Solutions to the 73rd William Lowell Putnam Mathematical Competition

Saturday, December 1, 2012

Kiran Kedlaya and Lenny Ng

A1 Without loss of generality, assume d1 ≤ d2 ≤ · · · ≤ d12 . Proof. We may assume #S ≥ 3, as otherwise S is trivially an
2 < d 2 + d 2 for some i ≤ 10, then d , d
If di+2 i i+1 i i+1 , di+2 arithmetic progression. Let a1 , a2 be the smallest and second-
are the side lengths of an acute triangle, since in this smallest elements of S, respectively, and put d = a2 − a1 . Let
case di2 < di+1
2 +d 2 and d 2 < d 2 +d 2 as well. Thus
i+2 i+1 i i+2 m be the smallest positive integer such that a1 + md ∈ / S. Sup-
we may assume di+2 2 ≥ d2 + d2 pose that there exists an integer n contained in S but not in
i i+1 for all i. But then
by induction, di2 ≥ Fi d12 for all i, where Fi is the i-th {a1 , a1 + d, . . . , a1 + (m − 1)d}, and choose the least such n.
Fibonacci number (with F1 = F2 = 1): i = 1 is clear, i = By the hypothesis applied with (a, b, c) = (a1 , a2 , n), we see
2 follows from d2 ≥ d1 , and the induction step follows that n − d also has the property, a contradiction.
from the assumed inequality. Setting i = 12 now gives
2 ≥ 144d 2 , contradicting d > 1 and d < 12.
d12 1 12
We now return to the original problem. By dividing
1
B, q, r by gcd(q, r) if necessary, we may reduce to the
Remark. A materially equivalent problem appeared on case where gcd(q, r) = 1. We may assume #S ≥ 3, as
the 2012 USA Mathematical Olympiad and USA Junior otherwise S is trivially an arithmetic progression. Let
Mathematical Olympiad. a1 , a2 , a3 be any three distinct elements of S, labeled
A2 Write d for a ∗ c = b ∗ c ∈ S. For some e ∈ S, d ∗ e = a, so that a1 < a2 < a3 , and write rai = bi + mi q with
and thus for f = c ∗ e, a ∗ f = a ∗ c ∗ e = d ∗ e = a and bi , mi ∈ Z and bi ∈ B. Note that b1 , b2 , b3 must also be
b ∗ f = b ∗ c ∗ e = d ∗ e = a. Let g ∈ S satisfy g ∗ a = b; distinct, so the differences b2 − b1 , b3 − b1 , b3 − b2 are
then b = g ∗ a = g ∗ (a ∗ f ) = (g ∗ a) ∗ f = b ∗ f = a, as all nonzero; consequently, two of them have the same
desired. sign. If bi − b j and bk − bl have the same sign, then we
must have
Remark. With slightly more work, one can show that
S forms an abelian group with the operation ∗. (ai − a j )(bk − bl ) = (bi − b j )(ak − al )
√
A3 We will prove that f (x) = 1 − x2 for √ all x ∈ [−1, 1]. because both sides are of the same sign, of absolute
Define g : (−1, 1)
√ → R by g(x) = f (x)/ 1 − x2 . Plug- value less than q, and congruent to each other modulo
2
ging f (x) = g(x) 1 − x into equation (i) and simplify- q. In other words, the points (a1 , b1 ), (a2 , b2 ), (a3 , b3 )
ing yields in R2 are collinear. It follows that a4 = a1 + a3 − a2
 2  also belongs to S (by taking b4 = b1 + b3 − b2 ), so S
x satisfies the conditions of the lemma. It is therefore an
g(x) = g (1)
2 − x2 arithmetic progression.
for all x ∈ (−1, 1). Now fix x ∈ (−1, 1) and define a Reinterpretations. One can also interpret this argu-
a2n ment geometrically using cross products (suggested by
sequence {an }∞
n=1 by a1 = x and an+1 = 2−a2n
. Then Noam Elkies), or directly in terms of congruences (sug-
|2
an ∈ (−1, 1) and thus |an+1 | ≤ |an for all n. It follows gested by Karl Mahlburg).
that {|an |} is a decreasing sequence with |an | ≤ |x|n for Remark. The problem phrasing is somewhat confus-
all n, and so limn→∞ an = 0. Since g(an ) = g(x) for ing: to say that “S is the intersection of [the interval]
all n by (1) and g is continuous at 0, we conclude that A with an arithmetic progression” is the same thing as
g(x) = g(0) = f (0) = 1. This holds for all x ∈ (−1, 1) saying that “S is the empty set or an arithmetic progres-
and thus for x = ±1 as well by continuity. The result sion” unless it is implied that arithmetic progressions
follows. are necessarily infinite. Under that interpretation, how-
Remark. As pointed out by Noam Elkies, condition ever, the problem becomes false; for instance, for
(iii) is unnecessary. However, one can use it to derive
a slightly different solution by running the recursion in q = 5, r = 1, A = [1, 3], B = [0, 2],
the opposite direction.
we have
A4 We begin with an easy lemma.
T = {· · · , 0, 1, 2, 5, 6, 7, . . . }, S = {1, 2}.
Lemma. Let S be a finite set of integers with the following
property: for all a, b, c ∈ S with a ≤ b ≤ c, we also have a + A5 The pairs (p, n) with the specified property are those
c − b ∈ S. Then S is an arithmetic progression. pairs with n = 1, together with the single pair (2, 2).
We first check that these do work. For n = 1, it is clear
that taking v = (1) and M = (0) has the desired effect.
 2
 

1 1 by hypothesis, so h(x, y + 1/c) = h(x, y). By the fundamental
For (p, n) = (2, 2), we take v = 0 1 and M =
0 1 theorem of calculus, we may differentiate both sides of this
and then observe that identity with respect to y to deduce that g(x, y + 1/c) = g(x, y).
        Differentiating this new identity with respect to x yields the
(k) 0 0 1 1 desired equality.
G (0) = , , , , k = 0, 1, 2, 3.
0 1 0 1 √
Lemma 2. Let C be a circle whose diameter d is at least 2,
We next check that no other pairs work, keeping in mind and let AB and A0 B0 be two diameters of C. Then f (A) +
that the desired condition means that G acts on Fnp as a f (B) = f (A0 ) + f (B0 ).
cyclic permutation. Assume by way of contradiction
that (p, n) has the desired property but does not appear Proof. By continuity, it suffices to check the case where α =
in our list. In particular, we have n ≥ 2. arcsin d22 is an irrational multiple of 2π. Let β be the ra-
Let I be the n × n identity matrix over F p . Decom- dian measure of the counterclockwise arc from A to A0 . By
pose Fnp as a direct sum of two subspaces V,W such that Lemma 1, the claim holds when β = α. By induction, the
M − I is nilpotent on V and invertible on W . Suppose claim also holds when β ≡ nα (mod 2π) for any positive in-
that W 6= 0. Split v as v1 +v2 with v1 ∈ V , v2 ∈ W . Since teger n. Since α is an irrational multiple of 2π, the positive
M − I is invertible on W , there exists a unique w ∈ W multiples of α fill out a dense subset of the real numbers mod-
such that (M − I)w = −v2 . Then G(k) (w) − w ∈ V for ulo 2π, so by continuity the claim holds for all β .
all nonnegative integers k. Let k be the least positive in-
teger such that G(k) (w) = w; then k is at most the cardi- Lemma 3. Let R be a rectangular region of arbitrary (posi-
nality of V , which is strictly less than pn because W 6= 0. tive) area with corners A, B,C, D labeled in counterclockwise
This gives a contradiction and thus forces W = 0. order. Then f (A) + f (C) = f (B) + f (D).
In other words, the matrix N = M − I is nilpotent; con- Proof. Let EF be a segment such that AEFD and√BEFC
sequently, N n = 0. For any positive integer k, we have are rectangles whose diagonals have length at least 2. By
Lemma 2,
G(k) (0) = v + Mv + · · · + M k−1 v
k−1 n−1  
j i f (A) + f (F) = f (D) + f (E)
=∑∑ Nv
f (C) + f (E) = f (B) + f (F),
j=0 i=0 i
n−1  
k yielding the claim.
=∑ N i v.
i=0 i + 1
Lemma 4. The restriction of f to any straight line is constant.
If n ≥ 2 and (p, n) 6= (2, 2), then pn−1
> n and so
Proof. We may choose coordinates so that the line in question
Gk (0) = 0 for k = pn−1 (because all of the binomial
is the x-axis. Define the function g(y) by
coefficients are divisible by p). This contradiction com-
pletes the proof. g(y) = f (0, y) − f (0, 0).
A6 First solution. Yes, f (x, y) must be identically 0. We
By Lemma 3, for all x ∈ R,
proceed using a series of lemmas.
Lemma 1. Let R be a rectangular region of area 1 with f (x, y) = f (x, 0) + g(y).
corners A, B,C, D labeled in counterclockwise order. Then
For any c > 0, by the original hypothesis we have
f (A) + f (C) = f (B) + f (D).
Z x+c Z y+1/c
Proof. We may choose coordinates so that for some c > 0, 0= f (u, v) du dv
x y
Z x+c Z y+1/c
A = (0, 0), B = (c, 0),C = (c, 1/c), D = (0, 1/c).
= ( f (u, 0) + g(v)) du dv
x y
Define the functions x+c Z y+1/c
1
Z
Z x+c = f (u, 0) du + c g(v) dv.
g(x, y) = f (t, y) dt c x y

Zx y
In particular, the function F(x) = xx+c f (u, 0) du is constant.
R
h(x, y) = g(x, u) du. By the fundamental theorem of calculus, we may differentiate
0
to conclude that f (x + c, 0) = f (x, 0) for all x ∈ R. Since c
For any x, y ∈ R, was also arbitrary, we deduce the claim.
Z x+c Z y+1/c
h(x, y + 1/c) − h(x, y) = f (t, u) dt du = 0
x y
 3

To complete the proof, note that since any two points in P, as otherwise the vanishing of the zero over any rect-
R2 are joined by a straight line, Lemma 4 implies that f angle would force f to vanish identically. By continuity,
is constant. This constant equals the integral of f over there must exist an open disc U such that f (P) > 0 for
any rectangular region of area 1, and hence must be 0 all P ∈ U. Choose a rectangle R of area 1 with sides
as desired. parallel to the coordinate axes with one horizontal edge
Second solution (by Eric Larson, communicated by contained in U. Since the integral of f over R is zero,
Noam Elkies). In this solution, we fix coordinates and there must exist a point Q ∈ R such that f (Q) < 0. Take
assume only that the double integral vanishes on each P to be the vertical projection of Q onto the edge of R
rectangular region of area 1 with sides parallel to the contained in U.
coordinate axes, and still conclude that f must be iden- By translating coordinates, we may assume that P =
tically 0. (0, 0) and Q = (0, a) for some a > 0. For s sufficiently
small, f is positive on the square of side length 2s cen-
Lemma. Let R be a rectangular region of area 1 with sides tered at P, which we call S, and negative on the square
parallel to the coordinate axes. Then the averages of f over of side length 2s centered at Q, which we call S0 . Since
any two adjacent sides of R are equal. the ratio 2s/(1−4s2 ) tends to 0 as s does, we can choose
s so that 2s/(1 − 4s2 ) = a/n for some positive integer n.
Proof. Without loss of generality, we may take R to have cor-
ners (0, 0), (c, 0), (c, 1/c), (0, 1/c) and consider the two sides For i ∈ Z, let Ai be the rectangle
adjacent to (c, 1/c). Differentiate the equality
1 − 4s2

Z x+c Z y+1/c (x, y) : s ≤ x ≤ s + ,
2s
0= f (u, v) du dv 
x y 2s 2s
−s + i ≤ y ≤ s+i
1 − 4s2 1 − 4s2
with respect to c to obtain
Z y+1/c Z x+c and let Bi be the rectangle
1
0= f (x + c, v) dv − 2 f (u, y + 1/c) du.
y c x

1 − 4s2
(x, y) : s ≤ x ≤ s + ,
Rearranging yields 2s

2s 2s
Z y+1/c
1
Z x+c s+i ≤ y ≤ −s + (i + 1) .
c f (x + c, v) dv = f (u, y + 1/c) du, 1 − 4s2 1 − 4s2
y c x
Then for all i ∈ Z,
which asserts the desired result.
S ∪ A0 , An ∪ S0 , Ai ∪ Bi , Bi ∪ Ai+1
Returning to the original problem, given any c > 0, we
can tile the plane with rectangles of area 1 whose ver- are all rectangles of area 1 with sides parallel to the co-
tices lie in the lattice {(mc, n/c) : m, n ∈ Z}. By re- ordinate axes, so the integral over f over each of these
peated application of the lemma, we deduce that for any rectangles is zero. Since the integral over S is positive,
positive integer n, the integral over A0 must be negative; by induction, for
all i ∈ Z the integral over Ai is negative and the integral
over Bi is positive. But this forces the integral over S0
Z c Z (n+1)c
f (u, 0) du = f (u, 0) du. to be positive whereas f is negative everywhere on S0 , a
0 nc
contradiction.
Replacing c with c/n, we obtain
B1 Each of the following functions belongs to S for the rea-
Z c/n Z c+1/n
sons indicated.
f (u, 0) du = f (u, 0) du.
0 c
f (x), g(x) given
Fixing c and taking the limit as n → ∞ yields f (0, 0) = ln(x + 1) (i)
f (c, 0). By similar reasoning, f is constant on any hor- ln( f (x) + 1), ln(g(x) + 1) (ii) plus two previous lines
izontal line and on any vertical line, and as in the first ln( f (x) + 1) + ln(g(x) + 1) (ii)
solution the constant value is forced to equal 0. ex − 1 (i)
( f (x) + 1)(g(x) + 1) − 1 (ii) plus two previous lines
Third solution. (by Sergei Artamoshin) We retain the
f (x)g(x) + f (x) + g(x) previous line
weaker hypothesis of the second solution. Assume by
f (x) + g(x) (ii) plus first line
way of contradiction that f is not identically zero.
f (x)g(x) (iii) plus two previous lines
We first exhibit a vertical segment PQ with f (P) > 0
and f (Q) < 0. It cannot be the case that f (P) ≤ 0 for all B2 Fix a face F of the polyhedron with area A. Suppose F
is completely covered by balls of radii r1 , . . . , rn whose
 4

volumes sum to V . Then on one hand, Second solution. Put bn = ean , so that bn+1 = bn e1/bn .
In terms of the bn , the problem is to prove that bn /n has
n
4 a limit as n → ∞; we will show that the limit is in fact
∑ 3 πri3 = V. equal to 1.
i=1
Expanding e1/bn as a Taylor series in 1/bn , we have
On the other hand, the intersection of a ball of radius r
with the plane containing F is a disc of radius at most r, bn+1 = bn + 1 + Rn
which covers a piece of F of area at most πr2 ; therefore
n
where 0 ≤ Rn ≤ c/bn for some absolute constant c > 0.
πri2 ≥ A. By writing
∑
i=1
n−1
By writing n as ∑ni=1 1 and applying Hölder’s inequality, bn = n + e + ∑ Ri ,
i=0
we obtain
2/3 !3 we see first that bn ≥ n + e. We then see that
n 
4 16 3
nV 2 ≥ ∑ πri3 ≥ A . bn
i=1 3 9π 0≤ −1
n
Consequently, any value of c(P) less than 16 3
9π A works. e n−1 Ri
≤ +∑
n i=0 n
B3 The answer is yes. We first note that for any collection
of m days with 1 ≤ m ≤ 2n − 1, there are at least m e n−1 c
distinct teams that won a game on at least one of those ≤ +∑
n i=0 nbi
days. If not, then any of the teams that lost games on
all of those days must in particular have lost to m other e n−1 c
≤ +∑
teams, a contradiction. n i=0 n(i + e)
If we now construct a bipartite graph whose vertices are e c log n
≤ + .
the 2n teams and the 2n − 1 days, with an edge linking n n
a day to a team if that team won their game on that day,
then any collection of m days is connected to a total of at It follows that bn /n → 1 as n → ∞.
least m teams. It follows from Hall’s Marriage Theorem Remark. This problem is an example of the general
that one can match the 2n − 1 days with 2n − 1 distinct principle that one can often predict the asymptotic be-
teams that won on their respective days, as desired. havior of a recursive sequence by studying solutions of
a sufficiently similar-looking differential equation. In
B4 First solution. We will show that the answer is yes. this case, we start with the equation an+1 − an = e−an ,
First note that for all x > −1, ex ≥ 1 + x and thus then replace an with a function y(x) and replace the dif-
ference an+1 − an with the derivative y0 (x) to obtain the
x ≥ log(1 + x). (2)
differential equation y0 = e−y , which indeed has the so-
We next claim that an > log(n + 1) (and in particular lution y = log x.
that an − log n > 0) for all n, by induction on n. For B5 Define the function
n = 0 this follows from a0 = 1. Now suppose that
an > log(n + 1), and define f (x) = x + e−x , which is f (x) = sup{x log g1 (s) + log g2 (s)}.
an increasing function in x > 0; then s∈R

an+1 = f (an ) > f (log(n + 1)) As a function of x, f is the supremum of a collection of

affine functions, so it is convex. The function e f (x) is
= log(n + 1) + 1/(n + 1) ≥ log(n + 2),
then also convex, as may be checked directly from the
where the last inequality is (2) with x = 1/(n + 1). This definition: for x1 , x2 ∈ R and t ∈ [0, 1], by the weighted
completes the induction step. AM-GM inequality
It follows that an − log n is a decreasing function in n: te f (x1 ) + (1 − t)e f (x2 ) ≥ et f (x1 )+(1−t) f (x2 )
we have
≥ e f (tx1 +(1−t)x2 ) .
(an+1 − log(n + 1)) − (an − log n)
For each t ∈ R, draw a supporting line to the graph of
= e−an + log(n/(n + 1)) e f (x) at x = t; it has the form y = xh1 (t) + h2 (t) for some
< 1/(n + 1) + log(n/(n + 1)) ≤ 0, h1 (t), h2 (t) ∈ R. For all x, we then have
where the final inequality is (2) with x = −1/(n + 1). sup{g1 (s)x g2 (s)} ≥ xh1 (t) + h2 (t)
Thus {an − log n}∞n=0 is a decreasing sequence of posi-
s∈R
tive numbers, and so it has a limit as n → ∞.
 5

with equality for x = t. This proves the desired equality p. Since p ≡ 2 (mod 3), by the law of quadratic reci-
(including the fact that the maximum on the right side procity we have χ(−3) = +1, so χ(c/3) = χ(−c). It
is achieved). p−1 (p+χ(−c))/2
thus remains to evaluate the product ∏c=1 c
Remark. This problem demonstrates an example of du- modulo p.
ality for convex functions. If p ≡ 3 (mod 4), this is easy: each factor is a quadratic
residue (this is clear if c is a residue, and otherwise
B6 First solution. Since fixed points do not affect the sig-
χ(−c) = +1 so p + χ(−c) is divisible by 4) and −1
nature of a permutation, we may ignore the residue class
is not, so we must get +1 modulo p.
of 0 and consider π as a permutation on the nonzero
residue classes modulo p. These form a cyclic group of If p ≡ 1 (mod 4), we must do more work: we choose a
order p − 1, so the signature of π is also the signature primitive root g modulo p and rewrite the product as
of multiplication by 3 as a permutation σ of the residue
p−2
classes modulo p − 1. If we identify these classes with i

the integers 0, . . . , p − 2, then the signature equals the ∏ gi(p+(−1) )/2 .

i=0
parity of the number of inversions: these are the pairs
(i, j) with 0 ≤ i < j ≤ p − 2 for which σ (i) > σ ( j). We The sum of the exponents, split into sums over i odd
may write and i even, gives
 
3i (p−3)/2 
(2 j + 1)(p − 1)

σ (i) = 3i − (p − 1) j(p + 1) +
p−1 ∑ 2
j=0
from which we see that (i, j) cannot be an inversion un-
3j 3i which simplifies to
less b p−1 c > b p−1 c. In particular, we only obtain in-
versions when i < 2(p − 1)/3. (p − 3)(p − 1)(p + 1) (p − 1)3 p2 − 1
 
p−1
+ = −p .
If i < (p − 1)/3, the elements j of {0, . . . , p − 2} for 8 8 2 2
which (i, j) is an inversion correspond to the elements
of {0, . . . , 3i} which are not multiples of 3, which are Hence the product we are trying to evaluate is congruent
2i in number. This contributes a total of 0 + 2 + · · · + to g(p−1)/2 ≡ −1 modulo p.
2(p − 2)/3 = (p − 2)(p + 1)/9 inversions. Third solution (by Mark van Hoeij). We compute the
If (p − 1)/3 < i < 2(p − 1)/3, the elements j of parity of π as the parity of the number of cycles of even
{0, . . . , p−2} for which (i, j) is an inversion correspond length in the cycle decomposition of π. For x a nonzero
to the elements of {0, . . . , 3i − p + 1} congruent to 1 residue class modulo p of multiplicative order d, the
modulo 3, which are (3i − p + 2)/3 = i − (p − 2)/3 in elements of the orbit of x under π also have order d
number. This contributes a total of 1+· · ·+(p−2)/3 = (because d divides p − 1 and hence is coprime to 3).
(p − 2)(p + 1)/18 inversions. Since the group of nonzero residue classes modulo p
is cyclic of order p − 1, the elements of order d fall
Summing up, the total number of inversions is (p − into ϕ(d)/ f (d) orbits under π, where ϕ is the Euler
2)(p + 1)/6, which is even if and only if p ≡ 3 phi function and f (d) is the multiplicative order of 3
(mod 4). This proves the claim. modulo d. The parity of π is then the parity of the sum
Second solution (by Noam Elkies). Recall that the sign of ϕ(d)/ f (d) over all divisors d of p−1 for which f (d)
of π (which is +1 if π is even and −1 if π is odd) can is even.
be computed as If d is odd, then ϕ(d)/ f (d) = ϕ(2d)/ f (2d), so the
π(x) − π(y) summands corresponding to d and 2d coincide. It thus
∏ suffices to consider those d divisible by 4. If p ≡ 3
0≤x<y<p x−y (mod 4), then there are no such summands, so the sum
is trivially even.
(because composing π with a transposition changes the
sign of the product). Reducing modulo p, we get a con- If p ≡ 1 (mod 4), then d = 4 contributes a summand
gruence with of ϕ(4)/ f (4) = 2/2 = 1. For each d which is a larger
multiple of 4, the group (Z/dZ)∗ is isomorphic to the
x3 − y3 product of Z/2Z with another group of even order,
∏ = ∏ (x2 + xy + y2 ). so the maximal power of 2 dividing f (d) is strictly
0≤x<y<p x − y 0≤x<y<p
smaller than the maximal power of 2 dividing d. Hence
ϕ(d)/ f (d) is even, and so the overall sum is odd.
It thus suffices to count the number of times each pos-
sible value of x2 + xy + y2 occurs. Each nonzero value Remark. Note that the second proof uses quadratic
c modulo p occurs p + 1 times as x2 + xy + y2 with 0 ≤ reciprocity, whereas the first and third proofs are sim-
x, y < p and hence (p + χ(c/3))/2 times with 0 ≤ x < ilar to several classical proofs of quadratic reciprocity.
y < p, where χ denotes the quadratic character modulo Abhinav Kumar notes that the problem itself is a spe-
cial case of the Duke-Hopkins quadratic reciprocity
 6

law for abelian groups (Quadratic reciprocity in a finite pdf).

group, Amer. Math. Monthly 112 (2005), 251–256; see
also http://math.uga.edu/~pete/morequadrec.