Solutions to the 60th William Lowell Putnam Mathematical Competition

Saturday, December 4, 1999

Manjul Bhargava, Kiran Kedlaya, and Lenny Ng

A–1 Note that if r(x) and s(x) are any two functions, then an (an−1 + an+1 ). Then

max(r, s) = (r + s + |r − s|)/2. 2b2n+1 + b2n = 2an an+1 + 2an−1 an + a2n−1 + a2n

Therefore, if F(x) is the given function, we have = 2an an+1 + an−1 an+1 + a2n
= a2n+1 + a2n = b2n+2 ,
F(x) = max{−3x − 3, 0} − max{5x, 0} + 3x + 2
= (−3x − 3 + |3x + 3|)/2 and similarly 2b2n + b2n−1 = b2n+1 , so that {bn } satis-
fies the same recurrence as {an }. Since further b0 =
− (5x + |5x|)/2 + 3x + 2
1, b1 = 2 (where we use the recurrence for {an } to cal-
1 culate a−1 = 0), we deduce that bn = an for all n. In
= |(3x + 3)/2| − |5x/2| − x + ,
2 particular, a2n + a2n+1 = b2n+2 = a2n+2 .
so we may set f (x) = (3x + 3)/2, g(x) = 5x/2, and Second solution: Note that
h(x) = −x + 12 .
1
A–2 First solution: First factor p(x) = q(x)r(x), where q has 1 − 2x − x2
all real roots and r has all complex roots. Notice that √ √ !
1 2+1 2−1
each root of q has even multiplicity, otherwise p would = √ √ + √
have a sign change at that root. Thus q(x) has a square 2 2 1 − (1 + 2)x 1 − (1 − 2)x
root s(x).
and that
Now write r(x) = ∏kj=1 (x − a j )(x − a j ) (possible be-
cause r has roots in complex conjugate pairs). Write 1 ∞ √
∏kj=1 (x − a j ) = t(x) + iu(x) with t, x having real coeffi-
√ = ∑ (1 ± 2)n xn ,
1 + (1 ± 2)x n=0
cients. Then for x real,
so that
p(x) = q(x)r(x)
1 √ √ 
= s(x)2 (t(x) + iu(x))(t(x) + iu(x)) an = √ ( 2 + 1)n+1 − (1 − 2)n+1 .
2 2
= (s(x)t(x))2 + (s(x)u(x))2 .
A simple computation (omitted here) now shows that
(Alternatively, one can factor r(x) as a product of a2n + a2n+1 = a2n+2 .
quadratic polynomials with real coefficients, write each
as a sum of squares, then multiply together to get a sum Third
 solution (by Richard Stanley): Let A be the matrix

of many squares.) 0 1
. A simple induction argument shows that
1 2
Second solution: We proceed by induction on the de-
gree of p, with base case where p has degree 0. As in  
n+2 an an+1
the first solution, we may reduce to a smaller degree in A = .
an+1 an+2
case p has any real roots, so assume it has none. Then
p(x) > 0 for all real x, and since p(x) → ∞ for x → ±∞,
The desired result now follows from comparing the top
p has a minimum value c. Now p(x) − c has real roots,
left corner entries of the equality An+2 An+2 = A2n+4 .
so as above, we deduce that p(x)−c is√ a sum of squares.
Now add one more square, namely ( c)2 , to get p(x)
A–4 Denote the series by S, and let an = 3n /n. Note that
as a sum of squares.
∞ ∞
A–3 First solution: Computing the coefficient of xn+1 in the 1
S= ∑ ∑ am (am + an )
identity (1 − 2x − x2 ) ∑∞ m
m=0 am x = 1 yields the recur- m=1 n=1
rence an+1 = 2an + an−1 ; the sequence {an } is then ∞ ∞
1
characterized by this recurrence and the initial condi- = ∑ ∑ an (am + an ) ,
m=1 n=1
tions a0 = 1, a1 = 2.
Define the sequence {bn } by b2n = a2n−1 + a2n , b2n+1 = where the second equality follows by interchanging m
 2

and n. Thus and Q(x)(R(x) − kx) has more roots in [0, 1] than does P

1 1
 (and has the same value at 0). Repeating this argument
2S = ∑ ∑ + shows that 01 |P(x)| dx is greater than the correspond-
R
m n am (am + an ) a n (a m + an )
ing integral for some polynomial with all of its roots in
1 [0, 1].
= ∑∑
m n m an
a
Under this assumption, we have
!2
∞
n 1999
= ∑ n .
n=1 3 P(x) = c ∏ (x − ri )
i=1
But
∞
n 3 for some ri ∈ (0, 1]. Since
∑ 3n = 4
n=1 P(0) = −c ∏ ri = 1,
since, e.g., it’s f 0 (1), where
we have
xn
∞
3
f (x) = ∑ n = , |c| ≥ ∏ |ri−1 | ≥ 1.
n=0 3 3 − x
and we conclude that S = 9/32. Thus it suffices to prove that if Q(x) is a monic polyno-
A–5 First solution: (by Reid Barton) Let r1 , . . . , r1999 be the mial of degree 1999 with all of its roots in [0, 1], then
R1
roots of P. Draw a disc of radius ε around each ri , 0 |Q(x)|Rdx ≥ D for some constant D > 0. But the in-
where ε < 1/3998; this disc covers a subinterval of tegral of 01 ∏1999
i=1 |x − ri | dx is a continuous function for
[−1/2, 1/2] of length at most 2ε, and so of the 2000 (or ri ∈ [0, 1]. The product of all of these intervals is com-
fewer) uncovered intervals in [−1/2, 1/2], one, which pact, so the integral achieves a minimum value for some
we call I, has length at least δ = (1 − 3998ε)/2000 > 0. ri . This minimum is the desired D.
We will exhibit an explicit lower bound for the integral Third solution (by Abe Kunin): It suffices to prove the
of |P(x)|/P(0) over this interval, which will yield such stronger inequality
a bound for the entire integral.
Z 1
Note that
sup |P(x)| ≤ C |P(x)| dx
−1
|P(x)| 1999 |x − ri | x∈[−1,1]
= .
|P(0)| ∏i=1 |ri | holds for some C. But this follows immediately from
Also note that by construction, |x − ri | ≥ ε for each x ∈ the following standard fact: any two norms on a finite-
dimensional vector space (here the polynomials of de-
I. If |ri | ≤ 1, then we have |x−r i|
|ri | ≥ ε. If |ri | > 1, then gree at most 1999) are equivalent. (The proof of this
|x − ri | statement is also a compactness argument: C can be
= |1 − x/ri | ≥ 1 − |x/ri | ≥= 1/2 > ε. taken to be the maximum of the L1-norm divided by
|ri |
R the sup norm over the set of polynomials with L1-norm
We conclude that I |P(x)/P(0)| dx ≥ δ ε, independent 1.)
of P.
Note: combining the first two approaches gives a con-
Second solution: It will be a bit more convenient to structive solution with a constant that is better than that
assume P(0) = 1 (which we may achieve by rescal- given by the first solution, but is still far from optimal. I
ing unless P(0) = 0, in which case there is nothing to don’t know offhand whether it is even known what the
prove) and to prove that there exists D > 0 such that optimal constant and/or the polynomials achieving that
R1 R1
−1 |P(x)| dx ≥ D, or even such that 0 |P(x)| dx ≥ D. constant are.
We first reduce to the case where P has all of its roots
A–6 Rearranging the given equation yields the much more
in [0, 1]. If this is not the case, we can factor P(x) as
tractable equation
Q(x)R(x), where Q has all roots in the interval and R has
none. Then R is either always positive or always neg- an an−1 an−2
ative on [0, 1]; assume the former. Let k be the largest =6 −8 .
an−1 an−2 an−3
positive real number such that R(x) − kx ≥ 0 on [0, 1];
then Let bn = an /an−1 ; with the initial conditions b2 =
Z 1 Z 1 2, b3 = 12, one easily obtains bn = 2n−1 (2n−2 − 1), and
|P(x)| dx = |Q(x)R(x)| dx so
−1 −1
Z 1 n−1
> |Q(x)(R(x) − kx)| dx, an = 2n(n−1)/2 ∏ (2i − 1).
−1 i=1
 3

To see that n divides an , factor n as 2k m, with m odd. P. Thus we also have that no roots of P0 lie on the sides
Then note that k ≤ n ≤ n(n − 1)/2, and that there exists of the convex hull of P, unless they are also roots of P.
i ≤ m − 1 such that m divides 2i − 1, namely i = φ (m) From this we conclude that if r is a root of P which is
(Euler’s totient function: the number of integers in a vertex of the convex hull of the roots, and which is
{1, . . . , m} relatively prime to m). not also a root of P0 , then f has a single pole at r (as r
cannot be a root of P00 ). On the other hand, if r is a root
B–1 The answer is 1/3. Let G be the point obtained by re-
of P which is also a root of P0 , it is a multiple root, and
flecting C about the line AB. Since ∠ADC = π−θ2 , we
π−θ then f has a double pole at r.
find that ∠BDE = π − θ − ∠ADC = 2 = ∠ADC =
π − ∠BDC = π − ∠BDG, so that E, D, G are collinear. If P has roots not all equal, the convex hull of its roots
Hence has at least two vertices.

|BE| |BE| sin(θ /2) B–3 We first note that

|EF| = = = , xy
|BC| |BG| sin(3θ /2) ∑ xm yn = .
m,n>0 (1 − x)(1 − y)
where we have used the law of sines in 4BDG. But by
l’Hôpital’s Rule, Subtracting S from this gives two sums, one of which is

sin(θ /2) cos(θ /2) x2n+1 x3 y

lim = lim = 1/3. ∑ xm yn = ∑ yn =
θ →0 sin(3θ /2) θ →0 3 cos(3θ /2) m≥2n+1 n 1−x (1 − x)(1 − x2 y)

and the other of which sums to xy3 /[(1 − y)(1 − xy2 )].
B–2 First solution: Suppose that P does not have n dis-
Therefore
tinct roots; then it has a root of multiplicity at least 2,
which we may assume is x = 0 without loss of general- xy x3 y
ity. Let xk be the greatest power of x dividing P(x), so S(x, y) = −
(1 − x)(1 − y) (1 − x)(1 − x2 y)
that P(x) = xk R(x) with R(0) 6= 0; a simple computation
yields xy3
−
(1 − y)(1 − xy2 )
P00 (x) = (k2 − k)xk−2 R(x) + 2kxk−1 R0 (x) + xk R00 (x). xy(1 + x + y + xy − x2 y2 )
=
(1 − x2 y)(1 − xy2 )
Since R(0) 6= 0 and k ≥ 2, we conclude that the great-
est power of x dividing P00 (x) is xk−2 . But P(x) = and the desired limit is
Q(x)P00 (x), and so x2 divides Q(x). We deduce (since Q
is quadratic) that Q(x) is a constant C times x2 ; in fact, lim xy(1 + x + y + xy − x2 y2 ) = 3.
(x,y)→(1,1)
C = 1/(n(n − 1)) by inspection of the leading-degree
terms of P(x) and P00 (x). B–4 (based on work by Daniel Stronger) We make repeated
Now if P(x) = ∑nj=0 a j x j , then the relation P(x) = use of the following fact: if f is a differentiable function
Cx2 P00 (x) implies that a j = C j( j − 1)a j for all j; hence on all of R, limx→−∞ f (x) ≥ 0, and f 0 (x) > 0 for all
a j = 0 for j ≤ n − 1, and we conclude that P(x) = an xn , x ∈ R, then f (x) > 0 for all x ∈ R. (Proof: if f (y) < 0
which has all identical roots. for some x, then f (x) < f (y) for all x < y since f 0 > 0,
but then limx→−∞ f (x) ≤ f (y) < 0.)
Second solution (by Greg Kuperberg): Let f (x) =
P00 (x)/P(x) = 1/Q(x). By hypothesis, f has at most From the inequality f 000 (x) ≤ f (x) we obtain
two poles (counting multiplicity).
f 00 f 000 (x) ≤ f 00 (x) f (x) < f 00 (x) f (x) + f 0 (x)2
Recall that for any complex polynomial P, the roots of
P0 lie within the convex hull of P. To show this, it suf- since f 0 (x) is positive. Applying the fact to the differ-
fices to show that if the roots of P lie on one side of a ence between the right and left sides, we get
line, say on the positive side of the imaginary axis, then 1 00
P0 has no roots on the other side. That follows because ( f (x))2 < f (x) f 0 (x). (1)
if r1 , . . . , rn are the roots of P, 2

On the other hand, since f (x) and f 000 (x) are both posi-
P0 (z) n
1
=∑ tive for all x, we have
P(z) i=1 z − ri
2 f 0 (x) f 00 (x) < 2 f 0 (x) f 00 (x) + 2 f (x) f 000 (x).
and if z has negative real part, so does 1/(z − ri ) for
i = 1, . . . , n, so the sum is nonzero. Applying the fact to the difference between the sides
yields
The above argument also carries through if z lies on the
imaginary axis, provided that z is not equal to a root of f 0 (x)2 ≤ 2 f (x) f 00 (x). (2)
 4

Combining (1) and (2), we obtain Second solution (by Mohamed Omar): Set x = eiθ and
2 write
1 f 0 (x)2

1
< ( f 00 (x))2 1 T 1 T 1 T T
u
 
2 2 f (x) 2 A = u u+ v v = u v
< f (x) f 0 (x), 2 2 2 v

or ( f 0 (x))3 < 8 f (x)3 . We conclude f 0 (x) < 2 f (x), as for

desired.
u = x x2 · · · xn , v = x−1 x−2 · · · xn .



Note: one can actually prove the result with a smaller
constant in place of 2, as follows. Adding 21 f 0 (x) f 000 (x)
to both sides of (1) and again invoking the original We now use the fact that for R an n × m matrix and S an
bound f 000 (x) ≤ f (x), we get m × n matrix,

1 0 1 det(In + RS) = det(Im + SR).

[ f (x) f 000 (x) + ( f 00 (x))2 ] < f (x) f 0 (x) + f 0 (x) f 000 (x)
2 2
3 This yields
≤ f (x) f 0 (x).
2
Applying the fact again, we get det(IN + A)
  
1
u
1 0 3 = det In + uT vT
f (x) f 00 (x) < f (x)2 . 2 v
2 4    
1 u
Multiplying both sides by f 0 (x) and applying the fact uT vT


= det I2 +
once more, we get 2 v
2 + uuT uvT
 
1
1 0 1 = det
( f (x))3 < f (x)3 . 4 vuT 2 + vvT
6 4
2 + (x2 + · · · + x2n )
 
1 n
From this we deduce f 0 (x) < (3/2)1/3 f (x) < 2 f (x), as = det
4 n 2 + (x−2 + · · · + x−2n )
desired.
n2
 
1 2 n
I don’t know what the best constant is, except that it = det = 1− .
4 n 2 4
is not less than 1 (because f (x) = ex satisfies the given
conditions).
B–6 First solution: Choose a sequence p1 , p2 , . . . of primes
B–5 First solution: We claim that the eigenvalues of A are as follows. Let p1 be any prime dividing an element of
0 with multiplicity n − 2, and n/2 and −n/2, each with S. To define p j+1 given p1 , . . . , p j , choose an integer
multiplicity 1. To prove this claim, define vectors v(m) , N j ∈ S relatively prime to p1 · · · p j and let p j+1 be a
0 ≤ m ≤ n − 1, componentwise by (v(m) )k = eikmθ , and prime divisor of N j , or stop if no such N j exists.
note that the v(m) form a basis for Cn . (If we arrange the
v(m) into an n × n matrix, then the determinant of this Since S is finite, the above algorithm eventually termi-
matrix is a Vandermonde product which is nonzero.) nates in a finite sequence p1 , . . . , pk . Let m be the small-
Now note that est integer such that p1 · · · pm has a divisor in S. (By the
n
assumption on S with n = p1 · · · pk , m = k has this prop-
(Av(m) ) j = erty, so m is well-defined.) If m = 1, then p1 ∈ S, and we
∑ cos( jθ + kθ )eikmθ are done, so assume m ≥ 2. Any divisor d of p1 · · · pm
k=1
in S must be a multiple of pm , or else it would also be
ei jθ n
e−i jθ n
=
2 ∑ eik(m+1)θ + 2 ∑ eik(m−1)θ . a divisor of p1 · · · pm−1 , contradicting the choice of m.
k=1 k=1 But now gcd(d, Nm−1 ) = pm , as desired.
Since ∑nk=1 eik`θ = 0 for integer ` unless n | `, we Second solution (from sci.math): Let n be the small-
conclude that Av(m) = 0 for m = 0 or for 2 ≤ est integer such that gcd(s, n) > 1 for all s in n; note
m ≤ n − 1. In addition, we find that (Av(1) ) j = that n obviously has no repeated prime factors. By the
n −i jθ
= n2 (v(n−1) ) j and (Av(n−1) ) j = 2n ei jθ = n2 (v(1) ) j , condition on S, there exists s ∈ S which divides n.
2e
so that A(v(1) ± v(n−1) ) = ± 2n (v(1) ± v(n−1) ). Thus On the other hand, if p is a prime divisor of s, then by
{v(0) , v(2) , v(3) , . . . , v(n−2) , v(1) + v(n−1) , v(1) − v(n−1) } is the choice of n, n/p is relatively prime to some ele-
a basis for Cn of eigenvectors of A with the claimed ment t of S. Since n cannot be relatively prime to t, t
eigenvalues. is divisible by p, but not by any other prime divisor of
Finally, the determinant of I + A is the product of (1 + n (as those primes divide n/p). Thus gcd(s,t) = p, as
λ ) over all eigenvalues λ of A; in this case, det(I + A) = desired.
(1 + n/2)(1 − n/2) = 1 − n2 /4.