Solutions to the 63rd William Lowell Putnam Mathematical Competition

Saturday, December 7, 2002

Kiran Kedlaya and Lenny Ng

A–1 By differentiating Pn (x)/(xk − 1)n+1 , we find that For i, j = 1, 2, 3, let Ai j denote the position in row i and
Pn+1 (x) = (xk − 1)Pn0 (x) − (n + 1)kxk−1 Pn (x); substi- column j. Without loss of generality, we may assume
tuting x = 1 yields Pn+1 (1) = −(n + 1)kPn (1). Since that Player 1’s first move is at A11 . Player 0 then plays
P0 (1) = 1, an easy induction gives Pn (1) = (−k)n n! for at A22 :
all n ≥ 0.  
1 ∗ ∗
Note: one can also argue by expanding in Taylor series ∗ 0 ∗
around 1. Namely, we have
∗ ∗ ∗
1 1 1
= = (x − 1)−1 + · · · , After Player 1’s second move, at least one of A23 and
xk − 1 k(x − 1) + · · · k
A32 remains vacant. Without loss of generality, assume
so A23 remains vacant; Player 0 then plays there.
dn 1 (−1)n n! After Player 1’s third move, Player 0 wins by playing
= at A21 if that position is unoccupied. So assume instead
dxn xk − 1 k(x − 1)−n−1
that Player 1 has played there. Thus of Player 1’s three
and moves so far, two are at A11 and A21 . Hence for i equal
to one of 1 or 3, and for j equal to one of 2 or 3, the
dn 1 following are both true:
Pn (x) = (xk − 1)n+1
dxn xk − 1

(−1)n n!
 (a) The 2 × 2 submatrix formed by rows 2 and i and
= (k(x − 1) + · · · )n+1 (x − 1)−n−1 + · · · by columns 2 and 3 contains two zeroes and two
k empty positions.
n
= (−k) n! + · · · .
(b) Column j contains one zero and two empty posi-
tions.
A–2 Draw a great circle through two of the points. There are
two closed hemispheres with this great circle as bound- Player 0 next plays at Ai j . To prevent a zero column,
ary, and each of the other three points lies in one of Player 1 must play in column j, upon which Player 0
them. By the pigeonhole principle, two of those three completes the 2 × 2 submatrix in (a) for the win.
points lie in the same hemisphere, and that hemisphere Note: one can also solve this problem directly by mak-
thus contains four of the five given points. ing a tree of possible play sequences. This tree can be
Note: by a similar argument, one can prove that among considerably collapsed using symmetries: the symme-
any n+3 points on an n-dimensional sphere, some n+2 try between rows and columns, the invariance of the
of them lie on a closed hemisphere. (One cannot get by outcome under reordering of rows or columns, and the
with only n + 2 points: put them at the vertices of a reg- fact that the scenario after a sequence of moves does
ular simplex.) Namely, any n of the points lie on a great not depend on the order of the moves (sometimes called
sphere, which forms the boundary of two hemispheres; “transposition invariance”).
of the remaining three points, some two lie in the same Note (due to Paul Cheng): one can reduce Determi-
hemisphere. nant Tic-Tac-Toe to a variant of ordinary tic-tac-toe.
A–3 Note that each of the sets {1}, {2}, . . . , {n} has the de- Namely, consider a tic-tac-toe grid labeled as follows:
sired property. Moreover, for each set S with inte- A11 A22 A33
ger average m that does not contain m, S ∪ {m} also
has average m, while for each set T of more than A23 A31 A12
one element with integer average m that contains m, A32 A13 A21
T \ {m} also has average m. Thus the subsets other than
{1}, {2}, . . . , {n} can be grouped in pairs, so Tn − n is Then each term in the expansion of the determinant oc-
even. curs in a row or column of the grid. Suppose Player
1 first plays in the top left. Player 0 wins by playing
A–4 (partly due to David Savitt) Player 0 wins with opti- first in the top row, and second in the left column. Then
mal play. In fact, we prove that Player 1 cannot prevent there are only one row and column left for Player 1 to
Player 0 from creating a row of all zeroes, a column of threaten, and Player 1 cannot already threaten both on
all zeroes, or a 2 × 2 submatrix of all zeroes. Each of the third move, so Player 0 has time to block both.
these forces the determinant of the matrix to be zero.
 2

A–5 It suffices to prove that for any relatively prime positive that for d ≥ 3,
integers r, s, there exists an integer n with an = r and
an+1 = s. We prove this by induction on r + s, the case 2d −1 Z 2d
1 1 1 dx
r +s = 2 following from the fact that a0 = a1 = 1. Given ∑ < d−1 − d +
n=2d−1
n 2 2 2d−1 x
r and s not both 1 with gcd(r, s) = 1, we must have r 6=
s. If r > s, then by the induction hypothesis we have 1
= + log 2
an = r − s and an+1 = s for some n; then a2n+2 = r and 2d
a2n+3 = s. If r < s, then we have an = r and an+1 = s −r 1
≤ + log 2 < 0.125 + 0.7 < 1.
for some n; then a2n+1 = r and a2n+2 = s. 8
Note: a related problem is as follows. Starting with the Put c = 81 + log 2 and L = 1 + 12 + 6(1−c)
1
. Then we can
sequence i
−1 1
prove that ∑2n=1 f (n) < L for all i ≥ 2 by induction on i.
0 1 The case i = 2 is clear. For the induction, note that by
, ,
1 0 (2),
repeat the following operation: insert between each pair 2i −1 i
a c a+c 1 1 1 1
b and d the pair b+d . Prove that each positive rational ∑ < 1+ + +c ∑
number eventually appears. n=1 f (n) 2 6 d=3 f (d)
Observe that by induction, if ba and dc are consecutive 1 1 1
< 1+ + +c
terms in the sequence, then bc−ad = 1. The same holds 2 6 6(1 − c)
for consecutive terms of the n-th Farey sequence, the 1 1
sequence of rational numbers in [0, 1] with denominator = 1+ + = L,
2 6(1 − c)
(in lowest terms) at most n.
1
A–6 The sum converges for b = 2 and diverges for b ≥ 3. We as desired. We conclude that ∑∞n=1 f (n) converges to a
first consider b ≥ 3. Suppose the sum converges; then limit less than or equal to L.
the fact that f (n) = n f (d) whenever bd−1 ≤ n ≤ bd − 1 Note: the above argument proves that the sum for b = 2
yields is at most L < 2.417. One can also obtain a lower
bound by the same technique, namely 1 + 21 + 6(1−c 1
0)
d
∞
1 ∞
1 b −1 1 with c0 = log 2. This bound exceeds 2.043. (By con-
∑ =∑ ∑ n. (1)
n=1 f (n) d=1 f (d) n=bd−1 trast, summing the first 100000 terms of the series only
yields a lower bound of 1.906.) Repeating the same ar-
However, by comparing the integral of 1/x with a Rie- guments with d ≥ 4 as the cutoff yields the upper bound
mann sum, we see that 2.185 and the lower bound 2.079.

bd −1 Z bd B–1 The probability is 1/99. In fact, we show by induction

1 dx
∑ n> on n that after n shots, the probability of having made
n=bd−1
bd−1 x any number of shots from 1 to n − 1 is equal to 1/(n −
= log(bd ) − log(bd−1 ) = log b, 1). This is evident for n = 2. Given the result for n,
we see that the probability of making i shots after n + 1
where log denotes the natural logarithm. Thus (1) yields attempts is
 
∞
1 ∞
1 i−1 1 i 1 (i − 1) + (n − i)
∑ > (log b) ∑ , + 1− =
f (n) n n−1 n n−1 n(n − 1)
n=1 n=1 f (n)
1
a contradiction since log b > 1 for b ≥ 3. Therefore the = ,
n
sum diverges.
as claimed.
For b = 2, we have a slightly different identity because
f (2) 6= 2 f (2). Instead, for any positive integer i, we B–2 (Note: the problem statement assumes that all polyhe-
have dra are connected and that no two edges share more than
one face, so we will do likewise. In particular, these are
2i −1 d
1 1 1 i
1 2 −1 1 true for all convex polyhedra.) We show that in fact
∑ = 1+ + + ∑ ∑ n. (2)
n=1 f (n) 2 6 d=3 f (d) n=2 d−1
the first player can win on the third move. Suppose the
polyhedron has a face A with at least four edges. If the
Again comparing an integral to a Riemann sum, we see first player plays there first, after the second player’s
first move there will be three consecutive faces B,C, D
adjacent to A which are all unoccupied. The first player
wins by playing in C; after the second player’s second
 3

move, at least one of B and D remains unoccupied, and than one of these guesses. If n ≡ 1 (mod 3), it will be
either is a winning move for the first player. guessed on an odd turn. If n ≡ 0 (mod 3), it will be
It remains to show that the polyhedron has a face with guessed on an even turn. If n ≡ 2 (mod 3), then n + 1
at least four edges. (Thanks to Russ Mann for suggest- will be guessed on an even turn, forcing a guess of n on
ing the following argument.) Suppose on the contrary the next turn. Thus the probability of success with this
that each face has only three edges. Starting with any strategy is 1335/2002 > 2/3.
face F1 with vertices v1 , v2 , v3 , let v4 be the other end- Note: for any positive integer m, this strategy wins
point of the third edge out of v1 . Then the faces ad- when the number is being guessed from [1, m] with
jacent to F1 must have vertices v1 , v2 , v4 ; v1 , v3 , v4 ; and probability m1 b 2m+1
3 c. We can prove that this is best
v2 , v3 , v4 . Thus v1 , v2 , v3 , v4 form a polyhedron by them- possible as follows. Let am denote m times the proba-
selves, contradicting the fact that the given polyhedron bility of winning when playing optimally. Also, let bm
is connected and has at least five vertices. (One can also denote m times the corresponding probability of win-
deduce this using Euler’s formula V − E + F = 2 − 2g, ning if the objective is to select the number in an even
where V, E, F are the numbers of vertices, edges and number of guesses instead. (For definiteness, extend the
faces, respectively, and g is the genus of the polyhe- definitions to incorporate a0 = 0 and b0 = 0.)
dron. For a convex polyhedron, g = 0 and you get the We first claim that am = 1 + max1≤k≤m {bk−1 + bm−k }
“usual” Euler’s formula.) and bm = max1≤k≤m {ak−1 + am−k } for m ≥ 1. To
Note: Walter Stromquist points out the following coun- establish the first recursive identity, suppose that our
terexample if one relaxes the assumption that a pair of first guess is some integer k. We automatically win if
faces may not share multiple edges. Take a tetrahedron n = k, with probability 1/m. If n < k, with probability
and remove a smaller tetrahedron from the center of an (k − 1)/m, then we wish to guess an integer in [1, k − 1]
edge; this creates two small triangular faces and turns in an even number of guesses; the probability of success
two of the original faces into hexagons. Then the sec- when playing optimally is bk−1 /(k − 1), by assumption.
ond player can draw by signing one of the hexagons, Similarly, if n < k, with probability (m − k)/m, then the
one of the large triangles, and one of the small trian- subsequent probability of winning is bm−k /(m − k). In
gles. (He does this by “mirroring”: wherever the first sum, the overall probability of winning if k is our first
player signs, the second player signs the other face of guess is (1 + bk−1 + bm−k )/m. For optimal strategy, we
the same type.) choose k such that this quantity is maximized. (Note
that this argument still holds if k = 1 or k = m, by our
B–3 The desired inequalities can be rewritten as definitions of a0 and b0 .) The first recursion follows,
   and the second recursion is established similarly.
1 1 1
1 − < exp 1 + n log 1 − < 1− . We now prove by induction that am = b(2m + 1)/3c and
n n 2n
bm = b2m/3c for m ≥ 0. The inductive step relies on the
By taking logarithms, we can rewrite the desired in- inequality bxc + byc ≤ bx + yc, with equality when one
equalities as of x, y is an integer. Now suppose that ai = b(2i + 1)/3c
    and bi = b2i/3c for i < m. Then
1 1
− log 1 − < −1 − n log 1 − 
2(k − 1)
 
2(m − k)

2n n 1 + bk−1 + bm−k = 1 + +

1
 3 3
< − log 1 − .  
2m
n ≤
3
Rewriting these in terms of the Taylor expansion of
− log(1 − x), we see that the desired result is also equiv- and similarly ak−1 + am−k ≤ b(2m + 1)/3c, with equal-
alent to ity in both cases attained, e.g., when k = 1. The induc-
∞ ∞ ∞
tive formula for am and bm follows.
1 1 1
∑ i2i ni < ∑ (i + 1)ni < ∑ ini , B–5 (due to Dan Bernstein) Put N = 2002!. Then for d =
i=1 i=1 i=1
1, . . . , 2002, the number N 2 written in base b = N/d − 1
which is evident because the inequalities hold term by has digits d 2 , 2d 2 , d 2 . (Note that these really are digits
term. because 2(2002)2 < (2002!)2 /2002 − 1.)
Note: David Savitt points out that the upper bound can Note: one can also produce an integer N which has base
be improved from 1/(ne) to 2/(3ne) with a slightly b digits 1, ∗, 1 for n different values of b, as follows.
more complicated argument. (In fact, for any c > 1/2, Choose c with 0 < c < 21/n . For m a large positive in-
one has an upper bound of c/(ne), but only for n above teger, put N = 1 + (m + 1) · · · (m + n)bcmcn−2 . For m
a certain bound depending on c.) sufficiently large, the bases

B–4 Use the following strategy: guess 1, 3, 4, 6, 7, 9, . . . until N −1

b= = (m + j)
the target number n is revealed to be equal to or lower (m + i)mn−2 ∏
j6=i
 4

for i = 1, . . . , n will have the properties that N ≡ 1 which is precisely the desired determinant.
(mod b) and b2 < N < 2b2 for m sufficiently large. Note: a simpler conceptual proof is as follows. (Every-
Note (due to Russ Mann): one can also give a “noncon- thing in this paragraph will be modulo p.) Note that
structive” argument. Let N be a large positive integer. for any integers a, b, c, the column vector [ax + by +
For b ∈ (N 2 , N 3 ), the number of 3-digit base-b palin- 2
cz, (ax + by + cz) p , (ax + by + cz) p ] is a linear com-
dromes in the range [b2 , N 6 − 1] is at least bination of the columns of the given matrix. Thus
 6 ax + by + cz divides the determinant. In particular, all
N − b2 N6

− 1 ≥ 2 − b − 2, of the factors of (3) divide the determinant; since both
b b (3) and the determinant have degree p2 + p + 1, they
agree up to a scalar multiple. Moreover, they have the
since there is a palindrome in each interval [kb, (k + 2
same coefficient of z p y p x (since this term only appears
1)b − 1] for k = b, . . . , b2 − 1. Thus the average num- in the expansion of (3) when you choose the first term
ber of bases for which a number in [1, N 6 − 1] is at least in each factor). Thus the determinant is congruent to
3 (3), as desired.
1 N −1 N 6
 
∑ − b − 2 ≥ log(N) − c Either argument can be used to generalize to a corre-
N 6 b=N 2 +1 b sponding n × n determinant, called a Moore determi-
nant; we leave the precise formulation to the reader.
for some constant c > 0. Take N so that the right side Note the similarity with the classical Vandermonde de-
exceeds 2002; then at least one number in [1, N 6 − 1] is
terminant: if A is the n × n matrix with Ai j = xij for
a base-b palindrome for at least 2002 values of b.
i, j = 0, . . . , n − 1, then
B–6 We prove that the determinant is congruent modulo p to
det(A) = ∏ (x j − xi ).
p−1 p−1 1≤i< j≤n
x ∏ (y + ix) ∏ (z + ix + jy). (3)
i=0 i, j=0

We first check that

p−1
∏ (y + ix) ≡ y p − x p−1 y (mod p). (4)
i=0

Since both sides are homogeneous as polynomials in x

and y, it suffices to check (4) for x = 1, as a congruence
between polynomials. Now note that the right side has
0, 1, . . . , p − 1 as roots modulo p, as does the left side.
Moreover, both sides have the same leading coefficient.
Since they both have degree only p, they must then co-
incide.
We thus have

p−1 p−1
x ∏ (y + ix) ∏ (z + ix + jy)
i=0 i, j=0
p−1
≡ x(y p − x p−1 y) ∏ ((z + jy) p − x p−1 (z + jy))
j=0
p−1
≡ (xy p − x p y) ∏ (z p − x p−1 z + jy p − jx p−1 y)
j=0

≡ (xy − x y)((z p − x p−1 z) p

p p

− (y p − x p−1 y) p−1 (z p − x p−1 z))

2 2 −p
≡ (xy p − x p y)(z p − x p zp)
− x(y p − x p−1 y) p (z p − x p−1 z)
2 2 2 −p+1 2
≡ xy p z p − x p yz p − x p y p z p + x p yz p
2 2 −p+1 2 2
− xy p z p + x p ypzp + xpyp z − xp ypz
2 2 2
≡ xy p z p + yz p x p + zx p y p
2 2 2
− xz p y p − yx p z p − zy p x p ,