Solutions to the 78th William Lowell Putnam Mathematical Competition

Saturday, December 2, 2017

Kiran Kedlaya and Lenny Ng

A1 We claim that the positive integers not in S are 1 and all Second solution. We establish directly that Qn (x) =
multiples of 5. If S consists of all other natural numbers, xQn−1 (x) − Qn−2 (x), which again suffices. From the
then S satisfies the given conditions: note that the only equation
perfect squares not in S are 1 and numbers of the form
(5k)2 for some positive integer k, and it readily follows 1 = Qn−1 (x)2 − Qn (x)Qn−2 (x) = Qn (x)2 − Qn+1 (x)Qn−1 (x)
that both (b) and (c) hold.
we deduce that
Now suppose that T is another set of positive integers
satisfying (a), (b), and (c). Note from (b) and (c) that if Qn−1 (x)(Qn−1 (x) + Qn+1 (x)) = Qn (x)(Qn (x) + Qn−2 (x)).
n ∈ T then n + 5 ∈ T , and so T satisfies the following
property: Since deg(Qn (x)) = n by an obvious induction, the
polynomials Qn (x) are all nonzero. We may thus
(d) if n ∈ T , then n + 5k ∈ T for all k ≥ 0. rewrite the previous equation as
The following must then be in T , with implications la-
Qn+1 (x) + Qn−1 (x) Qn (x) + Qn−2 (x)
beled by conditions (b) through (d): = ,
Qn (x) Qn−1 (x)
c c d b d b
2 ⇒ 49 ⇒ 542 ⇒ 562 ⇒ 56 ⇒ 121 ⇒ 11 Q (x)+Q (x)
meaning that the rational functions n Q (x) n−2
are all
d b d b n−1
11 ⇒ 16 ⇒ 4 ⇒ 9 ⇒ 3 equal to a constant value. By taking n = 2 and comput-
d b ing from the definition that Q2 (x) = x2 − 1, we find the
16 ⇒ 36 ⇒ 6
constant value to be x; this yields the desired recurrence.
Since 2, 3, 4, 6 ∈ T , by (d) S ⊆ T , and so S is smallest. Remark: By induction, one may also obtain the ex-
plicit formula
A2 First solution. Define Pn (x) for P0 (x) = 1, P1 (x) = x,
bn/2c
and Pn (x) = xPn−1 (x) − Pn−2 (x). We claim that Pn (x) =
 
n − k n−2k
k
Qn (x) for all n ≥ 0; since Pn (x) clearly is a polynomial Qn (x) = ∑ (−1) x .
k=0 k
with integer coefficients for all n, this will imply the
desired result.
Remark: In light of the explicit formula for Qn (x),
Since {Pn } and {Qn } are uniquely determined by their Karl Mahlburg suggests the following bijective inter-
respective recurrence relations and the initial conditions pretation of the identity Qn−1 (x)2 − Qn (x)Qn−2 (x) = 1.
P0 , P1 or Q0 , Q1 , it suffices to check that {Pn } satis- Consider the set Cn of integer compositions of n with
fies the same recurrence as Q: that is, (Pn−1 (x))2 − all parts 1 or 2; these are ordered tuples (c1 , . . . , ck ) such
Pn (x)Pn−2 (x) = 1 for all n ≥ 2. Here is one proof of that c1 + · · · + ck = n and ci ∈ {1, 2} for all i. For a given
this: for n ≥ 1, define the 2 × 2 matrices composition c, let o(c) and d(c) denote the number of
    1’s and 2’s, respectively. Define the generating function
P (x) Pn (x) x −1
Mn = n−1 , T=
Pn−2 (x) Pn−1 (x) 1 0 Rn (x) = ∑ xo(c) ;
c∈Cn
with P−1 (x) = 0 (this value being consistent with the
recurrence). Then det(T ) = 1 and T Mn = Mn+1 , so by then Rn (x) = ∑ j n−j j xn−2 j , so that Qn (x) =


induction on n we have
i−n/2 Rn (ix). (The polynomials Rn (x) are some-
(Pn−1 (x))2 − Pn (x)Pn−2 (x) = det(Mn ) = det(M1 ) = 1. times called Fibonacci polynomials; they satisfy
Rn (1) = Fn . This interpretation of Fn as the cardinality
of Cn first arose in the study of Sanskrit prosody,
Remark: A similar argument shows that any second-
specifically the analysis of a line of verse as a sequence
order linear recurrent sequence also satisfies a quadratic
of long and short syllables, at least 500 years prior to
second-order recurrence relation. A familiar example
the work of Fibonacci.)
is the identity Fn−1 Fn+1 − Fn2 = (−1)n for Fn the n-
th Fibonacci number. More examples come from var- The original identity is equivalent to the identity
ious classes of orthogonal polynomials, including the
Chebyshev polynomials mentioned below. Rn+1 (x)Rn−1 (x) − Rn (x)2 = (−1)n−1 .
 2
Rb
This follows because if we identify the composition c MSTang) Since a ( f (x) − g(x)) dx = 0, we have
with a tiling of a 1 × n rectangle by 1 × 1 squares and
( f (x))n+2 ( f (x))n+1
Z b 
1 × 2 dominoes, it is almost a bijection to place two
In+1 − In = − dx
tilings of length n on top of each other, offset by one a (g(x))n+1 (g(x))n
square, and hinge at the first possible point (which is
( f (x))n+1
Z b
the first square in either). This only fails when both = ( f (x) − g(x)) dx
tilings are all dominoes, which gives the term (−1)n−1 . (g(x))n+1
a
( f (x))n+1
Z b 
Remark: This problem appeared on the = − 1 ( f (x) − g(x)) dx
2012 India National Math Olympiad; see a (g(x))n+1
( f (x) − g(x))2 (( f (x))n + · · · + g(x)n )
https://artofproblemsolving.com/community/ Z b
c6h1219629. Another problem based on the same idea = dx.
a (g(x))n+1
is problem A2 from the 1993 Putnam.
The integrand is continuous, nonnegative, and not iden-
A3 First solution. Extend the definition of In to n = 0, so tically zero; hence In+1 − In > 0.
that I0 = ab f (x) dx > 0. Since ab ( f (x) − g(x)) dx = 0,
R R

we have To prove that limn→∞ In = ∞, note that we cannot

have f (x) ≤ g(x) identically, as then the equality
Z b Rb Rb
f (x)
I1 − I0 = ( f (x) − g(x)) dx a f (x) dx = a g(x) dx would imply f (x) = g(x) iden-
a g(x) tically. That is, there exists some t ∈ [a, b] such that
Z b
( f (x) − g(x))2 f (t) > g(t). By continuity, there exist a quantity c > 1
= dx > 0, and an interval J = [t0 ,t1 ] in [a, b] such that f (x) ≥ cg(x)
a g(x)
for all x ∈ J. We then have
where the inequality follows from the fact that the in-
( f (x))n+1
Z t1 Z t1
tegrand is a nonnegative continuous function on [a, b] n
In ≥ dx ≥ c f (x) dx;
that is not identically 0. Now for n ≥ 0, the Cauchy– t0 (g(x))n t0
Schwarz inequality gives
since f (x) > 0 everywhere, we have tt01 f (x) dx > 0 and
R
Z b
( f (x))n+1
 Z b
( f (x))n+3
 hence In is bounded below by a quantity which tends to
In In+2 = dx dx ∞.
a (g(x))n a (g(x))
n+2
2 Remark: One can also give a variation of the second
( f (x))n+2
Z b 
≥ dx 2
= In+1 . half of the solution which shows directly that In+1 −
n+1
a (g(x)) In ≥ cn d for some c > 1, d > 0, thus proving both asser-
tions at once.
It follows that the sequence {In+1 /In }∞n=0 is nondecreas-
ing. Since I1 /I0 > 1, this implies that In+1 > In for Third solution. (from David Savitt, via Art of Problem
all n; also, In /I0 = ∏n−1 n Solving) Extend the definition of In to all real n, and
k=0 (Ik+1 /Ik ) ≥ (I1 /I0 ) , and so
limn→∞ In = ∞ since I1 /I0 > 1 and I0 > 0. note that
Z b Z b
Remark: Noam Elkies suggests the following vari-
I−1 = g(x) dx = f (x) dx = I0 .
ant of the previous solution, which eliminates the need a a
to separately check that I1 > I0 . First, the proof that
2
In In+2 ≥ In+1 applies also for n = −1 under the con- By writing
vention that I−1 = ab g(x) dx (as in the fourth solution
R
Z b
below). Second, this equality must be strict for each In = exp((n + 1) log f (x) − n log g(x)) dx,
a
n ≥ −1: otherwise, the equality condition in Cauchy–
Schwarz would imply that g(x) = c f (x) identically for we see that the integrand is a strictly convex function of
some c > 0, and the equality ab f (x) dx = ab g(x) dx
R R
n, as then is In . It follows that In is strictly increasing
would then force c = 1, contrary to assumption. Con- and unbounded for n ≥ 1.
sequently, the sequence In+1 /In is strictly increasing; Fourth solution. (by David Rusin) Again, extend the
since I0 /I−1 = 1, it follows that for n ≥ 0, we again definition of In to n = −1. Now note that for n ≥ 0 and
have In+1 /In ≥ I1 /I0 > 1 and so on. x ∈ [a, b], we have
Second solution. (from Art of Problem Solving, user
 !
f (x) n+1 f (x) n
  
( f (x) − g(x)) − ≥0
g(x) g(x)

because both factors have the same sign (depending

on the comparison between f (x) and g(x)); moreover,
equality only occurs when f (x) = g(x). Since f and g
are not identically equal, we deduce that

In+1 − In > In − In−1

 3

and so in particular whence S1 ≤ 7.4N − 27.5 ≤ SN−4 . It follows that there

is some k such that Sk ∈ [7.4N − 40, 7.4N − 15], since
In+1 − In ≥ I1 − I0 > I0 − I−1 = 0. this interval has length 25 and 7.4N − 27.5 lies inside it.

This proves both claims. For this value of k, note that both Sk and 7.4N are inte-
gers (the latter since the sum of all scores in the class is
Remark: This problem appeared in 2005 on the integer (7.4)(2N) and so N must be divisible by 5).
an undergraduate math olympiad in Brazil. See Thus there is an integer m with 15 ≤ m ≤ 40 for which
https://artofproblemsolving.com/community/ Sk = 7.4N − m. By our first claim, we can choose five
c7h57686p354392 for discussion. scores from a1 , . . . , a10 whose sum is m. When we add
these to the sum of the N − 5 scores ak+10 , . . . , ak+N+4 ,
A4 First solution. Let a1 , . . . , a2N be the scores in non-
we get precisely 7.4N. We have now found N scores
decreasing order, and define the sums si = ∑i+N j=i+1 a j whose sum is 7.4N and thus whose average is 7.4.
for i = 0, . . . , N. Then s0 ≤ · · · ≤ sN and s0 + sN =
∑2N
j=1 a j = 7.4(2N), so s0 ≤ 7.4N ≤ sN . Let i be the
Third solution. It will suffices to show that given any
largest index for which si ≤ 7.4N; note that we can- partition of the students into two groups of N, if the
not have i = N, as otherwise s0 = sN = 7.4N and hence sums are not equal we can bring them closer together by
a1 = · · · = a2N = 7.4, contradiction. Then 7.4N − si < swapping one pair of students between the two groups.
si+1 − si = ai+N+1 − ai and so To state this symbolically, let S be the set of students
and, for any subset T of S, let ΣT denote the sum of
ai < si + ai+N+1 − 7.4N ≤ ai+N+1 ; the scores of the students in T ; we then show that if
S = A ∪ B is a partition into two N-element sets with
since all possible scores occur, this means that we can ΣA > ΣB, then there exist students a ∈ A, B ∈ B such
find N scores with sum 7.4N by taking ai+1 , . . . , ai+N+1 that the sets
and omitting one occurrence of the value si + ai+N+1 −
7.4N. A0 = A \ {a} ∪ {b}, B0 = A \ {b} ∪ {a}
Remark: David Savitt (via Art of Problem Solving)
satisfy
points out that a similar argument applies provided that
there are an even number of students, the total score is 0 ≤ ΣA0 − ΣB0 < ΣA − ΣB.
even, and the achieved scores form a block of consecu-
tive integers. In fact, this argument will apply at the same level of
Second solution. We first claim that for any integer m generality as in the remark following the first solution.
with 15 ≤ m ≤ 40, we can find five distinct elements To prove the claim, let a1 , . . . , an be the scores in A and
of the set {1, 2, . . . , 10} whose sum is m. Indeed, for let b1 , . . . , bn be the scores in B (in any order). Since
0 ≤ k ≤ 4 and 1 ≤ ` ≤ 6, we have ΣA − ΣB ≡ ΣS (mod 2) and the latter is even, we must
k
!
10
! have ΣA − ΣB ≥ 2. In particular, there must exist in-
+ (k + `) + = 34 − 5k + `, dices i, j ∈ {1, . . . , n} such that ai > b j . Consequently,
∑j ∑ j
if we sort the sequence a1 , . . . , an , b1 , . . . , bn into nonde-
j=1 j=k+7
creasing order, it must be the case that some term b j is
and for fixed k this takes all values from 35 − 5k to 40 − followed by some term ai . Moreover, since the achieved
5k inclusive; then as k ranges from 0 to 4, this takes all scores form a range of consecutive integers, we must in
values from 15 to 40 inclusive. fact have ai = b j + 1. Consequently, if we take a = ai ,
Now suppose that the scores are a1 , . . . , a2N , where we b = b j , we then have ΣA0 − Σ0 B = ΣA − ΣB − 2, which
order the scores so that ak = k for k ≤ 10 and the subse- proves the claim.
quence a11 , a12 , . . . , a2N is nondecreasing. For 1 ≤ k ≤ A5 First solution. Let an , bn , cn be the probabilities that
N − 4, define Sk = ∑k+N+4 j=k+10 a j . Note that for each k, players A, B, C, respectively, will win the game. We
Sk+1 − Sk = ak+N+5 − ak+10 and so 0 ≤ Sk+1 − Sk ≤ 10. compute these by induction on n, starting with the val-
Thus S1 , . . . , SN−4 is a nondecreasing sequence of inte- ues
gers where each term is at most 10 more than the previ-
ous one. On the other hand, we have a1 = 1, b1 = 0, c1 = 0.
2N
If player A draws card k, then the resulting game state
S1 + SN−4 = ∑ aj
j=11
is that of a deck of k − 1 cards with the players taking
turns in the order B,C, A, B, . . . . In this state, the proba-
10
= (7.4)(2N) − ∑ a j bilities that players A, B,C will win are ck−1 , ak−1 , bk−1
j=1
provided that we adopt the convention that
= (7.4)(2N) − 55, a0 = 0, b0 = 0, c0 = 1.
 4

We thus have by the residue of m mod 3). We use the convention that
p0 (0) = 1, p0 (m) = 0 for m > 0. Define the generating
1 n 1 n 1 n function Pn (X) = ∑nm=0 pn (m)xm . We will establish that
an = ∑ ck−1 , bn = ∑ ak−1 , cn = ∑ bk−1 .
n k=1 n k=1 n k=1
X(X + 1) · · · (X + n − 1)
Put Pn (X) =
n!
xn = an − bn , yn = bn − cn , zn = cn − an ; (which may be guessed by computing pn (m) for small
n by hand). There are several ways to do this; for in-
we then have stance, this follows from the recursion
n 1 1 (n − 1)
xn+1 = xn + zn , Pn (X) = XPn−1 (X) + Pn−1 (X).
n+1 n+1 n n
n 1
yn+1 = yn + xn , (In this recursion, the first term corresponds to condi-
n+1 n+1
n 1 tional probabilities given that the first card drawn is
zn+1 = zn + yn . n, and the second term corresponds to the remaining
n+1 n+1
cases.)
Note that if an+1 = bn+1 = cn+1 = 0, then Let ω be a primitive cube root of 1. With notation as in
the first solution, we have
xn = −nzn = n2 yn = −n3 xn = n4 zn
Pn (ω) = an + bn ω + cn ω;
and so xn = zn = 0, or in other words an = bn = cn . By
induction on n, we deduce that an , bn , cn cannot all be combining this with the explicit formula for Pn (X) and
equal. That is, the quantities xn , yn , zn add up to zero the observation that
and at most one
p of them vanishes; consequently, the √
quantity rn = xn2 + y2n + z2n is always positive and the 3
arg(w + n) = arctan
quantities 2n − 1
xn yn zn recovers the geometric description of an , bn , cn given in
xn0 = , y0n = , z0n =
rn rn rn the first solution (as well as the remark following the
first solution).
form the coordinates of a point Pn on a fixed circle C in
R3 . Third solution. For this argument, we use the auxiliary
quantities
Let Pn0 be the point (zn , xn , yn ) obtained from Pn by a
clockwise rotation of angle 2π 3 . The point Pn+1 then lies 1 1 1
on the ray through the origin passing through the point a0n = an − , b0n = bn − , c0n = cn − ;
3 3 3
dividing the chord from Pn to Pn0 in the ratio 1 : n. The
(clockwise) arc from Pn to Pn+1 therefore has a measure these satisfy the relations
of
1 n 0 1 n 0 1 n 0
√ √ a0n = ∑ ck−1 , b0n = ∑ ak−1 , c0n = ∑ bk−1
3 3 n k=1 n k=1 n k=1
arctan = + O(n−3 );
2n − 1 2n − 1
as well as
these measures form a null sequence whose sum di-
verges. It follows that any arc of C contains infinitely 1
a0n+1 = a0n + (c0 − a0n )
many of√
the Pn√
; taking a suitably short arc around the n+1 n
point ( 2 , 0, − 22 ), we deduce that for infinitely many
2 1
b0n+1 = b0n + (a0 − b0n )
n, A has the highest winning probability, and similarly n+1 n
for B and C. 1
c0n+1 = c0n + (b0 − c0n ).
Remark: From the previous analysis, we also deduce n+1 n
that We now show that ∑∞ 0
n=1 an cannot diverge to +∞ (and
√ 0
likewise for ∑n=1 bn and ∑∞
∞ 0
rn+1 n2 − n + 1 3 n=1 cn by similar reasoning).
= = 1− + O(n−2 ), Suppose the contrary; then there exists some ε > 0
rn n+1 2(n + 1)
and some n0 > 0 such that ∑nk=1 a0k ≥ ε for all n ≥ n0 .
For n > n0 , we have b0n ≥ ε; this in turn implies that
from which it follows that rn ∼ cn−3/2 for some c > 0. 0
∑∞ n=1 bn diverges to +∞. Continuing around the cir-
Second solution. (by Noam Elkies) In this approach, cle, we deduce that for n sufficiently large, all three of
we instead compute the probability pn (m) that the game a0n , b0n , c0n are positive; but this contradicts the identity
ends after exactly m turns (the winner being determined
 5

a0n + b0n + c0n = 0. We thus conclude that ∑∞ 0

n=1 an does Consequently, for n0 the smallest value for which xn0 6=
0
not diverge to +∞; in particular, lim infn→∞ an ≤ 0. xn , we must have xn0 = 2. By this and two similar argu-
By the same token, we may see that ∑∞ 0 ments, we deduce that
n=1 an cannot con-
verge to a positive limit L (and likewise for ∑∞ 0
n=1 bn and 1 ⇒ 5, 2 ⇒ 6, 3 ⇒ 4.
∞
c0 by similar reasoning). Namely, this would im-
∑n=1 n
ply that b0n ≥ L/2 for n sufficiently large, contradicting Suppose that xn = 4. By the earlier discussion, we must
the previous argument. have a0n0 < 0 for some n0 > n, and so we cannot have
0
By similar reasoning, ∑∞ n=1 an cannot diverge to −∞ or xn0 = 4 for all n0 > n. On the other hand, as long as
converge to a negative limit L (and likewise for ∑∞ 0
n=1 bn xn = 4, we have
∞ 0
and ∑n=1 cn by similar reasoning).
1
We next establish that there are infinitely many n for b0n+1 − a0n+1 = b0n − a0n + (2a0n − b0n − c0n )
n+1
which a0n > 0 (and likewise for b0n and c0n by similar
n−1 0 1
reasoning). Suppose to the contrary that for n suffi- = (b − a0n ) + (b0 − c0n ) > 0
ciently large, we have a0n ≤ 0. By the previous argu- n+1 n n+1 n
ments, the sum ∑∞ 0 1
n=1 an cannot diverge to ∞ or con- c0n+1 − a0n+1 = c0n − a0n + (a0 + b0n − 2c0n )
verge to a nonzero limit; it must therefore converge n+1 n
to 0. In particular, for n sufficiently large, we have n−1 0 1
= (c − a0n ) + (b0 − a0n ) > 0.
b0n = ∑nk=1 a0k−1 ≥ 0. Iterating the construction, we see n+1 n n+1 n
that for n sufficiently large, we must have c0n ≤ 0, a0n ≥ 0,
b0n ≤ 0, and c0n ≥ 0. As a result, for n sufficiently large Consequently, for n0 the smallest value for which xn0 6=
we must have a0n = b0n = c0n = 0; but we may rule this xn , we must have xn0 = 1. By this and two similar argu-
out as in the original solution. ments, we deduce that
By similar reasoning, we may deduce that there are in- 4 ⇒ 1, 5 ⇒ 2, 6 ⇒ 3.
finitely many n for which a0n < 0 (and likewise for b0n
and c0n by similar reasoning). We now continue us- Combining, we obtain
ing a suggestion of Jon Atkins. Define the values of
the sequence xn according to the relative comparison of 1⇒5⇒2⇒6⇒3⇒4⇒1
a0n , b0n , c0n (using the fact that these cannot all be equal):
and hence the desired result.
xn = 1 : a0n ≤ b0n < c0n
A6 The number of such colorings is 220 310 =
xn = 2 : b0n ≤ c0n < a0n 61917364224.
xn = 3 : c0n ≤ a0n < b0n First solution: Identify the three colors red, white, and
xn = 4 : a0n < c0n ≤ b0n blue with (in some order) the elements of the field F3 of
xn = 5 : b0n < a0n ≤ c0n three elements (i.e., the ring of integers mod 3). The set
of colorings may then be identified with the F3 -vector
xn = 6 : c0n < b0n ≤ a0n . space FE3 generated by the set E of edges. Let F be
the set of faces, and let FF3 be the F3 -vector space on
We consider these values as states and say that there is
the basis F; we may then define a linear transformation
a transition from state i to state j, and write i ⇒ j, if for
T : FE3 → FF3 taking a coloring to the vector whose com-
every n ≥ 2 with xn = i there exists n0 > n with xn0 = j.
ponent corresponding to a given face equals the sum of
(In all cases when we use this notation, it will in fact be
the three edges of that face. The colorings we wish to
the case that the first value of n0 > n for which xn0 6= i
count are the ones whose images under T consist of vec-
satisfes xn0 = j, but this is not logically necessary for
tors with no zero components.
our final conclusion.)
We now show that T is surjective. (There are many pos-
Suppose that xn = 1. By the earlier discussion, we must
sible approaches to this step; for instance, see the fol-
have a0n0 > 0 for some n0 > n, and so we cannot have
lowing remark.) Let Γ be the dual graph of the icosa-
xn0 = 1 for all n0 > n. On the other hand, as long as
hedron, that is, Γ has vertex set F and two elements of
xn = 1, we have
F are adjacent in Γ if they share an edge in the icosa-
1 hedron. The graph Γ admits a hamiltonian path, that is,
c0n+1 − b0n+1 = c0n − b0n + (2b0n − a0n − c0n ) there exists an ordering f1 , . . . , f20 of the faces such that
n+1
n−1 0 1 any two consecutive faces are adjacent in Γ. For exam-
= (cn − b0n ) + (c0 − a0n ) > 0 ple, such an ordering can be constructed with f1 , . . . , f5
n+1 n+1 n being the five faces sharing a vertex of the icosahedron
1
c0n+1 − a0n+1 = c0n − a0n + (a0 + b0n − 2c0n ) and f16 , . . . , f20 being the five faces sharing the antipo-
n+1 n dal vertex.
n−1 0 1
= (c − a0n ) + (b0 − a0n ) > 0. For i = 1, . . . , 19, let ei be the common edge of fi and
n+1 n n+1 n fi+1 ; these are obviously all distinct. By prescribing
 6

components for e1 , . . . , e19 in turn and setting the oth- the plane by a factor of λ with center P carries L1 to
ers to zero, we can construct an element of FE3 whose another line parallel to L1 and hence not parallel to L2 .
image under T matches any given vector of FF3 in the Let A2 be the unique intersection of L2 with the image
components of f1 , . . . , f19 . The vectors in FF3 obtained of L1 , and let A1 be the point on L1 whose image under
in this way thus form a 19-dimensional subspace; this −→ −→
the dilation is A2 ; then PA2 = λ PA1 .
subspace may also be described as the vectors for which
In the other direction, suppose that L1 and L2 are par-
the components of f1 , . . . , f19 have the same sum as the
allel. Let P be any point in the region between L1 and
components of f2 , . . . , f20 .
L2 and take λ = 1. Then for any point A1 on L1 and
By performing a mirror reflection, we can construct a −→ −→
any point A2 on L2 , the vectors PA1 and PA2 have com-
second hamiltonian path g1 , . . . , g20 with the property ponents perpendicular to L1 pointing in opposite direc-
that tions; in particular, the two vectors cannot be equal.
g1 = f1 , g2 = f5 , g3 = f4 , g4 = f3 , g5 = f2 . Reinterpretation: (by Karl Mahlburg) In terms of vec-
tors, we may find vectors ~v1 ,~v2 and scalars c1 , c2 such
Repeating the previous construction, we obtain a differ- that Li = {~x ∈ R2 : ~vi ·~x = ci }. The condition in the
ent 19-dimensional subspace of FF3 which is contained problem amounts to finding a vector ~w and a scalar t
in the image of T . This implies that T is surjective, as such that P + ~w ∈ L1 , P + λ w ∈ L2 ; this comes down to
asserted earlier. solving the linear system
Since T is a surjective homomorphism from a 30-
~v1 · (P + ~w) = c1
dimensional vector space to a 20-dimensional vector
space, it has a 10-dimensional kernel. Each of the 220 ~v2 · (P + λ ~w) = c2
elements of FF3 with no zero components is then the im-
which is nondegenerate and solvable for all λ if and
age of exactly 310 colorings of the desired form, yield-
only if ~v1 ,~v2 are linearly independent.
ing the result.
Remark: There are many ways to check that T is sur- B2 We prove that the smallest value of a is 16.
jective. One of the simplest is the following (from Art Note that the expression for N can be rewritten as
of Problem Solving, user Ravi12346): form a vector in k(2a + k − 1)/2, so that 2N = k(2a + k − 1). In this
FE with components 2, 1, 2, 1, 2 at the five edges around expression, k > 1 by requirement; k < 2a + k − 1 be-
some vertex and all other components 0. This maps to cause a > 1; and obviously k and 2a + k − 1 have oppo-
a vector in FF with only a single nonzero component; site parity. Conversely, for any factorization 2N = mn
by symmetry, every standard basis vector of FF arises with 1 < m < n and m, n of opposite parity, we obtain
in this way. an expression of N in the desired form by taking k = m,
Second solution: (from Bill Huang, via Art of Problem a = (n + 1 − m)/2.
Solving user superpi83) Let v and w be two antipodal We now note that 2017 is prime. (On the exam, solvers
vertices of the icosahedron. Let Sv (resp. Sw ) be the set would have had to verify this by hand. Since 2017 <
of five edges incident to v (resp. w). Let Tv (resp. Tw ) be 452 , this can be done by trial division by the primes up
the set of five edges of the pentagon formed by the op- to 43.) For 2N = 2017(2a + 2016) not to have another
posite endpoints of the five edges in Sv (resp. Sw ). Let U expression of the specified form, it must be the case that
be the set of the ten remaining edges of the icosahedron. 2a + 2016 has no odd divisor greater than 1; that is,
Consider any one of the 310 possible colorings of U. 2a + 2016 must be a power of 2. This first occurs for
The edges of Tv ∪ U form the boundaries of five faces 2a + 2016 = 2048, yielding the claimed result.
with no edges in common; thus each edge of Tv can be Reinterpretation: (by Karl Mahlburg) To avoid N hav-
colored in one of two ways consistent with the given ing another representation, for k = 2, . . . , 2016, we must
condition, and similarly for Tw . That is, there are 310 210 have
possible colorings of Tv ∪ Tw ∪ U consistent with the (
given condition. k/2 k ≡ 0 (mod 2)
N 6≡
To complete the count, it suffices to check that there are 0 k ≡ 1 (mod 2).
exactly 25 ways to color Sv consistent with any given
coloring of Tv . Using the linear-algebraic interpretation Consequently, N 6≡ 0 (mod p) for any odd prime p <
from the first solution, this follows by observing that 2017 and N ≡ 0 (mod 1024). Since N must be divisible
(by the previous remark) the map from FS3v to the F3 - by 2017, this again yields the claimed value of a.
vector space on the faces incident to v is surjective, and
hence an isomorphism for dimensional reasons. A di- B3 Suppose by way of contradiction that f (1/2) is rational.
Then ∑∞ −i is the binary expansion of a rational
rect combinatorial proof is also possible. i=0 ci 2
number, and hence must be eventually periodic; that is,
B1 Recall that L1 and L2 intersect if and only if they are not there exist some integers m, n such that ci = cm+i for all
parallel. In one direction, suppose that L1 and L2 inter-
sect. Then for any P and λ , the dilation (homothety) of
 7

i ≥ n. We may then write Finally, we have

n−1 ∞ ∞
xn m−1
f (x) = ∑ ci xi + ∑ cn+i xi .
1 − xm i=0
S= ∑ a2k + ∑ a2k+1 + (log 2)2
i=0 k=1 k=0
∞
Evaluating at x = 2/3, we may equate f (2/3) = 3/2 = ∑ ak + (log 2)2 = (log 2)2 .
with k=1

n−1 m−1
1 2n 3m Second solution. We start with the following observa-
ci 2i 3n−i−1 + n+m−1 m cn+i 2i 3m−1−i ;
3n−1 ∑ i=0 3 (3 − 2m ) ∑ i=0
tion: for any positive integer n,

d −s 
since all terms on the right-hand side have odd denomi- n  = −(log n)n−s .
nator, the same must be true of the sum, a contradiction. ds s=1
Remark: Greg Marks asks whether the assumption that (Throughout, we view s as a real parameter, but see the
f (2/3) = 3/2 further ensures that f (1/2) is transcen- remark below.) For s > 0, consider the absolutely con-
dental. We do not know of any existing results that vergent series
would imply this. However, the following result fol-
lows from a theorem of T. Tanaka (Algebraic indepen- ∞
dence of the values of power series generated by linear L(s) = ∑ (3(4k +2)−s −(4k +3)−s −(4k +4)−s −(4k +5)−s );
k=0
recurrences, Acta Arith. 74 (1996), 177–190), building
upon work of Mahler. Let {an }∞ n=0 be a linear recurrent in the same range we have
sequence of positive integers with characteristic poly-
nomial P. Suppose that P(0), P(1), P(−1) 6= 0 and that ∞ 
log(4k + 2) log(4k + 3)
no two distinct roots of P have ratio which is a root L0 (s) = ∑ 3 −
an k=0 (4k + 2)s (4k + 3)s
of unity. Then for f (x) = ∑∞ n=0 x , the values f (1/2) 
and f (2/3) are algebraically independent over Q. (Note log(4k + 4) log(4k + 5)
+ − ,
that for f as in the original problem, the condition on (4k + 4)s (4k + 5)s
ratios of roots of P fails.)
so we may interchange the summation with taking the
B4 We prove that the sum equals (log 2)2 ; as usual, we limit at s = 1 to equate the original sum with −L0 (1).
write log x for the natural logarithm of x instead of ln x. To make further progress, we introduce the Riemann
Note that of the two given expressions of the original zeta function ζ (s) = ∑∞ −s
n=1 n , which converges abso-
sum, the first is absolutely convergent (the summands lutely for s > 1. In that region, we may freely rearrange
decay as log(x)/x2 ) but the second one is not; we must sums to write
thus be slightly careful when rearranging terms.
First solution. Define ak = logk k − log(k+1) L(s) + ζ (s) = 1 + 4(2−s + 6−s + 10−s + · · · )
k+1 . The infinite
n
sum ∑∞ a
k=1 k converges to 0 since ∑ k=1 ak telescopes to = 1 + 22−s (1 + 3−s + 5−s + · · · )
log(n+1)
− n+1 and this converges to 0 as n → ∞. Note that = 1 + 22−s (ζ (s) − 2−s − 4−s − · · · )
ak > 0 for k ≥ 3 since logx x is a decreasing function of x = 1 + 22−s ζ (s) − 22−2s ζ (s).
for x > e, and so the convergence of ∑∞ k=1 ak is absolute.
Write S for the desired sum. Then since 3a4k+2 + In other words, for s > 1, we have
2a4k+3 + a4k+4 = (a4k+2 + a4k+4 ) + 2(a4k+2 + a4k+3 ),
we have L(s) = 1 + ζ (s)(−1 + 22−s − 22−2s ).
∞ s
Now recall that ζ (s) − s−1 extends to a C∞ function for
S= ∑ (3a4k+2 + 2a4k+3 + a4k+4 ) s > 0, e.g., by applying Abel summation to obtain
k=0
∞ ∞ s s
= ∑ a2k + ∑ 2(a4k+2 + a4k+3 ), ζ (s) − = ∑ n(n−s − (n + 1)−s ) −
k=1 k=0
s − 1 n=1 s−1
∞ Z n+1
s
where we are allowed to rearrange the terms in the =s∑n x−s−1 dx −
infinite sum since ∑ ak converges absolutely. Now n=1 n s−1
Z ∞
2(a4k+2 + a4k+3 ) = log(4k+2)
2k+1 − log(4k+4)
2k+2 = a2k+1 + = −s (x − bxc)x−s−1 dx.
1 1 1
(log 2)( 2k+1 − 2k+2 ), and summing over k gives
Also by writing 22−s = 2 exp((1 − s) log 2 and 22−2s =
∞ ∞
(−1)k+1 ∞
exp(2(1 − s) log 2), we may use the exponential series
∑ 2(a4k+2 + a4k+3 ) = ∑ a2k+1 + (log 2) ∑ k to compute the Taylor expansion of
k=0 k=0 k=1
∞
= ∑ a2k+1 + (log 2)2 . f (s) =
−1 + 22−s − 22−2s
k=0 s−1
 8

at s = 1; we get equalizers intersecting T on the sides AB and AC, we

want to solve g(x) = p/2 where g(x) = x + bc/(2x)
f (s) = −(log 2)2 (s − 1)2 + O((s − 1)3 ). and x ∈ [c/2, c]; since g is convex and g(c/2) < p/2,
g(c) < p/2, there are no such solutions.
Consequently, if we rewrite the previous expression for
L(s) as It follows that if T has exactly two equalizers, then it
must have exactly one equalizer intersecting T on the
−1 + 22−s − 22−2s sides AC and BC. Here we want to solve h(x) = p/2
L(s) = 1 + (s − 1)ζ (s) · , where h(x) = x + ab/(2x) and x ∈ [a/2, a]. Now h is
s−1
convex and h(a/2) > p/2, h(a) > p/2; thus h(x) = p/2
then we have an equality of C∞ functions for s > 1, and has exactly one solution x ∈ [a/2, a] if and only if there
hence (by continuity) an equality of Taylor series about is x0 ∈ [a/2, a] with h0 (x0 ) =p0 and h(x0 ) = p/2. The
s = 1. That is, first condition implies x0 = ab/2, and then p the sec-
ond condition gives 8ab = p2 . Note that ab/2 is in
L(s) = 1 − (log 2)2 (s − 1) + O((s − 1)2 ), [a/2, a] since a > b and a < b + c < 2b.
which yields the desired result. We conclude that T has two equalizers if and only if
8ab = (a + b + c)2 . Note that (a, b, c) = (9, 8, 7) works.
Remark: We claim that this is the only possibility when a > b > c
The use of series ∑∞ −s as functions of a real pa-
n=1 cn n are integers and a ≤ 9. Indeed, the only integers (a, b)
rameter s dates back to Euler, who observed that the di- such that 2 ≤ b < a ≤ 9 and 8ab is a perfect square
vergence of ζ (s) as s → 1 gives a proof of the infinitude are (a, b) = (4, 2), (6, 3), (8, 4), (9, 2), and (9, 8), and
of primes distinct from Euclid’s approach, and Dirich- the first four possibilities do not produce triangles since
let, who upgraded this idea to prove his theorem on the they do not satisfy a < 2b. This gives the claimed result.
distribution of primes across arithmetic progressions. It 2016!
was Riemann who introduced the idea of viewing these B6 First solution. The desired count is 1953! − 63! · 2016,
series as functions of a complex parameter, thus mak- which we compute using the principle of inclusion-
ing it possible to use the tools of complex analysis (e.g., exclusion. As in A2, we use the fact that 2017 is prime;
the residue theorem) and leading to the original proof this means that we can do linear algebra over the field
of the prime number theorem by Hadamard and de la F2017 . In particular, every nonzero homogeneous linear
Vallée Poussin. equation in n variables over F2017 has exactly 2017n−1
solutions.
In the language of complex analysis, one may handle
the convergence issues in the second solution in a dif- For π a partition of {0, . . . , 63}, let |π| denote the num-
ferent way: use the preceding calculation to establish ber of distinct parts of π, Let π0 denote the partition of
the equality {0, . . . , 63} into 64 singleton parts. Let π1 denote the
partition of {0, . . . , 63} into one 64-element part. For
L(s) = 1 + ζ (s)(−1 + 22−s − 22−2s ) π, σ two partitions of {0, . . . , 63}, write π|σ if π is a re-
finement of σ (that is, every part in σ is a union of parts
for Real(s) > 1, then observe that both sides are holo- in π). By induction on |π|, we may construct a collec-
morphic for Real(s) > 0 and so the equality extends to tion of integers µπ , one for each π, with the properties
that larger domain. that
(
B5 The desired integers are (a, b, c) = (9, 8, 7). 1 σ = π0
Suppose we have a triangle T = 4ABC with BC = a, ∑ µπ = 0 σ 6= π0 .
π|σ
CA = b, AB = c and a > b > c. Say that a line is
an area equalizer if it divides T into two regions of Define the sequence c0 , . . . , c63 by setting c0 = 1 and
equal area. A line intersecting T must intersect two of ci = i for i > 1. Let Nπ be the number of ordered 64-
the three sides of T . First consider a line intersecting tuples (x0 , . . . , x63 ) of elements of F2017 such that xi = x j
the segments AB at X and BC at Y , and let BX = x, whenever i and j belong to the same part and ∑63 i=0 ci xi is
BY = y. This line is an area equalizer if and only if
divisible by 2017. Then Nπ equals 2017|π|−1 unless for
xy sin B = 2 area(4XBY ) = area(4ABC) = 21 ac sin B,
each part S of π, the sum ∑i∈S ci vanishes; in that case,
that is, 2xy = ac. Since x ≤ c and y ≤ a, the area equal-
izers correspond to values of x, y with xy = ac/2 and Nπ instead equals 2017|π| . Since c0 , . . . , c63 are positive
x ∈ [c/2, c]. Such an area equalizer is also an equal- integers which sum to 1 + 63·64 2 = 2017, the second out-
izer if and only if p/2 = x + y, where p = a + b + c come only occurs for π = π1 . By inclusion-exclusion,
is the perimeter of T . If we write f (x) = x + ac/(2x), the desired count may be written as
then we want to solve f (x) = p/2 for x ∈ [c/2, c]. Now
note that f is convex, f (c/2) = a + c/2 > p/2, and ∑ µπ Nπ = 2016 · µπ1 + ∑ µπ 2017|π|−1 .
π π
f (c) = a/2 + c < p/2; it follows that there is exactly
one solution to f (x) = p/2 in [c/2, c]. Similarly, for
 9

Similarly, the number of ordered 64-tuples with no re- some i. Since then x j 6= 0 for all j 6= i, we may apply the in-
peated elements may be written as duction hypothesis to see that there are f (k − 1, p) solutions
  that arise this way for a given i (and these do not overlap).
2017 This proves the claim.
64! = ∑ µπ 2017|π| .
64 π
To compute f (k, p) explicitly, it is convenient to work
2016! with the auxiliary function
The desired quantity may thus be written as 1953! +
2016µπ1 .
p f (k, p)
It remains to compute µπ1 . We adopt an approach sug- g(k, p) = ;
gested by David Savitt: apply inclusion-exclusion to k!
count distinct 64-tuples in an arbitrary set A. As above, by the lemma, this satisfies g(1, p) = 0 and
this yields  
p
|A|(|A| − 1) · · · (|A| − 63) = ∑ µπ |A||π| . g(k, p) = − g(k − 1, p)
k
π    
p−1 p−1
= + − g(k − 1, p) (k > 1).
Viewing both sides as polynomials in |A| and comparing k k−1
coefficients in degree 1 yields µπ = −63! and thus the By induction on k, we deduce that
claimed answer.
    
Second solution. (from Art of Problem Solving, user p−1 k−1 p−1
g(k, p) − = (−1) g(1, p) −
ABCDE) We first prove an auxiliary result. k 1
Lemma. Fix a prime p and define the function f (k) on posi- = (−1)k (p − 1)
tive integers by the conditions p−1

and hence g(k, p) = k + (−1)k (p − 1).
f (1, p) = 0 We now set p = 2017 and count the tuples in ques-
(p − 1)! tion. Define c0 , . . . , c63 as in the first solution. Since
f (k, p) = − k f (k − 1, p) (k > 1). c0 + · · · + c63 = p, the translation action of F p preserves
(p − k)!
the set of tuples; we may thus assume without loss of
Then for any positive integers a1 , . . . , ak with a1 +· · ·+ak < p, generality that x0 = 0 and multiply the count by p at the
there are exactly f (p) solutions to the equation a1 x1 + · · · + end. That is, the desired answer is
ak xk = 0 with x1 , . . . , xk ∈ F p nonzero and pairwise distinct.
2017 f (63, 2017) = 63!g(63, 2017)
Proof. We check the claim by induction, with the base case
  
2016
k = 1 being obvious. For the induction step, assume the claim = 63! − 2016
63
for k − 1. Let S be the set of k-tuples of distinct elements of
p!
F p ; it consists of (p−k)! elements. This set is stable under the as claimed.
action of i ∈ F p by translation:

(x1 , . . . , xk ) 7→ (x1 + i, . . . , xk + i).

Since 0 < a1 · · · + ak < p, exactly one element of each orbit

gives a solution of a1 x1 + · · · + ak xk = 0. Each of these solu-
tions contributes to f (k) except for those in which xi = 0 for