3 + sin a3 sin α1 cos(α2 − a2 )}
[A1 A2 A3 ] sin a1 sin α1 sin(α2 − a2 ) sin(α3 − a3 )
= ,
sin a p1 sin a p2 sin α p3 sin(α p3 − a p3 )
570

where is the sum over the even permutations. Writing
x = sin a1 sin α1 sin(α2 − a2 ) sin(α3 − a3 ),
y = sin a2 sin α2 sin(α3 − a3 ) sin(α1 − a1 ),
z = sin a3 sin α3 sin(α1 − a1 ) sin(α2 − a2 ),
we have t1 = z/y, t2 = x/z, t3 = y/x. So
x yz = T1 T2 T3 sin(α1 − a1 ) sin(α2 − a2 ) sin(α3 − a3 ),
y + z = sin(α1 − a1 ) [sin a2 sin α2 sin(α3 − a3 ) + sin a3 sin α3 sin(α2 − a2 )] .
Expanding and rearranging gives
y + z = sin(α1 − a2 ) [sin a2 sin a3 sin α1 + sin a1 sin α2 sin α3 ] = sin(α1 − a1 ) S(123) .
Using this and similar formulas for z + x and x + y, we get
2T1 T2 T3
[B1 B2 B3 ] = [A1 A2 A3 ],
S
as desired.
Also solved by V. Schindler (Germany) and the proposer.
A Combinatorial Inequality
11269 [2007, 78]. Proposed by Hervé Moulin, Rice University, Houston, TX. Let n be a
positive integer and let a1 , . . . , an benonnegative p real numbers. For integers 1 ≤ k ≤

n, and for p ≥ 1, let A(k, p) = S j ∈S a j , where the outer summation extends
p
n
over all subsets of {1, . . . , n} having exactly k elements. Thus, A(n, p) = j =1 a j

while A(1, p) = nj =1 a jp .
(a) Show that if p ≥ n − 1 or p is a positive integer, then
A(n, p) n−1
(−1)k A(n − k, p)
≤ .
n k=0
k+1
(b) Show that if p ≥ n or p is a positive integer, then

n−1
(−1)k A(n − k, p) 1 n+1 p
≤ A(n, p).
k=0
k+1 n+1 n
(c)∗ Does the conclusion of part (a) hold also for nonintegral p with 1 < p < n − 1?
Solution by the editors. (a) More generally, let f be a function differentiable n times,
such that f (x), . . . , f (n) (x) ≥ 0 on (0, ∞) and f (0) = 0. For any subset S of [n] let

aS = as , bS = (−1)|S|−|T | f (aT ).
s∈S T ⊆S
Note that for nonempty S, b S is an |S|-fold iterated finite difference

of f . By the
(|S|)
mean value theorem there is some ξ ∈ (0, s∈S as ) with b S = s∈S as f (ξ ). In
particular, b S ≥ 0 for nonempty S, and b∅ = 0. Also note that by inclusion-exclusion,

f (a S ) = bT .
T ⊆S

Hence
(−1)n−|S| (−1)n−|S| (−1)n−|S|
f (a S ) = bT = bT
S⊆[n]
n + 1 − |S| S⊆[n]
n + 1 − |S| T ⊆S T ⊆[n] S⊃T
n + 1 − |S|
|
n−|T

(−1)k n − |T | bT
= bT =
T ⊆[n] k=0
k+1 k T ⊆[n]
n + 1 − |T |

1 1 n
≥ bT = f aj .
n T ⊆[n] n j =1
Here in the third equality we let k = n − |S|. For the inequality we have used bT ≥ 0
for nonempty T , and b∅ = 0. For part (a) we take f (x) = x p , which has the desired
nonnegativity property exactly when p ≥ n − 1 or p is a positive integer. For use in
part (b), note that similarly
(−1)n−|S| | (−1)k n − |T |
n−|T
f (a S ) = bT
S⊆[n]
n + 2 − |S| T ⊆[n] k=0
k+2 k
bT
= ≥0
T ⊆[n]
(n + 1 − |T |)(n + 2 − |T |)
(b) We show that if f is an n + 1-times differentiable function such that f (0) = 0

and f (x), . . . , f (n+1) (x) ≥ 0 on (0, ∞), then

(−1)n−|S| 1 n+1 n
f (a S ) ≤ f aj .
S⊆[n]
n + 1 − |S| n+1 n j =1
The desired result is then the special case f (x) = x p . Let F(a1 , . . . , an ) denote the
left side of this desired inequality and suppose that an > an−1 . Since
∂F (−1)n−|S|
= f (a S ),
∂ak S:k∈S
n + 1 − |S|
we have
∂F ∂F (−1)n−2−|S|
− =− f (an + a S ) − f (an−1 + a S ) ≤ 0.
∂an ∂an−1 S⊆[n−2]
(n − 2) + 2 − |S|
Here for the inequality we have used the result of the previous paragraph for the func-
tion g given by g(x) = f (an + x) − f (an−1 + x). Since the symmetric result holds
for the other variables, it follows that for fixed nj =1 a j , F attains its maximum when
a1 = · · · = an . Thus it suffices to prove the inequality when a1 = · · · = an = a. If we
take a1 = · · · = an = an+1 = a then

n+1−k n + 1 (−1)n−k (n + 1)
n+1 n
b[n+1] = (−1) f (ka) = f ((n + 1)a) − f (ka).
k=0
k k=0
n+1−k
Since b[n+1] ≥ 0,
n
(−1)n−k 1
f (ka) ≤ f ((n + 1)a).
k=0
n+1−k n+1
572
This is exactly the desired inequality in the case where a1 = · · · = an = a. Note that if
f (x) = x p with n − 1 < p < n, then f (n+1) (x) < 0 on (0, ∞). Hence b[n+1] < 0 and
the reverse inequality holds. Thus (b) does not hold for n − 1 < p < n.
Editorial comment. No solution of (c)∗ is available.
Parts (a) and (b) were also solved by the proposer.
The Euler Line After All

11272 [2007, 164]. Proposed by Vasile Mihai, Belleville, ON, Canada. Let ABC be
an acute nonequilateral triangle, and let H be its orthocenter, O its circumcenter, and
K its symmedian point (defined below). Let R be the circumradius of ABC. Let L be
the line through H and parallel to the line segment K O.
(a) Show that there are exactly two solution points V on L to the equations
|V A| |V B| |V C|
= = . (1)
|BC| |C A| |AB|
(b) For the solutions V1 and V2 in (a), show that |H V1 | · |H V2 | = 4R 2 .

(The symmedian point of a triangle is the point of concurrence of its three symmedian
lines. The symmedian line through A is the reflection of the line through A and the cen-
troid of ABC across the line bisecting angle B AC, and the symmedian lines through
B and C are defined similarly. For more about the symmedian point, see Ross Hons-
berger’s Episodes in Nineteenth and Twentieth Century Euclidean Geometry, MAA,
Washington, DC, 1995.)
Solution by Sin Hitotumatu, Kyoto University, Kyoto, Japan. The points V satisfying
(1) lie on the Apollonius circles

|V A| |BC| |V A| |BC|
V: = and V: = ,
|V B| |C A| |V C| |AB|
so there are at most two such points.
The problem is not correct as stated. We show that the two solution points that
satisfy (1) lie instead on the Euler line H O.
−→ −→ −→
Let |BC| = a, |C A| = b, |AB| = c, O A = a, O B = b, OC = c.
−→ −→ 2
Lemma 1. If O P = αa + βb + γ c, then O P = R 2 (α + β + γ )2 − (a 2 βγ +
b γ α + c2 αβ).
2
Proof. With ·, · denoting the inner product,

−→
O P2 = (α 2 + β 2 + γ 2 )R 2 + 2αβa, b + 2βγ b, c + 2γ αc, a.
Now a, b = a b cos 2C = R 2 (1 − 2 sin2 C) = R 2 − c2 /2 by the law of sines;
−→
similarly for the other two inner products. Therefore, O P2 = R 2 (α 2 + β 2 + γ 2 +
2αβ + 2βγ + 2γ α) − c2 αβ − b2 γ α − a 2 βγ .
Lemma 2. If ABC is an acute non-equilateral triangle, then 8R 2 < a 2 + b2 +
c < 9R 2 .
2
−→
Proof. Since O H = a + b + c, we conclude from Lemma 1 that |O H |2 = 9R 2 −
(a 2 + b2 + c2 ) ≥ 0, with equality only if H = O (that is, ABC is equilateral).
For the left inequality, writing s = (a + b + c)/2, we have
a 2 + b2 + c2 16s 2
= (a 2 + b2 + c2 )
R2 (4Rs)2
a 2 + b2 + c2
= (−a 4 − b4 − c4 + 2a 2 b2 + 2b2 c2 + 2c2 a 2 )
a 2 b2 c2

by Heron’s formula, and then
2
a + b2 + c2
2
− 8 a 2 b2 c2 = (a 2 + b2 + c2 ) − (a 2 + b2 + c2 )2
R
+ 4(a 2 + b2 + c2 )(a 2 b2 + b2 c2 + c2 a 2 ) − 8a 2 b2 c2
= (a 2 + b2 + c2 − 2a 2 )(a 2 + b2 + c2 − 2b2 )(a 2 + b2 + c2 − 2c2 ),
which is positive since ABC is acute.
−→ −→
Now let O V = O H − λ(a + b + c), where λ is a parameter. This is a parametric
description of the Euler line H O, since λ = 0 corresponds to V = H and λ = 1 cor-
−→ −→ −→
responds to V = O. Since AV = O V − O A = −λa + (1 − λ)b + (1 − λ)c, we see
by Lemma 1 that
|AV |2 = R 2 (−λ + 1 − λ + 1 − λ)2 − a 2 (1 − λ)2 + b2 λ(1 − λ) + c2 λ(1 − λ)
= R 2 (2 − 3λ)2 + λ(1 − λ)(a 2 + b2 + c2 ) + (λ − 1)a 2 .
Therefore, if we can take λ satisfying the equation
R 2 (2 − 3λ)2 + λ(1 − λ)(a 2 + b2 + c2 ) = 0, (2)
then we have |AV |2 = (λ − 1)a 2 . Similarly, for the same λ we have |BV |2 =
(λ − 1)b2 and |C V |2 = (λ − 1)c2 , so we obtain the relation (1).
Let p(λ) be the quadratic expression on the left in (2). Then

p(λ) = 9R 2 − (a 2 + b2 + c2 ) λ2 − 12R 2 − (a 2 + b2 + c2 ) λ + 4R 2 .
The leading coefficient 9R 2 − (a 2 + b2 + c2 ), call it ρ, is positive by Lemma 2. The

discriminant
2
12R 2 − (a 2 + b2 + c2 ) − 16R 2 ρ
= (16 − 24)R 2 (a 2 + b2 + c2 ) + (a 2 + b2 + c2 )2
= (a 2 + b2 + c2 )(a 2 + b2 + c2 − 8R 2 )
is positive by Lemma 2. This means that the equation (2) has two distinct roots λ1 and
λ2 . Since p(0) > p(1) > 0, we have λ1 , λ2 > 1. So λ1 and λ2 give the desired points
V1 and V2 , which lie on the Euler line H O. Finally,
4R 2 2
|H V1 | |H V2 | = λ1 λ2 a + b + c2 = 9R − (a 2 + b2 + c2 ) = 4R 2 .
ρ
Also solved by GCHQ Problem Solving Group (U. K.).
574

that address before December 31, 2008. Additional information, such as gen-
eralizations and references, is welcome. The problem number and the solver’s
name and address should appear on each solution. An asterisk (*) after the num-
ber of a problem or a part of a problem indicates that no solution is currently
available.
PROBLEMS
11376. Proposed by Bogdan M. Baishanski, The Ohio State University, Columbus OH.
Given a real number a and a positive integer n, let
1
Sn (a) = √ .
an<k≤(a+1)n kn − an 2
For which a does the sequence Sn (a) converge?

11377. Proposed by Christopher Hillar, Texas A&M University, College Station, TX
and Lionel Levine, Massachusetts Institute of Technology, Cambridge, MA. Given a
monic polynomial p of degree n with complex coefficients, let A p be the (n + 1) ×
(n + 1) matrix with p(−i + j) in position (i, j), and let D p be the determinant of A p .
Show that D p depends only on n, and find its value in terms of n.
11378. Proposed by Daniel Troy, Professor Emeritus, Purdue University Calumet,
Hammond, IN. Let n be a positive integer, and let U1 , . . . , Un be random variables
defined by one of the following two processes:
A: Select a permutation of {1, . . . , n} at random, with each permutation of equal
probability. Then take Uk to be the number of k-cycles in the chosen permutation.
B: Repeatedly select an integer at random from {1, . . . , M} with uniform distribu-
tion, where M starts at n and at each stage in the process decreases by the value of
the last number selected, until the sum of the selected numbers is n. Then take Uk to
be the number of times the randomly chosen integer took the value k.
Show that the probability distribution of (U1 , . . . , Un ) is the same for both processes.
11379. Proposed by Oskar Maria Baksalary, Adam Mickiewicz University, Poznań,
Poland, and Götz Trenkler, Technische Universität Dortmund, Dortmund, Germany.
Let A be a complex matrix of order n whose square is the zero matrix. Show that
R(A + A∗ ) = R(A) + R(A∗ ), where R(·) denotes the column space of a matrix ar-
gument.
664
11380. Proposed by Hugh Montgomery, University of Michigan, Ann Arbor, MI, and
Harold
x S.Shapiro, Royal Institute of Technology, Stockholm, Sweden. For x ∈ R, let
k−1
= 1
j =0 (x − j). For k ≥ 1, let ak be the numerator and qk the denominator of
k k!
the rational number −1/3 k
expressed as a reduced fraction with qk > 0.
(a) Show that qk is a power of 3.
(b) Show that ak is odd if and only if k is a sum of distinct powers of 4.
11381. Proposed by Jesús Guillera, Zaragoza, Spain, and Jonathan Sondow, New York,
NY. Show that if x is a positive real number, then
n 1/n
∞ n
(−1)k+1
e =
x
(kx + 1) k .
n=1 k=0
11382. Proposed by Roberto Tauraso, Università di Roma “Tor Vergata”,

Rome, Italy.
For k ≥ 1, let Hk be the kth harmonic number, defined by Hk = kj =1 1/j. Show that
if p is prime and p > 5, then

p−1
Hk2 p−1
Hk
≡ (mod p 2 ).
k=1
k k=1
k2
(Two rationals are congruent modulo d if their difference can be expressed as a reduced
fraction of the form da/b with b relatively prime to a and d.)
SOLUTIONS
A Group In Which Equations Are Solvable

11231 [2006, 567]. Proposed by Christopher J. Hillar, Texas A&M University, College
Station, TX. Find a non-Abelian group G with the following property: for each positive
integer n, each word W on the alphabet of n + 1 letters A1 , . . . , An and X , each list
a1 , . . . , an of elements of G, and each b in G there exists a unique x in G such that
W (a1 , . . . , an , x) = b. (Thus, in particular, ax 2 ax = b must have a unique solution
x.)
Netherlands. We define such a group on R × R. The group operation, denoted ⊕, is
defined by
(a, α) ⊕ (b, β) = (a + b, α + ea β).
This operation is associative:
((a, α) ⊕ (b, β)) ⊕ (c, γ ) = ((a + b) + c, (α + ea β) + ea+b γ )
= (a + (b + c), α + ea (β + eb γ ))
= (a, α) ⊕ ((b, β) ⊕ (c, γ )).
Clearly (0, 0) is an identity element. Inverses exist, with (a, α)−1 = (−a, −αe−a ). The
operation is not commutative, since in general α + βea = β + αeb .
Assume that W contains at least one X . We seek a unique solution to
W ((a1 , α1 ), . . . , (an , αn ), (x, ξ )) = (b, β).
August–September 2008] PROBLEMS AND SOLUTIONS 665

If W has length l, then (b, β) is the result of l − 1 applications of ⊕. Thus b is the sum
of the first coordinates of the pairs in W , and β is the sum of l terms such that the ith
term is the second coordinate of the ith pair in W times the exponential of the sum of
the first coordinates of the previous pairs in W .
We can thus write (b, β) as (mx + s, σ + μξ ), where m is the number of occur-
rences of X in W , and s is an additive combination of a1 , . . . , an . Thus x = (b − s)/m,
uniquely determined. Furthermore, μ is a sum of exponentials of additive combina-
tions of x and a1 , . . . , an and hence is positive. Also, σ is a linear combination of
α1 , . . . , αn with coefficients depending on x and a1 , . . . , an . Since μ is positive, we
must have ξ = (β − σ )/μ.
Also solved by R. Bagby, R. Chapman (U. K.), M. Goldenberg & M. Kaplan, J. Lockhart, D. Spellman,
R. Stong, C. T. Stretch (U. K.), T. Tam, N. Vonessen, BSI Problems Group (Germany), Szeged Problems Group
“Fejéntalátuka” (Hungary), GCHQ Problem Solving Group (U. K.), Microsoft Research Problems Group, and
the proposer.
A Special Prime
11235 [2006, 568]. Proposed by Lenny Jones and Rachel Keller, Shippensburg Uni-
versity, Shippensburg, PA. Find all primes p such that 128 p + 1 is composite and a
divisor of 264 p − 1.
Solution by Toni Ernvall and Anne-Maria Ernvall-Hytönen, University of Turku, Fin-
land. For a fixed prime p, let d be the smallest positive integer such that 2d ≡ 1
(mod 128 p + 1). Suppose that 128 p + 1 is composite and is a divisor of 264 p − 1.
Note that d | 64 p and d | φ(128 p + 1), where φ is Euler’s function. We divide the
problem into cases according to whether p divides d.
If p | d, then p | φ(128 p + 1). Since p does not divide 128 p + 1, there exists a
prime q such that q ≡ 1 (mod p) and q | 128 p + 1. Write q = r p + 1, so that
128 p + 1 = (r p + 1)(sp + 1) = r sp 2 + r p + sp + 1
for some positive integer s, and r sp + r + s = 128. If p = 2, then there are no so-
lutions, since 257 is prime. Hence p is odd, and r ≡ s ≡ 0 (mod 2). Let r = 2k
and s = 2. Now 2kp + k + = 64, and kp ≤ 31. We may assume that k ≤ . By
straightforward calculations we notice that the only positive integer solution is p = 3,
k = 1, and = 9, yielding 128 · 3 + 1 = 385 = 5 · 7 · 11. However, 11 does not divide
264·3 − 1, so this solution does not satisfy the original conditions.
Assume now that p does not divide d. Now d | 64, and 128 p + 1 | 264 − 1. We
compute
264 − 1 = (232 + 1)(216 + 1)(28 + 1)(24 + 1)(22 + 1)(2 + 1)
= 641 · 6700417 · 65537 · 257 · 17 · 5 · 3.
Checking through the possibilities gives 128 p + 1 = 65537 · 6700417. Thus p =
3430665851 is the only solution.
Also solved by M. Avidon, R. Chapman (U. K.), P. Corn, S. M. Gagola Jr., O. P. Lossers ((Netherlands)),
A. Nakhash, A. Stadler (Switzerland), R. Stong, M. Tetiva (Romania), D. B. Tyler, BSI Problems Group
(Germany), Microsoft Research Problems Group, NSA Problems Group, and the proposers.
An Equality for Parts in Partitions of n

11237 [2006, 655]. Proposed by Emeric Deutsch, Polytechnic University, Brooklyn,
NY. Prove that the number of 2s occurring in all partitions of n is equal to the number
666
of singletons occurring in all partitions of n − 1, where a singleton in a partition is a
part occurring once. (For example, partitions of 5 yield four 2s: one from (3, 2), two
from (2, 2, 1) and one from (2, 1, 1, 1); partitions of 4 yield four singletons: one from
(4), two from (3, 1) and one from (2, 1, 1).)
Solution I by John H. Smith, Needham, MA. Let p(n) denote the number of partitions
of n, with p(0) = 1 and p(n) = 0 for n < 0. The number of partitions of n with at
least i occurrences of j is p(n − i j). Hence the total number of 2s in all partitions of
∞
n is i=1 p(n − 2i), since a partition with exactly k occurrences of 2 is counted once
in each of the first k terms.
On the other hand, a partition of n with a singleton j is formed by appending j to
a partition of n − j that has no j. Since p(n − 2 j) partitions of n − j contain at least
one j, the number of partitions of n with a singleton j is p(n − j) − p(n − 2 j). Thus
the total number of singletons in partitions of n − 1 is ∞ j =1 ( p(n − 1 − j) − p(n −
1 − 2 j)). This sum equals
( p(n − 2) − p(n − 3)) + ( p(n − 3) − p(n − 5)) + ( p(n − 4) − p(n − 7)) + · · · .
Terms p(n − r ) occur
∞positively when r is even and with both signs when r is odd.
Cancellation yields i=1 p(n − 2i), which is the formula obtained for the total number
of 2s.
Solution II by Albert Stadler, Dübendorf, Switzerland. Euler obtained the generating
function for partitions:

∞
∞
1
P(x) = p(n)x n = .
n=1 n=1
(1 − x n )
The number of partitions of n with exactly k occurrences of 2 is the coefficient of x n in

x 2k (1 − x 2 )P(x). Hence the total number of 2s over all partitions of n is the coefficient
of x n in each expression below:

∞
x2 x2
kx 2k (1 − x 2 )P(x) = (1 − x 2
)P(x) = P(x).
k=1
(1 − x 2 )2 (1 − x 2 )
On the other hand, the number of partitions of n − 1 containing the singleton k is the
coefficient of x n−1 in x k (1 − x k )P(x). Hence the total number of singletons over all
partitions of n − 1 is the coefficient of x n in x ∞ k=1 x (1 − x )P(x). Note that
k k
∞

x x2 x2
x x k (1 − x k )P(x) = x − P(x) = P(x).
k=1
1−x 1−x 2 (1 − x 2 )
Thus the coefficients are equal, as claimed.

Also solved by M. Avidon, D. Beckwith, N. Caro (Brazil), R. Chapman (U. K.), J. Dalbac, P. P. Dályay (Hun-
gary), Y. Dumont (France), G. Keselman, M. Kidwell, S. C. Locke, O. P. Lossers ((Netherlands)), P. Massaro
(Italy), R. McCoart, M. D. Meyerson, A. Nijenhuis, I. Novakovic & A. Milan (Serbia), R. Pratt, R. Stong,
R. Tauraso (Italy), J. Taylor-Goodman, M. Tetiva (Romania), W. Watkins, BSI Problems Group (Germany),
Szeged Problems Group “Fejéntalátuka” (Hungary), GCHQ Problem Solving Group (U. K.), Microsoft Re-
search Problems Group, NSA Problems Group, and the proposers.
Sequences Built from the Golden Ratio

11238 [2006, 655]. Proposed by Aviezri S. Fraenkel, Weizmann Institute √
of Science,
Rohovot, Israel. Let n be (n + 1)φ φ − nφ φ, where φ = (1 + 5)/2 and

x denotes the integer part of x. Prove that the following hold for every positive
integer n:
(a)
n is either 22 or 3; 2
(b) (n + 1)φ 2 φ − nφ φ = 2 n − 1;
(c) nφ +
2 nφ = nφ 2
φ ;
(d) nφ φ = nφ φ + 1. 2
Solution by Reiner Martin, Teddington, U.K. Since n − 1 < (nφ − 1)/φ ≤ nφ /φ <
n, we have nφ /φ = n − 1. With φ = 1 + 1/φ, this implies
nφ φ = nφ (1 + 1/φ) = nφ + n − 1.
Thus n = (n + 1)φ − nφ + 1, and so (a) follows since 1 < φ < 2.
Note that nφ
φ = nφ + nφ
/φ ≤ nφ
+ n = n(φ
+ 1) = nφ2 . Thus

nφ ≤ nφ 2
2 2 /φ ≤ nφ, and so nφ 2
/φ = nφ. Thus nφ 2
φ = nφ 2
+
nφ /φ = nφ + nφ, which
is claim (c).
Moreover, it follows that nφ 2 φ = 2 nφ + n. Applying this to n + 1 and to n
and invoking n = (n + 1)φ − nφ + 1 yields (b).
Finally, using nφ φ = nφ + n − 1, (d) follows from
2
nφ φ = 2 nφ + n = nφ + nφ φ + 1

= nφ (1 + φ) + 1 = nφ φ 2 + 1.
Also solved by M. R. Avidon, D. Beckwith, N. Caro & O. López (Brazil), R. Chapman (U. K.), J. Christopher,
P. P. Dályay (Hungary), M. Goldenberg & M. Kaplan, G. Keselman, R. Lampe, J. H. Lindsey II, O. P. Lossers
((Netherlands)), A. Nijenhuis, S. Northshield, J. Rebholz, K. Schilling, A. Stadler (Switzerland), A. Stenger,
R. Stong, BSI Problems Group (Germany), Con Amore Problem Group (Denmark), GCHQ Problem Solving
Group (U. K.), Microsoft Research Problems Group, and the proposer.
Perfect Parity Patterns

11243 [2006, 759]. Proposed by Donald Knuth, Stanford University, Stanford, CA. An
m × n matrix of 0s and 1s is a parity pattern if every 0 is adjacent (horizontally or
vertically) to an even number of 1s and every 1 is adjacent to an odd number of 1s. It
is perfect if no row or column is entirely zero. Thus,
⎛ ⎞
⎛ ⎞ ⎛ ⎞ ⎛ ⎞ 0 1 1 1 0
0 0 1 1
1 1 0 1 0 1 0 ⎜1 0 1 0 1⎟
⎝0 0⎠ , ⎝0 1 0 0⎟
⎜
⎠ , ⎝1 1 0 1 1⎠ , ⎜1 1 0 1 1⎟
⎜ ⎟
1 1
1 1 0 1
0 1 0 1 0 ⎝ 1 0 1 0 1⎠
0 1 0 1
0 1 1 1 0
are parity patterns of sizes 3 × 2, 4 × 4, 3 × 5, and 5 × 5, respectively; only the 4 × 4

and 5 × 5 patterns are perfect.
(a) Determine the number c(n) of perfect parity patterns that have exactly n columns.
(b) Alone among these examples, the 5 × 5 parity pattern is invariant under rotation
by 90 degrees and under reflection across its central column. Thus it has eightfold
symmetry. Prove that a perfect n × n parity pattern with eightfold symmetry exists for
all n of the form n = 3 · 2k − 1 with k ≥ 1.
Solution of part (a) by Robin Chapman, University of Bristol, Bristol, U. K. Let c(n)
be the number of perfect parity patterns with n columns. To avoid a trivial addition of
668
1 to the formula, we require at least one row. We prove that

n + 1
c(n) = μ 2d−1 ,
d|(n+1)
d
where μ denotes the Möbius function.

Let M be an m × n matrix, with rows v1 , . . . , vm . Let 0 denote the all-zero vector
of length n, and set v0 = vm+1 = 0. In binary arithmetic, the condition that M is a
parity pattern is that the sum of entry (i, j) and its neighboring entries in M is 0 (that
is, even). Equivalently,
vi+1 + vi Tn + vi−1 = 0 (∗)
for 1 ≤ i ≤ m, where Tn is the n × n binary matrix whose (i, j) entry is 1 if and only
if |i − j| ≤ 1.
Note that (∗) is a recurrence that determines vi+1 from vi−1 and vi . Also, it is re-
versible: from vi and vi+1 , it determines vi−1 . Fixing v0 = 0, any n-vector v1 deter-
∞
mines a doubly infinite sequence (vi )i=−∞ of vectors via (∗). Since Fn2 is finite, some
pair (vi , vi+1 ) must repeat as (vi+ p , vi+1+ p ) for some p. Now (∗) implies that vi = vi+ p
for all i, and hence the sequence is periodic.
Call a parity pattern semi-perfect if it has no all-zero row (it may have all-zero
columns). A semi-perfect parity pattern is determined by its first row v1 , which is
nonzero. Since the sequence is periodic, there is a first positive integer q such that vq =
0. The parity pattern with rows v1 , . . . , vq−1 is the only semi-perfect parity pattern with
first row v1 . As there are 2n − 1 nonzero vectors of length n, there are 2n − 1 semi-
perfect parity patterns with n columns.
We now claim that in the sequence produced by v0 and v1 with q defined as above,
vi = 0 if and only if q | i. Suppose vk = 0. From vk+1 = vk Tn + vk−1 , we obtain
vk+1 = vk−1 . Applying the same computation to vk+ j shows inductively that vk+ j =
vk− j for all j. Thus the sequence v j is symmetric about any k for which vk = 0. Since
v0 = 0 and vq = 0, it follows that vmq = 0 for all m. Since v j = 0 for 0 < j < q, by
our symmetry observation, v j = 0 for any j other than multiples of q. As a conse-
quence, the number of rows in a parity pattern with first row v1 must be one less than a
multiple of q. Let Mk (v1 ) denote the parity pattern with kq − 1 rows and first row v1 ;
note that only k = 1 yields a semi-perfect parity pattern.
Now consider an m × n parity pattern M that is semi-perfect but not perfect; M has
an all-zero column. Let the first all-zero column be in position r . If r = 1 then by the
definition of a parity pattern, the other columns are forced to be zero, so r > 1.
The matrix N formed by the first r − 1 columns is a parity pattern with no all-zero
column. Its transpose N T is a semi-perfect parity pattern. Hence N T = M1 (w), where
w is the transpose of the first column of M. Now M T = Mk (w) for some integer k
with k ≥ 2. By the preceding paragraph, n = kr − 1. As M is semi-perfect, M T has
no zero column. Since every row of M T either equals a row of N T or is zero, it follows
from the lack of all-zero columns in M T that N T has no all-zero column. Hence N is
perfect.
On the other hand, from a perfect parity pattern N with r − 1 columns and first
column w, we can reverse this process to obtain a semi-perfect parity pattern M with
kr − 1 columns: let M = Mk (w)T , where N = M1 (w)T . Since N has no zero rows,
neither does M. Thus M is semi-perfect and its first all-zero column is in position r .
Counting semi-perfect parity patterns with n columns in groups according to the
associated perfect parity pattern N gives r |(n+1),r >1 c(r − 1) = 2n − 1. Setting j =

n + 1, a1 = 1, and ar = c(r − 1) for r > 1, this becomes r | j ar = 2 j −1 . By Möbius

inversion, a j = d| j μ( j/d)2d−1 . This is equivalent to the claimed result for c(n)
since c(n) = an+1 .
Solution of part (b) by Reiner Martin, Teddington, U. K. We recursively construct a
perfect parity pattern Pk with eightfold symmetry and 3 · 2k − 1 rows.
Let P1 be the 5 × 5 parity pattern in the problem statement. With ai, j denoting entry
(i, j) in Pk−1 , define entry (r, s) in Pk to be br,s , where
b2i,2 j = ai, j ,
b2i+1,2 j = ai, j + ai+1, j (mod 2),
b2i,2 j +1 = ai, j + ai, j +1 (mod 2), and
b2i+1,2 j +1 = 0,
where ai, j is understood to be 0 if i or j equals 0 or 3 · 2k−1 . It is straightforward to

verify the parity condition for each type of entry; hence Pk is indeed a parity pattern.
Also, eightfold symmetry is preserved by the construction. A zero row in Pk would
imply two adjacent identical rows in Pk−1 , which by induction does not occur. Thus
Pk is indeed perfect.
Editorial comment. Parity patterns arise in the study of a popular puzzle called
“Lights Out”, invented in the 1980s by Dario Uri and available as a Java applet at
http://www.whitman.edu/mathematics/lights_out/. Klaus Supner called such pat-
terns “even-parity covers” of the m × n grid; see Math. Intell. 11 No. 2 (Spring 1989)
49–53, and Theor. Comp. Sci. 230 (2000) 49–73.
Also solved by D. Beckwith, S. M. Gagola, D. Serre (France), Microsoft Research Problems Group, and the
proposer. Part (a) also solved by J. H. Lindsey II.
Double Integral with Log and Sinh

11275 [2007, 164]. Proposed by Michael S. Becker, University of South Carolina at
Sumter, Sumter, SC. Find
∞ ∞
(x − y)2 log((x + y)/(x − y))
d x dy.
y=0 x=y x y sinh(x + y)
Solution I by David Beckwith, Sag Harbor, NY. Let I denote the required integral.
Introduce the change of variables x = (z + t)/2, y = (z − t)/2 which gives
∞ z 2
1 t log(z/t)
I =2 dt dz.
z=0 sinh z t=0 z2 − t 2
In the inner integral, set t = zw to obtain the product of two known integrals,
∞ 1

z dz 1 π2 π2 π 2 (π 2 − 8)
I =2 1− log w dw = 2 − 1 = .
0 sinh z 0 1 − w2 4 8 16
Solution II by Hongwei Chen, Christopher Newport University, Newport News, VA.

Let I denote the required integral. Substitute x = uy and interchange the order of
integration to obtain
∞
∞
(u − 1)2 log((u + 1)/(u − 1)) y
I = dy du.
u=1 u y=0 sinh(y(u + 1))
670
Let Iu denote the inner integral in the previous formula for I . Substituting t = e−(u+1)y
gives
1 1 ∞
−2 ln t −2
Iu = dt = t 2k ln t dt
(1 + u)2 0 t 2 − 1 (1 + u)2 t=0 k=0
2 ∞
1 π2
= = .
(1 + u)2 k=0 (2k + 1)2 4(1 + u)2
So I takes the form
∞
π2 (u − 1)2 log((u + 1)/(u − 1))
I = du.
4 1 u(1 + u)2
Finally, substitute s = (u − 1)/(u + 1):

π 2 1 s 2 ln s π2 1 ln s π 2 (π 2 − 8)
I =− ds = − − ln s ds = .
2 0 1 − s2 2 0 1 − s2 16
Editorial comment. Many other interesting substitutions were submitted.

Also solved by R. Bagby, D. H. Bailey (USA) & J. M. Borwein (Canada), R. Chapman (U. K.), K. Dale
(Norway), B. E. Davis, P. De (Ireland), N. Eklund, C. Fleming, W. Fosheng (China), M. L. Glasser, M. R.
Gopal, R. Govindaraj & R Mythili & G. Anupama & G. Sudharsan (India), J. Grivaux (France), J. A. Grzesik,
E. A. Herman, G. Keselman, O. Kouba (Syria), G. Lamb, D. Lovit, K. McInturff, O. Padé (Israel), P. Perfetti
(Italy), T.-L. & V. Rădulescu (Romania), O. G. Ruehr, H.-Z. Seiffert (Germany), A. Stadler (Switzerland), R.
Stong, R. Tauraso (Italy), E. I. Verriest, M. Vowe (Switzerland), FAU Problem Solving Group, GCHQ Problem
Solving Group (U. K.), Microsoft Research Problems Group, University of Sharjah Problem Solving Group
(United Arab Emirates), and the proposer.
A Matrix Inequality
11296 [2007, 252]. Proposed by Said Amghibech, Quebec, Canada. Let A be a real
symmetric positive definite matrix, with entries ai, j , 1 ≤ i, j ≤ n. Let Ak be the k × k
matrix in the upper left corner of A. Show that

n
1 n
(det Ak )1/k ≤ 1 + Tr(A)
k=1
n
where Tr(A) denotes the trace of A.
Solution by Joanna Chachulska and Wojciech Matysiak, Polytechnica Warszawska,
Warsaw, Poland. Hadamard’s inequality says det Ak ≤ a1,1 a2,2 · · · ak,k . Carleman’s
inequality says

n
n
1/k 1 n
x1 x2 · · · xk ≤ 1+ xk
k=1
n k=1
for xk ≥ 0. Our result follows.

Note: Carleman’s inequality is usually stated in its limit form, with infinite sums and
the number e replacing the term (1 + 1/n)n by monotonicity. However, standard proofs
of the inequality prove the finite form stated above. For example, see pp. 146–152 of
G. Pólya, Mathematics and Plausible Reasoning, Princeton Univ. Press, Princeton, NJ,
1954.
Also solved by D. Chakerian, R. Chapman (U. K.), J.-P. Grivaux (France), E. A. Herman, C. J. Hillar, F.
Holland (Ireland), D. Jesperson, O. P. Lossers (Netherlands), M. Omarjee, H.-J. Seiffert (Germany), M. Tetiva
(Romania), NSA Problems Group, and the proposer.


that address before February 28, 2009. Additional information, such as general-
PROBLEMS
11383. Proposed by Michael Nyblom, RMIT University, Melbourne, Australia. Show
that
√ √
∞
1 + n 2 + 2n n 2 + 4n + 3 π
−1
cos = .
n=1
(n + 1)(n + 2) 3
11384. Proposed by Moubinool Omarjee, Lycée Jean-Lurçat, Paris, France. Let pn

denote the nth prime. Show that
√

∞
(−1) n
n=1
pn
converges.
11385. Proposed by José Luis Dı́az-Barrero, Universidad Politécnica de Cataluña,
Barcelona, Spain. Let α0 , α1 , and α2 be the radian measures of the angles of an acute
triangle, and for i ≥ 3 let αi = αi−3 . Show that

2
αi2 1/4 √
3 + 2 tan2 αi ≥ 3 3.
i=0
αi+1 αi+2
11386. Proposed by Greg Markowsky, Somerville, MA. Consider a triangle ABC. Let
O be the circumcircle of ABC, r the radius of the incircle, and s the semiperimeter.
Let arc(BC) be the arc of O opposite A, and define arc(C A) and arc(AB) similarly.
Let O A be the circle tangent to AB and AC and internally tangent to O along arc(BC),
and let R A be its radius. Define O B , OC , R B , and RC similarly. Show that
1 1 1 s2
+ + = .
a RA b RB c RC rabc
October 2008] PROBLEMS AND SOLUTIONS 757

11387. Proposed by Oskar Maria Baksalary, Adam Mickiewicz University, Poznań,
Poland, and Götz Trenkler, Technische Universität Dortmund, Dortmund, Germany.
Let Cn,n denote the set of n × n complex matrices. Determine the shortest interval
[a, b] such that if P and Q in Cn,n are nonzero orthogonal projectors, that is, Hermitian
idempotent matrices, then all eigenvalues of PQ + QP belong to [a, b].
11388. Proposed by M. Farrokhi D.G., University of Tsukuba, Tsukuba Ibaraki, Japan.
Given a group G, let G 2 denote the set of all squares in G. Show that for each natural
number n there exists a finite group G such that the cardinality of G is n times the
cardinality of G 2 .
11389. Proposed by Elizabeth R. Chen and Jeffrey C. Lagarias, University of Michi-
gan, Ann Arbor, MI. Given a multiset A = {a1 , . . . , an } of n real numbers (not neces-
sarily distinct), define the sumset S(A) of A to be {ai + a j : 1 ≤ i < j ≤ n}, a multi-
set with n(n − 1)/2 not necessarily distinct elements. For instance, if A = {1, 1, 2, 3},
then S(A) = {2, 3, 3, 4, 4, 5}.
(a) When n is a power of 2 with n ≥ 2, show that there are two distinct multisets A1
and A2 such that S(A1 ) = S(A2 ).
(b) When n is a power of 2 with n ≥ 4, show that if r distinct multisets A1 , . . . , Ar all
have the same sumset, then r ≤ n − 2.
(c∗ ) When n is a power of 2 with n ≥ 4, can there be as many as 3 distinct multisets
with the same sumset?
(Distinct multisets are known to have distinct sumsets when n is not a power of 2.)
SOLUTIONS
Recalling 11159 and Sixty-Sixth Putnam A5

11277 [2007, 259]. Proposed by Prithwijit De, University College Cork, Republic of
Ireland. Find
π/2 π/2
log(2 − sin θ cos φ) sin θ
dθ dφ.
θ=0 2 − 2 sin θ cos φ + sin θ cos φ
2 2
φ=0
Solution by E. Omey, EHSAL, Brussels, Belgium. The answer is (π 2 /16) log 2. The
integral to be evaluated is
π/2 π/2
I = f (sin θ cos φ) sin θ dθ dφ,
φ=0 θ=0
where f (x) = log(2 − x)/(2 − 2x + x 2 ). In an editorial comment attached to the so-

lution of Problem 11159 in this M ONTHLY 114 (2007) 167, it was noted that if g is
integrable on [0, 1] then
π/2 π/2
π 1
g(cos ψ cos φ) cos ψ dψ dφ = g(t) dt.
φ=0 ψ=0 2 0
1
After substituting θ = π/2 − ψ and setting g = f , we see that I = (π/2) 0 f (t)dt.
Write x = 1 − t. The required computation,
1
log(1 + x) π log 2
dx = ,
0 1+x 2 8
758
appears in various tables and its evaluation was given as Problem A5 on the Sixty-Sixth
William Lowell Putnam Competition (see this M ONTHLY 113 (2006) 733-743).
Also solved by R. Bagby, D. H. Bailey & J. M. Borwein (U. S. A., Canada), D. Beckwith, R. Chapman (U. K.),
H. Chen, K. Dale (Norway), B. E. Davis, G. de Marco (Italy), A. Fok (Hong Kong), O. Furdui, M. L. Glasser,
J. Grivaux (France), E. A. Herman, O. Kouba (Syria), G. Lamb, K. McInturff, P. Perfetti (Italy), O. G,
Ruehr, H.-J. Seiffert (Germany), J. G. Simmonds, A. Stadler (Switzerland), V. Stakhovsky, R. Stong, J. Sun,
R. Tauraso (Italy), GCHQ Problem Solving Group (U. K.), Microsoft Research Problems Group, University
of Sharjah Problem Solving Group (United Arab Emirates), and the proposer.
Entire Limit
11278 [2007, 259]. Proposed by Slavko Simic, Mathematical Institute SANU, Bel-
grade, Serbia. Let f be a nonconstant entire function with nonnegative Taylor series
coefficients. Prove that limr →∞ f (r )/r f (r ) exists and is rational.
Solution by John H. Lindsey II, Cambridge, MA. First consider the case that f =
a0 + · · · + an x n , a polynomial with an > 0. We have
f (r ) a0 + · · · + an r n a0 /r n + · · · + an 1
lim = lim = lim = .
r →∞ r f (r )
r →∞ a1 r + · · · + an nr n r →∞ a1 /r + · · · + an n
n n
∞
Now consider the case in which f = k=0 ak x k with all coefficients nonnegative, and
not eventually zero. For n ∈ N with an > 0,
n−1 ∞
f (r ) k=0 ak r
k
k=n ak r
k
lim sup ≤ lim sup + lim sup ∞
r →∞ r f (r ) an nr n k
r →∞ r →∞ k=n ak kr
∞
ak r k 1
≤ 0 + lim sup ∞k=n k
= .
r →∞ k=n ak nr n
Since n can be taken arbitrarily large, lim f (r )/r f (r ) = 0.

Editorial comment. It should be noted that the limit is intended in the sense of a real
variable r → +∞, not in the sense of a complex variable r → ∞.
Also solved by K. F. Andersen (Canada), J. Bak, R. Chapman (U. K.), G. De Marco (Italy), J. Grivaux (France),
E. A. Herman, F. Holland (Ireland), E. J. Ionascu, G. Keselman, K.-W. Lau (China), O. P. Lossers (Nether-
lands), A. Nakhash, N. C. Singer, R. Stong, J. Vinuesa (Spain), GCHQ Problem Solving Group (U. K.), Mi-
crosoft Research Problems Group, Northwestern University Math Problem Solving Group, and the proposer.
Sliding Beads
11279 [2007, 259]. Proposed by Vitaly Stakhovsky, Redwood City, CA. Two test-mass
beads are sliding along a vertical circular track under (Newtonian) constant gravity,
without friction. The first bead M1 starts from the highest point on the circle, at some
nonzero velocity. Some time later the second bead M2 starts from the same position at
the top of the circle and with the same initial velocity as M1 had. Prove that there is a
circle to which the line through the current positions of M1 and M2 is always tangent,
and find the center and radius of that circle in terms of the original circle.
Solution by the proposer. Assume that the original circle has unit radius and center
at the origin O. Let OY point in the upward direction. Describe the position of bead
i by its current location Pi and the angle ϕi = Pi OY . Assume both particles are
moving with ϕi increasing. Let vi = ϕ̇i be the ith bead’s velocity. The equation of

motion can be written as vi2 = 1 − μ cos ϕi for some 0 ≤ μ < 1. Note that v12 − v22 =
μ(cos ϕ2 − cos ϕ1 ) and v̇i = (μ/2) sin ϕi . Hence

v̇1 + v̇2 1 sin ϕ1 + sin ϕ2 1 ϕ1 − ϕ2
= · = cot .
v12 − v22 2 cos ϕ2 − cos ϕ1 2 2

Since |P1 P2 | = 2 sin ϕ1 −ϕ2
2
we have

d 1 ϕ1 − ϕ2 v̇1 + v̇2 d
log |P1 P2 | = cot (v1 − v2 ) = = log(v1 + v2 ).
dt 2 2 v1 + v2 dt
Therefore τ = |P1 P2 |/(v1 + v2 ) is constant over time. Note that

ϕ1 + ϕ2
v1 − v2 = μ sin
2 2
|P1 P2 |,
2
so we also have μτ = (v1 − v2 )/ sin((ϕ1 + ϕ2 )/2).
Let Q be the point on P1 P2 such that |P1 Q|/|P2 Q| = v1 /v2 . (Q is the instantaneous
center of rotation of the beads and will be the point of tangency of the circle.) Let a
perpendicular to P1 P2 at Q intersect the vertical axis at a point D = (0, d). Let M
be the midpoint of P1 P2 . Now Q = (v1 P2 + v2 P1 )/(v1 + v2 ) and M = (P1 + P2 )/2,
hence

(v1 − v2 )|P1 P2 | v1 − v2 1 2 ϕ1 + ϕ2
|M Q| = = τ = μτ sin .
2(v1 + v2 ) 2 2 2
Hence d = μτ 2 /2, and the point D is constant over time. Also |D Pi |2 = 1 + d 2 −
2d cos ϕi = |D Q|2 + |Q Pi |2 , so
v12 |D P2 |2 − v22 |D P1 |2
|D Q|2 =
v12 − v22
v12 cos ϕ2 − v22 cos ϕ1 2
= 1 + d 2 − 2d = 1 + d2 − d
v1 − v2
2 2
μ
is constant
over time. Thus P1 P2 is always tangent to the circle with center D and
radius 1 + d 2 − 2d/μ.
Note: In case P2 starts when P1 is at the lowest point, we get r = 0 and hence the
points P1 , P2 , D are always collinear.
Also solved by Ó. Ciaurri & E. Fernández & L. Roncal (Brazil), J. Freeman, J. A. Grzesik, J. H. Lindsey II,
J. B. Zacharias, and GCHQ Problem Solving Group (U. K.).
What We Mean, Not What We Say

11280 [2007, 259]. Proposed by Harris Kwong, SUNY Fredonia, Fredonia, NY. Let f
be a positive nondecreasing function on the real line that is twice differentiable
and
concave down. For any list x of positive real numbers x1 , . . . , xn , let S = nk=1 xk . In
terms of f and n, find

n
max ( f (S − xk ))xk /S .
x
k=1
Solution by GCHQ Problem Solving Group, Cheltenham, U. K. There is something

wrong with the problem statement: apart from constant functions, there are no positive
concave functions on the real line, let alone twice differentiable nondecreasing ones.
760
If n = 1 the answer is f (0) whatever properties f has. For general n, we show that
if f is a positive nondecreasing function on the real line (not necessarily concave or
differentiable, or even continuous), then

n
xk /S
sup f (S − xk ) = lim f (y). (1)
x y→+∞
k=1
We claim first that the limit exists. Indeed, if f is bounded above, its least upper bound
is its limit; if
f is not bounded above, the limit is +∞. Let this limit be ϕ. Next, if ϕ
is finite, then nk=1 ( f (S − xk ))xk /S ≤ nk=1 ϕ xk /S = ϕ. Finally, if x1 = · · · = xn = z,
then

n
xk /S 1/n
f (S − xk ) = f ((n − 1)z) = f ((n − 1)z) → ϕ
k=1
as z → ∞. This completes the proof of (1). The maximum is attained if and only if
the limit is.
There are of course functions on the nonnegative half of the real line with the full
range of properties given in the problem statement. The result given here applies to
them.
Editorial comment. It seems that in the intended problem f is defined on the positive
real line, and n and S are given values.
Then maximize over all lists x satisfying the
conditions. Answer: f (1 − 1/n)S .
Also solved by R. A. Agnew, K. F. Andersen (Canada), J.-P. Grivaux (France), E. A. Herman, J. H. Lindsey II,
V. Stakhovsky, R. Stong, M. Tetiva (Romania), and Microsoft Research Problems Group.
A Triangular Result
11285 [2007, 358]. Proposed by Yakub Aliyev, Qafqaz University and Baku State Uni-
versity, Baku, Azerbaijan. Let six points be chosen in cyclic order on the sides of
triangle ABC: A1 and A2 on BC, B1 and B2 on C A, and C1 and C2 on AB. Let K
denote the intersection of A1 B2 and C1 A2 , L the intersection of B1 C2 and A1 B2 , and
M the intersection of C1 A2 and B1 C2 . Let T , U , and V be the intersections of A1 B2
and B1 A2 , B1 C2 and B2 C1 , and C1 A2 and C2 A1 , respectively. Prove that lines AK ,
B L, and C M are concurrent if and only if points T , U , and V are collinear.
Solution I (one direction) by Apostolis Denis, Varvakeio High School, Athens, Greece.
We prove the first direction of the implication. Suppose AK , B L, C M are concur-
rent. Write u.v for the intersection of lines u and v (extended if necessary). Let X =
A1 C2 .B1 A2 , Y = B1 A2 .C1 B2 , and Z = C1 B2 .A1 C2 . Given triangles ABC and abc,
Desargues’ Theorem with its dual states that Aa, Bb, Cc are concurrent if and only
if AB.ab, BC.bc, C A.ca are collinear. We consider parallel lines to intersect at infinity
and allow for collinearity and concurrency at infinity (see Editorial comment below).
Application of Desargues’ Theorem to the following three pairs of triangles gives the
claimed implication:
(i) [ABC and K L M]: AK , B L, C M are concurrent if and only if AB.K L,
BC.L M, C A.M K are collinear. Since C1 lies on AB, and B2 lies on K L, we have
AB.K L = C1 B.B2 L. Similarly, BC.L M = B A2 .B1 L, and C A.M K = B1 B2 .C1 A2 .
(ii) [C1 B A2 and B2 L B1 ]: C1 B.B2 L, B A2 .L B1 , and A2 C1 .B1 B2 are collinear if and
only if C1 B2 , B L, A2 B1 are concurrent. Note: Since C1 B2 .A2 B1 = Y , this concurrency
is equivalent to B L = LY . Similarly, AK = K X and C M = M Z so K X , LY , M Z
are concurrent.

(iii) [K L M and X Y Z]: K X , LY, M Z are concurrent if and only if T = K L .X Y ,
U = L M.Y Z, V = M K .Z X are collinear.
Solution II by The GCHQ Problem Solving Group, Cheltenham, United Kingdom. Pas-
cal’s Theorem states that the intersection points of pairs of opposite sides of a hexagon
are collinear if and only if the vertices of the hexagon lie on a conic. With the notation
of Solution I:
(a) Applying Desargues’ Theorem to ABC and K L M, we have: AK , B L , C M are
concurrent if and only if AB.K L, BC.L M, C A.M K are collinear.
(b) {AB.K L , BC.L M, C A.M K } = {C1 C2 .A1 B2 , A1 A2 .B1 C2 , B1 B2 .C1 A2 }, so by
Pascal’s Theorem these are collinear if and only if C1 , C2 , A1 , A2 , B1 , B2 lie on a
conic.
(c) {T, U, V } = {A1 B2 .B1 A2 , B1 C2 .B2 C1 , C1 A2 .C2 A1 }, so again by Pascal’s The-
orem, these are collinear if and only if A1 , B2 , B1 , C2 , C1 , A2 lie on a conic.
Thus, AK , B L, C M are concurrent if and only if T, U, V are collinear.
Editorial comment. The statement of the problem is true projectively, but for affine (or
Euclidean) geometry, it degenerates when X , Y , or Z fails to exist, or if K L M is a
degenerate triangle.
Also solved by R. Chapman (U. K.), P. Dályay (Hungary), M. Englefield (Australia), W. Fosheng (China),
O. P. Lossers (Netherlands), M. Tetiva (Romania), L. Zhou, and the proposer.
Continuous Blackjack
11287 [2007, 359]. Proposed by Stephen J. Herschkorn, Highland Park, NJ. Players
1 through n play “continuous blackjack.” At his turn, Player k considers a random
number X k drawn from the uniform distribution on [0, 1]. He may either accept X k as
his score or draw a second number Yk from the same distribution, in which case his
score is X k + Yk if X k + Yk < 1 and 0 otherwise. The highest score wins. Give a rule
for when player k should draw a second number, in terms of k, n, the result of X k , and
the highest score attained so far.
Solution by BSI Problems Group, Bonn, Germany, and the editors. We renumber the
players P0 , P1 , . . . , Pn−1 in reverse order, so that Pk has k players yet to draw after
his turn. It turns out that the optimal strategy for Pk is the same whether the remaining
players play as a team or selfishly.
Let Z k be Pk ’s score; that is, Z k = X k if Pk accepts his first number, Z k = X k + Yk
if he draws a second number Yk and X k + Yk < 1, and Z k = 0 if he draws a second
number and X k + Yk ≥ 1. For k = 0, . . . , n − 1, let Mk = max j ≥k Z j , and let Mn = 0.
We assume selfish play. That is, Pk should take a second number if and only if
doing so improves his probability of winning, assuming Pk−1 through P0 do the same.
We ignore ties, which have zero probability. Note that Pk ’s situation depends only on
the number k of players yet to draw and on Mk+1 . How Mk+1 was produced, and how
many players came earlier to produce it, are irrelevant.
Let f k (z) denote the probability that if players P j for j ≥ k have already played and
the best score so far is Mk = z, then for all j < k, Z j < z—that is, all later players
achieve scores lower than z, so z is the winning score. We will prove that for all k:
1. f k is a continuous, nondecreasing function on [0, 1] and f k (z) > 0 for z > 0.
1
2. There is a unique number gk ∈ [0, 1) such that f k (gk ) = gk f k (t) dt, and Pk
should take a second number if and only if X k < max(Mk+1 , gk ).
We first observe that statement 2 follows from statement 1. To see why, suppose
1
statement 1 holds. Clearly z f k (t) dt is a strictly decreasing continuous function of z,
762
1 1
f k (t) dt ≥ f k (0), and 1 f k (t) dt = 0 < f k (1), so by the intermediate value theo-
0 1
rem there is a unique number gk ∈ [0, 1) such that f k (gk ) = gk f k (t) dt. Furthermore,
1 1
if z < gk then z f k (t) dt > f k (z), and if z > gk then z f k (t) dt < f k (z). To deter-
mine how player Pk should play, suppose that all players P j for j > k have already
played. Clearly if X k < Mk+1 then player Pk must take a second number to have any
hope of winning. If X k > Mk+1 , then player Pk ’s probability of winning is f k (X k ) if
1
he doesn’t take a second number, and X k f k (t) dt if he does. The latter probability is
larger than the former if and only if X k < gk , so Pk should take a second number if
and only if X k < max(Mk+1 , gk ).
To prove statement 1, we proceed by induction on k. In the case k = 0, if all of the
players have played and M0 = z, then z is definitely the winning score, so f 0 (z) = 1.
Now assume statement 1 (and therefore also statement 2) holds for k. Suppose that
players P j for j ≥ k + 1 have already played and Mk+1 = z. There are two ways for
player Pk to end up with a score z Z k < z: either X k + Yk < z, or X k + Yk > 1. The
probability of the first event is 0 (z − x) d x = z 2 /2, and the probability of the second
max(z,gk )
is 0 x d x = (max(z, gk ))2 /2. Thus, the probability that Z k < z is:

1
(g 2 + z 2 ), if z ≤ gk ;
h k (z) = 22 k
z , if z > gk .
If Z k < z then Mk = z, and therefore the probability that Z j < z for all j < k is
f k (z). Thus, we have f k+1 (z) = h k (z) f k (z), and combining our formula for h k (z) with
statement 1 for k, we see that statement 1 holds for k + 1 as well.
Finally, we wish to determine the numbers gk . We first observe that for all k, gk+1 >
gk . To see why, suppose that gk+1 ≤ gk . Then, on putting l = k + 1, we have
1
fl (gl ) = h k (gl ) f k (gl ) ≤ h k (gl ) f k (gk ) = h k (gl ) f k (t) dt
gk
1 1 1
≤ h k (gl ) f k (t) dt < h k (t) f k (t) dt = fl (t) dt = fl (gl ),
gl gl gl
a contradiction. It follows that gk+1 > gk , as claimed.

Next, we observe that for z ≥ gk , f k (z) = z 2k . This follows easily by induction
from the facts that f 0 (z) = 1, f k+1 (z) = h k (z) f k (z), and gk+1 > gk , and the formula
for h k (z).
Finally, we use this formula for f k (z) to get an equation that we can solve to find
gk :
1 1
1
gk = f k (gk ) =
2k
f (t) dt = t 2k dt = (1 − gk2k+1 ).
gk gk 2k + 1
In summary, the best strategy is for Pk to take a second number if and only if
X k < max(Mk+1 , gk ), where gk is the unique nonnegative solution to z 2k = (1 −
z 2k+1 )/(2k + 1).
Editorial comment. If the remaining players are not selfish, but cooperate to try to stop
Pk , this is still Pk ’s best strategy. For j < k, player P j now accepts any score X j that
is better than M j . The result is that h j (z) = z 2 and f j (z) = z 2 j for j < k. Thus Pk
should still take a second number if and only if X k < max(Mk+1 , gk ).

Also solved by D. R. Bridges, J. Chachulska & W. Matysiak (Poland), R. Chapman (U. K.), J. Freeman, J. H.
Lindsey II, G. Pastor (Mexico), T. Rucker, B. Schmuland (Canada), R. Staum, T. Tam, E. A. Weinstein, GCHQ
Problem Solving Group (U. K.), and the proposer.
Formulas Involving the Angles of a Triangle

11289 [2007, 359]. Proposed by Oleh Faynshteyn, Leipzig, Germany. Let ABC be a
triangle with sides a, b, and c, all different, and corresponding angles α, β, and γ .
Show that
(a) (a + b) cot(β + 12 γ ) + (b + c) cot(γ + 12 α) + (a + c) cot(α + 12 β) = 0.
(b) (a − b) tan(β + 12 γ ) + (b − c) tan(γ + 12 α) + (c − a) tan(α + 12 β) = 4(R + r ),
where R is the circumradius of the triangle and r the inradius.
Solution by Marian Tetiva, Birlad, Romania. Since α + β + γ = π, we have β +
1
2
γ = π2 − α−β
2
, so cot(β + 12 γ ) = tan α−β
2
and tan(β + 12 γ ) = cot α−β
2
. From the law
of sines and trigonometric identities, we get
α+β α−β
a + b = 2R(sin α + sin β) = 4R sin cos ,
2 2
α−β α+β
a − b = 2R(sin α − sin β) = 4R sin cos .
2 2
are obtained by cyclically permuting a, b, c and α, β, γ .
Similar equations
We will write “ ” for a sum with three terms obtained by cyclic permutation from
the one term given. For part (a):

1 α+β α−β α−β
(a + b) cot β + γ = 4R sin cos tan
2 2 2 2

α+β α−β α β
= 4R sin sin = 4R sin2 − sin2 = 0.
2 2 2 2
For part (b):

1 α−β α+β α−β
(a − b) tan β + γ = 4R sin cos cot
2 2 2 2
α+β α−β
= 4R cos cos = 4R cos α = 4(R + r ).
2 2
For the last step we used the formula
α β γ r
cos α = 1 + 4 sin sin sin = 1 + .
2 2 2 R
Also solved by Z. Ahmed (India), S. Amghibech (Canada), M. Bataille (France), R. Chapman (U.K.), P. P.
Dályay (Hungary), P. De (Ireland), A. Demis (Greece), A. Fok (China), W. Fosheng (China), V. V. Garcı́a
(Spain), M. Goldenberg & M. Kaplan, O. Kouba (Syria), H. Kwong, K.-W. Lau (China), O. P. Lossers (Nether-
lands), Y. Mikata, E. Mouroukos (Greece), R. D. Nelson (U.K.), J. H. Nieto (Venezuela), P. E. Nuesch (Switzer-
land), C. R. Pranesachar (India), J. Rooin & A. Alikhani (Iran), V. Schindler (Germany), L. Sega, H.-J. Seiffert
(Germany), M. Shattuck, R. A. Simon (Chile), A. Stadler (Switzerland), D. Vacaru (Romania), M. Vowe
(Switzerland), L. Zhou, GCHQ Problem Solving Group (U.K.), Microsoft Research Problems Group, North-
western University Math Problem Solving Group, and the proposer.
764
cioni, Frank B. Miles, Richard Pfiefer, Cecil C. Rousseau, Leonard Smiley, Kenneth
Stolarsky, Richard Stong, Walter Stromquist, Daniel Ullman, Charles Vanden Eynden,
and Fuzhen Zhang.

that address before March 31, 2009. Additional information, such as general-
PROBLEMS
11390. Proposed by Jeffrey C. Lagarias, University of Michigan, Ann Arbor, MI. Let
G be the undirected graph on the vertex set V of all pairs (a, b) of relatively prime
integers, with edges linking (a, b) to (a + kab, b) and (a, b + kab) for all integers k.
(a) Show that for all (a, b) in V , there is a path joining (a, b) and (1, 1).
(b)∗ Call an edge linking (a, b) to (a + kab, b) or (a, b + kab) positive if k > 0, and
negative if k < 0. Let the reversal number of a path from (1, 1) to (a, b) be one more
than the number of sign changes along the path, and let the reversal value of (a, b)
be the minimal reversal number over all paths from (1, 1) to (a, b). Are there pairs of
arbitrarily high reversal value?
Bı̂rlad, Romania. Let p be a positive prime and s a positive integer. Let n and k be
integers such that n ≥ k ≥ p s − p s−1 , and let x1 , . . . , xn be integers. For 1 ≤ j ≤ n,
let m j be the number of expressions of the form xi1 + · · · + xi j with 1 ≤ i 1 < · · · <
i j ≤ n that evaluate to 0 modulo p, and let n j denote the number of such expressions
that do not. (Set m 0 = 1 and n 0 = 0). Apart from the cases (s, k) = (1, p − 1) and
s = p = k = 2, show that

j n−k + j
k
(−1) m k− j ≡ 0 (mod p s ),
j =0
j
and show that the same congruence holds with n k− j in place of m k− j .

11392. Proposed by Omran Kouba, Higher Institute for Applied Science and Technol-
ogy, Damascus, Syria. Let the consecutive vertices of a regular n-gon P be denoted
A0 , . . . , An−1 , in order, and let An = A0 . Let M be a point such that for 0 ≤ k < n
the perpendicular projections of M onto each line Ak Ak+1 lie interior to the segment
(Ak , Ak+1 ). Let Bk be the projection of M onto Ak Ak+1 . Show that

n−1
1
Area((M Ak Bk )) = Area(P).
k=0
2
November 2008] PROBLEMS AND SOLUTIONS 855

11393. Proposed by Cosmin Pohoata (student), National College “Tudor Vianu”,
Bucharest, Romania. In triangle ABC, let M and Q be points on segment AB,
and similarly let N and R be points on AC, and P and S be points on BC. Let
d1 be the line through M and N , d2 the line through P and Q, and d3 the line
through R and S. Let ρ(X, Y, Z) denote the ratio of the length of X Z to that of
X Y . Let m = ρ(M, A, B), n = ρ(N , A, C), p = ρ(P, B, C), q = ρ(Q, B, A),
r = ρ(R, C, A), and s = ρ(S, C, B). Prove that the lines (d1 , d2 , d3 ) are concurrent if
and only if mpr + nqs + mq + nr + ps = 1.
11394. Proposed by K. S. Bhanu, Institute of Science, Nagpur, India, and M. N. Desh-
pande, Nagpur, India. A fair coin is tossed n times, with n ≥ 2. Let R be the resulting
number of runs of the same face, and X the number of isolated heads. Show that the
covariance of the random variables R and X is n/8.
11395. Proposed by M. Farrokhi D. G.,University of Tsukuba, Tsukuba, Japan. Prove
that if H is a finite subgroup of the group G of all continuous bijections of [0, 1] to
itself, then the order of H is 1 or 2.
11396. Proposed by Gérard Letac, Université Paul Sabatier, Toulouse, France. For
complex z, let Hn (z) denote the n × n Hermitian matrix whose diagonal elements all
equal 1 and whose above-diagonal elements all equal z. For n ≥ 2, find all z such that
Hn (z) is positive semi-definite.
SOLUTIONS
Generalizing (1 − 1)n = 0
11230 [2006, 567]. Proposed by Gregory Keselman, Oak Park, MI, formerly of Lvov
Polytechnic Institute, Ukraine. Let n be a positive integer, let [n] = {0, 1, . . . , n − 1},
and for a subset P of [n] let s(z, P) = j ∈P z j . With the usual understanding that
00 = 1, show that

(−1)|P| s k (z, P) = 0 (k < n),

P⊆[n]
(−1)|P| s n (z, P) = (−1)n (n!)z n(n−1)/2 ,
P⊆[n]
(n + 1)!(z n − 1)z n(n−1)/2
(−1)|P| s n+1 (z, P) = (−1)n .
P⊆[n]
2(z − 1)
Solution I by Aleksandar Ilić, student, University of Niš, Serbia. Let S[n, k] =

|P| k
P⊆[n] (−1) s (z, P). We prove the identities by induction on n. It holds by in-
spection that S[1, 0] = 0, S[1, 1] = −1, and S[1, 2] = −1. For the induction step,
place the subsets of [n + 1] into two groups: those that contain n and those that do
not. Using the binomial theorem,

S[n + 1, k] = (−1)|P| s k (z, P) + (−1)|P|+1 (s(z, P) + z n )k
P⊆[n] P⊆[n]

k
|P| k |P| k i
= (−1) s (z, P) − (−1) s (z, P)z n(k−i)
.
P⊆[n] P⊆[n] i=0
i
856
After canceling equal terms and interchanging sums,
k−1
k−1
k n(k−i) k n(k−i)
S[n + 1, k] = − z (−1)|P| s i (z, P) = − z S[n, i].
i=0
i P⊆[n] i=0
i
For 0 ≤ k ≤ n, the induction hypothesis yields S[n + 1, k] = 0. For k = n + 1,

n+1 n
S[n + 1, n + 1] = − z S[n, n] = (−1)n+1 (n + 1)! z n(n+1)/2 .
n
For k = n + 2, we simply extract common factors:

n + 2 2n n+2 n
S[n + 1, n + 2] = − z S[n, n] − z S[n, n + 1]
n n+1

(n + 2)! n n(n−1)/2 n z n − 1
= (−1)n+1 z z z +
2 z−1
(n + 2)! (z n+1 − 1)z n(n+1)/2
= (−1)n+1 .
2(z − 1)
Solution II by O. P. Lossers, Eindhoven University of Technology, Eindhoven, The

Netherlands. We have

n−1
m ∞
yk k
1 − e yz = (−1)|P| e ys(z,P) = (−1)|P| s (z, P)
m=0 P⊆[n] P⊆[n] k=0
k!

∞
yk
= (−1)|P| s k (z, P)
k=0
k! P⊆[n]
and also

n−1

n−1
m 1
1 − e yz = (−1)n yz m + y 2 z 2m + O(y 3 )
m=0 m=0
2
1
n−1
= (−1)n y n z n(n−1)/2 + (−1)n y n+1 z n(n−1)/2 z m + O(y n+2 ),
2 m=0
where O(y t ) indicates a series divisible by y t . Equating powers of y in these two

expressions gives the required equalities.
Also solved by U. Abel (Germany), S. Amghibech (Canada), M. R. Avidon, D. Beckwith, K. Bernstein,
N. Caro (Brazil), R. Chapman (U. K.), P. Corn, P. P. Dályay (Hungary), S. M. Gagola Jr., J. Grivaux (France),
E. A. Herman, J. H. Lindsey II, U. Milutinović (Slovenia), A. Nijenhuis, M. A. Prasad (India), N. C. Singer,
R. J. Snelling, A. Stadler (Switzerland), A. Stenger, R. Stong, M. Tetiva (Romania), BSI Problems Group (Ger-
many), GCHQ Problem Solving Group (U. K.), Microsoft Research Problems Group, NSA Problems Group,
and the proposer.
If AB Preserves A, then BA Preserves B

11239 [2006, 655]. Proposed by Michel Bataille, Rouen, France. Let A and B be
complex n × n matrices of the same rank. Show that if A2 B = A, then B 2 A = B.
Solution by John W. Hagood, Northern Arizona University, Flagstaff, AZ. If A2 B = A,
then rank A ≥ rank A2 ≥ rank A. Thus A, A2 , and B all have the same rank, and hence

their null spaces have the same dimension. Since the null spaces of A2 and B are
subspaces of the null space of A, the three null spaces are identical. If z ∈ Cn , then
A2 B(Az) = A(Az), so A2 (B Az − z) = 0. This yields B(B Az − z) = 0, since the null
spaces are identical. Since z is arbitrary, the conclusion follows.
Editorial comment. Jeff Stuart proved the following generalization: if Ak B = Ak−1
and rank (Ak−1 ) = rank (B k−1 ) for some integer k greater than 1, then B k A = B k−1 .
Also solved by A. Aguado & G. F. Seelinger, A. Alikhani & A. Dehkordi (Iran), S. Amghibech (Canada),
M. Barr (Canada), P. Budney, R. Chapman (U. K.), P. R. Chernoff, K. Dale (Norway), P. P. Dályay (Hungary),
L. M. DeAlba, G. Dospinescu (France), M. Goldenberg & M. Kaplan, J. Hartman, E. A. Herman, R. A. Horn,
A. K. Shaffie (Iran), G. Keselman, J. H. Lindsey II, O. P. Lossers (Netherlands), S. Rosset, K. Schilling,
N. C. Singer, J. H. Smith, A. Stadler (Switzerland), R. Stong, J. Stuart, T. Tam, X. Wang, BSI Problems
Group (Germany), Con Amore Problem Group (Denmark), GCHQ Problem Solving Group (U. K.), Microsoft
Research Problems Group, NSA Problems Group, and the proposer.
Fibonacci Numbers and Tiling a Board with Cuts

11241 [2006, 656]. Proposed by Roberto Tauraso, Università di Roma “Tor Vergata”,
Rome, Italy. Find a closed formula for

n
k+1
2n−k F1+2xi ,
k=0 x∈S[k,n] i=1
where Fn denotes the nth Fibonacci number (that is, F0 = 0, F1 = 1, and F j = F j −1 +

F j −2 when j ≥ 2) and S[k, n] is the set of all (k + 1)-tuples of nonnegative integers
that sum to n − k.
Solution I by O. P.
Lossers, Eindhoven University of Technology, Eindhoven, The
Netherlands. Since xi = n − k for x ∈ S[k, n], we can draw the factors of 2 into
k+1
the terms of the summation and obtain nk=0 x∈S[k,n] i=1 G xi , where G i = 2i F1+2i .
The Fibonacci recurrence yields Fn+2 = 3Fn − Fn−2 , so the sequence G i is defined
by G 0 = 1, G 1 = 4, and G n+2 = 6G n+1 − 4G n for n ≥ 0. The generating function
G(y) of this sequence is given by

∞
1 − 2y
G(y) = G i yi = .
i=0
1 − 6y + 4y 2

The desired expression is the coefficient of y n in ∞ k k+1
k=0 y G (y). Hence its gener-
G(y) 1−2y
ating function is 1−yG(y) , which equals 1−7y+6y 2 . This expands by partial fractions to
1
5
(1 − y)−1 + 45 (1 − 6y)−1 . Hence the coefficient of y n is (1 + 4 · 6n )/5.
Solution II by the proposer. Let an be the number of domino tilings of a 4-by-2n region
[0, 2n] × [0, 4] in which dominos are not allowed to cross the n − 1 vertical cuts from
(2i, 2) to (2i, 4) for 1 ≤ i ≤ n − 1. The cuts are indicated by heavy lines in the figure
below for n = 5.
858
We show first that the given expression equals an . Call each 4-by-2 rectangle a tooth. If
the top part of a tooth is not tiled with two horizontal dominos or two vertical dominos,
then the tooth is tiled like the second tooth shown above. Let k be the number of teeth
tiled in this way. For each of the n − k remaining teeth, the top part can be tiled in
two ways. This leaves regions that are 2-by-2xi rectangles, where x0 , . . . , xk is an
element of S[k, n]. It is well known (by induction using the Fibonacci recurrence) that
the number of domino tilings of a 2-by-m rectangle is F1+m . Hence we have proved
that the number of tilings equals the given sum.
Next we obtain a recurrence for the sequence an by considering how the last tooth
can be tiled. Consider whether the domino covering the lower right corner is horizontal
or vertical. If it is horizontal, then no domino crosses from the last tooth to the one
before it, and there are three ways to tile the rest of the last tooth. Hence there are
3an−1 tilings of this type.
If the lower right domino is vertical, then there are two ways to tile the top part of the
tooth, and what remains is a 4-by-2(n − 1) region with the two lower rows extended
by one unit. Let bn−1 be the number of ways to tile this region. With two ways to tile
the upper half of the last tooth, we have an = 3an−1 + 2bn−1 .
For bn−1 , we also consider whether the domino covering the lower right corner is
horizontal or vertical. If vertical, then there are an−1 ways to complete the tiling. If
horizontal, then there is another horizontal tile above it, two ways to tile the top half
of tooth n − 1, and bn−2 ways to complete the rest. Hence bn−1 = an−1 + 2bn−2 .
Using the first recurrence to substitute into the second and eliminate the auxiliary
sequence yields an = 7an−1 − 6an−2 . With a0 = 1 and a1 = 5, the solution is an =
(1 + 4 · 6n )/5.
Also solved by S. Amghibech (Canada), D. Beckwith, R. Chapman (U. K.), P. P. Dályay (Hungary), J. W. From-
meyer, C. C. Heckman, G. Keselman, K. McInturff, N. C. Singer, A. Stadler (Switzerland), A. Stenger,
R. Stong, BSI Problems Group (Germany), GCHQ Problem Solving Group (U. K.), Microsoft Research Prob-
lems Group, and the proposer.
“The Ballot Problem in Disguise”, in Disguise

11249 [2006, 760]. Proposed by David Beckwith, Sag Harbor, NY. A node-labeled
rooted tree is a tree such that any parent with label has + 1 children, labeled
1, 2, . . . , + 1, and such that the root vertex (generation 0) has label 1. Find the pop-
ulation of generation n.
2n
Solution I by Michael R. Avidon, Allston, MA. The answer is 2n n
− n+2 . Let P(n)
be the population of generation n and P j (n) be the number with label j. Thus P(n) =
n+1
j =1 P j (n). Children with label j come (one each) from parents with label at least
2n− j 2n− j
n
j − 1, so P j (n) = i= j −1 Pi (n − 1). We claim that P j (n) = n−1 − n+1 for n ≥
m i
1 and 1 ≤ j ≤ n + 1. Note that this holds for n = 1. Inductively, since i=k =
m+1 n k
k+1
, summing P
i= j −1 i (n − 1) yields

n
2n − 2 − i n
2n − 2 − i 2n − j 2n − j
P j (n) = − = − .
i= j −1
n−2 i= j −1
n n−1 n+1
n+1
Now evaluate j =1 P j (n) to find P(n).
Solution II by Li Zhou, Polk Community College, Winter Haven, FL. The answer is
1 2n+2
the Catalan number Cn+1 , which equals n+2 n+1
. Identify each member of genera-
tion n with the list (a0 , . . . , an ) of labels on the path to it from the root. For n ≥ 0,

let sn+1 = |Sn+1 |, where Sn+1 = {(a0 , . . . , an ) : a0 = 1 and 1 ≤ ai ≤ ai−1 + 1 for 1 ≤
i ≤ n}. We take s0 = 1, since S0 consists only of the empty 0-tuple. For n ≥ 1, consider
(a0 , . . . , an ) ∈ Sn+1 . When a1 = 1, let k = 0. When a1 ≥ 2, let k be the largest i such
that a j ≥ 2 for 1 ≤ j ≤ i. Now (a1 − 1, . . . , ak − 1) ∈ Sk and (ak+1 , . . . , an ) ∈ Sn−k .
Grouping by k yields sn+1 = nk=0 sk sn−k , which is the well-known Catalan recur-
rence.
Solution III by Richard Stong, Rice University, Houston, TX. The answer is the Catalan
number Cn+1 , which counts the paths from (0, 0) to (2n + 2, 0) with up-steps (+1, +1)
and down-steps (+1, −1) that never go below the x-axis (“Dyck paths”). We exhibit a
bijection from such paths to the vertices in generation n. From a path, we form a list
(a0 , . . . , an ) by letting ak be the height of the path after the (k + 1)th up-step. Note
that a0 = 1. If ak = , then ak+1 can be any of {1, . . . , + 1}, since between 0 and
down-steps before can occur before the next up-step. Furthermore, each element of
the set of path-labels to vertices of generation n arises exactly once under this map.
Editorial comment. Several solvers noted that P j (n) is the appropriate entry in the
Catalan triangle (see sequence A009766 in The On-Line Encyclopedia of Integer Se-
quences). After shifting the index by 1, that value is in fact what is requested in Prob-
lem E3402 of this M ONTHLY [1990, 612], with essentially the same solution (“The
Ballot Problem in Disguise”, [1992, 367–368]) as in Solution I here. Robin Chapman
mentioned that these trees appear in two papers by Julian West (Discrete Mathemat-
ics 146 (1995) 247–262 and 157 (1996) 363–374). Daniele Degiorgi pointed out that
a solution appears in Donald Knuth’s The Art of Computer Programming, Volume
4, page 13. The BSI Problems Group observed that this is essentially one of the 66
characterizations of the Catalan numbers given in Exercise 6.19 of Richard Stanley’s
Enumerative Combinatorics, Volume 2.
Also solved by T. Achenbach, M. R. Bacon & C. K. Cook, J. C. Binz (Switzerland), N. Caro (Brazil), R. Chap-
man (U. K.), A. Chaudhuri, K. Dale (Norway), P. P. Dályay (Hungary), D. Degiorgi (Switzerland), J.-P. Gri-
vaux (France), M. Hudelson, G. Keselman, R. A. Kopas, Y.-J. Kuo, H. Kwong, C. F. Letsche, J. H. Lindsey II,
O. P. Lossers (Netherlands), R. Martin (U. K.), K. McInturff, M. A. Prasad (India), R. E. Prather, V. Schindler
(Germany), E. Schmeichel, R. Tauraso (Italy), D. Walsh, BSI Problems Group (Germany), Szeged Problem
Solving Group “Fejéntaláltuka” (Hungary), GCHQ Problem Solving Group (U. K.), Microsoft Research Prob-
lems Group, NSA Problems Group, and the proposer.
A Permanent Lower Bound

11253 [2006, 847]. Proposed by David Beckwith, Sag Harbor, NY. Let n be a positive
integer and A be an n × n matrix with all entries ai, j positive. Let P be the permanent
of A. Prove that

1/n
P ≥ n! ai, j .
1≤i, j ≤n
Solution by Michel Bataille, Rouen, France. Let Sn be the set of permutations

n
of {1, . . . , n}. Since P = σ ∈Sn aσ , where aσ = i=1 ai,σ (i) > 0, the arithmetic-
1/n!
geometric mean inequality yields P/n! ≥ σ ∈Sn aσ . Now

n

n
n
aσ = ai,σ (i) = ai,(n−1)!
j ,
σ ∈Sn i=1 σ ∈Sn i=1 j =1
and the result follows.
860
Editorial comment. Equality holds if and only if all aσ are equal, which requires that
ai j = αi β j for all i and j, where all αi and β j are positive numbers.
Also solved by S. Amghibech (Canada), J. C. Binz (Switzerland), P. Budney, R. Chapman (U. K.), K. Dale
(Norway), P. P. Dályay (Hungary), S. M. Gagola Jr., J.-P. Grivaux (France), E. A. Herman, S. Hitotumatu
(Japan), D. Karagulyan (Armenia), P. T. Krasopoulos (Greece), O. López & N. Caro (Brazil), O. P. Lossers
(Netherlands), R. Martin (U. K.), J. Rooin & M. A. Maleki (Iran), K. Schilling, V. Schindler (Germany),
A. Stadler (Switzerland), A. Stenger, R. Stong, R. Tauraso (Italy), M. Tetiva (Romania), J. Vinuesa (Spain), BSI
Problems Group (Germany), GCHQ Problem Solving Group (U. K.), Microsoft Research Problems Group,
Missouri State University Problem Solving Group, NSA Problems Group, and the proposer.
Primes with Special Primitive Roots

11254 [2006, 847]. Proposed by Lenny Jones, Shippensburg University, Shippensburg,
PA. For a prime p greater than 3, let S p be the set of positive integers less than
( p − 1)/2 and relatively prime to p − 1. Characterize the primes p for which there
exists a primitive root g modulo p such that the product of g a , taken over all a in S p ,
is also a primitive root modulo p.
Solution by Toni Ernvall, University of Turku, Finland. Let g be a primitive
root mod-
ulo p. The product of g a , taken over all a in S p , is g s , where s = a∈S p a, and it is
a primitive root modulo p if and only if gcd(s, p − 1) = 1. We will prove that this
happens only when p = 5 and when p has the form 2q m + 1, where m is a positive
integer, q is a prime, and q ≡ 3 (mod 4).
Suppose first that p ≡ 1 (mod 4). If p = 5, then S p = {1} and any primitive
root works. For p = 5, we have a ∈ S p if and only if ( p − 1)/2 − a ∈ S p . Since
( p − 1)/4 ∈ / S p , this makes |S p | even. Since every element of S p is odd, s is even and
gcd(s, p − 1) = 1. Hence p is not such a prime.
Next consider p ≡ 3 (mod 4), with gcd(s, p − 1) = 1. The set of positive num-
bers less than p − 1 and relatively prime to p − 1 consists of S p and those a such that
p − 1 − a ∈ S p . Hence |S p | = 12 φ( p − 1). Since elements of S p are odd, |S p | ≡ s
(mod 2). Thus gcd(s, p − 1) = 1 forces |S p | to be odd, and hence φ( p − 1) ≡ 2
(mod 4).
Always φ(n) = n q∈P(n) q−1 q
, where P(n) is the set of distinct prime factors of n.
When n ≡ 2 (mod 4), the number of factors of 2 dividing φ(n) is thus the number of
distinct odd prime factors of n, plus 1 or more for each that is congruent to 1 modulo
4. Since φ( p − 1) ≡ 2 (mod 4), we conclude that p − 1 = 2q m , where m is a positive
integer, q is a prime, and q ≡ 3 (mod 4).
For sufficiency, we compute s for such a prime p. We sum the odd integers less than
( p − 1)/2 that are not divisible by q. Since q | ( p − 1)/2, we sum the odd numbers
up to ( p − 1)/2 and subtract q times the sum of the odd numbers up to ( p − 1)/2q.
Using p − 1 = 2q m , we obtain
2
p+1 2 1 p−1 1 q
s= −q + 1 = (q m + 1)2 − (q m−1 + 1)2 .
4 4 2q 4 4
After simplifying, s equals 14 (q − 1)(q 2m−1 − 1), which is relatively prime to 2q m .
Hence g s is a primitive root, as desired.
Also solved by M. R. Avidon, J. C. Binz (Switzerland), J. Christopher, S. M. Gagola Jr., K. Goldenberg &
M. Kaplan, D. E. Ianucci, O. P. Lossers (Netherlands), B. Schmuland (Canada), A. Stadler (Switzerland),
V. Stakhovsky, R. Stong, M. Tetiva (Romania), A. Wyn-Jones, Armstrong Problem Solvers, BSI Problems
Group (Germany), GCHQ Problem Solving Group (U. K.), Microsoft Research Problems Group, and the
proposer.

Never a Square
11259 [2006, 939]. Proposed by Nobuhisa Abe, NBU Attached Senior High School,
Saiki, Japan. For integers n greater than 2, let

n−2
f (n) = 2j k,
j =1 S k∈S
where the sum is over all j-element subsets S of the set {1, . . . , n − 1}. Show that
4(2n − 1)! + ( f (n))2 is never the square of an integer.

Solution by Marian Tetiva, Bı̂rlad, Romania. By expanding n−1k=1 (1 + ck x) and setting
n−1
ck = k and x = 2, we obtain f (n) = k=1 (2k + 1) − (n − 1)!2n−1 − 1. Letting a =
n−1 n−1
k=1 (2k + 1) and b = k=1 2k, we have f (n) = a − b − 1, and our task is to show
that 4ab + (a − b − 1)2 is not a square. This follows from
(a + b − 1)2 < 4ab + (a − b − 1)2 < (a + b)2 ,
which holds whenever a > b > 0. That condition holds when n ≥ 2.
Also solved by S. Amghibech (Canada), A. Bandeira & J. Moreira (Portugal), D. Beckwith, R. Chapman
(U. K.), P. P. Dályay (Hungary), O. Kouba (Syria), O. P. Lossers (Netherlands), C. R. Pranesachar (India),
A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), BSI Problems Group (Germany), FAU Problem Solving
Group, and the proposer.
Getting to Know You, Directly and Indirectly

11262 [2006, 940]. Proposed by Ashay Burungale, Satara, Maharashtra, India. In a
certain town of population 2n + 1, one knows those to whom one is known. For any
set A of n citizens, there is some person among the other n + 1 who knows everyone
in A. Show that some citizen of the town knows all the others.
Solution I by Kenneth Schilling, University of Michigan—Flint, Flint, MI. A clique is
a set of citizens who all know one another. If a clique B has fewer than n + 1 citizens,
then by hypothesis someone not in B knows everyone in B, so we may form a larger
clique by adding that person to B. Hence there is a clique of size n + 1. The set A of
citizens not in B has size n, so by hypothesis some person in B knows every member
of A and hence knows everyone.
Solution II by Christopher Carl Heckman, Arizona State University, Tempe, AZ. We
prove the contrapositive: If nobody knows everyone, then some set A of n citizens is
not completely known by anyone outside A. We first color each citizen red or green
as follows. If p is not yet colored, choose some p unknown to p. If p is already
colored, give p the opposite color. Otherwise, give p and p opposite colors. At the
end, everyone has a color opposite from that of someone he or she does not know. The
less-frequent color occurs at most n times; let A be a set of size n including all those
with that color. Each person outside A has the other color and hence does not know
everyone in A.
Editorial comment. The problem of course is a statement about graphs (perhaps the
popularity of the problem stemmed from not using that language). In the language of
graph theory, it states that a graph of odd order has a vertex adjacent to all others if
each set with fewer than half the vertices has a common neighbor. Several solvers did
use graph theoretic language and arguments.
862
with the collaboration of Mario Benedicty, Itshak Borosh, Paul Bracken, Ezra A.
Brown, Randall Dougherty, Tamás Erdélyi, Zachary Franco, Christian Friesen, Ira
M. Gessel, László Lipták, Frederick W. Luttmann, Vania Mascioni, Frank B. Miles,
Richard Pfiefer, Cecil C. Rousseau, Leonard Smiley, Kenneth Stolarsky, Richard
Stong, Walter Stromquist, Daniel Ullman, Charles Vanden Eynden, Sam Vandervelde,
and Fuzhen Zhang.

that address before April 30, 2009. Additional information, such as generaliza-
PROBLEMS
11397. Proposed by Grahame Bennett, Indiana University, Bloomington, IN. Let
a, b, c, x, y, z be positive numbers such that a + b + c = x + y + z and abc = x yz.
Show that if max{x, y, z} ≥ max{a, b, c}, then min{x, y, z} ≥ min{a, b, c}.
11398. Proposed by Stanley Huang, Jiangzhen Middle School, Huaining, China. Sup-
pose that acute triangle ABC has its middle-sized angle at A. Suppose further that the
incenter I is equidistant from the circumcenter O and the orthocenter H . Show that
angle A has measure 60 degrees and that the circumradius of I BC is the same as that
of ABC.
11399. Proposed by Biaggi Ricceri, University of Catania, Catania, Italy. Let
(, F , μ) be a measure space with finite nonzero measure M, and let p > 0. Let
f be a lower semicontinuous function on R with the property that f has no global
minimum, but for each λ > 0, the function t → f (t) + λ|t| p does have a unique
global minimum. Show that exactly one of the two following assertions holds:
(a) For every u ∈ L p () that is not essentially constant,
1/ p
1
Mf |u(x)| dμ
p
< f (u(x)) dμ,
M
and f (t) < f (s) whenever t > 0 and −t ≤ s < t.

(b) For every u ∈ L p () that is not essentially constant,
1/ p
1
Mf − |u(x)| dμ
p
< f (u(x)) dμ,
M
and f (−t) < f (s) whenever t > 0 and −t < s ≤ t.

11400. Proposed by Paul Bracken, University of Texas-Pan American, Edinburg, TX.
Let ζ be the Riemann zeta function. Evaluate ∞ n=1 ζ (2n)/(n(n + 1)) in closed form.
948
11401. Proposed by Marius Cavachi, “Ovidius” University of Constanţa, Constanţa,
Romania. Let A be a nonsingular square matrix with integer entries. Suppose that for
every positive integer k, there is a matrix X with integer entries such that X k = A.
Show that A must be the identity matrix.
11402. Proposed by Catalin Barboianu, Infarom Publishing, Craiova, Romania. Let
f : [0, 1] → [0, ∞) be a continuous function such that f (0) = f (1) = 0 and f (x) >
0 for 0 < x < 1. Show that there exists a square with two vertices in the interval (0,1)
on the x-axis and the other two vertices on the graph of f .
11403. Proposed by Yaming Yu, University of California Irvine, Irvine, CA. Let n be
an integer greater than 1, and let f n be the polynomial given by
n

i−1
n
f n (x) = (−x)n−i (x + j).
i=0
i j =0
Find the degree of f n .
SOLUTIONS
A Telescoping Fibonacci Sum

11258 [2006, 939]. Proposed by Manuel Kauers, Research Institute for Symbolic Com-
putation, Johannes Kepler
√ University, Linz, Austria. Let Fn denote the nth Fibonacci
number, and let i denote −1. Prove that
∞
F3k − 2F1+3k 1
√
=i+ 1− 5 .
k=0
F3k + i F2·3k 2
Solution by√Richard Stong, Rice University, Houston, TX. Let φ denote the golden
ratio (1 + 5)/2, and recall the Binet formula for the Fibonacci numbers:
√
Fn = φ n + (−1)n−1 φ −n / 5.
For odd m,
√
5(Fm + i F2m ) = iφ 2m + φ m + φ −m − iφ −2m = −iφ −2m (iφ m + 1)(iφ 3m + 1)
and
1 − 2φ m 1 + 2φ −1 −m
Fm − 2F1+m = √ φ + √ φ = −φ m + φ −m .
5 5
Hence
√ √
Fm − 2F1+m √ −iφ 3m + iφ m 5 5
= 5 m = − .
Fm + i F2m (iφ + 1)(iφ 3m + 1) iφ 3m + 1 iφ m + 1
Thus the desired sum telescopes as

√ √ √
∞
F3k − 2F1+3k ∞
5 5 5
= − 3k =− = i − φ −1 .
k + i F2·3k + + +
k+1
k=0
F 3 k=0 iφ 3 1 iφ 1 iφ 1
December 2008] PROBLEMS AND SOLUTIONS 949

Also solved by S. Amghibech (Canada), M. R. Avidon, M. Bataille (France), R. Chapman (U. K.), C. K. Cook,
M. Goldenberg & M. Kaplan, C. C. Heckman, G. C. Greubel, D. E. Iannucci, H. Kwong, O. P. Lossers (Nether-
lands), K. McInturff, C. R. Pranesachar (India), H. Roelants (Belgium), H.-J. Seiffert (Germany), A. Stadler
(Switzerland), GCHQ Problem Solving Group (U. K.), Microsoft Research Problems Group, and the proposer.
3-D Rotations and Translations

11276 [2007, 165]. Proposed by Eugene Herman, Grinnell College, Grinnell, IA. Let
T1 , . . . , Tn be translations in R3 with translation vectors t1 , . . . , tn , and let R be a
rotational linear transformation on R3 that rotates space through an angle of π/n about
an axis parallel to a vector r. Define a transformation C by C = (RTn · · · RT2 RT1 )2 .
Prove that C is a translation, find an explicit formula for its translation vector in terms
of r, n, and t1 , . . . , tn , and prove that there is a line in R3 , independent of t1 , . . . , tn ,
such that C translates space parallel to .
Solution by Mark D. Meyerson, US Naval Academy, Annapolis, MD. If Tv denotes
translation by vector v, then RTv = TRv R. Now R is a rotation through π/n, so R n
is a half turn about the axis of R and R 2n is the identity. If v is orthogonal to r, then
R n v = −v and Tv+Rn v is the identity.
Let pk be the vector projection of tk onto r, and let ok be the orthogonal component,
so that tk = pk + ok . Now Tk = Tpk Tok . Furthermore, R takes pk to itself, R and Tpk
commute, and any two translations commute.
Let p = 2(pn + · · · + p2 + p1 ). We compute
C = (RTn · · · RT2 RT1 )2 = (RTon · · · RTo2 RTo1 )2 (Tpn · · · Tp2 Tp1 )2 .
The second factor reduces to Tp . In the first factor, move the rotations to the right:
(RTon · · · RTo2 RTo1 )2
= TRon RTRon−1 R · · · TRo2 RTRo1 RTRon RTRon−1 R · · · TRo2 RTRo1 R
= TRon TR2 on−1 R · · · RTR2 o2 RTR2 o1 RTR2 on RTR2 on−1 R · · · RTR2 o2 RTR2 o1 R 2
= · · · = TRon +Rn+1 on TR2 on−1 +Rn+2 on−1 · · · TRn−1 o2 +R2n−1 o2 TRn o1 +R2n o1 R 2n ;
the last expression is the identity. Hence C = Tp , translation through twice the sum of
the projections of the tk onto r, and can be any line parallel to r.
Also solved by R. Bagby, M. Bataille (France), D. R. Bridges, R. Chapman (U. K.), K. Claassen, K. Dale
(Norway), M. Englefield (Australia), A. Fok (Hong Kong), J.-P. Grivaux (France), J. A. Grzesik, G. Janusz,
J. H. Lindsey II, O. P. Lossers (Netherlands), R. Stong, T. Tam, E. I. Verriest, Szeged Problem Solving Group
“Fejéntaláltuka” (Hungary), GCHQ Problem Solving Group (U. K.), Hofstra University Problem Solvers, and
the proposer.
A Sinh of a Series
11286 [2007, 78]. Proposed by M. L. Glasser, Clarkson University, Potsdam, NY. Show
that
∞
π/4
e − (−1)n sinh(n + 1/2)π e−n(n+1)π = 0.
n=0
Solution by O. P. Lossers, Eindhoven University of Technology, The Netherlands. The

C ∞ function
∞
2
F(x, y) = e−(k+x) y , y > 0, (1)
k=−∞
950
is even and periodic in x with period 1. Therefore it is the sum of a cosine series
1 ∞
F(x, y) = A0 + An cos(2nπ x),
2 n=1
∞ 2
with An = 2 −∞
e−x y cos(2πnx) d x. It follows that

π
∞
−π 2 n 2 /y
F(x, y) = 1+2 e cos(2πnx) . (2)
y n=1
Substituting x = 1/2 and y = π into the expression for F(x, y) given in (1) leads to

∞
2π

−1
2π

∞
2π
F(1/2, π) = e−(k+1/2) = e−(k+1/2) + e−π/4 + e−(k+1/2)
k=−∞ k=−∞ k=1

∞
2π

∞
2π

∞
2
= e−(k−1/2) + 2e−π/4 + e−(k+1/2) =2 e−(k+1/2) π .
k=2 k=1 k=0
Substituting into (2), we obtain

∞
2

∞
2

∞
2
F(1/2, π) = 1 + 2 (−1)n e−πn = (−1)n e−πn + (−1)n e−πn
n=1 n=0 n=1

∞
2

∞
2
= (−1)n e−πn + (−1)n+1 e−π(n+1) .
n=0 n=0
Equating these two expressions for F, we find that

∞
2π

∞
2

∞
2
2 e−(k+1/2) = (−1)n e−πn + (−1)n+1 e−π(n+1) .
k=0 n=0 n=0
Multiplication of both sides by eπ/2 leads to

∞
∞

2 eπ/4 e−k(k+1)π = (−1)n e−πn(n+1) enπ+π/2 − e−nπ−π/2 .
k=0 n=0
Applying the definition of the hyperbolic sine, we obtain the desired equation.
Also solved by R. Chapman (U.K.), J. Grivaux (France), O. Kouba (Syria), G. Lamb, M. A. Prasad (India),
O. G. Ruehr, A. Stadler (Switzerland), R. Stong, J. Sun, FAU Problem Solving Group, GCHQ Problem Solving
Group, and the proposer.
A Variant Intermediate Value

11290 [2007, 359]. Proposed by Cezar Lupu, student, University of Bucharest,
Bucharest, Romania, and Tudorel Lupu, Decebal Highschool, Constanza, Roma-
nia. Let f and g be continuous real-valued functions on [0, 1]. Prove that there exists
c in (0, 1) such that
1 c 1 c
f (x) d x xg(x) d x = g(x) d x x f (x) d x.
x=0 x=0 x=0 x=0

Solution by Kenneth F. Andersen, University of Alberta, Edmonton, x AB, Canada. Ob-
serve first that if h(x) is continouous on [0, 1] and H (x) = 0 yh(y) dy, then H (x) is
continuous on [0, 1] with limx→0+ H (x)/x = 0, so an integration by parts yields
1 1 1
xh(x) H (x) 1 H (x) d x
h(x) d x = dx = +
0 0 x x 0 0 x2
1 1
H (x) d x H (x) d x
= H (1) + 2
= lim H (x) + . (1)
0 x x→1 −
0 x2
1
Now suppose in addition that 0 h(x) = 0. By (1), H (x) cannot be positive for all
x in (0, 1), nor can it be negative for all x in (0, 1). Thus by the Intermediate Value
Theorem there is a ch ∈ (0, 1) such that H (ch ) = 0. Now the required result may be
1 1
deduced: if 0 f (x) d x = 0, then the result holds with c = c f ; if 0 g(x) d x = 0, then
the result holds with c = cg . Otherwise the result holds with c = ch , where
f (x) g(x)
h(x) = 1 − 1 .
0
f (y) dy 0
g(y) dy
Editorial comment. (i) The functions f and g need not be continuous—it is sufficient
that they be integrable. This was observed by Botsko, Pinelis, Schilling, and Schmu-
1 1
land. (ii) Keselman, Martin, and Pinelis noted that 0 x f (x) d x and 0 xg(x) d x can
1 1
be replaced with 0 φ(x) f (x) d x and 0 φ(x)g(x) d x, where φ(x) satisfies suitable
conditions—roughly speaking, that φ is differentiable and strictly monotonic, although
the specific conditions vary from one of these solvers to another.
Also solved by U. Abel (Germany), S. Amghibech (Canada), M. W. Botsko & L. Mismas, R. Chapman (U. K.),
J. G. Conlon & W. C. Troy, P. P. Dályay (Hungary), J. W. Hagood, E. A. Herman, S. J. Herschkorn, E. J.
Ionascu, G. L. Isaacs, G. Keselman, O. Kouba (Syria), J. H. Lindsey II, O. P. Lossers (Netherlands), G. Martin
(Canada), J. Metzger & T. Richards, M. D. Meyerson, A. B. Mingarelli & J. M. Pacheco & A. Plaza (Spain), E.
Mouroukos (Greece), P. Perfetti (Italy), I. Pinelis, M. A. Prasad (India), K. Schilling, B. Schmuland (Canada),
H.-J. Seiffert (Germany), J. Sun, R. Tauraso (Italy), M. Tetiva (Romania), L. Zhou, GCHQ Problem Solving
Group (U. K.), Microsoft Research Problems Group, NSA Problems Group, and the proposer.
Double Integral
11295 [2007, 452]. Proposed by Stefano Siboni, University of Trento, Trento, Italy.
For positive real numbers and ω, let M be the mapping of [0, 1) × [0, 1) into itself
defined by M(x, y) = ({2x}, {y + ω + x}), where {u} denotes u −
u, the fractional
part of u. For integers a and b, let ea,b (x, y) = e2πi(ax+by). Let
1 1
Cn (a, b; p, q) = ea,b (M n (x, y))e p,q (x, y) d x dy.
y=0 x=0
Show that Cn (a, b; p, q) = 0 if q = b, while Cn (a, b; p, b) is given by

n
sin π(a + b − 2−n ( p + b))
(−1) eb,b (ωn, n/2)
a
cos π(b − 2− j ( p + b) .
π (a + b − 2 ( p + b)) j =0
−n

Netherlands. Let e(z) = e2πi z . For n ≥ 1 we have

M n (x, y) = 2n x , y + nω + ({x} + {2x} + · · · + {2n−1 x}) .
952
Since a and b are integers, the outer braces do not matter for Cn (a, b; p, q) =
1 1

e (2n a − p)x + (b − q)y + bnω + b({x} + {2x} + · · · + {2n−1 x}) d x dy.
0 0
Interchanging the order of integration shows that Cn (a, b; p, q) = 0 if q = b and that

1

Cn (a, b; p, b) = e(bωn) e (2n a − p)x + b({x} + {2x} + · · · + {2n−1 x}) d x.
0
For n ≥ 1 define
1
Fn (α, β) = exp iαx + 2iβx + iβ({2x} + · · · + {2n−1 x}) d x,
0
so that Cn (a, b; p, b) = e(bωn)Fn (2π(2n a − p − b), 2πb). Now

ei(α+2β) − 1 ei(β+α/2) β α β α
F1 (α, β) = = cos + sin + .
i(α + 2β) β/2 + α/4 2 4 2 4
For n > 1, the last term in the exponent has period 1/2, so we split the integral into
two parts to obtain

i(α/2+β) 1/2
Fn (α, β) = e +1 exp i(α + 4β)x + iβ({4x} + · · · + {2n−1 x}) d x,
0
which gives the recurrence

α
β α
Fn (α, β) = ei(β/2+α/4) cos + Fn−1 ,β .
2 4 2
Repeated use of this, together with the formula for F1 , leads to

eiβ(n+1/2)+iα/2 β α n
β α
Fn (α, β) = sin + n+1 cos + j +1 .
β/2 + α/2n+1 2 2 j =1
2 2
Finally, Cn (a, b; p, b) = e2πibωn Fn (2π(2n a − p − b), 2πb) yields the required re-
sult after some simplifications based on the assumption that a and b are integers.
Also solved by R. Chapman (U. K.), D. Fleischman, GCHQ Problem Solving Group, and the proposer.
A Tricky Minimum
11297 [2007, 452]. Proposed by Marian Tetiva, Bı̂rlad, Romania. For positive a, b,
and c, let
a 2 b2 c2 − 64
E(a, b, c) = .
(a + 1)(b + 1)(c + 1) − 27
Find the minimum value of E(a, b, c) on the set D consisting of all positive triples
(a, b, c) such that abc = a + b + c + 2, other than (2, 2, 2).
Solution by John H. Lindsey II, Cambridge, MA. Let m be the geometric mean, de-
fined by m = (abc)1/3 . By the arithmetic-geometric mean inequality, a + b + c ≥ 3m,

with equality if and only if a = b = c. Thus m 3 = a + b + c + 2 ≥ 3m + 2, or
(m − 2)(m + 1)2 = m 3 − 3m − 2 ≥ 0 and hence m ≥ 2. Equality forces a = b =
c = 2, so in fact m > 2. Using the arithmetic-geometric mean inequality again to ob-
tain ab + bc + ca ≥ 3((ab)(bc)(ca))1/3 = 3m 2 , we see that (a + 1)(b + 1)(c + 1) >
m 3 + 3m 2 + 3m + 1 > 27. Thus the numerator and denominator of E are always pos-
itive on D.
For fixed (a, b, c) ∈ D, consider all triples (a , b , c ) ∈ D with a b c = abc. For
such triples the numerator of E is fixed and the denominator will be maximized (hence
E minimized) if we maximize a b + b c + c a . Since a + b + c = a + b + c is
fixed (at abc − 2), a , b , and c are bounded above; since also a b c is fixed and
positive, they are bounded away from zero. Thus they form a closed bounded set.
Hence we may choose a , b , c to maximize a b + b c + c a .
Suppose this maximum occurs for a , b , c distinct. By symmetry we may assume
that a < b < c . Let f (x) = (x − a )(x − b )(x − c ). When is positive and suf-
ficiently small, f (x) + x has three distinct positive roots with the same sum and
product as a, b, c (since f is a cubic polynomial), but this contradicts maximality of
the denominator. Thus two of a , b , c must be equal. Hence it suffices to minimize E
under the additional constraint a = b. In this case the condition abc = a + b + c + 2
gives c = a−12
and we compute

2 4(a 2 + 4a − 4) 17 1
E a, a, = =4− +
a−1 (a + 7)(a − 1) 2(a + 7) 2(a − 1)
and

d 2 17 1
E a, a, = − .
da a−1 2(a + 7)2 2(a − 1)2
√ √
The unique critical point occurs at a = 3+2 17 where E = 23+8 17 ≈ 3.390388. As a →
1+ , E → ∞ and as a → ∞, E → 4 so this is the minimum of E.
Also solved by A. Alt, J. Grivaux (France), E. A. Herman, G. I. Isaacs, K.-W. Lau (China), GCHQ Problem
Solving Group (U. K.), Microsoft Research Problems Group, and the proposer.
Errata and End Notes for 2008.
An Infinite Product Based on a Base

11222 [2006, 459]. Proposed by Jonathan Sondow, New York, NY. Fix an integer B ≥
2, and let s(n) denote the sum of the base-B digits of n. Prove that
∞
n B + k (−1)s(n) 1
=√ .
n=0 k odd
nB + k + 1 B
0<k<B
Solution by the proposer. Set (n) = (−1)s(n) . If B is odd, then s(n) ≡ n mod 2, since
all powers of B are odd. If B is even, then the constant term in the base-B expansion of
2m cannot be B − 1, and hence s(2m + 1) = s(2m) + 1. Hence in both cases (2m +
1) = −(2m). Let δk = 1 if k = 0, and otherwise δk = 0, and let
∞
n B + k (n)
Pk,B = Pk = .
n=δk
nB + k + 1
954
Then Pk converges because it is a product over m ≥ 0 of factors of the form

(2m)B + k + 1 (2m + 1)B + k ±1
· ,
(2m)B + k (2m + 1)B + k + 1
which simplifies to (1 + x(x+B+1)

B+1
)±1 , where x = 2m B + k. Thus products of the form

k∈S Pk converge for any finite subset S of [0, B − 1], and in particular, the original
product converges.
We now consider a product of Pk ’s that telescopes nicely:

B−1
n B + k (n) B−1
k (0) ∞ B−1
n B + k (n)
=
k=0 n=δk
nB + k + 1 k=1
k+1 n=1 k=0
nB + k + 1
(1)
∞
1 m (m)
= .
B m=1 m + 1
If 0 ≤ k < B, then s(n B + k) = s(n) + k, so (n B + k) = (−1)k (n). After splitting

the last product in (1) by collecting factors with the same residue modulo B, apply
(n B + k) = (−1)k (n) to obtain
∞
m (m) B−1 ∞
n B + k (−1)k (n)
= .
m=1
m+1 k=0 n=δk
nB + k + 1
Substitute this into (1). Since the infinite products are all nonzero (being convergent
and having no zero factors), the factors for even k are the same on both sides and
cancel out. This yields
∞
n B + k (n) 1 ∞
n B + k −(n)
= .
k odd n=0
nB + k + 1 B k odd n=0 n B + k + 1
0<k<B 0<k<B
All the products are positive, and the desired formula follows.
Editorial comment. The solution of Problem 11222 for odd values of B given in the
May, 2008 issue of the Monthly was selected at a time when the proposer’s solution had
become separated from the file of solutions. The previously published solution treated
odd B only by reference to the literature. The proposer’s elegant solution covers all
cases simultaneously and efficiently. Fortunately, it was recovered, and we are pleased
to present it.
The names of solver Apostolis Demis, of Athens, Greece (11285, [2007,358]), and
William Dickinson, (11201, [2008,73]) were misspelled. Our apologies.
Paolo Perfetti gives a counterexample to the
if-direction of part (a) in 11257,[2008,
269]. If z k = (−1)k−1 / ln(k + 1) (and sk = 1j z k ), then sn converges, but z k /sk
diverges, tending to −∞.

A M M P R o B L e M S 2 0 0 8

Caricato da

Informazioni sul documento

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

A M M P R o B L e M S 2 0 0 8

Caricato da

Copyright:

Formati disponibili

PROBLEMS AND SOLUTIONS

Edited by Gerald A. Edgar, Doug Hensley, Douglas B. West

Proposed problems and solutions should be sent in duplicate to the MONTHLY

11334. Proposed by Jonathan Bober and Jeffrey Lagarias, University of Michigan,

Furdui, Cluj, Romania. Let denote the classical gamma

January 2008] PROBLEMS AND SOLUTIONS 71

A Condition on the Commutator

January 2008] PROBLEMS AND SOLUTIONS 73

Differentiable and Discontinuous Densely

has the required properties.

To prove this finite sum is g (x0 ), let

Snapshots for Velocities

January 2008] PROBLEMS AND SOLUTIONS 75

with snapshots at times 1, . . . , P. Hence P(2n) ≥ P(n) + 1. It follows that P(n) ≥

Solution by Microsoft Research Problems Group, Redmond, WA. If a ≥ 1 and x ≥ 0,

0 ≤ f (x) ≤ (1 − x a1 )(1 − x a2 ) ≤ a1 a2 (1 − x)2 .

Solution by Minh Can, Irvine Valley College, Irvine, CA.

January 2008] PROBLEMS AND SOLUTIONS 77

Specializing to f (x) = sin(x) gives the desired result.

from which the above formula follows by taking k = n and m = n − 1.

Proposed problems and solutions should be sent in duplicate to the MONTHLY

11342. Proposed by Luis H. Gallardo, University of Brest, Brest, France. Let p be a

Orthogonal Sudoku Squares

February 2008] PROBLEMS AND SOLUTIONS 167

Maximum Velocity on a Car Trip

Goes Back to Boole

Solution by Eugene A. Herman, Grinnell College, Grinnell, IA. Let

February 2008] PROBLEMS AND SOLUTIONS 169

The last step follows from equation (3).

Coloring the Plane

degrees. Now D B E is an equilateral triangle√with sides of length 1 + 3, and AC D

February 2008] PROBLEMS AND SOLUTIONS 171

Since 0 ≤ (a − b)2 /c2 < 1 and

February 2008] PROBLEMS AND SOLUTIONS 173

Proposed problems and solutions should be sent in duplicate to the MONTHLY

11349. Proposed by Cezar Lupu (student), University of Bucharest, Bucharest, Roma-

converges with probability one.

March 2008] PROBLEMS AND SOLUTIONS 263

By approximation, we may assume that X 1 has finite support:

Proof. If C is nonsingular, then

= det C · det(D − XC −1 Y ) = det(C D − C XC −1 Y )

= det(C D − XCC −1 Y ) = det(C D − X Y ).

A Bessel Function Identity

Show that, for any positive b and any real λ,

Solution by Mourad E. H. Ismail, University of Central Florida, Orlando, FL. First

Formula (9.0.1) in the same reference is

This, together with

proves the identity claimed.

March 2008] PROBLEMS AND SOLUTIONS 265

Proof. Either h Y ≤ 1 or h Z ≤ 1, say h√ Y ≤ 1. By the Law of Sines, sin  X ≤

Double Sum Inequality

Expand f (x) and g(x) by Taylor’s theorem:

An Inequality Proved Without Computer Assistance

March 2008] PROBLEMS AND SOLUTIONS 267

Solution by Marian Tetiva, National College “Gheorghe Roşca Codreanu”, Bârlad,

A Myth About Infinite Products

March 2008] PROBLEMS AND SOLUTIONS 269

Proposed problems and solutions should be sent in duplicate to the MONTHLY

11357. Proposed by Mehmet Şahin, Ankara, Turkey. Let Ia , Ib , Ic and ra , rb , rc be

11358. Proposed by Marian Tetiva, National College “Gheorghe Roşca Codreanu”,

April 2008] PROBLEMS AND SOLUTIONS 365

A Vanishing Alternating Sum

Comparing coefficients of t n+1 yields the desired identity.

Proof. Either h Y ≤ 1 or h Z ≤ 1, say h√ Y ≤ 1. By the Law of Sines, sin X ≤

nβ = (nα + n(t − 1))α + 1 = (nα + n(t − 1) + 1)α − 1.

M 3 + k 2 M = M (a · b)2 − (a · a)(b · b) + a × b2 = 0.

A= √ (ba − abt )2 − a × b(bat − abt ) .

Note that for nonempty S, b S is an |S|-fold iterated finite difference

Proof. With ·, · denoting the inner product,

For which a does the sequence Sn (a) converge?

nφ φ = nφ (1 + 1/φ) = nφ + n − 1.