VOL. 97, NO.

2, APRIL 2024 225

Solutions

Find the maximum and minimum values April 2023

2166. Proposed by H. A. ShahAli, Tehran, Iran.
Let x1 , . . . , xn be nonnegative real numbers with x1 + · · · + xn = 1 and n ≥ 2. Deter-
mine the minimum and maximum values of the following function
x1 + x2 x2 + x3 xn + x1
+ + ··· + .
1 + x1 x2 1 + x2 x3 1 + xn x1
When do the extreme values occur?

Composite solution by Stan Dolan, Charmouth, UK, and the proposer.

Put
x1 + x2 x2 + x3 xn + x1
fn (x1 , . . . , xn ) = + + ··· +
1 + x1 x2 1 + x2 x3 1 + xn x1
with xi ≥ 0 and x1 + · · · + xn = 1. Since
fn (x1 , . . . , xn ) ≤ (x1 + x2 ) + (x2 + x3 ) + · · · + (xn + x1 ) = 2,

the maximum value is 2, which occurs if and only if x1 x2 = x2 x3 = · · · = xn x1 = 0.

To minimize fn , we consider three cases. If n = 2, then
2(x + y) 2
f2 (x, y) = = .
1 + xy 1 + xy
This is minimized, then xy is maximized, i.e., only when x = y = 1/2 and the mini-
mum value is thus 8/5.
Without loss of generality, when n = 3, we may consider f3 (x, y, z) with 0 ≤ x ≤
y ≤ z ≤ 1. Note that this forces 0 ≤ x ≤ 1/3. We claim that in this case
   
y+z y+z 1−x 1−x
f3 (x, y, z) ≥ f3 x, , = f3 x, , . (1)
2 2 2 2
After some computations, this is equivalent to

(1 − x)(y − z)2 (2 − 4x − x 2 − xyz(5 − 4x + 2x 2 )) ≥ 0. (2)

Since 0 ≤ x ≤ 1/3 and 0 ≤ xyz ≤ ((x + y + z)/3)3 = 1/27, we have
   2
1 1
49 > 104 + 29 ≥ 104x + 29x 2
3 3
⇒ 54 − 108x − 27x 2 > 5 − 4x + 2x 2
1
⇒ 2 − 4x − x 2 > (5 − 4x + 2x 2 ) ≥ xyz(5 − 4x + 2x 2 )
27
⇒ 2 − 4x − x 2 − xyz(5 − 4x + 2x 2 ) > 0.

Therefore (2) and equivalently (1) are true, and the equality holds if and only if y = z.
226 MATHEMATICS MAGAZINE
We have
   
y+z y+z 1−x 1−x
f3 x, , = f3 x, ,
2 2 2 2
9 x(1 − 3x)2
= + ,
5 5(2 − x)(5 − 2x + x 2 )
so the minimum value of f3 is 9/5, which occurs only when x = 0 or x = 1/3. This
corresponds to (x, y, z) = (1/3, 1/3, 1/3) or a permutation of (0, 1/2, 1/2).
We now consider the case when n ≥ 4. Each term of fn (x1 , . . . , xn ) involves neither
x1 nor x3 or involves just one of them (since n ≥ 4) and is a concave function of those
arguments. Therefore fn (x1 , . . . , xn ) is a concave function of x1 if x1 + x3 and all
other arguments are kept constant. Hence its minimum as x1 varies is attained at the
boundary values where either x1 or x3 is zero.
Without loss of generality, suppose that the minimum occurs when x1 = 0. Then
x2 + xn
fn (x1 , . . . , xn ) = fn−1 (x2 , . . . , xn ) + x2 + xn − ≥ fn−1 (x2 , . . . , xn )
1 + x2 xn
with x2 + · · ·xn = 1 and equality occurring if and only if x2 xn = 0. Repeating this
procedure leads to the case n = 3 with one of the arguments being zero. From the
results for the case n = 3, the three arguments must be a permutation of (1/2, 1/2, 0)
and the minimum value is 9/5. Hence the minimum value of fn is also 9/5 and this
occurs only when (x1 , . . . xn ) is a cyclic permutation of (1/2, 1/2, 0, . . . , 0).
Also solved by Paul Bracken, Prithwijit De (India), and Harris Kwong. There was one incom-
plete or incorrect solution.

The limit of an exponential product April 2023

2167. Proposed by Moubinool Omarjee, Lycée Henri IV, Paris, France.
Prove that

n  4  
i2 1 i 5 3ζ (3)
lim e n/2
e 1− 2 = π exp − + ,
n→∞
i=2
i 4 π2
∞
where ζ (3) = n=1 1/n3 .

Solution by Hongwei Chen, Christopher Newport University, Newport News, VA.

Let
n  4
i2 1 i
Pn := e n/2
e 1− 2 .
i=2
i

Taking logarithms yields

 
n  2  4
n n
1
ln Pn = + i + i ln 1 − 2 .
2 i=2 i=2
i

Applying the power series


∞
xk
ln(1 − x) = − for x ∈ [−1, 1),
k=1
k
VOL. 97, NO. 2, APRIL 2024 227
we have
n  2   i4
n n ∞
ln Pn = + i −
2 i=2 i=2 k=1
k i 2k

n  2  2 1 
n n ∞
1
= + i − i + + 2(k−2)
2 i=2 i=2
2 k=3 k i

1 
n ∞
1
= − 2(k−2)
(use k − 2 → k)
2 i=2 k=3 k i

1 
n ∞
1
= − .
2 i=2 k=1 (k + 2) i 2k

Hence
1 
∞ ∞
1
lim ln Pn = − .
n→∞ 2 i=2 k=1 (k + 2) i 2k

Exchanging the order of the summation gives


∞
1  1  1
∞ ∞
S := = (ζ (2k) − 1),
k=1
k + 2 i=2 i 2k k=1
k+2
∞
where ζ (2k) = n=1 1/n2k . Thus

∞ 1
S= 2 (ζ (2k) − 1)x 2k+3 dx
k=1 0

1  
2x 2
= 1 − πx cot(πx) − x 3 dx
0 1 − x2
1  
2
= 3 − πx cot(πx) − x 3 dx,
0 1 − x2

where we have used 2x 2 /(1 − x 2 ) = −2 + 2/(1 − x 2 ) and


∞
1
ζ (2k)x 2k = (1 − πx cot(πx))
k=1
2

(https://proofwiki.org/wiki/Riemann Zeta Function at Even Integers). Since x 3 /(1 −

x 2 ) = −x + 2x/(1 − x 2 ), we have
1 
7 2x
S= − πx cot(πx) +
4
dx
4 0 1 − x2
b 
7 2x
= − lim πx cot(πx) +
4
dx
4 b→1− 0 1 − x2
7   1
= − lim b4 ln(sin(πb)) − ln(1 − b2 ) + 4 x 3 ln(sin(πx)) dx
4 b→1− 0
228 MATHEMATICS MAGAZINE
1
7
= − ln(π/2) + 4 x 3 ln(sin(πx)) dx.
4 0

Using

∞
cos(2nx)
ln(sin x) = − ln 2 −
n=1
n

(https://proofwiki.org/wiki/Fourier Series/Logarithm of Sine of x over 0 to Pi), and

1
3
x 3 cos(2nπx) dx =
0 4n2 π 2
we find that
7 1 ∞
1 1
S = − ln(π/2) − 4 ln 2 x dx − 4
3
x 3 cos(2nπx) dx
4 0 n=1
n 0

3  1
∞
7
= − ln π − 4
4 4π 2 n=1 n3
7 3ζ (3)
= − ln π − .
4 π2
Hence
1 5 3ζ (3)
lim ln Pn = − S = − + ln π + .
n→∞ 2 4 π2
Taking exponent yields the desired result.
Also solved by Paul Bracken, Robert Calcaterra, Kee-Wai Lau (China), Raymond Mortini
(Luxembourg) & Rudolf Rupp (Germany), Albert Stadler (Switzerland), Seán Stewart (Saudi
Arabia), and the proposer. There was one incomplete or incorrect solution.

Polynomial identities for a recursive sequence April 2023

2168. Proposed by C. J. Hillar, San Francisco, CA.
Let α, r ∈ R and let an be the sequence

a0 = 0, a1 = α and an = ran−1 + an−2 for n > 1.

Prove that for each odd integer k, there are polynomials pk , qk ∈ R[x] such that for all
nonnegative integers n,
akn = pk (an ) for n even and akn = qk (an ) for n odd.

For example, for the Fibonacci sequence (where α = r = 1),

F3n = 5Fn3 + 3Fn for n even and F3n = 5Fn3 − 3Fn for n odd.
VOL. 97, NO. 2, APRIL 2024 229
Solution by Kyle Gatesman, Fairfax, VA.
If α = 0, then an = 0 for all whole numbers n, so for all k, the polynomials pk (z) =
qk (z) = z satisfy the desired criteria. Now suppose α = 0. The recurrence has the
characteristic polynomial λ2 − rλ − 1 = 0, yielding solutions
√ √
r + r2 + 4 r − r2 + 4
λ1 = and λ2 = .
2 2
The explicit form of the solution to the recurrence is then an = c1 λn1 + c2 λn2 . Taking
n = 0 gives c2 = −c1 , and taking n = 1 gives
α
c1 = √ ∈ R − {0}.
r +4
2

Since λ1 λ2 = −1,
   
1 n
an = c1 λ1 − −
n
.
λ1
We claim that for all odd integers k ≥ −1, there exist polynomials Pk and Qk with real
coefficients such that
   k    k
1 1 1 1
Pk x + =x +k
and Qk x − =x −
k
x x x x
for all x. We prove this claim by induction. The case k = −1 is satisfied
when P−1 (z) = z and Q−1 (z) = −z, and the case k = 1 is satisfied when
P1 (z) = Q1 (z) = z. Now suppose that for some odd k > 1, the polynomials Pk−2 ,
Pk−4 , Qk−2 , and Qk−4 exist. For s ∈ {−1, +1}, we have
 k−2
1 s 2
x k−2 + s x+ =
x x
 k  k−2  k−4
1 1 1
x +s
k
+ 2s x k−2
+s + x k−4
+s .
x x x

Therefore if we define polynomials Pk and Qk by

Pk (z) = Pk−2 (z)z2 − 2Pk−2 (z) − Pk−4 (z)

and

Qk (z) = Qk−2 (z)z2 + 2Qk−2 (z) − Qk−4 (z),

then
   k    k
1 1 1 1
Pk x + = xk + and Qk x − = xk − ,
x x x x
as desired.
For every odd k ≥ 1, let

pk (z) = c1 Pk (z/c1 ) and qk (z) = c1 Qk (z/c1 ).

230 MATHEMATICS MAGAZINE
This makes sense since c1 = 0. Note that pk and qk have real coefficients since Pk and
Qk do and c1 ∈ R. For odd n,
   k
1  n k 1
pk (an ) = c1 Pk λ1 + n = c1 λ1 +
n
λ1 λn1
 kn
1
= c1 λkn
1 − − = akn ,
λ1

and for even n,

   k
1  n k 1
qk (an ) = c1 Qk λ1 − n = c1 λ1 −
n
λ1 λn1
 
1 kn
= c1 λkn
1 − − = akn .
λ1

Also solved by Robert Calcaterra, Hongwei Chen, Eagle Problem Solvers (Georgia Southern
University), Dmitry Fleischman, Russell Gordon, Eugene A. Herman, Walther Janous (Austria),
Shin Hin Jimmy Pa (China), Harris Kwong, Northwestern University Math Problem Solving
Group, Angel Plaza (Spain), Randy K. Schwartz, Albert Stadler (Switzerland), and the proposer.

A functional equation that mirrors polynomial factorization April 2023

2169. Proposed by Marian Tetiva, National College “Gheorghe Roşca Codreanu,”
Bı̂rlad, Romania.
Let a, b, and c be distinct positive real numbers, which are not equal to 1, and let d be
one of them. Let f : R → R be a function satisfying the following conditions:
(i) f (f (f (x))) − (a + b + c)f (f (x)) + (ab + ac + bc)f (x) − abcx = 0 for all
x ∈ R.
(ii) f is continuous.
(iii) There exists an x0 ∈ (0, ∞) such that f (x0 ) = dx0 , and f (f (x0 )) = d 2 x0 .
Prove that f (x) = dx for all x ∈ [0, ∞).

Solution by Robert Calcaterra, University of Wisconsin-Platteville, Platteville, WI.

The notation f n will be used to denote f composed with itself n times. For example,
f 3 (x) = f (f (f (x))). Also, we may assume d = c due to the inherit symmetry in the
problem. Note that if f (x) = f (y), then condition (i) implies abcx = abcy and thus
x = y. Hence f is one-to-one. Since f is continuous, f must be strictly monotone on
R. Moreover,
f (f (x0 )) − f (x0 ) c2 x0 − cx0
= = c > 0.
f (x0 ) − x0 cx0 − x0
Thus f is strictly increasing. Assume f has an upper bound. Then limx→∞ f (x) = s,
the supremum of f , and by (i),
lim abcx = f 2 (s) − (a + b + c)f (s) + (ab + ac + bc)s,
x→∞
VOL. 97, NO. 2, APRIL 2024 231
which is finite. This is a contradiction, so f does not have an upper bound. Similarly,
f does not have a lower bound, so f is a bijection from R onto R. In particular, f −1
exists.
We claim that for all k ∈ Z and x ∈ R,

f k+3 (x) − (a + b + c)f k+2 (x) + (ab + ac + bc)f k+1 (x) − abcf k (x) = 0.

The hypotheses imply the claim is true for k = 0. Moreover, replacing the argument
x by f (x) shows we can increment k by 1 and replacing x by f −1 (x) shows we can
decrement k by 1. Therefore a bidirectional induction argument validates the claim.
We next claim that f k (x0 ) = ck x0 for all integers k. The claim is trivially true if
k = 0 and (iii) implies the claim is true for k = 1 and k = 2. Next assume

f k (x0 ) = ck x0 , f k+1 (x0 ) = ck+1 x0 , and f k+2 (x0 ) = ck+2 x0 ,

for some k ∈ Z. Then

f k+3 (x0 ) = (a + b + c)f k+2 (x0 ) − (ab + ac + bc)f k+1 (x0 ) + abcf k (x0 )
= (a + b + c)ck+2 x0 − (ab + ac + bc)ck+1 x0 + (abc)ck x0
= ck+3 x0
and

abcf k−1 (x0 ) = f k+2 (x0 ) − (a + b + c)f k+1 (x0 ) + (ab + ac + bc)f k (x0 )
= ck+2 x0 − (a + b + c)ck+1 x0 + (ab + ac + bc)ck x0
= abck x0 .

Hence f k−1 (x0 ) = ck−1 x0 and the claim is verified.

Finally, suppose t is an arbitrary positive real number. Denote the closed interval
with endpoints
  v (regardless of which is larger) by |u, v|. Choose L ∈ Z so
at u and
that t ∈ cL x0 , cL+1 x0  and let t0 = t/cL . For k ∈ Z, let

zk = f k (t0 ) − cf k−1 (t0 ).

Since
f k+2 (t0 ) − (a + b + c)f k+1 (t0 ) + (ab + ac + bc)f k (t0 ) − abcf k−1 (t0 ) = 0

implies that
 k+2     
f (t0 ) − cf k+1 (t0 ) − (a + b) f k+1 (t0 ) − cf k (t0 ) + ab f k (t0 ) − cf k−1 (t0 ) = 0,
we have zk+2 = (a + b)zk+1 − abzk . This recurrence relation has characteristic equa-
tion r 2 − (a + b)r + ab = 0 with roots r = a and r = b. Therefore there exist real
constants A and B such that
 k  k
zk a b
zk = Aa + Bb and hence k = A
k k
+B
c c c

for all k ∈ Z. Note that t0 ∈ |x0 , cx0 |. Since cf k−1 and f k are increasing functions,
   
cf k−1 (t0 ) ∈ cf k−1 (x0 ), cf k−1 (cx0 ) = ck x0 , ck+1 x0 
232 MATHEMATICS MAGAZINE
and
   
f k (t0 ) ∈ f k (x0 ), f k (cx0 ) = ck x0 , ck+1 x0  .
Therefore,
    z 
 k
|zk | = f k (t0 ) − cf k−1 (t0 ) ≤ ck+1 x0 − ck x0  , so  k  ≤ |cx0 − x0 | .
c
This implies that A(a/c)k + B(b/c)k is bounded as k ranges from −∞ to ∞. Since a,
b, and c are distinct positive constants, it follows that A = B = 0, zk = 0, and hence
f k (t0 ) = cf k−1 (t0 ) for every integer k. Since f 0 (t0 ) = t0 , a routine induction argument
yields f k (t0 ) = ck t0 for all k ∈ Z. Hence
   
f (t) = f cL t0 = f f L (t0 ) = f L+1 (t0 ) = cL+1 t0 = ct
for all t ∈ (0, ∞). In addition,
f (0) = lim f (t) = 0
t→0+

and the proof is complete.

Also solved by Anthony Kindness & Dylan Strohl and the proposer.

A partition of the k-element subsets of Zp April 2023

2170. Proposed by George Stoica, Saint John, NB, Canada.
For a fixed prime number p, let Ak be the set of all subsets of {0, . . . , p − 1} having k
elements, 1 ≤ k ≤ p − 1. Let
  k 

Ak,m = {i1 , . . . , ik } ∈ Ak  ij ≡ m (mod p)
j =1

with 0 ≤ m ≤ p − 1.
Prove that
 
1 p
|Ak,m | = .
p k

Solution by Harris Kwong, SUNY Fredonia, Fredonia, NY.  

It is obvious that Ak,s ∩ Ak,t = ∅ whenever
  s ≡ t (mod p). Thus
 A  
k,0 , . . . , A
 k,p−1
form a partition of Ak . Since |Ak | = pk , it suffices to show that Ak,s  = Ak,t  when-
ever s ≡ t (mod p). Consider {i1 , . . . , ik } ∈ Ak,s . Since t − s ≡ 0 (mod p), there
exists a positive integer k  such that kk  ≡ t − s (mod p). For j = 1, . . . , k, define
j ≡ ij + k  (mod p). Clearly |{1 . . . , k }| = k. We have
⎛ ⎞
k  k
j = ⎝ ij ⎠ + kk  ≡ s + (t − s) = t (mod p).
j =1 j =1
   
We deduce that {1 , . .. , k } ∈ A . This proves that Ak,s  ≤ Ak,t . In a similar man-
 k,t    
ner, we also have Ak,t  ≤ Ak,s . Therefore Ak,s  = Ak,t , which completes the proof.
Also solved by Robert Calcaterra, Dimitry Fleischman, Kyle Gatesman, Shing Hin Jimmy Pa
(China), Albert Stadler (Switzerland), and the proposer.
VOL. 97, NO. 2, APRIL 2024 233
Answers
Solutions to the Quickies from page 224.

A1139. Since
 
p−1 (p − 1)(p − 2) · · · (p − i) (−1)(−2) · · · (−i)
= ≡ ≡ (−1)i (mod p),
i i! i!
we have

p−1
1 
p−1
p−1 ≡ (−1)i ≡ 0 (mod p).
i=1 i i=1

A1140. Consider the generating function

∞ ν   
k m+k m+n+1
h(x) := (−1) xν .
ν=0 k=0
k ν − k

This is the product of the two series

 ∞  
m+j 1
f (x) := (−1)j x j =
j =0
j (1 + x)m+1

and
∞  
m+n+1 j
g(x) := x = (1 + x)m+n+1 .
j =0
j

Hence

h(x) = f (x)g(x) = (1 + x)n .


Thus the nth Taylor coefficient of h is nn = 1.
Solutions

Bitstrings which contain neither 11 nor 000 as substrings February 2023

2161. Proposed by Didier Pinchon, Toulouse, France, and George Stoica, Saint John,
NB, Canada.
Let xn denote the number of bitstrings of length n which contain neither 11 nor 000 as
substrings. Find a recursive formula for xn .

Solution by the Northwestern University Math Problem Solving Group, Northwestern

University, Evanston, IL.
The answer is xn = xn−2 + xn−3 , with initial conditions x1 = 2, x2 = 3, x3 = 4.
Assume n > 3 and divide the set of bitstrings with the desired property into two
subsets: those starting with 0 and those starting with 1. Let yn be the number in the
first set and zn the number in the second.
Bitstrings starting with 0 can be followed by a string starting with 1, or by another 0
and then a string starting with 1. Hence

yn = zn−1 + zn−2 .
VOL. 97, NO. 1, FEBRUARY 2024 83
Bitstrings starting with 1 can only be followed by a string starting with 0, hence
zn = yn−1 .
This, together with the first equation, implies yn = yn−2 + yn−3 and zn = zn−2 + zn−3 .
Since xn = yn + zn , we get the recursive formula:
xn = xn−2 + xn−3 .
The initial conditions can be obtained by enumerating the bitstrings with the desired
property. For n = 1 : 0, 1, for n = 2 : 00, 01, 10, and for n = 3 : 001, 010, 100, 101.
Hence x1 = 2, x2 = 3, and x3 = 4.
Also solved by Ricardo Bittencourt (Brazil), Robert Calcaterra, Kyle Calderhead, Eagle
Problem Solvers (Georgia Southern University), John Ferdinands, Eugene A. Herman, Aykhan
Ismayilov (Azerbaijan), Walther Janous (Austria), Kee Wai Lau (Hong Kong), Kent E. Morrison,
José Heber Nieto (Venezuela), Michelle Nogin, Shing Hin Jimmy Pa (China), Angel Plaza (Spain),
Rob Pratt, Edward Schmeichel, Randy Schwartz, Albert Stadler (Switzerland), Paul Stockmeyer,
and the proposers. There were two incomplete or incorrect solution.

An infinite series involving skew-harmonic numbers February 2023

2162. Proposed by Narendra Bhandari, Bajura, Nepal.
Prove that
∞  
H 2n 2n π2
= 4G + − 2π ln 2,
n=1
n(2n + 1)4n n 12

where
n
(−1)k−1
Hn =
k=1
k

is the nth skew-harmonic number and

∞
(−1)k−1
G=
k=1
(2k − 1)2

is Catalan’s constant.

Solution by the proposer.

It is well known that
∞  
x 2n+1 2n
= arcsin x − x.
n=1
4 (2n + 1) n
n

Dividing by x 2 and multiplying by ln(1 + x) on both sides, gives us

∞  
x 2n−1 ln(1 + x) 2n arcsin x ln(1 + x) ln(1 + x)
= − . (1)
n=1
4 (2n + 1)
n n x2 x

Now
 2n
1
1 − x 2n (−1)k−1
dx = = H 2n .
0 1+x k=1
k
84 MATHEMATICS MAGAZINE
Integrating by parts yields
 1  1
1 − x 2n
dx = 2n x 2n−1 ln(1 + x) dx.
0 1+x 0

Therefore,
 1
H 2n
x 2n−1 ln(1 + x) dx = .
0 2n
Integrating (1) from 0 to 1 and using the above, we obtain
∞    1  1
H2n 2n arcsin x ln(1 + x) ln(1 + x)
= dx − dx.
n=1
2n(2n + 1)4 n n 0 x 2
0 x

The second integral can be evaluated as

 1 ∞  ∞
ln(1 + x) (−1)k−1 1 k−1 (−1)k−1 π2
J = dx = x dx = = .
0 x k=1
k 0 k=1
k2 12

For the first integral I , substituting x = sin y and applying integration by parts (with
u = y ln (1 + sin y) and dv = cos y/ sin2 y), we get
 π  π
2 y ln(1 + sin y) cos y π 2 ln(1 + sin y)
I= dy = − ln(2) + dy
0 sin2 y 2 0 sin y
 π  π
2 y cos y 2 y
− dy + dy.
0 1 + sin y 0 tan y
We have
 
 π  ln 1 + 2t
2 ln(1 + sin y) y=2 arctan t
1
1+t 2 2tdt
A= dy =
0 sin y 0
2t
1+t 2
1 + t2
 ∞ ∞
1
2 ln(1 + t) − ln(1 + t 2 ) (−1)k−1 1 (−1)k−1
= dt = 2 −
0 t k=1
k2 2 k=1
k2
∞ k−1 2
3 (−1) π
= =
2 k=1
k2 8

and
 π  π
y cos y
2 IBP π 2
B= dy = ln(2) − ln(1 + sin y)dy
0 1 + sin y 2 0
 π  π
π 2 π 2 y sin y
= ln(2) − ln(1 + cos y)dy = ln(2) − dy
2 0 2 0 1 + cos y
 π  π 
π 2 y tan y π 2 y y
= ln(2) − dy = ln(2) − − dy
2 0 sec y + 1 2 0 sin y tan y
 π
π 2 y
= ln(2) − 2G + dy,
2 0 tan y
VOL. 97, NO. 1, FEBRUARY 2024 85
 π
1 2 y
where we have used the well-known fact that G = dy. Therefore,
2 0 sin y
 π
π 2 x π π2 π
I=− ln(2) + A − B + dx = − ln(2) + + 2G − ln(2)
2 0 tan x 2 8 2
π2
= 2G − π ln(2) + ,
8
and hence
∞  
H2n 2n π2
= 2 (I − J ) = 4G + − 2π ln(2).
n=1
n(2n + 1)4n n 12

Also solved by Paul Bracken, Hongwei Chen, Walther Janous, and Albert Stadler (Switzerland).
There was one incomplete or incorrect solution.

Sequences on which σ (respectively φ) is decreasing February 2023

2163. Proposed by Philippe Fondanaiche, Paris, France.
Recall that for a positive integer n, σ (n) denotes the sum of the positive divisors of
n, and φ(n) denotes the number of positive integers less than or equal to n that are
relatively prime to n. Show that the following hold.
(a) There are arbitrarily long sequences n1 < n2 < · · · < nk such that σ (n1 ) >
σ (n2 ) > · · · > σ (nk ).
(b) There are arbitrarily long sequences n1 < n2 < · · · < nk such that φ(n1 ) >
φ(n2 ) > · · · > φ(nk ).

Solution by José Heber Nieto, Universidad del Zulia, Maracaibo, Venezuela, (part (a))
and the proposer (part (b))
(a) We shall prove, by induction on k, that for any k > 1 there is a sequence
n1 < n2 < · · · < nk such that σ (n1 ) > σ (n2 ) > · · · > σ (nk ).
For k = 2, put n1 = 4, n2 = 5. Then n1 < n2 and σ (n1 ) = 7 > 6 = σ (n2 ). Now
assume that the result is true for k. Take a prime r > nk . It is well known that, given
 > 0, there exists an integer N such that, for any n > N, there is a prime p such that
n < p < (1 + )n. Take  = 1/r and a prime q > r such that q > N/r as well. Then
rq > N and there exists a prime p such that rq < p < rq(1 + 1/r), Put mi = rqni
for i = 1, . . . , k and mk+1 = nk p. Clearly
m1 < m2 < · · · < mk < mk+1 and σ (m1 ) > σ (m2 ) > · · · > σ (mk ).
Moreover
σ (mk+1 ) = σ (nk )(p + 1) < σ (nk )(rq + q + 1) < σ (nk )(r + 1)(q + 1) = σ (rqnk ) = σ (mk )
and we are done.

(b) Let us reason again by induction. For k = 2, put n1 = 5 and n2 = 6. Then n1 < n2
and φ(n1 ) = 4 > 2 = φ(n2 ). Assume that we have a sequence
n1 < n2 < · · · < nk such that φ(n1 ) > φ(n2 ) > · · · > φ(nk ).
86 MATHEMATICS MAGAZINE
We consider an integer y whose prime factors are strictly greater than all the prime fac-
tors of the ni , i = 1, 2, . . . , k. Since φ is multiplicative, φ(yni ) = φ(y)φ(ni ), hence
yn1 < yn2 < · · · < ynk and φ(yn1 ) > φ(yn2 ) > · · · > φ(ynk ).

We claim that there is an integer y having the properties above, which satisfies the
relation 4φ(yn1 ) < yn1 . To see this, note that if q1 , . . . , qr are the prime factors of y,
then
r  
φ(y)  1
= 1−
y i=1
qi

and this product tends to zero as r goes to infinity. In particular, we can find a y such
that φ(y)/y < n1 /(4φ(n1 )) and the claim follows.
Since, for such a y, 4φ(yn1 ) < yn1 , there is an integer m such that φ(yn1 ) < 2m <
2m+1 < yn1 . Taking z = 2m+1 , we have z < yn1 and φ(z) = 2m > φ(yn1 ).
Also solved by Robert Calcaterra (part (b)). There was one incomplete or incorrect solution.

A locus comprising (almost) two lines February 2023

2164. Proposed by Dixon J. Jones, Coralville, IA.
In the plane of a triangle ABC, let P1 and P2 be fixed points such that P1 P2 is not
perpendicular to AB, BC, or CA. Find the set of points P3 for which

L 1 L2 M 1 M 2 N 1 N 2
· · = +1 ,
L2 L3 M 2 M 3 N 2 N 3

where Li , Mi+1 , Ni+2 are the feet of the perpendiculars from Pi to BC, CA, AB,
respectively (indices taken modulo 3), and the quantities are directed distances.

Solution by the proposer.

The solution set comprises a line ∗1 from which two points have been removed, along
with a complete line 2 .
First, let ∗1 be the line through P1 and P2 , but with those two points excluded. If
P3 lies on ∗1 , then P1 L1 , P2 L2 , and P3 L3 are parallel, distinct, and of nonzero length.
Therefore,
L 1 L2 P1 P2
= ,
L2 L3 P2 P3
and similarly
M 1 M2 P3 P1 N1 N2 P2 P3
= , and = .
M2 M3 P1 P2 N2 N3 P3 P1
Multiplying these equalities and canceling like terms yields the claim.
Next, let P1 M2 meet P2 L2 at Q. Let 2 be the perpendicular to AB through Q. We
assert that if P3 lies on 2 , the claim follows. We have
L1 L2 = QP1 sin ∠P2 QP1 .
VOL. 97, NO. 1, FEBRUARY 2024 87

Similarly

L2 L3 = QP3 sin ∠P3 QP2 ,

M1 M2 = QP3 sin ∠P3 QP1 , M2 M3 = QP2 sin ∠P2 QP1 ,
N1 N2 = QP2 sin ∠P3 QP1 , N2 N3 = QP1 sin ∠P3 QP2 .
Again, forming quotients and canceling produces the claim.
We now show that if the claim holds, then P3 necessarily lies on ∗1 or 2 . Let P3 N2
meet ∗1 at P3 , and let the feet of the perpendiculars from P3 to BC and CA be L3 and
M1 , respectively. Let P3 L3 meet 2 at P3 , and let M1 be the foot of the perpendicular
from P3 to CA. Set
L1 L 2 M 1 M 2 N 1 N 2
x= · · . (1)
L2 L3 M 2 M 3 N 2 N 3
Since P3 lies on ∗1 , the claim applies; that is,
L1 L2 M1 M2 N1 N2
· · = +1 ,
L2 L3 M 2 M 3 N 2 N 3
which can be rearranged as
L2 L3 L1 L2 N1 N2
= · . (2)
M1 M2 M2 M3 N2 N3
Substituting (2) in (1) yields
M 1 M2 M1 M2
x = . (3)
L2 L3 L2 L3
88 MATHEMATICS MAGAZINE
Since M1 M2 = M1 M1 + M1 M2 , (3) can be rewritten as
L2 L3 M1 M1
x = +1. (4)
L2 L3 M1 M2
We have
L 2 L3 QP3 1
= = ; (5)
L2 L3 QP3 + P3 P3 P3 P3
1+
QP3
furthermore, it is clear that
P3 P3 M1 M1
= . (6)
QP3 M2 M1
Combining (4)–(6), we obtain
  
M 1 M1 M1 M 1
x= +1 +1
M1 M2 M2 M1
 
M 1 M1 + M1 M2 + M2 M1
= M1 M1 + 1. (7)
M 1 M 2 · M2 M 1
From (7) it follows that x = 1 if, and only if,
 
M 1 M 1 + M1 M 2 + M 2 M 1
M1 M1 = 0 or M1 M1 = 0.
M 1 M 2 · M2 M 1
If M1 M1 = 0, then P3 must lie on ∗1 . On the other hand,
0 = M1 M 1 + M 1 M 2 + M 2 M 1 = M 1 M 1
implies that P3 must lie on 2 . Thus, the only points P3 for which the claim holds must
lie on ∗1 or 2 .
Also solved by Volkhard Schindler (Germany).

Maximize α(G)χ(G) February 2023

2165. Proposed by Zion Hefty (student), Grinell College, Grinnell, IA, and Peter John-
son, Auburn University, Auburn, AL.
Let G be a graph. We will denote the vertex set of G by V (G). The independence
number of G, denoted α(G), is the cardinality of the largest subset of V (G) such that
no two vertices of that subset are connected by an edge. The chromatic number of G,
denoted χ(G), is the smallest number of colors needed to color each vertex of V (G)
so that no two vertices with the same color are connected by an edge.
If we let
g(n) = min ({α(G)χ(G)||V (G)| = n}) ,
it is well known that g(n) = n. This can be realized, for example, if G is the complete
graph on n vertices.
Let
f (n) = max ({α(G)χ(G)||V (G)| = n}) .
Determine f (n).
VOL. 97, NO. 1, FEBRUARY 2024 89
Note: This problem and its solution arose during the 2022 Research Experience
for Undergraduates in Algebra and Discrete Mathematics at Auburn University. This
research was supported by NSF(DMS) grant no. 1950563.

Solution by Edward Schmeichel, San José State University, San José, CA.
We claim that
 
n+1 2
f (n) = .
2

Note first that for any graph G with n vertices,

α(G) + χ(G) ≤ n + 1.
To prove this, let I be any maximum independent subset of V (G) (so |I | = α(G)). If
we color the vertices of I with a first color, the remaining vertices in V (G) − I can be
colored using at most
|V (G)| − |I | = n − α(G)

additional colors. Thus χ(G) ≤ 1 + (n − α(G)).

By the AM-GM inequality, we have
   
α(G) + χ(G) 2 n+1 2
α(G)χ(G) ≤ ≤
2 2
for any graph G with n vertices. Thus
 2
n+1
f (n) ≤ .
2

For the reverse inequality, consider the n-vertex graph

Gn = K n/2 ∪ K n/2 ,

where Km is the complete graph on m vertices and K m is the graph on m vertices

having no edges. We have
n n
α(Gn ) = + 1, χ(Gn ) = ,
2 2
and therefore
 n   n   
n+1 2
f (n) ≥ α(Gn )χ(Gn ) = +1 = .
2 2 2

The claim now follows.

Also solved by José Nieto and the proposers.
VOL. 97, NO. 1, FEBRUARY 2024 83
Bitstrings starting with 1 can only be followed by a string starting with 0, hence
zn = yn−1 .
This, together with the first equation, implies yn = yn−2 + yn−3 and zn = zn−2 + zn−3 .
Since xn = yn + zn , we get the recursive formula:
xn = xn−2 + xn−3 .
The initial conditions can be obtained by enumerating the bitstrings with the desired
property. For n = 1 : 0, 1, for n = 2 : 00, 01, 10, and for n = 3 : 001, 010, 100, 101.
Hence x1 = 2, x2 = 3, and x3 = 4.
Also solved by Ricardo Bittencourt (Brazil), Robert Calcaterra, Kyle Calderhead, Eagle
Problem Solvers (Georgia Southern University), John Ferdinands, Eugene A. Herman, Aykhan
Ismayilov (Azerbaijan), Walther Janous (Austria), Kee Wai Lau (Hong Kong), Kent E. Morrison,
José Heber Nieto (Venezuela), Michelle Nogin, Shing Hin Jimmy Pa (China), Angel Plaza (Spain),
Rob Pratt, Edward Schmeichel, Randy Schwartz, Albert Stadler (Switzerland), Paul Stockmeyer,
and the proposers. There were two incomplete or incorrect solution.

An infinite series involving skew-harmonic numbers February 2023

2162. Proposed by Narendra Bhandari, Bajura, Nepal.
Prove that
∞  
H 2n 2n π2
= 4G + − 2π ln 2,
n=1
n(2n + 1)4n n 12

where
n
(−1)k−1
Hn =
k=1
k

is the nth skew-harmonic number and

∞
(−1)k−1
G=
k=1
(2k − 1)2

is Catalan’s constant.

Solution by the proposer.

It is well known that
∞  
x 2n+1 2n
= arcsin x − x.
n=1
4 (2n + 1) n
n

Dividing by x 2 and multiplying by ln(1 + x) on both sides, gives us

∞  
x 2n−1 ln(1 + x) 2n arcsin x ln(1 + x) ln(1 + x)
= − . (1)
n=1
4 (2n + 1)
n n x2 x

Now
 2n
1
1 − x 2n (−1)k−1
dx = = H 2n .
0 1+x k=1
k
84 MATHEMATICS MAGAZINE
Integrating by parts yields
 1  1
1 − x 2n
dx = 2n x 2n−1 ln(1 + x) dx.
0 1+x 0

Therefore,
 1
H 2n
x 2n−1 ln(1 + x) dx = .
0 2n
Integrating (1) from 0 to 1 and using the above, we obtain
∞    1  1
H2n 2n arcsin x ln(1 + x) ln(1 + x)
= dx − dx.
n=1
2n(2n + 1)4 n n 0 x 2
0 x

The second integral can be evaluated as

 1 ∞  ∞
ln(1 + x) (−1)k−1 1 k−1 (−1)k−1 π2
J = dx = x dx = = .
0 x k=1
k 0 k=1
k2 12

For the first integral I , substituting x = sin y and applying integration by parts (with
u = y ln (1 + sin y) and dv = cos y/ sin2 y), we get
 π  π
2 y ln(1 + sin y) cos y π 2 ln(1 + sin y)
I= dy = − ln(2) + dy
0 sin2 y 2 0 sin y
 π  π
2 y cos y 2 y
− dy + dy.
0 1 + sin y 0 tan y
We have
 
 π  ln 1 + 2t
2 ln(1 + sin y) y=2 arctan t
1
1+t 2 2tdt
A= dy =
0 sin y 0
2t
1+t 2
1 + t2
 ∞ ∞
1
2 ln(1 + t) − ln(1 + t 2 ) (−1)k−1 1 (−1)k−1
= dt = 2 −
0 t k=1
k2 2 k=1
k2
∞ k−1 2
3 (−1) π
= =
2 k=1
k2 8

and
 π  π
y cos y
2 IBP π 2
B= dy = ln(2) − ln(1 + sin y)dy
0 1 + sin y 2 0
 π  π
π 2 π 2 y sin y
= ln(2) − ln(1 + cos y)dy = ln(2) − dy
2 0 2 0 1 + cos y
 π  π 
π 2 y tan y π 2 y y
= ln(2) − dy = ln(2) − − dy
2 0 sec y + 1 2 0 sin y tan y
 π
π 2 y
= ln(2) − 2G + dy,
2 0 tan y
VOL. 97, NO. 1, FEBRUARY 2024 85
 π
1 2 y
where we have used the well-known fact that G = dy. Therefore,
2 0 sin y
 π
π 2 x π π2 π
I=− ln(2) + A − B + dx = − ln(2) + + 2G − ln(2)
2 0 tan x 2 8 2
π2
= 2G − π ln(2) + ,
8
and hence
∞  
H2n 2n π2
= 2 (I − J ) = 4G + − 2π ln(2).
n=1
n(2n + 1)4n n 12

Also solved by Paul Bracken, Hongwei Chen, Walther Janous, and Albert Stadler (Switzerland).
There was one incomplete or incorrect solution.

Sequences on which σ (respectively φ) is decreasing February 2023

2163. Proposed by Philippe Fondanaiche, Paris, France.
Recall that for a positive integer n, σ (n) denotes the sum of the positive divisors of
n, and φ(n) denotes the number of positive integers less than or equal to n that are
relatively prime to n. Show that the following hold.
(a) There are arbitrarily long sequences n1 < n2 < · · · < nk such that σ (n1 ) >
σ (n2 ) > · · · > σ (nk ).
(b) There are arbitrarily long sequences n1 < n2 < · · · < nk such that φ(n1 ) >
φ(n2 ) > · · · > φ(nk ).

Solution by José Heber Nieto, Universidad del Zulia, Maracaibo, Venezuela, (part (a))
and the proposer (part (b))
(a) We shall prove, by induction on k, that for any k > 1 there is a sequence
n1 < n2 < · · · < nk such that σ (n1 ) > σ (n2 ) > · · · > σ (nk ).
For k = 2, put n1 = 4, n2 = 5. Then n1 < n2 and σ (n1 ) = 7 > 6 = σ (n2 ). Now
assume that the result is true for k. Take a prime r > nk . It is well known that, given
 > 0, there exists an integer N such that, for any n > N, there is a prime p such that
n < p < (1 + )n. Take  = 1/r and a prime q > r such that q > N/r as well. Then
rq > N and there exists a prime p such that rq < p < rq(1 + 1/r), Put mi = rqni
for i = 1, . . . , k and mk+1 = nk p. Clearly
m1 < m2 < · · · < mk < mk+1 and σ (m1 ) > σ (m2 ) > · · · > σ (mk ).
Moreover
σ (mk+1 ) = σ (nk )(p + 1) < σ (nk )(rq + q + 1) < σ (nk )(r + 1)(q + 1) = σ (rqnk ) = σ (mk )
and we are done.

(b) Let us reason again by induction. For k = 2, put n1 = 5 and n2 = 6. Then n1 < n2
and φ(n1 ) = 4 > 2 = φ(n2 ). Assume that we have a sequence
n1 < n2 < · · · < nk such that φ(n1 ) > φ(n2 ) > · · · > φ(nk ).
86 MATHEMATICS MAGAZINE
We consider an integer y whose prime factors are strictly greater than all the prime fac-
tors of the ni , i = 1, 2, . . . , k. Since φ is multiplicative, φ(yni ) = φ(y)φ(ni ), hence
yn1 < yn2 < · · · < ynk and φ(yn1 ) > φ(yn2 ) > · · · > φ(ynk ).

We claim that there is an integer y having the properties above, which satisfies the
relation 4φ(yn1 ) < yn1 . To see this, note that if q1 , . . . , qr are the prime factors of y,
then
r  
φ(y)  1
= 1−
y i=1
qi

and this product tends to zero as r goes to infinity. In particular, we can find a y such
that φ(y)/y < n1 /(4φ(n1 )) and the claim follows.
Since, for such a y, 4φ(yn1 ) < yn1 , there is an integer m such that φ(yn1 ) < 2m <
2m+1 < yn1 . Taking z = 2m+1 , we have z < yn1 and φ(z) = 2m > φ(yn1 ).
Also solved by Robert Calcaterra (part (b)). There was one incomplete or incorrect solution.

A locus comprising (almost) two lines February 2023

2164. Proposed by Dixon J. Jones, Coralville, IA.
In the plane of a triangle ABC, let P1 and P2 be fixed points such that P1 P2 is not
perpendicular to AB, BC, or CA. Find the set of points P3 for which

L 1 L2 M 1 M 2 N 1 N 2
· · = +1 ,
L2 L3 M 2 M 3 N 2 N 3

where Li , Mi+1 , Ni+2 are the feet of the perpendiculars from Pi to BC, CA, AB,
respectively (indices taken modulo 3), and the quantities are directed distances.

Solution by the proposer.

The solution set comprises a line ∗1 from which two points have been removed, along
with a complete line 2 .
First, let ∗1 be the line through P1 and P2 , but with those two points excluded. If
P3 lies on ∗1 , then P1 L1 , P2 L2 , and P3 L3 are parallel, distinct, and of nonzero length.
Therefore,
L 1 L2 P1 P2
= ,
L2 L3 P2 P3
and similarly
M 1 M2 P3 P1 N1 N2 P2 P3
= , and = .
M2 M3 P1 P2 N2 N3 P3 P1
Multiplying these equalities and canceling like terms yields the claim.
Next, let P1 M2 meet P2 L2 at Q. Let 2 be the perpendicular to AB through Q. We
assert that if P3 lies on 2 , the claim follows. We have
L1 L2 = QP1 sin ∠P2 QP1 .
VOL. 97, NO. 1, FEBRUARY 2024 87

Similarly

L2 L3 = QP3 sin ∠P3 QP2 ,

M1 M2 = QP3 sin ∠P3 QP1 , M2 M3 = QP2 sin ∠P2 QP1 ,
N1 N2 = QP2 sin ∠P3 QP1 , N2 N3 = QP1 sin ∠P3 QP2 .
Again, forming quotients and canceling produces the claim.
We now show that if the claim holds, then P3 necessarily lies on ∗1 or 2 . Let P3 N2
meet ∗1 at P3 , and let the feet of the perpendiculars from P3 to BC and CA be L3 and
M1 , respectively. Let P3 L3 meet 2 at P3 , and let M1 be the foot of the perpendicular
from P3 to CA. Set
L1 L 2 M 1 M 2 N 1 N 2
x= · · . (1)
L2 L3 M 2 M 3 N 2 N 3
Since P3 lies on ∗1 , the claim applies; that is,
L1 L2 M1 M2 N1 N2
· · = +1 ,
L2 L3 M 2 M 3 N 2 N 3
which can be rearranged as
L2 L3 L1 L2 N1 N2
= · . (2)
M1 M2 M2 M3 N2 N3
Substituting (2) in (1) yields
M 1 M2 M1 M2
x = . (3)
L2 L3 L2 L3
88 MATHEMATICS MAGAZINE
Since M1 M2 = M1 M1 + M1 M2 , (3) can be rewritten as
L2 L3 M1 M1
x = +1. (4)
L2 L3 M1 M2
We have
L 2 L3 QP3 1
= = ; (5)
L2 L3 QP3 + P3 P3 P3 P3
1+
QP3
furthermore, it is clear that
P3 P3 M1 M1
= . (6)
QP3 M2 M1
Combining (4)–(6), we obtain
  
M 1 M1 M1 M 1
x= +1 +1
M1 M2 M2 M1
 
M 1 M1 + M1 M2 + M2 M1
= M1 M1 + 1. (7)
M 1 M 2 · M2 M 1
From (7) it follows that x = 1 if, and only if,
 
M 1 M 1 + M1 M 2 + M 2 M 1
M1 M1 = 0 or M1 M1 = 0.
M 1 M 2 · M2 M 1
If M1 M1 = 0, then P3 must lie on ∗1 . On the other hand,
0 = M1 M 1 + M 1 M 2 + M 2 M 1 = M 1 M 1
implies that P3 must lie on 2 . Thus, the only points P3 for which the claim holds must
lie on ∗1 or 2 .
Also solved by Volkhard Schindler (Germany).

Maximize α(G)χ(G) February 2023

2165. Proposed by Zion Hefty (student), Grinell College, Grinnell, IA, and Peter John-
son, Auburn University, Auburn, AL.
Let G be a graph. We will denote the vertex set of G by V (G). The independence
number of G, denoted α(G), is the cardinality of the largest subset of V (G) such that
no two vertices of that subset are connected by an edge. The chromatic number of G,
denoted χ(G), is the smallest number of colors needed to color each vertex of V (G)
so that no two vertices with the same color are connected by an edge.
If we let
g(n) = min ({α(G)χ(G)||V (G)| = n}) ,
it is well known that g(n) = n. This can be realized, for example, if G is the complete
graph on n vertices.
Let
f (n) = max ({α(G)χ(G)||V (G)| = n}) .
Determine f (n).
VOL. 97, NO. 1, FEBRUARY 2024 89
Note: This problem and its solution arose during the 2022 Research Experience
for Undergraduates in Algebra and Discrete Mathematics at Auburn University. This
research was supported by NSF(DMS) grant no. 1950563.

Solution by Edward Schmeichel, San José State University, San José, CA.
We claim that
 
n+1 2
f (n) = .
2

Note first that for any graph G with n vertices,

α(G) + χ(G) ≤ n + 1.
To prove this, let I be any maximum independent subset of V (G) (so |I | = α(G)). If
we color the vertices of I with a first color, the remaining vertices in V (G) − I can be
colored using at most
|V (G)| − |I | = n − α(G)

additional colors. Thus χ(G) ≤ 1 + (n − α(G)).

By the AM-GM inequality, we have
   
α(G) + χ(G) 2 n+1 2
α(G)χ(G) ≤ ≤
2 2
for any graph G with n vertices. Thus
 2
n+1
f (n) ≤ .
2

For the reverse inequality, consider the n-vertex graph

Gn = K n/2 ∪ K n/2 ,

where Km is the complete graph on m vertices and K m is the graph on m vertices

having no edges. We have
n n
α(Gn ) = + 1, χ(Gn ) = ,
2 2
and therefore
 n   n   
n+1 2
f (n) ≥ α(Gn )χ(Gn ) = +1 = .
2 2 2

The claim now follows.

Also solved by José Nieto and the proposers.
568 MATHEMATICS MAGAZINE
Solutions

The quadrilateral must be a square December 2022

2156. Proposed by Cezar Lupu, Tsinghua University, Beijing, China.
Let ABCD be a convex quadrilateral in the plane with vertices having rational coor-
dinates. Let P be a point in its interior having rational coordinates such that
m∠P AB = m∠P BC = m∠P CD = m∠P DA = qπ, with q ∈ Q.
Show that ABCD is a square. Give an example to show that the condition that q ∈ Q
cannot be dropped.

Solution by Victor Pambuccian, Arizona State University, Phoenix, AZ.

Given that the coordinates
√ of A, B, and P are rational, the distances AB, AP , and BP
are all of the form r, where r ∈ Q. Therefore,
AP 2 + AB 2 − BP 2 √
cos qπ = cos ∠P AB = = s,
2 · AB · AP
where q, s ∈ Q.
Although it is a well-known result, we will show that
√ 1 1 3
cos qπ = s with q, s ∈ Q if and only if s = 0, , , , or 1.
4 2 4
To see this, note that if q = m/n with m, n ∈ Z, then exp(qπi) and exp(−qπi) √ are
algebraic integers, being roots of x 2n = 1. Therefore, their sum, 2 cos qπ = 2 s, is
 √ 2
also an algebraic integer. But then, 4s = 2 s is a rational algebraic integer, hence
an integer, and the candidates given are the only possibilities. To finish, we note that
s = 0, 1/4, 1/2, 3/4, 1 give q = 1/2, 1/3, 1/4, 1/6, 0, respectively.
The case s = 1 is impossible since P would have to lie on all sides of the polygon.
If s = 0, then ∠P AB would be a right angle. But this would make all of the angles of
ABCD obtuse.
To deal with the cases s = 1/4 and s = 3/4, we need a well-known result: no
triangle in the plane whose coordinates are rational can have an angle with measure
π/3 or π/6. To see this, we let the vertices of the triangle be
A = (x1 , y1 ), B = (x2 , y2 ), and C = (x3 , y3 ),
where xi , yi ∈ Q. If the side lengths of the triangle are a, b, and c, then a 2 , b2 , c2 ∈ Q.
Let θ denote the angle between the sides of length a and length b. If K is the area of
the triangle, then
  
1 1  x1 − x3 y1 − y3 
ab sin θ = K = det ∈ Q.
2 2 x2 − x3 y2 − y3 

By the law of cosines, c2 = a 2 + b2 − 2ab cos θ. Solving the displayed equation for
ab and substituting, we have
a 2 + b2 − c2
c2 = a 2 + b2 − 4K cot θ or cot θ = ∈ Q.
4K
√ √
But cot(π/3) = 1/ 3 and cot(π/6) = 3, which gives a contradiction.
VOL. 96, NO. 5, DECEMBER 2023 569
The result above eliminates the cases s = 1/4 and s = 3/4, so we are left with
s = 1/2 and q = 1/4. We will invoke the following result of Dmitriev and Dynkin
(see: Besenyei, Á. (2015). The Brocard angle and a geometrical gem from Dmitriev
and Dynkin. Amer. Math. Monthly 122(5): 495–499. https://doi.org/10.4169/amer.
math.monthly.122.5.495.)

Let P be an arbitrary point in the interior of a convex n-gon A1 A2 . . . An and

denote An+1 = A1 . Then

π π
min ∠P Ak Ak+1 ≤ − .
k=1,...,n 2 n
Equality occurs if and only if A1 A2 . . . An is a regular n-gon and P is its
center.

Applying this result with n = 4 and q = 1/4 gives the conclusion

√ we seek.
Here is a non-square parallelogram with q = arccos(3/ 10)/π  ∈ Q:
A = (0, 0), B = (2, 0), C = (6, 2), D = (4, 2), and P = (3, 1).

Also solved by Robert Calcaterra and the proposer.

An application of the Rayleigh-Beatty theorem December 2022

2157. Proposed by Philippe Fondanaiche, Paris, France.
Consider two sequences. One is the number of digits in the base 2 representation of
10k , k = 1, 2, . . ., and the other is the number of digits in the base 5 representation of
10k , k = 1, 2, . . .. Show that every integer greater than 1 appears in exactly one of the
two sequences. Which sequence contains 2023?

Solution by Brian D. Beasley, Presbyterian College, Clinton, SC.

We denote the first sequence by ak and the second sequence by bk . Then for each
positive integer k,
ak = log2 (10k ) + 1 = k(log2 10) + 1
and
bk = log5 (10k ) + 1 = k(log5 10) + 1.
Let c = log2 10 and d = log5 10. Then c and d are irrational with
1 1
+ = log10 2 + log10 5 = 1.
c d
Given an integer n ≥ 2, we have
n−1 n
ak = n ⇐⇒ n < kc + 1 < n + 1 ⇐⇒ <k<
c c
and
n−1 n
bk = n ⇐⇒ n < kd + 1 < n + 1 ⇐⇒ <k< ,
d d
 570 MATHEMATICS MAGAZINE
since each of (n − 1)/c, n/c, (n − 1)/d, and n/d is irrational.
(i) If there are positive integers i and j with n = ai = bj , then
n−1 n n−1 n
< i < and <j < .
c c d d
Adding the inequalities yields the contradiction n − 1 < i + j < n.
(ii) If neither sequence contains n, then there are positive integers p and q with
n−1 n n−1 n
p−1< < < p and q − 1 < < < q.
c c d d
Once again, adding the inequalities produces a contradiction, namely
p + q − 2 < n − 1 < n < p + q.
(iii) For n = 2023, we have
2022 2023
≈ 608.683 and ≈ 608.984,
c c
while
2022 2023
≈ 1413.317 and ≈ 1414.016.
d d
Thus b1414 = 2023.
Editor’s Note. Several solvers noted that the result follows from the Rayleigh-Beatty
theorem. Let
Nα = { nα : n ∈ N}.
If α and β are real numbers, then
Nα ∪ Nβ = N and Nα ∩ Nβ = ∅
if and only if α and β are irrational and
1 1
+ = 1.
α β

Also solved by Ulrich Abel & Vitaliy Kushnirevych (Germany), Jacob Boswell, Robert Cal-
caterra, Eagle Problem Solvers, Dmitry Fleischman, Walther Janous (Austria), José Heber Nieto
(Venezuela), Michelle Nogin, Mariam Obeidallah, Shing Hin Jimmy Pa (China), Celia Schacht,
Edward Schmeichel, Randy K. Schwartz, Paul K. Stockmeyer, Ertugrul Tarhan, Enrique Treviño,
Edward White & Roberta White, and the proposer.

Adjacent numbers sum to a perfect square December 2022

2158. Proposed by the Missouri State University Problem Solving Group, Missouri
State University, Springfield, MO.

(a) Arrange the integers from 1 to 15 (inclusive) in a row so that the sum of any two
adjacent numbers is a perfect square.
(b) Find the smallest positive integer n such that the integers from 1 to n can be
arranged in a circle so that the sum of any two adjacent numbers is a perfect square.
Justify your answer.
 VOL. 96, NO. 5, DECEMBER 2023 571
Solution by Jacob Boswell and Chip Curtis, Missouri Southern State University,
Joplin, MO.
Let n denote the graph whose vertices are the integers from 1 to n with an edge
between two vertices if their sum is a perfect square. Part (a) asks for a Hamiltonian
path in 15 and part (b) askes for the smallest n such that n has a Hamiltonian cycle.
(a) Since 8 and 9 are the only vertices of degree 1, a Hamiltonian path must start at one
of these and end at the other. It is not difficult to show that the following solution is
unique (up to reversal).
8 – 1 – 15 – 10 – 6 – 3 – 13 – 12 – 4 – 5 – 11 – 14 – 2 – 7 – 9.
(b) We claim that the smallest possible integer with the given property is n = 32. To
prove this, we first exhibit a solution for n = 32.
22 – 27 – 9 – 16 – 20 – 29 – 7 – 18 – 31 – 5 – 11 – 25 – 24 – 12
– 13 – 3 – 6 – 30 – 19 – 17 – 32 – 4 – 21 – 28 – 8 – 1 – 15
– 10 – 26 – 23 – 2 – 14 – 22
We next show that no n with 3 ≤ n ≤ 30 works. (We exclude n = 1 and n = 2
from consideration.) A necessary condition for n to have a Hamiltonian cycle is
that every vertex must have degree at least two.
• For 3 ≤ n ≤ 4, the only neighbor of 1 is 3.
• For 5 ≤ n ≤ 8, the only neighbor of 4 is 5.
• For 9 ≤ n ≤ 15, the only neighbor of 9 is 7.
• For 16 ≤ n ≤ 19, the only neighbor of 16 is 9.
• For 20 ≤ n ≤ 30, the only neighbor of 18 is 7.
Finally, we show that n = 31 does not work. Two edges emanating from a vertex
of degree two must be part of any Hamiltonian cycle. This gives the following
fragments.
22 – 27 – 9 – 16 – 20 – 29 – 7 – 18 – 31 – 5,

21 – 28 – 8 – 17 – 19 – 30 – 6,

10 – 26 – 23, and 11 – 25 – 24
Consider a vertex of degree three. Suppose one of its neighbors already has two
edges of the Hamiltonian cycle we are building emanating from it. Then the Hamil-
tonian cycle must contain the other two edges emanating from the given vertex.
This gives the fragments
14 – 2 – 23, 3 – 1 – 24, (and subsequently) 10 – 15 – 24.
Finally, the only possible remaining edge emanating
from 6 is 6 – 3, from 22 is 22 – 14, and from 11 is 11 – 5.
However, this gives a cycle that does not contain all of the vertices, yielding a
contradiction.
We note that this type of argument shows that there are two solutions when
n = 32 (up to reversal).
572 MATHEMATICS MAGAZINE
Also solved by Carl Axness (Spain), Brett Chiodo, Keon Cruz, Eagle Problem Solvers, Dmitry
Fleischman (part (a)), Evan Grahn, Shannon Heinig, Alyssa Janowski, Kelly McLenithan &
Stephen Mortenson, José Heber Nieto (Venezuela), Michelle Nogin, Pittsburg State University
Problem Solving Group, Rob Pratt, Zaur Rajabov (Azerbaijan), Mary Reil (part (a)), Edward
Schmeichel, Randy Schwartz, Paul Stockmeyer, and the proposers.

The winner is the one whose roll occurs first December 2022
2159. Proposed by George Stoica, Saint John, NB, Canada.
Two players, A and B, alternately throw a pair of dice with A going first. Let a, b ∈
{2, 3, . . . , 12} be fixed. Player A wins by having a roll worth a points before player B
has a roll worth b points. Otherwise, player B wins.
What is the probability that player A wins?

Solution by Michelle Nogin (student), Clovis North High School, Fresno, CA.
Let f (n) be the probability of having a roll worth n points. Observe that there are
36 possible outcomes when rolling two dice. Since there is only one way to get the
value 2 (1 + 1) and only one way to get the value 12 (6 + 6), f (2) = f (12) = 1/36.
Similarly, since there are two ways to get the value 3 (1 + 2 and 2 + 1) and two ways
to get the value 11 (5 + 6 and 6 + 5), f (3) = f (11) = 2/36, and so forth until there
are six ways to get the value 7, so f (7) = 6/36. From this, we can write an explicit
formula for f (n):
6 − |7 − n|
f (n) = .
36
The probability of player A winning on the first move is the probability of player A
having a roll worth a points on their first move, that is, f (a). The probability of player
A winning on the second move is the probability of player A and player B not having
rolls worth a and b points, respectively, on their first moves times the probability of
player A having a roll worth a points on their second move, that is,
(1 − f (a))(1 − f (b))f (a).

Similarly, the probability of player A winning on the third move is

((1 − f (a))(1 − f (b)))2 f (a),

and so forth. Thus, the probability of player A winning is the sum of the geometric
series
f (a) + (1 − f (a))(1 − f (b))f (a) + ((1 − f (a))(1 − f (b)))2 f (a) + · · ·

which is equal to
f (a)
.
1 − (1 − f (a))(1 − f (b))

Also solved by Robert A. Agnew, Jacob Boswell & Chip Curtis, Brian Bradie, Cal Poly Pomona
Problem Solving Group, Robert Calcaterra, Eagle Problem Solvers, Eugene A. Herman, Stephen
Herschkorn, Walther Janous (Austria), Kenneth Levasseur, Northwestern University Math Prob-
lem Solving Group, Pittsburgh State University Problem Solving Group, Rob Pratt, Gary Rad-
mus, Celia Schacht, Edward Schmeichel, Randy K. Schwartz, and the proposer. There were two
incomplete or incorrect solutions.
VOL. 96, NO. 5, DECEMBER 2023 573
Find the area of the checkerboard pattern December 2022
2160. Proposed by Gregory Dresden, Washington & Lee University, Lexington, VA.
Consider the lines

y = x/1, y = x/2, y = x/3, y = x/4, . . .

and the lines

y = (1 − x)/1, y = (1 − x)/2, y = (1 − x)/3, y = (1 − x)/4, . . . ,

which intersect to form an infinite number of quadrilaterals. Starting with the lozenge
at the top, shade every other quadrilateral, as shown in the figure.

Find the total area of all the shaded quadrilaterals.

Solution by Clayton Coe (student), Cal Poly Pomona, Pomona, CA.

The total area is 2 ln 2 − 54 .
Let y = x/n be the equation of line Ln , and y = (1 − x)/k be the equation of line
Mk , where n, k ≥ 1. Observe that
 
n 1
Ln ∩ Mk = , .
n+k n+k
The set of vertices of any of these tiles is
{Ln ∩ Mk , Ln ∩ Mk+1 , Ln+1 ∩ Mk+1 , Ln+1 ∩ Mk }.

Note that Ln+1 ∩ Mk and Ln ∩ Mk+1 have the same y-coordinate. Therefore, we can
calculate the area of a quadrilateral to be the sum of the area of two triangles with hori-
zontal bases. Doing so, we find the area to be hw/2, where h is the difference between
the y-coordinates of Ln ∩ Mk and Ln+1 ∩ Mk+1 , and w is the difference between the
x-coordinates of Ln ∩ Mk+1 and Ln+1 ∩ Mk .
Let A(n, k) denote the area of a single tile, with uppermost vertex Ln ∩ Mk . We
therefore have
  
1 1 1 n+1 n
A(n, k) = − −
2 n+k n+k+2 n+k+1 n+k+1
574 MATHEMATICS MAGAZINE
  
1 2 1
=
2 (n + k)(n + k + 2) n+k+1
1
= .
(n + k)(n + k + 1)(n + k + 2)
Note that each black quadrilateral has uppermost vertex Ln ∩ Mk with n + k even.
Letting n + k = 2m,
1
A(n, k) = .
2m(2m + 1)(2m + 2)
In each horizontal row of quadrilaterals, n + k is constant, and there are n + k − 1 =
2m − 1 quadrilaterals in that row. Consequently, the sum of the areas of all black
quadrilaterals is

∞
2m − 1
S=
m=1
2m(2m + 1)(2m + 2)

Observe that the above sum is absolutely convergent because it is comparable to a

p-series, with p = 2. We perform a partial fraction decomposition, yielding
∞  
−1/2 2 −3/2
S= + +
m=1
2m 2m + 1 2m + 2

Because of absolute convergence, we may shift the index of the first term of the sum-
mand, and obtain
∞  
1  −1/2 2 −3/2
S=− + + +
4 m=1 2(m + 1) 2m + 1 2m + 2
∞  
1  2 2
=− + −
4 m=1 2m + 1 2m + 2
∞    
1 1 1 1
=− +2 − −2 1−
4 m=0
2m + 1 2m + 2 2

5  (−1)i+1
∞
=− +2
4 i=1
i
5
= − + 2 ln 2
4

Also solved by Farrukh Rakhimjanovich Ataev (Uzbekistan), Chip Curtis, Eagle Problem
Solvers, Dmitry Fleischman, Eugene A. Herman, Walther Janous (Austria), Do Hyun Lee (South
Korea), Chrysostom G. Petalas (Greece), William Reil, Volkhard Schindler (Germany), Edward
Schmeichel, Paul K. Stockmeyer, Maria van der Walt, and the proposer.
470 MATHEMATICS MAGAZINE
Solutions

Two directly similar squares October 2022

2151. Proposed by Tran Quang Hung, Hanoi, Vietnam.
Let ABCD and XY ZT be two directly similar squares such that A and Y lie on the
lines XT and CD, respectively. Let M be the intersection of lines XZ and AC, and
let N be the intersection of lines XY and BC. Prove that the circumcenter of XAC
lies on the line MN.

Solution by Katherine Nogin, Clovis North High School (student), Fresno, CA.
Position the squares in the coordinate plane, so that D is at the origin and ABCD is a
unit square with A(0, 1), B(1, 1), and C(1, 0). Let the coordinates of X be (r, s) and
let P be the circumcenter of XAC.

We will make calculations in terms of r and s assuming that none of the resulting
denominators are zero. We will address the question of what happens if some of the
denominators are zero later.
It is straightforward to determine that the equation of line XY is
r
y= (x − r) + s.
1−s
Since N lies on XY , we have
r r + s − r 2 − s2
N = 1, (1 − r) + s = 1, .
1−s 1−s
VOL. 96, NO. 4, OCTOBER 2023 471
Since the y-coordinate of Y is 0, we can solve the equation
r
0= (x − r) + s
1−s
to obtain
r 2 + s2 − s
Y = ,0 ,
r
therefore
−→ s2 − s
XY = , −s .
r
−→ −→
Since XT is perpendicular to XY and has the same length,
−→ s − s2
XT = −s, .
r
−→ −→
As T Z = XY , we have
−→ −→ −→ s2 − s s − s2
XZ = XT + T Z = − s, −s .
r r
The slope of line XZ is thus
s−s 2
−s 1−s−r
mXZ = r
= ,
s 2 −s
−s s−1−r
r

and its equation is therefore

1−s−r
y= (x − r) + s.
s−1−r
The equation of line AC is y = −x + 1. Since M is the intersection of XZ and AC,
we can solve the corresponding system of linear equations to obtain
r 2 + s 2 − 2s + 1 2r + 2s − r 2 − s 2 − 1
M= , .
2r 2r
In order to find the coordinates of P , we will find the intersection of the perpendicular
bisectors of AC and AX. The equation of BD, which is the perpendicular bisector of
AC, is y = x. The midpoint of AX is
r 1+s
,
2 2
and the slope of the perpendicular bisector of AX, which is also the slope of XY , is
r/(1 − s). Thus, the equation of the perpendicular bisector of AX is
r  r 1+s
y= x− + .
1−s 2 2
Setting these two expressions for y equal to each other and solving, we find
r 2 + s2 − 1 r 2 + s2 − 1
P = , .
2(r + s − 1) 2(r + s − 1)
472 MATHEMATICS MAGAZINE
In order to prove that P lies on line MN, we will show that the slopes of line NP
and line MN are equal. We have
2r+2s−r 2 −s 2 −1 2 −s 2
− r+s−r
mMN = 2r
r 2 +s 2 −2s+1
1−s

2r
−1
(2r + 2s − r 2 − s 2 − 1)(1 − s) − (r + s − r 2 − s 2 )2r
= .
(r 2 + s 2 − 2r − 2s + 1)(1 − s)
Then
r 2 +s 2 −1 r+s−r 2 −s 2
2(r+s−1)
− 1−s
mNP = r 2 +s 2 −1
2(r+s−1)
−1
(r 2 + s 2 − 1)(1 − s) − (r + s − r 2 − s 2 )2(r + s − 1)
=
(r 2 + s 2 − 2r − 2s + 1)(1 − s)
(r 2 + s 2 − 1)(1 − s) − (r + s − r 2 − s 2 )2r + (r + s − r 2 − s 2 )2(1 − s)
=
(r 2 + s 2 − 2r − 2s + 1)(1 − s)
(2r + 2s − r 2 − s 2 − 1)(1 − s) − (r + s − r 2 − s 2 )2r
=
(r 2 + s 2 − 2r − 2s + 1)(1 − s)
= mMN .
It follows that the three points N, P , and M lie on the same line, thus the circum-
center of XAC lies on the line MN.
We now consider the cases where the denominators in the above expressions are
zero. Note that r  = 0, otherwise XZ AC, so point M would not be defined. Also,
1 − s  = 0, otherwise XY BC, so point N would not be defined. Next, r + s − 1  = 0,
otherwise point X lies on AC, so XAC is degenerate and does not have a circum-
center.
The case s − 1 − r = 0 is possible, however. In that case,
r +1 r +1
N = (1, 2r) , M = (r, 1 − r) , and P = , ,
2 2
which are collinear.
Finally, r 2 + s 2 − 2r − 2s + 1 = 0 is possible. This equation is equivalent to (r −
1) + (s − 1)2 = 1, so point X lies on a circle of radius 1 centered at B. It follows that
2

P = B, M = C, and N lies on BC. Thus, in all possible cases P lies on MN.

Also solved by Robert Calcaterra, Ivko Dimitric̀, Walther Janous (Austria), Michael Vowe
(Switzerland), and the proposer. There was one incomplete or incorrect solution.

Evaluate the double integral October 2022

2152. Proposed by Paul Bracken, University of Texas Rio Grande Valley, Edinburg,
TX.
Evaluate
 1  1
dy dx
√ .
0 0 1 − x 1 − y 2 (1 + xy)
2
VOL. 96, NO. 4, OCTOBER 2023 473
Solution by Omran Kouba, Higher Institute for Applied Sciences and Technology,
Damascus, Syria.
Denote the proposed integral by I . The change of variables x = sin θ, y = sin ϕ shows
that
 π/2  π/2
dθ dϕ
I= .
0 0 1 + sin ϕ sin θ
For ϕ ∈ (−π/2, π/2), we define F (ϕ) by
 π/2
dθ
F (ϕ) =
0 1 + sin ϕ sin θ
Clearly, for 0 < ϕ < π/2 we have
 π/2  π/2
dθ 1 dθ
F (ϕ) + F (−ϕ) = 2 =2
0 1 − sin ϕ sin θ
2 2
0 1 + cot θ − sin ϕ sin2 θ
2 2

 ∞
du π
=2 2 ϕ + u2
=
0 cos cos ϕ
 π/2  π/2
−2 sin ϕ sin θ −2 sin ϕ sin θ
F (ϕ) − F (−ϕ) = dθ = dθ
0 1 − sin ϕ sin θ
2 2
0 cos ϕ + sin2 ϕ cos2 θ
2

 0
2 du 2ϕ
= =− (u = tan ϕ cos θ).
cos ϕ tan ϕ 1 + u 2 cos ϕ
Thus
π − 2ϕ π
F (ϕ) = , 0<ϕ< .
2 cos ϕ 2
Integrating and using the change of variables ϕ = π
2
− 2x, we get
 π/2
I= F (ϕ) dϕ
0
 π/4  π/4
4x x
= dx = 2 dx
0 sin 2x 0 sin x cos x
 π/4  1
π/4 ln(u)
= 2x ln(tan x) −2 ln(tan x) dx = −2 du
0 0 0 1 + u2

∞  1 
∞ n
(−1)
=− (−1)n u2n ln(u) du = 2 = 2G.
n=0 0 n=0
(2n + 1)2

where G is the Catalan number.

Also solved by Ulrich Abel & Vitaliy Kushnirevych (Germany), Carl Axness, Michel Bataille
(France), Khristo N. Boyadzhiev, Brian Bradie, James Brewer, Charles Burnette, Robert Cal-
caterra, Hongwei Chen, Bruce Davis, Cal Poly Pomona Problem Solving Group, Eagle Problem
Solvers, Fejéntalàltuka Szeged Problem Solving Group (Hungary), Jan Grzesik, Walther Janous
(Austria), Stephen Kaczkowski, Kee-Wai Lau (Hong Kong, China), Muzahim Mamedov (Azer-
baijan), Moubinool Omarjee (France), Shing Hin Jimmy Pa (China), Paolo Perfetti, Volkhard
Schindler (Germany), Seán Stewart (Saudi Arabia), Michael Vowe (Switzerland), Haohao Wang,
and the proposer. There were two incomplete or incorrect solutions.
474 MATHEMATICS MAGAZINE
Series involving Fibonacci and Lucas numbers October 2022
2153. Proposed by Rex H. Wu, New York, NY.
Let Fn and Ln be the Fibonacci and Lucas numbers, respectively. Evaluate the follow-
ing for k ≥ 0.

∞
F2k
(a) arctan
n=0
F2n+1

∞
L2k+1
(b) arctan
n=0
L2n

Solution by Russell Gordon, Whitman College, Walla Walla, WA.

We first establish some common notation and make note of a simple trigonometric
identity. Let α = φ and β = −1/φ, where φ is the golden ratio. By the Binet formulas
for the Lucas and Fibonacci numbers, we have
αn − β n
Fn = √ and Ln = α n + β n
5
for all nonnegative integers n. In addition, the trigonometric identity
 u−v 
arctan u − arctan v = arctan ,
1 + uv
is valid for all positive real numbers u and v.
When k = 0, the Fibonacci series clearly sums to 0. For the other series involving
the Fibonacci numbers, we begin by noting that
α 2n+2k+1 − α 2n−2k+1 −α 2k + α −2k α 2k − β 2k F2k
= = =
1+α 4n+2 β 2n+1 −α 2n+1 α 2n+1 −β 2n+1 F2n+1
for all positive integers k and all nonnegative integers n. It follows that

∞  F  N 
   
2k
arctan = lim arctan α 2n+2k+1 − arctan α 2n−2k+1
n=0
F2n+1 N→∞
n=0


N
 
 2k−1  
= lim arctan α 2n+2k+1 − arctan α 2n−2k+1
N→∞
n=N−2k+1 n=0


2k
  
2k−1
 
= lim arctan α 2N−2k+2+1 − arctan α 2n−2k+1
N→∞
=1 n=0

π      
k−1 2k−1
= 2k · − arctan α −2(k−n−1)−1 − arctan α 2(n−k)+1
2 n=0 n=k

π      
k−1 k−1
= 2k · − arctan α −(2j +1) − arctan α 2j +1
2 j =0 j =0

π     
k−1
= 2k · − arctan α 2j +1 + arctan α −(2j +1)
2 j =0
VOL. 96, NO. 4, OCTOBER 2023 475
π π
= 2k · − k ·
2 2
kπ
= .
2
This gives the sums of the series involving the Fibonacci numbers.
For the series involving the Lucas numbers, we first observe that
α 2n+2k+1 − α 2n−2k−1 α 2k+1 − α −2k−1 α 2k+1 + β 2k+1 L2k+1
= = =
1+α 4n β +α
2n 2n α 2n + β 2n L2n
for all positive integers k and all nonnegative integers n. It follows that


∞ L  N 
    
2k+1
arctan = lim arctan α 2n+2k+1 − arctan α 2n−2k−1
n=0
L2n N→∞
n=0


N
  
2k
 
= lim arctan α 2n+2k+1
− arctan α 2n−2k−1
N→∞
n=N−2k n=0


2k
  
2k
 
= lim arctan α 2N−2k+2+1
− arctan α 2n−2k−1
N→∞
=0 n=0

π 
k
  
2k
 
−2(k−n+1)+1
= (2k + 1) · − arctan α − arctan α 2(n−k)−1
2 n=0 n=k+1

π   
k
 −(2j −1)  k
 
= (2k + 1) · − arctan α −(2k+1) − arctan α − arctan α 2j −1
2 j =1 j =1

k 
π       
= (2k + 1) · − arctan α −(2k+1) − arctan α 2j −1 + arctan α −(2j −1)
2 j =1

π  π   π
= 2k · + − arctan α −(2k+1) − k ·
2 2 2
kπ  
= + arctan α 2k+1
2
kπ  
= + arctan φ 2k+1 .
2
We have thus found the sums of all of the series involving the Lucas numbers.
Also solved by Michel Bataille (France), Brian Bradie, Hongwei Chen, Michael Goldenberg &
Mark Kaplan, Walther Janous (Austria), Won Kyun Jeong (South Korea), Omran Kouba (Syria),
Angel Plaza (Spain), Albert Stadler (Switzerland), and the proposer.

Ordered partitions with all parts odd October 2022

2154. Proposed by the Columbus State University Problem Solving Group, Columbus,
GA.
Let f (n) denote the number of ordered partitions of a positive integer n such that all of
the parts are odd. For example, f (5) = 5, since 5 can be written as 5, 3 + 1 + 1, 1 +
3 + 1, 3 + 1 + 1, and 1 + 1 + 1 + 1 + 1. Determine f (n).
476 MATHEMATICS MAGAZINE
Solution by Dominic Kozlowski and Anayelli Vigo (students), Seton Hall University,
South Orange, NJ.
We claim that f (n) = Fn , where Fn is the nth Fibonacci number. We will write indi-
vidual partitions as ordered k-tuples and denote the set of odd partitions of n by An .
For example,
A5 = {(5), (3, 1, 1), (1, 3, 1), (1, 1, 3), (1, 1, 1, 1, 1)}.
It will suffice to show that
|A1 | = 1, |A2 | = 1, and |An | = |An−1 | + |An−2 | for all n > 2.
That |A1 | = |A2 | = 1 can be calculated immediately. For n > 2, define the maps p1 :
An−1 → An and p2 : An−2 → An as
p1 (a1 , a2 , . . . , ak ) = (1, a1 , a2 , . . . , ak ),
p2 (a1 , a2 , . . . , ak ) = (a1 + 2, a2 , . . . , ak ).
Now let p : An−1 ∪ An−2 → An be the map consisting of the application of p1 to the
elements of An−1 and the application of p2 to the elements of An−2 .
To show p is onto, let b ∈ An , with b = (b1 , b2 , . . . , bk ). We have two cases.
Case 1: b1 = 1. In this case (b2 , b3 , . . . , bk ) is an odd partition of n − 1, and so
(b2 , b3 , . . . , bk ) ∈ An−1 . But then
p(b2 , b3 , . . . , bk ) = p1 (b2 , b3 , . . . , bk ) = b.
Case 2: b1 > 1. In this case since b1 is odd we must have b1 ≥ 3 and odd. But then
b1 − 2 ≥ 1 and is odd, and so (b1 − 2, b2 , . . . , bk ) is an odd partition of n − 2 and so
(b1 − 2, b2 , . . . , bk ) ∈ An−2 . But then
p(b1 − 2, b2 , . . . , bk ) = p2 (b1 − 2, b2 , . . . , bk ) = b.
To show that p is one-to-one, we note that both of the individual maps p1 and p2 are
clearly one-to-one. We need only check a  = a implies p(a)  = p(a ) for all a ∈ An−1 ,
a ∈ An−2 . But p(a) starts with a 1, and p(a ) starts with a number greater than 1, so
p(a)  = p(a ).
Since p is a bijection between An−1 ∪ An−2 and An , |An−1 | + |An−2 | = |An | for all
n > 2 and the claim follows.
Also solved by Alrich Abel & Vitaliy Kushnirevych (Germany), Ashland University Prob-
lem Solving Group, McCrea Black & the Texas State University Problem Solvers Group, Saham
Bhadra (India), Ricardo Bittencourt (Brazil), Charles Burnette, Robert Calcaterra, Rohan Dalal,
Eagle Problem Solvers, Fejéntalátuka Szeged Problem Solving Group (Hungary), Haydn Gwyn,
Brian Hopkins, Walther Janous (Austria), Kenneth Klinger, Kee-Wai Lau (Hong Kong, China),
S. C. Locke, Samuel Lucas Mazariegos, Katherine Nogin, Northwestern University Math Problem
Solving Group, Rob Pratt, Edward Schmeichel, Albert Stadler (Switzerland), Paul K. Stockmeyer,
and the proposers. There was one incomplete or incorrect solution.

A problem from ring theory October 2022

2155. Proposed by Ioan Băetu, Botoşani, Romania.
Let R be a ring with identity and U a subset of the units of R with |U | = p, where p
is an odd prime. Suppose that for all a ∈ R, there is a u ∈ U and a k ∈ Z+ such that
ua k = a k+1 . Show that
(a) For all a ∈ R, there is a u ∈ U such that ua = a 2 .
VOL. 96, NO. 4, OCTOBER 2023 477
(b) The ring R is commutative.

Solution by Robert Calcaterra, University of Wisconsin - Platteville, Platteville, WI.

The condition that p be prime is unnecessary; it suffices to assume that p is odd.
Suppose v is an invertible element of R. Then uv k = v k+1 for some u ∈ U and
k ∈ Z+ and so v = u ∈ U . Therefore, U must be the group of all the units of R. The
order of −1 in the group U is 1 as it is a divisor of both 2 and p. Hence, −1 = 1 and
R has characteristic 2.
Next we claim a 2 = 0 in the ring R if and only if a = 0 and that every finite subring
of R containing 1 is the direct sum of fields. To prove this, suppose a ∈ R and a 2 = 0.
Then (a + 1)2 = a 2 + 1 = 1. Thus, a + 1 is its own inverse in the group U and so
a + 1 = 1, and a = 0 and the first part of the claim is validated. Note that this implies
every finite subring of R containing 1 is semiprime. Since such a ring must also be
artinian, The Wedderburn-Artin theorem implies it is isomorphic to the direct sum
M 1 ⊕ M2 ⊕ · · · ⊕ M k ,
where each Mi is a full matrix ring over a division ring. Since these division rings are
necessarily finite, another theorem of Wedderburn implies they are fields. Finally, note
that the square of any n × n matrix with n > 1 that has a 1 as the upper right entry and
zeros elsewhere is the zero matrix. This is impossible by the first part of the claim, so
each Mi must consist of 1 × 1 matrices and therefore be isomorphic to its component
field. This verifies the second part of the claim.
Let a ∈ R and let Sa be the set of finite sums in R in which each term is either 1
or a raised to a positive integer. Then Sa is a subring of R. Since ua k = a k+1 for some
u ∈ U and k ∈ Z+ , it follows that
u2 a k = ua k+1 = a k+2 , u3 a k = ua k+2 = a k+3 ,
and so on. Therefore,
a k = up a k = a k+p , a k+1 = a k+p+1 ,
and so on. Consequently, Sa is finite, so by the claim above
Sa ∼
= F1 ⊕ F2 ⊕ · · · ⊕ Fn ,
where the Fi are finite fields. Let a ∈ Sa be represented by (a1 , a2 , . . . , an ) in that
direct sum and let uj be the identity of Fj if aj = 0 and aj if aj  = 0. Then

(u1 , u2 , . . . , un )(a1 , a2 , . . . , an ) = (a1 , a2 , . . . , an )2 ,

where (u1 , u2 , . . . , un ) is a unit in F1 ⊕ F2 ⊕ · · · ⊕ Fn . Hence, there is a unit u ∈ Sa
such that ua = a 2 . Since Sa and R share the same identity, a unit of Sa is also invertible
in R. This proves statement (a).
Suppose a ∈ R. Let Ta be the set of all finite sums in which each term is of the form
u or uav for some u, v ∈ U . Observe that if x, y, s, t ∈ U then
(xay)(sat) = x(ays)(ays)s −1 y −1 t = x(uays)s −1 y −1 t = xuat
for some u ∈ U by (a). Hence, Ta is a subring of R and is finite because U is finite.
Therefore, Ta is isomorphic to the direct sum of fields and is commutative. Since a is
an arbitrary element of R, U must be in the center of R.
Lastly, let a and b be elements of R. Let Qab be the set of all finite sums in which
each term is of the form u, ua, ub, uab, uba, uaba, or ubab for some u ∈ U . Note
478 MATHEMATICS MAGAZINE
that whenever two elements of these types are multiplied, their product is also of one
of these types. For example, if x, y ∈ U , then there exists u, v, w ∈ U such that
(xa)(yab) = xya 2 b = xy(ua)b,
(xb)(yaba) = xy(ba)2 = xy(vba), and,
(xab)(yaba) = xy(ab)2 a = xy(wab)a.

But Qab is finite, hence ab = ba for all a, b ∈ R by the claim above. This verifies
statement (b).
Also solved by the proposer.

Answers
Solutions to the Quickies from page 469.

A1133. We have

2n + 2  2kn+1  2k n
2n + 2 2n
= − = an+1 − an .
n+1 k=0
k k=0
k n+1 n

Since
2n (2n)! (n + 1)2 (2n + 2)! n + 1 2n + 2
= = · = ,
n (n!) 2 (2n + 2)(2n + 1) ((n + 1)!) 2 4n + 2 n + 1
we have
2n + 2 2n + 2 n + 1 2n + 2
= an+1 − an · .
n+1 n+1 4n + 2 n + 1
Therefore
n+1
an+1 = an + 1 with a0 = 1.
4n + 2

A1134. Let r be the rank of A. The condition we seek is that all of the entries in the
last m − r rows of A are zero.
Suppose the column spaces are equal. Since the last m − r rows of R consist entirely
of zeros, the last m − r rows of A must also have this property.
Now suppose that the last m − r rows of A consist entirely of zeros. The last m − r
rows of R must have the same property. Let V be the r-dimensional subspace of Fm
consisting of vectors whose final m − r entries are zero. Clearly the column space
of A and the column space of R are subspaces of V . But both column spaces have
dimension r, so both are equal to V .
 VOL. 96, NO. 3, JUNE 2023 361
Solutions

Good and bad integer pairs June 2022

2146. Proposed by Kenneth Fogarty, Bronx Community College (emeritus), Bronx, NY.
Let a and d be integers with d > 0. We say that (a, d) is good if there is an arithmetic
sequence with initial term a and difference d that can be split into two sequences of
consecutive terms with the same sum. In other words, there exist integers k and n with
0 < k < n such that

k−1 
n−1
(a + di) = (a + di) .
i=0 i=k

If there is no such arithmetic sequence, we say that (a, d) is bad.

(a) Show that if 2a > d, then (a, d) is good.
(b) Show that if 2a = d, then (a, d) is bad.
(c) Show that if a = 0 (and hence 2a < d), then (a, d) is good.
(d) Show that if 2a < d and a  = 0, then there is a d such that (a, d) is good and a d
such that (a, d) is bad.

Solution by Eagle Problem Solvers, Georgia Southern University, Statesboro, GA and

Savannah, GA.
It is straightforward to verify that


k−1
k(2a + d(k − 1)) 
n−1
(n − k)(2a + d(k + n − 1))
(a + di) = and (a + di) = .
i=0
2 i=k
2

One then readily deduces that (a, d) is good if and only if there exist integers k and n
with 0 < k < n such that

(2a − d)(2k − n) = d(n2 − 2k 2 ). (1)

For part (a), if 2a > d, then 2a − d ∈ N. Thus, to show (a, d) is good, it suffices
to show there is a solution to (2k − n) = d(n2 − 2k 2 ) since if (k1 , n1 ) is a solution to
(2k − n) = d(n2 − 2k 2 ) with 0 < k1 < n1 , then ((2a − d)k1 , (2a − d)n1 ) is a solution
to (2a − d)(2k − n) = d(n2 − 2k 2 ) with 0 < (2a − d)k1 < (2a − d)n1 . Multiplying
both sides of (2k − n) = d(n2 − 2k 2 ) by 4d and rearranging gives

4d 2 n2 + 4dn − 2(4d 2 k 2 + 4dk) = 0

(2dn + 1)2 − 2(2dk + 1)2 = −1
x 2 − 2y 2 = −1,

where x = 2dn + 1 and y = 2dk + 1. Positive integer solutions (x, y) to Pell’s equa-
tion x 2 − 2y 2 = −1 are given by
   j  
xj 3 4 1
= ,
yj 2 3 1
362 MATHEMATICS MAGAZINE
   
x 1
where j is a nonnegative integer. Since 0 = corresponds to k = n = 0, we
y0 1
 
3 4
seek a positive integer j such that xj ≡ 1 ≡ yj (mod 2d). Since A = has
2 3
determinant 1, it follows that A ∈ SL2 (Z2d ), which
 has finite
 order for each positive
1 0
integer d. Thus, there exists j ∈ N such that Aj ≡ (mod 2d), and
0 1
   j    
xj 3 4 1 1
= ≡ (mod 2d).
yj 2 3 1 1
So, xj = 2dn + 1 and yj = 2dk + 1 for positive integers k and n, and (k, n) is a
solution to (2k − n) = d(n2 − 2k 2 ). An easy induction argument shows that xj > yj
for j > 9 and hence, n > k > 0. Thus, (a, d) is good for all integers a and d with
2a > d > 0.
√
For part (b), if 2a = d, then we would√need n2 = 2k 2 or 2 = n/k, which is impos-
sible for positive integers n and k since 2 is irrational. Thus, (a, d) is bad if 2a = d.
For part (c), if a = 0, then equation (1) becomes 2k − n = 2k 2 − n2 , or n2 − n =
2(k 2 − k), which is satisfied for k = 3 and n = 4 (for all positive integers d). Thus
(a, d) is good if a = 0.
For part (d), we show that for each nonzero integer a, there exist positive integers
d1 > 2a and d2 > 2a for which (a, d1 ) is good and (a, d2 ) is bad. If 2a < d, then
equation (1) becomes
(d − 2a)(2k − n) = d(2k 2 − n2 ).
We consider two cases, depending on the sign of a.
Case 1: a > 0. Let d1 = 3a > 0. Letting k = 5 and n = 7,
(d1 − 2a)(2k − n) = a(10 − 7) = 3a = d1 = d1 (2k 2 − n2 ),
so that (a, 3a) is good for every positive integer a.
Let d2 = 4a > 0. Then equation (1) becomes 2k − n = 2(2k 2 − n2 ). Multiplying
both sides by 8 and rearranging gives us
(16n2 − 8n) − 2(16k 2 − 8k) = 0
(16n2 − 8n + 1) − 2(16k 2 − 8k + 1) = −1
x 2 − 2y 2 = −1,
where x = 4n − 1 ≡ 3 (mod 4) and y = 4k − 1 ≡ 3 (mod 4). From our earlier dis-
 2  
3 4 1 0
cussion of solutions to x − 2y = −1 in part (a), we see that
2 2
≡
2 3 0 1
(mod 4). So, if j is odd, then
   j    
xj 3 4 1 3
= ≡ (mod 4);
yj 2 3 1 1
meanwhile, if j is even, then
   j    
xj 3 4 1 1
= ≡ (mod 4).
yj 2 3 1 1
VOL. 96, NO. 3, JUNE 2023 363
Thus, there are no solutions (x, y) to x 2 − 2y 2 = −1 with x ≡ 3 ≡ y (mod 4). Hence,
(a, 4a) is bad for every positive integer a.
Case 2: a < 0. Then 2a < d for any positive integer d. Let d1 = 1. Then d1 − 2a =
1 − 2a > 0. Letting k = 3(1 − 2a) and n = 4(1 − 2a), we see that 0 < k < n and
(d1 − 2a)(2k − n) = (1 − 2a)2 · 2 = 2 (3(1 − 2a))2 − (4(1 − 2a))2 = d1 (2k 2 − n2 ),

and (a, 1) is good for every negative integer a.

Let d2 = −4a > 0. Then d2 − 2a = −6a and equation (1) becomes

−6a(2k − n) = −4a(2k 2 − n2 )
3(2k − n) = 2(2k 2 − n2 )
(16n2 − 24n + 9) − 2(16k 2 − 24k + 9) = −9
x 2 − 2y 2 = −9,

where x = 4n − 3 ≡ 1 (mod 4) and y = 4k − 3 ≡ 1 (mod 4). Positive solutions to

x 2 − 2y 2 = −9 are given by
   j  
xj 3 4 3
= ,
yj 2 3 3

where j is a nonnegative integer. If j is odd, then

   j    
xj 3 4 3 1
= ≡ (mod 4);
yj 2 3 3 3

meanwhile, if j is even, then

   j    
xj 3 4 3 3
= ≡ (mod 4).
yj 2 3 3 3

Thus, there is no solution (x, y) to x 2 − 2y 2 = −9 with x ≡ 1 ≡ y (mod 4); hence,

(a, −4a) is bad for every negative integer a.
Also solved by Eugene A. Herman, Dmitry Fleischman (partial solution), Fresno State Problem
Solving Group (partial solution), William Boyd & Ernest James (partial solution), Fejéntaláltuka
Szeged Problem Solving Group (Hungary) (partial solution), and the proposer.

Evaluate the infinite product June 2022

2147. Proposed by Lokman Gökçe, Istanbul, Turkey.
Evaluate


∞
n4 + 4
.
n=2
n4 − 1

Solution by Michel Bataille, Rouen, France.

2 sinh(π)
We prove that the value of the given infinite product is .
5π
364 MATHEMATICS MAGAZINE
From Sophie Germain’s identity
x 4 + 4 = ((x + 1)2 + 1)((x − 1)2 + 1),
we deduce that for any integer N ≥ 3, we have

N+1


N 
N (n2 + 1)
n4 + 4 ((n + 1)2 + 1)((n − 1)2 + 1) 2 n=3
PN := = = 2 · N ,
n4 − 1 (n2 − 1)(n2 + 1) N +1 
n=2 n=2 (n2 − 1)
n=2

that is,
N 

1+ 1
N 2 + 2N + 2 n=3 n2
PN = · N .
2(N 2 + 1) 
1− 1
n2
n=2

Now, from
N 
  
N
1 (n − 1)(n + 1) 1 N +1
1− 2 = = · ,
n=2
n n=2
n·n 2 N

we obtain
N 
 
1 1
lim 1− =
N→∞
n=2
n2 2

and from the well-known

∞ 
 
z2 sin(πz)
1− 2 = (z ∈ C, z  = 0),
n=1
n πz

we deduce that
N 
 
1 sin(πi) sinh(π)
lim 1+ 2 = = .
N→∞
n=1
n πi π

Since
N 2 + 2N + 2
lim = 1,
N→∞ N2 + 1
it follows that

∞
n4 + 4 sinh(π)/π 1 2 sinh(π)
= lim PN = · = .
n=2
n4 −1 N→∞ 2(1 + 112 )(1 + 1
22
) 1/2 5π

Also solved by Anthony J. Bevelacqua, Ricardo Bittencourt (Brazil), Paul Bracken, Brian
Bradie, Hongwei Chen, Junan Chen (China), Bruce E. Davis, Fejéntaláltuka Szeged Problem Solv-
ing Group (Hungary), Tasha Fellman, Shuyang Gao, Eugene A. Herman, Walther Janous (Aus-
tria), Warren P. Johnson, Sofia Lacerda (Brazil), Kee-Wai Lau (China), Isaac Venegas Macevschi,
Donald Jay Moore, Raymond Mortini (Luxembourg) & Rudolf Rupp (Germany), Northwestern
University Math Problem Solving Group, Peter Oman & Haohao Wang, Celia Schacht, Albert
VOL. 96, NO. 3, JUNE 2023 365
Stadler (Switzerland), Seán M. Stewart (Saudi Arabia), Michael Vowe (Switzerland), Mark Wildon
(UK), and the proposer. There were four incomplete or incorrect solutions.

An application of the Erdős-Mordell inequality June 2022

2148. Proposed by Quang Hung Tran, Hanoi, Vietnam.
Let P be an interior point of triangle ABC. Denote by δa , δb , and δc the distances from
the midpoints of segments P A, P B, and P C to the lines BC, CA, and AB. Prove that
P A + P B + P C ≥ δ a + δb + δ c .
Show that equality holds if and only if triangle ABC is equilateral and P is its center.

Solution by Fejéntaláltuka Szeged Problem Solving Group, University of Szeged,

Szeged, Hungary.
Let ha , hb , and hc be the altitudes of triangle ABC through the vertices A, B, and C
respectively. Denote by x, y, and z the distances from P to the lines BC, CA, and AB.

Consider the trapezoid defined by the points A, A , Pa , P . The segment of length

δa is the midline of this trapezoid, thus δa = (ha + x)/2, and ha = 2δa − x.
From the triangle inequality we have P A + x ≥ APa . As the segment of length ha
is also the altitude of triangle ABPa , we have APa ≥ ha . Thus,
P A + x ≥ ha
P A + x ≥ 2δa − x,
which implies
P A ≥ 2δa − 2x.
 366 MATHEMATICS MAGAZINE
Similarly, it can be shown that

P B ≥ 2δb − 2y,
P C ≥ 2δc − 2z.
The sum of these inequalities gives

P A + P B + P C ≥ 2δa + 2δb + 2δc − 2x − 2y − 2z.

The Erdős-Mordell inequality states that

P A + P B + P C ≥ 2(x + y + z).

Adding this to the inequality preceding it gives

2(P A + P B + P C) ≥ 2(δa + δb + δc ).

After dividing both sides by 2, we obtain the desired inequality.

It is clear from the proof that if equality holds in the proposed inequality, then we
must have equality in all the inequalities above. However, it is known that the Erdős-
Mordell equality holds if and only if triangle ABC is equilateral and P is its center.
One can easily show that if triangle ABC is equilateral and P is its center then equality
holds in the desired inequality. This finishes the solution.
Also solved by Nandan Sai Dasireddy (India), Celia Schacht, Michael Vowe (Switzerland), and
the proposer.

An irrational alternating sum June 2022

2149. Proposed by Ioan Băetu, Botoşani, Romania.
Let a1 , a2 , . . . be a sequence of integers greater than 1. The series
∞
(−1)k 1 1 1
k =1− + − + ···
k=0 i=1 ai
a1 a1 a2 a1 a2 a3

converges by the alternating series test.

(a) If the sequence a1 , a2 , . . . is unbounded, show that the sum of the series is irrational.
(b) Give an example of a bounded sequence of ai ’s such that the sum of the series is
irrational.

Solution by the Fresno State Journal Problem Solving Group, Fresno State University,
CA.
(a) Suppose to the contrary that for some unbounded sequence a1 , a2 , . . . , the sum of
the series is rational, say,
∞
(−1)k m
k = ,
k=0 i=1 ai
n

where m, n ∈ Z, n > 0. Since the sequence {ai } is unbounded, there exists r ∈ N

such that ar > n. Expanding the left-hand side of the above equation, we have
1 1 1 1
1− + − + · · · + (−1)r−1
a1 a1 a2 a1 a2 a3 a1 a2 . . . ar−1
 VOL. 96, NO. 3, JUNE 2023 367
1 1
+ (−1)r + (−1)r+1 + ···
a1 a2 . . . ar−1 ar a1 a2 . . . ar−1 ar ar+1
m
= .
n
Multiplying both sides by a1 a2 . . . ar−1 n gives
 
1 1 1 1
a1 a2 . . . ar−1 n 1 − + − + · · · + (−1) r−1
a1 a1 a2 a1 a2 a3 a1 a2 . . . ar−1
n n
+ (−1)r + (−1)r+1 + ···
ar ar ar+1
= a1 a2 . . . ar−1 m,
or, equivalently,
 
1 1 1 1
a1 a2 . . . ar−1 n 1 − + − + · · · + (−1)r−1
a1 a1 a2 a1 a2 a3 a1 a2 . . . ar−1
⎛ ⎞ S

⎜n n ⎟
+ (−1)r ⎜
⎝ ar − + · · ·⎟
⎠
ar ar+1
X

= a1 a2 . . . ar−1 m,
Observe that S and a1 a2 · · · ar−1 m are integers, therefore, X is an integer. How-
ever, X is the sum of an alternating series with decreasing terms approaching 0
n n
(i.e., lim = 0), therefore, 0 < X < < 1. In this case X cannot be an
q→∞ ar · · · aq ar
integer, so we have a contradiction.
(b) Consider the following sequence:
2, 5, 5, 2, 2, 5, 2, 5, 5, 2, 2, 5, 2, 5, 2, 5, 5, 2, . . . ,
where a block of 2,5 is followed by one block of 5,2, then two blocks of 2,5 are
followed by one block of 5,2, then three blocks of 2,5 are followed by one block of
5,2, and so on.
In this case, we have

∞    
(−1)k 1 1 1
k = 1 − + − + ···
k=0 i=1 ai
a1 a1 a2 a1 a2 a3
     
1 1 1 1 1
= 1− + 1− + 1− + ···
a1 a1 a2 a3 a1 a2 a3 a4 a5
1 1 1 1
= 0.5 + · 0.8 + 2 · 0.5 + 3 · 0.5 + 4 · 0.8 + · · ·
10 10 10 10
= 0.585585558 . . . .
Since the blocks of 5 in the resulting number increase in length, this is a non-
repeating decimal, so it represents an irrational number.
368 MATHEMATICS MAGAZINE
Also solved by Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia), Eugene A. Herman
(partial solution), Evin Liang, Northwestern University Problem Solving Group, Ioana Mihaila &
Ivan Ventura, Celia Schacht, and the proposer. There was one incomplete or incorrect solutions.

Maximize the area of a triangle in a cardioid June 2022

2150. Proposed by Matthew McMullen, Otterbein University, Westerville, OH.
Find the maximum area of a triangle whose vertices lie on the cardioid r = 1 + cos θ.

Editor’s Note. Unfortunately, the argument that the triangle having maximal area is
symmetric with respect to the x-axis is too long and involved to include here. The
statement of the problem should have included the condition that this symmetry held
in order to make the problem more tractable. We regret not having done so.
Solution by the proposer.
There are two cases to consider. First, assume the vertex lying on the x-axis is (0, 0). If
the other vertices are ((1 + cos θ) cos θ, ±(1 + cos θ) sin θ) with 0 ≤ θ ≤ π, the area
of the triangle is A(θ) = |f (θ)|, where

f (θ) = sin θ cos θ (1 + cos θ)2 .

Now

f (θ) = cos2 θ (1 + cos θ)2 − sin2 θ (1 + cos θ)2 − sin2 θ cos θ (1 + cos θ) .
Letting sin2 θ = 1 − cos2 θ and simplifying gives


f (θ) = (1 + cos θ)2 4 cos2 θ − 2 cos θ − 1 .
 √ 
The only critical points for A(θ) having a nonzero area are when cos θ = 1 ± 5 /4.
 √ 
The maximum value of A(θ) occurs when cos θ = 1 + 5 /4 and in that case


5 √
A(θ) = 50 + 22 5 ≈ 1.55619.
32
The second case is when the vertex lying on the x-axis is (2, 0). In this case,
A(θ) = sin θ(1 + cos θ)(2 − cos θ(1 + cos θ))
= (2 + cos θ) sin3 θ.

Therefore,
A (θ) = (4 cos2 θ + 6 cos θ − 1) sin2 θ.
√ 
The only critical point for A(θ) having a nonzero area is when cos θ = 13 − 3 /4.
In that case

3 √
A(θ) = 105 + 39 13 ≈ 2.07785.
32
This is therefore the overall maximum area.
192 MATHEMATICS MAGAZINE
Solutions

An improper logarithmic integral April 2022

2141. Proposed by Paul Bracken, University of Texas Rio Grande Valley, Edinburg,
TX.
Evaluate
 ∞  
ln 1 + 2x −2 cos ϕ + x −4 dx.
0

Solution by J. A. Grzesik, Torrance, CA.

Let
 ∞
I (ϕ) = ln(1 + 2x −2 cos ϕ + x −4 ) dx
0
 ∞
= ln(x 4 + 2x 2 cos ϕ + 1) − 4 ln x dx.
0

Since I has period 2π, we may take ϕ ∈ (−π, π]. We claim that I (ϕ) = 2π cos(ϕ/2).
One obtains this by first noting that
x 4 + 2x 2 cos ϕ + 1 = (x 2 + eiϕ )(x 2 + e−iϕ ),
whence
 ∞
I (ϕ) = 2Re ln(x 2 + eiϕ ) − 2 ln x dx
0
 ∞
= 2Re ln(x + ei(π+ϕ)/2 ) + ln(x − ei(π+ϕ)/2 ) − 2 ln x dx
0

= 2Re (x + ei(π+ϕ)/2 ) ln(x + ei(π+ϕ)/2 ) − x

 x=∞
+ (x − ei(π+ϕ)/2 ) ln(x − ei(π+ϕ)/2 ) − x − 2x ln x + 2x
x=0
 
= 2πIm ei(π+ϕ)/2 = 2π cos(ϕ/2).

These manipulations hold so long as ϕ = π. When ϕ = π, there is a singularity when

x = 1 and the integral must be split into two parts. Here one finds that
 ∞  
I (π) = 2 ln |x − 1|(x + 1) − 2 ln x dx
0
   
1 ∞ ∞
=2 ln(1 − x)dx + ln(x − 1)dx + ln(x + 1) − 2 ln x dx
0 1 0

⎡
1 ∞

= 2 ⎣ − (1 − x) ln(1 − x) + (1 − x) + (x − 1) ln(x − 1) − (x − 1)
0 1
VOL. 96, NO. 2, APRIL 2023 193
∞   ∞ 
(x + 1) ln(x + 1) − (x + 1) −2 x ln x − x
0 0

= 0,

in agreement with the claimed value of I (ϕ) = 2π cos(ϕ/2).

Note that replacing the definite integrals with indefinite integrals in the second set
of displayed equations allows us to find an elementary antiderivative

   
ln x −4 + 2x −2 cos ϕ + 1 dx = x ln x −4 + 2x −2 cos ϕ + 1

ϕ   
x 2 + 2x sin(ϕ/2) + 1
+ sin ln
2 x 2 − 2x sin(ϕ/2) + 1
ϕ   
2x cos(ϕ/2)
+ 2 cos arctan .
2 1 − x2

Also solved by Ulrich Abel & Vitaliy Kushnirevych (Germany), Carl Axness (Spain), Michel
Bataille (France), Robert Benim, Khristo N. Boyadzhiev, Brian Bradie, Bruce S. Burdick, Hong-
wei Chen, Bruce E. Davis, John N. Fitch, Fatima Gulieva (Azerbaijan), Eugene A. Herman,
Walther Janous (Austria), Warren P. Johnson, Stephen Kaczkowski, Omran Kouba (Syria), James
Magliano, Kelly D. McLenithan, Raymond Mortini (France) & Rudolph Rupp (Germany), North-
western University Math Problem Solving Group, Moubinool Omarjee (France), Shing Hin Jimmy
Pa (China), Paolo Perfetti (Italy), Didier Pinchon (France) Albert Stadler (Switzerland), Seán M.
Stewart (Saudi Arabia), Michael Vowe (Switzerland), and the proposer.

Constructing the axis and focus of a parabola April 2022

2142. Proposed by Roger Izard, Dallas, TX.
Given a parabola in the plane, find its axis and focus using compass and straightedge.

Solution by Michelle Nogin (student), Clovis North High School, Fresno, CA.
We will use the following facts about parabolas.
(1) If points A, B, C, and D lie on the parabola with AB CD, then the line through
the midpoints of segments AB and CD is parallel to the axis of symmetry.
Proof. Choose the coordinate system so that the vertex of the parabola is at
the origin and the axis of symmetry is the y-axis. Then the parabola is given by
y = ax 2 . Let the lines AB and CD be given by y = mx + b1 and y = mx + b2 ,
respectively. The x-coordinates of points A and B are the roots of ax 2 = mx + b1 .
By Vieta’s formulas, their sum is m/a. Thus, the x-coordinate of the midpoint of
AB is m/(2a). Similarly, the x-coordinate of the midpoint of CD is also m/(2a).
Therefore, the line going through the midpoints of AB and CD is parallel to the
y-axis, which is the axis of symmetry.
(2) If points G and H lie on the parabola and line GH is perpendicular to the axis of
symmetry, then the axis goes through the midpoint of segment GH .
(3) For the parabola y = ax 2 , the line y = x meets the parabola at the origin and
another point, whose y-coordinate is four times larger than the y-coordinate
of the focus. Note that the line y = x forms a 45◦ angle with the axis of the
parabola.
194 MATHEMATICS MAGAZINE
Proof. The x-coordinates of the intersection points satisfy the equation ax 2 =
x. Therefore, the points of intersection are (0, 0) and (1/a, 1/a). Since the focus
is at (0, 1/(4a)), the result follows.
We will also use the following well-known constructions using compass and straight-
edge:
(a) Construct a line through a given point parallel to a given line.
(b) Construct a line through a given point perpendicular to a given line.
(c) Construct the midpoint of a given line segment.
(d) Given a point that lies on a line, construct a line through the given point that forms
a 45◦ angle with the given line.

We first give the construction of the axis of the parabola. Take any two points A
and B on the given parabola and draw a line through them. Take another point C
on the parabola and draw a second line through C parallel to line AB. Let D be the
other intersection point of this line and the parabola. (If the line through C happens to
be tangent to the parabola, choose another point C.) Next, take points E and F , the
midpoints of line segments AB and CD, respectively, and draw line EF . By Fact 1,
this line is parallel to the axis of symmetry. Next, pick a point G on the parabola. Draw
a line perpendicular to EF through G and call the other intersection point of that line
and the parabola H . (If the line through G happens to be tangent to the parabola,
choose another point G.) Let I be the midpoint of GH . By Fact 2, the line parallel to
EF that goes through I is the parabola’s axis of symmetry.
We now construct the focus of the parabola. The vertex of the parabola is J , the
intersection point of the axis of symmetry and the parabola. Through J , draw a line
that forms a 45◦ angle with the axis of symmetry. Call the second intersection point of
this line and the parabola K. Next, draw a line perpendicular to the axis of symmetry
through K. Call the intersection point of that line and the axis of symmetry L. Con-
struct N, the midpoint of segment J L and M, the midpoint of segment J N. By Fact
3, M is the focus of the parabola.
Also solved by Michel Bataille (France), Bruce S. Burdick, Elton Bojaxhiu (Germany) & Enkel
Hysnelaj (Australia), Micah Fogel, Michael Goldenberg & Mark Kaplan, Shing Hin Jimmy Pa
(China), Randy K. Schwartz, and the proposer.

The limit of a binomial sum April 2022

2143. Proposed by Florin Stănescu, Şerban Cioculescu School, Găeşti, Romania.
VOL. 96, NO. 2, APRIL 2023 195
Evaluate
n n   
(−1)i+j n+i n+j
lim .
n→∞
i=0 j =0
2i + 2j + 1 n − i n − j

Solution by Brian Bradie, Christopher Newport University, Newport News, VA.

Let
n n   
(−1)i+j n+i n+j
Sn = .
i=0 j =0
2i + 2j + 1 n − i n − j

The Chebyshev polynomials of the second kind, Un (x), are given by

n/2  
n−j
Un (x) = (−1) j
(2x)n−2j ,
j =0
j
so
n   n  
2n − j n−j n + j
U2n (x) = (−1) j
(2x) 2n−2j
= (−1) (2x)2j
j =0
j j =0
n − j
n  
n+j
= (−1)n (−1)j (2x)2j .
j =0
n − j

Write
 1
1
= x 2i+2j dx.
2i + 2j + 1 0
Then
 ⎛ ⎞
 n    n  
1
n + i n + j
Sn = (−1)i x 2i ⎝ (−1)j x 2j ⎠ dx
0 i=0
n − i j =0
n − j
 1 x 
= 2
U2n dx.
0 2
With the substitution x = 2 cos θ, we get
 π/2  π/2
sin2 (2n + 1)θ
Sn = 2 2
U2n (cos θ) sin θ dθ = 2 dθ.
π/3 π/3 sin θ
Now,
2n 2n
1
sin θ sin(2j + 1)θ = [cos 2j θ − cos(2j + 2)θ]
j =0
2 j =0
1
= (1 − cos(4n + 2)θ) = sin2 (2n + 1)θ,
2
so
 π/2 2n 2n π/2
cos(2j + 1)θ
Sn = 2 sin(2j + 1)θ dθ = −2
π/3 j =0 j =0
2j + 1 π/3

2n 2n 2n  
cos(2j + 1) π3 cos((2j + 1) arccos 12 ) T2j +1 12
=2 =2 =2 ,
j =0
2j + 1 j =0
2j + 1 j =0
2j + 1
196 MATHEMATICS MAGAZINE
where Tn (x) is a Chebyshev polynomial of the first kind. Next, the generating function
for the Chebyshev polynomials of the first kind is
∞
1 − xt
Tj (x)t j = .
j =0
1 − 2xt + t 2

Separating the j = 0 term from the series, dividing by t, and integrating yields
∞
Tj (x) j 1
t = ln √ ,
j =1
j 1 − 2xt + t 2

from which it follows that

∞  
T2j +1 (x) 1 1 1
= ln √ − ln √
j =0
2j + 1 2 1 − 2xt + t 2 1 + 2tx + t 2 t=1
√
1 2 + 2x 1 1+x
= ln √ = ln .
2 2 − 2x 4 1−x
Finally,
∞  
T2j +1 12 1 1+ 1
1
lim Sn = 2 = ln 2
= ln 3.
n→∞
j =0
2j + 1 2 1− 1
2
2

Also solved by Ulrich Abel & Vitaliy Kushnirevych (Germany), Omran Kouba (Syria), Didier
Pinchon (France), Albert Stadler (Switzerland) Séan M. Stewart (Saudi Arabia) Michael Vowe
(Switzerland) and the proposer.

A ring with distinct ideals having distinct orders April 2022

2144. Proposed by Souvik Dey (graduate student), University of Kansas, Lawrence,
KS.
Let R be a finite commutative ring with unity such that distinct ideals of R have distinct
orders. Show that R is a principal ideal ring.

Solution by the Missouri State University Problem Solving Group, Missouri State Uni-
versity, Springfield, MO.
We will define a finite commutative ring with unity such that distinct ideals have dis-
tinct orders to be distinctive. It is well known that any finite ring R is a direct sum of
finite local rings, that ideals of R correspond to direct sums of ideals of the finite local
rings, and that a direct sum of principal ideals is principal. Clearly, any summand of
a distinctive ring must be distinctive. Therefore, it suffices to prove the result for R
local with maximal ideal m. Suppose, to the contrary, that a, b ∈ m are distinct ele-
ments of a minimal generating set for m, and let I = (a, m2 ) and J = (b, m2 ). Then I
and J are distinct ideals. Both I /m2 and J /m2 are one-dimensional vector spaces over
R/m, implying that I /m2 = |R/m| = J /m2 . Since R is finite, I /m2 = |I | / m2 ,
hence |I | = |R/m| |m2 |. Similarly, |J | = |R/m| |m2 |, which contradicts the fact that
I and J are distinct ideals. Thus, m is principal. It is well known that if the maximal
ideal of a finite local ring R is principal, then R is a principal ideal ring, and the result
follows.
VOL. 96, NO. 2, APRIL 2023 197
We note that if R is a finite principal ideal ring, then for R to be distinctive, it
is necessary that the cardinalities of all its summands are distinct, and it is sufficient
for the cardinalities of its summands to be pairwise relatively prime. The problem of
completely characterizing distinctive rings seems to be complicated.
Also solved by the proposer.

Determine L(L(S)) April 2022

2145. Proposed by the Missouri State University Problem Solving Group, Missouri
State University, Springfield, MO.
Given a set of points S, let L(S) be the set of all points lying on any line connecting two
distinct points in S. For example, if S is the disjoint union of a closed line segment and
a point not lying on the line containing the segment, then L(S) consists of two vertical
angles, their interiors, and the line containing the segment. In this case, L(L(S)) is the
entire plane.
Determine L(L(S)) when S consists of the vertices of a regular tetrahedron.

Solution by José Heber Nieto, Universidad del Zulia, Maracaibo, Venezuela.

Without loss of generality, we may assume that the vertices of the tetrahedron are
S = {A, B, C, D}, where
A = (1, 1, 1), B = (1, −1, −1), C = (−1, 1, −1), and D = (−1, −1, 1).
Let
A = (−1, −1, −1), B  = (−1, 1, 1), C  = (1, −1, 1), and D  = (1, 1, −1).
Note that A , B  , C  , and D  are the reflections of A, B, C, and D through the origin,
which is the centroid of the tetrahedron. We claim that
L(L(S)) = R − {A , B  , C  , D  }.
Clearly, L(S) consists of the lines through the vertices. The points on line AB are of
the form (1, s, s) and those on line CD are of the form (−1, t, −t). Given any point
(x, y, z) with x = ±1, we have (x, y, z) = λ(1, s, s) + (1 − λ)(−1, t, −t), where
x+1 y+z y−z
λ= ,s = , and t = .
2 x+1 1−x
Therefore, L(L(S)) contains all points with x = ±1. Similar arguments using the
other pairs of skew lines shows that L(L(S)) contains all points with y = ±1 and
all points with z = ±1. Hence,
L(L(S)) ⊇ R − {A, B, C, D, A , B  , C  , D  },
but L(L(S)) clearly contains A, B, C, and D, so L(L(S)) ⊇ R − {A , B  , C  , D  }. It
only remains to prove that A , B  , C  , and D  do not belong to L(L(S)). Suppose, on
the contrary, that A is on the line between P and Q with P , Q ∈ L(S). Then P must
be on the line determined by two vertices and Q on the line determined by the other
two (if P and Q were on two intersecting lines A would lie on the plane of a face,
and it does not). For example, suppose that P lies on line AB and Q lies on line CD.
But then the line through Q and A lies in the plane x = −1, which does not meet line
198 MATHEMATICS MAGAZINE
AB, whch lies in the plane x = 1. This gives a contradiction. Examining the other two
pairs of skew lines, shows that A ∈ L(L(S)). Similar arguments show the same for
B  , C  , and D  . Therefore, L(L(S)) = R − {A , B  , C  , D  }, as claimed.
Also solved by Robert Calcaterra, Eugene A. Herman, Didier Pinchon (France), and the pro-
posers. There were three incomplete or incorrect solutions.

Answers
Solutions to the Quickies from page 191.

A1129. The series equals 1.

Let
∞
1 1 1
xn = = 2+ + ··· .
k=n
k2 n (n + 1)2

We have
xn xn xn xn
1
n2
+ xn+1
= − = −
n(n + 1) n n+1 n n+1
xn xn+1 1
= − −
n n + 1 n2 (n + 1)
xn xn+1 1 1
= − − 2+
n n+1 n n(n + 1)
xn xn+1 1 1 1
= − − 2+ − .
n n+1 n n n+1
It follows that
∞ 1
+ 1
+ ··· ∞   ∞ ∞  
n2 (n+1)2 xn xn+1 1 1 1
= − − + −
n=1
n(n + 1) n=1
n n+1 n=1
n2 n=1
n n+1

= x1 − ζ (2) + 1 = 1,
as claimed.
A1130. Note that if gcd(i, n) = 1, then gcd(n − i, n) = 1. Hence,
i n−i
S= =
1≤i<n
n 1≤i<n
n
gcd(i,n)=1 gcd(i,n)=1

and
2S = 1 = φ(n),
1≤i<n
gcd(i,n)=1

where φ(n) is the Euler totient function. Therefore, S = φ(n)/2, and we must solve
φ(n) = 628318. Since 628319 is prime, and since for p an odd prime we have that
φ(p) = p − 1 and φ(2p) = φ(2)φ(p) = p − 1, we can immediately give two solu-
tions to the original equation: n = 628319 and n = 1256638. One readily verifies that,
in fact, these are the only solutions.
90 MATHEMATICS MAGAZINE
Solutions
A formula for ζ (3) February 2022
2136. Proposed by Necdet Batir, Nevşehir HBV University, Nevşehir, Turkey.
Evaluate
  

n
H2 H3
lim k
− n ,
n→∞
k=1
k 3
n
where Hn = 1
k=1 k is the nth harmonic number.

Solution by Kelly D. McLenithan, Los Alamos, NM.

The desired limit is
 n  
 H2 H 3
5
lim k
− n = ζ (3) ,
n→∞
k=1
k 3 3

where ζ (3) is Apéry’s constant given by

∞
1
ζ (3) = = 1.20205 69031 59594 . . . .
k=1
k3

This follows from an application of the summation-by-parts formula


n 
n
(ak+1 − ak )bk = an+1 bn+1 − a1 b1 − ak+1 (bk+1 − bk ) .
k=1 k=1

Letting a1 = 0, ak+1 − ak = 1/k, and bk = Hk2 , we find that ak = Hk−1 and

bk+1 − bk = Hk+1
2
− Hk2
= (Hk+1 − Hk )(Hk+1 + Hk )
 
1 1
= + 2Hk
k+1 k+1
2Hk 1
= + .
k + 1 (k + 1)2
By summation by parts, we have

n 
n  
H2 2Hk 1
k
= Hn Hn2 −0− Hk +
k=1
k k=1
k + 1 (k + 1)2
n
Hk2 
n
Hk
= Hn3 − 2 −
k=1
k + 1 k=1 (k + 1)2

n 2
Hk−1 
n
Hk−1
= Hn3 − 2 −
k=1
k k=1
k2
VOL. 96, NO. 1, FEBRUARY 2023 91
    
1 2  1
n n
1 1
= Hn3 − 2 Hk − − Hk −
k=1
k k k=1
k2 k
n
Hk2 n
Hk n
1 n
Hk  1
n
= Hn3 − 2 +4 − 2 − + .
k=1
k k=1
k2 k=1
k3 k=1
k2 k=1
k3

Collecting terms and rearranging, it follows that


n
H2 Hn3  Hk
n
1 1
n
k
− = 2
− .
k=1
k 3 k=1
k 3 k=1 k 3

After taking the limit, we obtain

 n  
 H2 Hn3 ∞
Hk 1 1
∞
lim k
− = −
n→∞
k=1
k 3 k=1
k2 3 k=1 k 3

∞
Hk 1
= − ζ (3) .
k=1
k2 3

In 1775, Euler showed that for integers q ≥ 2


∞
Hk 
q−2
2 = (q + 2)ζ (q + 1) − ζ (m + 1)ζ (q − m).
k=1
kq m=1

When q = 2, this gives


∞
Hk
= 2ζ (3).
k=1
k2

Therefore, our desired limit is

 n  
 H2 H 3 ∞
Hk 1
lim k
− n = 2
− ζ (3)
n→∞
k=1
k 3 k=1
k 3
1 5
= 2ζ (3) − ζ (3) = ζ (3),
3 3
as claimed.
Also solved by Michel Bataille (France), Jake Boswell & Chip Curtis, Paul Bracken, Brian
Bradie, Bruce S. Burdick, Hongwei Chen, Robert L. Doucette, Russell Gordon, Lixing Han, Eugene
A. Herman, Walther Janous (Austria), Kee-Wai Lau (Hong Kong, China), Shing Hin Jimmy Pa
(Canada), Paolo Perfetti (Italy), Didier Pinchon (France), Albert Stadler (Switzerland), Séan M.
Stewart (Saudi Arabia), and the proposer.

The gcd of terms in a recursive sequence February 2022

2137. Proposed by the Columbus State University Problem Solving Group, Columbus
State University, Columbus, GA.
For a positive integer n, let an and bn be the unique integers such that
√ √
(5 + 3)n = an + bn 3.
92 MATHEMATICS MAGAZINE
√
Find gcd(an , bn )√as a function of n. Solve the analogous problem when 5 + 3 is
replaced by 3 + 5.

Solution by Jacob Boswell and Chip Curtis, Missouri Southern State University,
Joplin, MO.
For the first version of the problem, we claim that
gcd (an , bn ) = 2n/2 .
To see this, set vn = [an , bn ]T . The sequence vn satisfies the recurrence
vn+1 = Mvn , and v0 = [1, 0]T .

5 3
where M = . We note that
1 5
 
28 30 170 234
M2 = and M 3 = .
10 28 78 170
Solving vn+1 = Mvn for an and bn gives
22an = 5an+1 − 3bn+1
22bn = −an+1 + 5bn+1 .
Set dn = gcd (an , bn ). Thus, any factor that divides an+1 and bn+1 must also divide
22dn . Noting that
v1 = [5, 1]T , v2 = [28, 10]T , and M 3 ∼
=M mod 11,
we find that 11 is not a factor of gcd (an , bn ) for any n. From vn+2 = M 2 vn , we see
that 2dn divides dn+2 , but 4dn does not divide dn+2 . Hence, dn+2 = 2dn . Consider the
subsequences of {dn } of even index and odd index separately, and note that
v1 = [5, 1]T and v2 = [28, 10]T ,
so d1 = 1 and d2 = 2. A simple induction completes the proof.
For the second case, we claim that gcd (an , bn ) = 2n−α(n) , where α(n) is 0 if n is a
multiple of 3 and 1 otherwise.
Here
  
3 5 14 30 72 160
M= , M = 2
, and M = 3
.
1 3 6 14 32 72

From vn+3 = M 3 vn , we find that 8dn divides dn+3 , and from

8an = 9an+3 − 20bn+3
8bn = −4an+3 + 9bn+3 ,

obtained by solving vn+3 = M 3 vn for an and bn , we find that dn+3 divides 8dn . Hence,
dn+3 = 8dn . Since v1 = [3, 1]T , v2 = [14, 6]T , and v3 = [72, 32]T , we have d1 = 1,
d2 = 2, and d3 = 8. The claim again follows by induction.
Also solved by Michel Bataille (France), Anthony J. Bevelacqua, Robert Calcaterra, Hongwei
Chen, John Christopher, Rohan Dalal, John Ferdinands, Michael Goldenberg & Mark Kaplan,
Russell Gordon, Eugene A. Herman, Northwestern University Math Problem Solving Group,
Michael Reid, Albert Stadler (Switzerland), and the proposers.
VOL. 96, NO. 1, FEBRUARY 2023 93
Find the locus of the circumcenter February 2022
2138. Proposed by Alexandru Girban, Constanta, Romania.
Let ABC be a triangle with circumcircle ω and let D be a fixed point on side BC. Let
E be a point on ω and let AE meet line BC at F . Find the locus of the circumcenter
of DEF as E varies along ω.
Solution by Michel Bataille, Rouen, France.
In what follows, the line AE will be taken to be the tangent line to ω at A when A = E.
Let the parallel to BC through A intersect ω again at X and let the line XD intersect
ω again at Y (see the figure). We show that the required locus is the perpendicular
bisector of DY with three points removed.
Let ∠( , ) denote the directed angle from line to line .
Let E be a point of ω, with E = X (so that AE does intersect BC). Assuming that
DEF is not degenerate, we have
∠(Y E, Y D) = ∠(Y E, Y X) = ∠(AE, AX) (since A, Y, E, X are concyclic)
= ∠(AF, AX) = ∠(F A, F D) (since F D AX),
hence ∠(Y E, Y D) = ∠(F E, F D). Therefore, Y lies on the circumcircle of DEF .
The circumcenter of DEF is on the perpendicular bisector m of DY , so the locus
we seek is a subset of m.
Conversely, Let U be any point of m and let γ be the circle with center U and radius
U D = U Y . Let BC intersect γ again at F and ω intersect γ again at E. Then U is a
point of the locus if A, E, F are collinear and DEF is not degenerate.
Now, we have
∠(AF, AE) = ∠(AF, AX) + ∠(AX, AE) = ∠(F A, F D) + ∠(Y X, Y E)
= ∠(F A, F D) + ∠(Y D, Y E) = ∠(F A, F D) + ∠(F D, F E)
= ∠(AF, F E).
Therefore, A, E, and F are collinear. Since DEF is degenerate if and only if F =
D, B, or C, the centers P , Q, and R of the circle tangent to BC at D, of the circum-
circle of BDY , and of the circumcircle of CDY (respectively) must be excluded.
Finally, the desired locus is m − {P , Q, R}.
94 MATHEMATICS MAGAZINE
Also solved by Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia), Eugene A. Herman,
Walther Janous (Austria), Albert Stadler (Switzerland), and the proposer.

Infinitely many “very good” Pythagorean triples February 2022

2139. Proposed by Philippe Fondanaiche, Paris, France.
Recall that a Pythagorean triple is a triplet of positive integers (a, b, c) such that a 2 +
b2 = c2 . We say that a Pythagorean triple is good if adding the same single digit to the
front of the decimal representations of a, b, and c yields another Pythagorean triple.
We will call a Pythagorean triple very good if it is good and it is not a nontrivial scalar
multiple of another good Pythagorean triple. For example (50, 120, 130) is good, since
(150, 1120, 1130) is also a Pythagorean triple, but it is not very good since it is a scalar
multiple of the very good triple (5, 12, 13).
Show that there are infinitely many very good Pythagorean triples.

Solution by Michael Reid, University of Central Florida, Orlando, FL.

Let n ≥ 2 be an integer, and put
a = 5 · 10n ,
b = 125 · 102n−2 − 5, and
c = 125 · 10 2n−2
+ 5.
We have
c2 − b2 = (c − b)(c + b)
= (10)(250 · 102n−2 )
= 25 · 102n = a 2 ,
so (a, b, c) is a Pythagorean triple. Let A, B, C be the integers obtained by prepending
the digit 1 to the decimal representations of a, b, c. Then
A = 15 · 10n ,
B = 1125 · 102n−2 − 5, and
C = 1125 · 10 2n−2
+ 5.
Hence,
C 2 − B 2 = (C − B)(C + B) = (10)(2250 · 102n−2 ) = 225 · 102n = A2 ,
so (A, B, C) is a Pythagorean triple. Thus, (a, b, c) is a good Pythagorean triple.
The good Pythagorean triples above are all very good, as we now show. Note that
the only prime divisors of a = 5 · 10n are 2 and 5. Also, b and c are divisible by 5 but
not by 52 . Since n ≥ 2, b and c are odd, so gcd(a, b, c) = 5. Thus, if (a, b, c) is not
very good, it is 5 times a good Pythagorean triple. Let
x = a/5 = 10n ,
y = b/5 = 25 · 102n−2 − 1, and
z = c/5 = 25 · 10 2n−2
+ 1,
and let X, Y, Z be the numbers obtained by prepending the nonzero digit d to x, y, z.
Since x, y, and z have n + 1, 2n, and 2n digits, respectively,
VOL. 96, NO. 1, FEBRUARY 2023 95
X = d · 10 n+1
+ x,
Y = d · 10 + y, 2n
and
Z = d · 10 + z. 2n

The equation X2 + Y 2 = Z 2 yields a quadratic equation in d whose roots are d = 0

and d = −4/25. This contradiction shows that (x, y, z) is not good, so (a, b, c) is
indeed very good.
Also solved by Arya Gupta & Amishi Gupta & Ethan Strubbe, and the proposer.

Minimize the exponential sum February 2022

2140. Proposed by Antonio Garcia, Strasbourg, France.
For a fixed integer n ≥ 2, find the minimum value of
⎛ ⎞

n 
f (x1 , . . . , xn ) = exp xi2 + exp ⎝ −xi xj ⎠ .
i=1 1≤i<j ≤n

Solution by Ulrich Abel and Vitaliy Kushnirevych, Technische Hochschule Mittel-

hessen, Friedberg, Germany.
Application of the AGM inequality

(a1 · · · an )1/n ≤ (a1 + · · · + an ) /n

for positive reals ai , yields
 n 1/n  n 

n  
exp xi2 ≥n exp xi2 = n exp xi2 /n .
i=1 i=1 i=1

By the Cauchy–Schwarz inequality, it follows that

 n 2  n 2
  
n
xi = xi · 1 ≤ n xi2 ,
i=1 i=1 i=1

which implies
 2
 
n 
n 
n
2 xi xj = xi − xi2 ≤ (n − 1) xi2 .
1≤i<j ≤n i=1 i=1 i=1

Combining both inequalities leads to

   
1 2 n−1 2
n n
f (x1 , . . . , xn ) ≥ n exp x + exp − x ,
n i=1 i 2 i=1 i

n(x1 , 2. . . , xn ) ∈ R , and equality holds if and only if x1 = · · · = xn . Putting

n
for all
t = i=1 xi , we want to find the minimum of
   
1 n−1
g (t) = n exp t + exp − t
n 2
96 MATHEMATICS MAGAZINE
for t ≥ 0. We have

  
1 n−1 n−1
g (t) = exp t − exp − t =0
n 2 2
if and only if
  
n−1 1 n−1
exp + t = ,
2 n 2
which occurs at
ln((n − 1)/2)
t0 = .
(n − 1)/2 + 1/n
We also have
     
1 1 n−1 2 n−1
g (t) = exp t + exp − t > 0,
n n 2 2
so g has an absolute minimum at t0 .
For n = 2, we have t0 < 0. Since t is restricted to nonnegative values, the minimum
occurs when t = 0 giving 3 as the minimum value in this case.
For n ≥ 3, we have t0 ≥ 0 and
      2
2 1 n2 − n + 2 n − 1 n2 −n+2
g (t0 ) = n + exp t0 =
n−1 n n−1 2
is the minimum value.
Also solved by Carl Axness (Spain), Jacob Boswell & Chip Curtis, Robert Calcaterra, Hongwei
Chen, Lixing Han, Eugene A. Herman, Kelly D. McLenithan, Michael Reid, Edward Schmeichel,
Albert Stadler (Switzerland), and the proposer. There were two incomplete or incorrect solutions.

Answers
Solutions to the Quickies from page 89.

A1127. Note that

(4 + i)(5 + i)(7 + i)(8 + i)(13 + i) = 11050 + 11050i.

Compare the arguments of the complex numbers on both sides of the equation. The
left-hand side is our sum, and the right-hand side must be π/4 + 2πk for some integer
k. But all the terms in our sum are greater than 0 and less than arctan(1) = π/4.
Therefore, our sum must lie between 0 and 5π/4, and π/4 is the only candidate.
A1128. There is a homomorphism from Z to R obtained by mapping 1 to 1R . The
image of this map is a subring of R, hence an ideal I . Therefore, for all r ∈ R, r =
r · 1R ∈ I , so R = I and the homomorphism is surjective. The kernel of this map is
nZ. By the first isomorphism theorem, R ∼ = Z/nZ.
Note that the condition that R be commutative is unnecessary.
A ring R with trivial multiplication clearly satisfies the condition. A nontrivial
example is R = 2Z/8Z. The only subrings of R are {0}, {0, 4}, and R and these are all
ideals.
VOL. 95, NO. 5, DECEMBER 2022 575

Solutions

A property of the symmedian point December 2021

2131. Proposed by Tran Quang Hung, Hanoi, Vietnam.
Recall that a symmedian is the reflection of a median through a vertex across the angle
bisector passing through that vertex. The three symmedians of a triangle meet in a point
known as the symmedian (or Lemoine or Grebe) point. Let ABC be a triangle with
symmedian point S. Let X, Y , and Z be points lying on segments SA, SB, and SC,
respectively, such that ∠XBA ∼ = ∠Y AB and ∠XCA ∼ = ∠ZAC. Prove that ∠ZBC ∼ =
∠Y CB.

Solution by Do Van Quyet, Vinh Phuc, Vietnam.

Recall that line 1 is said to be anti-parallel to line 2 with respect to lines m1 and m2
if the opposite angles in the quadrilateral formed by the four lines are supplementary.

Let the anti-parallel line to BC with respect to sides AC and AB passing through
X meet those sides at K and L, respectively.
Let the anti-parallel line to AC with respect to sides BA and BC passing through
Y meet those sides at M and N, respectively.
576 MATHEMATICS MAGAZINE
Let the anti-parallel line to AB with respect to sides CB and CA passing through
Z meet those sides at P and Q, respectively.
Note that quadrilaterals BCKL, CAMN, and ABP Q are cyclic.
A key property of a symmedian through a vertex is that it bisects any anti-parallel
to the opposite side with respect to the adjacent sides. Therefore, X, Y , and Z are the
midpoints of segments KL, MN, and P Q, respectively.
We have
∠ALK ∼
= ∠ACB (since BCKL is cyclic),
and
∠ACB ∼
= ∠BMN (since CAMN is cyclic).
Therefore, ∠ALK = ∼ ∠BMN and consequently, ∠BLX ∼
= ∠AMY (supplementary
angles). We are given that ∠XBA ∼
= ∠Y AB, so
AMY ∼ BLX by the AA criterion.

Since X and Y are the midpoints of KL and MN, respectively, we deduce that
AMN ∼ BLK. Therefore, ∠LBK ∼ = ∠MAN. Now
∠MAN ∼
= ∠MCN
since CAMN is cyclic and the angles are subtended by the same arc. Therefore,
∠LBK ∼ = ∠MCN.
A similar argument shows that ∠LCK ∼
= ∠QBP .
We have

∠LBK ∼
= ∠LCK,
since BCKL is cyclic and the angles are subtended by the same arc.
From the three congruences directly above, we obtain ∠MCN ∼ = ∠QBP . Now
∠QP C ∼
= ∠BAC (because CAMN is cyclic)
and

∠BAC ∼
= ∠MNB (because ABP Q is cyclic).
Thus

∠QP C ∼
= ∠MNB, and therefore, ∠BP Q ∼
= ∠MNC (supplementary angles).
Hence,
CMN ∼ BQP by the AA criterion.

Since Y and Z are the midpoints of MN and P Q, respectively, BZP ∼ CY N.

Therefore, ∠ZBC = ∠Y CB, as we wished to show.
Also solved by Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia), Nandan Sai
Dasireddy (India), Michael Goldenberg & Mark Kaplan, Volkhard Schindler (Germany), and the
proposer.
VOL. 95, NO. 5, DECEMBER 2022 577
Buffon’s tetrahedron December 2021
2132. Proposed by the Missouri State University Problem Solving Group, Missouri
State University, Springfield, MO.
A regular tetrahedral die with sides of length 1 is tossed onto a floor having a family of
parallel lines spaced 1 unit apart. What is the probability that the die lands on a line?

Solution by the Eagle Problem Solvers, Georgia Southern University, Statesboro, GA

and Savannah, GA.
Since the tetrahedron is regular, every configuration of the bottom triangular face on
the floor is equally likely. In other words, the probability we seek is the same as the
probability of a randomly tossed equilateral triangle landing on a line. Orient the par-
allel lines horizontally and use the usual cartesian coordinate system. We can give the
vertical coordinate of any point on the floor as a real number in the interval [0, 1),
representing the distance to the closest horizontal line below, or passing through, the
given point. Let y represent the vertical coordinate of the lowest point of the triangular
face. Let θ represent the angle with smallest nonnegative measure between the sides
of the triangle containing the lowest point and the positive x-axis. Then
2π
0≤y<1 and 0≤θ < .
3
Thus, a random toss of the equilateral triangle corresponds to a random selection of a
point (θ, y) from the rectangle
 2π 
0, × [0, 1).
3
If we rotate around a vertex fixed on a horizontal line, then the vertical coordinate of
the highest vertex will be
π  π π 2π
sin + θ for 0 ≤ θ ≤ , and sin θ for ≤ θ < .
3 3 3 3
Thus, the triangle will miss all horizontal lines if and only if
π 
0 < y < 1 − sin +θ
3
for 0 ≤ θ ≤ π
3
and

0 < y < 1 − sin θ

for π3 ≤ θ < 2π3 .
The area of this region in the rectangle is given by
 π/3  π   2π/3
2π
1 − sin +θ dθ + (1 − sin θ) dθ = − 2.
0 3 π/3 3
Thus, the probability that the tetrahedral die misses all lines is
2π
−2 3
3
2π
=1− ,
3
π
578 MATHEMATICS MAGAZINE
and the probability that the die lands on a line is
3
≈ 0.95493.
π
Editor’s Note. Michael Vowe points out that a more general result is known (and has
been rediscovered multiple times): if d is the distance between the lines, and p is
the perimeter of a convex polygon, then the probability the polygon lands on a line is
p/(πd) as long as the diameter of the polygon is less than or equal to d. See: Uspensky,
J. V. (1937). Introduction to Mathematical Probability. New York: McGraw-Hill, pp.
251–255.
Also solved by Jacob Boswell & Chip Curtis, Elton Bojaxhiu (Germany) & Enkel Hysnelaj
(Australia), Owen Byer and the Calculus II class at Eastern Mennonite University, Robert Cal-
caterra, Stephen J. Herschkorn, José Heber Nieto (Venezuela), Didier Pinchon (France), Volkhard
Schindler (Germany), Randy K. Schwartz, Michael Vowe (Switzerland), and the proposers. There
were three incomplete or incorrect solutions.

An infinite series involving the tangent function December 2021

2133. Proposed by Péter Kórus, University of Szeged, Szeged, Hungary.
Evaluate the infinite sum

∞

2−k tan 2−k .
k=1

Solution by Seán M. Stewart, King Abdullah University of Science and Technology,

Thuwal, Saudi Arabia.
Observe that for x ∈ (0, π/2) we have
cot2 (x) − 1 1
2 cot(2x) − cot(x) = 2 · − cot(x) = − = − tan(x).
2 cot(x) cot(x)
Setting x = 2−k in this trigonometric identity and multiplying both sides by 2−k we
obtain
1 1 1 1 1 1
k
tan k =− k
cot k − cot .
2 2 2 2 2k−1 2k−1
Consider the nth partial sum

n

Sn = 2−k tan 2−k .
k=1

From the equation above, we can write this partial sum as


n
1 1 1 1
Sn = k
cot k − cot
k=1
2 2 2k−1 2k−1

= − cot(1) + 2−n cot 2−n
since the sum telescopes. Therefore, the required sum is

∞
 
2−k tan 2−k = lim Sn = − cot(1) + lim 2−n cot 2−n .
n→∞ n→∞
k=1
VOL. 95, NO. 5, DECEMBER 2022 579
Letting u = 2−n , we have
 u
lim 2−n cot 2−n = lim u cot(u) = lim = 1.
n→∞ u→0+ u→0+ tan(u)
Therefore,

∞

2−k tan 2−k = 1 − cot(1).
k=1

Also solved by Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia), Paul Bracken,
Brian Bradie, Robert Calcaterra, Hongwei Chen, CMC 328, Bruce Davis, Prithwijit De (India),
Noah Garson (Canada), Subhankar Gayen (India), G. Greubel, Lixing Han, Mark Kaplan, Kelly
McLenithan, Albert Natian, José Nieto (Venezuela), Northwestern University Math Problem
Solving Group, Shing Hin Jimmy Pa (China), Didier Pinchon (France), Angel Plaza & Francisco
Perdomo (Spain), Michael Reid, Henry Ricardo, Celia Schacht, Volkhard Schindler (Germany),
Vishwesh Ravi Shrimali (India), Albert Stadler (Switzerland), Michael Vowe (Switzerland), and
the proposer. There were three incomplete or incorrect solutions.

Questions about nilpotent matrices December 2021

2134. Proposed by Antonio Garcia, Strasbourg, France.
Let N ∈ Mn (R) be a nilpotent matrix. In what follows, X ∈ Mn (R).
(a) Show that there is always an X such that N = X2 + X − I .
(b) Show that if n is odd, there is no X such that N = X2 + X + I .
(c) Show that if n = 2 and N = 0, there is no X such that N = X2 + X + I .
(d) Give examples, when n = 4, of an N = 0 and an X such that N = X2 + X + I
and of an N with no X such that N = X2 + X + I .

Solution by the Case Western Reserve University Problem Solving Group, Case West-
ern Reserve University, Cleveland, OH.
(a) We claim that if M is a nilpotent matrix, then I + M has a square root. Consider
the formal power series
√ ∞
1/2 i
1+x = x.
i=0
i

If M k = 0, we set x = M and obtain

√ 
k−1
1/2
1+M = Mi.
i=0
i

Returning to the problem, we may rewrite the condition as

 √ √ 2
4 2 5 5
I+ N= X+ I .
5 5 5

Since 45 N is nilpotent, we can solve for X using the claim above.

580 MATHEMATICS MAGAZINE
(b) Assume to the contrary that there exists such an X. Since N is nilpotent,
N k = (X2 + X + I )k = 0
for some k. This implies that (λ2 + λ + 1)k is a polynomial multiple of the min-
imal polynomial of X. Therefore X cannot have any real eigenvalues, since the
eigenvalues of X are the roots of the minimal polynomial, and (λ2 + λ + 1)k has
no real roots. However, n is odd, which guarantees that X has a real eigenvalue.
This is a contradiction.
(c) Suppose there exists such an X. Since we are in dimension two,
N 2 = (X2 + X + I )2 = 0.
This implies that (λ2 + λ + 1)2 is a polynomial multiple of the minimal poly-
nomial of X. Since N = X2 + X + I = 0, (λ2 + λ + 1)2 must be the minimal
polynomial of X. The characteristic polynomial of X must have degree 2, and also
must be a multiple of the minimal polynomial. But the minimal polynomial has
degree 4. This is a contradiction.
(d) Let
⎡ ⎤
−2 −3 −2 −1
⎢1 0 0 0⎥
X=⎣
0⎦
.
0 1 0
0 0 1 0
It is straightforward to verify that the characteristic polynomial of X is
 2
λ4 + 2λ3 + 3λ2 + 2λ + 1 = λ2 + λ + 1 .

Let N = X2 + X + I . One readily verifies that N = 0 and by the Cayley-

Hamilton theorem, N 2 = 0. This solves the first part of the problem.
For the second part of the problem, let
⎡ ⎤
0 1 0 0
⎢0 0 1 0⎥
N =⎣
0 0 0 1⎦
0 0 0 0

and assume there were an X such that N = X2 + X + I . Since N 4 = 0, the min-

imal polynomial of X must be a polynomial multiple of (λ2 + λ + 1). Because
N k = 0 for k < 4, (λ2 + λ + 1)4 must be the minimal polynomial. The character-
istic polynomial of X must be a multiple of the minimal polynomial and also must
have degree 4. But (λ2 + λ + 1)4 has degree 8. This is a contradiction.

Also solved by Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia), Jacob Boswell & Chip
Curtis, Paul Budney, Robert Calcaterra, Lixing Han, Eugene A. Herman, Sonebi Omar (Morroco),
Didier Pinchon (France), Michael Reid, and the proposer.

An exponential generating function December 2021

2135. Proposed by Băetu Ioan, “Mihai Eminescu” National College, Botoşani, Roma-
nia.
For k ∈ Z+ , let an (k) denote the number of elements σ ∈ Sn , the group of all permu-
tations on an n-element set, such that σ k = e, the identity element. We take a0 (k) = 1
VOL. 95, NO. 5, DECEMBER 2022 581
by convention. Find a closed form for the exponential generating function

∞
an (k)x n
fk (x) = .
n=0
n!

Solution by Jacob Boswell and Chip Curtis, Missouri Southern State University,
Joplin, MO.
Let N = {0, 1, 2, . . .}. A permutation σ satisfies σ k = e if and only if all of its disjoint
cycles have lengths which are factors of k. Let k1 , k2 , . . . , kr be the distinct factors
of k. We note that the number of permutations of j k objects that are a product of j
k-cycles is given by (j k)!/k j j !. Breaking permutations with σ k = e into a product
having ji ki -cycles, we see that
 n (j1 k1 )! (jr kr )!
an (k) = · · · jr
j1 k1 , j2 k2 , . . ., jr kr k1j1 j1 ! kr jr !
(ji )∈Nr

ji ki =n

 n!
= j j
,
(ji )∈Nr
k11 · · · kr r · j 1 ! · · · jr !

ji ki =n

where
n n!
=
i1 , i2 , . . . , ir i1 !i2 ! · · · ir !
is a multinomial coefficient. Thus,

⎛ ⎞
∞
⎜  1 ⎟ n
fk (x) = ⎜ ⎟x
⎝ k
j1
· · · k
jr
j ! · · · j ! ⎠
n=0 (j )∈N r 1 r 1 r
i
ji ki =n
⎛ ⎞ ⎛ ⎞

∞
x j1 k1 ∞
x jr kr
=⎝ j1
⎠···⎝
jr
⎠
j1 k1 j1 ! jr kr jr !
 
r
x ki  xd
= exp = exp .
i=1
ki d|k
d

Editor’s Note. Albert Stadler notes that this result appears in an old paper of Chowla,
Herstein, and Scott: Chowla, S., Herstein, I. N., Scott, W. R. (1952). The solutions of
x d = 1 in symmetric groups. Norske Vid. Selsk. 25: 29–31.
Also solved by Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia), CMC 328, Reiner
Martin (Germany), José Heber Nieto (Venezuela), Michael Reid, and the proposer.
VOL. 95, NO. 4, OCTOBER 2022 407

Solutions
Minimize the length of the tangent segment October 2021
2126. Proposed by M. V. Channakeshava, Bengaluru, India.
A tangent line to the ellipse
x2 y2
+ =1
a2 b2
meets the x-axis and y-axis at the points A and B, respectively. Find the minimum
value of AB.

Solution by Kangrae Park (student), Seoul National University, Seoul, Korea.

We may assume that a, b > 0 and that the point of tangency P = (α, β) lies in the first
quadrant. One readily verifies that the tangent line to the ellipse at P is
αx βy
2
+ 2 = 1.
a b
Therefore, A and B are (a 2 /α, 0) and (0, b2 /β), respectively. Note that
α2 β2
+ =1
a2 b2
408 MATHEMATICS MAGAZINE
since the point P is on the ellipse. Applying the Cauchy-Schwarz inequality with
 2 2  
a b α β
u= , and v = , ,
α β a b
we obtain
  
a4 b4 a4 b4 α2 β2
+ = + + = (u · u)(v · v) ≥ (u · v)2 = (a + b)2 .
α2 β2 α2 β2 a2 b2
It follows that

a4 b4
AB = 2
+ 2 ≥ a + b.
α β

This lower bound is attained if and only if u and v are linearly dependent. A straight-
forward calculation shows that this occurs if and only if
a3 b3
α2 = and β 2 = .
a+b a+b
This gives the esthetically pleasing result that when AB attains its minimum value of
a + b, we have P B = a and P A = b.
Also solved by Ulrich Abel & Vitaliy Kushnirevych (Germany), Yagub Aliyev (Azerbaijan),
Michel Bataille (France), Bejmanin Bittner, Khristo Boyadzhiev, Paul Bracken, Brian Bradie,
Robert Calcaterra, Hongwei Chen, Joowon Chung (South Korea), Robert Doucette, Rob Downes,
Eagle Problem Solvers (Georgia Southern University), Habib Y. Far, John Fitch, Dmitry Fleis-
chman, Noah Garson (Canada), Kyle Gatesman, Subhankar Gayen (India), Jan Grzesik, Emmett
Hart, Eugene A. Herman, David Huckaby, Tom Jager, Walther Janous (Austria), Mark Kaplan &
Michael Goldenberg, Kee-Wai Lau (Hong Kong), Lucas Perry & Alexander Perry, Didier Pinchon
(France), Ivan Retamoso, Celia Schacht, Randy Schwartz, Ioannis Sfikas (Greece), Vishwesh Ravi
Shrimali (India), Albert Stadler (Switzerland), Seán M. Stewart (Saudi Arabia), David Stone &
John Hawkins, Nora Thornber, R. S. Tiberio, Michael Vowe (Switzerland), Lienhard Wimmer
(Germany), and the proposer. There were seventeen incomplete or incorrect solutions.

Two idempotent matrices October 2021

2127. Proposed by Jeff Stuart, Pacific Lutheran University, Tacoma, WA and Roger
Horn, Tampa, FL.
Suppose that A, B ∈ Mn×n (C) is such that AB = A and BA = B. Show that
(a) A and B are idempotent and have the same null space.
(b) If 1 ≤ rank A < n, then there are infinitely many choices of B that satisfy the
hypotheses.
(c) A = B if and only if A − I and B − I have the same null space.

Solution by Michel Bataille, Rouen, France.

(a) The fact that A2 = A and B 2 = B follows from:
A2 = (AB)A = A(BA) = AB = A, B 2 = (BA)B = B(AB) = BA = B.

In addition, if X is a column vector and AX = 0, then BAX = 0, that is, BX = 0.

Thus, ker A ⊆ ker B. Similarly, if BX = 0, then ABX = 0. Hence AX = 0 so that
ker B ⊆ ker A. We conclude that ker A = ker B.
VOL. 95, NO. 4, OCTOBER 2022 409
(b) Let r = rank(A). Since A is idempotent, we have range(A) ⊕ ker A = Cn . Since
AX = X if X ∈ range(A) and dim(range(A)) = r, it follows that A = P Jr P −1 for
some invertible n × n matrix P and
 
Ir O
Jr = ,
O O
where Ir denotes the r × r unit matrix and O a null matrix of the appropriate size.
Consider the matrices B = P B P −1 with
 
Ir O
B = ,
C O
where C is an arbitrary (n − r) × r matrix with complex entries. There are infinitely
many such matrices B, and we calculate
AB = P Jr P −1 P B P −1 = P Jr B P −1 = P Jr P −1 = A,
and
BA = P B P −1 P Jr P −1 = P B Jr P −1 = P B P −1 = B.
(c) Clearly, A − I and B − I have the same null space if A = B. Conversely, suppose
that ker(A − I ) = ker(B − I ). Let X be a column vector. Since (A − I )A = O, the
vector AX is in ker(A − I ), hence is in ker(B − I ). This means that (B − I )AX =
0, that is, BX = AX (since BA = B). Since X is arbitrary, we can conclude that
A = B.
Also solved by Paul Budney, Robert Calcaterra, Hongwei Chen, Robert Doucette, Dmitry Fleis-
chman, Kyle Gatesman, Eugene A. Herman, Tom Jager, Rachel McMullan, Thoriq Muhammad
(Indonesia), Didier Pinchon (France), Michael Reid, Randy Schwartz, Omar Sonebi (Morroco),
and the proposer. There was one incomplete or incorrect solution.

Two exponential inequalities October 2021

2128. Proposed by George Stoica, Saint John, NB, Canada.
Let 0 < a < b < 1 and  > 0 be given. Prove the existence of positive integers m and
n such that (1 − bm )n <  and (1 − a m )n > 1 − .

Solution by Robert Doucette, McNeese State University, Lake Charles, LA.

It is well known that
lim (1 − x)1/x = e−1 .
x→0

Suppose 0 < α < 1. Then, since α x → 0+ as x → ∞,

−x
lim (1 − α x )α = e−1 .
x→∞

Hence,

−x −x (β/α)−x 0, if 0 < β < α < 1

lim (1 − α x )β = lim (1 − α x )α = .
x→∞ x→∞ 1, if 0 < α < β < 1
410 MATHEMATICS MAGAZINE
Choose c and d such that 0 < a < c < d < b < 1. Note that c−x − d −x → ∞ as
x → ∞.
By the limits established above, there exists a positive integer m such that
−m −m
(1 − bm )d < , (1 − a m )c > 1 − , and c−m − d −m > 1.
There also exists a positive integer n such that d −m < n < c−m . Therefore,
−m −m
(1 − bm )n < (1 − bm )d <  and (1 − a m )n > (1 − a m )c > 1 − .

Also solved by Levent Batakci, Michel Bataille (France), Elton Bojaxhiu (Germany) & Enkel
Hysnelaj (Australia), Bruce Burdick, Michael Cohen, Dmitry Fleischman, Kyle Gatesman, Michael
Goldenberg & Mark Kaplan, Eugene Herman, Miguel Lerma, Reiner Martin (Germany), Raymond
Mortini (France), Michael Nathanson, Moubinool Omajee (France), Didier Pinchon (France),
Albert Stadler (Switzerland), Omar Sonebi (Morroco), and the proposer.

Two improper integrals October 2021

2129. Proposed by Vincent Coll and Daniel Conus, Lehigh University, Bethlehem, PA
and Lee Whitt, San Diego, CA.
Determine whether the following improper integrals are convergent or divergent.
 1 ∞ 
 k
(a) exp x2 dx
0 k=0
  
1 
∞
3k
(b) exp x dx
0 k=0

Solution by Gerald A. Edgar, Denver, CO.

(a) The integral diverges. For 0 < x < 1 we have
⎛ ⎞
1 ∞
1 n  ⎝  1 n⎠
∞ 2k+1 −1
log = x = x
1−x n=1
n k=0 k
n
n=2
⎛ ⎞

∞  1 k
2k+1 −1 ∞  k   ∞
⎝ 2 ⎠ 2 2k k
≤ k
x = k
x = x2 .
k=0 k
2 k=0
2 k=0
n=2

Therefore,
∞ 
 1
2k
exp x ≥ .
k=0
1−x
1
The integral (a) diverges by comparison with the divergent integral 0
dx/(1 − x).
(b) The integral converges. We will need an estimate for a harmonic sum. The function
1/x is decreasing, so for k ≥ 1
3
k −1  3k
1 dx
> = log 3.
n 3k−1 x
n=3k−1
VOL. 95, NO. 4, OCTOBER 2022 411
Now, for 0 < x < 1 we have
⎛ ⎞
  3
∞ ∞ k −1
1 1 n ⎝ 1 n⎠
log = x = x
1−x n=1
n k=1 k−1
n
n=3

⎛ ⎞
 3
1 ⎠ 3k 
∞ k −1 ∞
> ⎝ x >
k
(log 3)x 3 .
k=1
n k=1
n=3k−1

Let r = 1/ log 3, so that 0 < r < 1. Then

1  k ∞
r log > x3 ,
1−x k=1

1  k ∞
log + 1 > x3 ,
(1 − x)r k=0
∞ 
e  k
> exp x3 .
(1 − x)r k=0

The integral (b) converges by comparison with the convergent integral

 1
e
dx.
0 (1 − x)r

Editor’s Note. A more detailed analysis shows that

 1 ∞ 
 k
exp xα dx
0 k=0

converges if α > e and diverges if 1 ≤ α ≤ e.

Also solved by Michael Bataille (France), Robert Calcaterra, Dmitry Fleischman, Eugene A.
Herman, Walther Janous (Austria), Albert Natian, Moubinool Omarjee (France), Didier Pinchon
(France), Albert Stadler (Switzerland), and the proposers. There was one incomplete or incorrect
solution.

When does the circumcenter lie on the incircle? October 2021

2130. Proposed by Florin Stanescu, Şerban Cioculescu School, Găeşti, Romania.
Given the acute ABC, let D, E, and F be the feet of the altitudes from A, B, and C,
←→ ←→ ←→
respectively. Choose P , R ∈ AB, S, T ∈ BC, Q, U ∈ AC so that
←→ ←
→ ←→ ←→ ← → ← → ← → ← → ← →
D ∈ P Q, E ∈ RS , F ∈ T U and P Q  EF , RS  DF , T U  DE.

Show that
P Q + RS − T U RS + T U − P Q T U + P Q − RS √
+ + =2 2
AB BC AC
if and only if the circumcenter of ABC lies on the incircle of ABC.
412 MATHEMATICS MAGAZINE
Solution by the Fejéntaláltuka Szeged Problem Solving Group, University of Szeged,
Szeged, Hungary.

Let O and I be the circumcenter and the incenter of ABC. Then Euler’s theorem
states that OI 2 = R(R − 2r), where R and r are the circumradius and the inradius of
the triangle, respectively. Now O lies on the incircle if and only if R(R − 2r) = r 2 ,
 2 √
which is equivalent to Rr + 2 Rr − 1 = 0. Therefore, Rr = 2 − 1 since Rr > 0. Since
cos α + cos β + cos γ = 1 + Rr in any triangle, we can reduce the original condition
√
to cos α + cos β + cos γ = 2 where α, β and γ are the angles of ABC.
We have
(1)
DE 2 = CD 2 + CE 2 − 2CD · CE cos γ
(2)
= (CA cos γ )2 + (BC cos γ )2 − 2(CA cos γ )(BC cos γ ) cos γ
(3)
= (CA2 + BC 2 − 2CA · BC cos γ ) cos2 γ = AB 2 cos2 γ ,
where (1) and (3) are the result of the law of cosines applied to CDE and ABC,
respectively, and (2) follows from the fact that CD and CE are altitudes. Since ABC
is acute, cos α > 0, so
DE = AB cos γ , and similarly EF = BC cos α and F D = CA cos β. (1)
Because ∠BF C and ∠BEC are right angles, E and F lie on the circle with diameter
BC, thus BCEF is a cyclic quadrilateral. Hence, m∠EF A = 180◦ − m∠BF E =
m∠ECB = γ and m∠AEF = 180◦ − m∠F EC = m∠CBF = β. We can similarly
see that m∠F DB = m∠CDE = α, m∠DEC = β and m∠BF D = γ . Since P Q 
EF , RS  F D and T U  DE we have
m∠RSB = m∠F DB = α = m∠CDE = m∠CT U,
m∠AQP = m∠AEF = β = m∠DEC = m∠T U C,
m∠BRS = m∠BF D = γ = m∠EF A = m∠QP A.
Therefore, the following triangles are all isosceles (because they all have two congru-
ent angles): DQE, EDS, ERF , F EU , F T D, and DF P . Therefore,
DQ = DE = ES, RE = EF = F U, and T F = F D = P D,
VOL. 95, NO. 4, OCTOBER 2022 413
which (by (1)) leads to
P Q = P D + DQ = F D + DE = CA cos β + AB cos γ ,
RS = RE + ES = EF + DE = BC cos α + AB cos γ ,
T U = T F + F U = F D + EF = CA cos β + BC cos α.
Substituting these into our original statement, we get that
P Q + RS − T U RS + T U − P Q T U + P Q − RS
+ + = 2 (cos γ + cos α + cos β) .
AB BC CA
√
In the first paragraph, we showed that the right side of the last equation equals 2 2 if
and only if the circumcenter lies on the incircle, which is exactly what we wanted to
prove.
Also solved by Michel Bataille (France), Kyle Gatesman, Volkhard Schindler (Germany), Albert
Stadler (Switzerland), and the proposer.

Answers
Solutions to the Quickies from page 407.

A1123. We will need the fact that if f satisfies P2 , then

 
n 1 n 1
f A1 + A2 = f (A1 ) + (A2 ). (1)
n+1 n+1 n+1 n+1
We proceed by induction. When n = 1 this is just condition P2 . Let
n+1 1 1 n+1
X= A1 + A2 and Y = A1 + A2 .
n+2 n+2 n+2 n+2
We have
n 1 1 n
X= A1 + Y and Y = X+ A2 ,
n+1 n+1 n+1 n+1
so, by the induction hypothesis,
n 1 1 n
f (X) = f (A1 ) + f (Y ) and f (Y ) = f (X) + f (A2 ).
n+1 n+1 n+1 n+1
Eliminating f (Y ) gives the desired result.
We will now use induction to show that P2 ⇒ Pn for all n ≥ 2, the case n = 2 being
immediate. Let

1  1
n+1 n
G= Ai and G = Ai .
n + 1 i=1 n i=1

Hence,
n 1
G= G + An+1 .
n+1 n+1
Therefore,
n 1
f (G) = f (G ) + f (An+1 ) (by (1))
n+1 n+1
414 MATHEMATICS MAGAZINE
 
n 1 
n
1
= f (Ai ) + f (An+1 ) (by induction)
n+1 n i=1
n+1

1 
n+1
= f (Ai ),
n+1 i=1

as desired.
To show that Pn ⇒ P2 , let M = (A1 + A2 )/2. Then,
     
1  n
1 n
f M +M + Ai =f A1 + A2 + Ai
n i=3
n i=3
   
1 n
1 n
2f (M) + f (Ai ) = f (A1 ) + f (A2 ) + f (Ai ) (by Pn ),
n i=3
n i=3

so f (M) = (f (A1 ) + f (A2 )) /2 as we wished to show.

A1124. The answer is yes. Note that if 1/Fn < x ≤ 1/Fn−1 with n ≥ 3, then
1 1 1 2 1 1
0<x− ≤ − ≤ − = .
Fn Fn−1 Fn Fn Fn Fn
For y ≤ 1, let g(y) denote the unique positive integer m such that
1 1
<y≤ .
Fm Fm−1
The relation above shows that g(x − 1/Fn ) > n. Now take x1 = 1, n1 = 3 and recur-
sively define
1
xk+1 = xk − and nk+1 = g(xk+1 ).
Fnk
This gives
1 1 1 1 1 1 1
1= + + + + + + + ....
F3 F4 F6 F9 F11 F21 F23
Note that the analogous result holds for any a such that
∞
1
0<a≤ = 3.35988 . . . .
n=1
Fn
244 MATHEMATICS MAGAZINE

Solutions

Evaluate the definite integral June 2021

2121. Proposed by Seán M. Stewart, Bomaderry, Australia.
Evaluate
 1
2 arctan x
dx.
0 x2 −x−1

Solution by Lixing Han, University of Michigan-Flint, Flint, MI and Xinjia Tang,

Changzhou University, Changzhou, China.
Using the substitution
1
−t 1 − 2t
x= 2
= ,
1+ 1
2
t 2+t

we obtain
1
  2 −t
1
2 arctan x arctan 1+
0 1t −5
dx =  
2
· dt
0 x2−x−1 1 1−2t 2
− 1−2t
− 1 (2 + t)2
 1 2+t  1  2+t
2
2 arctan − arctan t
= 2
dt
t −t − 1
2
 1
0
1  1
2 arctan 2 arctan t
= 2
dt − dt.
0 t −t −1 0 t −t −1
2 2

Thus, we have
 1  1
2 arctan x 1 dt 1 2
dx = arctan
0 x2 − x − 1 2 0 t 2 −t −1 2
√  1/2
1 1 2t − 5 − 1
1
= √ ln
arctan √
2 5 2t + 5− 1
2
√ 0
1 1 5+1
= − √ arctan ln √
2 5 2 5 − 1
√
1 1 5+1
= − √ arctan ln .
5 2 2

Also solved by Brian Bradie, Hongwei Chen, Herevé Grandmontagne (France), Eugene A. Her-
man, Omran Kouba (Syria), Kee-Wai Lau (China), Albert Natian, Moobinool Omarjee (France),
Didier Pichon (France), Albert Stadler (Switzerland), Fejéntaláltuka Szöged (Hungary), and the
proposer. There were four incomplete or incorrect solutions.
VOL. 95, NO. 3, JUNE 2022 245
Find the maximum gcd June 2021
2122. Proposed by Ahmad Sabihi, Isfahan, Iran.
Let

G(m, k) = max{gcd((n + 1)m + k, nm + k)|n ∈ N}.

Compute G(2, k) and G(3, k).

Solution by Michael Reid, University of Central Florida, Orlando, FL.

We show that for k ∈ Z, G(2, k) = |4k + 1|, and

27k 2 + 1 if k is even,
G(3, k) = 
27k 2 + 1 /4 if k is odd.

The polynomial identity

(2n + 3)(n2 + k) − (2n − 1)((n + 1)2 + k) = 4k + 1

shows that
gcd((n + 1)2 + k, n2 + k) divides 4k + 1,

and thus is at most |4k + 1|. Hence, G(2, k) ≤ |4k + 1|.

Suppose k > 0, and let n = 2k ∈ N. We have

n2 + k = k(4k + 1) and (n + 1)2 + k = (k + 1)(4k + 1),

both of which are divisible by 4k + 1. Thus

gcd((n + 1)2 + k, n2 + k) = 4k + 1 = |4k + 1|,

so G(2, k) = |4k + 1| in this case.

For k = 0, we have gcd((n + 1)2 , n2 ) = 1 for all n ∈ N, so G(2, 0) = 1 = |4k + 1|
in this case.
Suppose k < 0, and consider n = −(2k + 1) ∈ N. Then
n2 + k = (k + 1)(4k + 1) and (n + 1)2 + k = k(4k + 1)

are each divisible by 4k + 1. Thus

gcd((n + 1)2 + k, n2 + k) = |4k + 1|,

so G(2, k) = |4k + 1| in this case as well.
Now we consider G(3, k). The polynomial identity
(6n2 − 9nk − 3n + 9k + 1)((n + 1)3 + k)
− (6n2 − 9nk + 15n − 18k + 10)(n3 + k) = 27k 2 + 1
shows that

gcd((n + 1)3 + k, n3 + k) divides 27k 2 + 1. (1)

246 MATHEMATICS MAGAZINE
For all n, (n + 1)3 + k and n3 + k have opposite
 parity, so
 their greatest common divi-
sor is odd. If k is odd, then 27k 2 + 1 = 4 (27k 2 + 1)/4 is a product of two integers.
Since the greatest common divisor is odd, and divides this product,
27k 2 + 1
gcd((n + 1)3 + k, n3 + k) divides . (2)
4
For k = 0, we have gcd((n + 1)3 , n3 ) = 1 for all n, so G(3, 0) = 27k 2 + 1 = 1.
For nonzero k, take n = 3k(9k − 1)/2, which is a positive integer. We calculate
(729k 3 − 243k 2 + 8)k
n3 + k = (27k 2 + 1)
8
and
729k 4 − 243k 3 + 162k 2 − 28k + 8
(n + 1)3 + k = (27k 2 + 1) .
8
If k is even, each factor above is an integer, which shows that

27k 2 + 1 divides gcd((n + 1)3 + k, n3 + k).

With (1), we have

gcd((n + 1)3 + k, n3 + k) = 27k 2 + 1,

so G(3, k) = 27k 2 + 1 when k is even.
If k is odd, rewrite the above factorizations as
27k 2 + 1 (729k 3 − 243k 2 + 8)k
n3 + k =
4 2
and
27k 2 + 1 729k 4 − 243k 3 + 162k 2 − 28k + 8
(n + 1)3 + k = ,
4 2
again, all factors being integers. Therefore
27k 2 + 1
divides gcd((n + 1)3 + k, n3 + k).
4
With (2), we conclude that
27k 2 + 1
gcd((n + 1)3 + k, n3 + k) = ,
4
so G(3, k) = (27k 2 + 1)/4 when k is odd.
Also solved by Hongwei Chen, Eagle Problem Solvers (Georgia Southern University), Dmitry
Fleischman, George Washington University Math Problem Solving Group, Eugene A. Her-
man, Walther Janous (Austria), Didier Pinchon (France), Albert Stadler (Switzerland), Enrique
Treviño, and the proposer. There were two incomplete or incorrect solutions.
VOL. 95, NO. 3, JUNE 2022 247
Find the expected winnings June 2021
2123. Proposed by Albert Natian, Los Angeles Valley College, Valley Glen, CA.
An urn contains n balls. Each ball is labeled with exactly one number from the set
{a1 , a2 , . . . , an } , a1 > a2 > · · · > an

(so no two balls have the same number). Balls are randomly selected from the urn
and discarded. At each turn, if the number on the ball drawn was the largest num-
ber remaining in the urn, you win the dollar amount of that ball. Otherwise, you win
nothing. Find the expected value of your total winnings after n draws.

Solution by Enrique Treviño, Lake Forest College, Lake Forest, IL.

Let X be the random variable described. Then X = ai1 + ai2 + · · · + aij with 1 = i1 <
i2 < · · · < ij ≤ n. Therefore, the expected value will be

n
E[X] = ck ak ,
k=1

where ck is the probability that the summand ak appears in X. For ak to appear, the ball
labeled ak must be drawn after those labeled a1 , a2 , . . . , ak−1 , but this only happens if
the permutation of {a1 , . . . , ak } ends in ak . This occurs with probability 1/k. Therefore

1 1 1
E[X] = a1 + a2 + a3 + · · · + an .
2 3 n

Also solved by Robert A. Agnew, Alan E. Berger, Brian Bradie, Elton Bojaxhiu (Germany)
& Enkel Hysnelaj (Australia), Paul Budney, Michael P. Cohen, Eagle Problem Solvers (Geor-
gia Southern University), John Fitch, Dmitry Fleischman, Fresno State Journal Problem Solving
Group, GWstat Problem Solving Group, George Washington University Problems Group, Victoria
Gudkova (student) (Russia), Stephen Herschkorn, Shing Hin Jimmy Pa (Canada), David Huck-
aby, Walther Janous (Austria), Omran Kouba (Syria), Ken Levasseur, Reiner Martin (Germany),
Kelly D. McLenithan, José Nieto (Venezuela), Didier Pinchon (France), Michael Reid, Edward
Schmeichel, Albert Stadler (Switzerland), Fejéntaláltuka Szöged, and the proposer. There were
two incomplete or incorrect solutions.

A sum over the partitions of n June 2021

2124. Proposed by Mircea Merca, University of Craiova, Craiova, Romania.
For a positive integer n, prove that
    λk 
 λ1 λ2
λ λ
··· 0 1
(−1)n−λ1 2λ λ3 = ,
λ1 +λ2 +···+λk =n
1 12 2 · · · k λk n!
λ1 λ2 ···λk >0

where the sum runs over all the partitions of n.

Solution by José Heber Nieto, Universidad del Zulia, Maracaibo, Venezuela.

Put s1 = λ1 − λ2 , s2 = λ2 − λ3 ,. . . , sk−1 = λk−1 − λk , sk = λk . Clearly, we have si ≥
0, s1 + s2 + · · · + sk = λ1 , and s1 + 2s2 + 3s3 + · · · + ksk = n. Moreover, for fixed
λ1 , if we vary k and λ2 , λ3 ,. . . , λk satisfying the conditions λ1 ≥ λ2 ≥ · · · ≥ λk > 0
248 MATHEMATICS MAGAZINE
and λ1 + λ2 + · · · + λk = n, we obtain all the sequences of si ’s satisfying si ≥ 0,
s1 + s2 + · · · + sk = λ1 and s1 + 2s2 + 3s3 + · · · + ksk = n.
Now
λ1 λ2   
λ2 λ3
· · · λ0k λ1 !
= .
1 2 ···k
λ 1 λ 2 λ k s1 !s2 ! · · · sk !(1!)s1 (2!)s2 · · · (k!)sk
We note that
n!
s1 !s2 ! · · · sk !(1!)s1 (2!)s2 · · · (k!)sk
is the number of partitions of the set {1, 2, . . . , n} into si blocks of size i, for i =
1, 2, . . . , k. For fixed λ1 , if we sum these expressions for all values of the si ’s and
k such that si ≥ 0, s1 + s2 + · · · + sk = λ1 and s1 + 2s2 + 3s3 + · · · + ksk = n, we
  of the set {1, 2, . . . , n} into λ1 blocks, that is the Stirling
obtain the number of partitions
number of second kind λn1 . Therefore
    λk   
 λ1 λ2
··· 0 1 
n
n
n−λ1 λ2 λ3
(−1) = (−1) n−λ1
λ1 ! . (1)
λ1 +λ2 +···+λk =n
1λ1 2λ2 · · · k λk n! λ =1 λ1
1
λ1 ≥λ2 ≥···≥λk >0

It is well known that

 n  
n
x(x − 1)(x − 2) · · · (x − λ1 + 1) = x n .
λ =1
λ 1
1

Substituting −x for x we obtain

 n  
n−λ1 n
(−1) x(x + 1)(x + 2) · · · (x + λ1 − 1) = x n .
λ =1
λ 1
1

For x = 1, we have

n  
n
(−1) n−λ1
λ1 ! = 1,
λ1 =1
λ1

hence the right-hand side of (1) is 1/n! and we are done.

Also solved by Albert Stadler (Switzerland) and the proposer.

A graph involving a partition of 100 into ten parts June 2021

2125. Proposed by Freddy Barrera, Colombia Aprendiendo, and Bernardo Recamán,
Universidad Sergio Arboleda, Bogotá, Colombia.
Given a collection of positive integers, not necessarily distinct, a graph is formed as
follows. The vertices are these integers and two vertices are connected if and only if
they have a common divisor greater than 1. Find an assignment of ten positive integers
totaling 100 that results in the graph shown below.
VOL. 95, NO. 3, JUNE 2022 249

Solution by Eagle Problem Solvers, Georgia Southern University, Statesboro, GA and

Savannah, GA.
With the labeling above,
(a, b, c, d, e, f, g, h, i, j ) = (1, 1, 7, 9, 10, 11, 11, 14, 15, 21)
is a solution. Note that each of e, h, i, and j must have at least two prime divi-
sors, since each is adjacent to two vertices that are not adjacent to each other. The
simplest option is e = pq, h = qr, i = rs, and j = ps with p, q, r, and s prime.
Assuming {p, q, r, s} = {2, 3, 5, 7}, the vertices e, h, i, and j must consist of two
of the three pairs (6, 35), (10, 21), and (14, 15). The possibility with the small-
est sum is {e, h, i, j } = {10, 14, 15, 21}. If we take a = b = 1 and f = g = 11,
this forces c + d = 16. Assuming that c and d are powers of distinct primes from
{2, 3, 5, 7}, we must have (c, d) = (7, 9) or (c, d) = (9, 7). The former forces
(e, h, i, j ) = (10, 14, 15, 21), which yields the solution above. The latter gives a
solution with (e, h, i, j ) = (10, 15, 14, 21).
A more detailed analysis shows that, in fact, these are the only solutions.
Also solved by Brian D. Beasley, Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia),
Dmitry Fleischman, George Washington University Problems Group, Kelly D. McLenithan &
Stephen C. Mortenson, Lane Nielsen, José Heber Nieto (Venezuela), Didier Pinchon (France),
Randy K. Schwartz, Albert Stadler (Switzerland), and the proposers.

Answers
Solutions to the Quickies from page 243.

A1121. More generally, we will evaluate


∞
an bn
 ,
n=0
cn an2 + bn2

where an , bn , c, α, and β are real, |c| > 1, and

an + bn i = (α + βi)n .
Note that
an2 + bn2 = (an + bn i)(an − bn i) = (α + βi)n (α − βi)n = (α 2 + β 2 )n ,
and
1 1  
an bn = Im((an + bn i)2 ) = Im (α + βi)2n .
2 2
158 MATHEMATICS MAGAZINE

Solutions

Evaluating an improper integral April 2021

2116. Proposed by Fook Sung Wong, Temasek Polytechnic, Singapore.
Evaluate
 ∞
ecos x cos (αx + sin x)
dx,
0 x2 + β2
where α and β are positive real numbers.

Solution by Omran Kouba, Higher Institute for Applied Sciences and Technology,
Damascus, Syria.
π
We claim the answer is exp(e−β − αβ).
2β
VOL. 95, NO. 2, APRIL 2022 159
Consider the meromorphic function
g(z)
F (z) = , where g(z) = exp(eiz + iαz).
z2 + β2
If z = x + iy with x, y ∈ R and y ≥ 0, then
     
|g(z)| = exp Re(eiz + iαz) = exp e−y cos(x) − αy ≤ exp e−y − αy ≤ e.

For R > β, consider the closed contour R consisting of the line segment [−R, R]
followed by the semicircle γR parametrized by θ → Reiθ for θ ∈ [0, π]. The only
singularity that F has inside the domain bounded by R is a simple pole at z = iβ
with residue
g(iβ) exp(e−β − αβ)
Res (F, iβ) = = .
2iβ 2iβ
By the residue theorem we have

π
F (z)dz = 2iπ Res (F, iβ) = exp(e−β − αβ).
R β
But
  R 
F (z)dz = F (x)dx + F (z)dz
R −R γR
 R
ecos x cos(αx + sin x)
=2 dx + R,
0 x2 + β2
where

R = F (z)dz.
γR

Since R > β, we have

R
| R | ≤ πR sup |F (Reiθ )| ≤ πe .
θ∈[0,π] R2 − β 2
Thus lim R = 0. Therefore
R→∞
 ∞
ecos x cos(αx + sin x) π
2 dx = exp(e−β − αβ),
0 x2 + β2 β
as claimed.

Also solved by Khristo N. Boyadzhiev, Hongwei Chen, John Fitch, G. C. Greubel, Eugene A.
Herman, Rafe Jones, Kee-Wai Lau (Hong Kong), Kelly D. McLenithan, Raymond Mortini (France)
& Rudolf Rupp (Germany), Moubinool Omarjee (France), Didier Pinchon (France), Ahmad Sabihi
(Iran), Albert Stadler (Switzerland), Seán M. Stewart (Australia), and the proposer. There were
two incomplete or incorrect solutions.
160 MATHEMATICS MAGAZINE
A factorial Diophantine equation April 2021
2117. Proposed by Ahmad Sabihi, Isfahan, Iran.
Find all positive integer solutions to the equation

(m + 1)n = m! + 1.

Solution by Michael Kardos (student), East Carolina University, Greenville, NC.

Note that for any solution with m ≥ 2, we have m even. This follows from the fact that
m! is even for m ≥ 2 and so m! + 1 = (m + 1)n is odd. Thus (m + 1) must be odd
and m even.
We can reduce the pool of possible solutions by showing that m ≤ 4. Clearly a
solution with m > 4 and m even implies 2 < m/2 < m, so
m n  
n k−1
2 m m! ⇒ m2 | ((m + 1)n − 1) ⇒ m2 m m .
2 k=1
k

Thus
n  
n k−2
m n+m m ,
k=2
k

so m divides n and hence n ≥ m.

We will now show that n ≥ m and m > 4 yields no solutions. In that case,

m! + 1 < mm−1 + 1 < (m + 1)mm−1 < (m + 1)m ≤ (m + 1)n .

Thus, there are no positive integer solutions with m > 4. Now we can find all solutions
using the previously gathered information about m. For each possible m we have the
following.

m = 1 ⇒ 2n = 2, so n = 1,
m = 2 ⇒ 3n = 3, so n = 1,
m = 4 ⇒ 5n = 25, so n = 2.

Also solved by John Christopher, Michael P. Cohen, Charles Curtis & Jacob Boswell, Eagle
Problem Solvers (Georgia Southern University), John Fitch, Khaled Halaoua (Syria), Walther
Janous (Austria), Rafe Jones, Koopa Tak Lun Koo (Hong Kong), Seungheon Lee (South Korea),
Graham Lord, Kelly D. McLenithan, Stephen Meskin, Raymond Mortini (France) & Rudolf
Rupp (Germany) & Amol Sasane (UK), Sonebi Omar (Morocco), Didier Pinchon (France), Henry
Ricardo, Celia Schacht, Albert Stadler (Switzerland), Wong Fook Sung (Singapore), and the pro-
poser. There were two incomplete or incorrect solutions.

Does the series converge or diverge? April 2021

2118. Proposed by Moubinool Omarjee, Lycée Henri IV, Paris, France.
It is well known that the series

∞
sin k
k=1
k
VOL. 95, NO. 2, APRIL 2022 161
converges. Does the series

∞
e− ln k sin k
k=1

converge or diverge?

Solution by the Northwestern University Math Problem Solving Group, Northwestern

University, Evanston, IL.
Both series can be shown to be convergent using the following well-known result.

Dirichlet’s test: If an is a monotonic sequence of real numbers that tends to zero,

N
and bn is a sequence of complex numbers such that, for some M, bn ≤ M for
n=1

∞
every positive integer N, then the series an bn converges.
n=1
For this problem, we use ak = e− ln k , which tends monotonically to zero, and bk =
sin k, whose partial sums are


N N  
1 ik −ik 1 ei(N+1) − ei e−i(N+1) − e−i
sin k = (e − e ) = −
k=1 k=1
2i 2i ei − 1 e−i − 1
 
1 ei(N+1/2) − ei/2 + e−i(N+1/2) − e−i/2
=
2i ei/2 − e−i/2
   
1 ei/2 + e−i/2 /2 − ei(N+1/2) + e−i(N+1/2) /2
=  
2 ei/2 − e−i/2 /(2i)
1  
= 1
cos 12 − cos (N + 12 )
2 sin 2
hence

N
1
sin k ≤
k=1
sin 12

So, by Dirichlet’s test, the series converges.

Also solved by Hongwei Chen, Richard Daquila, Eagle Problem Solvers (Georgia Southern
University), John Fitch, Russell Gordon, Eugene A. Herman, Walther Janous (Austria), Mark
Kaplan & Michael Goldenberg, Raymond Mortini (France), Didier Pinchon (France), Omar Sonebi
(Morocco), Albert Stadler (Switzerland), Seán Stewart (Australia), and the proposer. There were
three incomplete or incorrect solutions.

Find the side length of the regular n-simplex April 2021

2119. Proposed by Viktors Berstis, Portland, OR.
A point in the plane is a distance of a, b, and c units from the vertices of an equilateral
triangle in the plane. Denote the side length of the equilateral triangle by s.
(a) Find a polynomial relation between a, b, c, and s.
 162 MATHEMATICS MAGAZINE
(b) Give a simple compass and straightedge construction of a segment of length s given
segments of lengths a, b, and c.
(c) Generalize part (a) to the case of a point at a distance of ai units, i = 1, . . . , n + 1,
from the vertices of a regular n-dimensional simplex having sides of length s.

Solution by Didier Pinchon, Toulouse, France.

(a) Given s > 0, the points
 √   √   √ 
3 1 3 1 3
A = 0, s , B = − s, − s , and C = s, − s
3 2 6 2 6
are the vertices of an equilateral triangle with side length s. For a point P = (x, y),
the relations P A2 = a 2 , P B 2 = b2 and P C 2 = c2 give the equations
 √ 2
3
E1 : x + y −
2
s − a 2 = 0,
3
   √ 2
1 2 3
E2 : x + s + y + s − b2 = 0,
2 6
   √ 2
1 2 3
E3 : x − s + y + s − c2 = 0.
2 6

From E2 − E √3 , we2 get 2x = (b2 − c )/(2s), and substituting this value into E1 − E2 ,
2 2

we get y = 3(b + c − 2a )/(6s). Finally, substituting these values into equation

E1 , we obtain

s 4 − (a 2 + b2 + c2 )s 2 + a 4 + b4 + c4 − a 2 b2 − a 2 c2 − b2 c2 = 0.
(b) Note that s 2 satisfies a second-degree polynomial with discriminant

= (a 2 + b2 + c2 )2 − 4(a 4 + b4 + c4 − a 2 b2 − a 2 c2 − b2 c2 )
= 3(a + b + c)(a + b − c)(a + c − b)(b + c − a).

Given positive real numbers a, b and c, ≥ 0 if and only if c ≤ a + b, b ≤ a + c and

a ≤ b + c. Indeed, when two factors of are negative, say for example a + c ≤ b,
b + c ≤ a, then c ≤ 0, which is impossible. Hence, ≥ 0 if and only a, b, and c are
the lengths of the sides of a triangle. Note that the triangle is degenerate if and only if
= 0. If a, b, and c are not all equal, then
1 
a 4 + b4 + c4 − a 2 b2 − a 2 c2 − b2 c2 = (a 2 − b2 )2 + (b2 − c2 )2 + (c2 − a 2 )2 > 0.
2
Therefore, the equation in s 2 has two different positive solutions, denoted by s1 and s2 ,
if > 0 and a, b, and c are not all equal, and one positive solution otherwise.
The two solutions will now be constructed using a compass and straightedge. It
is straightforward to construct a triangle ABC with side lengths a, b and c. The two
circles of centers B and C and radius a intersect in two points D and E such that the
triangles BCD and BCE are equilateral, with D and A being on opposite sides of line
BC. Because ABC is not an equilateral triangle, point E is distinct from A.
We claim the lengths of the segments AD and AE are the solutions s1 and s2 . The
images of the points B and A by the rotation of center D and angle −π/3 are C and
VOL. 95, NO. 2, APRIL 2022 163
F , and thus BA = CF = c. In a similar way, the images of A and B by the rotation
of center E and angle π/3 are C and G, and thus AB = c = CG.
When a = b = c, then√the first part of the construction is possible, and the unique
solution is s = DA = a 3, and C is the center of equilateral triangle ADF . When
= 0, A, B, and C are collinear, so A is equidistant from D and E and there is only
one solution.

(c) Editor’s Note. The solver uses the fact that if the distance between the ith and
j th vertices of an n-simplex is di,j , then the volume of the simplex is
0 1 1 1 ... 1
2 2 2
1 0 d1,2 d1,3 ... d1,n+1
2 2 2
(−1)n+1 1 d1,2 0 d2,3 ... d2,n+1
V2 = .. .. .. .. .
2n (n!)2 . . . .
2 2 2
1 d1,n d2,n 0 dn,n+1
2 2 2
1 d1,n+1 d2,n+1 ... dn,n+1 0
He applies this formula to the degenerate (n + 1)-simplex whose vertices are the ver-
tices of the regular n-simplex along with the additional point and performs a series of
row and column operations to derive the result.
Here is an alternative derivation. Let the vertices of the regular n-simplex be
√ √ √
(s/ 2, 0, 0, . . . , 0), (0, s/ 2, 0, . . . , 0), . . . , (0, 0, . . . , s/ 2)
in Rn+1 . Note that these vertices lie in the hyperplane whose equation is

n+1
s
xi = √ .
i=1 2
Let (x1 , . . . , xn+1 ) be a point in this hyperplane. We have
√ s2  2
n+1
ai2 = − 2sxi + + xi . (1)
2 i=1

Expanding and summing these equations as i = 1, . . . , n + 1, we obtain


n+1
√  n+1
s2 
n+1
ai2 = − 2s xi + (n + 1) + (n + 1) xi2
i=1 i=1
2 i=1
164 MATHEMATICS MAGAZINE
  
√ s s2
n+1
= − 2s √ + (n + 1) + (n + 1) xi2
2 2 i=1

s2  n+1
= (n − 1) + (n + 1) xi2 .
2 i=1

Therefore

n+1
n−1 2 1  2
n+1
xi2 =− s + a . (2)
i=1
2(n + 1) n + 1 i=1 i

Substituting into (1), we have

√ 1  2
n+1
s2 n−1 2
ai2 = − 2sxi + − s + a
2 2(n + 1) n + 1 i=1 i

√ 1  2
n+1
1 2
= − 2sxi + s + a .
n+1 n + 1 i=1 i

Solving for xi , we find that

1  2
n+1
1 1 2
xi = √ −ai2 + s + a .
2s n+1 n + 1 i=1 i

Substituting into (2) and letting s = a0 , we have

2
1  2 1  1  2
n+1 n+1 n+1
n−1 2
− a0 + ai = 2 −ai2 + a
2(n + 1) n + 1 i=1 2a0 i=1 n + 1 i=0 i
2
1  2  1  2
n+1 n+1 n+1
1
2a02 − a02 + a = −ai2 + a .
2 n + 1 i=0 i i=1
n + 1 i=0 i
n+1
Letting T = i=0 ai2 /(n + 1), we have


n+1 
n+1
−a04 + 2a02 T = ai4 − 2T ai2 + (n + 1)T 2
i=1 i=1


n+1 
n+1
0= ai4 − 2T ai2 + (n + 1)T 2
i=0 i=0


n+1
0= ai4 − 2T (n + 1)T + (n + 1)T 2
i=0
2

n+1
1 
n+1
ai4 = (n + 1)T = 2
a2
i=0
n + 1 i=0 i


n+1 
n+1 
(n + 1) ai4 = ai4 + 2 ai2 aj2
i=0 i=0 0≤i<j ≤n+1
VOL. 95, NO. 2, APRIL 2022 165

n+1 
n ai4 = 2 ai2 aj2 ,
i=0 0≤i<j ≤n+1

which is the desired relation.

Also solved by Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia), Albert Stadler
(Switzerland), and the proposer. There were two incomplete or incorrect solutions.

Find the normalizer April 2021

2120. Proposed by Gregory Dresden, Jackson Gazin (student), and Kathleen McNeill
(student), Washington & Lee University, Lexington, VA.
Recall that the normalizer of a subgroup H of G is defined as
 
NG (H ) = g ∈ G|ghg −1 ∈ H for all h ∈ H .
Determine NG (H ), when G = GL2 (R), the group of all invertible 2 × 2 matrices with
real entries, and
  
cos θ − sin θ
H = SO2 (R) = θ ∈R .
sin θ cos θ

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

More generally, for any n ≥ 1, let G = GLn (R) and H = SOn (R), the subgroup of
On (R), the group of orthogonal matrices, consisting of matrices whose determinant is
1. We will show that
NG (H ) = {aU |a ∈ R − {0}, U ∈ On (R) } .

Suppose A = aU , where a  = 0 and U is orthogonal. Then for any M ∈ SOn (R),

1 −1
AMA−1 = aU M U = U MU −1 .
a
Since

det(U MU −1 ) = det(U ) det(M)/ det(U ) = 1,

and the product of orthogonal matrices is orthogonal, we see that AMA−1 ∈ SOn (R).
For the converse, we use a polar decomposition. For A ∈ NG (H ), write A = P U ,
where P is positive-definite and U is orthogonal. For any M ∈ SOn (R), let N =
U −1 MU . Then N ∈ SOn (R), so ANA−1 ∈ SOn (R). But
ANA−1 = P (U NU −1 )P −1 = P MP −1 ,

so P ∈ NG (H ). Therefore, it remains only to determine which positive-definite

matrices are in the normalizer. Now every positive-definite matrix can be written
as P = V DV −1 , where D = diag(d1 , . . . , dn ) is a diagonal matrix with di > 0 and
V ∈ On (R). For any M ∈ SOn (R), let N = V MV −1 . Then B = P NP −1 ∈ SOn (R)
and

B = V DMD −1 V −1 ∈ SOn (R), so DMD −1 = V −1 BV ∈ SOn (R).

Therefore, D ∈ NG (H ).
166 MATHEMATICS MAGAZINE
For k > 1, let Mk = [mij ], where
m11 = 0, m1k = −1, mk1 = 1, mkk = 0, mii = 1 (i  = 1, k), and mi,j = 0 otherwise.
Then R ∈ SOn (R) and the first column of DRD −1 consists of zeros except the kth
entry, which is dk /d1 . Since DRD −1 is orthogonal, this column must have length 1,
which means that dk = d1 for all k > 1. Therefore D is a positive multiple of the
identity, and so A is a multiple of an orthogonal matrix.
Note: The same proof works for the complex version. In that case, G = GLn (C)
and H = SUn (C), where the latter is the group of n × n unitary matrices whose deter-
minant equals 1. Then NG (H ) is the group of all nonzero complex multiples of n × n
unitary matrices.
Also solved by Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia), Robert Calcaterra,
Eagle Problem Solvers (Georgia Southern University), John Fitch, Dmitry Fleischman, Mark
Kaplan & Michael Goldenberg, Koopa Tak Lun Koo (Hong Kong), Didier Pinchon (France),
Albert Stadler (Switzerland) and the proposers. There were two incomplete or incorrect solutions.

Answers
Solutions to the Quickies from page 158.

A1119. The aces divide the 48 other cards into 5 “urns”, with a, b, c, d, and e non-
aces in them, respectively. The position of the third ace is equal to a + b + c + 3, so
the expected value of its position is E[a + b + c + 3]. By linearity of expectation, this
is E[a] + E[b] + E[c] + 3. Because a non-ace is equally likely to be placed in any
of the five “urns”, E[a] = . . . = E[e]. Since E[a + b + c + d + e] = 48, we have
E[a] = . . . = E[e] = 485 .
Therefore the expected value is
48 159
3· +3= .
5 5

A1120. Let S, S1 , S2 , and S3 be the areas of ABC, XBC, XCA, and XAB,
respectively. Let h2 and h3 be the heights of XCA and XAB with AX as base. Let
←→ ← →
θ be the angle between AX and BC. Then
1 1 1
S2 + S3 = (h2 + h3 ) R1 = a sin θR1 ≤ aR1 .
2 2 2
Similar arguments give
1 1
S 3 + S1 ≤ bR2 and S1 + S2 ≤ cR3 .
2 2
Therefore
1 1 1
aR1 + bR2 + cR3 ≥ (S2 + S3 ) + (S3 + S1 ) + (S1 + S2 ) = 2S = r(a + b + c)
2 2 2
and the result follows.
Equality holds if and only if the line through a vertex and X and the line containing
the side opposite the vertex are perpendicular. In other words, X must be the ortho-
center of the triangle, which must be acute in order for X to lie in its interior.

Note. Let O and R be the circumcenter and the circumradius for a given acute
triangle. Since R1 = R2 = R3 = R, we obtain Euler’s inequality R ≥ 2r.
VOL. 95, NO. 1, FEBRUARY 2022 75
Solutions
A series involving central binomial coefficients December 2020
2111. Proposed by Enrique Treviño, Lake Forest College, Lake Forest, IL.
Evaluate
4n

∞
2n
.
n=0
42n (2n + 1)(2n + 2)

Solution by Hongwei Chen,√ Christopher Newport University, Newport News, VA.

The value of the series is 43 ( 2 − 1). To this end, recall the generating function for the
central binomial coefficients
∞ 
2n n 1
x =√ , for |x| < 14 .
n=0
n 1 − 4x

Replacing x by −x gives

∞ 
2n n 1
(−1)n x =√ , for |x| < 14 .
n=0
n 1 + 4x

Adding these series gives

 ∞  
4n 2n 1 1 1
x = √ +√ .
n=0
2n 2 1 − 4x 1 + 4x

Replacing x by x/4 yields

∞ 4n 
1 1 1
2n
x 2n
= √ +√ , for |x| < 1.
n=0
4 2n 2 1−x 1+x

Integrating this series on [0, x] with 0 < x < 1, we find

4n
∞
√ √
2n
2n (2n + 1)
x 2n+1
= 1 + x − 1 − x.
n=0
4

Integrating this series on [0, x] with 0 < x < 1 again, we find

4n
 ∞ x √ √
2n
x 2n+2
= ( 1 + t − 1 − t)dt
n=0
42n (2n + 1)(2n + 2) 0

2  4
= (1 + x)3/2 + (1 − x)3/2 − .
3 3
Applying Abel’s convergence theorem and letting x → 1, we conclude
4n
∞
4 √
2n
= ( 2 − 1)
n=0
42n (2n + 1)(2n + 2) 3

as claimed.
76 MATHEMATICS MAGAZINE
Also solved by Ulrich Abel & Vitaliy Kushnirevych (Germany), Farrukh Ataev (Uzbekistan),
Michel Bataille (France), Khristo Boyadzhiev, Paul Bracken, Brian Bradie, Cal Poly Pomona
Problem Solving Group, Robert Doucette, Gerald Edgar, Dmitry Fleischman, Mohit Hulse (India),
Dixon Jones & Marty Getz, Mark Kaplan & Michael Goldenberg, GWstat Problem Solving Group,
Omran Kouba (Syria), Sushanth Sathish Kumar, Elias Lampakis (Greece), Kee-Wai Lau (China),
James Magliano, Northwestern University Math Problem Solving Group, Moubinool Omarjee
(France), Shing Hin Jimmy Pa (Canada), Angel Plaza (Spain), Rob Pratt, Volkhard Schindler
(Germany), Edward Schmeichel, Randy Schwartz, Albert Stadler (Switzerland), Seán M. Stewart
(Australia), Ibrahim Suleiman (United Arab Emirates), Michael Vowe (Switzerland), and the
proposer. There were two incomplete or incorrect solutions.

A problem from commutative algebra December 2020

2112. Proposed by Souvik Dey, (graduate student), University of Kansas, Lawrence,
KS.
Let R be an integral domain and I and J be two ideals of R such that I J is a non-zero
principal ideal. Prove that I and J are finitely-generated ideals.

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

Let I J = x , where x is a nonzero element of R. Since x ∈ I J , there exist
i1 , . . . , in ∈ I and j1 , . . . , jn ∈ J
such that
x = i 1 j 1 + · · · in j n .
We claim that
I = i1 , . . . , in and J = j1 , . . . , jn .
In each of these two equations, it suffices to prove that the left side is contained in the
right. For any i ∈ I , there exist r1 , . . . , rn ∈ R such that
ijk = rk x, k = 1, . . . , n.
Multiply the kth equation by ik and add the the resulting equations to obtain

n
ix = rk ik x
k=1

Since R is an integral domain,


n
i= rk ik ,
k=1

and so I = i1 , . . . , in . Similarly, J = j1 , . . . , jn .
Also solved by Paul Budney, Noah Garson (Canada), Elias Lampakis (Greece), and the pro-
poser.

A condition for the nilpotency of a matrix December 2020

2113. Proposed by George Stoica, Saint John, NB, Canada.
Let A be an n × n complex matrix such that det(Ak + In ) = 1 for k = 1, 2, . . . , 2n − 1.
 VOL. 95, NO. 1, FEBRUARY 2022 77
(a) Prove that An = On .
(b) Show that the result does not hold if 2n − 1 is replaced by any smaller positive
integer.

Solution by Michael Reid, University of Central Florida, Orlando, FL.

(a) First we have a lemma.
Lemma. Suppose z1 , . . . , zm ∈ C are such that the power sums Sk = z1k + · · · + zm
k

vanish for k = 1, 2, . . . , m. Then each zj = 0.

Proof. For k = 1, 2, . . . , m, let σk denote the kth elementary symmetric function of
z1 , . . . , zm . By Newton’s identities,


k−1
kσk = (−1)k−1 Sk + (−1)i−1 σk−i Si = 0,
i=1

for k = 1, 2, . . . , m. Hence each σk = 0. Therefore,

(T + z1 )(T + z2 ) · · · (T + zm ) = T m + σ1 T m−1 + · · · + σm−1 T + σm = T m .

By unique factorization of polynomials, each factor, T + zj , on the left is a constant

multiple of T , so each zj = 0. 
The matrix A is similar to an upper triangular matrix M (for example, take M to
be a Jordan canonical form of A). Let d1 , . . . , dn be the diagonal entries of M. Then
Ak + In is similar to M k + In , which is an upper triangular matrix with diagonal entries
d1k + 1, . . . , dnk + 1, so
det(Ak + In ) = (d1k + 1) · · · (dnk + 1).

For each subset S ⊆ {1, 2, . . . , n}, let bS = j ∈S dj . Thus

n 
1 = det(Ak + In ) = (djk + 1) = bSk .
j =1 S⊆{1,2,...,n}

Let m = 2n − 1, and let S1 , . . . , Sm be the non-empty subsets of {1, 2, . . . , n}, and

put zj = bSj . Then, the equation above becomes z1k + · · · + zm k
= 0, which holds for
k = 1, 2, . . . , m. From the lemma, each zj = 0. In particular, for a singleton subset
{i}, we have di = b{i} = 0. Hence M is upper triangular, with all zeros on its diagonal,
so its characteristic polynomial is T n . Since A is similar to M, it has the same charac-
teristic polynomial, so by the Cayley–Hamilton theorem, An = 0n .

(b) Let m = 2n − 1, and let ζ ∈ C be a primitive mth root of 1. Let A be the diagonal
n−1
matrix with diagonal entries ζ, ζ 2 , ζ 4 , . . . , ζ 2 . Then, A is non-singular, so it is not
nilpotent. For k ∈ N, Ak + In is the diagonal matrix whose diagonal is
n−1 k
ζ k + 1, ζ 2k + 1, ζ 4k + 1, . . . , ζ 2 + 1.

Thus,
n−1 k
det(Ak + In ) = (ζ k + 1)(ζ 2k + 1) · · · (ζ 2 + 1).
78 MATHEMATICS MAGAZINE
If k is not divisible by m, then this product telescopes to give
n−1 j +1 n
ζ2 k − 1 ζ2 k − 1
det(Ak + In ) = = = 1,
j =0
ζ2 k − 1
j
ζk − 1
nk
because ζ 2 = ζ mk+k = ζ k . Hence,
det(Ak + In ) = 1

for k = 1, 2, . . . , 2n − 2.
Also solved by Lixing Han & Xinjia Tang, Koopa Tak Lan Koo (Hong Kong), Elias Lampakis
(Greece), Albert Stadler (Switzerland), and the proposer. There were two incomplete or incorrect
solutions.

Planar 2-distance sets having four points December 2020

2114. Proposed by Robert Haas, Cleveland Heights, OH.
Find all configurations of four points in the plane (up to similarity) such that the set of
distances between the points consists of exactly two lengths.

Solution by Robert L. Doucette, McNeese State University, Lake Charles, LA.

Suppose A, B, C, D are distinct points in the plane such that the list of six segment
distances, AB, AC, AD, BC, BD, and CD, has exactly two real values. For conve-
nience, we may suppose that one of these values is 1. We consider three cases.
Case 1. Exactly five of the six distances equal 1. Suppose AB = AC = AD =
BC = BD = 1, CD = 1. In this case, ABC and ABD must form equilateral
triangles. Since C = D, ADBC must form a rhombus with side length 1 and one
◦
√ 60 . This yields a configuration in which the dis-
pair of opposite angles measuring
tance not equal to 1 is CD = 3.

Case 2. Exactly four of the six distances equal 1. There are two subcases to consider.
(i) Suppose first that the two segments with length not equal to 1 do not have an
endpoint in common. Say AC = BD = 1. Since ABCD is a rhombus with congruent
diagonals, it must be a square. √ This yields a configuration in which the distances not
equal to 1 are AC = BD = 2.
(ii) Suppose next that the two segments with length not equal to 1 do share an
endpoint. Say BD = CD = 1. In this case, ABC forms an equilateral triangle of side
length 1. The point D must lie on the perpendicular bisector of segment BC. Either D
← →
lies on the same side of BC as A or on the opposite side. In the former case, BCD
√
is a 30◦ -75◦ -75◦ triangle, A is its circumcenter, and BD = CD = 2 + 3.
VOL. 95, NO. 1, FEBRUARY 2022 79

In the latter case, ABDC is a kite with opposite angles of measure 60◦ and 150◦ , and
√
BD = CD = 2 − 3.

Case 3. Exactly three of the six distances equal 1. Again we consider two subcases.
(i) Suppose that three of the segments of equal length have an endpoint in common.
We may assume that AB = AC = AD = 1 and BC = BD = CD = 1. In this case,
the points B, C and D lie on the circle with center A and radius 1 and form an equi-
lateral triangle. In other words, BCD√forms an equilateral triangle with circumcenter
A. In this case, BC = BD = CD = 3.

(ii) Next suppose that no three of the segments of equal length share a common
endpoint. We may assume that AB = AC = BD = 1 and AD = BC = CD = x >
1. Since ABC ∼ = BAD, ∠BAC ∼ = ∠ABD. If C and D are on opposite sides of
←→
AB, then ACBD is a parallelogram. But by the parallelogram law, AB 2 + CD 2 =
2AC 2 + 2AD 2 , implying that 1 + x 2 = 2 + 2x 2 , which is impossible. Therefore C and
←→
D lie on the same side of AB and ABDC is an isosceles trapezoid. Let m(∠ADC) =
α. Then m(∠BCD) = α (since ADC ∼ = BCD), m(∠ABC) = m(∠BAD) = α
(alternating interior angles), m(∠ACB) = m(∠ADB) = α (base angles of isosceles
triangles), and m(∠CAD) = m(∠CBD) = 2α (base angles of isosceles triangles).
The sum of the measure of the interior angles of a quadrilateral is 360◦ , so 10α = 360◦
and α = 36◦ . This means that A, B, C, and D are √ four of the five vertices of a regular
pentagon. In this case, AD = BC = CD = (1 + 5)/2.
 80 MATHEMATICS MAGAZINE

We have shown that there are six configurations of four points satisfying the require-
ments described in the problem statement: (1) a rhombus with one pair of opposite
angles measuring 60◦ , (2) a square, (3) an isosceles triangle with vertex angle of 30◦
and its circumcenter, (4) a kite with a pair of opposite angles measuring 60◦ and 150◦ ,
(5) an equilateral triangle and its circumcenter, and (6) four of the five vertices of a
regular pentagon.
Also solved by Diya Bhatt & Riley Platz & Tony Luo (students), Viera Cernanova (Slovakia),
M. V. Channakeshava (India), Seungheon Lee (Korea), Eagle Problem Solvers, Michael Reid, Celia
Schacht, Albert Stadler (Switzerland), Tianyue Ruby Sun (student), Randy K. Schwartz, and the
proposer. There were six incomplete or incorrect solutions.

Two compass and straightedge constructions December 2020

2115. Proposed by H. A. ShahAli, Tehran, Iran.
Let A and B be two distinct points on a circle and let k be a positive rational number.
(a) Give a compass and straightedge construction of a point C on the circle such that
AC/BC = k.
(b) Give a compass and straightedge construction of a point C on the circle such that
AC · BC = k. As part of your solution, find the restrictions on k in terms of AB
and the radius of the circle necessary for such a C to exist.

Solution by Enrique Treviño, Lake Forest College, Lake Forest, IL.

(a) It is well known that we can construct a point D on segment AB such that
AD/BD = k. Let M be a point of intersection of the perpendicular bisector of
AB with the given circle. Then AM = BM. Let C be the second point of intersec-
←→
tion of MD with the circle. Since AM = BM, then ∠ACM = ∠BCM. Therefore
D is on the angle bisector of ∠ACB and by the angle bisector theorem
AC AD
= = k.
BC BD
An alternative solution is to note that {X|AX/BX = k} is a circle or the perpen-
 VOL. 95, NO. 1, FEBRUARY 2022 81
dicular bisector of AB. This curve is readily constructible, and we then find its
intersection with the original circle.
(b) To solve this problem in full generality, we need a segment of length 1 to be given.
It is well known that given such a segment, λ ∈ Q+ , and a segment of length a, we
can use similar triangles to construct a segment of length λ/a.

Denote the center of the circle by O and let m∠AOB = 2α. Then x = 2r sin α.
For any point C on the circle, m∠ACB = α or m∠ACB = π − α. In either case
sin ∠ACB = sin α. We will denote the area of P QR by (P QR). We know
AC · BC · sin α
(ABC) = .
2
Let h be the height of ABC with base AB, then
xh
(ABC) = .
2
Therefore

AC · BC = 2rh.
Let  be the perpendicular bisector of AB. Using the facts stated above, we can
construct a point D on  such that the distance from D to AB is k/(2r). Next, we
←→
draw a line through D parallel to AB and let C be one of the points of intersection
of this line with the given circle. This point C satisfies AC · BC = 2rh = k.

The largest possible value of AC · BC occurs when C lies on the perpendicular

bisector of AB at a maximum distance from AB, namely when

x2
AC = BC = r + r 2 − .
4
Therefore

k = AC · BC
  2
x2
≤ r+ r2 −
4
 
= r 2r + 4r 2 − x 2 .

Also solved by Michel Bataille (France), Ivko Dimitrić, Elias Lampakis (Greece), Celia Schacht,
Albert Stadler (Switzerland), and the proposer. There were three incomplete or incorrect solutions.
VOL. 94, NO. 4, OCTOBER 2021 309

Solutions
Invariance of a ratio of sums of cotangents October 2020
2101. Proposed by Michael Goldenberg, The Ingenuity Project, Baltimore Polytechnic
Institute, Baltimore MD and Mark Kaplan, Towson University, Towson, MD.
Recall that the Steiner inellipse of a triangle is the unique ellipse that is tangent to each
side of the triangle at the midpoints of those sides. Consider the Steiner inellipse ES of
ABC and another ellipse, EA , passing through the centroid G of ABC and tangent
310 MATHEMATICS MAGAZINE
←→ ←→
to AB at B and to AC at C. If ES and EA meet at M and N, let ∠MAN = α. Construct
ellipses EB and EC , introduce their points of intersection with ES , and define angles β
and γ in an analogous way. Prove that
cot α + cot β + cot γ 11
= √ .
cot A + cot B + cot C 3 5

Solution by Albert Stadler, Herrliberg, Switzerland.

We first consider the equilateral triangle with vertices
√ √
A = (16, 0), B = (−8, 8 3), and C = (−8, −8 3),

whose centroid is the origin. In this case, ES is the circle whose equation is x 2 + y 2 =
82 and EA is the circle whose equation is (x + 16)2 + y 2 = 162 . Solving this system
of equations we find
√ √
M = (−2, 2 15) and N = (−2, −2 15).
Let ∠(−
→ v ) denote the angle between the vectors −
u ,−
→ →u and −
→
v . Then
 √ √   √ √ 
A = ∠ (−24, 8 3), (−24, −8 3) and α = ∠ (−18, 2 15), (−18, −2 15) .

Rotating the vectors above 120◦ and 240◦ counter-clockwise gives

 √ √ 
B = ∠ (0, −16 3), (24, −8 3) ,
 √ √ √ √ √ √ 
β = ∠ (9 − 3 5, −9 3 − 15), (9 + 3 5, −9 3 + 15) ,
 √ √ 
C = ∠ (24, 8 3), (0, 16 3) , and
 √ √ √ √ √ √ 
γ = ∠ (9 + 3 5, 9 3 − 15)), (9 − 3 5, 9 3 + 15) .

Now let A B C be any non-degenerate triangle whose centroid is at the origin.

There is an invertible linear map f (x, y) = (ax + by, cx + dy) such that A B C =
f ( ABC). This linear mapping preserves the centroid, all midpoints, all tangencies,
and it maps lines to lines and circles to ellipses. It remains to analyze how this lin-
ear mapping transforms the six numbers cot A, cot B, cot C, cot α, cot β, and cot γ to
cot A , cot B , cot C , cot α , cot β , and cot γ .
We will use the fact if φ = ∠((u1 , u2 ), (v1 , v2 )), then
u1 v1 + u2 v2
cot φ =
u1 v2 − u2 v1
by the difference formula for cotangent.
Now
 √ √ 
A = ∠ f (−24, 8 3), f (−24, −8 3) ,
 √ √ 
B = ∠ f (0, −16 3), f (24, −8 3) , and
 √ √ 
C = ∠ f (24, 8 3), f (0, 16 3) .
VOL. 94, NO. 4, OCTOBER 2021 311
This gives
3a 2 − b2 + 3c2 − d 2
cot A = √
2 3(ad − bc)
√ √
b2 − 3ab + d 2 − 3cd
cot B = √
3(ad − bc)
√ √
b2 + 3ab + d 2 + 3cd
cot C = √ .
3(ad − bc)
Therefore,
√  2
3 a + b2 + c2 + d 2
cot A + cot B + cot C = .
2(ad − bc)
A similar calculation yields

11 a 2 + b2 + c2 + d 2
cot α + cot β + cot γ = √ .
2 15(ad − bc)
Finally,
cot α + cot β + cot γ 11
= √
cot A + cot B + cot C 3 5
as desired.
Also solved by Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia) and the proposers.
There were two incomplete or incorrect solutions.

Trigonometric identities for the heptagonal triangle October 2020

2102. Proposed by Donald Jay Moore, Wichita, KS.
Let α = π/7, β = 2π/7, and γ = 4π/7. Prove the following trigonometric identities.
cos2 α cos2 β cos2 γ
+ + = 10,
cos2 β cos2 γ cos2 α
sin2 α sin2 β sin2 γ
+ + = 6,
sin2 β sin2 γ sin2 α
tan2 α tan2 β tan2 γ
2
+ 2
+ = 83.
tan β tan γ tan2 α

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

Denote the trigonometric expressions by C , S , T , respectively. The expansion

sin(7t) = sin t 64 cos6 t − 80 cos4 t + 24 cos2 t − 1
yields the key polynomial as follows. When t = α or t = β or t = γ , then sin(7t) = 0
but sin t  = 0. Hence the cubic polynomial
p(x) = 64x 3 − 80x 2 + 24x − 1
312 MATHEMATICS MAGAZINE
has the three zeros a = cos2 α, b = cos2 β, c = cos2 γ . Since
p(x) = 64(x − a)(x − b)(x − c),
we have values for the three elementary symmetric polynomials:
5 3 1
a+b+c = , ab + bc + ca = , abc = .
4 8 64
We use the double angle formula for sine as follows:
sin2 t sin2 t 1
2
= 2
= .
sin 2t 2
4 sin t cos t 4 cos2 t
Hence, since sin2 2γ = sin2 α,
sin2 α sin2 β sin2 γ 1 1 1 bc + ca + ab 3/8
S= 2
+ 2
+ 2
= + + = = = 6.
sin β sin γ sin α 4a 4b 4c 4abc 4/64
We use the double angle formula for cosine as follows:
cos2 t cos2 t
= .
cos2 2t (2 cos2 t − 1)2
Hence, since cos2 2γ = cos2 α,
cos2 α cos2 β cos2 γ a b c
C= + + = + + .
2
cos β 2
cos γ 2
cos α (2a − 1) 2 (2b − 1) 2 (2c − 1)2
Substituting x = (y + 1)/2 into the polynomial p(x) yields
q(y) = 8y 3 + 4y 2 − 4y − 1.
Since y = 2x − 1, the zeros of q(y) are a = 2a − 1, b = 2b − 1, c = 2c − 1 and
the elementary symmetric polynomial expressions are
1 1 1
a +b +c =− , ab +bc +ca =− , abc = .
2 2 8
Hence,
a +1 b +1 c +1 a b 2c 2 + b a 2c 2 + c a 2b 2 + b 2c 2 + a 1c 2 + a 2b 2
C= 2
+ 2
+ 2
=
2a 2b 2c 2(a b c )2
(a b c )(a b + b c + c a ) + (a b + b c + c a )2 − 2(a b c )(a + b + c )
=
2(a b c )2
−1/16 + 1/4 + 1/8
= = 10.
2/64
For the third identity, we use both double angle formulas:
tan2 t sin2 t cos2 2t (2 cos2 t − 1)2
= 2
=
tan2 2t cos2 t sin 2t 4 cos4 t
Thus, since tan2 2γ = tan2 α,
2 2 2
tan2 α tan2 β tan2 γ 2a − 1 2b − 1 2c − 1
T = + + = + + .
tan2 β tan2 γ tan2 α 2a 2b 2c
VOL. 94, NO. 4, OCTOBER 2021 313
Substituting x = 1/(2(1 − z)) into the polynomial p(x) and clearing fractions yields

r(z) = 8(z3 + 9z2 − z − 1).

Since z = (2x − 1)/(2x), the zeros of r(z) are

2a − 1 2b − 1 2c − 1
a = , b = , c =
2a b c
and the elementary symmetric polynomial expressions are

a + b + c = −9, a b + b c + c a = −1, a b c = 1.
Hence,

T = a 2 + b 2 + c 2 = (a + b + c )2 − 2(a b + b c + c a ) = 92 − 2(−1) = 83.

Also solved by Michel Bataille (France), Anthony J. Bevelacqua, Brian Bradie, Robert Cal-
caterra, Hongwei Chen, John Christopher, Robert Doucette, Habib Y. Far, J. Chris Fisher, Dmitry
Fleischman, Michael Goldenberg & Mark Kaplan, Russell Gordon, Walther Janous (Austria), Kee-
Wai Lau (Hong Kong), James Magliano, Ivan Retamoso, Volkhard Schindler (Germany), Randy
Schwartz, Allen J.Schwenk, Albert Stadler (Switzerland), Seán M. Stewart (Australia), Enrique
Treviño, Michael Vowe (Switzerland), Edward White & Roberta White, Lienhard Wimmer (Ger-
many), and the proposer. There were two incomplete or incorrect solutions.

How many tickets to buy to guarantee three out of four? October 2020
2103. Proposed by Péter Kórus, University of Szeged, Szeged, Hungary.
In a soccer game there are three possible outcomes: a win for the home team (denoted
1), a draw (denoted X), or a win for the visiting team (denoted 2). If there are n games,
betting slips are printed for all 3n possible outcomes. For four games, what is the
minimum number of slips you must purchase to guarantee that at least three of the
outcomes are correct on at least one of your slips?

Solution by Northwestern University Math Problem Solving Group, Northwestern Uni-

versity, Evanston, IL.
The answer is nine.
First, we prove that it is impossible to guarantee at least three correct outcomes with
fewer than nine slips.
Let T be the set of all possible outcomes, i.e., all 4-tuples of 1, X, and 2. There are
34 = 81 such 4-tuples. In that set, we define the Hamming distance d as the number of
places in which two tuples differ. For example, d(1X21, 2X12) = 3 because 1X21 and
2X12 differ in three places, namely the first, third and fourth places. The Hamming
distance satisfies the usual axioms for a metric, and we can define balls in T in the
usual way, i.e., a ball with center c ∈ T and radius r ∈ R is
Br (c) = {t ∈ T | d(t, c) ≤ r}.

Given a tuple c ∈ T , the set of tuples that coincide with c in at least three places
consists of those that differ from c in no more than one place. In other words, this set
is B1 (c). Note that B1 (c) contains exactly 9 elements: the center c, the two tuples that
differ from c exactly in the first element, the two that differ in the second, the two that
differ in the third, and the two that differ in the fourth.
314 MATHEMATICS MAGAZINE
In order to ensure that our slips c1 , c2 , . . . , cn contain at least three correct entries,
the balls B1 (ci ), i = 1, 2, . . . , n must cover T , i.e.,
n
T = B1 (ci ).
i=1

Since |B1 (c)| = 9 and |T | = 81, we will need at least 81/9 = 9 slips.
Next, we will prove that nine slips suffice. That can be accomplished by exhibit-
ing nine 4-tuples c1 , . . . , c9 such that Bi (c1 ), . . . , Bi (c9 ) cover T , i.e., such that every
element in T has a Hamming distance of at most 1 from at least one of the ci . The
following 4-tuples satisfy the condition:

1111 1XXX 1222 X1X2 XX21 X21X 2X12 212X 22X1

One (somewhat tedious) way to check it is to verify that each of the 81 elements in
T differ from at least one of these tuples in no more one place.
A slightly easier way to verify the assertion is to observe that these tuples differ
from each other in exactly three places, so the Hamming distance between any two
of them is 3. Because of the triangle inequality, it is impossible for balls of radius 1
centered on the ci to overlap. Therefore the total number of elements contained in the
union of these balls is 9 · 9 = 81, so the union must be all of T .
This completes the proof.
Also solved by Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia), Eagle Problem
Solvers, Fresno State Problem Solving Group, Dan Hletko, Rob Pratt, Allen J. Schwenk, and
the proposer. There were seven incomplete or incorrect solutions.

Vector spaces as unions of proper subspaces October 2020

2104. Proposed by the Missouri State University Problem Solving Group, Missouri
State University, Springfield, MO.
It is well known that no vector space can be written as the union of two proper sub-
spaces. For which m does there exist a vector space V that can be written as a union of
m proper subspaces with this collection of subspaces being minimal in the sense that
no union of a proper subcollection is equal to V ?

Solution by Paul Budney, Sunderland, MA.

Such a decomposition exists for any m > 2.
Let V = Fn2 , where F2 is the field with two elements. Let
Vi = { (x1 , . . . , xn ) ∈ V | xi = 0}
for 1 ≤ i ≤ n and let
W = {(0, 0, . . . , 0), (1, 1, . . . , 1)}.
Clearly W and the Vi are proper subspaces of V . Since (1, 1, . . . , 1) is the only vector
not in V1 ∪ V2 ∪ . . . ∪ Vn ,
W ∪ V1 ∪ V2 ∪ . . . ∪ Vn = V .
Deleting W from this union excludes (1, 1, . . . , 1). Deleting Vi from this union
excludes (1, . . . , 1, 0, 1, . . . , 1), with 0 for the ith component and 1’s elsewhere.
Thus, there is no proper subcollection of these subspaces whose union is V . There are
VOL. 94, NO. 4, OCTOBER 2021 315
n + 1 subspaces, and since n ≥ 2 is arbitrary, the desired decomposition exists for any
m > 2.
Also solved by Anthony Bevelacqua, Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia),
Robert Doucette, Eugene Herman, and the proposer. There was one incomplete or incorrect solu-
tion.

An asymptotic formula for a definite integral October 2020

2105. Proposed by Marian Tetiva, National College “Gheorghe Ro̧sca Codreanu”,
Bı̂rlad, Romania.
Let f : [0, 1] → R be a function that is k times differentiable on [0, 1], with the kth
derivative integrable on [0, 1] and (left) continuous at 1. For integers i ≥ 1 and j ≥ 0
let

σj(i) = 1j1 2j2 · · · i ji ,
j1 +j2 ···+ji =j

where the sum is extended over all i-tuples (j1 , . . . , ji ) of nonnegative integers that
sum to j . Thus, for example, σ0(i) = 1, and σ1(i) = 1 + 2 + · · · + i = i(i + 1)/2 for all
i ≥ 1. Also, for 0 ≤ j ≤ k let
(j −1) (j ) (j +1)
aj = σj(1) f (1) + σj(2)
−1 f (1) + · · · + σ1 f (1) + σ0 f (j ) (1).

Prove that
 1
a0 a1 ak 1
x n f (x)dx = − 2 + · · · + (−1)k k+1 + o k+1
,
0 n n n n

for n → ∞. As usual, we denote by f (s) the sth derivative of f (with f (0) = f ), and
by o(xn ) a sequence (yn ) with the property that limn→∞ yn /xn = 0.

Solution by Michel Bataille, Rouen, France.

For x ∈ [0, 1], let f0 (x) = f (x) and
d 
fj (x) = xfj −1 (x) , 1 ≤ j ≤ k.
dx
An easy induction shows that for 0 ≤ j ≤ k, the function fj is a linear combination of
the functions f (x), xf (x), . . . , x j f (j ) (x). It follows that f0 , f1 , . . . , fk−1 are differ-
entiable on [0, 1] and that fk is integrable on [0, 1] and continuous at 1.
Integrating by parts, we obtain the following recursion that holds for 1 ≤ j ≤ k − 1:
 1 1 
xn 1 1 n
x n fj −1 (x) dx = · (xfj −1 (x)) − x fj (x) dx
0 n 0 n 0

fj −1 (1) 1 1 n
= − x fj (x) dx.
n n 0
With the help of this recursion, we are readily led to
 1  1
x f (x) dx =
n
x n f0 (x) dx
0 0
316 MATHEMATICS MAGAZINE

k−1  1
fj (1) (−1)k
= (−1)j + x n fk (x) dx.
j =0
nj +1 nk 0

Now, if g : [0, 1] → R is integrable on [0, 1] and continuous at 1, then

 1
lim n · x n g(x) dx = g(1)
n→∞ 0

(Paulo Ney de Souza, Jorge-Nuno Silva, Berkeley Problems in Mathematics, Springer,

2004, Problem 1.2.13). With g = fk , this yields
 1
fk (1) 1
x n fk (x) dx = +o
0 n n
and therefore
 1 
k−1
fj (1) (−1)k fk (1) 1
x f (x) dx =
n
(−1)j + +o
0 j =0
nj +1 nk n n


k
fj (1) 1
= (−1)j j +1
+o k+1
.
j =0
n n

Comparing this with the statement of the problem, it remains to prove that aj = fj (1)
for 0 ≤ j ≤ k. Clearly, it is sufficient to prove that for x ∈ [0, 1]


j
fj (x) = σj(i+1) i (i)
−i x f (x). (Ej )
i=0

We use induction. Since f0 (x) = f (x) = 1 · x 0 f (0) (x), (E0 ) holds. Before addressing
the induction step, we establish two results about the numbers σj(i) . The first result is


j
σj(i+1) = (1 + i)r σj(i)
−r . (1)
r=0

Proof. When j1 + · · · + ji + ji+1 = j , then ji+1 can take the values 0, 1, . . . , j . It

follows that

σj(i+1) = 1j1 2j2 · · · i ji (i + 1)ji+1
j1 +···+ji+1 =j


j

= (1 + i)r 1j1 2j2 · · · i ji
r=0 j1 +···+ji =j −r


j
= (1 + i)r σj(i)
−r .
r=0

The second result is

+1 = σj +1 + (1 + i)σj
σj(i+1) (i) (i+1)
. (2)
VOL. 94, NO. 4, OCTOBER 2021 317
Proof. Applying (1),
j +1

+1 =
σj(i+1) (1 + i)r σj(i)
+1−r
r=0

j +1

= σj(i)
+1 + (1 + i) (1 + i)r−1 σj(i)
−(r−1)
r=1


j
= σj(i)
+1 + (1 + i) (1 + i)r σj(i)
−r
r=0

+1 = σj +1 + (1 + i)σj
and applying (1) again we conclude that σj(i+1) (i) (i+1)
.
Now, assume that (Ej ) holds for some integer j such that 0 ≤ j ≤ k − 1. Then, we
calculate
 j 
d  (i+1) i+1 (i)
fj +1 (x) = σ x f (x)
dx i=0 j −i


j

j
= −i (i + 1)x f (x) +
σj(i+1) i (i)
σj(i+1)
−i x
i+1 (i+1)
f (x)
i=0 i=0

j +1

j

= −i (i + 1)x f (x) +
σj(i+1) i (i)
σj(i) i (i)
−i+1 x f (x)
i=0 i=1

j  
 (j +1) j +1 (j +1)
= σj(1) f (x) + [σj(i)
−i+1 + (i + 1)σ (i+1) i (i)
j −i ]x f (x) + σ0 x f (x).
i=1

(j +1) (j +2)
Using (2) and σj(1) = σj(1)
+1 = 1 = σ0 = σ0 , we see that
j +1

fj +1 (x) = σj(i+1) i (i)
+1−i x f (x)
i=0

so that (Ej +1 ) holds. This completes the induction step and the proof.
 
Note. The number σj(i) is the Stirling number of the second kind S(i + j, i) = i+j
i
(see L. Comtet, Advanced Combinatorics, Reidel, 1974, Theorem D p. 207).
Also solved by Albert Stadler (Switzerland) and the proposer.
 230 MATHEMATICS MAGAZINE
Solutions

The number of isosceles triangles in various polytopes June 2020

2096. Proposed by H. A. ShahAli, Tehran, Iran.
Any three distinct vertices of a polytope P form a triangle. How many of these
triangles are isosceles if P is (a) a regular n-gon? (b) one of the Platonic solids? (c) an
n-dimensional cube?

Solution by Robert Calcaterra, University of Wisconsin-Platteville, Platteville, WI.

Let m denote the number of vertices of P . For a fixed vertex A of P , let F (P ) denote
the number of unordered triplets of distinct vertices A, B, and C of P for which AB =
AC, G(P ) is the number of such triplets for which AB = AC = BC, and I (P ) the
number of isosceles triangles that can be formed using the vertices of P . Note that
since all of the polytopes under consideration are uniform, F (P ) and G(P ) do not
depend on A. Since each equilateral triangle is counted in F (P ) for three different
choices of A,
m 2
I (P ) = m (F (P ) − G(P )) + G(P ) = mF (P ) − mG(P ).
3 3
(a) If P is a regular n-gon, then F (P ) = (n − 1)/2. Moreover, G(P ) = 1 if n is a
multiple of 3 and G(P ) = 0 if not. Therefore,
  
n n−1 if 3  n
I (P ) =  2  2n
n n−1
2
− 3
if 3|n

(b) Let P be a Platonic solid. If A and B are vertices of P , the minimum number of
edges of the solid that must be traversed to get from A to B will be called the span
from A to B. For the Platonic solids, the spans for two pairs of vertices are the same
if and only if the Euclidean distances are the same.
• If P is a tetrahedron, every triplet of distinct vertices forms an isosceles (in

fact, equilateral) triangle. Therefore I (P ) = 43 = 4.
• If P is a cube, then the numbers of vertices with spans 1, 2, and   3 from
  the
fixed vertex A are 3, 3, and 1, respectively. Therefore, F (P ) = 32 + 32 = 6.
Moreover, 0 pairs of the vertices with span 1 from A have span 1 from each
other, and 3 pairs with span 2 from A have span 2 from each other. Thus
G(P ) = 3 and I (P ) = 8 · 6 − 23 · 8 · 3 = 32. (This also follows from part (c)
below).
• If P is an octahedron, every   triplet of distinct vertices forms an isosceles tri-
angle. Therefore I (P ) = 63 = 20.
• If P is an icosahedron, then the numbers of vertices with spans 1, 2,and  3from

the fixed vertex A are 5, 5, and 1, respectively. Therefore, F (P ) = 52 + 52 =
20. Moreover, 5 pairs of the vertices with span 1 from A have span 1 from
each other, and 5 pairs with span 2 from A have span 2 from each other; thus
G(P ) = 10 and I (P ) = 12 · 20 − 23 · 12 · 10 = 160.
• If P is a dodecahedron, then the numbers of vertices with spans   1, 2,3, 4,
and 5 from A are 3, 6, 6, 3, and 1, respectively. So, F (P ) = 32 + 62 + 62 +
3
2
= 36. Moreover, 0 pairs of vertices with span 1 from A have span 1 from
each other, 3 pairs with span 2 from A have span 2 from each other, 6 pairs
VOL. 94, NO. 3, JUNE 2021 231
with span 3 from A have span 3 from each other, and 0 pairs with span 4 from
A have span 4 from each other; thus, G(P ) = 9 and I (P ) = 20 · 36 − 23 · 20 ·
9 = 600.
(c) Let P be a cube in Rn . We may view the vertices of P as binary n-tuples, so that
the distance between two vertices is the square root of the number
√ of components

at which they differ. The number of vertices of P at distance k from A is nk for
k = 0, 1, . . . , n. Recall that
n
n n
n 2
2n
= 2n and = .
k=0
k k=0
k n

Therefore,

n−1
1 n n 1  n
n−1 2 
n−1
n
F (P ) = −1 = −
k=1
2 k k 2 k=1 k k=1
k

1 2n
= − 2 − (2n − 2)
2 n
1 2n
= − 2n
2 n
For the vertices A, B, and C to form an equilateral triangle with sides of length
√
k, three disjoint subsets, say X, Y , and Z, must be chosen from {1, 2 . . . , n} in
such a way that the components of A differ from those of B at precisely the posi-
tions in X ∪ Y , the components of A differ from those of C at precisely the posi-
tions in X ∪ Z, and the components of B differ from those of C at precisely the
positions in Y ∪ Z. This forces |X ∪ Y | = |X ∪ Z| = |Y ∪ Z| = k, which yields
|X| = |Y | = |Z| =  and k = 2. There will be n − 3 positions at which the com-
ponents of A, B, and C all agree (the positions in the complement of X ∪ Y ∪ Z).
Note that each equilateral triangle will be generated twice using this procedure
because interchanging Y and Z will reverse the roles of B and C. Therefore (using
multinomial coefficients), we have
n/3
1 n
G(P ) = and
2 =1 n − 3, , , 
n/3
2n 2n  n
I (P ) = 2n−1 − 2n −
n 3 =1 n − 3, , , 

Also solved by Allen J. Schwenk, Albert Stadler (Switzerland), and the proposer. There were
two incomplete or incorrect solutions.

A series involving the floor, ceiling, and round functions June 2020
2097. Proposed by Omran Kouba, Higher Institute for Applied Sciences and Technol-
ogy, Damascus, Syria.
For a real number x ∈ / 12 + Z, denote the nearest integer to x by x . For any real
number x, denote the largest integer smaller than or equal to x and the smallest integer
 232 MATHEMATICS MAGAZINE
larger than or equal to x by x and x , respectively. For a positive integer n let
2 1 1
an = √ − √ − √ .
n  n n
∞
(a) Prove that the series n=1 an is convergent and find its sum L.
(b) Prove that the set
√  n 
n( ak − L) : n ≥ 1
k=1

is dense in [0, 1].

Solution by Hongwei Chen, Christopher Newport University, Newport News, VA.

(a) We show that the sum converges to zero. To see this, first, we can easily check the
following facts:
√
n = k,for n ∈ [k(k − 1) + 1, k(k + 1)],
√
 n = k,for n ∈ [k 2 , (k + 1)2 ),
√
n = k + 1,for n ∈ (k 2 , (k + 1)2 ].
These imply that ak2 = 0 and
2 1 1 1
an = − − = , for n ∈ (k 2 , k(k + 1)],
k k k+1 k(k + 1)
2 1 1 1
an = − − =− , for n ∈ (k(k + 1), (k + 1)2 ).
k+1 k k+1 k(k + 1)
k2
Therefore, for k 2 ≤ n ≤ (k + 1)2 , we have m=1 am = 0 and

n
1 1
0≤ am ≤ · [k(k + 1) − k 2 ] =
m=1
k(k + 1) k+1

As n → ∞, we have k → ∞ and so

∞ 
n
an = lim am = 0.
n→∞
n=1 m=1

√ x ∈ [0, 1]. We show that there exists a subsequence from the set
(b) Let
{ n nm=1 am }, which converges to x. Notice that there exist two integer sequences
pk and qk with 0 ≤ pk ≤ qk such that pk /qk → x, as k → ∞. Let nk = qk2 + pk .
Then
2
1
qk2 ≤ nk ≤ qk2 + qk < qk + .
2
This implies that
√ √ √
nk = qk ,  nk  = qk , nk = qk + 1.
VOL. 94, NO. 3, JUNE 2021 233
Therefore, as k → ∞, we have
√
√ 
nk
√ nk − qk2 pk nk
nk am = nk · = · → x.
m=1
q (q
k k + 1) qk qk + 1
√ n
This proves that the set { n m=1 am } is dense in [0, 1].

Also solved by Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia), Brian Bradie, Robert
Calcaterra, Dmitry Fleischman, Maxim Galushka (UK), GWstat Problem Solving Group, Eugene
A. Herman, Walter Janous (Austria), Donald E. Knuth, Sushanth Sathish Kumar, Elias Lampakis
(Greece), Shing Hin Jimmy Pa (Canada), Allen Schwenk, Albert Stadler (Switzerland), and the
proposer. There was one incorrect or incomplete solution.

A zigzag sequence of random variables June 2020

2098. Proposed by Albert Natian, Los Angeles Valley College, Valley Glen, CA.
Let Z0 = 0, Z1 = 1, and recursively define random variables Z2 , Z3 , . . . , taking
values in [0, 1] as follows: For each positive integer k, Z2k is chosen uniformly in
[Z2k−2 , Z2k−1 ], and Z2k+1 is chosen uniformly in [Z2k , Z2k−1 ].
Prove that, with probability 1, the limit Z ∗ = limn→∞ Zn exists and find its distri-
bution.

Solution by Northwestern University Math Problem Solving Group, Northwestern Uni-

versity, Evanston, IL.
We will prove:
1. The limit Z ∗ exists.
2. The limit Z ∗ has probability density f (x) = 2x on [0, 1].
Proof of 1. We have that [Z0 , Z1 ] ⊇ [Z2 , Z1 ] ⊇ [Z2 , Z3 ] ⊇ [Z4 , Z3 ] ⊇ . . . is a
sequence of nested closed intervals. By the nested interval theorem, their intersec-
tion will be non-empty, and will consist of a unique point precisely if the sequence of
lengths of the nested intervals tends to zero. We prove that this happens with probabil-
ity 1.
Let In (n = 0, 1, 2, , . . . ) be the nth interval in the sequence, and Ln = length of
In , i.e., L2k = Z2k+1 − Z2k and L2k+1 = Z2k+1 − Z2k+2 . Pick δ > 0. We will prove by
induction that the probability of Ln > δ is P (Ln > δ) ≤ (1 − δ)n . Since P (Ln > 1) =
0 the result is trivially true for δ ≥ 1, so we may assume 1 > δ > 0.
Base case: For n = 0 the inequality P (L0 > δ) ≤ (1 − δ)0 obviously holds because
L0 = 1, hence P (L0 > δ) = P (1 > δ) = 1 and (1 − δ)0 = 1.
Induction step: Assume P (Ln > δ) ≤ (1 − δ)n . Then
P (Ln+1 > δ) = P (Ln ≤ δ) · P (Ln+1 > δ | Ln ≤ δ) + P (Ln > δ) · P (Ln+1 > δ | Ln > δ).
Note that the first term is zero because if Ln ≤ δ then Ln+1 > δ is impossible. On the
other hand, if Ln > δ then we only have Ln+1 > δ if the next endpoint Zn+2 is selected
at a distance less than Ln − δ from the right or left (depending on the parity of n)
endpoint of In . The probability is
Ln − δ δ
P (Ln+1 > δ | Ln > δ) = =1− ≤ 1 − δ.
Ln Ln
Hence
P (Ln+1 > δ) ≤ (1 − δ)n (1 − δ) = (1 − δ)n+1 ,
234 MATHEMATICS MAGAZINE
and this completes the induction.
From here we get limn→∞ P (Ln+1 > δ) = 0 for every δ > 0, hence Ln → 0 as
n → ∞ with probability 1.
Proof of 2. For each n ≥ 0 define the new random variable Un , chosen between Z2n
and Z2n+1 with probability density
2(x − z2n )
fUn |Z2n =zn ,Z2n+1 =z2n+1 (x) =
(z2n+1 − z2n )2
on [z2n , z2n+1 ], where “Un |Z2n = z2n , Z2n+1 = z2n+1 ” means the random variable Un
given Z2n = z2n and Z2n+1 = z2n+1 (we ignore the case z2n+1 = z2n because its proba-
bility is zero).
Since Un is between Z2n and Z2n+1 , its limit U ∗ will coincide with Z ∗ .
Next, we will prove by induction that for every n ≥ 0, the probability density of Un
is always the same, namely fUn (x) = 2x on [0, 1].
2(x − 0)
Base case: For n = 0 we have Z0 = 0, Z1 = 1, hence fU0 (x) = = 2x on
(1 − 0)2
[0, 1].
Induction step: Assume fUn (x) = 2x. Next, note that Un+1 is defined like Un but
with starting points Z2 and Z3 in place of Z0 and Z1 . So, Un+1 given Z2 = z2 and Z3 =
z3 is just Un mapped from [0, 1] to [z2 , z3 ] with the transformation (z3 − z2 )Un + z2 .
By induction hypothesis we have fUn (x) = 2x, and its transformation to [z2 , z3 ] will
have probability density
2(x − z2 )
fUn+1 |Z2 =z2 ,Z3 =z3 (x) =
(z3 − z2 )2
on [z2 , z3 ].
The cumulative distribution function of Un+1 is FUn+1 (x) = P (Un+1 ≤ x). By def-
inition Un+1 must be in the interval [Z2 , Z3 ], while x may be in any of two different
intervals, namely [Un+1 , Z3 ) or [Z3 , 1]. So, the event Un+1 ≤ x can be expressed as
the union of Z2 ≤ Z3 ≤ x and Z2 ≤ Un+1 ≤ x < Z3 . Since they are disjoint we have
P (Un+1 ≤ x) = P (Z2 ≤ Z3 ≤ x) + P (Z2 ≤ Un+1 ≤ x < Z3 ) .

We have that X2 is random uniform on [0, 1], and X3 is random uniform on [Z2 , 1],
so
1
fZ3 |Z2 =z2 (x) = ,
1 − z2
hence
 x
x − z2
P (Z2 ≤ Z3 ≤ x) = dz2 = x + (1 − x) log(1 − x) .
0 1 − z2
The second term can be computed as follows:
 x 1 x
P (Z2 ≤ Un+1 ≤ x < Z3 ) = fUn+1 |Z2 =z2 ,Z3 =z3 (t)fZ3 |Z2 =z2 (x) dt dz3 dz2
0 x z2
 x 1 x
2(t − z2 ) 1
= dt dz3 dz2
0 x z2 (z3 − z2 )2 1 − z2
= (x − 1)(x + log(1 − x)) ,
VOL. 94, NO. 3, JUNE 2021 235
hence
FU2n+1 (x) = x + (1 − x) log(1 − x) + (x − 1)(x + log(1 − x)) = x 2 .
Differentiating we get fU2n+1 (x) = 2x on [0, 1], and this completes the induction.
Since the distribution of Un is the same for every n we have that the limit U ∗ will
have the same distribution too. And since U ∗ = Z ∗ , the same will hold for Z ∗ , hence
fZ∗ (x) = 2x.
Also solved by Robert A. Agnew, Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia),
Robert Calcaterra, Shuyang Gao, John C. Kieffer, Omran Kouba (Syria), Kenneth Schilling, and
the proposer.

An almost linear functional equation June 2020

2099. Proposed by Russ Gordon, Whitman College, Walla Walla, WA and George Sto-
ica, Saint John, NB, Canada.
Let r and s be distinct nonzero rational numbers. Find all functions f : R → R that
satisfy
x+y f (x) + f (y)
f =
r s
for all real numbers x and y.

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

Clearly the zero function is always a solution and, when s = 2, all constant functions
are solutions. We show that there are no others. First assume s  = 2. Substituting 0
for both x and y yields f (0) = 0. Substituting y = 0 and y = −x yield these two
identities:
 x  f (x)
f = , f (−x) = −f (x) for all x ∈ R.
r s
Given any x ∈ R, we use induction to show that f (nx) = nf (x) for all n ∈ N. The
base case is a tautology. If f (nx) = nf (x) for some n ∈ N, then
f ((n + 1)x) (n + 1)x nx + x f (nx) + f (x) (n + 1)f (x)
=f =f = =
s r r s s
and so f ((n + 1)x) = (n + 1)f (x). It follows that f (x/n) = f (x)/n for all n ∈ N
and hence that f ((m/n)x) = (m/n)f (x) for all m, n ∈ N. Since f (−x) = −f (x),
this last statement is also true for m negative. Choose m, n so that r = n/m. Therefore
f (x)  x  f (x)
=f =
s r r
and so f (x) = 0.
Now assume s = 2, and let t = 2/r. Thus t  = 1 and
t f (x) + f (y)
f (x + y) = , for all x, y ∈ R.
2 2
Substituting y = x and y = −x yield
f (x) + f (−x)
f (tx) = f (x), = f (0) for all x ∈ R.
2
236 MATHEMATICS MAGAZINE
Thus f (−x/t) = f (−x), and so
t −1 t f (x) + f (−x/t) f (x) + f (−x)
f x =f (x − x/t) = = = f (0).
2 2 2 2
Therefore f is a constant function.
Also solved by Michel Bataille (France), Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Aus-
tralia), Paul Budney, Robert Calcaterra, Walther Janous (Austria), Sushanth Sathish Kumar,
Omran Kouba (Syria), Elias Lampakis (Greece), Albert Natian, Kangrae Park (South Korea),
Kenneth Schilling, Jacob Siehler, Albert Stadler (Switzerland), Michael Vowe (Switzerland), and
the proposers.

Two congruent triangles on the sides of an arbitrary triangle June 2020

2100. Proposed by Yevgenya Movshovich and John E. Wetzel, University of Illinois,
Urbana, IL.
Given ABC and an angle θ, two congruent triangles ABP and QAC are con-
structed as follows: AQ = AB, BP = AC, m∠ABP = m∠CAQ = θ, B and Q are on
←→ ←→
opposite sides of AC, and C and P are on opposite sides of AB, as shown in the
figure. Let X, Y , and Z be the midpoints of segments AP, BC, and CQ, respectively.
Show that ∠XYZ is a right angle.

Solution by Sushanth Sathish Kumar (student), Portola High School, Irvine, CA.

Let M be the midpoint of segment AB. Note that YZ is a midline of triangle CBQ, and
←→ ←
→ ←
→ ←
→
so BQ is parallel to YZ . Thus, it suffices to show that XY is perpendicular to BQ.
Since MX and MY are midlines of triangles APB and ABC, we have that MX =
←→ ← →
BP/2 = AC/2 = MY. Hence, triangle MXY is isosceles. Moreover, since MX|| BP and
←→ ← →
MY|| AC, we have

m∠XMY = m∠XMA + m∠AMY = θ + 180◦ − α,

where we set α = m∠BAC. It follows that m∠MXY = m∠XYM = (α − θ)/2.
←→ ← →
We wish to calculate m∠(XM, BQ), where m∠(1 , 2 ) denotes the measure of the
non-obtuse angle between 1 and 2 . Note that
←→ ←→
m∠(XM, BQ) = m∠PBQ = m∠PBA + m∠ABQ.

Since AB = AQ and m∠BAQ = α + θ, we find that m∠ABQ = 90◦ − (α + θ)/2.

←→ ← → ←
→ ← →
Thus, m∠(XM, BQ) = 90◦ − (α − θ)/2. But since m∠(MX, XY ) = (α − θ)/2, we
←→ ← →
find that m∠( BQ, XY ) = 90◦ , and we are done.
152 MATHEMATICS MAGAZINE
Solutions

A geometric inequality April 2020

2091. Proposed by Marian Tetiva, National College “Gheorghe Roşca Codreanu,”
Bârlad, Romania.
Let ABC be a triangle with sides of lengths a, b, c, altitudes ha , hb , hc , inradius r, and
circumradius R. Prove that the following inequality holds:
a 2 + b2 + c2 − ab − ac − bc
ha + hb + hc ≥ 9r + ,
4R
with equality if and only if ABC is equilateral.

Solution by Robert Calcaterra, University of Wisconsin-Platteville, Platteville, WI.

Let K denote the area of ABC. We have
2K
r= ,
a+b+c
abc
R= ,
4K
2K
ha = ,
a
2K
hb = , and
b
2K
hc = .
c
Note that
abc
(ha + hb + hc ) = 2(ab + ac + bc),
K
and
 
abc a 2 + b2 + c2 − ab − ac − bc
9r +
K 4R
18abc
= + a 2 + b2 + c2 − ab − ac − bc.
a+b+c
Therefore, it will suffice to show that
18abc
2(ab + ac + bc) ≥ + a 2 + b2 + c2 − ab − ac − bc,
a+b+c
or equivalently,
f (a, b, c) = 2a 2 b + 2ab2 + 2a 2 c + 2ac2 + 2b2 c + 2bc2 − a 3 − b3 − c3 − 9abc ≥ 0.
Without loss of generality, we may assume that c ≥ b ≥ a. Note that
f (a, b, c) = (a + b − c)(c − a)(c − b) + (3c − a − b)(b − a)2 .
Since a, b, and c are the side lengths of a triangle, a + b − c > 0. Also,
3c − a − b = c + c − a + c − b > 0
VOL. 94, NO. 2, APRIL 2021 153
as well. Hence f (a, b, c) > 0 if c > b or b > a, and consequently f (a, b, c) = 0 can
only occur when a = b = c. This concludes the proof.
Also solved by Arkady Alt, Farrukh Rakhimjanovich Ataev (Uzbekistan), Herb Bailey,
Michel Bataille (France), Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia), Scott H.
Brown, Habib Y. Far, Subhankar Gayen & Vivekananda Mission Mahavidyalaya & Haldia Purba
Medinipur (India), Finbarr Holland (Ireland), Walther Janous (Austria), Parviz Khalili, Koopa
Tak Lun Koo (Hong Kong), Omran Kouba (Syria), Sushanth Sathish Kumar. Elias Lampakis
(Greece), Kee-Wai Lau (China), Antoine Mhanna (Lebanon), Quan Minh Nguyen (Canada),
Sang-Hoon Park (Korea), Volkhard Schindler (Germany), Albert Stadler (Switzerland), Daniel
Văcaru (Romania), Michael Vowe (Switzerland), John Zacharias, and the proposer.

An integral involving the tail of a Maclaurin series April 2020

2092. Proposed by Seán M. Stewart, Bomaderry, Australia.
Let n be a non-negative integer. Evaluate
 ∞ n
1 (−1)k x 2k+1
sin x − dx.
0 x 2n+3 k=0
(2k + 1)!

Solution by Omran Kouba, Higher Institute for Applied Sciences and Technology,
Damascus, Syria.
The answer is
π
(−1)n+1 .
2(2n + 2)!
We define

n
(−1)k x 2k
F2n (x) = (−1)n cos x − , and
k=0
(2k)!


n
(−1)k x 2k+1
F2n+1 (x) = (−1)n sin x −
k=0
(2k + 1)!

One easily sees that Fm = Fm−1 . Further,

Fm (x) = O(x m ) as x → ∞, and
Fm (x) = O(x m+2 ) as x → 0,
so the integral
 ∞
Fm (x)
Im = dx
0 x m+2
is convergent. A straightforward integration by parts shows that
 ∞
−Fm (x) ∞ 1 Fm−1 (x)
Im = + dx
(m + 1)x m+1
x=0 m+1 0 x m+1
1
= Im−1 .
m+1
This implies that
I0
Im = .
(m + 1)!
154 MATHEMATICS MAGAZINE
Another integration by parts gives
 ∞
cos x − 1
I0 = dx
0 x2
 ∞
1 − cos x ∞ sin x
= − dx
x x=0 0 x
 ∞
sin x
=− dx
0 x
π
=− .
2
Thus,
π
Im = − .
2(m + 1)!
In particular,
 ∞ 
n
1 (−1)k x 2k+1
sin x − dx = (−1)n I2n+1
0 x 2n+3 k=0
(2k + 1)!
π
= (−1)n+1 ,
2(2n + 2)!
which is the desired conclusion.
Also solved by Michel Bataille (France), Paul Bracken, Brian Bradie, David M. Bradley, Robert
Calcaterra, William Chang, Robin Chapman (UK), Hongwei Chen, G.A. Edgar, Russell Gordon,
Lixing Han, Eugene A. Herman, Finbarr Holland (Ireland), Sushanth Sathish Kumar, Elias Lam-
pakis (Greece), Kee-Wai Lau (China), Quan Minh Nguyen (Canada), and the proposer. There
were three incomplete or incorrect solutions.

A permutation probability April 2020

2093. Proposed by Jacob Siehler, Gustavus Adolphus College, Saint Peter, MN.
Suppose π is a permutation of {1, 2, . . . , 2m}, where m is a positive integer. Consider
the (possibly empty) subsequence of π(m + 1), π(m + 2), . . . , π(2m) consisting of
only those values which exceed max{π(1), . . . , π(m)}. Let P (m) denote the probabil-
ity that this subsequence never decreases (note that the empty sequence has this prop-
erty), when π is a randomly chosen permutation of {1, . . . , 2m}. Evaluate lim P (m).
m→∞

Solution by José
√ Heber Nieto, Universidad del Zulia, Maracaibo, Venezuela.
The limit is e/2. Let
k = max{π(1), . . . , π(m)}.
Clearly m ≤ k ≤ 2m. A permutation π with a given k satisfies the condition if and only
if k + 1, k + 2, . . . , 2m is a (possibly empty, if k = 2m) subsequence of π(m + 1),
π(m + 2), . . . , π(2m). In the sequence π(1), . . . , π(2m) the number k may occupy any
of the first m positions. The numbers k +  1, k + 2, . . . , 2m may occupy any 2m − k
places among the last m places (i.e., m−k m
possibilities), and the 2m − 1 − (m − k) =
m + k − 1 remaining elements may be distributed in (m + k − 1)! ways. Therefore
 
1 
2m
m
P (m) = m (m + k − 1)!.
(2m)! k=m 2m − k
VOL. 94, NO. 2, APRIL 2021 155
Putting j = k − m we have
 
1 
m
m
P (m) = m (2m − j − 1)!.
(2m)! j =0 j

Now
 
1 m
aj,m = m (2m − j − 1)!
(2m)! j
m(m − 1)(m − 2) · · · (m − j + 1)
= .
2j !(2m − 1) · · · (2m − j )
For fixed j , we have
j −1
(1 − m1 )(1 − m2 ) · · · (1 − )
lim aj,m = lim m
m→∞ m→∞ 2j !(2 − m1 ) · · · (2 − mj )
1
= .
j ! 2j +1
Also
mj 1
aj,m < =
2j !(2m − m)j 2j !
and
∞
1
= e/2.
j =0
2j !

Hence by the dominated convergence theorem we have


m
lim P (m) = lim aj,m
m→∞ m→∞
j =0


∞
= lim aj,m
m→∞
j =0


∞
1
=
j =0
j ! 2j +1
√
e
= ,
2
as claimed.
Also solved by Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia), Robert Calcaterra,
Robin Chapman (UK), Kenneth Schilling, Edward Schmeichel, Albert Stadler (Switzerland), and
the proposer. There was one incomplete or incorrect solution.

An upper bound for a vector sum April 2020

2094. Proposed by George Stoica, Saint John, NB, Canada.
Find the smallest number f (n) such that for any set of unit vectors x1 , . . . , xn in Rn ,
there is a choice of ai ∈ {−1, 1} such that |a1 x1 + · · · + an xn | ≤ f (n).
156 MATHEMATICS MAGAZINE
Solution by Sushanth Sathish √ Kumar, student, Portola High School, Irvine, CA.
We claim that f (n) = n. To see that this is minimal, consider the unit vectors xi =
(0, . . . , 1, . . . , 0), where the ith term is 1 and the rest are 0. Then,
a1 x1 + · · · + an xn = (±1, . . . , ±1)
√
has magnitude n regardless of choice of the ai ’s.
√
We now show that f (n) = n does indeed work. Randomly and independently
choose each ai to be 1 or −1, both with probability 1/2. We will prove that
 
E |a1 x1 + · · · + an xn |2 = n.
To see this, note that
⎡ ⎤
n 
n
E[|a1 x1 + · · · + an xn |2 ] = E ⎣ ai xi · aj xj ⎦
i=1 j =1


n
 2 2 
n 
n
= E ai |xi | + 2 E[ai xi · aj xj ],
i=1 i=1 j =i+1

by the dot product and linearity of expectation. Since ai2 = 1, and xi is a unit vector,
the first sum is just n. To compute the second sum, we note that
E[ai xi · aj xj ] = E[ai aj |xi ||xj | cos θij ]
= E[ai aj cos θij ]
= 0,
where θij is the angle between vectors xi and xj . It follows that
 
E |a1 x1 + · · · + an xn |2 = n,
as claimed. Hence, there is a choice of a1 , . . . , an for which
|a1 x1 + · · · + an xn |2 ≤ n,
and we are done.
Also solved by Elton Bojaxhiu (Germany) & Enkel Hysnelaj (Australia), Robert Calcaterra,
William Chang, Lixing Han, Eugene Herman, Omran Kouba (Syria), Miguel A. Lerma, José Nieto
(Venezuela), Celia Schacht, Albert Stadler (Switzerland), Edward Schmeichel, and the proposer.
There was one incomplete or incorrect solution.

A floor function sum April 2020

2095. Proposed by Mircea Merca, University of Craiova, Romania.
Show that
⎧
⎪ (n + 1)(n − 1)(2n + 3)/24 if d =2
n   ⎪⎨(n + 1)2 (n − 2)/18
⎪
n+1−k if d =3
k = .
d ⎪ (n + 1)(2n + 1)(n − 3)/48
⎪ if d =4
k=1 ⎪
⎩
(n + 1)n(n − 4)/30 if d =5
VOL. 94, NO. 2, APRIL 2021 157
Solution by Russell Gordon, Whitman College, Walla Walla, WA.
We first observe that these four formulas can be combined into one formula by noting
that

n    
n+1−k (n + 1)(n + 1 − d)(2n + 5 − d)
k =
k=1
d 12d

is equivalent to the equation above for d = 2, 3, 4, 5. We will also show that the anal-
ogous formula holds when d = 1. It is easy to verify that the formulas are valid for
n = 1, 2, . . . , d for each of these values of d; we omit the simple arithmetic compu-
tations that generate 0’s and 1’s for these values of n and d. Hence, by induction, it is
sufficient to show that the equation for a given d is valid for n + d when it is valid for
n. To verify this, we will use the fact that

m + x = m + x and m + x = m + x

for any positive integer m and positive number x. We then have

n+d 
 
n+d +1−k
k
k=1
d
n   
n+1−k
= k 1+ + (n + 1)
k=1
d


n+1  n  
n+1−k
= k+ k
k=1 k=1
d
 
(n + 1)(n + 2) (n + 1)(n + 1 − d)(2n + 5 − d)
= +
2 12d
 
(n + 1)(n + 2) (n + 1)(n + 1 − d)(2n + 5 − d)
= +
2 12d
 
(n + 1) 6dn + 12d + 2n2 + (7 − 3d)n + (1 − d)(5 − d)
=
12d
 
(n + 1) 2n2 + (7 + 3d)n + (1 + d)(5 + d)
=
12d
 
(n + 1)(n + 1 + d)(2n + 5 + d)
= ,
12d

as desired.

Remark. The analogous formulas do not hold for d ≥ 6. For example, when d = 6 the
two sides agree for all n, except when n ≡ 0 (mod 6). In that case, we must subtract
1 from the right-hand side to maintain equality.
Also solved by Robert Calcaterra, William Chang, Dmitry Fleischman, Walther Janous (Aus-
tria), Elias Lampakis (Greece), Jacob Petry, Albert Stadler (Switzerland), and the proposer.
 SOLUTIONS

Note that this section includes the solutions to Problems 1241–1244, which would
normally have appeared in the January 2024 issue. The solution to Problem 1245 will
appear in a later issue.

An inequality for a nonincreasing sequence on (0, 1]

1241. Proposed by Reza Farhadian, Razi University, Kermanshah, Iran.

Consider a finite sequence 1 = a0 ≥ a1 ≥ · · · ≥ an+1 > 0 of real numbers. Prove the
following inequality:
√ √
n+1
a0 + a1 + · · · + an+1 < n a0 + a1 + · · · + an .

Solution by Shing Hin Jimmy Pa, China.

We introduce x = a0 + a1 + · · · + an , and apply the AM–GM inequality:
 1 + n(x + an+1 )
n+1
1 · (x + an+1 )n < .
n+1
We also observe that nan+1 + 1 ≤ x. Add nx to both sides and divide by n + 1:
1 + n(x + an+1 )
≤ x.
n+1
Thus,

n+1
(x + an+1 )n < x,
√ √
which is equivalent to x + an+1
n+1
< n x, as desired.
Also solved by Michel Bataille, Rouen, France; Walther Janous, Ursulinengymnasium, Innsbruck, Austria;
Lau Kee-Wai, Hong Kong, China; JHSLPN Group, North Carolina School of Science and Mathematics, Durham,
NC; Naı̈m Mégarbané, Lycée Stanislas High School, Paris, France; Albert Stadler, Herrliberg, Switzerland;
and the proposer.

176 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 Solution to a transcendental equation

1242. Proposed by Adam Glesser, California State University, Fullerton, CA.

Find all solutions to the following transcendantal equation (via proof, not by computer
calculation):
 4   
6 x + 48 3
arcsin exp 1 + x + √ = 1.
πe6 64x 3
x−1

Solution by Robert Doucette, McNeese State University, Lake Charles, LA.

√
Let α(x) = (x 4 + 48)/(64x), β(x) = 1 + x + 3/ 3 x − 1, and φ(x) =
arcsin(α(x)) exp(β(x)). Since φ(2) = (π/6)e6 , the given equation can be rewritten as
φ(x) = φ(2). Since φ(x) > 0 only if x > 0, we need only consider positive numbers
as possible solutions. Since
 
3 x2 1 1
α (x) = − 2 and β (x) = 1 − ,
4 16 x (x − 1)4/3

the function α is strictly decreasing on (0, 2) and strictly increasing on (2, ∞), while
the function β is strictly decreasing on [0, 1) and on (1, 2) (note the pole at x = 1!) and
strictly increasing on (2, ∞). Since α(x) → ∞ as x → 0+ and α(1) < 1, there exists a
unique x0 ∈ (0, 1) such that α(x0 ) = 1. Also since α(2) = 1/2 and α(x) → ∞ as x →
∞, there exists a unique x1 ∈ (2, ∞) such that α(x1 ) = 1. The set [x0 , 1) ∪ (1, x1 ] are
the positive numbers for which φ is defined.
For x ∈ [x0 , 1), β(x) < β(0) = −2. This implies that φ(x) < (π/2)e−2 < φ(2), so
there are no solutions to the given equation in the interval [x0 , 1).
Since both α and β are positive and strictly decreasing on the interval (1, 2), the
functions arcsin(α) and exp(β) are both positive and strictly decreasing on (1, 2). It
follows that φ is strictly decreasing on (1, 2) and that φ(x) > φ(2) for x ∈ (1, 2).
In a similar way we may show that φ is strictly increasing on (2, x1 ], so that φ(x) >
φ(2) for x ∈ (2, x1 ].
It follows that 2 is the unique solution to the given equation.

Also solved by Naı̈m Mégarbané, Lycée Stanislas High School, Paris, France; Albert Stadler, Switzerland;
and the proposer. Received one incomplete solution.

An integral inequality

1243. Proposed by Cezar Lupu, Yanqi Lak Bimsa and Tsinghua University, Beijing,
China.
 1
Let f : [0, 1] → R be an integrable function such that f (x) dx = 1 and
 1  1 0

x 2 f (x) dx = 1. Prove that f 2 (x) dx ≥ 6.

0 0

VOL. 55, NO. 2, MARCH 2024 THE COLLEGE MATHEMATICS JOURNAL 177
Solution by Mark Sand, College of St. Mary, Omaha, NE.
Given such a function f (x), we know that for any real number r,
 1
1+r = (1 + r · x 2 )f (x) dx
0
 1
≤ (1 + r · x 2 )f (x) dx
0
 1 1/2  1 1/2
≤ (1 + r · x 2 )2 dx f 2 (x) dx ,
0 0

where we have used the Cauchy-Schwarz inequality in the last step. The integral
that includes the number r has the value 1 + 23 r + 15 r 2 . Dividing by this and then
squaring and simplifying, we see that
 1
15(r + 1)2
f 2 (x) dx ≥ .
0 3r 2 + 10r + 15

Since the left side is fixed once f (x) is given, this inequality must be true for all values
of the fraction on the right, including the maximum value of the fraction. We note here
that the denominator is never zero, which can be seen by completing the square.
2 60(r 2 +6r+5)
2 +10r+15 , we find G (r) = (3r 2 +10r+15)2 , so the derivative is zero when
Letting G(r) = 3r15(r+1)
r is −1 or −5. The only term in the derivative that changes sign is (r 2 + 6r + 5), and
we can easily see that G (r) is positive on (−∞, −5) ∪ (−1, ∞) and negative on
(−5, −1). This, along with lim G(r) = 5, tells us that G(−5) = 6 is the maximum
r→∞
value of the fraction we have been investigating.
 1
Thus, f 2 (x) dx ≥ 6, as desired.
0
Also solved by Michel Bataille, Rouen, France; Russell Gordon, Whitman University; Tom Jager, Calvin
University; Walther Janous, Ursulinengymnasium, Innsbruck, Austria; Kee-Wai Lau, Hong Kong, China;
Michael Lavigne, North Carolina School of Math and Science; Kelly McLenithan, Los Alamos, NM;
Albert Stadler, Herrliberg, Switzerland; and the proposer.

Distribution of fractional parts of uniform variables

1244. Proposed by Albert Natian, Los Angeles Valley College, Valley Glen, CA.
Suppose (Xk )n1 is a sequence of n independent random variables uniformly distributed
over the interval [0, 1]. Prove that the fractional part of the random variable nk=1 Xi
is uniformly distributed over [0, 1].

Solution by Eagle Problem Solvers, Georgia Southern University, Statesboro, GA.

If n = 1, then X1 is equal to its fractional part, and the statement is trivially true.
We claim that if X and Y are independent random variables uniformly distributed
over [0, 1], then the fractional component of the sum W = X + Y is also uniformly
distributed on [0, 1]. First, we express the fractional component of the sum as

W = (X + Y )1{X+Y <1} + (X + Y − 1)1{X+Y >1} .

178 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Then for w ∈ [0, 1],

Pr ( W ≤ w) = Pr (((X + Y ≤ w) ∩ (X + Y < 1)) ∪ ((X + Y − 1 ≤ w) ∩ (X + Y > 1)))

= Pr ((X + Y ≤ w) ∩ (X + Y < 1)) + Pr (0 < X + Y − 1 ≤ w)
 
1 2 1 1
= w + − (1 − w) 2
2 2 2
1 2 1
= w + w − w2
2 2
= w.

The probabilities are illustrated geometrically in the following diagram.

Therefore, the fractional part of W = X + Y is uniformly distributed over [0, 1], and
the general statement follows by induction with X = nk=1 Xk and Y = Xn+1 .
Also solved by Michael P. Cohen, Fairfax, VA; Jan Grzesik, Torrance, CA; Shing Hin Jimmy Pa, China;
Albert Stadler, Herrliberg, Switzerland; and the proposer.

VOL. 55, NO. 2, MARCH 2024 THE COLLEGE MATHEMATICS JOURNAL 179
 SOLUTIONS
To our valued contributors: CMJ Solutions is in transition.
Charles N. Curtis, who has served as Solutions editor for nearly 10 years, is retiring
from this position. I am thankful for his valuable service on the CMJ board over these
many years. I’m certain that everyone associated with CMJ has been grateful for his
leadership and expertise.
I am pleased to announce that Katherine Thompson and Matyas Sustik are joining
the CMJ editorial board as our new Solutions editors. Both bring significant experience
with problem solving competitions, and I am looking forward to working with them.
Dr. Thompson is currently Assistant Professor of Mathematics at the U.S. Naval
Academy. She is a regular instructor and grader for the Art of Problem Solving and is
the former chair of the Question Writing Committee for MATHCOUNTS.
Dr. Sustik works in industry as a mathematician and software engineer. With an
active and successful high school math contest participation behind him (that included
the IMO) now he gives back by developing, grading, and evaluating mathematical
contest problems for AMC, AIME, and BAMO.
I currently expect CMJ Solutions to return in the March issue. It will take a couple
of issues for us to catch up and resume our typical schedule. I ask for your patience as
we complete this transition.

— Tamara Lakins, Editor

VOL. 55, NO. 1, JANUARY 2024 THE COLLEGE MATHEMATICS JOURNAL 65

1265. Proposed by Narendra Bhandari, Bajura District, Nepal.
Calculate the following sum:
2 n

∞ 
ζ (n) − ζ (n + 1)
,
n=2 k=1
k

where [ x ] denotes the floor function of x and ζ denotes the Riemann zeta function.

SOLUTIONS

The centroid of a tetrahedron

1236. Proposed by Tran Quang Hung, Vietnam National University, Hanoi, Vietnam.
Let ABCD be a tetrahedron in 3-space, and let P , Q and R be three collinear points.
Assume that lines P A, P B, P C, and P D are not parallel to planes (BCD), (CDA),
(DAB), and (ABC), respectively. Line P A meets plane (BCD) at point A1 . In the
plane (AP R), assume that the two lines AR and A1 Q intersect at A2 . Point A3 lies
on line P A2 such that RA3 is parallel to line AA1 . Define similarly the points B1 , B2 ,
B3 , C1 , C2 , C3 , D1 , D2 , and D3 . Prove that R is the centroid of tetrahedron A3 B3 C3 D3
(see figure).

Solution by the proposer.

Proof. Let AB and BA denote signed lengths of segments. Apply the theorem of
Menelaus to AP R with transversal A1 A2 Q to get

A1 P A2 A QR
·· = 1. (1)
A1 A A2 R QP
Let the Euclidean vector connecting an initial point X with a terminal point Y be
−→
denoted by XY . Let x, y, z, t, not all zero, such that

x P A + y P B + zP C + t P D = 0. (2)

VOL. 54, NO. 5, NOVEMBER 2023 THE COLLEGE MATHEMATICS JOURNAL 493
If x + y + z + t = 0, then equation (2) becomes

−(y + z + t)P A + y P B + zP C + t P D = 0,

which implies

y AB + zAC + t AD = 0.
 
This would mean that AB, AC, AD is linearly dependent, which is impossible since
A, B, C, and D are not coplanar. Therefore, we have x + y + z + t = 0.
Note that applying projections parallel to line AP onto plane (BCD), one has

A, P → A1 , B → B, C → C, D → D. (3)

Since parallel projection is an affine transformation, it follows from (2) and (3),

y A1 B + zA1 C + t A1 D = 0. (4)

From (4), we may deduce y P B + zP C + t P D = (y + z + t)P A1 . Combining with

(2), we obtain

−x P A = (y + z + t)P A1 ,

or

x A1 A = (x + y + z + t)P A1 .

From this,

A1 P −x
= . (5)
A1 A x+y+z+t

Let

QR
= k. (6)
QP

It follows from (1), (5), and (6),

A2 R −kx
= .
A2 A x+y+z+t

Since RA3 P A, applying the theorem of Thales yields

RA3 A2 R kx · P A
RA3 = · PA = · PA = . (7)
PA AA2 x+y+z+t

Similarly, we have

ky · P B kz · P C kt · P D
RB3 = , RC3 = , RD3 = . (8)
x+y+z+t x+y+z+t x+y+z+t

494 © THE MATHEMATICAL ASSOCIATION OF AMERICA

From (7) and (8), we get
 
k x P A + y P B + zP C + t P D
RA3 + RB3 + RC3 + RD3 = = 0,
x+y+z+t
which implies R is the centroid of A3 B3 C3 D3 . This completes the proof. 
No other solutions were received.

Two polygons
1237. Proposed by Tran Quang Hung, Vietnam National University, Hanoi, Vietnam.
Let A1 A2 . . .A2n and A1 A2 . . .A2n (n ≥ 2) be two directly 2n-regular polygons. Prove
 n−1
that ni=1 A2i A22i =
2
i=0 A2i+1 A2i+1 (see figure).

Solution by Albert Stadler, Herrliberg, Switzerland.

We may assume (without loss of generality) that the vertices of the two polygons are
given by
π ik +iω π ik +iω
Ak = re n , and Ak = 1 + r  e n , for k = 1, 2, . . ., 2n.
Then

n
2

n−1
2
A2i A2i − A2i+1 A2i+1
i=1 i=0


n
2 2
= A2k A2k − A2k−1 A2k−1
k=1


n  π i(2k)  2
  n +iω π i(2k) 
= r e + 1 − re n +iω 
k=1
 π i(2k−1)  2 
  +iω π i(2k−1)
+iω 
− r e n + 1 − re n 

VOL. 54, NO. 5, NOVEMBER 2023 THE COLLEGE MATHEMATICS JOURNAL 495
 
n
2π ik −iω π i(2k−1) 2π ik +iω
= −re− n + re− n −iω
− re n

k=1

π i(2k−1)
+iω 2π ik −iω
+re n + r  e− n

π i(2k−1)
−iω 2π ik +iω π i(2k−1)
+iω
−r  e− n + r e n − r e n = 0,

n 2π ik
since k=1 e n = 0.
Also solved by Dmitry Fleischman, Santa Monica, CA; Eugene Herman, Grinnell C.; and the proposer.

Rotated squares
1238. Proposed by Jacob Siehler, Gustavus Adolphus College, St. Peter, MN.
Consider the intersection of a unit square with a copy of itself rotated through an angle
of θ about their mutual center. Note that in general, this region is an octagon. Evaluate
the average area of the intersection as θ ranges from 0 to π2 .
Solution by Kyle Calderhead, Malone University, Canton, Ohio.

By extending lines from the mutual center to the midpoints of each side of each
square, as well as to the points of intersection of their sides, we can decompose the
octagonal intersection into sixteen right triangles—eight with a leg of length 12 and
adjacent angle of θ2 , and eight more with a leg of length 12 and adjacent angle of π4 − θ2 .
In the figure above, one of each of these types of triangles has been highlighted.
Using right-triangle trigonometry, we see that the length of the other legs of these
triangles are 12 tan θ and 12 tan π4 − θ2 , respectively. Hence the areas of each type of
triangle are 12 · 12 · 12 tan θ2 and 12 · 12 · 12 tan π4 − θ2 , respectively. With eight of each,
we have a total area of
   
θ π θ
A = tan + tan − .
2 4 2

496 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Note that this formula is consistent with the situation where the squares coincide, cor-
responding to θ being equal to either 0 or π2 .
π
Before taking the average value, note that the integral 0 2 tan π4 − θ2 dθ can be
π
shown to be equivalent to the integral 0 2 tan θ2 dθ by means of the substitution
u = π2 − θ. This simplifies the average value calculation to
      π
 θ  2
π
1 2 θ 4 4 ln 2
2 tan dθ = −2 ln cos = ,
π/2 0 2 π 2  0 π
or approximately 0.8825.
Note: Using the same dissection technique, we can show that in the more gen-
eral case of two overlapping regular n-gons with unit area, the average area of their
intersection will be πn cot πn ln sec2 πn .
Also solved by Ricardo Alfaro, U. of Michigan - Flint; Andrew Bauman, U. of Arkansas at Little Rock;
Nate Belgard, The Barrie School; Brian Bradie, Christopher Newport U.; Rob Downes, Newark Academy;
Bill Dunn , Montgomery C.; Eagle Problem Solvers, Georgia Southern U.; Habib Far, Lone Star C. -
Montgomery; Dmitry Fleischman, Santa Monica, CA; Michael Goldenberg, Reiserstown, MD and Mark
Kaplan, U. of Maryland Global Campus (jointly); Aakash Gurung, Asahi Nago, and Xuan Pham (jointly);
Spencer Harris, Westmont C. (graduate); Eugene Herman, Grinnell C.; Stephen Herschkorn, Rutgers U.;
Liam Mauck and Clayton Coe, Cal Poly Pomona Problem Solving Group; Kelly McLenithan, Los Alamos,
NM; Peter Oman and Haohao Wang, Southeast Missouri St. U.; Leah Ramos (student), Seton Hall U.;
Volkhard Schindler, Berlin, Germany; Skidmore C. Problem Group; Albert Stadler, Herrliberg, Switzer-
land; and the proposer.

An explicit formula for a sequence from a recursion

1239. Proposed by Moubinool Omarjee, Lycée Henry IV, Paris, France.
√
u3n −3un − 5
Let u0 be a positive real number, and for every n ∈ N, define un+1 := 2
√
3un +3 5un +4
.
Find a closed-form expression for un in terms of u0 and n.
Solution by the Stephen Locke, Florida Atlantic University.

x3
Lemma 1. Let g(x) = . Then, the kth iterate g (k) of g is given by
(x + 1)3 − x 3
k
x3
g (x) =
(k)
.
(x + 1)3k − x 3k
k
x3
Proof. We note that g (1) = g and assume that for some k, g (k) (x) = .
(x + 1)3k − x 3k
Then,
 k

x3
g (k+1)
(x) = g
(x + 1)3k − x 3k
 k
3 ⎛ 3  3 ⎞−1
3k 3k
x3 ⎝ x x ⎠
= +1 −
(x + 1)3k − x 3k (x + 1)3k − x 3k (x + 1)3k − x 3k

VOL. 54, NO. 5, NOVEMBER 2023 THE COLLEGE MATHEMATICS JOURNAL 497
   3  k 3 −1
3k+1 3k 3k 3k
=x x + (x + 1) − x − x3

 3 −1
3k+1 3k 3k+1
=x x + 1) −x ,

establishing the inductive proof. 

√ √
w3 − 3w − 5 1− 5
Now, let f (w) = √ , so that un+1 = f (un ), and let τ = .
3w2 + 3 5w + 4 2
3 3
w w
Note that f (τ + w) = τ + =τ+ = τ + g(w). Hence,
3w + 3w + 1
2 (w + 1)3 − w3
for w = u0 − τ ,
n
w3
un = f (n) (τ + w) = τ + g (n) (w) = τ + ,
(w + 1)3n − w3n

providing a closed form for un in terms of u0 and n.

Also solved by Brian Bradie, Christopher Newport U.; Michael Goldenberg, Reistertown, MD and Mark
Kaplan, U. of Maryland Globan Campus (jointly); Albert Stadler, Herrliberg, Switzerland; and the proposer.

Fields for which the collection of additive subgroups and the collection
of multiplicative subgroups are isomorphic
1240. Proposed by Greg Oman, University of Colorado at Colorado Springs, Colorado
Springs, CO.
Let S be a set. Recall that a partial order on S is a binary relation ≤ which is re-
flexive, anti-symmetric, and transitive. If S, T are sets and ≤,  are partial orders on
S and T , respectively, then we say that the partially ordered set (S, ≤) and (T , )
are isomorphic if there is a bijection f : S → T such that for all s1 , s2 ∈ S: s1 ≤ s2
iff f (s1 )  f (s2 ). Now let F be a field, and let P + (F ) be the collection of additive
subgroups of F , partially ordered by set-theoretic inclusion, and let P × (F ) be the col-
lection of multiplicative subgroups of F × := F \{0}, partially ordered by inclusion.
Find all fields F for which P + (F ) and P × (F ) are isomorphic.
Solution by Anthony Bevelacqua, University of North Dakota, Grand Forks, North
Dakota.
Any subgroup H of F × corresponds to an additive subgroup A of F in such a
way that the subgroup lattices of H and A are isomorphic. Consequently the trivial
subgroup 1 of F × must correspond to the trivial subgroup 0 of F . Since a group is
finite if and only if it has finitely many subgroups, finite subgroups of F × correspond
to finite additive subgroups of F . Since a field of characteristic zero has a nontrivial
finite multiplicative subgroup (namely {1, −1}) and every nontrivial additive subgroup
of a field of characteristic zero is infinite, F must have characteristic p > 0. Thus Zp ,
the field with p elements, is a subfield of F . We note that the additive subgroups of F
are precisely the Zp -subspaces of F .
Assume dimZp F > 1. Then F contains a subspace A of dimension two. A contains
exactly p + 1 proper, nontrivial subgroups, no one of which is contained in another.
Now A corresponds to a finite subgroup H of F × with exactly p + 1 proper, nontrivial
subgroups, no one of which is contained in another. Recall that J → |J | gives an

498 © THE MATHEMATICAL ASSOCIATION OF AMERICA

isomorphism between the lattice of subgroups of a cyclic group of order n and the
lattice of positive divisors of n ordered by divisibility. Since H has p + 1 ≥ 3 proper,
nontrivial subgroups, no one of which is contained in another, |H | must be divisible
by (at least) three distinct primes q, r, and s. Now q is a proper divisor of qr and qr
is a proper divisor qrs, so H contains a pair of nested proper, nontrivial subgroups, a
contradiction.
Thus F = Zp . Since the additive group Zp has exactly two subgroups, Z× p has
exactly two subgroups. Therefore p − 1 = |Z× p | is a prime, and so p = 3. Hence Z3 is
the only field F for which P + (F ) and P × (F ) are isomorphic.
Also solved by the proposer.

VOL. 54, NO. 5, NOVEMBER 2023 THE COLLEGE MATHEMATICS JOURNAL 499
1260. Proposed by Nick Fiala, St. Cloud State University, St. Cloud, MN and Greg
Oman, University of Colorado, Colorado Springs, CO.
Recall that an associative ring R is a division ring provided R is a ring with identity
1  = 0 and every nonzero element of R is invertible. Consider dropping the axiom that
every member of R has an additive inverse. Let’s call a division ring for which we
don’t assume additive inverses negative poor. Prove that if R is a finite negative poor
division ring with more than two elements, then every member of R has an additive
inverse.

SOLUTIONS

An inequality for the angles of a triangle

1231. Proposed by George Apostolopoulos, Messolongi, Greece.
 √
Let ABC be a triangle. Show that α=A,B,C sin3 (α) cos(α) ≤ 9163 .
Solution by John Christopher, California State University, Sacramento.
Note that each angle of triangle ABC lies in the interval (0, π). Using first semester
calculus, it is easily shown that in the interval (0, π), the function f (x) = sin3 x cos x
√ 3 √
3 1 3 3
attains its maximum value when x = π/3. Since f (π/3) = · = ,
2 2 16
√ √ √ √
3 3 3 3 3 3 9 3
we have f (∠A) + f (∠B) + f (∠C) ≤ + + = . Equality is at-
16 16 16 16
tained when the triangle is equilateral and each angle is π/3.
Also solved by Ulrich Abel and Vitaliy Kushnirevych, Technische Hochschule Mittelhessen, Germany;
Michel Bataille, Rouen, France; Paul Bracken, U. of Texas, Edinburg; Brian Bradie, Christopher New-
port U.; Charles Burnette, Xavier U. of Louisiana; M. V. Channakeshava, Bengaluru, India; Ritabrato
Chaterjee (student), Western Michigan U.; Danko Dmitry (student), RUDN U., Moscow, Russia; Eagle
Problem Solvers, Georgia Southern U.; The Episcopal Academy Problem Solvers; Habib Far, Lone Star
C. - Montgomery; Meagan Fisher, Anna Phillips, Juan Martinez, and William French (students), U.
of Arkansas at Little Rock; Fresno State Journal Problem Solving Group; Shubham Goel, GGSIPU,
Uttar Pradesh, India; Michael Goldenberg, Reierstown, MD and Mark Kaplan, U. of Maryland Global
Campus (jointly); Russ Gordon, Whitman C.; Jacob Guerra, Lowell, MA; Eugene Herman, Grinnell C.;
Walther Janous, Ursulinengymnasium, Innsbruck, Austria; A. Bathi Kasturiararchi, Kent St. U. at Stark;
Hidefumi Katsuura, San Jose St. U.; Parviz Khalili, Newport News, VA; Joseph Klaips (student), North
Central C.; Panagiotis Krasopoulos, Athens, Greece; Wei-Kai Lai, U. of South Carolina Salkehatchie; Kee-
Wai Lau, Hong Kong, China; Shing Hin Jimmy Pa; Paolo Perfetti, Universitá degli studi di Tor Vergata
Roma; Chrysostom Petalas, Ioannina, Greece; Volkhard Schindler, Berlin, Germany; Joel Schlosberg,
Bayside, NY; Digby Smith, Waterton Lakes Mathematics Guild; Southeast Missouri State U. Math Club,
; Albert Stadler, Herrliberg, Switzerland; Michael Vowe, Therwil, Switzerland; and the proposer. One in-
complete solution was received.

The Catalan numbers

1232. Proposed by Jacob Guerra, Salem State University, Salem, MA.

VOL. 54, NO. 4, SEPTEMBER 2023 THE COLLEGE MATHEMATICS JOURNAL 401
 1 2n
Define, for every nonnegative integer n, the nth Catalan number by Cn := n+1 n
.
Consider the sequence of complex polynomials in z defined by zk := zk−1 2
+ z for
every nonnegative integer k, where z0 := z. It is clear that zk has degree 2k and thus
k
has the representation zk = 2n=1 Mn,k zn , where each Mn,k is a positive integer. Prove
that Mn,k = Cn−1 for 1 ≤ n ≤ k + 1.
Solution by Charles Burnette, Xavier University of Louisiana, New Orleans, LA.
We proceed by induction on k, noting that for the base case k = 0, we have M1,0 =
1 = C0 . For the induction step, suppose that Mn,r = Cn−1 for 1 ≤ n ≤ r + 1, where r
is a nonnegative integer. Observe that
 2r
 2r 2r+1
 n−1
zr+1 = Mn,r z n
Mn,r z n
+z= Mm,r Mn−m,r zn + z,
n=1 n=1 n=2 m=1

= 1 = C0 . Furthermore, because the Catalan numbers satisfy the

so that then M1,r+1 
recurrence Cn+1 = nm=0 Cm Cn−m , we find that

n−1 n−1 n−2

Mn,r+1 = Mm,r Mn−m,r = Cm−1 Cn−m−1 = Cm Cn−m−2 = Cn−1
m=1 m=1 m=0

for 1 ≤ n ≤ r + 2. Also solved by Ulrich Abel, Technische Hochschule Mittelhessen, Germany; Cal
Poly Pomona Problem Solving Group; Hongwei Chen, Christopher Newport U.; Eagle Problem Solvers,
Georgia Southern U.; Michael Goldenberg, Reisterstown, MD and Mark Kaplan, U. of Maryland Global
Campus (jointly); Eugene Herman, Grinnell C.; Walther Janous, Innsbruck, Austria; Panagiotis Krasopou-
los, Athens, Greece; Shing Hin Jimmy Pa; John Quintanilla, U. of North Texas; Ajay Srinivasan, U. of
Southern California; Albert Stadler, Herrliberg, Switzerland; Dan Swenson, Black Hills St. U.; and the pro-
poser. One incomplete solution was received.

Uniform random variables

1233. Proposed by Albert Natian, Los Angeles Valley College, Valley Glen, CA.
Suppose that X and Y are independent, uniform random variables over [0, 1]. Define
UX , VX , and BX as follows: UX is uniform over [0, X], VX is uniform over [X, 1], and
BX ∈ {0, 1}, with P (BX = 1) = X, and P (BX ) = 0 = 1 − X. Now define random
variables Z and WX as follows:

Z = Y − X1{Y ≥ X} + (1 − X + Y )1{Y < X}, and

WX = BX · UX + (1 − BX )VX .

Prove that both Z and WX are uniform over [0, 1]. Here, 1[S] is the indicator function
that is equal to 1 if S is true and 0 otherwise. Solution by John Quintanilla, University
of North Texas, Denton, Texas.
We proceed by induction on k. The statement clearly holds for k = 1:

z1 = z02 + z = z + z2 = C0 z + C1 z2 .

402 © THE MATHEMATICAL ASSOCIATION OF AMERICA

We now assume that, for some k ≥ 1, Mn,k = Cn−1 for all 1 ≤ n ≤ k + 1, and we
define

k
2
zk+1 = z + M1,k z + M2,k z2 + M3,k z3 + · · · + M2k ,k z2

Our goal is to show that Mn,k+1 = Cn−1 for n = 1, 2, . . . , k + 2.

For n = 1, the coefficient M1,k+1 of z in zk+1 is clearly 1, or C0 . For 2 ≤ n ≤ k + 2,
the coefficient Mn,k+1 of zn in zk+1 can be found by expanding the above square; every
product of the form Mj,k zj · Mn−j,k zn−j will contribute to the term Mn,k+1 zn . Since
n ≤ k + 2 ≤ 2k + 1 (since k ≥ 1), the values of j that will contribute to this term
will be j = 1, 2, . . . , n − 1. (Ordinarily, the z0 and zn terms would also contribute;
however, there is no z0 term in the expression being squared). Therefore,

n−1
Mn,k+1 = Mj,k Mn−j,k
j =1

n−1
= Cj −1 Cn−j −1 by induction hypothesis
j =1

n−2
= Cj Cn−2−j after reindexing
j =0

= Cn−1 ,

where we used a well-known recursive relationship for the Catalan numbers in the last
step. 
Also solved by Robert Agnew, Palm Coast, FL; Charles Burnette, Xavier U. of Louisiana; Dmitry Fleis-
chman, Santa Monica, CA; Missouri St. U. Problem Solving Group; Northwestern U. Math Problem
Solving Group; Rob Pratt, Apex, NC; Ajay Srinivasan, U. of Southern California; Dan Swenson, Black
Hills St. U.; and the proposer.

The limit of a quotient of sequences defined by sums

1234. Proposed by Moubinool Omarjee, Lycée Henry IV, Paris, France.
 
For every positive integer n, set an := nk=1 k14 and bn := nk=1 (2k−1) 1
4 . Compute

limn→∞ ( an − 16 ).
bn 15

Solution by Russelle Guadalupe (student), University of the Philippines, Diliman, Que-

zon City, Philippines.
We note that for integers n ≥ 1,

2n n n
1 1 1 an
= + = + bn , and
k=1
k4 k=1
(2k)4 k=1
(2k − 1) 4 16
2n n 2n n
1 1 1 1
= + = an + .
k=1
k4 k=1
k 4
k=n+1
k4 k=1
(n + k)4

VOL. 54, NO. 4, SEPTEMBER 2023 THE COLLEGE MATHEMATICS JOURNAL 403
Thus, we have
n
1 an 15
= bn + − an = bn − an
k=1
(n + k) 4 16 16

and
  n n
bn 15 n3 1 1 1 1
lim n3 − = lim = lim · .
n→∞ an 16 n→∞ an
k=1
(n + k)4 n→∞ an n k=1
(1 + k/n)4

π4
Since it is well-known that an approaches 90
as n → ∞ and
n
1 1
lim
n→∞ n
k=1
(1 + k/n)4
2
is the limit of a Riemann sum, which is given by the definite integral 1 x −4 dx, we
obtain
    
bn 15 90 2 dx 30 1 105
lim n3
− = 4 4
= 4 1− = .
n→∞ an 16 π 1 x π 8 4π 4
Also solved by Robert Agnew, Palm Coast, FL; Michel Bataille, Rouen, France; Paul Bracken, U. of Texas,
Edinburg (2 solutions); Brian Bradie, Christopher Newport U. (2 solutions); Ritabrato Chaterjee, Western
Michigan U. (2 solutions); Hongwei Chen, Christopher Newport U. ; Giuseppe Fera, Vicenza, Italy; Dmitry
Fleischman, Santa Monica, CA; Michael Goldenberg, Reistertown, MD and Mark Kaplan, U. of Mary-
land Globan Campus (jointly); Russ Gordon, Whitman C.; Eugene Herman, Grinnell C.; Eugen Ionaşcu,
; Walther Janous, Ursulinengymnasium, Innsbruck, Austria; Stephen Kaczkowski, South Carolina Gover-
nor’s S. for Science and Mathematics; A. Bathi Kasturiararchi, Kent St. U.; Yoodam Kim, Seoul National
U. of Science and Technology; Kee-Wai Lau, Hong Kong, China; Missouri St. U. Problem Solving Group;
Ángel Plaza, Universidad de Las Palmas de Gran Canaria, Spain; Mark Sand, C. of Saint Mary; Kenneth
Schilling, U. of Michigan - Flint; Volkhard Schindler, Berlin, Germany; Ajay Srinivasan, U. of Southern
California; Albert Stadler, Herrliberg, Switzerland; Seán Stewart, King Abdullah U. of Science and Tech-
nology, Saudi Arabia; Southeast Missouri St. U. Math Club; Michael Vowe, Therwil, Switzerland; and the
proposer. Three incorrect solutions were received.

Non-finitely generated sets whose proper subsets closed under a given

function are all finitely generated
1235. Proposed by Greg Oman, University of Colorado at Colorado Springs, Colorado
Springs, CO.
Let S be a set, and let f : S → S be a function. For s ∈ S, the orbit of s is defined by
O(s) := {f n (s) : n ≥ 0}, where f0 : S → S is the identity map and f n is the n-fold
composition of f with itself for n > 0. A subset X ⊆ S is closed under f provided that
for all x ∈ X, also f (x) ∈ X. Finally, if X is closed under
 F , we say that X is finitely
generated if there is a finite F ⊆ X such that X = x∈F O(x). Find all structures
(S, f ) up to isomorphism where S is not finitely generated, but every proper subset of
S closed under f is finitely generated. Note that (S, f ) and (T , g) are isomorphic if
there is a bijection ϕ : S → T such that ϕ(f (s)) = g(ϕ(s)) for all s ∈ S. Solution by
Kenneth Schilling, University of Michigan-Flint.

404 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 Let f : S → S be as described in the proposal.
First note that f : S → S is surjective, for if t ∈ S \ f (S), then S \ {t} is invariant,
and if S \ {t} were finitely generated, then so would be S.
Second, say s ∈ S is of finite order there exists k > 0 with f k (s) = s, and of infinite
order if no such k exists. We claim that there exists s ∈ S of infinite order. Suppose
to the contrary that every s ∈ S is of finite order. Then it is clear that every orbit is
finite. Furthermore, the orbits are disjoint, for if O(s) ∩ O(t)  = ∅, then there exist
i, j, k such that f i (s) = f j (t) and f k (s) = s. Then s = f ik (s) = f i(k−1) ◦ f i (s) =
f i(k−1) ◦ f j (t), so s ∈ O(t), and by symmetry t ∈ O(s), so O(s) = O(t). Now S is
a disjoint union of finite orbits, so any union of infinitely many but not all orbits is a
proper closed but not finitely generated subset of S, contrary to hypothesis.
Let s0 ∈ S be of infinite order. For n > 0, let sn = f n (s). Choose s−1 so that
f (s−1 ) = s0 , then choose s−2 so that f (s−2 ) = s−1 , then choose s−3 so that f (s−3 ) =
s−2 , and so on. The doubly infinite sequence s = {sn : n ∈ Z} is closed under f . For
n < 0, sn is of infinite order, for if f k (sn ) = sn , then s0 = f −n (sn ) = f k−n (s0 ), con-
trary to the fact that s0 is of infinite order. It follows that s is not finitely generated; for
n < 0, sn is not in the orbit of sm for m > n. Therefore s = S.
The structure (S, f ) is isomorphic to one of the following:

S=Z  and f (z) = z + 1, or for some positive integer m, S = {z ∈ Z : z ≤ m} and

z + 1 for z < m
f (z) = .
1 for z = m
Also solved by Eugen Ionaşcu, Columbus St. U.; Dan Swenson, Black Hills St. U.; and the proposer.

VOL. 54, NO. 4, SEPTEMBER 2023 THE COLLEGE MATHEMATICS JOURNAL 405
 SOLUTIONS

An easy logarithmic inequality

1226. Proposed by George Apostolopoulos, Messolongi, Greece.
(a−b)2 +(b−c)2 +(c−a)2
3 ≤
27abc
Let a, b, and c be positive real numbers. Prove that ln (a+b+c) 3
.
Solution by Shing Hin Jimmy Pa.

 
27abc abc
ln = ln
(a + b + c)3 a+b+c 3
3
 
abc
≤ ln (AM-GM Inequality)
(abc)3/3
=0
(a − b)2 + (b − c)2 + (c − a)2
≤ .
3
Also solved by F. R. Ataev, Westminster International U. in Tashkent; Michel Bataille, Rouen, France; So-
ham Bhadra (student), Patha Bhavan, India; Connor Chambers, Rohan Dalal, Jonathan Hong, Kassidy
Kryukov, Dylan Lorello (students), Tommy Goebeler, and Molly Konopka, The Episcopal Academy;
Carson Dorough, Cuesta C.; Habib Far, Lone Star C. - Montgomery; Dmitry Fleischman, Santa Monica,
CA; Philip Wagala Gwanyama, Northeastern Illinois U.; Eugene Herman, Grinnell C.; Donald Hooley,
Bluffton, OH; Walther Janous, Ursulinengymnasium, Innsbruck, Austria; A. Bathi Kasturiarachi, Kent

VOL. 54, NO. 4, SEPTEMBER 2023 THE COLLEGE MATHEMATICS JOURNAL 395
St. U. at Stark; Hidefumi Katsuura, San Jose St. U.; Panagiotis Krasopoulos, Athens, Greece; Wei-Kai
Lai, U. of S. Carolina Salkehatchie and JOhn Risher (graduate student), C. of Charleston; Mihat Mammadli;
Kelly McLenithan, Los Alamos, NM; Antoine Mhanna, Lebanon; Paolo Perfetti, Universitá degli studi
di Tor Vergata Roma; Benjamin Phillabaum; Henry Ricardo, Westchester Area Math Circle; Digby Smith,
Waterton Lakes Mathematics Guild; Southeast Missouri St. U. Math Club; Albert Stadler, Herrliberg,
Switzerland; Kwame Yeboah and Fatema Ruhi, Southeast Missouri St. U.; and the proposer.

Nonexistence of a pair of functions with intertwined inequalities

1227. Proposed by Albert Natian, Los Angeles Valley College, Valley Glen, CA.
Do there exist functions f : (0, 1) → R and g : (0, 1) → R such that for all x ∈ (0, 1),
the following two conditions are satisfied:
1. f (x) < g(x), and
2. if x < y, then g(x) < f (y)?
Either find examples of such f and g or prove that no such f and g exist.
Solution by Bruce Burdick, retired, Providence, RI.
Suppose functions f and g satisfy the given properties. Since x < y implies f (x) <
g(x) < f (y), we see that f is strictly increasing. Therefore, f can only have count-
ably many points of discontinuity in (0, 1). We choose x ∈ (0, 1) with f (x) =
limy→x f (y). By property 2, we must have

g(x) ≤ lim f (y) = f (x).

y→x +

But that contradicts property 1. So, no such pair of functions can exist.
Also solved by Jesús Sistos Barron (student) and Eagle Problem Solvers, Georgia Southern U.; Bobby
Benim, U. of Colorado - Boulder; Soham Bhadra (student), Patha Bhavan, India; Michael Ecker (retired),
Penn. St. U.; Kaitlyn Gibson and Arthur Rosenthal, Salem St. U.; Lixing Han, U. of Michigan - Flint; Eu-
gene Herman, Grinnell C.; Eugen Ionaşcu, Columbus St. U.; Juniata C. Problem Solving Group Ioana
Mihaila and Ivan Ventura, Cal Poly Pomona; Charlie Mumma, Seattle, WA; Katherine Nogin, Clovis
North High School; Northwestern U. Math Problem Solving Group; Paolo Perfetti, Universitá degli
studi di Tor Vergata Roma; Lawrence Peterson, U. of N. Dakota; Mark Sand, C. of St. Mary; Stephen
Scheinberg, Corona del Mar; Joel Schlosberg, Bayside, NY; Omar Sonebi; Nora Thornber; and the pro-
poser.

Rings with few multiplicative maps are rare.

1228. Proposed by Greg Oman, University of Colorado at Colorado Springs, Colorado
Springs, CO.
Let R be a ring, and let f : R → R be a function. Say that f is multiplicative if
f (xy) = f (x)f (y), f (0) = 0, and (if R has an identity) f (1) = 1. Find all commu-
tative rings R (not assumed to have an identity) with the following two properties:
1. There exists an element a ∈ R which is not nilpotent, and
2. every multiplicative map f : R → R is either the identity map or the zero map.

Solution by Kevin Byrnes.

Claim: The only commutative ring R satisfying conditions 1 and 2 is R = F2 .

396 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Proof. We will prove the claim by showing it is true when |R| = 2, and that no com-
mutative ring R with |R| ≥ 3 satisfies the conditions (observe, no ring of size 1 has a
non-nilpotent element). If |R| = 2 then the distinguished non-nilpotent element a ∈ R
serves as the identity element 1, and applying Cauchy’s Theorem to the additive group
of R forces R = F2 . Trivially, the only multiplicative map f : F2 → F2 with f (0) = 0
and f (1) = 1 is f = id.
Now suppose |R| ≥ 3 and R has a distinguished non-nilpotent element a, we’ll
demonstrate the existence of a multiplicative function f : R → R that is neither 0
nor id.

Case 1: 1 ∈ R
Recall that if 1 ∈ R then R contains at least one maximal ideal M and M is also a
prime ideal (see Dummit and Foote Chapter 7, Prop. 11–13). Now define f : R → R
0, if x ∈ M
by f (x) = . Observe that f (0) = 0, f (1) = 1 (as M cannot con-
1, otherwise
tain 1). Furthermore, f is multiplicative since for any x, y ∈ R: if x or y ∈ M then
xy ∈ M so f (xy) = 0 = f (x)f (y); if neither x nor y ∈ M then xy  ∈ M as M is
prime, so f (xy) = 1 = f (x)f (y). Finally, f must map some x ∈ R − {0, 1} to 0 or
1, so f  = id. Thus we have demonstrated the desired function f .

Case 2: 1  ∈ R
We will show that one of the two functions: g(x) = x 2 or h(x) = ax is multiplicative,
maps 0 to 0, and is not 0 or id. Clearly g is multiplicative and g(0) = 0; if g  = id we
are done, so suppose that g = id, hence x 2 = x ∀x ∈ R. In particular, a 2 = a and thus
a 2 x = ax ∀x ∈ R, implying a(ax − x) = 0 ∀x ∈ R. Since 1  ∈ R we have a x̃  = x̃ for
some x̃ ∈ R and thus ∃b ∈ R − {0} (specifically b = a x̃ − x̃) such that ab = 0. In this
case, for any x, y ∈ R: h(xy) = axy = a 2 xy = axay = h(x)h(y) since a 2 = a and
R is commutative, so h is multiplicative. Clearly h(0) = 0, and h(b) = ab = 0  = b,
so h  = id. 

But wait, there’s more! Even if condition 1 is dropped it is still possible to find
multiplicative functions  = 0 or id for commutative rings of size ≥ 3. Consider the
subring S = {0, 3, 6} of Z9 . There we have xy = 0 ∀x, y ∈ S, hence any function f :
S → S with f (0) = 0 is multiplicative. In particular, f (0) = 0, f (3) = 6, f (6) = 3
is multiplicative (even stronger, it is a ring homomorphism).

Also solved by Ioana Mihaila and Ivan Ventura, Cal Poly Pomona; and the proposer.

A bound on the spectral radius of a matrix

1229. Proposed by George Stoica, Saint John, New Brunswick, Canada.
Let A = (aij ) be an n × n matrix such that aii = 0 and aij = bi cj for i  = j , where
bi > 0 and cj ≥ 0 for 1 ≤ i, j ≤ n. Prove that the spectral radius of A is strictly less

than 1 if and only if ni=1 bibciic+1
i
< 1.

Solution by Lixing Han, University of Michigan - Flint.

Denote the spectral radius of matrix A by ρ(A). We will use the following well-known
result about nonnegative matrices.

VOL. 54, NO. 4, SEPTEMBER 2023 THE COLLEGE MATHEMATICS JOURNAL 397
 If A is an n × n nonnegative matrix and x ∈ Rn is a (entry-wise) positive vector,
then
(Ax)i (Ax)i
min ≤ ρ(A) ≤ max .
1≤i≤n xi 1≤i≤n xi

For the matrix A in the problem, consider A − I . where I is the n × n identity matrix.
Then

A − I = bcT − diag ([b1 c1 + 1, . . ., bn cn + 1]) ,

where b = [b1 , . . ., bn ]T and c = [c1 , . . ., cn ]T are the column vectors and diag
([b1 c1 + 1, . . ., bn cn + 1]) is the diagonal matrix whose diagonal entries are b1 c1 +
1, . . ., bn cn + 1. Choose the positive vector
 T
b1 bn
y= , . . ., .
b1 c1 + 1 bn cn + 1

Then we have
 

n
bi ci
(A − I )y = − 1 b.
i=1
bi ci + 1
n
If bi ci
i=1 bi ci +1 < 1, then from (2) we have (A − I )y. Thus Ay < y. This implies

(Ay)i
max < 1.
1≤i≤n yi

Therefore by (1) we musthave ρ(A) < 1.

On the other hand, if ni=1 bibciic+1
i
≥ 1, then from (2) we have (A − I )y ≥ 0. Thus
Ay ≥ y, which implies

(Ay)i
min ≥ 1.
1≤i≤n yi

By (1) we obtain ρ(A) ≥ 1. 

We thus conclude that ρ(A) < 1 if and only if ni=1 bi ci
bi ci +1
< 1.
Also solved by Michel Bataille, Rouen, France; Soham Bhadra (student), Patha Bhavan, India; and the pro-
poser.

Primitive Heronian triangles with equivalent rectangles

1230. Proposed by Jason Zimba, Amplify, New York, NY.
A Heronian triangle is a triangle with positive integer side lengths and positive integer
area. Denoting the side lengths of a Heronian triangle by a, b, and c, the triangle is
called primitive if gcd(a, b, c) = 1. We shall say that a primitive Heronian triangle has
an equivalent rectangle if there exists a rectangle with integer length and width that
shares the same perimeter and area as the triangle. Show that infinitely many primitive
Heronian triangles have equivalent rectangles.

398 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Solution by Kyle Calderhead, Malone University, Canton, OH.
We provide a constructive solution by parameterizing an infinite family of such
triangles.
Consider the triangles with sides

a = n3 + 2n2 + 2n + 1,
b = n3 + 2n2 + 2n, and
c = 2n2 + 2n + 1,

where n is a positive integer. This must be primitive, since a = b + 1.

This gives us a perimeter of P = 2n3 + 6n2 + 6n + 2. Calculating the area (us-
ing Heron’s formula, of course), it is straightforward to verify that it simplifies to
A = n(n + 1)2 (n2 + n + 1). The equivalent rectangle has dimensions n(n + 1) × (n +
1)(n2 + n + 1). We can immediately see that the area is the same, and another straight-
forward calculation shows that the perimeter is the same as well.
We should note, however, that this parameterization does not cover all such
triangles—for example, those with sides (a, b, c) equal to (56, 53, 53) or (95, 87, 68).
Also solved by John Christopher, California St. U., Sacramento; Rohan Dalal (student) and Tommy Goe-
beler, The Episcopal Academy; Habib Far, Lone Star C. - Montgomery; Eugen Ionaşcu, Columbus St. U.;
Michael Vowe, Therwil, Switzerland; and the proposer.

VOL. 54, NO. 4, SEPTEMBER 2023 THE COLLEGE MATHEMATICS JOURNAL 399
 ··· b

··· b
··· b
..
.

SOLUTIONS
(Note that this section includes solutions that would normally have appeared in the
January issue, together with all solutions slated for the March issue.)

Tiling a square with small squares and narrow rectangles

1216. Proposed by Oluwatobi Alabi, Government Science Secondary School Pyakasa
Abuja, Abuja, Nigeria.
For an integer n ≥ 3, find a closed form for the number of ways to tile an n × n
square with 1 × 1 squares and (n − 1) × 1 rectangles (each of which may be placed
horizontally or vertically).

148 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Solution by Rob Pratt, Apex, NC.
Each tiling is uniquely determined by its placement of h horizontal and v vertical
rectangles. We consider nine cases.
• h = 0, v = 0: There is clearly 1 such tiling with no rectangles.
• h = 1, v = 1: If the horizontal rectangle H is in row 1 or n, there are 2 ways to
place H and n + 1 ways to place the vertical rectangle V . If H is in one of the other
n − 2 rows, there are 2 ways to place H and 2 ways to place V . This case yields
4(n + 1) + 4(n − 2) = 8n − 4 tilings.
• h = 2, v = 2: There are 2 such tilings, with the horizontal rectangles in rows 1 and
n and the vertical rectangles in columns 1 and n.
• h = 0, v > 0: Each column has 3 choices for a vertical rectangle (upper, lower, or
empty), but v > 0 implies that not all columns are empty. This case yields 3n − 1
tilings.
• h > 0, v = 0: Same count as h = 0, v > 0.
• h = 1, v > 1: The horizontal rectangle H must be in row 1 or n, and for each row
there are 2 ways to place H . If the remaining column contains a vertical rectangle V ,
there are 2 ways to place V and 2n−1 − 1 nonempty placements of vertical rectangles
in the n − 1 columns shared with H . If the remaining column does not contain a
vertical rectangle, there are 2n−1 − 1 − (n − 1) placements of at least 2 vertical rect-
angles. This case yields 4[2(2n−1 − 1) + (2n−1 − n)] = 4(3 · 2n−1 − n − 2) tilings.
• h > 1, v = 1: Same count as h = 1, v > 1.
• h ≥ 2, v > 2: There are 0 such tilings because the horizontal rectangles block at
least n − 2 columns.
• h > 2, v ≥ 2: Same count as h ≥ 2, v > 2.

Hence, the total number of tilings is

1 + (8n − 4) + 2 + 2(3n − 1) + 2[4(3 · 2n−1 − n − 2)] = 2 · 3n + 12 · 2n − 19.

Each tiling is uniquely determined by its placement of h horizontal and v vertical
rectangles. We consider nine cases.
• h = 0, v = 0: There is clearly 1 such tiling with no rectangles.
• h = 1, v = 1: If the horizontal rectangle H is in row 1 or n, there are 2 ways to
place H and n + 1 ways to place the vertical rectangle V . If H is in one of the other
n − 2 rows, there are 2 ways to place H and 2 ways to place V . This case yields
4(n + 1) + 4(n − 2) = 8n − 4 tilings.
• h = 2, v = 2: There are 2 such tilings, with the horizontal rectangles in rows 1 and
n and the vertical rectangles in columns 1 and n.
• h = 0, v > 0: Each column has 3 choices for a vertical rectangle (upper, lower, or
empty), but v > 0 implies that not all columns are empty. This case yields 3n − 1
tilings.
• h > 0, v = 0: Same count as h = 0, v > 0.
• h = 1, v > 1: The horizontal rectangle H must be in row 1 or n, and for each row
there are 2 ways to place H . If the remaining column contains a vertical rectangle V ,
there are 2 ways to place V and 2n−1 − 1 nonempty placements of vertical rectangles
in the n − 1 columns shared with H . If the remaining column does not contain a
vertical rectangle, there are 2n−1 − 1 − (n − 1) placements of at least 2 vertical rect-
angles. This case yields 4[2(2n−1 − 1) + (2n−1 − n)] = 4(3 · 2n−1 − n − 2) tilings.

VOL. 54, NO. 2, MARCH 2023 THE COLLEGE MATHEMATICS JOURNAL 149
• h > 1, v = 1: Same count as h = 1, v > 1.
• h ≥ 2, v > 2: There are 0 such tilings because the horizontal rectangles block at
least n − 2 columns.
• h > 2, v ≥ 2: Same count as h ≥ 2, v > 2.

Hence, the total number of tilings is

1 + (8n − 4) + 2 + 2(3n − 1) + 2[4(3 · 2n−1 − n − 2)] = 2 · 3n + 12 · 2n − 19.

Also solved by Kyle Calderhead, Malone U.; Vincent and Owen Zhang high school students from
MathILy summer program; Ethan Curb, Peyton Matheson, Aiden Milligan, Cameron Moening, Vir-
ginia Rhett Smith and Ell Torek, high school students at The Citadel; Eagle Problem Solvers,
Georgia Southern U.; Dmitri Fleishman, Santa Monica, CA; Walther Janous, Ursulinengymnasium,
Innsbruck, Austria; Lawrence Peterson, U. of N. Dakota; and the proposer. Two incorrect solutions
were received.

Fibonacci numbers from the solution to an integral equation

1217. Proposed by Eugen Ionascu, Columbus State University, Columbus, GA.
Prove the following:
1. There exists a unique function f : R → R which satisfies the following equation
for every x ∈ R:
 x
f (−x) = 1 + cos(t)f (x − t)dt.
0

Moreover, express f explicitly in terms of elementary functions.

k+1
2. For every nonnegative integer k, f k (0) = (−1) 2 Fk , where F0 = 0, F1 =
1, Fk+2 = Fk + Fk+1 , and x denote the greatest integer less than or equal to
a real number x.

Solution by Russ Gordon, Whitman College, Walla Walla, WA.

Using some simple substitutions, it is easy to verify that
 x
g(−x) = 1 + g(t) cos(x − t) dt
0

and
 x
g(x) = 1 − g(−t) cos(x − t) dt.
0

It then follows that

 x
u(x) ≡ g(x) + g(−x) = 2 + v(t) cos(x − t) dt;
0

 x
v(x) ≡ g(x) − g(−x) = − u(t) cos(x − t) dt.
0

150 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Taking Laplace transforms (with the obvious notation and noting the convolution op-
erator), we find that
2 s
U (s) = + 2 V (s)
s s +1
and
s
V (s) = − U (s).
s2 + 1
Letting α = φ and β = −1/φ (the two solutions to the equation x 2 = x + 1), where
phi represents the golden mean, we find that
2 s 4 + 3s 2 + 1 − s 2 2 2s
U (s) = · = − 2  
s s + 3s + 1
4 2 s s + α2 s 2 + β 2
 
2 2 s s
= +√ − 2 ,
s 5 s 2 + α2 s + β2
√
where we have used the simple facts αβ = −1, α + β = 1, and α − β = 5. Taking
the inverse Laplace transform, it follows that
2
u(x) = 2 + √ (cos(αx) − cos(βx)) .
5
The function V (s) satisfies
 2 2
s 2 s +1 s2 + 1
V (s) = − 2 · · 4 = −2 ·  2  
s + 1 s s + 3s + 1 2
s + α2 s 2 + β 2
 
2 α β
= −√ − 2 ,
5 s 2 + α2 s + β2
and thus
2
v(x) = − √ (sin(αx) − sin(βx)) .
5
Combining these results gives
u(x) + v(x) 1
g(x) = = 1 + √ (cos(αx) − cos(βx) − sin(αx) + sin(βx)) .
2 5
Using simple derivative properties of the sine and cosine functions, along with the
Binet formula for the Fibonacci numbers, we see that
α 2k−1 − β 2k−1
g (2k−1) (0) = (−1)k · √ = (−1)k f2k−1
5
and
α 2k − β 2k
g (2k) (0) = (−1)k · √ = (−1)k f2k
5
for each positive integer k. This completes the solution.

VOL. 54, NO. 2, MARCH 2023 THE COLLEGE MATHEMATICS JOURNAL 151
Also solved by Michel Bataille, Rouen, France; Brian Bradie, Christopher Newport U.; Bruce Burdick (re-
tired), Providence, RI; Hongwei Chen, Christopher Newport U.; Russ Gordon (additional solution), Whitman
C.; Eugene Herman, Grinnell C.; Walther Janous, Ursulinengymnasium, Innsbruck, Austria; Kee-Wai Lau,
Hong Kong, China; Albert Natian, Los Angeles Valley C.; and the proposer.

Pell numbers and Pell-Lucas numbers

1218. Ángel Plaza, Universidad de Las Palmas de Gran Canaria, Las Palmas de Gran
Canaria, Spain.
The Pell and Pell-Lucas numbers, {Pn : n ∈ N} and {Qn : n ∈ N}, respectively, are
defined recursively as follows: P0 = 0, P1 = 1, Q0 = Q1 = 2, and (for each sequence)
un+1 = 2un + un−1 for n ≥ 1. Next, let n ∈ N, and let An (x) and Bn (x) be polynomials
of degre n with real coefficients such that for 0 ≤ i ≤ n, we have An (i) = Pi and
Bn (i) = Qi . Find An (n + 1) and Bn (n + 1) in terms of Pn+1 and Qn+1 , respectively.
Solution by Brian Bradie, Christopher Newport University, Newport News, VA.
Solution: For each i = 0, 1, 2, . . . , n, let xi = i and define

n
x−j (−1)n−i 
n
Ln,i (x) = = (x − j ).
j =0,j =i
i−j i!(n − i)! j =0,j =i

Note Ln,i (x) is the Lagrange interpolating polynomial associated with the node xi = i
which satisfies
  
0, j = i n−i n + 1
Ln,i (j ) = and Ln,i (n + 1) = (−1) .
1, j = i i

The Lagrange form for the interpolating polynomials An (x) and Bn (x) is then


n 
n
An (x) = Ln,i (x)Pi and Bn (x) = Ln,i (x)Qi ;
i=0 i=0

consequently,

n   
n+1  
n+1 n+1
An (n + 1) = (−1)n (−1)i Pi = Pn+1 + (−1)n (−1)i Pi
i=0
i i=0
i

and


n   
n+1  
n+1 i n+1
Bn (n + 1) = (−1) n
(−1) Qi = Qn+1 + (−1)
i n
(−1) Qi .
i=0
i i=0
i

Now, the Binet forms for Pi and Qi are

√ i √
(1 + 2) − (1 − 2)i √ √ i
Pi = √ and Qi = (1 + 2)i + (1 − 2) ,
2 2

152 © THE MATHEMATICAL ASSOCIATION OF AMERICA

so, by the binomial theorem,

 √ √

n+1
n+1 (− 2)n+1 − ( 2)n+1
(−1) i
Pi = √
i=0
i 2 2
√ 
( 2)n 0, √ n odd
= ((−1) − 1) =
n+1
2 −( 2)n , n even

and
  
n+1
n+1 √ √
(−1)i Qi = (− 2)n+1 + ( 2)n+1
i=0
i
 √
√ 2( 2)n+1 , n odd
= ( 2)n+1 (1 + (−1)n+1 ) = .
0, n even

Finally,

0,√ n odd
An (n + 1) = Pn+1 −
( 2)n , n even

and
 √
2( 2)n+1 , n odd
Bn (n + 1) = Qn+1 − .
0, n even

Also solved by Michel Bataille, Rouen, France; Eugene Herman, Grinnell C.; Northwestern U. Math
Problem Solving Group; Albert Stadler, Herrliberg, Switzerland; and the proposer. One incorrect solution
was received.

A criterion for a commutative ring to be a field

1219. Proposed by Greg Oman, University of Colorado at Colorado Springs, Colorado
Springs, CO.
Let R be a commutative ring with identity 1 = 0. Recall that if I and J are ideals of
R, then the product of I and J is defined as follows:

I J := {i1 j1 + · · · + in jn : ik ∈ I, jk ∈ J, n ∈ Z+ }.

Prove that R is a field if and only if for every ideal I and J of R, we have I J ∈ {I, J }.
Solution by Missouri State Problem Solving Group.
Sufficiency follows directly since if R is a field, then the only ideals of R are 0
and R. For necessity, let x, y ∈ R. Then the assumption implies that either (xy) =
(x)(y) = (x) or (xy) = (x)(y) = (y), where (z) denotes the ideal of R generated by
z ∈ R. Now if xy = 0, then either (x) = (0) or (y) = (0), that is either x = 0 or y = 0,
so R is an integral domain. Let a be a nonzero element of R. Then we have (a 2 ) =
(a)2 = (a)(a) ∈ {(a), (a)}, that is, (a 2 ) = (a). Since R is a domain, then a 2 = ua for
some unit u ∈ R, and by cancelation we get a = u. So all nonzero elements are units
and hence R is a field.

VOL. 54, NO. 2, MARCH 2023 THE COLLEGE MATHEMATICS JOURNAL 153
Also solved by Anthony Bevelacqua, U. of N. Dakota; Paul Budney, Sunderland, MA; Bill Dunn, Mont-
gomery C.; Eugene Herman, Grinnell C.; Scheilla Raffaelli, Indiana U. East; Diego Vurgait; and the
proposer.

Cofactors of cofactors
1220. Proposed by Jeff Stuart, Pacific Lutheran University, Tacoma, WA.
Let A be an n × n real or complex matrix with n ≥ 2. Let co(A) denote the matrix of
cofactors of A, that is, for each i and j , (co(A))ij is the product of (−1)i+j and the
determinant of the matrix obtained by deleting the ith row and j th column of A. Prove
the following:
1. If n = 2, then co(co(A)) = A for every A.
2. If n > 2, show that there is a unique singular A such that co(co(A)) = A.
3. If n > 2, find a condition on det(A) that is satisfied exactly when A is invertible
and co(co(A)) = A.

Solution by Mark Wildon, Royal Holloway, Egham, UK.

Say that a ring R with unit element 1 = 0 is small if no proper nontrivial subring of
R has an identity.
The subring of R generated by 1 is {m1 : m ∈ Z}. Clearly it contains the identity of
R. Therefore, if R is small, R is generated as an abelian group by 1. Hence R has Z-
rank 1 as an abelian group and so either R = Z or R = Z/NZ for some N ∈ N with
N ≥ 2. Since m2 = m for m ∈ Z if and only if m = 0 or m = 1, the only possible
identity in a subring of Z is 1. Hence, Z is small. If N is composite, with N = AB
where gcd(A, B) = 1 then, by the Chinese Remainder Theorem,

Z ∼ Z Z
= ×
NZ AZ BZ

and {(x, 1) : x ∈ Z/AZ} is a proper subring with identity of the right-hand side. (In
this case the identity is not the identity of Z/NZ.) Hence, Z/NZ is small only if N
is a power of a prime. In this case Z/NZ is small, since m2 ≡ m mod p a if and only
if m(m − 1) ≡ 0 mod p a , and since m and m − 1 are coprime integers, either p a | m
which implies that m ≡ 0 mod p a , or p a | m − 1, which implies that m ≡ 1 mod p a .
We conclude that the small rings are precisely Z and Z/p a Z for p a prime and a ≥ 1.
Also solved by Michel Bataille, Rouen, France; Missouri State Problem Solving Group, ; Albert
Stadler, Herrliberg, Switzerland; and the proposer. One incorrect solution was received.

Area of a polar graph

1221. Proposed by Gregory Dresden, Washington and Lee University, Lexington, VA.
Shown below (from left to right) are the graphs of r = sin 4θ/3 and r = sin 6θ/5,
where every other adjacent region (starting from the outside) is shaded black. Find the
total shaded area for any such graph r = sin(k + 1)θ/k, where k > 0 is an odd integer
and θ ranges from 0 to 2kπ.

154 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Solution by Guiseppe Fera, Vincenza, Italy.
We prove that the total shaded area is π2 , regardless of k.
The symmetry center of the 2(k + 1)-petalled rose
  
k+1
r = sin θ
k

is the pole of a polar coordinate system, and the polar axis passes through one of the
common points to two black shaded regions on the border of the curve.
First, we evaluate the total shaded area of a petal. Consider the petal symmetric to
the line θ = 2(k+1)
π
. Looking at the solutions of

x = r cos θ > 0
y = r sin θ = 0

we get k − 1 intersection point (other than the pole) between the polar axis and
the
 curve, for θ = mπ, with m = mπ 1, 2, . . ., k − 1. Their cartesian coordinates are
sin mπ
k m=1,2,...,k−1
. The identity sin k = sin (k−m)π k
for m = 1, 2, . . ., k−1
2
shows that
every intersection is double. Indeed, these intersection points (and the pole) are the
start-points of the black shaded regions inside the petal. The slope of the tangent line
to the curve at such points is less than π2 for m = 1, 2, . . ., k−12
and greater than π2 for
m = 2 , 2 , . . ., k − 1. Since the petal contains 2 shaded regions, symmetric with
k+1 k+3 k+1

respect to the line θ = 2(k+1)

π
, the shaded half area of the petal is

 k−1     
1
π 2 π
mπ+ 2(k+1) k−1 +m π+ π
1 2(k+1) 2 2(k+1)
S= r 2 dθ + r dθ −
2
 
2
r dθ .
2 0 2 m=1 mπ k−1 +m π
2

Set n = k−1
2
+ m. The integration is elementary and gives

  k−1
π − k sin πk 2

S= + s,
8(k + 1) m=1

VOL. 54, NO. 2, MARCH 2023 THE COLLEGE MATHEMATICS JOURNAL 155
where
    
k 2mπ (2m + 1)π
s= sin − sin
8(k + 1) k k
    
2nπ (2n + 1)π
− sin − sin .
k k

Reintroducing n and simplifying, we get

  k−1
2  
π − k sin πk k  π  π 
S= + sin 2m − 1) − sin (2m + 1) .
8(k + 1) 8(k + 1) m=1 k k

Using a prosthaphaeresis identity, we have

  
k−1
 π    
π − k sin πk k 2
2mπ
S= + −2 sin cos .
8(k + 1) 8(k + 1) k m=1 k

Using the exponential representation of the cosine,

     
2mπ exp 2mπi + exp −2mπi
cos = k k
,
k 2

the sum becomes a geometric series, so the value of the sum is

  πi(k+1)   2πi   −πi(k+1)   −2πi  
1 exp − exp exp − exp
k
  k
+ k
 −2πi  k
.
2 exp 2πi k
− 1 exp k
− 1

Simplifying, this is − 21 so that

   
π − k sin πk k sin πk π
S= + = .
8(k + 1) 8(k + 1) 8(k + 1)
Finally, since the rose has 2(k + 1) petals, the total shaded area is 4(k + 1)S = π2 .
Also solved by J. A. Grzesik, Allwave Corp.; Paul Stockmeyer, C. of William & Mary; and the proposer.

Properties of a general parabola

1222. Proposed by Kent Holing, Trondheim, Norway.
Consider the parabola f (x, y) = Ax 2 + 2Bxy + Cy 2 + 2Dx + 2Ey + F = 0 with
real coefficients, B = 0 and A, C > 0.
1. Show that the parabola is nondegenerate if and only if β = BE − CD = 0.
2. Show that in the degenerate√case, the parabola can be given by the formula
f (x, y) = Ax + By + D ± √ α1 = 0 for α1 = D 2 − AF or (equivalently) by
f (x, y) = Bx + Cy + E ± α2 = 0 for α2 = E 2 − CF and α1,2 ≥ 0.
3. When β = 0, show that (A + C)(BxT + CyT ) + BD + CE = 0 for the coordi-
nates xT and yT of the vertex T .

156 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 4. Using 3., show that xT = − 2β
α2
+ At for t = β
2C(A+C)2
.
5. Show that the coordinates of the focus F of the parabola are xF = xT + Ct and
yF = yT − Bt.

Solution by Michel Bataille, Rouen, France.

Let P be the given parabola. The discriminant of the second degree part Ax 2 +
2Bxy + Cy 2 must vanish, hence, B 2 = AC, an equality that will be used freely in
what follows. The equation of the parabola is equivalent to

(Bx + Cy)2 + 2CDx + 2CEy + CF = 0

or

(Ax + By)2 + 2DAx + 2EAy + AF = 0.

The equation Bx + Cy = 0 (or equivalently Ax + By = 0) gives the direction of the

diameters of P .
1. P is nondegenerate if and only if every diameter intersects the parabola in a
unique point. Let d, with equation Bx + Cy + k = 0, be a diameter. From (1),
a point (x, y) is in d ∩ P if and only the two equations Bx + Cy + k = 0 and
2CDx + 2CEy + CF + k 2 = 0 are satisfied. This system has a unique solu-
tion if and only if 2CD · C − 2CE · B = 0, that is, if and only if BE − CD = 0.

2. If P degenerates into two parallel lines, then its equation can be written
as (Ax + By + p)(Ax + By + q) = 0. Comparing with (1) leads to pq =
AF, p + q = 2D (note that DB = EA because CD = BE) and p, q are so-
lutions of the quadratic
√ X2√− 2DX + F A = 0. Thus, α1 = D 2 − F A ≥ 0 and
{p, q} = {D + α1 , D − α1 }. In a similar way, comparing (1) with √ (Bx +
   
Cy
√ + p )(Bx + Cy + q ) = 0 gives α 2 ≥ 0 and {p , q } = {E + α2 , E −
α2 }. The required results follow.
 
3. The vector ∂f ∂x
(xT , yT ), ∂f
∂y
(xT , yT ) is orthogonal to the tangent at the vertex
T , hence, is collinear to the direction vector (C, −B) of the diameters. It follows
that B ∂f
∂x
(xT , yT ) + C ∂f∂y
(xT , yT ) = 0.
Since ∂x (x, y) = 2B(Bx + Cy) + 2CD and ∂f
∂f
∂y
(x, y) = 2C(Bx + Cy) + 2CE,
an easy calculation yields (A + C)(BxT + CyT ) + BD + CE = 0.

4. Let λ = BD+CE
A+C
so that BxT + CyT + λ = 0. Since the equation of P can be
written as

(Bx + Cy + λ)2 + 2x(CD − λB) + 2Cy(E − λ) + F C − λ2 = 0,

expressing that T is on P we obtain 2(CD − λB)xT − 2(E − λ)(λ + BxT ) +

F C − λ2 = 0 so that

−2βxT = 2λE − F C − λ2 = α2 − (λ − E)2 .

Aβ 2
Since an easy calculation gives (λ − E)2 = C(A+C)2
, we get xT = − 2β
α2
+ At.

VOL. 54, NO. 2, MARCH 2023 THE COLLEGE MATHEMATICS JOURNAL 157
 5. Let δ be the line C(x − xT ) − B(y − yT ) + C(A + C)t = 0. The parabola with
directrix δ and focus F = (xT + Ct, yT − Bt) is the locus of all the points
P (x, y) such that (d(P , δ))2 = P F 2 . Thus, to answer the question, it is suffi-
cient to show that the equation f (x, y) = 0 is equivalent to

[C(x − xT ) − B(y − yT ) + C(A + C)t]2

C(A + C)
= (x − xT − Ct)2 + (y − yT + Bt)2

But (2) writes as ((B(y − yT ) − C(x − xT ))2 − 2C(A + C)t ((B(y − yT ) −

C(x − xT )) = (B 2 + C 2 )((x − xT )2 + (y − yT )2 ) + 2C(A + C)t ((B(y − yT ) −
C(x − xT )), that is, (Bx + Cy + λ)2 + 4C(A + C)t[By − Cx + CxT + BC (λ +
BxT )] = 0, hence, we have to show that
 
λB
4C(A + C)t By − Cx + (A + C)xT +
C
= 2x(CD − λB) + 2Cy(E − λ) + F C − λ2

for all x, y. Simple calculations give CD − λB = −2C 2 (A + C)t, E −  λ=

2(A + C)tB  and a slightly longer one gives F C − λ 2
= 4C(A + C)t (A +
C)xT + λBC
so we are done.

Also solved by Hongwei Chen, Christopher Newport U.; Eugene Herman, Grinnell C.; Walther Janous,
Ursulinengymnasium, Innsbruck, Austria; and the proposer.

Which rectangular numbers are squares - again?

1223. Don Redmond, Southern Illinois University, Carbondale, IL.
Let h be a positive integer and define the nth rectangular number of order h, denoted
by Rh (n), as Rh (n) = n(n + h). Determine all positive integer values of h for which
the equation Rh (n) = m2 has a solution for some positive integers n and m.
Solution by Kathleen Lewis, University of the Gambia, Brikama, Republic of the Gam-
bia.
All positive integers except 1, 2, and 4 are possible values for h. First notice why
these three values are excluded. When h = 1, n(n + h) = n(n + 1) = n2 + n, which
lies between n2 and (n + 1)2 , so it cannot be a perfect square. The same problem occurs
when h = 2 and n(n + 2) = n2 + 2n. When h = 4, the integers n and n + h have the
same parity, so n(n + h) also has the same parity as n2 . That means that n(n + 4)
cannot be equal to (n + 1)2 . But it’s too small to be (n + 2)2 . Therefore, 4 is also an
impossible choice for h.
To see that all other values of h are possible, consider the cases h = 2k + 1, h =
4k + 2 and h = 4k + 4, with k ∈ N. All positive integers other than 1, 2 and 4 fall into
one of these cases.
• If h = 2k + 1, let n = k 2 . Then n(n + h) = (k 2 )(k 2 + 2k + 1) = [k(k + 1)]2 .
• If h = 4k + 2, let n = 2k 2 . Then n(n + h) = 2k 2 (2k 2 + 4k + 2) = 4k 2 (k 2 + 2k +
1) = [2k(k + 1)]2 .
• If h = 4k + 4, let n = k 2 . Then n(n + h) = k 2 (k 2 + 4k + 4) = [k(k + 2)]2 .

158 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Editor’s note: Bataille pointed out that this problem appeared as number 871 in the
March 2008 issue and provided two new solutions. Stone and Hawkins, provided an
algorithm for producing, for a given h, all pairs (n, m) such that Rh (n) = m. Indeed,
j2
writing m = n + j , with 1 ≤ j < h2 , and setting n = h−2j produces a solution for each
such n that is a positive integer.
Also solved by Michel Bataille, Rouen, France; Anthony Bevelacqua, U. of N. Dakota; Kyle Calder-
head, Malone U.; John Christopher, California St. U., Sacramento; Eagle Problem Solvers, Georgia South-
ern U.; Habib Far, Lone Star C. - Montgomery; Dmitry Fleischman, Santa Monica, CA; Donald Hooley,
Bluffton, Ohio; Tom Jager, Calvin U.; Graham Lord, Princeton, NJ; Matthew McMullen, Otterbein U.;
Northwestern U. Math Problem Solving Group; Mark Sand, C. of Saint Mary; David Stone and
John Hawkins, Georgia Southern U. (retired); Michael Vowe, Therwil, Switzerland; Owen Zhang, (student)
MathILy summer math program; and the proposer. Two incomplete solutions were received.

A criterion for a group to be cyclic

1224. Proposed by George Stoica, Saint John, New Brunswick, Canada.
Let G be a finite group, and suppose that for any subgroups H and K of G, we have
|H ∩ K| = gcd(|H |, |K|). Prove that G is cyclic.
Solution by Anthony Bevelacqua, University of North Dakota.
Suppose a, b ∈ G have order d. Then

|a ∩ b| = gcd(|a|, |b|) = d

and so a = b. Since a cyclic group of order d has exactly φ(d) generators, we see
that G has exactly φ(d) elements of order d.
Let N be the order of G, and let Nd be the number of elements of order d in G. By
the last paragraph we have either Nd = 0 or Nd = φ(d). Thus,
 
N= Nd ≤ φ(d)
d|N d|N


Since N = d|N φ(d) for any positive integer N and Nd ≤ φ(d) for each d, we must
have Nd = φ(d) for each d|N. In particular, G must contain an element of order N,
and so G is cyclic.
Also solved by Paul Budney, Sunderland, MA; Aran Bybee and Sam Lowery; Kevin Byrnes, Glen Mills,
PA; Michael Goldenberg, Baltimore Polytechnic Inst. and Mark Kaplan, U. of Maryland Global Campus;
Eugene Herman, Grinnell C.; Tom Jager, Calvin U.; Joel Scholosberg, Bayside, NY; Ed Enochs, U. of
Kentucky (retired) and David Stone, Georgia Southern U. (retired); and the proposer.

A reduced ring with all subrings chained is a field

1225. Proposed by Greg Oman, University of Colorado at Colorado Springs, Colorado
Springs, CO.
All rings R throughout are commutative with 1 = 0 and all subrings S of R are unital
(that is, 1 ∈ S). Recall that a ring R is chained provided that for any ideals I and J of
R, either I ⊆ J or J ⊆ I .

VOL. 54, NO. 2, MARCH 2023 THE COLLEGE MATHEMATICS JOURNAL 159
 1. Give an example of a ring R which is not a field with the property that every
subring of R is chained.
2. Suppose now that R is reduced, that is, R has no nonzero nilpotents. Prove that
if every subring of R is chained, then R is a field.

Solution by Anthony Bevelacqua, University of North Dakota.

Z4 , the ring of integers modulo 4, is not a field, but it is chained as the only ideals in
Z4 are 0Z4 ⊆ 2Z4 ⊆ Z4 .
We note that the following rings are not chained: Z (consider 2Z and 3Z), k[t] the
ring of polynomials over a field k (consider tk[t] and (t + 1)k[t]), and S ⊕ T the direct
sum of rings S and T (consider S ⊕ 0 and 0 ⊕ T ). As special cases of the last example,
Zm , the ring of integers modulo m, if m = st for relatively prime s, t > 1 and k[t]/(g)
if g ∈ k[t] is the product of relatively prime polynomials of positive degree are not
chained.
Now suppose R is reduced and every subring of R is chained. Z = 1Z is a subring
of R isomorphic to either Z or Zm for some m ≥ 2. Since Z is chained we must
Z∼ = Zpe for some prime p and some e ≥ 1, and since Z is reduced we must have
e = 1. We can suppose Zp is a subring of R.
Zp [a] is a subring of R for any a ∈ R. Since the ring of polynomials over Zp is not
chained, a must be algebraic over Zp . Thus Zp [a] ∼ = Zp [t]/(g) for some monic g ∈
k[t] of positive degree. Since Zp [a] is chained we have g = π e for a monic irreducible
π ∈ k[t] and some e ≥ 1, and since Zp [a] is reduced we have e = 1. Thus, Zp [a] ∼ =
k[t]/(π) is a field. Since every nonzero a ∈ R is invertible, R is a field.
Also solved by Eugene Herman, Grinnell C.; Tom Jager, Calvin U.; and the proposer.

160 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 SOLUTIONS

The limit of a difference of harmonic sums

1211. Proposed by Needet Batir, Nevs, ehir Haci Bektas, Veli University, Nevs, ehir,
Turkey.

VOL. 53, NO. 5, NOVEMBER 2022 THE COLLEGE MATHEMATICS JOURNAL 401
  limit, where below, H0 = 0 and for n > 0, Hn denotes the nth
Evaluate the following
haromic number nk=1 k1 :

 
n 
Hn−k
lim (Hn ) −
2
.
n→∞
k=1
k

Solution by Henry Ricardo, Westchester Area Math Circle, Purchase, NY.

First we establish that

n
Hn−k
= Hn2 − Hn(2) ,
k=1
k
n
where Hn(2) = k=1 1/k 2 .
The formula is clearly true for n = 1. Now suppose that the formula holds for
some integer N > 1. Then, noting that Hm+1 = Hm + 1/(m + 1) and Hm+1
(2)
= Hm(2) +
1/(m + 1) ,
2


N+1
HN+1−k  HN−k+1
N
H0
= +
k=1
k k=1
k N +1


N
HN−k 
N
1
= +
k=1
k k=1
k(N − k + 1)
N  
1  1 1
= HN2 − HN(2) + +
N + 1 k=1 k N − k + 1
2HN
= HN2 − HN(2) +
N +1
 2  
1 1 2HN
= HN+1 − − HN+1 −
(2)
+
N +1 (N + 1) 2 N +1
 
2 HN + N+1
1
2 2HN
= HN+1
2
− HN+1
(2)
− + +
N +1 (N + 1)2 N +1
= HN+1
2
− HN+1
(2)
.

Therefore,


n
Hn−k π2
Hn2 − = Hn2 − (Hn2 − Hn(2) ) = Hn(2) → ζ (2) = as n → ∞.
k=1
k 6

Also solved by Robert Agnew, Palm Coast, FL; Paul Bracken, U. of Texas at Austin; Brian Bradie,Christopher
Newport U.; Bruce Burdick, Providence, RI; Hongwei Chen, Christopher Newport U.; Russ Gordon, Whit-
man C.; G. C. Greubel, Newport News, VA; Jacob Guerra, Salem St. U.; GWStat Problem Solving Group;
Stephen Kaczkowski, South Carolina Governor’s School for Science and Mathematics; Kee-Wai Lau, Hong
Kong, China; Shing Hin Jimmy Pa; Henry Ricardo, Westchester Area Math Circle, Purchase, NY (2 additional

402 © THE MATHEMATICAL ASSOCIATION OF AMERICA

solutions); Abhishek Sinha, Tata Institute of Fundamental Research, Mumbai, India; Albert Stadler, Her-
rliberg, Switzerland; Seán Stewart, King Abdullah U. of Science and Technology; Michael Vowe, Therwil,
Switzerland; Mark Wildon, Royal Holloway, Egham, UK; and the proposer.

Two trig sum identities

1212. Proposed by Paul Bracken, University of Texas, Edinburg, TX.
Let n be an odd natural number and let θ ∈ R be such that cos(nθ) = 0. Prove the
following:


n−1
sin θ n sin(nθ)
=− , and (1)
k=0
sin θ − cos ( n )
2 2 kπ cos θ cos(nθ)


n−1
(−1)k+1 cos( kπ ) n sin( nπ )
n
= 2
. (2)
k=0
sin2 θ − cos2 ( kπ
n
) cos θ cos(nθ)

Solution by Michel Bataille, Rouen, France.

We will apply the following formula: if n ∈ N and x, y, x − y are not a multiple of
π, then


n−1
1 n sin(x − y)
= (3)
k=0
sin( x−kπ
n
) sin( y−kπ
n
) sin(x) sin(y) sin( x−y
n
)

(see a proof at the end).

Proof of (1). (1) is obvious if sin(θ) = 0 so we suppose sin(θ) = 0 in what follows.
We notice that
kπ 1 − cos(2θ) 1 + cos( 2kπ ) kπ kπ
sin2 θ − cos2 ( )= − n
= − cos( + θ) cos( − θ),
n 2 2 n n
(4)
hence sin2 θ − cos2 ( kπ
n
) = − sin( x−kπ
n
) sin( y−kπ
n
) with x = n( π
2
− θ), y = n( π
2
+ θ).
Formula (3) yields


n−1
1 n sin(−2nθ) n sin(nθ)
=− =−
k=0
sin θ − cos ( n )
2 2 kπ sin(n( 2 − θ)) sin(n( 2 + θ)) sin(−2θ)
π π
sin θ cos θ cos(nθ)

(note that, n being odd, sin(n( π2 ± θ)) = (−1)(n−1)/2 cos(nθ).) The identity (1) fol-
lows.
Proof of (2). First, we consider (3) with x = y + n π2 and obtain


n−1
1 n sin(n π2 )
=
k=0
cos( y−kπ
n
) sin( y−kπ
n
) (−1)(n−1)/2 sin(y) cos(y)

or

n−1
1 n sin(n π2 )
= . (5)
k=0
cos( π2 − 2y
n
+ 2kπ
n
) (−1)(n−1)/2 sin(2y)

VOL. 53, NO. 5, NOVEMBER 2022 THE COLLEGE MATHEMATICS JOURNAL 403
Now, using 2 cos( kπ
n
) cos(θ) = cos( kπ
n
+ θ) + cos( kπ
n
− θ) and (4), we see that we
have to prove

2n sin(n π2 )
S= ,
cos(nθ)


n−1 
where S = (−1)k 1
cos( kπ
+ 1
cos( kπ
. Setting n = 2m + 1, we have
k=0 n +θ) n −θ)


m
1 1 
m−1
1 1
S= + − +
j =0
cos( 2jnπ + θ) cos( 2jnπ − θ) j =0
cos( (2j +1)π
n
+ θ) cos( (2j +1)π
n
− θ)

and
1 1 1 1
+ = +
− cos( (2j +1)π
n
+ θ) − cos( (2j +1)π
n
− θ) cos( 2(m+jn+1)π + θ) cos( 2(m+jn+1)π − θ)

so that


n−1
1 1
S= + .
k=0
cos( 2kπ
n
+ θ) cos( 2kπ
n
− θ)

With the help of (5), we obtain


n−1
1 n sin(n π2 ) n sin(n π2 )
= =
k=0
cos( 2kπ
n
+ θ) (−1)(n−1)/2 sin(n π2 − nθ) cos(nθ)

and therefore
n sin(n π2 ) n sin(n π2 ) 2n sin(n π2 )
S= + = ,
cos(nθ) cos(n(−θ)) cos(nθ)
as desired.
Proof of (3). From sinsin(x−y)
x·sin y
= 2i
e2iy −1
− 2i
e2ix −1
(easily checked) and the decomposition
into partial fractions

1  wk 1
n−1 n−1
1 1
= =
z −1
n n k=0 z − w k
n k=0 zw − 1
k

2π i
where w = e− n we deduce that
n−1    
1 1  sin x−kπ − y−kπ
n−1
sin(x − y) 2i 2i
= − 2i(x−kπ ) =  n
 n

sin x · sin y n k=0 e 2i(y−kπ
n
)
−1 e n −1 n k=0 sin x−kπ
n
· sin y−kπ
n

and therefore


n−1
1 n sin(x − y)
 x−kπ   y−kπ  =  
k=0
sin n
· sin n
sin x · sin y · sin x−y
n

404 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Also solved by Brian Bradie, Christopher Newport U.; Hongwei Chen, Christopher Newport U.; Shing Hin
Jimmy Pak; Albert Stadler, Herrliberg, Switzerland; Michael Vowe, Therwil, Switzerland; and the proposer.
One incomplete solution was received.

The limit of a product of powers of sums

1213. Proposed by Rafael Jakimczuk, Universidad National de Lujá, Buenos Aires,
Argentina.
Let (an ) be a sequence of positive integers, and for every positive integer n, define
Pn := (1 + a11n )a1 · (1 + a21n )a2 · · · (1 + an1n )an . Find limn→∞ Pn .
Solution by Ulrich Abel, Technische Hochschule, Mittelhessen, Germany.
Let n and a1 , a2 , a3 , . . . be positive
 integers. The Bernoulli inequality (1 + x)n ≥
ak
1 + nx (x ≥ −1) implies that 1 + ≥ 1 + 1/n. On the other hand, the well-
1

ak n
ak
known inequality (1 + x/n)n ≤ ex (x ≥ 0) implies that 1 + ak1n ≤ e1/n . Conse-
quently,
 n n  ak
1 1
1+ ≤ 1+ ≤ e,
n k=1
ak n

which implies
 n 
1 ak
lim 1+ = e.
n→∞
k=1
ak n

Several solvers pointed out that this problem, by a different proposer, appeared as
problem 12256 in The American Mathematical Monthly.
Also solved by Robert Agnew, Palm Coast, FL; Michel Bataille, Rouen, France; Paul Bracken, U. of
Texas, Edinburg; Brian Bradie,Christopher Newport U.; Hongwei Chen, Christopher Newport U.; Dmitri
Fleischman, Santa Monica, CA; Michael Goldenberg, Baltimore Polytechnic Inst. and Mark Kaplan, U.
of Maryland Global Campus; Lixing Han, U. of Michigan - Flint; Jim Hartman, C. of Wooster; Eugene Her-
man, Grinnell C.; Walther Janous, Innsbruck, Austria; Stephen Kaczkowski, S. Carolina Governor’s School
for Science and Mathematics; Kee-Wai Lau, Hong Kong, China; Kelly McLenithan, Los Alamos, NM; Al-
bert Natian, Los Angeles Valley C.; Edward Omey, KULeuven @ Campus Brussels; Shing Hin Jimmy Pak;
Mark Sand, C. of Saint Mary; Randy Schwartz (emeritus), Schoolcraft C.; Abhishek Sinha, Tata Inst. of
Fundamental Research, Mumbai, India; Albert Stadler, Herrliberg, Switzerland; Michael Vowe, Therwil,
Switzerland; and the proposer. One incorrect solution was received.

A closed form expression for a sequence

1214. Proposed by Luis Moreno, SUNY Broome Community College, Binghampton,
NY.
The following sequence can be found in the text Intermediate Analysis by John Olm-
sted: (1, 2, 2 21 , 3, 3 31 , 3 23 , 4, 4 41 , 4 24 , 4 34 , 5, . . .). Now let n be a positive integer. Find a
closed-form expression for an , the nth term of the above sequence.
Solution by Habib Far, Lone Star College – Montgomery, Conroe, Texas.

VOL. 53, NO. 5, NOVEMBER 2022 THE COLLEGE MATHEMATICS JOURNAL 405
 k(k + 1)
We realize that an = k + 1 when n = Tk + 1, where T + k = is the trian-
2
gular number for some positive integer k. If Tk+1 < n ≤ Tk+1 , then

n − Tk − 1
an = k + 1 + .
k+1

Let n = Tj + 1, for some positive integer j . Solve j (j + 1) = 2(n − 1) yields

√ √
−1 + 8n − 7 −1 + 8n − 7
j = . Let k = j = , where x is the greatest
2 2
integer function. Thus

n − Tk − 1
an = k + 1 + .
k+1

Also solved by Ulrich Abel, Technische Hochschule, Mittelhessen, Germany; Robert Agnew, Palm Coast,
FL; Ashland U Problem Solving Group; Michel Bataille, Rouen, France; Brian Beasley, Presbyterian
C.; Hudson Bouw, Braxton Green, Dillon King (students), Taylor U.; Brian Bradie, Christopher New-
port U.; Case Western Reserve U. Problem Solving Group; Hongwei Chen, Christopher Newport U.;
John Christopher, California St. U.; Gregory Dresden, Washington & Lee U.; Skye Fisher, (student) U.
of Arkansas at Little Rock; Dmitry Fleischman, Santa Monica, CA; Natacha Fontes-Merz, Westminster C.;
Dominique Frost (student) U. of Arkansas at Little Rock; Rohan Dalal, (student) and Tommy Goebeler, The
Episcopal Academy; Lixing Han, U. of Michigan - Flint and Xinjia Tang, Changzhou U., Changzhou, China;
Walther Janous, Innsbruck, Austria; Kelly McLenithan, Los Alamos, NM; Northwestern U Math Prob-
lem Solving Group; Lawrence Peterson, U. of North Dakota; Bill Reil, Philadelphia, PA; Mark Sand, C.
of St. Mary; Tyler Sanders, (student) U. of Arkansas at Little Rock; Randy Schwartz (emeritus), Schoolcraft
C.; Doug Serfass, (student) U. of Arkansas at Little Rock; Vishwest Ravi Shrimali; Albert Stadler, Her-
rliberg, Switzerland; Seán Stewart, King Abdullah U. of Science and Technology; Robert Vallin, Lamar U.;
Michael Vowe, Therwil, Switzerland; Edward White and Roberta White, Frostburg, MD; and the proposer.

Rings for which no proper subring has an identity

1215. Proposed by Greg Oman, University of Colorado at Colorado Springs, Colorado
Springs, CO.
Let R be a ring (assumed only to be associative but not to contain an identity unless
stated). Recall that a subring of R is a nonempty subset of R closed under addition,
negatives, and multiplication. Find all rings R with identity 1 = 0 with the property
that no proper, nontrivial subring of R has an identity (which need NOT be the identity
of R).
Solution by Mark Wildon, Royal Holloway, Egham, UK.
Say that a ring R with unit element 1 = 0 is small if no proper nontrivial subring of
R has an identity.
The subring of R generated by 1 is {m1 : m ∈ Z}. Clearly it contains the identity of
R. Therefore if R is small, R is generated as an abelian group by 1. Hence R has Z-
rank 1 as an abelian group and so either R = Z or R = Z/NZ for some N ∈ N with
N ≥ 2. Since m2 = m for m ∈ Z if and only if m = 0 or m = 1, the only possible

406 © THE MATHEMATICAL ASSOCIATION OF AMERICA

identity in a subring of Z is 1. Hence Z is small. If N is composite, with N = AB
where gcd(A, B) = 1 then, by the Chinese Remainder Theorem,

Z ∼ Z Z
= ×
NZ AZ BZ
and {(x, 1) : x ∈ Z/AZ} is a proper subring with identity of the right-hand side. (In
this case the identity is not the identity of Z/NZ.) Hence Z/NZ is small only if N
is a power of a prime. In this case Z/NZ is small, since m2 ≡ m mod p a if and only
if m(m − 1) ≡ 0 mod p a , and since m and m − 1 are coprime integers, either p a | m
which implies that m ≡ 0 mod p a , or p a | m − 1, which implies that m ≡ 1 mod p a .
We conclude that the small rings are precisely Z and Z/p a Z for p a prime and a ≥ 1.
Also solved by Anthony Bevelacqua, U. of N. Dakota; Paul Budney, Sunderland, MA; Francisco Perdomo
and Ángel Plaza, Universidad de Las Palmas de Gran Canaria, Spain; and the proposer.

VOL. 53, NO. 5, NOVEMBER 2022 THE COLLEGE MATHEMATICS JOURNAL 407
 SOLUTIONS

Harmonic, Fibonacci, and triangular numbers

1206. Proposed by Seán M. Stewart, Bomaderry, NSW, Australia.


Let Hn := nk=1 k1 denote the nth harmonic number, let Fn denote the nth Fibonacci
number, where F0 := 0, F1 := 1, and Fn := Fn−1 + Fn−2 for n ≥ 2. Further, let Tn be
the nth triangular
√
number defined by T0 := 0 and Tn := n + Tn−1 for n ≥ 1, and let
ϕ := 1+2 5 be the golden ratio. Prove the following:


∞
Tn Hn Fn 232
= 52 log(2) + √ log(ϕ) + 73.
n=1
2n 5

Solution by Hongwei Chen, Christopher Newport University, Newport News, Virginia.

Recall the generating function of the harmonic numbers:

∞
log(1 − x)
Hn x n = − .
n=1
1−x

Let
x log(1 − x)
f (x) := − .
1−x
Differentiating

∞
Hn x n+1 = f (x)
n=1

twice leads to

∞
n(n + 1)Hn x n−1 = f (x). (1)
n=1

320 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Using this fact and Tn = n(n + 1)/2 we find the generating function of {Tn Hn }∞
n=1 :


∞
1
Tn Hn x n = xf (x) := g(x).
n=1
2

Direct computation gives

 
x 2 3x 2 log(1 − x) 2x log(1 − x)
g(x) = + − − .
2 (1 − x)2 (1 − x)3 (1 − x)2 (1 − x)3

Using the well-known Binet formula

   
1 1 n
Fn = √ φ − −
n
,
5 φ

and with some simplifications, we have


∞     
Tn Hn Fn 1 φ 1
=√ g −g −
n=1
2n 5 2 2φ
√   √  
130 − 58 5 1 130 + 58 5 φ
= 73 − log 1 + − log 1 − (1)
5 2φ 5 2

Notice that
        
1 φ 1 φ 1
log 1 + + log 1 − = log 1 + 1− = log = −2 log(2)
2φ 2 2φ 2 4

and
     
1 φ 1 + 1/2φ
log 1 + − log 1 − = log = log(φ 4 ) = 4 log(φ).
2φ 2 1 − φ/2

From (1) we consequently find


∞
Tn Hn Fn 232
= 73 + 52 log(2) + √ log(φ),
n=1
2n 5

as desired.
Also solved by Narendra Bhandari, Bajura, Nepal; Brian Bradie,Christopher Newport U.; Bruce Burdick,
Providence, RI; Nandan Sai Dasireddy, Hyderabad, Telangana, India; Russ Gordon, Whitman C.; Eugene
Herman, Grinnell C.; Walther Janous, Ursulinengymnasium, Innsbruck, Austria; Volkhard Schindler,
Berlin, Germany; Albert Stadler, Herrliberg, Switzerland; Enrique Treviño, Lake Forest C.; and the pro-
poser.

A sum of a product of sums

1207. Ovidiu Furdui and Alina Sı̂ntămărian, Technical University of Cluj-Napoca,
Cluj-Napoca, Romania.

VOL. 53, NO. 4, SEPTEMBER 2022 THE COLLEGE MATHEMATICS JOURNAL 321
Establish the following:

∞ ∞  ∞ 
1 1
(2n − 1) = ζ (2) + ζ (3),
n=1 k=n
k2 k=n
k3
∞
where for a positive integer k, we have ζ (k) = 1
n=1 nk .

Solution by Brian Bradie, Christopher Newport University, Newport News, Virginia.

First, write

∞
1 ∞
1 ∞
1  1
∞ 
∞
1  1
∞
= + .
k=n
k2 k=n
k3 j =n
j 2 =j 3 j =n j 3 =j +1 2

Next,

∞ ∞
1  1
∞ ∞
1 
j
∞
1 ∞  ∞
1
(2n − 1) 2 3
= 2
(2n − 1) 3
= 3
n=1 j =n
j =j
 j =1
j n=1 =j
 j =1 =j


∞
1 
 ∞
1
= 3
1= 2
= ζ (2),
=1
 j =1 =1


and


∞ ∞
1  1
∞ ∞
1 
j
∞
1 ∞
1  1
∞
(2n − 1) = (2n − 1) =
n=1 j =n
j 3 =j +1 2 j =1
j 3 n=1 =j +1
2 j =1
j =j +1 2

∞
1  1  H −
−1 ∞ 1 
∞
H ∞
1
= = 
= −
=2
2 j =1 j =2
2 =1
2 =1
3

= 2ζ (3) − ζ (3) = ζ (3),


where Hn = nj=1 1
j
denotes the nth harmonic number, and we have used the well-
known identity


∞
H
= 2ζ (3).
=1
2

Finally,


∞ ∞
1 ∞
1
(2n − 1) 2 3
= ζ (2) + ζ (3).
n=1 k=n
k k=n
k

Also solved by Narendra Bhandari, Bajura, Nepal; Paul Bracken, U. of Texas, Edinburgh; Bruce Burdick,
Providence, RI; Hongwei Chen, Christopher Newport U.; Eugene Herman, Grinnell C.; Walther Janous,
Ursulinengymnasium, Innsbruck, Austria; Kee-Wai Lau, Hong Kong, China; Shing Hin Jimmy Pak; Seán
Stewart, King Abdullay U. of Sci. and Tech., Thuwal, Saudi Arabia; and the proposer.

322 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 An integral of logarithms
1208. Proposed by Marián S̆tofka, Slovak University of Technology, Bratislava, Slo-
vakia.
Prove that

ln(1 − x) ln(1 + x)
1
5
dx = − ζ (3),
0 x 8

where as above, for a positive integer k, we have ζ (k) = ∞ n=1
1
nk
.
Solution by Didier Pinchon, Toulouse, France.
Let I be the integral to evaluate. Using identity
1
ln(1 − x) ln(1 + x) = (ln(1 − x) + ln(1 + x))2 − (ln(1 − x) − ln(1 + x))2
4
  
1 1−x
= ln (1 − x ) − ln
2 2 2
,
4 1+x

it follows that I = (I1 − I2 )/4, with

 1−x 
1
ln2 (1 − x 2 ) 1 ln2
I1 = dx, I2 = 1+x
dx.
0 x 0 x
√
The substitutions x = u in I1 and x = (1 − u)/(1 + u) in I2 give

1 1 1
ln2 (u) 1 1
ln2 (u)
I1 = du, I2 = 2 du.
2 0 0 1−u 0 0 1 − u2
The dominated convergence theorem allows to permute the series expansion of 1/(1 −
u) (resp. 1/(1 − u2 ) ) with the integration in I1 (resp. I2 ), and therefore

1 1  1
I1 = ln2 (u) un du, I2 = 2 ln2 (u) u2n du.
2 n≥0 0 n≥0 0

For any nonnegative integer k, two successive integrations by parts provide the result
1
2
ln2 (u) un du = ,
0 (n + 1)3
and it follows that
 1
I1 = = ζ (3),
n≥0
(n + 1)3

 
 1  1  1 7
I2 = 4 =4 − = ζ (3).
n≥0
(2n + 1) 3
n≥0
(n + 1) 3
n≥0
(2n + 2) 3 2

In conclusion, I = 1
4
(I1 − I2 ) = − 85 ζ (3).

VOL. 53, NO. 4, SEPTEMBER 2022 THE COLLEGE MATHEMATICS JOURNAL 323
 Several solvers pointed out that this problem, by a different proposer, appeared as
problem 12256 in The American Mathematical Monthly.
Also solved by F. R. Ataev, Uzbekistan; Khristo Boyadzhiev, Ohio Northern U.; Brian Bradie,Christopher
Newport U.; Bruce Burdick, Providence, RI; Hongwei Chen, Christopher Newport U.; Kyle Gatesman
(student), Johns Hopkins U.; Subhankar Gayen, West Bengal, India; Walther Janous, Ursulinengymnasium,
Innsbruck, Austria; Moubinool Omarjee, Lycée Henri IV, Paris, France; Henry Ricardo, Westchester Area
Math Circle; Albert Stadler, Herrliberg, Switzerland; Seán Stewart, King Abdullay U. of Sci. and Tech.,
Thuwal, Saudi Arabia; Michael Vowe, Therwil, Switzerland; and the proposer. One incomplete solution was
received.

The rank of a matrix

1209. Proposed by George Stoica, Saint John, New Brunswick, Canada.
For non-negative integers i and j , define

⎧
⎪
⎨i(i − 1) · · · (i − j + 1) if 1 ≤ j ≤ i,
aij := 1 if i = 0 and j ≥ 0, or j = 0 and i ≥ 0, and
⎪
⎩0 if j > i ≥ 1.

Now let m be a positive integer. Prove that every m × m submatrix of the infinite
matrix
m (a2i,j ) with 0 ≤ j ≤ m − 1 and i ≥ 0 has rank m and, in addition, that
i=0 (−1) i m
i
a2k+2i,j = 0 for 0 ≤ j ≤ m − 1 and any k ∈ N.
Solution by the proposer.
Introduce the polynomials

f0 (x) = 1, f1 (x) = x, f2 (x) = x(x − 1), . . . , fj (x) = x(x − 1) · · · (x − j + 1).

Then a2i,j = fj (2i). Since for any j

j −1

x j = fj (x) + cn fn (x)
n=0

for some constants cn , it is clear that any matrix of the form

 
fj (xi ) with 0 ≤ j, i ≤ m − 1, and where all xi are distinct,

can be transformed into a Vandermonde matrix by elementary row operations, so its

determinant must be different from zero.
For the second statement, start by observing that the identity

m  
m
(−1)i f (i) = 0
i=0
i

must be valid whenever f (x) is a polynomial of degree at most m − 1. Indeed, let us

define (f (x)) = f (x) − f (x + 1), and note that
m  
i m
(−1) f (i) = m (f (x))(0).
i=0
i

324 © THE MATHEMATICAL ASSOCIATION OF AMERICA

The difference operator decreases the degree of the polynomial, and the equation can
be proved inductively, using Pascal’s identity.
As we saw above, the function i → a2i,j is a polynomial of degree j . Hence


m  
m
(−1) i
a2k+2i,j = 0 for 0 ≤ j ≤ m − 1.
i=0
i

This completes the solution.

No other solutions were received.

The existence of a countable commutative integral domain with a

sum-free collection of ideals
1210. Proposed by Greg Oman, University of Colorado at Colorado Springs, Colorado
Springs, CO.
Let R be a commutative ring with identity, and let I and J be ideals of R. Recall that
the sum of I and J is the ideal defined by I + J := {i + j : i ∈ I, j ∈ J }. Prove or
disprove: there exists a countable commutative integral domain D with identity and a
collection S of 2ℵ0 ideals of D such that for all I = J in S , we have I + J ∈
/ S.
Solution by Anthony Bevelacqua, University of North Dakota.
Let D = Z[x1 , x2 , . . .] be the polynomial ring in countably many indeterminates
with coefficients in Z. Since D is the countable union of the countable Z[x1 , . . . , xn ]
for each n ∈ N, D is a countable commutative integral domain with identity.
For any A ⊆ N let IA be the ideal of D generated by {xi | i ∈ A}. For all A, B ⊆ N
we have (i) IA = IB if and only if A = B and (ii) IA + IB = IA∪B . Thus it suffices to
find a collection S of 2ℵ0 subsets of N such that for all A = B in S we have A ∪ B ∈ / S.
It’s well-known (see below for sketch of proof) that for any countable set X there
exists a collection T of 2ℵ0 subsets of X such that each U ∈ T is infinite and for
all U = V in T we have U ∩ V is finite. Since each element of T is an infinite set,
U ∩V ∈ / T . So there exists T a family of 2ℵ0 subsets of N such that for all U = V in
T we have U ∩ V ∈ / T . Now S = {N − U | U ∈ T } has the desired properties: S is a
family of 2ℵ0 subsets of N such that for all A = B in S we have A ∪ B ∈ / S.
Thus D = Z[x1 , x2 , . . .] is a countable commutative integral domain with identity
containing a collection S = {IA | A ∈ S} of 2ℵ0 ideals such that for all I = J in S we
have I + J ∈ / S.
Sketch of a standard proof of above claim: Without loss of generality we can sup-
pose X = Q. There are 2ℵ0 real irrational numbers. For each real irrational r let (un )∞
n=1
be a sequence of rational numbers converging to r, and let Ur = {un | n ∈ N}. Each Ur
is infinite and Ur ∩ Us is finite for any distinct real irrationals r and s.
Also solved by Northwestern U. Math Problem Solving Group; and the proposer.

Correction: In the featured solution to problem 1195 in the January 2022 issue, two
numerators were missing in the second line. The second line as provided by the solver
should have been

∞ 
∞
hn ∞ ∞
hn
= .
n=1 k=n+2
(n + 1)k 2
n=1 k=1
(n + 1)(n + k + 1) 2

The editor apologizes for the error.

VOL. 53, NO. 4, SEPTEMBER 2022 THE COLLEGE MATHEMATICS JOURNAL 325
 SOLUTIONS

Polynomials of degree n tangent to a circle at n − 1 points

1196. Proposed by Ferenc Beleznay, Mathleaks, Budapest, Hungary, and Daniel
Hwang, Wuhan Britain-China School, Wuhan, China.
Prove or disprove: for every positive integer n, there exists a polynomial of degree
n + 1 with real coefficients whose graph is tangent to some circle at n points.
Solution by Mark Wildon, Royal Holloway, Egham, UK.
Such polynomials exist. Shifting n, we shall prove that for each n ∈ N with n ≥ 3
there exists a polynomial Pn of degree n with coefficients in the integers such that the
graph of Pn (x) is tangent to the unit circle at exactly n − 1 points in the open interval
(−1, 1). For n = 2 we may simply take P2 (x) = 1, which is tangent to the unit circle
at 0 and has degree 0.
To define the Pn for n ≥ 3, we need the Chebyshev polynomials of the second kind.
Recall that, in the usual notation, Um is the unique polynomial with real coefficients of
degree m such that (sin θ)Um (cos θ) = sin(m + 1)θ. For instance U0 (x) = 1, U1 (x) =
2x, and since sin 3θ = − sin3 θ + 3 sin θ cos2 θ = sin θ(− sin2 θ + 3 cos2 θ) = sin θ(−1 +
4 cos2 θ) we have U2 (x) = 4x 2 − 1. In fact each Un has integer coefficients. For each
n ∈ N with n ≥ 4, define

Pn (x) = x 2 Un−2 (x) − 2xUn−3 + Un−4 .

As shown in [1, Theorem 5], the defining property of Um and the relation 2 cos θ sin rθ =
sin(r + 1)θ + sin(r − 1)θ imply that if n ≥ 4 then

(sin θ)Pn (cos θ)

= (cos2 θ sin θ)Un−2 (cos θ) − 2(cos θ sin θ)Un−3 (cos θ) + (sin θ)Un−4 (cos θ)
= cos2 θ sin(n − 1)θ − 2 cos θ sin(n − 2)θ + sin(n − 3)θ
= (1 − sin2 θ) sin(n − 1)θ − sin(n − 1)θ − sin(n − 3)θ + sin(n − 3)θ
= − sin2 θ sin(n − 1)θ.

Hence, Pn (cos θ) = − sin θ sin(n − 1)θ for each such n. Setting P3 (x) = 2x 3 − 2x
we have P3 (cos θ) = 2 cos3 θ − 2 cos θ = 2(cos2 θ − 1) cos θ = −2 sin2 θ cos θ =
− sin θ sin 2θ. Therefore,

Pn (cos θ) = − sin θ sin(n − 1)θ if n ≥ 3. ()

Since each Um has integer coefficients, so does each Pn .

Observe that, by (),

(cos θ)2 + Pn (cos θ)2 = cos2 θ + sin2 θ sin2 (n − 1)θ ≤ cos2 θ + sin2 θ = 1.

Hence, the graph of Pn (x) for −1 ≤ x ≤ 1 lies inside the closed unit disc. Moreover,
we have (cos θ)2 + Pn (cos θ)2 = 1 if and only if sin2 (n − 1)θ = 1, so if and only if
θ = (2k−1)π
n−1
for some k ∈ N. Thus if x = cos (2k−1)π
n−1
and x ∈ (−1, 1), the graph of
Pn (x) is tangent to the unit circle.

154 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 To get distinct values of cos θ, we may assume that θ ∈ [0, π]. If n = 2m is even
then there are 2m − 1 distinct tangent points, obtained by taking k = 1, . . . , m −
1, m, m + 1, . . . 2m − 1 to get x-coordinates

π (2m − 3)π (2m − 1)π 2π

cos , . . . , cos , cos = −1, − cos ,
2m − 1 2m − 1 2m − 1 2m − 1
(2m − 2)π
. . . , − cos .
2m − 1
If n = 2m + 1 is odd, then there are 2m distinct tangent points, obtained by taking
k = 1, . . . , m − 1, m to get x coordinates

π (2m − 3)π (2m − 1)π

cos , . . . , cos , cos
2m 2m 2m
and then k = m + 1, . . . , 2m to get x coordinates

π (2m − 3)π (2m − 1)π

− cos , . . . , − cos , − cos .
2m 2m 2m
This completes the proof.
Remark. We remark that since Pn (1) = Pn (cos 0) = 0 and Pn (−1) = Pn (cos π) = 0
by (), the graph of Pn (x) meets the graph of the unit circle at x = ±1; of course since
the unit circle has a vertical asymptote at these points, the graph is not tangent. Thus,
Pn is tangent to the unit circle at n − 1 points and has two further intersection points.
Since tangent points have multiplicity (at least) 2, this meets the bound in Bezout’s
Theorem, that the intersection multiplicity between the algebraic curves y = Pn (x)
and x 2 + y 2 = 1 of degrees n and 2, respectively, is 2n, and shows that each tangent
point has degree exactly 2.

References

[1] Janjić, M. (2008). On a class of polynomials with integer coefficients. J. Integer Seq. 11(5): Article 08.5.2, 9.

Also solved by the proposer. We received one incomplete solution.

VOL. 53, NO. 2, MARCH 2022 THE COLLEGE MATHEMATICS JOURNAL 155
 Matrices with presistently unequal rows
1197. Proposed by Valery Karachik and Leonid Menikhes, South Ural State University,
Chelyabinsk, Russia
Let A be an arbitrary n × m matrix that has no equal rows. Find a necessary sufficient
condition relating n and m so that there exists a column of A, after removal of which,
all rows remain different.
Solution by Eugene Herman, Grinnell College, Grinnell, Iowa.
The given property holds in a trivial sense when n = 1 or m = 1. In both cases, after
a column has been removed there do not exist two rows that are equal. Otherwise, the
necessary and sufficient condition is 2 ≤ n ≤ m. Suppose first that m + 1 = n ≥ 2.
Let A = [aij ], where aij = 0 when j ≥ i and aij = 1 when j < i. If column j of A
is removed then rows j and j + 1 are equal; hence the given property fails to hold.
If n ≥ m + 2, construct the first m + 1 rows of A as before and fill in the rest of the
matrix so all rows are different.
Suppose 2 ≤ n ≤ m and suppose the given property does not hold. Thus, for each
j ∈ {1, 2, . . . , m}, there exists a pair of rows Pj = {r, s} such that r and s are unequal
but become equal when the j th entry is removed from each. We create an undirected
graph as follows. Each vertex corresponds to a row, and so the number of vertices is
n. The edges correspond to the sets Pj ; specifically, (r, s) is an edge if and only if
{r, s} = Pj for some j . Hence the number of edges is m. No vertex is joined to it-
self by an edge and no two vertices are joined by more than one edge. We show that
the graph contains no cycles. Suppose (r1 , . . . , rk ) is a cycle; that is, r1 , . . . , rk are
distinct vertices and (r1 , r2 ), . . . , (rk−1 , rk ), (rk , r1 ) are edges. The edges correspond
to different columns, which we may assume are columns 1 through k (by permut-
ing columns, if necessary). Let r1 = (a1 , a2 , . . . , am ). Thus, r2 = (b1 , a2 , a3 , . . . , am )
where b1  = a1 and r3 = (b1 , b2 , a3 , . . . , am ) where b2  = a2 , and so on until rk =
(b1 , b2 , . . . , bk−1 , ak , . . . , an ) where bk−1  = ak−1 . Then (rk , r1 ) cannot be an edge since
rk and r1 differ in in k − 1 entries and k − 1 > 1. Our graph is therefore a tree. In a
tree, the number of vertices is always larger than the number of edges, and so m < n.
This contradiction establishes our necessary and sufficient condition.
Also solved by the proposer.

156 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 The cardinality of a set of maximal ideals
1198. Proposed by Alan Loper, The Ohio State University, Newark OH, and Greg
Oman, The University of Colorado, Colorado Springs, CO.
Let n be a nonnegative integer, and consider the ring R := Q[X0 , . . . , Xn ] of polyno-
mials (via usual polynomial addition and multiplication) in the (commuting) variables
X0 , . . . Xn with coefficients in Q. It is well known that R is a Noetherian ring, and so
every ideal of R is finitely generated. Since R is countable, and there are but countably
many finite subsets of a countable set, we deduce that R has but countably many ide-
als and thus, in particular, countably many maximal ideals. Next, let X0 , X1 , X2 , . . .
be a countably infinite collection of indeterminates. Observe that (to within iso-
morphism) Q[X0 ] ⊆ Q[X0 , X1 ] ⊆ Q[X0 , X1 , X2 ] ⊆ · · · . Let Q[X0 , X1 , X2 , . . .] be
the union of the this increasing chain. How many maximal ideals does the ring
Q[X0 , X1 , X2 , . . .] have? (More precisely, what is the cardinality of the set of maxi-
mal ideals of Q[X0 , X1 , X2 , . . .]?)
Solution by Kenneth Schilling, University of Michigan-Flint, Flint, Michigan.
Since Q [X0 , X1 , X2 ....] has countably many elements, it has at most 2ℵ0 maximal
ideals. We shall exhibit 2ℵ0 maximal ideals, proving that this is the exact cardinality.
Let p0 (t) = t and p1 (t) = t − 1. For each infinite sequence α : N → {0, 1}, let Iα
be the ideal of Q [X0 , X1 , X2 ....] generated by the set of polynomials

{pα(k) (Xk ) : k = 1, 2, 3, ...}.

Since pα(k) (α(k)) = 0, for any q (X1 , X2 , ..., Xn ) ∈ Iα ,

q (α(0), α(1), ..., α(n)) = 0.

It follows that Iα is a proper ideal of Q [X0 , X1 , X2 ....], and so is contained in a maxi-

mal ideal Mα .
Now consider any pair α, β of distinct infinite sequences from {0, 1}. For some k,
{α(k), β(k)} = {0, 1}, so {pα(k) (Xk ), pβ(k) (Xk )} = {Xk , Xk − 1}. Therefore the ideal
generated by Iα ∪ Iβ is the whole ring Q [X0 , X1 , X2 ....]. It follows that the union
Mα ∪ Mβ of maximal ideals must also generate the whole ring, and so, in particular,
Mα  = M β .
We conclude that the set of ideals Mα over all infinite sequences α : N → {0, 1} is
of cardinality 2ℵ0 , and the proof is complete.
Also solved by Paul Budney, Sunderland, MA; and the proposer. We received one incomplete solution.

VOL. 53, NO. 2, MARCH 2022 THE COLLEGE MATHEMATICS JOURNAL 157
 An oscillating function with prescribed zeros
1199. Proposed by Corey Shanbrom, Sacramento State University, Sacramento, CA.
Find a smooth, oscillating function whose periods form a bi-infinite geometric se-
quence. More precisely, given a positive λ  = 1, find a smooth function f on an open
half-line whose root set R is given by

1 1 1 1 1 1
R = · · · − 3 − 2 − . − 2 − , − , 0,
λ λ λ λ λ λ

1, 1 + λ, 1 + λ + λ2 , 1 + λ + λ2 + λ3 , · · · .

Editor’s note: The problem statement in the March 2021 issue omitted one of the zeros.
The functions defined in the submitted solutions included this value in their root set.
Solution by Albert Natian, Los Angeles Valley College, Valley Glen, California..
   
Answer: f (x) = sin π ln[(λ−1)x+1] defined on [1 − λ]−1 , ∞ if λ > 1 and defined
  ln λ
on −∞, [1 − λ]−1 if λ < 1.

Justification It’s clear that sin θ = 0 ⇐⇒ θ = nπ, n ∈ Z. So

 
π ln [(λ − 1) x + 1]
f (x) = 0 ⇐⇒ sin =0
ln λ
π ln [(λ − 1) x + 1]
⇐⇒ = nπ, n ∈ Z
ln λ
⇐⇒ ln [(λ − 1) x + 1] = n ln λ, n ∈ Z
⇐⇒ ln [(λ − 1) x + 1] = ln λn , n ∈ Z
⇐⇒ (λ − 1) x + 1 = λn , n ∈ Z
 −n
λn − 1 1 λ1 −1
⇐⇒ x = if n ≥ 0, x = − ·  1  if n < 0, n ∈ Z
λ−1 λ λ
−1

n−1 −n  j
1
⇐⇒ x = λ if n ≥ 0, x = −
j
if n < 0, n ∈ Z.
j =0 j =1
λ

Also solved by Albert Stadler, Herrliberg, Switzerland; and the proposer.

158 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 A recurrence satisfied by a sequence with a given generating function
1200. Proposed by Russ Gordon, Whitman College, Walla Walla, Washington, and
George Stoica, St. John, New Brunswick, Canada
Let c be an arbitrary real number. Prove that the sequence (an )n≥0 defined by

∞
1
an x n =
n=0
1 − cx + cx 2 − x 3

satisfies an (an − 1) = an+1 an−1 for all n ≥ 1.

Solution 1 by Michel Bataille, Rouen, France.
Since 1 − cx + cx 2 − x 3 = (1 − x)(1 + (1 − c)x + x 2 ), the sequence (an ) is the
unique sequence satisfying

∞
1  ∞
(1 + (1 − c)x + x 2 ) · an x n = = xn.
n=0
1−x n=0

Multiplying out on the left, we obtain a0 = 1, a1 + (1 − c)a0 = 1 and for n ≥ 2

an + an−1 (1 − c) + an−2 = 1. (1)

Now, we prove that an (an − 1) = an+1 an−1 for all n ≥ 1 by induction.

Since a1 (a1 − 1) = c(c − 1) and (using (1)), a2 a0 = a2 = 1 − a1 (1 − c) − a0 = c(c −
1), the relation holds for n = 1.
Assume that an (an − 1) = an+1 an−1 for some integer n ≥ 1. Then, we have

an an+2 = an (1 − an − (1 − c)an+1 ) (using (1))

= an (1 − an ) − an an+1 (1 − c)
= −an+1 an−1 − an an+1 (1 − c) (by assumption)
= −an+1 (an−1 + an (1 − c))
= −an+1 (1 − an+1 ) (using (1)),

hence an+1 (an+1 − 1) = an an+2 . This completes the induction step and the proof.
Solution 2 by Kee-Wai Lau, Hong Kong, China.
Denote the recurrence relation an (an − 1) = an+1 an−1 by *.
• If c = −1, then

∞
1 1 1
an x n = + + ,
n=0
4(1 − x) 4(1 + x) 2(1 + x)2

so that an = 14 [1 + (−1)n (2n + 3)], and * holds.

• If c = 3, then

∞
1
an x n = ,
n=0
(1 − x)3

so that an = (n+1)(n+2)
2
, and * again holds.

VOL. 53, NO. 2, MARCH 2022 THE COLLEGE MATHEMATICS JOURNAL 159
 √
In what follows, we assume that c  = −1, 3. Let α = c−1+ (c−3)(c+1)
2
, so that
1 + α + α2
α  = −1, 0, 1. We have c = , and
α
1 α
=
1 − cx + cx − x
2 3 (1 − x)(α − x)(1 − αx)
 
α 1 α2 1
= + − .
(1 − α)2 (1 + α)(α − x) (1 + α)(1 − αx) 1 − x

Hence
    
α 1 α n+2 1 − α n+1 1 − α n+2
an = + −1 = ,
(1 − α)2 (1 + α)α n+1 1+α (1 + α) (1 − α)2 α n

and it is easy to check that * holds in this case as well.

Solution 3 by Graham Lord, Princeton, New Jersey.
That a0 = 1 is immediate from the substitution x = 0 in the equation. The lat-
ter’s first and second derivatives at 0 show a1 = c and a2 = c(c − 1), respectively.
Note, c − 1 = a2 +aa10 −1 and a1 (a1 − 1) = a2 a0 . For convenience, set a−1 = 0, so
a0 (a0 − 1) = a1 a−1 .

The equation’s RHS denominator, 1 − cx + cx 2 − x 3 factors into (1 − x) and (1 −

(c − 1)x + x 2 ). So multiplication of the equation through by the latter factor gives:
1+ ∞ n=1 (an − (c − 1)an−1 + an−2 )x = 1−x = 1 + x + x + ... .
n 1 2

Hence for all n ≥ 1, as the coefficients of x n on both sides of this last equation
a +an−2 −1
are equal: (an − (c − 1)an−1 + an−2 ) = 1. Equivalently: c − 1 = n an−1 . That is,
an +an−2 −1
for any n ≥ 1 the ratio, an−1
is constant, independent of n, and equal to c − 1.
an +an−2 −1 an+1 +an−1 −1
In particular: an−1
= an
. The latter simplified is the sought after identity
an (an − 1) = an+1 an−1 .

Also solved by Ulrich Abel and Vitaliy Kushnirevych, Technische Hochschule, Mittelhessen, Germany;
Paul Bracken, U. of Texas, Edinburg; Brian Bradie, Christopher Newport U.; Kyle Calderhead, Malone
U.; Hongwei Chen, Christopher Newport U.; FAU Problem Solving Group, Florida Atlantic U.; Geuseppe
Fera, Vicenza, Italy; Dmitry Fleischman, Santa Monica, CA; Michael Goldenberg, Baltimore Polytechnic
Inst. and Mark Kaplan, U. of Maryland Global Campus (jointly); G. C. Greubel, Newport News, VA; GWstat
Problem Solving Group, The George Washington U.; Eugene Herman, Grinnell C.; Walther Janous, Ursu-
linengymnasium, Innsbruck, Austria; Omran Kouba, Higher Inst. for Applied Sci. and Tech., Damascus, Syria.
Northwestern U. Math Problem Solving Group; Carlos Shine, São Paulo, Brazil; Albert Stadler,
Herrliberg, Switzerland; Enrique Treviño, Lake Forest C.; Michael Vowe, Therwil, Switzerland; and the pro-
poser.

160 © THE MATHEMATICAL ASSOCIATION OF AMERICA

It was brought to our attention that CMJ problem 1208 has already appeared as prob-
lem 12256 in the May 2021 issue of the Monthly (by a different proposer). Accord-
ingly, we will not be featuring a solution to this problem. We apologize for the error.

SOLUTIONS

An equilateral triangle in an isosceles triangle

1191. Proposed by Herb Bailey, Rose-Hulman Institute of Technology, Terre Haute,
IN.
An isosceles triangle has incenter I , circumcenter O, side length S, and base length W .
Show that there is a unique value of WS so that there exists a point P on one of the two
sides of length S such that triangle I OP is equilateral. Find this value. Solution by the
Eagle Problem Solvers, Georgia Southern University, Statesboro, GA and Savannah,
GA.
The unique value is
 √
S 1 + 3 + 2 7/3
= ≈ 1.7303506.
W 2

√ with A = (0, a), B = (W/2, 0) and C = (−W/2, 0).

Position the isosceles triangle ABC
W2 4S 2 −W 2
Then a 2 = S 2 − 4
and a = 2
. The circumcenter O lies at the intersection of

70 © THE MATHEMATICAL ASSOCIATION OF AMERICA

the perpendicular bisectors of AB and BC. The midpoint of AB is (W/4, a/2) and
the slope of AB is −2a
W
W
, so the perpendicular bisector of AB has slope 2a and equation
 
a W W
y= + x− .
2 2a 4

Since the perpendicular bisector of BC is the y-axis, then the circumcenter O has
y-coordinate

a W2 4a 2 − W 2 2S 2 − W 2
yO = − = = √ .
2 8a 8a 2 4S 2 − W 2

Since the y-axis bisects ∠A, then the incenter I also lies on the y-axis; its y-coordinate
is given by
√
Wa W 4S 2 − W 2
yI = = .
2S + W 2(2S + W )

Thus, the distance between I and O is given by

√
2S 2 − W 2 W 4S 2 − W 2 S(S − W )
I O = yO − yI = √ − =√ .
2 4S − W
2 2 2(2S + W ) 4S 2 − W 2

If a point P is equidistant from O and I , then its y-coordinate must be given by

yO + yI
yP =
2
√
2S 2 − W 2 W 4S 2 − W 2
= √ +
4 4S 2 − W 2 4(2S + W )
2S 2 − W 2 + W (2S − W )
= √
4 4S 2 − W 2
S 2 + SW − W 2
= √ .
2 4S 2 − W 2

If P also lies on AB, then its x-coordinate must be given by

W  yP 
xP = 1−
2 a
 
W S 2 + SW − W 2
= 1−
2 4S 2 − W 2
SW (3S − W )
= .
2(4S 2 − W 2 )

Thus, the square of the distance between O and P is

 2
IO
OP = 2
+ xP2
2

VOL. 53, NO. 1, JANUARY 2022 THE COLLEGE MATHEMATICS JOURNAL 71

 S 2 (S − W )2 S 2 W 2 (3S − W )2
= +
4(4S 2 − W 2 ) 4(4S 2 − W 2 )2

S 3 S 3 − 2S 2 W + 3SW 2 − W 3
=  2
.
4S 2 − W 2

If triangle I OP is equilateral, then I O 2 = OP 2 ; that is,


S 2 (S − W )2 S 3 S 3 − 2S 2 W + 3SW 2 − W 3
= 
4S 2 − W 2 4S 2 − W 2
2


(S − W )2 4S 2 − W 2 = S 4 − 2S 3 W + 3S 2 W 2 − SW 3
3S 4 − 6S 3 W + 3SW 3 − W 4 = 0.

Dividing by W 4  = 0 and letting x = S/W gives the equation

3x 4 − 6x 3 + 3x − 1 = 0.

Substituting x = z + 1/2 and multiplying by 16, we get

48z4 − 72z2 − 1 = 0,

so that
√ √ √
72 ± 16 21 3 21 3 ± 2 7/3
z =
2
= ± = ,
96 4 6 4

 √
± 3 ± 2 7/3
z= ,
2

and
 √
1± 3 ± 2 7/3
x= .
2

Since x = S
W
is a positive real number, there is a unique solution:

 √
S 1+ 3 + 2 7/3
= ≈ 1.7303506.
W 2

Also solved by Michel Bataille, Rouen, France; James Duemmel, Bellingham, WA; Jeffrey Groah, Lone
Star C. - Montgomery; Eugene Herman, Grinnell C.; Elias Lampakis, Kiparissia, Greece; Volkhard
Schindler, Berlin, Germany; Randy Schwartz, Schoolcraft C. (retired); Albert Stadler, Herrliberg,
Switzerland; Enrique Treviño, Lakeforest C.; Michael Vowe, Therwil, Switzerland; and the proposer.

72 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 Ubiquitous zero divisors without nontrivial nilpotent elements implies
infinite
1192. Proposed by Greg Oman, University of Colorado at Colorado Springs, Colorado
Springs, CO.
Let R be a commutative ring (not assumed to have an identity). Recall that an element
x ∈ R is a zero divisor if there is some nonzero y ∈ R such that xy = 0; x is nilpotent
if x n = 0 for some positive integer n (note that we do not require a zero divisor to be
nonzero).
(a) Prove or disprove: there exists a finite commutative ring R for which
1. every element of R is a zero divisor, and
2. the only nilpotent element of R is 0.

(b) Does your answer change if “finite” is replaced with “infinite”?

Solution by Northwestern University Math Problem Solving Group.
1. The answer is negative, i.e., there is no finite commutative ring satisfying 1 and
2. If R = {0} (the trivial ring), then 0 is not a zero divisor, sit fails to satisfy 1.
Hence, we may assume that R is non-trivial, and the proof proceeds as follows.
Let S = {x1 , x2 , ..., xn } be a maximal set (n maximum) of distinct non-zero
elements of R with the property xi xj = 0 for every i  = j . Denote s = x1 + x2 +
· · · + xn its sum. Then
(a) We h ave s  = 0 because otherwise x1 = −x2 − · · · − xn , hence x12 =
−x2 x1 − · · · − xn x1 = 0, contradicting the assumption that 0 is the only
nilpotent element.
(b) Since all elements of R are zero divisors, there must be a non-zero r such
that 0 = rs = rx1 + rx2 + · · · + rxn . Hence, for each i = 1, ..., n, we have
n n
rxi = − → (rxi )2 = r (rxi ) xi = −r rxj xi = 0 → rxi = 0.
j =1 j =1
j =i j =i

This implies that the set S = {r, x1 , x2 , ..., xn } also has the property that ev-
ery pair of distinct elements in it has product zero, but S has n + 1 elements,
contradicting the maximality of S.
2. For infinite rings, the answer is affirmative. An example is the ring R of infinite
sequences of integers with finitely many non-zero elements (and term-wise ad-
dition and multiplication). This ring satisfies the required properties, as shown
below.
• Property 1: If {an }n∈N is in R, then there will be some (in fact infinitely
many) m ∈ N such that am = 0. Given a fixed m such that am = 0, let
bm = 1 and bn = 0 for n  = m. Then we have that {bn }n∈N is not zero, but
an bn = 0 for every n, so that {an }n∈N is a zero divisor.
• If k ≥ 1, then, for each n, ank = 0 if and only if an = 0. Hence the zero
element of R, consisting of the sequence with all terms zero, is the only
nilpotent element in R.
Also solved by Egle Bettio and Liceo Benedetti-Tommaseo, Venezia, Italy; Anthony Bevelacqua, U. of
N. Dakota; Paul Budney, Sunderland, MA; Elias Lampakis, Kiparissia, Greece; and the proposer.

VOL. 53, NO. 1, JANUARY 2022 THE COLLEGE MATHEMATICS JOURNAL 73

A function that is a polynomial over the rationals in each slot separately
need not be a polynomial over Q2

1193. Proposed by George Stoica, Saint John, New Brunswick, Canada.

Let f : Q × Q → Q be a function such that y → f (a, y) is a polynomial over Q for
every a ∈ Q and x → f (x, b) is a polynomial over Q for every b ∈ Q. Is it true that
f (x, y) is a polynomial in (x, y) ∈ Q2 ?
Solution by Paul Budney, Sunderland, Massachusetts.
Such functions exist which cannot be defined by a polynomial in Q[x, y]. Let r1 , r2 , ...
be a faithfully-indexed sequence of the rationals. Define f : Q2 → Q by
∞ k
f (x, y) = (x − ri ) (y − ri ) .
k=1 i=1

For each (x, y) = (rm , rn ) ∈ Q2 , this series has only finitely many non-zero terms, so
it converes on Q2 . For any rational x = rn , if n > 1,
n−1 k
f (rn , y) = (rn − ri ) (y − ri ) ∈ Q[y],
k=1 i=1

a polynomial of degree n − 1. If n = 1, f (r1 , y). Similarly, for n > 1, f (x, rn ) ∈

Q[x], a polynomial of degree n − 1. Also, f (x, r1 ) = 0. Now, if f is defined by a
polynomial f (x, y) ∈ Q[x, y], we can choose a positive integer n > deg[f (x, y)] =
d > 0. But then f (rn+1 , y) is a polynomial of degree n and also a polynomial of
degree at most d < n. This is impossible since non-constant polynomials have only
finitely many zeros. Thus f (x, y) can’t be defined by a polynomial in Q[x, y].
Also solved by Gerald Edgar, Denver, CO; Albert Natian, Los Angeles Valley C.; Kenneth Schilling, U.
of Michigan - Flint; and the proposer. One incomplete solution and one incorrect solution were received.

A two-variable inequality over the integers

1194. Proposed by Andrew Simoson, King University, Bristol, TN.

Let a and b be positive
√ integers with a ≥ b. Prove the following:
b
(a) a+b + a+b > 5, and
b √ √
a
(b) either a+b + a+b
a
> 5 or ab + ab > 5.
Solution by Charlie Mumma, Seattle, Washington.
√ √
For convenience, set c = ( 5 − 1)/2, d = ( 5 + 1)/2, and f (x) = x + 1/x. Ob-
serve that√f is strictly decreasing on (0, 1), strictly increasing on (1, ∞), and f (c) =
f (d) = 5. Since a ≥ b, (a + b)/b ≥ 2 >√d, which proves (a) [f ((a + b)/b) >
f (d)]. Next notice that when a/b + b/a ≤ 5, c ≤ b/a ≤ 1. Hence √ (a + b)/a =
1 + b/a ≥ 1 + c = d. If a = b, a/(a + b) + (a √ + b)/a = 5/2 > 5. However, for
b = ca, a/(a + b) + (a + b)/a = a/b + b/a = 5. Thus (b) is true so long as a/b
is not the golden ratio (a condition less stringent than the requirement that both a and
b be integers).
Also solved by Ulrich Abel, Technische Hochschule Mittelhessen, Germany and Georg Arends, Eschweiler,
Germany (jointly); Farrukh Rakhimjanovich Ataev, Westminster International U., Tashkent, Uzbekistan;

74 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Michel Bataille, Rouen, France; Brian Beasley, Presybterian C.; Brian Bradie, Christopher Newport U.;
Kyle Calderhead, Malone U.; John Christopher, California St. U., Sacramento; Christopher Newport
U. Problem Solving Seminar; Matthew Creek, Assumption U.; Richard Daquila, Muskingum U.; Eagle
Problem Solvers, Georgia Southern U.; Habin Far, Lone Star C. - Montgomery; Dmitry Fleischman, Santa
Monica, CA; Davide Fusi, U. of South Florida Beaufort; Russ Gordon, Whitman C.; Lixing Han, U. of Michi-
gan - Flint and Xinjia Tang, Chang Zhou U.; Eugene Herman, Grinnell C.; Donald Hooley, Bluffton, OH;
Tom Jager, Calvin U.; A. Bathi Kasturiarachi, Kent St. U. at Stark; Elias Lampakis, Kiparissia, Greece;
Kee-Wai Lau, Hong Kong, China; Seungheon Lee, Yonsei U.; Graham Lord, Princeton, NJ; Rhea Malik;
Northwestern U. Math Problem Solving Group; Ángel Plaza and Francisco Perdomo, Universidad de
Las Palmas de Gran Canaria, Las Palmas, Spain; Mark Sand; Randy Schwartz, Schoolcraft C. (retired); Al-
bert Stadler, Herrliberg, Switzerland; Enrique Treviño, Lake Forest C.; Michael Vowe, Therwil, Switzer-
land; Roy Willits; Lienhard Wimmer; and the proposer.

A sum of harmonic sums

1195. Proposed by Marián Štofka, Slovak University of Technology, Bratislava, Slovak

Republic.
Prove the following:

∞  2 
Hk π π4
− Hk+1,2 = ,
k=1
k+1 6 90

where Hk = ki=1 1i is the kth harmonic number and Hk,2 = k 1

i=1 i 2 is the kth gener-
alized harmonic number.
Solution by Russ Gordon, Whitman College, Walla Walla, WA.

Since 16 π 2 = ∞ 2
k=1 1/k , we can express the given sum as

n=1 ∞ ∞ ∞
= .
∞ k=n+2
(n + 1)k 2 n=1 k=1
(n + 1)(n + k + 1)2

Using integration by parts, it is not difficult to verify that

 1  1
1 2
−x n−1 ln x dx = 2 and x n−1 (ln x)2 dx = 3
0 n 0 n
for each positive integer n. We also make note of the following Macluarin series:
∞ ∞
1 n (ln(1 − x))2 hn n
− ln(1 − x) = x and = x .
n=1
n 2x n=1
n+1

Using this information, we find that

∞ ∞ ∞ ∞  1
hn hn
= −x n+k ln x dx
n=1 k=1
(n + 1)(n + k + 1)2 n=1
n+1 k=1 0

∞ 
hn 1
−x n+1
= ln x dx
n=1
n+1 0 1−x

VOL. 53, NO. 1, JANUARY 2022 THE COLLEGE MATHEMATICS JOURNAL 75

  ∞
1
− ln x hn n+1
= x dx
0 1−x n=1
n+1
 1
− ln x (ln(1 − x))2
= · dx
0 1−x 2

1 1
− ln(1 − x)(ln x)2
= dx
2 0 x
∞ 
1 1 1 n−1
= x (ln x)2 dx
2 n=1
n 0
∞
1 2
=
2 n=1
n4

π4
= ,
90
the desired result.
Also solved by Michel Bataille, Rouen, France; Gerald Bilodeau, Boston Latin School; Khristo Boy-
adzhiev, Ohio Northern U.; Paul Bracken, U. of Texas, Edinburg; Brian Bradie, Christopher Newport U.;
Bruce Burdick, Roger Williams U.; Hongwei Chen, Christopher Newport U.; Lixing Han, U. of Michigan-
Flint and Xinjia Tang, Chang Zhou U.; Eugene Herman, Grinnell C.; Omran Kouba, Higher Inst. for Applied
Sci. and Tech., Damascus, Syria. Elias Lampakis, Kiparissia, Greece; Albert Stadler, Herrliberg, Switzer-
land; Seán Stewart, Bomaderry, NSW, Australia; Michael Vowe, Therwil, Switzerland; and the proposer.

Editor’s note: The name of James Brenneis was omitted from the list of solvers of
problem 1183 in the November 2021 issue. We apologize for the omission.

76 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 SOLUTIONS

A continued fraction given by Fibonacci

1186. Proposed by Gregory Dresden, Washington and Lee University, Lexington,
VA and ZhenShu Luan (high school student), St. George’s School, Vancouver, BC,
Canada.
Find a closed-form expression for the continued fraction [1, 1, . . . , 1, 3, 1, 1, . . . ,
1], which has n ones before, and after, the middle three.
Solution by Walther Janous, Ursulinengymnasium, Innsbruck, Austria.
In order to get the desired expression, we recall the following elegant way of evaluat-
ing the convergents of a continued fraction. [See, for instance,
https://de.wikipedia.org/wiki/Kettenbruch, particularly the paragraph “matrixdarstel-
lung.”] We have to evaluate the product

 n    n
1 1 3 1 1 1
· ·
1 0 1 0 1 0

Let Fn be the nth Fibonacci number. From the familiar representation

Fn+1
[1, 1, ..., 1] = ,
Fn
(with n 1’s), we get
 n  
1 1 F Fn
= n+1 ,
1 0 Fn Fn−1

whence
 n    n      
1 3 1 1 1 F Fn 3 1 F Fn
· · = n+1 · · n+1 ;
1 0 1 0 1 0 Fn Fn−1 1 0 Fn Fn−1

VOL. 52, NO. 5, NOVEMBER 2021 THE COLLEGE MATHEMATICS JOURNAL 389
that is
 n    n
1 3 1 1 1
· ·
1 0 1 0 1 0
 
Fn+1 · (3Fn+1 + 2Fn ) Fn+1 · Fn−1 + Fn (3Fn+1 + Fn )
= .
Fn+1 · Fn−1 + 3Fn Fn+1 + Fn2 Fn (2Fn−1 + 3Fn )

This leads to the desired closed-form expression of [1, ..., 1, 3, 1, ..., 1]:

Fn+1 (3Fn+1 + 2Fn ) Fn+1 (3Fn+1 + 2Fn )

=
Fn+1 · Fn−1 + 3Fn · Fn+1 + Fn
2 Fn+1 (Fn+1 − Fn ) + Fn (3F − n + 1 + Fn )
Fn+1 (3Fn+1 + 2Fn )
=
Fn+1 + 2Fn+1 · Fn + Fn2
2

Fn+1 (3Fn+1 + 2Fn )

=
(Fn+1 + Fn )2
Fn+1 (3Fn+1 + 2Fn )
= 2
Fn+2
Fn+1 (Fn+1 + 2Fn+2 )
= 2
.
Fn+2

This and

Fn+1 + 2Fn+2 = Fn+3 + Fn+2 = Fn+4

yield the closed-form result

Fn+1 Fn+4
2
.
Fn+2

Also solved by Brian Beasley, Presbyterian C.; Anthony Bevelacqua, U. of N. Dakota; Brian Bradie,
Christopher Newport U.; James Brenneis, Penn State - Shenango; Hongwei Chen, Christopher Newport
U.;Giuseppe Fera, Vicenza, Italy; Eugene Herman, Grinnell C.; Donald Hooley, Bluffton, OH; Joel Iiams,
U. of N. Dakota; Harris Kwong, SUNY Fredonia; Seungheon Lee, Yonsei U.; Carl Libis, Columbia Southern
U.; Graham Lord, Princeton, NJ; Ioana Mihaila, Cal Poly Pomona; Missouri State U. Problem Solving
Group; Northwestern U. Math Problem Solving Group; Randy Schwartz, Schoolcraft C. (retired); Al-
bert Stadler, Herrliberg, Switzerland; Paul Stockmeyer, C. of William and Mary; David Terr, Oceanside,
CA; Enrique Treviño, Lakeforest C.; Michael Vowe, Therwil, Switzerland; and the proposer.

A limit of maxima
1187. Proposed by Reza Farhadian, Lorestan University, Khorramabad, Iran.
Let α > 1 be a fixed real number, and consider the function M : [1, ∞) → N defined
by M(x) = max{m ∈ N : m! ≤ α x }. Prove the following:
√
n
M(1)M(2) · · · M(n)
lim = e−1 .
n→∞ M(n)

390 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Solution by Randy Schwartz, Schoolcraft College (retired), Ann Arbor, Michigan.
From the definition of the function M, we have [M(n) + 1]! > α n for α > 1, so
limn→∞ M(n) = ∞, and thus limn→∞ ln M(n) = ∞. Also from the definition, we
have
ln([M(n)]!)
[M(n)]! ≤ α n ⇒ ln([M(n)]!) ≤ n ln α ⇒ ≤ ln α,
n
and thus
ln([M(n)]!)
lim ≤ ln α. (1)
n→∞ n
Applying Stirling’s approximation to (1) leads to

M(n) + 12 ln M(n) − M(n) + 12 ln 2π
lim ≤ ln α
n→∞ n
 
M(n) ln M(n) ln 2π
lim (ln M(n) − 1) + + ≤ ln α
n→∞ n 2n 2n
 
M(n) ln M(n)
lim (ln M(n) − 1) + ≤ ln α
n→∞ n 2n

The last term inside the brackets is nonnegative and, from the foregoing, the factor
ln M(n) − 1 increases without bound; thus, M(n) n
must vanish, since otherwise the
above limit could not be a finite number such as ln α. Thus, we have established
M(n)
∈n→∞ = 0.
n
We can deduce more the definition of the function M:

[M(n) + 1]!α n
[M(n) + 1]M(n)! > α n
αn
[M(n)]! >
M(n) + 1
ln([M(n)!) > n ln α − ln[M(n) + 1]
ln([M(n)!]) ln[M(n) + 1]
> ln α −
n n
ln([M(n)]!)
lim ≥ ln α,
n→∞ n
and combining this with (1) yields

ln)[M(n)!])
lim = ln α
n→∞ n
and then
ln)[M(n)!])
lim = 1. (2)
n→∞ n

VOL. 52, NO. 5, NOVEMBER 2021 THE COLLEGE MATHEMATICS JOURNAL 391
 Using Stirling again, we have

ln([M(n)!) M(n) + 1
ln M(n) − M(n) + 12 ln 2π
lim = lim 2
n→∞ M(n) ln M(n) n→∞ M(n) ln M(n)
M(n) + 1
1 ln 2π
= lim 2
− +
n→∞ M(n) ln M(n) 2M(n) ln M(n)

= 1 − 0 + 0 = 1.
and combining this with (2) yields
M(n) ln M(n)
lim = 1. (3)
n→∞ n ln α
We can now calculate the requested value, L. We have

 n
n n
h=1 M(h)
 M(h)
L = lim = lim n
,
n→∞ M(n) n→∞
h=1
M(n)

and then
n  
1 M(h)
ln L = lim ln .
n→∞
h=1
n M(n)

There are many repeated terms in the above summation. The interval between (j −
1)! and j !, involving as it does a multiplication by j , encloses approximately logα j
powers of α, each one of them associated with the same value of the function M. In
other words, the number of integer solutions of M(n) = j is asymptotically logα j =
ln j
ln α
. Using that as a weighting factor to gather the repeated terms, we can rewrite the
above summation as
 1 ln j  j 
M(n)
ln L = lim · ln
n→∞
j =1
n ln α M(n)


M(n)
(ln j )2 − ln j · ln M(n)
= lim
n→∞
j =1
n ln α


M(n)
(ln j )2 − ln j · ln M(n)
= lim , using (3),
n→∞
j =1
M(n) ln M(n)

and thus
⎡ ⎤
1 
M(n)
1 
M(n)
ln L = lim ⎣ (ln j )2 − ln j ⎦ . (4)
n→∞ M(n) ln M(n) j =1
M(n) j =1

Using inscribed and circumscribed rectangles, we have

 k 
k  k+1
ln x dx < ln j < ln x dx
1 j =1 2

392 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 
k  k
ln j ≈ ln x dx = k ln k − k + 1
j =1 1

1
k
lim ln j = ln k − 1,
n→∞ k
j =1

and similarly


k  k
(ln j )2 ≈ (ln x)2 dx = k(ln k)2 − 2k ln k + 2k − 2
j =1 1

1 
k
lim (ln j )2 = ln k − 2.
n→∞ k ln k
j =1

Applying these to (4) yields

ln L = lim [(ln M(n) − 2) − (ln M(n) − 1)] = −1,

n→∞

and thus

L = e−1 .

Also solved by Dmitry Fleischman, Santa Monica, CA; Lixing Han, U. of Michigan-Flint and Xinjia Tang,
Chang Zhou U.; Albert Stadler, Herrliberg, Switzerland; and the proposer.

A recursively defined sequence of trigonometric functions

1188. Proposed by Ángel Plaza, Universidad de Las Palmas de Gran Canaria, Las
Palmas de Gran Canaria, Spain.
 f (x)
Let {fn (x)}n≥1 be the sequence offunctions recursively defined by fn (x) = 0 n−1 sin tdt,
x
with initial condition f1 (x) = 0 sin tdt. For each n ∈ N, find the value of pn
such that Ln = lim fxnp(x)
n ∈ R\{0} and the corresponding value Ln . Prove also that
x→0
−1 −1
log2 (L−1
n ) = 3 log2 (Ln−1 ) − 2 log2 (Ln−2 ) for n ≥ 3.

Solution by Michael Vowe, Therwil, Switzerland.

We have
 x
x2 
f1 (x) = sin t dt = 1 − cos x = + O x4
0 2!

and hence p1 = 2, L1 = 12 . Further

f2 (x) = 1 − cos(1 − cos x)

   
1 − cos x 2 1 − cos x 4 x4 
= − + ··· = + O x6 ,
2! 4! 2!4

VOL. 52, NO. 5, NOVEMBER 2021 THE COLLEGE MATHEMATICS JOURNAL 393
which means that p2 = 4, L2 = 18 .
Since
1
fn (x) = 1 − cos (fn−1 (x)) , p1 = 2, L1 = ,
2
we obtain

pn = 2pn−1 = 2 · 2pn−2 = · · · = 2n−1 p1 = 2n

and
1 1 1 1 n−1
Ln = (Ln−1 )2 = · 2
(Ln−1 )4 = · · · = 1+2+4+···+2n−2 (L1 )2
2! 2! (2!) 2!
1 1 1
= 2n−1 −1 · 2n−1 = 2n −1 .
2 2 2
Now
  −1  n−1 
3 log2 L−1
n−1 − 2 log2 Ln−2 = 3 2 − 1 − 2 2n−2 − 1
n 
= 2 · 2n−1 − 1 = 2n − 1 = log2 22 −1 = log2 L−1
n .

Also solved by Michel Bataille, Rouen, France; Brian Bradie, Christopher Newport U.; Paul Budney, Sun-
derland, MA; Hongwei Chen, Christopher Newport U.; Christopher Newport U. Problem Solving Sem-
inar; Gerald Edgar, Denver, CO; Lixing Han, U. of Michigan-Flint; Justin Haverlick, State U. of New
York at Buffalo; Eugene Herman, Grinnell C.; Christopher Jackson, Coleman, Florida; Elias Lampakis,
Kiparissia, Greece; Albert Natian, Los Angeles Valley C.; Mark Sand, C. of Saint Mary; Randy Schwartz,
Schoolcraft C. (retired); Albert Stadler, Herrliberg, Switzerland; Seán Stewart, Bomaderry, NSW, Aus-
tralia; and the proposer. One incomplete solution and one incorrect solution were received.

A sum of harmonic sums

1189. Proposed by Seán Stewart, Bomaderry, NSW, Australia.
Evaluate the following sum:


∞
Hn+1 + Hn − 1
,
n=1
(n + 1)(n + 2)
n
where Hn = 1
k=1 k denotes the nth harmonic number.
Solution by Robert Agnew, Palm Coast, Florida.
The sum

∞
Hn+1 + Hn − 1
S=
n=1
(n + 1)(n + 2)

can be written as
 

∞
1 1 n
1
S= −1 + +2·
n=1
(n + 1)(n + 2) n+1 k=1
k

394 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 
∞
1  ∞
1  1
∞ 1 n
=− + + 2 · .
n=1
(n + 1)(n + 2) n=1 (n + 1)2 (n + 2) n=1
(n + 1)(n + 2) k=1 k

Evaluating each of these sums in turn gives


∞ ∞  
1 1 1 1
= − = ;
n=1
(n + 1)(n + 2) n=1
n+1 n+2 2


∞ ∞  
1 1 1 1
= − + +
n=1
(n + 1)2 (n + 2) n=1
n + 1 n + 2 (n + 1)2
∞    ∞
1 1 1
=− − +
n=1
n+1 n+2 n=1
(n + 1)2
 2 
1 π
=− + −1
2 6
3 π2
=− + ;
2 6
and

∞
1 1 1
n ∞ ∞
1
=
n=1
(n + 1)(n + 2) k=1 k k=1
k n=k (n + 1)(n + 2)
 ∞  
1 
∞ ∞
1 1 1
= − =
k=1
k n=k n + 1 n + 2 k=1
k(k + 1)

= 1.

Hence
π2
S= .
6

Also solved by Arkady Alt, San Jose, CA; Farrukh Rakhimjanovich Ataev, Westminster International U.,
Tashkent, Uzbekistan; Michel Bataille, Rouen, France; Necdet Batir, Nevşehir Haci Bektaş Veli U.; Khristo
Boyadzhiev, Ohio Northern U.; Paul Bracken, U. of Texas, Edinburg; Brian Bradie, Christopher Newport
U.; Hongwei Chen, Christopher Newport U.; Geon Choi, Yonsei U.; Nandan Sai Dasireddy, Hyderabad,
India; Bruce Davis, St. Louis Comm. C. at Florissant Valley; Giuseppe Fera, Vicenza, Italy; Subhankar
Gayen, West Bengal, India; Michael Goldenberg, Baltimore Polytechnic Inst. and Mark Kaplan, U. of
Maryland Global Campus; GWStat Problem Solving Group; Lixing Han, U. of Michigan - Flint and Xinjia
Tang, Chang Zhou U.; Eugene Herman, Grinnell C.; Walther Janous, Innsbruck, Austria; Kee-Wai Lau,
Hong Kong, China; Seungheon Lee, Yonsei U.; Graham Lord, Princeton, NJ; Missouri State U. Problem
Solving Group; Shing Hin Jimmy Pa; Ángel Plaza and Francisco Perdomo, Universidad de Las Palmas
de Gran Canaria, Las Palmas, Spain; Rob Pratt, Apex, NC; Arnold Saunders, Arlington, VA; Volkhard
Schindler, Berlin, Germany; Randy Schwartz, Schoolcraft C. (retired); Allen Schwenk, Western Michigan
U. Albert Stadler, Herrliberg, Switzerland; Marián Ŝtofka, Slovak U. of Technology; Enrique Treviño,
Lake Forest C.; Michael Vowe, Therwil, Switzerland; and the proposer.

VOL. 52, NO. 5, NOVEMBER 2021 THE COLLEGE MATHEMATICS JOURNAL 395
 A second-order differential equation
1190. Proposed by George Stoica, Saint John, New Brunswick, Canada.
Find all twice differentiable functions y = y(x) such that (y + x)y = y (y + 1).
Solution by Eugene Herman, Grinnell College, Grinnell, Iowa.
Substituting z(x) = y(x) + x into the differential equation yields zz = (z − 1)z .
This has solutions z = k and z = x + k. Other than these, we have
 
d z −1 zz − (z − 1)z
= =0
dx z z2
kecx −1
and so z − 1 = cz, where c  = 0. Separating variables yields z = c
. Therefore,
the solutions for y are

kecx − 1
k − x, k, − x (where c  = 0).
c

Editor’s note: Solvers exercised various degrees of care in ensuring the existence of an
interval on which one could safely avoid dividing by zero. In the interests of space, we
have not incorporated that analysis here.
Also solved by Robert Agnew, Palm Coast, FL; Arkady Alt, San Jose, CA; Tomas Barajas, U. of Arkansas at
Little Rock; Michel Bataille, Rouen, France; Paul Bracken, U. of Texas, Edinburg; Brian Bradie, Christo-
pher Newport U.; Hongwei Chen, Christopher Newport U.; Richard Daquila, Muskingham U.; Bruce Davis,
St. Louis Comm. C. at Florissant Valley; Michael Goldenberg, Baltimore Polytechnic Inst. and Mark Ka-
plan, U. of Maryland Global Campus; Anna DePoyster, Missie Bogard, Rylee Buck, and Chanty Gray,
(students) U. of Arkansas, Little Rock; Raymond Greenwell, Hofstra U.; Lixing Han, U. of Michigan-Flint
and Xinjia Tang, Chang Zhou U.; Justin Haverlick, State U. of New York at Buffalo; Logan Hodgson;
Walther Janous, Innsbruck, Austria; Harris Kwong, SUNY Fredonia; Seungheon Lee, Yonsei U.; William
Littlejohn, Jason Pearson, and Cole Stillman (students) U. of Arkansas, Little Rock; James Magliano,
Union Country C. (emeritus); Albert Natian, Los Angeles C.; Randy Schwartz, Schoolcraft C. (retired); Al-
bert Stadler, Herrliberg, Switzerland; Seán Stewart, Bomaderry, NSW, Australia; Nora Thornber, Raritan
Valley Comm. C.; and the proposer.

396 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 SOLUTIONS

Two limits of integrals

1181. Proposed by Ovidiu Furdui and Alina Sı̂ntămărian, Technical University of
Cluj-Napoca, Cluj-Napoca, Romania.
Let k > 0 be a real number. Calculate the following:

√
n
1 x+k−1 n
1. L := limn→∞ 0 k
dx, and
√
n
1 n
2. limn→∞ n 0
x+k−1
k
dx − L .

Solution by Seán Stewart, Bomaderry, NSW, Australia.

We will show that for k > 0,
 1 √ n
n
x+k−1 k
1. L = lim dx = , and
n→∞ 0 k k+1
 1  √ n 
n
x+k−1 k(k − 1)
2. lim n dx − L = .
n→∞ 0 k (k + 1)3

VOL. 52, NO. 4, SEPTEMBER 2021 THE COLLEGE MATHEMATICS JOURNAL 307
 We first find an asymptotic expansion for the term
 1
n
xn +k −1
J = ,
k

for large n. For x ∈ (0, 1), from the Maclaurin series expansion for the exponential
function as y → 0 we have

1
exp (y log x) = 1 + y log(x) + y 2 log2 (x) + O(y 3 ).
2

Setting y = 1
n
then as n → ∞, we have
   
1 log(x) log2 (x)
1 1
exp log x =x =1+ n + 2
+O .
n n 2n n3

Thus
1  
xn −1 log(x) log2 (x) 1
= + + O .
k nk 2n2 k n3

Now
 1
    
xn −1 log(x) log2 (x) 1
log J = n log 1 + = n log 1 + + 2
+O . (1)
k nk 2n k n3

From the Maclaurin series expansion for log(1 + x), as x → 0, we have

x2
log(1 + x) = x − + O(x 3 ).
2
Using this result we can write (1) as
  
log(x) (k − 1) log2 (x) 1
log J = n + 2 2
+O
nk 2n k n3
 1  (k − 1) log2 (x)  
1
= log x k + +O .
2nk 2 n2

Thus
    
(k − 1) log2 (x)
1 1
J =e = exp log x +
log J
2
k +O
2nk n2
  
1 (k − 1) log2 (x) 1
= x k exp + O . (2)
2nk 2 n2

From the Maclaurin series expansion for the exponential function, as x → 0, we have

ex = 1 + x + O(x 2 ).

308 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Using this result we can write (2) as
  
(k − 1) log2 (x)
1 1
J =x 1+ k
2
+O
2nk n2
1  
1 (k − 1)x k log2 (x) 1
=x +
k
2
+O , (3)
2nk n2

as n → ∞ and is the asymptotic expansion we sought for the term J .

Let
 1 √ n
n
x+k−1
In = dx.
0 k

From the result given for the asymptotic expansion in (3), an asymptotic expansion for
the integral In as n → ∞ is
   
1
1 k−1 1
1 1
In = x dx +
k x log (x) dx + O
k 2
. (4)
0 2nk 2 0 n2

The first of the integrals to the right of the equality is elementary. The result is
 1
1 k
x k dx = .
0 k+1

For the second of the integrals to the right of the equality, enforcing a substitution of
x → x k produces
 1  1
1
x log (x) dx = k
k 2 3
x k log2 (x) dx.
0 0

Integrating by parts twice leads to

 1
1 2k 3
x k log2 (x) dx = .
0 (k + 1)3

Thus (4) becomes

 
k k(k − 1) 1
In = + +O . (5)
k + 1 n(k + 1) 3 n2

Using the result given in (5), we are now in a position to answer the questions asked
in each part. For the first part
  
k k(k − 1) 1 k
L = lim In = lim + +O = ,
n→∞ n→∞ k + 1 n(k + 1) 3 n 2 k+1

as announced. And for the second part.

   
k k(k − 1) 1 k
lim n(In − L) = lim n + +O −
n→∞ n→∞ k + 1 n(k + 1)3 n2 k+1

VOL. 52, NO. 4, SEPTEMBER 2021 THE COLLEGE MATHEMATICS JOURNAL 309
   
k(k − 1) 1
= lim + O
n→∞ (k + 1) 3 n
k(k − 1)
= ,
(k + 1)3

as announced.
Also solved by Robert Agnew, Palm Coast, FL (part 1 only); Paul Brracken, U. of Texas, Edinburg; Brian
Bradie, Christopher Newport U.; Hongwei Chen, Christopher Newport U.; James Duemmel, Bellingham,
WA; Giuseppe Fera, Vicenza, Italy; Dmitry Fleischman, Santa Monica, CA (part 1 only); Russ Gordon,
Whitman C.; Walther Janous, Innsbruck, Austria (part 1 only); Albert Stadler, Herrliberg, Switzerland;
and the proposer. One incorrect solution was received.

The edge of convergence

1182. Proposed by Adam Hammett, Cedarville University, Cedarville, OH.
Let c ∈ R, let {ak }k≥1 be a sequence of real numbers satisfying ak − ak−1 ≥ ak+1 −
ak ≥ 0 for all k ≥ 2, and introduce the power series

 (−1)n
χ(c, {ak }, x) := (an−1 − c) .
n≥2
xn

1. Find a real number r > 0 such that χ(c, {ak }, x) converges absolutely for x > r
and all choices of c and {ak }, but χ(c, {ak }, r) diverges for for some choice of c
or {ak }, and
2. Prove that there exists a function f (c, {ak }) ≥ r and a threshold value c∗ such
that χ(c, {ak }, x) > 0 for each c < c∗ and x > f (c, {ak }), and χ(c, {ak }, x) < 0
for each c > c∗ and x > f (c, {ak }). Give an explicit formula for f (c, {ak }) and
value for c∗ .

Solution by the proposer.

Since the constant sequence {ak } = {1} satisfies the sequence condition, and χ(0, {1}, 1)
diverges, it becomes clear that r ≥ 1. Let’s show that we actually have r = 1. For this,
assuming x > 0, note that by the triangle inequality and the condition on {ak } we have
 1  1
|an−1 − c| n
= |an−1 − an−2 + an−2 − an−3 + · · · + a2 − a1 + a1 − c| n
n≥2
x n≥2
x
 1
≤ (an−1 − an−2 + an−2 − an−3 + · · · + a2 − a1 + |a1 − c|)
n≥2
xn
 (n − 1)M(c, {ak })
≤ , M(c, {ak }) := max{a2 − a1 , |a1 − c|}.
n≥2
xn

Applying the ratio test to this last series, we obtain

 
nM(c, {ak })/x n+1 n 1 1
= → , n → ∞,
(n − 1)M(c, {ak })/x n n−1 x x

310 © THE MATHEMATICAL ASSOCIATION OF AMERICA

and so absolute convergence of χ(c, {ak }, x) is guaranteed for 1/x < 1, i.e. x > 1. So
r = 1 and part (a) is solved. Consequently, below we will assume x > 1.
Now on to (b). By appropriately “shifting” the terms in the series (c, {ak }, x)
and taking advantage of their alternating nature, we can remove c from all but one
term, making our analysis far simpler. To this end, introduce ψ(c, {ak }, x) = (x 2 +
x)χ(c, {ak }, x) and note that
 (−1)n
ψ(c, {ak }, x) = x 2 + x (an−1 − c)
n≥2
xn

(a2 − c) (a3 − c) (a4 − c)

= (a1 − c) − + − + ···
x x2 x3
(a1 − c) (a2 − c) (a3 − c)
+ − + − ···
x x2 x3
 (−1)n−1
= (a1 − c) + (an − an−1 ) . (6)
n≥2
x n−1

Since x 2 + x > 0 for x > 1, it follows that ψ(c, {ak }, x) and χ(c, {ak }, x) have the
same sign for x > 1. So let’s analyze ψ(c, {ak }, x) as defined in (6), which will involve
a careful case analysis for various c-values.
To start, what if consecutive terms of the sequence {ak } are ever equal? If, say,
am = am+1 for some minimal m ≥ 1, then the sequence condition implies

0 = am+1 − am ≥ ak+1 − ak ≥ 0, k ≥ m,

that is ak = ak+1 for all k ≥ m. So the sequence {ak } is constant from the mth term
onward, and hence in this case ψ(c, {ak }, x) is a finite polynomial:

 (−1)n−1
ψ(c, {ak }, x) = (a1 − c) + (an − an−1 ) . (7)
2≤n≤m
x n−1

Here, the sum in (7) may well be empty (i.e. m = 1), and this would correspond to the
case where {ak } is a constant sequence. If the sum is nonempty with at least two terms
(i.e. m ≥ 3), then the magnitude of the ratio of consecutive terms in the sum is
 
(an+1 − an )/x n an+1 − an 1
= <1
(an − an−1 )/x n−1 an − an−1 x

for 2 ≤ n < m and x > 1, since (an+1 − an )/(an − an−1 ) ≤ 1 due to the sequence
condition. Hence, this alternating sum has terms that decrease in magnitude, and so

(a2 − a1 )
a1 − c − ≤ ψ(c, {ak }, x) ≤ a1 − c for x > 1, c ∈ R. (8)
x
So from the right-hand side of (8), it follows that

ψ(c, {ak }, x) ≤ a1 − c < 0 for c > a1 , x > 1.

Consequently we have χ(c, {ak }, x) < 0 for c > a1 and x > 1.

VOL. 52, NO. 4, SEPTEMBER 2021 THE COLLEGE MATHEMATICS JOURNAL 311
 And what happens when c < a1 ? Notice that for x > 1 we have
(a2 − a1 ) a2 − a1
a1 − c − >0 ⇐⇒ (a1 − c)x − (a2 − a1 ) > 0 ⇐⇒ x> .
x a1 − c
(9)
Here, the last algebraic manipulation requires that c < a1 in order to safely divide
through and preserve the direction of the inequality. So, invoking the left-hand side
inequality in (8) we see that
 
(a2 − a1 ) a2 − a1
ψ(c, {ak }, x) ≥ a1 − c − >0 for x > max 1, , c < a1 .
x a1 − c
This means that χ(c, {ak }, x) > 0 for x > max{1, (a2 − a1 )/(a1 − c)} and any fixed
c < a1 .
Finally, it remains to check the case where the sequence has all consecutive terms
differing. Clearly, the condition on {ak } implies that the sequence is non-decreasing,
and so in this case we would have a strictly increasing sequence a1 < a2 < · · · . But
careful examination of the argument just given for an eventually constant sequence
shows that the same analysis goes through. So in summary, given a sequence {ak }
satisfying our condition we’ve shown that for fixed c > a1 , χ(c, {ak }, x) < 0 for x >
1, and that for fixed c < a1 , (c, {ak }, x) > 0 for x > max{1, (a2 − a1 )/(a1 − c)}. For
the sake of simplicity, it is worth noting that this latter condition on x reduces to x > 1
for c ≤ a1 − (a2 − a1 ), and x > (a2 − a1 )/(a1 − c) for c ∈ (a1 − (a2 − a1 ), a1 ). In
other words, our threshold value c∗ = a1 , and

a2 −a1
, for c ∈ (a1 − (a2 − a1 ), a1 )
f (c, {ak }) = a1 −c
1 / (a1 − (a2 − a1 ), a1 )
, for c ∈

Moreover, by taking, for example, {ak } = {1, 2, 2, 2, . . .} we see that the left-hand
bound in (8) is actually equality, and hence the algebraic manipulation in (9) involves
ψ(c, {ak }, x) itself. This means that our choice of f (c, {ak }) is optimal.
No other solutions were received.

Circular sums
1183. Proposed by Eugen Ionascu, Columbus State University, Columbus, GA.
Let n be an odd positive integer. Suppose that the integers 1, 2, . . . , 2n are placed
around a circle in arbitrary order.
1. Show that there exist n of these numbers, placed in successive locations around
the circle, having sum S1 satisfying S1 ≥ n2 + n+1
2
,
2. Show that there exist n of these numbers, placed in successive locations around
the circle, having sum S2 satisfying S2 ≤ n2 + n−1
2
, and
3. Show that it is possible to place the 2n numbers around the circle in such a way
that the sum of every n of these numbers, placed in successive locations around
the circle, has sum S3 satisfying n2 + n−1
2
≤ S ≤ n2 + n+12
.

Solution by Andie Rawson (undergraduate), Smith College.

Let x1 , x2 , . . . , x2n be an arbitrary ordering of the integers 1, 2, . . . , 2n around a
circle. Then
2n
xi = 1 + 2 + · · · + 2n = 2n2 + n.
i=1

312 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Then let Si be the sum of n numbers placed in successive locations around the circle
starting from xi . That is,

S1 = x1 + x2 + · · · + xn
S2 = x2 + x3 + · · · + xn+1
.
.
S2n = x2n + x1 + · · · + xn−1

As each xi occurs in n sums,


2n
Si = n(x1 + x2 + · · · + x2n ) = 2n3 + n2
i=1

The mean of the Si s is then S = n2 + n2 .

1. As n is
 odd, Sis not an integer. Thus at least one of the Si satisfies the inequality
Si ≥ n2 + n2 , so there exist n numbers in successive locations with sum S
satisfying S ≥ n2 + n+12
.
2. Again
 2 as  S is not an integer, at least one of the Si satisfies the inequality Si ≤
n + n2 , so there exist n numbers in successive locations with sum S satisfying
S ≤ n2 + n−1
2
.
3. Consider the ordering where
⎧
⎪
⎨i i ∈ 1, 3, . . . , 2n − 1
xi = i + (n + 1) i ∈ 2, 4, . . . , n − 1 .
⎪
⎩i − (n − 1) i ∈ n + 1, n + 3, . . . , 2n

Then

S1 = 1 + 3 + · · · + n + (n + 3) + (n + 5) + · · · + 2n
n + 1 n + 1 n − 1 3n + 3
= +
2 2 2 2
n−1
= n2 + .
2

For i ∈ 1, 2, . . . , n we have that

Si+1 = Si + xn+i − xi = Si + (−1)i+1 ,

and for i ∈ n + 1, n + 2, . . . , 2n − 1 we have that

Si+1 = Si + xi−n − xi = Si + (−1)i+1 .

Therefore for all even i, Seven = S1 + 1 = n2 + n+1

2
and for all odd i, Sodd = S1 =
n + 2 . So every sum S of n successive numbers in this ordering satisfies n2 + n−1
2 n−1
2
≤
S ≤ n2 + n+1
2

VOL. 52, NO. 4, SEPTEMBER 2021 THE COLLEGE MATHEMATICS JOURNAL 313
Also solved by Levent Batakci and Paramjyoti Mohapatra, Case Western Reserve U.; Brian Beasley,
Presbyterian C.; Cal Poly Pomona Problem Solving Group; Michael Goldenberg, Baltimore Polytechnic
Inst. and Mark Kaplan, U. of Maryland Global Campus; Eugene Herman, Grinnell C.; Walther Janous,
Innsburck, Austria; Missouri State U. Problem Solving Group; Mooez Muhammad, (student) Bloor Colle-
giate Inst.; Albert Natian, Los Angeles Valley C.; Northwestern U. Math Problem Solving Group; Joel
Schlosberg, Bayside, NY; Philip Straffin; Texas State U. Problem Solving Group; Janet Lai-Yu Wang
and Nicole Yuen-Yi Pa; and the proposer.

A double integral of a product

1184. Proposed by Seán Stewart, Bomaderry, NSW, Australia.
Evaluate the following integral:
 ∞ ∞
sin x sin(x + y)
dxdy.
0 0 x(x + y)

Solution by the Missouri State University Problem Solving Group.

We will show that, more generally,

 ∞  ∞  ∞ 2
1
f (x)f (x + y) dy dx = f (t) dt .
0 0 2 0

Since it is well known that

 ∞
sin t π
dt = ,
0 t 2

the value of the original integral is π 2 /8.

Letting u = x and v = x + y, reversing the order of integration, and then revers-

ing the roles of u and v, we have
 ∞ ∞
I= f (x)f (x + y) dy dx
0 0
 ∞  ∞
= f (u)f (v) dv du
0 u
 ∞  v
= f (u)f (v) du dv
0 0
 ∞  u
= f (u)f (v) dv du.
0 0

Therefore
 ∞  ∞  ∞  u
2I = f (u)f (v) dv du + f (u)f (v) dv du
0 u 0 0
 ∞  ∞
= f (u)f (v) dv du
0 0

314 © THE MATHEMATICAL ASSOCIATION OF AMERICA

  ∞ 2
= f (t) dt ,
0

and the result follows.

We note that similar techniques show that

 ∞  ∞
··· f (x1 )f (x1 + x2 ) . . . f (x1 + x2 + . . . + xn ) dxn . . . dx1
0 0
 ∞ n
1
= f (t) dt .
n! 0

Also solved by U. Abel and V. Kushnirevych, Technische Hochschule Mittelhesen, Germany; Radouan
Boukharfane, Kaust, Thuwal, KSA; Khristo Boyadzhiev, Ohio Northern U.; Paul Bracken, U. of Texas,
Edinburg; Brian Bradie, Christopher Newport U.; Hongwei Chen, Christopher Newport U.; Bruce Davis,
St. Louis Comm. C. at Florissant Valley; Giuseppe Fera, Vicenza, Italy; Lixing Han, U. of Michigan-Flint;
Eugene Herman, Grinnell C.; Walther Janous, Innsbruck, Austria; John Kampmeyer, (student), Elizabeth-
town C.; Kee-Wai Lau, Hong Kong, China; Moubinool Omarjee, Lycé e Henri IV, Paris, France; Volkhard
Schindler, Berlin, Germany; Albert Stadler, Herrliberg, Switzerland; Justin Turner, (Ph. D student) U. of
Arkansas at Little Rock; Stan Wagon, Macalester C.; and the proposer.

The non-existence of ’special’ rings

1185. Proposed by Greg Oman, University of Colorado, Colorado Springs, Colorado
Springs, CO.
Suppose that S is a commutative ring with identity 1. A subring R of S is called
unital if 1 ∈ R. For the purposes of this problem, call S special if S has the following
properties:
(a) S has a proper unital subring,
(b) there exists a prime ideal of S which is not maximal, and
(c) if R is any proper unital subring of S, then every prime ideal of R is maximal.
Prove the existence of a special ring or show that no such ring exists.
Solution by Anthony Bevelacqua, U. of North Dakota.
Assume such a ring S exists. Then S contains a prime but not maximal ideal P . Since
Z = Z · 1S has no proper unital subrings we have Z  S. Since Z ∩ P is a prime (and
therefore maximal) ideal of Z we must have Z ∩ P = pZ for some prime p. Hence
Zp ∼= Z/pZ, the field of p elements, embeds in S/P .
Suppose a ∈ S and Z[a]  S. Then Z[a] ∩ P is a prime (and hence maximal)
ideal of Z[a]. Thus Z[a]/(Z[a] ∩ P ) is a field, and since Z[a]/(Z[a] ∩ P ) naturally
embeds in S/P , we see that a = a + P is either zero or a unit in S/P . Therefore if
a ∈ S is such that a is a nonzero nonunit in S/P then S = Z[a].
Now S/P is an integral domain but not a field so there exists a ∈ S such that a is a
nonzero nonunit in S/P . Thus we have S = Z[a]. Since a 2 is another nonzero nonunit
we must have S = Z[a 2 ] as well.
Whenever S = Z[w] for some w ∈ S we have S/P = Zp [w]. Thus Zp [a] =
Zp [a 2 ], and so a must be algebraic over Zp . Now S/P = Zp [a] is an integral do-
main algebraic over Zp . Hence S/P is a field, and so P is maximal, a contradiction.
Therefore no such ring S exists.
Also solved by the proposer.

VOL. 52, NO. 4, SEPTEMBER 2021 THE COLLEGE MATHEMATICS JOURNAL 315
 SOLUTIONS

An inequality involving the trace

1176. Proposed by Xiang-Qian Chang, MCPHS University, Boston, MA.
Let An×n be an n × n positive semidefinite Hermitian matrix. Prove that the following
inequality holds for any pair of integers p ≥ 1 and q ≥ 0:

Tr(Ap ) + Tr(Ap+1 ) + · · · + Tr(Ap+q ) rA

≤ ,
Tr(A ) + Tr(A ) + · · · + Tr(A
p+1 p+2 p+q+1 ) Tr(A)

where rA is the rank of A and Tr is the trace function.

Solution by Michel Bataille, Rouen, France.
We assume that A is a non-zero matrix.
The matrix A is similar to a diagonal matrix D = diag(λ1 , λ2 , . . . , λk , 0, . . . , 0) where
λ1 , λ2 , . . . , λk are the positive eigenvalues of A. Since similar matrices have the same
rank and the same trace, we have k = rA and Tr(A) = λ1 + λ2 + · · · + λk . Also, for
any positive integer m, Am is similar to D m , hence Tr(Am ) = λm 1 + λ2 + · · · + λ k .
m m

Without loss of generality, we suppose that λ1 ≤ λ2 ≤ · · · ≤ λk . Then, from Cheby-

chev’s inequality, we have

1 + λ2 + · · · + λk )(λ1 + λ2 + · · · + λk ) ≤ k(λ1 + λm+1 + · · · + λm+1

m+1
(λm m m
2 k )

so that
k
Tr(Am ) ≤ Tr(Am+1 ).
Tr(A)

It is immediately deduced that

Tr(Ap ) + Tr(Ap+1 ) + · · · + Tr(Ap+q )

k
≤ (Tr(Ap+1 ) + Tr(Ap+2 ) + · · · + Tr(Ap+q+1 )),
Tr(A)

and the required result follows (since k = rA ).

Also solved by James Duemmel, Bellingham, WA; Dmitry Fleischman, Santa Monica, CA; Jim Hartman,
The College of Wooster; Justin Haverlick, Keene Valley, NY; Eugene Herman, Grinnell C.; Koopa Koo,
Hong Kong STEAM Academy; Omran Kouba, Damascus, Syria; Elias Lampakis, Kiparissia, Greece; Pi’ilani
Noguchi; Northwestern U. Math Problem Solving Group; Sunghee Park, Seoul, Korea; Michael
Vowe, Therwil, Switzerland; and the proposer.

VOL. 52, NO. 3, MAY 2021 THE COLLEGE MATHEMATICS JOURNAL 229
 Small maximal ideals
1179. Proposed by Greg Oman, University of Colorado, Colorado Springs, Colorado
Springs, CO.
Let R be a ring, and let I be an ideal of R. Say that I is small provided |I | < |R| (i.e.,
I has a smaller cardinality than R). Suppose now that R is an infinite commutative
ring with identity that is not a field. Suppose further that R possesses a small maximal
ideal M0. Prove the following:
1. there exists a maximal ideal M1 of R such that M1  = M0 , and
2. M0 is the unique small maximal ideal of R.

Solution by Anthony Bevelacqua, University of North Dakota, Grand Forks, ND.

We will need the following basic result about cardinality: If A or B is infinite then
|A × B| = max(|A|, |B|).
Since R is not a field there exists a non-zero non-unit a ∈ R. Let Ra = {ra | r ∈ R}
and R[a] = {r ∈ R | ra = 0}. It’s clear that both Ra and R[a] are ideals of R. Since
a is a non-unit we have 1 ∈ / Ra, and since a is not zero we have 1 ∈ / R[a]. Thus
both Ra and R[a] are proper ideals of R. The map R → Ra given by r  → ra is a
ring epimorphism with kernel R[a] so, by the first isomorphism theorem, we have
Ra ∼ = R/R[a]. Hence |R| = |Ra × R[a]| = max(|Ra|, |R[a]|). Thus R possesses a
proper ideal I of cardinality |R|. Let M1 be a maximal ideal of R containing I . Then
|I | ≤ |M1 | ≤ |R| so M1 has cardinality |R|. Since |M0 | < |R| we have M1  = M0 .
Thus we’ve shown 1.
Now assume M0 and N are distinct small maximal ideals of R. Then, since they
are distinct maximal ideals, we have R = M0 + N. Since M0 + N = {x + y | (x, y) ∈
M0 × N } and R is infinite we have M0 or N is infinite. Now

|R| ≤ |M0 × N| = max(|M0 |, |N|) < |R|,

a contradiction. This establishes 2.

Also solved by Paul Budney, Sunderland, MA; Eagle Problem Solvers, Georgia Southern U.; Elias
Lampakis, Kiparissia, Greece; and the proposer.

VOL. 52, NO. 3, MAY 2021 THE COLLEGE MATHEMATICS JOURNAL 231
 Ideals in ideals
1180. Proposed by Luke Harmon, University of Colorado, Colorado Springs, Col-
orado Springs, CO.
In both parts, R denotes a commutative ring with identity. Prove or disprove the
following:
1. there exists a ring R with infinitely many ideals with the property that every
nonzero ideal of R is a subset of but finitely many ideals of R, and
2. there exists a ring R with infinitely many ideals with the property that every
proper ideal of R contains (as a subset) but finitely many ideals of R.

Solution by Bill Dunn, Lone Star College Montgomery, Conroe, TX.

For 1, let R be the ring of integers. Every ideal I of R is principal, I = (n), for some
positive integer n. Suppose I is nonzero, n  = 0. Then I is a subset of any other ideal
J = (m) if and only if m divides n. Because there are only finitely many positive
integer divisors of n, there are only finitely many ideals of R that contain I .

For 2, suppose such a ring R existed. Because R has infinitely many ideals, it must
have infinitely many proper ideals. Also, R must be Artinian because, by hypothesis
on every proper ideal containing but finitely many ideals of R, any decreasing
sequence of ideal must terminate.

However, an Artinian ring has only finitely many maximal ideals. Because every
proper ideal is contained in some maximal ideal, one of these maximal ideals must
contain infinitely many ideals of R, contradicting the hypothesis.

Therefore, such a ring R does not exist.

Also solved by Anthony Bevelacqua, U. of N. Dakota; Paul Budney, Sunderland, MA; Eagle Problem
Solvers, Georgia Southern U.; and the proposer.

232 © THE MATHEMATICAL ASSOCIATION OF AMERICA

 SOLUTIONS

Roots of a cubic equation

1171. Proposed by George Apostolopoulos, Messolonghi, Greece.
Let a, b, and c be the roots of the equation x 3 − 2x 2 − x + 1 = 0, with a < b < c.
Find the value of the expression ( ab )2 + ( bc )2 + ( ac )2 .
Solution by Robert Doucette, McNeese State University, Lake Charles, LA.
Let S = x + y and P = x · y, where
 a 2  2  
b c 2
x= + + ,
b c a

and
 2  
b a 2  c 2
y= + + .
a c b

Since abc = −1,

S =a 4 c2 + b4 a 2 + c4 b2 + b4 c2 + a 4 b2 + c4 a 2
= a 2 + b2 + c2 a 2 b2 + b2 c2 + c2 a 2 − 3

and

P = a 4 c2 + b4 a 2 + c4 b2 b4 c2 + a 4 b2 + c4 a 2
= a 6 c6 + a 6 b6 + b6 c6 + a 6 + b6 + c6 + 3
3 3
= a 2 b2 + b2 c2 + c2 a 2 + a 2 + b2 + c2 − 6S − 9.

We also have a + b + c = 2 and ab + bc + ca = −1, so that

a 2 + b2 + c2 = (a + b + c)2 − 2(ab + bc + ca) = 6,

and

a 2 b2 + b2 c2 + c2 a 2 = (ab + bc + ca)2 − 2abc(a + b + c) = 5.

Therefore S = 6 · 5 − 3 = 27, and P = 53 + 63 − 6 · 27 − 9 = 170. The system x +

y = 27, xy = 170 has two solutions: (x, y) = (10, 17) and (x, y) = (17, 10).
Letting p(x) = x 3 − 2x 2 − x + 1, we find that p(−1)p(−0.8), p(0)p(1), and
p(2)p(3) are all negative. By the intermediate value theorem, y > c4 a 2 > 24 (0.8)2 >
10. Therefore x = 10 is the desired value.
Also solved by Robert Agnew, Palm Coast, FL; Yagub Aliyev, ADA U., Baku, Azerbaijan; Hatef Arshagi,
Guilford Tech. Comm. C.; Farrukh Rakhimjanovich Ataev, WIUT, Uzbekistan; Dione Bailey, Elsie Camp-
bell, and Charles Diminnie (jointly), Angelo St. U.; Michel Bataille, Rouen, France; Rich Bauer, Shore-
line, WA; Anthony Bevelacqua, U. of N. Dakota; Brian Bradie, Christopher Newport U.; James Brenneis;
Scott Brown, Auburn U. Montgomery; Jiakang Chen; John Christopher, California St. U., Sacramento;
Satvik Dasariraju, (student), Lawrenceville S., Princeton, NJ; Gregory Dresden, Washington & Lee U.;
James Duemmel, Bellingham, WA; G. A. Edgar, Ohio St. U.; Michael Goldenberg, Baltimore Polytechnic

144 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Inst. and Mark Kaplan, Towson U. (jointly); G. C. Greubel, Newport News, VA; Lixing Han, U. of Michi-
gan - Flint; Justin Haverlick, Keene Valley, New York; Eugene Herman, Grinnell C.; Timmy Hodges and
Sean Parnell (jointly); Walther Janous, Ursulinengymnasium, Innsbruck, Austria; Benjamin Klein, David-
son C.; Sushanth Satish Kumar, Portola H. S.; Elias Lampakis, Kiparissia, Greece; Kee-Wai Lau, Hong
Kong, China; Math for America Teachers (2 solutions); Missouri State Problem Solving Group; Don-
ald Moore, Wichita, KS; Bob Newcomb, U. of Maryland; Joel Schlosberg, Bayside, NY; Randy Schwartz,
Schoolcraft C.; Ioannis Sfikas, Athens, Greece; Seán Stewart, Bomaderry, NSW, Australia; Georges Vidi-
ani, Les Dijon, France; Michael Vowe, Therwil, Switzerland; Stan Wagon, Macalester C.; and the proposer.
We received two incorrect solutions.

Asymptotic behavior of the solution of a first-order differential equation

1172. Proposed by Xiang-Qian Chang, MCPHS University, Boston, MA.
Suppose that a function y = y(x) satisfies the following first-order differential equa-
tion:

y  + x 6 − x 4 − 2yx 3 − 3x 2 + yx + y 2 − 1 = 0,
1+x 4
with initial value y(0) = π
2
. Show that y(x) ∼ x
as x tends to infinity.
Solution by Kee-Wai Lau, Hong Kong, China.
By the substitution z = y − x 3 + x2 , we transform the differential equation to

x2 3
z = −z2 + + , (1)
4 2
π 1 + x4
with initial value z(0) = . To show that y(x) ∼ , it suffices to show that
2 x
x 1
z(x) ∼ + . (2)
2 x
A particular solution to (1) is z = x2 + x1 . By using formula a ◦ on p. 7 of reference [1]
, we readily obtain the exact solution
  2
x −t 2 /2
x
π
2
+ 0
e dt ex /2
z= + x ,
2 π
2
+ 0 e−t 2 /2 dt xex 2 /2 + 1

and (2) follows.

Reference
[1] Polyanin, A. D., Zaitsev, V. F. (2003). Handbook of Exact Solutions for Ordinary
Differential Equations, 2nd ed. Boca Raton, London, New York: Chapman &
and Hall, CRC Press.

Also solved by U. Abel and V. Kushnirevych, Technische Hochschule Mittelhessen, Germany; Robert Ag-
new, Palm Coast, FL; Michel Bataille, Rouen, France; Paul Bracken, U. of Texas, Edinburg; William
Chang, U. of Southern California; G. C. Greubel, Newport News, VA; Elias Lampakis, Kiparissia, Greece;
Ioannis Sfikas, Athens, Greece; and the proposer.

VOL. 52, NO. 2, MARCH 2021 THE COLLEGE MATHEMATICS JOURNAL 145
 An infinite integral domain has the same cardinality as the set of units
of an integral domain which is integral over it
1173. Proposed by Greg Oman, University of Colorado Colorado Springs, Colorado
Springs, CO.
All rings in this problem are assumed commutative with identity. Now, let R and S
be rings and suppose that R is a subring of S (we assume that the identity of R is the
identity of S). An element s ∈ S is integral over R if s is a root of a monic polynomial
f (x) ∈ R[x]. If we set R := {s ∈ S : s is integral over R}, then it is well-known that
R is a subring of S containing R. The ring R is called the integral closure of R in
S. In case R = S, then we say that S is integral over R. For a ring R, let R × denote
the multiplicative group of units of R. Prove or disprove: for every infinite integral
domain D1 , there exists an integral domain D2 such that D2 is integral over D1 and
|D2× | = |D1 | (that is, the set of units of D2 has the same cardinality as that of D1 ).
Solution by Anthony Bevelacqua, University of North Dakota.
Let F be the quotient field of D1 . Since D1 is infinite we have |D1 − {0}| = |D1 | and so
|D1 × (D1 − {0})| = |D1 |2 = |D1 |. Since D1 × (D1 − {0}) → F given by (a, b) →
a/b is surjective, we have |F | ≤ |D1 × (D1 − {0})|. Thus |F | ≤ |D1 |.
Let  be the algebraic closure of F and let D2 be the integral closure of D1 in .
Then D2 is integral over D1 . Since  is an algebraic extension of F and F is infinite
we have || ≤ |F |. Indeed, for each d ≥ 1 the set of elements of  with minimal
polynomial over F of degree d has cardinality ≤ d|F |d = |F |, and so || ≤ ℵ0 |F | =
|F |. Combining this with the first paragraph we have || ≤ |D1 |.
Now for each a ∈ D1 , x 2 + ax − 1 has a root ua ∈ , and, since x 2 + ax − 1 is
monic, we have ua ∈ D2 . Since a ∈ D1 ⊆ D2 we have ua + a ∈ D2 as well. Thus
ua (ua + a) = 1 so ua ∈ D2× . We note that if ua = ub for a, b ∈ D1 then

u2a + aua − 1 = u2b + bub − 1 ⇒ a = b.

Thus |D1 | = |{ua : a ∈ D1 }| ≤ |D2× |.

Finally we have

|D1 | ≤ |D2× | ≤ || ≤ |D1 |

where the first inequality is given by the previous paragraph, the second follows from
D2× ⊆ , and the last is given by the second paragraph. Therefore D2 is integral over
D1 and |D2× | = |D1 |.
Also solved by Tom Jager, Calvin U.; and the proposer.

Criterion for convergence of an infinite product

1174. Proposed by George Stoica, New Brunswick, Canada.
Let a1 , . . . , ak and b1 , . . . , bk be complex numbers which are not integers. Prove that
 
the infinite product below converges if and only if ki=1 ai = ki=1 bi . What is the
value of the product?


∞
(n − a1 )(n − a2 ) · · · (n − ak )
n=1
(n − b1 )(n − b2 ) · · · (n − bk )

146 © THE MATHEMATICAL ASSOCIATION OF AMERICA

Solution by Eugene Herman, Grinnell College, Grinnell, Iowa.

The gamma function identity (1 + z) = z(z) holds for all complex numbers z that
are not integers. Hence


m
(1 + z) (n + z) = (m + 1 + z).
n=1

Therefore
k k
(1 − ai )  (n − a1 )(n − a2 ) · · · (n − ak )
m
(m + 1 − ai )
i=1
k · = i=1 .
i=1 (1 − bi ) n=1 (n − b1 )(n − b2 ) · · · (n − bk ) k
i=1 (m + 1 − bi )

Furthermore,

(n + z)
lim =1
n→∞ (n)nz

for all complex numbers z that are not integers. Therefore the mth partial product of
the given infinite product converges as m → ∞ if and only if the following expression
converges:
k
i=1 (m + 1)−ai k k
k = (m + 1) i=1 bi − i=1 ai .
i=1 (m + 1)−bi

Therefore, a necessary and sufficient condition for convergence of the product is

k k
i=1 ai = i=1 bi . Also, the limit is

k
i=1 (1 − bi )
k .
i=1 (1 − ai )

Editor’s note: Janous and Lampakis pointed out that this problem and its solution are
known, with both of these solvers providing reference [1] and Lampakis also providing
reference [2].
References
[1] Whittaker, E. T., Watson, G. N. (1927). Modern Analysis: An Introduction to
the General Theory of Infinite Processes and of Analytic Functions, with an Ac-
count of the Principal Transcendental Functions, 4th ed. Cambridge: Cambridge
University Press, p. 238.
[2] Nimbran, A. S. (2016). Interesting infinite products of rational functions moti-
vated by Euler. Math. Stud. 85(1–2): 122, Theorem 3.1.

Also solved by Michel Bataille, Rouen, France; Paul Bracken, U. of Texas, Edinburg; James Duemmel,
Bellingham, WA; Walther Janous, Ursulinengymnasium, Innsbruck, Austria; Elias Lampakis, Kiparissia,
Greece; Michael Vowe, Therwil, Switzerland; and the proposer. There were three solutions that were either
incomplete or incorrect

VOL. 52, NO. 2, MARCH 2021 THE COLLEGE MATHEMATICS JOURNAL 147
 Nonexistence of a sign-preserving field isomorphism between distinct
proper subfields of the reals
1175. Proposed by George Stoica, New Brunswick, Canada.
Let F1 and F2 be distinct proper subfields of the field R of real numbers. Is there a field
isomorphism f : F1 → F2 preserving signs, that is, for all real x: x ∈ F1 and x > 0 if
and only if f (x) ∈ F2 , f (x) > 0?
Solution by Northwestern University Math Problem Solving Group.
First note that every subfield of R contains the field of rational numbers Q. This follows
from the fact that every subfield of R contains 1, and Q is the subfield of R generated
by 1. On the other hand, every isomorphism f between subfields of R restricted to
Q is the identity on Q, i.e., if r ∈ Q, then f (r) = r. This can be proved as follows:
f(0) = 0; f(1) = 1; for integers n, f (n) = f (1 + · · · + 1) = nf (1), f (−n) = −f (n) =
−n; and for integers m and n, with n  = 0, f (m/n) = f (m)/f (n) = m/n.
Next, since F1 and F2 are distinct, f cannot be the identity on F1 , so there is some
u ∈ F1 such that f (u)  = u. Assume u < f (u) (the case u > f (u) is analogous).
Since the rational numbers are dense in the reals, there is some number r ∈ Q such
that u < r < f (u); hence,

u − r < 0 < f (u) − r = f (u) − f (r) = f (u − r).

Letting x = u − r, we have x < 0 and f (x) > 0, implying that f does not preserve
signs.

Also solved by Anthony Bevelacqua, U. of N. Dakota; Paul Budney, Sunderland, MA; William Chang,
U. of Southern California; Dmitry Fleischman, Santa Monica, CA; Eugene Herman, Grinnell C.; Tom Jager,
Calvin C.; Sushanth Sathish Kumar, Portola High S.; Elias Lampakis, Kiparissia, Greece; Missouri State
Problem Solving Group; Lawrence Peterson, U. of N. Dakota; Stephen Scheinberg, Corona del Mar, CA;
and the proposer.

148 © THE MATHEMATICAL ASSOCIATION OF AMERICA

72 MATHEMATICS MAGAZINE

Solutions

The largest divisor of nk − n February 2020

2086. Proposed by David M. Bradley, University of Maine, Orono, ME.
Let f (k) denote the largest integer that is a divisor of nk − n for all integers n. For
example, f (2) = 2 and f (3) = 6. Determine f (k) for all integers k > 1.

Solution by the Northwestern University Math Problem Solving Group, Northwestern

University, Evanston, IL.
To simplify notation, we write gk (n) = nk − n.
First, we prove two lemmas.
VOL. 94, NO. 1, FEBRUARY 2021 73
Lemma 1. For every k > 1, f (k) is square-free, i.e., if p is a prime then p 2 does not
divide f (k).
Proof. Note that gk (p) = p(p k−1 − 1). If p 2 divided gk (p) then p would divide
p k−1 − 1. But this would imply that p divides 1, giving a contradiction. 
Lemma 2. If p is a prime, then p divides f (k) if and only if p − 1 divides k − 1.
Proof. (⇐) If k − 1 = (p − 1) for some  ≥ 1, then
gk (n) = n((n )p−1 − 1).
If p divides n then it divides gk (n) too. If p does not divide n then by Fermat’s little
theorem p divides (n )p−1 − 1. Hence p divides gk (n) for every n, and this implies
that p divides f (k).
(⇒) Assume a prime p divides f (k). This means that p divides gk (n) = n(nk−1 −
1) for every n. Pick n to be a primitive root modulo p (which, by a well-known result
in number theory, always exists). Then 1, n, n2 , . . . , np−2 , are distinct modulo p. Since
p does not divide n, it must divide nk−1 − 1. Using the Euclidean algorithm we write
k − 1 = (p − 1) + i, with  ≥ 0, 0 ≤ i < p − 1. By Fermat’s little theorem np−1 ≡ 1
(mod p), hence
nk−1 = n(p−1)+i ≡ ni (mod p).
Since p divides nk−1 − 1 we have nk−1 ≡ 1 (mod p), hence ni ≡ 1 (mod p). Since
1, n, . . . , np−2 are distinct modulo p, we must have i = 0. Therefore k − 1 = (p −
1), i.e., p − 1 divides k − 1. 
Lemmas 1 and 2 allow us to determine f (k):

f (k) = (d + 1).
d|k−1
d+1 is prime

Example: To compute f (19) we find the divisors of 19 − 1 = 18 : 1, 2, 3, 6, 9, 18,

add 1 to each of them: 2, 3, 4, 7, 10, 19, then multiply the primes appearing on this
list: 2 · 3 · 7 · 19 = 798. Thus f (19) = 798.
Editor’s Note. It is immediate that f (2j ) = 2. The proposer points out that by the
von Staudt–Clausen theorem, f (2j + 1) is the denominator of B2j , the 2j th Bernoulli
number.
Also solved by Elijah Bland & Brooke Mullins, Elton Bojaxhiu (Germany) & Enkel Hys-
nelaj (Australia), Robert Calcaterra, William Chang, John Christopher, Prithwijit De & B. Sury
(India), Joseph DiMuro, Dmitry Fleischman, George Washington University Problems Group,
Justin Haverlick, Omran Kouba (Syria), Sushanth Satish Kumar, Elias Lampakis (Greece), László
Lipták, José Heber Nieto (Venezuela), Joel Schlosberg, Randy K. Schwartz, Doga Can Sert-
bas (Turkey), Jacob Siehler, John H. Smith, Albert Stadler (Switzerland), David Stone & John
Hawkins, Edward White & Roberta White, and the proposer. There was one incomplete or incor-
rect solution.

A limit involving a recursively defined sequence February 2020

2087. Proposed by Florin Stanescu, Şerban Cioiculescu School, Găeşti, Romania.
Consider the sequence defined by x1 = a > 0 and
 
x1 + x2 + · · · + xn−1
xn = ln 1 + for n ≥ 2.
n−1
74 MATHEMATICS MAGAZINE
Compute limn→∞ xn ln n.

Solution by Omran Kouba, Higher Institute for Applied Sciences and Technology,
Damascus, Syria.
The answer is 2. A simple induction argument shows that xn > 0 for all n ≥ 1. Now,
let Sn = x1 + x2 + · · · + xn and define σn = Sn /n. Using the well-known inequality
ln(1 + x) ≤ x which is valid for x > −1 (with equality if and only if x = 0), we
conclude that
 
Sn−1 Sn−1
Sn − Sn−1 = xn = ln 1 + ≤
n−1 n−1
or equivalently σn ≤ σn−1 for n ≥ 2. So, the sequence (σn )n≥1 is positive and decreas-
ing, and since xn = ln(1 + σn−1 ) the sequence (xn )n≥1 is also positive decreasing. Let
 = limn→∞ xn . By Cezáro’s lemma we know that  = limn→∞ σn and the equality
xn = ln(1 + σn−1 ) implies that  = ln(1 + ), and consequently  = 0.
Now, because
lim ln(1 + x)/x = 1
x→0

we conclude that limn→∞ xn /σn−1 = 1 On the other hand

1 σn−1 − ln(1 + σn−1 )
σn = σn−1 − (σn−1 − xn ) = σn−1 − .
n n
But ln(1 + x) = x − (1/2)x 2 + O(x 3 ) for small x, so
 3 
1 2 σn−1
σn = σn−1 − σn−1 + O .
2n n
In particular, σn , σn−1 , xn , and xn+1 are all equivalent as n → ∞. Now
  3 
1 2 σn−1
1 + σn = (1 + σn−1 ) 1 − σn−1 + O .
2n n
So
  3   3 
1 2 σn−1 1 2 σn−1
xn+1 = xn + ln 1 − σn−1 + O = xn − σn−1 + O .
2n n 2n n
Hence
  2
1 1 1 σn−1
n − = + O (σn−1 ) .
xn+1 xn 2 xn xn+1
Thus
 
1 1 1
lim n − = .
n→∞ xn+1 xn 2
Consequently, the Stolz–Cesáro theorem implies that
1 1 1
lim · = ,
n→∞ Hn xn 2
where Hn = nk=1 1/k is the nth harmonic number. Finally, recalling that Hn = ln n +
O(1) we conclude that limn→∞ xn ln n = 2, as claimed.
VOL. 94, NO. 1, FEBRUARY 2021 75
Also solved by Robert A. Agnew, Brian Bradie, Robert Calcaterra, Hongwei Chen, Kee-Wai
Lau (Hong Kong), Albert Stadler (Switzerland), and the proposer.

A Fibonacci sum February 2020

2088. Proposed by Mircea Merca, University of Craiova, Romania.
Let n and t be nonnegative integers. Prove that

2n
Ft
(−1)k Ftk F2tn−tk = − F2tn ,
k=0
Lt

where Fi denotes the ith Fibonacci number and Li denotes the ith Lucas number.

Solution by G. C. Greubel, Newport News, VA.

More generally let
μn − ν n
φn = and θn = μn + ν n ,
μ−ν
where μ + ν = a and μ ν = −b. Note that when a = b = 1, φn = Fn and θn = Ln by
the Binet formulas.
We have
 tk
ν μ tk
(μ − ν) φtk φt (2n−k) = θ2tn − μ
2 2tn
− ν 2tn .
μ ν

Using the sums


2n
(−1)k = 1
k=0


2n  tk  t (2n+1) 
ν μt ν
(−1) k
= 1+
k=0
μ θt μ


2n  
μ tk νt μ t (2n+1)
(−1)k = 1+
k=0
ν θt ν

we find that


2n  
1 2 θt (2n+1)
(−1)k φtk φt (2n−k) = θ2tn −
k=0
(μ − ν)2 θt
1 1   φt
=− 2 θt (2n+1) − θt θ2tn = − φ2tn .
θt (μ − ν) 2 θt
Letting a = b = 1 gives the desired result.
A similar argument shows that


2n
(2 n φt θ2tn − θt φ2tn )
φtk φt (2n−k) =
k=0
(a 2 + 4 b) φt
76 MATHEMATICS MAGAZINE
and hence


2n
(2 n Ft L2tn − Lt F2tn )
Ftk Ft (2n−k) = .
k=0
5 Ft

Also solved by Michel Bataille (France), Brian Bradie, Robert Calcaterra, Dmitry Fleishman,
Harris Kwong, Abhisar Mittal, José Heber Nieto (Venezuela), Angel Plaza (Spain), Albert Stadler
(Switzerland), Michael Vowe (Switzerland), and the proposer.

A product of ratios for nested polygons February 2020

2089. Proposed by Rick Mabry, LSU Shreveport, Shreveport, LA.
Let A1 , A2 , . . . , An be the vertices of a convex n-gon in the plane. Identifying the
indices modulo n, define the following points: Let Bi and Ci be vertices on Ai Ai+1
such that Ai Bi = C i Ai+1 < Ai Ai+1 /2 and let Di be the intersection of Bi−1 Ci and
Bi Ci+1 . Prove that ni=1 (Bi Di )/(Di Ci ) = 1.

Solution by José Heber Nieto, Universidad del Zulia, Maracaibo, Venezuela.

Let βi = ∠Ci Bi Di and γi = ∠Di Ci Bi . Applying the law of sines to triamgles
Bi Ci Di and Bi−1 Ai Ci leads to

Bi Di sin γi Bi−1 Ai sin γi

= and = .
D i Ci sin βi Ai Ci sin βi−1

Also, Ai Bi = Ci Ai+1 implies that Ai Ci = Bi Ai+1 . Using these equations, we obtain

n n n n
Bi Di sin γi i=1 sin γi sin γi
= = n = n i=1
i=1
D i Ci i=1
sin β i i=1 sin β i i=1 sin βi−1
n n n
sin γi Bi−1 Ai Bi−1 Ai
= = = i=1 n
i=1
sin βi−1 i=1
Ai Ci i=1 Ai Ci
n n
Bi−1 Ai Bi−1 Ai
= i=1
n = i=1n = 1.
i=1 Bi Ai+1 i=1 Bi−1 Ai

Also solved by Robert Calcaterra, William Chang, Elton Bojaxhiu (Germany) & Enkel Hys-
nelaj (Australia), George Washington University Problems Group, Joel Schlosberg, and the pro-
poser.
VOL. 94, NO. 1, FEBRUARY 2021 77
Matchings in a certain family of graphs February 2020
2090. Proposed by Gregory Dresden, Washington & Lee University, Lexington, VA.
Recall that a matching of a graph is a set of edges that do not share any vertices. For
example, C4 , the cyclic graph on four vertices (i.e., a square), has seven matchings: the
empty set, single edges (four of these), or pairs of opposite edges (two of these).
Let Gn be the (undirected) graph with vertices xi and yi , 0 ≤ i ≤ n − 1, and edges
{xi , xi+1 }, {xi , yi }, and {yi , xi+1 }, 0 ≤ i ≤ n − 1, where the indices are to be taken
modulo n. For example, G4 is shown below. Determine the number of matchings of
Gn .

Solution by the George Washington University Problems Group, George Washington

University, Washington, DC.
The answer is 3n . To see this, let S = {−1, 0, 1}n , a set whose cardinality is clearly
3n . We show that there is a bijection φ from S to the set of matchings of Gn . Let
a = (a1 , . . . , an ) be an element of S. We define φ(a) as follows:

{xi , xi+1 } ∈ φ(a) if and only if ai = 1 and ai+1 = −1,

{xi , yi } ∈ φ(a) if and only if ai = 1 and ai+1 = −1, and
{xi+1 , yi } ∈ φ(a) if and only if ai = 1 and ai+1 = −1.

We now check that φ(a) is indeed a matching. The edges incident to yi are not both
in φ(a), since {xi , yi } ∈ φ(a) requires ai = 1 but {xi+1 , yi } ∈ φ(a) requires ai = 1.
Also, among the four edges incident to xi , at most one can be chosen for φ(a), since
including {xi , xi−1 }, {xi , yi−1 }, {xi , yi }, and {xi , xi+1 } require, respectively, the four
mutually exclusive conditions (1) ai = −1 and ai−1 = 1, (2) ai = −1 and ai−1 = 1,
(3) ai = 1 and ai 1 = −1, and (4) ai = 1 and ai 1 = −1.
+ +
Given a matching M, there is a unique a ∈ S so that M is φ(a). To see this, let
ai = 1 if M contains {xi , xi+1 } or {xi , yi }, let ai = −1 if M contains {xi−1 , xi } or
{xi , yi−1 }, and let ai = 0 if xi is not the endpoint of any edge in M. This element a ∈ S
is the only element in φ −1 (M). Hence φ is bijective.
Also solved by Elton Bojaxhiu (Germany) and Enkel Hysnelaj (Australia), Robert Calcaterra,
Jiakang Chen, Eddie Cheng; Serge Kruk; Li Li & László Lipták (jointly), José H. Nieto (Venezuela),
Kishore Rajesh, Edward Schmeichel, John H. Smith, and the proposer. There was one incomplete
or incorrect solution.
SOLUTIONS

A Nilpotent Commutator
12339 [2022, 686]. Proposed by Cristian Chiser, Elena Cuza College, Craiova, Romania.
Let A and B be complex n-by-n matrices for which A2 + xB 2 = y (AB  − BA), where x
is a positive real number and y is a real number such that (1/π ) cos−1 (y 2 − x)/(y 2 + x)
is irrational. Prove that (AB − BA)n is the zero matrix.
√ √
Solution by√ Kyle Gatesman, Fairfax, VA. Let U = A + i xB and V = A − i xB. Note
that y ± i x = 0 because y is real and x is positive. Since
√ √
U V = A2 + xB 2 − i x(AB − BA) = (y − i x)(AB − BA)
and
√ √
V U = A2 + xB 2 + i x(AB − BA) = (y + i x)(AB − BA),
we have
√ √
y+i x y 2 − x + 2yi x
VU = √ UV = UV.
y−i x y2 + x
√ √
Let (y + i x)/(y − i x) = cos θ + i sin θ = eiθ . The spectrum of V U is eiθ times that of
U V . By hypothesis, θ is not a rational multiple of π , so einθ = 1 for all nonzero integers n.
It is well known for complex n-by-n matrices U and V , that U V and V U have the same
characteristic polynomial. Hence any eigenvalue of U V or V U is an eigenvalue of the

446 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131

other. Thus the spectrum of U V is invariant under multiplication by eiθ . Since the complex
numbers eiθ , e2iθ , e3iθ , . . . are distinct and the spectrum of U V has cardinality at most
n, we conclude that the only eigenvalue of U V is zero. It follows that the characteristic
polynomial of AB − BA is λn . By the Cayley–Hamilton Theorem, (AB − BA)n is the
zero matrix.
Also solved by C. P. Anil Kumar (India), S. Bhadra, E. A. Herman, O. P. Lossers (Netherlands), M. Omarjee
(France), R. Stong, L. Zhou, and the proposer.

A Nascent Delta Function

12340 [2022, 686]. Proposed by Antonio Garcia, Strasbourg, France. Let g : [0, 1] → R
be continuous. Prove that
1
n g(x)
lim dx = Cg(1/2)
n→∞ 2n 0 x n + (1 − x)n
for some constant C (independent of g), and determine the value of C.
Solution by Missouri State University Problem Solving Group, Missouri State University,
Springfield, MO. Substituting u = n(2x − 1) and letting χ[−n,n] denote the characteristic
function of [−n, n] gives

n 1 g(x) 1 ∞ g 12 + 2n u
χ[−n,n] (u)
dx =   du.
2 0 x + (1 − x)
n n n 2 −∞ 1 + n + 1 − un n
u n

Since g is continuous, we may choose a K > 0 such that |g(x)| ≤ K on [0, 1]. Further, for
n ≥ 2, the binomial theorem gives
   2  
u n u n n u u2
1+ + 1− ≥2 1+ ≥2 1+ .
n n 2 n2 4
Therefore for n ≥ 2,
  
1  g 12 + u
χ[−n,n] (u)  K
 2n
 ≤ .
2  1 + un n
+ 1− n u n 4 + u2

This upper bound has finite integral, so the dominated convergence theorem applies, and
we get

n 1 g(x) 1 ∞ g 12 + 2nu
lim n dx = lim   du
n→∞ 2 0 x + (1 − x)
n n 2 −∞ n→∞ 1 + un + 1 − un n
n

∞
1 g(1/2)
= −u
du
−∞ e + e
2 u

∞
1  π
= g(1/2)arctan(eu ) = g(1/2).
2 −∞ 4

Also solved by M. Aassuka (France), A. Berkane (Algeria), S. Bhadra (India), H. Chen (US), W. J. Cowieson,
M.-C. Fan (China), K. Gatesman, R. Guadalupe (Philippines), E. A. Herman, N. Hodges (UK), F. Holland
(Ireland), E. J. Ionascu, S. Kaczkowski, O. Kouba (Syria), C. Krattenthaler (Germany), G. Lavau (France),
J. H. Lindsey II, P. W. Lindstrom, O. P. Lossers (Netherlands), F. Masroor, R. Mortini (Luxembourg) &
R. Rupp (Germany), M. Omarjee (France), D. Pascuas (Spain), P. Perfetti (Italy), K. Schilling, A. Stadler
(Switzerland), A. Stenger, R. Stong, R. Tauraso (Italy), E. I. Verriest, J. Vukmirović (Serbia), J. H. Yan (China),
and the proposer.

May 2024] PROBLEMS AND SOLUTIONS 447

 A Product Inequality
12341 [2022, 686]. Proposed by George Apostolopoulos, Messolonghi, Greece. Let
x1 , . . . , xn be positive real numbers with ni=1 xi2 ≤ n, and let S = ni=1 xi . Prove
n  xi2
1 2
1+ ≥ 2S /n ,
i=1
xi xi+1
where xn+1 is taken to be x1 .
Solution by Roberto Tauraso, Tor Vergata University of Rome, Rome, Italy. We prove the
more general inequality
n  2
1 xi n S 2 /n
1+ ≥ 1+ , (∗)
i=1
yi T

where x1 , . . . , xn and y1 , . . . , yn are positive real numbers, S = ni=1 xi , and T = ni=1 yi .

The required inequality follows from (∗) by letting yi = xi xi+1 and noting that, by the
rearrangement inequality,

n 
n 
n
T = yi = xi xi+1 ≤ xi2 ≤ n.
i=1 i=1 i=1
To prove (∗), we compute
 n   2   
 1 xi
n
1
log 1+ = xi2 log 1 +
i=1
yi i=1
yi


n 1
dt 1 
n
xi2
= xi2 = dt.
i=1 0 yi + t 0 i=1
yi + t
For 0 ≤ t ≤ 1, the Cauchy–Schwarz inequality implies
 n 2
√ x i
n n
xi2 n
xi2
S2 = yi + t · √ ≤ (yi + t) · = (T + nt) ,
i=1
yi + t i=1
y +t
i=1 i
y +t
i=1 i
so

n
xi2 S2
≥ .
i=1
yi + t T + nt
Therefore
 n   2
 1 xi 1 
n
xi2 1
S2 S2 n
log 1+ = dt ≥ dt = log 1 + .
i=1
yi 0 i=1
yi + t 0 T + nt n T
Inequality (∗) follows.
Also solved by P. Bracken, W. J. Cowieson, O. P. Lossers (Netherlands), S. Patra, A. Stadler (Switzerland),
R. Stong, and the proposer.

Characterizing Cyclic Quadrilaterals

12343 [2022, 785]. Proposed by Tran Quang Hung, Hanoi, Vietnam. Let ABCD be a
convex quadrilateral with AB = a, BC = b, CD = c, DA = d, AC = e, and BD = f .
Prove that ABCD is a cyclic quadrilateral (i.e., the four vertices lie on a circle) if and
only if  2 
f 2 − e2 a − c2 b2 − d 2
= .
ac + bd (ab + cd)(ad + bc)

448 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131

Solution by Prithwijit De, Mumbai, India. Denote the angles of ABCD at the four vertices
by A, B, C, and D. Let
d 2 + a2 − f 2 b2 + c2 − f 2
T1 = cos A + cos C = + ,
2ad 2bc
a 2 + b2 − e2 c2 + d 2 − e2
T2 = cos B + cos D = + .
2ab 2cd
Algebraic manipulation yields

2abcd (ab + cd)T1 − (ad + bc)T2 =
(ac + bd)(a 2 − c2 )(b2 − d 2 ) − (ab + cd)(ad + bc)(f 2 − e2 ).
It therefore suffices to show that ABCD is cyclic if and only if
(ab + cd)T1 − (ad + bc)T2 = 0.
By the sum-to-product formula for the cosine function and the fact that B + D = 2π −
(A + C), we have

(ab + cd)T1 − (ad + bc)T2 =

      
A−C B −D A+C
2 (ab + cd) cos + (ad + bc) cos cos .
2 2 2
Since |A − C| and |B − D| are less than π , cos((A − C)/2) and cos((B − D)/2) are
strictly positive. Hence (ab + cd)T1 − (ad + bc)T2 = 0 if and only if cos((A + C)/2) = 0,
which happens if and only if A + C = π , which is equivalent to ABCD being cyclic.
Also solved by G. Fera (Italy), O. Geupel (Germany), M. Goldenberg & M. Kaplan, N. Hodges (UK),
O. P. Lossers (Netherlands), C. R. Pranesachar (India), C. Schacht, A. Stadler (Switzerland), R. Stong,
R. Tauraso (Italy), L. Zhou, Fejéntaláltuka Szeged Problem Solving Group (Hungary), and the proposer.

Linear Combinations of Powers That Are Not Perfect Squares

12346 [2022, 785]. Proposed by Nguyen Quang Minh, Hwa Chong Institution, Bukit
Timah, Singapore. Prove that there are infinitely many integers A such that, for every
nonzero integer x and distinct positive odd integers m and n, the integer x m + Ax n is
not a perfect square.
Solution by Yury J. Ionin, Central Michigan University, Mount Pleasant, MI. We claim that
the infinite family consisting of the negatives of primes congruent to 3 modulo 8 satisfies
the requirements of the problem.
Let A = −p for such a prime p. Factoring out the perfect square x min{m,n}−1 , we see
that it suffices to show that no x m − px n is a perfect square when m and n are odd and
either m = 1 or n = 1. Suppose otherwise.
First consider m = 1 and set k = (n − 1)/2. With x − px n = x(1 − px 2k ), both factors
are negative. Since also 1 − px 2k is relatively prime to x, both −x and px 2k − 1 must
be squares. Modulo p, the equation px 2k − 1 = a 2 for a positive integer a reduces to
a 2 ≡ −1. However, when p ≡ 3 (mod 8) (indeed, whenever p ≡ 3 (mod 4)) the value
−1 is not a square modulo p, a contradiction.
Now consider n = 1 and set k = (m − 1)/2, so x m − px = x(x 2k − p). The greatest
common divisor of x and x 2k − p is 1 or p. Since x m − px is a square, we have either
(i) x = ±a 2 and x 2k − p = ±b2 or (ii) x = ±pa 2 and x 2k − p = ±pb2 , for some integers
a and b.

May 2024] PROBLEMS AND SOLUTIONS 449

 Note that squares are congruent to 0, 1, or 4 modulo 8, and recall that p ≡ 3 mod 8. In
case (i), if a is odd, then x 2k − p ≡ 6 (mod 8). If a is even, then x 2k − p ≡ 5 (mod 8).
In both subcases, this value cannot be a square or its negative, so we move on to case (ii).
Substituting for x and simplifying, we have p2k−1 a 4k − 1 = ±b2 . The left side is positive.
However, again because −1 cannot be a square modulo p, the alternative p2k−1 a 4k − 1 =
b2 is also impossible.
Editorial comment. All solvers had roughly similar approaches. We generalize some of
their families. Using the fact that −2 is a quadratic nonresidue for primes p congruent to 5
or 7 modulo 8, one can show that the family A = pr satisfies the condition of the problem
for such primes p and even r. Another family is given by A = pr , where p is a prime
congruent to 7 modulo 16 and r is odd. This can be proved by the method of descent.
Also solved by J. Boswell & C. Curtis, W. J. Cowieson, K. Gatesman, P. W. Lindstrom, R. Stong, R. Tauraso
(Italy), H. von Eitzen (Germany), and the proposer.

A Functional Equation With Piecewise Linear Solutions

12347 [2022, 786]. Proposed by Marian Tetiva, Gheorghe Roşca Codreanu National Col-
lege, Bı̂rlad, Romania. Let a and b be real numbers with 0 < a < 1 < b. Find all continu-
ous functions f : R → R such that f (0) = 0 and f (f (x)) − (a + b)f (x) + abx = 0 for
all x ∈ R.
Solution by Omran Kouba, Higher Institute for Applied Sciences and Technology, Damas-
cus, Syria. We show that there are exactly four solutions, given by
 
ax, if x ≥ 0, bx, if x ≥ 0,
f (x) = ax, f (x) = bx, f (x) = and f (x) =
bx, if x < 0, ax, if x < 0.
Clearly these four functions are solutions. Now let f : R → R be continuous and satisfy
f (0) = 0 and f (f (x)) − (a + b)f (x) + abx = 0 for all x ∈ R. For all x ∈ R,
(a + b)f (x) − f (f (x))
x= .
ab
This implies that x = y if f (x) = f (y), so f is one-to-one. Since f is continuous, it
follows that f is monotonic, and consequently f ◦ f is increasing. Moreover, the equality
f (f (x)) + abx
f (x) = (1)
a+b
shows that f is increasing. Since f (0) = 0, the sign of f (x) is the same as the sign of x.
By (1), we have f (x) > abx/(a + b) for all x > 0 and f (x) < abx/(a + b) for all x < 0.
This implies that limx→∞ f (x) = +∞ and limx→−∞ f (x) = −∞. Hence f is bijective.
Let g = f −1 . Applying the functional equation to g(g(x)) leads to
 
1 1 1
g(g(x)) − + g(x) + x = 0.
a b ab
Thus g satisfies the same functional equation as f , but with a and b replaced by 1/a and
1/b.
Suppose x > 0. We define two sequences {xn }n≥0 and {yn }n≥0 by x0 = x, y0 = f (x),
and xn+1 = f (xn ) and yn+1 = g(yn ) when n ≥ 0. By the functional equations of f and g,
{xn }n≥0 and {yn }n≥0 satisfy the following second-order linear recurrence relations:
x0 = x, x1 = f (x), xn+2 − (a + b)xn+1 + abxn = 0,
 
1 1 1
y0 = f (x), y1 = x, yn+2 − + yn+1 + yn = 0.
a b ab

450 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131

Solving these recurrence relations, we find that for all n ≥ 0,
f (x) − bx n f (x) − ax n
xn = a + b , (2)
a−b b−a
f (x) − bx 1−n f (x) − ax 1−n
yn = a + b . (3)
a−b b−a
We now consider two cases. If f (x) ≤ x, then because f is increasing, we have xn ≥
xn+1 > 0 for all n. Thus the sequence (xn )n≥0 is nonincreasing and bounded below, so it
must be convergent. Since b > 1, the coefficient of bn in (2) must be zero, which implies
that f (x) = ax.
On the other hand, if f (x) > x, then similar reasoning shows that the sequence (yn )n≥0
converges, the coefficient of a 1−n in (3) is zero, and f (x) = bx.
Thus for all x > 0, either f (x) = ax or f (x) = bx, so f (x)/x can take only the two
values a and b on (0, ∞). However, since f is continuous, it cannot take both values. We
conclude that either f (x) = ax for all x > 0 or f (x) = bx for all x > 0.
Applying the above analysis for x > 0 to the function −f (−x), we conclude that either
f (x) = ax for all x < 0 or f (x) = bx for all x < 0. Thus there are no solutions other than
the four listed earlier.
Also solved by J. Boswell & C. Curtis, H. Chen (China), W. J. Cowieson, H. von Eitzen (Germany), D. Hen-
derson, N. Hodges (UK), O. P. Lossers (Netherlands), R. Mortini (Luxembourg), K. Schilling, R. Stong,
R. Tauraso (Italy), and the proposer.

A Variation on the Josephus Problem

12348 [2022, 786]. Proposed by Erik Vigren, Uppsala, Sweden, and Hans Rullgård,
Kungälv, Sweden. We have n people in a circle, numbered from 1 to n clockwise. They
are removed one at a time as follows, until just one remains. At each step, remove the
nth person among those remaining, where the count starts at the lowest-numbered person
remaining and proceeds clockwise. Let W (n) be the number of the last person remaining.
For example, with n = 5, we remove in order the people numbered 5, 1, 3, and 2, and so
W (5) = 4. (This is a variation of the classic Josephus problem.)
(a) What is W (1012 )?
(b) For n ≥ 5, show that W (n) = n − 4 if and only if n/2 is a Sophie Germain prime (i.e.,
n/2 and n + 1 are prime).
(c) Find the smallest even number that does not equal W (n) for any n.
Composite solution by Roberto Tauraso, Tor Vergata University of Rome, Rome, Italy, and
the proposers. 
(a) By reversing the procedure, we show W 1012 = 671,046,354,072. As in the problem
statement, the number of a person is that person’s original index and remains unchanged.
The position of a person at a given time is that person’s index among the remaining people;
it counts the remaining people with smaller numbers (plus 1).
Consider the point in the process when m people remain. In the next step, skipping
n − 1 people means passing through the entire list r times before stopping at the person
to be removed, where r = (n − 1)/m. The person removed will be in position n − rm.
We say that removals whose associated value of r are the same occur in the same round,
and we label this round with the value r. For example, in round 0 we remove person n, and
in round 1 we remove all the remaining odd-numbered people, starting with person 1. The
rounds occur in increasing order, but the round numbers are not consecutive. For example,
when n = 9 there is no round 3, because 8/3 = 2 and 8/2 = 4. Rather than reversing
the procedure one removal at a time, the computation is quicker if we reverse it one round
at a time. This will also be useful in part (c).

May 2024] PROBLEMS AND SOLUTIONS 451

 Now consider the time when a round has just been finished and k rounds remain to be
completed. Let mk denote the number of people remaining at this time, and let pk denote
the position at this time of the person P who will be the last person remaining. Thus m0 = 1
and p0 = 1, since P is never removed. For k ≥ 1, let rk denote the number of the round
about to start. By definition, rk = (n − 1)/mk .
The last removal in round rk+1 occurs with mk + 1 people remaining, so
rk+1 = (n − 1)/(mk + 1). (1)
When rk+1 > 0, the number of people remaining at the start of round rk+1 is the largest m
such that rk+1 = (n − 1)/m; that is,
mk+1 = (n − 1)/rk+1 . (2)
During round rk+1 , when m people remain, the person in position n − rk+1 m will be
removed. This position strictly increases throughout round rk+1 as m decreases from mk+1
to mk + 1. Meanwhile, the position of P decreases from pk+1 to pk . Since P reaches pk ,
the position of P must decrease on the step that starts with m people remaining if and
only if
n − rk+1 m ≤ pk . (3)
By (2), we have (n − 1)/rk+1 < mk+1 + 1, which yields n − rk+1 (mk+1 + 1) < 1.
Also, the definition of rk implies (n − 1)/mk ≥ rk ≥ rk+1 + 1, from which we obtain
n − rk+1 mk ≥ mk + 1. Together, these inequalities yield
n − rk+1 (mk+1 + 1) < 1 ≤ pk < mk + 1 ≤ n − rk+1 mk .
It follows that there is some integer j with 0 ≤ j ≤ mk+1 − mk such that
n − rk+1 (mk+1 − (j − 1)) ≤ pk < n − rk+1 (mk+1 − j ).
By (3), there will then be exactly j steps during round rk+1 on which the position of P
decreases by 1. Therefore
 
pk + rk+1 (mk+1 + 1) − n
pk+1 = pk + j = pk + . (4)
rk+1
We now have a recursive procedure, starting from m0 = p0 = 1. Given mk and pk , we
use mk to compute rk+1 by (1), rk+1 to compute mk+1 by (2), and then all of {pk , rk+1 , mk+1 }
to compute pk+1 by (4). We run the recursion until reaching k such that mk equals n − 1.
The original position (and number) of P is then pk . In the particular instance n = 1012 , we
obtain k = 1999997, leading to W (n) as claimed.
(b) Assume n ≥ 5. Because all people with odd numbers will have been removed by the
end of round 1, W (n) is an even number less than n. In particular, n − 4 is removed by
then if n is odd, so we need only consider even n. When n is even, the person with the
larger number will be removed when only two people remain. Therefore W (n) = n − 4 if
and only if the last two people are numbered n − 4 and n − 2.
Suppose that m people remain, where m ≤ n/2 − 1. Recall that n is removed first and
then all odd numbers. If both n − 4 and n − 2 remain, then they occupy positions m − 1
and m. To avoid removing either, n must not be congruent to m − 1 or m modulo m. That
is, we avoid removing person n − 2 if and only if n is not divisible by any number from
3 to n/2 − 1, meaning that n/2 is prime. Similarly, we avoid removing person n − 4 if
and only if n − 1 is not divisible by any number from 3 to n/2 − 1, meaning that n + 1 is
prime.
(c) We show that the smallest even number that does not equal W (n) for any n is 34. The
table below gives the smallest value of n yielding each value of W (n) less than 34, by
explicit computation.

452 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131

 W (n) 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
n 3 5 7 16 11 13 50 17 19 76 23 56 248 29 31 424

We need only consider n > 34 and show that in all cases person 34 is removed at some
point in the process. We have observed that person n is removed in round 0, and all smaller
odd numbers are removed in round 1. Person 34 is then in position 17.
Since round r is defined as {m : (n − 1)/m = r}, the number of people remaining
when round r ends is min{m : (n − 1)/m = r} − 1. This number is (n − 1)/(r + 1).
Let ar+1 be the integer such that
(n − 1)/(r + 1) = (n − ar+1 )/(r + 1).
The first person removed in round r + 1 is in position ar+1 at the start of the round. For
each subsequent removal in round r + 1, the removed element pushes the round-starting
position of the next person removed up by r + 2. That is, the key additional observation is
that positions at the start of round r + 1 of the people removed in round r + 1 are
ar+1 , ar+1 + r + 2, ar+1 + 2r + 4, . . . .
For even n, those removed in round 2 start the round in positions 2, 5, 8, 11, 14, 17, . . . .
Hence we may assume n is odd.
For odd n, those removed in round 2 start the round in positions 1, 4, 7, 10, 13, 16, . . . .
Thus after round 2, person 34 is in position 11.
When n ≡ 3 (mod 6), those removed in round 3 start the round in positions 3, 7, 11,
15, . . . , so we may forbid this case.
When n ∈ {1, 5, 7, 11} (mod 12), getting (n − a3 )/3 to be an integer requires a3 ∈
{1, 2}. Those removed in round 3 start the round in positions 1, 5, 9, 13, . . . , or positions
2, 6, 10, 14, . . . . In both cases, person 34 ends round 3 in position 8.
When n ∈ {7, 11} (mod 12), we have a4 = 3, and those starting round 4 in positions 3,
8, . . . are removed. Hence we may forbid this case.
When n ∈ {1, 5} (mod 12), we have a4 = 1, and those starting round 4 in positions
1, 6, . . . are removed. Hence person 34 occupies position 6 at the end of round 4. Since
a5 ∈ {1, 2, 3, 4, 5}, round 5 removes exactly one person from the first five positions, so
person 34 ends round 5 in position 5.
When n ≡ 5 (mod 12), we have a6 = 5, so round 6 removes person 34.
Hence we may assume n ≡ 1 (mod 12). If also n ≥ 73, then at least 12 people remain
at the end of round 5. When the number of people remaining is in {12, 6, 4, 3, 2}, the person
occupying the first position at that time will be removed. This means that person 34, who
is already as early as position 5 when at least 12 people remain, is removed while a person
still remains.
To complete the proof, it remains only to check explicitly that W (n) = 34 when n ∈
{37, 49, 61}.
Editorial comment. Reasoning like that for part (b) shows that W (n) = n − 1 if and only
if n is an odd prime. Round r actually eliminates one or more√people if (n − 1)/(r + 1) <
(n − 1)/r. This holds for all r with r ≤ r ∗ , where r ∗ = ( 4n − 3 − 1)/2. Thereafter,
at most one person is removed per round. As a result, the number of rounds in which people
are removed is r ∗ + (n − 1)/(r ∗ + 1).
Also solved by O. P. Lossers (Netherlands). Parts (b) and (c) also solved by K. Schilling and Eagle Problem
Solvers.

A Lobachevsky-type Formula
12351 [2022, 886]. Proposed by Seán Stewart, King Abdullah University of Science and

May 2024] PROBLEMS AND SOLUTIONS 453

Technology, Thuwal, Saudi Arabia. Evaluate

∞ ln cos2 x sin3 x
 dx.
0 x 3 1 + 2 cos2 x

Solution by Mohammed Aassila, Strasbourg, France. Let I denote the requested integral.
We prove that
 √ 
π ln(1 + 3)
I =− ln 2 + √ .
4 3
We have
 ∞ 
1 ∞
ln cos2 x sin3 x 1  (k+1)π ln cos2 x sin3 x
I=  dx =  dx
−∞ x 1 + 2 cos x x 3 1 + 2 cos2 x
2 3 2 2 k=−∞ kπ
∞ 
1  π (−1)k ln cos2 x sin3 x
=  dx
2 k=−∞ 0 (x + kπ )3 1 + 2 cos2 x
 ∞  
1 π  (−1)k ln cos2 x sin3 x
= dx,
2 0 k=−∞
(x + kπ )3 1 + 2 cos2 x

where the final interchange of integration and summation can be justified by the dominated
convergence theorem.
To evaluate the summation in the last formula, we start with the equation
∞
(−1)k 1
= .
k=−∞
x + kπ sin x

(See I. S. Gradshteyn, I. M. Ryzhik (2007), Table of Integrals, Series, and Products, 7th
ed., Burlington, MA: Academic Press, equation 1.422.6.) Differentiating twice, we get
∞
 (−1)k 1 + cos2 x
= ,
k=−∞
(x + kπ ) 3
2 sin3 x
so this gives

1 π (1 + cos2 x) ln cos2 x π/2
(1 + cos2 x) ln (cos x)
I= dx = dx
4 0 1 + 2 cos2 x 0 1 + 2 cos2 x
π/2 π/2
1 1 ln (cos x)
= ln(cos x) dx + dx.
2 0 2 0 1 + 2 cos2 x
Both of these integrals are special cases of equation 4.385.3 in Gradshteyn and Ryzhik:
π/2  
ln(cos x) π b
dx = ln
0 b2 sin2 x + a 2 cos2 x 2ab a+b
√
for a, b > 0. Applying this with b = 1 and both a = 1 and a = 3 leads to the claimed
answer.
Editorial comment. As several solvers noted, the beginning of this argument proves a
Lobachevsky-type result: For any continuous function f (x) that is periodic with period
π,
∞
sin3 x 1 π

3
f (x) dx = (1 + cos2 x)f (x) dx.
−∞ x 2 0

454 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131

Also solved by T. Amdeberhan, A. Berkane (Algeria), P. Bracken, B. Bradie, C. Burnette, H. Chen (US),
B. E. Davis, M. L. Glasser, G. C. Greubel, N. Hodges (UK), W. Janous (Austria), L. Kempeneers & J. V. Cast-
eren (Belgium), O. Kouba (Syria), K. Nelson, M. Omarjee (France), A. Stadler (Switzerland), A. Stenger,
R. Stong, R. Tauraso (Italy), Y. Zhang (China), and the proposer.

CLASSICS
C25. Let w0 , w1 , . . . be the sequence of Fibonacci words, defined by w0 = 0, w1 = 1, and,
for n ≥ 2, wn = wn−2 wn−1 , the concatenation of wn−2 and wn−1 . Thus the sequence begins
0, 1, 01, 101, 01101, 10101101, 0110110101101, . . . . Show that, for n ≥ 3, removing the
first two symbols from wn yields a palindrome.

The Tennis Ladder

C24. Due to Colin L. Mallows. Over the history of a certain tennis club, every player has
played at least one match against every other player. Matches are played one at a time, and
after each match a ranking of the players in the club is computed as follows. Starting with
the most recent match and working backwards through time, use the match results to build
up a partial order. Ignore any match that is inconsistent with more recent results. The final
result is guaranteed to be a linear order, since any incomparability between a pair of players
is resolved when a match between them is encountered. This linear order becomes the new
club ranking. Prove or disprove: A player cannot rise in the club ranking by intentionally
losing a match.
Solution. The assertion is false. Suppose that the results of the last nine matches among six
players are as follows, where we write a > b for a match where player a defeats player b
and we list the matches from oldest to most recent.
2 > 3, 6 > 1, 2 > 4, 1 > 2, 6 > 4, 4 > 5, 3 > 4, 3 > 6, 5 > 6
The ranking at this moment is 1 > 2 > 3 > 4 > 5 > 6, with player 3 in third place. How-
ever, if player 3 loses the next match to player 5, the ranking becomes 5 > 3 > 6 > 1 >
2 > 4, with player 3 in second place. So player 3 ranks higher after losing.
Editorial comment. The problem appeared as E3240 [1987, 996; 1989, 530] in this
Monthly. The problem statement has two interpretations. The strong form asks if a player
can rank higher immediately after throwing a match. The weak form asks if a player can
rank higher today by deciding to forfeit a match that took place in the past. No solution to
the strong form of the problem was received from the Monthly readership other than the
proposer’s solution, which involved seven players. The example here involves six players.
This raises the question of whether there is an example with five players.
One can show that any time a player defeats a lower-ranked opponent (or loses to a
higher-ranked opponent), the ranking remains unchanged. However, reversing the outcome
of each match in the example above shows that defeating a higher-ranked opponent can
lower one’s overall ranking.
Say that a ranking algorithm respects duality if changing all wins to losses reverses the
resulting ranking. A familiar algorithm for ranking tennis club members is as follows: If
a lower-ranked player A defeats a higher-ranked player B, the new ranking is formed by
replacing B with A in the prior ranking and moving B and all the players ranked between A
and B down one spot. If a higher-ranked player defeats a lower-ranked player, the ranking
remains unchanged. One concern with this usual algorithm is that it fails to respect duality.
The algorithm of this problem is an alternative that does respect duality. The existence
of the example above, however, shows that this ranking system violates a certain kind of
monotonicity and suggests that it is an unreasonable system for actual use.

May 2024] PROBLEMS AND SOLUTIONS 455

 −x −α

SOLUTIONS
A Sufficient Condition for Generalized Commuting
12331 [2022, 588]. Proposed by WeChat Group on Matrix Analysis, Nova Southeastern
University, Fort Lauderdale, FL. Let A and B be complex m-by-n matrices, and let C
be a complex n-by-m matrix. Prove that if there are nonzero scalars x and y such that
ACB = xA + yB, then ACB = BCA.
Solution by Li Zhou, Polk State College, Winter Haven, FL. Since CACB = xCA + yCB,
we have (CA − yIn )(CB − xIn ) = xyIn . Thus CB − xIn has an inverse P that satisfies
CA − yIn = xyP . Hence
B(CA − yIn ) = xyBP = x(ACB − xA)P = xA(CB − xIn )P = xA,
so BCA = xA + yB = ACB.
Also solved by C. P. Anil Kumar (India), M. Bataille (France), M. R. Elgersma, K. Gatesman, E. A. Herman,
O. Kouba (Syria), O. P. Lossers (Netherlands), M. Omarjee (France), P. Oman & H. Wang, M. Reid, K. Sarma
(India), K. Schilling, R. Stong, R. Tauraso (Italy), Southeast Missouri State University Math Club, UM6P Math
Club (Morocco), and the proposer.

A Symmetric Decomposition into Icosahedra

12333 [2022, 588]. Proposed by Moshe Rosenfeld, University of Washington, Seattle, WA,
and Tacoma Institute of Technology, Tacoma, WA. Let G be the multigraph obtained by
replacing each edge of the complete graph K12 by five edges. Show that the 330 edges
of G can be partitioned into 11 sets such that each set forms a graph isomorphic to the
icosahedron.
Solution by Eagle Problem Solvers, Georgia Southern University, Statesboro and Savan-
nah, GA. We assign the label z to one vertex of K12 , and label the other 11 vertices with
the elements of Z11 . Notice that each edge of K12 can be written uniquely as either (z, v),
for some v ∈ Z11 , or (v, v + j ), for some v ∈ Z11 and j ∈ {1, 2, 3, 4, 5}. In the latter case,
we refer to j as the difference of the edge.

354 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131

 The icosahedron H is a 5-regular graph with 12 vertices and 30 edges. For each i ∈ Z11 ,
we define a copy Hi of H in K12 and then show that these copies together use each edge of
K12 five times. These copies of H can then be used to define the required partition of the
edges of G.
We draw Hi with z at the top and vertex i as the vertex at distance 3 from z. The five
vertices adjacent to z in Hi are i + 1, i + 2, i + 3, i + 4, and i + 5, and the five vertices
adjacent to i are i − 1, i − 2, i − 3, i − 4, and i − 5. There is a 10-cycle alternating between
the vertices adjacent to z and those adjacent to i. (The “wraparound” edge from i − 2 to
i + 1 has been left broken to emphasize the symmetry.) In the drawing of Hi below, we
have labeled each edge not incident to z with its difference. For example, the edge from
i + 1 to i − 5 has difference 5 because (i + 1) + 5 = i + 6 = i − 5 in Z11 .

It remains to show that each edge of K12 occurs in Hi for exactly five values of i in Z11 .
An edge of the form (z, v) for some v ∈ Z11 occurs in Hi if and only if
v ∈ {i + 1, i + 2, i + 3, i + 4, i + 5},
in other words, if and only if
i ∈ {v − 1, v − 2, v − 3, v − 4, v − 5}.
Now consider an edge (v, v + j ) with difference j . If j = 1, then the edge occurs in
Hi if and only if v ∈ {i + 1, i + 2, i + 3, i + 4, i − 1}. These values correspond to the five
edges in the figure that are labeled 1, and they determine five values of i for which the edge
occurs in Hi . Similarly, there are five edges in the figure labeled with each of the other
differences 2, 3, 4, and 5, and therefore for each of these values of j , an edge of the form
(v, v + j ) occurs in Hi for five values of i.
Editorial comment. Several solvers described a decomposition using a rotation modulo 11.
Rob Pratt obtained a decomposition via integer linear programming.
The proposer used the fact that the graph H whose edges are the pairs of vertices
separated by distance 2 in a copy H of the icosahedron is isomorphic to the icosahedron.
Since each vertex has exactly one antipodal vertex at distance 3, the union of H and H
omits exactly a perfect matching in K12 . Said another way, the complement of a perfect
matching in K12 decomposes into two copies of the icosahedron. The icosahedron is 5-
edge-colorable, meaning that it decomposes into five perfect matchings. This can be seen
by drawing h with a central vertex and 5-fold rotational symmetry and forming a perfect
matching using one edge from each of the six orbits of five edges. For each of these five
matchings, the remaining edges of K12 decompose into two copies of the icosahedron. The
resulting ten copies of the icosahedron together cover each edge outside H five times and
cover each edge in H four times. Together with H itself, we obtain 11 copies of H covering
each edge in K12 five times.

April 2024] PROBLEMS AND SOLUTIONS 355

 The decomposition argument generalizes in a straightforward way to yield the following
result: If H is an n-vertex k-regular graph that decomposes into k perfect matchings, and
Kn decomposes into t copies of H plus a leftover perfect matching, then the multigraph
kKn with k copies of each vertex pair as edges decomposes into kt + 1 copies of H . Other
applications besides the problem here include decomposing 4K6 into five octahedra and
decomposing 3K8 into seven cubes, which were proved earlier in the literature.
Also solved by K. Gatesman, O. P. Lossers (Netherlands), R. Pratt, A. J. Schwenk, R. Stong, and the proposer.

A Recursively Defined Sequence

12334 [2022, 588]. Proposed by Florin Stanescu, Şerban Cioculescu School, Găeşti,
Romania. Let f be a real-valued function on [0, 1] with a continuous second derivative.
Assume that f (0) = 0, f (0) = 1, f (0) = 0, and 0 < f (x) < 1 for all x ∈ (0, 1]. Let
x1 , x2 , . . . be a sequence with 0 < x1 ≤ 1 and with
 
x1 + x2 + · · · + xn−1
xn = f
n−1
for n ≥ 2. Prove lim xn ln n = −2/f (0).
n→∞
Solution by Jinhai Yan, Fudan University, Shanghai, China. Let
x1 + x2 + · · · + xn
sn = .
n
Since 0 < f (x) < 1 for all x ∈ (0, 1], we find that f is increasing, 0 < f (x) < x on
(0, 1], and sn ∈ (0, 1]. Moreover,
nsn − (n − 1)sn−1 = xn = f (sn−1 ) < sn−1 .
It follows that 0 < sn < sn−1 . Hence sn is monotone decreasing, so it converges. Let
limn→∞ sn = A. By the Stolz–Cesàro theorem and the continuity of f ,
(n + 1)sn+1 − nsn
f (A) = lim f (sn ) = lim xn+1 = lim = lim sn = A,
n→∞ n→∞ n→∞ n+1−n n→∞

which implies that A = 0. By assumption,

f (0) 2
f (x) = x + x + o(x 2 ) (x → 0+ ).
2
Hence
nsn + f (sn ) nsn + sn + o(sn )
sn+1 = = ∼ sn , and
n+1 n+1
xn = f (sn−1 ) = sn−1 + o(sn−1 ) ∼ sn .
Notice that 1/sn → ∞ monotonically. Applying the Stolz–Cesàro theorem again, together
with ln(1 + 1/n) ∼ 1/n ∼ 1/(n + 1), we find
xn ln n ln(n + 1) − ln n
lim xn ln n = lim · = lim
n→∞ n→∞ sn 1/sn n→∞ 1/sn+1 − 1/sn

ln(1 + 1/n)sn+1 sn sn2 /(n + 1)

= lim = lim
n→∞ sn − sn+1 n→∞ sn − (nsn + f (sn ))/(n + 1)

sn2 sn2 2
= lim = lim =− ,
n→∞ sn − f (sn ) n→∞ −f (0)s 2 /2 + o(s 2 ) f (0)
n n

as claimed.

356 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131

Editorial comment. This problem is a generalization of problem 12079 [2018, 944; 2020,
568] from this Monthly.
Also solved by K. F. Andersen (Canada), A. Berkane (Algeria), H. Chen (US), C. Chiser (Romania),
H. von Eitzen (Germany), K. Gatesman, M. Goldenberg & M. Kaplan, E. A. Herman, O. Kouba (Syria),
J. H. Lindsey II, O. P. Lossers (Netherlands), M. Omarjee (France), N. D. Phuoc (Vietnam), K. Sarma (India),
K. Schilling, A. Stadler (Switzerland), A. Stenger, R. Stong, R. Tauraso (Italy), and the proposer.

Gaussian Integers
12335 [2022, 685]. Proposed by Tom Karzes, Sunnyvale, CA, Stephen Lucas, James Madi-
son University, Harrisonburg, VA, and James Propp, University of Massachusetts, Lowell,
MA. A Gaussian integer is a complex number z such that z = a + bi for integers a and b.
Show that every Gaussian integer can be written in at most one way as a sum of distinct
powers of 1 + i, and that the Gaussian integer z can be expressed as such a sum if and only
if i − z cannot.
Solution by William J. Cowieson, Fullerton College, Fullerton, CA. Let N and N0 denote
the
 sets of positive  integers and nonnegative integers, respectively. Suppose first that
s∈S (1 + i) s
= t∈T (1 + i) t
for distinct sets S, T ⊂ N0 . Subtract one side from the
other, divide by the lowest remaining power of 1 + i, and isolate 1 to obtain 1 = (1 + i)w,
where
 
w= (1 + i)s−1 − (1 + i)t−1
s∈S t∈T

for some disjoint sets S , T ⊂ N. Letting N (z) = zz = a 2 + b2 when z = a + bi, we

conclude
 
1 = N (1) = N (1 + i)w = N (1 + i)N (w) = 2N (w),
which is impossible. Thus equality requires S = T , so there is at most one way to write
any Gaussian integer as a sum of distinct powers of 1 + i.
Let G denote the set Z[i] of Gaussian integers. For z = a + ib ∈ G, we have
z 1 a+b b−a
= (1 − i)z = +i .
1+i 2 2 2
Thus z/(1 + i) is also in G if and only if a + b is even, which holds if and only if a 2 + b2 is
even. On the other hand, if a + b is odd, then (a − 1) + b is even, so (z − 1)/(1 + i) ∈ G.
Writing these conditions in terms of N, we have a mapping F : G → G defined by
z/(1 + i) if N (z) is even
F (z) = .
(z − 1)/(1 + i) if N (z) is odd
We now establish various properties of the mapping F .
Claim 1: For all z ∈ G and n ∈ N0 , we have F n (i − z) = i − F n (z).

Proof. If N(z) is even, then N (i − z) is odd, and

F (i − z) = (i − z − 1)/(1 + i) = i − z/(1 + i) = i − F (z).
The computation when N (z) is odd is similar, yielding F (i − z) = i − F (z) for all z ∈ G.
This is the base case for a proof by induction on n. For the induction step, assuming the
result for n = k − 1, we obtain
F k (i − z) = F (F k−1 (i − z)) = F (i − F k−1 (z)) = i − F (F k−1 (z)) = i − F k (z),
which is the result for n = k. 

April 2024] PROBLEMS AND SOLUTIONS 357

Claim 2: The Gaussian integer z is a sum of distinct powers of 1 + i if and only if F (z) is
also.

Proof. If F (z) is such a sum, then so are (1 + i)F (z) and (1 + i)F (z) + 1, one of which is
z. Conversely, if z is such a sum, then either all powers are positive and F (z) = z/(1 + i)
is such a sum, or 1 is a summand and F (z) = (z − 1)/(1 + i) is such a sum. 

Claim 3: For all z ∈ G, there exists n ∈ N0 such that either F n (z) = 0 or F n (z) = i.
 
Proof. If N(z) is even, then N (F (z)) = N z/(1 + i) = N (z)/2. If N (z) is odd, then
   
N(F (z)) = N (z − 1)/(1 + i) = N (z − 1)/2 = (a − 1)2 + b2 /2,
which is at least a 2 + b2 if and only if (a + 1)2 + b2 ≤ 2. Thus N (F (z)) < N (z) for
all z ∈ G except z ∈ {0, i, −i, −1, −2 + i, −2 − i}, so for every z ∈ G there exists m ∈ N
with F m (z) equal to a member of this set. Furthermore, F 3 (−i) = F 2 (−1) = i = F (i) and
F 6 (−2 − i) = F 5 (−2 + i) = 0 = F (0), so always F n (z) ∈ {0, i} for some n ∈ N. 

Finally, observethat 0 is (vacuously) a sum of distinct powers of 1 + i, while i is not

such a sum: If i = s∈S (1 + i)s for some S ⊂ N0 , then
  
(1 + i)s = i = 1 + (1 + i)i = 1 + (1 + i) (1 + i)s = (1 + i)s .
s∈S s∈S s∈(S+1)∪{0}

By uniqueness, |S| = |(S + 1) ∪ {0}| = |S| + 1, which is impossible for finite S. Since no
such infinite sum converges, i is not such a sum.
It follows from this and Claims 1–3 that z is a sum of distinct powers of 1 + i if and
only if F n (z) = 0 for some n. This is further equivalent to F n (i − z) = i − F n (z) = i for
some n, which holds if and only if i − z is not a sum of distinct powers of 1 + i.
Editorial comment. Gagola, Ionaşcu, Meyerson, Tauraso, Wildon, and the Eagle Problem
Solvers all mentioned the fractal nature of the Gaussian integers shaded in one of two
colors depending on whether the Gaussian integer can or cannot be expressed as a sum of
distinct powers of 1 + i, and they attached graphics showing this property. See W. J. Gilbert
(1982), Fractal geometry derived from complex bases, Math. Intelligencer 4(2): 78–86.
Also solved by J. Boswell & C. Curtis, T. Eisenkölbl (Austria), H. von Eitzen (Germany), S. M. Gagola Jr.,
K. Gatesman, F. Gesmundo (Germany) & T. M. Mazzoli (Austria), N. Hodges (UK), E. J. Ionaşcu, Y. J. Ionin,
S. Lee, O. P. Lossers (Netherlands), M. D. Meyerson, K. Schilling, A. Stadler (Switzerland), A. Stenger,
R. Stong, R. Tauraso (Italy), E. I. Verriest, M. Wildon (UK), Eagle Problem Solvers, Fejéntaláltuka Szeged
Problem Solving Group (Hungary), and the proposers.

Four Concurrent Euler Lines

12336 [2022, 685]. Proposed by Szilárd András, Babeş-Bolyai University, Cluj-Napoca,
Romania. Let N be the center of the nine-point circle of triangle ABC, and let D, E, and
F be the orthogonal projections of N onto the sides BC, CA, and AB, respectively. Prove
that the Euler lines of triangles ABC, AEF , BF D, and CDE are concurrent. Prove also
that the point of concurrency is equidistant from the circumcenters of AEF , BF D, and
CDE.
Solution by Kyle Gatesman, MITRE Corporation, Fairfax, VA. Let H be the orthocenter of
triangle ABC, and let O be its circumcenter. Let the midpoints of the sides opposite A, B,
and C be MA , MB , and MC , respectively. Let the feet of the altitudes from A, B, and C (the
orthic points) be XA , XB , and XC , respectively. Let the midpoints of AH , BH , and CH
(the halfway points) be HA , HB , and HC , respectively. The nine-point circle passes through

358 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131

all three midpoints, all three orthic points, and all three halfway points. It is the image of
the circumcircle of triangle ABC under a dilation centered at H with scaling factor 1/2, so
its center N is the midpoint of the segment OH .
Let OA , OB , and OC be the circumcenters of triangles AEF , BF D, and CDE, respec-
tively. By definition of the points D, E, and F , quadrilaterals AEN F , BF N D, and
CDN E are cyclic, and their circumcircles have diameters AN , BN , and CN , respec-
tively. Therefore OA , OB , and OC are the midpoints of these diameters. It follows that
the circumcircle of triangle OA OB OC is the image of the circumcircle of triangle ABC
under a dilation centered at N with scaling factor 1/2, and its center P is the midpoint
of NO. The point P is equidistant from OA , OB , and OC , and it lies on the line H O,
which is the Euler line of triangle ABC, so it suffices now to show that the Euler lines
of triangles AEF , BF D, and CDE are concurrent at P . We show that the Euler line of
triangle AEF passes through P ; the claims for the other two triangles follow by symmetry
of the construction.

Note that HA is the orthocenter of triangle AMB MC , because AMB HA MC is the image
of ACH B under a dilation centered at A with scaling factor 1/2. Let J be the orthocenter of
triangle AXB XC . Letting α = ∠BAC, we have AXC = AC cos α and AXB = AB cos α.
Therefore triangle AXB XC is the image of triangle ABC under first a dilation centered
at A with scale factor cos α and then a reflection across the line m that bisects ∠BAC. It
follows that J lies on the reflection of AH across m. Since
π 1 π π
∠OAB = − ∠AOB = − ∠ACB = − ∠ACXA = ∠XA AC = ∠H AC,
2 2 2 2
O also lies on the reflection of AH across m. Thus A, O, and J are collinear.
For any pair of parallel lines 1 and 2 , we say that the line that is parallel to both 1
and 2 and halfway between them is their midline. This midline contains the midpoint of
every line segment with an endpoint on each of 1 and 2 . The nine-point circle of triangle
ABC passes through MB , XB , MC , and XC , so E and F are the midpoints of MB XB and
MC XC , respectively. The midline of the altitudes from MB and XB to AB is the altitude
from E to AB, and the midline of the altitudes from MC and XC to AC is the altitude from
F to AC. Since HA lies on the altitudes from both MB and MC , and J lies on the altitudes
from both XB and XC , the midpoint of the segment HA J lies on the altitudes from both E
and F , and therefore it is the orthocenter of triangle AEF .

April 2024] PROBLEMS AND SOLUTIONS 359

 Since HA is the midpoint of H A and N is the midpoint of H O, the line AO is parallel
to HA N . Since OA is the midpoint of N A and P is the midpoint of N O, the line OA P is
parallel to AO, and it is the midline of AO and HA N . The orthocenter of triangle AEF is
the midpoint of the segment HA J , whose endpoints lie on the lines HA N and AO, so this
orthocenter lies on OA P . This shows that OA P is the Euler line of triangle AEF , and it
passes through P , as required.
Editorial comment. N. S. Dasireddy pointed out that this problem was posed as part
of the 2015 IMO preparation program in Vietnam. Several solutions can be found at
artofproblemsolving.com/community/c6h1087710.
Also solved by M. Bataille (France), S. Bhadra (India), N. S. Dasireddy (India), G. Fera (Italy), O. Geupel
(Germany), N. Hodges (UK), W. Janous (Austria), K.-W. Lau (China), G.-H. Liu (Taiwan), O. P. Lossers
(Netherlands), F. Masroor, C. R. Pranesachar (India), C. Schindler (Germany), R. Stong, B. D. Suceavă, and
the proposer.

Beta Integrals and Partial Fractions

12337 [2022, 685]. Proposed by Hideyuki Ohtsuka, Saitama, Japan. For k ∈ {0, 1, 2}, let
 (−4)n 2n−1
Sk = ,
2n + 1 n
where the sum is taken over all nonnegative integers n that are congruent to k modulo 3.
Prove
√
ln 1 + 2 π
(a) S0 = √ + ;
3 2 6
√ √
ln 1 + 2 ln 2 + 3 π
(b) S1 = √ − √ − ; and
3 2 2 3 12
√ √
ln 1 + 2 ln 2 + 3 π
(c) S2 = √ + √ − .
3 2 2 3 12
Solution by Roberto Tauraso, Tor Vergata University of Rome, Rome, Italy. The presence
 −1
of 2n
n
suggests expressing the sums using beta function integrals. We have
   1
(−4)n 2n −1
= (−4)n B(n + 1, n + 1) = (−4)n t n (1 − t)n dt
2n + 1 n 0
 1  1
1
= (s 2 − 1)n ds = (s 2 − 1)n ds
2 −1 0

under the change of variable s = 2t − 1. Evaluating a geometric series and using partial
fractions yields
 1 ∞  1   
ds 1 1 1 s2 + 1
S0 = (s − 1) ds =
2 3r
= + ds. (∗)
0 1 − (s − 1) 3 0 2 − s2 s4 − s2 + 1
2 3
0 r=0

To study this and similar expressions, let

 1  1 
ds s2 − 1 1
s2 + 1
I0 = , I1 = ds, and I2 = ds.
0 2−s 0 s −s +1 s4 − s2 + 1
2 4 2
0
Using partial fractions and
√ √
s 4 − s 2 + 1 = (s 2 + 3s + 1)(s 2 − 3s + 1),

360 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131

we obtain
   √
1 1
1 1 ln(1 + 2)
I0 = √ √ +√ ds = √ ,
2 2 0 2+s 2−s 2
 √ √ √
1 1
2s + 3 2s − 3 ln(2 + 3)
I1 = − √ √ + √ ds = − √ ,
2 3 0 s 2 + 3s + 1 s 2 − 3s + 1 3
and
  
1 1 1 1
I2 = √ + √ ds
2 0 s 2 + 3s + 1 s 2 − 3s + 1
 √ √ 1 π
= arctan(2s + 3) + arctan(2s − 3) = .
0 2
From (∗), we have S0 = (I0 + I2 )/3, which completes (a). For the other sums, we com-
pute
 1∞
 2 
S1 − S2 = (s − 1)2k+1 − (s 2 − 1)3k+2 ds
0 k=0
 
1
(s 2 − 1)(2 − s 2 ) 1
s2 − 1
= ds = ds = I1
0 1 − (s 2 − 1)3 0 s4 − s2 + 1
and
 ∞
1
S0 + S1 + S2 = (s 2 − 1)n ds = I0 .
0 n=0

This yields S1 = (2I0 + 3I1 − I2 )/6 for (b) and S2 = (2I0 − 3I1 − I2 )/6 for (c).
Editorial comment. Michel Bataille based his solution on the general formula
∞
(2x)2n 2x arcsin(x)
2n = √ ,
n=1
n n 1 − x2
proved in D. H. Lehmer (1985), Interesting series involving the central binomial coeffi-
cient, this Monthly 92(7): 449–457.
Also solved by T. Amdeberhan & V. H. Moll, M. Bataille (France), A. Berkane (Algeria), P. Bracken, B. Bradie,
H. Chen (US), W. J. Cowieson, K. Gatesman, M. L. Glasser, N. Hodges (UK), W. Janous (Austria), O. Kouba
(Syria), C. Krattenthaler (Austria), P. Lalonde (Canada), G. Lavau (France), O. P. Lossers (Netherlands),
R. Molinari, A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), D. Terr, M. Vowe (Switzerland), T. Wiandt,
M. Wildon (UK), Y. Zhang (China), and the proposer.

A Trigonometric Exponential Integral by the Leibniz Integral Rule

12338 [2022, 686]. Proposed by István Mező, Nanjing, China. Prove
 ∞
cos(x) − 1 1  
dx = ln π csch(π ) .
0 x (e − 1)
x 2

Solution by Tewodros Amdeberhan and Victor H. Moll, Tulane University, New Orleans,
LA. We start by writing the integral in the form
 ∞  ∞ ∞  ∞
cos x − 1 cos x − 1 (cos x − 1)e−(n+1)x
dx = dx = dx.
0 x(ex − 1) 0 xex (1 − e−x ) n=0 0
x

April 2024] PROBLEMS AND SOLUTIONS 361

For a > 0, let
 ∞
(cos x − 1)e−ax
I (a) = dx.
0 x
Differentiation under the integral sign (an application of what is known as the Leibniz
integral rule) yields
 ∞  
a cos x − sin x 1 ∞ 1
I (a) = (1 − cos x)e−ax dx = e−ax −  = a(1 + a 2 ) ,
0 1 + a 2 a 0

and therefore
  
da 1 a2
I (a) = = ln + C.
a(1 + a 2 ) 2 1 + a2
Since lima→∞ I (a) = 0, we have C = 0, so the requested integral is given by
 ∞ ∞   ∞
cos x − 1 1 (n + 1)2 1  k2
dx = ln = ln .
0 x(ex − 1) 2 n=0 1 + (n + 1)2 2 k=1
1 + k2

To evaluate the infinite product, we use the known formula

∞  
sinh z  z2
= 1+ 2 2
z k=1
π k

(see I. S. Gradshteyn, I. M. Ryzhik (2007), Table of Integrals, Series, and Products, 7th
ed., Burlington, MA: Academic Press, equation 1.431.2, p. 45). Applying this formula
with z = π , we obtain
 ∞ ∞  
cos x − 1 1 1 −1 1 π
dx = ln 1 + = ln .
0 x(e − 1)
x 2 k=1
k 2 2 sinh π

Also solved by U. Abel & V. Kushnirevych (Germany), A. Berkane (Algeria), S. Bhadra (India), R. Bit-
tencourt (Brazil), K. N. Boyadzhiev, P. Bracken, B. Bradie, C. Burnette, H. Chen (US), W. J. Cowieson,
B. E. Davis, M.-C. Fan (China), G. Fera (Italy), P. Fülöp (Hungary), M. L. Glasser, H. Grandmontagne
(France), N. Hodges (UK), F. Holland (Ireland), W. Janous (Austria), S. Kaczkowski, O. Kouba (Syria),
K.-W. Lau (China), G. Lavau (France), O. P. Lossers (Netherlands), J. Magliano, M. Maniquiz, F. Masroor,
R. Mortini (Luxembourg) & R. Rupp (Germany), K. Nelson, M. Omarjee (France), D. Pascuas (Spain), P. Per-
fetti (Italy), S. Sharma (India), A. Stadler (Switzerland), S. M. Stewart (Saudi Arabia), R. Stong, R. Tauraso
(Italy), E. I. Verriest, M. Vowe (Switzerland), T. Wiandt, Y. Zhang (China), L. Zhou, and the proposer.

CLASSICS
C24. Due to Colin L. Mallows. Over the history of a certain tennis club, every player has
played at least one match against every other player. Matches are played one at a time, and
after each match a ranking of the players in the club is computed as follows. Starting with
the most recent match and working backwards through time, use the match results to build
up a partial order. Ignore any match that is inconsistent with more recent results. The final
result is guaranteed to be a linear order, since any incomparability between a pair of players
is resolved when a match between them is encountered. This linear order becomes the new
club ranking. Prove or disprove: A player cannot rise in the club ranking by intentionally
losing a match.

362 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131

 Decomposing Space into Disjoint Circles
C23. Due to John H. Conway and Hallard T. Croft. Determine whether it is possible to
partition R3 into circles.
Solution. We first show that any sphere with two points deleted can be partitioned into
circles. This is clear when the two
deleted points are antipodal, wit-
nessed by the circular latitude lines
that cover the surface of the earth
apart from the two poles. When the
two deleted points are not antipodal,
let L be the line in space common
to the two planes that are tangent
to the sphere at the deleted points.
The cross-sections of the sphere with
the planes through L decompose the
sphere with two deleted points into
circles. This is illustrated at right.
Let B be the family of unit circles in the xy-plane centered at (4k + 1, 0, 0) for some
integer k. These circles are disjoint, and one of them contains the origin. For r > 0,
let Sr be the sphere of radius r centered at the origin. The key observation is that, for
every r > 0, Sr intersects B in exactly two points. This is illustrated in the figure here:

Let Tr be a set of disjoint circles whose union is Sr with those two points deleted. The
union of B and Tr for all r > 0 gives the desired decomposition of R3 .
Editorial comment. The problem first appeared in J. H. Conway and H. T. Croft (1964),
Covering a sphere with congruent great-circle arcs, in Math. Proc. Cambridge Phil. Soc.,
60: 787–800, where it was solved using the axiom of choice. While that solution gives the
stronger result that all the circles can be of unit radius and no two circles are linked, it does
not give an explicit construction. The solution here is due to Andrzej Szulkin (1983), R3 is
the union of disjoint circles, this Monthly, 90: 640–641. For a more detailed treatment,
see J. B. Wilker (1989), Tiling R3 with circles and disks, Geom. Dedicata 32: 203–209.
It is not possible to partition R2 into circles. In fact, if S is a family of disjoint circles in
the plane, then in the interior of every circle in S is a point not contained in any circle in S.
To see this, let C be such a circle. If the center of C is not part of any circle in S, then we
are done. Otherwise, let C be a circle in S containing the center of C. Note that the radius
of C is less than half the radius of C. If the center of C is not part of any circle in S, then
we are again done. Otherwise, in the same way, let C be a circle in S containing the center
of C , and continue in this way to form an infinite family of nested circles in S whose radii
converge to 0. The point in the interior of all these circles cannot be part of any circle in S.

April 2024] PROBLEMS AND SOLUTIONS 363

SOLUTIONS

Two Tangent Circles

12325 [2022, 487]. Proposed by Dong Luu, Hanoi University of Education, Hanoi, Viet-
nam. Let ABCD be a quadrilateral with a circumscribed circle ω and an inscribed circle
γ . Prove that there are two circles α and β with the following property: For any triangle
MEF with (1) M on ω, (2) E and F on the line AB, and (3) the lines ME and MF
tangent to γ , the circumcircle of MEF is tangent to α and β.
Solution by Faraz Masroor, New York, NY. We first address a more general situation. Let ω
and γ be circles, with γ inside ω, and let  be a line tangent to γ . We show that there are
circles α and β as in the statement of the problem, with  playing the role of the line AB,
except that one of the circles may degenerate to a point. At the end, we show that if ω and
γ are the circumscribed and inscribed circles of a quadrilateral, then the degeneracy can
be ruled out.
Let W be the point where  is tangent to γ . Let I be the center of γ , and let R be its
radius. Let ω be the image of ω under inversion in γ . The circle ω is inside γ and contains
I in its interior. Let J be the center of ω , and let r be its radius. Let S be the image of I
under reflection in J , and let T be the midpoint of W S.
Consider any triangle MEF as in the statement of the problem, and let μ be its cir-
cumcircle. Let M , E , and F be the images of M, E, and F under inversion in γ , and let
μ be the image of μ. The circle μ circumscribes M E F ; let P be its center. We now
make two claims about μ and P :
Claim 1: The radius of μ is R/2.
Claim 2: T P = r.
Before proving these claims, we show that they imply the desired conclusion. Let α be
a circle centered at T with radius |R/2 − r|, and let β be a circle centered at T with radius
R/2 + r. (It is possible that r = R/2, in which case α degenerates to a point.) The claims
above imply that the circle μ is tangent to both α and β . The point of tangency with α
is the point on μ that is closest to T , and the point of tangency with β is the point on μ
furthest from T . It follows that, as long as neither α nor β passes through I , the images
of α and β under inversion in γ are circles α and β that are tangent to μ, as required.

March 2024] PROBLEMS AND SOLUTIONS 263

 To confirm that neither α nor β passes through I , note first that since J is the midpoint
of I S and T is the midpoint of W S, we have J T = (1/2)I W = R/2. Also, since I is inside
ω , which is inside γ , we must have I J < r and I J < R − r. It follows that T I > R/2 − r
and T I > R/2 − (R − r) = r − R/2, so T I > |R/2 − r|, and therefore I lies outside of
α . Also, T I ≤ J T + I J < R/2 + r, so I lies inside of β . Thus neither α nor β passes
through I .
To prove the claims, let ME and MF be tangent to γ at G and H , respectively. We use
the fact that if the tangent lines to γ at two points X and Y intersect at Z, then the image of
Z under inversion in γ is the midpoint of XY . Applying this fact three times, we see that
E is the midpoint of W G, F is the midpoint of W H , and M is the midpoint of GH . Let
V and Q, respectively, be the images of W and I under reflection in M , and let σ be the
image of γ under this reflection. The circle σ has radius R passing through G, H , and V ,
and its center is Q. Also, M is the midpoint of both I Q and W V .
Since the midpoints of W V , W G, and W H are M , E , and F , respectively, the image
of σ under a dilation with ratio 1/2 centered at W is μ , the circumcircle of M E F . This
proves the Claim 1. Also, the image of Q under this dilation is P , so P is the midpoint of
W Q. Since T and P are the midpoints of W S and W Q, respectively, and J and M are
the midpoints of I S and I Q, respectively, we have T P = (1/2)SQ = J M . But M is on
ω, so M is on ω , which is the circle of radius r centered at J . Therefore T P = J M = r,
which proves Claim 2.
Finally, we show that if ω and γ are the circumscribed and inscribed circles of a quadri-
lateral ABCD, then the case r = R/2 can be ruled out, thus eliminating the possibility that
one of the circles is degenerate. In this case, by Poncelet’s porism, there is another quadri-
lateral A B C D such that ω and γ are the circumscribed and inscribed circles of A B C D
and, in addition, A C is a diameter of ω and contains a diameter of γ . A straightforward
calculation now shows that r = R(cos θ + sin θ )/2 > R/2, where θ = ∠A C B .
Editorial comment. Since ω is a circle centered at J with radius r and J T = R/2, the
circles α and β are both tangent to ω . It follows that α and β are both tangent to ω. Also,
the line through the centers of α and β passes through I .
Also solved by O. Kouba (Syria), O. P. Lossers (Netherlands), C. R. Pranesachar (India), R. Stong, and the
proposer.

264 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
 A Function with Polynomial Differences is a Polynomial
12326 [2022, 487]. Proposed by George Stoica, Saint John, NB, Canada. Let f : R → R
be a continuous function such that, for every fixed y ∈ R, f (x + y) − f (x) is a polynomial
in x. Prove that f is a polynomial function.
Solution by Jinhai Yan, Fudan University, Shanghai, China. We first note that, for every
nonzero real number c, all polynomials in x can be expressed in the form p(x + c) − p(x)
for some other polynomial p. To see this, note that (x + c)n − x n has degree n − 1, and
hence {(x + c)n − x n : n ≥ 1} forms a basis for the vector space of all polynomials in x.
Taking y = 1 in the hypothesis, we see that f (x + 1) − f (x) is a polynomial, so
we can find a polynomial p such that p(x + 1) − p(x) = f (x + 1) − f (x). Therefore
f (x + 1) − p(x + 1) = f (x) − p(x). If we let T (x) = f (x) − p(x), then T is periodic
with period 1.
Now let n be any positive integer. It is easy to check that T satisfies the same hypoth-
esis as f , and taking y = 1/n in that hypothesis we conclude that T (x + 1/n) − T (x)
is a polynomial. Therefore we can find a polynomial q such that q(x + 1/n) − q(x) =
T (x + 1/n) − T (x). It follows that
q(x + 1) − q(x) = T (x + 1) − T (x) = 0,
so q is periodic. But q is a polynomial, and therefore it is a constant function. Thus

T (x + 1/n) − T (x) = q(x + 1/n) − q(x) = 0,

so T is periodic with period 1/n.

Since n was arbitrary, T has period 1/n for every positive integer n, so it is constant
on the rationals. It is also continuous, so it must be a constant function. Finally, since
f (x) = p(x) + T (x), we conclude that f is a polynomial.
Editorial comment. The problem appears as Lemma 2.5 in F. Kühn and R. L. Schilling
(2021), For which functions are f (Xt ) − Ef (Xt ) and g(Xt )/Ef (Xt ) martingales?,
Theor. Prob. and Math. Statist. 105: 79–91. The problem was submitted to the Monthly
without mention of this reference. We regret publishing the problem without proper attri-
bution to the source.
Omran Kouba pointed out that it is sufficient to assume that f (x + y) − f (x) is a poly-
nomial for two values of y that are independent over the rationals.
The hypothesis that f is continuous is necessary. We can prove this by imitating the
reasoning used in the solution to classic problem C7 [2022, 694; 2022, 794]. Let B be a
basis for R as a vector space over Q, and let f : B → R be any function that takes the
value 0 infinitely many times but is not identically 0. Now extend f to a function n from
R to R as follows: for any real number x, write x (uniquely)
 as a finite sum i=1 qi bi ,
where qi ∈ Q \ {0} and bi ∈ B, and define f (x) to be ni=1 qi f (bi ). It is easy to verify
that for any fixed y, f (x + y) − f (x) is equal to f (y), which is a constant function and
therefore a polynomial. Since f takes the value 0 infinitely many times but is not identically
0, it cannot be a polynomial. As explained in the editorial comment of problem C7, the
reasoning here requires the axiom of choice, and indeed the result cannot be proved without
using the axiom of choice.
Also solved by J. Boswell & C. Curtis, N. Caro-Montoya (Brazil), J.-P. Grivaux (France), D. A. Hejhal,
N. Hodges (UK), Y. J. Ionin, O. Kouba (Syria), O. P. Lossers (Netherlands & ELTE (Hungary), F. Mas-
roor, R. Mortini (France) & P. Pflug (Germany) & A. Sasane (UK), M. Omarjee (France), K. Sarma (India),
K. Schilling, A. Stenger, R. Stong, R. Tuaraso (Italy), D. J. Velleman, UM6P Math Club (Morocco), and the
proposer.

March 2024] PROBLEMS AND SOLUTIONS 265

 A Gaussian Binomial Identity
12327 [2022, 487]. Proposed by Mircea Merca, University of Craiova, Craiova, Romania.
Let
⎧k−1
 ⎪  1 − q n−i
⎨ if 1 ≤ k ≤ n;
n
= i=0 1 − q k−i
k q ⎪ ⎩
1 if k = 0.
Prove
n 
 n 

n n 2 −n(n−1)/2
q =
k
q k(k−1)+(n−k)
k q2
k q2
k=0 k=0

for n ≥ 0.
Solution by Doyle Henderson, Omaha, NE. Let Sn and Tn , respectively,
n denote kthe sums
on the left and on the right, respectively.
  We show that both equal k=1 (1 + q ). Using
n
(1 − q ) k q = (1 − q
n+1 n+1−k
) kn+1 2
with q replacing q, we obtain
q

n 
 n+1
(1 − q 2n+2 )Sn = (1 − q 2n+2−2k )q k
k=0
k q2

n 
 n 

n+1 n+1
= q −
k
q 2n+2−k .
k=0
k q2 k=0
k q2
n  n 
Using the well-known identity k = and reversing the index of summation yields
  n+1 
q n−k q

Sn+1 = n+1k=0 k 2
q n+1−k , so
q

n 
 n+1
q 2n+2−k = q n+1 (Sn+1 − 1).
k=0
k q2

We have now proved

(1 − q 2n+2 )Sn = Sn+1 − q n+1 − q n+1 (Sn+1 − 1) = (1 − q n+1 )Sn+1 ,

so Sn+1 = (1 + q n+1 )Sn . This and S0 = 1 yield Sn = nk=1 (1 + q k ).
To evaluate Tn , let f (n, k) = k(k − 1) + (n − k)2 − n(n − 1)/2. Proceeding as for Sn
yields
n  n 
n+1 n+1
(1 − q 2n+2
)Tn = q f (n,k)
− q f (n,k) q 2n+2−2k .
k=0
k q 2
k=0
k q 2

 n   
Since f (n + 1, n + 1 − k) = f (n, k), using n−k = nk q and reversing the index of
  n+1 
q
summation in Tn+1 yields Tn+1 = n+1 k=0 k 2
q f (n,k) . Computing also that f (n, k) +
q
n + 1 − 2k = f (n + 1, k), we obtain
n   n 
n+1 n+1
q f (n,k) 2n+2−2k
q =q n+1
q f (n+1,k)
k=0
k q 2
k=0
k q 2

 
=q n+1
Tn+1 − q f (n+1,n+1) .

266 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
Since f (n, n + 1) = n + 1 + f (n + 1, n + 1), we have now proved
 
(1 − q 2n+2 )Tn = Tn+1 − q f (n,n+1) − q n+1 Tn+1 − q f (n+1,n+1) = (1 − q n+1 )Tn+1 .

Hence Tn+1 = (1 + q n+1 )Tn . This and T0 = 1 yield Tn = nk=1 (1 + q k ), finishing the
proof.
Also solved by T. Amdeberhan & S. B. Ekhad, N. Hodges (UK), W. P. Johnson, P. Lalonde (Canada),
O. P. Lossers (Netherlands), R. Stong, R. Tauraso (Italy), and the proposer.

The Value Set of an Integer Quadratic Form

12328 [2022, 587]. Proposed by Peter Koymans and Jeffrey Lagarias, University of Michi-
gan, Ann Arbor, MI. An integer binary quadratic form is a function f : Z2 → Z defined by
f (m, n) = am  + bmn + cn for some
2 2
 a, b, c ∈ Z. The value set V(f) of such a form is
defined to be f (m, n) : (m, n) ∈ Z2 .
(a) Prove that if f1 (m, n) = m2 − mn − 3n2 and f2 (m, n) = m2 − 13n2 , then V (f1 ) =
V (f2 ).
(b) Prove that if f1 (m, n) = m2 − mn − 4n2 and f2 (m, n) = m2 − 17n2 , then V (f2 ) ⊆
V (f1 ) but V (f1 ) = V (f2 ).
Solution by Jacob Boswell and Charles Curtis, Missouri Southern State University, Joplin,
MO. In both (a) and (b), f2 (m, n) = f1 (m + n, 2n), so V (f2 ) ⊆ V (f1 ).
(a) It suffices to show V (f1 ) ⊆ V (f2 ). We have
 
11m − 25n −3m + 7n
f1 (m, n) = f2 , ,
2 2
so f1 (m, n) ∈ V (f2 ) when m and n have the same parity. When m is even and n is odd,
we use f1 (m, n) = f1 (m − n, −n) to see that f1 (m, n) ∈ V (f2 ), since in this case m − n
and −n have the same parity. Finally, when m is odd and n is even, we have f1 (m, n) =
f2 (m − n/2, n/2) ∈ V (f2 ).
(b) It suffices to show V (f1 ) ⊆ V (f2 ). We have f2 (m, n) = m2 − 17n2 ≡ 2 (mod 4).
However f1 (3, 1) = 2 ≡ 2 (mod 4).
Editorial comment.
√ The identity used in√(a) arises from the multiplicative property of the
norm N in Z[ 13] defined by N (a + b 13) = f2 (a, b). Thus
   
11 3 n n 11m − 25n 7n − 3m
f1 (m, n) = f2 ,− f2 m − , = f2 , .
2 2 2 2 2 2
Several solvers considered more generally f1 (m, n) = m2 − mn − an2 and f2 (m, n) =
m2 − (4a + 1)n2 , for integers a. The argument given for (a) applies when f2 (m, n) = 4
has a solution in odd integers. For example, when
 a = 7 we have 272 − 29 · 52 = 4, so
f1 (m, n) = f2 (27m + 59n)/2, (5m + 11n)/2 .
Also solved by U. Abel & V. Kushnirevych (Germany), A. J. Bevelacqua, P. Corn, T. Eisenkölbl (Austria),
G. Fera (Italy), K. Gatesman, Y. J. Ionin, O. P. Lossers (Netherlands), B. Phillabaum, C. R. Pranesachar (India),
M. Reid, J. P. Robertson, A. Stadler (Switzerland), A. Stenger, R. Stong, R. Tauraso (Italy), D. Terr, L. Zhou,
and the proposer.

Equally Spaced Unit Vectors

12329 [2022, 587]. Proposed by Leonard Giugiuc, Drobeta-Turnu Severin, Romania. Let n
be a positive integer with n ≥ 3. For each positive integer m with m ≥ 2, find all real values
λm such that there are m distinct unit vectors v1 , . . . , vm in Rn satisfying vi · vj = λm for
all i, j with 1 ≤ i < j ≤ m.

March 2024] PROBLEMS AND SOLUTIONS 267

Solution by Kuldeep Sarma, Tezpur University, Tezpur, India. If m ≤ n, then the allowed
values for λm are all numbers in the interval [−1/(m − 1), 1). If m = n + 1, then the only
allowed value for λm is −1/n. If m > n + 1, then no such λm exists.
Assume that such vectors exist, and let them be the columns of a real n × m matrix V .
The Gram matrix V T V has ones on the diagonal and λm in every off-diagonal position.
Thus V T V = (1 − λm )Im + λm Jm , where Jm denotes the m-by-m matrix of all 1s. The
vector (1, . . . , 1)T is an eigenvector for this matrix with eigenvalue 1 + (m − 1)λm , and
there are m − 1 linearly independent vectors whose coordinates sum to zero, all of which
are eigenvectors with eigenvalue 1 − λm . Thus the eigenvalues of this matrix are 1 − λm
with multiplicity m − 1 and 1 + (m − 1)λm with multiplicity 1. Since V T V is positive
semidefinite, these eigenvalues must be nonnegative, and hence −1/(m − 1) ≤ λm ≤ 1.
However, the case λm = 1 is excluded, since it would force all of the vi to be equal. Also
note that V T V has rank at most n. Hence if m = n + 1, then V T V must have 0 as an eigen-
value, which implies λm = −1/n. If m > n + 1, then V T V must have 0 as an eigenvalue
with multiplicity at least 2, which is impossible since λm = 1 has been excluded.
Conversely, suppose that the conditions on the eigenvalues λm are satisfied. Let A =
(1 − λm )Im + λm Jm . The matrix A is a symmetric positive semidefinite m-by-m matrix
with rank at most n. Let r = rank(A). Using either the Cholesky decomposition or the
orthogonal diagonalizability of real symmetric matrices, we can write A = XT X for some
r-by-m matrix X, and padding with n − r extra rows of 0s, we can write A = V T V for
some n-by-m matrix V . The columns of V are then the desired unit vectors vi . They are
distinct since vi · vj = λm < 1 for 1 ≤ i < j ≤ m.
Also solved by C. P. Anil Kumar (India), N. Caro-Montoya (Brazil), M. Elgersma, K. Gatesman, Y. J. Ionin,
O. P. Lossers (Netherlands), R. Stong, L. Zhou, Eagle Problem Solvers, and the proposer.

One Concurrency Leads To Another

12330 [2022, 587]. Proposed by Oleh Faynshteyn, Leipzig, Germany. In the acute and
scalene triangle ABC, let G be the centroid, H be the orthocenter, D, E, and F be the feet
of the altitudes from A, B, and C, respectively, and K, L, and M be the midpoints of BC,
CA, and AB, respectively. Let P be the
intersection of DG and KH , let Q be
the intersection of EG and LH , and
let R be the intersection of F G and
MH .
(a) Prove that AP , BQ, and CR are
concurrent.
(b) Let X, Y , and Z be the points
where GH intersects AP , BQ, and
CR. Prove
HX HY HZ
+ + = 3.
XG YG ZG
Solution by Faraz Masroor, New York, NY. Neither statement depends on the triangle being
acute or H being the orthocenter of ABC. We can let H be any point in the interior of
ABC not lying on any of the lines AG, BG, or CG, as long as we redefine D, E, and F
to be the intersections of AH , BH , and CH with BC, CA, and AB, respectively.
(a) Let S, T , and U be the intersections of AP , BQ, and CR with BC, CA, and AB,
respectively. We must prove that AS, BT , and CU are concurrent. By Ceva’s theorem it
suffices to show
BS CT AU
· · = 1. (1)
SC T A U B

268 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
 Let HD and HK be the reflec-
tions of H across D and K,
and let A be the point such
that AHD HK A is a parallelogram.
Note that A A  HK HD  KD, so
∠A AG = ∠DKG. Also, A A =
HK HD = 2KD, and since G is
the centroid of ABC, AG =
2KG. Therefore triangles A AG
and DKG are similar. It follows
that ∠A GA = ∠DGK, so A , G,
P , and D are collinear.
Since SDP ∼ AA P and DH P ∼ A HK P ,
SD PD DH DHD
= = = . (2)
AA PA A HK AHD
Let S be the intersection of AHK with BC. Since S D HK ∼ AA HK ,
SD D HK DHD
= = . (3)
AA A HK AHD
Combining (2) and (3), we see that SD = S D . Also, D D = HK HD = 2KD, so K is
the midpoint of DD , and therefore KS = KS . We conclude BS = CS and SC = S B.
Let HL and HM be the reflections of H across L and M, respectively, let T be the
intersection of BHL with CA, and let U be the intersection of CHM with AB. Imitating
the reasoning above, we can show CT = AT , T A = T C, AU = BU , and U B = U A.
Therefore (1) is equivalent to
CS AT BU
· · = 1.
SB T C UA
By another application of Ceva’s theorem, to prove this it suffices to show that AHK , BHL ,
and CHM are concurrent.
Since K is the midpoint of
both H HK and BC, BH CHK is a
parallelogram. Similarly, CH AHL
and AH BHM are parallelograms.
We have AHL  H C  BHK and
AHL = H C = BHK , so AHL HK B
is also a parallelogram. There-
fore the midpoints of the diago-
nals AHK and BHL coincide. Sim-
ilarly, this common midpoint coin-
cides with the midpoint of CHM , so
the three segments are concurrent,
as required.
(b) The law of sines implies
DS AD sin ∠DAP
= · .
SK AK sin ∠KAP
This is sometimes known as the ratio lemma. A second application of the ratio lemma
yields
HX AH sin ∠DAP
= · .
XG AG sin ∠KAP

March 2024] PROBLEMS AND SOLUTIONS 269

Combining these, and applying the fact that AK = (3/2)AG, we obtain
HX AH AK DS 3 AH DS
= · · = · · . (4)
XG AG AD SK 2 AD SK
By Ceva’s theorem applied to ADK,
DS KG AH
· · = 1,
SK GA H D
and therefore
DS GA H D HD
= · =2· .
SK KG AH AH
Substituting into (4), we obtain
HX 3 AH HD HD [H BC]
= · ·2· =3· =3· ,
XG 2 AD AH AD [ABC]
where for any points α, β, and γ , [αβγ ] denotes the area of αβγ . Similarly, H Y /Y G =
3[H CA]/[ABC] and H Z/ZG = 3[H AB]/[ABC], so
HX HY HZ [H BC] + [H CA] + [H AB] [ABC]
+ + =3· =3· = 3.
XG YG ZG [ABC] [ABC]

Editorial comment. Let O be the point on GH such that G is between O and H and
GH = 2OG. It can be shown that O is the common midpoint of AHK , BHL , and CHM .
Also solved by M. Bataille (France), H. Chen (China), C. Chiser (Romania), I. Dimitrić, G. Fera (Italy),
M. Goldenberg & M. Kaplan, K. Gatesman, J.-P. Grivaux (France), K.-W. Lau (China), C. R. Pranesachar
(India), V. Schindler (Germany), R. Stong, D. E. Türköz (Turkey), L. Zhou, and the proposer.

A Hyperbolic Integral
12332 [2022, 588]. Proposed by Finbarr Holland, University College, Cork, Ireland. Prove
 ∞
tanh2 x 14 ζ (3)
2
dx = ,
0 x π2

where ζ (3) is Apéry’s constant ∞ k=1 1/k .
3

Solution by Kuldeep Sarma, Tezpur University, Tezpur, India. We begin with the Weier-
strass product formula for the hyperbolic cosine,
∞  
x2
cosh x = 1+ .
n=0
(n + 1/2)2 π 2

Applying logarithmic differentiation, we obtain

∞
tanh x 1
=2 ,
x n=0
(n + 1/2)2π 2 + x2

and therefore
∞
tanh2 x 1
= 4   . (1)
x2 n,m=0
(n + 1/2)2 π 2 + x 2 (m + 1/2)2 π 2 + x 2

Next we claim that for all positive a and b,

 ∞
dx π
= . (2)
0 (a + x )(b + x )
2 2 2 2 2ab(a + b)

270 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
For distinct a and b, this follows from the calculation
 ∞  ∞ 
dx 1 1 1
= 2 − 2 dx
0 (a 2 + x 2 )(b2 + x 2 ) b − a2 0 a2 + x 2 b + x2
1 π π π
= − = ,
b2 − a 2 2a 2b 2ab(a + b)
but it is easily verified that (2) also holds when a = b.
Combining (1) and (2), we have
 ∞ ∞  ∞

tanh2 x 1
dx = 4    dx
0 x 2
n,m=0 0
(n + 1/2) π + x (m + 1/2)2 π 2 + x 2
2 2 2

∞
16  1
= 2
π n,m=0 (2n + 1)(2m + 1)(2n + 2m + 2)
∞  1
16  1 dx
= x (2n+2m+2)
π n,m=0 (2n + 1)(2m + 1) 0
2 x
  ∞ ∞
16 1  x 2n+1  x 2m+1 dx
= 2 ·
π 0 n=0 2n + 1 m=0 2m + 1 x
 1  
4 2 1−x dx
= 2 ln .
π 0 1+x x
Finally, to evaluate the last integral, we substitute u = (1 − x)/(1 + x) and obtain
 ∞  1 ∞ 
tanh2 x 8 ln2 u 8  1 2n 2
dx = du = u ln u du
0 x2 π 2 0 1 − u2 π 2 n=0 0
∞
16  1 16 7ζ (3) 14ζ (3)
= = 2· = .
π 2 n=0 (2n + 1)3 π 8 π2

Editorial comment. Several solvers noted a relationship between this problem and problem
12317, which asked for a proof of
 π/2
sin(4x) ζ (3)
dx = −14 2 .
0 ln(tan x) π
Let I denote the integral in problem 12317, and J the integral in this problem. Using the
substitution u = tan x, we have
 ∞
4u(1 − u2 )
I= du.
0 (1 + u2 )3 ln u
We can reexpress J by applying integration by parts, recognizing that the resulting inte-
grand is odd, and expressing the hyperbolic functions in terms of exponentials:
 ∞  ∞ 2x 2x
sinh x 4e (e − 1)
J =2 dx = dx.
−∞ x(e + 1)
3 2x 3
0 x cosh x
Finally, using the substitution u = ex , we obtain
 ∞
4u(u2 − 1)
J = du = −I.
0 (1 + u2 )3 ln u

March 2024] PROBLEMS AND SOLUTIONS 271

Also solved by U. Abel & V. Kushnirevych (Germany), T. Amdeberhan & V. Moll, K. F. Andersen (Canada),
M. Bataille (France), A. Berkane (Algeria), P. Bracken, B. Bradie, H. Chen, B. E. Davis, M.-C. Fan (China),
G. Fera (Italy), D. Fleischman, K. Gatesman, M. L. Glasser, H. Grandmontagne (France), G. C. Greubel,
E. A. Herman, N. Hodges (UK), F. Holland (Ireland), S. Kaczkowski, L. Kempeneers & J. Van Casteren (Bel-
gium), O. Kouba (Syria), O. P. Lossers (Netherlands), C. Maniquiz, K. Nelson, M. Omarjee (France), P. Per-
fetti (Italy), S. Sharma (India), A. Stadler (Switzerland), A. Stenger, S. M. Stewart (Saudi Arabia), R. Stong,
R. Tauraso (Italy), T. Wiandt, H. Widmer (Germany), M. Wildon (UK), Y. Zhang (China), Fejéntaláltuka
Szeged Problem Solving Group (Hungary), UM6P Math Club (Morocco), and the proposer.

CLASSICS
C23. Due to John H. Conway and Hallard T. Croft. Determine whether it is possible to
partition R3 into circles.

The Erdős–Mordell Inequality

C22. Due to Paul Erdős. Prove that from any point in any triangle, the sum of the distances
to the vertices of the triangle is at least twice as large as the sum of the distances to the
sides of the triangle.
Solution. Let P be a point in ABC, let a = BC, b = CA, c = AB, x = P A, y = P B,
and z = P C, and let d, e, and f be the distances from P to the sides BC, CA, and AB,
respectively. The area of ABC is given by (ad + be + cf )/2. Since f + z is no less
than the altitude of ABC dropped to base AB, the area of ABC is no greater than
c(f + z)/2. It follows that c(f + z) ≥ ad + be + cf , or cz ≥ ad + be.
We apply this inequality not to the original point P , but to the point in ABC that is
the reflection of P across the angle bisector from C. We obtain the same inequality but
with the roles of d and e reversed: cz ≥ ae + bd.
The analogous inequalities ax ≥ bf + ce and by ≥ cd + af are obtained in the same
way. We obtain
   
b c c a a b c b c a b a
x +y +z ≥ f + e+ d + f + e+ d = + d+ + e+ + f.
a a b b c c b c a c a b
By the AM-GM inequality, the three quantities in parentheses are all at least 2, and the
desired inequality x + y + z ≥ 2(d + e + f ) follows.
Editorial Comment: Equality holds if and only if ABC is equilateral and P is its center.
The problem appeared as problem 3740 [1935, 396; 1937, 252] in this Monthly, proposed
by Paul Erdős and solved by Louis Mordell and independently by David F. Barrow. It has
come to be known as the Erdős–Mordell inequality. Barrow’s solution proved the stronger
claim that the inequality holds even if d, e, and f are the distances from P to the points
where the angle bisectors meet the sides of the triangle.
Numerous alternative proofs and generalizations have appeared over the decades. For
example, a proof that is more elementary than those of Mordell and Barrow appears in
V. Komornik (1997), A short proof of the Erdős–Mordell theorem, this Monthly, 104(1):
57–60. Our proof here follows roughly that of C. Alsina and R. Nelsen (2007), A visual
proof of the Erdős–Mordell inequality, Forum Geom. 7: 99–102. Peter Walker (2016), The
Erdős–Mordell theorem in the exterior domain, Internat. J. Geom. 5(1): 31–38 examines
the extent to which the result generalizes to points outside a triangle.

272 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
SOLUTIONS

A Logarithmic Trigonometric Integral

12317 [2022, 385]. Proposed by Seán Stewart, King Abdullah University of Science and
Technology, Thuwal, Saudi Arabia. Prove
 π/2
sin(4x) ζ (3)
dx = −14 2 ,
0 log(tan x) π
∞
where ζ (3) is Apéry’s constant n=1 1/n3 .
Composite solution by Hongwei Chen, Christopher Newport University, Newport News,
VA, and Thomas Dickens, Houston, TX. Let I denote the requested integral. Using the

February 2024] PROBLEMS AND SOLUTIONS 171

identity sin(4x) = 2 sin(2x) cos(2x) = 4 sin x cos x(cos2 x − sin2 x) and the substitutions
u = tan x and t = u2 , we obtain
 π/2  ∞
sin(4x) 4u(1 − u2 )
I= dx = du
0 log(tan x) 0 (log u)(1 + u2 )3
 ∞
t −1
= −4 dt.
0 (log t)(1 + t)3
We now use parametric integration to compute this integral. First observe that for t > 0,
 1
t −1
t p dp = .
0 log t
Thus
 ∞  1  1  ∞ 
tp tp
I = −4 dp dt = −4 dt dp.
0 0 (1 + t)3 0 0 (1 + t)3
We evaluate the inner integral using the substitution s = 1/(1 + t) and then recognizing the
beta function. Using the identities B(z1 , z2 ) = (z1 )(z2 )/ (z1 + z2 ), (z + 1) = z(z),
and (z)(1 − z) = π/ sin(zπ ) yields
 ∞  1
tp
dt = s 1−p (1 − s)p ds = B(2 − p, p + 1)
0 (1 + t) 3
0
(2 − p)(p + 1) p(1 − p)(1 − p)(p) p(1 − p)π
= = = ,
(3) 2 2 sin(pπ )
for p ∈ (0, 1). Thus
 
1
p(1 − p) 2 π
x(π − x)
I = −2π dp = − 2 dx, (∗)
0 sin(pπ ) π 0 sin x
where we have used the substitution x = pπ in the second step. Applying the Fourier sine
series of x(π − x) on [0, π ] yields
∞
8 1  
x(π − x) = sin (2n + 1)x ,
π n=0 (2n + 1)3

and so
∞  π  
16  1 sin (2n + 1)x
I =− 3 dx.
π n=0 (2n + 1)3 0 sin x

We now use the formula

 π
sin(nx) 0, if n is even,
dx =
0 sin x π, if n is odd,

for n ∈ N, which follows from the fact that, for n ≥ 2,

 π  π    π  
sin(nx) sin (n − 2)x sin(nx) − sin (n − 2)x
dx − dx = dx
0 sin x 0 sin x 0 sin x
 π
 
=2 cos (n − 1)x dx = 0.
0

172 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
This yields
∞ ∞ ∞
16  π 16  1  1 16 7 ζ (3)
I =− =− 2 − =− · ζ (3) = −14 2 .
π n=0 (2n + 1)
3 3 π n=1
n3
n=1
(2n)3 2
π 8 π

Editorial comment. Most solvers converted the proposed integral into (∗) and then used
either the power series of csc x or a contour integral.
Also solved by A. Berkane (Algeria), N. Bhandari (Nepal), P. Bracken, B. Bradie, B. S. Burdick, B. E. Davis,
M.-C. Fan (China), G. Fera (Italy), M. L. Glasser (Spain), H. Grandmontagne (France), E. A. Herman,
N. Hodges (UK), W. Janous (Austria), O. Kouba (Syria), K.-W. Lau (China), O. P. Lossers (Netherlands),
M. Maniquiz, K. Nelson, M. Omarjee (France), P. Perfetti (Italy), A. Stadler (Switzerland), A. Stenger,
R. Stong, R. Tauraso (Italy), T. Wiandt, H. Widmer (Switzerland), Y. Zhang (China), Fejéntaláltuka Szeged
Problem Solving Group (Hungary), UM6P Math Club (Morocco), and the proposer.

Two Operator Norms

12318 [2022, 386]. Proposed by Mohammadhossein Mehrabi, University of Gothenburg,
Gothenburg, Sweden. Let a be a positive real number, and let Sa be the set of func-
a a
tions f : [−a, a] → R such that −a (f (x))2 dx = 1. Let A(f ) = −a f (x) dx, B(f ) =
a a
−a x f (x) dx, and C(f ) = −a x f (x) dx.
2

(a) What is sup A(f )2 + B(f )2 : f ∈ Sa ?

(b) What is sup A(f )2 + B(f )2 + C(f )2 : f ∈ Sa ?
Solution by Kenneth Andersen, University of Alberta, Edmonton, AB, Canada. Applying
the Gram–Schmidt process to the basis {1, x, x 2 , . . .} in L2 [−a, a], with inner product
a
f, g = −a f (x)g(x) dx, produces an orthonormal basis {e1 , e2 , . . .} with
 
1 3 −3/2 5 −5/2 2
e1 = √ , e2 = a x, and e3 = a (3x − a 2 ).
2a 2 8
 ∞ 2
For f ∈ Sa we have f = ∞
a
n=1 fn = f, f = −a (f (x)) dx = 1, where
2
n=1 fn en and
fn = en , f . Note that
 √  
√ 2 3/2 2 5/2 2
1 = 2ae1 , x = a e2 , and x = 2
a √ e3 + e1 ,
3 3 5
so
 a √
A(f ) = f (x) dx = 1, f = 2af1 ,
−a
 a

2 3/2
B(f ) = xf (x) dx = x, f =
a f2 ,
−a 3
 a √  
2 5/2 2
C(f ) = x f (x) dx = x , f =
2 2
a √ f3 + f1 .
−a 3 5
(a) For any f ∈ Sa we have
2a 3 2
A(f )2 + B(f )2 = 2af12 + f = [f1 f2 ] S [f1 f2 ]T ,
3 2
where
 
2a 0
S= .
0 2a 3 /3

February 2024] PROBLEMS AND SOLUTIONS 173

When computing the requested supremum, we may restrict attention to functions f with
f12 + f22 = 1 and fn = 0 for n ≥ 3. The supremum is the largest eigenvalue of S, so
√
 2a, if a ≤ 3,
sup A(f )2 + B(f )2 : f ∈ Sa = √
2a 3 /3, if a > 3.

(b) For any f ∈ Sa , we have

  √
2a 5 2a 3 2 8a 5 2 8 5a 5
A(f ) + B(f ) + C(f ) = 2a +
2 2 2
f1 +
2
f + f + f1 f3
9 3 2 45 3 45
= [f1 f2 f3 ] U [f1 f2 f3 ]T ,

where
⎛ √ ⎞
2a + 2a 5 /9 0 4 5a 5 /45
U =⎝ √ 0 2a 3 /3 0 ⎠.
4 5a 5 /45 0 8a 5 /45
As in part (a), the requested supremum is the largest eigenvalue of U . The eigenvalues of
U are
2a 3 a5 a 8
and a + ± 9a + 10a 4 + 225.
3 5 15
√ √ √
Since 5/a 2 + a 2 / 5 ≥ 2, we have a + a 5 /5 ≥ 2a 3 / 5 > 2a 3 /3. Therefore
 a5 a 8
sup A(f )2 + B(f )2 + C(f )2 : f ∈ Sa = a + + 9a + 10a 4 + 225.
5 15

Also solved by A. Berkane (Algeria), H. Chen (US), O. Kouba (Syria), B. Lai (China), J. H. Lindsey II,
P. W. Lindstrom, O. P. Lossers (Netherlands), M. Omarjee (France), K. Schilling, A. Stadler (Switzerland),
R. Stong, R. Tauraso (Italy), J. Yan (China), and the proposer. Part (a) also solved by E. A. Herman.

Counting Rectangles with Prime Area

12320 [2022, 386]. Proposed by Enrique Treviño, Lake Forest College, Lake Forest, IL.
Consider the grid of n2 lattice points {1, . . . , n}2 . Let S1 (n) be the number of rectangles
with corners in the grid (though not necessarily with horizontal and vertical sides) that have
area equal to a prime integer congruent to 1 (mod 4). Define S3 (n) similarly using primes
congruent to 3 (mod 4). Prove that there is a value n0 such that S1 (n) > S3 (n) for n ≥ n0 .
Solution by Nigel Hodges, Cheltenham, UK. We say that a rectangle is aligned if it has
horizontal and vertical sides; otherwise it is unaligned. First consider aligned rectangles.
When the area is a prime p, the sidelengths must be 1 and p. For p ≤ n − 1, there are
2(n − 1)(n − p) aligned rectangles with area p. Therefore the total number of aligned
rectangles with prime area is


n−1
2(n − 1)(n − p),
p=1
p prime

which is less than 2n3 . √

Now consider an unaligned rectangle with area p. Because each side has length a for
√
some integer a that is at least 2, unaligned rectangles must be squares of side-length p.

174 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
Since an integer congruent to 3 modulo 4 cannot be the sum of two squares, no unaligned
rectangles will contribute to S3 .
On the other hand, every prime p congruent to 1 mod 4 can be written as the sum of two
squares (uniquely up to order). Let p = k 2 + m2 be one such representation. The smallest
aligned square that contains an unaligned square of area p has side-length k + m, and such
an aligned square contains two unaligned squares of area p. There are (n − (k + m))2
aligned squares with side-length k + m within the grid, so there are at least 2(n − (k + m))2
unaligned rectangles of area p.
Restricting to p ≤ n2 /8, we have
 
k + m = 2p − (k − m)2 < 2p ≤ n/2.
Thus there are at least n2 /2 unaligned squares of area p in the grid. Therefore
 n2
S1 (n) ≥ π(n2 /8; 4, 1) · ,
2
where π(x; q, r) denotes the number of primes up to x that are congruent to r mod q.
By the prime number theorem for arithmetic progressions (see, for instance, H. Davenport
(1980), Multiplicative Number Theory, 2nd ed., Springer-Verlag, Berlin, Ch. 20),
x x
π(x; 4, 1) ∼ = .
φ(4) ln x 2 ln x
Therefore, for ε ∈ (0, 1) and sufficiently large n,
ε n2 /8 n2 ε n4
S1 (n) ≥ 2
· = > 2n3 > S3 (n).
2 ln(n /8) 2 32 ln(n2 /8)

Also solved by N. Caro-Montoya (Brazil), N. Fellini (Canada), D. Fleischman, K. Gatesman, A. Stadler

(Switzerland), R. Tauraso (Italy), and the proposer.

Primes are Rarely Squares Modulo Squares

12321 [2022, 486]. Proposed by Mohammadamin Sharifi, Sharif University of Technology,
Tehran, Iran. Let p be a prime number. Prove that the number of perfect squares m such
that the least nonnegative remainder of p (mod m) is a perfect square is less than 2p1/3 .
Solution by Richard Stong, Center for Communications Research, San Diego, CA. Writ-
ing m = b2 and letting a 2 be the least nonnegative residue of p (mod m), we can write
p = kb2 + a 2 for integers a, b, k with b > a ≥ 0 and k > 0 (since p is not a square).
Conversely, any such expression gives a solution m = b2 . We prove the following:
Claim. For each positive integer k, there is at most one pair (a, b) with b > a ≥ 0 such
that p = kb2 + a 2 .
Proof: Suppose that (a, b) and (c, d) are two such pairs. Since b2 determines a, we must
have d = b. We write
0 = (b2 − d 2 )p = b2 (kd 2 + c2 ) − d 2 (kb2 + a 2 ) = b2 c2 − a 2 d 2 = (bc + ad)(bc − ad).
Hence bc + ad or bc − ad is a nonzero multiple of p. Since b/a > 1 > c/d, the vectors
(b, a) and (c, d) are not parallel. By the Cauchy–Schwarz inequality (with strict inequality
for nonparallel vectors),
0 < |bc − ad| ≤ bc + ad < (b2 + a 2 )1/2 (c2 + d 2 )1/2 ≤ p,
so neither factor is a multiple of p. This contradiction yields the claim. 

February 2024] PROBLEMS AND SOLUTIONS 175

 Given p = kb2 + a 2 , either k ≤ p1/3 or b2 ≤ p2/3 , since otherwise kb2 > p. Since k
determines the solution (by the claim), the first case gives at most p1/3  solutions. In the
second case, since b determines a and hence determines the solution, we also have at most
p1/3  solutions. Therefore the total number of solutions is at most 2p1/3 .
Also solved by N. Hodges (UK), O. P. Lossers (Netherlands), R. Tauraso (Italy), and the proposer.

A Skew-Symmetric Determinant
12322 [2022, 486]. Proposed by Askar Dzhumadil’daev, Kazakh-British Technical Univer-
sity, Almaty, Kazakhstan. Given real numbers x1 , . . . , x2n , let A be the skew-symmetric
2n-by-2n matrix with entries ai,j = (xi − xj )2 for 1 ≤ i < j ≤ 2n. Prove
 2
det(A) = 4n−1 (x1 − x2 )(x2 − x3 ) · · · (x2n−1 − x2n )(x2n − x1 ) .

Solution by Kuldeep Sarma, Tezpur University, Tezpur, India. Let p(x1 , . . . , x2n ) be the
desired determinant as a polynomial in x1 , . . . , x2n . It is a homogeneous polynomial of
degree 4n. We claim (xk − xk+1 )2 | p for 1 ≤ k < 2n and (x2n − x1 )2 | p, which implies

(x1 − x2 )2 (x2 − x3 )2 · · · (x2n−1 − x2n )2 (x2n − x1 )2 | p.

Noting the degree of p, we conclude that p is a scalar multiple of the desired polynomial.
To prove the claim for k = 1, fix arbitrary real numbers x2 , . . . , x2n and let x1 = x2 + ,
where may vary. It suffices to show that p(x2 + , x2 , x3 , . . . , x2n ) = O( 2 ). With this
expression for x1 , the matrix A becomes
⎡ ⎤
0 2
(x2 + − x3 )2 (x2 + − x4 )2 · · ·
⎢ − 2 0 (x2 − x3 )2 (x2 − x4 )2 · · ·⎥
⎢ ⎥
⎢−(x2 + − x3 )2 −(x2 − x3 )2 · · · · ·⎥
⎢ ⎥.
⎢−(x2 + − x4 )2 −(x2 − x4 )2 · · · · ·⎥
⎣ ⎦
.. .. .. .. ..
. . . . .

Subtracting the second row from the first and then the second column from the first (which
does not change the determinant) gives
⎡ ⎤
0 2
2(x2 − x3 ) + 2 2(x2 − x4 ) + 2 · · ·
⎢ − 2 0 (x2 − x3 )2 (x2 − x4 )2 · · ·⎥
⎢ ⎥
⎢−2(x2 − x3 ) − 2
−(x − x ) 2
· · · · ·⎥
⎢ 2 3 ⎥.
⎢−2(x2 − x4 ) − 2 −(x2 − x4 )2 · · · · ·⎥
⎣ ⎦
.. .. .. .. ..
. . . . .

Factors of can now be taken out from the first row and the first column. This yields
p(x2 + , x2 , . . . , x2n ) = O( 2 ) and thus (x1 − x2 )2 | p.
The same argument works when 1 < k < 2n: set xk = xk+1 + and perform the oper-
ations on the kth and (k + 1)th rows and columns instead of the first and second. For
(x2n − x1 )2 | p, set x2n = x1 + and add the 2nth row to the first and the 2nth column to
the first; the rest of the argument is identical.
It remains only to find the scalar coefficient. We do this by evaluating p(x1 , . . . , x2n )
when xk = (−1)k−1 /2. In this case,

(x1 − x2 )2 (x2 − x3 )2 · · · (x2n−1 − x2n )2 (x2n − x1 )2 = 1,

so we need to show that the determinant of A is 4n−1 . Changing the order of the rows and
columns to put the odd-indexed rows and columns in order before the even-indexed rows

176 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
and columns does not change the sign of the determinant and yields
 
 0 Bn 
det A =  ,
−BnT 0 
where Bn is the n-by-n matrix with all entries 1 on and above the main diagonal and all
entries −1 below the main diagonal. Thus det A = (det Bn )2 .
We prove by induction on n that det Bn = 2n−1 . Note det B1 = 1. For n > 1, subtracting
the second row of Bn from the first gives the block matrix
 
2 0
,
−1 Bn−1
where the second block of rows or columns has length n − 1. Thus det Bn = 2 det Bn−1 =
2n−1 . Hence det A = 4n−1 , and
p(x1 , . . . , x2n ) = 4n−1 (x1 − x2 )2 (x2 − x3 )2 · · · (x2n−1 − x2n )2 (x2n − x1 )2 ,
as desired.
Also solved by N. Caro-Montoya (Brazil), H. Chen (US), D. Fleischman, J.-P. Grivaux (France), N. Hodges
(UK), O. Kouba (Syria), P. Lalonde (Canada), O. P. Lossers (Netherlands), M. Omarjee (France), C. R. Prane-
sachar (India), R. Stong, R. Tauraso (Italy), Fejéntaláltuka Szeged Problem Solving Group (Hungary), and the
proposer.

Beyond Bell Numbers

12323 [2022, 486]. Proposed by Erik Vigren, Swedish Institute of Space Physics, Uppsala,
Sweden, and Andreas Dieckmann, Physikalisches Institut der Universität Bonn, Bonn, Ger-
many.
(a) Find integers c0 , c1 , and c2 such that
∞ ∞
k 11 c0 + c1 k + c2 k 2
= .
k=0
(k!)3 k=0
(k!)3

(b) Prove that for any integers n and b with 1 ≤ b ≤ n, there are integers cm for 0 ≤ m ≤
b − 1 such that
∞ ∞
1 
b−1
kn
b
= b
cm k m .
k=0
(k!) k=0
(k!) m=0

(c) Prove that the integers cm from part (b) are unique.
Solution by Kenneth Schilling, University of Michigan, Flint, MI. We first prove the claim
in (b). For 1 ≤ b ≤ n, we construct a sequence {pi } of polynomials as follows: let p0 (x) =
x n , and if pi (x) = dm=0 am x m , then let

b−1 
d
pi+1 (x) = am x m + am (x + 1)m−b .
m=0 m=b

An easy induction shows that the polynomials pi (x) have integer coefficients. Since
degree(pi+1 ) ≤ max{b − 1, degree(pi ) − b}, the sequence of degrees of pi decreases until
we reach the first polynomial pr with degree less than b (at which point the sequence
repeats pr indefinitely). For m ≥ b we have
∞ ∞ ∞
km k m−b (k + 1)m−b
= = ,
k=0
(k!)b k=1
((k − 1)!)b k=0
(k!)b

February 2024] PROBLEMS AND SOLUTIONS 177

and it follows that for all i,
∞
 ∞

pi+1 (k) pi (k)
= .
k=0
(k!)b k=0
(k!)b

Hence
∞ ∞
kn pr (k)
b
=
k=0
(k!) k=0
(k!)b

and the required integers c0 , . . . , cb−1 are the coefficients of pr (x).

For part (a), following the procedure above, we have
p0 (x) = x 11 ,
p1 (x) = (x + 1)8 = 1 + 8x + 28x 2 + 56x 3 + 70x 4 + 56x 5 + 28x 6 + 8x 7 + x 8 ,
p2 (x) = 1 + 8x + 28x 2 + 56 + 70(x + 1) + 56(x + 1)2 + 28(x + 1)3 +
8(x + 1)4 + (x + 1)5 = 220 + 311x + 226x 2 + 70x 3 + 13x 4 + x 5 ,
p3 (x) = 220 + 311x + 226x 2 + 70 + 13(x + 1) + (x + 1)2 = 304 + 326x + 227x 2 .
Hence the required integers are given by c0 = 304, c1 = 326, and c2 = 227.
For part (c), suppose the contrary. Taking the difference of two distinct solutions, we
get
∞a nonzero polynomial p(x) with integer coefficients and degree at most b − 1 such that
k=0 p(k)/(k!) b
= 0. Assume without loss of generality that the leading coefficient of p
is positive. For sufficiently large N we have p(k) > 0 for all k > N , and it follows that in
the equation
∞ N
p(k) p(k)
(N !) b
b
= −(N !) b
b
,
k=N+1
(k!) k=0
(k!)

the left side is positive and the right side is an integer. Hence this quantity is a positive
integer and in particular it is at least 1.
Let C be the sum of the absolute values of all the coefficients of p. For x ≥ 1 we have
p(x) ≤ Cx b−1 and hence for every positive integer k,
p(N + k) ≤ C(N + k)b−1 ≤ C(kN + k)b−1 = Ck b−1 (N + 1)b−1 .
We also have
N! 1 1
= ≤ .
(N + k)! (N + 1) · · · (N + k) (N + 1)k!
Thus
∞ ∞
p(k) p(N + k)(N !)b
1 ≤ (N !)b =
k=N+1
(k!) b
k=1
((N + k)!)b

∞ ∞
Ck b−1 (N + 1)b−1 C  1
≤ ≤ →0 as N → ∞,
k=1
(N + 1)b (k!)b N + 1 k=1 k!

so we have a contradiction.
Editorial comment. The problem can be viewed as saying that for all n ≥ 0 and b ≥ 1,
there is a unique polynomial Pn,b (x) with integer coefficients and of degree at most b − 1
such that
∞ ∞
kn Pn,b (k)
b
= .
k=0
(k!) k=0
(k!)b

178 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
The polynomial Pn,1 is constant and equal to the Bell number Bn , the number of partitions
of the set {1, . . . , n}. Hence the polynomials Pn,k can be viewed as a generalization of the
Bell numbers. The solution above gives a recurrence for these polynomials, and the pro-
posers have shown that a similar recurrence holds when the factorials are replaced by mul-
tifactorials (see www-elsa.physik.uni-bonn.de/ dieckman/InfProd/InfProd.html#Sumsxinv
olvingxreciprocalxmultifactorialsxorxfactorials).
One can also give a summation formula for Pn,b analogous to the formula for the Bell
numbers as a sum of Stirling numbers of the second kind. Specifically, let hk (x0 , . . . , xm )
denote the complete homogeneous symmetric polynomial of degree k and define gener-
alized Stirling numbers Sb (n, m) to be hn−m evaluated at xi = i/b for 0 ≤ i ≤ m. For
b = 1 these are the Stirling numbers of the second kind. One can prove the polynomial
identity
n 
m−1
Xn = Sb (n, m) (X − i/b),
m=0 i=0

from which it follows that 

n
Pn,b (x) = Sb (n, m)x (m mod b) .
m=0

Also solved by N. Hodges (UK), O. Kouba (Syria), K.-W. Lau (China), O. P. Lossers (Netherlands), M. Omar-
jee (France), C. R. Pranesachar (India), R. Tauraso (Italy), Eagle Problem Solvers, and the proposer. Parts (a)
and (b) were solved by P. Bracken, H. Chen (US), N. Grivaux (France), E. A. Herman, W. Janous (Austria),
and D. Terr.

A Symmetrical Integral
12324 [2022, 486]. Proposed by Albert Stadler, Herrliberg, Switzerland. Let a and b be
positive real numbers. Prove
 ∞  ∞
1 1
 dx =  dx.
0 ax + 2(2b − a)x + a
4 2 0 bx + 2(2a − b)x 2 + b
4

Solution by Giuseppe Fera, Vicenza, Italy. Let

 ∞
1
f (a, b) =  dx.
0 ax 4 + 2(2b − a)x 2 + a
Splitting the integral at 1 and then making the change of variables y = 1/x in the second
integral, we get
 1  ∞
1 1
f (a, b) =  dx +  dx
0 ax + 2(2b − a)x + a
4 2 1 ax + 2(2b − a)x 2 + a
4

 1  1
1 1
=  dx +  dy
0 ax 4 + 2(2b − a)x 2 + a 0 ay 4 + 2(2b − a)y 2 + a
 1
1
=2  dx. (∗)
0 ax + 2(2b − a)x 2 + a
4

Substituting x = (1 − u)/(1 + u) and dx = −2 du/(1 + u)2 , we obtain

 1
1
f (a, b) = 2  du = f (b, a),
0 bu4 + 2(2a − b)u2 + b
which completes the proof.

February 2024] PROBLEMS AND SOLUTIONS 179

Editorial comment. Many solvers used the substitution x = tan(θ/2) in (∗) to get
 π/2
dθ
f (a, b) =  ,
0 a cos2 θ + b sin2 θ
which also easily yields the desired symmetric property with respect to a and b.
Also solved by K. F. Andersen (Canada), M. Bataille (France), A. Berkane (Algeria), P. Bracken, H. Chen
(US), K. Gatesman, M. L. Glasser, D. Henderson, O. Kouba (Syria), B. Lai (China) & R. Wang (China),
S. Lee, O. P. Lossers (Netherlands), J. Magliano, F. Masroor, M. Omarjee (France), C. R. Pranesachar (India),
V. Schindler (Germany), S. M. Stewart (Saudi Arabia), R. Stong, R. Tauraso (Italy), T. Wiandt, H. Widmer
(Switzerland), L. Zhou, UM6P Math Club (Morocco), and the proposer.

CLASSICS
C22. Due to Paul Erdős. Prove that from any point in any triangle, the sum of the distances
to the vertices of the triangle is at least twice as large as the sum of the distances to the
sides of the triangle.

The Infamous Pentagon Problem

C21. From the 1986 International Mathematical Olympiad. An integer is assigned to each
vertex of a regular pentagon in such a way that the sum of the five integers is positive. If
three consecutive vertices are assigned the numbers x, y, z in order, and y is negative, then
one may replace x, y, and z by x + y, −y, and z + y, respectively. Such an operation is
performed repeatedly as long as at least one of the five numbers is negative. Determine
whether this procedure necessarily comes to an end after a finite number of steps.
Solution. The procedure must end. To see this, let

F (a, b, c, d, e) = (a − c)2 + (b − d)2 + (c − e)2 + (d − a)2 + (e − b)2 .

When the integers a, b, c, d, and e are assigned to the pentagon, we call F (a, b, c, d, e)
the score of that assignment. The score of any assignment is nonnegative, but the effect of
the replacement move is to lower the score, since, with c < 0, we have

F (a,b, c, d, e) − F (a, b + c, −c, d + c, e)

= (a − c)2 + (b − d)2 + (c − e)2 + (d − a)2 + (e − b)2
 
− (a + c) + (b − d) + (−c − e) + (d + c − a) + (e − b − c)
2 2 2 2 2

= −2c(a + b + c + d + e),

which is positive. Since there is no infinite strictly decreasing sequence of nonnegative

integers, no infinite sequence of replacement moves is possible.
Editorial Comment: Although the solution above appears simple, this was the hardest prob-
lem on the 1986 International Mathematical Olympiad, because the function F is difficult
to find. The problem has seen many incarnations and generalizations, and it has spawned
many papers in research journals. The solution in P. Winkler (2003) Mathematical Puz-
zles: A Connoisseur’s Collection, A K Peters, is attributed to B. Chazelle. It allows for the
generalization to an n-gon for any n and further yields the surprising conclusion that the
number of replacement moves until no negative integers remain and the final assignment
of these integers is independent of which sequence of moves is selected.

180 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
SOLUTIONS

Minimizing an Integral
12308 [2022, 285]. Proposed by Cezar Lupu, Yanqi Lake BIMSA and Tsinghua Univer-
1 2
sity, Beijing, China. What is the minimum value of 0 f  (x) dx over all continuously
1 1
differentiable functions f : [0, 1] → R such that 0 f (x) dx = 0 x 2 f (x) dx = 1?
Solution by Raymond Mortini, University of Luxembourg, Esch-sur-Alzette, Luxembourg.
The minimum value is 105/2.
By the Cauchy–Schwarz inequality,
 1 2  1  1
f  (x)(x 3 − x) dx ≤ (f  (x))2 dx (x 3 − x)2 dx. (∗)
0 0 0

Integrating by parts in the integral on the left in (∗), we obtain

 1 1
 1
f  (x)(x 3 − x) dx = f (x)(x 3 − x) − f (x)(3x 2 − 1) dx
0 0 0
 1  1
= −3 x 2 f (x) dx + f (x) dx = −3 + 1 = −2.
0 0

Moreover,
 1  1
8
(x − x) dx =
3 2
(x 6 − 2x 4 + x 2 ) dx = .
0 0 105
Therefore (∗) implies
 1
4 105
(f  (x))2 dx ≥ = .
0 8/105 2

January 2024] PROBLEMS AND SOLUTIONS 77

 Equality holds in (∗) when f  (x) is a scalar multiple of x 3 − x, or equivalently when
ax 4 ax 2
− +c f (x) =
4 2
1 1
for some real numbers a and c. The constraint 0 f (x) dx = 0 x 2 f (x) dx = 1 leads to
the values a = −105/4 and c = −33/16. Thus the minimum value 105/2 is attained when
f (x) = −(105/16)x 4 + (105/8)x 2 − 33/16.
Also solved by R. A. Agnew, K. F. Andersen (Canada), M. Bataille (France), A. Berkane (Algeria), P. Bracken,
H. Chen (US), C. Chiser (Romania), P. J. Fitzsimmons, K. Gatesman, L. Han, K. T. L. Koo (China), O. Kouba
(Syria), B. Lai & R. Wang (China), K.-W. Lau (China), J. H. Lindsey II, P. W. Lindstrom, O. P. Lossers
(Netherlands), R. Nandan, M. Omarjee (France), P. Perfetti (Italy), A. D. Pı̂rvuceanu (Romania), K. Schilling,
A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), E. I. Verriest, F. Visescu (Romania), J. Vukmirović
(Serbia), T. Wiandt, J. Yan (China), L. Zhou, UM6P Math Club (Morocco), and the proposer.

Two Inequalities Involving Power Means

12311 [2022, 286]. Proposed by Hideyuki Ohtsuka, Saitama, Japan. Let m and n be posi-
tive integers, and let r, x1 , x2 , . . . , xn be positive real numbers.


m n r n (m+1
2 )
1 j 1
(a) Prove xk ≥ xkr when r ≤ m/2.
j =0
n k=1
n k=1


m n r n (m+1
2 )
1 j 1
(b) Prove xk ≤ xkr when r ≥ m.
j =0
n k=1
n k=1

Solution by Faraz Masroor, New York, NY. For t > 0, let

n 1/t
1
St = xkt ,
n k=1

which is the power mean with exponent t of the numbers x1 , . . . , xn . By the power mean
inequality, Sa ≤ Sb when 0 < a ≤ b.
(a) Suppose r ≤ m/2. By the Cauchy–Schwarz inequality, for any i and j in {0, . . . , m},
n n n 2
1 1 j 1 (i+j )/2
xki · xk ≥ xk .
n k=1
n k=1
n k=1

Therefore
r r/2 r/2

m
1
n
j

m
1
n
j

m
1
n
m−j
xk = xk · xk
j =0
n k=1 j =0
n k=1 j =0
n k=1

r/2 r

m
1
n
j 1
n
m−j

m
1
n
m/2
= xk · xk ≥ xk
j =0
n k=1
n k=1 j =0
n k=1


m n (m+1
2 )
1
= (Sm/2 )mr/2 ≥ (Sr )(m+1)mr/2 = xkr .
j =0
n k=1

(b) Suppose r ≥ m. Since the j = 0 term in the product is 1, we have


m n r

m 
m n (m+1
2 )
1 j r m j 1
xk = (Sj )j r ≤ (Sr )j r = (Sr ) j =1 = xkr .
j =0
n k=1 j =1 j =1
n k=1

78 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
Editorial comment. Some solvers observed that the inequality in (a) actually holds for
r ≤ (m + 1)/2. This can be proven by letting j run from 1 to m in the product (as in the
m−j m+1−j
solution to (b)) and then changing xk in the solution above to xk , with appropriate
modifications in the later steps.
Also solved by K. F. Andersen (Canada), M. Bataille (France), O. Geupel (Germany), O. Kouba (Syria),
H. Kwong, O. P. Lossers (Netherlands), A. Mhanna (Lebanon), P. Perfetti (Italy), M. Reid, R. Stong, R. Tauraso
(Italy), L. Zhou, and the proposer.

An Above Average Function

12312 [2022, 286]. Proposed by Martin Tchernookov, University of Wisconsin, Whitewater,
WI. Find all continuous functions f : [0, ∞) → R such that, for all positive x,
  
1 x  2
f (x) f (x) − f (t) dt ≥ f (x) − 1 .
x 0

Solution by Edward Schmeichel, San Jose State University, San Jose, CA. Clearly the con-
stant function defined by f (x) = 1 satisfies the given inequality. We show that it is the
only continuous function that does so.
Let f be a continuous function satisfying the inequality. For x ≥ 0, let
 x
1
f (t) dt, if x > 0,
A(x) = x 0
f (0), if x = 0.

Note that A is continuous from the right at 0. Letting x → 0+ in the given inequality, we
obtain 0 = f (0)(f (0) − f (0)) ≥ (f (0) − 1)2 , and therefore f (0) = 1.
Since f (x) = 0 for any x > 0 gives a contradiction, the intermediate value theorem
implies f (x) > 0 for all x ∈ [0, ∞). If follows that f (x) − A(x) ≥ (f (x) − 1)2 /f (x) ≥ 0
and hence f (x) ≥ A(x). Thus A (x) = (f (x) − A(x))/x ≥ 0 for all x > 0, so A(x) is
nondecreasing, and we obtain f (x) ≥ A(x) ≥ A(0) = 1 for all x ≥ 0.
The given inequality can be rearranged to read f (x)(2 − A(x)) ≥ 1, so A(x) < 2. Thus
A(x) is both nondecreasing and bounded above, so as x tends to infinity, A(x) approaches
a limit L from below, where 1 ≤ L ≤ 2. If L = 1, then A(x) = 1 and hence f (x) = 1 for
all x, and we are done. Thus we may assume L > 1.
Say L = 1 + , where 0 <  ≤ 1. Let a be any number with 1/(1 + ) < a < 1, and
choose b large enough that A(x) ≥ 1 + a for x ≥ b. For x ≥ b,
1 1
f (x) ≥ ≥ ,
2 − A(x) 1 − a
and therefore
 
1 x
1 x
1 x−b
A(x) = f (t) dt ≥ dt = .
x 0 x b 1 − a x(1 − a)
It follows that
x−b 1 1
L = lim A(x) ≥ lim = > = 1 +  = L,
x→∞ x→∞ x(1 − a) 1 − a 1 − /(1 + )
a contradiction.
Also solved by K. F. Andersen (Canada), P. Bracken, J. Boswell & C. Curtis, H. Chen (China), C. Chiser
(Romania), P. J. Fitzsimmons, L. Han, D. A. Hejhal, D. Henderson, E. A. Herman, G. Herzog (Germany) &
R. Mortini (France), N. Hodges (UK), K.-W. Lau (China), O. P. Lossers (Netherlands), M. Omarjee (France),
L. J. Peterson, A. Sinha (India), R. Stong, R. Tauraso (Italy), J. Vukmirović (Serbia), J. Yan (China), and the
proposer.

January 2024] PROBLEMS AND SOLUTIONS 79

 Trees with Pairwise Isomorphic Subtrees
12313 [2022, 286]. Proposed by Douglas B. West, University of Illinois, Urbana IL. For all
n ∈ N, determine all n-vertex trees having the property that the connected (n − 2)-vertex
subgraphs that can be obtained by deleting two vertices are pairwise isomorphic.
Solution by O. P. Lossers, Eindhoven University of Technology, Netherlands and Eötvös
Loránd University, Hungary. For all n, the trees with this property are the star and the path,
plus when n = 5 the one tree that is not a star or path. In each of these cases, the subtrees
all are stars or all are paths.
For n ≥ 6, let T be an n-vertex tree, and let  be the number of leaves of T . If  = 2,
then T is a path. If  = 3, then T consists of three paths with a common endpoint, and any
tree with three leaves is determined by the multiset of lengths of those paths. Let a, b, and c
be their lengths in T , with a ≥ b ≥ c ≥ 1. Note that a + b + c = n − 1 ≥ 5. If c ≤ 2, then
T has both paths and non-paths as connected (n − 2)-vertex subgraphs. If c ≥ 3, then the
sets {a, b, c − 2} and {a − 1, b − 1, c} are different and determine nonisomorphic subtrees.
Hence we may assume  ≥ 4. The diameter of a tree is the maximum length of its paths,
and any longest path connects two leaves. Let d be the diameter of T . Deleting two leaves
outside a longest path produces a subtree having diameter d. If some two leaves together
cover all longest paths, then deleting them produces a subtree with smaller diameter. Hence
we may assume that T has no such pair of leaves, which implies that every connected
subgraph of T with n − 2 vertices has diameter d.
If d is odd, then T has a central edge e belonging to all paths of length d. Since all con-
nected subgraphs with n − 2 vertices have diameter d, both components of T − e contain
at least two leaves of T . The tree obtained by deleting any two leaves still has diameter d
and the same central edge e. Hence a subtree obtained by deleting two of the leaves of T
from the smallest component of T − e is not isomorphic to a subtree obtained by delet-
ing one leaf of T from each component of T − e; in these two subtrees the sizes of the
components obtained by deleting the central edge e are different.
The argument for even d with d ≥ 4 is similar. In this case the tree has a unique central
vertex z at the middle of every longest path. Let k be the degree of z. Note that k ≥ 2, and
k is the number of components of T − z. Let n1 , . . . , nk be the numbers of vertices in the
components of T − z, in nonincreasing order. Every connected subgraph of T with n − 2
vertices has diameter d and the same central vertex z. If nk ≤ 2, then T has (n − 2)-vertex
subtrees whose central vertices have different degrees. If nk ≥ 3, then a subtree obtained
by deleting two vertices from a smallest component of T − z is not isomorphic to a subtree
obtained by deleting one vertex each from two largest components of T − z.
Editorial comment. Motivated by this problem, Stan Wagon conjectured that among all
graphs, the graphs whose connected subgraphs obtained by deleting two vertices are pair-
wise isomorphic are the stars, paths, cycles, complete graphs, and five other graphs with at
most five vertices. Wagon’s conjecture was proved by the proposer.
Also solved by K. Gatesman, O. Geupel (Germany), Y. Ionin, R. Stong, R. Tauraso (Italy), Texas State Problem
Solvers, and the proposer.

Rotating Devices
12314 [2022, 385]. Proposed by Gregory Galperin, Eastern Illinois University, Charleston,
IL, and Yury J. Ionin, Central Michigan University, Mount Pleasant, MI. Let n, m, and k
be positive integers with k ≤ n − 1. Consider n devices each of which can be in any of m
states denoted 0, 1, . . . , m − 1. A move consists of selecting a set of k devices and adding
1 (mod m) to each of their states. Prove that for any n, m, k as specified and any initial

80 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
states of the n devices, there exists a sequence of moves that leaves each device in the state
0 or 1.
Solution by the Eagle Problem Solvers, Georgia Southern University, Statesboro, GA and
Savannah, GA. Fixing k, we prove the claim by induction on n − k. For the induction step,
suppose n > k + 1. Let (a1 , . . . , an ) be the initial states. First, add 1 modulo m to the first
k devices until the state of the first device reaches 0. Next apply the induction hypothesis
to the last n − 1 devices, leaving a1 unchanged, to bring all those devices to state 0 or 1.
Thus it suffices to prove the claim when n = k + 1. For this, we begin with a lemma.
Lemma. Given any distinct indices i and j , moves can be made that result in replacing
ai with ai − s, replacing aj with aj + s, and leaving all other states unchanged, for any s
with 1 ≤ s ≤ m − 1.
Proof. By symmetry, we may assume i = 1 and j = k + 1. Add 1 to the last k devices s
times, and add 1 to the first k devices m − s times. 
Given initial states (a1 , . . . , ak+1 ), using the lemma k times with (i, j, s) = (i, k + 1, ai )
for 1 ≤ i ≤ k brings the first k devices to 0 while accumulating k+1 i=1 ai in position k + 1.
That is, we obtain (0, . . . , 0, ), where  ≡ k+1 i=1 ia (mod m) and 0 ≤  ≤ m − 1.
If  < k, then we apply the lemma  times with (i, j, s) = (k + 1, j, 1) for 1 ≤ j ≤ 
to obtain (1, . . . , 1, 0, . . . , 0), with 1 in the first  devices and 0 in the others.
If  ≥ k, then write  = qk + r for integers q and r with 0 ≤ r < k. Using the lemma
k times with (i, j, s) = (k + 1, j, 1) for 1 ≤ j ≤ k yields (1, . . . , 1,  − k). Now adding 1
to the first k devices m − 1 times produces (0, . . . , 0,  − k).
Repeating this process q times brings the states to (0, . . . , 0, r), where 0 ≤ r < k. The
argument in the case  < k then allows us to reach (1, . . . , 1, 0, . . . , 0), with 1 in the first
r devices and 0 in the others.
Also solved by J. Boswell & C. Curtis, B. S. Burdick, N. Caro-Montoya (Brazil), H. Chen (China), P. Corn,
K. Gatesman, A. Goel, E. A. Herman, N. Hodges (UK), O. P. Lossers (Netherlands), A. Mandal (India),
F. Masroor, G. Raduns, T. Song, A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), O. Zhang, and the
proposer.

Determinants and Rooted Trees

12315 [2022, 385]. Proposed by Mikael P. Sundqvist and Victor Ufnarovski, Lund Uni-
versity, Sölvegatan, Sweden. Suppose ai,j ∈ [0, 1] for 1 ≤ i ≤ n and 1 ≤ j ≤ n. Let
B be the n-by-n matrix with i, j -entry bi,j defined by bi,j = ai,j when j = i − 1 and
bi,j = − nk=1 ai,k when j = i − 1.
(a) Evaluate det(B) in the case where ai,j = 1 for all i and j .
(b) Show that the value in part (a) is the maximum possible value of det(B).
(c) Show that det(B) ≥ 0 in all cases.
Solution I by Richard Stong, Center for Communications Research, San Diego, CA.
(a) We denote by A the matrix with i, j -entry ai,j . The matrix B in this case has a 1 as
every entry except just below the main diagonal, where the entries are −n. Subtracting the
top row from each other row yields a matrix whose only nonzero entry in row i for i > 1 is
−n − 1 in column i − 1. The cofactor expansion of the determinant along column n yields
det(B) = (−1)n+1 (−n − 1)n−1 = (n + 1)n−1 .
(b) and (c) Clearly, det(B) is a homogeneous
 polynomial of degree n in the values ai,j ,
where each monomial has the form ni=1 ai,f (i) for some function f : [n] → [n] (gener-
ally not a permutation). We prove a stronger version of (c) that implies (b), namely that the
coefficient of each such monomial is nonnegative. Increasing all entries of A to 1 therefore
maximizes each monomial and the value of det(B). The coefficient of a particular mono-
mial is the value of det(B) in the case where the corresponding n entries of A equal 1 and

January 2024] PROBLEMS AND SOLUTIONS 81

all other entries are 0. These n entries are one from each row of A. Thus it suffices to prove
(c) in the special case where A has a single 1 in each row and the rest of A is 0. This is the
statement proved in Lemma 2, since the corresponding matrix B always has −1 in each
subdiagonal entry.
Lemma 1. If M is an n-by-n matrix with all entries 0 except for diagonal entries 1 and at
most one entry in each row equal to −1, then det(M) ≥ 0.
Proof. We proceed by induction on n, with trivial base case n = 1. If any row of M lacks
a −1, then the expansion of det(M) along that row reduces to the case for n − 1 and the
induction hypothesis suffices. If every row of M has a −1, then the row sums of M are all
zero, so det(M) = 0. 
Lemma 2. If B is an n-by-n matrix with all entries 0 except for subdiagonal entries −1
and at most one entry equal to 1 in each row, then det(B) ≥ 0.
Proof. We again proceed by induction on n, with trivial base case n = 1. If B has no 1 in
the top row, then already det(B) = 0.
If b1,j = 1 with j < n, then obtain B  from B by deleting row 1 and column j . Expand-
ing along row 1 yields det(B) = (−1)j +1 det(B  ). Obtain B  from B  by moving row j
of B  to the top, so det(B  ) = (−1)j −1 det(B  ) = det(B). Since the −1 in row j + 1 of
B was deleted before moving the row, B  satisfies the conditions of Lemma 2, and hence
det(B  ) ≥ 0 by the induction hypothesis.
If b1,n = 1, then move row 1 to the bottom, introducing a factor of (−1)n−1 , and negate
the top n − 1 rows, introducing a second factor of (−1)n−1 . The resulting matrix M has
the form described in Lemma 1, so det(B) = det(M) ≥ 0, as desired. 
Solution II by Pierre Lalonde, Plessisville, QC, Canada. We express det(B) as the sum
of the weights of certain rooted spanning trees in a directed graph. To do this, we locate
−B as a submatrix in a particular matrix obtained from a weighted digraph with n + 1
vertices. Let G be the directed graph on vertices v1 , . . . , vn+1 in which every ordered pair
of distinct vertices is an edge. Define the weight of edge vi vj with i = j as follows: let
w(vi v1 ) = ai,i−1 for 2 ≤ i ≤ n, let w(vi vj ) = ai,j −1 for 1 ≤ i ≤ n and 2 ≤ j ≤ n + 1,
and let w(vn+1 vj ) = 0 for 1 ≤ j ≤ n.
The Laplacian matrix L(G) of G is defined by letting the entry in position (i, j ) be
−w(vi vj ) when i = j and in position (i, i) be j =i w(vi vj ). Thus
⎡ n ⎤
⎢ j =1 a 1,j −a 1,1 −a 1,2 · · · −a 1,n−1 −a 1,n ⎥
⎢ ⎥
⎢ n ⎥
⎢ −a2,1 a2,j −a2,2 · · · −a2,n−1 −a2,n ⎥
⎢ ⎥
⎢ j =1 ⎥
⎢ n ⎥
⎢ −a −a3,1 a3,j · · · −a3,n−1 −a3,n ⎥
L(G) = ⎢ ⎢
3,2 ⎥.
⎥
j =1
⎢ ⎥
⎢ .. .. .. .. .. .. ⎥
⎢ . . . . . . ⎥
⎢ ⎥
⎢ n ⎥
⎢ −an,n−1 −an,1 −an,2 · · · an,j −an,n ⎥
⎣ j =1 ⎦
0 0 0 ··· 0 0
The n-by-n submatrix obtained by deleting row n + 1 and column 1 is −B.
A tree is an acyclic connected graph. We consider orientations of trees, in which each
edge is given a direction. A rooted tree is an orientation of a tree in which all edges are
oriented along paths to a distinguished vertex called the root. The weight of a rooted tree
is the product of the weights of its edges. A special case of a deep generalization of Tutte’s

82 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
directed matrix tree theorem (S. Chaiken and D. J. Kleitman (1978), Matrix tree theorems,
J. Combinatorial Theory (A) 24:3, 377–381) states that the sum of the weights of the rooted
trees in a directed graph G that are rooted at a particular vertex vj is given by any cofactor
in the row for that vertex in the Laplacian matrix L(G) defined above. (Their proof is also
presented in D. B. West (2021), Combinatorial Mathematics, Cambridge, p. 750.) Thus
(−1)n+2 det(−B), which equals det(B), is the sum of the weights of all spanning trees of
G rooted at vn+1 .
When ai,j = 1 for all i and j , we are just counting spanning trees in a complete graph
with n + 1 vertices, which by Cayley’s formula yields det(B) = (n + 1)n−1 , solving (a).
Because the weights are nonnegative, det(B) is always nonnegative, solving (c), and det(B)
is maximized when all weights are maximized at 1, solving (b).
Also solved by L. Han & J. Xu, O. P. Lossers (Netherlands), R. Tauraso (Italy), and the proposer.

Organizing a Row of Coins

12316 [2022, 385]. Proposed by H. A. ShahAli, Tehran, Iran, and Manija Shahali, Bakers-
field, CA. For each i in {1, 2, . . . , C}, we have 2i coins with color i. Place these C(C + 1)
coins in a line. A move consists of the transposition of two adjacent coins. Let m be the
minimum number of moves required to reach a configuration where all coins of the same
color are together in a run of consecutive coins. Show that the maximum value of m over
all initial configurations is (C − 1)C(C + 1)(3C + 2)/12.
Solution by Nigel Hodges, Cheltenham, UK. We show first that, from any configuration,
(C − 1)C(C + 1)(3C + 2)/12 moves suffice to complete the task.
We proceed by induction on C. When C = 1 there is only one configuration, and the
number of moves needed is 0, which is the value of the specified formula at C = 1.
Now consider C > 1. Index the positions from 1 at the left end of the row to C(C + 1)
at the right end. We first move all coins of color C to one end, whichever end requires
fewer moves. Consider the leftmost coin of color C. Since an optimal set of moves will
never swap two coins of the same color, we may assume that this coin ends at position 1 or
position C 2 − C + 1. The sum of the number of moves involving it if it goes left plus the
number if it goes right is C(C − 1).
This sum is independent of which coin of color C we consider, so the total number of
moves spent on all color C coins if moved left, plus the number of moves spent if moved
right, equals 2C 2 (C − 1). Therefore at most C 2 (C − 1) moves are needed to gather the
coins of color C at one end.
With all coins of color C coins at one end, what remains is an instance with C − 1
colors. By the induction hypothesis, the number of moves remaining to complete the task
now is at most (C − 2)(C − 1)C(3C − 1)/12, and adding this to 2C 2 (C − 1) yields the
desired formula.
Consider the initial configuration whose right half consists of i coins of color i together
for each i, from color 1 in the middle of the row through color C at the right end. The left
half is the reflection of this, with half the coins of color C on the left end. We now show
that this configuration requires the full (C − 1)C(C + 1)(3C + 2)/12 moves. Again we
use induction on C, and again the case C = 1 is trivial.
With C > 1, the two blocks of color C have all other C(C − 1) coins between them,
so forming a single block of color C will require each of at least C coins to move at least
C(C − 1) times, requiring a total of at least C 2 (C − 1) moves involving coins of color
C. Moves that involve a coin of color C do not change the order among the coins with
earlier colors. Hence the number of moves that must be made not involving a coin with
color C must be at least the number required to solve the problem with C − 1 colors that is
obtained by ignoring the coins with color C. That ordering is the instance of the specified

January 2024] PROBLEMS AND SOLUTIONS 83

configuration when the number of colors is C − 1. By the induction hypothesis, there must
be at least (C − 2)(C − 1)C(3C − 1)/12 moves not involving a coin with color C. Again
the sum of the two contributions is the desired formula.
Also solved by H. Chen (China), A. De la Fuente, L. Gualá & S. Leucci & R. Tauraso (Italy), Y. J. Ionin,
Y. Kim (Korea), O. P. Lossers (Netherlands), K. Schilling, R. Stong, and the proposer.

A Fermat Point Inequality

12319 [2022, 386]. Proposed by Mihály Bencze, Braşov, Romania. Let ABC be a triangle
with all angles less than 120◦ , and let F be the Fermat point of ABC (the point in the
interior that minimizes the sum of the distances to A, B, and C). Prove
F A4 F B4 F C4 F A3 + F B 3 + F C 3
+ + ≥ .
AB 2 BC 2 CA2 FA + FB + F C

Solution by Nandan Sai Dasireddy, Hyderabad, India. Write F A = x, F B = y, and F C =

z. It is well known that all of the angles ∠AF B, ∠BF C, and ∠CF A are equal to 120◦ .
Therefore, by the law of cosines, AB 2 = x 2 + xy + y 2 , and similarly for BC 2 and CA2 .
Thus the desired inequality is
x4 x 3 + y 3 + z3
≥ ,
cyc
x 2 + xy + y 2 x+y+z

where we use f (x, y, z) as a shorthand for f (x, y, z) + f (y, z, x) + f (z, x, y). Since
cyc

x 3 + y 3 + z3 3xyz
= + (x 2 − xy),
x+y+z x+y+z cyc

it suffices to show
 
3xyz x4 xy 3
≤ − (x − xy) =
2
.
x+y+z cyc
x 2 + xy + y 2 cyc
x 2 + xy + y 2

By the Cauchy–Schwarz inequality,

xy 3 y2 (x + y + z)2 xyz(x + y + z)
= ≥ = .
cyc
x 2 + xy + y 2 cyc
x/y + y/x + 1 cyc (x/y + y/x + 1) xy + yz + zx

Therefore it suffices to prove

xyz(x + y + z) 3xyz
≥ .
xy + yz + zx x+y+z
But this is equivalent to (x + y + z)2 ≥ 3(xy + yz + zx), which is true because
(x − y)2
(x + y + z)2 − 3(xy + yz + zx) = ≥ 0.
cyc
2

Also solved by O. Geupel (Germany), K. T. L. Koo (China), C. G. Petalas (Greece), A. Stadler (Switzerland),
R. Stong, R. Tauraso (Italy), A. Tzavellas (Greece), L. Zhou, UM6P Math Club (Morocco), and the proposer.

84 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
CLASSICS
C21. From the 1986 International Mathematical Olympiad, suggested by the editors. An
integer is assigned to each vertex of a regular pentagon in such a way that the sum of
the five integers is positive. If three consecutive vertices are assigned the numbers x, y,
z in order, and y is negative, then one may replace x, y, and z by x + y, −y, and z + y,
respectively. Such an operation is performed repeatedly as long as at least one of the five
numbers is negative. Determine whether this procedure necessarily comes to an end after
a finite number of steps.

Period of the Fibonacci Sequence Modulo m

C20. Due to Peter Freyd, suggested by the editors. Given a positive integer m, let f (m) be
the period of the Fibonacci sequence taken modulo m. Prove f (m) ≤ 6m and that equality
holds for infinitely many m.  
Solution. Let A denote the Fibonacci matrix 01 11 , and let I denote the 2-by-2 identity

F Fn
 F0 = 0 and F1 = 1 and the relation Fn+1 = Fn + Fn−1 imply that
matrix. The values
An = Fn−1 n F
. Hence f (m) is the multiplicative order of A modulo m. From the first
n+1
few Fibonacci numbers we learn that A3 ≡ I (mod 2), so f (2) = 3. Similarly, A4 ≡ −I
(mod 3), so f (3) divides 8; it does not divide 4, so f (3) = 8. To compute f (5), note that
A5 ≡ 3I (mod 5); hence A20 ≡ I (mod 5), and so f (5) divides 20. Because neither A4
nor A10 equals I modulo 5, we infer f (5) = 20.
We next note that f (m1 m2 ) = lcm (f (m1 ), f (m2 )) when gcd (m1 , m2 ) = 1. This is
a consequence of the fact that a and b are congruent modulo m1 m2 if and only if they
are congruent modulo both m1 and m2 . This reduces the problem of bounding f (m) for
general m to the problem of finding such bounds when m is a prime power. For example,
f (10) = lcm(f (2), f (5)) = 60, so f (m) = 6m when  m = 10.
pf pa−1  p
When p is prime and a ≥ 2, the calculation A   = I + pa−1 M ≡ I (mod pa )
for some matrix M shows that f (pa ) divides pf pa−1 . Applying this a − 1 times yields
that f (pa ) divides pa−1 f (p).
We need some special information about the case p = 5.
Claim 1. f (5a ) = 4 · 5a .
Proof. We have A20 = I + 5K for some matrix K, and it is easily checked that K is
nonzero modulo 5. This is the base case for an induction proof of the stronger claim that
a
A4·5 ≡ I + 5a K (mod 5a+1 ). If this holds for some a, then
a+1
A4·5 = (I + 5a K + 5a+1 M)5 ≡ I + 5a+1 K (mod 5a+2 ),
for some matrix M. This completes the induction. 
For odd primes p not equal to 5, we say that p is type 1 if p ≡ ±1 (mod 5) and type 2
if p ≡ ±2 (mod 5). Let p be the Legendre symbol, equal to 0 if a ≡ 0 (mod p), equal
a
 
to 1 if a is a quadratic residue modulo p, and equal to −1 otherwise. Thus p5 is 1 if p is
type 1 and is −1 if p is type 2.

Claim 2. If p is type 1, then f (p) divides p − 1. If p is type 2, then f (p) divides 2(p + 1).

Proof. All congruences here are modulo p. Expand the Binet formula for the Fibonacci
numbers to obtain
√ √ m/2−1  
(1 + 5)m − (1 − 5)m 1 m
Fm = √ = m−1 5 i
.
2m 5 2 i=0
2i + 1

January 2024] PROBLEMS AND SOLUTIONS 85

We examine this formula for m = p − 1 and m = p, reducing modulo the prime p and
using 2p−1 ≡ 1 (mod p). When m = p, all terms in the sum except   the last one are divisi-
ble by p, and so Fp is congruent modulo p to 5(p−1)/2 , which is p5 . By the law of quadratic
   
reciprocity, p5 = p5 , and so Fp ≡ 1 in the type 1 case and Fp ≡ −1 in the type 2 case.
 
When m = p − 1, note that p−1 k
≡ (−1)k (mod p). Hence
(p−3)/2  (p−1)/2 
5 −1
Fp−1 ≡ 2 Fp−1 ≡ 2
p−1 i
5 (−1) 2i+1
≡ −2 ,
i=0
5−1
and so Fp−1 ≡ 0 in the type 1 case and Fp−1 ≡ 1 in the type 2 case. In the type 1 case, we
have (Fp−1 , Fp ) ≡ (0, 1) and hence Fp−2 ≡ 1 and Ap−1 = I . In the type 1 case, we have
(Fp−1 , Fp ) ≡ (1, −1) and hence Fp+1 ≡ 0, Fp+2 = −1, and Ap+1 = −I . 
We now combine the preceding facts to get the desired bound on f (m). Consider an
 c  d
arbitrary modulus m with prime factorization m = 2a 5b pi i qj j , where a, b ≥ 0 and
pi and qj range, respectively, over all type 1 and type 2 primes that divide m. We have
      d −1 
c −1
f (m) ≤ lcmi,j 3 · 2a−1 2 , 4 · 5b 5 , (pi − 1) pi i , 2 qj + 1 qj j ,
where [x]p is x if p divides m and 1 otherwise. It follows that
   
  dj −1
3 · 2a−1 2  pi − 1 lcmj 4 · 5 , 4 qj + 1 /2 · qj
b
f (m)
≤ · ·  d
m 2a i
pi 5b j qj j
3  pi − 1  qj + 1
≤ ·4· · ≤ 6. (∗)
2 i
pi j
2qj

This proves that f (m) ≤ 6m.

The inequality in (∗) is strict if either 2 or 5 does not divide m or if m has any type 1
a b
divisors pi or type 2 divisors qj . Thus
 f (m) can equal
 6m only if m has the
 form 2 5 for
a, b ≥ 1. In that case, f (m) ≤ lcm 3 · 2 , 4 · 5 = 3 · 5 · lcm 2 , 4 . If a ≥ 2, then
a−1 b b a−1

lcm 2a−1 , 4 /2a ≤ 1, so 6m cannot be reached. If a = 1, then f (m) = lcm 3, 4 · 5b =
3 · 4 · 5b = 6m, so f (m) = 6m exactly when m is one of the infinitely many values 2 · 5b
with b ≥ 1.
Editorial Comments: The present problem appeared as part of Problem E3410 [1990, 916;
1992, 278] in this Monthly, with solution by Kevin Brown.
The first detailed investigation into Fibonacci periodicity was D. D. Wall (1960),
Fibonacci series modulo m, this Monthly 67, 525–532. Included there is an alternative
proof of Claim 2 that uses linear algebra, along the following lines: For type 1 primes p, the
characteristic polynomial of A factors into distinct factors as x 2 − x − 1 = (x − α)(x − β)
in the field with p elements. Hence A is similar modulo p to the diagonal matrix diag(α, β),
and so Ap−1 ≡ I (mod p), because both α and β have multiplicative order dividing p − 1.
A similar argument exists for type 2 primes, though then α and β reside in the field with
p2 elements.
It is easy to see that the Fibonacci sequence modulo m is in fact periodic (and not merely
eventually periodic), which  explains the implicit assumption in the problem statement.
When p is prime, f p2 equals either pf (p) or f (p). Primes p with f (p2 ) = f (p)
are known as Wall–Sun–Sun primes. Surprisingly, it is unknown if any such primes exist.
There are none less than 1014 (A.-S. Elsenhans and J. Jahnel, The Fibonacci sequence
modulo p2 —an investigation by computer for p < 1014 , arxiv.org/pdf/1006.0824.pdf).
The editors thank Joe Buhler for his contribution in producing the solution here.

86 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 131
SOLUTIONS

An Exponential Field Homomorphism

12301 [2022, 186]. Proposed by Jan Mycielski, University of Colorado, Boulder, Colorado.
Suppose that α : C → C respects addition and exponentiation, in the sense that α(x + y) =
α(x) + α(y) and α (ex ) = eα(x) for all complex numbers x and y. (An example is complex
α (z) =√z̄.)
conjugation: √
(a) Prove α( 2) = 2.
(b)∗ Must it be the case that α(21/3 ) = 21/3 ? What about α(21/4 ) = 21/4 or α(ln 2) = ln 2?
Solution to (a) by Jayanta Manoharmayum, University of Sheffield, Sheffield, UK. We
first show that α is a field homomorphism. To see this, note first that α(0) = α(0 + 0) =
α(0) + α(0), so α(0) = 0. Hence α(1) = α(e0 ) = eα(0) = e0 = 1. To check that α respects
multiplication, let x, y ∈ C. If either x = 0 or y = 0, then α(xy) = 0 = α(x)α(y). If not,

December 2023] PROBLEMS AND SOLUTIONS 953

then we can find complex numbers u and v such that x = eu and y = ev , and therefore
α(xy) = α(eu+v ) = eα(u+v) = eα(u)+α(v) = eα(u) eα(v) = α(x)α(y).
Being a field homomorphism, the map α is injective and fixes every rational number.
From −1 = α(−1) = α(eiπ ) = eα(iπ) we conclude that α(iπ ) = inπ for some odd integer
n. Since
α(eiπ/n ) = eα(iπ)α(1/n) = einπ/n = −1 = α(−1),
√
we must have eiπ/n = −1, so n = ±1, and hence α(iπ ) = ±iπ . From 2 = eiπ/4 +
e−iπ/4 , it follows that
√ √
α( 2) = α(eiπ/4 + e−iπ/4 ) = e±iπ/4 + e∓iπ/4 = 2.

Editorial comment. Let K ⊆ C be the maximal cyclotomic extension of Q (that is, the field
extension of Q generated by the roots of unity eirπ for r ∈ Q). The argument above shows
that α(eirπ ) = eα(r)α(iπ) = e±irπ for all rational r. Hence either α acts on K trivially, or
α acts on K as complex conjugation. The Kronecker–Weber theorem then √ implies
√ that
α(θ ) = θ for any θ ∈ R with Q(θ )/Q an abelian extension. In particular, α( r) = r for
any positive rational r.
A sketch of a solution to part (a) of this problem appeared in J. Mycielski (1985),
Remarks on infinite systems of equations, Alg. Univ. 21, 307–309.
No correct solutions to (b) were received.
Part (a) also solved by J. Boswell & C. Curtis, N. Caro-Montoya (Brazil), G. A. Edgar, N. Grivaux (France),
E. A. Herman, Y. J. Ionin, O. P. Lossers (Netherlands), G. Plumpton & R. Su (Canada), M. Reid, K. Schilling,
A. Stenger, R. Stong, Missouri State University Problem Solving Group, and the proposer.

Where Angle Bisectors Meet Opposite Sides

12303 [2022, 186]. Proposed by George Apostolopoulos, Messolonghi, Greece. Let R and
r be the circumradius and inradius, respectively, of triangle ABC. Let D, E, and F be
chosen on sides BC, CA, and AB so that AD, BE, and CF bisect the angles of ABC.
Prove
 
FD DE EF 3 R
+ + ≤ 1+ .
AB + BC BC + CA CA + AB 8 2r

Composite solution by Richard Stong, Center for Communications Research, San Diego,
CA, and Tamas Wiandt, Rochester Institute of Technology, Rochester, NY. We prove
FD DE EF 3
+ + ≤ ,
AB + BC BC + CA CA + AB 4
which, by Euler’s inequality R ≥ 2r, implies the stated inequality.
Let a, b, and c denote the lengths of sides BC, CA, and AB, respectively. By the angle
bisector theorem, we have BD = ac/(b + c) and BF = ac/(a + b). Therefore, by the law
of cosines (twice),
      
ac 2 ac 2 ac ac
FD = 2
+ −2 cos(∠ABC)
b+c a+b b+c a+b
       2 
ac 2 ac 2 ac ac a + c2 − b2
= + −2
b+c a+b b+c a+b 2ac
abc(b3 + ab2 + cb2 + 3abc − a 2 b − c2 b + a 2 c + c2 a − a 3 − c3 )
= .
(a + b)2 (b + c)2

954 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
Combining this with similar formulas for DE 2 and EF 2 , we obtain
 2  2  2
FD DE EF
+ + =
AB + BC BC + CA CA + AB
abc(9abc + a 2 b + a 2 c + b2 c + b2 a + c2 a + c2 b − a 3 − b3 − c3 )
.
(a + b)2 (b + c)2 (c + a)2
By Schur’s inequality, a 2 b + a 2 c + b2 c + b2 a + c2 a + c2 b − a 3 − b3 − c3 ≤ 3abc, so
 2  2  2
FD DE EF 12(abc)2
+ + ≤
AB + BC BC + CA CA + AB (a + b)2 (b + c)2 (c + a)2
12(abc)2 3
≤ = .
(4ab)(4bc)(4ac) 16
The desired conclusion now follows by the Cauchy–Schwarz inequality.
Editorial comment. Several solvers noted that this problem is related to problem 12182
[2020, 461; 2022, 92] from this Monthly. Indeed, the inequality EF ≤ (2a + b + c)/8,
derived in the published solution there, can be used as the basis for another solution to this
problem.
Also solved by M. Bataille (France), N. S. Dasireddy (India), M. Drăgan & N. Stanciu (Romania), G. Fera
(Italy), O. Geupel (Germany), N. Hodges (UK), W. Janous (Austria), C. G. Petalas (Greece), C. R. Pranesachar
(India), V. Schindler (Germany), A. Stadler (Switzerland), R. Tauraso (Italy), M. Vowe (Switzerland), L. Zhou,
and the proposer.

An Interpolation Identity
12304 [2022, 186]. Proposed by Michel Bataille, Rouen, France. Let m and n be positive
integers with m < n. Prove
m   m   m  
m (−1)k n (−1)k m (−1)k
= .
k=0
k n−k k=0
k k+1 k=0
k (n − k)(k + 1)

Solution by Pierre Lalonde, Plessisville, QC, Canada. For a nonnegative integer k, extend
the binomial coefficient in the usual way as
 
x x(x − 1) · · · (x − k + 1)
= .
k k!
We prove the polynomial identity
m   m  x 
x (−1)k m + 1 m+1
= (−1) m−k
. (1)
k=0
k k+1 k=0
k+1 x−k
 x 
(The right side is a polynomial since x − k divides m+1 .) Evaluating the left side of (1) at
x = j for 0 ≤ j ≤ m yields
m   j  
j (−1)k −1 j +1 −1   1
= (−1)k+1 = (1 − 1)j +1 − 1 = .
k=0
k k+1 j + 1 k=0 k + 1 j +1 j +1
When we evaluate the right side of (1) at x = j , the only term in the sum that is nonzero is
the one with k = j . Thus
m   x   
m−k m + 1 m−j m + 1 j !(−1) (m − j )!
m−j
m+1 1
(−1) = (−1) = .
k=0
k+1 x−k j +1 (m + 1)! j +1

December 2023] PROBLEMS AND SOLUTIONS 955

Since both sides of (1) are polynomials of degree m that agree on m + 1 values, they are
equal.  x 
Dividing both sides of (1) by (−1)m m+1 (m + 1), we have
m   m  
(−1)m x (−1)k m (−1)k
 x  = . (2)
m+1
(m + 1) k=0 k k + 1 k=0
k (k + 1)(x − k)
Expanding the coefficient on the left side of (2) by partial fractions yields
m
(−1)m (−1)m m! ak
 x  = = ,
m+1
(m + 1) x(x − 1) · · · (x − m) k=0
x−k
for some coefficients a0 , . . . , am . To compute these coefficients, clear fractions and set
x = j . Only the term for k = j survives, and so we obtain
m
ak j (j − 1) · · · (j − m)
(−1)m m! = = aj j !(−1)m−j (m − j )! .
k=0
j −k
 
Thus aj = (−1)j mj . Substituting this expansion into (2) gives
m   m   m  
m (−1)k x (−1)k m (−1)k
= .
k=0
k x−k k=0
k k+1 k=0
k (x − k)(k + 1)
Evaluating at x = n when n > m yields the result.
Editorial comment. Lalonde notes that the left and right sides of (1) are the Newton and
Lagrange interpolation polynomials, respectively, for the points {(j, j +1
1
) : 0 ≤ j ≤ m}.
Most solvers evaluated the three sums individually, either in terms of Euler’s beta inte-
grals or by induction using binomial identities.
Also solved by U. Abel (Germany), A. Berkane (Algeria), P. Bracken, C. Curtis, G. Fera (Italy), O. Geupel
(Germany), N. Hodges (UK), W. Janous (Austria), O. Kouba (Syria), O. P. Lossers (Netherlands), F. Masroor,
E. Schmeichel, A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), L. Zhou, and the proposer.

Laplace Simplifies an Integral

12305 [2022, 187]. Proposed by Shivam Sharma, Delhi University, New Delhi, India. Prove
x − 1 − x ln x
1
dx = γ ,
0 x ln x − x ln2 x

where γ is Euler’s constant limn→∞ (− ln n + nk=1 1/k).
Solution by Seán Stewart, King Abdullah University of Science and Technology, Thuwal,
Saudi Arabia. Denote the integral to be calculated by I . The substitution x = e−t produces
∞
1 − e−t (1 + t)
I= dt.
0 t (1 + t)
To evaluate this integral, we use the Laplace transform L, defined by L{g}(t) =
∞ −st
0 g(s)e ds. In particular, we use the property that, for positive functions f and g,

∞ ∞
f (t) · L{g}(t) dt = L{f }(s) · g(s) ds, (∗)
0 0
as long as both improper integrals are defined. This property is proved by expressing both
sides of (∗) as double integrals and reversing the order of integration.

956 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
 From elementary properties of the Laplace transform we have
1 1 1
L{1 − e−t (1 + t)}(s) = L{1}(s) − L{e−t }(s) − L{te−t }(s) = − −
s s + 1 (s + 1)2
and
1 1 1
L{1 − e−s }(t) = L{1}(t) − L{e−s }(t) = − = .
t t +1 t (t + 1)
Applying (∗) and then integration by parts, we get
∞
I= (1 − e−t (1 + t)) · L{1 − e−s }(t) dt
0
∞
= L{1 − e−t (1 + t)}(s) · (1 − e−s ) ds
0
∞  
1 1 1
= − − · (1 − e−s ) ds
0 s s + 1 (s + 1)2
  ∞ ∞ 
1 
−s  1
= ln s − ln(s + 1) + (1 − e ) − ln s − ln(s + 1) + e−s ds
s+1 0 0 s + 1
∞ ∞  
−s −s 1  ∞
=− e ln s ds − e − ln(s + 1) ds = γ − e−s ln(s + 1) 0 = γ ,
0 0 s+1
∞
where in the last line we use the well-known integral representation γ = − 0 e−x ln x dx.
(See, for example, F. W. J. Olver et al. (2010), NIST Handbook of Mathematical Functions,
Cambridge Univ. Press, p. 140, Eq. 5.9.17.)
Also solved by E. Alan, T. Amdeberhan & V. H. Moll, M. Bataille (France), A. Berkane (Algeria), N. Bhandari
(Nepal), P. Bracken, B. Bradie, B. S. Burdick, W. Chang, H. Chen, H. Chen & F. Zhuang (Canada), B. E. Davis,
G. Fera (Italy), D. Fleischman, M. L. Glasser (Spain), R. Gordon, J.-P. Grivaux (France), L. Han, E. A. Herman,
N. Hodges (UK), F. Holland (Ireland), W. Janous (Austria), S. Kaczkowski, A. M. Karparvar (Iran), O. Kouba
(Syria), O. P. Lossers (Netherlands), F. Masroor, M. Omarjee (France), H. Ricardo, V. Schindler (Germany),
T. P. Sharma (India), A. Stadler (Switzerland), M. S̆tofka (Slovakia), R. Stong, R. Tauraso (Italy), M. Vowe
(Switzerland), J. Vukmirović (Serbia), T. Wiandt, H. Widmer, J. Yan (China), L. Zhou, Fejéntaláltuka Szeged
Problem Solving Group (Hungary), and the proposer.

A Sum of Euler and von Mangoldt Functions

12306 [2022, 187]. Proposed by Amrit Awasthi, Amritsar, India. For a positive integer n,
evaluate
φ(a) ln a + φ(a)(b),
a|n a|n b|(n/a)

where φ is the Euler phi function (φ(m) is the number of integers k with 1 ≤ k ≤ m that
are relatively prime to m) and  is the von Mangoldt function ((m) equals ln p when m
is a power of the prime number p and equals 0 when m is not a prime power).
Solution by Richard Stong, Center for Communications  Research, San Diego,
 CA. The
sum equals n ln n. We use the well-known identities b|n (b) = ln n and a|n φ(a) = n.
m
The first identity follows from the prime factorization n = p1 1 · · · prmr : for each i, the mi
powers of pi each contribute ln pi to the sum, and all other terms are zero. The second
follows because φ(a) counts the elements k of {1, 2, . . . , n} such that gcd(k, n) = n/a for
each divisor a of n.

December 2023] PROBLEMS AND SOLUTIONS 957

 Using these identities, we obtain
 n
φ(a) ln a + φ(a)(b) = φ(a) ln a + ln = ln n φ(a) = n ln n.
a|n a|n b|(n/a) a|n
a a|n

Also solved by B. S. Burdick, C. Burnette, N. Caro-Montoya (Brazil), W. Chang, C. Curtis, T. Dickens,

O. Geupel (Germany), N. Hodges (UK), Y. J. Ionin, W. Janous (Austria), A. M. Karparvar (Iran), K. T. L. Koo
(China), O. Kouba (Syria), O. P. Lossers (Netherlands), R. Molinari, M. Reid, H. Ricardo, A. Stadler (Switzer-
land), R. Tauraso (Italy), L. Zhou, Missouri State University Problem Solving Group, and the proposer.

Double-Loading Six-Pack
12307 [2022, 285]. Proposed by Stuart Boersma, Central Washington University,
Ellensburg, WA, Kim Ruhland, Breckenridge,
CO, and Bruce Torrence, Randolph-Macon
College, Ashland, VA. Consider a ski lift with
n chairs attached to a cable loop. Let m be an
integer such that 1 ≤ m ≤ n. At each loading
stage at the bottom, the lowest m descending
chairs are detached from the cable in order,
loaded with skiers, and then reattached in the
reverse order but otherwise at the same loca-
tions around the cable from which they were
removed (see figure for the case n = 107 and
m = 2; a lift of this type is used at the Breck-
enridge ski area). At the next stage, the same
steps are carried out with the next m descend-
ing chairs; the process continues indefinitely.
After how many loading stages are the chairs
returned to the same cyclic order they had at
the beginning?
Solution by the Eagle Problem Solvers, Georgia Southern University, Statesboro, GA and
Savannah, GA. Let f (n, m) denote the (minimum) number of loading stages required
before the chairs are returned to the same cylic order they had at the beginning. If m = 1,
then there is no change in the order of the chairs after each loading stage, so f (n, 1) = 1
for each n. Henceforth we assume m > 1.
Let q and r be the unique integers such that n = qm + r and 0 ≤ r < m. We prove
⎧
⎪
⎨2q if r = 0;
f (n, m) = q(q + 1) if r = 1 and q is odd, or r = m − 1 and q is even;
⎪
⎩
2q(q + 1) otherwise.
If r = 0, then after the first q loading stages each successive group of m chairs is inter-
nally reversed. Since m > 1, this differs from the original order. After q more stages, the
original order within each group is restored, and the order of the groups is unchanged, so
f (n, m) = 2q.
Now suppose r > 0. When k is a nonnegative integer, let Ak = (mk + 1, . . . , mk + r)
and Bk = (mk + r + 1, . . . , mk + m). Let Ak and B k denote the reversals of Ak and Bk ,
respectively, and let Sk denote the order of the chairs after k loading stages. We assume the
initial order S0 is (1, . . . , n), so
S0 = A0 B0 A1 B1 · · · Aq−1 Bq−1 Aq .

958 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
 After each loading stage, the first (leftmost) m chairs are reversed and moved to the end
of the cable, so
S1 = A1 B1 A2 B2 · · · Aq B 0 A0
and
Sq+1 = A0 B 1 A1 · · · B q−1 Aq−1 B0 Aq .
Notice that in each configuration, exactly one pair of Ak groups are adjacent (possibly with
one or both being reversed). In the initial configuration, A0 and Aq are adjacent. After one
loading stage, A1 and A0 are adjacent (and A0 has been reversed); after two loading stages,
the adjacent pair is A2 and A1 . After q + 1 loading stages, the adjacent pair is again A0
and Aq , and then the pattern repeats. It follows that f (n, m) must be a multiple of q + 1.
After the first q + 1 loading stages, the groups Ak are back in their original positions, but
reversed, while the positions of the Bk have shifted by one position. The group B0 has
been reversed twice, so it is back in its original order, and every other Bk has been reversed
once. Each set of q + 1 loading stages has a similar effect, so the first time that the Ak and
Bk groups are back in their original cyclic order is after q sets of q + 1 loading stages, at
which point each Ak has been reversed q times and each Bk has been reversed q + 1 times.
Therefore

A0 B0 A1 B1 · · · Aq−1 Bq−1 Aq if q is odd
Sq(q+1) =
A0 B 0 A1 B 1 · · · Aq−1 B q−1 Aq if q is even.

If r = 1, then Ak = Ak , while if r = m − 1 then B k = Bk . Thus Sq(q+1) = S0 when r = 1

and q is odd, and when r = m − 1 and q is even. Otherwise, we need another q(q + 1)
loading stages for the Ak and Bk groups to once again return to their original cyclic order,
and S2q(q+1) = S0 . This establishes the desired result for f (n, m).
Since 107 = 53 · 2 + 1, for the Breckenridge ski lift r = 1 and q is odd, so the number
of loading stages is f (107, 2) = q(q + 1) = 53 · 54 = 2862.
Editorial comment. The Quicksilver Super6 ski lift at Breckenridge seats 6 skiers per chair
and is the only ski lift in North America to have double loading. This arrangement is known
to the locals as a “double-loading six-pack”.
The problem is a discrete version of classic problem C4 [2022, 394; 2022, 495] from
this Monthly, and the methods of solution are similar.
Also solved by C. Farnsworth, K. Gatesman, O. Geupel (Germany), N. Hodges (UK), Y. J. Ionin, O. P. Lossers
(Netherlands), A. Mandal (India), R. Stong, L. Zhou, and the proposers.

Eliminating Tiles
12309 [2022, 286]. Proposed by Joseph DeVincentis, Salem, MA, Thomas C. Occhipinti,
Luther College, Decorah IA, and Daniel J. Velleman, Amherst College, Amherst, MA, and
University of Vermont, Burlington, VT. Consider a square grid that is infinite in all direc-
tions, with tiles placed on finitely many squares of the grid. Two grid squares are called
adjacent if they share an edge. There are two types of legal moves:
(A) If two tiles are on adjacent squares, then they can both be removed.
(B) If a tile is on a square and all adjacent squares are unoccupied, then the tile can be
removed with four new tiles then placed on the four adjacent squares.
For which initial configurations is it possible to eliminate all tiles from the grid?
Solution by José Heber Nieto, University of Zulia, Maracaibo, Venezuela. Color the squares
of the grid alternating black and white, as in an infinite chessboard. Given a distribution of

December 2023] PROBLEMS AND SOLUTIONS 959

tiles, let w and b denote the number of tiles placed on white and black squares, respectively.
We prove that it is possible to eliminate all of the tiles if and only if w ≡ b (mod 5).
The necessity of w ≡ b (mod 5) holds because A-moves do not change w − b and B-
moves increase or decrease w − b by 5. Hence w − b (mod 5) is an invariant and, if it is
possible to eliminate all of the tiles, then w − b ≡ 0 (mod 5).
To prove sufficiency, we apply induction on max{w, b}, beginning with the obvious case
w = b = 0. If any tiles are adjacent, we can apply an A-move and then apply the induction
hypothesis. If no tiles are adjacent and w = b, say w > b without loss of generality, we
apply a B-move to any tile on a white square, leaving w − 1 tiles on white squares and b + 4
tiles on black squares. Because w ≥ b + 5, we may again apply the induction hypothesis.
Thus we may assume w = b > 0 and that no tiles are adjacent.
Now choose a tile on a white square and one on a black square so that the Euclidean
distance between the centers of the squares is minimal. We may assume that the squares are
unit squares, assign integral coordinates to their centers, and name each square by its center.
By symmetry, we may assume one tile is on (0, 0) and the other on (a, b), where a >
b ≥ 0. Observe that a ≥ 2 and that a and b have opposite parity. Squares (a ± 1, b) and
(a, b ± 1) are adjacent to (a, b), so they are empty. Squares (a − 2, b) and (a − 1, b − 1)
are also empty because they are closer to (0, 0) than is (a, b).
Case 1: (a − 1, b + 1) is empty. Apply a B-move to (a, b), a B-move to (a − 1, b), and
A-moves to (a − 1, b + 1) and (a, b + 1), to (a, b) and (a + 1, b), and to (a − 1, b − 1)
and (a, b − 1). The net effect is to move the tile on (a, b) to (a − 2, b).

Case 2: (a − 1, b + 1) is occupied. Again first apply a B-move to (a, b), this time
followed by applying an A-move to (a − 1, b + 1) and (a, b + 1). Now apply a B-move to
(a − 1, b) and A-moves to (a, b) and (a + 1, b) and to (a − 1, b − 1) and (a, b − 1). The
net effect is again to move the tile on (a, b) to (a − 2, b).

Case 1 and Case 2 both reduce the distance between (0, 0) and the nearest square of
the opposite color. Hence iterating the appropriate case brings a tile next to (0, 0). At that
point an A-move removes the pair, and the induction hypothesis applies.
Also solved by J. Boswell & C. Curtis, V. Chen & O. Zhang, K. Gatesman, O. Geupel (Germany), N. Hodges
(UK), Y. J. Ionin, O. P. Lossers (Netherlands), A. Martin & P. Martin & R. Martin (Germany), K. Schilling,
R. Stong, R. Tauraso (Italy), and the proposer.

Parallel Segments and Concurrent Cevians

12310 [2022, 286]. Proposed by Thanos Kalogerakis, Kiato, Greece, Dan-Stefan Mari-
nescu, Hunedoara, Romania, and Mehmet Şahin, Ankara, Turkey. Let P be a point inside
triangle ABC, and let D, E, and F be points on BC, CA, and AB, respectively, such that
P E, P F , and P D are parallel to AB, BC, and CA, respectively. Let A , B  , and C  be
points on BC, CA, and AB, respectively, such that AA , BB  , and CC  are parallel to DE,
EF , and F D, respectively.
(a) Prove Area(ABC) ≥ 3 · Area(DEF ), and determine conditions for equality.
(b) Prove that AA , BB  , and CC  are concurrent.
(c) Prove P D · P E · P F ≥ 8 · A D · B  E · C  F , and determine conditions for equality.

960 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
Solution by Faraz Masroor, New York, NY.
(a) Suppose that AP , BP , and CP intersect BC, CA, and AB at X, Y , and Z, respectively.
Let x = P X/AX, y = P Y /BY , and z = P Z/CZ. For any polygon I J KL, denote its area
by [I J KL]. We have x = [P BC]/[ABC], y = [P CA]/[ABC], and z = [P AB]/[ABC].
Thus x + y + z = 1. By the Cauchy–Schwarz inequality,
 
(1 + 1 + 1) x 2 + y 2 + z2 ≥ (x + y + z)2 = 1,
with equality if and only if
x = y = z = 1/3. Suppose
that P D, P E, and P F inter-
sect AB, BC, and CA at U ,
V , and W , respectively. We
have [P V D]/[ABC] = x 2 ,
[P W E]/[ABC] = y2,
and [P U F ]/[ABC] = z2 .
Also, [P DE] = [P DC] =
[P DCW ]/2. Likewise,
[P EF ] = [P EAU ]/2 and
[P F D] = [P F BV ]/2. There-
fore
[DEF ] = [P DE] + [P EF ] + [P F D]
[ABC] − [P V D] − [P W E] − [P U F ]
=
2
[ABC]   [ABC]
= 1 − x 2 − y 2 − z2 ≤ ,
2 3
with equality if and only if P is the centroid of triangle ABC.
(b) Since ∠BAA = ∠P ED and ∠A AC = ∠EDP , applying the law of sines in P DE
we get sin ∠BAA / sin ∠A AC = P D/P E. Multiplying this by the other two analogous
equations, we have
sin ∠BAA sin ∠CBB  sin ∠ACC  PD PE PF
· · = · · = 1.
sin ∠A AC sin ∠B  BA sin ∠C  CB PE PF PD
Therefore, by the trigonometric form of Ceva’s theorem, AA , BB  , and CC  concur at
some point Q.
(c) Let a = BC, b = CA, and c = AB. Notice that P D = bx, P E = cy, and P F =
az. Also, AE = U P = bz, DC = P W = ay, and EC = EW + P D = by + bx. From
A D/AE = DC/EC we have A D = ayz/(x + y). Likewise, B  E = bzx/(y + z) and
C  F = cxy/(z + x). The AM–GM inequality now yields
PD · PE · PF (x + y)(y + z)(z + x)
  
= ≥ 8.
AD·B E·C F xyz
Equality holds if and only if x = y = z = 1/3, that is, exactly when P is the centroid of
triangle ABC.
Editorial comment. The problem statement here corrects a typographical error in the orig-
inal statement of the problem.
Also solved by M. Bataille (France), H. Chen (China), C. Curtis, G. Fera (Italy), K. Gatesman, O. Geupel
(Germany), O. Kouba (Syria), O. P. Lossers (Netherlands), C. Petalas (Greece), C. R. Pranesachar (India),
V. Schindler (Germany), R. Stong, L. Zhou, Davis Problem Solving Group, and the proposer. Parts (a) and (b)
also solved by J. P. Grivaux.

December 2023] PROBLEMS AND SOLUTIONS 961

CLASSICS
C20. Due to Peter Freyd, suggested by the editors. Given a positive integer m, let f (m) be
the period of the Fibonacci sequence taken modulo m. Prove f (m) ≤ 6m and that equality
holds for infinitely many m.

Conway’s Solitaire Army

C19. Due to John H. Conway, suggested by the editors. A battlefield is modeled by an
infinite grid of unit squares, whose centers are indexed by {(a, b) : a, b ∈ Z}. The soldiers
are modeled by pegs, which are placed initially at a finite number of squares (a, b) with
b ≤ 0. The soldiers advance by jumping in the style of peg solitaire: A jump is permitted
when there are three squares in the grid forming a 1-by-3 rectangle with one end square
of this rectangle empty while the other two squares are occupied by pegs. Where this
configuration exists, the peg on the end may jump over the peg in the middle and move to
the empty end, while the peg in the middle is removed. How many pegs are needed in an
initial configuration to allow a peg to advance to the square (0, 5)?

A possible initial configuration A possible initial jump

Solution. It is impossible to advance a peg to the square (0, 5), no matter how many pegs
|a|−b
are in the initial
√ configuration. To see this, assign to the square (a, b) the2 weight λ ,
where λ = ( 5 − 1)/2. This λ is chosen so that 0 < λ < 1 and λ + λ = 1. For any
position of pegs on the grid, define the weight of the position to be the sum of the weights
of the squares that are occupied by pegs. The weight of the entire halfplane with b ≤ 0 is
∞ 0 ∞ ∞ ∞ ∞
|a|−b
λ = λ a
λ +
b
λ a
λb
a=−∞ b=−∞ a=0 b=0 a=1 b=0

1 1 λ 1 1+λ 1+λ 1
= + = = = 5.
1−λ1−λ 1−λ1−λ (1 − λ)2 λ 4 λ
Since there are only finitely many soldiers at the start, the weight of the original configu-
ration is strictly less than 1/λ5 .
Any jump involves eliminating pegs in squares of weight λn and λn+1 , respectively,
for some n, and adding a peg to an empty square of weight λn−1 or λn+1 or λn+2 . Since
λn + λn+1 = λn−1 , while λn+1 and λn+2 are smaller, we see that no jump can increase the
weight of the position. Yet the weight of the target square (0, 5) is 1/λ5 , which exceeds the
weight of the original configuration. Thus no finite number of initial pegs allows a peg to
reach the square (0, 5).
Editorial Comment: The number of pegs required in an initial configuration to advance a
peg to the square (0, n) for n = 1, 2, 3, and 4 is 2, 4, 8, and 20, respectively. The problem
has seen many reincarnations and generalizations. It originated with John H. Conway in
1961 and appears in R. Honsberger (1976), A problem in checker jumping, in Mathemati-
cal Gems II, Mathematical Association of America, pp. 23–28.

962 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
SOLUTIONS

Two Cyclic Quadrilaterals

12294 [2022, 86]. Proposed by Tran Quang Hung, Hanoi, Vietnam. Let A1 A2 A3 A4 be a
quadrilateral inscribed in a circle with center O. Let B1 B2 B3 B4 be the quadrilateral that
contains A1 A2 A3 A4 in its interior such that, for 1 ≤ i ≤ 4 and with subscripts taken cycli-
cally, Bi Bi+1 is parallel to Ai Ai+1 and at distance |Ai Ai+1 | from it. Because B1 B2 B3 B4
has the same angles as A1 A2 A3 A4 , there is a circle in which it is inscribed. Let P be the
center of that circle. Show that A1 A3 , A2 A4 , and OP are concurrent.
Solution by Richard Stong, Center for Communications Research, San Diego, CA. Lay
down complex coordinates with the circumcircle of A1 A2 A3 A4 as the unit circle and
A1 A2 A3 A4 oriented clockwise. The coordinate of the circumcenter O is 0. We use low-
ercase letters to denote the coordinates of the corresponding uppercase points. Hence the
complex numbers ai , for 1 ≤ i ≤ 4, have modulus one, so their complex conjugates are
their respective inverses.
Let X be the intersection point of the diagonals A1 A3 and A2 A4 . Its coordinate can be
found by treating the equation

x = ta1 + (1 − t)a3 = ua2 + (1 − u)a4

and its complex conjugate as two linear equations in the (real) unknowns t and u. The
result is
a1 (a2 − a3 )(a3 − a4 ) a2 (a3 − a4 )(a4 − a1 )
t= , u= .
(a1 − a3 )(a1 a3 − a2 a4 ) (a2 − a4 )(a2 a4 − a1 a3 )
Hence
a1 a2 a3 + a1 a3 a4 − a1 a2 a4 − a2 a3 a4
x= .
a1 a3 − a2 a4

November 2023] PROBLEMS AND SOLUTIONS 863

 The line parallel to A1 A2 , at a distance |a2 − a1 | from A1 A2 and outside A1 A2 A3 A4 ,
can be parametrized (by real t) as
12 (t) = ta1 + (1 − t)a2 + i(a2 − a1 ),
and symmetrically for the other sides. Treating the equation b1 = 12 (t) = 41 (u) and its
complex conjugate as two linear equations in the unknowns t and u, we get
a1 a2 − a1 a4 + 2a1 a2 i + 2a1 a4 i − 4a2 a4 i
b1 = ,
a2 − a4
and symmetrically for the other points.
Since P is equidistant from B1 , B2 , and B3 , we have
(p − b1 )(p − b1 ) = (p − b2 )(p − b2 ) = (p − b3 )(p − b3 ).
This gives two linear equations in the two unknowns p and p, and solving these equations
we find the coordinate of P to be
−4i(a1 a2 a3 + a1 a3 a4 − a1 a2 a4 − a2 a3 a4 ) −4i(a1 a3 − a2 a4 )
p= = · x.
(a1 − a3 )(a2 − a4 ) (a1 − a3 )(a2 − a4 )
The coefficient of x in the last equation is easily checked to be its own complex conjugate,
so it is real. Thus X and P lie on the same line through the origin O, as desired.
Editorial comment. Using a similar argument, Roberto Tauraso showed that the same con-
clusion holds if each Bi Bi+1 is at a distance r|Ai Ai+1 | from Ai Ai+1 , for any r > 0.
Also solved by J.-P. Grivaux (France), C. R. Pranesachar (India), A. Stadler (Switzerland), R. Tauraso (Italy),
and the proposer.

A Geometric Progression as a Sum of Two Squares

12295 [2022, 86]. Proposed by Koopa Tak Lun Koo, Chinese STEAM Academy, Hong
Kong, China.
(a) Show that when n is an odd positive integer, 1 + 7n + 72n + 73n + 74n + 75n + 76n is a
sum of two squares.
(b)∗ Show that when n is even, the expression in part (a) is not a sum of two squares.
Solution by Michael Reid, University of Central Florida, Orlando, FL. (a) The expression
equals (7n − 1)6 + 7n+1 (72n − 7n + 1)2 . When n is odd, 7n+1 is a square, so the number is
the sum of a sixth power and a square.
(b) When n is even, the expression is congruent to 7 modulo 8, so it is not even the sum of
three squares.
Editorial comment. Several solvers used quadratic reciprocity to solve (b), which yields a
generalization. For a prime p congruent to 3 modulo 4,
1 + pn + p2n + · · · + p(p−1)n
is the sum of two squares if and only if n is odd. As in the proof above, for even n the sum
is congruent to p mod 8, so it is not the sum of two squares. Conversely, factor
ppn − 1 = (pn − 1)(1 + pn + p2n + · · · + p(p−1)n ).
For a prime q dividing the first factor, the second factor is p modulo q, so the two factors are
relatively prime. Thus for a prime q dividing the second factor, the multiplicative order of
p modulo q is divisible by p. By Lagrange’s theorem, p divides q − 1, so q ≡ 1 (mod p)
and q is a square modulo p. Since p has odd order modulo q, it further follows that p is

864 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
a square modulo q. Therefore quadratic reciprocity yields (−1)(p−1)(q−1)/4 = 1, implying
q ≡ 1 (mod 4). Such primes q are the sum of two squares, as are the products of sums of
two squares. See, for instance, G. H. Hardy and E. M. Wright (2008), An Introduction to
the Theory of Numbers, 6th ed., Oxford University Press, for an explanation of quadratic
reciprocity and the needed facts about sums of two squares.
Other solvers used a theorem of Aurifeuille, Le Lasseur, and Lucas from the 1870s
to give another existence proof for (a). The theorem implies that, for m odd and square-
free, the cyclotomic polynomial m (x) may be expressed as p2 (x) − (−1)(m−1)/2 mxq 2 (x),
where the polynomials p and q have integer coefficients. See A. Schinzel (1962), On prim-
itive prime factors of a n − bn , Proc. Cambridge Phil. Soc., 58(4), 555–562. When m ≡ 3
(mod 4) and n is odd, setting x = mn generalizes (a).
Also solved by M. Chamberland, R. Dietmann & M. Widmer (UK), A. Dixit (India), S. Fan, N. Fellini
(Canada), P. Lalonde (Canada), O. P. Lossers (Netherlands), J. Manoharmayum (UK), R. Martin (Germany),
J. P. Robertson, J. Silverberg, A. Stenger, R. Stong, Eagle Problem Solvers, and the proposer. Part (a) also
solved by C. Curtis and O. Geupel (Germany). Part (b) also solved by C. Degenkolb, B. Finkel, N. Hodges
(UK), and B. Sury (India).

Coloring the Complement of a Matching

12296 [2022, 86]. Proposed by David A. Kalarkop and R. Rangarajan, University of
Mysore, Mysuru, India, and Douglas B. West, University of Illinois, Urbana, IL. For t ≤
n/2, let H (n, t) be the graph obtained from the complete graph on n vertices by deleting
t pairwise disjoint edges. Determine the number of ways to assign each vertex of H (n, t)
a color from a set of k available colors so that vertices forming an edge receive distinct
colors.
Solution by Oliver Geupel, Brühl, Germany. Such a coloring of vertices is known as a
proper coloring of a graph. We show that the number of proper colorings of H (n, t) from
k available colors is
t  
t
k(n−t+r) ,
r=0
r

where x(m) denotes the falling factorial m−1 i=0 (x − i).

Let e1 , . . . , et denote the deleted edges, which we view as pairs of vertices. The set
of vertices of H (n, t) is then the disjoint union of the sets e1 , . . . , et and a set U of size
n − 2t. An induced subgraph consisting of all of U and exactly one vertex from each ej is
the complete graph on its n − t vertices. Hence the vertices of any such subgraph require
distinct colors. Given a proper coloring of H (n, t), let C1 , . . . , Ct and C denote the disjoint
  Note that |Cj | ∈ {1, 2}
sets of colors assigned to vertices in e1 , . . . , et and U , respectively.
for each j and |C| = |U | = n − 2t. For 0 ≤ r ≤ t, there are rt ways to choose r among
C1 , . . . , Ct to have two elements. Each such choice defines a partition of the vertices into
n − (t − r) sets receiving distinct colors, and then there are k(n−t+r) ways to assign colors
to these sets. Summing over r counts all the proper colorings using k available colors.
Also solved by N. Caro-Montoya (Brazil), K. Gatesman, O. Geupel (Germany), S. C. Locke, O. P. Lossers
(Netherlands), R. Martin (Germany), J. H. Nieto (Venezuela), E. Schmeichel, A. Stadler (Switzerland),
R. Stong, R. Tauraso (Italy), Missouri State University Problem Solving Group, and the proposers.

A Hyperbolic Trigonometric Integral

12297 [2022, 86]. Proposed by Narendra Bhandari, Bajura District, Nepal. Prove
 
π/2 sinh−1 (sin x) 2
π π
2
dx = − ln 2 .
0 sin x 2 2

November 2023] PROBLEMS AND SOLUTIONS 865

Solution by John E. Kampmeyer III, Springfield, PA. We first use integration by parts and
the identities
   
−1 −1 x −1 1 1+z 1
z
sinh x = tanh √ and tanh z = ln = dy
1+x 2 2 1−z 0 1−z y
2 2

to compute
 2  2 π/2
π/2 sinh−1 (sin x) π/2 sinh−1 (sin x) 
cos2 x sinh−1 (sin x)
dx = 2 dx − 
0 sin2 x 0 sin x 1 + sin2 x tan x 
0
 
π/2 2
cos x sin x
=2 tanh−1 dx
0 sin x 1 + sin x
2
1 + sin2 x
π/2 1 1 π/2
cos2 x 1
=2 dy dx = 2 dx dy.
0 0 1 + (1 − y 2 ) sin x
2
0 0 1 + (2 − y 2 ) tan2 x

Using the substitutions t = tan x and then u = t 2 − y 2 we obtain

 
π/2 sinh−1 (sin x) 2 1 ∞
1 1
dx = 2 · dt dy
0 sin x2
0 0 1 + (2 − y 2 )t 2 1 + t 2

1   ∞ ∞ 
1 dt 1 dt
=2 1− 2 + 2 dy
0 y − 1 0 1 + (2 − y 2 )t 2 y − 1 0 1 + t2
 
1   ∞ ∞
1 1 du 1 dt
=2 1− 2 + 2 dy
0 y −1 2 − y 2 0 1 + u2 y − 1 0 1 + t2
 
1  
1 1 1
=π 1− 2 + 2 dy
0 y −1 2 − y2 y −1
   
1 1
dy 1 1
=π + 1− dy . (∗)
0 y −1
2
0 2 − y2 2 − y2
1 √ 1
The first integral 0 dy/ 2 − y 2 in (∗) is equal to sin−1 y/ 2  = π/4. For the second
0
we use the substitution u = y/ 2 − y 2 to compute
   
1 t t
1 1 dy dy
1− dy = lim −
0 y 2 − 1 2−y 2 t→1 −
0 y 2−1
0 (y 2 − 1) 2 − y 2
⎛ √ ⎞
t t/ 2−t 2
dy du ⎠
= lim ⎝ 2−1
− 2−1
t→1− 0 y 0 u
    
1 t t
= lim ln(1 − t) − ln(1 + t) − ln 1 − √ + ln 1 + √
2 t→1− 2 − t2 2 − t2
    
1 1−t 1+t 1
= lim ln √ − ln √ = − ln 2.
2 t→1− 1 − t/ 2 − t 2 1 + t/ 2 − t 2 2
Substituting these values into (∗) yields the desired result.
Editorial comment. Many solvers used the Maclaurin series for (sinh−1 x)2 and then
reversed the order of the integration and summation.

866 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
Also solved by A. Berkane (Algeria), P. Bracken, H. Chen (US), B. E. Davis, S. Fan, G. Fera (Italy),
M. L. Glasser, H. Grandmontagne (France), F. Holland (Ireland), O. Kouba (Syria), J. Magliano, J. Manohar-
mayum (UK), M. Omarjee (France), K. Sarma (India), V. Schindler (Germany), A. Stadler (Switzerland),
A. Stenger, S. M. Stewart (Saudi Arabia), M. S̆tofka (Slovakia), R. Stong, R. Tauraso (Italy), M. Vowe
(Switzerland), M. Wildon (UK), UM6P Math Club (Morocco), and the proposer.

A Disappearing Root of Unity

12298 [2022, 87]. Proposed by George Stoica, Saint John, NB, Canada. Let n be a positive
integer, Sn be the group of all permutations of {1, 2, . . . , n}, and z be a primitive complex
nth root of unity. Prove
 
 n
  n
1 − xj z τ (j )
= n! 1 − xi
τ ∈Sn j =1 i=1

for any x1 , . . . , xn ∈ C.
Solution by José Heber Nieto, University of Zulia, Maracaibo, Venezuela. Since

n 
n 
xn − 1 = (x − zj ) = x n + (−1)k x n−k zi1 +···+ik ,
j =1 k=1 1≤i1 <···<ik ≤n

we have
  
z i1 = zi1 +i2 = · · · = zi1 +···+in−1 = 0
1≤i1 ≤n 1≤i1 <i2 ≤n 1≤i1 <···<in−1 ≤n

and

zi1 +···+in = z1+···+n = (−1)n−1 .
1≤i1 <···<in ≤n

Hence for any nonempty set J ⊆ {1, . . . , n},

   
z j ∈J τ (j ) = zi1 +···+i|J |
τ ∈Sn 1≤i1 <···<i|J | ≤n τ (J )={i1 ,...,i|J | }

 (−1)n−1 n!, if J = {1, . . . , n},
i1 +···+i|J |
= |J |! (n − |J |)! z =
1≤i1 <···<i|J | ≤n
0, otherwise.

We can now compute


n
   
1 − xj zτ (j ) = 1 − xi1 zτ (i1 ) + xi1 xi2 zτ (i1 )+τ (i2 ) − · · ·
j =1 1≤i1 ≤n 1≤i1 <i2 ≤n

+ (−1)n xi1 · · · xin zτ (i1 )+τ (i2 )+···+τ (in ) ,
1≤i1 <i2 <···<in ≤n

and therefore
 n
     
1 − xj zτ (j ) = n! − xi1 zτ (i1 ) + xi1 xi2 zτ (i1 )+τ (i2 ) − · · ·
τ ∈Sn j =1 1≤i1 ≤n τ ∈Sn 1≤i1 <i2 ≤n τ ∈Sn
 
+ (−1)n xi1 · · · xin zτ (i1 )+···+τ (in )
1≤i1 <···<in ≤n τ ∈Sn
 

n
= n! + (−1) 2n−1
n! x1 · · · xn = n! 1− xi .
i=1

November 2023] PROBLEMS AND SOLUTIONS 867

Also solved by N. Caro-Montoya (Brazil), R. Dietmann (UK) & M. Widmer (Switzerland), K. Gatesman,
O. Geupel (Germany), E. A. Herman, N. Hodges (UK), Y. J. Ionin, O. Kouba (Syria), P. Lalonde (Canada),
J. H. Lindsey II, O. P. Lossers (Netherlands), F. Gesmundo (Germany) & T. M. Mazzoli (Austria), T. Amde-
berhan & V. H. Moll, M. Omarjee (France), D. Pinchon (France), G. Plumpton (Canada), M. Reid, K. Sarma
(India), A. Stadler (Switzerland), A. Stenger, R. Stong, R. Tauraso (Italy), T. Wiandt, UM6P Math Club
(Morocco), and the proposer.

Wait Till You See Them All Together

12299 [2022, 87]. Proposed by Erik Vigren, Swedish Institute of Space Physics, Uppsala,
Sweden. For n a positive integer, let x0,n = x1,n = 1 and, for integers k with 2 ≤ k ≤ n − 1,

let xk,n = nxk−1,n − k−1 j =1 xj,n /k. Let Tn = n xn−1,n − n + 1. The first few values of
2

Tn are 1, 3, 7, 47/3, 427/12, 416/5. Prove that Tn is the expected number of throws of an
n-sided die until the last n throws contain all possible face values. For example, if throws
of a 6-sided die give the sequence 12345266426351, then it took 14 throws for the event to
occur.
Solution by Haoran Chen, Xi’an Jiaotong–Liverpool University, Suzhou, China. We prove
that Tn and the expected value both equal Un , where

n−1
nj
Un = 1 + j
.
j =1 i=1 (n − i)

For n = 1, the assertion is true, so we assume n ≥ 2.

Say that the process is in state Sk when the maximum number of distinct throws at the
end of the current list is k. For 0 ≤ k ≤ n, let ek be the expected number of additional
throws when in state Sk until the event occurs. Note that en = 0 and e0 is the desired
expected number of throws. Also e0 = e1 + 1, since there must be a throw to reach S1
from S0 .
We derive a recurrence for ek . When in Sk with 1 ≤ k ≤ n − 1, if the next throw differs
from the previous k, then the process moves to Sk+1 . This event occurs with probability
(n − k)/n. On the other hand, if the next throw equals one of the previous k numbers (each
with probability 1/n), then the process enters one of S1 , . . . , Sk . Hence for 1 ≤ k ≤ n − 1,

1
k
n−k
ek = 1 + ej + ek+1 .
n j =1 n

We solve this recurrence by considering differences:

n−k
ek−1 − ek = (ek − ek+1 ), for 1 ≤ k ≤ n − 1.
n
From e0 − e1 = 1, we derive ek − ek+1 = kj =1 n/(n − j ). By summing these differences
up to k = n − 1, the telescoping sum yields the desired result e0 = Un .
We next obtain the same formula for  Tn . Keeping n fixed, we simplify notation by
writing xj for xj,n . Let s0 = 0 and sk = kj =1 xj for k ≥ 1, so
sk = sk−1 + xk . (1)
We rewrite the given recurrence for xk as
kxk = nxk−1 − sk−1 . (2)
Multiplying (1) by k and combining that with (2) yields
ksk = (k − 1)sk−1 + nxk−1 . (3)

868 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
Writing (3) with index j instead of k and then summing produces


k 
k 
k 
k−1
j sj = (j − 1)sj −1 + nxj −1 = j sj + nsk−1 ,
j =2 j =2 j =2 j =1

and hence

ksk = s1 + nsk−1 = 1 + nsk−1 . (4)

We now prove

1
k
nj
sk = ,
n j =1 k(k − 1) · · · (k + 1 − j )

by induction on k. This formula holds for k = 1 since s1 = 1. For k ≥ 2, (4) yields

1 1 1 1
k−1 k−1
nj nj +1
sk = + = +
k k j =1 (k − 1) · · · (k − j ) k n j =1 k(k − 1) · · · (k − j )

1 n 1 1
k k
nj nj
= · + = .
n k n j =2 k(k − 1) · · · (k + 1 − j ) n j =1 k(k − 1) · · · (k + 1 − j )

Rewriting (3) and (4) as ksk + nxk = (k + 1)sk+1 = 1 + nsk , we have

⎛ ⎞
1 n−k 1 ⎝  k
nj
xk = + sk = 2 n + (n − k) ⎠.
n n n j =1
k(k − 1) · · · (k + 1 − j )

Setting k = n − 1 yields Tn = Un .
Also solved by P. Lalonde (Canada), A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), and the proposer.

Forbidden Permutations
12300 [2022, 186]. Proposed by H. A. ShahAli, Tehran, Iran. Let n be an integer such that
n ≥ 3. Prove that there is no permutation π of {1, 2, . . . , n} such that π(1), 2π(2), . . . ,
nπ(n) are distinct modulo n.
Solution by Allen Stenger, Boulder, CO. Assume that π is such a permutation. Note that
nπ(n) ≡ 0 (mod n), so π(k) = n for k < n. Thus π fixes n. Now restrict k so that
1 ≤ k ≤ n − 1. Define r(k) by kπ(k) ≡ r(k) (mod n) with 1 ≤ r(k) ≤ n − 1. Both π
and r permute {1, . . . , n − 1}.
Write gcd(a, b) for the greatest common divisor of a and b. Note that ab ≡ c (mod n)
implies gcd(a, n) | gcd(c, n). Applying this observation when {a, b} = {k, π(k)} and c =
r(k) yields

gcd(k, n) | gcd(r(k), n) and gcd(π(k), n) | gcd(r(k), n).

The first divisibility gives gcd(k, n) ≤ gcd(r(k), n). Summing over k and observing that k
and r(k) run through the same values yields


n−1 
n−1 
n−1
gcd(k, n) ≤ gcd(r(k), n) = gcd(k, n).
k=1 k=1 k=1

November 2023] PROBLEMS AND SOLUTIONS 869

Thus we have gcd(k, n) = gcd(r(k), n) for each k. Applying the same argument to π(k)
and r(k) yields
gcd(k, n) = gcd(π(k), n) = gcd(r(k), n). (∗)
We now prove that n is squarefree. Suppose that p2 | n for some prime p. When k = p,
we have gcd(k, n) = p, and then (∗) implies also gcd(π(k), n) = p and gcd(r(k), n) = p.
Now π(k) is a multiple of p, so kπ(k) is a multiple of p2 . Since p2 | n and kπ(k) ≡ r(k)
(mod n), we have p2 | r(k). Thus p2 | gcd(r(k), n) = p, a contradiction.
Since n ≥ 3 and n is squarefree, n cannot be a power of 2. Thus n is divisible by some
odd prime p, and p is relatively prime to n/p. Let S = {n/p, 2n/p, . . . , (p − 1)n/p}. The
set S is the set of values of k such that gcd(k, n) = n/p.
By (∗), {k : k ∈ S} = {π(k) : k ∈ S} = {r(k) : k ∈ S}. Writing A for k∈S k, we then
have
  
   
A= r(k) ≡ kπ(k) = k π(k) = A2 (mod p).
k∈S k∈S k∈S k∈S

However, A = (n/p)p−1 (p − 1)! ≡ 1 · (−1) (mod p), where we have used Fermat’s little
theorem in the first factor and Wilson’s theorem in the second factor. Now A ≡ A2 becomes
−1 ≡ 1 (mod p), which is false when p is an odd prime.
Also solved by C. P. Anil Kumar (India), T. Beran & F. Fürnsinn & F. Lang & S. Schneider & M. Reibnegger
& S. Yurkevich (Austria), J. Boswell & C. Curtis, N. Caro-Montoya (Brazil), W. Chang, A. De la Fuente,
C. Farnsworth, N. Fellini (Canada), K. Gatesman, O. Geupel (Germany), N. Hodges (UK), Y. J. Ionin,
W. Janous (Austria), Y. Kim (Korea), O. P. Lossers (Netherlands), J. Manoharmayum (UK), M. Reid, T. Song,
A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), M. Velásquez (Colombia), and the proposer.

A Skew-Symmetric Determinant of Sines

12302 [2022, 186]. Proposed by Moubinool Omarjee, Lycée Henri IV, Paris, France. Let
n be a positive integer, and let A be the 2n-by-2n skew-symmetric matrix with (j, k)-entry
sin(j − k)/ sin(j + k). Prove
  sin(j − k) 2
det (A) = .
1≤j <k≤2n
sin(j + k)

Solution by Pierre Lalonde, Plessisville, QC, Canada. We have

sin(j − k) ei(j −k) − e−i(j −k) e2ij − e2ik
= i(j +k) = .
sin(j + k) e − e−i(j +k) e2ij e2ik − 1
For j, k ≥ 1 define aj,k = (xj − xk )/(xj xk − 1), where each x is an indeterminate. For
a positive integer r, let Ar = (aj,k )rj,k=1 , which we observe is skew-symmetric. We prove
the more general result

det(A2n ) = 2
aj,k ,
1≤j <k≤2n

which yields the desired conclusion when x is set to e2i for all .
We prove the claim by induction on n. The case n = 1 is easily checked. Elementary
algebra yields
(xj − xk )(xk − xm )(xm − xj )
aj,k + ak,m + am,j = − = −aj,k ak,m am,j .
(xj xk − 1)(xk xm − 1)(xm xj − 1)

870 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
We now operate on Ar for r = 2n with n > 1. Subtract the last row from every other row.
Subtract the resulting last column from every other column. This does not change the value
of the determinant. The (j, k)-entry for j, k < r becomes
(aj,k − ar,k ) − (aj,r − ar,r ) = aj,k + ak,r + ar,j = −aj,k ak,r ar,j = aj,k ak,r aj,r .
Hence for j < r, each element of row j contains aj,r as a factor and, for k < r, every
element of column k contains ak,r (which equals −ar,k ) as a factor. Thus
⎡ ⎤
1
⎢ .. ⎥ 
⎢ Ar−1 . ⎥
det(Ar ) = det ⎢ ⎥ 2
aj,r .
⎣ 1 ⎦ 1≤j <r
−1 . . . −1 0
Continuing the process, we subtract the penultimate row from rows 1 to r − 2 and the
penultimate column from columns 1 to r − 2. Again, this does not change the determinant.
We have
⎡ ⎤
1 0
⎢ .. .. ⎥
⎢ . ⎥
⎢ Ar−2 . ⎥  
det(Ar ) = det ⎢ 1 0 ⎥ 2
aj,r−1 2
aj,r .
⎢ ⎥
⎣ −1 . . . −1 0 1 ⎦ 1≤j <r−1 1≤j <r

0 ... 0 −1 0
Expansion of the determinant along the last row and the last column yields
 
det(Ar ) = det(Ar−2 ) 2
aj,r−1 2
aj,r .
1≤j <r−1 1≤j <r

With r = 2n, the induction hypothesis completes the proof.

Also solved by T. Amdeberhan & S. B. Ekhad, N. Caro-Montoya (Brazil), O. P. Lossers (Netherlands), B. Ly,
A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), and the proposer.

CLASSICS
C19. Due to John H. Conway, suggested by the editors. A battlefield is modeled by an
infinite grid of unit squares, whose centers are indexed by {(a, b) : a, b ∈ Z}. The soldiers
are modeled by pegs, which are placed initially at a finite number of squares (a, b) with
b ≤ 0. The soldiers advance by jumping in the style of peg solitaire: A jump is permitted
when there are three squares in the grid forming a 1-by-3 rectangle with one end square
of this rectangle empty while the other two squares are occupied by pegs. Where this
configuration exists, the peg on the end may jump over the peg in the middle and move to
the empty end, while the peg in the middle is removed. How many pegs are needed in an
initial configuration to allow a peg to advance to the square (0, 5)?

November 2023] PROBLEMS AND SOLUTIONS 871

 Guessing Which of Two Numbers is Larger
C18. Due to Thomas Cover; suggested by Richard Stanley. Alice chooses two distinct
numbers and writes each of them on a slip of paper. Bob selects one of the two slips
at random and looks at the number on it. He must then choose to either keep that slip or
switch to the other slip. Bob wins if he ends up with the slip with the larger number. Is there
anything Bob can do to ensure that, no matter what numbers Alice chooses, his probability
of winning is greater than 1/2?
Solution. Yes, there is a randomized strategy that achieves Bob’s goal. Let f : R → (0, 1)
be a strictly increasing function. For example, f could be defined by
1 1
f (x) = arctan x + .
π 2
If Bob sees the number x, he keeps his initial selection with probability f (x) and switches
to the other slip with probability 1 − f (x).
To see that this works, suppose Alice chooses numbers a and b, with a < b. Bob ends
up with the larger number if he either chooses a initially and switches or chooses b initially
and keeps it. Therefore his probability of success is (1/2)(1 − f (a)) + (1/2)f (b), which
equals 1/2 + (1/2)(f (b) − f (a)). This is strictly larger than 1/2.
Editorial Comment. One way to understand Bob’s strategy is to imagine that f is the
cumulative distribution function of some continuous random variable whose support is the
entire real line. Bob selects a real number at random according to the cumulative distribu-
tion function. He then imagines that this number is on the other slip and acts accordingly,
keeping the first slip if the number on the slip exceeds the random real and switching if the
random real exceeds the number on the slip. If the random real is less than both of Alice’s
numbers, then no matter which slip Bob selects initially, he keeps it, so he wins with prob-
ability 1/2. Similarly, if the random real is greater than both of Alice’s numbers, then Bob
always switches, again winning with probability 1/2. But if the random real lands between
Alice’s numbers, an event of positive probability, then Bob wins no matter which slip he
selects first.
How Alice chooses her numbers is irrelevant. Bob’s strategy triumphs for all choices
that Alice can make. The probability mentioned in the problem statement is not to be
misread as a probability over Alice’s possible choices.
If one does not permit Bob to randomize his responses—that is to say, if Bob’s space of
possible actions is limited to deterministic strategies governed by a choice function from R
to {keep, switch}—then the answer to the question is negative. This illustrates the power
of randomized strategies and also may explain why the affirmative answer seems to be
paradoxical.
The fact that f is strictly increasing means that the larger the number that Bob sees, the
more likely he is to keep it. This is an intuitively reasonable guideline for Bob to follow if
he wants to end up with the larger number.
Notice that limx→∞ (f (x + 1) − f (x)) = 0. Thus for every positive , if Alice chooses
the numbers x and x + 1 for sufficiently large x, then Bob’s probability of success will be
less than 1/2 + .
An extensive review of this problem appears in A. Gnedin (2016), Guess the larger
number, Mathematica Applicanda, 44(1): 183–207, where it is suggested that the phrasing
of the problem as a guessing game is due to T. M. Cover (1987), Pick the largest number,
Open Problems in Communication and Computation, New York, NY: Springer, p. 152.

872 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
12417. Proposed by Mohsen Maesumi, Lamar University, Beaumont, TX. Consider the
sphere S given by x 2 + y 2 + (z − 1)2 = 1, with north pole N at (0, 0, 2). The stereographic
projection of a point P at (x, y, 0) is the point, different from N , that is on the intersection
of NP with S. Consider the region H in the xy-plane given by 0 ≤ xy ≤ c2 , where c > 0.
What is the area of the stereographic projection of H to S?
12418. Proposed by Vladimir Lucic, Imperial College, London, UK. Let  be the cumu-
lative distribution function of a standard normal random variable, defined by (x) =
√ x 2
(1/ 2π ) −∞ e−t dt. 
(a) For positive real numbers σ1 , . . . , σn and w1 , . . . , wn with ni=1 wi = 1, determine
 n  
1 −1  x
lim  wi  .
x→∞ x σi
i=1

(b) Let L be the limit in (a). Determine

  n   
1 −1  x
lim x 2
 wi  −L .
x→∞ x i=1
σi

SOLUTIONS

Commuting Orthogonal Projections

12283 [2021, 856]. Proposed by Yongge Tian, Shanghai Business School, Shanghai, China.
∗
√ matrices that are orthogonal projections, that is, A = A = A
2
Let A and B be two n-by-n
∗
and B = B = B . Let A + B denote the positive semidefinite square root of A + B.
2

Prove
√ √
trace(A + B)−(2 − 2)rank(AB) ≤ trace A + B
√ √
≤ ( 2 − 1)trace(A + B) + (2 − 2)rank(A + B),
and show that equality holds simultaneously if and only if AB = BA.
Solution by Kyle Gatesman, student, Johns Hopkins University, Baltimore MD. For an n-
by-n Hermitian matrix H and an integer k ∈ {1, . . . , n}, let λk (H ) be the kth smallest
eigenvalue of H , with repetitions according to algebraic multiplicity. All eigenvalues of
a Hermitian matrix are real, by the spectral theorem, so the ordering of these eigenvalues
is well defined. Extend this notation to all integers k by letting λk (H ) = ∞ for k > n
and λk (H ) = −∞ for k < 1. The spectral theorem also says that Hermitian matrices are
diagonalizable, so algebraic and geometric multiplicity are the same for all eigenvalues.
Thus the rank of a Hermitian matrix H equals |{k ∈ {1, . . . , n} : λk (H ) = 0}|.
A projection matrix P satisfies P 2 − P = 0, so any eigenvalue λ of P satisfies
λ − λ = 0, which implies λ ∈ {0, 1}. The matrices A and B are Hermitian with solely
2

nonnegative eigenvalues, so they are positive semidefinite. The sum of any two n-by-n
positive semidefinite matrices is also positive semidefinite, so A + B is positive semidefi-
nite.
For an n-by-n orthogonal projection matrix P and an x ∈ Cn ,
(P x)∗ (x − P x) = x ∗ P ∗ x − x ∗ P ∗ P x = x ∗ (P − P 2 )x = 0,

so (P x) ⊥ (x − P x), and thus P x 2 = x 22 − x − P x 22 ≤ x 2 . All eigenvalues of

A + B are at most 2 because, for any nonzero vector x ∈ Cn ,
(A + B)x 2 ≤ Ax 2 + Bx 2 ≤ x 2 + x 2 = 2 x 2.

766 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
 The next few lemmas build up to a critical theorem about the eigenvalues of (A + B)
that are strictly between 0 and 2.
Lemma 1. If (x, λ) is an eigenvector-eigenvalue pair for A + B with 0 < λ < 2, then
((A − B)x, 2 − λ) is also an eigenvector-eigenvalue pair for A + B.
Proof. Given Ax + Bx = λx, multiplying on the left by A yields A2 x + ABx = A(λx),
which implies Ax + ABx = λAx, or ABx = (λ − 1)Ax. Similarly, multiplying on the
left by B yields BAx + B 2 x = B(λx), which implies BAx + Bx = λBx, or equivalently
BAx = (λ − 1)Bx. Letting y = (A − B)x, we obtain
(A + B)y = (A + B)(A − B)x = (A − B + BA − AB)x
= (A − B)x + (BAx − ABx) = y + ((λ − 1)Bx − (λ − 1)Ax)
= y − (λ − 1)(A − B)x = y + (1 − λ)y = (2 − λ)y,
so ((A − B)x, 2 − λ) is an eigenvector-eigenvalue pair. Note that (A − B)x cannot be the
zero vector, because (A − B)x = 0 ⇐⇒ Ax = Bx ⇐⇒ λx = (A + B)x = 2Ax, and
the only possible eigenvalues of 2A are 0 and 2, which by assumption cannot equal λ. 
Definition: For eigenvectors x and y of A + B associated with eigenvalues strictly between
0 and 2, let y be a dual of x if y is a nonzero scalar multiple of (A − B)x. By Lemma 1, if
y is a dual of x, then the associated eigenvalues of x and y sum to 2.
Lemma 2. For any two eigenvectors x and y of A + B associated with eigenvalues strictly
between 0 and 2, if y is a dual of x, then x is a dual of y.
Proof. Let λ be the eigenvalue associated with x. If y is a dual of x, then for some nonzero
scalar γ we have y = γ · (A − B)x. From the proof of Lemma 1, (A + B)x = λx implies
ABx = (λ − 1)Ax and BAx = (λ − 1)Bx. Now
(A − B)y = γ (A − B)2 x = γ (A2 + B 2 )x − (AB + BA)x
= γ ((A + B)x − (λ − 1)(A + B)x)
= γ (λx − (λ − 1)λx) = γ λ(2 − λ)x.
Since λ ∈
/ {0, 2}, the quantity γ λ(2 − λ) is nonzero. Thus x equals 1/ γ λ(2 − λ)
(A − B)y and is a dual of y. 
Lemma 3. Let u1 , . . . , uN and v1 , . . . , vN be eigenvectors of A + B corresponding to
eigenvalues strictly between 0 and 2, such that uk and vk are duals of each other for all
k ∈ {1, . . . , N }. The vectors u1 , . . . , uN are linearly independent if and only if v1 , . . . , vN
are linearly independent.
Proof. By the symmetry of the duality relationship between uk and vk , it suffices to show
that if u1 , . . . , uN are linearly dependent, then v1 , . . . , vN are also linearly dependent.
Given u1 , . . . , uN and v1 , . . . , vN , let r1 , . . . , rN be the (unique) nonzero scalars satisfying
vk = rk · (A − B)uk for allk ∈ {1, . . . , N}. Suppose some nonzero vector (c1 , . . . , cN )
expresses dependence by N 
k=1 ck uk = 0. At least one entry in (c1 /r1 , . . . , cN /rN ) is
nonzero, and

N
ck 
N
ck 
N
· vk = · rk · (A − B)uk = (A − B) 
ck uk = 0,
k=1
rk k=1
rk k=1

so v1 , . . . , vN are linearly dependent. 

Let m = min{k ∈ Z : λk (A + B) > 0}, and let s = max{k ∈ Z : λk (A + B) < 2}.
Since the rank of a positive semidefinite matrix equals the number of positive eigenvalues
(counted with multiplicity), rank(A + B) = n − m + 1. This implies that the nullity of
A + B is m − 1.

October 2023] PROBLEMS AND SOLUTIONS 767

Theorem 1. If m ≤ k ≤ s, then λk (A + B) + λm+s−k (A + B) = 2.
Proof. Let W be the span of the eigenvectors of A + B corresponding to eigenvalues in
(0, 2), and let I denote the interval of integers from m to s. Given a basis {um , . . . , us } of
W whose members are eigenvectors of A + B, for k ∈ I let vk be any dual of uk . Since
um , . . . , us are linearly independent eigenvectors of A + B, by Lemma 3, vm , . . . , vs are
also linearly independent eigenvectors and thus also form a basis of W .
For an eigenvector x of A + B, let λ(x) denote the eigenvalue of A + B associated with
x. Index {um , . . . , us } so that λ(uk ) = λk (A + B) for k ∈ I . For indices j, k ∈ I with j < k,
we have λ(uj ) ≤ λ(uk ) , so
λ(vj ) = 2 − λ(uj ) ≥ 2 − λ(uk ) = λ(vk ) .
Thus the list (λ(vm ) , . . . , λ(vs ) ) is precisely the reverse of the list (λ(um ) , . . . , λ(us ) ). We
conclude λ(vk ) = λ(um+s−k ) for k ∈ I , so
2 = λ(uk ) + λ(vk ) = λ(uk ) + λ(um+s−k ) = λk (A + B) + λm+s−k (A + B). 
s
Corollary 1. k=m λk (A + B) = s − m + 1.

Proof. By Theorem 1,
s 
s 
s
2 λk (A+B) = (λk (A+B)+λm+s−k (A+B)) = 2 = 2(s −m+1). 
k=m k=m k=m

By Corollary 1,

n 
s 
m−1
trace(A + B) = 2+ λk (A + B) + 0
k=s+1 k=m k=1

= 2(n−s) + (s −m+1) = (n−s) + (n−m+1)

= (n − s) + rank(A + B).
Similarly,
√ n
√ 1
s 
trace A+B = 2+ λk (A+B) + λm+s−k (A+B)
k=s+1
2 k=m

√ 1
s 
= 2(n − s) + λk (A+B) + 2 − λk (A+B) .
2 k=m
Also,
√ √
( 2 − 1)trace(A + B) + (2 − 2)rank(A + B)
√ √ 
= ( 2 − 1) (n − s) + rank(A + B) + 2 rank(A + B)
√ √ √
= ( 2 − 1)(n − s) + ( 2 − 1)( 2 + 1)rank(A + B)
√
= ( 2 − 1)(n − s) + rank(A + B)
√ √
= 2(n − s) − (n − s) + (n − m + 1) = 2(n − s) + s − m + 1.
Hence the inequality on the right in the problem statement is equivalent to
s √ √
λk (A + B) + 2 − λk (A + B)
≤ s − m + 1.
k=m
2

768 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
Because the square root function is strictly concave over (0, ∞), we in fact have the
stronger inequality
√ √ 
λk (A + B) + 2 − λk (A + B) 1 1
≤ λk (A + B) + 2 − λk (A + B) = 1
2 2 2
for m ≤ k ≤ s. Thus the inequality on the right is true, and equality occurs if and only if
λk (A + B) = 1 for m ≤ k ≤ s, which holds if and only if all eigenvalues of A + B belong
to the set {0, 1, 2}.
We next prove the inequality on the left. We use the well-known fact that if X and Y
are n-by-n complex matrices, then XY and Y X have the same characteristic polynomial
and therefore the same spectrum. (See W. V. Parker (1953), The matrices AB and BA, this
Monthly, 60(5): 316.)
Lemma 4. All eigenvalues of AB are nonnegative real numbers.
√ √ √
Proof. Since A is positive
√ semidefinite,
√ AB is well-defined. Since √AB = A( AB),
we know that√AB and √ ( ∗ AB) A have the same eigenvalues. Since A is also positive
semidefinite, A = A . Hence for x ∈ Cn ,
√ √ √ √
x ∗ ( AB A)x = ( Ax)∗ B( Ax) ≥ 0,
√ √
where the last step follows from the fact that B is positive semidefinite. Thus AB A
is positive semidefinite, so all its eigenvalues are nonnegative real numbers. Hence all
eigenvalues of AB are also nonnegative real numbers. 
Even if AB is not Hermitian, Lemma 4 implies that all eigenvalues of AB are real and
thus have a well-defined ordering. Hence we can designate the kth smallest eigenvalue of
the matrix AB as λk (AB).
Lemma 5. All eigenvalues of AB are at most 1.
Proof. We showed earlier that P x 2 ≤ x 2 for any n-by-n orthogonal projection matrix

P and vector x ∈ Cn . Thus for nonzero x ∈ Cn with Bx = 0,
ABx 2 A(Bx) 2 Bx 2
= · ≤ 1 · 1 = 1.
x 2 (Bx) 2 x 2
 then ABx 2 / x 2 = 0 ≤ 1. For any eigenvector-eigenvalue pair (x, λ) of AB,
If Bx = 0,
we know ABx 2 / x 2 = λx 2 / x 2 = λ, so λ ≤ 1. 
Lemma 6. rank(AB) ≥ trace(AB).
Proof. The nullity of AB is the geometric multiplicity of 0 as an eigenvalue, which
is at most its algebraic multiplicity. Letting m = min{k ∈ Z : λk (AB) > 0}, we have
rank(AB) = n − nullity(AB) ≥ n − m + 1. By Lemma 5,

n 
n
n − m + 1 ≥ λk (AB) = λk (AB) = trace(AB),
k=m k=1

so rank(AB) ≥ trace(AB). Equality holds if and only if the geometric and algebraic mul-
tiplicities of the eigenvalue 0 of AB are equal and all eigenvalues of AB are 0 or 1. 
√ √
By Lemma 6, the result trace(A + B) − (2 − 2)rank(AB) ≤ trace A + B follows
from the stronger
√ √
(2 − 2)trace(AB) ≥ trace(A + B) − trace A + B, (1)

October 2023] PROBLEMS AND SOLUTIONS 769

 √
which we now prove. Our expressions for trace(A + B) and trace A + B yield
√ √ 
s 
trace(A+B) − trace A+B = (2 − 2)(n − s) + λk (A+B) − λk (A+B)
k=m

√ 1 
s 
= (2 − 2)(n − s) + 2− λk (A + B) − 2 − λk (A + B)
2 k=m

√ 1
s 
= (2 − 2)(n − s) + 2− 1 + αk − 1 − αk ,
2 k=m
where αk = λk (A + B) − 1 ∈ (−1, 1) for m ≤ k ≤ s. Since trace(AB) = trace(BA) and
(A + B)2 = A2 + B 2 + AB + BA = A + B + AB + BA , we have
1
n
1
trace(AB) = trace((A+B)2 ) − trace(A+B) = λk (A+B)2 − λk (A+B)
2 2 k=1
 
1  s  s
= (2 −2)(n−s) +
2
λk (A+B) −
2
λk (A+B) + (0 −0)(m−1)
2
2 k=m k=m

1
s
= (n − s) + λk (A + B)2 + (2 − λk (A + B))2 − 2
4 k=m

1 1 2
s s
= (n − s) + (1 + αk )2 + (1 − αk )2 − 2 = (n − s) + α ,
4 k=m 2 k=m k
where the third line follows from the second by Lemma 1. Thus (1) is equivalent to
√ s 
√ 2− 2  2 √ 1
s
(2 − 2)(n − s) + αk ≥ (2 − 2)(n − s) + 2− 1+αk − 1−αk ,
2 k=m
2 k=m
which in turn is equivalent to

s
√ 
(2 − 2)αk2 + 1 + αk + 1 − αk − 2 ≥ 0.
k=m
√ √ √
It suffices to prove the stronger inequality (2 − 2)α 2 + 1 + α + 1 − α − 2 ≥ 0
for α ∈ (−1, 1). This stronger inequality is equivalent to
√ √ √ √
2 + (2 − 2)(1 − α 2 ) ≤ 1 + α + 1 − α.
Since both sides√of this new
√ inequality are
√ nonnegative, we can square√ both sides to obtain
the equivalent 2 2(2 − 2)β 2 + (2 − 2)2 β 4 ≤ 2β, where β = 1 − α 2 ∈ (0, 1]. Since
β is positive, we can divide by β to reduce to the equivalent inequality
√ √ √
2 2(2 − 2)β + (2 − 2)2 β 3 ≤ 2. (2)
√ √ √ 2 3
The function mapping β to 2 2(2 − 2)β + (2 − 2) β is strictly increasing on [0, 1],
and at β = 1 the value is 2, so (2) holds. Thus the desired inequality
√ √
(2 − 2)trace(AB) ≥ trace(A + B) − trace A + B
holds for all orthogonal projection matrices A and B, with equality if and only if αk = 0
for m ≤ k ≤ s, which happens if and only if all eigenvalues of A + B lie in the set {0, 1, 2}.
This implies the original inequality on the left in the problem statement,
√ √
trace(A + B) − (2 − 2)rank(AB) ≤ trace A + B,

770 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
with equality if and only if (a) all eigenvalues of A + B lie in {0, 1, 2}, (b) all eigenvalues of
AB lie in {0, 1}, and (c) 0 has equal algebraic and geometric multiplicity as an eigenvalue
of AB. As we showed earlier, the original inequality on the right holds with equality if
and only if (a) holds. We now show that (a) is equivalent to AB = BA. Finally, we show
that AB = BA implies (b) and (c) to complete the proof that AB = BA is necessary and
sufficient for equality in each inequality in the problem statement.
We analyze the three subspaces {x : (A + B)x = λx} for λ ∈ {0, 1, 2}. First, consider
λ = 0. Since A + B is positive semidefinite, (A + B)x = 0 implies 0 = x ∗ (A + B)x =
x ∗ Ax + x ∗ Bx, which implies 0 = x ∗ Ax = (Ax)∗ (Ax) and 0 = x ∗ Bx = (Bx)∗ (Bx), or
Ax = Bx = 0.  Thus {x : (A + B)x = 0}  = ker A ∩ ker B.
Next, consider λ = 2. If (A + B)x = 2x, then x satisfies equality in both halves of
(A + B)x 2 ≤ Ax 2 + Bx 2 ≤ x 2 + x 2,
which occurs if and only if Ax = Bx = x. Thus {x : (A + B)x = 2x} = imA ∩ imB.
Finally, consider λ = 1. Note that
(A + B)x = x ⇐⇒ Bx = (I − A)x

=⇒ A(Bx) = (A − A2 )x = 0 ⇐⇒ Bx ∈ ker A.
Similarly, (A + B)x = x implies Ax ∈ ker B. Thus a necessary condition for (A + B)x = x
is the existence of vectors xA ∈ imA ∩ ker B and xB ∈ imB ∩ ker A satisfying xA + xB = x.
This condition is also sufficient, since if x admits such a decomposition, then
(A + B)x = (A + B)(xA + xB )

= AxA + BxA + AxB + BxB = xA + 0 + 0 + xB = x.

Thus {x : (A + B)x = x} is the direct sum of the orthogonal subspaces imA ∩ ker B
and ker A ∩ imB. The orthogonality of these subspaces crucially means that the subspace
{x : (A + B)x = x} has an orthonormal basis that can be partitioned into orthonormal
bases of imA ∩ ker B and ker A ∩ imB.
Since A + B is unitarily diagonalizable (by the spectral theorem), A + B admits an
orthonormal basis of n eigenvectors, so the eigenvectors of A + B are all associated with
eigenvalues in the set {0, 1, 2} if and only if there is an orthonormal basis of Cn that
partitions into orthonormal bases of the pairwise orthogonal subspaces ker A ∩ ker B,
imA ∩ imB, imA ∩ ker B, and ker A ∩ imB. This condition holds if and only if there is an
n-by-n unitary matrix U whose columns all lie in (ker A ∪ imA) ∩ (ker B ∪ imB), which
is equivalent to saying that the columns of U are eigenvectors of both A and B.
Thus the eigenvalues of A + B all belong to {0, 1, 2} if and only if A and B are simul-
taneously unitarily diagonalizable, meaning that there are diagonal matrices DA and DB
and a single unitary matrix U satisfying A = U DA U ∗ and B = U DB U ∗ . This holds if
and only if AB = BA, by a well-known result in matrix analysis stating that two diag-
onalizable matrices commute if and only if they are simultaneously diagonalizable. (See
R. A. Horn and C. R. Johnson (2013), Matrix Analysis, 2nd ed., Cambridge University
Press.)
It remains to show that AB = BA implies the conditions (b) and (c) stated earlier.
As just mentioned, AB = BA implies the existence of diagonal matrices DA and DB
and a single unitary matrix U satisfying A = U DA U ∗ and B = U DB U ∗ , so AB =
(U DA U ∗ )(U DB U ∗ ) = U (DA DB )U ∗ . Thus the matrix DA DB displays the eigenvalues of
AB along its main diagonal, which is the elementwise product of the main diagonals of DA
and DB . Since all entries in DA and DB are 0 or 1, all such elementwise products are also
0 or 1, so all eigenvalues of AB belong to {0, 1}, which implies (b). Finally, AB = BA

October 2023] PROBLEMS AND SOLUTIONS 771

implies that AB is diagonalizable, which means that every eigenvalue of AB has equal
algebraic and geometric multiplicity, implying (c).
Also solved by E. A. Herman, R. A. Horn, and the proposer.

Alternating Coefficients of Powers of Polynomials

12286 [2021, 946]. Proposed by Ira Gessel, Brandeis University, Waltham, MA. Let p be a
prime number, and let m be a positive integer not divisible by p. Show that the coefficients
of (1 + x + · · · + x m−1 )p−1 that are not divisible by p are alternately 1 and −1 modulo p.
For example, (1 + x + x 2 + x 3 )4 ≡ 1 − x + x 4 − x 6 + x 8 − x 11 + x 12 (mod 5).
Solution by Jayanta Manoharmayum, University of Sheffield, Sheffield,
 UK. Consider a
formal power series with  coefficients in Zp given by f (x) = ∞ k=0 ak x with a0 = 1.
k
−1 ∞ n
Letting (1 − x) f (x) = n=0 bn x , we have bn = k=0 ak . Hence b0 = 1 and bn =
n

bn−1 + an for n ≥ 1. We conclude that the nonzero coefficients of f alternate between 1

and −1 if and only if each coefficient of (1 − x)−1 f (x) is 0 or 1.
It therefore suffices to show that each nonzero coefficient in the expansion of
(1 − x)−1 (1 + x + · · · + x m−1 )p−1
is 1. Using (a + b)p = a p + bp (modulo p), we compute
 
1 − xm p 1 − x (1 − x pm )(1 − x)
(1 + x + · · · + x m−1 )p−1 = · = ,
1−x 1−x m (1 − x p )(1 − x m )
and thus
1 + x m + · · · + x m(p−1)
(1 − x)−1 (1 + x + · · · + x m−1 )p−1 =
1 − xp

p−1
= x km 1 + x p + x 2p + · · · .
k=0

Since m and p are relatively prime, the exponents 0, m, 2m, . . . , (p − 1)m are distinct
modulo p. It follows that each power of x appears in the series at most once, so each
coefficient is either 0 or 1.
Editorial comment. The proof generalizes directly to show that, for m and n relatively
prime, the nonzero coefficients of
(1 − x mn )(1 − x)
(1 − x n )(1 − x m )
alternate between 1 and −1.
Also solved by T. Amdeberhan & V. H. Moll, N. Caro-Montoya (Brazil), J.-P. Grivaux (France), N. Hodges
(UK), Y. J. Ionin, J. H. Lindsey II, P. W. Lindstrom, O. P. Lossers (Netherlands), A. Pathak (India), M. Reid,
A. Stadler (Switzerland), A. Stenger, R. Stong, B. Sury (India), R. Tauraso (Italy), M. Tetiva (Romania),
M. Wildon (UK), L. Zhou, and the proposer.

An Application of the Jacobi Triple Product

12289 [2021, 946]. Proposed by George E. Andrews, Pennsylvania State University, Uni-
versity Park, PA, and Mircea Merca, University of Craiova, Craiova, Romania. Prove
∞   ∞
(2n + 1)π
2 cos q n(n+1)/2 = (1 − q n )(1 − q 6n−1 )(1 − q 6n−5 ),
n=0
3 n=1

when |q| < 1.

772 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
Solution by Rishabh Sarma, University of Florida, Gainesville, FL. Let L denote the left
side of the desired identity. With ω = e2πi/3 , we have
     
ωn−1 + ω−(n−1) 2π(n − 1) (2n + 1)π (2n + 1)π
= cos = cos π − = − cos .
2 3 3 3
When m = −n − 1, we have 2m + 1 = −(2n + 1) and m(m + 1)/2 = n(n + 1)/2. This
and the computation above yield
∞   ∞
(2n + 1)π 1 
L= cos q n(n+1)/2 = − ωn−1 + ω−(n−1) q n(n+1)/2 . (1)
n=−∞
3 2 n=−∞

For z = 0 and |q| < 1, the Jacobi triple product identity states
∞
 ∞

2
zn q n = (1 − q 2n+2 )(1 + zq 2n+1 )(1 + z−1 q 2n+1 ). (2)
n=−∞ n=0
1 1
Writing (2) using q 2 instead of q and then letting z = ωq 2 and multiplying by ω−1 , we
obtain
∞ ∞
ωn−1 q n(n+1)/2 = ω−1 (1 − q n+1 )(1 + ωq n+1 )(1 + ω−1 q n )
n=−∞ n=0
∞

= ω−1 (1 + ω−1 ) (1 − q n+1 )(1 + ωq n+1 )(1 + ω−1 q n+1 ). (3)
n=0
1 1
Similarly, writing (2) using q 2 instead of q and then letting z = ω−1 q 2 and multiplying
by ω yields
∞
 ∞

ω−(n−1) q n(n+1)/2 = ω (1 − q n+1 )(1 + ω−1 q n+1 )(1 + ωq n )
n=−∞ n=0
∞

= ω(1 + ω) (1 − q n+1 )(1 + ω−1 q n+1 )(1 + ωq n+1 ). (4)
n=0

Note that ω−1 (1 + ω−1 ) + ω(1 + ω) = −2. In addition,

(1 + ωx n )(1 + ω−1 x n ) = 1 − x n + x 2n = (1 + x 3n )/(1 + x n ).
Thus substituting (3) and (4) into (1) yields
∞
1 −1
L=− ω (1 + ω−1 ) + ω(1 + ω) (1 − q n+1 )(1 + ω−1 q n+1 )(1 + ωq n+1 )
2 n=0
∞
 ∞
 1 + q 3n
= (1 − q n+1 )(1 + ω−1 q n+1 )(1 + ωq n+1 ) = (1 − q n ) .
n=0 n=1
1 + qn
Finally, the claimed identity follows from
∞
 ∞
1 + q 3n (1 − q 6n )(1 − q n )
=
n=1
1 + qn n=1
(1 − q 3n )(1 − q 2n )
∞
 ∞

1 − q 2n−1
= = (1 − q 6n−1 )(1 − q 6n−5 ),
n=1
1 − q 6n−3 n=1

October 2023] PROBLEMS AND SOLUTIONS 773

where in the second line we have in two stages canceled common factors in the numerator
and denominator.
Also solved by T. Amdeberhan & V. H. Moll, K. Banerjee & M. G. Dastidar (Austria), A. Berkane (Algeria),
H. Chen (US), R. Hemmecke (Austria), N. Hodges (UK), W. P. Johnson, P. Lalonde (Canada), K.-W. Lau
(China), J. Manoharmayum (UK), R. Molinari, M. Omarjee (France), Z. Shen (China), A. Stadler (Switzer-
land), A. Stenger, R. Stong, R. Tauraso (Italy), M. Wildon (UK), L. Zhou, and the proposer.

Analytic Solutions of a Functional Equation

12290 [2021, 946]. Proposed by Walther Janous, Ursulinengymnasium, Innsbruck, Austria.
Find all analytic functions f : C → C that satisfy
|f (x + iy)|2 = |f (x)|2 + |f (iy)|2
for all real numbers x and y.
Solution by Raymond Mortini, Université du Luxembourg, Esch-sur-Alzette, Luxembourg,
and Rudolf Rupp, Technische Hochschule Nürnberg, Nürnberg, Germany. We show that
the solutions are given by f (z) = az, f (z) = a sin(kz), and f (z) = a sinh(kz), where
a ∈ C and k ∈ R. It is easy to check that each of these satisfies the given equation, using
the identities
sin(x + iy) = sin x cosh y + i cos x sinh y
and
sinh(x + iy) = sinh x cos y + i cosh x sin y.
Conversely, suppose that f is an analytic function satisfying the given equation. By
setting x = y = 0 we see that f (0) = 0. Now let h(x + iy) = |f (x + iy)|2 = (f f )
(x + iy). Note that
∂2
hxy (x + iy) = |f (x)|2 + |f (iy)|2 = 0.
∂y∂x
Also, fx = f  , fy = ifx = if  , fxy = (f  )y = if  , and f y = (fy ) = ifx = −if x . Hence
0 = hxy = fxy f + f f xy + fx f y + fy f x = 2 Re(fxy f ) = −2 Im(f  f ).
Since |f |2 = f f is a real-valued function, it follows that the meromorphic function
f  /f = f  f /|f |2 is real-valued where defined. If we assume that f is not identically
zero, then this can happen only when f  /f is a real constant λ. √ √
The differential√ equation f√ = λf in C has solutions az + d if λ = 0, αe λz + βe− λz
if λ > 0, and√αei |λ|z + βe−i |λ|z if λ < 0. Since f (0) = 0, we have d = 0 or β = −α.
Setting k = |λ| yields the claimed expressions for f (z).
Also solved by O. Kouba (Syria), J. Manoharmayum (UK), R. Stong, and the proposer.

A Perpendicularity Involving the Incenter and Nagel Point

12291 [2021, 947]. Proposed by Leonard Giugiuc, Drobeta Turnu Severin, Romania, and
Petru Braica, Satu Mare, Romania. The Nagel point of a triangle is the point common to
the three segments that join a vertex of the triangle to the point at which an excircle touches
the opposite side. Let ABC be a triangle with incenter I and Nagel point J . Prove that AJ
is perpendicular to the line through the orthocenters of triangles I AB and I AC.
Solution by Li Zhou, Polk State College, Winter Haven, FL. Suppose that the incircle ω of
ABC is tangent to BC at D and AB at E. Extend DI to intersect ω again at K, and let

774 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
AK intersect ω again at L. The tangent line to ω at K is parallel to BC, and therefore there
is a homothety centered at A that sends this tangent line to BC. The image of ω under
this homothety is the excircle tangent to BC, and the image of K is the point where this
excircle is tangent to BC. Since the image of K is on the line AL, it follows that the Nagel
point J lies on AL.
A
Let H be the intersection point
of EI and LD. Since KD is a
diameter of ω, KL ⊥ LD, and E K
since AB is tangent to ω at E,
EI ⊥ AB. It follows that A, I
E, L, and H lie on the circle
with diameter AH . Therefore C
L
∠AH E = ∠ALE = ∠KDE,
D
and since I DE is isosce- B H
les, ∠KDE = ∠DEH . Thus
AH ED. Since BI is the
perpendicular bisector of DE, we have BI ⊥ AH . Combining this with I H ⊥ AB, we
conclude that H is the orthocenter of triangle I AB. Likewise, the orthocenter of I AC is
on LD as well, completing the proof.
Also solved by M. Bataille (France), C. Chiser (Romania), N. S. Dasireddy (India), I. Dimitrić, G. Fera
(Italy), O. Geupel (Germany), J.-P. Grivaux (France), N. Hodges (UK), W. Janous (Austria), O. Kouba (Syria),
J. H. Lindsey II, N. Osipov (Russia), C. G. Petalas (Greece), C. R. Pranesachar (India), V. Schindler (Germany),
A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), T. Wiandt, T. Zvonaru (Romania), Davis Problem Solv-
ing Group, UM6P Math Club (Morocco), and the proposer.

The Congruence Class of a Trigonometric Power

12292 [2021, 947]. Proposed by Nikolai Osipov, Siberian Federal University, Krasnoyarsk,
Russia. Let p be a prime number, and let r = 1/(2 cos(4π/7)). Evaluate r p+2  modulo
p.
Solution by UM6P Math Club, Mohammed VI Polytechnic University (UM6P), Ben Guerir,
Morocco. For p ∈ / {2, 7}, when p is congruent to ±1, ±2, or ±3 modulo 7, the value of
r p+2  is congruent modulo p to −12, −5, or 2, respectively. For the exceptions, r 4  ≡ 1
(mod 2) and r 9  ≡ 2 (mod 7).
Let θ = 2π/7, so r = 1/(2 cos 2θ ). Also let s = 1/(2 cos θ ) and t = 1/(2 cos 4θ ). For
x ∈ {θ, 2θ, 4θ }, we have 3x ≡ −4x mod 2π , so sin 3x + sin 4x = 0. Using the multiple-
angle formulas sin 3x = 3 sin x − 4 sin3 x and
sin 4x = 2 sin 2x cos 2x = 4 sin x cos x(cos2 x − sin2 x),
we compute
sin 3x + sin 4x
0= = 3 − 4 sin2 x + 4 cos3 x − 4 cos x sin2 x
sin x
= 8 cos3 x + 4 cos2 x − 4 cos x − 1.
Thus r, s, and t satisfy z−3 + z−2 − 2z−1 − 1 = 0 and hence z3 + 2z2 − z − 1 = 0. Let
Sn = r n + s n + t n . From the equation in z satisfied by r, s, and t, we obtain the recurrence
Sn+3 = −2Sn+2 + Sn+1 + Sn . Since (z − r)(z − s)(z − t) = z3 + 2z2 − z − 1, the initial
conditions are S0 = 3, S1 = −2, and
S2 = S12 − 2(rs + st + tr) = 4 − 2(−1) = 6.

October 2023] PROBLEMS AND SOLUTIONS 775

The recurrence then implies that Sn is an integer for n ≥ 0. Since s ≈ 0.80 and t ≈ −0.55,
we have s n + t n ∈ (0, 1) and r n  = Sn − 1 when n ≥ 1.
Let ω = e2πi/7 . Since cos x = (eix + e−ix )/2, we have
1/r = ω2 + ω−2 , 1/s = ω + ω−1 , and 1/t = ω4 + ω−4 .
In the ring Fp [ω] for prime p, we have (x + y)p = x p + y p , so
1/r p = ω2p + ω−2p , 1/s p = ωp + ω−p , and 1/t p = ω4p + ω−4p .
In order to compute Sp+2 , we need the pth powers of r, s, and t. These are obtained by
first “rationalizing the denominator” to express these quantities as polynomials in ω. For
example,
1 ω ω(ω2 − 1) ω(ω2 − 1)(ω4 + 1) ω(ω2 − 1)(ω4 + 1)
s= = 2 = = = .
ω + 1/ω ω +1 ω4 − 1 ω8 − 1 ω−1
Canceling ω − 1 and expanding yields s = ω6 + ω5 + ω2 + ω. Similarly, or by substituting
ω2 or ω4 for ω in the expression for s, we have r = ω5 + ω4 + ω3 + ω2 and t = ω6 + ω4 +
ω3 + ω.
Raising these expressions to the pth power multiplies the exponents by p, and the expo-
nents then reduce modulo 7. We need only consider three cases, since negating the expo-
nents in these polynomials does not change the values.
(i) If p ≡ ±1 (mod 7), then r p ≡ r, s p ≡ s, and t p ≡ t modulo p, reducing the computa-
tion to Sp+2 ≡ S3 = −2S2 + S1 + S0 = −12 − 2 + 3 = −11 and r p+2  ≡ −12.
(ii) If p ≡ ±2 (mod 7) with p = 2, then r p ≡ t, s p ≡ r, and t p ≡ s modulo p, reducing
the computation to Sp+2 ≡ tr 2 + rs 2 + st 2 = −4 (see comment below) and r p+2  ≡ −5.
(iii) If p ≡ ±3 (mod 7), then r p ≡ s, s p ≡ t, and t p ≡ r modulo p, reducing the compu-
tation to Sp+2 ≡ sr 2 + ts 2 + rt 2 = 3 (see comment below) and r p+2  ≡ 2.
Finally, if p = 2, then r 4  = S4 − 1 = 25 ≡ 1 mod p, and if p = 7, then r 9  =
S9 − 1 = −1461 ≡ 2 mod p.
Editorial comment. The sums of powers of roots of a polynomial f (denoted Sk above)
are called Newton sums owing to Newton’s identities,  which compute them in terms of
the coefficients of f . In particular, Sk = −(kck + k−1 i=1 ci Sk−i ), where f (x) = ck x n−k
with c0 = 1 and ck = 0 for k > n. More than half a dozen proofs are known; those in this
Monthly include 75 (1968), 396–397, 99 (1992), 749–751, and 110 (2003), 232–234. For
further discussion, see artofproblemsolving.com/wiki/index.php/Newton’s Sums.
As in the solution above, most solvers needed to calculate α and β, where α = rs 2 +
st + tr 2 and β = rt 2 + sr 2 + ts 2 . Richard Stong noted that α + β = S1 S2 − S3 = −1 and
2

αβ = (rs + st + tr)3 + rst · S13 = 12, yielding {α, β} = {−4, 3}. Allen Stenger noted that
α and β must be integers and computed them to the nearest integer. Michael Reid expanded
them in terms of ω. Roberto Tauraso noted that the sequence {Sn } (OEIS A094648) is
related to Catalan’s constant.
Also solved by G. Fera (Italy), K. T. L. Koo (China), O. P. Lossers (Netherlands), M. Omarjee (France),
M. Reid, A. Stenger, R. Stong, R. Tauraso (Italy), and the proposer.
A Real Identity
12293 [2022, 86]. Proposed by Hideyuki Ohtsuka, Saitama, Japan, and Roberto Tauraso,
University of Rome Tor Vergata, Rome, Italy. Let n be a positive integer and r be a positive
real number. Prove
⎛ ⎞⎛ ⎞
n  k    k    
n ⎠⎝ n ⎠ (r + 1)n + (r − 1)n 2
(−1)k ⎝ rj (−r)j = .
k=0 j =0
j j =0
j 2

776 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
Solution by Martin Widmer, University of London, London, UK. Let a0 = 0, let
     
n n 2 n
ak = + r + ··· + r k−1
0 2 k−1
when 0 < k ≤ n + 1 and k is odd, and let
     
n n 3 n
ak = r+ r + ··· + r k−1
1 3 k−1
when 0 < k ≤ n + 1 and k is even. We compute
⎛ ⎞⎛ ⎞
 k    k  
n n
(−1)k ⎝ rj ⎠ ⎝ (−r)j ⎠ = (ak+1 + ak )(ak+1 − ak ) = ak+1
2
− ak2 .
j =0
j j =0
j

Finally,
n  
(r + 1)n + (r − 1)n 2
2
(ak+1 − ak2 ) = an+1
2
= .
k=0
2

Also solved by T. Amdeberhan & V. H. Moll, A. Berkane (Algeria), N. Caro-Montoya (Brazil), C. Curtis,
K. Gatesman, O. Geupel (Germany), N. Hodges (UK), P. Lalonde (Canada), O. P. Lossers (Netherlands),
J. Manoharmayum (UK), J. H. Nieto (Venezuela), M. Omarjee (France), E. Schmeichel, A. Stadler (Switzer-
land), R. Stong, M. Wildon (UK), L. Zhou, Fejéntaláltuka Szeged Problem Solving Group (Hungary), UM6P
Math Club (Morocco), and the proposers.

Exploring a Planet, Revisited, Revisited

12342 [2022, 785]. Proposed by George Stoica, Saint John, NB, Canada. Let v1 , . . . , vn
be unit vectors in Rd . Prove that if u maximizes ni=1 |vi · u| over all unit vectors u ∈ Rd ,
then for all i, |vi · u| ≥ sin(π/(2n)).
Editorial comment. Several readers pointed out that this problem was presented and solved
in Y. Zhao (2022), Exploring a planet, revisited, this Monthly 129(7): 678–680. As
Zhao’s note explains, the problem is connected to a conjecture of László Fejes Tóth (1973),
Research problems: Exploring a planet, this Monthly 80(9): 494–498. A preprint of
Zhao’s note was posted to arxiv.org before we received this submitted problem, and the
problem here is taken verbatim from Zhao’s note. We regret the repetition without proper
attribution.

CLASSICS
C18. Due to Thomas Cover; suggested by Richard Stanley. Alice chooses two distinct
numbers and writes each of them on a slip of paper. Bob selects one of the two slips
at random and looks at the number on it. He must then choose to either keep that slip or
switch to the other slip. Bob wins if he ends up with the slip with the larger number. Is there
anything Bob can do to ensure that, no matter what numbers Alice chooses, his probability
of winning is greater than 1/2?

Choosing n Numbers With Sum 0 Modulo n

C17. Due to Paul Erdős, Abraham Ginzburg, and Abraham Ziv; suggested by Gabriel
Carroll and Yuri Ionin, independently. Given 2n − 1 integers, show that it is possible to
choose n of them that sum to a multiple of n.
Solution. Let the integers be a1 , . . . , a2n−1 . We first address the case of prime n. For
 n−1 
I ⊂ {1, . . . , 2n − 1} with |I | = n, let SI = i∈I ai , and let S = I SI . Thinking

October 2023] PROBLEMS AND SOLUTIONS 777

momentarily of {a1 , . . . , a2n−1 } as a set of indeterminates, the expressions SI and S are
homogenous polynomials of degree n − 1. Each monomial in S will have k of the inde-
terminates represented for some k with 1 ≤ k ≤ n − 1, and each such monomial will
arise with equal coefficient in SI for precisely 2n−1−kn−k
choices of I . Since 2n−1−k
n−k
≡0
(mod n), the coefficient of each monomial in S when like terms are gathered is a multiple
of n. Hence S ≡ 0 (mod n).
On the other hand, if the result of the problem is false, then for every I , we have SI ≡ 0
(mod n) and so SI ≡ 1 (mod n) by Fermat’s little theorem. Since there are 2n−1 n
such
sets I , and since
 
2n − 1 (2n − 1)(2n − 2) · · · (n + 1)
≡ ≡ 1 (mod n),
n (n − 1)(n − 2) · · · 1
we conclude that S ≡ 1 (mod n), a contradiction. This proves the result when n is prime.
Finally, we extend the result to the case where n is any positive integer. We argue
by induction on n. Suppose that n = pm where p is prime, and suppose we are given
a multiset {a1 , . . . , a2n−1 }. Repeatedly applying the prime case of the result to p, we
j −1
extract sets I1 , . . . , I2m−1 with Ij ⊂ {1, . . . , 2n − 1} \ s=1 Is such that |Ij | = p and

i∈Ij ai ≡ 0 (mod p) for all j ∈ {1, . . . , 2m − 1}. The reason this is possible is that, as
long as j ≤ 2m − 1,
j −1 
  
 
{1, . . . 2n − 1} \ Is  = 2n − 1 − (j − 1)p ≥ 2n − 1 − (2m − 2)p = 2p − 1,
s=1

and so the  result for p ensures a choice for Ij . Now, for j ∈ {1, . . . , 2m − 1}, let
bj = (1/p) i∈Ij ai . Applying the induction hypothesis to m and {b1 , . . . , b2m−1 }, we

obtain a set J ⊂ {1, . . . , 2m − 1} with |J | = m and j ∈J bj ≡ 0 (mod m). The set
{ai : i ∈ Ij , j ∈ J } provides the subset of size n whose elements sum to a multiple of n.
Editorial comment. The result is from P. Erdős, A. Ginzburg, and A. Ziv (1961), Theorem
in the additive number theory, Bull. Res. Council Israel 10F, 41–43. The number 2n − 1
in the problem statement is optimal, as the multiset with n − 1 zeroes and n − 1 ones has
size 2n − 2 but no subset of size n summing to a multiple of n. Five separate proofs of the
result for prime n appear in N. Alon and M. Dubiner (1993), Zero-sum sets of prescribed
size, in Combinatorics, Paul Erdős is Eighty (Vol. 1), eds. D. Miklós, V. T. Sós, and T.
Szőnyi, Keszthely, 33–50, where the argument given here is attributed to N. Zimmerman.

778 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
SOLUTIONS
Four Concyclic Points
12280 [2021, 856]. Proposed by Nguyen Duc Toan, Da Nang, Vietnam. Let ABC be an
acute scalene triangle with circumcenter O and orthocenter H . Let M and R be the mid-
points of segments BC and OH , respectively, let S be the reflection across BC of the
circumcenter of triangle BOC, and let T be the second point of intersection of the circum-
circle of triangle BH C and line OH . Prove that M, R, S, and T are concyclic.
Solution by O. P. Lossers, Eindhoven University of Technology, Eindhoven, Netherlands.
The line OM passes through S, and the line OH passes through both R and T , so by the
power law, to prove that the points M, S, R, and T are concyclic it suffices to show that
OR · OT = OM · OS and that O lies between R and T if and only if O lies between M
and S.
Let α, β, and γ be the angles
at vertices A, B, and C, respec-
tively, of ABC, and let a = BC.
By the inscribed angle theorem,
∠BOC = 2α, and therefore
∠BOM = α. Since ∠BMO is a
right angle, we have
a
OM = BM cot(∠BOM) = cot α.
2
Since CH is perpendicular to AB,
we have ∠BCH = π/2 − β. Like-
wise, ∠CBH = π/2 − γ , so

∠BH C = π − (π/2 − β)
− (π/2 − γ ) = β + γ = π − α.

680 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

Let P and Q be the intersection points of the circumcircle C of BH C and the line
OM, with P on the same side of BC as A and Q on the opposite side. Since ∠BP C =
∠BH C = π − α, we have ∠BP M = (π − α)/2 and
  α
a π −α a
P M = cot = tan .
2 2 2 2
Thus
a α
OP = |OM − P M| = cot α − tan .
2 2
Similarly, ∠BQC = α, ∠BQM = α/2, QM = (a/2) cot(α/2), and
a α
OQ = OM + QM = cot α + cot .
2 2
Thus by the power law for the circle C,
1 1 a2 α  α
OR · OT = OH · OT = OP · OQ = cot α − tan cot α + cot . (1)
2 2 8 2 2
Next, we claim that
a
OS = | cot α + cot(2α)|, (2)
2
and therefore
a2
OM · OS = cot α | cot α + cot(2α)|. (3)
4
Let S be the circumcenter of BOC, so that S is the reflection of S across BC. The
calculation of OS depends on how α compares to π/4. Suppose first that α < π/4. In that
case, S lies on the same side of BC as A, so S lies on the opposite side. By the inscribed
angle theorem, ∠BSC = ∠BS C = 2∠BOC = 4α and ∠BSM = 2α. Therefore SM =
BM cot(∠BSM) = (a/2) cot(2α), so
a 
OS = OM + SM = cot α + cot(2α) .
2
Since cot α and cot(2α) are positive in this case, this agrees with (2).
If α = π/4, then S = S = M, so OS = OM = (a/2) cot α, which again agrees with
(2) because cot(2α) = 0. Finally, suppose α > π/4. In that case, S lies on the opposite
side of BC from A and S lies on the same side. Therefore ∠BSC = ∠BS C = 2π − 4α,
∠BSM = π − 2α, SM = (a/2) cot(π − 2α) = −(a/2) cot(2α), and
a
OS = |OM − SM| = | cot α + cot(2α)|,
2
again confirming (2).
Combining (1) and (3), we see that establishing OR · OT = OM · OS reduces to show-
ing
α  α
cot α − tan cot α + cot = 2 cot α | cot α + cot(2α)|.
2 2
After some rewriting (using 2 cot(2α) = cot α − tan α), we see that both sides are equal to
|3 cot2 α − 1|.
Finally, we must confirm the betweenness condition. For O to lie between R and T ,
it must be inside C, which happens if and only if ∠BOC > ∠BH C. Using the equations
∠BOC = 2α and ∠BH C = π − α, we see that this is equivalent to α > π/3. For O to

August–September 2023] PROBLEMS AND SOLUTIONS 681

lie between M and S, we must have S on the same side of BC as O and ∠BOC greater
than ∠BSC. By previous reasoning, this holds if and only if α > π/4 and 2α > 2π − 4α,
which again reduces to α > π/3, thus completing the proof.
Editorial comment. The requirements that ABC be acute and scalene are unnecessary,
as long as some exceptional situations are accounted for. For example, if AB = AC, then
the points M, R, S, and T are collinear, since all lie on the perpendicular bisector of BC.
If ∠BAC is a right angle, then O = M, H = A, the points M, R, and T are collinear, and
S is undefined.
Also solved by M. Bataille (France), H. Chen (China), K. Gatesman, O. Geupel (Germany), J.-P. Grivaux
(France), N. Hodges (UK), W. Janous (Austria), O. Kouba (Syria), K.-W. Lau (China), C. R. Pranesachar
(India), V. Schindler (Germany), A. Stadler (Switzerland), R. Stong, T. Wiandt, Davis Problem Solving Group,
and the proposer.

Generating the Positive Rational Numbers

12282 [2021, 856]. Proposed by George
√ Stoica, Saint John, NB, Canada. Prove that the
multiplicative group generated by { 2n/n : n ∈ Z+ } is the group of positive rational
numbers.
Solution by Stephen M. Gagola Jr., Kent State University, Kent, OH. Let M be the multi-
plicative group generated by the set of rationals in the problem statement. We first√ show
that 3, 5, 7, 11, and 2 lie in M. To do this, we calculate here some small values of 2n/n.
n 3 4 5 7 8 11
√
2 n/n 4/3 5/4 7/5 9/7 11/8 15/11
The product of the first four entries in the second row is 3, so 3 ∈ M. The equa-
tions 5 = 3(4/3)(5/4), 7 = 3(4/3)(5/4)(7/5), 11 = 3(3)(4/3)(5/4)(11/15), and 2 =
3(3)(4/3)(4/3)(11/8)(1/11) then show that 2, 5, 7, and 11 are also in M.
Let√p be an odd prime number,
√ and suppose that all primes less than p lie in M. Set
q= 2 p. Since p < q < 2 p, if q is composite then √ all prime factors of q are less
than p and belong to M. Therefore q ∈ M and p = q/( 2 p/p) ∈ M. √
Hence we may assume √ prime. The definition of q yields q < 2 p <
√ that q is an odd
q + 1, which implies√ 2 q/2 √ < p < 2 (q + 1)/2. Note that (q + 1)/2 is an integer.
Since the interval√[ 2 q/2, 2 (q + 1)/2] has length less than 1 and contains the integer
p, we conclude 2 (q + 1)/2 = p, and thus
√
p 2 (q + 1)/2
= ∈ M.
(q + 1)/2 (q + 1)/2
Since (q + 1)/2 < p, all prime factors of (q + 1)/2 belong to M, and it follows that so
does (q + 1)/2 itself. Therefore p ∈ M. √
Since M contains all prime numbers, and 1 = 2, it follows that M is the group of
all positive rationals.
Editorial
√ comment. Gagola commented further that the set of “numerators,” namely
{ 2 n : n ∈ Z+ }, also √generates the √
group of positive rationals. First note inductively
that if n ∈ Z+ and (1 + 2)n = a + b 2, then a 2 − 2b2 = (−1) n
√ . Let p be prime,
√ and
choose n odd√and large enough so that√b > p. We have a − 2 b = −1/(a + 2 b),
so a + ε =√ 2 b, where ε = 1/(a + 2 √ b) < 1/b. Therefore when k < b we have
ka + kε = 2 kb, where kε < 1. Thus 2 kb = ka, and all these integers belong
to the group generated by the specified set. In particular, both a and pa belong to this
group, and hence so does p.

682 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

 Celia Schacht observed that the complete solution of this problem
√ is the topic of I. Kátai
and B. M. Phong, On the multiplicative group generated by {[ 2n]/n : n ∈ N}, Acta Math-
ematica Hungarica 145 (2015), no. 1, 80–87. She also cited two subsequent papers by the
same authors with the same title (II and III) dealing with deeper questions: (2015), Acta
Scientiarum Mathematicarum (Szeged) 81 no. 3–4, 431–436, and (2015), Acta Mathemat-
ica Hungarica 147 no. 1, 247–254.
Also solved by J. Boswell & C. Curtis, A. Dixit & S. Pathal (India), K. Gatesman, N. Hodges (UK),
O. P. Lossers (Netherlands), M. Reid, C. Schacht, A. Stadler (Switzerland), D. Terr, and the proposer.

A Triangle with Perimeter 2021

12284 [2021, 857]. Proposed by Zachary Franco, Houston, TX. Let ABC be a triangle
with circumcenter O, incenter I , orthocenter H , sides of integer length, and perimeter
2021. Suppose that the perpendicular bisector of OH contains A and I . Find the length of
BC.
Solution by José Heber Nieto, Universidad del Zulia, Maracaibo, Venezuela. We show that
BC = 679.
Let a, b, and c denote the lengths of sides BC, CA, and AB, respectively. We first show
that a 2 = b2 + c2 ± bc.
Since either ∠ABC or ∠ACB is acute, by symmetry we may assume that ∠ABC is
acute. Let P be the foot of the altitude from A, and let M be the midpoint of AC. Let R
be the circumradius of ABC, and let h = AP and u = BP . Note that since A lies on the
perpendicular bisector of OH , AH = AO = R.
We first consider the case in which ABC is acute. Triangles ABP , CH P , and AOM
are similar, since they all have a right angle and an angle equal to ∠ABC. Therefore
AP CP AP AM
= and = . (1)
BP HP AB AO
Since ABC is acute, H is between A and P and P is between B and C, so CP = a − u
and H P = h − R. Hence the equations in (1) become
h a−u h b/2
= and = . (2)
u h−R c R
From the first of these equations we get hR + au = h2 + u2 = c2 , and the second yields
hR = bc/2; combining these, we have au = c2 − bc/2.
Applying the Pythagorean theorem to ACP yields h2 = b2 − (a − u)2 , which implies
c2 − u2 = b2 − (a − u)2 and therefore a 2 = b2 − c2 + 2au. Substituting au = c2 − bc/2,
we conclude a 2 = b2 + c2 − bc.
If ∠ACB is obtuse, then similar reasoning can be used, except that now P is between
A and H and C is between B and P . Therefore CP = u − a and H P = R − h. The
equations in (2) still hold, so we again obtain a 2 = b2 + c2 − bc.
If ∠ACB is a right angle, then H = C and O is the midpoint of AB, so c = 2R = 2b.
Thus by the Pythagorean theorem, a 2 = c2 − b2 = b2 + c2 − 2b2 = b2 + c2 − bc.
It is not possible for ∠BAC to be a right angle, because that would imply H = A,
contradicting AH = R. Finally, we consider the case in which ∠BAC is obtuse. In this
case, A is between P and H and P is between B and C, so CP = a − u, H P = h + R, and
the first equation in (1) becomes h/u = (a − u)/(h + R). Imitating the earlier algebraic
reasoning then leads to the equation a 2 = b2 + c2 + bc. This completes the proof that
a 2 = b2 + c2 ± bc.
Since a + b + c = 2021, we have (2021 − b − c)2 = b2 + c2 ± bc, which simplifies to
tbc − 2 · 2021b − 2 · 2021c + 20212 = 0, (3)

August–September 2023] PROBLEMS AND SOLUTIONS 683

where t is either 1 or 3. We see that 2021 | tbc. Since 2021 = 43 · 47, either 47 | b and
43 | c or 43 | b and 47 | c (neither b nor c can be divisible by both 43 and 47, because b
and c are both less than 2021). By symmetry, we may assume b = 47x and c = 43y, where
0 < x < 43 and 0 < y < 47. Substituting into (3) gives
txy − 94x − 86y + 2021 = 0. (4)
If t = 1, then equation (4) can be rearranged to read
(x − 86)(y − 94) = 3 · 43 · 47.
Since 43  x and 47  y, we must have either x − 86 = ±3 · 47 or y − 94 = ±3 · 43. But
these are all inconsistent with 0 < x < 43 and 0 < y < 47, so this case is impossible.
Thus t = 3, and equation (4) is equivalent to
(3x − 86)(3y − 94) = 43 · 47.
In this case the only possibilities are 3x − 86 = ±47 and 3y − 94 = ±43. Since 3x − 86 =
47 does not give an integer value for x, we must have 3x − 86 = −47 and 3y − 94 = −43,
so x = 13, y = 17, b = 611, c = 731, and finally a = 2021 − b − c = 679.
Editorial comment. An alternative proof of the equation a 2 = b2 + c2 ± bc uses the well-
known equation AH = 2R| cos α|, where α denotes the measure of ∠BAC. This implies
that the perpendicular bisector of OH passes through A if and only if cos α = ±1/2, which
means α is either π/3 or 2π/3. The law of cosines then gives a 2 = b2 + c2 ± bc.
Note that the condition that the perpendicular bisector of OH contains I was not used
in the solution above. In fact, this condition is implied by α = π/3 but contradicted by
α = 2π/3. To see this, let V be the midpoint of the arc BC of the circumcircle of ABC
−− −−
that does not contain A. If α = π/3, then OV = AH , and therefore AH V O is a rhombus.
It follows that the perpendicular bisector of OH is AV , which bisects ∠BAC and therefore
−− −−
passes through I . On the other hand, if α = 2π/3, then OV = −AH . It follows that AV
is parallel to OH , and therefore the perpendicular bisector of OH , which passes through
A, does not pass through I .
Also solved by F. R. Ataev (Uzbekistan), M. Bataille (France), H. Chen (China), C. Chiser (Romania) &
N. Ivaschescu (Canada), G. Fera (Italy), K. Gatesman, O. Geupel (Germany), N. Hodges (UK), W. Janous
(Austria), K.-W. Lau (China), O. P. Lossers (Netherlands), C. R. Pranesachar (India), A. Stadler (Switzerland),
R. Stong, R. S. Tiberio, M. Vowe (Switzerland), T. Wiandt, H. Widmer (Switzerland), Davis Problem Solv-
ing Group, Eagle Problem Solvers, Fejéntaláltuka Szeged Problem Solving Group (Hungary), and the proposer.

A Sum and Integral That Cannot Be Interchanged

12285 [2021, 857]. Proposed by Atul Dixit, Indian Institute of Technology, Gandhinagar,
India. Prove
∞  ∞  ∞ ∞
t cos t π t cos t
2 + m2 u2
dt = − cos t + 2 + m2 u2
dt
m=1 0
t 0 2u m=1
t

for u > 0.
Solution by Albert Stadler, Herrliberg, Switzerland. Integrating by parts twice, always tak-
ing the antiderivative that vanishes at 0, we get
 ∞  ∞ 2  ∞
t cos t (t − m2 u2 ) sin t 2t (t 2 − 3m2 u2 )
dt = dt = (1 − cos t)dt.
0 t 2 + m2 u2 0 (t 2 + m2 u2 )2 0 (t 2 + m2 u2 )3
Since this last integrand satisfies the bound
2t (t 2 − 3m2 u2 ) 12t
(1 − cos t) ≤ 2
(t + m u )
2 2 2 3 (t + m2 u2 )2

684 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

and
∞  ∞ ∞
t 1
dt = < ∞,
m=1 0
(t + m u )
2 2 2 2
m=1
2m2 u2

the dominated convergence theorem applies. Hence we can interchange summation and
integration to get
∞  ∞  ∞ ∞
t cos t 2t (t 2 − 3m2 u2 )
dt = (1 − cos t) dt.
m=1 0
t 2 + m2 u2 0 m=1 (t + m u )
2 2 2 3

Next we integrate by parts twice “in the other direction,” this time choosing in each case
the antiderivative that vanishes as t → ∞. To choose the right antiderivative in the second
integration by parts, we use the partial fraction decomposition of coth (see I. S. Gradshteyn
and I. M. Ryzhik (2007), Table of Integrals, Series, and Products, 7th ed., Burlington, MA:
Academic Press, p. 44, equation 1.421.4), to compute
∞    
t π πt 1 π
lim = lim coth − = .
t→∞
m=1
t +m u
2 2 2 t→∞ 2u u 2t 2u

This leads to the calculation

∞  ∞  ∞ ∞
t cos t t 2 − m2 u2
dt = (1 − cos t) d −
m=1 0
t +m u
2 2 2
0 m=1
(t 2 + m2 u2 )2
 ∞ ∞
t 2 − m2 u2
= sin t dt
0 m=1
(t 2 + m2 u2 )2
 ∞ ∞
π t
= sin t d −
0 2u m=1 t + m2 u2
2

 ∞ ∞
π t cos t
= − cos t + dt.
0 2u m=1
t + m2 u2
2

Editorial comment. As the solution above shows, one integration by parts is sufficient to get
a situation where it is valid to pull the sum inside the integral, but to justify the interchange
one must give more careful bounds (as was done by O. P. Lossers). The proposer and
N. Hodges showed that the given integral I can be evaluated explicitly in terms of the
digamma function ψ as
 u     
1 1 iu iu
I = log − ψ +ψ − .
2 2π 4 2π 2π
This follows from identifying the final sum as in the solution above and using Gradshteyn
& Ryzhik (3.951.6).
Also solved by N. Hodges (UK), O. P. Lossers (Netherlands), J. Van Casteren & L. Kempeneers (Belgium),
and the proposer.

Another Consequence of Euler’s Identity

12287 [2021, 946]. Proposed by Ovidiu Furdui and Alina Sı̂ntămărian, Technical Univer-
sity of Cluj-Napoca, Cluj-Napoca, Romania. Prove
⎛ ⎞
∞ ∞ 2
⎝n 1 1 3 1 3
2
− ⎠ = − ζ (2) + ζ (3),
n=1 k=n
k n 2 2 2

August–September 2023] PROBLEMS AND SOLUTIONS 685

 ∞
where ζ is the Riemann zeta function, defined by ζ (s) = k=1 1/k s .
Solution by Omran Kouba,
 Higher Institute for Applied Sciences and Technology, Damas-
cus, Syria. Let Sn = ∞ 2
k=n 1/k . We rewrite the summand of the series in the form

1
nSn2 − = Tn − Tn−1 + Un ,
n

where T0 = 0 and we can evaluate both Un and limn→∞ Tn . By the telescoping of the
partial sums, the desired series converges to lim Tn + Un .
In order to obtain a suitable Tn , we define Tn in terms of the sequence S. Let Tn =
2
an Sn+1 + bn Sn+1 , where an and bn will be chosen later. Since Sn+1 = Sn − 1/n2 ,
   
1 2 1
Tn − Tn−1 = an Sn − 2 + bn Sn − 2 − an−1 Sn2 − bn−1 Sn
n n
−2an an bn
= (an − an−1 )Sn2 + Sn + 4 + (bn − bn−1 )Sn − 2 .
n2 n n
To make the coefficient on Sn2 be n, set an = n(n + 1)/2. Now
−(n + 1) n+1 bn
Tn − Tn−1 = nSn2 + Sn + 3
+ (bn − bn−1 )Sn − 2
n 2n n
1
= nSn2 − + (bn − bn−1 − (1 + 1/n))Sn + En ,
n
where En = (n + 1)/(2n3 ) − bn /n2 + 1/n. To eliminate the coefficient on Sn , set b0 = 0
and bn =
bn−1 + 1 + 1/n for n ≥ 1. Thus bn = n + Hn , where Hn is the nth harmonic
number ni=1 1/ i. Now
1 1 1 Hn
Tn − Tn−1 = nSn2 − + + 3− 2.
n 2n2 2n n
Since T0 = 0, summing this identity yields
m   m m m
1 Hn 1 1
nSn2 − = Tm + 2
− 2
− 3
.
n=1
n n=1
n n=1
2n n=1
2n

Letting m → ∞ and using the Euler identity ∞
n=1 Hn /n = 2ζ (3), we obtain
2

∞  
1 1 1
nSn2 − = lim Tm + 2ζ (3) − ζ (2) − ζ (3).
n=1
n m→∞ 2 2

Returning to the definition of Tm , we now have

m(m + 1) 2
Tm = Sm+1 + (m + Hm )Sm+1 . (∗)
2
To compute the limit, we use 1/k − 1/(k + 1) < 1/k 2 < 1/(k − 1) − 1/k to obtain
1/m < Sm < 1/(m − 1), and thus limm→∞ mSm = 1. Hence the first term in (∗) tends to
12 /2 and the second term in (∗) tends to 1 (since Hm /m → 0). This gives limm→∞ Tm =
1/2 + 1 = 3/2, and the desired result follows.
Editorial comment. A simple proof of Euler’s formula for ζ (3) appears in an editorial
comment following problem 12091 [2019, 180; 2020, 853] in this Monthly. The solution
to problem 2136 from Mathematics Magazine by Kelly D. McLenithan (2023) Math. Mag.

686 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

96, 90–91 uses similar methods, in particular the same identity of Euler, which it observes
is the q = 2 case of Euler’s 1775 result
∞ q−2
Hk
2 = (q + 2)ζ (q + 1) − ζ (m + 1)ζ (q − m).
k=1
kq m=1

Also solved by T. Amdeberhan & V. H. Moll, K. F. Andersen (Canada), M. Bataille (France), A. Berkane
(Algeria), O. Bordellés (France), P. Bracken, B. Bradie, B. S. Burdick, H. Chen, A. De la Fuente, G. Fera
(Italy), O. Geupel (Germany), E. A. Herman, N. Hodges (UK), M. Hoffman, K.-W. Lau (China), O. P. Lossers
(Netherlands), J. Manoharmayum (UK), C. Morin (France), M. Omarjee (France), C. Sanford, A. Stadler
(Switzerland), A. Stenger, S. M. Stewart (Saudi Arabia), R. Tauraso (Italy), T. Wiandt, UM6P Math Club
(Morocco), and the proposer.

The Dirichlet Integral in Disguise

12288 [2021, 946]. Proposed by Seán Stewart, Bomaderry, Australia. Prove
 ∞  2
1 π
1 − x 2 sin2 dx = .
0 x 5

Solution by Ming-Can Fan, Huizhou University, Guangdong, China. The substitution x =

1/t yields
 ∞  2  ∞ 2 
2 1 t − sin2 t 2
1 − x sin
2
dx = dt
0 x 0 t3
 ∞ 2  ∞
2t − 2 sin2 t sin4 t − t 4
= dt + dt.
0 t4 0 t6
Using 2 sin2 t = 1 − cos(2t) and integration by parts three times yields
 ∞ 2  ∞ 2
2t − 2 sin2 t 2t − 1 + cos(2t)
4
dt = dt
0 t 0 t4
∞  ∞
2t 2 − 1 + cos(2t) 4t − 2 sin(2t)
=− + dt
3t 3 0 0 3t 3
 ∞
2t − sin(2t) ∞ 2 − 2 cos(2t)
=− + dt
3t 2 0 0 3t 2
 ∞
2 − 2 cos(2t) ∞ 4 sin(2t) 4 π 2π
=− + dt = · = ,
3t 0 0 3t 3 2 3
∞
where in the last step we have used the well-known fact that 0 sin(at)/t dt = π/2 for
all a > 0. (The case a = 1 is known as the Dirichlet integral, and the general formula
follows via the substitution u = at.) Similarly, using sin4 t = (3 + cos(4t) − 4 cos(2t))/8
and integration by parts five times, we get
 ∞  ∞
sin4 t − t 4 (3 + cos(4t) − 4 cos(2t))/8 − t 4
dt = dt
0 t6 0 t6
 ∞
2 sin(2t) − 16 sin(4t) 2 π 16 π 7π
= dt = · − · =− .
0 15t 15 2 15 2 15
Hence
 ∞  2
1 2π 7π π
1 − x sin2 2
dx = − = .
0 x 3 15 5

August–September 2023] PROBLEMS AND SOLUTIONS 687

Also solved by U. Abel & V. Kushnirevych (Germany), T. Amdeberhan & V. H. Moll, K. F. Andersen (Canada),
M. Bataille (France), A. Berkane (Algeria), P. Bracken, H. Chen, Ó. Ciaurri (Spain), G. A. Edgar, G. Fera
(Italy), M. L. Glasser, G. C. Greubel, J.-P. Grivaux (France), N. Grivaux (France), J. A. Grzesik, E. A. Herman,
N. Hodges (UK), F. Holland (Ireland), W. Janous (Austria), O. Kouba (Syria), K.-W. Lau (China), J. Londoño
& J. Quintero (Colombia), O. P. Lossers (Netherlands), J. Manoharmayum (UK), K. D. McLenithan, R. Mortini
(Luxembourg) & R. Rupp (Germany), A. Natian, M. Omarjee (France), A. Stadler (Switzerland), R. Stong,
R. Tauraso (Italy), J. Van Casteren & L. Kempeneers (Belgium), E. I. Verriest, M. Wildon (UK), UM6P Math
Club (Morocco), and the proposer.

CLASSICS
C17. Due to Paul Erdős, Abraham Ginzburg, and Abraham Ziv; suggested by Gabriel
Carroll and Yuri Ionin, independently. Given 2n − 1 integers, show that it is possible to
choose n of them that sum to a multiple of n.

Parallel Mountain Climbers

C16. Suggested by the editors. Two hikers start together at the bottom of a mountain and
climb to the summit but along different trails, which may go up and down along the way.
Show that it is possible for them to complete their respective hikes in such a way that they
are at the same elevation at every moment.
Solution. Suppose that the length of the first trail is a and the length of the second trail
is b. For s ∈ [0, a], let f (s) represent the elevation of the first trail at distance s from the
start, and for t ∈ [0, b], let g(t) represent the elevation of the second trail at distance t from
the start. (We take “bottom” and “summit” in the problem statement to mean that neither
trail dips below the initial elevation or rises above the final elevation.) The ordered pair
(s, t) describes simultaneous positions of the two hikers. We require a plan that maintains
f (s) = g(t). The initial position is (0, 0), and we want to show that (a, b) can be reached.
When f (s) = g(t) but s is not a local extremum of f and t is not a local extremum of
g, the two hikers can both move higher or both move lower. Hence the crucial points are
those pairs (s, t) such that at least one coordinate is a local extremum of the corresponding
function. We call the local extrema other than (0, 0) or (a, b) internal.
We define a graph with these crucial points as its vertices. Two crucial points are adja-
cent if the hikers can move from one to the other without encountering any other crucial
point. In particular, if one hiker is at an internal local extremum, then that hiker can move
forward or backward while the other matches the change in elevation, until one of them
reaches another local extremum. Thus such vertices have degree 2 in the graph. If both
hikers are at an internal local extremum, then the degree is 0 or 4.
The vertices (0, 0) and (a, b) have degree 1 in the graph; they are the only vertices with
odd degree. Since every graph has an even number of vertices of odd degree, (0, 0) and
(a, b) must lie in the same component of the graph, and hence there is a path for the two
hikers to reach the summit while maintaining the same elevation.
Editorial Comment. It is necessary that the functions f and g achieve their common global
minimum at 0 and their common global maximum at a and b, respectively. If the trail for
one hiker but not the other dips below the starting elevation (or passes above the summit),
the hikers cannot achieve the required goal. In addition, we assume that f and g have only
finitely many critical points.
The problem has appeared in many places and in various forms. The earliest formulation
in terms of mountain climbers may be in J. V. Whittaker (1966), A mountain-climbing
problem, Canadian J. Math., 18: 873–882, but the essential idea appears in T. Homma
(1952), A theorem on continuous functions, Kodai Math. Sem. Reports, 4(1): 13–16.

688 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

SOLUTIONS

Four Inequalities, One Proof

12275 [2021, 755]. Proposed by Yun Zhang, Xi’an, China. Let x, y, and z be positive real
numbers with x + y + z = 3. Prove each of the following inequalities.
(a) x 5 y 5 z5 (x 4 + y 4 + z4 ) ≤ 3. (c) x 11 y 11 z11 (x 6 + y 6 + z6 ) ≤ 3.
(b) x 8 y 8 z8 (x 5 + y 5 + z5 ) ≤ 3. (d) x 16 y 16 z16 (x 7 + y 7 + z7 ) ≤ 3.
Solution by Kyle Gatesman, Johns Hopkins University, Baltimore, MD. More generally,
we are interested in finding the maximum value of the function f defined by f (x, y, z) =
x p y p zp (x q + y q + zq ) subject to the constraints x, y, z > 0 and x + y + z = 3, where
we allow p and q to be arbitrary real numbers with p > 0 and q > 2. We prove that the
condition
(p + q)q (q − 1)q−1 ≤ (2p + q)pq−1 (q + 1)q−1 (1)
is sufficient to guarantee that the unique optimal solution is (x, y, z) = (1, 1, 1), which
implies that the maximum value of f (x, y, z) is 3. The inequalities (a), (b), (c), and (d) in
the problem statement follow from this result.
Let S denote the closed simplex {(x, y, z) ∈ R3 : x, y, z ≥ 0 and x + y + z = 3}, so
that our optimization domain is the relative interior of S. By the extreme value theorem,
the continuous function f attains its supremum on S at one or more points in S. Since the
value of f is positive in the interior of S and zero on the boundary, the supremum of f over
S is attained in the interior. Therefore, any global maximizer (x, y, z) of f over S satisfies
∇f (x, y, z) = λ∇(x + y + z − 3) = (λ, λ, λ) for some λ. Letting α = x q + y q + zq , we
have
 q 
qx + pα qy q + pα qzq + pα
∇f (x, y, z) = x p y p zp , , ,
x y z

588 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

so a necessary condition for optimality is
qx q + pα qy q + pα qzq + pα
= = . (2)
x y z
Temporarily fix α > 0, and let g(t) = (qt q + pα)/t. Since
g (t) = q(q − 1)(q − 2)t q−3 + 2pα/t 3
and since p > 0 and q > 2, we have g (t) > 0 for all t > 0, so g is strictly convex over
(0, ∞). Thus, for any constant c, the equation g(t) = c admits at most two distinct solu-
tions in t. The numbers x, y, and z must be solutions to such an equation, so x, y, and z
cannot all be distinct. By symmetry, we may assume that z = x.
Let u = y/x, so that α = x q (uq + 2). Condition (2) is equivalent to
qx q + px q (uq + 2) quq x q + px q (uq + 2)
= ,
x ux
which simplifies to
puq+1 − (p + q)uq + (2p + q)u − 2p = 0.
This condition is satisfied when u = 1, which corresponds to y = x. To show that u = 1
is the only solution when (1) holds, it suffices to show that the function h defined by
h(u) = puq+1 − (p + q)uq + (2p + q)u − 2p is strictly increasing (and therefore injec-
tive) over (0, ∞).
Observe that
h (x) = p(q + 1)uq − (p + q)quq−1 + 2p + q and
 
(p + q)(q − 1)
h (x) = pq(q +1)uq−1 −(p+q)q(q −1)uq−2 = pq(q +1)uq−2 u − .
p(q + 1)
Let u0 = (p + q)(q − 1)/(p(q + 1)). Clearly h (u) is negative for u ∈ (0, u0 ) and positive
for u ∈ (u0 , ∞), so h (u) attains its minimum value at u = u0 . Therefore, h is strictly
increasing if and only if h (u0 ) ≥ 0. This is equivalent to
 
(p + q)(q − 1) q−1  
(p + q)(q − 1) − (p + q)q + 2p + q ≥ 0,
p(q + 1)
which is equivalent to (1). Hence, when (1) holds, u = 1 is the only value of u for which
(x, ux, x) can be a maximizer of f over S. It follows that the only possible maximizer of
f over all of S is (1, 1, 1).
Editorial comment. It is not hard to show that, for fixed q > 2, inequality (1) holds for all
sufficiently large p. In fact, in each of (a)–(d), the value of p is the smallest positive integer
for which (1) holds.
There are several other ways to prove these inequalities. As indicated by multiple
solvers, one could use the pqr-method, which involves rewriting the inequalities in
terms of x + y + z, xy + yz + zx, and xyz (often denoted p, q, and r; see Chapter
14 in Z. Cvetkovski, (2012), Inequalities: Theorems, Techniques, and Selected Problems,
Berlin: Springer). Alternatively, one can rewrite all four inequalities in the form f ≥ 0,
where f (x, y, z) = (x + y + z)k+1 − 3k (xyz)p (x q + y q + zq ). Assuming without loss
of generality that x ≥ y ≥ z, we can write x = u + v + w, y = u + v, and z = u for
u, v, w ≥ 0. For inequalities (a)–(d), taking k = 18, 28, 38, and 54, respectively, Albert
Stadler used Mathematica to verify that f (u + v + w, u + v, u) ≥ 0. These are the small-
est values of k for which Stadler’s method works.

June–July 2023] PROBLEMS AND SOLUTIONS 589

Also solved by P. Bracken, D. Henderson, N. Hodges (UK), W. Janous (Austria), K.-W. Lau (China),
P. W. Lindstrom, A. Stadler (Switzerland), R. Stong, J. Vukmirović (Serbia), J. Yan (China), L. Zhou,
and Fejéntaláltuka Szeged Problem Solving Group (Hungary).

A Complicated Way to Write 1

12276 [2021, 755]. Proposed by Joe Santmyer, Las Cruces, NM. Prove
∞
 1 
n/2
1
= 1.
n=2
n + 1 i=1 2i−1 (i − 1)!(n − 2i)!

Solution by Allen Stenger, Boulder, CO. Letting an denote the inner sum, we see that an is
the coefficient of x n−2 in the product
∞  ∞ 
 (x 2 /2)k  xm
.
k=0
k! m=0
m!
2 /2 x
Since the product equals ex e , we have
∞
 2 /2 + x
an x n = x 2 ex .
n=2

Integrating both sides from 0 to 1 yields

∞ 1
an 2 /2 + x 2 /2 + x 1
= x 2 ex dx = (x − 1)ex = 1,
n=2
n +1 0 0

2 /2 + x 2 /2 + x
justified by computing f (x) = x 2 ex when f (x) = (x − 1)ex .
Also solved by T. Amdeberhan & V. H. Moll, M. Bataille (France), A. Berkane (Algeria), C. Burnette, Ó. Ciau-
rri (Spain), A. De la Fuente, G. Fera (Italy), K. Gatesman, M. L. Glasser, J. W. Hagood, E. A. Herman,
N. Hodges (UK), W. Janous (Austria), O. Kouba (Syria), O. P. Lossers (Netherlands), D. Pinchon (France),
E. Schmeichel, A. Stadler (Switzerland), S. M. Stewart (Saudi Arabia), R. Stong, R. Tauraso (Italy), M. Vowe
(Switzerland), L. Zhou, and the proposer.

A Matrix Rank Restriction

12277 [2021, 756]. Proposed by Cristian Chiser, Elena Cuza College, Craiova, Romania.
Let A, B, and C be three pairwise commuting 2-by-2 real matrices. Show that if at least
one of the matrices A − B, B − C, and C − A is invertible, then the matrix
A2 + B 2 + C 2 − AB − AC − BC
cannot have rank 1.
Solution by Jacob Boswell & Chip Curtis, Missouri Southern State University, Joplin, MO.
Set M = A2 + B 2 + C 2 − AB − AC − BC, D = A − B, and E = A − C. By symmetry,
we may assume that D is invertible. All of the named matrices pairwise commute. Thus
M = D 2 − DE + E 2 .
Multiplying on the left by (D −1 )2 yields
N = I − X + X2 , (∗)
where N = (D −1 )2 M and X = D −1 E. Since D is invertible, M has rank 1 if and only if
N has rank 1.

590 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

 We conclude by showing that N cannot have rank 1. To the contrary, assume that N
has 0 as an eigenvalue with multiplicity 1. Let v be an eigenvector of N with eigenvalue
0. Since N and X commute, N Xv = XNv = 0, so Xv must be a multiple of v, since
the eigenspace of 0 for N is one-dimensional. Thus, v is an eigenvector of X for some 
eigenvalue λ. Multiplying both sides of (∗) on the right by v gives 0 = 1 − λ + λ2 v.
It follows that λ2 − λ + 1 = 0. Since X is a real matrix, its complex eigenvalues occur
in conjugate pairs, so both roots of the polynomial p given by p(x) = x 2 − x + 1 are
eigenvalues of X, and p is the characteristic polynomial of X. This yields I − X + X2 = 0,
or N = 0, a contradiction.
Also solved by G. Bourgeois (France), S. M. Gagola Jr., K. Gatesman, J.-P. Grivaux (France), J. W. Hagood,
E. A. Herman, E. J. Ionaşcu, K. T. L. Koo (China), J. H. Lindsey II, O. P. Lossers (Netherlands),
K. D. McLenithan, M. Omarjee (France), A. Pathak, A. Stadler (Switzerland), R. Stong, J. Stuart & R. Horn,
R. Tauraso (Italy), L. Zhou, UM6P Math Club (Morocco), and the proposer.

An Equilateral Triangle and a Circle

12278 [2021, 756]. Proposed by Dao Thanh Oai, Thai Binh, Vietnam. Let ABC be a sca-
lene triangle, and let its external angle bisectors at A, B, and C meet BC, CA, and AB at
D, E, and F , respectively. Let l, m, and n be lines through D, E, and F that (internally)
trisect angles ∠ADB, ∠BEC, and ∠CF A, respectively, with the angle between l and AD
equal to 1/3 of ∠ADB, the angle between m and BE equal to 1/3 of ∠BEC, and the angle
between n and CF equal to 1/3 of ∠CF A.
(a) Show that l, m, and n form an equilateral triangle.
(b) The lines l, m, and n each intersect AD, BE, and CF . Of these nine points of intersec-
tion, three are the points D, E, and F . Show that the other six lie on a circle.
Solution by Li Zhou, Polk State College, Winter Haven, FL.
(a) We use A, B, and C to denote both
the vertices of ABC and the interior
angles at those vertices. We may assume
A < B < C. We also use D and E
to denote ∠CDA and ∠BEC, respec-
tively. Let J be the intersection of AD
and BE, K the intersection of BE and
CF , and L the intersection of CF and
AD. Let P be the intersection of l and
m, Q the intersection of m and n, and R
the intersection of n and l.
By construction, the three triangles
J AB, CKB, and CAL are all
similar to J KL, with interior angles
(π − C)/2, (π − A)/2, and (π − B)/2
at J , K, and L, respectively. Also,
π −A B +C C−B
D = π − ∠ACD − ∠DAC = π − (π − C) − =C− = ,
2 2 2
and similarly E = (C − A)/2. Therefore,
D E π −C C−B C−A 3π − (A+B +C) π
∠RP Q = ∠LJ K + + = + + = = .
3 3 2 6 6 6 3
Similarly, ∠P QR = ∠QRP = π/3, so P QR is equilateral.

June–July 2023] PROBLEMS AND SOLUTIONS 591

(b) Suppose that l intersects BE at U and CF at X, m intersects CF at V and AD at Y ,
and n intersects AD at W and BE at Z. Applying the law of sines to J BA, we get
JA JB
= , (1)
sin((π − B)/2) sin((π − A)/2)
and applying it to J BD and J EA yields
JD JB JE JA
= , = . (2)
sin((π − B)/2) sin D sin((π − A)/2) sin E
Also, ∠J U D = π − ∠P U E = ∠EP U + ∠U EP = (2π + E)/3, and similarly ∠EY J =
(2π + D)/3. Therefore applying the law of sines to J U D and EY J gives us
JU JD JY JE
= and = . (3)
sin(D/3) sin((2π + E)/3) sin(E/3) sin((2π + D)/3)
Combining (1), (2), and (3) yields
JU sin E sin(D/3) sin((2π + D)/3)
= .
JY sin D sin(E/3) sin((2π + E)/3)
By the triple-angle formula,

sin D = sin(D/3)(3 cos2 (D/3) − sin2 (D/3))

= 4 sin(D/3) sin((2π + D)/3) sin((π + D)/3),

and the same is true if angle D is replaced with angle E. Thus,

JU sin((π + E)/3)
= .
JY sin((π + D)/3)
Finally,

∠ZW J = π − ∠DW R = ∠W RD + ∠RDW = (π + D)/3,

and similarly ∠J ZW = (π + E)/3. Therefore, the law of sines applied to J ZW yields

JW JZ
= ,
sin((π + E)/3) sin((π + D)/3)
and hence
JW sin((π + E)/3) JU
= = .
JZ sin((π + D)/3) JY
We conclude that U , Y , W , and Z lie on a circle. Likewise, V , Z, U , and X lie on a
circle, and W , X, V , and Y lie on a circle. If the three circles are distinct, then the three
radical axes U Z, V X, and W Y are concurrent. But these axes are J K, KL, and LJ , which
are not concurrent. Therefore, two of the three circles are the same, so the six points are all
on the same circle.
Editorial comment. Zhou points out that applying Pascal’s theorem to the hexagon
U XV Y W Z shows that D, E, and F are collinear. Thus, by the theorem of Desargues,
J P , KQ, and LR are concurrent.
Also solved by C. R. Pranesachar (India), R. Stong, and the proposer.

592 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

 A Stirling Identity
12279 [2021, 856]. Proposed by Brad Isaacson, Brooklyn, NY. Let S(m, k) denote the
number of partitions of a set with m elements into k nonempty blocks. (These are the
Stirling numbers of the second kind.) Let j and n be positive integers of opposite parity
with j < n. Prove

 n
(−1)r (r − 1)! jr S(n, r)
= 0.
r=j
2r

Solution I by Omran Kouba, Higher Institute for Applied Sciences and Technology, Dam-
ascus, Syria. Note that r!S(n, r) is the number of surjective mappings from a set with n
elements onto a set with r elements. Therefore, by inclusion-exclusion,
     
dn 
r r
r−k r r−k r dn t
r!S(n, r) = (−1) k =
n
(−1) e kt
= (e − 1) r
.
k=1
k dt n k=0 k dt n t=0
t=0

Let a(n, j ) denote the sum in question. Since S(n, r) = 0 for r > n,
⎡ ⎤ ⎡ ⎤
n ∞ r    ∞   t r
d (−1) r d n
1 r − 1 1 − e
a(n, j ) = ⎣ n et − 1 ⎦ = ⎣ n ⎦
r
.
dt r=j r2r j dt j r=j j − 1 2
t=0 t=0

(The interchange of the derivative and summation can be justified by showing that the series
of derivatives
 converges
r−1 r uniformly on an interval around 0.) From the negative binomial
expansion, ∞ r=j j −1 x = x j
/(1 − x)j . Hence,
 j
1 dn 1 − et
a(n, j ) = · .
j dt n 1 + et
t=0

Since (1 − et )/(1 + et ) is odd, so is

 j
dn 1 − et
dt n 1 + et
when j and n have opposite parity. Therefore, a(n, j ) = 0 in this case.
Solution II by Tewodros Amdeberhan and Victor H. Moll, Tulane University, New Orleans,
LA. Let a(n, j ) denote the sum in question. We proceed by induction on n, beginning
with a(2, 1) = −1/2 + 2/4 = 0. Grouping the partitions by whether n is a part  by itself,
S(n, r) = rS(n − 1, r) + S(n − 1, r − 1). With the standard conventions that jr = 0 for
j < 0 or j > r and that S(n, r) = 0 for r > n, we use the recurrence and reindexing to
obtain

n
(−1)r (r − 1)! jr S(n, r)
a(n, j ) =
r=0
2r
  
n
(−1)r (r − 1)! jr rS(n − 1, r) + S(n − 1, r − 1)
=
r=0
2r
    
n
(−1)r r! jr − r+1 j
/2 S(n − 1, r)
= r
.
r=0
2

June–July 2023] PROBLEMS AND SOLUTIONS 593

Via three applications of the binomial recurrence,
         
r 1 r +1 r 1 r 1 r
− = − −
j 2 j j 2 j 2 j −1
           
1 r −1 1 r −1 1 r −1 1 r −1 1 r −1 1 r −1
= + − − = − .
2 j 2 j −1 2 j −1 2 j −2 2 j 2 j −2
Substituting this identity into the previous expression for the sum yields
    
 n
(−1)r r! r−1
j
/2 − jr−1−2
/2 S(n − 1, r)
a(n, j ) =
r=0
2r
  r  j −1  r 
 n
(−1)r (r − 1)! j +1
2 j +1
− 2 j −1 S(n − 1, r)
= r
r=0
2
j +1 j −1
= a(n − 1, j + 1) − a(n − 1, j − 1).
2 2
By convention a(n − 1, n) = 0, and the rightmost term is 0 when j = 1. In all other cases,
when j and n have opposite parity, the induction hypothesis implies a(n − 1, j ± 1) = 0.
We conclude a(n, j ) = 0.
Also solved by U. Abel & V. Kushnirevych (Germany), A. Berkane (Algeria), A. De la Fuente, O. P. Lossers
(Netherlands), J. H. Nieto (Venezuela), A. Stadler (Switzerland), R. Tauraso (Italy), M. Wildon (UK),
UM6P Math Club (Morocco), and the proposer.

A Hyperbolic Logarithmic Integral

12281 [2021, 856]. Proposed by Paolo Perfetti, University of Rome Tor Vergata, Rome,
Italy. Evaluate
∞ 
cosh x 1  2
2
− 2 ln x dx.
0 sinh x x
Solution by Michel Bataille, Rouen, France. Let I be the integral to be evaluated. We show
that I = (ln 2)(2γ − ln 2 − 2 ln π ), where γ is Euler’s constant.
Suppose 0 < a < b. Integrating by parts gives
b  b  
cosh x 1 ln x 1 1
− (ln x) 2
dx = F (b) − F (a) − 2 − dx,
a sinh2 x x2 a x x sinh x
where F (x) = (ln x)2 (1/x − 1/ sinh x). Since
 
(ln b)2 2(ln b)2
lim F (b) = lim − b =0−0=0
b→∞ b→∞ b e − e−b
and
sinh a − a 1
lim F (a) = lim a(ln a)2 · = 0 · = 0,
a→0+ a→0+ a 2 sinh a 6
we conclude
∞  
ln x 1 1
I = −2 − dx.
0 x x sinh x
It is known that for x = 0,
∞
1 1  2x
= + (−1)n 2
sinh x x n=1 x + n2 π 2

594 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

(see I. S. Gradshteyn, I. M. Ryzhik (2007), Table of Integrals, Series, and Products, 7th
ed., Burlington, MA: Academic Press, p. 27, equation 1.217.2). Hence,
∞ 
∞ 
4(−1)n ln x
I= dx.
0 n=1
x 2 + n2 π 2
Next we show that we can reverse the order of the integration and summation in this
formula. For 0 < x ≤ 1 and N a positive integer, we have
   ∞
4 ln x  1
N N N
4(−1)n ln x 4(−1)n ln x −4 ln x 2 ln x
≤ ≤ ≤− 2 =− ,
n=1
x 2 + n2 π 2 n=1
x 2 + n2 π 2
n=1
n2π 2 π n=1
n2 3
1
and 0 −(2/3) ln x dx < ∞. It follows, by the dominated convergence theorem, that
∞
1  ∞ 1
4(−1)n ln x 4(−1)n ln x
dx = dx. (4)
0 n=1
x 2 + n2 π 2 n=1 0 x 2 + n2 π 2
Similarly, for x ≥ 1 and N a positive integer,

N
4(−1)n ln x 4 ln x 4 ln x
≤ ≤
n=1
x 2 + n2 π 2 x2 + π 2 x2
∞
and 1 4(ln x)/x dx < ∞, so
2

∞
∞  ∞ ∞
4(−1)n ln x 4(−1)n ln x
dx = dx. (5)
1 n=1
x +n π
2 2 2
n=1 1 x 2 + n2 π 2
Combining (4) and (5), we have
∞
∞ ∞ ∞
4(−1)n ln x ln x
I= dx = 4 (−1)n dx. (6)
0 n=1
x 2 + n2 π 2 n=1 0 x2 + n2 π 2
To evaluate the integral on the right side of (6), we first use the substitution u = x/(nπ ),
as follows:
∞ ∞
ln x 1 ln x
2 + n2 π 2
dx = dx
0 x n2π 2
0 (x/(nπ ))2 + 1
 ∞ ∞ 
1 ln(nπ ) ln u
= du + du
nπ 0 u2 + 1 0 u2 + 1
∞
ln(nπ ) 1 ln u
+ = du.
2n nπ 0 u + 1
2

The last integral above vanishes, as can be seen by making the substitution t = 1/u:
∞ ∞ ∞
ln u − ln t 1 ln u
du = · 2 dt = − du.
0 u +1
2
0 1/t + 1 t
2
0 u2 +1
Substituting into (6), we obtain
∞
∞ ∞

  (−1)n ln n  (−1)n
n ln(nπ )
I =4 (−1) =2 + ln π .
n=1
2n n=1
n n=1
n

Finally, we use the formulas ∞ n=1 (−1)
n−1
/n = ln 2 and
∞

(−1)n−1 (ln n)/n = (ln 2)2 /2 − γ ln 2
n=1

June–July 2023] PROBLEMS AND SOLUTIONS 595

(see the solution to problem 873 in Coll. Math. J. 40(2), March 2009, pp. 136–137) to
conclude
 
(ln 2)2
I = 2 γ ln 2 − − ln π ln 2 = (ln 2)(2γ − ln 2 − 2 ln π ).
2

Also solved by U. Abel & V. Kushnirevych (Germany), T. Amdeberhan & V. H. Moll, A. Berkane (Alge-
ria), N. Bhandari (Nepal), K. N. Boyadzhiev, P. Bracken, H. Chen, G. Fera (Italy), M. L. Glasser, N. Hodges
(UK), J. E. Kampmeyer, L. Kempeneers & J. Van Casteren (Belgium), O. Kouba (Syria), M. Omarjee (France),
A. Stadler (Switzerland), A. Stenger, S. M. Stewart (Saudi Arabia), M. S̆tofka (Slovakia), R. Stong, R. Tauraso
(Italy), Fejéntaláltuka Szeged Problem Solving Group (Hungary), UM6P Math Club (Morocco), and the
proposer.

CLASSICS
C16. Suggested by the editors. Two hikers start together at the bottom of a mountain and
climb to the summit but along different trails, which may go up and down along the way.
Show that it is possible for them to complete their respective hikes in such a way that they
are at the same elevation at every moment.

Costly Positive Integers

C15. Suggested by Joel Spencer, New York University, New York, NY. A construction chain
for n is a sequence a1 , . . . , ak where a1 = 1, ak = n, and each entry in the sequence
is either the sum or the product of two previous, possibly identical, elements from the
sequence. The cost of a construction chain is the number of entries that are the sum (but
not the product) of preceding entries. For example, 1, 2, 3, 6, 12, 144, 1728, 1729 is a con-
struction chain for 1729; its cost is 3, because the elements 2, 3, and 1729 require addition.
Let c(n) be the minimal cost of a construction chain for n. Prove that c is unbounded.
Solution. We show that, given n, the total number of construction chains for numbers less
2
than or equal to n and with cost K or less is at most K(1 + log2 n)2K . Since this is less
than n for large n, some integer does not have a construction chain with cost K or less.
Suppose that a1 , . . . , ak is a construction chain for m with m ≤ n having cost s, with
0 ≤ s ≤ K. Let b1 , . . . , bs+1 be the subsequence of a1 , . . . , ak with b1 = a1 = 1 consisting
of all entries that were produced using addition. For 2 ≤ i ≤ s + 1,

i−1
e 
i−1
f
bi = bjj + bj j ,
j =1 j =1

where ej and fj are nonnegative integers. Note that ej and fj are in {0, 1, . . . , log2 n }.
Hence, the number of choices for bi with 2 ≤ i ≤ s + 1 is bounded above by
(1+log2 n)2(i−1) . This is at most (1+log2 n)2s . Hence, the number of possible sequences
2 2
b1 , . . . , bs+1 is at most (1 + log2 n)2s , which in turn is bounded by (1 + log2 n)2K .
2
Summing over all costs s from 1 to K yields at most K(1 + log2 n)2K , as claimed.
Editorial Comment. We do not know the origin of this problem.
If the number of primes were finite, we could calculate them all with finitely many addi-
tions of 1, and then any composite could be computed with zero additional cost. Therefore
a corollary of the problem is that the number of primes is infinite. It is challenging to com-
pute c(n). Work of Joseph DeVincentis, Stan Wagon, and Alan Zimmermann has led to
results on the cost function for n beyond one million. For k ≥ 0, let Mk be the least n
such that c(n) = k. The sequence M0 , M1 , . . . begins 1, 2, 3, 7, 23, 719, 1169951. See
oeis.org/A355015 and also the related oeis.org/A354914.

596 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

SOLUTIONS

A Recurrence Yielding Factorials

12265 [2021, 658]. Proposed by Ross Dempsey, student, Princeton University, Princeton,
NJ. For a fixed positive integer k, let a0 = a1 = 1 and an = an−1 + (k − n)2 an−2 for n ≥ 2.
Show that ak = (k − 1)!.
Solution by Jovan Vukmirović, Belgrade, Serbia, and UM6P Math Club, Mohammed
VI Polytechnic University, Ben Guerir, Morocco, independently. Let
bn = an + (k − n − 1)an−1 . Note bk−1 = ak−1 . In general, an = an−1 + (k − n)2 an−2
implies
 
bn = an − an−1 + (k − n)an−1 = (k − n) (k − n)an−2 + an−1 = (k − n)bn−1 .

Therefore,
bn = (k − n)bn−1
= (k − n)(k − n + 1)bn−2 = · · ·
= (k − n)(k − n + 1) · · · (k − 2)b1 .

In particular, bk = 0. Since b1 = k − 1,

ak = bk + ak−1 = ak−1 = bk−1 = (k − 1)!.

Also solved by M. R. Bacon & C. K. Cook, B. Bradie, A. C. Castrillón (Colombia), H. Chen (China),
A. De la Fuente, H. Y. Far, K. Gatesman, J. F. Gonzalez & F. A. Velandia (Colombia), J.-P. Grivaux (France),
E. A. Herman, N. Hodges (UK), E. J. Ionaşcu, O. Kouba (Syria), P. Lalonde (Canada), O. P. Lossers (Nether-
lands), R. Martin (Germany), A. Natian, M. Omarjee (France), C. R. Pranesachar (India), M. Reid, J. L. Guerra
& A. J. Rosenthal, K. Sarma (India), A. Stadler (Switzerland), A. Stenger, R. Stong, R. Tauraso (Italy),
M. Tetiva (Romania), J. Vinuesa (Spain), M. Wallner (Austria), H. Widmer (Switzerland), M. Wildon (UK),
L. Zhou, Davis Problem Solving Group, and the proposer.

486 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
 Arbitrarily Disconnectable Polyominos
12266 [2021, 658]. Proposed by Haoran Chen, Xi’an Jiaotong–Liverpool University,
Suzhou, China. A union of a finite number of squares from a grid is called a polyomino if
its interior is simply connected. Given a polyomino P and a subpolyomino Q, we write
c(P , Q) for the number of components that remain when Q is
removed from P . Let f (k) = maxP minQ c(P , Q), where the
maximum is taken over all polyominoes and the minimum is
taken over all subpolyominoes Q of P of size k. For example,
f (2) ≥ 3, because any domino removed from the pentomino at
right breaks the pentomino into 3 pieces. Is f bounded?
Solution by Richard Stong, Center for Communications Research, San Diego, CA. We show
that f is unbounded. With any polyomino P we can associate a graph G by taking a vertex
for each square of P and making vertices adjacent when their squares share a side. We use
only polyominos P where the resulting graph G is a tree. The removed subpolyomino Q
will correspond to a subtree H , so that the graph associated with Q − P will be G − V (H ),
and they will have the same number of components.
We use a polyomino whose associated graph is a subdivision of a complete binary tree.
Let Gh,N be the subdivision of the complete binary tree with height h in which each edge
is replaced by a path of length N . For fixed h, we prove that Gh,N is the graph associated
with some polyomino when N is sufficiently large. It then suffices to show that when m is
fixed, for sufficiently large h and N there is a choice of k such that deleting the vertices of
any k-vertex subtree of Gh,N results in at least m components.
Let Th be the complete binary tree of height h, with 2h leaves. We initially represent a
subdivision of Th and can then lengthen paths appropriately to obtain Gh,N . The vertices of
Th at distance j from the leaves will initially be on the line y = 2j , and the root will be at
(0, 2h). For h = 0, place the root at the origin. For h ≥ 1, having embedded a subdivision
of Th−1 with leaves on the horizontal axis (with consecutive leaves separated by 2), take
two copies and shift one rightward to have leaves at odd points (1, 0) through (2h − 1, 0),
and shift the other leftward to have leaves at odd points (−1, 0) through (−2h + 1, 0). The
roots of the two copies will now be at (2h−1 , 2h − 2) and (−2h−1 , 2h − 2). Place the root of
Th at (0, 2h). The edge from (0, 2h) to its right child is represented by a path from (0, 2h)
to (2h−1 , 2h) and then down two steps to (2h−1 , 2h − 2); the path to (−2h−1 , 2h − 2) is the
reflection of this. Here is T3 :

This construction requires N ≥ 2h−1 + 2. The vertical steps involved in a given level
can be lengthened by the same amount to produce an embedding of Gh,N associated with
a polyomino Ph,N . The vertical steps have length at least 2 to avoid unwanted edges in the
associated graph. 
Let a 2-power sum be an integer of the form i εi 2ai , where εi ∈ {1, −1} and ai is
a nonnegative integer for all i. When we consider deleting the vertices of a subtree, the
following claim is helpful.

May 2023] PROBLEMS AND SOLUTIONS 487

Claim: Given a positive integer m, there is a positive integer t such that if |t − u| ≤ 2m,
then any expression of u as a 2-power sum has more than m terms.
To prove the claim, we show that when R is sufficiently large, there are congruence classes
modulo 2R that can serve as t. Powers of 2 and their negations take on only 2R + 1 distinct
values modulo 2R . Hence sums of m such terms take on at most (2R + 1)m values mod-
ulo 2R . Within 2m units of such values there are at most (4m + 1)(2R + 1)m congruence
classes. Since a polynomial in R grows more slowly than 2R , when R is sufficiently large
we can pick t from any of the remaining congruence classes.
Fix m, and let t be an integer as guaranteed by the claim. Choose h so that 2h+1 − 2 > t.
Let G = Gh,N for some large N , and let k = tN . We claim that for any subtree H of G
with k vertices, G − V (H ) has at least m components. Since m is arbitrary, this makes f
unbounded.
Let v be a vertex of H closest to the root of G. Let the distance from v to the leaves
below it be rN + s, where 0 ≤ s < N . The subtree of G rooted at v has 1 + s + (2r+1 − 2)N
vertices. Let S be the set of vertices w in G such that w is not in H but the parent of w is
in H . Let ri N + si be the distance from the ith vertex of S to the leaves below it, where
0 ≤ si < N. The vertices of H are precisely the descendants of v that are not descendants
of vertices in S. Thus
 
N t = k = |V (H )| = 1 + s + (2r+1 − 2)N − 1 + si + (2ri +1 − 2)N .
i

Let u = 2r+1 − i 2ri +1 . The difference between t and u is

(1 + s − 2N )/N − (1 + si − 2N )/N.
i

Since 0 ≤ si < N and 0 ≤ s < N , each term lies between −2 and 2. Hence |t − u| ≤ 2m
if |S| < m. Since u is a 2-power sum with |S| + 1 terms, the choice of t yields |S| ≥ m.
That is, G − V (H ) has at least m components.
Also solved by the proposer.

Balanced Colorings of Graphs

12268 [2021, 658]. Proposed by Samina Boxwala Kale, Nowrosjee Wadia College, Pune,
India, Vas̆ek Chvátal, Concordia University, Montreal, Canada, Donald E. Knuth, Stanford
University, Stanford, CA, and Douglas B. West, University of Illinois, Urbana, IL.
(a) Show that there is an easy way to decide whether the edges of a graph can each be
colored red or green so that at each vertex the number of incident edges with one color
differs from the number having the other color by at most 1.
(b) Show that it is NP-hard to decide whether the vertices of a graph can each be colored
red or green so that at each vertex the number of neighboring vertices with one color differs
from the number having the other color by at most 1.
Solution by Edward Schmeichel, San Jose State University, San Jose, CA. In both (a) and
(b), we call a coloring of the specified type a balanced coloring. The existence of balanced
colorings in one component does not affect their existence in others, so we can apply the
criterion for connected graphs to each component.
(a) A connected graph G fails to have a balanced edge-coloring if and only if all vertices
have even degree and the number of edges is odd.
If all vertices have even degree and G has a balanced edge-coloring, then the subgraphs
in the two colors have the same degree at each vertex and hence the same number of edges,
which is impossible when the number of edges is odd.

488 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
 If the vertices have even degree and the number of edges is even, then assigning colors
alternately along an Eulerian circuit gives half of the edges at each vertex to each color.
If some vertex has odd degree, then the number of vertices with odd degree is even, and
adding one vertex v and making it adjacent to all the vertices of odd degree produces a
connected graph G with all vertex degrees even. In G there is an Eulerian circuit starting
and ending at v. Assigning colors alternately along the circuit gives each vertex other than
v the same number of edges of each color, and then deleting the edges at v produces a
balanced edge-coloring of G.
(b) We show that if there is a polynomial-time algorithm to test whether a balanced vertex
coloring exists, then there is a polynomial-time algorithm for the following well-known
NP-hard problem.
NOT-ALL-EQUAL 3SAT: Given variables x1 , . . . , xn and clauses c1 , . . . , cm , where
each clause is a set of three “literals” (variables or their complements), is there a truth
assignment to the variables so that each clause contains both a true literal and a false
literal?
Given an instance I of NOT-ALL-EQUAL 3SAT, we construct a graph G such that I
is satisfiable if and only if G has a balanced vertex coloring. For each clause ci , create a
set Si of three independent vertices labeledby the literals in ci , together with a vertex σi
adjacent to all three vertices in Si . Let S = i Si . Note that S is an independent set of size
3m; labels may appear on more than one vertex.
Next we add vertices and edges to G to ensure that in a balanced vertex coloring, ver-
tices in S having the same label will have the same color, while vertices with comple-
mentary labels will have opposite colors. Think of green as representing TRUE and red as
representing FALSE.
For each instance of two vertices v and w in S with identical labels, add a star with four
edges, with each of v and w adjacent to two leaves of the star, giving those leaves degree
2. The leaves of the star need neighbors of opposite colors, so v and w must have the same
color in a balanced vertex coloring.
For each instance of two vertices v and w in S with complementary labels, add two new
vertices, with v and w adjacent to both. The new vertices have degree 2, and hence v and
w must have opposite colors in a balanced vertex coloring.
If G has a balanced vertex coloring, then the balance condition at each σi guarantees
that each clause has a vertex of each color. Thus a balanced vertex coloring of G converts
to a satisfying truth assignment for I .
Conversely, given a satisfying truth assignment for I , using green on vertices labeled
with true literals and red on vertices labeled with false literals fulfills the balance condition
at each σi . Each vertex of S is adjacent to an even number of added vertices, and we can
color the added vertices so that each vertex of S has the same number of neighbors of each
color among the added vertices. Since each vertex of S is adjacent to only one vertex of
the form σi , we can then color the vertices of that form arbitrarily to complete a balanced
vertex coloring of G.
Editorial comment. In G. P. Cornuéjols (1988), General Factors of Graphs, J. Comb. Th. B
45, 185–198, it is shown that for any nonnegative integer k, there is a polynomial-time
algorithm to decide whether the edges of a graph can be colored red or green so that
at each vertex the numbers of incident edges of the two colors differ by at most k. For
part (b), Mark Wildon reduced a variant of the Subset Sum problem to the given coloring
problem.
Also solved by R. Stong, M. Wildon (UK), and the proposers.

May 2023] PROBLEMS AND SOLUTIONS 489

 Integrating an Absolute Value
12271 [2021, 659]. Proposed by Steven Deckelman, University of Wisconsin–Stout,
Menomonie, WI. Let n be a positive integer. Evaluate
 2π   
 π 
sin (n − 1)θ − cos(nθ ) dθ.
0 2n

Solution by Jovan Vukmirović, Belgrade, Serbia. Let In denote the requested integral. We
show that
⎧  π  4(n − 1)
⎪
⎪ 4n π
⎨ cot − cot , if n is even;
In = 2n − 1 2n 2n − 1 2(n − 1)
  4(n − 1)
⎪
⎪ 4n π π
⎩ csc − csc , if n is odd.
2n − 1 2n 2n − 1 2(n − 1)
Since the integrand is periodic with period π , the substitution θ = x − π/(2n) gives
 π/2
 
In = 2 cos((n − 1)x) sin(nx) dx.
−π/2
 
Let fn (x) = cos (n − 1)x sin(nx). Since |fn (x)| is an even function, we have
 π/2
In = 4 |fn (x)| dx.
0

Note that the function Fn defined by

1 1  
Fn (x) = − cos x + cos (2n − 1)x
2 2n − 1
is an antiderivative of fn . When x ∈ [0, π/2] we have f1 (x) ≥ 0, so
 π/2
 
I1 = 4 f1 (x) dx = 4 F1 (π/2) − F1 (0) = 4.
0
 
Now suppose n ≥ 2. The positive  values of x where cos (n − 1)x changes sign are
given by ck = (2k − 1)π/ 2(n − 1) , and the values where sin(nx) changes sign are given
by dk = kπ/n, for k = 1, 2, . . .. Setting m = n/2 , we have
π
0 < c1 < d1 < c2 < · · · < cm ≤ dm ≤ < dm+1 < cm+1 ,
2
so fn (x) is negative for ck < x < dk , k = 1, . . . , m, and nonnegative at all other points in
[0, π/2]. Hence
 π/2 m
 
|fn (x)| dx = Fn (π/2) − Fn (0) − 2 Fn (dk ) − Fn (ck ) ,
0 k=1

and the desired integral is given by

m
4n 8n kπ 8(n − 1) (2k − 1)π
In = + cos − cos .
2n − 1 k=1
2n − 1 n 2n − 1 2(n − 1)

To simplify the sum, we apply the identity

 
sin a + (2m + 1)b/2 − sin(a + b/2)
cos(a + b) + cos(a + 2b) + · · · + cos(a + mb) =
2 sin(b/2)

490 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
(easily verified by induction on m) to get
   
4n sin (2m + 1)π/(2n) 4(n − 1) sin mπ/(n − 1)
In = ·   − ·  .
2n − 1 sin π/(2n) 2n − 1 sin π/(2(n − 1))
Since m is equal to n/2 if n is even and (n − 1)/2 if n is odd, we obtain the desired formula
for In .
Note that In → 8/π as n → ∞.
Also solved by G. Fera (Italy), D. Henderson, N. Hodges (UK), O. Kouba (Syria), O. P. Lossers (Netherlands),
A. Natian, A. Stadler (Switzerland), M. S̆tofka (Slovakia), R. Stong, E. I. Verriest, and the proposer.

Lists Whose Consecutive Terms Sum to Powers of 2

12272 [2021, 755]. Proposed by H. A. ShahAli, Tehran, Iran, and Stan Wagon, Macalester
College, St. Paul, MN.
(a) For which integers n with n ≥ 3 do there exist distinct positive integers a1 , . . . , an such
that ai + ai+1 is a power of 2 for all i ∈ {1, . . . , n}? (Here subscripts are taken modulo n,
so that an+1 = a1 .)
(b) What is the answer if the word “positive” is removed from part (a)?
Solution by Rory Molinari, Michigan. For (a) there is no such n, but for (b) there exist such
lists for all n except n = 4.
(a) Suppose that a1 , . . . , an is such a list. By symmetry, we may assume a1 < a2 . Let
ai + ai+1 = 2ci for all i. Since ai−1 = ai+1 , we have ci−1 = ci . If ai−1 < ai and ai > ai+1 ,
then
ai > max{2ci−1 /2, 2ci /2} ≥ min{2ci−1 , 2ci },
from which min{ai−1 , ai+1 } is negative. Hence a1 < · · · < an < a1 , a contradiction.
(b) Suppose that distinct integers a1 , . . . , a4 exist such that ai + ai+1 = 2ci . By symmetry,
we may assume that a1 is the smallest. Now
0 < a3 − a1 = 2c2 − 2c1 = 2c3 − 2c4 .
Consequently, c1 and c4 are both the exponent of the greatest power of 2 dividing a3 − a1 .
Hence c1 = c4 , which yields a2 = a4 , a contradiction.
For n = 3, one such list is (3, −1, 5).
Let αi = 1 − 2i and βi = 3 · 2i − 1. For even n at least 6, with k = n/2, consider the
list
(1, 3, α1 , β1 , . . . , αk−2 , βk−2 , αk−1 , 2k − 1).
Since 1 + 3 = 4, 3 + α1 = 2, αi + βi = 2i+1 , βi + αi+1 = 2i , αk−1 + 2k − 1 = 2k−1 , and
2k − 1 + 1 = 2k , every sum of two cyclically consecutive elements is a power of 2. Since
0 > α1 > · · · > αk−1 and 3 < β1 < · · · < βk−2 < 2k − 1, the terms are distinct.
When n = 2k − 1 ≥ 5, it suffices to use the list for 2k with the term α1 deleted, since
3 + β1 = 8.
Editorial comment. Yuri Ionin strengthened the conclusion in part (a), using induction to
prove that positive integers a1 , . . . , an chosen so that cyclically ai + ai+1 is always a power
of 2 has at most (n + 1)/2 distinct elements and that this bound is sharp.
Also solved by C. Curtis & J. Boswell, S. M. Gagola Jr., K. Gatesman, O. Geupel (Germany), N. Hodges
(UK), Y. J. Ionin, M. D. Meyerson, M. Reid, A. Stadler (Switzerland), R. Tauraso (Italy), F. A. Velandia &
J. F. Gonzalez (Colombia), J. Yan (China), Fejéntaláltuka Szeged Problem Solving Group (Hungary), and the
proposer. Part (a) also solved by H. Chen (China), O. P. Lossers (Netherlands), R. Martin (Germany), L. Zhou,
and the UM6P Math Club (Morocco).

May 2023] PROBLEMS AND SOLUTIONS 491

 Zeta Function Inequalities from Convexity
12273 [2021, 755]. Proposed by Hideyuki
 Ohtsuka, Saitama, Japan. Let ζ be the Riemann
zeta function, defined by ζ (s) = ∞ s
k=1 1/k . For s > 1, prove the following inequalities:

1 1 ζ (s) 1 ζ (s)
< log ζ (s), < log √ , < log .
prime p
ps − 0.5 prime p
p s ζ (2s) prime p
ps + 0.5 ζ (2s)

Composite solution by Allen Stenger, Boulder, CO, and Li Zhou, Polk State College, Winter
Haven, FL. We prove the more general inequality
1 ζ (s)
< log , (∗)
p
ps + α (ζ (2s))α+1/2

where −1/2 ≤ α ≤ 1/2 and the sum is over all primes. The three requested inequalities
are for α ∈ {−1/2, 0, 1/2}. 
The Euler product formula for ζ (s) with s > 1 is ζ (s) = p 1/(1 − p−s ), where the
product is taken over all primes. Hence the right side of (∗) is the logarithm of

   α+1/2
1 1
,
p
1 − p−s p
1 − p−2s

which simplifies to p (1 − p−2s )α+1/2 /(1 − p−s ), where the products are over all primes.
 
Letting R = log (1 − p−2s )α+1/2 /(1 − p−s ) , we obtain the desired inequality term-by-
term by proving R > 1/(ps + α). We compute
 
R = (α + 1/2) log (1 − p−s )(1 + p−s ) − log(1 − p−s )
ps − 1 ps + 1
= (α − 1/2) log + (α + 1/2) log
ps ps
1 − 2α   1 + 2α  
= log ps − log(ps − 1) + log(ps + 1) − log ps
2 2
 ps  ps +1
1 − 2α 1 + 2α
= dx + dx.
ps −1 2x ps 2x
We obtain lower bounds on these integrals using the left side of the Hermite–Hadamard
inequality
 b
a+b 1 f (a) + f (b)
f ≤ f (x)dx ≤
2 b−a a 2
for convex f , with the inequalities being strict when f is strictly convex. Applying the
Hermite–Hadamard inequality to both integrals in the final expression for R yields
1 − 2α 1 + 2α
R> + s .
2p − 1 2p + 1
s

Letting u = ps , it now suffices to prove

1/2 − α 1/2 + α 1
+ ≥
u − 1/2 u + 1/2 u+α
for −1/2 ≤ α ≤ 1/2 and u ≥ 2. Letting g(α) denote the left side minus the right side
in this inequality, we compute g  (α) = −2/(u + α)3 < 0. Thus g is a concave function,

492 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
and its minimum on the interval [−1/2, 1/2] occurs at an endpoint. Since g(−1/2) =
g(1/2) = 0, we have g(α) ≥ 0 throughout the interval, and the result follows.
Editorial comment. The proof above uses only the left side of the Hermite–Hadamard
inequality. Applying the right side to the convex function ex yields
eb − ea eb + ea
< .
b−a 2
For b = 2/u and a = 0, this reduces to e2/u − 1 < (e2/u + 1)/u. For u > 1, we can rear-
range and take logarithms to obtain 2/u < log (u + 1)/(u − 1) . The proposer used this
last inequality to show that one can start from any of the specified sums in the problem and
build up to the desired expression in terms of the zeta function without a decrease at any
step of the process. For example,
2 2ps − 1 + 1 ps
< log = log = log ζ (s).
p
2p − 1
s
p
2ps − 1 − 1 p
ps − 1

This solution proceeds in the opposite direction from the solution presented above.
Also solved by H. Chen, D. Fleischman, K. Gatesman, O. Kouba (Syria), K.-W. Lau (China), O. P. Lossers
(Netherlands), K. Nelson, M. Omarjee (France), D. Pinchon (France), A. Stadler (Switzerland), R. Stong,
R. Tauraso (Italy), J. Vinuesa (Spain), M. Vowe (Switzerland), T. Wiandt, J. Yan (China), Fejéntaláltuka Szeged
Problem Solving Group (Hungary), UM6P Math Club (Morocco), and the proposer.

A Trigonometric Logarithmic Integral

12274 [2021, 755]. Proposed by Roberto Tauraso, University of Rome Tor Vergata, Rome,
Italy. Evaluate
 1 2
arctan x 2x
ln dx.
0 1+x 1 − x2
2

Solution by Michel Bataille, Rouen, France. Let I be the integral to be evaluated. We show
that I = π 4 /128.
The change of variables x = tan(u/2) readily leads to

1 π/2
I= u(ln tan u)2 du.
4 0
Using the substitution u = π/2 − v we obtain
 π/2  π/4    π/4  
π π
u(ln tan u) du =
2
− v (ln(cot v)) dv =
2
− v (ln tan v)2 dv,
π/4 0 2 0 2
from which we deduce
 π/4  π/4   
1 π π π/4
I= u(ln tan u) du +
2
− u (ln tan u) du =
2
(ln tan u)2 du.
4 0 0 2 8 0
Finally, the substitution u = arctan t gives
 ∞ 
π 1 (ln t)2 π 1
I= dt = (−1)n t 2n (ln t)2 dt
8 0 1 + t2 8 n=0 0

∞
π (−1)n π π π3 π4
= = β(3) = · = ,
4 n=0
(2n + 1)3 4 4 32 128

where β is the Dirichlet beta function.

May 2023] PROBLEMS AND SOLUTIONS 493

Editorial comment. Seán M. Stewart derived the more general formula
 1 2n
arctan x 2x π
ln dx = (2n)!β(2n + 1).
0 1+x 1−x
2 2 8
1
The integral 0 (ln t)2 /(1 + t 2 ) dt also made an appearance in the solution of Problem
12158 [2020, 86; 2021, 757] from this Monthly.
Also solved by T. Amdeberhan & V. H. Moll, K. F. Andersen (Canada), A. Berkane (Algeria), N. Bhandari
(Nepal), P. Bracken, J. V. Casteren & L. Kempeneers (Belgium), H. Chen (China), H. Chen, A. Dixit (India),
G. Fera (Italy), K. Gatesman, M. L. Glasser, H. Grandmontagne (France), N. Grivaux (France), J. A. Grzesik
(Canada), E. A. Herman, N. Hodges (UK), F. Holland (Ireland), W. Janous (Austria), J. E. Kampmeyer III,
O. Kouba (Syria), O. P. Lossers (Netherlands), J. Magliano, K. D. McLenithan & S. C. Mortenson, A. Natian,
M. Omarjee (France), D. Pinchon (France), A. Stadler (Switzerland), A. Stenger, S. M. Stewart (Saudi Arabia),
M. S̆tofka (Slovakia), R. Stong, M. Vowe (Switzerland), T. Wiandt, H. Widmer (Switzerland), J. Yan (China),
L. Zhou, Fejéntaláltuka Szeged Problem Solving Group (Hungary), Missouri Problem Solving Group, UM6P
Math Club (Morocco), and the proposer.

CLASSICS
C15. Suggested by Joel Spencer, New York University, New York, NY. A construction chain
for n is a sequence a1 , . . . , ak where a1 = 1, ak = n, and each entry in the sequence
is either the sum or the product of two previous, possibly identical, elements from the
sequence. The cost of a construction chain is the number of entries that are the sum (but
not the product) of preceding entries. For example, 1, 2, 3, 6, 12, 144, 1728, 1729 is a con-
struction chain for 1729; its cost is 3, because the elements 2, 3, and 1729 require addition.
Let c(n) be the minimal cost of a construction chain for n. Prove that c is unbounded.

Coprimality in Pascal’s Triangle

C14. Due to Paul Erdős and George Szekeres; suggested by the editors. Show that no two
entries chosen from the interior of any row of Pascal’s triangle are relatively prime.
Solution. Suppose 0 < a < b < n. The identity
n n−a n b
= (∗)
a b−a b a
is easily verified (both sides count committees of size b with
 a subcommittee of size a
chosen from a set of n people). It follows that if na and nb are relatively prime, then na
  
divides ab . This contradicts ab < na .
Editorial Comment. The result is from Paul Erdős and George Szekeres (1978), Some num-
ber theoretic problems on binomial coefficients, Aust. Math. Soc. Gazette 597–99 (available
on-line at combinatorica.hu/∼p erdos/1978-46.pdf).
    There the following stronger result is
proved: If 0 < a < b ≤ n/2 and d = gcd na , nb , then d ≥ 2a . To see this, note that (∗)
    
implies na /d divides ab , which in turn implies d ≥ na ab . Since this last expression is
equal to
n n − 1 n−a+1
··· ,
b b−1 b−a+1
and since each of these factors is at least 2, we have d ≥ 2a . This inequality is strict when
a > 1.

494 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
SOLUTIONS

The Laplace Transform Simplifies an Integral

12260 [2021, 563]. Proposed by Seán M. Stewart, Bomaderry, Australia. Prove
 ∞
sin2 x − x sin x 1
3
dx = − log 2.
0 x 2

Solution by Tewodoros Amdeberham, Tulane University, New Orleans, LA, and Akalu
Tefera, Grand Valley State University, Allendale, MI. The Laplace transform L defined
∞
by L[f ](s) = 0 f (t)e−st dt has the property
 ∞  ∞
f (x)g(x) dx = L[f ](s) · L−1 [g](s) ds.
0 0

Applying this with f (x) = sin x − x sin x = 1/2 − (1/2) cos(2x) − x sin x and g(x) =
2

1/x 3 leads to
 ∞  ∞
sin2 x − x sin x 1 1 1
3
dx = L − cos(2x) − x sin x (s) · L−1 3 (s) ds
0 x 0 2 2 x
 ∞  2
1 1 s 2s s
= − 2+4
− 2 · ds
0 2s 2 s (s + 1)2 2
 ∞
s s s
= − + ds
0 s 2 + 4 s 2 + 1 (s 2 + 1)2
∞
log(s 2 + 4) − log(s 2 + 1) 1 1
= − = − log 2.
2 2(s + 1)
2
0 2

Also solved by U. Abel & V. Kushnirevych (Germany), K. F. Andersen (Canada), M. Bataille (France),
A. Berkane (Algeria), G. E. Bilodeau, K. N. Boyadzhiev, P. Bracken, B. Bradie, A. C. Castrillón, H. Chen,
C. Degenkolb, A. De la Fuente, H. Y. Far, G. Fera (Italy), A. Garcia (France), M. L. Glasser, R. Gordon,
H. Grandmontagne (France), G. C. Greubel, N. Grivaux (France), P. Haggstrom (Australia), L. Han (US) &

386 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

X. Tan (China), D. Henderson, E. A. Herman, N. Hodges (UK), F. Holland (Ireland), W. Janous (Austria),
W. P. Johnson, A. M. Karparvar (Iran), O. Kouba (Syria), K.-W. Lau (China), O. P. Lossers (Netherlands),
J. Magliano, K. McLenithan, I. Mező (China), M. Omarjee (France), D. Pinchon (France), S. Sharma (India),
P. Shi (China), A. Stadler (Switzerland), J. L. Stitt, R. Stong, R. Tauraso (Italy), Y. Tsyban (Saudi Arabia),
J. Van Casteren & L. Kempeneers (Belgium), E. I. Verriest, M. Vowe (Switzerland), S. Wagon, T. Wiandt,
H. Widmer (Switzerland), M. Wildon (UK), L. Zhou, Fejéntaláltuka Szeged Problem Solving Group (Hun-
gary), UM6P Math Club (Morocco), and the proposer.

Counting Equilateral Triangles in Hypercubes

12261 [2021, 563]. Proposed by Albert Stadler, Herrliberg, Switzerland. Let an be the num-
ber of equilateral triangles whose vertices are chosen from the vertices of the n-dimensional
cube. Compute limn→∞ nan /8n .
Solution by √ Richard Stong, Center for Communications Research, San Diego, CA. The
limit is 1/(3 3 π ).
Let the n-dimensional hypercube have vertex set {0, 1}n . For vertices A, B, C chosen
from this set, let I be the set of coordinates where A differs from both B and C, let J be the
set of coordinates where B differs from both A and C, and let K be the set of coordinates
where C differs from both A and B. Since A − B 2 = |I | + |J |, B − C 2 = |J | + |K|,
and C − A 2 = |K| + |I |, the vertices in {A, B, C} form an equilateral triangle if and
only if |I | = |J | = |K|. Conversely, choose a vertex A and three disjoint sets of indices
I , J , K, each of positive size k. Define B to differ from A in coordinates I ∪ J and C to
differ from A in coordinates I ∪ K. The resulting triangle ABC is equilateral, and each
equilateral triangle arises in 3! ways. Thus,
n/3  
2n  n (3k)!
an = . (∗)
6 k=1 3k (k!)3
Stirling’s formula gives
√   
(3k)! 3 1
3
= · 33k
1 + O ,
(k!) 2π k k
which we can write equivalently as
√   
(3k)! 3 3 1
= ·33k
1+O .
(k!)3 2π(3k + 1) k
Since 3kn
≤ 2n and (3k)!/(k!)3 ≤ 33k , any term in the sum (∗) with k < n/6 contributes
√
less than 2n · 2n · 3n/2 to an . This value, which simplifies to (4 3)n , is o(8n ). Therefore,
in computing limn→∞ nan /8n , the sum of the estimates has relative error O(1/n). Also,
starting the sum at k = 0 has no impact on the limit. Thus
   √  n/3 
nan (n + 1)an 1 3  n+1 n  
1
= 1+O = n+1 33k
1+O
8 n 8 n n 4 π k=0
3k + 1 3k n
 n/3     
1  n+1 1
= √ 33k+1 1+O .
4 n+1 3π k=0
3k + 1 n
√
Letting ω = e2πi/3 and using |3ω + 1| = |3ω−1 + 1| = 7 < 4, it follows that
  
nan 1 (3 + 1)n+1 + ω−1 (3ω + 1)n+1 + ω(3ω−1 + 1)n+1 1
n
= √ · 1+O
8 4n+1 3π 3 n
  
1 1
= √ 1+O .
3 3π n

April 2023] PROBLEMS AND SOLUTIONS 387

 √
Therefore, the requested limit is 1/(3 3 π ).
Also solved by U. Abel & V. Kushnirevych (Germany), H. Chen (China), H. Chen (US), R. Dempsey, G. Fera
& G. Tescaro (Italy), N. Hodges (UK), M. Omarjee (France), D. Pinchon (France), R. Tauraso (Italy), L. Zhou,
and the proposer.

A Trigonometric Generating Function

12262 [2021, 563]. Proposed by Li Zhou, Polk State College, Winter Haven, FL. For a
nonnegative integer m, let
∞  
1 1
Am = − .
k=0
(6k + 1)2m+1 (6k + 5)2m+1
√
Prove A0 = π 3/6 and, for m ≥ 1,
√
 (−1)m (4m + 1) 3  π 2m+1
m
(−1)n π 2n
2Am + Am−n = .
n=1
(2n)! 2(2m)! 3

Solution by Omran Kouba, Higher Institute for Applied Science and Technology, Damas-
cus, Syria. The sequence (Am )m≥0 is bounded, so for x ∈ (−1, 1) we may define
∞ ∞  ∞  
x 2m x 2m
F (x) = Am x 2m = −
m=0 m=0 k=0
(6k + 1)2m+1 (6k + 5)2m+1
∞  ∞  
x 2m x 2m
= −
k=0 m=0
(6k + 1)2m+1 (6k + 5)2m+1
∞  
6k + 1 6k + 5
= − .
k=0
(6k + 1)2 − x 2 (6k + 5)2 − x 2

Setting α = (1 + x)/6 and β = (1 − x)/6, we have

6k + 1 6k + 5
−
(6k + 1)2 − x 2 (6k + 5)2 − x 2
 
1 1 1 1 1
= + − −
2 6k + 1 + x 6k + 1 − x 6k + 5 + x 6k + 5 − x
 
1 1 1 1 1
= + + + .
12 α + k β +k β −k−1 α−k−1
Next we use the partial fraction expansion of the cotangent, which is
∞  
1 1
π cot(π z) = + ,
k=0
z+k z−k−1

when z is not an integer. Applying this with z = α and z = β gives

π π sin π(α + β)
F (x) = cot(π α) + cot(πβ) = ·
12 12 sin(π α) sin(πβ)
π sin(π(α + β)) π sin(π/3)
= · = ·
6 cos π(α − β) − cos π(α + β) 6 cos(π x/3) − cos(π/3)
√
π 3 1
= · .
6 2 cos(π x/3) − 1

388 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

 From (cos(2θ ) + cos θ )(2 cos θ − 1) = cos(3θ ) + 1, with θ = π x/3, we conclude
√     π x 
π 3 2π x
1 + cos(π x) F (x) = cos + cos ,
6 3 3
and hence
 ∞
 ∞ √ ∞
 (−1)n π 2n  π 3  (−1)m (4m + 1)π 2m 2m
2+ x 2n
An x 2n
= x .
n=1
(2n)! n=0
6 m=0 32m (2m)!
√
Comparing the coefficients of x 2m on both sides, we get A0 = π 3/6 and, for m ≥ 1,
√
 (−1)m (4m + 1) 3  π 2m+1
m
(−1)n π 2n
2Am + Am−n = ,
n=1
(2n)! 2(2m)! 3

as desired.
Editorial comment. Omran Kouba also noted that by using
 πx   √
π 3
2 cos − 1 F (x) = ,
3 6
we obtain the alternative recurrence
 2(−1)n−1  π 2n
m
Am = Am−n .
n=1
(2n)! 3

Also solved by K. F. Andersen (Canada), P. Bracken, H. Chen, G. Fera (Italy), M. L. Glasser, G. C. Greubel,
E. A. Herman, N. Hodges (UK), O. P. Lossers (Netherlands), K. Nelson, A. Stadler (Switzerland), M. S̆tofka
(Slovakia), R. Tauraso (Italy), and the proposer.

A Concurrency from A Conic Inscribed in A Triangle

12263 [2021, 564]. Proposed by Dong Luu, Hanoi National University of Education,
Hanoi, Vietnam. In triangle ABC, let D, E, and F be the points at which the incircle
of ABC touches the sides BC, CA, and AB, respectively. Let D  , E  , and F  be three
other points on the incircle with E  and F  on the minor arc EF and D  on the major arc
EF and such that AD  , BE  , and CF  are concurrent. Let X, Y , and Z be the intersections
of lines EF and E  F  , lines F D and F  D  , and lines DE and D  E  , respectively. Prove
that AX, BY , and CZ are either concurrent or parallel.
Solution by O. P. Lossers, Eindhoven University of Technology, Eindhoven, Netherlands.
It is well known that AD, BE, and CF intersect at a point G, the Gergonne point of
ABC. We choose homogeneous coordinates such that A = (1 : 0 : 0), B = (0 : 1 : 0),
C = (0 : 0 : 1), and G = (1 : 1 : 1). It follows that D = (0 : 1 : 1), E = (1 : 0 : 1), and
F = (1 : 1 : 0), and the equation of the incircle is x 2 + y 2 + z2 − 2xy − 2xz − 2yz = 0.
Since the point of intersection of the lines AD  , BE  , and CF  lies in the interior of
ABC, we can take its coordinates to be (a 2 : b2 : c2 ), with a, b, c > 0. This gives D  =
(x : b2 : c2 ) for some x satisfying the quadratic equation

x 2 + b4 + c4 − 2xb2 − 2xc2 − 2b2 c2 = 0.

Of its two solutions x = (b − c)2 and x = (b + c)2 , we must choose x = (b − c)2 for D 
to be on the major arc EF . Note that since D = D  , we have b = c. In the same way we

April 2023] PROBLEMS AND SOLUTIONS 389

find E  = (a 2 : (c − a)2 : c2 ) and F  = (a 2 : b2 : (a − b)2 ), and a, b, and c are distinct. A
somewhat tedious but elementary computation gives

X = (a(c − b) : b(c − a) : c(a − b)),

Y = (a(b − c) : b(a − c) : c(a − b)),
Z = (a(b − c) : b(c − a) : c(b − a)),

so the lines AX, BY , and CZ intersect at the point (a(b − c) : b(c − a) : c(a − b)).
Editorial comment. Lossers observed that the solution above works if the incircle is
replaced with any ellipse tangent to the sides of the triangle. Li Zhou generalized the prob-
lem further by showing that the result holds for any conic tangent to the lines containing
the sides of the triangle, with suitable adjustments to the restrictions on the positions of
D  , E  , and F  .
Also solved by L. Zhou and the proposer.

Irreducible Polynomials in Two Variables

12264 [2021, 564]. Proposed by Navid Safaei, Sharif  University of Technology, Tehran,
Iran. Let Pd be the set of all polynomials of the form 0≤i,j ≤d ai,j x i y j with ai,j ∈ {1, −1}
for all i and j . Prove that there is a positive integer d such that more than 99 percent of the
elements of Pd are irreducible in the ring of polynomials with integer coefficients.
Solution by Richard Stong, Center for Communications Research, San Diego, CA. The
number 2 is a primitive root modulo the prime p when the smallest value of m such that p
divides 2m − 1 is p − 1. Hence the field F2p−1 is the extension of F2 of lowest degree that
contains a primitive pth root of unity modulo 2. It follows that the minimal polynomial
of any primitive pth root of unity modulo 2 has degree at least p − 1. Since the primitive
pth roots of unity are the roots of the polynomial (x p − 1)/(x − 1) (which equals x p−1 +
· · · + x + 1 and has degree p − 1) it follows that this polynomial is irreducible modulo
2. Thus all polynomials of the form a0 + a1 x + · · · + ap−1 x p−1 with all ai ∈ {−1, 1} (or
indeed with all ai odd) are irreducible over Z.
If 0≤i,j ≤p−1 ai,j x i y j ∈ Pp−1 is reducible, say as F (x, y)G(x, y), then

F (x, 0)G(x, 0) = a0,0 + a1,0 x + · · · + ap−1,0 x p−1 .

Since this polynomial in x is irreducible, F (x, 0) or G(x, 0) (we may assume F (x, 0))
has degree p − 1 as a polynomial in x. Looking at the term with highest degree in x in
F (x, y)G(x, y), we conclude that G(x, y) is a constant polynomial in x, and hence we can
write G(x, y) as G(y). Swapping the roles of x and y, we find symmetrically that (since
G(y) cannot be constant), G(y) has degree p − 1 and F (x, y) is constant in y, so we
write it as F (x). Thus all reducible polynomials in Pp−1 have the form F (x)G(y). Since
F (0)G(0) = ±1, we conclude F (0), G(0) ∈ {−1, 1}, Looking at the terms with degree 0
in x and y yields that all coefficients of F (x) are in {1, −1}.
Finally, there are 2p choices for each of F and G, but this double counts the product F G
as the product (−F )(−G). Thus there are exactly 22p−1 reducible polynomials in Pp−1 .
In particular, taking p = 5 and noting that 2 is a primitive root modulo 5, we see that
only 29 of the 225 elements of P4 are reducible, which is less than 1% of the total number
of polynomials in P4 . The fraction only decreases as p increases.
Also solved by S. M. Gagola Jr., O. P. Lossers (Netherlands), D. Pinchon (France), and the proposer.

390 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

 Combining the Cauchy–Schwarz and AM–GM Inequalities
12267 [2021, 658]. Proposed by Michel Bataille, Rouen, France. Let x, y, and z be non-
negative real numbers such that x + y + z = 1. Prove
   √
(1−x) x(1−y)(1−z) + (1−y) y(1−z)(1−x) + (1−z) z(1−x)(1−y) ≥ 4 xyz.

Solution by Tamas Wiandt, Rochester Institute of Technology, Rochester, NY. It is clear that
the required inequality holds if any of x, y, or z is zero; it is an equality if two of them are
√
zero. Now suppose that x, y, and z are all positive. Dividing by xyz and using the fact
that x + y + z = 1, we see that the inequality is equivalent to
√ √ √
(y + z) (x + z)(x + y) (x + z) (x + y)(y + z) (x + y) (y + z)(x + z)
√ + √ + √ ≥ 4.
yz xz xy
√ √
The Cauchy–Schwarz inequality
√ gives (x + z)(x + y) ≥ x + yz, and by the AM–
GM inequality, y + z ≥ 2 yz. Applying these, we obtain
√ √
(y + z) (x + z)(x + y) (y + z)(x + yz) (y + z)x
√ ≥ √ = √ + y + z ≥ 2x + y + z = x + 1.
yz yz yz

Combining this with similar inequalities for the other two terms, we get
√ √ √
(y + z) (x + z)(x + y) (x + z) (x + y)(y + z) (x + y) (y + z)(x + z)
√ + √ + √
yz xz xy
≥ (x + 1) + (y + 1) + (z + 1) = 4,

as required. When x, y, and z are positive, equality holds only if x = y = z = 1/3.

Also solved by A. Alt, F. R. Ataev (Uzbekistan), A. Berkane (Algeria), P. Bracken, H. Chen (China), H. Chen,
C. Chiser (Romania), N. S. Dasireddy (India), M. Dinc̆a (Romania), H. Y. Far, G. Fera (Italy), A. Garcia
(France), O. Geupel (Germany), P. Haggstrom (Australia), D. Henderson, N. Hodges (UK), F. Holland (Ire-
land), E. J. Ionaşcu, W. Janous (Austria), A. M. Karparvar (Iran), P. Khalili, K. T. L. Koo (Hong Kong),
O. Kouba (Syria), K.-W. Lau (Hong Kong), S. Lee (Korea), O. P. Lossers (Netherlands), J. F. Loverde,
A. Mhanna (Lebanon), M. Reid, V. Schindler (Germany), A. Stadler (Switzerland), R. Stong, R. Tauraso
(Italy), M. Tetiva (Romania), J. F. Gonzalez & F. A. Velandia (Colombia), M. Vowe (Switzerland), J. Vuk-
mirović (Serbia), H. Widmer (Switzerland), L. Wimmer (Germany), L. Zhou, UM6P MathClub (Morocco),
and the proposer.

A Triangle Inscribed in a Similar Triangle

12269 [2021, 659]. Proposed by Mehmet Şahin and Ali Can Güllü, Ankara, Turkey. Let
ABC be an acute triangle. Suppose that D, E, and F are points on sides BC, CA, and
AB, respectively, such that F D is perpendicular to BC, DE is perpendicular to CA, and
EF is perpendicular to AB. Prove
AF BD CE
+ + = 1.
AB BC CA

Solution I by Michael Reid, University of Central Florida, Orlando, FL. For a polygon
P Q · · · Z, let (P Q · · · Z) denote its area. Let H be the orthocenter of ABC. Since the
triangle is acute, H lies in its interior. Both CH and EF are perpendicular to AB, so they
are parallel, and therefore (CEF ) = (H EF ). Thus

April 2023] PROBLEMS AND SOLUTIONS 391

 AF (AF C) (AF E) + (CEF ) (AF E) + (H EF ) (H EAF )
= = = = .
AB (ABC) (ABC) (ABC) (ABC)
Similarly, BD/BC = (H F BD)/(ABC) and CE/CA = (H DCE)/(ABC), so
AF BD CE (H EAF ) + (H F BD) + (H DCE) (ABC)
+ + = = = 1.
AB BC CA (ABC) (ABC)

Solution II by Li Zhou, Polk State College, Winter Haven, FL. By Miquel’s theorem, the
circumcircles of triangles AF E, BDF , and CED concur at a point, the Miquel point M.
Note that since ∠AF E is a right angle, AE is a diameter of the circumcircle of AF E,
and therefore ∠AME is also a right angle. Similarly, ∠BMF and ∠CMD are right angles.
Since ∠MF E and ∠MAE are subtended by the same arc of the circumcircle of AF E,
they are equal. Similarly, ∠MED = ∠MCD and ∠MDF = ∠MBF . Also, ∠MAE =
∠MED, since both are complementary to ∠MEA, and similarly ∠MCD = ∠MDF .
We conclude that all six of the angles ∠MF E, ∠MAE, ∠MED, ∠MCD, ∠MDF , and
∠MBF are equal. This means that M is a Brocard point of both ABC and DEF . Let
ω denote the measure of all six angles, which is the Brocard angle. It is well known that
cot ω = cot A + cot B + cot C.
Triangles MEF and MAB are similar, since corresponding sides are perpendicular.
Hence EF /AB = EM/AM, so
AF AF EF EM
= · = cot A · = cot A tan ω.
AB EF AB AM
Similarly, BD/BC = cot B tan ω and CE/CA = cot C tan ω, so
AF BD CE
+ + = (cot A + cot B + cot C) tan ω = cot ω tan ω = 1.
AB BC CA

Editorial comment. Several readers noted that the result can be extended to obtuse triangles
by allowing one of the points D, E, and F to lie on an extension of a side of ABC and
using signed distances.
It was not required to construct DEF , or even to show that such a triangle exists.
However, Solution II shows how to construct the unique such triangle. Let M be the Bro-
card point of ABC such that ∠MAC, ∠MBA, and ∠MCB all have the same measure
ω. Triangle DEF is the image of triangle CAB under a rotation of π/2 radians about M
followed by a dilation centered at M with ratio tan ω.
Also solved by M. Bataille (France), R. B. Campos (Spain), H. Chen (China), C. Chiser (Romania), M. Dincă,
G. Fera (Italy), D. Fleischman, K. Gatesman, O. Geupel (Germany), E. A. Herman, N. Hodges (UK),

392 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

E. J. Ionaşcu, Y. J. Ionin, W. Janous (Austria), W. Ji (China), M. Goldenberg & M. Kaplan, A. M. Karpar-
var (Iran), P. Khalili, O. Kouba (Syria), K.-W. Lau (China), J. H. Lindsey II, O. P. Lossers (Netherlands),
J. McHugh, M. D. Meyerson, J. Minkus, M. R. Modak (India), C. G. Petalas (Greece), C. R. Prane-
sachar (India), I. Retamoso, V. Schindler (Germany), A. Stadler (Switzerland), R. Stong, R. Tauraso
(Italy), M. Vowe (Switzerland), J. Vukmirović (Serbia), T. Wiandt, H. Widmer (Switzerland), L. Wimmer
(Germany), T. Zvonaru (Romania), Davis Problem Solving Group, Fejéntaláltuka Szeged Problem Solving
Group (Hungary), UM6P Math Club (Morocco), and the proposer.

A Refinement of a Putnam Problem

12270 [2021, 659]. Proposed by Moubinool Omarjee, Lycée Henri IV, Paris, France. Let
a0 = 1, and let an+1 = an + e−an for n ≥ 0. Show that the sequence whose nth term is
ean − n − (1/2) ln n converges.
Solution by Kuldeep Sarma, Tezpur University, Tezpur, India. Define un = ean , and note
that un+1 = un e1/un . Since the sequence {un } is positive and strictly increasing, it must
either converge to a positive limit or diverge to +∞. If the sequence converges to L,
then the recurrence relation gives L = Le1/L , which is impossible; therefore limn→∞ un =
+∞.
Note that limn→∞ (un+1 − un ) = limn→∞ un (e1/un − 1) = 1. Therefore, by the Stolz–
Cesàro theorem, limn→∞ un /n = 1. It follows that
un+1 − un − 1 u2 (e1/un − 1 − 1/un ) 1/2 1
lim = lim n = = .
n→∞ 1/n n→∞ un /n 1 2
By the Stolz–Cesàro theorem again,
un − n (un+1 − (n + 1)) − (un − n)
lim = lim
n→∞ ln n n→∞ ln(n + 1) − ln n
un+1 − un − 1 1/n 1 1
= lim · = ·1= .
n→∞ 1/n ln(1 + 1/n) 2 2
Combining the recurrence relation for un with the Maclaurin series for the exponential
function, for n ≥ 1 we have
   
1 1 1 un − n 1
un+1 = un + 1 + +O 2
= un + 1 + − +O .
2un un 2n 2nun u2n
From previous observations, we know that
un − n ln n 1 1
∼ 2 and 2
∼ 2,
2nun 4n un n
so
 
1 ln n
un+1 = un + 1 + +O .
2n n2
 N−1
Since ∞ 2
n=1 ln n/n converges, we conclude that n=1 (un+1 − un − 1 − 1/(2n)) con-
verges as N → ∞. For N ≥ 2,

N−1
1

HN−1
un+1 − un − 1 − = uN − u1 − (N − 1) − ,
n=1
2n 2

where we write Hk for the kth harmonic number ki=1 1/ i. Therefore

1 
N−1
1

1 1
e − N − ln N =
aN
un+1 − un − 1 − + u1 − 1 − + (HN − ln N ).
2 n=1
2n 2N 2

April 2023] PROBLEMS AND SOLUTIONS 393

The desired result follows, since HN − ln N → γ as N → ∞.
Editorial comment. Several solvers noted similarities between this problem and Monthly
Problem 11837 [2015, 391; 2017, 91], which asks for a proof that the sequence {an − ln n}
decreases monotonically to 0. The earlier Monthly problem is a refinement of Problem
B4 of the 73rd William Lowell Putnam Mathematical Competition, which simply asks
whether {an − ln n} has a finite limit. Indeed, since an − ln n = ln(un /n), it follows from
the above solution that limn→∞ (an − ln n) = 0. This solves the Putnam problem and part
of the earlier Monthly problem.
Also solved by M. Bataille (France), A. Berkane (Algeria), P. Bracken, H. Chen, N. Grivaux (France), X. Tang
(China) & L. Han (US), E. A. Herman, N. Hodges (UK), E. J. Ionaşcu, O. Kouba (Syria), K.-W. Lau (China),
J. H. Lindsey II, O. P. Lossers (Netherlands), S. Omar (Morocco), E. Omey (Belgium), A. Stadler (Switzer-
land), A. Stenger, R. Stong, R. Tauraso (Italy), J. Vukmirović (Serbia), J. Yan (China), UM6P Math Club
(Morocco), and the proposer.

CLASSICS
C14. Due to Paul Erdős and George Szekeres; suggested by the editors. Show that no two
entries chosen from the interior of any row of Pascal’s triangle are relatively prime.

Visiting Every Region on a Sphere Exactly Once

C13. Due to Leo Moser; suggested by the editors. Let n be a multiple of 4, and consider an
arrangement of n great circles on the sphere, no three concurrent, dividing the sphere into
regions. Show that there is no path on the sphere that visits each region once and only once
and never passes through an intersection point of two of the great circles.
Solution. The great circles define a graph G: the vertices are the intersection points of the
circles, and the edges are the arcs of the circles joining vertices. Let H be the graph of the
corresponding map: the vertices are the regions of G, and edges connect adjacent regions
across an edge of G. Because any two great circles intersect twice, G has n(n − 1) vertices.
Because every vertex of G has four neighbors, G has 2n(n − 1) edges. By Euler’s formula
V − E + F = 2 relating the numbers of vertices, edges, and faces of a connected graph on
the sphere, G has n(n − 1) + 2 faces. This is the number of vertices of H and is even.
Since every edge in H crosses a great circle, and every cycle in H must cross each great
circle an even number of times to return to the original region, every cycle in H has even
length. Hence H is bipartite, meaning that we can color each vertex of H red or blue in
such a way that all edges connect a red vertex and a blue vertex.
The regions of G containing diametrically opposite points on the sphere lie on opposite
sides of every great circle. Hence every path joining the vertices for these points crosses
every great circle an odd number of times. Since n is even, this implies that such a path has
even length, so the vertices representing antipodal regions are colored the same. It follows
that H has an even number of vertices of each color.
If H has a path that visits each vertex, then H must have the same number of vertices
of each color. Since the two color classes have the same even size, the number of vertices
in H is a multiple of 4. However, that number is n(n − 1) + 2, which is not divisible by 4.
Editorial comment. This problem appeared in this Monthly as problem E788 [1947, 471;
1948, 366] and is due to Leo Moser. There is an essentially unique arrangement of n great
circle arcs on a sphere when n ≤ 5, and for n ∈ {2, 3, 5} each of these arrangements does
permit a Hamiltonian path, in fact a Hamiltonian circuit. When n = 6, some arrangements
permit Hamiltonian paths and some do not.

394 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

SOLUTIONS

Two Zeta Sums that Sum to Zeta of Two

12246 [2021, 376]. Proposed by Seán Stewart, Bomaderry,
∞ Australia. Let ζ be the Rie-
mann zeta function, defined
 for n ≥ 2 by ζ (n) = k=1 1/k n
. Let Hn be the nth harmonic
number, defined by Hn = nk=1 1/k. Prove
∞
 ∞
ζ (n) ζ (n)Hn π2
+ (−1)n = .
n=2
n2 n=2
n 6

Composite solution by Khristo N. Boyadzhiev, Ohio Northern University, Ada, OH, and
Stephen Kaczkowski, South Carolina Governor’s School for Science and Mathematics,
2
Hartsville, SC. The factor 1/n in the first sum suggests relevance of the dilogarithm func-
tion Li2 , defined by Li2 (x) = ∞ n=1 x /n . Henceforth let L(x) = Li2 (x). It is well known
n 2

that L(1) = π /6.

2

For |x| ≤ 1/2, let

x 1 2
M(x) = L(x) + L + ln(1 − x) .
x−1 2
From the power series expansions of ln(1 − x), we find that M  (x) = 0 = M(0) whenever
|x| < 1/2. Thus we have the functional equation M(x) = 0, known as Landen’s identity.
For x = −1/k with k ≥ 2, this becomes
1 1 1 1
L − +L + ln2 1 + = 0. (∗)
k k+1 2 k
For the first sum in the desired statement, we obtain
∞ ∞ ∞  ∞ ∞
1  1  1 1 n  1 1
2
= = L − .
n=2
n k=1 k n
k=1 n=2
n2 k k=1
k k

Here the interchange of summations is valid since every summand is positive. Note that
the subtraction of 1/k is essential for the convergence.
If the second sum in the statement is the similarly convergent sum
∞
1 1
−L ,
k=1
k k + 1

then the result follows, since the combined sum over k telescopes to L(1).
From the power series of − ln(1 − x) and 1/(1 − x), we have
∞
− ln(1 − x) 
= Hn x n .
1−x n=1

286 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

Integration then yields
∞
 1 x n+1 1
Hn+1 − = ln2 (1 − x),
n=1
n+1 n+1 2
which we rewrite as
∞
 Hn x n 1 2
= ln (1 − x) + L(x) − x.
n=2
n 2
Pending justification of the interchange of summations, we compute
∞ ∞ ∞ ∞ 
∞
ζ (n)Hn Hn  1 n  Hn 1 n
(−1)n = − = −
n=2
n n=2
n k=1 k k=1 n=2
n k

∞ ∞
1 2 1 1 1 1 1
= ln 1 + +L − + = −L .
k=1
2 k k k k=1
k k + 1
Here the last step uses the functional equation (∗) for L.
It remains to justify the interchange of summations. The double summation with the
inner sum over k may be written as
∞
 ∞ 
 ∞
Hn Hn 1 n
(−1)n + − .
n=2
n n=2 k=2
n k
Since Hn /n is decreasing, the first sum converges. Next,
∞  ∞ ∞  ∞ ∞
Hn 1 n 1 n 1 2 k
< = < ∞.
n=2 k=2
n k k=2 n=2
k k=2
k k−1
Thus the double summation is absolutely convergent. It follows that the interchange is
valid, which completes the proof.
Editorial comment. Many solvers (including the proposer) relied on some version of the
known identity
∞
 ζ (n)
ln (1 − x) = γ x + xn,
n=2
n
where γ is Euler’s constant. The proposer also showed that the two sums are, respectively,
−γ + J and π 2 /6 + γ − J , where
 1
ln (1 − x)
J = dx.
0 x
T. Apostol famously proved ζ (2) = π 2 /6 by making a change of variable in a double
integral for ζ (2). Solvers Hervé Grandmontagne and Richard Stong, independently,
showed that each of the two sums here summing to π 2 /6 has a usable representation
as a double integral. Grandmontagne used well-known integrals for ζ (n), Hn , and 1/n2 to
write the two sums as
 ∞  1  1 n−1
y ln(1 − y)(ln x)n−1
dx dy
n=2 0 0 (1 − x)(n − 1)!
and
∞ 
 1  1
(−y)n−1 ln(1/y)(ln x)n−1
dx dy.
n=2 0 0 (1 − x)(n − 1)!

March 2023] PROBLEMS AND SOLUTIONS 287

After interchanging summation and integration, some simplification leads to
 1  1 1−y  1  1 −y
(x − 1) ln y (x − 1) ln(1/y)
dx dy + dx dy
0 0 1−x 0 0 1−x
 1 1
= x −y ln(1/y) dx dy.
0 0
The two double integrals on the left are the two sums in the posed problem, and the double
integral on the right equals ζ (2). However, while the interchange of summation and inte-
gration needed to complete this proof can be justified, it does require a fair amount of work,
especially for the first double integral.
Also solved by F. R. Ataev (Uzbekistan), A. Berkane (Algeria), P. Bracken, B. Bradie, B. S. Burdick, H. Chen
(US), G. Fera (Italy), M. L. Glasser, R. Gordon, H. Grandmontagne (France), G. C. Greubel, A. M. Karparvar
(Iran), O. Kouba (Syria), Z. Lin (China), C. Sanford, K. Sarma (India), A. Stadler (Switzerland), R. Stong,
R. Tauraso (Italy), M. Wildon (UK), and the proposer.

An Angle Bisector That Bisects a Segment

12253 [2021, 467]. Proposed by Alexandru Gı̂rban, Constanţa, Romania, and Bogdan
D. Suceavă, Fullerton, CA. Let ABC be a triangle, and let D and E be the contact points of
the incircle of ABC with the segments BC and CA, respectively. Let M be the intersection
of the line DE and the line through A parallel to BC. Prove that the bisector of ∠ABC
passes through the midpoint of DM.
Solution by Haoran Chen, Suzhou, China. Let F be the tangency point of the incircle with
AB, and let N be the intersection of the bisector of ∠ABC with AM. By three applications
of the tangent segment theorem, AE = AF , BF = BD, and CD = CE. Since AM is
parallel to BC, CDE and AME are similar, and therefore AM = AE. Also, ∠AN B =
∠CBN = ∠ABN , so ABN is isosceles and AN = AB > AF = AE = AM. Thus M
is between A and N , and MN = AN − AM = AB − AF = BF = BD. It follows that
BN and DM intersect at a point P such that P BD and P N M are congruent, and
hence P D = P M.
Also solved by M. Bataille (France), J. Cade, C. Chiser (Romania), P. De (India), C. de la Losa (France), I. Dim-
itrić, M. Dobrescu, G. Fera (Italy), D. Fleischman, K. Gatesman, O. Geupel (Germany), J.-P. Grivaux (France),
E. A. Herman, N. Hodges (UK), W. Janous (Austria), M. Getz & D. Jones, A. M. Karparvar (Iran), K. T. L. Koo
(China), O. Kouba (Syria), K.-W. Lau (China), O. P. Lossers (Netherlands), E. Mika & I. Adams & L. Loprieno
& R. McMullen & D. Schmitz, J. Minkus, D. Pinchon (France), C. R. Pranesachar (India), V. Schindler (Ger-
many), A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), M. Vowe (Switzerland), T. Wiandt, L. Zhou,
T. Zvonaru (Romania), Davis Problem Solving Group, and the proposer.

Sum of Squares Modulo 6

12255 [2021, 467]. Proposed by Besfort Shala, student, University of Primorska, Koper,
Slovenia. Given a positive integer a0 , define a1 , . . . , an recursively by ai = 12 + 22 + · · · +
2
ai−1 for i ≥ 1. Is it true that, given any subset A of {1, . . . , n}, there is a positive integer a0
such that, for 1 ≤ i ≤ n, 6 divides ai if and only if i ∈ A?
Solution by Nigel Hodges, Cheltenham, UK. The answer is yes. We prove the following
more general result: Given a list b1 , . . . , bn of integers, there is a positive integer a0 such
that ai ≡ bi (mod 6) for 1 ≤ i ≤ n. We may assume that each bi lies in {0, 1, . . . , 5}.
Since ai = ai−1 (ai−1 + 1)(2ai−1 + 1)/6 for i ≥ 1, it is reasonable to extend the definition
by letting the sequence be identically 0 when a0 = 0. The identity
(ai−1 + 6r )(ai−1 + 6r + 1)(2ai−1 + 2 · 6r + 1) ai−1 (ai−1 + 1)(2ai−1 + 1)
−
6 6
= 2 · 63r−1 + 62r−1 (6ai−1 + 3) + 6r−1 (6ai−1
2
+ 6ai−1 + 1)

288 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

describes the change in ai when ai−1 increases by 6r . Modulo 6r , the change is 6r−1 . This
allows the following inductive algorithm to find a0 to satisfy the given conditions.
Start with a0 = 0, so a1 = 0 as well. Add 6 to a0 exactly b1 times, adding 6b1 overall.
Since 60 = 1, applying the identity with r = 1 yields a1 ≡ b1 (mod 6).
Recalculate a2 from the revised a0 . Choose δ2 nonnegative so that δ2 ≡ b2 − a2
(mod 6), and then add 62 · δ2 to a0 . This increases a1 by a multiple of 6, so still a1 ≡ b1
(mod 6). Also, a2 increases by δ2 modulo 6, so a2 ≡ b2 (mod 6).
Continue in this manner. At stage j , recalculate a1 , . . . , aj from the revised a0 . Choose
δj nonnegative so that δj ≡ bj − aj (mod 6), and add 6j · δj to a0 . This increases each
of a1 , . . . , aj −1 by a multiple of 6, so still ai ≡ bi (mod 6) for 1 ≤ i ≤ j − 1. Also, aj
increases by δj , so aj ≡ bj (mod 6).
Repeat this process until j = n. If the resulting value of a0 is still 0, set a0 = 6n+1 to
make it positive, as required. This does not affect any of a1 , . . . , an modulo 6, so each
required congruence is still satisfied, finishing the proof.
Also solved by Y. J. Ionin, O. P. Lossers (Netherlands), D. Pinchon (France), M. A. Prasad (India), K. Sarma
(India), R. Stong, R. Tauraso (Italy), and the proposer.

An Integral Formula for Apéry’s Constant

12256 [2021, 468]. Proposed by Paul Bracken, University of Texas, Edinburg, TX. Prove
 1
log(1 + x) log(1 − x) 5
dx = − ζ (3),
0 x 8
∞
where ζ (3) is Apéry’s constant n=1 1/n3 .
Solution by Giuseppe Fera, Vicenza, Italy. With A = log(1 − x) and B = log(1 + x), the
algebraic identity AB = (1/4)((A + B)2 − (A − B)2 ) yields
 1  1  1 
log(1 + x) log(1 − x) 1 log2 (1 − x 2 ) 1 2 1−x
dx = dx − log dx .
0 x 4 0 x 0 x 1+x

To evaluate the first integral on the right side, we use the substitution y = 1 − x 2 , obtaining
 1   ∞
log2 (1 − x 2 ) 1 1 log2 (y) 1 1 2
dx = dy = log (y) y n−1 dy
0 x 2 0 1−y 2 0 n=1
∞  ∞
1 1  1
= y n−1 log2 y dy = = ζ (3),
2 n=1 0 n=1
n3

where the last integral is computed using integration by parts twice. Similarly, the substi-
tution y = (1 − x)/(1 + x) in the second integral yields
 1  1  1 ∞
1 1−x log2 (y)
log2 dx = 2 dy = 2 log 2
(y) y 2(n−1) dy
0 x 1+x 0 1−y
2
0 n=1
∞ 
 1 ∞
 1
=2 y 2n−2 log2 y dy = 4
n=1 0 n=1
(2n − 1)3
∞ ∞

 1  1 1 7
=4 3
− = 4 ζ (3) − ζ (3) = ζ (3).
n=1
n n=1
(2n)3 8 2

March 2023] PROBLEMS AND SOLUTIONS 289

Thus
 1
log(1 + x) log(1 − x) 1 7 5
dx = ζ (3) − ζ (3) = − ζ (3).
0 x 4 2 8

Editorial comment. Several solvers pointed out that this integral appears in C. I. Vălean
(2019), (Almost) Impossible Integrals, Sums, and Series, Cham, Switzerland: Springer.
This integral played a role in some submitted solutions to problem 12206 [2020, 722;
2022, 492] from this Monthly.
Also solved by T. Amdeberhan & A. Tefera, F. R. Ataev (Uzbekistan), M. Bataille (France), A. Berkane (Alge-
ria), N. Bhandari (Nepal), B. Bradie, V. Brunetti & D. B. Malesani & A. Aurigemma (Denmark), H. Chen,
N. S. Dasireddy (India), B. E. Davis, J. Fu (China), A. Garcia (France), S. Gayen (India), M. L. Glasser, R. Gor-
don, H. Grandmontagne (France), G. C. Greubel, J.-P. Grivaux (France), R. Guadalupe (Philippines), L. Han
(US) & X. Tang (China), D. Henderson, E. A. Herman, N. Hodges (UK), F. Holland (Ireland), W. Janous (Aus-
tria), A. M. Karparvar (Iran), O. Kouba (Syria), O. P. Lossers (Netherlands), R. Mortini (France) & R. Rupp
(Germany), M. Omarjee (France), D. Pinchon (France), M. A. Prasad (India), C. Sanford, K. Sarma (India),
V. Schindler (Germany), S. Sharma (India), A. Stadler (Switzerland), S. M. Stewart (Australia), R. Stong,
R. Tauraso (Italy), J. Van Casteren & L. Kempeneers (Belgium), M. Vowe (Switzerland), T. Wiandt, H. Wid-
mer (Switzerland), T. Wilde (UK), M. Wildon (UK), FAU Problem Solving Group, The Logic Coffee Circle
(Switzerland), UM6P Math Club (Morocco), Westchester Area Math Circle, and the proposer.

A Saturated Arrangement of Equilateral Triangles

12257 [2021, 468]. Proposed by Erich Friedman, Stetson University, DeLand, FL, and
James Tilley, Bedford Corners, NY. An arrangement of equilateral triangles in the plane
is called saturated if the intersection of any two is either empty or is a common vertex
and every vertex is shared by exactly two triangles. What is the smallest positive integer n
such that there exists a saturated arrangement of n equilateral triangles with integer length
sides?
Solution by the Davis Problem Solving Group, Davis, CA. The smallest such n is 10, with
an example given by Figure 1.
C 65 D
• ...........................................................................
....................................................................................................
•
......................................................................................................................... .........................
...................... ........................................................................ ............................
. . ....................................................... .................................................................................................................................... ............................................................................
91 . .
.......................... ...................................... ..............................
........................................................... ........................................................................ .........................................................................
91
..................................................... ........................................................................................ ..................................................................................................................................
........................................................................... .................................... ....................................................................................................
...................................................................................................................................... .........................................
.. .. .
. . . . . . . . . ............... .............................. .................................................................................................................
.
.......................................................................................................... ..................... ............................................................................................. B
................................................................................................................................................................................. ...........................................................................................................................................................................................................
A .. . . .
.....................................................................................................
......................................................................................................................................................................................................................... •Q
.. .
. .
.....
.
. . .
.. ..
.........................................................................................................
.
.......................................................................................................................................................................................
•
•
.... . ..................................
........................................................................................................................................................................................................................................................
...
..............
....................................
.................................................................................................................
......................................................................................................................................................................................................................
. .
...................................................................................................................................................................................................... ..................................................................... ..........................................................................................................................................................................................................................................
.................................................................... ......................................................... ..................................................... .......................................... ............................................................................................
....................................................................................................... ............................... ................................................................ ...............................................................................
...................................................................................................
.............................................................................................................................................. • • .................................................................................. ..................................................................................................................
...................... ................................................................................................
.........................................................................................................
................................................................................................................................................................... E 39 F ................................................................................................................
.
........................................................................................................................................................................................................................................................................ .
. .
...
..........................................................................................................................................................................................................................................................................................
............................................
P
............................................................................................................................................................................................................................................................................................................................................................................................................................................................................
112 •
..................................................................................................................................................................................................................... .......................................................................................................................................................
.............................................................................................................................................................. ...................................................................................................................................................................................................................................................................................... 128
............................................................................................................................................................................... ........................................................................................................................................................................
 ......................................................................................................................................................................................................................................................................
........ ..... .... .... ... ........
........................................................................................................................................................................................ 
.........................................................................................................
.................................................................................................................
E • 39 F • . . . . .
.................................
.................................................................................................... ......................................................................................................................................................................................
. . . . . .
......................................................................................... . . . .. . . .
. . ............................................ ......................................................................... .................................... .............................................................................................................................
. .... ............ ............ ......... .. .......................
............................................................................................................................................................................................................................... .......................................................................... .............................................................................................................................................................................................................................................
............................................................................................................................................ ....................... ................................................................................................................................................
................................................................................................................................................................................... ........................ ...................................................................................................................................................
•
..... ....................................................................................................................................................................................
•Q
.............
.... ......................... . ........
 ...................................................................................................................................................................................................................................................................
A
.....................................................................................................
...........................................................................................................................
....................................................................................... .
.. . .. ..
........................... .................................................................................................................................................................................
.................................................................................................................................... •
B
.................................................................................. .. ...................... .................................................................................................................................................
............................................................................................... .....................................
................................................................... ..........................................................................................................
............................................................................. ..................................................................................... ......................
.................................................... ............................................... ..................................................................................................................................
.......................................................... ....................................................................... .........................................................................
...................................... ............................................................ .............................................
91 ...................................... ....................................................................................... ..................................................
......................... ........................................................................ ..............................
.................... .....................................................................................................................................
91
.............................................................................................................
• .................................................................................................
....... •
C 65 D
Figure 1
First, we show that n ≥ 10 for a saturated arrangement of n equilateral triangles,
whether or not the sides have integer lengths. The total number of vertices is 3n/2, so
n must be even. Consider the simple polygon consisting of the edges of the triangles

290 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

bordering the unbounded region outside the arrangement. Because the triangles intersect
in at most a vertex, each interior angle of this polygon is greater than 120 degrees. Thus
the polygon has at least seven edges, corresponding to distinct boundary triangles in the
arrangement.
If no two boundary triangles share a vertex inside the polygon, then we have at least
seven interior vertices and hence at least three additional triangles. In this case n ≥ 7 + 3 =
10. If two boundary triangles have a common interior vertex, then they cannot be adjacent
on the polygon, so there must be an interior vertex on each side of their union. Hence there
must also be an interior triangle on each side of their union. Therefore, n ≥ 7 + 2 = 9 and
n is even, so n ≥ 10.
Returning to our example, we establish that such an arrangement does indeed exist.
We begin with two equilateral triangles as in Figure 2, where ∠AP B = 120◦ and a ≤ b.
Applying the law of cosines to ABP , we find AB 2 = a 2 + b2 + ab.
B C D
A. . . . . . . . . . . . . . . . .
. . . . . . ................. • x
•............................. ...................
.....................................
• •
.........................................................................................................
......................................................................................................................
............................ ...................................... ........................ ............................................................ ......................
.
.
.
.
.
.
..................................................
.
.
.
.
.
.
.
.
.
.
.
.
. . . . . . .. . . .. ............................................................................................................
. z . .. . .......................................................... ................................................................................................... ................................................................ z
.................................................. ................................................................ ......................... ............................ .........................
.......................................................................................................................... ........................................................................ ............................................................. ................................................. ............................................................
............................................ ........................................................................................................... ........................................................................ ....................................... ........................................................................
.................................................................................................................................................................... .......................................................................................... ............................................................... .......................................... ..................................................................................................................................
P ...................................................................................................................................................................................................................................... ......................................................................................................................................... ................ .
............................................................................................................................... . . . .
.. . .
..................................... . . . . . .
...........................................
. . . . . . . . . . ............ ......................................................................................................
........................................................................................................................................................................................................... ...................................................................................
a .............................................................................................................................................................................
•
................................................................................................................................................................................................... ................................................................................................................................................................................................................................
..................................................................................................................................
b A•
...................................................................................................................................................................................................
......................................................................... •Q .......
............................................................... ......................... .......................................................................................................................................
•B
............................................................................................................................................
......................................................................................................................................... .......................................................................................... ................................................... ..................... ....................................................
.................................................................................................................. .................................................................................. .......................................... ........................... ..........................................
............................................................................... ............................................................................................... ..................................... .......................................... ...................................
..........................................................................
.............................................................................
.......................................................
.................................................................
........................................................................
............................................. • • .................................................
.
....................................... ................................................
............................ E y F
• ..... .......................
..........
A • Figure 3
.
B
Figure 2
To obtain a saturated arrangement, we combine two copies of the configuration of four
equilateral triangles in Figure 3 (one upside down) with that in Figure 2 to obtain the satu-
rated configuration in Figure 1. Here ∠DQF = 120◦ and y ≤ x. There are three conditions
on the integers a, b, x, y, z that together are necessary and sufficient for the construction
to yield a saturated configuration. Applying the law of cosines to DF Q yields the first:
x 2 + y 2 + xy = z2 .
The second is that AB has the same length in both figures. To compute AB in Fig-
ure 3, observe that the quadrilateral BDQF has opposite angles summing to 180◦ , so these
four points lie on a circle. Angles BDF and BQF subtend the same arc of the circle, so
∠BQF = ∠BDF = 60◦ . Similarly ∠AQE = 60◦ , so A, Q, and B are collinear. Apply-
ing the law of sines to DF Q, we find
√ 
3y 3y 2 2x + y
sin ∠F DQ = , and so cos ∠F DQ = 1 − 2 = .
2z 4z 2z
Applying the law of sines and the addition formula for sines to ∠BDQ, we find that BQ
and AQ have length x + y and AB has length 2x + 2y. Therefore, the second condition is
a 2 + b2 + ab = 4(x + y)2 .
The third and final condition is that the triangles do not overlap when we combine
the pieces. Because a ≤ b, it follows that ∠ABP ≤ ∠BAP . Thus, the requirement
becomes ∠F BQ < ∠ABP , which, because both angles are acute, is equivalent √ to
sin
√ ∠F BQ
 < sin ∠ABP . Because BDQF is cyclic, also sin ∠F BQ = 3y/(2z) =
3/(2(x/y)2 + 1 + x/y ). Applying the law of sines to ABP yields sin ∠ABP =
√
3/(2 (1 + (b/a)2 + b/a ). We conclude that the third condition is equivalent to x/y >
b/a.
It is easy to check that the necessary and sufficient set of two equalities and one inequal-
ity holds when a = 112, b = 128, x = 65, y = 39, and z = 91.
Editorial comment. In the above solution, one can also prove that BQ = x + y geometri-
cally. Rotate DF Q by 60◦ counterclockwise about Q. The image of F is a point R on

March 2023] PROBLEMS AND SOLUTIONS 291

BQ with QR = y. The image of D is C and CR = z. Therefore, BDCR is a parallelo-
gram and BR = CD = x.
Solvers presented several other constructions, including some with seven boundary
triangles and three interior triangles.
Also solved by T. Fujita & S. Kim, S. M. Gagola Jr., O. P. Lossers (Netherlands), A. Martin & R. Martin
(Germany), R. Stong, R. Tauraso (Italy), L. Zhou, and the proposer.

Factorials That Are Not the Sum of Three Squares

12258 [2021, 563]. Proposed by Jeffrey C. Lagarias, University of Michigan, Ann Arbor,
MI. Let S be the set of positive integers n such that n! is not the sum of three squares.
Show that S has bounded gaps, i.e., there is a positive constant C such that for every
positive integer n, there is an element of S between n and n + C.
Solution by Michael Reid, University of Central Florida, Orlando, FL. We prove that the
difference between any two consecutive elements of S is at most 77.
Legendre proved that a positive integer is not a sum of three squares if and only if it
has the form 4c (8q + 7) for some nonnegative integers c and q. We claim that for every
nonnegative integer m, there is an integer t with 1 ≤ t ≤ 14 such that 64m + t ∈ S. Write
(64m)! uniquely as 2a (8q + r), where a is a nonnegative integer and r ∈ {1, 3, 5, 7}. When
r = 5 and a is odd, with a = 2b + 1, we take t = 3. This yields
(64m + 3)! = 22b+2 (8q + 5)(64m + 1)(32m + 1)(64m + 3) = 4b+1 (8k + 7)
for some positive integer k. Hence in this case (64m + 3)! is not a sum of three squares.
When r = 1 and a is odd, with a = 2b + 1, we take t = 5. This yields
(64m + 5)! = 22b+4 (8q + 1)(64m + 1)(32m + 1)(64m + 3)(16m + 1)(64m + 5)
= 4b+2 (8k + 7)
for some positive integer k. Hence in this case (64m + 5)! is not a sum of three squares.
Similar computations for the other six cases of the parity of a and the value of r yield
the following table of values of t such that 64m + t ∈ S.
r=1 r=3 r=5 r=7
a odd 5 14 3 2
a even 10 6 7 1
This establishes that consecutive elements of S differ by at most 64 + 13.
Editorial comment. Michael Reid and John Robertson independently showed that the dif-
ference between consecutive elements of S never exceeds 42. To prove this, one can con-
sider 15 consecutive values of n, having the form 64m + 16j + t with 1 ≤ t ≤ 15 for
fixed m and fixed j ∈ {1, 2, 3}. Like (64m)! as discussed above, one writes (64m + 16j )!
as 2a (8q + r) with eight cases for a and r. For each j in {1, 2, 3}, there is thus a table
like that above in whose cells are listed the values of t such that (64m + 16j + t)! can be
expressed in the form 4c (8k + 7). If all cells were nonempty, then consecutive members of
S would differ by at most 16 + 14.
In fact, there are two empty cells, for 64m + 32 with (64m + 32)! = 22b+1 (8q + 3)
and for 64m + 48 with (64m + 48)! = 22b (8q + 5). For these cases one must go farther
than t = 15. In the first case, 64m + 49 or 64m + 58 is in S, depending on the parity of
m. Since also 64m + 16 + t ∈ S for some t ≥ 5, these consecutive members of S differ
by at most 58 − 21, which equals 37. In the second case, 64m + 36 or 64m + 47 is in
S, depending on the parity of m. The chart above shows 64(m + 1) + t ∈ S for some t

292 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

with 1 ≤ t ≤ 14. Thus in this case 78 − 36 is an upper bound on the difference between
consecutive members of S, and hence in all cases the bound is at most 42.
Furthermore, differences of 42 occur infinitely often. A computer search shows that the
first such difference occurs for  the 2932nd and 2933rd elements in S, which are 23268
and 23310. Using (2n)! = 2n n! ni=1 (2i − 1), it is easy to show by induction that (9 · 2t )!
has the form 4b (8q + 1) for t ≥ 2. When t is sufficiently large and 1 ≤ j ≤ 23310, the
factors j and 9 · 2t + j are divisible by the same power of 2 and have odd parts that are
congruent modulo 8 (in fact, t ≥ log2 23310 + 3 = 17 is sufficient). More precisely,
for j in this range, j ! = 2b (8q + r) and (9 · 2t + j )! = 2B (8Q + R) with b ≡ B mod 2
and r ≡ R mod 8. Thus j ∈ S if and only if 9 · 2t + j ∈ S. Therefore, 9 · 2t + 23268 and
9 · 2t + 23310 are consecutive elements of S when t ≥ 17, differing by 42.
A proof that the density of S is 1/8 can be found in J.-M. Deshouillers and F. Luca, How
often is n! the sum of three squares?, K. Alladi, J. R. Klauder, and C. R. Rao, Eds. (2010),
The Legacy of Alladi Ramakrishnan in the Mathematical Sciences, Springer, 243–251.
Also solved by R. Dietmann (UK), A. Goel, N. Hodges (UK), O. P. Lossers (Netherlands), R. Martin (Ger-
many), J. P. Robertson, C. Schacht, A. Stadler (Switzerland), R. Stong, M. Tang, R. Tauraso (Italy), L. Zhou,
and the proposer.

Supplementary Pairs of Heronian Triangles

12259 [2021, 563]. Proposed by Giuseppe Fera, Vicenza, Italy. A triangle is Heronian if
it has integer sides and integer area. A pair of noncongruent Heronian triangles is called a
supplementary pair if the triangles have the same perimeter and the same area and some
interior angle of one is the supplement of some interior angle of the other. Prove that there
are infinitely many supplementary pairs of Heronian triangles.
Solution by Eagle Problem Solvers, Georgia Southern University, Statesboro, GA, and
Savannah, GA. We claim that for each integer n ≥ 2, the triangles with side lengths
(a1 , b1 , c1 ) and (a2 , b2 , c2 ) given by
(a1 , b1 , c1 ) = (n4 + n2 + 1, n6 + n4 + 2n2 + 1, n6 + 2n4 + n2 )
and
(a2 , b2 , c2 ) = (n4 + 2n2 + 1, n6 + n4 + n2 , n6 + 2n4 + n2 + 1)
form a supplementary pair of Heronian triangles. Note that ai < bi < ci for i ∈ {1, 2}.
Also, ai + bi > ci , so there is indeed a triangle for each triple. Since c2 = c1 + 1, the two
triangles are not congruent.
Since a1 + b1 + c1 = 2(n6 + 2n4 + 2n2 + 1) = a2 + b2 + c2 , the two triangles have the
same perimeter. Let s be the common semiperimeter;√note that s = (n2 + 1)(n4 + n2 + 1).
By Heron’s formula, the area of the ith triangle is s(s − ai )(s − bi )(s − ci ). Thus the
area of the first triangle is

(n2 + 1)(n4 + n2 + 1) · n2 (n4 + n2 + 1) · n4 · (n2 + 1),
and the area of the second triangle is

(n2 + 1)(n4 + n2 + 1) · n4 (n2 + 1) · (n4 + n2 + 1) · n2 .
Therefore, each triangle has area n3 (n2 + 1)(n4 + n2 + 1).
Finally, let B1 be the angle opposite the side of length b1 , and let C2 be the angle
opposite the side of length c2 . By the law of cosines, after some calculation, we find
a12 + c12 − b12 n2 − 1
cos B1 = = 2 ,
2a1 c1 n +1

March 2023] PROBLEMS AND SOLUTIONS 293

and
a22 + b22 − c22 n2 − 1
cos C2 = =− 2 = − cos B1 .
2a2 b2 n +1
Thus B1 and C2 are supplementary, and for n ≥ 2, we have a supplementary pair of Hero-
nian triangles. As n runs through the integers greater than 1, we obtain infinitely many
distinct values for cos B1 , so this method produces infinitely many such pairs.
Also solved by J. Keadey & J. Boltz & S. Kompella & S. Vemuru, P. Lalonde (Canada), C. R. Pranesachar
(India), R. Stong, R. Tauraso (Italy), M. Vowe (Switzerland), and the proposer.

CLASSICS
C13. Due to Leo Moser; suggested by the editors. Let n be a multiple of 4, and consider an
arrangement of n great circles on the sphere, no three concurrent, dividing the sphere into
regions. Show that there is no path on the sphere that visits each region once and only once
and never passes through an intersection point of two of the great circles.

The Unilluminable Room

C12. Due to Lionel Penrose and Roger Penrose; suggested by the editors. Is there a plane
region bounded by a differentiable Jordan curve with the property that no matter where a
light source is placed inside it, some part of the region remains unilluminated? Assume
that the curve acts as a perfect mirror.
Solution. An unilluminable region is shown below. It has a horizontal line of symmetry.
The arc AD is the upper half of the ellipse with foci B and C. The remaining portion
of the boundary curve may be constructed from circular arcs, although any differentiable
curve with the approximate shape of the diagram will suffice. Any light ray that starts below
the segment BC might visit the part of the region
above BC, but to do so it will have to pass through
BC and strike the elliptical arc AD. By a well-
known property of the ellipse, a light ray from B
that strikes the elliptical arc AD will reflect back
to C. It follows that a ray that passes through BC
and strikes the elliptical arc will be reflected back
between B and C, and therefore it cannot visit the
two shaded parts of the region. Similarly, a light
ray that starts above the reflection of BC across
the line of symmetry will never visit the reflections
of the shaded regions across the line of symmetry.
Editorial comment. The question was raised by E. G. Straus in the early 1950s and solved in
L. S. Penrose and R. Penrose (1958), Puzzles for Christmas, The New Scientist, 1580–1581,
1597. Victor Klee, in V. Klee (1979), Some unsolved problems in plane geometry, Math.
Mag. 52, 131–145, asked if a polygonal solution was possible and, somewhat surprisingly,
the answer is yes. In G. W. Tokarsky (1995), Polygonal rooms not illuminable from every
point, this Monthly, 102, 867–879, a 26-gon is constructed that cannot be illuminated
from a point. This was later improved by D. Castro to a 24-gon. In the polygonal examples,
only a single point stays dark.

294 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130

 

SOLUTIONS

Constructing a Tangent to a Circle

12245 [2021, 376]. Proposed by Jiahao Chen, Tsinghua University, Beijing, China. Sup-
pose that two circles α and β, with centers
P and Q, respectively, intersect orthogo- E
nally at A and B. Let CD be a diameter
F
of β that is exterior to α. Let E and F be A P
points on α such that CE and DF are tan- B
gent to α, with C and E on one side of P Q S
T
and D and F on the other side of P Q. Let
C D
S be the intersection of CF and QA, and Q
let T be the intersection of DE and QB.
Prove that ST is parallel to CD and is tan-
gent to α.
Solution by Davis Problem Solving Group, Davis, CA. Let Y be the intersection point of
lines BC and AD. We claim that Y lies on circle α and that the tangent line  to α at
Y is parallel to CD. To prove the claim, we assume for ease of exposition that A and
C are on the same side of P Q, with B and D on the other side, as in the figure that
accompanies the problem statement; however, the argument also works if the roles of A and
B are switched, as long as we view all angles as directed. Note that ∠BY A = ∠CY D =
180◦ − ∠DCB − ∠ADC, while ∠AP B = 180◦ − ∠BQA = 2∠DCB + 2∠ADC. Thus
∠BY A is inscribed in circle α and Y lies on α. Now let Z denote the intersection of AQ
and . Since ZA and ZY are both tangent to α, ∠ZY A = ∠Y AZ = ∠DAQ = ∠QDA,
and therefore  is parallel to CD. This proves the claim.

188 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
 Now let D  denote the second intersection point of line P D and circle β. Since inversion
in circle α preserves circle β, this inversion sends D to D  . Since ∠DD  C = 90◦ , it follows
that C is on the polar line d of point D with respect to circle α. The circumcircle of P DF
has diameter P D and thus maps to d under inversion in α. Thus line F C is the polar line
d of point D. Similarly, line ED is the polar line of point C with respect to α.
The polar lines of points A and B with respect to α are QA and QB, respectively, so S
is the intersection of the polar lines of A and D, and T is the intersection of the polar lines
of B and C. By duality, the polar lines of S and T are lines AD and BC, respectively. By
our initial claim, these polar lines intersect in Y . It follows that line ST is the polar line of
point Y , which is just the tangent line  to α at Y . Thus ST is parallel to CD and tangent
to α, as desired.
Also solved by M. Bataille (France), E. Bojaxhiu (Albania) & E. Hysnelaj (Australia), J. Cade, G. Fera (Italy),
D. Fleischman, K. Gatesman, N. Hodges (UK), A. M. Karparvar (Iran), K.-W. Lau (China), C. R. Pranesachar
(India), A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), L. Zhou, and the proposer.

An Integral over the Sphere

12247 [2021, 377]. Proposed by Prathap Kasina Reddy, Bhabha Atomic Research Centre,
Mumbai, India. For positive real constants a, b, and c, prove
 π ∞
a x 1
2 + a 2 )3/2
√ dx dθ =  .
0 0 π(x x + b + c − 2cx cos θ
2 2 2 (a + b)2 + c2
Solution by Giuseppe Fera, Vicenza, Italy. Let f (a, b, c) be the left side of the desired
equation. With the substitution x = a tan(ϕ/2), we obtain
√  π π
2 sin ϕ dϕ dθ
f (a, b, c) =  .
4π 0 0 a + b + c + (b + c2 − a 2 ) cos ϕ − 2ac cos θ sin ϕ
2 2 2 2

Since the integrand is invariant under the substitution θ → 2π − θ , we can write

√  2π  π
2 sin ϕ dϕ dθ
f (a, b, c) =  .
8π 0 0 a 2 + b2 + c2 + (b2 + c2 − a 2 ) cos ϕ − 2ac cos θ sin ϕ
Interpret ϕ and θ as the spherical coordinates for a point
r = (cos θ sin ϕ, sin θ sin ϕ, cos ϕ)
on the unit sphere S, and let v = (−2ac, 0, b2 + c2 − a 2 ). We see that
√ 
2 1
f (a, b, c) = √ dS.
8π S a + b + c2 + v · r
2 2

To evaluate this integral, we write it in cylindrical coordinates z and θ , with the positive
z-axis aligned with the vector v. Setting t = a 2 + b2 + c2 and
  
v = v = 4a 2 c2 + (b2 + c2 − a 2 )2 = (a 2 + b2 + c2 )2 − 4a 2 b2 = t 2 − 4a 2 b2 ,
this yields
√  1  2π √
2 dθ dz 2 √ √ 
f (a, b, c) = √ = t +v− t −v
8π −1 0 t + vz 2v
√  √ √
2 √ √ 2 t − t 2 − v2 t − 2ab
= t +v− t −v = =√
2v v t 2 − 4a 2 b2
1 1 1
=√ =√ = .
t + 2ab a + b + c + 2ab
2 2 2 (a + b)2 + c2

February 2023] PROBLEMS AND SOLUTIONS 189

Also solved by M. L. Glasser, O. Kouba (Syria), M. Omarjee (France), K. Sarma (India), A. Stadler (Switzer-
land), R. Tauraso (Italy), and the proposer.

An Identity from the Pfaffian

12248 [2021, 377]. Proposed by Askar Dzhumadil’daev, Almaty, Kazakhstan. Let n be a
positive integer, and let xk be a real number for 1 ≤ k ≤ 2n. Let C be the 2n-by-2n skew-
symmetric matrix with i, j -entry cos(xi − xj ) when 1 ≤ i < j ≤ 2n. Prove

det(C) = cos2 (x1 − x2 + x3 − x4 + · · · + x2n−1 − x2n ).

Solution by Richard Ehrenborg, University of Kentucky, Lexington, KY. The determinant

of a skew-symmetric matrix A is equal to the square of the Pfaffian of the matrix A. The
Pfaffian Pf(A) of a 2n-by-2n skew-symmetric matrix A with entries ai,j for 1 ≤ i, j ≤ 2n
is defined by

Pf(A) = (−1)c(M) · aij .
M (i,j )∈M

Here the sum is over all perfect matchings M on the set {1, . . . , 2n}, where an edge (i, j )
is written with i < j . Also c(M) is the number of pairs of crossing edges in M, where
two edges (i  , j  ) and (i  , j  ) in M form a crossing if i  < i  < j  < j  . The sign of a
matching M is (−1)c(M) . Our goal is to prove

Pf(C) = cos(x1 − x2 + x3 − · · · − x2n ).

Using the identity 2 cos(α) cos(β) = cos(α + β) + cos(α − β) and the fact that cosine
is an even function, a straightforward induction yields
n 
2 ·
n
cos(αi ) = cos(ε1 α1 + · · · + εn αn ).
i=1 (ε1 ,...,εn )∈{±1}n

Thus we express Pf(C) as follows, where we denote the edges of a matching M by

(i1 , j1 ), . . . , (in , jn ).
 
2n · Pf(C) = (−1)c(M) cos(ε1 (xi1 − xj1 ) + · · · + εn (xin − xjn )).
M (ε1 ,...,εn )∈{±1}n

By reordering the terms in the argument to cos, we can express each term on the right
side in the form cos(±x1 ± · · · ± x2n ), with n numbers weighted positively and n numbers
weighted negatively.
Consider a term for a matching M in which xk and xk+1 have the same coefficients, that
is, cos(· · · + εxk + εxk+1 · · · ), where ε ∈ {±1}. Since any two indices forming an edge of
M are given different signs, k and k + 1 do not form an edge in M.
Hence we can obtain another matching M  by switching the mates of k and k + 1 in
M. Always |c(M  ) − c(M)| = 1, and hence this mapping τk is a sign-reversing involution
on the set of matchings. The fixed points of τk are exactly those matchings that pair k and
k + 1. Hence the contributions of M and M  to the coefficient of any term of the form
cos(· · · + xk + xk+1 + · · · ) cancel.
Thus for each M the only terms that remain uncanceled under all τk are the two terms
with alternating signs: cos(x1 − x2 + x3 − · · · − x2n ) and cos(−x1 + x2 − x3 + · · · + x2n ).
Since cosine is an even function, these two terms are equal. We conclude

Pf(C) = cn · cos(x1 − x2 + x3 − · · · − x2n ), (∗)

190 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
where cn is a constant depending on n. To determine cn , set x1 = · · · = x2n = 0. The left
side of (∗) is now the Pfaffian of a skew-symmetric matrix having all entries above the
diagonal equal to 1. Expressing it in terms of matchings reduces it to M (−1)c(M) , where
the sum is over all matchings on {1, . . . , 2n}.
We prove M (−1)c(M) = 1 by induction on n. The base case n = 1 is easy: there
is exactly one matching on {1, 2}, with no crossings. For the induction step, define an
involution on the set of matchings on {1, . . . , 2n} by switching the elements 2n − 1 and
2n. If the result is a new matching, then the numbers of crossings in these two matchings
differ by 1, and the terms for these two matchings cancel in the sum. What remains are the
matchings where 2n − 1 and 2n form an edge. This edge crosses no other, so the sum for
these matchings is the same as the sum for all matchings on {1, . . . , 2n − 2}, which by the
induction hypothesis is 1.
Also solved by F. R. Ataev (Uzbekistan), H. Chen, N. Hodges (UK), P. Lalonde (Canada), O. P. Lossers
(Netherlands), M. Omarjee (France), C. R. Pranesachar (India), K. Sarma (India), A. Stadler (Switzerland),
M. Tang, R. Tauraso (Italy), J. Van hamme (Belgium), T. Wiandt, M. Wildon (UK), and the proposer.

Simplifying a Sum
12249 [2021, 377]. Proposed by Florin Stanescu, Serban Cioculescu School, Gaesti,
Romania. Prove
n−1  n−k  
m−1 k + m k + 1 k−m n
(−1) 2 =
k=n/2 m=1
k+1 m−1 2

for any positive integer n.

Solution by Rory Molinari, Beverly Hills, MI. Call the desired sum T (n), and let S(n) =
2T (n)/n. We prove S(n) = 1 for n > 0. For n > 0 and n/2 ≤ k ≤ n − 1, set
 
m−1 k + m k + 1
tm = (−1) 2k−m .
k+1 m−1
Letting
 
2(m − 1) k
sm = − tm = (−1) m
2k−m+1 ,
k+m m−2
it can easily be verified that tm = sm+1 − sm . Note that S(n) = n−1k=n/2 f (n, k), where
n−k
f (n, k) = (2/n) m=1 tm for n > 0 and n/2 ≤ k ≤ n − 1. Using s1 = 0, we have
 
2 2sn−k+1 (−1)n−k+1 k
f (n, k) = (sn−k+1 − s1 ) = = 22k−n+1 .
n n n n−k−1
 k 
Noting that n−k−1 is taken to be 0 unless n/2 ≤ k ≤ n − 1, it is natural to extend
f (n, k) by letting it be 0 unless n/2 ≤ k ≤ n − 1. Now

n−1 
S(n) = f (n, k) = f (n, k),
k=n/2 k

where k ranges over all integers. Let

(2k − n + 1)(2k − n)
R(n, k) = ,
2(n − k)(n + 1)
and put g(n, k) = R(n, k)f (n, k). Direct manipulation yields
f (n + 1, k) − f (n, k) = g(n, k + 1) − g(n, k)

February 2023] PROBLEMS AND SOLUTIONS 191

for all k and positive n. When summed over k, the right side telescopes to 0, so
 
S(n + 1) − S(n) = f (n + 1, k) − f (n, k) = 0.
k k

Thus S(n) is constant, and S(1) = f (1, 0) = 1, as required.

Editorial comment. The factors −2(m − 1)/(k + m) and R(n, k) come from Gosper’s algo-
rithm and the WZ algorithm, respectively (see M. Petkovšek, H. S. Wilf, and D. Zeilberger
(1997), A=B, A K Peters). In particular, R(n, k) is the certificate showing that (f, g) is a
Wilf-Zeilberger pair, meaning that f and g satisfy the properties needed to ensure that the
sum of f over k telescopes.
Also solved by J. Boswell & C. Curtis, P. Bracken, G. Fera (Italy), K. Gatesman, G. C. Greubel, D. Henderson,
N. Hodges (UK), O. Kouba (Syria), O. P. Lossers (Netherlands), E. Schmeichel, A. Stadler (Switzerland),
R. Stong, R. Tauraso (Italy), and the proposer.

A Polygon Inequality
12250 [2021, 377]. Proposed by Dorin Mărghidanu, Colegiul National A. I. Cuza, Cora-
bia, Romania. With n ≥ 4, let a1 , . . . , an be the lengths of the sides of a polygon.
Prove
a1 a2 an 2n
+ + ··· + > .
−a1 + a2 + · · · + an a1 − a2 + · · · + an a1 + a2 + · · · − an n−1

Solution by UM6P Math Club, Mohammed VI Polytechnic University, Ben Guerir, Morocco.
Since the left side is unaffected when the ai are scaled by a constant factor, we may assume
that the perimeter of the polygon is 1. Therefore, we need to show

n
ak 2n
> .
k=1
1 − 2ak n−1

By the triangle inequality, each ak belongs to the interval (0, 1/2), so by the AM–GM
inequality,

ak ak2 ak2 2ak

= ≥ = .
1 − 2ak ak (1 − 2ak ) (1 − ak ) /4
2 1 − ak
Note that this inequality is strict unless ak = 1/3. Since n ≥ 4, the inequality is strict for
some k, and therefore it suffices to show
n
2ak 2n
≥ .
k=1
1 − a k n −1

Let g(x) = 2x/(1 − x). Since g is convex on (0, 1/2), by Jensen’s inequality
 n n  n 
2ak k=1 g(ak ) k=1 ak 2n
=n· ≥n·g = n · g(1/n) = ,
k=1
1 − a k n n n −1

as required.
Solution II by Nigel Hodges, Cheltenham, UK. Denote the left side of the inequality √ by
T (a1 , . . . , an ). Since n ≥ 4, we have 4(n − √ 1) ≥ 3n, so 2n/(n − 1) ≤ 8/3 < 2 2. We
prove the stronger result T (a1 , . . . , an ) ≥ 2 2.

192 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
 As in the first solution above, we may assume nj=1 aj = 1, and hence 0 < aj < 1/2
for all j . Set aj = (1/2) sin2 θj with θj ∈ (0, π/2). This yields
√
aj aj tan θj 2 sin2 θj √ √
n = = √ = ≥ 2 sin2 θj = 2 2aj .
−2aj + t=1 at 1 − 2aj 2 sin(2θj )
Therefore

n
aj √ n
√
T (a1 , . . . , an ) = n ≥2 2 aj = 2 2.
j =1
−2aj + t=1 at j =1

√It is easy to see that this result is the best possible in that no larger constant can replace
2 2. Set a1 = a2 = a3 = a4 = 1 and √ aj = for 5 ≤ j ≤ n, where is a small positive
√
constant. We have T√(a1 , . . . , an ) = 2 2 + O( ), and so T (a1 , . . . , an ) can be made
arbitrarily close to 2 2 by choosing small enough.
Also solved by K. F. Andersen (Canada), M. Bataille (France), M. V. Channakeshava (India), H. Chen (China),
H. Chen (US), C. Chiser (Romania), K. Gatesman, C. Geon (Korea), W. Janous (Austria), O. Kouba (Syria),
S. S. Kumar, J. H. Lindsey II, O. P. Lossers (Netherlands), M. Lukarevski (North Macedonia), M. Omarjee
(France), E. Schmeichel, A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), D. Văcaru (Romania), F. Vis-
escu (Romania), L. Zhou, Westchester Area Math Circle, and the proposer.

Forcing Monochromatic Convex Pentagons with Fixed Area

12251 [2021, 467]. Proposed by Roberto Tauraso, Università di Roma “Tor Vergata,”
Rome, Italy. Each point in the plane is colored either red or blue. Show that for any positive
real number S, there is a proper convex pentagon of area S all five of whose vertices have
the same color. (By a proper convex pentagon we mean a convex pentagon whose internal
angles are less than π .)
Solution by Michael Tang, University of Washington, Seattle, WA. Denote the area of a
polygon by placing brackets around a list of its vertices, and let XY denote both the seg-
ment with endpoints X and Y and its length. Let B and R be the sets of blue and red points,
respectively. We begin with three observations that follow from assuming that the coloring
yields no such pentagon.
(i) Both B and R are unbounded. If B is bounded, then we find five acceptable red
vertices; similarly for bounded R.
(ii) Both R and B are dense in the plane. If R is not dense in the plane, then B contains
a disk D of some radius r centered at some point O. Also, B contains a point P with
OP > 2S/r. Choose X, Y ∈ D such that XY contains O and XY ⊥ OP and OX =
OY = (S − )/OP . Thus OX = OY < r/2 and [P XY ] = S − . We choose small
enough to guarantee the existence of a chord W Z on the circumcircle of P XY close and
parallel to XY (but farther from P than XY is) so that the isosceles trapezoid XY ZW has
area . Now XZW Y P is a proper convex pentagon with area S.
(iii) Every line segment contains points of both colors. If segment X1 X2 is all red, then
we construct such a pentagon. Choose Y, W, Z, V so that XY W ZV is proper convex for
all X ∈ X1 X2 and [X1 Y ZW V ] < S < [X2 Y ZW V ]. Since R is dense in the plane, we
may choose Y  , W  , Z  , V  in R arbitrarily close to Y, W, Z, V preserving convexity and
the inequality [X1 Y  Z  W  V  ] < S < [X2 Y  Z  W  V  ]. By continuity of the area function,
[XY ZW V ] = S for some X ∈ R, and this is our desired pentagon.
Given X, Y ∈ B, take Z ∈ B from a parallel line segment on one side of XY at a
distance (2S − 4 )/XY from it. Thus [XY Z] = S − 2 . From the other side of XY choose
W ∈ B. For sufficiently small , we can chose W inside the circumcircle of XY Z so that

February 2023] PROBLEMS AND SOLUTIONS 193

[XY W ] = . Similarly chose V inside the circumcircle of XY Z (but outside the triangle
XY Z near the edge XZ) so that [XV Z] = . Now [XV ZY W ] = S, and the construction
guarantees that XV ZY W is proper convex.
Editorial comment. Most solvers constructed a class of monochromatic quadrilaterals and
used casework to obtain a pentagon. The proposer started with a monochromatic rectangle
(similar to Problem 8.5 of the 1991 Colorado Math Olympiad). Many extended the result
to proper convex n-gons.
Also solved by J. Barát (Hungary), H. Chen (China), K. Gatesman, N. Hodges (UK), Y. J. Ionin, M. Reid,
C. Schacht, R. Stong, and the proposer.

Some Floors and Ceilings

12252 [2021, 467]. Proposed by Nguyen Quang Minh, Saint Joseph’s Institution, Singa-
pore. Let k, q, and n be positive integers with k ≥ 2, and let P be the set of positive integers
less than k n that are not divisible by k. Prove
  n − log p   k q−1 (k n−1 − 1)(k − 1) 
k
= + 1.
p∈P
q kq − 1
 
n−logk (p)
Solution by M. A. Prasad, Navi Mumbai, India. Write p∈P q
= T1 − T2 ,
where      n−1 kj +1   
 n − logk (p) n   n − logk (p)
T1 = = +
0<p<k n
q q j =0
q
j p=k +1
  n−1  
n n−j −1
= + k (k − 1)
j
q j =0
q

and    
 n − logk (j k)  n − 1 − logk (j )
T2 = =
q q
0<j ≤k n−1 0<j ≤k n−1
  
n−2  
n−1 n−j −2
= + k (k − 1)
j
.
q j =0
q

Combining these yields

    n−2    
n n−1 n−j −1 n−j −2
T1 − T2 = − + k j (k − 1) − .
q q j =0
q q

Let n − 2 = q + r with 0 ≤ r < q. If r = q − 1, then the only terms that contribute to

the right side are those with j ≡ r (mod q), so we obtain


(k − 1)k r (k (+1)q − 1)
T1 − T2 = k iq+r (k − 1) =
i=0
kq − 1

(k − 1)k q−1 (k n−1 − 1) (k − 1)(k q−1 − k r )

= + .
kq − 1 kq − 1
Since 0 < (k − 1)(k q−1 − k r )/(k q − 1) < 1, the result follows. If r = q − 1, we similarly
obtain
 
(k − 1)k q−1 (k n−1 − 1)
T1 − T 2 = 1 + k iq+r (k − 1) = 1 + .
i=0
kq − 1

194 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
Since the sum is an integer, the right side is an integer, and again we have the desired value.
Also solved by N. Hodges (UK), Y. J. Ionin, O. P. Lossers (Netherlands), K. Sarma (India), A. Stadler (Switzer-
land), A. Stenger, R. Stong, R. Tauraso (Italy), and the proposer.

An Arctangent Integral Solves a Summation

12254 [2021, 467]. Proposed by Cezar Lupu, Texas Tech University, Lubbock, TX, and
Tudorel Lupu, Constanţa, Romania. Prove
∞
 
 (−1)n  1
n
3π
= log 2 − G,
n=0
2n + 1 k=1
n + k 8
∞
where G is Catalan’s constant k
k=0 (−1) /(2k + 1)2 .
Composite solution by Michel Bataille, Rouen, France, and Omran Kouba, Higher Institute
for Applied Sciences and Technology, Damascus, Syria. Let S denote the requested sum.
We first compute


n
1 1
2n+1  n
1 1  (−1)k−1
2n+1
1
= −2 − = −
k=1
n+k k=1
k k=1
2k 2n + 1 k=1
k 2n + 1
 1 2n+1
 
1 1
1 + x 2n+1 1
= (−x)k−1 dx − = dx −
0 k=1 2n + 1 0 1 + x 2n +1
 1 2n+1
x 1
= dx + log 2 − .
0 1 + x 2n +1
It follows that
∞  1 ∞ ∞
(−1)n x 2n+1 (−1)n (−1)n
S= dx + log 2 −
n=0 0
(2n + 1)(1 + x) n=0
2n + 1 n=0 (2n + 1)2
∞  1
(−1)n x 2n+1 π
= dx + log 2 − G. (∗)
n=0 0
(2n + 1)(1 + x) 4
To evaluate the last sum, first note that
∞  1 
∞ 
 (−1)n x 2n+1  1
x 2n+1
  dx = dx
 (2n + 1)(1 + x)  (2n + 1)(1 + x)
n=0 0 n=0 0
∞  1  ∞
x 2n+1 1
≤ dx = < ∞.
n=0 0 2n + 1 n=0
(2n + 1)(2n + 2)
Hence we can reverse the order of the summation and integration to obtain
∞  1  1 ∞  1
(−1)n x 2n+1 (−1)n x 2n+1 arctan x
dx = dx = dx.
n=0 0
(2n + 1)(1 + x) 0 n=0 (2n + 1)(1 + x) 0 1+x

Using the change of variables x = (1 − t)/(1 + t) and the fact that for 0 ≤ t ≤ 1,
arctan((1 − t)/(1 + t)) = π/4 − arctan t we get
 1  1
arctan x π/4 − arctan t
dx = dt,
0 1 + x 0 1+t
and therefore  
1 1
arctan x π dt π
2 = = log 2.
0 1+x 4 0 1+t 4

February 2023] PROBLEMS AND SOLUTIONS 195

We conclude that the sum in (∗) equals (π/8) log 2, and therefore S = (3π/8) log 2 − G,
as required.
Also solved by A. Berkane (Algeria), N. Bhandari (Nepal), P. Bracken, B. Bradie, A. C. Castrillón (Colombia),
H. Chen, B. E. Davis, G. Fera (Italy), M. L. Glasser, R. Gordon, H. Grandmontagne (France), G. C. Greubel,
N. Grivaux (France), N. Hodges (UK), L. Kempeneers & J. Van Casteren (Belgium), O. P. Lossers (Nether-
lands), J. R. McCrorie (Scotland), M. Omarjee (France), D. Pinchon (France), M. A. Prasad (India), J. Song
(China), A. Stadler (Switzerland), S. M. Stewart (Australia), R. Stong, R. Tauraso (Italy), M. Vowe (Switzer-
land), T. Wiandt, M. Wildon (UK), FAU Problem Solving Group, and the proposer.

CLASSICS
C12. Due to Lionel Penrose and Roger Penrose; suggested by the editors. Is there a plane
region bounded by a differentiable Jordan curve with the property that no matter where a
light source is placed inside it, some part of the region remains unilluminated? Assume
that the curve acts as a perfect mirror.

Guessing When a Playing Card is Red

C11. Suggested by Richard Stanley, University of Miami, Coral Gables, FL. A standard
deck of cards has 26 red cards and 26 black cards. Deal out the cards in a shuffled standard
deck, one card at a time. At any point before the last card is dealt, you can guess that the
next card is red. For example, you may guess that the very first card is red, and your guess
will be correct with probability 1/2. Or you may watch some cards go by, noting their color
in order to decide when to guess. What strategy maximizes the probability that your guess
is correct?
Solution I. It is not possible to improve on 1/2. In fact, all stopping strategies have success
probability exactly 1/2. To see this, compare the game to a variant in which, after the guess
is made, the revealed card is the bottom card in the deck rather than the next card. When
any strategy is applied to this variant, the chance of success is clearly 1/2, since the bottom
card in a shuffled deck is red with probability 1/2. The key observation is that, no matter
when the guess is made, the next card has the same probability of being red as does the
bottom card. The probabilities are r/(r + b), where r and b are the number of red cards
and the number of black cards, respectively, in the deck following the specified position.
Since these probabilities determine the probability of success, the original game and the
variant have the same probability of success, independent of the strategy that is applied.
Solution II. We use induction on the size of the deck, proving the more general result that
any strategy wins with probability r/(b + r) when the deck starts with r red cards and b
black cards. If you guess that the first card is red, your probability of success is r/(b + r). If
you don’t, then consider two cases depending on the color of the first card. With probability
b/(b + r), the first card is black, and you are facing b − 1 black cards and r red cards in
the remaining deck. With probability r/(b + r), the first card is red, and you are facing b
black cards and r − 1 red cards. By the induction hypothesis, the probability of success,
independent of how the strategy continues, is
b r r r −1
+ ,
b+r b+r −1 b+r b+r −1
which equals r/(b + r).
Editorial comment. The problem is folklore, and appears on p. 67 of P. Winkler (2003),
Mathematical Puzzles, A Connoisseur’s Collection, A K Peters/CRC Press.

196 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
SOLUTIONS

Golden Eigenvalues of Special Matrices

12240 [2021, 276]. Proposed by Yue Liu, Fuzhou University, Fuzhou, China, and Fuzhen
Zhang, Nova Southeastern University, Fort Lauderdale, FL. We denote by A∗ the conjugate
transpose of the matrix A.
(a) Let x ∈ Cm be a unit column vector. Find the eigenvalues of the (m + 1)-by-(m + 1)
matrices
x∗x x∗ xx ∗ x
and .
x 0 x∗ 0

(b) More generally, let X be an m-by-n complex matrix, and let ρ be any real number. Find
the eigenvalues of the (m + n)-by-(m + n) matrices
X∗ X X∗ XX∗ X
and .
X ρIm X∗ ρIn

Solution to part (a) by Jean-Pierre Grivaux, Paris, France. Let M and N be the two spec-
ified matrices. Since x is a unit vector, x ∗ x = 1. The rank of M is two. Thus it has two
nonzero eigenvalues λ1 and λ2 , plus 0 with multiplicity m − 1. Note λ1 + λ2 = tr(M) = 1.
We calculate M 2 :

2 x∗
M2 = .
x xx ∗

January 2023] PROBLEMS AND SOLUTIONS 87

 With the entries of x indexed as x1 , . . . , xm , the m-by-m matrix xx ∗ has diagonal entries
|x1 |2 , . . . , |xm |2 . Thus tr(M 2 ) = 2 + |xi |2 = 3, so λ21 + λ22 =√3. Substituting√ λ2 =
1 − λ1 yields a quadratic equation, and we obtain {λ1 , λ2 } = {(1 − 5)/2, (1 + 5)/2}.
The argument for N is similar; it also has rank 2 and trace 1. Now
2xx ∗ xx ∗ x
N2 = ,
x ∗ xx ∗ 1
√ √
so tr(N 2 ) = 3. Again the two nonzero eigenvalues are (1 − 5)/2 and (1 + 5)/2.
Solution to part (b) by Kuldeep Sarma, Tezpur University, Tezpur, India. Again let M and N
be the two specified matrices. We use the singular value decomposition (SVD). The SVD
factors the m-by-n complex matrix X as U V ∗ , where U is an m-by-m complex unitary
matrix, V is an n-by-n complex unitary matrix, and  is an m-by-n rectangular diagonal
matrix with nonnegative real numbers σ1 , . . . , σs on the diagonal, where s = min{m, n}.
We can then write
V  ∗ V ∗ V ∗U ∗ V 0 ∗ ∗ V∗ 0
M= = .
U V ∗ U [ρIm ]U ∗ 0 U  ρIm 0 U∗
Since multiplication by a unitary matrix does not change eigenvalues, it suffices to find the
eigenvalues of the matrix S given by
∗ ∗
S= .
 ρIm
We consider a simultaneous permutation of the rows and columns of S, which does not
change the eigenvalues. Since  is nonzero only on its diagonal, many entries in S are
0. Index the first n rows (and columns) of S as 1 through n, and index the last m rows
(and columns) as 1 through m . Let s = min{m, n}. Reorder the rows (and columns) in the
order (1, 1 , 2, 2 , . . . , s, s ), followed by the remaining m + n − 2s rows (and columns) in
their original order. This converts S to a block-diagonal matrix S in which the ith block,
for 1 ≤ i ≤ s, is the 2-by-2 matrix
σi2 σi
,
σi ρ
and the final m + n − 2s blocks are 1-by-1 blocks that are all [ρ] if m > n and are all [0] if
m < n (there are none of these 1-by-1 blocks if m = n). Note that m + n − 2s = |m − n|.
The eigenvalues are the eigenvalues of the blocks: 0 or ρ with the stated multiplicity
|m − n|, plus
 2
ρ + σi2 ± ρ − σi2 + 4σi2
2
from the block for σi , where 1 ≤ i ≤ s. Note that if σi = 0, then the block for σi reduces
to two extra 1-by-1 blocks [0] and [ρ], but this is in fact described by the formula given
above for the eigenvalues of the block for σi .
The matrix N is generated in the same way as the matrix M, using X∗ instead of X.
It follows that the spectrum of N is the same as the spectrum of M, except that the mul-
tiplicities of 0 and ρ generated by the 1-by-1 blocks are, respectively, max{m − n, 0} and
max{n − m, 0}, obtained by interchanging the roles of m and n.
Also solved by D. Fleischman, K. Gatesman, L. Han (US) & X. Tang (China), E. A. Herman, C. P. A. Kumar
(India), O. P. Lossers (Netherlands), A. Stadler (Switzerland), R. Stong, E. I. Verriest, T. Wiandt, and the
proposer.

88 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
 An Integral Limit for This Year—Or, As It Turns Out, Any Year
12242 [2021, 277]. Proposed by Elena Corobea, Technical College Carol I, Constanţa,
Romania. For n ≥ 1, let
2022
 1 n
x k
/(2k + 1)
k=0
In = 2021 dx.
n+1 k
0
k=0 x /(2k + 1)

Let L = limn→∞ In . Compute L and limn→∞ n(In − L).

Solution by Kyle Gatesman (student), Johns Hopkins University, Baltimore, MD. We show
that L = 2 ln 2 and limn→∞ n(In − L) = −1/2.
For integers n ≥ 1 and p ≥ 0, let
n  1
xk (Sn (x))p+1
Sn (x) = and In (p) = dx.
k=0
2k + 1 0 (Sn+1 (x))
p

For p ≥ 1,
 1
(Sn (x))p Sn (x)
In (p) = ·
p−1 S
dx
0 (Sn+1 (x)) n+1 (x)
 1  
(Sn (x))p x n+1
= · 1− dx
0 (Sn+1 (x))
p−1 (2n + 3)Sn+1 (x)
 1 
Sn (x) p x n+1
= In (p − 1) − · dx.
0 Sn+1 (x) 2n + 3
For x ∈ [0, 1], we have
 p
Sn (x) x n+1 x n+1
0≤ · ≤ ,
Sn+1 (x) 2n + 3 2n + 3
so
 1
x n+1 1
0 ≤ In (p − 1) − In (p) ≤ dx = .
0 2n + 3 (n + 2)(2n + 3)
Therefore limn→∞ (In (p − 1) − In (p)) = 0, and by a straightforward induction on p we
conclude that limn→∞ (In (0) − In (p)) = 0 for all p ∈ Z+ . Moreover, for any constant
c ∈ R,
n
0 ≤ n(In (p − 1) − c) − n(In (p) − c) ≤ ,
(n + 2)(2n + 3)
and so lim (n(In (p − 1) − c) − n(In (p) − c)) = lim (n(In (0) − c) − n(In (p) − c)) = 0.
n→∞ n→∞
Because
 1  1 n n
xk 1
In (0) = Sn (x) dx = dx = ,
0 0 k=0 2k + 1 k=0
(k + 1)(2k + 1)

we conclude

n  ∞
1 1
lim In (p) = lim In (0) = lim =
n→∞ n→∞ n→∞
k=0
(k + 1)(2k + 1) k=0
(k + 1)(2k + 1)
∞
 ∞   ∞
1 1 1 (−1)k−1
=2 =2 − =2 = 2 ln 2.
k=0
(2k + 2)(2k + 1) k=0
2k + 1 2k + 2 k=1
k

January 2023] PROBLEMS AND SOLUTIONS 89

In particular, in the case p = 2021, we obtain L = 2 ln 2.
Similarly, observe that
lim n(In (p) − L) = lim n(In (0) − L)
n→∞ n→∞
 n ∞

 1  1
= lim n −
n→∞
k=0
(k + 1)(2k + 1) k=0 (k + 1)(2k + 1)
 ∞

 1
= lim n − .
n→∞
k=n+1
(k + 1)(2k + 1)

For every n ∈ Z+ we have

∞ ∞
 ∞
1 1 1
≤ ≤ .
k=n+1
(k + 1)(2k + 4) k=n+1
(k + 1)(2k + 1) k=n+1 (k + 1)2k
Since
∞
 ∞  
1 1  1 1 1
= − =
k=n+1
(k + 1)(2k + 4) 2 k=n+1 k + 1 k + 2 2(n + 2)
and
∞
 ∞  
1 1  1 1 1
= − = ,
k=n+1
(k + 1)2k 2 k=n+1
k k + 1 2(n + 1)
we conclude
 ∞

n  1 n
− ≤n − ≤− .
2(n + 1) k=n+1
(k + 1)(2k + 1) 2(n + 2)
Thus, by the squeeze theorem,
 ∞

1 1
lim n(In (p) − L) = lim n − =− ,
n→∞ n→∞
k=n+1
(k + 1)(2k + 1) 2
and setting p = 2021 completes the solution of the stated problem.
Editorial comment. The solution shows that the answers are the same if 2021 and 2022 are
replaced by p and p + 1 for any nonnegative integer p. Indeed, since In (p) is a decreasing
function of p, the answers are the same if 2021 and 2022 are replaced by x and x + 1 for
any nonnegative real number x.
Also solved by K. F. Andersen (Canada), P. Bracken, H. Chen, G. Fera (Italy), D. Fleischman, L. Han (USA)
& X. Tang (China), E. A. Herman, N. Hodges (UK), J. H. Lindsey II, O. P. Lossers (Netherlands), M. Omarjee
(France), K. Sarma (India), A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), T. Wiandt, J. Yan (China),
and the proposer.

A Hyperbolic Integral
12243 [2021, 277]. Proposed by M. L. Glasser, Clarkson University, Potsdam, NY. For
a > 0, evaluate
 a
t
 dt.
0 sinh t 1 − csch2 a · sinh2 t
Solution by Kuldeep Sarma, Tezpur University, Tezpur, India. Let I (a) be the desired value.
First, we observe that
1 − csch2 a sinh2 t = cosh2 t (1 − coth2 a tanh2 t).

90 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
Using this, we obtain
 a  a
t dt t dt
I (a) =  =  .
0 sinh t 1 − csch2 a · sinh2 t 0 sinh t cosh t 1 − coth2 a · tanh2 t
Now using the substitution cos x = coth a tanh t, we have
 π/2
tanh−1 (tanh a cos x)
I (a) = dx
0 cos x
and hence
 π/2
sech2 a π/2 π
I (a) = dx = sech a tan−1 (cosh a tan x)0 = sech a.
0 1 − tanh a cos x
22 2
Thus
 a  a
π π
I (a) = I (s) ds = sech s ds = tan−1 (sinh a).
0 2 0 2
Editorial comment. Several solvers noted that the requested integral can be reduced to
integral (3.535) from I. S. Gradshteyn, I. M. Ryzhik, et al. (2014), Table of Integrals, Series,
and Products, 8th edition, Cambridge, MA: Academic Press.
Also solved by U. Abel & V. Kushnirevych (Germany), P. Bracken, H. Chen, G. Fera (Italy), L. Han (US) &
X. Tang (China), N. Hodges (UK), O. P. Lossers (Netherlands), T. M. Mazzoli (Austria), M. Omarjee (France),
A. Stadler (Switzerland), S. M. Stewart (Saudi Arabia), R. Tauraso (Italy), UM6P Math Club (Morocco), and
the proposer.

Equitable Polyominos in a Box

12244 [2021, 376]. Proposed by Rob Pratt, SAS Institute Inc., Cary, NC, Stan Wagon,
Macalester College, St. Paul, MN, Douglas B. West, University of Illinois, Urbana, IL,
and Piotr Zielinski, Cambridge, MA. A polyomino is a region in the plane with connected
interior that is the union of a finite number of squares from a grid of unit squares. For which
integers k and n with 4 ≤ k ≤ n does there exist a polyomino P contained entirely within
an n-by-n grid such that P contains exactly k unit squares in every row and every column
of the grid? Clearly such polyominos do not exist when k = 1 and n ≥ 2. Nikolai Beluhov
noticed that they do not exist when k = 2 and n ≥ 3, and his Problem 12137 [2019, 756;
2021, 381] shows that they do not exist when k = 3 and n ≥ 5.
Solution by Jacob Boswell, Missouri Southern State University, Joplin, MO. Polyominos
with the desired properties, which we call (k, n)-equitable polyominos, exist whenever
4 ≤ k ≤ n.
Denote the n-by-n grid by G n . We call its unit squares cells and specify their positions in
matrix notation. We call the three cells (1,1), (1,2), and (2,1) the top left guard. Similarly,
we define top right, bottom left, and bottom right guards.
We argue by induction on k that in Gn there is a (k, n)-equitable polynomino that con-
tains two diagonally opposite guards such that removing the corner square from one of
those guards leaves the remainder connected. Let Ck,n denote the class of such polyomi-
nos. We postpone the discussion of the base cases.
For the induction step, consider (k, n) with n ≥ k ≥ 9. Cover Gn using two diagonally
opposite copies of G n/2 and two diagonally opposite copies of Gn/2 . When n is odd, the
two larger subgrids share one cell in the center, but other than that the subgrids share no
cells.
We describe a uniform construction for all cases except when n is odd and k is even. In
the two opposite copies of G n/2 , place members of C k/2 , n/2 , with one of the guards that

January 2023] PROBLEMS AND SOLUTIONS 91

are inductively guaranteed to exist placed in the center of Gn . In the two opposite copies of
Gn/2 , similarly place members of Ck/2,n/2 with their guaranteed guards in the center of
Gn .
When n is odd and k is even, use members of Ck/2+1, n/2 in the larger subgrids and
Ck/2−1,n/2 in the smaller subgrids, and (in this case) delete the central cell from the result-
ing polyomino. The use of Ck/2−1,n/2 here is the reason we need k = 8 in the basis.
In each case, the guards from each subpolyomino retain a cell adjacent to a cell retained
from the guard in a neighboring subpolyomino, so the resulting full polyomino is con-
nected. The polyomino also retains diagonally opposite complete guards, and deleting the
corner cell from one of those guards does not disconnect the polyomino, because it does
not disconnect the subpolyomino (even when the central cell is deleted, the two neighbors
of the central cell are connected through the other subpolyominos).
When n is even, the number of cells in each row and column of the final polyomino
is k/2 + k/2. When n is odd and k is odd, the computation is the same except for
the central row and column, where it is k/2 + k/2 − 1 as desired, since the central
cell contributes only once. When n is odd and k is even, we have k/2 + 1 + k/2 − 1
cells in each noncentral row and column, and in the central row and column we have
k/2 + 1 + k/2 + 1 − 2 cells, since the central cell was deleted. (Keeping the larger subgrid
connected in this case is the reason for the special condition on the subgrid.) Below we
show the construction of a member of C10,12 from four members of C5,6 .

Now we return to the base cases. Because the induction step for k needs the induction
hypothesis for (k − 1)/2 and (k, n)-equitable polyominos do not generally exist when
k ≤ 3, we need base cases for 4 ≤ k ≤ 8. Below we show members of C4,5 and C4,12 . The
general construction shown for (k, n) = (4, 12) is valid when n ≥ 6, which completes the
proof for k = 4.

92 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
 For k ≥ 5, we show first that a special construction for n = 2k + 2 yields constructions
for all larger n. Say that a member of Ck,2k+2 is a butterfly if its portion in the upper left and
lower right quadrants consists precisely of triangular arrays of cells with side-length k/2
touching the center of G2k+2 , as indicated on the left below. Suppose that Ck,2k+2 contains a
butterfly Bk . Note that the polyomino A in the upper right quadrant of Bk can be assumed
to be the transpose of A.
From Bk one can obtain a member of Ck,n whenever n > 2k + 2 by enlarging the central
portion of the butterfly and spreading A and A farther apart, as shown on the right below.
When k is even, the central diagonal of the added portion is omitted, but when k is odd it
is present. The correct counts in the rows and columns occupied by A and A are inherited
from Bk .

Below we show butterflies for 5 ≤ k ≤ 8. One issue in these constructions is ensuring

that the polyomino is connected; this is the reason we provided a different construction for
k = 4.

January 2023] PROBLEMS AND SOLUTIONS 93

 At this point the proof is completed by exhibiting explicit examples for k ≤ n ≤ 2k + 1
when 5 ≤ k ≤ 8. General constructions for n = k and n = k + 1 are trivial. What remains
is a finite problem, exhibiting 26 polynominos. We leave the constructions to the reader.
Editorial comment. The constructions are far from unique. For example, there is a con-
struction similar to the butterfly that exists when n = 2k and expands like the butterfly,
reducing the finite problem to 18 polyominos.
Also solved by K. Gatesman, R. Stong, and the proposer.

CLASSICS
C11. Suggested by Richard Stanley, University of Miami, Coral Gables, FL. A standard
deck of cards has 26 red cards and 26 black cards. Deal out the cards in a shuffled standard
deck, one card at a time. At any point before the last card is dealt, you can guess that the
next card is red. For example, you may guess that the very first card is red, and your guess
will be correct with probability 1/2. Or you may watch some cards go by, noting their color
in order to decide when to guess. What strategy maximizes the probability that your guess
is correct?

Repetitions in the Interior of Pascal’s Triangle

C10. Due to Douglas Lind, suggested by the editors. Show that there are infinitely many
numbers that appear at least six times in Pascal’s triangle.
   m 
Solution. For m ≥ 3, m occurs twice as m1 and m−1 . By symmetry, it will suffice to
find infinitely many values of m with at least two more occurrences in the left half of the
triangle.   16
There are several small examples of such pairs of occurrences: 120 = 10 = 2 ,
10 21 22 56 15 14 3
210 = 4 = 2 , 1540 = 3 = 2 , and 3003 = 5 = 6 . The last of these exhibits
  
the intriguing relationship nk = n−1 k+1
. To solve the problem, we will find infinitely many
solutions of this equation
   with k > 1 and k + 1 < (n − 1)/2.
The equation nk = n−1 k+1
is equivalent to n(k + 1) − (n − k)(n − k − 1) = 0. We claim
that for every positive integer j , this equation is satisfied by the values n = F2j +2 F2j +3

94 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 130
and k = F2j F2j +3 , where Fi is the ith Fibonacci number. To see why, note that with these
values we have n − k = (F2j +2 − F2j )F2j +3 = F2j +1 F2j +3 , and therefore
n(k + 1) − (n − k)(n − k − 1) = F2j +2 F2j +3 (F2j F2j +3 + 1) − F2j +1 F2j +3 (F2j +1 F2j +3 − 1)

= F2j +3 (F2j +2 F2j F2j +3 + F2j +2 − F2j

2
+1 F2j +3 + F2j +1 )

= F2j +3 (F2j +2 F2j F2j +3 − F2j

2
+1 F2j +3 + F2j +3 )

= F2j
2
+3 (F2j +2 F2j − F2j +1 + 1) = 0,
2

where the last step uses the well-known identity Fi+1 Fi−1 − Fi2 = (−1)i .
The case j = 1 yields n = 15 and k = 5, the
103example
 we found earlier. When j = 2 we
get n = 104 and k = 39, and indeed 104 39
= 40
= 61218182743304701891431482520.
Editorial comments. The appearance of the Fibonacci numbers in this solution can be
explained by reference to classic problem C2 (this Monthly, Feb. 2022, p. 194). View-
ing the equation n(k + 1) − (n − k)(n − k − 1) = 0 as a quadratic in n and applying the
quadratic formula yields
√
3k + 2 ± 5k 2 + 8k + 4
n= .
2
For n to be an integer, we need 5k 2 + 8k + 4 to be a perfect square. Setting 5k 2 + 8k + 4 = t 2
and solving for k by the quadratic formula, we get
√
−4 ± 5t 2 − 4
k= .
5
For k to be an integer, 5t 2 − 4 must be a perfect square, and the solution to classic problem
C2 (March 2022, pp. 293–294) shows that this happens if and only if t is an odd-indexed
Fibonacci number. Setting t = F2i+1 and applying Fibonacci identities leads to the values
(−1)i+1 − 1 4((−1)i+1 − 1)
n = Fi+1 Fi+2 + , k = Fi−1 Fi+2 + .
5 5
These are integers when i is odd, and setting i = 2j + 1 leads to the values used in the
solution. √
This result is due to Lind (D. Lind, The quadratic field Q( 5) and a certain Dio-
phantine equation, Fib. Quart. 6 (1968) 86–94, fq.math.ca/Scanned/6-3/lind.pdf). See also
C. A. Tovey, Multiple occurrences of binomial coefficients, Fib. Quart. 23 (1985) 356–358.
It is related to a 1971 conjecture of Singmaster (D. Singmaster, How often does an integer
occur as a binomial coefficient?, this Monthly 78 (1971) 385–386). For an integer m with
m ≥ 2, let Sm be the number of times m appears in Pascal’s triangle. Singmaster conjec-
tured that Sm is bounded, and suggested that 10 or 12 might be a bound. The problem shows
that 5 cannot be an asymptotic bound. It turns out that S3003 = 8; there are no other known
values of m for which Sm ≥ 8. The sequence of binomial coefficients for which Sm ≥ 6
starts 120, 210, 1540, 3003, 7140, 11628, 24310, 61218182743304701891431482520 (see
the OEIS sequences: oeis.org/A003015, oeis.org/A003016, and oeis.org/A090162). See
also K. Matomäki, M. Radziwiłł, X. Shao, T. Tao, and J. Teräväinen, Singmaster’s conjec-
ture in the interior of Pascal’s triangle, arxiv.org/abs/2106.03335.

January 2023] PROBLEMS AND SOLUTIONS 95

 +

SOLUTIONS

Counting Sets Without Consecutive Elements

12233 [2021, 178]. Proposed by C. R. Pranesachar, Indian Institute of Science, Bengaluru,
India. Let n and k be positive integers with 1 ≤ k ≤ (n + 1)/2. For 1 ≤ r ≤ n, let h(r)
be the number of k-element subsets of {1, . . . , n} that do not contain consecutive elements
but that do contain r. For example, with n = 7 and k = 3, the string h(1), . . . , h(7) is
6, 3, 4, 4, 4, 3, 6. Prove
(a) h(r) = h(r + 1) when r ∈ {k, . . . , n − k}.
(b) h(k − 1) = h(k) ± 1.
(c) h(r) > h(r + 2) when r ∈ {1, . . . , k − 2} and r is odd.
(d) h(r) < h(r + 2) when r ∈ {1, . . . , k − 2} and r is even.
Composite solution by Kyle Gatesman, Johns Hopkins University, Baltimore, MD, and
Roberto Tauraso, University of Rome Tor Vergata, Rome, Italy. The problem statement
requires correction in parts (c) and (d), where in the special case k = (n + 1)/2 we have
h(r) = h(r + 2) for all r.
For a proof by induction, we make the dependence on n and k explicit. Let hn,k (r) =
h(r), and extend the definition to give 0 when n, k, or r is outside its natural domain. For
1 ≤ r ≤ n − 1, partition the k-element subsets containing r by whether they contain n,
obtaining

hn,k (r) = hn−1,k (r) + hn−2,k−1 (r). (1)

Similarly, for 1 < r ≤ n, partition the k-element subsets containing r by whether they
contain 1. After shifting indices to start at 2 or 3, this yields

hn,k (r) = hn−1,k (r − 1) + hn−2,k−1 (r − 2). (2)

(a) We use induction on n. Note that hn,1 (r) = 1 for all r and n, from which (a) follows
for k = 1, including all cases with n ≤ 3. Now suppose n > 3 and k > 1. By symmetry,

986 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
hn,k (r) = hn,k (n + 1 − r), so we need only consider k ≤ r ≤ (n − 1)/2. In that case,
r ≤ (n − 1) − k = (n − 2) − (k − 1). Now (1) and the induction hypothesis imply
hn,k (r) = hn−1,k (r) + hn−2,k−1 (r) = hn−1,k (r + 1) + hn−2,k−1 (r + 1) = hn,k (r + 1).
(b) We use induction on k to prove that hn,k (k − 1) − hn,k (k) = (−1)k , for all positive
integers n beginning with hn,1 (0) = 0 and hn,1 (1) = 1. By (1) and (2),
   
hn,k (r) − hn,k (r + 1) = hn−1,k (r) + hn−2,k−1 (r) − hn−1,k (r) + hn−2,k−1 (r − 1)
 
= − hn−2,k−1 (r − 1) − hn−2,k−1 (r) . (3)
With r = k − 1 ≤ ((n − 2) + 1)/2, the induction hypothesis completes the proof.
(c, d) We use induction on  r. The number of k-element subsets of {1, . . . , n} having no
consecutive elements is n−k+1 , corresponding to insertions of k balls in distinct posi-
k    
tions between or outside n − k markers in a row. Thus hn,k (1) = n−k , hn,k (2) = n−k−1 ,
n−k−2 n−k−1 k−1 k−1
and, by (2), hn,k (3) = k−1 + k−2 . Using Pascal’s formula for binomial coefficients
 
twice, hn,k (1) − hn,k (3) = n−k−2
k−2
. Thus hn,k (1) − hn,k (3) > 0 unless k = (n + 1)/2, in
which case the difference is 0. This completes the proof for r = 1.
Now suppose r ≥ 2. If k = (n + 1)/2, then n is odd, and hn,k (r) is 1 when r is odd
and 0 when r is even, so the desired difference is 0. Hence we may restrict our attention to
k ≤ n/2, which yields k − 1 ≤ (n − 3 + 1)/2. Using (1) and (2), then (3), and finally (1)
and (2) again, we find
hn,k (r) − hn,k (r + 2) = hn−1,k (r) + hn−2,k−1 (r) − hn−1,k (r + 1) − hn−2,k−1 (r)
 
= − hn−3,k−1 (r − 1) − hn−3,k−1 (r)
 
= − hn−2,k−1 (r − 1) − hn−2,k−1 (r + 1) .
Now the induction hypothesis completes the proof.
Editorial comment. Nigel Hodges conditioned on the number j of selected elements pre-
ceding r to prove

k−1
r −1−j n−r −k+1+j
h(r) = .
j =0
j k−1−j

He then
used induction
 and  Pascal’s formula to prove for r ≤ n − k + 1 that this expression
equals r−1
j =0 (−1) j n−k−j
k−1−j
, from which (a)–(d) all follow quickly.
Also solved by H. Chen (China), C. Curtis & J. Boswell, N. Hodges (UK), Y. J. Ionin, O. P. Lossers (Nether-
lands), L. J. Peterson, R. Stong, and the proposer.

A Congruence for a Product of Quadratic Forms

12234 [2021, 179]. Proposed by Nicolai Osipov, Siberian Federal University, Krasnoyarsk,
Russia. Let p be an odd prime, and let Ax 2 + Bxy + Cy 2 be a quadratic form with A, B,
and C in Z such that B 2 − 4AC is neither a multiple of p nor a perfect square modulo p.
Prove that

(Ax 2 + Bxy + Cy 2 )
0<x<y<p

is 1 modulo p if exactly one or all three of A, C, and A + B + C are perfect squares

modulo p and is −1 modulo p otherwise.

December 2022] PROBLEMS AND SOLUTIONS 987

Solution by O. P. Lossers, Eindhoven University of Technology, Eindhoven, Netherlands.
All expressions below involving x and y take place in the finite field Fp with p elements.
We first study the desired product in general, leaving until later a consideration of how
many elements of {A, C, A + B + C} are squares. For convenience, define

Q(x, y) = Ax 2 + Bxy + Cy 2 .

Since we are given that B 2 − 4AC is a nonsquare, A and C must be nonzero,  and it follows
that Q(x, y) = 0 when (x, y) = (0, 0). In order to evaluate the product 0<x<y<p Q(x, y),
we want to group the factors by the value of Q(x, y). That is, for each D we seek the
number of solutions of Q(x, y) = D such that 0 < x < y < p.
For D = 0, since Q(x, y) − Dz2 = 0 determines a nondegenerate quadric, there are
altogether p2 − 1 solution triples (x, y, z) to Q(x, y) − Dz2 = 0. (See Lemma 7.23 on
p. 142 of J. W. P. Hirschfeld (1979), Projective Geometries over Finite Fields, Clarendon
Press.) The set of solution triples is invariant under multiplication by any nonzero ele-
ment of Fp . Hence the solutions come in p + 1 multiplicative classes of size p − 1, each
containing one triple of the form (x, y, 1), yielding p + 1 solutions to Q(x, y) = D.
This partitions the set of nonzero pairs (x, y) by the value of Q(x, y), with each value
D occurring exactly p + 1 times. Note that Q(x, y) = Q(p − x, p − y), so for fixed D
the number of pairs satisfying Q(x, y) = D with x < y equals the number of pairs with
x > y. Hence we will need to divide the number of occurrences of D by 2.
Since we require 0 < x < y < p in the stated product, we must also exclude occur-
rences of D that arise when x = 0, y = 0, or x = y. Two nonzero elements of Fp have the
same quadratic character if they are both squares or both nonsquares, equivalent to their
ratio being a square. Occurrences of D on the line x = 0 have Cy 2 − D = 0, or y 2 = D/C,
so there are two such pairs yielding D when D and C have the same quadratic character;
otherwise none. Similarly, there are two occurrences of D on y = 0 if and only if A and
D have the same quadratic character (satisfying x 2 = D/A), and two occurrences of D
on x = y if and only if A + B + C and D have the same quadratic character (satisfying
x 2 = D/(A + B + C)). Also, such occurrences on the three lines are distinct.
Let the number of squares among {A, C, A + B + C} be s. Starting with the p + 1 pairs
(x, y) ∈ F2p − (0, 0) that generate D, we subtract the occurrences with x = 0, y = 0, or
x = y and then divide the remaining occurrences by 2, as discussed above. We thus com-
pute that each square D occurs
  (p + 1 − 2s)/2 times, while each nonsquare
in the product
D occurs in the product p + 1 − 2(3 − s) /2 times.
This tells us how many times we have the product of all the squares and how many
times we have the product of all the nonsquares. It is well known that the product of all
the squares is (−1)(p+1)/2 , and the product of all the nonsquares is (−1)(p−1)/2 , because
an element and its reciprocal have the same quadratic character. After canceling reciprocal
pairs and ignoring 1, we are left with −1, which is a square if and only if p ≡ 1 mod 4.
We thus compute
 1 1 1 1
Q(x, y) = (−1) 2 (p+1) 2 (p+1−2s) (−1) 2 (p−1) 2 (p+1+2s−6)
0<x<y<p
 
1 (p+1)2 +(p 2 −1)−4s−6(p−1)
= (−1) 4
 
1 2 −2p+3−2s) 1 (p−1)2 +2−2s
= (−1) 2 (p = (−1) 2 = (−1)1−s .

This equals 1 or −1 when the number s of squares in {A, C, A + B + C} is odd or even,

respectively, as desired.
Also solved by C. Curtis & J. Boswell, Y. J. Ionin, R. Tauraso (Italy), and the proposer.

988 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
 An Application of Liouville’s Theorem
12235 [2021, 179]. Proposed by George Stoica, Saint John, NB, Canada. Let a0 , a1 , . . .
be a sequence of real numbers tending to infinity, and let f : C → C be an entire function
satisfying
|f (n) (ak )| ≤ e−ak
for all nonnegative integers k and n. Prove f (z) = ce−z for some constant c ∈ C with
|c| ≤ 1.
Solution by Kenneth F. Andersen, Edmonton, AB, Canada. We prove that the entire function
g(z) = ez f (z) satisfies
|g(z)| ≤ 1 (∗)
for all z. From this, Liouville’s theorem yields g(z) = c for some constant c, and then (∗)
yields |c| ≤ 1. Hence, f (z) = ce−z with |c| ≤ 1, as claimed.
Since f (z) is entire, for z = x + iy and k ≥ 0 we have
∞ 
 f (n) (a )  ∞
|f (n) (ak )|
 k 
|g(z)| = |ez |  (z − ak )n  ≤ ex |z − ak |n
 n!  n!
n=0 n=0
∞
 |z − ak |n
≤ ex e−ak = ex−ak +|z−ak | .
n=0
n!

Since limk→∞ ak = ∞, we have x < ak for sufficiently large k. Thus, for such k,
y2
|g(z)| ≤ e|z−ak |−|x−ak | = exp .
|z − ak | + |x − ak |
Taking the limit as k → ∞, we obtain (∗), which completes the proof.
Also solved by P. Bracken, L. Han (USA) & X. Tang (China), E. A. Herman, K. T. L. Koo (China), O. Kouba
(Syria), K. Sarma (India), A. Sasane (UK), A. Stadler (Switzerland), J. Yan (China), and the proposer.

The Googolth Term of a Sequence

12237 [2021, 276]. Proposed by Donald E. Knuth, Stanford University, Stanford, CA. Let
3/10
x0 = 1 and xn+1 = xn + xn for n ≥ 0. What are the first 40 decimal digits of xn when
n = 10 ?
100

Solution by Richard Stong, Center for Communications Research, San Diego, CA. The first
40 digits are 43236 87954 44259 51263 21573 91617 78825 77073.
Let f (x) = (10/7)x 7/10 , and let ak = f (xk ) for all k. Applying the mean value theorem
to f yields cn ∈ (xn , xn+1 ) such that
an+1 − an = cn−3/10 (xn+1 − xn ) = cn−3/10 xn3/10 .
Since cn > xn , this implies an+1 − an < 1. Computing x6 = 7 and a6 = 10 · 7−3/10 < 6,
we obtain an < n and hence xn < (7n/10)10/7 for n ≥ 6. Putting n = 10100 , we obtain an
upper bound for xn less than
4.3236 87954 44259 51263 21573 91617 78825 77073 38123 × 10142 .
We now provide a lower bound for xn . Applying the mean value theorem to g(x) =
x 3/10 yields bn ∈ (xn , xn+1 ) such that
3/10 3 −7/10 3 −7/10 3/10
cn3/10 − xn3/10 < xn+1 − xn3/10 = b (xn+1 − xn ) = b xn < 1.
10 n 10 n

December 2022] PROBLEMS AND SOLUTIONS 989

Hence
3/10 3/10
cn − xn 2
an+1 − an = 1 − 3/10
>1− 3/10
. (∗)
cn xn
By direct iteration, x45 = 102 > 410/3 . Since xn  is increasing, an+1 ≥ an + 1/2 whenever
n ≥ 45. From a45 > 45/2, for n ≥ 45 we conclude that an > n/2, hence xn > (7n/20)10/7 .
Explicit computation shows that this lower bound for xn also holds for n < 45. Therefore,
summing (∗) from 1 through n − 1 gives


n−1
2 
n−1
2 7
an > a1 + (n − 1) − 3/10
>n− 3/7
>n− n4/7 ,
k=1 xk k=1
(7k/20) 2(7/20)3/7

where at the last step we used the standard integral bound


n−1
1 n
1 7
3/7
≤ dt = n4/7 .
k=1
k 0 t 3/7 4

For n = 10100 , this yields a lower bound for xn greater than

4.3236 87954 44259 51263 21573 91617 78825 77073 37651 × 10142 .

Therefore, the first 40 digits of xn when n = 10100 are as claimed.

Also solved by O. P. Lossers (Netherlands), A. Stadler (Switzerland), R. Tauraso (Italy), E. Treviño, T. Wilde
(UK), The Logic Coffee Circle (Switzerland), and the proposer.

Collinear Midpoints from a Glide Reflection

12238 [2021, 276]. Proposed by Tran Quang Hung, Hanoi, Vietnam. Let ABCD be a
convex quadrilateral with AD = BC. Let P be the intersection of the diagonals AC and
BD, and let K and L be the circumcenters of triangles P AD and P BC, respectively. Show
that the midpoints of segments AB, CD, and KL are collinear.
Solution by Michel Bataille, Rouen, France. Let E and F be the midpoints of AB and CD,
respectively. Let m be the line through D that is parallel to EF , and let m be the image of
m under reflection through EF . Since F is the midpoint of CD, the point C must lie on
m . Let  be the circle centered at B with radius AD. Since AD = BC, the point C also
lies on .
Consider the 180◦ rotation of the plane centered at E. This rotation sends A to B and
D to some point D  . The rotation sends m to m , so D  lies on m , and since BD  = AD,
the point D  also lies on . However, D  cannot be C, because the midpoint of D  D is E,
whereas the midpoint of CD is F . Thus  and m intersect at two points, and those two
points are C and D  . It follows that if n is the line through B that is perpendicular to EF ,
then C is the reflection of D  through n.
Let g be the transformation of the plane consisting of rotation by 180◦ centered at E
followed by reflection through n. One sees easily that g is an orientation-reversing isometry
that sends A to B and D to C. (The transformation g can also be described as a glide
reflection with axis EF .)
For any lines  and  , let ∠(,  ) denote the directed angle from  to  . Let AD and
BC be the circumcircles of P AD and P BC, respectively, and let Q = g(P ).
Since g is orientation-reversing, ∠(QB, QC) = ∠(P D, P A) = ∠(P B, P C). There-
fore Q lies on BC . However, also Q, B, and C lie on g(AD ), so g(AD ) = BC . It
follows that g(K) = L, and therefore the midpoint of KL lies on EF .

990 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
Editorial comment. This solution shows that the quadrilateral need not be convex. Indeed,
it need not even be simple, as long as the lines AC and BD intersect.
Also solved by A. Ali (India), J. Cade, H. Chen (China), P. De (India), G. Fera (Italy), D. Fleischman, K. Gates-
man, O. Geupel (Germany), J.-P. Grivaux (France), W. Janous (Austria), D. Jones & M. Getz, O. Kouba
(Syria), K.-W. Lau (China), J. H. Lindsey II, O. P. Lossers (Netherlands), C. R. Pranesachar (India), A. Stadler
(Switzerland), R. Stong, R. Tauraso (Italy), M. Tetiva (Romania), T. Wiandt, L. Wimmer (Germany), L. Zhou,
Davis Problem Solving Group, and the proposer.

Factorials and Powers of 2

12239 [2021, 276]. Proposed by David Altizio, University of Illinois, Urbana, IL. Deter-
mine all positive integers r such that there exist at least two pairs of positive integers (m, n)
satisfying the equation 2m = n! + r.
Solution by Celia Schacht, North Carolina State University, Raleigh, NC. There are two
such values of r. They are r = 2, with 23 = 3! + 2 and 22 = 2! + 2, and r = 8, with
27 = 5! + 8 and 25 = 4! + 8. We show that there are no other values.
If 2m1 = n1 ! + r and 2m2 = n2 ! + r, then 2m1 − n1 ! = 2m2 − n2 !. For x ∈ N, let 2v(x) be
the highest power of 2 dividing x. Note that x can be uniquely written as 2v(x) times an odd
number, which we call the odd part of x. Since r > 0, we have 2mi > ni !, so mi > v(ni !)
for i ∈ {1, 2}. Therefore,
v(n1 !) = v(2m1 − n1 !) = v(2m2 − n2 !) = v(n2 !).
Given that (m1 , n1 ) = (m2 , n2 ), we may assume m1 > m2 and n1 > n2 . If there are any
even numbers from n2 + 1 to n1 , then v(n1 !) > v(n2 !), so v(n1 !) = v(n2 !) implies that n2
is even and n1 = n2 + 1. Let n2 = 2k. Thus
2m1 − 2m2 = n1 ! − n2 ! = (2k) · (2k)!. (4)
The odd part of the left side is 2m1 −m2 − 1. It equals the product of the odd parts of 2k and
(2k)!, so it is at least the odd part of (2k)!, which we write as 2q + 1. That is, 2m1 −m2 − 1 ≥
2q + 1.
By dividing out all the factors of 2 from (2k)!, we obtain
∞    ∞
2k 2k
v((2k)!) = i
< = 2k.
i=1
2 i=1
2i

First consider the case k ≥ 5. By induction, (2k)! > 24k for k ≥ 5. Therefore,
24k < (2k)! = 2v((2k)!) (2q + 1) < 22k (2q + 1),
so 22k − 1 < 22k < 2q + 1 ≤ 2m1 −m2 − 1. Also (2k)! = n2 ! < n2 ! + r = 2m2 , which yields
   
(2k)! 22k − 1 < 2m2 22k − 1 < 2m1 − 2m2 = (2k) · (2k)!.
Dividing by (2k)! yields 22k − 1 < 2k, which is false for all positive k. This contradiction
eliminates the possibility k ≥ 5.
It remains to check the cases of the form (n1 , n2 ) = (2k + 1, 2k) for k ∈ {1, 2, 3, 4}.
According to (4), we need powers of 2 differing by 2k(2k)!. For 1 ≤ k ≤ 4, the values
of 2k(2k)! are 4, 96, 4320, and 322560, respectively. Examining powers of 2 yields the
solutions for k ∈ {1, 2} listed at the start, but no solution for k ∈ {3, 4}.
Also solved by A. Ali (India), F. R. Ataev (Uzbekistan), C. Curtis & J. Boswell, S. M. Gagola Jr., K. Gates-
man, M. Ghelichkhani (Iran), N. Hodges (UK), P. Komjáth (Hungary), O. P. Lossers (Netherlands), S. Omar
(Morocco), J. Polo-Gómez (Canada), K. Sarma (India), A. Stadler (Switzerland), R. Stong, M. Tang,
R. Tauraso (Italy), E. Treviño, T. Wilde (UK), L. Zhou, and the proposer.

December 2022] PROBLEMS AND SOLUTIONS 991

 Harmonic Sums: Euler Once, Abel Twice
12241 [2021, 276]. Proposed by Ovidiu Furdui and Alina Sı̂ntămărian, Technical Univer-
sity of Cluj-Napoca, Cluj-Napoca, Romania. Prove
∞
 
 1 2n
1 ln 2 − 1
(−1)n n − ln 2 + = .
n=1
4n k=n+1
k 8
Solution by Kee-Wai Lau, Hong Kong, China. We first address the partial sum of the series
on the left side and show
 
N
1 2n
1
8 (−1) nn
− ln 2 + (1)
n=1
4n k=n+1
k
 
2N
1 N
(−1)n
= 2(−1) (2N + 1)
N
− ln 2 + + (−1)N − 1 + 2 ln 2.
k=N+1
k n=1
n
Since ln 2 is irrational, it must have the same coefficient on both sides, requiring

N
8 (−1)n n = 2(−1)N (2N + 1) − 2.
n=1
This equality is easily verified by considering odd
 and evenn N separately. NLet K(N ) denote
the quantity on both sides. In addition, since 8 N n=1 (−1) (1/4) = (−1) − 1, the sum of
the N initial terms on the left in (1) equals the sum of two terms on the right. It remains to
prove
N 2n
1 2N
1  (−1)n
N
8(−1)n n = 2(−1)N (2N + 1) + .
n=1 k=n+1
k k=N+1
k n=1 n
Let L(N ) denote the left side in this equation. Rewrite that double sum as

N
L(N ) = (K(n) − K(n − 1))J (n),
n=1
2n
where J (n) = k=n+1 1/k and K(0) = 0. By partial summation,

N−1
L(N) = K(N )J (N ) + K(n)(J (n) − J (n + 1)).
n=1

Now
1 1 1 −1
J (n) − J (n + 1) = − − = .
n + 1 2n + 1 2n + 2 2(n + 1)(2n + 1)
Hence
  
N−1
 1
L(N ) = 2(−1)N (2N + 1) − 2 J (N ) + (−1)n+1 (2n + 1) + 1
n=1
(n + 1)(2n + 1)


N−1
(−1)n+1  N−1
1
= 2(−1)N (2N + 1)J (N ) + − 2J (N ) + . (2)
n=1
n+1 n=1
(n + 1)(2n + 1)
Restoring the expression involving J in the last summand, the last two terms in (2) simplify
by telescoping as

N−1
−2J (N ) − 2 (J (n) − J (n + 1)) = −2J (N ) − 2(J (1) − J (N )) = −1.
n=1

992 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
Now the expression for L(N ) reduces to the right side of (1), completing the proof of the
identity. 
Let HN denote the harmonic number N n=1 1/n. By Euler–Maclaurin summation,

1
HN = ln N + γ + + O(N −2 ),
2N
where γ is Euler’s constant. Thus
2N
1 1
= H2N − HN = ln 2 − + O(N −2 ).
n=N+1
n 4N

Hence the first term on the right side of (1) simplifies as

−1
2(−1)N (2N + 1) + O(N −2 ) = −(−1)N + O(N −1 ).
4N
Also,
∞
 (−1)n
= − ln 2.
n=1
n

Thus the right side of (∗) converges to −1 + ln 2, which completes the proof.
Editorial comment. Another approach to evaluating the left side is to introduce the factor x n
for 0 < x < 1 into the sum, expand, and let x approach 1. This is an application of Abel’s
limit theorem, known as Abel summation. Ulrich Abel (fittingly) and Vitaliy Kushnirevych
used this method. With
∞

1 − ln(1 − x)
an = − ln 2 + H2n − Hn and g(x) = Hn x n = ,
4n n=1
1−x

let
∞
 √ √
− ln(1 + x) x ln 2 g(i x) + g(−i x)
f (x) = an (−x) = n
− + − g(−x).
n=1
4 1+x 2

Upon differentiating f (x), we obtain a power series for (−1)n nan , and Abel summation
yields the result.
Many solvers used a method somewhat akin to Abel summation, that of integral repre-
sentation. For example, Richard Stong used
1 1
1 − x 2n−1
an = x dx.
2 0 1+x
Upon interchange of summation and integration (justified by dominated convergence), the
desired sum then becomes the readily evaluated integral
1 1
1−x x
− dx.
2 0 1 + x (1 + x 2 )2

Also solved by U. Abel & V. Kushnirevych (Germany), A. Berkane (Algeria), P. Bracken, B. Bradie, H. Chen,
G. Fera (Italy), K. Gatesman, M. L. Glasser, G. C. Greubel, L. Han (US) & X. Tang (China), E. A. Herman,
N. Hodges (UK), S. Kaczkowski, O. Kouba (Syria), P. W. Lindstrom, O. P. Lossers (Netherlands), M. Omarjee
(France), K. Sarma (India), A. Stadler (Switzerland), S. M. Stewart (Australia), R. Stong, R. Tauraso (Italy),
M. Vowe (Switzerland), T. Wiandt, and the proposer.

December 2022] PROBLEMS AND SOLUTIONS 993

CLASSICS
C10. Due to Douglas Lind, suggested by the editors. Show that there are infinitely many
numbers that appear at least six times in Pascal’s triangle.

How Much of a Parabolic Arc Can Fit in a Unit Disk?

C9. From the 2001 Putnam Competition. Can an arc of a parabola inside a circle of radius
1 have a length greater than 4?
Solution. The answer is yes. For a positive real number A, the parabola
√ y = Ax 2 intersects
the√circle x + (y − 1) = 1 at the origin and at the points ( 2A − 1/A, 2 − 1/A) and
2 2

(− 2A − 1/A, 2 − 1/A). The length L(A) of the parabolic arc between these points con-
sists of two congruent parts, one in each quadrant. Expressing the length of one of these
parts as an integral with respect to the variable y and then letting u = Ay, we obtain
 
2−1/A
1 2 2A−1 1
L(A) = 2 1+ dy = 1+ du.
0 4Ay A 0 4u

It suffices to find a value of A so that L(A) is greater than 4. This occurs when
 
2A−1
1
1+ − 1 du ≥ 1.
0 4u

Since
   
1 1 1
1+ −1 1+ +1 = ,
4u 4u 4u

when u > 1/12 we have


1 1
1+ −1≥ .
4u 12u
Therefore
   
2A−1 2A−1 2A−1
1 1 1
1+ − 1 du ≥ 1+ − 1 du ≥ du.
0 4u 1 4u 1 12u
∞
Because 1 (1/x) dx diverges, we may choose A so large that this last integral exceeds 1.

Editorial comments. Numerical calcu-

lation shows that the longest arc is
achieved when A is approximately 94.1,
at which point the length is approximately
4.00267. The figure shows this longest
parabolic arc. Not until A is approxi-
mately 37 does the arc length exceed 4.
In the 2001 Putnam Competition, just
one participant (out of approximately
3000) earned full credit for solving this
problem.

994 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
SOLUTIONS

Making Equality Improbable with Two Dice

12223 [2021, 88]. Proposed by Michael Elgersma, Plymouth, MN, and James R. Roche,
Ellicott City, MD. Two weighted m-sided dice have faces labeled with the integers 1 to m.
The first die shows the integer i with probability pi , while the second die shows the integer
i with probability ri . Alice rolls the two dice and sums the resulting integers; Bob then
independently does the same.
(a) For each m with m ≥ 2, find the probability vectors (p1 , . . . , pm ) and (r1 , . . . , rm ) that
minimize the probability that Alice’s sum equals Bob’s sum.
(b)* Generalize to n dice, with n ≥ 3.
Composite solution to part (a) by the proposers and Shuyang Gao, George Washington
University, Washington, DC. The minimum probability is 3/(6m − 4), achieved only by
the two distributions
1 1 1
, 0, 0, . . . , 0, 0, and (2, 3, 3, . . . , 3, 3, 2).
2 2 3m − 2
We start with some notation. We write v for a probability (row) vector (v1 , . . . , vm )
associated with the faces of an m-sided die; that is, the probability that a toss of such a die
turns up value i is vi (similarly with other letters). The reverse R(v) of v is (vm , . . . , v1 ).
We say that v is symmetric if v = R(v). For symmetrization and antisymmetrization,
let Sv = (v + R(v))/2 and Av = (v − R(v))/2. Thus v = Sv + Av , R(Sv ) = Sv , and
R(Av ) = −Av .
Let p and r denote the probability vectors for the two dice. Let X and Y be the sums
rolled by Alice and Bob, respectively. Note that X and Y have the same distribution. Let
s = (s2 , . . . , s2m ), where

m
sk = P(X = k) = P(Y = k) = pi rk−i ,
i=1

with the understanding that rj = 0 unless 1 ≤ j ≤ m. With ∗ denoting convolution of

vectors, we write s as p ∗ r.
Our first task is to show that the probability is minimized only when p and r are sym-
metric. The tool for this is the claim
P(X = Y ) ≥ (Sp ∗ Sr ) · (Sp ∗ Sr ),

November 2022] PROBLEMS AND SOLUTIONS 887

with equality holding if and only if p and r are both symmetric probability vectors. Given
this, let p and r be minimizing probability vectors. If we replace p and r by their sym-
metrizations Sp and Sr , then the new resulting probability P(X = Y ) will be equal to
(Sp ∗ Sr ) · (Sp ∗ Sr ), which will be strictly smaller than the original probability unless
p = Sp and r = Sr .
Hence we proceed to the claim. Since the players’ rolls are independent,

2m 
2m 
m
2
P(X = Y ) = P(X = k) P(Y = k) = pi rk−i .
k=2 k=2 i=1

We write this using convolution and inner product as

 
P(X = Y ) = (p ∗ r) · (p ∗ r) = (Sp +Ap ) ∗ (Sr +Ar ) · (Sp +Ap ) ∗ (Sr +Ar ) .
By linearity of convolution and inner product, this expression expands into sixteen terms of
the form (fp ∗ gr ) · (hp ∗ ir ) with f, g, h, i ∈ {S, A}. We show that the contribution from
the terms other than (Sp ∗ Sr ) · (Sp ∗ Sr ) is nonnegative and is 0 if and only if p and r are
symmetric.
Since Sp ∗ Sr and Ap ∗ Ar are symmetric and Sp ∗ Ar and Ap ∗ Sr are antisymmetric,
each of the eight terms having one or three factors in {Ap , Ar } is the dot product of a
symmetric and an antisymmetric vector and hence vanishes.
With f, g ∈ {S, A}, we find four terms of the form (fp ∗ gr ) · (fp ∗ gr ). Each is non-
negative, since it is the dot product of a vector with itself, and it equals 0 if and only if
fp ∗ gr = 0. The convolution is 0 when f = A and p is symmetric, since then Ap = 0.
However, if p is not symmetric, then Ap ∗ Sr = 0. The corresponding statements hold also
for g. Hence the contribution from these four terms is at least (Sp ∗ Sr ) · (Sp ∗ Sr ), with
equality if and only if both p and r are symmetric.
The remaining four terms use each factor in {Sp , Sr , Ap , Ar }. They sum to

2 (Sp ∗ Sr ) · (Ap ∗ Ar ) + (Sp ∗ Ar ) · (Ap ∗ Sr ) . (1)
We claim that this sum is 0. We have

(Sp ∗ Sr ) · (Ap ∗ Ar ) = Sp (k)Sr ()Ap (k )Ar ( ) (2)
and

(Sp ∗ Ar ) · (Ap ∗ Sr ) = Sp (k)Ar ( )Ap (k )Sr (), (3)
where the sum in (2) is over choices of k, , k ,  in {1, . . . , m} such that k +  = k +  ,
and the sum in (3) is over choices such that k +  = k + . Note that k +  = k + 
if and only if k − k =  −  and that k +  = k +  if and only if k − k =  −  . By
symmetry and antisymmetry,
Sr () = Sr (m −  + 1) and Ar ( ) = −Ar (m −  + 1).
Thus Sp (k)Sr ()Ap (k )Ar ( ) = −Sp (k)Sr (m −  + 1)Ap (k )Ar (m −  + 1). When we
require k − k =  − , at the same time we have k − k = (m −  + 1) − (m −  + 1).
Hence terms in the sum in (3) negate corresponding terms in the sum in (2), and the expres-
sion in (1) is 0. This completes the proof of the claim.
The claim implies the desired result in the case m = 2, giving p = r = (1/2, 1/2). For
the remainder of the argument, we assume m ≥ 3. With p and r symmetric, the convolu-
tion s is also a symmetric probability vector, and the desired probability is 2m 2
k=2 sk . By
symmetry,

m
sm+1 = pi rm−i+1 ≥ p1 rm + pm r1 = 2p1 r1 = 2s2 . (4)
i=1

888 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
This suggests that we consider the following nonlinear optimization problem:

minimize 2(s22 + · · · + sm2 ) + sm+1

2

subject to the constraints

2(s2 + s3 + · · · + sm ) + sm+1 = 1, 2s2 ≤ sm+1 , and si ≥ 0 for 2 ≤ i ≤ m + 1.

Extending (s2 , . . . , sm+1 ) by letting s2m−i = s2+i for 0 ≤ i ≤ m − 2 relates this optimiza-
tion problem to the symmetric probability vector s considered earlier. This problem incor-
porates the constraint (4), but it ignores the requirement in the original problem that s be
realizable as the convolution of two probability vectors. It then suffices to show that we
can realize the resulting optimum by such a convolution.
Such constrained optimization problems can be solved using the Karush-Kuhn-Tucker
(KKT) conditions (see for example S. Boyd and L. Vandenberghe (2004), Convex Opti-
mization, Cambridge University Press). Satisfying the conditions is sufficient for a global
optimum. The method starts with a generalized Lagrangian incorporating the objective
function, the inequality constraints, and the equality constraints:

L = 2(s22 + · · · + sm2 ) + sm+1
2
+ μ(2s2 − sm+1 ) + λ 2(s2 + · · · + sm ) + sm+1 − 1 .

The KKT conditions require partial derivatives with respect to the original variables and
the multipliers for equality constraints to be 0, while for the multipliers of the inequal-
ity constraints we must have nonnegativity (see (9)) and “complementary slackness” (see
(10)). That is,
∂L
= 4s2 + 2μ + 2λ = 0; (5)
∂s2
∂L
= 4si + 2λ = 0 for 3 ≤ i ≤ m; (6)
∂si
∂L
= 2sm+1 − μ + λ = 0; (7)
∂sm+1
2(s2 + · · · + sm ) + sm+1 − 1 = 0. (8)
μ ≥ 0; and (9)
μ(2s2 − sm+1 ) = 0. (10)

We also require si ≥ 0 for all i in {2, . . . , m + 1}.

We show first that λ must be negative. If λ > 0, then by (6) each si with i ≥ 3 is
negative, which is forbidden. If λ = 0, then (6) requires s3 = · · · = sm = 0. Since (5) now
reads 4s2 + 2μ = 0, it forbids μ > 0, so μ = 0 by (9). Now s2 = 0 by (5) and sm+1 = 0
by (7), but that contradicts (8).
Hence λ < 0. Note that subtracting (5) from (7) gives 2sm+1 − 4s2 = 3μ + λ. Since we
require 2s2 ≤ sm+1 and have λ < 0, we must have μ > 0. Now (10) requires 2s2 = sm+1 .
With these restrictions, (5)–(7) reduce to
3
λ = −3μ, s2 = μ, sm+1 = 2μ, and si = μ for 3 ≤ i ≤ m.
2
m
Using sm+1 + 2 i=2 si = 1, we obtain μ = 1/(3m − 2), and consequently
1 2 3
s2 = , sm+1 = , and si = for 3 ≤ i ≤ m.
3m − 2 3m − 2 6m − 4

November 2022] PROBLEMS AND SOLUTIONS 889

Extending back to the probability vector s with indices 2 through 2m, we obtain
1
s= (2, 3, 3, . . . , 3, 3, 4, 3, 3, . . . , 3, 3, 2), (11)
6m − 4
yielding the minimum probability 2m k=2 sk = 3/(6m − 4).
2

This solution to the optimization problem is achievable as the convolution of the two
probability vectors
1 1 1
, 0, 0, . . . , 0, 0, and (2, 3, 3, . . . , 3, 3, 2).
2 2 3m − 2
Our final task is to show that these are the only probability vectors whose convolution
is (11). To achieve s2 = s2m > 0, we have p1 = pm > 0 and r1 = rm > 0. Since we must
satisfy

m−1
2s2 = sm+1 = p1 rm + pm r1 + pi rm+1−i ,
i=2

we obtain pi rm+1−i = 0 for 2 ≤ i ≤ m − 1. Consequently, for each i with 2 ≤ i ≤ m − 1,

pi = pm+1−i = 0 or rm+1−i = ri = 0.
By symmetry, we may take p2 = 0. Now let k be the least integer in {2, . . . , m} such that
pk > 0. It suffices to show that k = m, which yields p = (1/2, 0, . . . , 0, 1/2), whereupon
the known convolution (11) yields r as claimed.
Suppose k < m. By (11),
3
= si = p1 ri−1 + 0 + 0 + · · · + 0 for 3 ≤ i ≤ k.
6m − 4
Since p1 r1 = 2/(6m − 4), we obtain ri−1 = 3r1 /2 > 0 for 3 ≤ i ≤ k.
Next, sk+1 = p1 rk + pk r1 . Since pk rk = pk rm+1−k = 0 and pk > 0, we have rk = 0.
Now pk r1 = sk+1 = 3/(6m − 4) and p1 r1 = s2 = 2/(6m − 4). Thus, pk = 3p1 /2. Finally,
 
3 3 4
sk+2 ≥ pk r2 = p1 r1 > 2s2 = ,
2 2 6m − 4
contradicting sk+2 ≤ 4/(6m − 4). Thus k = m, completing the proof.
Editorial comment. The problem arose as an extension of Problem 1290 in Stan Wagon’s
Problem of the Week, which in turn was inspired by a problem on Tanya Khovanova’s
blog: blog.tanyakhovanova.com/2018/12/two-dice.
No solutions to part (b) or other correct solutions to part (a) were received.

A Lower Bound on Average Squared Acceleration

12229 [2021, 89]. Proposed by Moubinool Omarjee, Lycée Henri IV, Paris, France. Let
f : [0, 1] → R be a function that has a continuous second derivative and that satisfies
1
f (0) = f (1) and 0 f (x) dx = 0. Prove
 1 2  1
 2
30240 xf (x) dx ≤ f (x) dx.
0 0
Solution by Rory Molinari, Beverly Hills, MI. Applying integration by parts twice, and
1 1
using 0 f (x) dx = 0 and 0 f (x) dx = f (1) − f (0) = 0, we get
 1  1   1 
1 x2 x
xf (x) dx = x− f (x) dx = − − f (x) dx
0 0 2 0 2 2
 1   1 
x2 x 1 x3 x2 x
=− − + f (x) dx = − + f (x) dx.
0 2 2 12 0 6 4 12

890 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
Thus, by the Cauchy–Schwarz inequality,
 1 2  1  2
x3 x2 x
xf (x) dx = − + f (x) dx
0 0 6 4 12
 2
  1   1
1
x3 x2 x 1
≤ − + dx · (f (x)) dx =
2
(f (x))2 dx,
0 6 4 12 0 30240 0

and the desired conclusion follows.

Editorial comment. Justin Freeman generalized the problem by proving
 1 2  1
(2n + 2)!
xf (x) dx ≤ (f (n) (x))2 dx,
|B2n+2 | 0 0

where Bk is the kth Bernoulli number.

Also solved by U. Abel & V. Kushnirevych (Germany), K. F. Andersen (Canada), M. Bataille (France),
A. Berkane (Algeria), P. Bracken, B. Bradie, H. Chen, G. Fera (Italy), J. Freeman (Netherlands), K. Gates-
man, G. Góral (Poland), N. Grivaux (France), L. Han, E. A. Herman, L. T. L. Koo (China), O. Kouba (Syria),
K.-W. Lau (China), Z. Lin (China), J. H. Lindsey II, O. P. Lossers (Netherlands), I. Manzur (UK) & M. Graczyk
(France), T. M. Mazzoli (Austria), A. Natian (UK), A. Pathak (India), B. Shala (Slovenia), A. Stadler (Switzer-
land), R. Stong, R. Tauraso (Italy), E. I. Verriest, M. Vowe (Switzerland), J. Vukmirović (Serbia), T. Wiandt,
J. Yan (China), L. Zhou, U. M. 6. P. MathClub (Morocco), and the proposer.

Families of Permutations with Equal Size

12230 [2021, 178]. Proposed by David Callan, University of Wisconsin, Madison, WI. Let
[n] = {1, . . . , n}. Given a permutation (π1 , . . . , πn ) of [n], a right-left minimum occurs at
position i if πj > πi whenever j > i, and a small ascent occurs at position i if πi+1 =
πi + 1. Let An,k denote the set of permutations π of [n] with π1 = k that do not have right-
left minima at consecutive positions, and let Bn,k denote the set of permutations π of [n]
with π1 = k that have no small ascents.
(a) Prove |An,k | = |Bn,k | for 1 ≤ k ≤ n.
(b) Prove |An,j | = |An,k | for 2 ≤ j < k ≤ n.
Solution by Richard Stong, Center for Communications Research, San Diego, CA. For
n = 1, we have |A1,1 | = |B1,1 | = 1. Hence it suffices to show that both cn,k = |An,k | and
cn,k = |Bn,k | satisfy the recurrence
⎧
⎪
⎪
n−1
⎪
⎨ cn−1,j if k = 1,
j =2
cn,k = n−1
⎪
⎪
⎪ cn−1,j if k > 1.
⎩
j =1

The common recurrence then shows (a), and its form implies (b).
To a permutation π of [n], associate the permutation σ of [n − 1] obtained by deleting
π1 and decreasing all entries exceeding π1 by 1. From π1 and σ , we can reconstruct π
uniquely. In addition, σ has a right-left minimum at position i if and only if π has a right-
left minimum at position i + 1.
For k > 1, any permutation σ of [n − 1] with no right-left minima in consecutive posi-
tions arises from a permutation π ∈ An,k , and permutations in An,k generate such σ , since
position 1 in π is not a right-left minimum. Thus, the recursive formula holds for |An,k |
when k > 1. When k = 1, π has a right-left minimum in position 1, so we must ensure

November 2022] PROBLEMS AND SOLUTIONS 891

that the corresponding σ has no right-left minimum in position 1, which is equivalent to
σ1 = 1. Thus, the formula holds also for |An,1 |.
We show that this recurrence also holds for Bn,k . Again consider the same map, with
π ∈ Bn,k . If σ has no small ascents, then also π has none, unless σ1 = k. On the other
hand, if π has no small ascents, then σ has at most one small ascent, with equality exactly
when πj = k − 1 and πj +1 = k + 1 for some j . Let En−1,k be the set of permutations of
[n − 1] with a small ascent involving entries k − 1 and k and no other small ascents. We
obtain
⎧n−1
⎪
⎪ |Bn−1,j | if k = 1,
⎪
⎪
⎪
⎪
⎨j =2
|Bn,k | = |En−1,k | + |Bn−1,j | if 2 ≤ k ≤ n − 1,
⎪
⎪ j =k
⎪
⎪
⎪
⎪
n−1
⎩ |Bn−1,j | if k = n.
j =1

We now prove |En−1,k | = |Bn−1,k | when n ≥ 3, which reduces this expression to the
desired recurrence. Suppose σ ∈ En−1,k . Since σ has only one small ascent, the value k + 1
does not follow k in σ . Hence collapsing the pair (k − 1, k) of consecutive values to k − 1
and decreasing larger values by 1 gives a permutation of [n − 2] with no small ascent, and
n−2
the map is reversible. Hence |En−1,k | = |Bn−2,j |. We now have a proof of the desired
j =1
recurrence by induction on n, since the induction hypothesis yields |En−1,k | = |Bn−1,k |.
Editorial comment. The proposer constructed a bijection from An,k to Bn,k iteratively as
follows. If the current permutation has a small ascent, choose the left-most small ascent and
move the larger value j + 1 so that it immediately follows the largest right-left minimum m
that it exceeds. For example, π = (10, 11, 12, 2, 3, 1, 6, 7, 4, 8, 9, 5) has right-left minima
at values 5, 4, and 1 (no two consecutive), and it has small ascents ending in the values 11,
12, 3, 7, and 9. The first iteration moves 11 to immediately after 5 and the fourth and final
iteration yields (10, 12, 2, 1, 3, 6, 4, 8, 5, 7, 9, 11).
Yury Ionin observed that exchanging the values k and k + 1 in π ∈ An,k yields a bijec-
tion between An,k and An,k+1 for k > 1. This is implicit in the featured solution.
Also solved by K. Gatesman, A. Goel, Y. J. Ionin, and the proposer. Part (b) also solved by N. Hodges (UK).

Complete Elliptic Integrals and Watson’s Integrals

12232 [2021, 178]. Proposed by Seán Stewart, Bomaderry, Australia. Prove
 1 1  ∞ 4
1 1
√ √ √ dx dy = e−t t −3/4 dt .
0 0 x(1 − x) y(1 − y) 1 − xy 4π 0

Solution I by Tamas Wiandt, Rochester Institute of Technology, Rochester, NY. Let I denote
the integral on the left side of the desired equation. Substituting x = k 2 and y = sin2 t, we
get
 1  π/2  1
1 1 K(k) dk
I =4 √  dt dk = 4 √ , (1)
0 1−k 0 2
1 − k 2 sin t
2 0 1 − k2
where K(k) is the complete elliptic integral of the first kind given by the formula
 π/2
dt
K(k) =  .
0 1 − k 2 sin2 t

892 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
The last integral in (1) is given by equation 6.143 on page 632 of I. S. Gradshteyn and
I. M. Ryzhik (2007), Table of Integrals, Series, and Products, 7th ed., Burlington, MA:
Academic Press. Filling in its value, we obtain
  ∞ 4
 √
4
(1/4) 1
e−t t −3/4 dt .
2
I = 4 K( 2/2) = =
4π 4π 0

Solution II by Lixing Han, University of Michigan, Flint, MI, and Xinjia Tang, Changzhou
University, Changzhou, China. Let I be as in Solution I. Substituting x = cos2 u,
y = cos2 v, we get
 π/2  π/2  π π
du dv du dv
I =4 √ = √ . (2)
0 0 1 − cos u cos v
2 2 0 0 1 − cos2 u cos2 v
For |a| < 1, the substitution s = tan(t/2) yields
   ∞
π
dt 2 ∞
ds 2 −1 1 + a  π
= = √ tan s  =√ .
0 1 − a cos t 1−a 0 1+ 1+a
1−a
s2 1 − a2 1−a  1 − a 2
0

Setting a = cos u cos v leads to

 π
dt π
=√ .
0 1 − cos u cos v cos t 1 − cos2 u cos2 v
Substituting into (2), we obtain
  
1 π π π dt du dv
I= = π 2 I1 ,
π 0 0 0 1 − cos u cos v cos t
where I1 is one of Watson’s triple integrals (see I. J. Zucker (2011), 70+ years of the
Watson Integrals, J. Stat. Phys. 145: 591–612, inp.nsk.su/∼silagadz/Watson Integral.pdf).
Filling in the known value of I1 gives the desired result.
Also solved by U. Abel & V. Kushnirevych (Germany), A. Berkane (Algeria), N. Bhandari (India), P. Bracken,
H. Chen, B. E. Davis, G. Fera (Italy), M. L. Glasser, J.-P. Grivaux (France), J. A. Grzesik, N. Hodges (UK),
Z. Lin (China), O. P. Lossers (Netherlands), M. Omarjee (France), K. Sarma (India), A. Stadler (Switzerland),
A. Stenger, R. Stong, R. Tauraso (Italy), M. Vowe (Switzerland), M. Wildon (UK), and the proposer.

Squarefree Sums
12236 [2021, 179]. Proposed by Navid Safaei, Sharif  University of Technology, Tehran,
Iran. Let pk be the kth prime number, and let an = nk=1 pk . Prove that for n ∈ N every
positive integer less than an can be expressed as a sum of at most 2n distinct divisors of an .
Solution by Rory Molinari, Beverly Hills, MI. The divisors of an are exactly the positive
squarefree integers whose largest prime factor is no bigger than pn . We need the claim that
every positive integer r can be written as the sum of at most two distinct positive squarefree
integers.
It is easy to verify the claim for r ≤ 9, so assume r ≥ 10. Let A(r) be the set of positive
squarefree integers not greater than r. If r ∈ A(r), we are done. Otherwise, it is known
that |A(r)| ≥ 53r/88 for all r (see K. Rogers (1964), The Schnirelmann density of the
squarefree integers, Proc. Am. Math. Soc. 15(4): 515–516). Thus |A(r)| > 1 + r/2 for
r ≥ 10, and the pigeonhole principle implies that A(r) and {r − k : k ∈ A(r)} share at
least two elements. At least one of them is not r/2, yielding an expression of r as the sum
of two elements of A(r).

November 2022] PROBLEMS AND SOLUTIONS 893

 To prove the problem statement, we use induction on n. The claim holds trivially for
n = 1. For n > 1, consider m such that 1 ≤ m < an . Write m as q · pn + r with 0 ≤
q < an−1 and 0 ≤ r < pn . By the claim, r is the sum of at most two positive squarefree
numbers. These numbers cannot have pn as a factor since r < pn , so they are factors of
an−1 . By the induction hypothesis, q is the sum of at most 2(n − 1) distinct factors of an−1 .
Hence, q · pn + r is the sum of at most 2(n − 1) distinct divisors of an , all of which are
multiples of pn , plus at most two distinct divisors of an−1 . It follows that m is the sum of
at most 2n distinct divisors of an .
Editorial comment. The problem statement above corrects a typo in the original printing.
All solvers used similar proofs. Some used bounds such as
∞

|A(r)| ≥ r − r pk−2 > .54r
k=1

in the proof of the initial claim.

Also solved by O. Geupel (Germany), N. Hodges (UK), M. Hulse (India), Y. J. Ionin, O. P. Lossers (Nether-
lands), C. Schacht, A. Stadler (Switzerland), M. Tang, R. Tauraso (Italy), and the proposer.

CLASSICS
C9. From the 2001 Putnam Competition, suggested by the editors. Can an arc of a parabola
inside a circle of radius 1 have a length greater than 4?
Flipping Coins Until They are All Heads
C8. Due to Leonard Räde, suggested by the editors. Start with n fair coins. Flip all of them.
After this first flip, take all coins that show tails and flip them again. After the second flip,
take all coins that still show tails and flip them again. Repeat until all coins show heads.
Let qn be the probability that the last flip involved only a single coin. What is limn→∞ qn ?
Solution. Let L = 1/ ln 4. Rough computation suggests that qn converges to L, but we
show that qn oscillates around L with an asymptotic amplitude of about 10−5 , and so the
limit does not exist. Here at left we display the graph of qn for 1 ≤ n ≤ 20, illustrating
the apparent convergence. At right we graph the same sequence, zooming in and using a
logarithmic horizontal axis. That view reveals what appears to be a persistent asymptotic
oscillation.

To prove that the limit does not exist, take n ≥ 2, let C be one of the coins, and let k be a
positive integer. Consider the event that C shows heads for the first time on flip k + 1, and
all other coins show heads earlier. This occurs only if C shows tails for each of the first k
flips and then heads on flip k + 1. This has probability 2−(k+1) . For each of the other n − 1
coins, it must not be the case that all of the first k flips show tails. This has probability
1 − 2−k . So the probability of the event is 2−(k+1) (1 − 2−k )n−1 .

894 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
 Because there are n possibilities for C, and because k can be any positive integer,
∞ 
n 1 n−1
qn = 1 − . (∗)
k=1
2k+1 2k

We show that the sequence q1 , q2 , . . . does not converge by showing that it has different
subsequences that converge but to different limits.
k
Let ck = (1 − 2−k )2 . It is well known and easy to show that c1 , c2 , . . . is an increasing
sequence and limk→∞ ck = 1/e.
We have
∞
  k n/2k  ∞ 
 n 1 2 1 −1  n n/2k 2k
qn = 1 − 1 − = c .
k=1
2k+1 2k 2k k=1
2k+1 k 2k − 1

Now fix an odd integer m, and let aj = qm2j for j ≥ 1. We have

∞  ∞ 
m2j m2j /2k 2k m m/2k 2k+j
aj = c = ck+j .
k=1
2k+1 k
2 −1
k
k=1−j
2k+1 2k+j − 1

k
The kth term of this series is bounded above by (m/2k )e−m/2 , whose sum over k from
−∞ to ∞ is finite. Hence, by the dominated convergence theorem,
∞  ∞
m m/2k 2k+j m −m/2k
lim aj = lim k+1 ck+j = e .
j →∞
k=−∞
j →∞ 2 2k+j −1 k=−∞
2k+1

With m = 1, this last sum can be approximated by letting k run from −5 to 27, giving
an approximation of L + 4.58 · 10−6 for the sum, and the error in this approximation is
seen by a simple integration to be less than 10−8 . Similarly, when m = 3, the last sum is
approximately L − 1.17 · 10−6 , again with an error of less than 10−8 . The distinct limits
prove that limn→∞ qn does not exist.
Editorial comment. One can approximate the sum in (∗) by
 ∞
n2−(x+1) (1 − 2−x )n−1 dx,
0

which is L, independent of n. The error in this approximation does not vanish with n,
however.
The problem appeared in this Monthly [1991, 366; 1994, 78]. A version of the same
problem appeared almost a decade earlier in the 1982 Can. Math. Bull. as Problem P322
by George Szekeres, who asked whether
n 
i n
lim (−1)i−1 i
n→∞
i=1
2 −1 i

equals 1/ ln 2. It turns out that the nth term here is just 2qn in disguise, so the answer to the
Szekeres problem is negative.
In N. J. Calkin, E. R. Canfield, and H. S. Wilf (2000), Averaging sequences, deranged
mappings, and a problem of Lambert and Slater, J. Comb. Th., Ser. A 91(1–2): 171–190,
a general class of sequences is found to exhibit the oscillating sequence phenomenon. In
particular, they answer an open question in D. E. Lampert and P. J. Slater (1998), Parallel
knockouts in the complete graph, this Monthly 105: 556–558.

November 2022] PROBLEMS AND SOLUTIONS 895

SOLUTIONS

A Double Sum for Apéry’s Constant

12222 [2020, 945]. Proposed by Roberto Tauraso, Università di Roma “Tor Vergata,”
Rome, Italy. Prove
 ∞ ∞
(−1)k  1 13 ζ (3)
2 n
=− ,
k=1
k n=k
n2 24

where ζ (3) is Apéry’s constant ∞ 3
k=1 1/k .
Composite solution by Brian Bradie andHongwei Chen, Christopher Newport University,
Newport News, VA. In general, ζ (m) = ∞ m
k=1 1/k . In working with expressions involving
reciprocal powers, it is useful to have the gamma function integral and its logarithmic
version
 ∞  1
n! −kt n
= e t dt = (−1) n
x k−1 (ln x)n dx, (1)
k n+1 0 0

where the latter integral is obtained from the former by setting t = − ln x.

Let S be the desired double sum. After interchanging the order of summation, we invoke
(1) with n = 1 to obtain
∞ ∞  1
1  (−1)k  1 
n n
S= n 2
= n
(−1) k+1
x k−1 ln x dx
n=1
n2 k=1
k n=1
n2 k=1 0

∞   n ∞  1
 1 1   1 1 − (−x)n
= (−1) k+1 k−1
x ln x dx = ln x dx.
n=1
n
n2 0 k=1 n=1
n
n2 0 1+x

Because the integrand in this last expression is nonpositive for every x in [0, 1] and every
n, one can interchange the summation and integration to obtain
 1  1
− ln(1 − 1/2) + ln(1 + x/2) ln(2 + x) ln x
S= ln x dx = dx.
0 1+x 0 x+1
We break the integral for S into three integrals by applying the polarization identity
ab = 12 (a 2 + b2 − (a − b)2 ) to the numerator of the integrand, using a = ln x and b =
ln(2 + x). Letting

786 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
  1
(ln f (x))2
J (f (x)) = dx,
0 1+x
we obtain
2S = J (x) + J (x + 2) − J (x/(2 + x)). (2)
Expanding 1/(1 + x) into a geometric series and applying (1) with n = 2 yields
∞  1 ∞
(−1)k
J (x) = (−1)k x k (ln x)2 dx = 2 .
k=0 0 k=0
(k + 1)3

To evaluate J (x + 2), we substitute t = 1/(x + 2). Since 1/(x + 1) = t/(1 − t), we

obtain dx/(1 + x) = dt/(t (t − 1)). Using partial fraction expansion and then another geo-
metric series,
 1/2 ∞ 
1 1 (ln 3)3 − (ln 2)3  1/2 k
J (x + 2) = + (ln t)2 dt = + t (ln t)2 dt.
1/3 t 1 − t 3 k=0 1/3

Integrating by parts twice yields

 1/2 1/2
(ln t)2 2 ln t 2
t k (ln t)2 dt = t k+1 − + . (3)
1/3 (k + 1) (k + 1)2 (k + 1)3 1/3

Summing over k, we now have J (x + 2) expressed

 in terms of polylogarithms, where the
polylogarithm Lis (z) is defined by Lis (z) = ∞ k=1 z /k . Note that Li1 (z) = − ln(1 − z).
k s

The function Li2 is called the dilogarithm, and Li3 is called the trilogarithm. In particular,
J (x) = −2 Li3 (−1) and
∞
(ln 3)3 − (ln 2)3  (ln(1/2))2 2 ln(1/2) 2
J (x + 2) = + (1/2)k − + 3
3 k=1
k k2 k
∞
 (ln(1/3))2 2 ln(1/3) 2
− (1/3)k − + 3
k=1
k k2 k

(ln 3)3 − (ln 2)3

= + (ln 2)2 Li1 (1/2) + 2 ln 2 Li2 (1/2) + 2 Li3 (1/2)
3
− (ln 3)2 Li1 (1/3) − 2 ln 3 Li2 (1/3) − 2 Li3 (1/3)
(ln 2)3 − (ln 3)3
= + 2 ln 2 Li2 (1/2) + 2 Li3 (1/2)
3/2
− 2 ln 3 Li2 (1/3) − 2 Li3 (1/3) + (ln 3)2 ln 2,
where the last step uses Li1 (z) = − ln(1 − z).
To evaluate J (x/(2 + x)), we substitute t = x/(2 + x), which yields x = 2t/(1 − t),
1 + x = (1 + t)/(1 − t), dx = 2 dt/(1 − t)2 , and dx/(1 + x) = 2 dt/(1 − t 2 ). Integrating
as we did in (3) after expanding a geometric sum yields
 1/3
1
J (x/(2 + x)) = 2 (ln t)2 dt
0 1 − t 2

∞ 2k+1
1 (ln 3)2 2 ln 3 2
=2 + + .
k=0
3 2k + 1 (2k + 1)2 (2k + 1)3

October 2022] PROBLEMS AND SOLUTIONS 787

  
The odd terms in a Taylor series T (x) at 0 sum to T (x) − T (−x) /2, so
   
J (x/(2 + x)) = (ln 3)2 ln 2 + 2 ln 3 Li2 (1/3) − Li2 (−1/3) + 2 Li3 (1/3) − Li3 (−1/3) .
Substituting these expressions for J (x), J (x + 2), and J (x/(2 + x)) into (2) and com-
bining like terms yields
(ln 2)3 − (ln 3)3    
S= − ln 3 2 Li2 (1/3) − Li2 (−1/3) − 2 Li3 (1/3) − Li3 (−1/3)
3
+ ln 2 Li2 (1/2) − Li3 (−1) + Li3 (1/2).
The following are known evaluations of dilogarithms and trilogarithms at −1, 1/2, and
±1/3:
3
Li3 (−1) = − ζ (3)
4
π2 (ln 2)2
Li2 (1/2) = −
12 2
−π 2 ln 2 (ln 2)3 7
Li3 (1/2) = + + ζ (3)
12 6 8
π2 (ln 3)2
2 Li2 (1/3) − Li2 (−1/3) = −
6 2
π 2 ln 3 (ln 3)3 13
2 Li3 (1/3) − Li3 (−1/3) = − + + ζ (3).
6 6 6
After substituting these evaluations into the last expression for S, remarkably all terms not
involving ζ (3) cancel, leaving
3 7 13 13
S= ζ (3) + ζ (3) − ζ (3) = − ζ (3).
4 8 6 24

Editorial comment. The generation of many terms not involving ζ (3), which then can-
cel, suggests that there should be a shorter solution not involving polylogarithms, but no
solver was able to contribute such a solution. Some solvers replaced the original 2 by 1/x,
differentiated, summed, integrated, and thereby reduced the desired sum to
 1/2
Li2 (−x)
dx.
0 x(1 − x)
However, this also does not seem to lead to a shorter solution.
A standard reference for polylogarithms and their evaluations is L. Lewin (1981), Poly-
logarithms and Associated Functions, Amsterdam: North-Holland. For further examples
of series summing to ζ (3) and historical background, see A. van der Poorten (1979), A
proof that Euler missed, Math. Intelligencer 1: 195–203, and W. Dunham (2021), Euler
and the cubic Basel problem, this Monthly 128: 291–301.
Also solved by N. Bhandari (Nepal), R. Boukharfane (Morocco), G. Fera (Italy), M. L. Glasser, P. W. Lind-
strom, M. Omarjee (France), A. Stadler (Switzerland), S. M. Stewart (Australia), R. Stong, and the proposer.

Collinear Intersection Points

12224 [2021, 88]. Proposed by Cherng-tiao Perng, Norfolk State University, Norfolk, VA.
Let ABC be a triangle, with D and E on AB and AC, respectively. For a point F in the
plane, let DF intersect BC at G and let EF intersect BC at H . Furthermore, let AF

788 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
intersect BC at I , let DH intersect EG at J , and let BE intersect CD at K. Prove that I ,
J , and K are collinear.
Solution I by Nigel Hodges, Cheltenham, UK. We use XY.ZW to denote the intersection
of lines XY and ZW . Let L = AG.DI , M = AH.EI , and N = BC.DE. Lines EH ,
AI , and GD concur at F . Therefore, by the theorem of Desargues, the points EA.H I ,
EG.H D, and AG.I D are collinear. Since E lies on AC, and since H and I lie on BC, we
have EA.H I = C, and by definition, EG.H D = J and AG.I D = L. Thus, we have

C, J , and L are collinear. (1)

Similarly, applying the theorem of Desargues to EH , I A, and GD we conclude that

M, J , and B are collinear, (2)

and using EH , I A, and DG we get

M, N , and L are collinear. (3)

Statement (3) implies that lines LM, DE, and CB concur at N , so by one more applica-
tion of the theorem of Desargues we conclude that LD.ME, LC.MB, and DC.EB are
collinear. But L lies on DI and M lies on EI , so LD.ME = I , (1) and (2) imply that
LC.MB = J , and DC.EB = K by definition. Thus I , J , and K are collinear.
Solution II by O. P. Lossers, Eindhoven University of Technology, Eindhoven, Netherlands.
We use homogeneous coordinates with A = (1 : 0 : 0), B = (0 : 1 : 0), C = (0 : 0 : 1),
and K = (1 : 1 : 1). This gives D = (1 : 1 : 0) and E = (1 : 0 : 1). Let F = (a : b : c).
Since G lies on BC and DF , we have G = (0 : b − a : c). Similarly,

H = (0 : b : c − a), I = (0 : b : c), and J = (a : a − b : a − c),

so it follows that I , J , and K are collinear.

Also solved by M. Bataille (France), J. Cade, C. Curtis, I. Dimitrić, G. Fera (Italy), R. Frank (Germany),
O. Geupel (Germany), J.-P. Grivaux (France), E. A. Herman, W. Janous (Austria), J. H. Lindsey II, C. R. Prane-
sachar (India), C. Schacht, V. Schindler (Germany), A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy),
T. Wiandt, L. Zhou, Davis Problem Solving Group, The Zurich Logic-Coffee (Switzerland), and the proposer.

Gamma at Reciprocals of Positive Integers

12225 [2021, 88]. Proposed by Pakawut Jiradilok, Massachusetts Institute of Technology,
Cambridge, MA, and Wijit Yangjit, University of Michigan, Ann Arbor, MI. Let  denote
∞
the gamma function, defined by (x) = 0 e−t t x−1 dt for x > 0.
(a) Prove that  (1/n) = n for every positive integer n, where y denotes the smallest
integer greater than or equal to y.
(b) Find the smallest constant c such that  (1/n) ≥ n − c for every positive integer n.
Solution by Missouri State University Problem Solving Group, Springfield, MO. We use
three facts about the gamma function: (i) (x + 1) = x(x), (ii)  (1) = −γ , where γ is
the Euler–Mascheroni constant, and (iii) the gamma function is convex on (0, ∞).
(a) The equation of the line tangent to y = (x + 1) at the point (0, 1) is
y = 1 +  (1)x = 1 − γ x.

Since the gamma function is convex, this implies that for x > −1,

(x + 1) ≥ 1 − γ x.

October 2022] PROBLEMS AND SOLUTIONS 789

Applying this with x = 1/n yields
(1/n) = n(1/n + 1) ≥ n(1 − γ /n) = n − γ .
Also, since (1) = (2) = 1, by convexity (x + 1) ≤ 1 for 0 ≤ x ≤ 1. Hence
(1/n) = n(1/n + 1) ≤ n.
Since n − γ ≤ (1/n) ≤ n and γ < 1, we conclude that (1/n) = n.
(b) The solution to part (a) shows that γ satisfies the required condition. Now let c be any
constant such that (1/n) ≥ n − c for all n. We have
(1 + 1/n) − 1
c ≥ n − (1/n) = n − n(1/n + 1) = − .
1/n
Letting n approach ∞ yields
(1 + 1/n) − 1
c ≥ lim − = − (1) = γ .
n→∞ 1/n
Thus, γ is the smallest such c.
Also solved by R. A. Agnew, K. F. Andersen (Canada), P. Bracken, H. Chen, G. Fera (Italy), D. Fleischman,
J.-P. Grivaux (France), J. A. Grzesik (Canada), L. Han, N. Hodges (UK), O. Kouba (Syria), O. P. Lossers
(Netherlands), I. Manzur (UK) & M. Graczyk (France), R. Molinari, M. Omarjee (France), A. Stadler (Switzer-
land), R. Stong, R. Tauraso (Italy), J. Vinuesa (Spain), M. Vowe (Switzerland), T. Wiandt, J. Yan (China),
L. Zhou, and the proposer.

A Recursive Sequence That Is Convergent or Eventually Periodic

12226 [2021, 88]. Proposed by Jovan Vukmirovic, Belgrade, Serbia. Let x1 , x2 , and x3 be
real numbers, and define xn for n ≥ 4 recursively by xn = max{xn−3 , xn−1 } − xn−2 . Show
that the sequence x1 , x2 , . . . is either convergent or eventually periodic, and find all triples
(x1 , x2 , x3 ) for which it is convergent.
Solution by O. P. Lossers, Eindhoven University of Technology, Eindhoven, Netherlands.
Let λ1 be the unique real root of λ3 + λ − 1, so
 √ 1/3  √ 1/3
9 + 93 9 − 93
λ1 = + = 0.682327803828 . . . .
18 18

The sequence converges if and only if (x1 , x2 , x3 ) = (x1 , x1 λ1 , x1 λ21 ) with x1 > 0 or
(x1 , x2 , x3 ) = (x1 , 0, 0) with x1 ≤ 0. Otherwise, it is eventually periodic with period 4.
Given such a sequence x1 , x2 , . . . , let i ∈ N be of type A if xi ≤ xi+2 and type B if
xi > xi+2 . We claim that if i is of type A and i + 1 is of type B, then xj = xj +4 for
j ≥ i + 3. To see this, let (a, b, c) = (xi , xi+1 , xi+2 ). We have a ≤ c and xi+3 = c − b, so
b > c − b and xi+4 = b − c.
If c ≤ b − c, which with b > c − b implies b > c, then the sequence continues
xi+5 = 2b − 2c, xi+6 = b − c, xi+7 = c − b, xi+8 = b − c, xi+9 = 2b − 2c.
With (xi+7 , xi+8 , xi+9 ) = (xi+3 , xi+4 , xi+5 ), the claim follows. If c > b − c, then the
sequence continues
xi+5 = b, xi+6 = c, xi+7 = c − b,
yielding (xi+5 , xi+6 , xi+7 ) = (xi+1 , xi+2 , xi+3 ). In both cases, the sequence has period 4
beginning no later than xi+3 and hence does not converge.

790 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
 If i of type A is never followed by i + 1 of type B, then either all i are of type B or there
exists some integer k ≥ 1 such that i is of type A if and only if i ≥ k. If all i are of type
B, then xn = −xn−2 + xn−3 for n ≥ 4. The characteristic polynomial λ3 + λ − 1 is strictly
increasing with unique real root λ1 between 0 and 1. The complex conjugate roots λ2 and
λ3 have magnitude greater than 1.
It follows that xn = c1 λn1 + (c2 λn2 ) for some real c1 and complex c2 , where (z)
denotes the real part of z. Since |λ2 | > 1 and xn−3 > xn−1 for n ≥ 4, we conclude c2 = 0
and therefore xn = c1 λn1 , where c1 > 0 to satisfy xn > xn+2 . This is a strictly decreasing
convergent solution, not eventually periodic.
Finally, if i is of type A if and only if i ≥ k, then xk+1 , xk+2 , . . . satisfies xn = xn−1 −
xn−2 for n ≥ k + 3. Therefore,
xk+3 = xk+2 − xk+1 ≥ xk+1 ,
xk+4 = −xk+1 ≥ xk+2 ,
xk+5 = −xk+2 ,
xk+6 = xk+1 − xk+2 ≥ −xk+1 ,
xk+7 = xk+1 ≥ −xk+2 .
From −xk+1 ≥ xk+2 and xk+1 ≥ −xk+2 we conclude xi = 0 for i ≥ k + 1. Since k is of
Type A, also xk ≤ 0. If k > 1, then xk+2 = xk−1 − xk > xk+1 − xk = −xk ≥ 0, which
contradicts xk+2 = 0. Therefore, k must equal 1, and the convergent sequences that are
also eventually periodic are given by (x1 , x2 , x3 ) = (x1 , 0, 0) with x1 ≤ 0.
Also solved by C. Curtis & J. Boswell, G. Fera (Italy), N. Hodges (UK), Y. J. Ionin, P. Lalonde (Canada),
M. Reid, R. Stong, L. Zhou, and the proposer.

Sum of Reciprocals of Consecutive Integers

12227 [2021, 88]. Proposed by Gregory Galperin, Eastern Illinois University, Charleston,
IL, and Yury J. Ionin, Central Michigan University, Mount Pleasant, MI. Prove that for
any integer n with n ≥ 3 there exist infinitely many pairs (A, B) such that A is a set of
n consecutivepositive integers,
 B is a set of fewer than n positive integers, A and B are
disjoint, and k∈A 1/k = k∈B 1/k.
Solution by Rory Molinari, Beverly Hills, MI. For positive integers t and n, let

{t − m, t − m + 1, . . . , t + m} if n = 2m + 1,
An (t) =
{t − m, t − m + 1, . . . , t + m − 1} if n = 2m,

where m is an integer. For a set X of nonzero numbers, let S(X) = i∈X 1/ i.
First consider the odd case: n = 2m + 1 ≥ 3. Fix a positive integer p. Using 1/(np) =
1/p − (n − 1)/(np), we compute
n−1 
m
1 1 1
S(An (np)) = − + +
p np i=1
np − i np + i

1 
m
1 1 2
= + + −
p i=1
np − i np + i np

1  1 
m m
2i 2 1
= + = + ,
p i=1
np(n p − i )
2 2 2 p i=1
b(np, i)

where b(x, y) = x(x 2 − y 2 )/(2y 2 ). If we choose p to be a multiple of 2m!, then b(np, i)

is an integer for 1 ≤ i ≤ m. By taking A = An (np) and B = {p, b(np, 1), . . . , b(np, m)},

October 2022] PROBLEMS AND SOLUTIONS 791

we see that B is a set of fewer than n distinct positive integers and S(A) = S(B). Since
b(np, i) = (p3 ), the sets A and B are disjoint for sufficiently large p.
The case n = 2m is similar. We compute

n−1 
m−1
1 1 1 1
S(An (np)) = + − + +
p np − m np i=1
np − i np −i

1 1 1 
m−1
1 1 2
= + − + + −
p np − m np i=1
np − i np − i np

1 1 
m−1
1
= + + .
p np(2p − 1) i=1
b(np, i)

Setting A = Ap (np) and B = {p, np(2p − 1), b(np, 1), . . . , b(np, m − 1)} suffices when
we take p to be a sufficiently large multiple of 2(m − 1)!.
Also solved by C. Curtis & J. Boswell, K. Gatesman, J.-P. Grivaux (France), N. Hodges (UK), P. Lalonde
(Canada), O. P. Lossers (Netherlands), I. Manzur (UK) & M. Graczyk (France), A. Pathak (India), C. R. Prane-
sachar (India), M. Reid, E. Schmeichel, A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), M. Tetiva
(Romania), L. Zhou, Missouri State University Problem Solving Group, and the proposers.

An Integral for Catalan Squared

12228 [2021, 89]. Proposed by Hervé Grandmontagne, Paris, France. Prove
 1  √ 
(ln x)2 ln 2 x/(x 2 + 1)
dx = 2G2 ,
0 x2 − 1

where G is Catalan’s constant ∞ n=0 (−1) /(2n + 1) .
n 2

Solution by Li Zhou, Polk State College, Winter Haven, FL. It is well known that 2G =
∞
0 (x/ cosh x) dx. (See, e.g., I. S. Gradshteyn and I. M. Ryzhik (2015), Table of Integrals,
Series, and Products, 8th ed., Waltham, MA: Academic Press, equation 3.521(2).) There-
fore, using the change of variables u = x + y, v = x − y, we have
 
1 ∞ ∞ xy
2G2 = dx dy
2 0 0 cosh x cosh y
 
1 ∞ ∞ (x + y)2 − (x − y)2
= dx dy
4 0 0 cosh(x + y) + cosh(x − y)
 ∞ u  
1 u2 − v 2 1 ∞ u u2 − v 2
= dv du = dv du
8 0 −u cosh u + cosh v 4 0 0 cosh u + cosh v
 ∞  u  ∞  ∞ 
1 1 1
= u 2
dv du − v 2
du dv
4 0 0 cosh u + cosh v 0 v cosh u + cosh v
 ∞  u  ∞ 
1 dv dv
= u2 − du.
4 0 0 cosh u + cosh v u cosh u + cosh v
To evaluate the inner integrals, we use
 
dv tanh((u + v)/2) + tanh((u − v)/2)
= dv
cosh u + cosh v 2 sinh u
1 cosh((u + v)/2)
= ln + C,
sinh u cosh((u − v)/2)

792 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
which yields
 u  ∞
dv ln cosh u dv u − ln cosh u
= and = .
0 cosh u + cosh v sinh u u cosh u + cosh v sinh u
Hence
 ∞   √ 
1 u2 (2 ln cosh u − u) 1 (ln x)2 ln 2 x/(x 2 + 1)
2G2 = du = dx,
4 0 sinh u 0 x2 − 1
where the last equality follows from the substitution u = − ln x.
Also solved by F. R. Ataev (Uzbekistan), A. Berkane (Algeria), N. Bhandari (Nepal), H. Chen, G. Fera (Italy),
M. L. Glasser, D. Henderson, N. Hodges (UK), O. Kouba (Syria), A. Stadler (Switzerland), S. M. Stewart
(Australia), R. Stong, R. Tauraso (Italy), M. Wildon (UK), and the proposer.

A Sum of Secants from a Triangle

12231 [2021, 178]. Proposed by George Apostolopoulos, Messolonghi, Greece. For an
acute triangle ABC with circumradius R and inradius r, prove
A−B B −C C−A R
sec + sec + sec ≤ + 1.
2 2 2 r

Solution by UM6P Math Club, Mohammed VI Polytechnic University, Ben Guerir, Morocco.
 2
Since cos((B − C)/2) − 2 sin(A/2) ≥ 0, we have
B −C B −C A A
cos2 ≥ 4 cos sin − 4 sin2 .
2 2 2 2
Using the well-known formula 4 sin(A/2) sin(B/2) sin(C/2) = r/R, we obtain
B −C A A A B −C A
4 cos sin − 4 sin2 = 4 sin cos − sin
2 2 2 2 2 2
A B −C B +C
= 4 sin cos − cos
2 2 2
A B C 2r
= 8 sin sin sin = .
2 2 2 R
  √
Thus sec (B − C)/2 ≤ R/(2r). Similarly
     
sec (A − B)/2 ≤ R/(2r) and sec (C − A)/2 ≤ R/(2r),
and summing these inequalities yields

A−B B −C C−A R
sec + sec + sec ≤3 .
2 2 2 2r
To complete the proof, it suffices to show

R R
3 ≤ + 1.
2r r
√
Setting t = R/(2r), the required inequality becomes 3t ≤ 2t 2 + 1, or (2t − 1)(t − 1) ≥ 0.
This holds because t ≥ 1, by Euler’s inequality R ≥ 2r.
Editorial comment. The assumption that the triangle is acute is not needed.

October 2022] PROBLEMS AND SOLUTIONS 793

Also solved by M. Bataille (France), H. Chen (China), C. Chiser (Romania), C. Curtis, N. S. Dasireddy (India),
P. De (India), H. Y. Far, G. Fera (Italy), O. Geupel (Germany), N. Hodges (UK), W. Janous (Austria), K.-W. Lau
(China), M. Lukarevski (North Macedonia), C. R. Pranesachar (India), V. Schindler (Germany), A. Stadler
(Switzerland), R. Stong, R. Tauraso (Italy), M. Tetiva (Romania), M. Vowe (Switzerland), T. Wiandt, L. Wim-
mer (Germany), and the proposer.

CLASSICS
C8. (Due to Leonard Räde, suggested by the editors). Start with n fair coins. Flip all of
them. After this first flip, take all coins that show tails and flip them again. After the second
flip, take all coins that still show tails and flip them again. Repeat until all coins show heads.
Let qn be the probability that the last flip involved only a single coin. What is limn→∞ qn ?

Are R and C Isomorphic Under Addition?

C7. Contributed by Alan D. Taylor, Union College, Schenectady, NY. Are the additive group
of real numbers and the additive group of complex numbers isomorphic?
Solution. Each of the given groups is a vector space over the set Q of rational numbers.
Because every vector space has a basis, we can let B1 be a basis for R and B2 a basis
for C. Because Q is countable while R is not, in order for B1 to span R, the cardinality
of B1 must equal the cardinality of R. The same holds for C. Because R and C have the
same cardinality, there is a bijection f : B1 → B2 . The bijection can be extended to an
isomorphism
n of the groups as follows: for each x ∈ R write x (uniquely)
n as a finite sum
i=1 qi bi , where qi ∈ Q \ {0} and bi ∈ B1 and define f (x) to be i=1 qi f (bi ). It is easy
to verify that f (x + y) = f (x) + f (y), so f is a group isomorphism.
Editorial comment. The result of the problem is folklore. The theorem that every vector
space has a basis relies on the axiom of choice (denoted AC). A simple proof uses Zorn’s
lemma to show that there is a maximal linearly independent set of vectors; such a set
must be a basis. It is well known that Zorn’s lemma is equivalent to AC. It turns out
that the statement that every vector space has a basis is also equivalent to AC (A. Blass
(1984), Existence of bases implies the axiom of choice, Contemp. Math. 31, 31–33). The
question therefore arises whether the existence of an isomorphism from R to C can be
proved without using AC. We sketch a proof that it cannot.
A set of reals has the property of Baire if it differs from an open set by a meager set
(i.e., a countable union of nowhere dense sets), and a function has the property of Baire if
the inverse image of any open set has the property of Baire (so it is “almost continuous”).
Let ZF be the axiomatic theory whose axioms are the Zermelo–Fraenkel axioms (AC not
included). Let PB be the assertion that “all sets of reals have the property of Baire” and
let ZF + PB be the theory in which PB is added to ZF as an additional axiom. The theory
ZF + PB is known to be consistent, assuming ZF is consistent (S. Shelah (1984), Can you
take Solovay’s inaccessible away?, Isr. J. Math. 48, 1–47).
We now show that, in ZF + PB, the additive groups (R, +) and (C, +) are not iso-
morphic. An involution of a group is an automorphism of order 2. The complex numbers
admit at least two involutions: z → −z and z → z̄. Any automorphism f of R satisfies
f (x + y) = f (x) + f (y), and it is a classic result (W. Sierpiński (1924), Sur un propriété
des fonctions de M. Hamel, Fund. Math. 5, 334–336) that any function with the property
of Baire that satisfies this functional equation has the form x → cx for some real c. There-
fore, by PB, the only involution of R is x → −x. Because C has more than one involution,
C cannot be isomorphic to R.

794 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
SOLUTIONS

A Common Coefficient
12209 [2020, 852]. Proposed by Li Zhou, Polk State College, Winter Haven, FL. Prove
 n     n   
k n m + 2n − 2k + 1 n m+k+1
(−1) =
k=0
k m k=0
k m−n
for all integers m and n with m ≥ n ≥ 0.
Solution by Michel Bataille, France. We show that both sides equal the coefficient of x m in
the polynomial P defined by
P (x) = (1 + x)m+1 (2x + x 2 )n = (1 + x)m+1 ((1 + x)2 − 1)n .
Using the binomial theorem twice yields
n   n  
n n
P (x) = (1 + x)m+1 (−1)k (1 + x)2(n−k) = (−1)k (1 + x)2n−2k+m+1
k=0
k k=0
k

n   2n−2k+m+1
 2n − 2k + m + 1
k n
= (−1) xj .
k=0
k j =0
j

This expresses the left side of the identity as the coefficient of x m in the expansion of P (x).
Also,
P (x) = (1 + x)m+1 (x(2 + x))n = x n (1 + x)m+1 (1 + (1 + x))n ,
so another two uses of the binomial theorem yield
n    n   m+k+1  
n n  m + k + 1 n+j
P (x) = x (1 + x)
n m+1
(1 + x) =
k
x .
k=0
k k=0
k j =0 j

686 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
This shows that the coefficient of x m in the expansion of P (x) is also the right side of the
identity, completing the proof.
Also solved by R. Boukharfane (Saudi Arabia), Ó. Ciaurri (Spain), J. Boswell & C. Curtis, G. Fera (Italy),
N. Hodges (UK), M. Kaplan & M. Goldenberg, O. Kouba (Syria), P. Lalonde (Canada), O. P. Lossers
(Netherlands), M. Maltenfort, E. Schmeichel, A. Stadler (Switzerland), R. Stong, F. A. Velandia (Colombia),
M. Vowe (Switzerland), J. Vukmirović (Serbia), J. Wangshinghin, M. Wildon (UK), X. Ye (China), and the
proposer.

A Median Inequality
12214 [2020, 853]. Proposed by George Apostolopoulos, Messolonghi, Greece. Let x, y,
and z be the lengths of the medians of a triangle with area F . Prove
xyz(x + y + z) √
≥ 3F.
xy + zx + yz

Solution by Oliver Geupel, Brühl, Germany. The Cauchy–Schwarz inequality implies that
x 2 + y 2 + z2 ≥ xy + yz + zx, and therefore
(x + y + z)2 = x 2 + y 2 + z2 + 2(xy + yz + zx) ≥ 3(xy + yz + zx). (1)
It is well known that the medians of a triangle with area F are the sides of a triangle with
area K = 3F /4 (see, for example, sections 91–93 in N. Altschiller-Court (1952), College
Geometry, New York: Barnes and Noble). Moreover, it is known that a triangle with sides
x, y, and z and area K satisfies the inequality
9xyz √
≥ 4 3K (2)
x+y+z
(see item 4.13 on p. 45 of O. Bottema et al. (1969), Geometric Inequalities, Groningen:
Wolters-Noordhoff). Combining (1) and (2), we obtain
√
xyz(x + y + z) 3xyz(x + y + z) 3xyz 4 3K √
≥ = ≥ = 3F.
xy + yz + zx (x + y + z) 2 x+y+z 3

Editorial comment. Inequality (2) appeared as part of elementary problem E1861 [1966,
199; 1967, 724] from this Monthly, proposed by T. R. Curry and solved by Leon Bankoff.
The equation K = 3F /4 is also featured as Theorem 10.4 on p. 165 of C. Alsina and
R. B. Nelsen (2010), Charming Proofs: A Journey Into Elegant Mathematics, Washington,
DC: Mathematical Association of America.
Also solved by A. Alt, H. Bai (Canada), M. Bataille (France), E. Bojaxhiu (Albania) & E. Hysnelaj (Australia),
I. Borosh, R. Boukharfane (Saudi Arabia), P. Bracken, S. H. Brown, C. Curtis, N. S. Dasireddy (India), A. Dixit
(India) & S. Pathak (UK), H. Y. Far, G. Fera (Italy), N. Hodges (UK), W. Janous (Austria), M. Kaplan &
M. Goldenberg, P. Khalili, O. Kouba (Syria), K.-W. Lau (China), O. P. Lossers (Netherlands), M. Lukarevski
(Macedonia), A. Pathak (India), C. R. Pranesachar (India), C. Schacht, V. Schindler (Germany), A. Stadler
(Switzerland), N. Stanciu & M. Drăgan (Romania), R. Stong, B. Suceavă, M. Vowe (Switzerland), J. Vuk-
miroviıc (Serbia), T. Wiandt, X. Ye (China), M. R. Yegan (Iran), Davis Problem Solving Group, and the pro-
poser.

Another Incenter-Centroid Inequality

12217 [2020, 944]. Proposed by Giuseppe Fera, Vicenza, Italy. Let I be the incenter and
G be the centroid of a triangle ABC. Prove
3 AI BI CI
< + + ≤ 3.
2 AG BG CG

August–September 2022] PROBLEMS AND SOLUTIONS 687

Solution by Haoran Chen, Suzhou, China. Let a = BC, b = CA, and c = AB. Also let
s = (a + b + c)/2. Let ma be the length of the median from A, r the radius of the incircle,
and K the point of tangency of the incircle with AB. By the triangle inequality,
a  a 
2ma < +b + + c = 2s.
2 2
Also, AG = 2ma /3 and AI > AK = s − a. Therefore
AI 3AI 3(s − a)
= > .
AG 2ma 2s
Summing this with the other two analogous inequalities establishes the strict lower bound
of 3/2.
For the upper bound, note that
bc sin A
rs = area of ABC = ,
2
and therefore
AK r (s − a)r bc(s − a)
AI 2 = · = = .
cos(A/2) sin(A/2) (1/2) sin A s
Also, by Apollonius’s theorem,
4m2a = 2b2 + 2c2 − a 2 = (b + c + a)(b + c − a) + (b − c)2 ≥ 4s(s − a).
Therefore
√
AI 3AI 3 bc 3(b + c)
= ≤ ≤ .
AG 2ma 2s 4s
Summing this with the other two analogous inequalities establishes the upper bound of 3.
Editorial comment. Problem 12175 [2020, 372; 2021, 952] establishes
AI 2 BI 2 CI 2
2
+ 2
+ ≤ 3.
AG BG CG2
This can be used to give an alternative proof of the upper bound: By the Cauchy–Schwarz
inequality,
 
AI BI CI AI 2 BI 2 CI 2
+ + ≤ 3 + + ≤ 3.
AG BG CG AG2 BG2 CG2

Also solved by A. Alt, S. Gayen (India), P. Khalili, S. Lee (Korea), C. R. Pranesachar (India), A. Stadler
(Switzerland), R. Stong, R. Tauraso (Italy), T. Wiandt, and the proposer.

Composing All Permutations of [n] to Do Nothing

12218 [2020, 944]. Proposed by Richard Stong, Center for Communications Research, La
Jolla, CA, and Stan Wagon, Macalester College, St. Paul, MN. For which positive integers
n does there exist an ordering of all permutations of {1, . . . , n} so that their composition in
that order is the identity?
Solution by S. M. Gagola Jr., Kent State University, Kent, OH. Such an ordering of permu-
tations is possible for n = 1 (trivially) and for all n at least 4.
When n is 2 or 3, the number of permutations with odd parity is odd, so no composition
in these cases can have even parity like the identity. Note, however, that when n = 3 the
product of the three distinct transpositions always equals the middle factor (t1 t2 t3 = t2 ).

688 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
 Before considering n ≥ 4, it is useful to note that any group of even order has an odd
number of elements of order 2. To see this, pair the elements of the group with their
inverses. The identity element and the elements of order two (involutions) are self-paired,
while the remaining elements form sets of size 2. Since the group has even order, the num-
ber of involutions is therefore odd.
If in a group of even order a product of the involutions (in some order) can be shown
to equal the identity, then the remaining elements can be paired with their inverses to
yield a product of all the elements equaling the identity. Hence it suffices to show that for
n ≥ 4, the involutions of the symmetric group Sn can be ordered so that their product is the
identity.
The nine involutions in S4 can be partitioned into three triples as follows:
{(12), (34), (12)(34)}, {(13), (24), (13)(24)}, {(14), (23), (14)(23)}.
The product of the three involutions in any one subset (in any order) equals the identity;
this completes the n = 4 case.
For n = 5, we partition the involutions in S5 into sets I1 , . . . , I5 and order each set to
obtain a product yielding the identity. For I1 we take the nine involutions on {2, 3, 4, 5}.
By the n = 4 case, there is a product of these yielding the identity. For j ≥ 2, let Ij con-
sist of all involutions that exchange 1 and j . One element is (1j ), and each of the other
three elements is the product of (1j ) and a transposition of two of the three elements of
{2, 3, 4, 5} − {j }. Each of the four elements of Ij transposes 1 and j , and we have noted
that the product of the three transpositions on a set of size 3 can be ordered to yield any
one of the three transpositions. We can therefore choose orderings of each of I2 , I3 , I4 ,
and I5 so that their products are (45), (45), (23), and (23), respectively. Combining these
orderings completes the n = 5 case.
The solutions for n = 4 and n = 5 provide a basis for a proof by induction. We write [n]
for {1, . . . , n}. For n ≥ 6, partition the involutions of Sn into the n sets I1 , . . . , In , where
I1 consists of all the involutions on [n] − {1}, and Ij for j ≥ 2 consists of all involutions
exchanging 1 and j . The n − 1 case yields an ordering of I1 that produces the identity.
For j ≥ 2, each element of Ij consists of the transposition (1j ) times an element of the
symmetric group on [n] − {1, j } that is the identity or an involution. As noted earlier, Ij
thus has even size, and hence any product of the elements of Ij leaves 1 and j in place.
Furthermore, the n − 2 case guarantees that the elements of Ij other than (1j ) can be
ordered so that their effect on [n] − {1, j } is the identity. Doing this independently for all
Ij completes the proof.
Editorial comment. The problem is a special case of a result from J. Dénes and P. Hermann
(1982), On the product of all elements in a finite group, in E. Mendelsohn, ed., Algebraic
and geometric combinatorics, North-Holland Math. Stud. 65, Amsterdam: North-Holland,
pp. 105–109. A special case of their theorem that still includes the problem here is proved
more simply in M. Vaughan-Lee and I. M. Wanless (2003), Latin squares and the Hall–
Paige conjecture. Bull. London Math. Soc. 35, no. 2, 191–195.
The solver Gagola noted that if a group G of even order has a cyclic Sylow 2-subgroup,
then there is a normal 2-complement N, and the product of the elements of G taken in
any order always represents a coset of order 2 in the factor group G/N. Therefore, this
product can never equal the identity element. He then asked whether a group of even order
that does not have a cyclic Sylow 2-subgroup always has an ordering of the elements so
that the resulting product produces the identity. As Vaughan-Lee and Wanless wrote, “The
Hall–Paige conjecture deals with conditions under which a finite group G will possess a
complete mapping, or equivalently a Latin square based on the Cayley table of G will

August–September 2022] PROBLEMS AND SOLUTIONS 689

possess a transversal. Two necessary conditions are known to be: (i) that the Sylow 2-
subgroups of G are trivial or noncyclic, and (ii) that there is some ordering of the elements
of G which yields a trivial product. These two conditions are known to be equivalent, but
the first direct, elementary proof that (i) implies (ii) is given here.” Thus the answer to
Gagola’s question is yes.
Also solved by F. Chamizo & Y. Fuertes (Spain), D. Dima (Romania), O. Geupel (Germany), N. Hodges (UK),
Y. J. Ionin (USA) & B. M. Bekker (Russia), O. P. Lossers (Netherlands), M. Reid, A. Stadler (Switzerland),
R. Tauraso (Italy), T. Wilde (UK), and the proposers.

A Vanishing Sum of Stirling Numbers

12219 [2020, 944]. Proposed by Brad Isaacson, New York City College of Technology,
New York, NY. Let k and m be positive integers with k < m. Let c(m, k) be the number
of permutations of {1, . . . , m} consisting of k cycles. (The numbers c(m, k) are known as
unsigned Stirling numbers of the first kind.) Prove

 m
(−2)j mj c(j, k)
=0
j =k
(j − 1)!

whenever m and k have opposite parity.

Solution by Roberto Tauraso, University of Rome Tor Vergata, Rome, Italy. Let

m  m
(−2)j mj c(j, k)
Fm (x) = (−x) k
.
k=1 j =k
(j − 1)!

Here Fm (x) is a generating function for the desired sum, evaluated at the negative of the
formal variable. We aim to show that the coefficients of odd powers of x are 0 when m is
even, and the coefficients of even powers of x are 0 when m is odd. For this it suffices to
show
Fm (−x) = (−1)m Fm (x).
The well-known generating function for the unsigned Stirling numbers of the first
j j −1
kind is given by k=1 c(j, k)y k = i=0 (y + i) (easily proved combinatorially). Setting
j −1
y = −x yields k=1 (−1)j −k c(j, k)x k = i=0 (x − i).
j

We interchange the order of summation to take advantage of this identity. Let x be an

integer with x ≥ m. We compute
 
 m
2j mj  j
j −k
 m
2j mj j −1
Fm (x) = (−1) c(j, k)x =
k
(x − i)
j =1
(j − 1)! k=1 j =1
(j − 1)! i=0


m     m   
m−1 x m−1 x j
=m 2j =m 2
j =1
j −1 j j =1
m−j j

 x
z
= m[z ](1 + z)
m m−1
(1 + 2z) = m[z ](1 + z)
x m x+m−1
1+ ,
1+z
where [zm ] is the “coefficient operator” extracting the coefficient of zm in the expression
that follows it.
To extract the coefficient of zm in a different way, we apply the binomial theorem twice
to obtain

690 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
  x 
x  
z x+m−j −1 x j
(1 + z) x+m−1
1+ = (1 + z) z
1+z j =0
j

x  
 x+m−j −1 
 
x x+m−j −1 k
= z j
z .
j =0
j k=0
k
m
To extract all the contributions to the coefficient of z , restrict j to run from 0 to m, and
set k = m − j in the inner sum. This leads to the formula
 x  m   
z x+m−j −1 x
Fm (x) = m[zm ](1 + z)x+m−1 1 + =m .
1+z j =0
m−j j

Viewing xj as a polynomial in x, this is a polynomial equation that holds for every integer
x with x ≥ m. It therefore holds for all real numbers x. Thus, by reversing the index of
summation and using
   
−y r y +r −1
= (−1) ,
r r
we obtain
 m     m   
−x + m − j − 1 −x −(x − j + 1) −x
Fm (−x) = m =m
j =0
m−j j j =0
j m−j


m    
x x+m−j −1
=m (−1)j · (−1)m−j = (−1)m Fm (x),
j =0
j m−j

as desired.
Editorial comment. In addition to the polynomials studied above, solvers used induction,
contour integration, generating function manipulations, or primitive Dirichlet characters.
There is a direct combinatorial proof of the needed identity
 m     m   
m−1 x x+m−j −1 x
2j =
j =1
j −1 j j =0
m−j j

in the proof given above. Both sides count the distinguishable ways to place m balls in x
boxes, where balls may be black or white, with each box having at most one white ball but
any number of black balls. On the left side, j is the number of boxes that have balls: Pick
the boxes, distribute the balls with a positive number in each chosen box, and decide for
each chosen box whether to make one of the balls white. On the right side, j is the number
of white balls: Pick boxes for them, and independently distribute m − j black balls into
the x boxes with repetition allowed.
Also solved by N. Hodges (UK), O. Kouba (Syria), P. Lalonde (Canada), A. Stadler (Switzerland), J. Wangsh-
inghin (Canada), and the proposer.

A Limit Related to the Basel Problem

12220 [2020, 944]. Proposed by D. M. Bătineţu-Giurgiu, “Matei Basarab” National Col-
lege, Bucharest, Romania, and Neculai Stanciu, “George Emil Palade” School, Buzău,
Romania. Let an = nk=1 1/k 2 and bn = nk=1 1/(2k − 1)2 . Prove
 
bn 3 3
lim n − = 2.
n→∞ an 4 π

August–September 2022] PROBLEMS AND SOLUTIONS 691

Solution by Charles Curtis, Missouri Southern State University, Joplin, MO. Note that
2n
1 1 1
n
3 1
n 2n
1 3 2n
1
bn = 2
− 2
= 2
+ 2
= an + .
k=1
k 4 k=1 k 4 k=1 k k=n+1
k 4 k=n+1
k2

Therefore
   n 
n  1 n  1 1
2n n
bn 3 1 1
n − = = = .
an 4 an k=n+1 k 2 an k=1 (n + k)2 an n k=1 (1 + k/n)2

It is well known that an converges to π 2 /6 (this is often called the Basel problem). The
expression in square brackets can be interpreted as a Riemann sum, yielding
 2
1
n
1 1 1
lim = dx = .
n→∞ n
k=1
(1 + k/n) 2
1 x
2 2
Hence we get the desired result.
Also solved by U. Abel & V. Kushnirevych (Germany), K. F. Andersen (Canada), F. R. Ataev (Uzbek-
istan), M. Bataille (France), N. Batir (Turkey), A. Berkane (Algeria), N. Bhandari (Nepal), R. Boukharfane
(Morocco), P. Bracken, B. Bradie, V. Brunetti & J. Garofali & A. Aurigemma (Italy), F. Chamizo (Spain),
H. Chen, C. Chiser (Romania), G. Fera (Italy), D. Fleischman, O. Geupel (Germany), D. Goyal (India), N. Gri-
vaux (France), J. A. Grzesik, L. Han, J.-L. Henry (France), E. A. Herman, N. Hodges (UK), F. Holland (Ire-
land), R. Howard, W. Janous (Austria), O. Kouba (Syria), H. Kwong, P. Lalonde (Canada), G. Lavau (France),
S. Lee, P. W. Lindstrom, O. P. Lossers (Netherlands), C. J. Lungstrom, J. Magliano, R. Molinari, A. Natian,
S. Omar (Morocco), M. Omarjee (France), M. Reid, S. Sharma (India), J. Singh (India), A. Stadler (Switzer-
land), S. M. Stewart (Australia), R. Stong, M. Tang, R. Tauraso (Italy), D. Terr, D. B. Tyler, D. Văcaru (Roma-
nia), J. Vinuesa (Spain), M. Vowe (Switzerland), J. Wangshinghin (Canada), T. Wiandt, Q. Zhang (China),
Missouri State University Problem Solving Group, and the proposer.

A Logarithmic Integral Evaluated by Residues

12221 [2020, 945]. Proposed by Necdet Batır, Nevşehir Hacı Bektaş Veli University,
Nevşehir, Turkey. Prove
 1
log(x 6 + 1) π
dx = log 6 − 3G,
0 x +1
2 2
∞
where G is Catalan’s constant k
k=0 (−1) /(2k + 1)2 .
Solution by Kenneth F. Andersen, Edmonton, AB, Canada. Let I denote the requested inte-
gral. Writing I as a sum of two integrals and then making the change of variable t = 1/x
in the first integral, we obtain
 1  1  ∞  1
log(1 + 1/x 6 ) log x log(1 + t 6 ) log x
I= dx + 6 dx = dt + 6 dx,
0 1 + x 2
0 1 + x 2
1 1 + t 2
0 1 + x2
and therefore  
∞
log(1 + x 6 ) 1
log x
2I = dx + 6 dx.
0 1 + x2 0 1 + x2
To evaluate the last integral, we express 1/(1 + x ) as an infinite series:
2

 1  1 ∞ 
log x
dx = (−1) k 2k
x log x dx.
0 1+x
2
0 k=0

Since the partial sums of the series are bounded in absolute value by 1, the dominated
convergence theorem justifies interchanging the order of summation and integration, and

692 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
then an integration by parts yields
 1 ∞  1 ∞
log x (−1)k+1
dx = (−1) k
x 2k
log x dx = = −G.
0 1+x (2k + 1)2
2
k=0 0 k=0

Thus,  ∞
log(1 + x 6 )
2I = dx − 6G,
0 1 + x2
so the required result follows from
 ∞
log(1 + x 6 )
dx = 2π log 6, (1)
−∞ 1 + x2
which we now prove using the method of residues.
For z = |z|eiθ with |z| > 0 and −π < θ ≤ π , define Log z = log |z| + iθ . The func-
tion Log z is analytic on the open upper half-plane. For R > 1 let CR denote the contour
z = Reiθ , 0 ≤ θ ≤ π . Let
√ √
P1 (z) = z + i, P2 (z) = z − 3/2 + i/2, and P3 (z) = z + 3/2 + i/2.
For j ∈ {1, 2, 3}, the function Log Pj (z) is analytic on the closed upper half-plane, and
therefore the residue theorem yields
 R   
Log Pj (x) Log Pj (z) Log Pj (z)
dx + dz = 2π i Res , i
−R 1 + x2 CR 1 + z2 1 + z2
= π Log Pj (i). (2)
Since
 
 Log Pj (z)  (log(R + 1) + π )
 dz ≤ πR ,
 1+z 2  R2 − 1
CR

letting R → ∞ in (2) and then taking the real part of the resulting identity yields
 ∞
log |Pj (x)|
dx = π log |Pj (i)|.
−∞ 1 + x2
Finally, since
  √  √
x6 + 1 = x2 + 1 x2 −3x + 1 x 2 + 3x + 1
  √  √
= x2 + 1 (x − 3/2)2 + 1/4 (x + 3/2)2 + 1/4
= |P1 (x)|2 |P2 (x)|2 |P3 (x)|2 ,
we have
 ∞  3  ∞
log(1 + x 6 ) 2 log |Pj (x)|
dx = dx
−∞ 1 + x2 j =1 −∞ 1 + x2


3
 √ √
= 2π log |Pj (i)| = 2π log 2 + log 3 + log 3)
j =1

= 2π log 6,
which completes the proof of (1).
Editorial comment. Several solvers noted that a similar problem appeared as problem 2107
in Math. Mag. 93 (2020), p. 389.

August–September 2022] PROBLEMS AND SOLUTIONS 693

Also solved by U. Abel & V. Kushnirevych (Germany), F. R. Ataev (Uzbekistan), M. Bataille (France),
A. Berkane (Algeria), N. Bhandari (Nepal), K. N. Boyadzhiev, P. Bracken, B. Bradie, V. Brunetti & J. Garo-
fali & J. D’Aurizio (Italy), H. Chen, B. E. Davis, G. Fera (Italy), M. L. Glasser, R. Gordon, H. Grandmon-
tagne (France), J. A. Grzesik, L. Han, D. Henderson, E. A. Herman, N. Hodges (UK), F. Holland (Ireland),
P. Khalili, O. Kouba (Syria), Z. Lin (China), O. P. Lossers (Netherlands), T. M. Mazzoli (Austria), M. Omar-
jee (France), V. Schindler (Germany), J. Singh (India), A. Stadler (Switzerland), S. M. Stewart (Australia),
R. Stong, R. Tauraso (Italy), D. Văcaru (Romania), T. Wiandt, M. R. Yegan (Iran), and the proposer.

CLASSICS
We solicit contributions of classics from readers, who should include the problem state-
ment, solution, and references with their submission. The solution to the classic problem
published in one issue will appear in the subsequent issue.
C7. Contributed by Alan D. Taylor, Union College, Schenectady, NY. Are the additive group
of real numbers and the additive group of complex numbers isomorphic?

Random Tetrahedra Inscribed in a Sphere

C6. Contributed by David Aldous, University of California, Berkeley, CA. Consider four
random points on the surface of a sphere, chosen uniformly and independently. Prove that
the probability that the tetrahedron determined by the points contains the center of the
sphere is 1/8.
Solution. Assume the sphere is in R3 centered at the origin O. Fix the point P4 and then
choose P1 , P2 , P3 by randomly choosing three diameters, D1 , D2 , and D3 , and then choos-
ing, randomly, an end of each. There are eight ways to choose the endpoints. The prob-
ability conclusion follows from the observation that, for almost all choices of diameters,
exactly one of the eight choices of endpoints yields a tetrahedron containing O.
To see this, assume that P1 , P2 , and P3 are chosen so that no three of the points
P1 , P2 , P3 , P4 are linearly dependent as vectors in R3 . (The opposite case has probability
0.) The equation −P4 = xP1 + yP2 + zP3 has a unique solution in nonzero real numbers
x, y, and z. Write this as O = xP1 + yP2 + zP3 + P4 . The eight choices of endpoints now
correspond to the eight choices of signs in the expression O = ±xP1 ± yP2 ± zP3 + P4 .
The tetrahedron contains O if and only if there is a representation O = a1 P1 + a2 P2 +
a3 P3 + a4 P4 where ai > 0 for all i. This happens if and only if the coefficients ±x, ±y, ±z
are all positive, and that occurs for exactly one of the eight equally likely choices.
Editorial comment. This was problem A6 on the 1992 Putnam Competition. For a geomet-
ric explanation of what is happening, see the 3blue1brown video “The hardest problem on
the hardest test” at youtube.com/watch?v=OkmNXy7er84. In J. G. Wendel (1962), A prob-
lem in geometric probability, Math. Scand. 11: 109–111, it is proved that for k points on
the sphere in Rn , the probability pn,k that the convex hull of the points contains the origin
k−1 k−1  k−1
is j =n j 2 . A corollary is the surprising duality formula pm,m+n + pn,m+n = 1.
According to Wendel, the problem goes back to R. E. Machol and was first solved by L. J.
Savage.
Some further generalizations can be found in R. Howard and P. Sisson (1996), Capturing
the origin with random points: Generalizations of a Putnam problem, College Math. J.,
27(3): 186–192.

694 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
 

SOLUTIONS

Dominated Convergence of an Integral

12207 [2020, 753]. Proposed by Ovidiu Furdui and Alina Sı̂ntămărian, Technical Univer-
sity of Cluj-Napoca, Cluj-Napoca, Romania. Let f : [0, 1] → R be a continuous function
1
satisfying 0 f (x) dx = 1. Evaluate
 1
n
lim x n f (x n ) ln(1 − x) dx.
n→∞ ln n 0

Solution by Roberto Tauraso, Università di Roma “Tor Vergata,” Rome, Italy. Substituting
t = x n , we get
 1  1
n
x f (x ) ln(1 − x) dx = −
n n
f (t)un (t) dt,
ln n 0 0

where
t 1/n ln(1 − t 1/n )
un (t) = − .
ln n
For fixed t ∈ (0, 1), letting y = 1/n and applying L’Hôpital’s rule twice yields
ln(1 − t y ) t y ln t/(t y − 1) y ln t ln t
lim un (t) = lim = lim = lim y = lim y = 1.
n→∞ y→0+ ln y y→0 + 1/y y→0 t − 1
+ +
y→0 t ln t

Moreover, by Bernoulli’s inequality, for n ≥ 3 we have

ln(1 − t 1/n ) ln((1 − t)/n) ln(1 − t)
0 ≤ un (t) ≤ − ≤− =1− ≤ 1 − ln(1 − t).
ln n ln n ln n

588 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

 1
Since f is bounded and 0 (1 − ln(1 − t)) dt = 2 < ∞, the dominated convergence
theorem applies, and we conclude that
 1  1  1
n
lim x n f (x n ) ln(1 − x) dx = − lim f (t)un (t) dt = − f (t) dt = −1.
n→∞ ln n 0 n→∞ 0 0

Also solved by U. Abel & V. Kushnirevych (Germany), K. F. Andersen (Canada), C. Antoni (Italy),
R. Boukharfane (Saudi Arabia), N. Caro (Brazil), R. Gordon, N. Grivaux (France), L. Han (USA) & X. Tang
(China), E. A. Herman, N. Hodges (UK), F. Holland (Ireland), E. J. Ionaşcu, Y. Jinhai, O. Kouba (Syria),
O. P. Lossers (Netherlands), M. Omarjee (France), A. Stadler (Switzerland), R. Stong, T. Wilde (UK), Y. Xiang
(China), and the proposer.

Three Wise Women

12208 [2020, 753]. Proposed by Gregory Galperin, Eastern Illinois University, Charleston,
IL, and Yury J. Ionin, Central Michigan University, Mount Pleasant, MI. (In memory of
John Horton Conway, 1937–2020.) Three wise women, Alice, Beth, and Cecily, sit around
a table. A card with a positive integer on it is attached to each woman’s forehead, so she
can see the other two numbers but not her own. The women know that one of the three
integers is equal to the sum of the other two. The same question, “Can you determine the
number on your forehead?”, is addressed to the wise women in the following order: Alice,
Beth, Cecily, Alice, Beth, Cecily, . . . . The answer is either “No” or “Yes, the number is
,” and the other wise women hear the answer. The questioning ends as soon as the posi-
tive answer is obtained. (Assume that the women are logical and honest, they all know this,
they all know that they all know this, and so on.)
(a) Prove that whichever numbers are assigned to the wise women, an affirmative answer
is obtained eventually.
(b) Suppose that Alice’s second answer is “Yes, the number is 50.” Determine the numbers
assigned to Beth and Cecily.
(c) Suppose the numbers assigned to Alice, Beth, and Cecily are 1492, 1776, and 284,
respectively. Determine who will give the affirmative answer and how many negative
answers she will give before that.
Solution by Mark D. Meyerson, US Naval Academy, Annapolis, MD. We describe each
assignment of numbers with a triple (a, b, c) giving Alice’s, Beth’s, and Cicely’s positive
numbers in that order. Note that one of the entries must be the sum of the other two.
We claim that for all triples, if a woman says “Yes” on some turn, then her number
must be the largest. Suppose not, and choose a counterexample (a, b, c) for which the
“Yes” answer occurs as early as possible. Suppose, for example, Alice says “Yes” on turn
n, but Beth has the largest number, so b = a + c. (Other cases are similar.) Alice, seeing
the numbers a + c and c, knows from the beginning that her number must be either a or
a + 2c. To say “Yes” on turn n, she must be able to rule out the triple (a + 2c, a + c, c)
for the first time on that turn, and this will happen only if either Cicely or Beth would have
said “Yes” on turn n − 1 or n − 2 on that triple. But this is ruled out by the minimality of
n, since neither Beth nor Cicely has the largest number in that triple.
Let f be the function that assigns to a triple the number of the turn on which the answer
“Yes” occurs. Part (a) asks us to show that f is defined for every triple. If the triple has the
form (2x, x, x), for some positive integer x, then Alice will say “Yes” on her first turn, so
f (2x, x, x) = 1. If it has the form (x, 2x, x), then Alice will think she could have either
x or 3x, so she will say “No,” and then Beth will say “Yes.” Therefore f (x, 2x, x) = 2.
Similarly, for triples of the form (x, x, 2x), Cicely will say “Yes” on her first turn, and
f (x, x, 2x) = 3.
Now consider triples in which the numbers are distinct. If some triple never yields an
affirmative answer, then let (a, b, c) be such a triple whose largest element is as small as

June–July 2022] PROBLEMS AND SOLUTIONS 589

possible. If c = a + b, then (a, b, |a − b|) has a smaller largest element, so f (a, b, |a − b|)
is defined. If f (a, b, |a − b|) = n, then on turn n + 1 or n + 2, depending on which of a
or b is larger, Cecily can eliminate the triple (a, b, |a − b|), since Alice or Beth would pre-
viously have said “Yes.” Cecily then answers “Yes” with a + b on her turn. The argument
is similar when a or b is the largest entry in (a, b, c). This completes the solution to (a).
(b) Using the reasoning from part (a), we can now determine, for every n, the triples
(a, b, c) for which f (a, b, c) = n. If f (a, b, c) = 1, then (a, b, c) must have the form
(2x, x, x), for some positive integer x. For f (a, b, c) = 2, we must have b = a + c. If
a = c then (a, b, c) has the form (x, 2x, x). If not, then f (a, |a − c|, c) must be 1, so
(a, |a − c|, c) has the form (2x, x, x), and therefore (a, b, c) = (2x, 3x, x). Thus, the
triples (a, b, c) such that f (a, b, c) = 2 are those of the form (x, 2x, x) or (2x, 3x, x). If
f (a, b, c) = 3, then c = a + b, and either (a, b, c) has the form (x, x, 2x) or f (a, b, |a −
b|) is either 1 or 2, in which case (a, b, c) has the form (2x, x, 3x), (x, 2x, 3x), or
(2x, 3x, 5x). A similar argument shows that the triples (a, b, c) with f (a, b, c) = 4
are those of the form (3x, 2x, x), (4x, 3x, x), (3x, x, 2x), (4x, x, 3x), (5x, 2x, 3x), or
(8x, 3x, 5x). Since 50 is not divisible by any number in {3, 4, 8}, the only way Alice will
say “Yes, my number is 50” on her second turn (n = 4) is for x to be 10 in the fifth triple,
so Beth has 20 and Cecily has 30.
(c) Working from (1492, 1776, 284) to determine the turn on which that
triple will be resolved, we iteratively replace the biggest number by the dif-
ference of the other two to undo the decision process. The successive triples
after (1492, 1776, 284) are these: (1492, 1208, 284), (924, 1208, 284), (924, 640, 284),
(356, 640, 284), (356, 72, 284), (212, 72, 284), (212, 72, 140), (68, 72, 140), (68, 72, 4),
(68, 64, 4), (60, 64, 4), (60, 56, 4), (52, 56, 4), (52, 48, 4), (44, 48, 4), (44, 40, 4),
(36, 40, 4), (36, 32, 4), (28, 32, 4), (28, 24, 4), (20, 24, 4), (20, 16, 4), (12, 16, 4),
(12, 8, 4), (4, 8, 4). The last triple would be resolved by Beth on turn 2, the one before it
by Alice on turn 4. Working backward, Yes comes on the following turns for these triples:
2, 4, 5, 7, 8, 10, 11, 13, 14, 16, 17, 19, 20, 22, 23, 25, 26, 27, 28, 30, 31, 32, 34, 35, 37, 38.
Since 38 = 3 · 12 + 2, the affirmative answer is by Beth after giving 12 negative answers.
(Tracking only the two smaller entries in each triple, the decision process parallels the
Euclidean algorithm.)
Also solved by E. Curtin, J. Boswell & C. Curtis, N. Hodges (UK), E. J. Ionaşcu, G. Lavau (France),
O. P. Lossers (Netherlands), K. Schilling, E. Schmeichel, R. Stong, F. A. Velandia & J. F. Gonzalez (Colombia),
T. Wilde (UK), Eagle Problem Solvers, The Zurich Logic Coffee (Switzerland), and the proposer.

Asymptotics of a Recursively Defined Sequence

12210 [2020, 852]. Proposed by Paul Bracken, University of Texas Rio Grande Valley,
Edinburg, TX. Let x1 = 1, and let
 
√ 1 2
xn+1 = xn + √
xn
when n ≥ 1. For n ∈ N, let an = 2n + (1/2) log n − xn . Show that the sequence a1 , a2 , . . .
converges.
Solution by Peter W. Lindstrom, Saint Anselm College, Manchester, NH. By the recurrence
for xn , we have xn+1 = xn + 2 + 1/xn > xn + 2, and therefore by induction xn ≥ 2n when
n > 1.
Let zk = xk − 2k. Since zk+1 − zk = xk+1 − xk − 2 = 1/xk , we have
n−1 n−1
1
zn = z 1 + (zk+1 − zk ) = −1 +
k=1 k=1
xk

590 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

for n > 1. Thus
k−1 k−1 
1 1 zk j =1 1/xj − 1 j =2 1/xj (1/2) k−1j =2 1/j log k
0≤ − = = ≤ 2
≤ 2
<
2k xk 2kxk 2kxk (2k) 4k 8k 2
∞ 
for k > 2. Since k=1 log k/(8k 2 ) is convergent, so is ∞ k=1 (1/(2k) − 1/xk ). Let
∞  
1 1
ζ = − .
k=1
2k xk
For n > 1,
n−1
log n log n 1 log n
an = 2n + − xn = −zn + =1− +
2 2 x
k=1 k
2
n n−1  
1 1 1 1 1
=1− − log n + − + .
2 k=1 k k=1
2k xk 2n
Thus limn→∞ an = 1 − γ /2 + ζ , where γ is the Euler–Mascheroni constant.
Also solved by G. Aggarwal (India), K. F. Andersen (Canada), M. Bataille (France), R. Boukharfane (Saudi
Arabia), H. Chen, C. Chiser (Romania), Ó. Ciaurri (Spain), C. Degenkolb, A. Dixit (India) & S. Pathak (USA),
G. Fera (Italy), J. Freeman (Netherlands), R. Gordon, J.-P. Grivaux (France), L. Han, R. Hang, D. Hen-
derson, E. A. Herman, N. Hodges (UK), Y. Jinhai (China), O. Kouba (Syria), Z. Lin (China), J. H. Lind-
sey II, O. P. Lossers (Netherlands), S. Omar (Morocco), M. Omarjee (France), P. Palmieri & C. Antoni (Italy),
A. Pathak (India), R. K. Schwartz, A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), D. Terr, D. B. Tyler,
E. I. Verriest, J. Vukmirović (Serbia), T. Wiandt, L. Wimmer (Germany), L. Zhou, and the proposer.

A Truncated Tetrahedron
12211 [2020, 852]. Proposed by Leonard Giugiuc, Drobeta Turnu Severin, Romania. On
each of the six edges of a tetrahedron, identify the point that is coplanar with the incenter
of the tetrahedron and with the two vertices incident to the opposite edge. Prove that the
volume of the octahedron formed by these six points is no more than half the volume of
the tetrahedron, and determine the conditions for equality.
Solution by Elton Bojaxhiu, Tirana, Albania, and Enkel Hysnelaj, Sydney, Australia. Let
A, B, C, and D be the vertices of the tetrahedron, and let w, x, y, and z denote the areas
of ABC, ABD, ACD, and BCD, respectively.
Let pAB be the plane passing through C, D, and the incenter of the tetrahedron, and let
PAB denote the intersection of pAB with AB. Let hA and hB be the altitudes from A and B,
respectively, to the line CD, and let dA and dB be the distances from A and B, respectively,
to the plane pAB . Since pAB bisects the angle between the planes containing ACD and
BCD, we have
APAB dA hA y
= = = .
BPAB dB hB z
Similarly, if PAC , PAD , PBC , PBD , and PCD are the vertices of the octahedron that lie on
the other edges of the tetrahedron, then we have
APAC x APAD w BPBC x BPBD w CPCD w
= , = , = , = , and = .
CPAC z DPAD z CPBC y DPBD y DPCD x
The octahedron is constructed from the tetrahedron ABCD by removing the four
smaller tetrahedra APAB PAC PAD , BPAB PBC PBD , CPAC PBC PCD , and DPAD PBD PCD . If
t is the volume of the tetrahedron ABCD and tA is the volume of APAB PAC PAD , then
tA APAD APAC APAB w x y
= · · = · · .
t AD AC AB w+z x+z y+z

June–July 2022] PROBLEMS AND SOLUTIONS 591

Combining this with similar formulas for the other small tetrahedra, we see that it suffices
to show
wxy wxz
+
(w + z)(x + z)(y + z) (w + y)(x + y)(z + y)
wyz xyz 1
+ + ≥ . (∗)
(w + x)(y + x)(z + x) (x + w)(y + w)(z + w) 2
Let a, b, c, and d denote the elementary symmetric polynomials in w, x, y, and z:
a = w + x + y + z,
b = wx + wy + wz + xy + xz + yz,
c = wxy + wxz + wyz + xyz,
d = wxyz.
By multiplying out and rearranging, we find that (∗) is equivalent to
abc − 5a 2 d ≥ c2 .
From Newton’s inequalities for the elementary symmetric polynomials, we have
(a/4)(c/4) ≤ (b/6)2 and (b/6)d ≤ (c/4)2 . Consequently,
√
3 ac 3c2 3c2 c3/2
b≥ and d ≤ ≤ √ = √ .
2 8b 12 ac 4 a
√ √
Also, by Maclaurin’s inequality, a/4 ≥ 3 c/4, so a 3/2 ≥ 4 c. Therefore
√ √
3 ac c3/2 a 3/2 c3/2 4 c · c3/2
abc − 5a d ≥ ac ·
2
− 5a · √ =
2
≥ = c2 ,
2 4 a 4 4
as required.
Equality holds if and only if w = x = y = z; that is, all faces of the tetrahedron have
the same area. It is well known that this is true precisely when the tetrahedron is isosceles,
which means that each pair of opposite edges have the same length.
Editorial comment. There are several other ways to establish (∗), as indicated by multiple
solvers. For instance, one could cite Muirhead’s inequality; alternatively, assume without
loss of generality that w ≤ x ≤ y ≤ z, write x = w + s, y = w + s + t, and z = w + s +
t + u for s, t, u ≥ 0, and note that expanding and rearranging (∗) yields f (w, s, t, u) ≥ 0,
where f is a polynomial with all nonnegative coefficients.
Also solved by C. Curtis, G. Fera (Italy), O. P. Lossers (Netherlands), A. Stadler (Switzerland), R. Stong,
J. Vukmirović, and the proposer.

An Application of Farkas’s Lemma

12212 [2020, 852]. Proposed by George Stoica, Saint John, NB, Canada. Let x1 , . . . , xm
and y1 , . . . , ym be two lists of m vectors in Rn , and suppose
xi − xj , yi − yj  ≥ 0
for all i and j in {1, . . . , m}. Prove that there exists a vector y in Rn such that
xi , yi  ≥ xi , y
for all i in {1, . . . , m}.

592 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

Solution by O. P. Lossers, Eindhoven University of Technology, Eindhoven, Netherlands.
The following is a variant of Farkas’s lemma (see for example Corollary 7.1(e) in
A. J. Schrijver, Theory of Linear and Integer Programming, John Wiley and Sons, Chich-
ester, UK, 1986).

If A is a p-by-q matrix, and b ∈ Rp , then exactly one of the following two assertions
is true:
(1) The system Au ≤ b has a solution u ∈ Rq .
(2) The system v T A = 0 has a solution v ∈ Rp with v ≥ 0 and v T b < 0.

Let X and Y be the n-by-m matrices that have the vectors xi and yi , respectively, for their
columns. Let A = XT Y ; in particular, the (i, j )-entry of A is xi , yj . Let b be the vector
consisting of the main diagonal entries of A. If some vector u satisfies Au ≤ b, then the
vector y defined by
m
y = Yu = uj yj
j =1

has the desired property, because

xi , y = uj xi , yj  = uj ai,j = (Au)i ≤ bi = xi , yi .

j j

If there is no such vector u, then by the variant of Farkas’s lemma there exists v ∈ Rm
such that v T A = 0 with v ≥ 0 and v T b < 0. The condition xi − xj , yi − yj  ≥ 0 expands
to the condition aii − aij − aj i + ajj ≥ 0 on the entries of A. Hence,

0≤ vi vj (aii − aij − aj i + ajj )

i,j

= vj vi aii − vj vi aij − vi vj aj i + vi vj ajj

j i j i i j i j

= vj v T b − vj 0 − vi 0 + vi v T b = 2v T b vi < 0,
j j i i i

which is a contradiction.
Also solved by R. Stong and the proposer.

A Sum of Tails of the Zeta Function

12215 [2020, 853]. Proposed by Ovidiu Furdui and Alina Sı̂ntămărian, Technical Univer-
sity of Cluj-Napoca, Cluj-Napoca, Romania. Calculate
∞   
1 1 1 1
+ + + ··· − .
n=1
n2 (n + 2)2 (n + 4)2 2n

Solution by Gaurav Aggarwal, student, Guru Nanak Dev University, Amritsar, India. The
sum equals π 2 /16 + 1/2. Let
N   
1 1 1 1
SN = + + + ··· − .
n=1
n2 (n + 2)2 (n + 4)2 2n

June–July 2022] PROBLEMS AND SOLUTIONS 593

The term
 
1 1 1 1
+ + + ··· −
n2 (n + 2)2 (n + 4)2 2n
clearly
∞ approaches 0 as n approaches infinity, since the part in parentheses is bounded by
2
k=n 1/k , which itself goes to 0. Therefore, it suffices to prove

lim S2N = π 2 /16 + 1/2.

N→∞

We compute
N   ∞ 2N
1 1 1 1
S2N = i + +N −
i=1
(2i − 1)2 (2i)2 i=2N+1
i 2
i=1
2i
N   ∞
i i 1 1 1
= + − − +N
i=1
(2i − 1)2 (2i) 2 2(2i − 1) 2(2i) i=2N+1
i2
N ∞
1 1
= +N .
i=1
2(2i − 1)2
i=2N+1
i2

Noting that ζ (2) = π 2 /6, where ζ is the Riemann zeta function, we have
N  
1 1 1 π2
lim = 1 − ζ (2) = .
N→∞
i=1
2(2i − 1)2 2 22 16

We use telescoping series again and the squeeze theorem to show that the remaining term
tends to 1/2:
∞   ∞ ∞
N 1 1 1 1
=N − =N <N
2N + 1 i=2N+1
i i+1 i=2N+1
i(i + 1) i=2N+1
i2
∞ ∞  
1 1 1 N 1
<N =N − = = .
i=2N+1
(i − 1)i i=2N+1
i − 1 i 2N 2

Hence lim SN = lim S2N = π 2 /16 + 1/2.

N→∞ N→∞

Also solved by U. Abel & V. Kushnirevych (Germany), K. F. Andersen (Canada), M. Bataille (France),
A. Berkane (Algeria), R. Boukharfane (Saudi Arabia), K. N. Boyadzhiev, P. Bracken, B. Bradie, V. Brunetti
& A. Aurigemma & G. Bramanti & J. D’Aurizio & D. B. Malesani (Italy), B. S. Burdick, H. Chen, C. Curtis,
T. Dickens, G. Fera (Italy), M. L. Glasser, H. Grandmontagne (France), J.-P. Grivaux (France), J. A. Grzesik,
E. A. Herman, N. Hodges (UK), F. Holland (Ireland), Y. Jinhai (China), O. Kouba (Syria), K.-W. Lau (China),
G. Lavau (France), O. P. Lossers (Netherlands), R. Molinari, A. Natian, M. Omarjee (France), P. Palmieri
(Italy), K. Schilling, A. Stadler (Switzerland), S. M. Stewart (Australia), R. Stong, R. Tauraso (Italy), D. Terr,
D. B. Tyler, J. Vukmirović (Serbia), T. Wiandt, Y. Xiang (China), FAU Problem Solving Group, Missouri State
Problem Solving Group, and the proposer.

Rotating an Icosahedron
12216 [2020, 944]. Proposed by Zachary Franco, Houston, TX. A regular icosahedron with
volume 1 is rotated about an axis connecting opposite vertices. What is the volume of the
resulting solid?

594 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

Solution by Albert Stadler, Herrliberg, Switzerland. It is known (see for example
en.wikipedia.org/wiki/Regular icosahedron) that if the edge length of a regular icosahe-
dron is a, then the radius of the circumscribed sphere is
a √
R= 10 + 2 5,
4
while the volume is
5 √
V = (3 + 5)a 3 .
12
We place the icosahedron in R3 in such a way that its 12 vertices have the following
coordinates:
P1 : (0, 0, R),
     
R 2kπ 2kπ
P2 –P6 : √ 2 cos , 2 sin , 1 , k ∈ {0, . . . , 4},
5 5 5
     
R (2k + 1)π (2k + 1)π
P7 –P11 : √ 2 cos , 2 sin , −1 , k ∈ {0, . . . , 4},
5 5 5
P12 : (0, 0, −R).
The segment connecting the two points P2 and P7 is given by
R    π  π  
s(t) = √ t 2, 0, 1 + (1 − t) 2 cos , 2 sin , −1 , 0 ≤ t ≤ 1.
5 5 5
This segment generates the boundary of the middle part of the solid formed when the
icosahedron is rotated about the z-axis. The other two parts are cones whose boundaries
are generated by rotating the segment connecting P1 and P2 and the segment connecting
P7 and P12 .
The distance of s(t) from the z-axis equals
 √
R  
   π   π  
 4 − 2(3 − 5)t (1 − t)
√ t 2, 0, 0 + (1 − t) 2 cos , 2 sin ,0  = R .
5 5 5 5
Therefore, the volume of the rotated icosahedron equals
    1 √
2 R 2R 2 2 2R 4 − 2(3 − 5)t (1 − t)
Vrot = π R − √ √ + πR √ dt.
3 5 5 5 0 5

The first term in this formula is the volume of the two cones, and the second is the volume
of the middle part. Evaluating the integral and simplifying we obtain
√ 
2 √ 2 √ 5/2 3
Vrot = (5 + 5)π R = 3
5+ 5 πa .
15 240
If the volume of the icosahedron is 1, then a is determined by
12
a3 = √ .
5(3 + 5)
Substituting this into our formula for Vrot gives a volume of
 √
π 5+ 5
Vrot = ≈ 1.19513.
5 2

June–July 2022] PROBLEMS AND SOLUTIONS 595

Also solved by F. Chamizo (Spain), C. Curtis & J. Boswell, G. Fera (Italy), O. Geupel (Germany), J.-P. Grivaux
(France), N. Hodges (UK), M. J. Knight, G. Lavau (France), O. P. Lossers (Netherlands), M. D. Meyerson,
R. Stong, D. Terr, T. Wiandt, L. Zhou, Davis Problem Solving Group, Eagle Problem Solvers, and the proposer.

CLASSICS
We solicit contributions of classics from readers, who should include the problem state-
ment, solution, and references with their submission. The solution to the classic problem
published in one issue will appear in the subsequent issue.
C6. Due to R. E. Machol and L. J. Savage, contributed by David Aldous, University of
California, Berkeley, CA. Consider four random points on the surface of a sphere, chosen
uniformly and independently. Prove that the probability that the tetrahedron determined by
the points contains the center of the sphere is 1/8.

The Affine Hull of Four Points in Space

C5. Contributed by the editors. Given a set S in Rn , let L(S) be the set of all points lying
on some line determined by two points in S. For example, if S is the set of vertices of an
equilateral triangle in R2 , then L(S) is the union of the three lines that extend the sides of
the triangle, and L(L(S)) is all of R2 . If S is the set of vertices of a regular tetrahedron,
then what is L(L(S))?
Solution. There are precisely four points that are not in L(L(S)). Inscribe the tetrahedron
in a cube with the vertices of the tetrahedron at four of the corners of the cube. The four
other corners of the cube are the missing points.
To see that these points are missed, observe that L(S) consists of all the points on the
extended edges of the tetrahedron. A line through points on adjacent extended edges lies
in the plane of a tetrahedral face and so misses the unused corners. Also, a line connecting
one such corner to a nearby extended edge of the tetrahedron lies in the plane of a face of
the cube and so misses any of the skew edges.
We now show that all other points in R3 are included. Let P1 be the plane containing
the top face of the cube and let P2 be the plane containing the bottom face. Let l1 and l2
be the tetrahedral edges lying in P1 and P2 , respectively. Notice that P1 is the unique plane
containing l1 that is parallel to l2 , and similarly for P2 . Suppose that Q is a point that does
not lie on either P1 or P2 . Let P be the plane containing Q and l1 . Since Q does not lie
on P1 , P is not equal to P1 , so it is not parallel to l2 . Therefore it intersects l2 , say at R.
The line QR lies in the plane P , which contains l1 . Since Q does not lie on P2 , QR is not
parallel to l1 . Therefore QR must intersect l1 , say at T . But now Q, R, and T are collinear,
so Q is in L(L(S)).
This argument shows that L(L(S)) contains all points that do not lie in either the plane
of the top of the cube or the plane of the bottom. Similarly, it contains all points that do
not lie on either the plane of the left side or the right side, and all points that do not lie on
either the plane of the front or back. This means that the only points that can be missed are
the corners of the cube.
Editorial comment. The problem was proposed by Victor Klee as Problem 1413 in
Math. Mag. 66 (1993) 56, with solution in Math. Mag. 67 (1993) 68–69. See also V. Klee
(1963), The generation of affine hulls, Acta Scient. Math. (Szeged) 24, 60–81.

596 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

SOLUTIONS

Non-divisors of Translated Sums of Squares

12200 [2020, 660]. Proposed by Ibrahim Suat Evren, Denizli, Turkey. Prove that for every
positive integer m, there is a positive integer k such that k does not divide m + x 2 + y 2 for
any positive integers x and y.
Solution by Peter W. Lindstrom, Saint Anselm College, Manchester, NH. We prove that
4m2 has the desired property. Let k = 4m2 , and let c be a positive integer, so ck − m =
m(4cm − 1). Since 4cm − 1 ≡ −1 (mod 4), the prime factorization of 4cm − 1 must have
an odd power of a prime p with p ≡ −1 (mod 4). Also, since m and 4cm − 1 are relatively
prime, p cannot divide m, so the prime factorization of ck − m has p to an odd power.
The “sum of two squares” theorem in number theory states that the prime factoriza-
tion of a number of the form x 2 + y 2 has even exponent for each prime congruent to −1
(mod 4). Hence no integers c, x, and y satisfy x 2 + y 2 + m = ck. This makes it impossible
for k to divide x 2 + y 2 + m for any integers x and y.
Also solved by R. Boukharfane (Saudi Arabia), R. Chapman (UK), C. Curtis & J. Boswell, S. M. Gagola Jr.,
N. Hodges (UK), E. J. Ionaşcu, Y. J. Ionin, J. S. Liu, O. P. Lossers (Netherlands), S. Miao (China), C. R. Prane-
sachar (India), A. Stadler (Switzerland), A. Stenger, R. Stong, R. Tauraso (Italy), M. Tetiva (Romania),
K. Williams (Canada), L. Zhou, FAU Problem Solving Group, and the proposer.

A Large Vector Sum from Probability or Polygons

12202 [2020,752]. Proposed by Koopa Tak Lun Koo, ChineseSTEAM Academy, Hong
Kong, China. Let V be a finite set of vectors
 in R2 such that v∈V |v| = π . Prove that
there exists a subset U of V such that | v∈U v| ≥ 1.
Solution I by Oliver Geupel, Brühl, Germany. Choose at random a ray h starting from the
origin. For v ∈ V , let Xv be the length of the projection of v onto h if the angle between

May 2022] PROBLEMS AND SOLUTIONS 487

them is acute, and 0 otherwise. The expected value of Xv is
 π/2
1 |v|
E[Xv ] = |v| cos φ dφ = .
2π −π/2 π
 
Therefore E[ v∈V Xv ] =
 v∈V E[Xv ] = 1, so there is some ray h such that
v∈V Xv ≥ 1. We can now let U = {v ∈ V : the angle between h and v is acute}.
Solution II by Elton Bojaxhiu, Tirana, Albania, and Enkel Hysnelaj, Sydney, Australia. Let
V = {v1 , . . . , vn }, and define vn+1 so that v1 + · · · + vn+1 = 0. For any vector v, let θ (v)
be the angle from the positive x-axis to v, with 0 ≤ θ (v) < 2π , and let v1 , . . . , vn+1 be a
permutation
 of v1 , . . . , vn+1 such that θ (v1 ) ≤ · · · ≤ θ (vn+1 ). The endpoints of the partial
sums ri=1 vi form the vertices of a (possibly degenerate) convex polygon. Let p and d be
the perimeter and diameter of this polygon; it is known that p < π d. Thus
 
n+1
π= |v| ≤ |vk | = p < π d,
v∈V k=1

so d > 1. The set U can be chosen to be a collection of vectors (not including vn+1 ) whose
sum gives a diameter of the polygon.
Editorial comment. Kevin Byrnes and Nicolás Caro pointed out that this problem appears
as exercise 14.9 in J. Michael Steele (2004), The Cauchy–Schwarz Master Class: An Intro-
duction to the Art of Mathematical Inequalities, Cambridge: Cambridge Univ. Press, and
also in W. W. Bledsoe (1970), An inequality about complex numbers, this Monthly 77,
pp. 180–182. If p and d are the perimeter and diameter of a convex m-gon, then the inequal-
ity p < π d follows from p ≤ 2m sin(π/(2m))d, proved in H. Sedrakyan and N. Sedrakyan
(2017), Geometric Inequalities: Methods of Proving, Cham, Switzerland: Springer, p. 379.
Radouan Boukharfane
√ and Tom Wilde extended the problem to Rn , where the constant π
generalizes to 2 π ((n + 1)/2)/ (n/2).
Also solved by R. Boukharfane (Saudi Arabia), K. M. Byrnes, N. Caro (Brazil), R. Chapman (UK), R. Frank
(Germany), Y. J. Ionin, Y. Jeong (Korea), J. H. Lindsey II, O. P. Lossers (Netherlands), M. D. Meyerson,
K. Schilling, E. Schmeichel, R. Stong, R. Tauraso (Italy), T. Wilde (UK), and the proposer.

A Family of Sums with Logarithmic Powers

12203 [2020, 752]. Proposed by Roberto Tauraso, Università di Roma “Tor Vergata,”
Rome, Italy. Let m be a nonnegative integer, and let μ be the Möbius function on Z+ ,
defined by setting μ(k) equal to (−1)r if k is the product of r distinct primes and equal to
0 if k has a square prime factor. Evaluate

1  μ(k) m+1  n 
n
lim ln .
n→∞ lnm (n) k k
k=1

Solution by Albert Stadler, Herrliberg, Switzerland. The limit is m + 1.

For a fixed j ≥ 1, we show that there is a positive constant c such that
  √ 
n
μ(k) d j 1 
(−1) ln k = j
j j
 + O e−c ln n , (1)
k=1
k ds ζ (s) s=1

where ζ (s) is the Riemann zeta function. We start with


n ∞
dj 1 μ(k) μ(k)
j
− s
(−1) j
lnj
(k) = s
(−1)j lnj (k) (2)
ds ζ (s) k=1 k k=n+1
k

488 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
for s > 1, which follows from Dirichlet’s expansion of 1/ζ (s). We now show that (2) holds
s = 1.
also in the case 
Let M(k) = ki=1 μ(i). The function M is known as the Mertens function. Partial sum-
mation yields
∞ ∞ ∞
μ(k) M(k) M(k)
(−1) j
lnj
(k) = (−1) j
lnj
(k) − (−1)j lnj (k + 1)
k=n+1
k s
k=n+1
k s
k=n
(k + 1) s

∞
 
 j
lnj (k+1)
M(n) j +1 j j ln (k)
= (−1) ln (n) + M(k)(−1) − .
ns k=n
ks (k + 1)s

For s ≥ 1 and x > ej ,

 
d lnj (x) lnj (x) j
s
= s+1 −s < 0.
dx x x ln x
Moreover,
d lnj (x) lnj (x)
s
> −s s+1 ,
dx x x
with the latter increasing in x. Thus, by the mean value theorem,
 
 lnj (k) lnj (k + 1)  lnj (k) lnj (k)
 
 s −  < s ≤ 2
 k (k + 1)s  k s+1 k s+1
 √ 
for 1 ≤ s ≤ 2 and k > ej . Since M(k) = O ke−2c ln k for a suitable positive con-
stant c (see, for instance, E. Landau (1974), Handbuch der Lehre von der Verteilung der
j +2
 √  v. 2, AMS Chelsea Publishing: Providence, p. 570) and since ln (k) =
Primzahlen,
O ec ln k , we have
   √ 
 lnj (k) lnj (k + 1) 
 −c ln k 1
M(k)(−1) j
−  = O e .
 ks (k + 1)s  k ln2 (k)
From this
 we deduce ∞
 
 M(n)  lnj (k) lnj (k + 1) 
 j +1 j
 ns (−1) ln (n) + M(k)(−1) −
j

k=n
ks (k + 1)s 
 √   ∞  √ 
1
= O e−c ln n + O e−c ln k
k=n
k ln2 (k)
 √   √   √ 
−c ln n −c ln n 1
=O e +O e = O e−c ln n .
ln n
The convergence of the series is uniform for s ∈ [1, 2], so both sides of (2) are continuous
on [1, 2]. Therefore, (2) is valid at s = 1, proving (1).
We conclude
1  μ(k) m+1  n  1  μ(k)
n n
ln = (ln n − ln k)m+1
lnm (n) k=1 k k lnm (n) k=1 k
m+1  
1  m+1 n
μ(k)
= m lnm+1−j
(n) (−1)j lnj (k)
ln (n) j =0 j k=1
k

m+1     
1  m+1 d j
1   √ 
= m lnm+1−j (n)  + O e−c ln n .
ln (n) j =0 j ds j ζ (s) s=1

May 2022] PROBLEMS AND SOLUTIONS 489

As n → ∞, all error terms have limit 0. Since ζ (s) is meromorphic with a simple pole
of residue 1 at s = 1, the function 1/ζ (s) is holomorphic at s = 1, and its Taylor series
expansion begins (s − 1) + · · · . The main term vanishes for j = 0 and has limit 0 for
j > 1 as n → ∞. Therefore,
  
1  μ(k) m+1  n  m + 1 d 1 
n
lim ln = = m + 1.
n→∞ lnm (n)
k=1
k k 1 ds ζ (s) s=1

Editorial comment. The proof of the bound on the Mertens function is similar to one for
the prime number theorem. Some
 solvers used other bounds, shortening their solutions.
Bounds on sums of the form nk=1 μ(k) lnq (k)/k (Landau, pp. 568–570, 594–595) allow
one to begin with the binomial expansion of ln n − ln k. For m > 0, the solution follows
immediately from


n
μ(k) n 
m−1
lnm+1 = (m + 1) lnm (n) + ck (m) lnk (n) + O(1),
k=1
k k k=1

which appears on p. 489 of H. N. Shapiro (1950), On a theorem of Selberg and generaliza-

tions, Ann. Math., 485–497.
Also solved by W. Janous (Austria), A. Stenger, R. Stong, and the proposer.

The Sum of Cosines in a Convex Quadrilateral

12204 [2020, 752]. Proposed by Florentin Visescu, Bucharest, Romania. Prove that the
absolute value of the sum of the cosines of the four angles in a convex quadrilateral is less
than 1/2.
Solution by O. P. Lossers, Eindhoven University of Technology, Eindhoven, Netherlands.
Denote
 the angles by αi for i ∈ {1, 2, 3, 4}, with 0 < α1 ≤ α2 ≤ α3 ≤ α4 < π . We have
αi = 2π . Let a = α1 +  ≤ π and α3 + α4 = 2π − a. If a = π , then
α2 , and note that a 
all four angles are π/2, so cos(αi ) = 0, so αi = 0 and the required inequality holds.
We may therefore assume a < π .
For the sum of the first two cosines,
a   
α2 − α1
cos α1 + cos α2 = 2 cos cos . (1)
2 2
Since 0 < α1 ≤ α2 , we have
α2 − α1 α1 + α2 a π
0≤ < = < ,
2 2 2 2
and therefore
a   
α2 − α1
cos < cos ≤ 1.
2 2
Multiplying by 2 cos(a/2), which is positive, we conclude
a  a    a 
α2 − α1
2 cos2 < 2 cos cos ≤ 2 cos ,
2 2 2 2
which by (1) implies
a  a 
2 cos2 < cos α1 + cos α2 ≤ 2 cos . (2)
2 2

490 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
 Since 0 < π − α4 ≤ π − α3 < π and
(π − α4 ) + (π − α3 ) = 2π − (α3 + α4 ) = a,
we can apply the same reasoning to π − α4 and π − α3 to obtain
a  a 
2 cos2 < cos(π − α4 ) + cos(π − α3 ) ≤ 2 cos ,
2 2
or equivalently
a  a 
−2 cos ≤ cos α3 + cos α4 < −2 cos2 . (3)
2 2
Adding (2) and (3), and putting x = cos(a/2), we get

2x 2 − 2x < αi < 2x − 2x 2 .
Since the quadratic 2x − 2x 2 has maximum value 1/2 at x = 1/2, this proves the inequality.
Editorial comment. The problem statement assumes that all angles are strictly less than π .
If one allows an angle to equal π , then one can achieve a cosine sum of 1/2 by beginning
with an equilateral triangle and adding a fourth vertex along one side, obtaining a four-
sided figure with angles π/3, π/3, π/3, and π . One can obtain quadrilaterals with all angles
less than π and cosine sum arbitrarily close to 1/2 by using angles π/3 + , π/3 + ,
π/3 + , and π − 3 . 
Nicolás Carosolved the more general problem of bounding ni=1 cos xi , given that
0 < xi < π and ni=1 xi = j π ; the stated problem is the case n = 4, j = 2.
Also solved by E. Bojazhiu (Albania) & E. Hysnelaj (Australia), R. Boukharfane (Saudi Arabia), N. Caro
(Brazil), R. Chapman (UK), C. Chiser (Romania), G. Fera & G. Tescaro (Italy), L. Giugiuc (Romania), J.-
P. Grivaux (France), N. Hodges (UK), E. J. Ionaşcu, Y. J. Ionin, W. Janous (Austria), A. B. Kasturiarachi,
O. Kouba (Syria), K.-W. Lau (China), Z. Lin (China), J. H. Lindsey II, K. Park (Korea), C. Schacht, E. Schme-
ichel, A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), E. I. Verriest, L. Zhou, and the proposer.

Minimizing a Ratio of Integrals

12205 [2020, 752]. Proposed by Christian Chiser, Elena Cuza College, Craiova, Romania.
Find the minimum value of
1 2 2
0 x (f (x)) dx
1
2 2
0 x (f (x)) dx
over all nonzero continuously differentiable functions f : [0, 1] → R with f (1) = 0.
Solution by Jinhai Yan, Fudan University, Shanghai, China. We show that the minimum
value is π 2 .
Let 
sin(π x)/x, if x = 0,
g(x) =
π, if x = 0.
Note that g ∈ C ∞ [0, 1], g(1) = 0, and g satisfies the Euler–Lagrange equation
d  2 
x g (x) = −π 2 x 2 g(x).
dx
Therefore, for any f as in the problem statement,
   
d x 2 g (x) 2g (x) g (x)2
f (x)2 = x 2 f (x)f (x) − π 2 f (x)2 − f (x) 2
dx g(x) g(x) g(x)2
 2
  g (x)
= x 2 f (x)2 − π 2 f (x)2 − x 2 f (x) − f (x) .
g(x)

May 2022] PROBLEMS AND SOLUTIONS 491

Note that the singularity at x = 1 on both sides of this equation is removable, because
f (x) f (x) f (1)
lim = lim =− ∈ R.
x→1− g(x) x→1− g (x) π
It follows that
   2  1
1   g (x) x 2 g (x) 
x f (x) − π f (x) − x f (x) −
2 2 2 2 2
f (x) dx = f (x)2  = 0.
0 g(x) g(x) 0

Thus
 1  1  1  2
g (x)f (x)
x f (x) dx − π
2 2 2
x f (x) dx =
2 2
x 2
f (x) − dx ≥ 0,
0 0 0 g(x)
with equality if f = g, and the desired conclusion follows.
Also solved by K. F. Andersen (Canada), R. Boukharfane (Saudi Arabia), P. Bracken, H. Chen, T. Dick-
ens, L. Han, O. Kouba (Syria), P. W. Lindstrom, A. Natian, M. Omarjee (France), A. Stadler (Switzerland),
R. Stong, R. Tauraso (Italy), E. I. Verriest, and the proposer.

A Skew-Harmonic Formula for Apéry’s Constant

12206 [2020, 752]. Proposed by Seán Stewart, Bomaderry, Australia. Prove
∞
 H 2n 3
= ζ (3),
n=1
n2 4
n k+1
∞ H n 3is the nth skew-harmonic number
where k=1 (−1) /k and ζ (3) is Apéry’s constant
k=1 1/k .
n
Solution by Michel Bataille, Rouen, France. With H0 = 0 and Hn = k=1 1/k,
m
1 m
1
H 2m = H2m − 2 = H2m − Hm = . (1)
k=1
2k k=1
m + k

Also note that


m+N   
2m+N
1 1 1
H2m−1 − Hm−1 − − = H2m+N − Hm+N = .
j =m
j j +m j =m+N+1
j

As N tends to ∞, the right side tends to 0, so

 ∞  
1 1
− = H2m−1 − Hm−1 . (2)
j =m
j j +m
∞
Let S = n=1 H 2n /n2 . By (1),
∞ ∞  
1  1  11 1
n n
1
S= = −
n=1
n2 k=1 n + k n=1
n k=1 k n n + k
∞
 ∞ 
 n
Hn 1
= − . (3)
n=1
n2 n=1 k=1
nk(n + k)

492 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
 We consider the two terms in this expression separately. First
∞ ∞    ∞
Hn Hn−1 1 Hn−1
2
= 2
+ 3
= + ζ (3) = 2ζ (3)
n=1
n n=1
n n n=1
n2

by Euler’s formula ∞ n=1 Hn−1 /n = ζ (3).
2

To evaluate the double sum in the second term of (3), interchange the order of sum-
mation, use (2), and then manipulate the harmonic terms and use the first part of (1) to
obtain
∞  ∞ ∞    ∞
1  1
n
1 1 H2k−1 − Hk−1
= − =
n=1 k=1
nk(n + k) k=1
k n=k n n + k
2
k=1
k2

∞ ∞
H2k − Hk + 1/(2k)  H 2k ζ (3) ζ (3)
= 2
= 2
+ =S+ .
k=1
k k=1
k 2 2

Thus
 
ζ (3)
S = 2ζ (3) − S + ,
2
and the result follows.
Editorial comment. A simple proof of Euler’s formula for ζ (3) appears in this Monthly
127 (2020), 855. That issue contains the solutions to Problem 12091 and Problem 12102,
both of which also link ζ (3) to infinite series involving harmonic sums.
Many solvers expressed harmonic numbers as integrals from 0 to 1 of the formula for
the sum of a finite geometric series and then performed interchanges. This led to various
integrals with logarithmic integrands and/or dilogarithms. Two known definite integrals
that played a role in many solutions were
 1
log2 (1 − x)
dx = 2ζ (3)
1 x
and
 1
log(1 − x) log(1 + x) 5
dx = − ζ (3).
0 x 8

Also solved by A. Berkane (Algeria), N. Bhandari (Nepal), R. Boukharfane (Saudi Arabia), K. N. Boyadzhiev,
P. Bracken, B. Bradie, N. Caro (Brazil), A. C. Castrillón (Colombia), H. Chen, N. S. Dasireddy (India), G. Fera
(Italy), M. L. Glasser, R. Gordon, H. Grandmontagne (France), L. Han, E. A. Herman, N. Hodges (UK),
F. Holland (Ireland), W. Janous (Austria), O. Kouba (Syria), K.-W. Lau (China), O. P. Lossers (Netherlands),
I. Mezö (China), R. Molinari, V. H. Moll & T. Amdeberhan, K. Nelson, M. Omarjee (France), S. Sharma
(India), A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), J. Wangshinghin (Canada), T. Wiandt, Y. Xiang
(China), and the proposer.

A Fibonacci Inequality
12213 [2020, 853]. Proposed by Hideyuki Ohtsuka, Saitama, Japan. Let Fn be the nth
Fibonacci number, defined by F0 = 0, F1 = 1, and Fn = Fn−1 + Fn−2 for n ≥ 2. Prove

n
  √
Fk−1 Fk+2 ≤ Fn+1 Fn+4 − 5.
k=1

May 2022] PROBLEMS AND SOLUTIONS 493

Solution by Rory Molinari, Beverly Hills, MI. More generally, consider a sequence a
of nonnegative real numbers such that an = an−1 + an−2 for n ≥ 2. For n ≥ 2 and d a
nonnegative integer, we prove


n−1
√ √ √
ak−1 ak+d−1 ≤ an an+d − a1 ad+1 .
k=1

Setting an = Fn+1and d = 3 proves the desired inequality.

m
The identity k=j ak = am+2 − aj +1 is easily shown by induction on m. By the
Cauchy–Schwarz inequality,
 n−1 1/2  n−1 1/2

n−1
√   
ak−1 ak+d−1 ≤ ak−1 ak+d−1 = (an − a1 )(an+d − ad+1 ).
k=1 k=1 k=1

By the AM-GM inequality,

(an − a1 )(an+d − ad+1 ) = an an+d + a1 ad+1 − a1 an+d − ad+1 an
√
≤ an an+d + a1 ad+1 − 2 a1 an+d ad+1 an
√ √ 2
= an an+d − a1 ad+1 .

Editorial comment. The majority of solvers proved the inequality by induction, showing
  
Fn+1 Fn+4 + Fn Fn+3 ≤ Fn+2 Fn+5
by squaring both sides and applying the AM-GM inequality. Doyle Henderson used this
approach to generalize to a sequence of real numbers satisfying an ≥ an−1 + an−2 for n ≥ 2
√ √ √
and a0 a3 ≤ a2 a5 − a5 , obtaining

n
√ √ √
ak−1 ak+2 ≤ an+1 an+4 − a5 .
k=1

Also solved by K. F. Andersen (Canada), M. Bataille (France), B. D. Beasley, R. Boukharfane (Saudi Ara-
bia), P. Bracken, B. Bradie, Ó. Ciaurri (Spain), C. Curtis, A. Dixit (India) & S. Pathak (USA), G. Fera (Italy),
D. Fleischman, O. Geupel (Germany), R. Gordon, D. Henderson, N. Hodges (UK), Y. J. Ionin, W. Janous (Aus-
tria), M. Kaplan & M. Goldenberg, K. T. L. Koo (China), O. Kouba (Syria), W.-K. Lai, P. Lalonde (Canada),
K.-W. Lau (China), O. P. Lossers (Netherlands), R. Nandan, M. Omarjee (France), J. Pak (Canada), A. Pathak
(India), Á. Plaza (Spain), E. Schmeichel, A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), D. B. Tyler,
J. Van hamme (Belgium), M. Vowe (Switzerland), J. Vukmirović (Serbia), T. Wiandt, L. Wimmer (Germany),
X. Ye (China), A. Zaidan, L. Zhou, FAU Problem Solving Group, and the proposer.

CLASSICS
We solicit contributions of classics from readers, who should include the problem state-
ment, solution, and references with their submission. The solution to the classic problem
published in one issue will appear in the subsequent issue.
C5. Due to Victor Klee, contributed by the editors. Given a set S in Rn , let L(S) be the set
of all points lying on some line determined by two points in S. For example, if S is the set
of vertices of an equilateral triangle in R2 , then L(S) is the union of the three lines that
extend the sides of the triangle, and L(L(S)) is all of R2 . If S is the set of vertices of a
regular tetrahedron, then what is L(L(S))?

494 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
 Returning the Icing to the Top
C4. From the 1968 Moscow Mathematical Olympiad, contributed by the editors. A round
cake has icing on the top but not the bottom. Cut a piece of the cake in the usual shape of
a sector with vertex angle one radian and with vertex at the center of the cake. Remove the
piece, turn it upside down, and replace it in the cake to restore roundness. Next, move one
radian around the cake, cut another piece with the same vertex angle adjacent to the first,
remove it, turn it over, and replace it. Keep doing this, moving around the cake one radian
at a time, inverting each piece. Show that, after a finite number of steps, all the icing will
again be on the top.
Solution. We solve the general problem in which the central angle of every slice is θ radi-
ans. If 2π/θ is an integer n, then clearly n flips put all the icing on the bottom, and n more
flips return it all to the top. Otherwise, let n = 2π/θ . We show that the icing returns
to the top for the first time after 2n(n + 1) steps. In the case θ = 1, we have n = 6, and
therefore it takes 84 steps for the icing to return to the top.
Let α = 2π − nθ . Clearly 0 < α < θ . Let β = θ − α, so that α + β = θ . Cut n
consecutive pieces with angle θ (these are the first n pieces to be flipped), leaving a piece
with angle α. Cut each of the n pieces into two
pieces of angle α and β, as in the figure.
Reading counterclockwise, you now have
pieces of width α, β, α, β, . . . , α, with the last
α adjacent to the first. Let A1 , . . . , An+1 be
the pieces with angle α, and let B1 , . . . , Bn be
the pieces with angle β, with Bi between Ai
and Ai+1 , as shown here. You may now dis-
card the knife; no further cutting is necessary.
Imagine that the cake is on a rotating cake
plate and we rotate the cake plate clockwise
through an angle of θ after each piece is
flipped. In the first step, we flip the piece con-
sisting of A1 and B1 and then rotate the plate
clockwise. Piece A1 is now upside down in the original location of piece An+1 , and B1
is now upside down in the original location of piece Bn . All other pieces simply rotate
clockwise without being flipped, so for 2 ≤ i ≤ n + 1, Ai moves to the original location
of Ai−1 , and for 2 ≤ i ≤ n, Bi moves to the original location of Bi−1 . At the end of this
operation the cuts are in the same positions as they were in originally; the net effect of one
step is simply to permute the A and B pieces cyclically, with one of each being flipped.
It is now clear that after n steps the B pieces have completed a full rotation, with each
piece being flipped once, so they are back in their original positions upside down, and after
another n steps they are in their original positions right side up again. Similarly, it takes
2(n + 1) steps for all the A pieces to return to right side up, in their original positions. It
follows that the number of steps needed to return all icing to the top is the least common
multiple of 2n and 2(n + 1), which is 2n(n + 1). Indeed, after this many steps, not only is
the icing on top, but the cake is fully restored to its original configuration.
Editorial comment. This problem appeared, in a somewhat different form, as problem
31.2.8.3 in the 1968 Moscow Mathematical Olympiad. The version given here appears in
P. Winkler (2007), Mathematical Mind-Benders, A K Peters/CRC Press, Wellesley, MA.

May 2022] PROBLEMS AND SOLUTIONS 495

SOLUTIONS
An Euler–Mascheroni Sum
12194 [2020, 564]. Proposed by Marian Tetiva,
 Gheorghe Roşca Codreanu National Col-
lege, Bı̂rlad, Romania. Let γn = − ln n + nk=1 1/k, and let γ be the Euler–Mascheroni
constant limn→∞ γn . Evaluate
∞ 
1
γn − γ − .
n=1
2n

Solution by Abdelhak Berkane, Université Frères Mentouri, Constantine, Algeria. We show

that the answer is 1 + γ − ln(2π ) /2. Let Hn denote the nth harmonic number, so that
γn = − ln n + H n , and let Sn denote the nth partial sum of the series in the problem. Apply-
ing the formula nk=1 Hk = (n + 1)Hn − n (which is easily verified by induction), we find
that
n  
1 1
Sn = Hk − ln k − γ − = n+ Hn − n − ln(n!) − nγ .
k=1
2k 2

Using the known asymptotic formulas


1 1
Hn = ln n + γ + +O and
2n n2

ln(2π n) 1
ln(n!) = n ln n − n + +O ,
2 n
we obtain
    
1 1 1 ln(2π n) 1
Sn = n+ ln n + γ + +O 2 − n − n ln n − n + +O − nγ
2 2n n 2 n

1 + γ − ln(2π ) 1
= +O .
2 n
Let n → ∞ to get the desired sum.

386 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

Also solved by U. Abel & V. Kushnirevych (Germany), T. Akhmetov (Russia), K. F. Andersen (Canada),
M. Bataille (France), N. Bhandari (Nepal), R. Boukharfane (Saudi Arabia), P. Bracken, B. Bradie, N. Caro
(Brazil), R. Chapman (UK), H. Chen, C. Chiser (Romania), B. E. Davis, A. Dixit (India) & S. Pathak (US),
S. P. I. Evangelou (Greece), G. Fera (Italy), D. Fleischman, S. Gayen (India), O. Geupel (Germany), J. A. Grze-
sik, E. A. Herman, N. Hodges (UK), W. Janous (Austria), M. Kaplan & M. Goldenberg, K. T. L. Koo (China),
O. Kouba (Syria), S. S. Kumar, K.-W. Lau (China), G. Lavau (France), O. P. Lossers (Netherlands), I. Mezo
(Canada), R. Molinari, A. Natian, K. Nelson, M. Omarjee (France), N. Osipov (Russia), A. Pathak, Á. Plaza
(Spain), K. Sarma (India), K. Schilling, S. Sharma (India), S. Singhania (India), A. Stadler (Switzerland),
S. M. Stewart (Australia), R. Stong, R. Tauraso (Italy), M. Vowe (Switzerland), T. Wiandt, H. Widmer (Switzer-
land), Y. Xiang (China), L. Zhou, and the proposer.

A Mean Inequality
12196 [2020, 659]. Proposed by Vasile Mircea Popa, Lucian Blaga University, Sibiu,
Romania. Determine which positive integers n have the following property: If a1 , . . . , an
are n real numbers greater than or equal to 1, and A, G, and H are their arithmetic mean,
geometric mean, and harmonic mean, respectively, then
1 1
G−H ≥ − .
G A

Composite solution by Radouan Boukharfane, Extreme Computing Research Center,

Thuwal, Saudi Arabia, Nigel Hodges, Cheltenham, UK, the proposer, and the editors.
The property holds for n ≤ 5 but fails for n ≥ 6.
If a1 = a2 = · · · = an−1 = 1 and an = n + 1, then the inequality becomes
√ n+1 1 1
n+1− ≥ √ − .
n
(1)
n n
n+1 2

√ inequality is false for n ≥ 6. To see why, we first √

We claim that this note that (5/4)12 > 13,
and√therefore 12 13 < 5/4. It is easily verified that the sequence { n n + 1} is decreasing,
so n n + 1 < 5/4 for n ≥ 12, and therefore
√ n+1 5 1 1 1 4 1 3 1
n+1− < −1= √ − > − =
n
and > .
n 4 4 n
n+1 2 5 2 10 4
Thus, (1) is false for n ≥ 12. One can check numerically that it is also false for n =
6, . . . , 11, so the property in the problem does not hold for n ≥ 6.
To prove that it holds for n ≤ 5, let
1 1
F (a1 , . . . , an ) = G − −H + .
G A
Suppose C > 1. We show that if n ≤ 5 and 1 ≤ a1 ≤ · · · ≤ an ≤ C, then F (a1 , . . . , an ) ≥ 0.
Since C is arbitrary, this will establish that the property holds for n ≤ 5.
Since we have restricted our attention to a compact domain, F achieves a minimum
value on that domain. We need the following fact about where the minimum occurs.
Lemma. If the minimum value of F (a1 , . . . , an ) for 1 ≤ a1 ≤ · · · ≤ an ≤ C is negative,
and F achieves that minimum value at a sequence (a1 , . . . , an ), then aj = 1 whenever
1 ≤ j ≤ n/2 + 1.

Proof. Suppose that the minimum value is negative. Note that if a1 = · · · = an , then
F (a1 , . . . , an ) = 0, so the minimum must occur at a nonconstant sequence. We proceed
now by induction on j .

April 2022] PROBLEMS AND SOLUTIONS 387

 For the base case, suppose that F achieves its minimum at a sequence (a1 , . . . , an )
with 1 < a1 ≤ · · · ≤ an ≤ C. Since the sequence is not constant, H < G < A. With bi =
ai /a1 , we have
 
1 1 1 1 1
F (b1 , . . . , bn ) = (G − H ) − a1 − < G−H − − = F (a1 , . . . , an ),
a1 G A G A
contradicting the assumption that F achieves its minimum at (a1 , . . . , an ). This establishes
the base case.
For the induction step, assume that j ≥ 1, the claim holds for 1, . . . , j , and j + 1 ≤
n/2 + 1; that is, j ≤ n/2. Suppose F achieves its minimum at a sequence (a1 , . . . , an )
with aj +1 > 1. By the induction hypothesis, a1 = · · · = aj = 1. We have A = S/n and
H = n/T , where
1 1 1 1
S = a1 + · · · + an = j + aj +1 + · · · + an , T = +···+ =j+ +···+ .
a1 an aj +1 an
√
Let bi = ai for i ∈
/ {j, j + 1}, and let bj = bj +1 = aj +1 . The sequence (b1 , . . . , bn )
has the same geometric mean as (a1 , . . . , an ), and its arithmetic and harmonic means are
S /n and n/T , respectively, where
√ √
S = S − 1 − aj +1 + 2 aj +1 = S − ( aj +1 − 1)2 ,
√
1 2 ( aj +1 − 1)2
T =T −1− +√ =T − .
aj +1 aj +1 aj +1
Therefore
 
1 n n 1 n n
F (a1 , . . . , an ) − F (b1 , . . . , bn ) = G − − + − G− − +
G T S G T S
 
n n n n
= √ − − √ −
T − ( aj +1 − 1) /aj +12 T S − ( aj +1 − 1) 2 S

√ 1 1
= n( aj +1 − 1)2 √ − √ . (2)
T (T aj +1 − ( aj +1 − 1) ) S(S − ( aj +1 − 1)2 )
2

Clearly, T ≤ S, and using the fact that j ≤ n/2 we obtain

 
1 1 n−j
T aj +1 = j + + ··· + aj +1 ≤ j + aj +1 = j aj +1 + n − j
aj +1 an aj +1
= n + j (aj +1 − 1) ≤ n + (n − j )(aj +1 − 1) = j + (n − j )aj +1
≤ j + aj +1 + · · · + an = S.
Combining this with (2), we conclude F (a1 , . . . , an ) − F (b1 , . . . , bn ) ≥ 0, which implies
F (b1 , . . . , bn ) ≤ F (a1 , . . . , an ) and hence F achieves its minimum at (b1 , . . . , bn ). But
√
bj = aj +1 > 1, so this contradicts the induction hypothesis. 

We are now ready to complete the solution. The case n = 1 is trivial. If n = 2 and the
minimum of F is negative, then by the lemma this minimum must occur at the sequence
(1, 1). But F (1, 1) = 0, so this is impossible.
If n = 3 and the minimum of F is negative, then by the lemma the minimum occurs at
some sequence (1, 1, a3 ). Writing a3 = (x + 1)3 for some x ≥ 0, we have
F (1, 1, (x + 1)3 ) < 0.

388 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

On the other hand,
1 3 3
F (1, 1, (x + 1)3 ) = (x + 1) − − +
x + 1 2 + 1/(x + 1)3 2 + (x + 1)3
x 3 (2 + x)(6 + 12x + 15x 2 + 9x 3 + 2x 4 )
= ≥ 0,
(1 + x)(3 + 3x + 3x 2 + x 3 )(3 + 6x + 6x 2 + 2x 3 )
so this is a contradiction.
Similarly, if n = 4 and the minimum of F is negative, then by the lemma we have
F (1, 1, 1, (x + 1)4 ) < 0 for some x ≥ 0, and we get a contradiction from the calculation

F (1, 1, 1,(x + 1)4 )

x 3 (2 + x)(8 + 24x + 60x 2 + 80x 3 + 56x 4 + 20x 5 + 3x 6 )
= ≥ 0.
(1 + x)(4 + 4x + 6x 2 + 4x 3 + x 4 )(4 + 12x + 18x 2 + 12x 3 + 3x 4 )
Finally, if n = 5 and the minimum of F is negative, then by the lemma we have
F (1, 1, 1, (x + 1)5 , (x + y + 1)5 ) < 0 for some x, y ≥ 0. A calculation similar to those in
the previous cases shows that F (1, 1, 1, (x + 1)5 , (x + y + 1)5 ) is a rational function with
all coefficients positive, which is a contradiction.
Editorial comment. When n = 6, F (1, 1, 1, 1, 1, (x + 1)6 ) is a rational function whose
numerator is

x 3 (2 + x)(−30 − 150x − 111x 2 + 456x 3 + 1328x 4

+ 1758x 5 + 1431x 6 + 764x 7 + 264x 8 + 54x 9 + 5x 10 ),

which is negative for x positive and close to 0.

No other complete solutions were received.

A Pell-type Equation in Disguise

12197 [2020, 659]. Proposed by Nicolai Osipov, Siberian Federal University, Krasnoyarsk,
Russia. Prove that the equation

(a 2 + 1)(b2 − 1) = c2 + 3333

has no solutions in integers a, b, and c.

Solution by Richard Stong, Center for Communications Research, San Diego, CA. We may
clearly assume a, b, c ≥ 0. If a = 0, then b2 − c2 = 3334, which has no solutions since
3334 ≡ 2 (mod 4). If b ∈ {0, 1}, then the left side is nonpositive and there are no solutions.
Thus we may assume a > 0 and b > 1. Hence neither a 2 + 1 nor b2 − 1 is a perfect square.
We rewrite the equation as c2 − da 2 = b2 − 3334, where d = b2 − 1, in order to apply
known results about Pell-type equations.
In the Pell-type equation x 2 − dy 2 = n, where d > 0 and d is not a perfect square, √
with any solution
√(x, y) we can associate
√ an algebraic number α by setting α = x + y d.
Since α = x + x 2 − n, and x + x 2 − n increases with x for x 2 > n, minimizing x is
equivalent to minimizing α.
integers to u2 − dv 2 = 1 we
With a solution (u, v) in positive √ √associate another alge-
braic number β by setting β = u + v d. Note that β −1 = u − v d. We compute
√ √ √
αβ −1 = (x + y d)(u − v d) = (xu − dyv) + (yu − xv) d.

April 2022] PROBLEMS AND SOLUTIONS 389

 Setting x = xu − dyv and y = yu − xv gives another solution to x 2 − dy 2 = n. Sup-
pose that (x, y) is the solution in nonnegative integers that minimizes x and hence also
minimizes α. Since β > 1, we have αβ −1 < α, so x or y must be negative. They cannot
both be negative, because αβ −1 > 0. Since (x )2 − d(y )2 = n, we have
√ √ √ √
nβα −1 = (x + y d)(x − y d)βα −1 = αβ −1 (x − y d)βα −1 = x − y d.
√ √
Since exactly one of√x and y is negative, |x − y d| = |x | + |y | d, and hence
|n|βα −1 = |x | + |y | d. Since (|x |, |y |) is√a solution to x 2 − dy 2 = n, the minimal-
ity of α implies α ≤ |n|βα −1 , and hence α ≤ |n|β.
Now consider a solution (a, b, c) to the original equation that minimizes c. Write the
equation as
c2 − (b2 − 1)a 2 = b2 − 3334 = n,
√
√ (u, v) = (b, 1) satisfies u − (b − 1)v = 1. Letting α = c + a b − 1 and
2 2 2 2
and note that
β = b + b − 1, we obtain
2

c + a b2 − 1 = α ≤ |n|β = |b2 − 3334|(b + b2 − 1) < |b2 − 3334|(2b).

We next prove that b < 117. If b ≥ 117 (in fact, whenever b ≥ 58), then
c2 − (b2 − 1)a 2 = b2 − 3334 > 0,
√
so c > a b2 − 1. Hence,

2a b2 − 1 < 2b|b2 − 3334|. (∗)

Now rewrite the original equation as
c2 − (a 2 + 1)b2 = −a 2 − 3334.
√
Note that (u, v) = (2a 2 +
√ 1, 2a) satisfies u2 − (a 2 + 1)v 2 = 1. Take α = c√
+ b a2 + 1
and β = (2a 2 + 1) + 2a a 2 + 1 in the preceding, and note that β = (a + a 2 + 1)2 <
4(a 2 + 1). We obtain

c + b a2 + 1 = α ≤ |n|β = (a 2 + 3334)β < 2 (a 2 + 3334)(a 2 + 1).

√
Since c > 0, we conclude b < 2 a 2 + 3334. Combining this with (∗), we obtain
2b(b2 − 3334)
b2 < 4(a 2 + 3334) < + 13336.
b2 − 1
The largest real root of t 2 (t 2 − 1) − 2t (t 2 − 3334) − 13336(t 2 − 1) is less than 117, so
b < 117.
Thus the problem is reduced to checking values of b up to 116 and values of a up
to b|b2 − 3334|/(2(b2 − 1)) and then evaluating c. This is easily done on a computer,
yielding no solutions with integral c.
Also solved by R. Chapman (UK), A. Stenger, and the proposer.

Dilating Kimberling’s Center X65 from the Incenter

12198 [2020, 659]. Proposed by Michel Bataille, Rouen, France. Let A1 A2 A3 be a
nonequilateral triangle with incenter I , circumcenter O, and circumradius R. For i ∈
{1, 2, 3}, let Bi be the point of tangency of the incircle of A1 A2 A3 with the side of the
triangle opposite Ai , and let Ci be the point of intersection between the circle centered at

390 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

I of radius R and the ray I Bi . Let K be the orthocenter of C1 C2 C3 . Prove that I is the
midpoint of OK.
Solution by Lienhard Wimmer, Isny im Allgäu, Germany. For i ∈ {1, 2, 3}, let Di be the
reflection of Ci through I . It suffices
to show that O is the orthocenter
of D1 D2 D3 , because this orthocen-
ter is the reflection of K through I .
Extend A1 I to intersect the circum-
circle of A1 A2 A3 at X.
Since A1 X bisects ∠A2 A1 A3 ,
arcs A2 X and A3 X are equal. There-
fore OX is the perpendicular bisec-
tor of A2 A3 , so OX  D1 I . By con-
struction, D1 I = R = OX. Thus
D1 I XO is a parallelogram, which
implies D1 O  A1 X. The isosce-
les triangles I B2 B3 and I D2 D3
are similar, and therefore B2 B3 
D2 D3 . Since A1 X ⊥ B2 B3 , we con-
clude that D1 O ⊥ D2 D3 . Like-
wise, D2 O ⊥ D3 D1 , completing the
proof.
Editorial comment. Oliver Geupel and Nigel Hodges point out that the orthocenter of
B1 B2 B3 is center X65 in Clark Kimberling’s Encyclopedia of Triangle Centers
(faculty.evansville.edu/ck6/encyclopedia/etc.html), and I divides OX65 in the ratio of
R : r. The result in the problem follows immediately, because C1 C2 C3 is the image of
B1 B2 B3 under a dilation with center I and ratio R/r.
Also solved by R. Boukharfane (Saudi Arabia), H. Chen (China), G. Fera (Italy), O. Geupel (Germany),
N. Hodges (UK), E. J. Ionaşcu, W. Janous (Austria), M. Kaplan & M. Goldenberg, L. Kiernan, O. Kouba
(Syria), J. H. Lindsey II, O. P. Lossers (Netherlands), C. R. Pranesachar (India), V. Schindler (Germany),
A. Stadler (Switzerland), R. Stong, T. Wiandt, L. Zhou, and the proposer.

The Basel Problem in Disguise

12199 [2020, 660]. Proposed by Shivam Sharma, Delhi University, New Delhi, India. Prove
 ∞
x sinh(x) π2
dx = .
0 3 + 4 sinh2 (x) 24

Solution by Robin Chapman, University of Exeter, Exeter, UK. Observe that for x > 0,
2 sinh x ex − e−x 1 (ex − e−x )2
= −x
= x −x
·
3 + 4 sinh x 3 + (e − e )
2 x 2 e −e 3 + (ex − e−x )2

1 3
= x −x
1−
e −e 3 + (e − e−x )2
x

1 3 1 3
= − 3x = − .
ex − e−x e − e−3x 2 sinh x 2 sinh(3x)
Therefore
 ∞  ∞  ∞
x sinh x 1 x dx 3 x dx
dx = − .
0 3 + 4 sinh x2 4 0 sinh x 4 0 sinh(3x)

April 2022] PROBLEMS AND SOLUTIONS 391

A simple substitution gives
 ∞  ∞
x dx 1 x dx
= ,
0 sinh(3x) 9 0 sinh x
so
 ∞  ∞  ∞
x sinh x 1 x dx 1 x dx
dx = =
0 3 + 4 sinh x
2 6 0 sinh x 3 0 ex − e−x
 ∞
∞ ∞ 
1 1 ∞
= xe−(2k+1)x dx = xe−(2k+1)x dx
3 0 k=0
3 k=0 0

∞
∞ ∞

1  1 1  1  1
= = −
3 k=0
(2k + 1)2 3 k=1 k 2 k=1 (2k)2
 ∞
1 1  1 1 π2 π2
= 1− = · = .
3 4 k=1 k 2 4 6 24

Also solved by Z. Ahmed (India), T. Akhmetov (Russia), K. F. Andersen (Canada), F. R. Ataev (Uzbekistan),
S. Attaoui & M. Slimane (Algeria), M. Bataille (France), N. Batir (Turkey), A. Berkane (Algeria), N. Bhan-
dari (Nepal), R. Boukharfane (Saudi Arabia), P. Bracken, B. Bradie, V. Brunetti (India), C. Burnette, H. Chen,
B. E. Davis, T. Dickens, G. A. Edgar, G. Fera (Italy), P. Fulop (Hungary), M. L. Glasser, H. Grandmon-
tagne (France), N. Grivaux (France), J. A. Grzesik, E. A. Herman, N. Hodges (UK), F. Holland (Ireland),
E. J. Ionaşcu, W. Janous (Austria), J. E. Kampmeyer III, O. Kouba (Syria), K.-W. Lau (China), G. Lavau
(France), J. Magliano, S. Miao (China), A. Natian, K. Nelson, Q. M. Nguyen (Canada), C. R. Pranesachar
(India), V. Schindler (Germany), A. Stadler (Switzerland), S. M. Stewart (Australia), R. Stong, R. Tauraso
(Italy), D. Terr, D. B. Tyler, A. Tzarellas, E. I. Verriest, T. Wiandt, H. Widmer (Switzerland), Y. Xiang (China),
M. R. Yegan (Iran), L. Zhou, FAU Problem Solving Group, and the proposer.

Group Algebras With Invariant Subsets

12201 [2020, 660]. Proposed by Stephen M. Gagola, Jr., Kent State University, Kent, Ohio.
Let F be a field, and
 let G be a finite group. The group algebra F [G] is the vector space
of all formal sums g∈G ag g, where ag ∈ F , with multiplication defined by extending the
multiplication in G via the distributive laws. A subset S of F [G] is G-invariant if s ∈ S
 g ∈ G imply sg ∈ S. In particular, the subset G is G-invariant, as is the singleton set
and
{ g∈G g}. Find all fields F and groups G such that there exists an F -linear transformation
φ : F [G] → F [G] that is not right multiplication by an element of G but that nevertheless
sends every G-invariant subset to itself.
Solution by Kenneth Schilling, University of Michigan, Flint, MI. The field F must be the
field of order 2, and the group G must be a cyclic group of order 3, 4, or 5.
Let F [G] be a group algebra, and let φ : F [G] → F [G] be an F -linear transformation
that preserves G-invariant sets but is not right-multiplication by an element of G. Let e
be the identity element of G. It follows that the map ψ : F [G] → F [G] given by ψ(x) =
φ(x)(φ(e))−1 is also an F -linear transformation of F [G] that preserves G-invariant sets
but is not right-multiplication by an element of G and has the additional property that
ψ(e) = e. We may therefore assume henceforth without loss of generality that φ(e) = e.
Claim 1: For every finite subset {g1 , . . . , gk } of G, there exists h ∈ G such that

{φ(g1 ), . . . , φ(gk )} = {g1 h, . . . , gk h}.

In particular, φ maps G injectively into itself, and hence φ is injective on F [G].

392 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

Proof. Since G is G-invariant, φ(G) ⊂ G. Since the set {g1 h + · · · + gk h : h ∈ G} is also
G-invariant and contains g1 + · · · + gk , there exists h ∈ G such that

φ(g1 + · · · + gk ) = φ(g1 ) + · · · + φ(gk ) = g1 h + · · · + gk h .

The claim now follows from the fact that G is a linearly independent set in the vector space
F [G] and φ(gi ) and gi h belong to G for all i. 
Claim 2: For all g ∈ G, φ(g) ∈ {g, g −1 }.
Proof. For g ∈ G − {e}, Claim 1 implies that {e, φ(g)} = {h, gh} for some h ∈ G. Thus
either e = h and φ(g) = gh, in which case φ(g) = g, or e = gh and φ(g) = h, in which
case φ(g) = g −1 . 
−1
Claim 3: If φ(g1 ) = g1 and φ(g2 ) = g2 for distinct elements g1 , g2 ∈ G, then g1 = g2
or g1 = g22 or g2 = g12 .
Proof. By Claims 1 and 2, {φ(e), φ(g1 ), φ(g2 )} = {e, g1−1 , g2−1 } = {h, g1 h, g2 h} for some
h ∈ G. If e = h, then {e, g1−1 , g2−1 } = {e, g1 , g2 }, and g1 = g2−1 follows from φ(g2 ) =
g2−1 = g2 . If e = g1 h, then {e, g1−1 , g2−1 } = {g1−1 , e, g2 g1−1 }, so g2−1 = g2 g1−1 , which yields
g1 = g22 . By symmetry, g2 = g12 when e = g2 h. 
Claim 4: If φ(g1 ) = g1 and φ(g2 ) = g2 for g1 , g2 ∈ G − {e}, then g1 and e are the only
elements of G fixed by φ. Also, g12 = e, and g 2 = g1 for all g ∈ G − {e, g1 }.
Proof. By Claims 1 and 2, {e, φ(g1 ), φ(g2 )} = {e, g1 , g2−1 } = {h, g1 h, g2 h} for some h ∈ G.
If e = h, then g2−1 = g2 , which contradicts φ(g2 ) = g2 . If e = g1 h, then {e, g1 , g2−1 } =
{g1−1 , e, g2 g1−1 }. Since g2−1 = g1−1 , we have g1 = g1−1 and g2−1 = g2 g1−1 , so g12 = e and
g22 = g1 . If e = g2 h, then {e, g1 , g2−1 } = {g2−1 , g1 g2−1 , e}, so g1 = g1 g2−1 , which contradicts
g2 = e.
We conclude g12 = e and g1 = g22 . This implies that g1 is the only element of G − {e}
that is fixed by φ. Furthermore, g 2 = g1 for all g ∈ G − {e, g1 }. 
Claim 5: F is the field of order 2.
Proof. If F has an element a that is neither 0 nor 1, then let g be any element of G − {e}.
The set {h + agh : h ∈ G} is G-invariant, and e + ag is one of its elements, so there exists
h ∈ G such that φ(e + ag) = e + aφ(g) = h + agh. It follows that e = h and φ(g) = gh,
so φ(g) = g. In other words, φ is the identity transformation on G, and so also on F [G],
contrary to hypothesis. 
We now find all possible groups G.
First, suppose that G − {e} has elements g1 and g2 such that φ(g1 ) = g1 and φ(g2 ) =
g2−1 = g2 . By Claim 4, g24 = g12 = e, so the group g2  generated by g2 is a cyclic group of
order 4 and contains g1 , which equals g22 . Furthermore, we claim G = g2 . If there exists
h ∈ G − g2 , then φ(h) = h by Claim 4. Applying Claim 3 to g2 and h now yields either
g22 = h (forbidden by h ∈ g2 ) or h2 = g2 (forbidden by Claim 4 implying h2 = g1 ). With
G being a cylic group of order 4, it is easy to check that φ(g) = g −1 satisfies the required
conditions.
A second case is G = {e, g1 , g1−1 }, where φ(g) = g −1 for g ∈ G. Here, G is a cylic
group of order 3, and it is easy to check that φ(g) = g −1 satisfies the required conditions.
The only remaining case is that no element of G − {e} is fixed by φ, and G contains
at least two distinct pairs of inverse elements. Let g1 , g1−1 , g2 , g2−1 be distinct elements of
G. Assume without loss of generality that g2 = g12 . We know that g1−1 = g22 or g2 = g1−2 .
The second option is impossible (if true, then g2 = g2−1 , which would imply φ(g2 ) = g2 ),
so g1−1 = g22 . Therefore, g1−1 = g22 = g14 , and the order of g1 in G is 5. Furthermore, since
g2 = g12 and g1 = g2−2 , each of g1 , g2 belongs to the group generated by the other. Since
g1 , g2 were chosen arbitrarily, the entire group G is the group generated by g1 , a cyclic

April 2022] PROBLEMS AND SOLUTIONS 393

group of order 5. Once again it is easy to check that φ(g) = g −1 satisfies the required
conditions.
Editorial comment. Kenneth Schilling observed that the hypothesis that G is finite is not
needed, although the reference to the singleton set { g∈G g} in the problem statement does
not make sense without that hypothesis.
Also solved by N. Caro (Brazil), R. Chapman (UK), J. H. Lindsey II, and the proposer.

CLASSICS
We solicit contributions of classics from readers, who should include the problem state-
ment, solution, and references with their submission. The solution to the classic problem
published in one issue will appear in the subsequent issue.
C4. From the 1968 Moscow Mathematical Olympiad, contributed by the editors. A round
cake has icing on the top but not the bottom. Cut a piece of the cake in the usual shape of
a sector with vertex angle one radian and with vertex at the center of the cake. Remove the
piece, turn it upside down, and replace it in the cake to restore roundness. Next, move one
radian around the cake, cut another piece with the same vertex angle adjacent to the first,
remove it, turn it over, and replace it. Keep doing this, moving around the cake one radian
at a time, inverting each piece. Show that, after a finite number of steps, all the icing will
again be on the top.

The Game of Chomp

C3. Attributed to Frederik Schuh, contributed by the editors. Alice and Bob play a game
in which they take turns removing squares from an m-by-n grid of squares. We label the
square in row i and column j with the pair (i, j ). A legal move in this game consists
of selecting one of the remaining squares (i, j ) and removing all the squares (a, b) with
i ≤ a ≤ m and j ≤ b ≤ n that were not were not already removed by a previous move.
The players alternate moves, with Alice going first, and the player who removes the square
(1, 1) loses. Show that Alice has a winning strategy.
Solution. Since the game is finite, either Alice or Bob has a winning strategy. Suppose it is
Bob who has a winning strategy. If Alice removes just the single square (m, n) on her first
move, then Bob has a winning response (i, j ), leading to a position P from which Alice
has no winning response. But Alice could have selected square (i, j ) on her first move, and
this would have been a winning move for Alice, since it leaves Bob to play from position
P . This contradicts the assumption that Bob has a winning strategy, so it must be Alice
who has a winning strategy.
Editorial comment. The solution illustrates the concept of strategy stealing from combi-
natorial game theory. It demonstrates that Alice has a winning move to open the game,
although it does not tell her what that move is. Indeed, little is known about how Alice
should play. It is easy to see that Alice’s only winning opening move in the case m = 1 is
(1, 2) and in the case m = 2 is (2, n). When m = n, Alice’s only winning opening move
is (2, 2). Some progress on the m = 3 case is given in D. Zeilberger (2001), Three-rowed
Chomp, Adv. Appl. Math. 26, 168–179.
The game goes back to Frederick Schuh, whose version of the game is played on the
positive integers, with players alternately choosing divisors of a given integer, subject to
the restriction that no choice can be a multiple of a previous choice. The version of the
game that we have given here is due to David Gale. It is isomorphic to Schuh’s game in the
case that the integer is 2m 3n .

394 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

SOLUTIONS
The Polytope of Parking Functions
12191 [2020, 563]. Proposed by Richard Stanley, University of Miami, Coral Gables, FL. A
parking function of length n is a list (a1 , a2 , . . . , an ) of positive integers whose increasing
rearrangement b1 ≤ b2 ≤ · · · ≤ bn satisfies bi ≤ i. It is well known that the number of
parking functions of length n is (n + 1)n−1 . Let Pn denote the convex hull in Rn of all
parking functions of length n.
(a) Find the number of vertices of the convex polytope Pn .
(b) Find the number of (n − 1)-dimensional faces of Pn .
(c)* Find the number of integer points in Pn , i.e., the number of elements of Zn ∩ Pn . For
n ≤ 8 these numbers are 1, 3, 17, 144, 1623, 22804, 383415, 7501422.
(d)* Find the volume of Pn . For n ≤ 5 these volumes are 0, 1/2, 4, 159/4, 492.

286 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

Solution by Richard Stong, Center for Communications Research, San Diego, CA.
(a) Let a tight parking function be one whose increasing rearrangement consists of k copies
of 1 followed by the numbers k + 1 through n. Since parking functions remain park-
ing functions when coordinates are reordered, there are n!/k! parking functions with this
increasing rearrangement and hence nk=1 n!/k! tight parking functions of length n. The
sum evaluates to n!(e − 1) . We prove by induction on n that the tight parking functions
are exactly the vertices of Pn .
Suppose first that n occurs in the parking function a. By the reordering criterion, n can
only occur once. Every vertex of a face containing a in its interior must also have n in
the same place. Deleting n from a parking function of length n always leaves a parking
function of length n − 1. This applies both to a and to the vertices of any face containing
a. Thus if the parking function obtained by deleting n from a is a vertex in Pn−1 , then a
is a vertex of Pn . The converse holds as well. By the induction hypothesis, a is a vertex if
and only if it is tight.
Suppose next that n does not occur in a. If every position in a is 1, then a minimizes
the sum of entries over all parking functions. Hence it is a vertex; also it is tight. If some
position in a is not 1, then a is not tight. Pick a largest entry of a, and let a + and a − be the
results of replacing this entry with n or 1, respectively. These are both parking functions:
for a − we have only lowered b, and for a + we have only changed bn to n. Since the largest
entry was not 1 or n, a is in the interior of the segment joining a − and a + and hence is not
a vertex.
(b) There are 2n − 1 such faces. The faces of a polytope are the sets of points where some
linear function is maximized, and such a set is the convex hull of the vertices that achieve
the maximum. By the reordering property of parking functions, when a linear function
x → α · x is maximized at a the coordinate values for α and a will be in the same order.
That is, α · a = β · b, where β is the increasing rearrangement of α and b is the increasing
rearrangement of a.
Furthermore, if the first r entries of β are negative, then at a maximum the first r entries
of b are all 1. Similarly, if the last s entries of β are positive, then at a maximum the last s
entries of b are (n + 1 − s, . . . , n) (after possibly re-sorting the places where β has a run
of equal entries). That is, if β has m equal positive entries, then those m entries of b are m
consecutive integers in some order; in particular, the sum of those m entries is fixed.
Putting this together, we see that if α has r negative entries and t distinct positive values,
then the set of points a maximizing α · x has codimension at least r + t (we fix one entry
for each negative entry in α and one sum of entries for each positive value).
Thus (n − 1)-dimensional faces must correspond to α with r + t = 1. Up to rescaling,
faces must correspond either to α being 1 on some set S of coordinates and 0 elsewhere
(which we denote by αS ), or to α being −1 in one coordinate and 0 elsewhere.  
If |S| = n − 1, then the codimension-1 hyperplane αS · x = n|S| − |S| 2
passes through
all the vertices whose coordinates in the S positions are a permutation of {n+1−|S|, . . . , n}
and in the other positions are any parking function of length n − |S|. Thus we obtain a
codimension-1 face that is isometric to the product Pn−|S| × Q|S| , where Qk is the convex
hull of the points that are permutations of (1, 2, . . . , k). Note that Pn−|S| has dimension
n − |S|, since n − |S| = 1, and Q|S| has dimension |S| − 1. The product has dimension
n − 1. If |S| = n − 1, then since P1 is only a single point we obtain a face of codimension
2 and dimension n − 2, contributing nothing to our count.
If α has a single −1 and zeroes elsewhere, then we get the face of codimension 1 (and
dimension n − 1) where that coordinate is fixed to 1.
The first case gives 2n − n − 1 faces (corresponding to nonempty subsets of coordinates
with size other than n − 1), and the second case gives n faces. Hence the number of faces
of Pn is 2n − 1.

March 2022] PROBLEMS AND SOLUTIONS 287

(c) No solution is available.
(d) Letting Vn denote the n-dimensional volume of Pn , we prove
n
n − 1 nn−s 
s
s
Vn = (−1)s−m (2m − 1)!!.
s=1
s − 1 2 m=0
s m
Let Wk denote the (k − 1)-dimensional volume of thepolytope
 Qk in part (b). We first
derive a closed formula for Wk . The polytope Qk has kr faces isometric to Qr × Qk−r ,
corresponding to fixing r coordinates with sum r(r + 1)/2, leaving the remaining k − r
coordinates to sum to k(k + 1)/2 − r(r + 1)/2, which equals (k − r)(k + r + 1)/2. (The
proof of this is essentially the same as part (b) above.)
The distance from the center ((k + 1)/2, (k + 1)/2, . . . , (k + 1)/2) of Qk to the plane
of such a face is
r(k − r)2 /4 + (k − r)r 2 /4 = kr(k − r)/2.
Hence
1  k
k−1
Wk = Wr Wk−r kr(k − r).
2(k − 1) r=1 r

This recurrence yields Wk = k k−3/2 using induction and the identity


k−1
k r−1
2(k − 1)k k−2 = r (k − r)k−r−1 . (∗)
r=1
r
The identity (∗) appears in the book of Lovász (Combinatorial Problems and Exercises,
North-Holland, 1979). It has both an analytic proof using generating functions and a bijec-
tive proof (due to L. Smiley) using Cayley’s formula, which states that there are k k−2 trees
with vertex set [k], where [k] = {1, . . . , k}. With n ways to distinguish one vertex as a root,
there are k k−1 rooted trees with vertex set [k]. Both sides of the identity count the ordered
pairs of rooted trees whose vertex sets have union [k].
Splitting Pn into cones with vertex at the point (1, , . . . , 1), and invoking  the solution
of part (b), we see that Pn is the union, over values of k other than n − 1, of nk cones with
√
base Pn−k × Qk and height k(2n − k − 1)/2. Thus
√
1 n 1  n
n n
k(2n − k − 1)
Vn = Vn−k Wk = Vn−k k k−1 (2n − k − 1).
n k=1 k 2 2n k=1 k
In this sum, we have included the term for k = n − 1, but the computation remains correct
since V1 = 0. Let V be the exponential generating function of the sequence Vn , so
∞
 ∞
1  n
n
Vn zn
V (z) = zn = Vn−k k k−1 (2n − k − 1) . (∗∗)
n=0
n! n=1
2n k=1 k n!

Let F (z) = ∞ n=1 n z /n!. Differentiating (∗∗) and breaking the factor 2n − k − 1
n−1 n

into three pieces, we obtain

∞  n
n zn−1
2V  (z) = Vn−k k k−1 (2n − 2k + k − 1)
n=1 k=1
k n!

F (z)V (z)
= 2F (z)V  (z) + F  (z)V (z) −
z
F (z)
= 2F (z)V  (z) + F  (z) − V (z). (∗∗∗)
z

288 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

We next study F  − F /z. Again using Cayley’s formula, F is the exponential generating
function (EGF) for rooted labeled trees: there are nn−1 with vertex set [n]. To form such a
rooted tree, one chooses a root label and a rooted forest on the remaining labels, with any
number of components. The EGF for choosing the root is just z, and the two choices are
enumerated by the product of the EGFs, which yields the standard relation F = zeF (from
which Cayley’s formula can be obtained by Lagrange inversion).
Taking the logarithm of F = zeF yields log F = F + log z, and differentiating yields
F /F = F  + 1/z, or F  − F /z = F F  . Equation (∗∗∗) then becomes


V FF
= .
V 2(1 − F )
Integrating yields log V = − (F + log(1 − F )) /2, and hence
e−F (z)/2
V (z) = √ .
1 − F (z)
We now have both a recurrence and an EGF for Vn , and we have left the realm of geom-
etry. A more explicit formula for Vn as a double sum can be derived from the generating
function. The standard expansions of ex and (1 − 4x)−1/2 yield
∞ ∞ ∞ 
(−1)k F k  2m F m  s
e−F /2 (−1)s−m 2m s
√ = = F .
1−F k=0
k!2k
m=0
m 4m
s=0 m=0
(s − m)!2s+m m

Also, the series expansion for F s is known to be

∞
sr r−s−1 r
F s (z) = z,
r=s
(r − s)!

since the coefficient [zr ]F s (z) of zr in F s (z) is given by

 
1 F s (z) 1 (1 − F )erF
r+1
dz = dF = [F r−s ](1 − F )erF
2π i z 2π i F r+1−s
r r−s r r−s−1 sr r−s−1
= − = .
(r − s)! (r − s − 1)! (r − s)!
Finally, set r√= n and plug this expression for the coefficient of zn in F s into the expan-
sion of e−F /2 / 1 − F in terms of F . Since we defined V to be an EGF, we seek the
coefficient of zn /n! and hence must introduce n! also into the numerator. After a little
algebra, we read off the formula
n
n − 1 nn−s 
s
s
Vn = (−1)s−m (2m − 1)!!.
s=1
s − 1 2 m=0
s m

Editorial comment. The inner sum in the formula for Vn is the well-known inclusion-
exclusion formula for the number of ways to form s couples into pairs of people with
no couple paired (sequence A053871 in the OEIS). Also, the generating function for V
and standard techniques yield
21/4 π 1/2  
Vn ∼ · nn−1/4 1 + O(n−1/2 ) .
(1/4)e1/2

Parts (a) and (b) also solved by A. Amanbayeva & D. Wang and the proposer.

March 2022] PROBLEMS AND SOLUTIONS 289

 An Integral Bound
12193 [2020, 564]. Proposed by Florin Stanescu, Serban Cioculescu School, Gaesti,
Romania. Suppose that f : [0, 1] → R has a continuous third derivative and f (0) = f (1).
Prove
 1 
  (k − 1)k!(k − 1)!  
 f (x)x (1 − x) dx  ≤
 k−1 k−1
max f  (x)
 6 (2k + 1)! 0≤x≤1
0

where k is a positive integer.

Solution by Koopa Tak Lun Koo, Chinese STEAM Academy, Hong Kong, China. We pro-
ceed by induction on k. For the base case k = 1, the left side is |f (1) − f (0)| = 0 and the
inequality is immediate.
1
For the inductive step, let gk (x) = x k (1 − x)k and Ik = 0 f  (x)gk−1 (x) dx. One easily
checks that
gk (x) = −2k(2k − 1)gk−1 (x) + k(k − 1)gk−2 (x). (∗)
For k ≥ 2, gk (0) = gk (1) = gk (0) = gk (1) = 0, so integrating by parts twice yields
1  1  1 1
f  (x)gk (x) dx = f  (x)gk (x) − f  (x)gk (x) + f  (x)gk (x) dx
0 0 0 0
1
= f  (x)gk (x) dx.
0

Using (∗) this yields

1
−2k(2k − 1)Ik + k(k − 1)Ik−1 = f  (x)gk (x) dx.
0

From the triangle inequality and gk (x) ≥ 0 for x ∈ [0, 1], we get
 1 
 

2k(2k − 1) |Ik | ≤ k(k − 1) |Ik−1 | +  f (x)gk (x) dx 

0
 
 1 
≤ k(k − 1) |Ik−1 | + max |f (x)|  
gk (x) dx  .
0≤x≤1 0
1
Recognizing 0 gk (x) dx as a beta integral, we have
1
gk (x) dx = B(k + 1, k + 1) = (k!)2 /(2k + 1)!.
0

Using this together with the induction hypothesis gives

(k − 2)k!(k − 1)! (k!)2
2k(2k − 1) |Ik | ≤ max |f  (x)| + max |f  (x)|,
6(2k − 1)! 0≤x≤1 (2k + 1)! 0≤x≤1
which after simplifying becomes
(k − 1)k!(k − 1)!
|Ik | ≤ max |f  (x)|,
6(2k + 1)! 0≤x≤1

completing the induction.

Also solved by K. F. Andersen (Canada), A. Berkane (Algeria), P. Bracken, R. Chapman (UK), C. Chiser
(Romania), R. Guadalupe (Philippines), F. Holland (Ireland), O. Kouba (Syria), K.-W. Lau (China), J. H. Lind-
sey II, M. Omarjee (France), A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), E. I. Verriest, L. Zhou, and
the proposer.

290 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

 Regular Polygons Inscribed in a Cube
12195 [2020, 659]. Proposed by Joseph DeVincentis, Salem, MA, James Tilley, Bedford
Corners, NY, and Stan Wagon, Macalester College, St. Paul, MN. For which integers n
with n ≥ 3 can a regular n-gon be inscribed in a cube? The vertices of the n-gon must all
lie on the cube but may not all lie on a single face.
Composite solution by Eugen J. Ionaşcu, Columbus State University, Columbus, GA, and
Yury J. Ionin, Champaign, IL. An inscribed n-gon exists if and only if 3 ≤ n ≤ 9 or n = 12.
We work in the standard unit cube. We first show that regular n-gons embed in the cube
for n ∈ {3, 4, 6, 8, 12}.
For n = 3, corners (1, 0, 0), (0, 1, 0), and (0, 0, 1) yield an equilateral triangle.
For n = 4, points (0, 0, 1/2), (1, 0, 1/2), (1, 1, 1/2), and (0, 1, 1/2) determine a square
embedded in the cube. Truncating it yields a regular octagon with all vertices on the faces
of the cube, which takes care of n = 8.
For n = 6, points (1/2, 0, 1), (0, 1/2, 1), (0, 1, 1/2), (1/2, 1, 0), (1, 1/2, 0), and
(1, 0, 1/2) determine a regular hexagon embedded in the cube. Truncating it yields a
regular 12-gon with all vertices on the faces of the cube, which takes care of n = 12.
Next, we give constructions for n ∈ {5, 7, 9}, showing that a regular n-gon can be
inscribed in a polygon embedded in the cube.

For n = 5, we start with a regular pentagon ABCDE. Let lines AB and DE intersect
at Q, and let the line through C perpendicular to CQ intersect lines AB and DE at R
and S, respectively, as in Figure 1a. The isosceles triangle QRS has apex angle π/5. With
Q = (0, 0, a), R = (0, b, 0), and S = (b, 0, 0), the apex angle of QRS has cosine equal to
a 2 /(a 2 + b2 ). We may choose real numbers a, b ∈ (0, 1) such that this equals cos(π/5).
For n = 7, we start with a regular heptagon ABCDEF G. Let the line through A paral-
lel to DE intersect lines F G and BC at Q and R, respectively. Let line DE intersect lines
BC and F G at S and T , respectively, as in Figure 1b. Now QRST is an isosceles trapezoid
with acute angles 3π/7. An isosceles trapezoid is uniquely determined, up to similarity, by
the measure of its acute angles and the ratio k of the shorter base to the longer base. By the
law of sines,
QR 2 sin(2π/7)/ sin(4π/7) 2 sin(2π/7)
k= = = .
ST 1 + 2 sin(2π/7)/ sin(3π/7) sin(3π/7) + 2 sin(2π/7)
Set T = (a, 0, 0), S = (0, a, 0), Q = (ka, 0, 1), and R = (0, ka, 1). It is required that
a ∈ (0, 1) satisfies
−→ −→
TQ·TS (1 − k)a
cos(3π/7) = =√ .
TQ·TS 2 · (1 − k)2 a 2 + 1

March 2022] PROBLEMS AND SOLUTIONS 291

Solving for a yields
√
2 cos(3π/7)
a= ≈ 0.8633.
(1 − k) 1 − 2 cos2 (3π/7)
For n = 9, first observe that for 0 < a < 1, the plane x − y + z = a intersects the
cube in a hexagon QRST U V , where Q = (0, 0, a), R = (0, 1 − a, 1), S = (a, 1, 1),
T = (1,√1, a), U = (1, 1 − a, 0), and V √
= (a, 0, 0). We compute that QR = ST = U V =
(1 − a) 2, that QV = RS = T U = a 2, and that all six angles are equal. Hence they
measure 2π/3. Let B, E, and H be the midpoints of RS, T U , and QV , respectively. Let
points A and I on QR, C and D on ST , and F and G on U V be such that angles ABR,
CBS, DET , F EU , GH V , and I H Q, each measure π/9 (see Figure 1c). All the angles
of nonagon ABCDEF√GH I measure 7π/9. Finally, by the law of sines, six sides of the
nonagon have length a 6/(4 sin(2π/9)) and three sides have length
√ a sin(π/9)
(1 − a) 2 − 2 · √ .
2 sin(2π/9)
Setting the two lengths equal and solving for a yields
4 sin(2π/9)
a=√ ≈ 0.4534.
3 + 4 sin(π/9) + 4 sin(2π/9)
We conclude by showing the impossibility of inscribing a regular n-gon for n > 12 and
n ∈ {10, 11}.
The vertices of a regular n-gon inscribed in the unit cube lie in the intersection of the
cube with the plane containing the n-gon. Thus a face of the cube contains at most two
vertices of the n-gon, which yields n ≤ 12.
Next, we exclude n = 11. Since no two sides of a regular 11-gon are parallel, opposite
faces of the cube together contain at most three vertices of the 11-gon, but this limits the
number of vertices to 9.
Finally, for n = 10, consider a regular inscribed
10-gon ABCDEF GH I J . Since it lies in the
intersection of a plane P with the cube, opposite
faces of the cube cannot together contain exactly
three vertices. Any four vertices on opposite faces
must form opposite sides of the 10-gon. Also, the
vertices of opposite sides of the 10-gon must form
a rectangle. Thus the intersection of P with the
cube must be a hexagon QRST U V with opposite
sides parallel. We may assume that P intersects the
plane z = 1 in QR and the xy-plane in T U with
Figure 2.
the 10-gon inscribed as in Figure 2.
The distance between sides QV and ST is the same as between sides RS and U V .
This shows that the dihedral angle α between P and the yz-plane equals the dihedral angle
β between P and the xz-plane. Consequently,
√ P is symmetric with respect to the plane
x = y. This symmetry implies SV = 2. Because triangles CDS and H I V are isosceles,
the center O of the 10-gon is at the midpoint of SV .
We calculate the circumradius r = OC using the law of sines as follows:
sin(3π/10) 1
r = OS = √ .
sin(3π/5) 2 2 cos(3π/10)

292 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

Let γ denote the dihedral angle between P and the xy-plane. It satisfies
2
√
1 5−1
cos γ = 1 − sin γ = 1 −
2 2
= 1 − 2 cos (3π/10) = cos(2π/5) =
2
.
2r 4
The distance between sides RS and U V is 2r cos(π/10), so
2 cos2 (3π/10)  2 √
cos2 α = cos2 β = 1−sin2 α = 1− 2
= 1−2 4 cos2 (π/10)−3 = 5−2.
cos (π/10)
It is well known and easy to prove that if a plane has dihedral angles α, β, and γ with the
yz-, xz-, and xy-planes, then

cos2 α + cos2 β + cos2 γ = 1.

This yields a contradiction, because

√
9 5 − 17
2 cos α + cos γ =
2 2
= 1.
4

Editorial comment. A few solvers interpreted the problem as requiring that the entire n-gon
be embedded in the cube, which is possible if and only if n = 3, 4, 6.
Also solved by R. Stong and the proposers.

CLASSICS
We solicit contributions of classics from readers, who should include the problem state-
ment, solution, and references with their submission. The solution to the classic problem
published in one issue will appear in the subsequent issue.
C3. Attributed to Frederik Schuh, contributed by the editors. Alice and Bob play a game
in which they take turns removing squares from an m-by-n grid of squares. We label the
square in row i and column j with the pair (i, j ). A legal move in this game consists
of selecting one of the remaining squares (i, j ) and removing all the squares (a, b) with
i ≤ a ≤ m and j ≤ b ≤ n that were not were not already removed by a previous move.
The players alternate moves, with Alice going first, and the player who removes the square
(1, 1) loses. Show that Alice has a winning strategy.

A Curious Characterization of the Fibonacci Numbers

C2. Ira Gessel [1972], contributed by the editors. Prove that a positive integer n is a
Fibonacci number if and only if either 5n2 + 4 or 5n2 − 4 is a perfect square.
Solution. The Fibonacci numbers are defined by: F0 = 0, F1 = 1, and Fk+2 = Fk + Fk+1
when k ≥ 0. Using the well-known identity Fk−1 Fk+1 = Fk2 + (−1)k , we obtain

5Fk2 + (−1)k 4 = 5Fk2 + 4(Fk−1 Fk+1 − Fk2 )

= (Fk+1 − Fk−1 )2 + 4Fk−1 Fk+1 = (Fk+1 + Fk−1 )2 .

This shows that 5n2 + 4 or 5n2 − 4 is a perfect square when n is Fibonacci.

For the converse, we prove that if m and n are positive integers satisfying 5n2 ± 4 = m2 ,
then there exists some positive integer k such that n = Fk and m = Fk−1 + Fk+1 .

March 2022] PROBLEMS AND SOLUTIONS 293

 The proof is by induction on n. For n = 1, there are two cases: Either m = 1, in which
case n = F1 and m = F0 + F2 , or m = 3, in which case n = F2 and m = F1 + F3 .
For the induction step, suppose n ≥ 2, the result holds for smaller values of n, and for
some positive integer m, 5n2 ± 4 = m2 . Note that
m2 ≤ 5n2 + 4 ≤ 5n2 + n2 = 6n2 < 9n2 ,
so m < 3n. Also
m2 ≥ 5n2 − 4 ≥ 5n2 − n2 = 4n2 ,
so m ≥ 2n.
Let n1 = (m − n)/2. Since the parities of n and m are the same, n1 is an integer, and
from 2n ≤ m < 3n we get n/2 ≤ n1 < n. Let m1 = (5n − m)/2. Again we see that m1
is an integer and m1 > (5n − 3n)/2 = n. So n1 and m1 are positive integers and n1 < n.
Also:
5(n2 − 2nm + m2 ) 5(6n2 ± 4 − 2nm) 15n2 − 5nm
5n21 = = = ± 5,
4 4 2
25n2 − 10nm + m2 30n2 ± 4 − 10nm 15n2 − 5nm
m21 = = = ± 1.
4 4 2
It follows that 5n21 ∓ 4 = m21 . By the induction hypothesis, there is a positive integer k such
that n1 = Fk and m1 = Fk−1 + Fk+1 .
From the equations n1 = (m − n)/2 and m1 = (5n − m)/2, we get
n1 + m1 Fk + Fk−1 + Fk+1 2Fk+1
n= = = = Fk+1 and
2 2 2
5n1 + m1 5Fk + Fk−1 + Fk+1 2Fk + 2Fk+2
m= = = = Fk + Fk+2 .
2 2 2
Editorial comment. The problem appeared as Problem H-187 in Fibonacci Quarterly 10
(1972) 417–419. The equation 5n2 ± 4 = m2 can be rearranged to read m2 − 5n2 = ±4,
which is a variant of Pell’s equation, and our proof that n in this equation must be a
Fibonacci number is based on a standard method for solving Pell’s equation. An alter-
native way to prove that n is a Fibonacci number is to let j = (m + n)/2 and then √show
that gcd(j, n) = 1 and |j/n − φ| < 1/(2n2 ), where φ is the golden mean (1 + 5)/2.
It follows that j/n is a convergent of the continued fraction for φ, and it is well known
that these convergents are ratios of successive Fibonacci numbers (see G. H. Hardy and
E. M. Wright (2008), An Introduction to the Theory of Numbers, 6th ed., Oxford: Oxford
√ Yet another proof begins by rewriting 5n ± 4 = m in the form
2 2
Univ. √Press, pp. 190, 196).
(m + 5n)/2 · (m − 5)/2 = ±1 and then using the fact that any unit in the ring Z[φ] is
of the form ±φ k .
There is a connection to Hilbert’s tenth problem about Diophantine equations. A set
X ⊂ Nr is called Diophantine if there is a polynomial p with integer coefficients in r + s
variables such that a ∈ X if and only if there exists b ∈ Ns such that p(a, b) = 0. This
problem shows that the set of Fibonacci numbers is Diophantine, by setting p(x, y) =
(5x 2 + 4 − y 2 )(5x 2 − 4 − y 2 ). In 1961, Martin Davis, Hilary Putnam, and Julia Robinson
showed that a negative answer to Hilbert’s tenth problem follows from the existence of
a Diophantine set of the form {(n, f (n)) : n ∈ N}, where f has exponential growth. In
1970, Y. V. Matiyasevich showed that the set {(n, F2n ) : n ∈ N} is Diophantine, settling
Hilbert’s problem. It is not hard to use this to prove that {(n, Fn ) : n ∈ N} is Diophantine.
The full story can be found in M. R. Davis (1973), Hilbert’s tenth problem is unsolvable,
Amer. Math. Monthly 80, 233–269.

294 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129

SOLUTIONS

Evaluating an Integral with Leibniz’s Help

12184 [2020, 461]. Proposed by Paolo Perfetti, Universitá degli Studi di Roma “Tor Ver-
gata,” Rome, Italy. Prove
 ∞ √
ln(x 4 − 2x 2 + 2)
√ dx = π ln(2 + 2).
1 x x2 − 1

Solution by Warren P. Johnson, Connecticut College, New London, CT. For positive num-
bers a and b, we consider the integral
 π/2
I (a, b) = ln(a 2 cos2 θ + b2 sin2 θ ) dθ.
0

By substituting θ = π/2 − φ, we see that I (a, b) = I (b, a). The Leibniz integral rule
yields
 π/2  π/2
∂I 2a cos2 θ dθ ∂I 2b sin2 θ dθ
= and = , (1)
∂a 0 a 2 cos2 θ + b2 sin2 θ ∂b 0 a 2 cos2 θ + b2 sin2 θ
and it follows that
 π/2
∂I ∂I
a +b = 2 dθ = π. (2)
∂a ∂b 0

Also, using the substitution b tan θ = a tan φ we see that

 π/2
∂I ∂I 2ab dθ
b +a =
∂a ∂b 0 a cos θ + b2 sin2 θ
2 2
 π/2  π/2
2ab sec2 θ dθ
= = 2 dφ = π. (3)
0 a 2 + b2 tan2 θ 0

When a = b, the solution to (2) and (3) is

∂I ∂I π
= = ,
∂a ∂b a+b
and it is easily checked from (1) that this is also correct when a = b.

February 2022] PROBLEMS AND SOLUTIONS 187

 Since I is symmetric in a and b, integrating with respect to either a or b gives I (a, b) =
π ln(a + b) + K for some constant K. Setting b = a we find
K = I (a, a) − π ln(2a) = π ln a − π ln(2a) = −π ln 2,
so   
π/2
a+b
I (a, b) = ln a 2 cos2 θ + b2 sin2 θ dθ = π ln . (4)
0 2
From this we can derive the well-known integral
 π/2 
1 π/2
ln(cos θ ) dθ = lim ln(cos2 θ + b2 sin2 θ ) dθ
b→0+ 2 0
0
 
π 1+b π
= lim ln = − ln 2. (5)
b→0+ 2 2 2
(We omit the justification of this limit calculation, since the result is well known.) Com-
bining (4) and (5) we have
 π/2  π/2
ln(a 2 + b2 tan2 θ ) dθ = ln(a 2 cos2 θ + b2 sin2 θ ) − 2 ln(cos θ ) dθ
0 0
 
a+b
= π ln + π ln 2 = π ln(a + b). (6)
2
With this in hand,√
we turn to the integral in the problem, which we denote by P . Using
the substitution u = x 2 − 1, we obtain
 ∞
ln 1 + u4
P = du.
0 u2 + 1
The further substitution v = 1/u shows that we also have
 ∞
ln(1 + 1/u4 )
P = du,
0 u2 + 1
and averaging these two expressions yields
 ∞
ln(u2 + 1/u2 )
P = du.
0 u2 + 1
Now substitute v = u − 1/u to get
 ∞  ∞
ln v 2 + 2 ln v 2 + 2
P = dv = 2 dv.
−∞ v +4
2
0 v2 + 4
Finally, substituting v = 2 tan θ yields
 π/2
P = ln 2 + 4 tan2 θ dθ,
0
√
which by (6) is π ln(2 + 2).
Also solved by Z. Ahmed (India), K. F. Andersen (Canada), F. R. Ataev (Uzbekistan), M. Bataille (France),
N. Batir (Turkey), A. Berkane (Algeria), N. Bhandari (Nepal), K. N. Boyadzhiev, P. Bracken, B. Bradie,
B. S. Burdick, W. Chang, R. Chapman (UK), H. Chen, Ó. Ciaurri (Spain), B. E. Davis, P. De & B. Sury
(India), A. Eydelzon, G. Fera (Italy), P. Fulop (Hungary), M. L. Glasser, H. Grandmontagne (France), N. Gri-
vaux (France), J. A. Grzesik, L. Han, E. A. Herman, N. Hodges (UK), E. J. Ionaşcu, W. Janous (Austria),
O. Kouba (Syria), K.-W. Lau (China), G. Lavau (France), K. Mahanta (India), L. Matejı́c̆ka (Slovakia), K. Nel-
son, Q. M. Nguyen (Canada), M. Omarjee (France), M. A. Prasad (India), K. Sarma (India), V. Schindler

188 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
(Germany), S. Sharma (India), F. Sinani (Kosovo), A. Stadler (Switzerland), A. Stenger, S. M. Stewart (Aus-
tralia), R. Stong, R. Tauraso (Italy), E. I. Verriest, M. Vowe (Switzerland), T. Wiandt, H. Widmer (Switzerland),
Y. Xiang (China), M. R. Yegan (Iran), FAU Problem Solving Group, and the proposer.

A Class of Matrices with Determinant 1

12185 [2020, 659]. Proposed by George Stoica, Saint John, NB, Canada. Let n1 , . . . , nk
be pairwise relatively prime odd integers greater than 1. For i ∈ {1, . . . k}, let fi (x) =
ni
m=1 x
m−1
. Let A be a 2k-by-2k matrix with real entries such that det fj (A) = 0 for all
j ∈ {1, . . . , k}. Prove det A = 1.
Solution by Nicolás Caro, Universidade Federal de Pernambuco, Recife, Brazil. For each i,
the set Ui of complex roots of the polynomial fi consists precisely of the ni th roots of unity
other than 1. When i = j , there exist integers r and s such that rni + snj = 1, and so if
λ ∈ C satisfies λni = λnj = 1, then λ = (λni )r (λnj )s = 1. Thus the sets U1 , . . . , Uk are
pairwise disjoint. Moreover, λ ∈ Ui implies λ ∈ Ui and λ = λ (because ni is odd and
greater than 1), and of course λλ = 1.
By the spectral mapping theorem, for each j there exists an eigenvalue λj of A such
that fj (λj ) = 0, that is λj ∈ Uj . Since A is a real matrix, λj is also an eigenvalue of A,
and therefore the 2k values λ1 , . . . , λk , λ1 , . . . , λk are precisely the eigenvalues of A. Since
det A is equal to the product of these eigenvalues, the determinant is 1.
Also solved by K. F. Andersen (Canada), R. Chapman (UK), J.-P. Grivaux (France), E. A. German, R. A. Horn,
O. Kouba (Syria), G. Lavau (France), S. Miao (China), É. Pité, K. Sarma (India), A. Stadler (Switzerland),
A. Stenger, R. Stong, B. Sury (India), E. I. Verriest, and the proposer.

A Median and Symmedian Produce Perpendicular Lines

12187 [2020, 462]. Proposed by Khakimboy Egamberganov, Sorbonne University, Paris,
France. Given a scalene triangle ABC, let M be the midpoint of BC, and let m and s
denote the median and symmedian lines, respectively, from A. (The symmedian line from
A is the reflection of the median from A across the angle bisector from A.) Let K be the
projection of C onto m, and let L be the projection of B onto s. Let P be the intersection
of BL and CK, and let Q be the intersection of KL and BC. Prove that P M and AQ are
perpendicular.
Solution by Haoran Chen, Jiangsu, China. We use a coordinate system in which A is the
origin and the bisector of the angle at A is the positive x-axis. Thus the coordinates of B and
C are (b, kb) and (c, −kc), respectively, for some b, c, and k, where b, c > 0 and k = 0.
Since the triangle is scalene, b = c. The coordinates of M are ((b + c)/2, k(b − c)/2), so
the equations of m and s are y = λx and y = −λx, respectively, where
k(b − c)
λ= .
b+c
The line through C perpendicular to m has slope −1/λ, and therefore its equation is
x−c
y + kc = − . (1)
λ
Intersecting this line with m, we find that
 
c(1 − kλ) cλ(1 − kλ)
K= , .
λ2 + 1 λ2 + 1
Similarly, the equation of the line through B perpendicular to s is
x−b
y − kb = , (2)
λ

February 2022] PROBLEMS AND SOLUTIONS 189

and therefore
 
b(1 − kλ) bλ(1 − kλ)
L= ,− .
λ2 + 1 λ2 + 1
Equations (1) and (2) are the equations of the lines CK and BL, and intersecting them we
find that
 
(b + c)(1 − kλ) (c − b)(1 − kλ)
P = , .
2 2λ
If kλ = 1, then K = L = A = (0, 0), but the statement of the problem presupposes that
K and L determine a line. We therefore assume kλ = 1. Intersecting the lines KL and BC
we obtain, after some calculation,
 
2bc 2bck 2 λ
Q= ,− .
(b + c)(λ2 + 1) (b + c)(λ2 + 1)
Finally, using the coordinates for P , M, A, and Q, we compute
b−c 1
slope of P M = = 2 ,
kλ2 (b + c) k λ
slope of AQ = −k 2 λ,

and the conclusion follows.

Editorial comment. It is not necessary that ABC be scalene; all that is required is the
condition AB = AC.
There are some other interesting geometrical relationships in the configuration in this
problem. Using the coordinates given above, we can compute
(b + c)λ
slope of KL = = −k,
c−b
(c − b) 1
slope of AP = =− .
λ(b + c) k
It follows that KL AC and AP ⊥ AB.
The case kλ = 1, which was excluded in the solution above, occurs when ∠CAM is a
right angle. The configuration of the points and lines in this problem varies significantly
depending on whether ∠CAM is acute or obtuse and whether or not m and s are perpen-
dicular. A few solvers gave synthetic solutions that were not completely general because
they did not take into account the full range of possible configurations. Most solvers used
analytic methods.
The proposer’s solution shows that AQ is the radical axis of the circles with diameters
AP and AM. This implies that AQ is perpendicular to the line through the centers of these
two circles, which is parallel to P M.
Also solved by J. Chen (China), C. Curtis, G. Fera (Italy), J.-P. Grivaux (France), N. Hodges (UK), W. Janous
(Austria), J. H. Lindsey II, C. R. Pranesachar (India), V. Schindler (Germany), A. Stadler (Switzerland),
R. Stong, T. Wiandt, L. Zhou, and the proposer.

Perfect Paths through the Positive Integers

12188 [2020, 563]. Proposed by H. A. ShahAli, Tehran, Iran.
(a) Is there a permutation of the positive integers with the property that every pair of con-
secutive elements sums to a perfect square?

190 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
(b)* Is there a permutation of the positive integers with the property that every pair of
consecutive elements sums to a perfect cube?
Solution by Texas State University Problem Solvers, San Marcos, TX. The answer to both
questions is yes. We prove the more general claim that for every k ∈ N there is a permuta-
tion of N such that every pair of consecutive elements sums to a perfect kth power. This is
trivial for k = 1, so consider k ≥ 2.
Let G be the graph with vertex set N in which u and v are adjacent when u + v is a
kth power. It suffices to find an infinite path n1 , n2 , . . . through G that visits every vertex
exactly once. For u ∈ N, let Gu be the subgraph of G induced by {n ∈ N : n ≥ u}. For
x, y ∈ N, write x → y when Gx has a path from x to y.
We first prove the xyz-property: If y, z ∈ V (Gx ), then x → z and y → z imply x → y.
This holds because the actual edges of G in a path in Gu witnessing u → v are undirected.
Following a path from x to z and then a path from z to y in G yields a walk from x to y
in G, which contains a path from x to y. Furthermore, since the edges came from Gx and
Gy , they all lie in Gx , so x → y.
We prove v → v + k! for every positive integer v and then use this to show v → v + 1
as well, establishing that Gv is connected for every v ∈ N. We then inductively construct
the desired path.
Define polynomials g1 , . . . , gk by g1 (m) = (m + 1)k − mk and gj (m) =
gj −1 (m + 1) − gj −1 (m) for 2 ≤ j ≤ k. Note inductively that gj is a polynomial of
j −1
degree k − j with leading coefficient i=0 (k − i), and all of its coefficients are non-
negative. In particular, gk (m) = k!. Also define polynomials f1 , . . . , fk by f1 (m) = 0 and
j
fj (m) = i=2 gi (m) for 2 ≤ j ≤ k. Note that fj +1 is a polynomial of degree k − 2 when
1 ≤ j < k. Since gi (n) ≥ 0 for all n ∈ N, we have 0 ≤ fj (m) ≤ fj +1 (m) for m ∈ N and
1 ≤ j ≤ k − 1. Choose M ∈ N so that g1 (m) > 2fk (m + 1) when m ≥ M, which we can
do since g1 has higher degree than fk .
Given 1 ≤ i ≤ k, we now prove by induction on i that v → v + gi (m) when m and v are
distinct positive integers such that m ≥ M and mk > 2v + 2fi (m). For i = 1, the condition
is mk > 2v, and the list (v, mk − v, (m + 1)k − mk + v) provides a path of length 2 from
v to v + g1 (v) in Gv , yielding v → v + g1 (m).
Now consider i > 1, with m ≥ M and mk > 2v + 2fi (m). Since fi (m) ≥ 0 and
g1 (m) > fk (m + 1) ≥ fi−1 (m + 1), we have
(m + 1)k = mk + g1 (m) > 2v + 2fi−1 (m + 1),
so v → v + gi−1 (m + 1) by applying the hypothesis for i − 1 to m + 1 and v. Also,
mk > 2v + 2fi (m)) = 2(v + gi (m)) + 2fi−1 (m).
This allows us to apply the hypothesis for i − 1 to m and v + gi (m) to obtain (v +
gi (m)) → (v + gi (m) + gi−1 (m)). Since gi (m) + gi−1 (m) = gi−1 (m + 1), this becomes
(v + gi (m)) → (v + gi−1 (m + 1). Now the xyz-property yields v → v + gi (m), estab-
lishing the claim.
Given v ∈ N, we can choose m with m ≥ M and mk > v + fk (m), because fk is
a polynomial of degree k − 2. We then have v → v + gk (m) = v + k!. It follows that
v → v + n · k! for all n ∈ N. Let r be a multiple of k! such that r k > 2v. Since also
(r + 1)k > 2(2k − v), the list (v, r k − v, (r + 1)k − (r k − v)) provides a path of length 2
in Gv showing v → (r + 1)k − (r k − v). Since (r + 1)k − (r k − v) − (v + 1) is a multi-
ple of r, it is also a multiple of k!, so v + 1 → (r + 1)k − (r k − v). Now the xyz-property
yields v → v + 1. Hence Gv is connected.
Finally, we construct the required path through the positive integers inductively. Let
S1 = (1). For j ∈ N, let Sj be a finite list of distinct positive integers such that the sum

February 2022] PROBLEMS AND SOLUTIONS 191

of any two consecutive elements in the list is a kth power. We extend Sj to a longer such
list Sj +1 containing the smallest positive integer p not in Sj as follows. Let q be the last
element of Sj , and let r be the largest element of Sj . Choose positive integers n and m
such that mk − p > nk − q > r. Let u = nk − q and v = mk − p. Choose a path P in
Gu from u to v. Obtain Sj +1 from Sj by appending P and then p. Since u > r, all the
integers appended to the list have not previously occurred in the list, the first element that
was missing is now included, and any two consecutive elements in the list sum to a kth
power. Since we iteratively extend the list in a way that includes the least integer missing
from the previous list, each positive integer appears eventually.
Also solved by E. J. Ionaşcu, J. R. Roche, K. Schilling, and R. Stong. Part (a) also solved by O. P. Lossers
(Netherlands) and the proposer.

Integrating a Rational Function

12189 [2020, 563]. Proposed by Hidefumi Katsuura, San Jose State University, San Jose,
CA. Evaluate
 1
(k + 1)x k − km=0 x mk
dx,
0 x k(k+1) − 1
where k is a positive integer.
Solution by Giuseppe Fera and Giorgio Tescaro, Vicenza, Italy. The value of the integral is
ln(k + 1)/k. To prove this, we start with the fact that for x = 1,


k
x k(k+1) − 1
x mk = .
m=0
xk − 1

Substituting this formula in the integrand, using a limit to avoid the singularity at x = 1,
and then making the change of variable y = x k+1 , we see that
 1  a  a 
(k + 1)x k − km=0 x mk (k + 1)x k dx dx
dx = lim −
0 x k(k+1) − 1 a→1− 0 x k(k+1) − 1 0 x −1
k
 k+1  a 
a
dy dx
= lim −
a→1− 0 yk − 1 0 x −1
k

 a
dx
= lim .
a k+1 1 − x
a→1 − k

To evaluate this limit, consider any a with 0 < a < 1. When a k+1 ≤ x ≤ a, set


k−1
1 − xk
g(x) = xm = .
m=0
1−x
k+1
Note that g is increasing on [a , a], so

(1 − x)g(a k+1 ) ≤ (1 − x)g(x) ≤ (1 − x)g(a).

Inverting and substituting for g(x), we get

1 1 1 1 1
· ≤ ≤ · ,
g(a) 1 − x 1 − xk g(a k+1 ) 1 − x

192 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
and integrating yields
 a  a  a
1 dx dx 1 dx
≤ ≤ .
g(a) a k+1 1 − x a k+1 1 − x
k g(a ) a k+1 1 − x
k+1

Since
    k 
a
dx 1 − a k+1 
= ln = ln a m
,
a k+1 1−x 1−a m=0

we arrive at the bounds

 k    k 
1  a
dx 1 
ln a m
≤ ≤ ln a m
.
a k+1 1 − x
g(a) k g(a k+1 )
m=0 m=0

k−1
Finally, we have lima→1− g(a k+1 ) = lima→1− g(a) = lima→1− m=0 a m = k, and
 k 

lim ln a m
= ln(k + 1).
a→1−
m=0

Therefore, by the squeeze theorem,

 1  a
(k + 1)x k − km=0 x mk dx ln(k + 1)
dx = lim = .
0 x k(k+1) − 1 a→1− a k+1 1 − x k k

Also solved by U. Abel & V. Kushnirevych (Germany), T. Akhmetov (Russia), K. F. Andersen (Canada),
N. Batir (Turkey), A. Berkane (Algeria), R. Boukharfane (Saudi Arabia), P. Bracken, B. Bradie, N. Caro
(Brazil), R. Chapman (UK), H. Chen, R. Dempsey, A. Dixit (India) & S. Pathak (US), S. P. I. Evan-
gelou (Greece), M. L. Glasser, E. A. Herman, N. Hodges (UK), F. Holland (Ireland), W. Janous (Austria),
K. T. L. Koo (China), O. Kouba (Syria), H. Kwong, K.-W. Lau (China), G. Lavau (France), O. P. Lossers
(Netherlands), L. Matejı́c̆ka (Slovakia), M. Omarjee (France), Á. Plaza (Spain), K. Sarma (India), V. Schindler
(Germany), A. Stadler (Switzerland), S. M. Stewart (Australia), R. Stong, R. Tauraso (Italy), T. Wiandt,
M. Wildon (UK), L. Zhou, and the proposer.

An Incenter is an Orthocenter
12190 [2020, 563]. Proposed by Leonard Giugiuc, Drobeta-Turnu Severin, Romania, and
Gabriela Negutescu, Telea, Romania. Let ABC be a triangle, and let D, E, and F be points
on BC, CA, and AB, respectively, such that AD, BE, and CF are concurrent at P . It is
well known that if P is the orthocenter of ABC, then P is the incenter of DEF . Prove the
converse.
Solution by Titu Zvonaru, Comăneşti, Romania. We show that if AD is the angle bisector
of ∠EDF , then AD is perpendicular to BC. Combining this with similar statements about
BE and CF , it then follows that if P is the incenter of DEF , then P is the orthocenter of
ABC, as desired.
Let be the line through A parallel to BC, and let M and N be the points where DE and
DF , respectively, intersect . Since CDE is similar to AME and BDF is similar to
AN F , we have
CE DC AF AN
= and = .
EA AM FB BD
By Ceva’s theorem,
BD CE AF
· · = 1.
DC EA F B

February 2022] PROBLEMS AND SOLUTIONS 193

Combining these three equations yields AM = AN . Consequently, in MDN, DA is both
the angle bisector and median at D. It follows that MDN is isosceles, with MD = N D,
and hence AD is perpendicular to and thus to BC.
Editorial comment. The problem statement here corrects a typographical error that
appeared in the original problem statement.
Also solved by R. Boukharfane (Saudi Arabia), R. B. Campos (Spain), H. Chen (China), C. Chiser (Romania),
P. De (India), G. Fera (Italy), N. Hodges (UK), I. Patrascu & I. Cotoi (Romania), Y. Ionin, M. Kaplan &
M. Goldenberg, K. T. L. Koo (China), O. Kouba (Syria), S. S. Kumar, Y. Lee (Korea), J. H. Lindsey II, M. Mihai
& D. Ş. Marinescu (Romania), C. R. Pranesachar (India), A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy),
M. Tetiva (Romania), T. Wiandt, L. Zhou, and the proposers.

Fermat Strikes Twice

12192 [2020, 564]. Proposed by Péter Kórus, University of Szeged, Szeged, Hungary. Find
all triples (a, b, c) of positive integers such that (c, c2 ) is a point on the graph of y = x 2
with minimum sum of distances to (0, a) and (0, b).
Solution by Nigel Hodges, Gloucestershire, UK. There are no such triples.
Let f (x) = x 2 + (x 2 − a)2 + x 2 + (x 2 − b)2 . We want to have f minimized at
x = c, so we must have f  (c) = 0. The derivative of f is given by
x + 2(x 2 − a)x x + 2(x 2 − b)x
f  (x) = + .
x 2 + (x 2 − a)2 x 2 + (x 2 − b)2
If a = b, then f  (c) = 0 implies 2c2 = 2a − 1, which cannot happen when a and c are
integers. Hence we may assume a = b. The condition f  (c) = 0 with c > 0 becomes
2c2 − 2a + 1 2c2 − 2b + 1
=− .
c2 + (c2 − a)2 c2 + (c2 − b)2
Squaring both sides and simplifying yields
(b − a) 4c4 + 2c2 + a + b − 4ab = 0.
Since a = b, this equation is equivalent to (4c2 + 1)2 = (4a − 1)(4b − 1). Since 4a − 1 ≡
3 (mod 4), the right side must have a prime factor p congruent to 3 modulo 4. This prime
p must also divide the left side, so (2c)2 ≡ −1 (mod p). Now Fermat’s little theorem and
the fact that (p − 1)/2 is odd yield the contradiction
1 ≡ (2c)p−1 ≡ (−1)(p−1)/2 ≡ −1 (mod p).

Editorial comment. Allen Stenger invoked Fermat in a different way, expressing the prob-
lem in terms of Fermat’s principle of least time in optics, which corresponds to the angle
of incidence equaling the angle of reflection.
Also solved by N. Caro (Brazil), R. Chapman (UK), H. Chen (China), K. Gatesman, E. J. Ionaşcu, O. Kouba
(Syria), A. Stadler (Switzerland), A. Stenger, R. Stong, R. Tauraso (Italy), T. Wiandt, H. Widmer (Switzerland),
L. Zhou, and the proposer.

CLASSICS
We solicit contributions of classics from readers, who should include the problem state-
ment, solution, and references with their submission. The solution to the classic problem
published in one issue will appear in the subsequent issue.
C2. Ira Gessel [1972], contributed by the editors. Prove that a positive integer n is a
Fibonacci number if and only if 5n2 + 4 or 5n2 − 4 is a perfect square.

194 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
 The Lion and the Man
C1. Attributed to Richard Rado in the 1930s, contributed by the editors. A lion and a man
are in an enclosure. The maximum speed of the lion is equal to the maximum speed of the
man. Can the lion catch the man?
Solution. We assume that the lion and the man start at different locations, and we show that
the man can evade capture forever.
If the man starts on the boundary of the enclosure, then he first moves into the interior.
As long as he does this by traveling less than half the distance to the lion, he won’t be
caught during this step. Once he is in the interior, we can let D be an open disk centered
at the man’s location that is entirely contained in the enclosure. We now give a strategy
that the man can follow to evade capture while staying inside D and therefore inside the
enclosure.
Let the unit of distance be chosen so that D has radius 2, and let the unit of time be
chosen so that the maximum speed of both lion and man is 1. The strategy proceeds in
stages. In stage 1, the man starts running directly away from the lion and runs at maximum
speed in a straight line for 1 unit of time. Since the lion cannot run faster than the man, the
man cannot be caught during stage 1. For n ≥ 2, at stage n the man travels at maximum
speed a distance 1/n in a direction that is perpendicular to the line L that passes through
his location at the beginning of the stage and the center of D. There are two such directions
to choose from, and the man chooses based on the location of the lion. If the lion is in one
of the half planes determined by L, then the man runs into the other half plane. The man
can run either way if the lion is on L. Every point that the man visits during stage n is
closer to the man’s position at the beginning of the stage than it is to the lion’s position, so
the man evades capture during stage n.
The time elapsed during the first n stages is nk=1 1/k, which diverges as n approaches
infinity. On the other hand, the distance between the man and the center of D after n
n 2
stages, by repeated use of the Pythagorean theorem, is k=1 1/k , which converges as
n approaches infinity and in particular is bounded (generously) by 2. Thus the man evades
capture forever while remaining inside D.
Editorial comment. We have treated the lion and man as points and assumed that to cap-
ture the man, the lion must reduce the distance between them to zero in finite time. The
solution given shows that certain details of the problem don’t matter, such as the shape of
the enclosure or the initial positions of the man and lion (as long as they are distinct).
The problem has a colorful history. It was proposed by Richard Rado in the 1930s,
with the enclosure being a disk, and solved as above by Abram Besicovitch in 1952. The
problem was popularized by John Littlewood in his book A Mathematician’s Miscellany
(see B. Bollobás, ed. (1986), Littlewood’s Miscellany, Cambridge: Cambridge Univ. Press,
pp. 114–117). For further details and generalizations see Bollobás, B., Leader, I., and Wal-
ters, M. (2012), Lion and man—can both win?, Israel J. Math. 189: 267–286.
It is tempting to think that the man’s best strategy is to stay as far from the lion as
possible, and in the case of a circular enclosure this means that the man would run to
the boundary and then run around the boundary (perhaps sometimes changing direction).
However, if the man stays on the boundary, then the lion can catch the man by running
outward from the center of the enclosure while staying on the radius from the center to
the man. Thus, in order to avoid capture, the man must step into the interior of the enclo-
sure. This gives him the freedom to move in any direction—a freedom that is exploited in
Besicovitch’s solution.

February 2022] PROBLEMS AND SOLUTIONS 195

SOLUTIONS

Brianchon’s Theorem on a Hidden Conic

12177 [2020, 372]. Proposed by Dao Thanh Oai, Thai Binh, Vietnam, and Cherng-tiao
Perng, Norfolk, VA. Let C be a nondegenerate conic, and let l be a line. Suppose that
A1 , . . . , A2n and B1 , . . . , B2n are points on C such that Ai Ai+1 and Bi Bi+1 intersect at a
point on l for i = 1, . . . , 2n − 1.
(a) Show that A2n A1 and B2n B1 intersect at a point on l.
(b) Let n = 3 and take subscripts modulo 6. For i = 1, . . . , 6, suppose that Ai Bi and
Ai+1 Bi+1 intersect at a point Di . Prove that the three lines D1 D4 , D2 D5 , and D3 D6 are
concurrent.
Solution by Richard Stong, Center for Communications Research, San Diego, CA.
(a) The case n = 1 is trivial. For the case n = 2, we note that applying Pascal’s
theorem to the hexagon A1 A2 A3 B1 B2 B3 shows that the intersection points A1 A2 ∩ B1 B2 ,
A2 A3 ∩ B2 B3 , and A3 B1 ∩ B3 A1 are collinear. Since the first two are on l, it follows that
the third is as well. Applying Pascal’s theorem again to the hexagon A1 B3 B4 B1 A3 A4 shows
that A1 B3 ∩ B1 A3 , B3 B4 ∩ A3 A4 , and B4 B1 ∩ A4 A1 are collinear. Again since the first two
are on l, it follows that the third is as well, proving the case n = 2. The cases n > 2 fol-
low immediately from the n = 2 case and induction. Using the n = 2 case we conclude
that A1 A4 and B1 B4 meet on l, and therefore we can drop the indices 2 and 3 and use the
induction hypothesis.
(b) By a projective transformation, we may assume l is the line at infinity. If C is disjoint
from l, then C is an ellipse, and by a further affine transformation we may assume C is a
circle. It suffices to prove the result in this case: If C is tangent to l, then C is a parabola.
The result for parabolas follows from continuity by treating them as limits of ellipses. If
C meets l in two points, then C is a hyperbola. This case follows from the circle case by
an argument using analytic continuation. The result for the circle x 2 + y 2 = 1 means that
a certain analytic function of the x-coordinates of the points A1 , . . . , A6 , B1 (which deter-
mine the remaining coordinates) vanishes for all real values of these coordinates between
−1 and 1. By analytic continuation, the same function is 0 for purely imaginary values
of these coordinates, which implies the result for the hyperbola y 2 − x 2 = 1; an affine
transformation reduces any hyperbola to this case.

January 2022] PROBLEMS AND SOLUTIONS 87

 If C is a circle and l is the line at infinity, then the statement that Ai Ai+1 and Bi Bi+1
meet on l says that Ai Ai+1 and Bi Bi+1 are parallel and hence ∠Ai Ai+1 Bi = ∠Ai+1 Bi Bi+1 .
Thus the chords Ai Bi and Ai+1 Bi+1 subtend the same arc of the circle and hence are
congruent. It follows that there is a smaller concentric circle simultaneously tangent to all
the chords Ai Bi at their midpoints. Thus the hexagon D1 D2 D3 D4 D5 D6 has an inscribed
circle. Brianchon’s theorem then states that the principal diagonals of this hexagon are
concurrent, which is the desired conclusion.
Also solved by L. Zhou and the proposer.

Every Function has Some Continuity

12178 [2020, 373]. Proposed by Stephen Portnoy, University of Illinois, Urbana, IL. Given
any function f : R → R, show that there is a real number x and a sequence x1 , x2 , . . . of
distinct real numbers such that xn → x and f (xn ) → f (x) as n → ∞.
Solution by Supravat Sarkar, Indian Statistical Institute, Bangalore, India. Let A =
{(x, f (x)) : x ∈ R}. The set A is an uncountable subset of R2 , which implies that some
point of A must be a limit point of A. To see this, suppose it is not true. Now every point
in A has an open neighborhood in R2 that contains no other point in A. Thus the subspace
topology of A is discrete. Any uncountable set with discrete topology is not second count-
able, but being a subspace of the second countable space R2 , A must be second countable.
This is a contradiction.
Let (x, f (x)) be an element of A that is a limit point of A. There exist distinct points
(xn , f (xn )) in A converging to (x, f (x)) as n → ∞. Hence the numbers x1 , x2 , . . . are
distinct and xn → x and f (xn ) → f (x) as n → ∞.
Editorial comment. Jacob Boswell and Charles Curtis gave an example showing that the
analogous result for functions from Q to Q need not hold. Jean-Pierre Grivaux, Klaas Pieter
Hart, Kenneth Schilling, and Richard Stong all proved the stronger statement that for all
but countably many x, such a sequence can be found. Celia Schacht pointed out that a proof
of this stronger statement can be found in W. H. Young (1907), A theorem in the theory of
functions of a real variable, Rendiconti del Circolo Matematico di Palermo 24(1), 187–192.
Éric Pité, Stephen Scheinberg, and George Stoica observed that the result in the problem
follows from a theorem of H. Blumberg saying that for every function f : R → R, there is
a dense subset D of R such that the restriction of f to D is continuous; see H. Blumberg
(1922), New properties of all real functions, Trans. Amer. Math. Soc. 24(2), 113–128.
Also solved by K. F. Andersen (Canada), J. Boswell & C. Curtis, R. Chapman (UK), H. Chen (China), T. Corso
(Germany), G. A. Edgar, G. Fera & G. Tescaro (Italy), O. Geupel (Germany), J.-P. Grivaux (France), K. P. Hart
(Netherlands), D. Hensley, E. A. Herman, E. J. Ionaşcu, B. Karaivanov (USA) & T. S. Vassilev (Canada),
J. C. Kieffer, L. Matejı́c̆ka (Slovakia), A. Natian, J. Nieto (Venezuela), J. Olson, M. Omarjee (France),
A. Pathak, L. J. Peterson, É. Pité, K. Sarma (India), C. Schacht, S. Scheinberg, K. Schilling, E. Schmeichel,
A. Stadler (Switzerland), G. Stoica (Canada), R. Stong, R. Tauraso (Italy), Northwestern University Math
Problem Solving Group, and the proposer.

Factorials are Rarely Good

12179 [2020, 373]. Proposed by Nick MacKinnon, Winchester College, Winchester, UK.
A positive integer n is good if its prime factorization 2a1 3a2 · · · pm
am
has the property that
ai /ai+1 is an integer whenever 1 ≤ i < m. Find all n greater than 2 such that n! is good.
Solution by Celia Schacht, North Carolina State University, Raleigh, NC. The values of n
such that n! is good are 3, 4, 5, 6, 7, 10, and 11.

88 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
 
We have n! = pαp (n) , where Pn is the set of primes less than or equal to n and
p∈Pn
 
logp n  
 n
αp (n) = .
k=1
pk

We focus on the relative sizes of α5 (n) and α7 (n). Note that n/5 ≥ n/7 + 1 implies
n/5 > n/7 and holds when n ≥ 17.5. Explicit checking
 shows that n/5 > n/7
also holds for n ∈ {15, 16, 17}. Since n/5k ≥ n/7k for all k, for n ≥ 15 we conclude
α5 (n) > α7 (n). (1)
We complete the argument by showing
α5 (n) < 2α7 (n) (2)
for n ≥ 28. When (1) and (2) both hold, n cannot be good, since α5 (n)/α7 (n) is strictly
between 1 and 2. Hence these inequalitites reduce the problem to checking explicitly which
n less than 28 are good, and these turn out to be only 3, 4, 5, 6, 7, 10, and 11.
To prove (2), we need an upper bound on α5 (n) and a lower bound on α7 (n). We com-
pute

 n
log5 n log5 n
 n n 1 − 1/5 log5 n n(1 − 1/n) n−1
α5 (n) = ≤ = · ≤ =
k=1
5k
k=1
5k 5 1 − 1/5 4 4

and
 n log7 n
log7 n
 n  
α7 (n) = k
≥ k
− log7 n
k=1
7 k=1
7

n 1 − 1/7 log7 n   n(1 − 7/n)  

= · − log7 n ≥ − log7 n
7 1 − 1/7 6
n−7  
= − log7 n .
6
 
 it suffices to show (n − 1)/4 < (n − 7)/3 − 2 log7 n , which simpli-
Hence to prove (2)
fies to 24 log7 n + 25 < n and holds when n ≥ 74. It is also easily checked that (2) holds
when 28 ≤ n ≤ 73.
Also solved by S. Chandrasekhar (India), R. Chapman (UK), W. Chang, G. Fera (Italy), D. Fleischman,
O. Geupel (Germany), N. Hodges (UK), Y. J. Ionin, W. Janous (Austria), M. Kaplan & M. Goldenberg,
O. Kouba (Syria), S. S. Kumar, J. H. Lindsey II, O. P. Lossers (Netherlands), R. Martin (Germany) J. H. Nieto
(Venezuela), S. Omar (Morocco), É. Pité, C. R. Pranesachar (India), M. A. Prasad (Inda), A. Stadler (Switzer-
land), R. Stong, R. Tauraso (Italy), D. Terr, F. A. Velandia & J. F. González (Columbia), L. Zhou, Eagle Prob-
lem Solvers, and the proposer.

A Combination of Betas
12180 [2020, 373]. Proposed by Pablo Fernández Refolio, Madrid, Spain. Prove
∞ 4n 2 √ 2 √
 2 2C 2π
2n
8n (2n + 1)
= − 3/2 + 2
,
n=0
2 π π 2C
∞
where C = 0 t −1/4 e−t dt.

January 2022] PROBLEMS AND SOLUTIONS 89

Solution by Quan Minh Nguyen, William Academy, Toronto, ON, Canada. Let S denote the
requested sum. Using Wallis’s integral, we see that
∞ 4n 2 ∞
 
  4n
2 π/2 4n
S= 2n
= 2n
· sin x dx
n=0
28n (2n + 1) n=0
24n (2n + 1) π 0

∞
 2n
π/2  4n
2 sin2 x
= 2n
dx.
π 0 n=0
(2n + 1) 4

Recall the generating function for the Catalan numbers:

∞ ∞ √
  2n
1 − 1 − 4t 1
Cn t n = n
tn = , 0 < |t| ≤ .
n=0 n=0
n+1 2t 4

(The singularity at t = 0 is removable.) Replacing t with −t in this equation and then

averaging the two equations yields
∞ √ √
 4n
1 + 4t − 1 − 4t 1
2n
t =
2n
, 0 < |t| ≤ .
n=0
2n + 1 4t 4

Setting t = (sin2 x)/4 in this equation, we obtain

 
2 π/2 1 + sin2 x − 1 − sin2 x
S= dx
π 0 sin2 x
2 π/2 
= 1 + sin2 x − cos x csc2 x dx.
π 0
To evaluate the integral, we begin by using integration by parts to get
π/2  
2   2 π/2 cos2 x
S = − cot x 2 
1 + sin x − cos x  +  + cos x dx
π 0 π 0 1 + sin2 x
π/2
2 2 cos2 x
= +  dx.
π π 0 1 + sin2 x
Substituting u = sin x and then t = u4 , and recognizing Beta functions, we obtain
√  1 
2 2 1 1 − u2 2 2 1 1
u2
S= + √ du = + √ du − √ du
π π 0 1 + u2 π π 0 1 − u4 0 1 − u4
 1 1 
2 1 −3/4 −1/2 −1/4 −1/2
= + t (1 − t) dt − t (1 − t) dt
π 2π 0 0
    
2 1 1 1 3 1
= + B , −B , .
π 2π 4 2 4 2
√
Using Euler’s reflection formula (3/4)(1/4) = π 2 and recognizing that C = (3/4),
we compute
  √ 3/2
1 1 (1/4)(1/2) 2π
B , = =
4 2 (3/4) C2
and
 
3 1 (3/4)(1/2) 23/2 C 2
B , = = √ .
4 2 (5/4) π

90 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
Hence √ 2 √
2 2C 2π
S = − 3/2 + .
π π 2C 2

Also solved by A. Berkane (Algeria), P. Bracken, R. Chapman (UK), H. Chen, G. Fera (Italy), P. Fulop (Hun-
gary), L. Glasser, O. Kouba (Syria), K.-W. Lau (China), A. D. Pirvuceanu (Romania), V. Schindler (Germany),
F. Sinani (Kosovo), A. Stadler (Switzerland), S. M. Stewart (Australia), R. Stong, R. Tauraso (Italy), M. Vowe
(Switzerland), T Wiandt, and the proposer.

A Sum of an Integral of a Fractional Part Yields Gamma

12181 [2020, 461]. Proposed by Shivam Sharma, University of Delhi, New Delhi, India.
Prove
 ∞  
1 1 1
√k
dx = γ ,
k=2
k 0 x

where {x} equals x − x , the fractional part of x, and γ is limn→∞ − ln n + ni=1 (1/ i) ,
the Euler–Mascheroni constant.
Solution by Gérard Lavau,
√ Fontaine lès Dijon, France. For integers n and k with n ≥ 1 and
k ≥√2, we have √ 1/ k
x = n if and only if 1/(n + 1)k < x ≤ 1/nk . For such x, we have
{1/ k x} = 1/ k x − n, so
1   ∞ 1/nk  
1 1
√k
dx = √
k
− n dx
0 x n=1 1/(n+1)
k x

∞     
k 1 1 1 1
= − −n −
n=1
k − 1 nk−1 (n + 1)k−1 nk (n + 1)k

∞    
1 1 1 1
= − −
n=1
k − 1 nk−1 (n + 1)k−1 (n + 1)k
∞    ∞
1  1 1 1 1
= − − = − (ζ (k) − 1),
k − 1 n=1 nk−1 (n + 1)k−1
n=2
nk k−1

where the first sum in the last line is a telescoping series and ζ is the Riemann zeta function.
Therefore
 ∞   ∞  
1 1 1 1 ζ (k) − 1
√ dx = − .
k=2
k 0 k
x k=2
k(k − 1) k

The desired result now follows from the formulas

∞
 ∞

1 ζ (k) − 1
=1 and = 1 − γ.
k=2
k(k − 1) k=2
k

The first of these formulas can be derived by using partial fractions to rewrite the sum as
a telescoping series. The second was proved by Euler (see page 111 in J. Havil (2003),
Gamma: Exploring Euler’s Constant, Princeton: Princeton University Press).
Also solved by Z. Ahmed (India), K. F. Andersen (Canada), M. Bataille (France), A. Berkane (Algeria),
N. Bhandari (Nepal), G. E. Bilodeau, R. Boukharfane (Saudi Arabia), J. Boswell & C. Curtis, P. Bracken,
B. Bradie, B. S. Burdick, F. Cardona (Columbia), J. N. Caro Montoya (Brazil), W. Chang, R. Chapman (UK),
H. Chen, C. Chiser (Romania), B. E. Davis, M. Dinca & D. S. Marinescu (Romania), A. Dixit (Canada) &

January 2022] PROBLEMS AND SOLUTIONS 91

S. Pathak (USA), A. Eydelzon, G. Fera (Italy), M. L. Glasser, N. Grivaux (France), J. A. Grzesik, E. A. Herman,
N. Hodges (UK), W. Janous (Austria), S. Kaczkowski, M. Kaplan, K. T. L. Koo (China), O. Kouba (Syria),
S. S. Kumar, P. Lalonde (Canada), K.-W. Lau (China), R. Molinari, S. E. Muñoz (Venezuela), K. Nelson,
Q. M. Nguyen (Canada), M. Omarjee (France), S.-H. Park (Korea), Á. Plaza (Spain), C. R. Pranesachar (India),
M. A. Prasad (India), K. Sarma (India), E. Schmeichel, B. Shala (Slovenia), F. Sinani (Kosovo), S. Singhania
(India), A. Stadler (Switzerland), S. M. Stewart (Australia), R. Tauraso (Italy), H. Vinuesa (Spain), T. Wiandt,
H. Widmer (Switzerland), M. Wildon (UK), Y. Xiang (China), L. Zhou, and the proposer.

Bounding Circumradii of Corner Triangles

12182 [2020, 461]. Proposed by George Apostolopoulos, Messolonghi, Greece. Let R and
r be the circumradius and inradius, respectively, of triangle ABC. Let D, E, and F be
chosen on sides BC, CA, and AB so that AD, BE, and CF bisect the angles of ABC. Let
RA , RB , and RC denote the circumradii of triangles AEF , BF D, and CDE, respectively.
Prove RA + RB + RC ≤ 3R 2 /(4r).
Solution by Michel Bataille, Rouen, France. Let a, b, and c be the sides of ABC
opposite angles A, B, and C, respectively. The law of sines gives a = 2R sin A and
EF = 2RA sin A, and hence RA = R · EF /a. Similar results hold for RB and RC , so the
requested inequality is equivalent to
EF FD DE 3R
+ + ≤ .
a b c 4r
Since BE bisects ∠ABC, AE/c = EC/a = (EC + AE)/(a + c) = b/(a + c), so
AE = bc/(a + c). Similarly, AF = bc/(a + b), and using the law of cosines, we obtain
b2 c2 b2 c2 bc(b2 + c2 − a 2 )
EF 2 = AE 2 + AF 2 − 2AE · AF · cos A = + −
(a + c)2 (a + b)2 (a + b)(a + c)
bc
= · a 2 (a + b)(a + c) − a(a + b + c)(b − c)2
(a + b)2 (a + c)2
bc a 2 bc
≤ · a 2
(a + b)(a + c) = .
(a + b)2 (a + c)2 (a + b)(a + c)
By the AM–GM inequality,
√ √ √ √ √
a bc a bc a4b4c 2a + b + c
EF ≤ √ ≤ √ = ≤ .
(a + b)(a + c) √ 2 8
2 ab · 2 ac
Similarly, F D ≤ (2b + c + a)/8 and DE ≤ (2c + a + b)/8. Therefore
 
EF FD DE 3 1 b c c a a b
+ + ≤ + + + + + +
a b c 4 8 a a b b c c
3 (a + b + c)(ab + bc + ca)
= + .
8 8abc
With s = (a + b + c)/2, we have ab + bc + ca = s 2 + r 2 + 4rR and abc = 4srR. Apply-
ing Gerretsen’s inequality s 2 ≤ 4R 2 + 3r 2 + 4rR and Euler’s inequality R ≥ 2r, we obtain
EF FD DE 3 2s(s 2 + r 2 + 4rR) s 2 + r 2 + 10rR
+ + ≤ + =
a b c 8 32srR 16rR
2R 2 + 2r 2 + 7rR 6R 2 − (R − 2r)(4R + r) 6R 2 3R
≤ = ≤ = ,
8rR 8rR 8rR 4r
which completes the proof.

92 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
Also solved by M. Dinc̆a & M. Ursărescu (Romania), G. Fera (Italy), N. Hodges (UK), W. Janous (Austria),
C. R. Pranesachar (India), V. Schindler (Germany), S. Singhania (India), A. Stadler (Switzerland), R. Stong,
R. Tauraso (Italy), M. Vowe (Switzerland), T. Wiandt, and the proposer.

A Gaussian Binomial Identity

!n"
12183 [2020, 461]. Proposed by Hideyuki Ohtsuka, Saitama, Japan. Let k q
denote the
Gaussian binomial coefficient
(1 − q n )(1 − q n−1 ) · · · (1 − q n−k+1 )
.
(1 − q k )(1 − q k−1 ) · · · (1 − q)
For integers m, n, and r with m ≥ 1 and n ≥ r ≥ 0, prove
 
 (−1)k q ( 2 )−rk # n $ m + n −1
n k+1
q rm
= .
k=0
1 − q k+m k q 1 − qm m q

Solution I by Albert Stadler, Herrliberg, Switzerland. We use induction on n. For n = r =

! n " equal 1/ (1 − q ).
m
0, both sides of the proposed identity
We extend the definition of k q in the standard way by setting it to 0 when k ∈/ [0, n].
We use the well-known and easily-proved analogues of the Pascal identities for the Gaus-
sian binomial coefficients:
  #n$   # $  
n+1 n k n n
= +q n−k+1
=q + . (∗)
k q k q k−1 q k q k−1 q
For the induction step, we obtain the truth of the statement for n + 1 from its truth for
n. First suppose 0 ≤ r ≤ n. Using the first equation in (∗) and then reindexing the second
sum,
     
  (−1)k q ( 2 )−rk # n $
n+1 k+1 n+1 k+1
(−1)k q ( 2 )−rk n + 1 n
= +q n−k+1

k=0
1 − q k+m k q k=0
1 − q k+m k q k−1 q

 (−1)k q ( 2 )−rk # n $  (−1)k+1 q ( 2 )−r(k+1)−k # n $

n k+1 n k+2

= + q n

k=0
1 − q k+m k q k=0
1 − q k+1+m k q

 (−1)k q ( 2 )−rk # n $  (−1)k q ( 2 )−rk # n $

n k+1 n k+1

= −q n+1−r

k=0
1 − q k+m k q k=0
1 − q k+1+m k q
   
q rm m + n −1 q r(m+1) m + n + 1 −1
= − q n+1−r
·
1 − qm m q 1 − q m+1 m+1 q
 −1  rm 
m+n+1 q 1−q m+n+1
q rm
1 − q m+1
= · −q n+1
· ·
m q 1 − qm 1 − q n+1 1 − q m+1 1 − q n+1
 
m + n + 1 −1 q rm
= ,
q 1−q
m m

as required.
It remains to prove the statement for r = n + 1. The computation is similar:
     

n+1 k+1
(−1)k q ( 2 )−(n+1)k n + 1 
n+1 k+1
(−1)k q ( 2 )−(n+1)k # $
k n n
= q +
k=0
1 − q k+m k q k=0
1 − q k+m k q k−1 q

January 2022] PROBLEMS AND SOLUTIONS 93

  (−1)k q ( 2 )−nk # n $  (−1)k q ( 2 )−nk # n $
n k+1 n k+1
−n
= − q
k=0
1 − q k+m k q k=0
1 − q k+1+m k q
   
q nm m + n −1 −n q n(m+1) m + n + 1 −1
= − q ·
1 − qm m q 1 − q m+1 m+1 q
 −1  nm 
m+n+1 q 1−q m+n+1
q nm
1 − q m+1
= · − ·
m q 1 − qm 1 − q n+1 1 − q m+1 1 − q n+1
 
m + n + 1 −1 q (n+1)m
= ,
q 1−q
m m

completing the induction step.

Solution II by Pierre Lalonde, Plessisville, QC, Canada. More generally, we see that
 (−1)k q ( 2 )−rk # n $ xr  1 − qk
n k+1 n
= ,
k=0
1 − xq k k q 1 − x k=1 1 − xq k

for n ≥ r ≥ 0. The proposed problem is the case x = q m .

The right side of the identity is a rational function in x with numerator of degree r and
denominator of degree n + 1, which is greater than r. The zeros of the denominator are
1/q k for 0 ≤ k ≤ n, which are formally distinct. Therefore, the partial fraction decompo-
sition of the right side is
xr  1 − qk 
n n
Ak
= ,
1 − x k=1 1 − xq k k=0
1 − xq k

where
n
xr  1 − qi
n
i=1 (1 − q )
i
−rk
Ak = lim (1 − xq ) · k
= q  
x→1/q k 1 − x i=1 1 − xq i k −i
i=1 (1 − q )
n−k
i=1 (1 − q )
i

n #n$
i=n+1−k (1 − q )
 i
k ( ki=1 i)−rk k (k+12 )−rk
= (−1) q ·  k
= (−1) q .
i=1 (1 − q )
i k q
Hence
xr  1 − qk  (−1)k q ( 2 )−rk # n $
n n k+1

= ,
1 − x k=1 1 − xq k k=0
1 − xq k k q

as claimed.
Editorial comment. Several solvers used the method of the first solution. It can be adapted
to prove the generalization in the second solution. Hacer Bozdag mentioned a still more
general result, with two additional parameters and implying the claim, from E. Kılıç and
H. Prodinger (2016), Evaluation of sums involving Gaussian q-binomial coefficients with
rational weight functions, Int. J. Number Theory 12, 495–504.
Also solved by H. Bozdag (Turkey), R. Chapman (UK), N. Hodges (UK), W. P. Johnson, H. Kwong,
M. A. Prasad (India), R. Stong, R. Tauraso (Italy), L. Zhou, and the proposer.

The Distance Between Norms

12186 [2020, 462]. Proposed by Anatoly Eydelzon, University of Texas at Dallas,
n p 1/p
Richardson, TX. For v = x1 , . . . , xn  in Rn , let vp = i=1 |xi | and v∞ =

94 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 129
max1≤i≤n |xi |; these are the usual p-norm and ∞-norm on Rn . For what v does the
series
∞
vp − v∞
p=1
converge?
Solution by Óscar Ciaurri, Universidad de La Rioja, Logroño, Spain. When v is the zero
vector, the terms of the series are identically zero, and hence the series converges. Exclud-
ing this trivial case, we show that the given series S converges if and only if there is a
unique j ∈ {1, . . . , n} such that |xj | = v∞ .
Suppose there is a unique such j . By symmetry, we may assume j = 1. We have
⎛ 1/p ⎞
n
|x | p
vp − v∞ = v∞ ⎝ 1 + − 1⎠
i

i=2
|x1 | p

and, by Bernoulli’s inequality (1 + z)r ≤ 1 + rz for 0 ≤ r ≤ 1 and z > −1, we have

n
|xi |p
vp − v∞ ≤ v∞ .
i=2
p|x1 |p

Summing over p, we obtain

n  ∞
|xi |p
S ≤ v∞ .
i=2 p=1
p|x1 |p

The inner series all converge since |xi |/|x1 | < 1, and hence S converges.
Now suppose that there are at least two values j, k ∈ {1, . . . , n} such that v∞ =
|xj | = |xk |. In this case, vp ≥ (|xj |p + |xk |p )1/p = 21/p v∞ , so vp − v∞ ≥
v∞ (21/p − 1). Since
21/p − 1 2t − 1
lim = lim = log 2 > 0,
p→∞ 1/p t→0 t
∞
the series p=1 (2
1/p
− 1) diverges by comparison to the harmonic series, and hence S
diverges.
Also solved by K. F. Andersen (Canada), N. Caro (Brazil), R. Chapman (UK), H. Chen (China), C. Curtis
& A. Appuhamy & J. Boswell, J. Freeman (Netherlands), J.-P. Grivaux (France), L. Han, E. A. Herman,
N. Hodges (UK), E. J. Ionaşcu, K. T. L. Koo (China), O. Kouba (Syria), J. H. Lindsey II, U. Milutinović
(Slovenia), M. Omarjee (France), Á. Plaza & K. Sasdarangani (Spain), M. A. Prasad (India), K. Sarma (India),
K. Schilling, A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), T. Wiandt, M. Wildon (UK), C.-Y. Wu,
and the proposer.

CLASSICS
Here each month we feature one classic problem, whose solution will appear in the
subsequent issue. Classics are problems of unusual elegance that are not new but deserve
to be better known. We solicit contributions of Classic problems from readers, who should
include the problem statement, solution, and references with their submission. We will not
be soliciting or publishing reader solutions to Classic problems that appear here.
C1. Attributed to Richard Rado in the 1930s, contributed by the editors. A lion and a man
are in an enclosure. The maximum speed of the lion is equal to the maximum speed of the
man. Can the lion catch the man?

January 2022] PROBLEMS AND SOLUTIONS 95

SOLUTIONS

Optimizing an Inequality
12169 [2020, 274]. Proposed by Leonard Giugiuc, Drobeta Turnu Severin, Romania. Let
n be an integer with n ≥ 2. Find the least positive real number α such that
  n n
(n − 1) · 1 + α (xi − xj )2 + xi ≥ xi
1≤i<j ≤n
i=1 i=1

for all nonnegative real numbers x1 , . . . , xn .

Solution by Richard Stong, Center for Communications Research, San Diego, CA. If x1 =
x2 = · · · = xn−1 = R and xn = 0, then the inequality states

(n − 1) · 1 + (n − 1)αR 2 ≥ (n − 1)R,
or
1 1
α≥ − .
n − 1 (n − 1)R 2
Letting R → ∞ we find α ≥ 1/(n − 1). We claim that the given inequality holds for
α = 1/(n − 1), and therefore this is the smallest possible value of α. Thus, we must show

1  
n n
(n − 1) · 1 + (xi − xj ) +
2 xi ≥ xi . (∗)
n−1 1≤i<j ≤n
i=1 i=1

Note that since both sides of this inequality are continuous in each xk , it suffices to prove
the inequality for xk > 0. 
Let S = ni=1 xi and P = ni=1 xi . Since 1≤i<j ≤n (xi − xj )2 = n ni=1 xi2 − S 2 , the
inequality can be written
  n 

 1 
(n − 1) · 1 + n xi2 − S 2 + P ≥ S.
n−1 i=1

Thus, for fixed S and P it suffices to prove (∗) when the xi are chosen to minimize ni=1 xi2 .
Let g(x1 , . . . , xn ) = ni=1 xi and h(x1 , . . . , xn ) = ni=1 log xi . Since our two constraints
g = S and h = log P define a smooth compact manifold, this minimum must exist, and it
must occur either at a point that satisfies the Lagrange multiplier equations
μ
2xi = λ +
xi
for some λ and μ or at a point where ∇g and ∇h are linearly dependent. The Lagrange
multiplier equations are quadratic in xi , so they can be satisfied only at points where the

December 2021] PROBLEMS AND SOLUTIONS 947

xi take at most two distinct values. Also, ∇g and ∇h are linearly dependent only at points
where the xi are all equal. Therefore, it will suffice to prove (∗) at all points where the xi
take either one or two distinct positive values.
If the xi all have the same value y, then inequality (∗) becomes n − 1 + y n ≥ ny, which
is precisely Bernoulli’s inequality. Now assume that k of the xi equal y and n − k of them
equal z, where 0 < k < n and z < y. In that case, inequality (∗) becomes

k(n − k)
(n − 1) · 1 + (y − z)2 + y k zn−k ≥ ky + (n − k)z.
n−1
Bernoulli’s inequality gives y k ≥ 1 + k(y − 1) and zn−k ≥ 1 + (n − k)(z − 1). Thus,
if y > z ≥ 1 then

y k zn−k ≥ (1 + k(y − 1))(1 + (n − k)(z − 1))

≥ 1 + k(y − 1) + (n − k)(z − 1) = ky + (n − k)z − (n − 1),

and the desired inequality follows.

Next, suppose z < y ≤ 1. If 1 + k(y − 1) and 1 + (n − k)(z − 1) are both nonnegative,
then we can reason as in the previous paragraph. If either one of them is negative, say
1 + (n − k)(z − 1) < 0, then (∗) follows from

ky + (n − k)z ≤ ky + (n − k − 1) ≤ k + (n − k − 1) = n − 1.

Thus, for the rest of the solution we may assume z < 1 < y. Suppose

z ≤ (n − k − 1)/(n − k).

Inequality (∗) would follow from


k(n − k)
(n − 1) · 1+ (y − z)2 ≥ ky + (n − k)z.
n−1
Since the left side is a decreasing function of z and the right side is an increasing function
of z, it suffices to prove this in the case z = (n − k − 1)/(n − k); that is, it suffices to prove

k(n − k) n−k−1 2
(n − 1) · 1 + y− ≥ ky + (n − k − 1).
n−1 n−k
After squaring and canceling a factor of k/(n − k) this reduces to

(n − k − 1)(n − k)n(y − 1)2 + (n − 1) ≥ 0,

which is clearly true.

Finally, suppose z > (n − k − 1)/(n − k). By Bernoulli’s inequality, it suffices to show

k(n − k)
(n − 1) · 1 + (y − z)2 + (1 + k(y − 1))(1 + (n − k)(z − 1)) ≥ ky + (n − k)z.
n−1
After some simplification, this reduces to
 
(n − 1) − k(n − k)(1 − z)2 (y − 1)2 + (n − 1)(1 − z)2 ≥ 0,

which is true since (1 − z)2 < 1/(n − k)2 and n − 1 > k/(n − k), so the coefficient of
(y − 1)2 is positive.
Also solved by A. Stadler (Switzerland) and the proposer.

948 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 128
 The Base-5 Expansion of a Reciprocal
12170 [2020, 274]. Proposed by Jeffrey C. Lagarias, University of Michigan, Ann Arbor,
MI. Let p be a prime number congruent to 1 modulo 15. Show that the minimal period of
the base 5 expansion of 1/p cannot be equal to (p − 1)/15.
Solution by Joel Schlosberg, Bayside, NY. Let n = (p − 1)/30. Since p is odd, (p − 1)/15
is even, so n ∈ N. Since p ≡ 12 (mod 5), the quadratic reciprocity theorem implies that 5
is a quadratic residue modulo p. Therefore, 5 ≡ z2 (mod p) for some z ∈ Z. Since p  5,
also p  z. Therefore, by Fermat’s little theorem,

515n ≡ 5(p−1)/2 ≡ zp−1 ≡ 1 (mod p).

Suppose that 2n is the minimal period of the base 5 expansion of 1/p. This means
that 2n is the least positive integer m such that p | (5m − 1). Since also p | (515n − 1),
the multiplicative order of 5 modulo p must divide gcd(2n, 15n), which equals n. Now
p | (5n − 1), a contradiction.
Also solved by R. Chapman (UK), A. Dixit (Canada) & S. Pathak (USA), S. M. Gagola, Jr., K. T. L. Koo
(China), O. P. Lossers (Netherlands), A. Nakhash, M. A. Prasad, A. Stadler (Switzerland), A. Stenger, R. Stong,
D. Terr, E. White & R. White, and the proposer.

A Tetrahedron and the Midpoints of its Edges

12172 [2020, 275]. Proposed by Hidefumi Katsuura, San Jose State University, San Jose,
CA. Let A, B, C, and D be four points in three-dimensional space, and let U , V , W , X, Y ,
and Z be the midpoints of AB, AC, AD, BC, BD, and CD, respectively.
(a) Prove
 
4 U Z 2 + V Y 2 + W X2 = AB 2 + AC 2 + AD 2 + BC 2 + BD 2 + CD 2 .

(b) Prove
 
4 (AB · CD · U Z)2 + (AC · BD · V Y )2 + (AD · BC · W X)2

≥ (AB · BC · CA)2 + (BC · CD · DB)2 + (CD · DA · AC)2 + (DA · AB · BD)2 ,

and determine when equality holds.

Solution by Li Zhou, Polk State College, Winter Haven, FL.
(a) By Apollonius’s theorem,

4W B 2 = 2AB 2 + 2BD 2 − AD 2 ,
4W C 2 = 2AC 2 + 2CD 2 − AD 2 ,

and

4W X2 = 2W B 2 + 2W C 2 − BC 2 = AB 2 + AC 2 + BD 2 + CD 2 − AD 2 − BC 2 .

Adding the last equation to the analogous expressions for 4U Z 2 and 4V Y 2 establishes the
identity.
(b) Using the expressions above for 4W X2 , 4V Y 2 , 4U Z 2 , a computation shows that
 
4· (AB · CD · U Z)2 + (AC · BD · V Y )2 + (AD · BC · W X)2
− (AB ·BC ·CA)2 − (BC ·CD·DB)2 − (CD·DA·AC)2 − (DA·AB ·BD)2 (∗)

December 2021] PROBLEMS AND SOLUTIONS 949

is equal to
⎡ ⎤
2AD 2 AD 2 + BD 2 − AB 2 AD 2 + CD 2 − AC 2
1 ⎣
det AD + BD 2 − AB 2
2
2BD 2 BD 2 + CD 2 − BC 2 ⎦ .
2 AD 2 + CD 2 − AC 2 BD + CD 2 − BC 2
2
2CD 2
The determinant here is the Cayley–Menger determinant for the tetrahedron ABCD and
its value is 2882 , where  is the volume of ABCD. Hence (∗) is equal to 1442 , which
is clearly nonnegative. This yields the desired inequality, and equality holds if and only if
 = 0, in other words A, B, C, and D are coplanar.
Editorial comment. The Cayley–Menger determinant generalizes Heron’s formula for the
area of a triangle to simplices of higher dimension.
Also solved by M. Bataille (France), R. Chapman (UK), G. Fera & G. Tescaro (Italy), D. Fleischman,
E. A. Herman, W. Janous (Austria), M. Kaplan & M. Goldenberg, B. Karaivanov (USA) & T. S. Vassilev
(Canada), K. T. L. Koo (China), A. Stadler (Switzerland), R. Stong, T. Wiandt, and the proposer.

A Matrix Equation
12173 [2020, 275]. Proposed by Florin Stanescu, Serban Cioculescu School, Gaesti,
Romania. Suppose that X and Y are n-by-n complex matrices such that 2Y 2 = XY − Y X
and the rank of X − Y is 1. Prove Y 3 = Y XY .
Solution by Roger A. Horn, Tampa, FL. Let z and w be nonzero complex n-vectors such
that X − Y = zw ∗ . It suffices to show that if

2Y 2 = zw ∗ Y − Y zw ∗ , (1)

then Y zw ∗ Y = 0. Jacobson’s lemma (see page 126 of R. Horn and C. Johnson (2018),
Matrix Analysis, 2nd ed., New York: Cambridge University Press) states that if BC −
CB commutes with C, then BC − CB is nilpotent. Consequently, Y 2 (and hence Y ) is
nilpotent. The rank of Y 2 is at most 2, since it is the sum of two matrices whose ranks are
at most 1. Therefore, the Jordan canonical form of Y is a direct sum of nilpotent Jordan
blocks that are not larger than 4-by-4. There are three cases.
Case (a): Y 2 = 0 (no block larger than 2-by-2). If Y 2 = 0, then Y 2 z = 0 and

0 = 2Y 3 = Y 2Y 2 = Y zw ∗ Y − Y 2 zw ∗ = Y zw ∗ Y. (2)

Case (b): Y 2 = 0 and Y 3 = 0 (the largest block is 3-by-3). We compute

0 = 2Y 4 = Y 2 2Y 2 = Y 2 zw ∗ Y − Y 3 zw ∗ = (Y 2 z)(w ∗ Y ).
Either w ∗ Y = 0 and we are done, or w ∗ Y = 0 and Y 2 z = 0. In the latter case, (2) also
holds, and it ensures that Y zw∗ Y = 0.
Case (c): Y 3 = 0 and Y 4 = 0 (the largest block is 4-by-4). Let v be a complex n-vector
such that Y 3 v = 0. Suppose Y z = 0. We compute
0 = 2Y 5 v = 2Y 2 Y 3 v = zw ∗ Y 4 v − Y zw ∗ Y 3 v = −(w ∗ Y 3 v)Y z,

so w ∗ Y 3 v = 0. We also have

0 = 2Y 4 v = 2Y 2 Y 2 v = zw ∗ Y 3 v − Y zw ∗ Y 2 v = −(w ∗ Y 2 v)Y z,

so w ∗ Y 2 v = 0 as well. Now compute

2Y 3 v = 2Y 2 Y v = zw ∗ Y 2 v − Y zw ∗ Y v = −(w ∗ Y v)Y z, (3)

950 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 128
which ensures w ∗ Y v = 0 since Y 3 v = 0. Multiply (3) by Y to obtain
0 = 2Y 4 v = −(w ∗ Y v)Y 2 z,
so Y 2 z = 0. Finally, use (1) to compute
2Y 3 v = Y 2Y 2 v = Y zw ∗ Y v − Y 2 zw ∗ v = (w ∗ Y v)Y z,
which contradicts (3). We conclude Y z = 0 and hence Y zw∗ Y = 0.
Editorial comment. Kyle Gatesman observed that the result holds when the hypothesis
2Y 2 = XY − Y X is replaced by the more general kY 2 = XY − Y X for some nonzero
k ∈ C. Several solvers noted that the conclusion Y 3 = Y XY can be strengthened to Y 3 =
0 = Y XY .
Also solved by M. Bataille (France), C. Chiser (Romania), K. Gatesman, N. Grivaux (France), L. Han,
E. A. Herman, N. Hodges (UK), K. T. L. Koo (China), C. P. A. Kumar (India), J. H. Lindsey II, O. P. Lossers
(Netherlands), M. Omarjee (France), K. Sarma (India), A. Stadler (Switzerland), R. Stong, J. Stuart, R. Tauraso
(Italy), E. I. Verriest, and the proposer.

Powers of 4 and 5 with the Same Leading Digits

12174 [2020, 372]. Proposed by Gregory Galperin, Eastern Illinois University, Charleston,
IL, and Yury J. Ionin, Central Michigan University, Mount Pleasant, MI.
(a) Let n be a positive integer, and suppose that the three leading digits of the decimal
expansion of 4n are the same as the three leading digits of 5n . Find all possibilities for
these three leading digits.
(b) Prove that for any positive integer k there exists a positive integer n such that the k
leading digits of the decimal expansion of 4n are the same as the k leading digits of 5n .
Solution by Oliver Geupel, Brühl, Germany. We prove the stronger claim that for any pos-
itive integer k there are exactly two k-digit numbers a that occur as the k leading digits of
the decimal expansions of 4n and 5n for some positive integer n. In particular, the condi-
tion for such a number a is the existence of a positive integer n and nonnegative integers
p and q such that 10p a ≤ 5n < 10p (a + 1) and 10q a ≤ 4n < 10q (a + 1). Notice that if
10p a = 5n then a ≤ 10q a ≤ 4n < 5n = 10p a, so p > 0. This implies that 10p a is even
and 5n is odd, which is a contradiction. Therefore we can strengthen the first inequality to
10p a < 5n < 10p (a + 1). The product of the second inequality with the square of the first
yields
102p+q a 3 < 52n 22n < 102p+q (a + 1)3 .
Thus a power 10m lies between a 3 and (a + 1)3 . Since a has k digits, we have
103k−3 ≤ a 3 < 10m < (a + 1)3 ≤ 103k .
Thus m ∈ {3k − 2, 3k − 1}, leaving only two candidates for a: 10k−2/3  and 10k−1/3 . In
the case k = 3, these numbers are 215 and 464.
Now suppose that a = 10β , where β ∈ {k − 2/3, k − 1/3}. We prove that a occurs
as the k leading digits of the decimal expansions of 4n and 5n for some positive integer n.
This confirms that 215 and 464 are solutions to part (a) and proves the claim in part (b).
The inequality 10p a < 5n < 10p (a + 1) can be rewritten as
p + log10 a < n log10 5 < p + log10 (a + 1).
Since log10 a < β < log10 (a + 1), to satisfy the inequality we need to have p + β close to
n log10 5. Thus we begin by finding p and n for which these numbers are close.

December 2021] PROBLEMS AND SOLUTIONS 951

 Let ε = min{β − log10 a, log10 (a + 1) − β}, which is positive. Kronecker’s approxima-
tion theorem asserts that the positive integer multiples of an irrational number modulo 1 are
dense in (0, 1) (see, for example, Chapter XXIII of G. H. Hardy and E. M. Wright (1975),
An Introduction to the Theory of Numbers, 4th ed., Oxford: Clarendon Press). Therefore,
there exist infinitely many pairs of positive integers n and p such that |n log10 5 − p − β| <
ε/2. Consider such pairs (n, p).
Taking logarithms in 4 = 100/25 yields log10 4 = 2(1 − log10 5), so
|n log10 4 − (2n − 2p − 3β) − β| = 2|p + β − n log10 5| < ε,
which can be rewritten as |n log10 4 − q − β| < ε, where q is the integer 2n − 2p − 3β.
Among the pairs (n, p) satisfying the restriction involving ε, choose a pair with n large
enough so that q is positive. We obtain
q + log10 a ≤ q + β − ε < n log10 4 < q + β + ε ≤ q + log10 (a + 1)
and, analogously,
p + log10 a < n log10 5 < p + log10 (a + 1).
Thus 10 a < 4 < 10q (a + 1) and 10p a < 5n < 10p (a + 1). Consequently, the k-digit
q n

number a occurs as the k leading digits of the decimal expansions of 4n and 5n .

Also solved by R. Chapman (UK), G. Fera (Italy), N. Hodges (UK), O. P. Lossers (Netherlands), A. Stadler
(Switzerland), R. Stong, R. Tauraso (Italy), L. Zhou, and the proposer. Part (a) also solved by D. Terr.

An Incenter-Centroid Inequality
12175 [2020, 372]. Proposed by Giuseppe Fera, Vicenza, Italy. Let I be the incenter and
G be the centroid of a triangle ABC. Prove
I A2 I B2 I C2
2< + + ≤ 3.
GA2 GB 2 GC 2
Solution by Arkady Alt, San Jose, CA. Let a, b, and c be the lengths of the sides opposite
A, B, and C, let ma , mb , and mc be the corresponding median lengths, and let lA , lB , and lC
be the corresponding angle bisector lengths. Let r be the inradius and s the semiperimeter.
By the Pythagorean theorem, I A2 = r 2 + (s − a)2 . From Heron’s formula and the inra-
dius/semiperimeter formula for the area of a triangle, we have
(s − a)(s − b)(s − c)
r2 = .
s
Using 2s = a + b + c, we obtain
(s − a)[(s − b)(s − c) + (s − a)s] bc(s − a)
I A2 = = .
s s
It is well known that GA = (2/3)ma . By Apollonius’s theorem,
m2a = (2b2 + 2c2 − a 2 )/4.
Therefore GA2 = (2b2 + 2c2 − a 2 )/9, so
I A2 9bc(s − a)
= .
GA 2 s(2b2 + 2c2 − a 2 )
To establish the upper bound, we observe that
2b2 + 2c2 − a 2 = (b + c)2 + (b − c)2 − a 2 ≥ (b + c)2 − a 2
= (b + c + a)(b + c − a) = 4s(s − a),

952 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 128
and therefore
I A2 9bc(s − a) 9bc
≤ 2 = 2.
GA 2 4s (s − a) 4s
Similarly, I B 2 /GB 2 ≤ 9ac/(4s 2 ) and I C 2 /GC 2 ≤ 9ab/(4s 2 ), so
I A2 I B2 I C2 9(ab + bc + ca)
2
+ 2
+ ≤ .
GA GB GC 2 4s 2
By the Cauchy–Schwarz inequality, a 2 + b2 + c2 ≥ ab + bc + ca, so
4s 2 = (a + b + c)2 = a 2 + b2 + c2 + 2(ab + bc + ca) ≥ 3(ab + bc + ca).
Therefore
I A2 I B2 I C2 9(ab + bc + ca)
+ + ≤ = 3.
GA 2 GB 2 GC 2 3(ab + bc + ca)
For the lower bound, we start with
2b2 + 2c2 − a 2 = (b + c)2 − (a 2 − (b − c)2 ) < (b + c)2 ,
which holds because a 2 > (b − c)2 , which follows from the triangle inequality. Therefore
I A2 9bc(s − a) 9lA2
> = ,
GA2 s(b + c)2 4s 2
where in the last step we have used the known formula lA2 = 4bcs(s − a)/(b + c)2 . Simi-
larly,
I B2 9l 2
2
> B2
GB 4s
and
I C2 9lC2
> ,
GC 2 4s 2
so
I A2 I B2 I C2 9
2
+ 2
+ 2
> 2 (lA2 + lB2 + lC2 ).
GA GB GC 4s
The required lower bound now follows from the inequality
lA2 + lB2 + lC2 > (8/9)s 2
(see page 218, inequality 11.7 in D. S. Mitrinović, J. E. Pečarić, V. Volenec (1989), Recent
Advances in Geometric Inequalities, Dordrecht: Springer).
Editorial comment. Li Zhou cited experimental evidence from Geometer’s Sketchpad
for the following conjectures: The order of I A/GA, I B/GB, and I C/GC corresponds
inversely to the order of a, b, and c, and hence also to the order of angles A, B, and C.
Moreover, the sum of the two largest of I A2 /GA2 , I B 2 /GB 2 , and I C 2 /GC 2 is already
at least 2.
Walter Janous strengthened the inequality to
r I A2 I B2 I C2 41 7r
2+ ≤ 2
+ 2
+ 2
≤ + ,
8R GA GB GC 16 8R
where R is the circumradius of ABC. That this upper bound is stronger follows from
2r ≤ R.

December 2021] PROBLEMS AND SOLUTIONS 953

Also solved by H. Bailey, S. Gayen (India), W. Janous (Austria), M. Kaplan & M. Goldenberg, P. Khalili,
K.-W. Lau (China), J. H. Lindsey II, C. R. Pranesachar (India), V. Schindler (Germany), A. Stadler
(Switzerland), R. Stong, R. Tauraso (Italy), T. Wiandt, T. Zvonaru (Romania), and the proposer.

A Diophantine Equation
12176 [2020, 372]. Proposed by Nikolai Osipov, Siberian Federal University, Krasnoyarsk,
Russia. Solve

xy 3 + y 2 − x 5 − 1 = 0

in positive integers.
Solution by Mandyam A. Prasad, Mumbai, India. We show that the only solution in posi-
tive integers is (x, y) = (1, 1). When x = 1, the equation becomes y 3 + y 2 − 2 = 0, whose
only solution is y = 1. When y = 1, the equation becomes x − x 5 = 0, whose only posi-
tive solution is x = 1.
Hence we may assume x ≥ 2 and y ≥ 2. The polynomial x 4 − x − 1 is positive at x = 2
and has positive derivative for x ≥ 2, so x 4 − x − 1 > 0 for x ≥ 2. Therefore

(x − 1)(x 4 − x − 1) > 0.

Expanding yields x 4 + x 2 < x 5 + 1. If y ≤ x, then

xy 3 + y 2 ≤ x 4 + x 2 < x 5 + 1,

which contradicts the original equation. Therefore, we may assume x < y.

If x 2 ≤ y, then

xy 3 + y 2 = x 5 + 1 ≤ xy 2 + 1,

which yields xy 2 (y − 1) ≤ 1 − y 2 , contradicting y > 1. Hence y < x 2 .

Rewritten as (y 2 − 1)(xy + 1) = x(x 4 − y), the original equation implies x(x 4 − y) ≡
0 (mod xy + 1). Multiplying by −x(y + y 3 ) yields
−x 6 y − x 6 y 3 + x 2 y 2 + x 2 y 4 ≡ −x(y + y 3 )x(x 4 − y) ≡ 0 (mod x(xy + 1)).

The extra factor of x in the modulus is allowed because we multiplied by a multiple of x.

Using x 2 y ≡ −x (mod x(xy + 1)) and the original equation, we obtain

x 5 + x 3 − xy − xy 3 = x 3 − xy − 1 + y 2 ≡ 0 (mod x(xy + 1)).

Since
x 3 − xy − 1 + y 2 > x 2 − xy + y 2 − 1 > 0

when x, y ≥ 2, we must have

x 3 − xy − 1 + y 2 ≥ x(xy + 1),

because the left side is a multiple of the right side. This inequality can be rewritten as

(x 2 − y)(x − y) − x − 1 ≥ 0.

Since x < y < x 2 , the left side is negative, which is a contradication. This forbids all
solutions with x, y > 1.
Also solved by the proposer.

954 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 128
 π

SOLUTIONS

Strengthening the Cauchy–Schwarz Inequality

12163 [2020, 179]. Proposed by Thomas Speckhofer, Attnang-Puchheim, Austria. Let Rn
have the usual dot product and norm. When v = (x1 , . . . , xn ) ∈ Rn , let v = x1 + · · · +xn .
Prove
1
v 2
w 2
≥ (v · w)2 + ( v |w| − w |v|)2
n
for all v, w ∈ Rn .
Solution by the Davis Problem Solving Group, Davis, CA. If either v = 0 or w = 0 then
both sides of the requested inequality are zero, so we may assume v = 0 and w = 0.
First assume v = 0 and w = 0. By homogeneity, we may assume v = w = 1.
We have (v · w)2 = v 2 w 2 cos2 θ , where θ is the angle between the vectors v and w.
Thus we must prove v 2 w 2 ≥ v 2 w 2 cos2 θ + (1/n)( v − w )2 , or
1
v 2
w 2
sin2 θ ≥ ( v − w )2 . (1)
n
Let S denote the area of the triangle whose vertices√are the origin, √ v, and w. If h is
the altitude of the triangle from the origin, then h ≥ 1/ n, since 1/ n is the minimum
distance from the origin to a point in the hyperplane x1 + · · · + xn = 1. Thus
1
v w sin θ = 2S = h v − w ≥ √ v − w ,
n
and squaring yields
1 1
v 2
w 2
sin2 θ ≥ v−w 2
≥ ( v − w )2 ,
n n
where the final inequality is a consequence of the triangle inequality. This establishes (1).
Equality holds if and only v and w are linearly dependent.
If v = 0 and w = 0, then the inequality reduces to the Cauchy–Schwarz inequality,
and once again equality holds if and only if v and w are linearly dependent. Finally, assume
that one of v or w is zero and the other is nonzero. It suffices to consider the case
where w = 0 and v = 0, and again we may assume v = 1. As before, if θ is the
angle between v and w then the inequality to be proved reduces to v 2 w 2 sin2 θ ≥
(1/n) w 2 , and since we have assumed w = 0, this is equivalent to
1
v sin θ ≥ √ . (2)
n

November 2021] PROBLEMS AND SOLUTIONS 857

The left side of (2) is the distance from v, which is in the hyperplane x1 + · ·√· + xn = 1, to a
point in the hyperplane x1 + · · · + xn = 0. This distance must be at least 1/ n, the distance
between the two parallel hyperplanes, showing that (2) is true. In this case, equality is
attained if and only if λv = μw + (1/n, . . . , 1/n) for some real λ and μ; that is, if and
only if (1, . . . , 1) is in the span of v and w.
Also solved by R. A. Agnew, K. F. Andersen (Canada), J. N. Caro Montoya (Brazil), R. Chapman (UK),
L. Giugiuc (Romania), L. Han, E. A. Herman, W. Janous (Austria), K. T. L. Koo (China), O. Kouba (Syria),
J. H. Lindsey II, O. P. Lossers (Netherlands), R. Mansuy (France), M. Omarjee (France), E. Schmeichel,
A. Stadler (Switzerland), G. Stoica (Canada), R. Stong, Florida Atlantic University Problem Solving Group,
and the proposer.

A Pell-Type Diophantine Equation

12164 [2020, 179]. Proposed by Nikolai Osipov, Siberian Federal University, Krasnoyarsk,
Russia. Characterize the positive integers d such that (d 2 + d)x 2 − y 2 = d 2 − 1 has a
solution in positive integers x and y.
Solution by Richard Stong, Center for Communications Research, San Diego, CA. There
are solutions exactly when d + 1 is a square. Write d + 1 = gm2 , where g is squarefree.
If g = 1, then d = m2 − 1, and (x, y) = (1, m) is a solution. We show that there is no
solution (x, y) when g ≥ 2.
Fix a squarefree g with g ≥ 2. Let d be minimal such that the specified equation has a
solution (x, y) when d has the form gm2 − 1. Note that d > 1, since when d = 1 the equa-
tion is 2x 2 − y 2 = 0, which famously has no positive integer solutions.
√ With d minimized,
we reduce to checking finitely many cases by first showing x < 2 d for the solution with
smallest positive x and then showing d < √ 14.
It is convenient to work in √the ring Z[ D], where D = d(d + 1), which is the set of
real numbers of the form a + b√ D, where a, b ∈ Z and elements multiply as real numbers.
The norm of an element a + b D is defined to be
√ √
(a + b D)(a − b D),

which equals a 2 − b2 D. With this definition, it is easy to confirm that the norm of a product
is the product of the norms of the factors.
√ v) to an equation of the form u − kv = c corresponds
2 2
A solution
√ (u, to an element
u + v k in Z[ k] with norm c. In particular, the Pell equation u2 − Dv 2√= 1 has the
solution (u, v) = (2d + 1, 2), which corresponds to the number 2d + 1 + 2 D of norm
1. Let α be this number. √
√Now choose β = y + x D with x, y > 0 so that β is the smallest real number in
Z[ D] having norm 1 − d 2 . Thus (x, y) is a solution to y 2 − Dx 2 = 1 − d 2 with minimal
positive x and y. √ √
Because the norm of α is 1, we have α −1 = 2d + 1 − 2 D, and hence α −1 is in Z[ D]
and has norm 1. For suitable integers x and y , we have
√ √ √
α −1 β = (2d + 1 − 2 D)(y + x D) = y + x D.

By the multiplicativity of the norm, α −1 β has norm 1 − d 2 . Also α −1 β < β, since α −1 < 1.
By the minimality of the positive coefficients in β, at least one of x and y is nonpositive.
Furthermore, since α −1 β is a positive real number, x or y is positive. We compute
√ √ √
α −1 β(−y + x D) = (y + x D)(−y + x D) = d 2 − 1,

858 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 128

where the final equality holds because the middle expression is the negative of the norm of
α −1 β. Thus
(d 2 − 1)α d2 − 1 √
= −1 = −y + x D.
β α β
√
Since d 2 − 1 and α −1 β are positive, so is −y + x D. With the restrictions above on x
and y , we conclude y ≤ 0 < x . Since (y )2 − (x )2 D = 1 − d 2 , setting y = 0 would
give (x )2 = (d 2 −√1)/(d(d + 1)) = (d − 1)/d < 1; hence y < 0.
Since −y + x D has norm 1 − d 2 with −y and x both positive, the minimality of β
implies that (d 2 − 1)α/β is at least β, so
√ √
x d(d + 1) < β ≤ (d 2 − 1)α = d 2 − 1( d + d + 1).
Therefore,
√ √
x< d −1+ (d 2 − 1)/d < 2 d.
Next we bound d. Write the original equation as
y 2 = (d + 1)((x 2 − 1)d + 1) = gm2 ((x 2 − 1)d + 1).
It follows that (x 2 − 1)d + 1 = gn2 for some positive integer n. Since g ≥ 2, we have
x = 1 and
(x 2 − 1)d + 1 − (x 2 − 1)(d + 1) 2 − x2
n2 − (x 2 − 1)m2 = = .
g g
√
In the ring Z[ x 2 − 1], consider γ and δ given by

γ =x+ x 2 − 1 and δ = n + m x 2 − 1,

√ 1 and (2 − x )/g, respectively. Let n1 and m1 be positive integers such that

2
with norms
n1 + m1 x 2 − 1 has norm (2 − x 2 )/g in this ring. Setting (x, y) = (x, gm1 n1 ) yields a
solution to the original equation with d + 1√= gm21 . The minimality of d for this g implies
that δ is minimal among all elements of Z[ x 2 − 1] having positive coefficients and norm
(2 − x 2 )/g.
The same argument given earlier for (d√ 2
− 1)α/β shows that (x 2 − 2)γ /(gδ) has norm
(2 − x )/g and can be written as n + m x 2 − 1 with n and m being positive integers.
2

The minimality of δ now implies

gm2 (x 2 − 1) < gδ 2 ≤ (x 2 − 2)γ < 2x(x 2 − 1),
and hence
√
d + 1 = gm2 < 2x < 4 d.
√
Treating this as an inequality in d and applying the quadratic formula yields
√
d < (2 + 3)2 < 14.
√
Since these minimal solutions require d < 14 and x < 2 d, there remain only finitely
many cases to consider. The casework is streamlined by reducing the equation modulo
d − 1, requiring 2x 2 ≡ y 2 (mod d − 1). If d − 1 has as a factor any prime congruent to
±3 modulo 8 (such as 3, 5, or 11), then x must also be√a multiple of this factor, since
2 is not a square modulo any such number. Since x < 2 d, these possibilities are easily
eliminated. For example, when d − 1 = 9, we need only consider 3 and 6 for x in the

November 2021] PROBLEMS AND SOLUTIONS 859

original equation, and neither 110 · 9 − y 2 = 99 nor 110 · 36 − y 2 = 99 has an integer
solution. If d − 1 has 4 as a factor, then x must be even because 2 is not a square mod 4,
and these possibilities can similarly be checked quickly.
For d < 14, in each case d − 1 is a multiple of some element of {3, 4, 5, 11} except for
the remaining cases where d is 2, 3, or 8. The last two of these have g = 1, so only d = 2
needs to be analyzed. In this case, the equation reads 6x 2 − y 2 = 3, hence y = 3z for some
integer z, and 2x 2 − 3z2 = 1. Taking the equation modulo 3 shows that this also fails.
Also solved by the proposer.

An Unexpected Bisection
12165 [2020, 180]. Proposed by Tran Quang Hung and Nguyen Minh Ha, Hanoi, Viet-
nam. Let MN P Q be a square with center K inscribed in triangle ABC with N and
P lying on sides AB and
AC, respectively, while M
and Q lie on side BC. Let
the incircle of BMN touch
side BM at S and side BN
at F , and let the incircle of
CQP touch side CQ at T
and side CP at E. Let L be
the point of intersection of
lines F S and ET . Prove that
KL bisects the segment ST .
Solution I by Haoran Chen, Suzhou, China. Let G and H be the feet of the altitudes to
BC from L and K, respectively. Let J be the intersection of KL and ST , and let I be the
midpoint of ST . Our goal is to show that I and J are the same point.
Let s be the side length of the square MN P Q. Let α = ∠CT E = ∠ST L and β =
∠BSF = ∠T SL. We establish formulas for cot α and cot β. To derive these formulas, let
D be the foot of the perpendicular from E to CT , so that cot α = DT /DE. Let x = QT ,
y = CT = CE, and z = P E. This gives x + z = P Q = s. Since CDE ∼ CQP , we
have DE/CE = QP /CP , so
DE = CE · QP /CP = y(x + z)/(y + z).
Similarly, CD = y(x + y)/(y + z), so
DT = y − CD = y(z − x)/(y + z).
We conclude
DT y(z − x)/(y + z) 2x 2x
cot α = = =1− =1− .
DE y(x + z)/(y + z) x+z s
Similarly, if we let u = MS, then cot β = 1 − 2u/s.
If x = u, then cot α = cot β, so α = β, and the desired conclusion follows by symmetry.
Now assume without loss of generality that x > u, so cot α = 1 − 2x/s < 1 − 2u/s =
cot β. Letting t = GL, we have GT = t cot α < t cot β = GS, so GT < ST /2. Also,
H T = x + s/2 > u + s/2 = H S, so H T > ST /2. Thus G lies between H and T , and
I lies between G and H . Clearly J is also between G and H , so to show that I = J it
suffices to prove I G/I H = J G/J H .
By similar triangles, we have
JG LG t 2t
= = = .
JH KH s/2 s

860 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 128

Also,
ST GS + GT GS − GT t (cot β − cot α)
IG = − GT = − GT = =
2 2 2 2
and
x+s+u s x−u
IH = IS − HS = − +u = .
2 2 2
Therefore
IG t (cot β − cot α)/2 t[(1 − 2u/s) − (1 − 2x/s)] 2t
= = = ,
IH (x − u)/2 x−u s
which completes the proof.
Solution II by L. Richie King, Davidson, NC. Let the bisector of MN P Q parallel to MQ
and NP intersect line ET at U and line F S at V . We show that K is the midpoint of U V .
The result follows from this, since LK is the median of LU V from L, and so it bisects
every section parallel to U V , including ST .
Let O be the center of the incircle of P QC. Note that QO bisects ∠P QC. Let P ,
E , and T be the reflections of P , E, and T in QO. The line P Q is tangent to the incircle
at T , and the lines P E and P T are also tangent to the incircle.
We use some known results about polars. The polar of a point Z with respect to the
incircle of P QC is the line perpendicular to ZO that passes through the image of Z
under inversion in the incircle. A fundamental fact about polars is that if the polar of Z
passes through a point Y then the polar of Y passes through Z.
Since E is fixed under inversion in the incircle, the polar of E is P C, the line tangent
to the incircle at E. Similarly, the polar of T is P Q. Since the polars of both E and T
pass through P , the polar of P must pass through both E and T , so it must be the line
ET . Similarly, the polar of P is E T . Let X be the point of intersection of ET and E T .
Then X lies on the polars of both P and P , so the polar of X is the line P P , which is
perpendicular to QO.
The point X is one of the vertices of the diagonal triangle of the concyclic quadrilateral
ET T E . The other two vertices are the point Y where the lines ET and E T intersect,
which lies on QO, and the point Z at infinity on the lines EE and T T . We now use
one more known fact about polars: the polar of each vertex of the diagonal triangle of a
concyclic quadrilateral is the line through the other two vertices (see H. S. M. Coxeter,
(1998), Non-Euclidean Geometry, 6th ed., Washington, DC: Mathematical Association of
America, p. 57). In particular, P P , which is the polar of X, passes through Y , and therefore
Y is the intersection point of P P and QO. We conclude that P QY is an isosceles right
triangle, with right angle at Y . Therefore Y lies on the bisector of MN P Q parallel to MQ
and N P , so U = Y and U K has length equal to the side length of the square. Similar
reasoning shows that V K has the same length, which establishes our claim that K is the
midpoint of U V .
Editorial comment. Marty Getz and Dixon Jones generalized the problem to a rectangle
inscribed in a triangle, as did the Davis Problem Solving Group. Giuseppe Fera and Giorgio
Tescaro generalized to an inscribed parallelogram.
Also solved by W. Burleson & C. Helms & L. Ide & A. Liendo & M. Thomas, W. Chang, P. De (India),
G. Fera & G. Tescaro (Italy), M. Getz & D. Jones, O. Geupel (Germany), M. Goldenberg & M. Kaplan,
J.-P. Grivaux (France), N. Hodges (UK), W. Hu (China), E.-Y. Jang (Korea), W. Janous (Austria), K. T. L. Koo
(China), O. Kouba (Syria), K.-W. Lau (China), J. H. Lindsey II, C. R. Pranesachar (India), V. Schindler (Ger-
many), A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), M. Vowe (Switzerland), T. Wiandt, T. Zvonaru
(Romania), Davis Problem Solving Group, and the proposers.

November 2021] PROBLEMS AND SOLUTIONS 861

 Asymptotics of a Recursive Sequence
12166 [2020, 180]. Proposed by Erik Vigren, Swedish Institute of Space Physics, Uppsala,
Sweden. Let a0 = 0, and define ak recursively by ak = eak−1 −1 for k ≥ 1.
(a) Prove k/(k + 2) < ak < k/(k + 1) for all k ≥ 1.
(b) Is there a number c such that ak < (k + c)/(k + c + 2) for all k?
Solution by Jean-Pierre Grivaux, Paris, France. We prove part (a) by induction on k. The
base case k = 1 follows from 2 < e < 3. For the induction step, the inductive hypothesis
implies that

e−2/(k+2) < ak+1 < e−1/(k+1) .

Thus it suffices to show that

k+1 k+1
e−2/(k+2) > and e−1/(k+1) < .
k+3 k+2
The first of these is a rearrangement of the inequality
2n x n 1+x
e2x = 1 + 2x + 2x 2 + · · · + + · · · < 1 + 2x + 2x 2 + 2x 3 + · · · =
n! 1−x
for 0 < x < 1 applied at x = 1/(k + 2), and the second is a rearrangement of the inequality
ex > 1 + x for x = 0 applied at x = 1/(k + 1).
The answer to part (b) is no. To establish this, we first study the asymptotics of ak more
carefully. Let vk = ak − 1 = evk−1 − 1. From part (a) we conclude that vk tends to 0 as
k → ∞. Thus we compute
1/vk+1 − 1/vk 1 1 1 + vk − evk −vk2 /2 1
= − = ∼ =− .
(k + 1) − k vk+1 vk vk (e k − 1)
v vk2 2
Hence by the Stolz–Cesàro theorem we have limk→∞ (1/vk )/k = −1/2, or equivalently
vk ∼ −2/k.
Now we compute
1 1 1 evk (vk − 2) + vk + 2 vk3 /6 vk 1
− + = ∼ = ∼− .
vk+1 vk 2 2vk (evk − 1) 2vk2 12 6k
Therefore
1
+ k+1
− 1
+ k  
vk+1 2 vk 2 1 1 1 1
=k − + ∼− ,
Hk − Hk−1 vk+1 vk 2 6
where Hk is the kth harmonic number. Applying the Stolz–Cesàro theorem again, we obtain
1 k Hk−1 ln k
+ ∼− ∼− .
vk 2 6 6
Thus
     
2 2 2 1 k 2 2 ln k 2 ln k
ak − 1 + = vk + = vk · · + ∼ − − = ,
k k k vk 2 k k 6 3k 2
and therefore
 
2
lim k ak − 1 +
2
= ∞.
k→∞ k

862 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 128

However, if a bound of the type given in part (b) held, we would have
 
2 2(c + 2)k
k 2 ak − 1 + < ,
k k+c+2
which is bounded above. Thus no such bound can hold.
Also solved by K. F. Andersen (Canada), R. Chapman (UK), L. Han (USA) & X. Tang (China), N. Hodges
(UK), M. Kaplan, O. Kouba (Syria), G. Lavau (France), J. H. Lindsey II, O. P. Lossers (Netherlands), A. Stadler
(Switzerland), and A. Stenger. Part (a) only solved by P. Bracken, D. Fleischman, O. Geupel (Germany),
W. Janous (Austria), A. Natian, and the proposer.

Bounds on a Function of the Angles and Sides of a Triangle

12168 [2020, 274]. Proposed by Martin Lukarevski, University “Goce Delcev,” Stip, North
Macedonia. Let a, b, and c be the side lengths of a triangle ABC with circumradius R and
inradius r. Prove
2 sec(A/2) sec(B/2) sec(C/2) 1
≤ + + ≤ .
R a b c r

Solution by S. S. Kumar, Portola High School, Irvine, California. Let s and K denote the
semiperimeter and area of ABC, respectively. We first prove the second inequality. Note
that by the half-angle formula and the law of cosines,
  
2 4bc bc
sec(A/2) = = = .
1 + cos A (b + c)2 − a 2 s(s − a)
√ √
By the AM-GM inequality, we have 2 bc ≤ b + c and 2 (s − b)(s − c) ≤ a. Applying
Heron’s formula and the relation K = rs, it follows that

sec(A/2) 1 bc(s − b)(s − c) b+c b+c
= ≤ = .
a a s(s − a)(s − b)(s − c) 4K 4rs
Combining this with similar formulas for the other angles, we have
sec(A/2) sec(B/2) sec(C/2) b+c c+a a+b 4s 1
+ + ≤ + + = = .
a b c 4rs 4rs 4rs 4rs r
To prove the first inequality, we note that by the law of sines, a = 2R sin A, and simi-
larly for the other sides, so the inequality is equivalent to
sec(A/2) sec(B/2) sec(C/2)
+ + ≥ 4.
sin A sin B sin C
Define f (x) = sec(x/2)/ sin x. It is tedious but straightforward to compute that on (0, π ),
1  
f (x) = sec(x/2) csc(x) 4 csc2 (x) + (2 cot(x) − tan(x/2))2 + sec2 (x/2) > 0.
4
Hence, by Jensen’s inequality, we obtain
 
sec(A/2) sec(B/2) sec(C/2) A+B +C
+ + ≥ 3f = 4,
sin A sin B sin C 3
as desired.
Editorial comment. As noted by Omran Kouba, one can also deduce the first inequality by
applying Jensen’s inequality to the function g(x) = − log(cos2 (x) sin(x)) on the interval

November 2021] PROBLEMS AND SOLUTIONS 863

(0, π/2), which is more easily computed to be convex than is f (x). In fact this yields the
stronger inequality

2 3 sec(A/2) sec(B/2) sec(C/2)
≤3 · · ,
R a b c
which along with the AM-GM inequality implies the first inequality.
Also solved by A. Alt, M. Bataille (France), H. Chen, C. Chiser (Romania), G. Fera (Italy), S. Gayen
(India), O. Geupel (Germany), N. Hodges (UK), M. Kaplan & M. Goldenberg, P. Khalili, K. T. L. Koo
(China), O. Kouba (Syria), K.-W. Lau (China), V. Schindler (Germany), A. Stadler (Switzerland), N. Stan-
ciu & M. Drăgan (Romania), R. Stong, R. Tauraso (Italy), M. Vowe (Switzerland), T. Wiandt, L. Wimmer,
M. R. Yegan (Iran), T. Zvonaru (Romania), and the proposer.

Estimating the Logarithmic Derivative of a Chebyshev Polynomial

12171 [2020, 275]. Proposed by Ulrich Abel and Vitaliy Kushnirevych, Technische
Hochschule Mittelhessen, Giessen, Germany. Let Tn be the nth Chebyshev polynomial,
defined by Tn (cos θ ) = cos(nθ ). Prove
Tn (1/z) nz  
=√ + O z2n+1
Tn (1/z) 1 − z2
as z → 0.
Solution by Kenneth F. Andersen, Edmonton, Canada. We prove the equivalent statement,
with x = 1/z,
 
Tn (x) n 1
=√ +O as x → ∞.
Tn (x) x2 − 1 x 2n+1
−n
√ begin with the fact that for x ≥ 1, Tn (x) = (A(x) + A(x) )/2, where A(x) =
n
We
x + x 2 − 1. This can be proved by induction, using the well-known recurrence Tn+1 (x) =
2xTn (x) − Tn−1 (x). Alternatively, if we extend A(x) to x < 1 by an appropriate choice
of a branch of the square root function in the complex numbers, then with x = cos θ for
0 ≤ θ ≤ π we have A(x) = cos θ + i sin θ = eiθ , and therefore
einθ + e−inθ A(x)n + A(x)−n
Tn (x) = Tn (cos θ ) = cos(nθ ) = = .
2 2
This equation can then be√extended to x ≥ 1√by analytic continuation.
Since A (x) = 1 + x/ x 2 − 1 = A(x)/ x 2 − 1 for x > 1, we have
nA(x)n−1 − nA(x)−n−1 n(A(x)n − A(x)−n )
Tn (x) = · A (x) = √ .
2 2 x2 − 1
Therefore
   
 Tn (x) n  n  A(x)n − A(x)−n  2n
   
 T (x) − √ 2 = √ 2  A(x)n + A(x)−n − 1 = √ .
n x −1 x −1 (A(x) + 1) x 2 − 1
2n
√
The desired conclusion now follows because A(x) ∼ 2x and x 2 − 1 ∼ x as x → ∞.
Editorial comment. The problem statement above corrects a typographical error from the
original printing.
Also solved by A. Berkane (Algeria), R. Chapman (UK), H. Chen, O. Geupel (Germany), J.-P. Grivaux
(France), L. Han (USA) & X. Tang (China), N. Hodges (UK), K. T. L. Koo (China), O. Kouba (Syria),
M. Omarjee (France), A. Stadler (Switzerland), R. Tauraso (Italy), D. Terr, E. I. Verriest, T. Wiandt, and the
proposer.

864 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 128

SOLUTIONS

An (Almost) Impossible Integral

12158 [2020, 86]. Proposed by Hervé Grandmontagne, Paris, France. Prove
 1
(ln x)2 arctan x 21 1 1
dx = π ζ (3) − π 2 G − π 3 ln 2,
0 1+x 64 24 32

∞ 
∞
where ζ (3) is Apéry’s constant 1/k 3 and G is Catalan’s constant (−1)k /(2k + 1)2 .
k=1 k=0

Solution by the proposer. Let R be the function defined by

 x 2  1
ln t x ln2 (tx)
R(x) = dt = dt.
0 1+t 0 1 + tx
Integrating the given integral J by parts yields
 1
 1 R(x)
J = R(x) arctan x 0 − dx
0 1 + x2
  1 1
π 1 ln2 x x ln2 (tx)
= dx − dt dx.
4 0 1+x 0 0 (1 + x )(1 + tx)
2

Observe that
x t x t
+ = + .
(1 + tx)(1 + x ) (1 + tx)(1 + t )
2 2 (1 + t )(1 + x ) (1 + t )(1 + x 2 )
2 2 2

Multiplying by ln2 (tx), integrating both sides, and exploiting symmetry under interchange
of x and t gives
 1 1  1 1
x ln2 (tx) x ln2 (tx)
dt dx = dt dx.
0 (1 + x )(1 + tx) 0 (1 + t )(1 + x )
2 2 2
0 0

756 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 128

Thus after rewriting ln2 (tx) as (ln t + ln x)2 we find
  1 1
π 1 ln2 x x ln2 (tx)
J = dx − dt dx
4 0 1+x 0 0 (1 + t )(1 + x )
2 2

  1  1  1  1
π 1 ln2 x ln2 t x ln t x ln x
= dx − dt dx − 2 dt dx
4 0 1+x 0 1 + t 2
0 1 + x 2
0 1 + t 2
0 1 + x2
 1  1
1 x ln2 x
− dt dx.
0 1+t 0 1+x
2 2

The component integrals of this last expression are all fairly standard. The nonelementary
ones are
 1  1
ln t ln2 t π3
dt = −G, dt = ,
0 1+t 0 1+t
2 2 16
 1  1
x ln x 1 ln y 1 π2
dx = dy = Li2 (−1) = − ,
0 1+x 2 4 0 1+y 4 48
and
 
1
x ln2 x 1 1
ln2 y 1 3
dx = dy = − Li3 (−1) = ζ (3),
0 1+x 2 8 0 1+y 4 16
where we have substituted y = x in the last two integrals. Plugging these all in, we get
2

π 3 π 3 ln 2 −π 2 π 3
J = · ζ (3) − · − 2(−G) · − · ζ (3)
4 2 16 2 48 4 16
21 1 1
= π ζ (3) − π 2 G − π 3 ln 2.
64 24 32
Editorial comment. Several solvers noted that this integral appears in Section 1.24, pp.
14–15 of C. I. Vălean (2019), (Almost) Impossible Integrals, Sums, and Series, Cham:
Springer, both explicitly and as the special case n = 1 of the more general integral
 1
(ln x)2n arctan x π 1
dx = (1 − 2−2n )ζ (2n + 1)(2n)! + β(2n + 2)(2n)!
0 1 + x 4 2
 2n      
π d πs 3 s 1 s
− lim csc ψ − − ψ −
16 s→0 ds 2n 2 4 4 4 4
   
πs s 1 s
+ sec ψ 1− −ψ − − 2π csc(π s) ,
2 4 2 4
where ζ is the Riemann zeta function, ψ is the digamma function, and β is the Dirichlet
beta function.
Also solved by A. Berkane (Algeria), P. Bracken, H. Chen, G. Fera (Italy), B. Huang, K. T. L. Koo (China),
O. Kouba (Syria), K.-W. Lau (China), M. A. Prasad (India), S. Sharma (India), F. Sinani (Kosovo), A. Stadler
(Switzerland), S. M. Stewart (Australia), R. Stong, R. Tauraso (Italy), C. I. Vălean (Romania), J. Van Casteren
& L. Kempeneers (Belgium), T. Wiandt, T. Wilde (UK), and Y. Zhou & M. L. Glasser.

The Neyman–Pearson Lemma in Disguise

12159 [2020, 86]. Proposed by Rudolf Avenhaus, Universität der Bundeswehr München,
Neubiberg, Germany, and Thomas Krieger, Forschungszentrum Jülich, Jülich, Germany.
Let  denote the distribution function of a standard normal random variable, and let U

October 2021] PROBLEMS AND SOLUTIONS 757

denote its inverse function. Let n be a positive integer, and suppose 0 < α < 1 and μ ≥ 0.
Prove
√ √ n
 U (α) − nμ ≤  U ( n α) − μ .

Solution by the proposers. The inequality in the problem is an equality if μ = 0. Thus we

may assume μ > 0.
Consider the following hypothesis testing problem: Let X1 , . . . , Xn be independent and
identically normally distributed random variables with variance 1, where under the null
hypothesis H0 their expected values are all zero, and under the alternative hypothesis H1
they are μ. In other words,

N (0, 1) under H0 ,
Xi ∼
N (μ, 1) under H1 .
We consider two decision procedures for testing these hypotheses: a simple intuitive
test and the Neyman–Pearson test. In the simple test, we reject the null hypothesis if
maxi=1,...,n Xi is larger than a constant k, in other words, if the sample (x1 , . . . , xn ) belongs
to the critical region Cs defined by
 
Cs = (x1 , . . . , xn ) : max xi > k .
i=1,...,n

We choose the threshold k so that the probability of a type I error is 1 − α; that is,
PH0 (Cs ) = 1 − α. This means
   n
α = PH0 (Cs ) = PH0 max Xi ≤ k = PH0 (Xi ≤ k) = ((k))n ,
i=1,...,n
i=1
√
and solving for k yields k = U ( n α). If we let βs denote the probability of a type II error
for the simple test, then
 
√
βs = PH1 max Xi ≤ k = ((k − μ))n = ((U ( n α) − μ))n . (1)
i=1,...,n

The Neyman–Pearson test uses the critical region CNP defined by

 
φH1 (x1 , . . . , xn )
CNP = (x1 , . . . , xn ) : >k ,
φH0 (x1 , . . . , xn )
for some positive constant k , where the joint density functions φH0 under H0 and φH1
under H1 are given by

n
1 2 
n
1 2
φH0 (x1 , . . . , xn ) = √ e−xi /2 and φH1 (x1 , . . . , xn ) = √ e−(xi −μ) /2 .
i=1
2π i=1
2π
Using these joint density functions, the critical region can be rewritten as


n
CNP = (x1 , . . . , xn ) : xi > k
i=1

for some constant k . Once againwe choose k , and therefore k , so that the probability of
a type I error is 1 − α. Because ni=1 Xi is normally distributed, with distribution given by

n
N (0, n) under H0 ,
Xi ∼
i=1
N (nμ, n) under H1 ,

758 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 128

we obtain
 n   
 k
α = PH0 (CNP ) = PH0 Xi ≤ k = √ ,
i=1
n
√
and therefore k = nU (α). The probability βNP of a type II error is then given by the
formula
 n   
 k − nμ √
βNP = PH1 (CNP ) = PH1 Xi ≤ k = √ = (U (α) − nμ). (2)
i=1
n

According to the Neyman–Pearson lemma, βNP ≤ βs , and by (1) and (2), this is equiv-
alent to the required inequality.
Editorial comment. The proposers’ solution shows that the inequality can be proved with-
out performing any calculations on the formulas on the two sides of the inequality. Richard
Stong showed that the
√ inequality can also be proved by direct calculations with these for-
mulas. Letting y = nμ, the requested inequality reads
√
(U (α) − y) ≤ ((U (α 1/n ) − y/ n))n .
Since this inequality is an equality when n = 1, it suffices to show that the right side is a
√ of n for all real n ≥ 1. Taking a logarithmic derivative and letting
nondecreasing function
x = U (α 1/n ) − y/ n, we find that this is equivalent to
(x) log (x) x U (α 1/n )
− ≥ α 1/n log(α 1/n )U (α 1/n ) − , (3)
φ(x) 2 2
where φ is the density function for the standard normal distribution, that is,
1 2
φ(x) = √ e−x /2 .
2π
Next we note that x ≤ U (α 1/n ), with equality if μ = 0 and y = 0, and in this case (3) is
an equality. Thus it suffices to show that the left side is a nonincreasing function of x, or
equivalently, taking a derivative, that
1 x(x) log (x)
+ log (x) + ≤ 0.
2 φ(x)
At this point, all of the parameters n, α, and μ have been eliminated, and the problem
has been reduced to an inequality involving the standard normal distribution and density
functions. Some further elaborate calculations verify this inequality.
No solutions were received other than the proposers’ solution and the solution of R. Stong.

Fibonacci and Lucas: A Golden Braid

12160 [2020, 179]. Proposed by Hideyuki Ohtsuka, Saitama, Japan, and Roberto Tauraso,
Univerità di Roma “Tor Vergata,” Rome, Italy. Let Fn be the nth Fibonacci number, and
let Ln be the nth Lucas number. (These numbers are defined recursively by F1 = F2 = 1
and Fn+2 = Fn+1 + Fn when n ≥ 1, and by L1 = 1, L2 = 3, and Ln+2 = Ln+1 + Ln when
n ≥ 1.) Prove
 n    n    n  
2n + 1 2n + 1 2k n−k
F2k+1 = 5n
and L2k+1 = 5
k=0
n−k k=0
n−k k=0
k

for all n ∈ N.

October 2021] PROBLEMS AND SOLUTIONS 759

Solution
√ by Robin Chapman, University of Exeter, Exeter, UK. Let φ be the golden ratio
( 5 + 1)/2. The familiar formulas for the Fibonacci numbers (Binet’s formula) and the
Lucas numbers are
φ n − (−φ)−n
Fn = √ and Ln = φ n + (−φ)−n .
5
√
Combining the two formulas, we get φ n = (Ln + 5 Fn )/2.
Let
n   √ n   n  
 2n + 1 L2k+1 + 5 F2k+1  2n + 1 2k+1  2n + 1 2n−2k+1
Sn = = φ = φ .
k=0
n−k 2 k=0
n−k k=0
k

Pascal’s formula yields 2n+1

k
= 2n−1
k
+ 2 2n−1
k−1
+ 2n−1
k−2
, a formula that holds even for
k ∈ {0, 1} if we take j = 0 when j is negative. We use this to compute
m

n 
  n 
  n 
 
2n − 1 2n − 1 2n − 1
Sn = φ 2n−2k+1
+2 φ 2n−2k+1
+ φ 2n−2k+1
k=0
k k=1
k−1 k=2
k−2
n 
  n−1 
  n−2  
2n − 1 2n−2k+1 2n − 1 2n−2k−1  2n − 1 2n−2k−3
= φ +2 φ + φ
k=0
k k=0
k k=0
k
n−1 
     
2n − 1 −22n − 1 2n − 1 −1
= φ 2n−2k−1
(φ + 2 + φ ) +
2
φ− φ
k=0
k n n−1
 
1 2n
= 5Sn−1 + .
2 n
√
In the last step, we used 2n−1 n
= 2n−1
n−1
= 12 2n
n
, along with φ + φ −1 = 5 and
φ − φ −1 = 1. With the initial condition S0 = φ, the recurrence gives
n   √ n  
 2n + 1 L2k+1 + 5 F2k+1 1 n√  2k n−k
Sn = = 5 5+ 5 .
k=0
n − k 2 2 k=0
k
√
As 5 is irrational, this separates into the two required identities.
Also solved by U. Abel & G. Arends (Germany), A. Berkane (Algeria), B. Bradie, B. Burdick, W. Chang,
H. Chen (China), G. Fera (Italy), P. Fulop (Hungary), J. Grivaux (France), N. Hodges (UK), Y. Ionin,
K. T. L. Koo (China), O. Kouba (Syria), P. Lalonde (Canada), G. Lavau (France), O. P. Lossers (Netherlands),
C. Pranesachar (India), L. Shapiro, A. Stadler (Switzerland), R. Stong, B. Sury (India), D. Terr, J. Van hamme
(Belgium), M. Vowe (Switzerland) M. Wildon (UK), and the proposer.

Integer Pairs on an Ellipse

12161 [2020, 179]. Proposed by José Hernández Santiago, Guerrero, Mexico. Let N (C)
be the number of pairs (u, v) ∈ Z × Z satisfying u2 + uv + v 2 = C. Prove that 6 divides
N(C) for every positive integer C.
Solution by Allen Stenger, Boulder, CO. The number of pairs (u, v) satisfying the given
equation is the same as the number of pairs satisfying u2 − uv + v 2 = C due to the map-
ping of (u, v) to (u, −v). We work with the second equation.
We work in the ring Z[ω], where ω = e2πi/3 . The elements of this ring have the form
u + vω, where u and v are integers, and the norm of this element is u2 − uv + v 2 . Thus our
number N(C) is equal to the number of elements of Z[ω] whose norm is C. The ring has

760 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 128

six units, namely ±1, ±ω, and ±ω2 , and so each nonzero ring element has six associates
(including itself). All associates have the same norm, so the total number of elements with
a given norm is a multiple of 6.
The number N (C) is finite, since 4C = (2u − v)2 + 3v 2 , which implies that v is
bounded and then also u is bounded. Since u and v are integers, the number of solutions
(u, v) is finite.
Editorial comment. A related result is mentioned in H. L. Keng (1982), Introduction to
Number Theory, Berlin: Springer. Exercise 2 on p. 308 states, “The number of solutions to
x 2 + xy + y 2 = k is 6E(k), where E(k) is the number of divisors of k of the form 3h + 1
minus the number of divisors of the form 3h + 2.” An anonymous solver noted that the
result is given with three solutions as Problem 195 in M. I. Krusemeyer, G. T. Gilbert, and
L. C. Larson (2012), A Mathematical Orchard: Problems and Solutions, Washington, DC:
MAA, 338–340.
Solvers used various techniques, such as (a) showing that if (u, v) is a solution to
u2 + uv + v 2 = C, then so is (v, −(u + v)), and that iterating this observation yields six
distinct solutions, (b) bringing in group actions, linear algebra, and/or the ring of integers
Z(ω), where ω = exp(2π i/3), and (c) using automorphisms of binary quadratic forms.
Most solvers tacitly assumed that N (C) is finite.
Also solved by K. F. Andersen (Canada), A. Berkane (Algeria), A. J. Bevelacqua, J. N. Caro Montoya (Brazil),
N. Caro (Brazil), W. Chang, R. Chapman (UK), C. Curtis & J. Boswell, R. Dempsey, A. Dixit (Canada)
& S. Pathak (USA), G. Fera (Italy), N. Garson (Canada), K. Gatesman, O. Geupel (Germany), J.-P. Grivaux
(France), J. W. Hagood, Y. J. Ionin, W. Janous (Austria), K. T. L. Koo (China), O. Kouba (Syria), C. P. A. Kumar
(India), P. Lalonde (Canada), G. Lavau (France), O. P. Lossers (Netherlands), C. Moe, A. Natian, A. Pathak,
L. J. Peterson, C. R. Pranesachar (India), J. Schlosberg, E. Schmeichel, J. H. Smith, A. Stadler (Switzerland),
D. Stone & J. Hawkins, R. Stong, R. Tauraso (Italy), D. Terr, M. Vowe (Switzerland), the Missouri State
University Problem Solving Group, and the proposer.

A Triangle Inequality from the Triangle Inequality

12162 [2020, 179]. Proposed by Dao Thanh Oai, Thai Binh, Vietnam, and Leonard
Giugiuc, Drobeta Turnu Severin, Romania. Consider a triangle with sides of lengths a,
b, and c and with area S. Prove
  
a 2 + b2 − 4S + a 2 + c2 − 4S ≥ b2 + c2 − 4S,

and determine when equality holds.

Solution by Yagub Aliyev, Baku, Azerbaijan. In the figure, ABDE is a square and
BDF ∼ = ABC. Applying the law of cosines
in ACE, we get c
 a
CE = b2 + c2 − 2bc cos ∠CAE
  b
= b2 + c2 − 2bc sin A = b2 + c2 − 4S.

Similar calculations show √that F E =

√
a 2 + c2 − 4S and CF = a 2 + b2 − 4S.
By the triangle inequality, CF + F E ≥ EC,
and equality holds if and only if F lies on the
segment CE.
Editorial comment. Most solvers used analytical approaches and provided one of various
equivalent conditions for equality:

October 2021] PROBLEMS AND SOLUTIONS 761

 • With notation as in the diagram above, C lies on the upper-left quarter of the circle with
diameter DE;
• a is the shortest side and 5(a 4 + b4 + c4 ) = 6(a 2 b2 + b2 c2 + c2 a 2 );
• a is the shortest side and a 4 + b4 + c4 = 24S 2 ;
• a is the shortest side and a 2 + b2 + c2 = 8S;
• A is the smallest angle and cot A + cot B + cot C = 2;
• A is the smallest angle and cot ω = 2, where ω is the Brocard angle;
√ √ √
• cot A = cot B + cot C; or
• for some real k > 0,
(k + 1)2 1 k2
cot A = , cot B = , cot C = .
k2 + k + 1 k2 + k + 1 k2 + k + 1

Also solved by M. Bataille (France), R. Chapman (UK), C. Curtis, G. Fera & G. Tescaro (Italy), K. Gatesman,
N. Hodges (UK), W. Janous (Austria), B. Karaivanov (USA) & T. S. Vassilev (Canada), P. Khalili, K. T. L. Koo
(China), O. Kouba (Syria), K.-W. Lau (China), D. J. Moore, K. S. Palacios (Peru), C. R. Pranesachar (India),
J. Schlosberg, A. Stadler (Switzerland), R. Stong, R. Tauraso (Italy), F. Visescu (Romania), T. Wiandt, L. Zhou,
T. Zvonaru (Romania), Davis Problem Solving Group, and the proposer.

Arithmetic Progressions and Fibonacci Numbers

12167 [2020, 274]. Proposed by Nick MacKinnon, Winchester College, Winchester, UK.
Let S be the set of positive integers expressible as the sum of two nonzero Fibonacci
numbers. Show that there are infinitely many six-term arithmetic progressions of numbers
in S but only finitely many such seven-term arithmetic progressions.
Solution by Richard Stong, Center for Communications Research, San Diego, CA. Since
2Fn = Fn+1 + Fn−2 , we may view each element of S as a sum of two distinct Fibonacci
numbers. Also note that any sum Fn + Fk with k < n lies in the interval (Fn , Fn+1 ]. Hence
the elements of S in this interval are precisely the sums of Fn with smaller Fibonacci
numbers. In particular, the expression of any given s ∈ S as a sum of two distinct Fibonacci
numbers is unique, and the larger is the largest Fn with Fn < s (except for s = 2).
To find 6-term arithmetic progressions, start with Fn (for some n ≥ 3, so that Fn =
Fn−1 + Fn−2 ∈ S) and let the common difference in the progression be Fn+3 . The resulting
6-term arithmetic progression with its terms shown to be in S is

Fn = Fn−1 + Fn−2 ,
Fn + Fn+3 = Fn + Fn+3 ,
Fn + 2Fn+3 = Fn+4 + Fn+2 ,
Fn + 3Fn+3 = Fn+5 + Fn+2 ,
Fn + 4Fn+3 = Fn+5 + Fn+4 ,
Fn + 5Fn+3 = Fn+6 + Fn+3 .

Such a progression cannot be extended to seven terms, since (a) the preceding term
Fn − Fn+3 is negative, and (b) the next term Fn+6 + 2Fn+3 , being smaller than Fn+7 , can
only be in S if 2Fn+3 is a Fibonacci number. Since Fn+4 < 2Fn+3 < Fn+5 , it is not a
Fibonacci number.
To complete the solution, we prove a stronger statement, namely that except for
small values, these progressions are the only 6-term progressions in S. (The exceptions

762 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 128

are subsets of the 10-term progression 2, 3, . . . , 11 and the two 7-term progressions
2, 6, 10, . . . , 26 and 3, 5, 7, . . . , 15; this requires checking small cases.)
For any 6-term progression {a0 + kd}5k=0 , we have
a0 + 5d 5 Fn+1
< ≤
a0 + 3d 3 Fn
when n ≥ 4. Thus at least two of the last three terms in this progression lie in the same
interval of the form (Fn , Fn+1 ]. Since we may ignore cases with n ≤ 8, we may assume
we have a 5-term arithmetic progression {aj }5j =1 whose last two terms lie in the interval
(Fn+5 , Fn+6 ] for some n ≥ 4. (We have chosen the indices here to match the example
above.) We now consider two cases.
Case 1: The top three terms lie in the interval (Fn+5 , Fn+6 ]. These terms (a3 , a4 , a5 )
must be (Fn+5 + Fj , Fn+5 + Fk , Fn+5 + Fl ), where j < k < l ≤ n + 4. Since the terms
are in progression, Fj + Fl = 2Fk = Fk−2 + Fk+1 . Because representations as the sum
of two Fibonacci numbers are unique, l = k + 1 and j = k − 2. Hence k ≤ n + 3, the
common difference is Fk−1 , and the preceding term a2 must satisfy

a2 = Fn+5 + Fk−2 − Fk−1 = Fn+5 − Fk−3 = Fn+4 + Fm

for some m. This forces Fk−3 + Fm = Fn+3 = Fn+2 + Fn+1 , which cannot hold since
k − 3 ≤ n and expressions as sums of distinct Fibonacci numbers are unique.
Case 2: Only the top two terms of the 5-term progression lie in (Fn+5 , Fn+6 ]. Those
terms a4 and a5 must be Fn+5 + Fk and Fn+5 + Fl , where k < l ≤ n + 4. The previous
term a3 is Fn+5 + 2Fk − Fl ; it must satisfy
a1 + a5 Fn+5 + Fl
a3 = > .
2 2
Eliminating Fl (by summing 1/3 of the equality and 2/3 of the inequality for a3 )
yields a3 > 23 Fn+5 + 23 Fk . By several applications of the Fibonacci recurrence, 23 Fn+5 =
Fn+4 + 13 Fn+1 , so
1 2
a3 > Fn+4 + Fn+1 + Fk .
3 3
Since a3 exceeds Fn+4 , we conclude a3 = Fn+4 + Fj for some j ≤ n + 3. Fur-
thermore, since 13 Fn+1 + 23 Fk > max(Fk−1 , 2), we have j ≥ max(k, 4). From a3 =
Fn+5 + 2Fk − Fl = Fn+4 + Fj , we conclude
Fn+3 + 2Fk = Fl + Fj ,

and hence at least one of j and l is at least as large as n + 3.

Since Fn+1 /Fn < 2 < Fn+2 /Fn whenever n ≥ 3, one Fibonacci number is twice
another only for the initial values 1, 1, 2. If j = n + 3, then Fl = 2Fk , so Fk = 1, and the
last three terms of the progression are Fn+5 , Fn+5 + 1, and Fn+5 + 2, but Fn+5 − 1 ∈ / S. If
l = n + 3, then Fj = 2Fk , but we already have Fj > 2.
Thus l = n + 4, which yields Fj + Fn+2 = 2Fk = Fk+1 + Fk−2 . If j = n + 2 = k, then
we obtain the family described earlier, extending to 6-term progressions. Otherwise, Fj and
Fn+2 are distinct, and hence one of j and n + 2 must equal k − 2. It is not j because j ≥ k,
and it is not n + 2 because k < n + 4. Hence we cannot produce such a 6-term arithmetic
progression outside the family described earlier.
Also solved by J. Christopher, N. Hodges (UK), Y. J. Ionin, J. H. Nieto (Venezuela), A. Pathak (India),
A. Stadler (Switzerland), R. Tauraso (Italy), T. Wilde (UK), and the proposer.

October 2021] PROBLEMS AND SOLUTIONS 763