Solutions to the 71st William Lowell Putnam Mathematical Competition

Saturday, December 4, 2010

Kiran Kedlaya and Lenny Ng

A–1 The largest such k is b n+1 n

2 c = d 2 e. For n even, this value Since 10n ≥ n ≥ 2m ≥ m + 1, 10n is divisible by 2n and
n n
is achieved by the partition hence by 2m+1 , and similarly 1010 is divisible by 210
and hence by 2 m+1 . It follows that
{1, n}, {2, n − 1}, . . . ;
m
N ≡ 1 + 1 + (−1) + (−1) ≡ 0 (mod 102 + 1).
for n odd, it is achieved by the partition
n m
Since N ≥ 1010 > 10n + 1 ≥ 102 + 1, it follows that N
{n}, {1, n − 1}, {2, n − 2}, . . . . is composite.
One way to see that this is optimal is to note that the A–5 We start with three lemmas.
common sum can never be less than n, since n itself
belongs to one of the boxes. This implies that k ≤ (1 + Lemma 1. If x, y ∈ G are nonzero orthogonal vectors, then
· · · + n)/n = (n + 1)/2. Another argument is that if k > x ∗ x is parallel to y.
(n + 1)/2, then there would have to be two boxes with
one number each (by the pigeonhole principle), but such Proof. Put z = x × y 6= 0, so that x, y, and z = x ∗ y are nonzero
boxes could not have the same sum. and mutually orthogonal. Then w = x × z 6= 0, so w = x ∗ z is
nonzero and orthogonal to x and z. However, if (x∗x)×y 6= 0,
Remark. A much subtler question would be to find
then w = x∗(x∗y) = (x∗x)∗y = (x∗x)×y is also orthogonal
the smallest k (as a function of n) for which no such
to y, a contradiction.
arrangement exists.
Lemma 2. If x ∈ G is nonzero, and there exists y ∈ G nonzero
A–2 The only such functions are those of the form f (x) =
and orthogonal to x, then x ∗ x = 0.
cx + d for some real numbers c, d (for which the prop-
erty is obviously satisfied). To see this, suppose that f
Proof. Lemma 1 implies that x ∗ x is parallel to both y and
has the desired property. Then for any x ∈ R,
x × y, so it must be zero.
2 f 0 (x) = f (x + 2) − f (x) Lemma 3. If x, y ∈ G commute, then x × y = 0.
= ( f (x + 2) − f (x + 1)) + ( f (x + 1) − f (x))
= f 0 (x + 1) + f 0 (x). Proof. If x × y 6= 0, then y × x is nonzero and distinct from
x×y. Consequently, x∗y = x×y and y∗x = y×x 6= x∗y.
Consequently, f 0 (x + 1) = f 0 (x).
We proceed now to the proof. Assume by way of con-
Define the function g : R → R by g(x) = f (x + 1) −
tradiction that there exist a, b ∈ G with a × b 6= 0. Put
f (x), and put c = g(0), d = f (0). For all x ∈ R,
c = a × b = a ∗ b, so that a, b, c are nonzero and lin-
g0 (x) = f 0 (x + 1) − f 0 (x) = 0, so g(x) = c identically,
early independent. Let e be the identity element of G.
and f 0 (x) = f (x+1)− f (x) = g(x) = c, so f (x) = cx+d
Since e commutes with a, b, c, by Lemma 3 we have
identically as desired.
e×a = e×b = e×c = 0. Since a, b, c span R3 , e×x = 0
A–3 If a = b = 0, then the desired result holds trivially, so for all x ∈ R3 , so e = 0.
we assume that at least one of a, b is nonzero. Pick Since b, c, and b × c = b ∗ c are nonzero and mutually
any point (a0 , b0 ) ∈ R2 , and let L be the line given by orthogonal, Lemma 2 implies
the parametric equation L(t) = (a0 , b0 ) + (a, b)t for t ∈
R. By the chain rule and the given equation, we have b ∗ b = c ∗ c = (b ∗ c) ∗ (b ∗ c) = 0 = e.
d
dt (h ◦ L) = h ◦ L. If we write f = h ◦ L : R → R, then
f 0 (t) = f (t) for all t. It follows that f (t) = Cet for some Hence b ∗ c = c ∗ b, contradicting Lemma 3 because b ×
constant C. Since | f (t)| ≤ M for all t, we must have c 6= 0. The desired result follows.
C = 0. It follows that h(a0 , b0 ) = 0; since (a0 , b0 ) was
an arbitrary point, h is identically 0 over all of R2 . A–6 First solution. Note that the hypotheses on f imply
that f (x) > 0 for all x ∈ [0, +∞), so the integrand is a
A–4 Put continuous function of f and the integral makes sense.
10n n
Rewrite the integral as
N = 1010 + 1010 + 10n − 1. Z ∞ 
f (x + 1)
1− dx,
Write n = 2m k with m a nonnegative integer and k a 0 f (x)
positive odd integer. For any nonnegative integer j,
mj m
102 ≡ (−1) j (mod 102 + 1).
 2

and suppose by way of contradiction that it converges If there exist arbitrarily large values of y for which
to a finite limit L. For n ≥ 0, define the Lebesgue mea- f (y) > 2 f (y + 1), we deduce that the original integral is
surable set greater than any multiple of 1/7, and so diverges. Oth-
erwise, for x large we may argue that
f (x + n + 1)
In = {x ∈ [0, 1] : 1 − ≤ 1/2}.
f (x + n) f (x) − f (x + 1) 3 f (x)
> log
1
f (x) 5 f (x + 1)
Then L ≥ ∑∞ n=0 2 (1 − µ(In )), so the latter sum con-
verges. In particular, there exists a nonnegative integer as in the above solution, and again get divergence using
N for which ∑∞ n=N (1 − µ(In )) < 1; the intersection a telescoping sum.
∞ ∞ Second solution. (Communicated by Paul Allen.) Let
b > a be nonnegative integers. Then
[ \
I= In = [0, 1] − ([0, 1] − In )
n=N n=N
Z b b−1 Z 1
f (x) − f (x + 1) f (x + k) − f (x + k + 1)
then has positive Lebesgue measure. dx = ∑ dx
a f (x) k=a 0 f (x + k)
By Taylor’s theorem with remainder, for t ∈ [0, 1/2], Z 1 b−1
f (x + k) − f (x + k + 1)
= ∑ dx
t2 f (x + k)
 
1 0 k=a
− log(1 − t) ≤ t + sup
2 t∈[0,1/2] (1 − t)2 Z 1 b−1
f (x + k) − f (x + k + 1)
≥ ∑ dx
= t + 2t 2 ≤ 2t. 0 k=a f (x + a)
Z 1
f (x + a) − f (x + b)
For each nonnegative integer n ≥ N, we then have = dx.
0 f (x + a)
Z n 
f (x + 1)
L≥ 1− dx Now since f (x) → 0, given a, we can choose an in-
N f (x)
teger l(a) > a for which f (l(a)) < f (a + 1)/2; then
n−1 Z 1  
f (x + i + 1) f (x+a)− f (x+l(a)) (l(a))
≥ 1 − ff (a+1) > 1/2 for all x ∈ [0, 1].
=∑ 1− dx f (x+a)
i=N 0 f (x + i) Thus if we define a sequence of integers an by a0 = 0,
n−1 Z 
f (x + i + 1)
 an+1 = l(an ), then
≥∑ 1− dx
i=N I f (x + i) Z ∞
f (x) − f (x + 1) ∞ Z an+1
f (x) − f (x + 1)
dx = ∑ dx
1 n−1 f (x + i) f (x) f (x)
Z
0 n=0 an
≥ ∑ log dx
2 i=N I f (x + i + 1) ∞ Z 1
! > ∑ (1/2)dx,
n−1
1 f (x + i) n=0 0
Z
= ∑ log f (x + i + 1) dx
2 I i=N and the final sum clearly diverges.
1 f (x + N)
Z
= log dx. Third solution. (By Joshua Rosenberg, communicated
2 I f (x + n) by Catalin Zara.) If the original integral converges, then
on one hand the integrand ( f (x) − f (x + 1))/ f (x) = 1 −
For each x ∈ I, log f (x + N)/ f (x + n) is a strictly f (x + 1)/ f (x) cannot tend to 1 as x → ∞. On the other
increasing unbounded function of n. ByR the mono- hand, for any a ≥ 0,
tone convergence theorem, the integral I log( f (x +
N)/ f (x + n)) dx grows without bound as n → +∞, a f (a + 1)
contradiction. Thus the original integral diverges, as 0<
f (a)
desired. Z a+1
1
Remark. This solution is motivated by the commonly- < f (x) dx
f (a) a
used fact that an infinite product (1 + x1 )(1 + x2 ) · · ·
1
Z ∞
converges absolutely if and only if the sum x1 + x2 + · · · = ( f (x) − f (x + 1)) dx
converges absolutely. The additional measure-theoretic f (a) a
f (x) − f (x + 1)
Z ∞
argument at the beginning is needed because one cannot
≤ dx,
bound − log(1−t) by a fixed multiple of t uniformly for a f (x)
all t ∈ [0, 1).
and the last expression tends to 0 as a → ∞. Hence by
Greg Martin suggests a variant solution that avoids use
the squeeze theorem, f (a + 1)/ f (a) → 0 as a → ∞, a
of Lebesgue measure. √ Note first that if f (y) > 2 f (y +
contradiction.
1),
√ then either f (y) > 2 f (y + 1/2) or f (y + 1/2) >
2 f (y + 1), and in either case we deduce that
Z y+1/2  
f (x) − f (x + 1) 1 1 1
dx > 1− √ > .
y−1/2 f (x) 2 2 7
 3

B–1 First solution. No such sequence exists. If it did, then Remark. Manjul Bhargava points out it is easy to con-
the Cauchy-Schwartz inequality would imply struct sequences of complex numbers with the desired
property if we drop the condition of absolute conver-
8 = (a21 + a22 + · · · )(a41 + a42 + · · · ) gence. Here is an inductive construction (of which sev-
≥ (a31 + a32 + · · · )2 = 9, eral variants are possible). For n = 1, 2, . . . and z ∈ C,
define the finite sequence
contradiction.  
1 2πi j/n
Second solution. (Communicated by Catalin Zara.) sn,z = e : j = 0, . . . , n − 1 .
z
Suppose that such a sequence exists. If a2k ∈ [0, 1] for
all k, then a4k ≤ a2k for all k, and so This sequence has the property that for any positive in-
teger j, the sum of the j-th powers of the terms of sn,z
4 = a41 + a42 + · · · ≤ a21 + a22 + · · · = 2, equals 1/z j if j is divisible by n and 0 otherwise. More-
over, any partial sum of j-th powers is bounded in ab-
contradiction. There thus exists a positive integer k for solute value by n/|z| j .
which a2k ≥ 1. However, in this case, for m large, a2m
k >
The desired sequence will be constructed as follows.
2m and so a12m + a2m
2 + · · · 6
= 2m.
Suppose that we have a finite sequence which has the
Third solution. We generalize the second solution to correct sum of j-th powers for j = 1, . . . , m. (For in-
show that for any positive integer k, it is impossible for stance, for m = 1, we may start with the singleton
a sequence a1 , a2 , . . . of complex numbers to satisfy the sequence 1.) We may then extend it to a new se-
given conditions in case the series ak1 + ak2 + · · · con- quence which has the correct sum of j-th powers for
verges absolutely. This includes the original problem j = 1, . . . , m + 1, by appending k copies of sm+1,z for
by taking k = 2, in which case the series a21 + a22 + · · · suitable choices of a positive integer k and a complex
consists of nonnegative real numbers and so converges number z with |z| < m−2 . This last restriction ensures
absolutely if it converges at all. that the resulting infinite sequence a1 , a2 , . . . is such that
Since the sum ∑∞ k
i=1 |ai | converges by hypothesis, we
for each positive integer m, the series am m
1 + a2 + · · · is
can find a positive integer n such that ∑∞ k
i=n+1 |ai | < 1.
convergent (though not absolutely convergent). Its par-
For each positive integer d, we then have tial sums include a subsequence equal to the constant
value m, so the sum of the series must equal m as de-
n ∞ sired.
kd − ∑ akd
i ≤ ∑ |ai |kd < 1.
i=1 i=n+1 B–2 The smallest distance is 3, achieved by A = (0, 0), B =
(3, 0), C = (0, 4). To check this, it suffices to check
We thus cannot have |a1 |, . . . , |an | ≤ 1, or else the sum that AB cannot equal 1 or 2. (It cannot equal 0 because
∑ni=1 akd
i would be bounded in absolute value by n inde- if two of the points were to coincide, the three points
pendently of d. But if we put r = max{|a1 |, . . . , |an |} > would be collinear.)
1, we obtain another contradiction because for any ε >
The triangle inequality implies that |AC − BC| ≤ AB,
0,
with equality if and only if A, B,C are collinear. If AB =
n 1, we may assume without loss of generality that A =
lim sup(r − ε)−kd ∑ akdi > 0. (0, 0), B = (1, 0). To avoid collinearity, we must have
d→∞ i=1 AC = BC, but this forces C = (1/2, y) for some y ∈ R,
a contradiction. (One can also treat this case by scaling
For instance, this follows from applying the root test to by a factor of 2 to reduce to the case AB = 2, treated in
the rational function the next paragraph.)
!
n
1 ∞ n If AB = 2, then we may assume without loss of gener-
∑ 1 − ak z = ∑ ∑ ai zd ,
kd
ality that A = (0, 0), B = (2, 0). The triangle inequal-
i=1 i d=0 i=1 ity implies |AC − BC| ∈ {0, 1}. Also, for C = (x, y),
AC2 = x2 + y2 and BC2 = (2 − x)2 + y2 have the same
which has a pole within the circle |z| ≤ r−1/k . (An ele- parity; it follows that AC = BC. Hence c = (1, y) for
mentary proof is also possible.) some y ∈ R, so y2 and y2 + 1 = BC2 are consecutive
Fourth solution. (Communicated by Noam Elkies.) perfect squares. This can only happen for y = 0, but
Since ∑k a2k = 2, for each positive integer k we have then A, B,C are collinear, a contradiction again.
a2k ≤ 2 and so a4k ≤ 2a2k , with equality only for a2k ∈ Remark. Manjul Bhargava points out that more gener-
{0, 2}. Thus to have ∑k a4k = 4, there must be a single ally, a Heronian triangle (a triangle with integer sides
index k for which a2k = 2, and the other ak must all equal and rational area) cannot have a side of length 1 or 2
0. But then ∑k a2m m
k = 2 6= 2m for any positive integer
(and again it is enough to treat the case of length 2).
m > 2. The original problem follows from this because a tri-
angle whose vertices have integer coordinates has area
 4

equal to half an integer (by Pick’s formula or the ex- Then r(x) = a, s(x) = c for all x ∈ Z, and hence identi-
plicit formula for the area as a determinant). cally; consequently, p(x) = ax + b, q(x) = cx + d for all
x ∈ Z, and hence identically. For p and q of this form,
B–3 It is possible if and only if n ≥ 1005. Since
p(x)q(x + 1) − p(x + 1)q(x) = bc − ad,
2009 × 2010
1 + · · · + 2009 = = 2010 × 1004.5,
2 so we get a solution if and only if bc − ad = 1, as
for n ≤ 1004, we can start with an initial distribution claimed.
in which each box Bi starts with at most i − 1 balls (so Second solution. (Communicated by Catalin Zara.)
in particular B1 is empty). From such a distribution, no Again, note that p and q must be nonzero. Write
moves are possible, so we cannot reach the desired final
distribution. p(x) = p0 + p1 x + · · · + pm xm
Suppose now that n ≥ 1005. By the pigeonhole prin- q(x) = q0 + q1 x + · · · + qn xn
ciple, at any time, there exists at least one index i for
which the box Bi contains at least i balls. We will de- with pm , qn 6= 0, so that m = deg(p), n = deg(q). It
scribe any such index as being eligible. The following is enough to derive a contradiction assuming that
sequence of operations then has the desired effect. max{m, n} > 1, the remaining cases being treated as in
the first solution.
(a) Find the largest eligible index i. If i = 1, proceed Put R(x) = p(x)q(x+1)− p(x+1)q(x). Since m+n ≥ 2
to (b). Otherwise, move i balls from Bi to B1 , then by assumption, the coefficient of xm+n−1 in R(x) must
repeat (a). vanish. By easy algebra, this coefficient equals (m −
(b) At this point, only the index i = 1 can be eligi- n)pm qn , so we must have m = n > 1.
ble (so it must be). Find the largest index j for For k = 1, . . . , 2m − 2, the coefficient of xk in R(x) is
which B j is nonempty. If j = 1, proceed to (c).
   
Otherwise, move 1 ball from B1 to B j ; in case this j i
makes j eligible, move j balls from B j to B1 . Then ∑ − (pi q j − p j qi )
i+ j>k, j>i k−i k− j
repeat (b).
(c) At this point, all of the balls are in B1 . For i = and must vanish. For k = 2m − 2, the only summand is
2, . . . , 2010, move one ball from B1 to Bi n times. for (i, j) = (m − 1, m), so pm−1 qm = pm qm−1 .
After these operations, we have the desired distribution. Suppose now that h ≥ 1 and that pi q j = p j qi is known to
vanish whenever j > i ≥ h. (By the previous paragraph,
B–4 First solution. The pairs (p, q) satisfying the given we initially have this for h = m − 1.) Take k = m + h − 2
equation are those of the form p(x) = ax + b, q(x) = and note that the conditions i + j > h, j ≤ m force i ≥
cx + d for a, b, c, d ∈ R such that bc − ad = 1. We will h − 1. Using the hypothesis, we see that the only possi-
see later that these indeed give solutions. ble nonzero contribution to the coefficient of xk in R(x)
Suppose p and q satisfy the given equation; note that is from (i, j) = (h − 1, m). Hence ph−1 qm = pm qh−1 ;
neither p nor q can be identically zero. By subtracting since pm , qm 6= 0, this implies ph−1 q j = p j qh−1 when-
the equations ever j > h − 1.
By descending induction, we deduce that pi q j = p j qi
p(x)q(x + 1) − p(x + 1)q(x) = 1 whenever j > i ≥ 0. Consequently, p(x) and q(x) are
p(x − 1)q(x) − p(x)q(x − 1) = 1, scalar multiples of each other, forcing R(x) = 0, a con-
tradiction.
we obtain the equation Third solution. (Communicated by David Feldman.)
As in the second solution, we note that there are no so-
p(x)(q(x + 1) + q(x − 1)) = q(x)(p(x + 1) + p(x − 1)).
lutions where m = deg(p), n = deg(q) are distinct and
The original equation implies that p(x) and q(x) have m+n ≥ 2. Suppose p, q form a solution with m = n ≥ 2.
no common nonconstant factor, so p(x) divides p(x + The desired identity asserts that the matrix
1) + p(x − 1). Since each of p(x + 1) and p(x − 1) has  
p(x) p(x + 1)
the same degree and leading coefficient as p, we must
q(x) q(x + 1)
have
has determinant 1. This condition is preserved by re-
p(x + 1) + p(x − 1) = 2p(x).
placing q(x) with q(x) − t p(x) for any real number t. In
If we define the polynomials r(x) = p(x + 1) − p(x), particular, we can choose t so that deg(q(x) − t p(x)) <
s(x) = q(x + 1) − q(x), we have r(x + 1) = r(x), and m; we then obtain a contradiction.
similarly s(x + 1) = s(x). Put

a = r(0), b = p(0), c = s(0), d = q(0).

 5

B–5 First solution. The answer is no. Suppose otherwise. x = g( f (x)), and we may differentiate to find that
For the condition to make sense, f must be differen- 1 = g0 ( f (x)) f 0 (x) = g0 ( f (x)) f ( f (x)). It follows that
tiable. Since f is strictly increasing, we must have g0 (y) = 1/ f (y) for yR≥ y0 ; since g takes arbitrarily large
f 0 (x) ≥ 0 for all x. Also, the function f 0 (x) is strictly values, the integral y∞0 dy/ f (y) must diverge. One then
increasing: if y > x then f 0 (y) = f ( f (y)) > f ( f (x)) = gets a contradiction from any reasonable lower bound
f 0 (x). In particular, f 0 (y) > 0 for all y ∈ R. on f (y) for y large, e.g., the bound f (x) ≥ αx2 from the
For any x0 ≥ −1, if f (x0 ) = b and f 0 (x0 ) = a > 0, then second solution. (One can also start with a linear lower
f 0 (x) > a for x > x0 and thus f (x) ≥ a(x − x0 ) + b for bound f (x) ≥ β x, then use the integral expression for g
x ≥ x0 . Then either b < x0 or a = f 0 (x0 ) = f ( f (x0 )) = to deduce that g(x) ≤ γ log x, which in turn forces f (x)
f (b) ≥ a(b − x0 ) + b. In the latter case, b ≤ a(x0 + to grow exponentially.)
1)/(a + 1) ≤ x0 + 1. We conclude in either case that B–6 For any polynomial p(x), let [p(x)]A denote the n × n
f (x0 ) ≤ x0 + 1 for all x0 ≥ −1. matrix obtained by replacing each entry Ai j of A by
It must then be the case that f ( f (x)) = f 0 (x) ≤ 1 for p(Ai j ); thus A[k] = [xk ]A. Let P(x) = xn + an−1 xn−1 +
all x, since otherwise f (x) > x + 1 for large x. Now · · · + a0 denote the characteristic polynomial of A. By
by the above reasoning, if f (0) = b0 and f 0 (0) = the Cayley-Hamilton theorem,
a0 > 0, then f (x) > a0 x + b0 for x > 0. Thus for
x > max{0, −b0 /a0 }, we have f (x) > 0 and f ( f (x)) > 0 = A · P(A)
a0 x + b0 . But then f ( f (x)) > 1 for sufficiently large x,
= An+1 + an−1 An + · · · + a0 A
a contradiction.
Second solution. (Communicated by Catalin Zara.) = A[n+1] + an−1 A[n] + · · · + a0 A[1]
Suppose such a function exists. Since f is strictly = [xp(x)]A.
increasing and differentiable, so is f ◦ f = f 0 . In
particular, f is twice differentiable; also, f 00 (x) = Thus each entry of A is a root of the polynomial xp(x).
f 0 ( f (x)) f 0 (x) is the product of two strictly increasing Now suppose m ≥ n + 1. Then
nonnegative functions, so it is also strictly increasing
and nonnegative. In particular, we can choose α > 0 0 = [xm+1−n P(x)]A
and M ∈ R such that f 00 (x) > 4α for all x ≥ M. Then
= A[m+1] + an−1 A[m] + · · · + a0 A[m+1−n]
for all x ≥ M,

f (x) ≥ f (M) + f 0 (M)(x − M) + 2α(x − M)2 . since each entry of A is a root of xm+1−n P(x). On the
other hand,
In particular, for some M 0 > M, we have f (x) ≥ αx2 for
all x ≥ M 0 . 0 = Am+1−n · P(A)

Pick T > 0 so that αT 2 > M 0 . Then for x ≥ T , f (x) > = Am+1 + an−1 Am + · · · + a0 Am+1−n .
M 0 and so f 0 (x) = f ( f (x)) ≥ α f (x)2 . Now
Therefore if Ak = A[k] for m + 1 − n ≤ k ≤ m, then
1 1
Z 2T 0
f (t)
Z 2T Am+1 = A[m+1] . The desired result follows by induction
− = dt ≥ α dt; on m.
f (T ) f (2T ) T f (t)2 T
Remark. David Feldman points out that the result is
however, as T → ∞, the left side of this inequality tends best possible in the following sense: there exist ex-
to 0 while the right side tends to +∞, a contradiction. amples of n × n matrices A for which Ak = A[k] for
Third solution. (Communicated by Noam Elkies.) k = 1, . . . , n but An+1 6= A[n+1] .
Since f is strictly increasing, for some y0 , we can de-
fine the inverse function g(y) of f for y ≥ y0 . Then