Solutions to the 76th William Lowell Putnam Mathematical Competition

Saturday, December 5, 2015

Manjul Bhargava, Kiran Kedlaya, and Lenny Ng

A1 First solution: Without loss of generality, assume that any integer m and any prime p dividing am , p also di-
A and B lie in the first quadrant with A = (t1 , 1/t1 ), B = vides a−m ; on the other hand, p cannot divide a−m+1 ,
(t2 , 1/t2 ), and t1 < t2 . If P = (t, 1/t) with t1 ≤ t ≤ t2 , as otherwise p would also divide a−m+2 , . . . , a0 = 1, a
then the area of triangle APB is contradiction. We can thus find an integer k such that
  am+1 ≡ ka−m+1 (mod p); by induction on n, we see
 1 1 1  that an ≡ kan−2m (mod p) for all n. In particular, if k is
1  t2 − t1
t 1 t t 2 =
 (t1 + t2 − t − t1t2 /t). odd, then p also divides akm ; we thus conclude (again)
2 1/t 1/t 1/t 
 2t1t2
1 2 that a2015 is divisible by a5 = 362 and thus by 181.
Remark: Although it was not needed in the solution,
When t1 ,t2 are fixed, this is maximized when t + t1t2 /t
we note in passing that if an ≡ 0 (mod p), then a2n+k ≡
is minimized,
√ which by AM-GM exactly holds when
−ak (mod p) for all k.
t = t1t2 .
Remark: One can find other odd prime factors of a2015
The line AP is given by y = t1 +t−x tt1 , and so the area of in the same manner. For example, a2015 is divisible by
the region bounded by the hyperbola and AP is each of the following quantities. (The prime factoriza-
Z t    tions were computed using the Magma computer algebra
t1 + t − x 1 t t1 t
− dx = − − log , system.)
t1 tt 1 x 2t 1 2t t1
√ a13 = 2 × 6811741
2 −t1
which at t = t1t2 is equal to 2t√
p
t1 t2 − log( t2 /t1 ). a31 = 2 × 373 × 360250962984637
Similarly, the area of the region bounded by the√hyper-
bola and PB is 2t t2
− 2tt 2 − log tt2 , which at t = t1t2 is a5·13 = 2 × 181 × 6811741
2 −t1 × 3045046274679316654761356161
also 2t√
p
t t − log( t2 /t1 ), as desired.
12
a5·31 = 1215497709121 × 28572709494917432101
Second solution: For any λ > 0, the map (x, y) 7→
× 13277360555506179816997827126375881581
(λ x, λ −1 y) preserves both areas and the hyperbola xy =
1. We may thus rescale the picture so that A, B are sym- a13·31 = 2 × 373 × 193441 × 6811741 × 360250962984637
metric across the line y = x, with A above the line. As × 16866100753000669
P moves from A to B, the area of APB increases until P × 79988387992470656916594531961 × p156
passes through the point (1, 1), then decreases. Conse-
quently, P = (1, 1) achieves the maximum area, and the where p156 is a prime of 156 decimal digits. Dividing
desired equality is obvious by symmetry. Alternatively, a2015 by the product of the primes appearing in this list
since the hyperbola is convex, the maximum is uniquely yields a number N of 824 decimal digits which is defi-
achieved at the point where the tangent line is parallel nitely not prime, because 2N 6≡ 2 (mod N), but whose
to AB, and by symmetry that point is P. prime factorization we have been unable to establish.
Note that N is larger than a 2048-bit RSA modulus, so
A2 First solution: One possible√answer is 181. √ nBy in- the difficulty of factoring it is not surprising.
duction, we have an = ((2 + 3)n + (2 − √ 3) )/2 =
(α n √
+ β n )/2 for all n, where α = 2 + 3 and β = One thing we can show is that each prime factor of N is
congruent to 1 modulo 6 × 2015 = 12090, thanks to the
2 − 3. Now note that if k is an odd positive integer
kn +β kn following lemma.
and an 6= 0, then aaknn = αα n +β n =α
(k−1)n −α (k−2)n β n +

· · · − α n β (k−2)n + β (k−1)n . This expression is both ra- Lemma. Let n be an odd integer. Then any odd prime factor
tional√(because an and akn are integers) and of the form p of an which does not divide am for any divisor m of n is
a + b 3 for some integers a, b by the expressions for congruent to 1 modulo lcm(6, n). (By either solution of the
α, β ; it follows that it must be an integer, and so akn is original problem, p also does not divide am for any positive
divisible by an . Applying this to n = 5 and k = 403, we integer m < n.)
find that a2015 is divisible by a5 = 362 and thus by 181. √
Proof. We first check that p ≡ 1 (mod 3). In Fq = F p ( 3)
Second solution: By rewriting the formula for an as we have (α/β )n ≡ −1. If p ≡ 2 (mod 3), then q = p2 and
an−2 = 4an−1 − an , we may extend the sequence back- n
wards to define an for all integers n. Since a−1 = 2,
α and β are conjugate in p; consequently,
n n
√ n
√ equality α =
the
−β in Fq2 means that α = c 3, β = −c 3 for some c ∈
we may see by induction that a−n = an for all n. For
F p . But then −3c2 = α n β n = 1 in Fq and hence in F p , which
contradicts p ≡ 2 (mod 3) by quadratic reciprocity.
 2

By the previous paragraph, α and β may be identified with and note (as above) that ω 2 , ω 4 , . . . , ω 2(n1 −1) is a per-
elements of F p , and we have (α/β )n ≡ −1, but the same does mutation of ω, . . . , ω n1 −1 , so the two products in the
not hold with n replaced by any smaller value. Since F× p is fraction are equal.
a cyclic group of order p − 1, this forces p ≡ 1 (mod n) as Remark: The function f (n) = ∑d|n d · φ (n/d) is multi-
claimed. plicative: for any two coprime positive integers m, n, we
have f (mn) = f (m) f (n). This follows from the fact that
A3 The answer is 13725. We first claim that if n is odd, f (n) is the convolution of the two multiplicative func-
then ∏nb=1 (1 + e2πiab/n ) = 2gcd(a,n) . To see this, write tions n 7→ n and n 7→ φ (n); it can also be seen directly
d = gcd(a, n) and a = da1 , n = dn1 with gcd(a1 , n1 ) = using the Chinese remainder theorem.
1. Then a1 , 2a1 , . . . , n1 a1 modulo n1 is a permutation
of 1, 2, . . . , n1 modulo n1 , and so ω a1 , ω 2a1 , . . . , ω n1 a1 A4 The answer is L = 4/7. For S ⊂ N, let F(S) =
is a permutation of ω, ω 2 , . . . , ω n1 ; it follows that for ∑n∈S 1/2n , so that f (x) = F(Sx ). Note that for T =
ω = e2πi/n1 , {1, 4, 7, 10, . . .}, we have F(T ) = 4/7.
n1 n1 n1 We first show by contradiction that for any x ∈ [0, 1),
∏ (1 + e2πiab/n ) = ∏ (1 + e2πia1 b/n1 ) = ∏ (1 + ω b ). f (x) ≥ 4/7. Since each term in the geometric series
b=1 b=1 b=1 ∑n 1/2n is equal to the sum of all subsequent terms, if
S, S0 are different subsets of N and the smallest positive
Now since the roots of zn1 − 1 are ω, ω 2 , . . . , ω n1 , it fol- integer in one of S, S0 but not in the other is in S, then
lows that zn1 − 1 = ∏nb=11
(z − ω b ). Setting z = −1 and F(S) ≥ F(S0 ). Assume f (x) < 4/7; then the smallest
using the fact that n1 is odd gives ∏nb=1
1
(1 + ω b ) = 2. integer in one of Sx , T but not in the other is in T . Now
Finally, ∏nb=1 (1 + e2πiab/n ) = (∏nb=1
1
(1 + e2πiab/n ))d = 1 ∈ Sx for any x ∈ [0, 1), and we conclude that there
d are three consecutive integers n, n + 1, n + 2 that are not
2 , and we have proven the claim.
in Sx : that is, bnxc, b(n + 1)xc, b(n + 2)xc are all odd.
From the claim, we find that Since the difference between consecutive terms in nx,
! (n + 1)x, (n + 2)x is x < 1, we conclude that bnxc =
2015 2015
2πiab/2015 b(n + 1)xc = b(n + 2)xc and so x < 1/2. But then 2 ∈ Sx
log2 ∏ ∏ (1 + e )
and so f (x) ≥ 3/4, contradicting our assumption.
a=1 b=1
2015 2015
! It remains to show that 4/7 is the greatest lower bound
2πiab/2015 for f (x), x ∈ [0, 1). For any n, choose x = 2/3 − ε with
= ∑ log2 ∏ (1 + e )
a=1 b=1 0 < ε < 1/(9n); then for 1 ≤ k ≤ n, we have 0 < mε <
2015 1/3 for m ≤ 3n, and so
= ∑ gcd(a, 2015).
a=1 b(3k − 2)xc = b(2k − 2) + 2/3 − (3k − 2)εc = 2k − 2
b(3k − 1)xc = b(2k − 1) + 1/3 − (3k − 1)εc = 2k − 1
Now for each divisor d of 2015, there are φ (2015/d)
integers between 1 and 2015 inclusive whose gcd with b(3k)xc = b(2k − 1) + 1 − 3kεc = 2k − 1.
2015 is d. Thus
It follows that Sx is a subset of S = {1, 4, 7, . . . , 3n −
2015 2, 3n + 1, 3n + 2, 3n + 3, . . .}, and so f (x) = F(Sx ) ≤
∑ gcd(a, 2015) = ∑ d · φ (2015/d). f (S) = (1/2 + 1/24 + · · · + 1/23n+1 ) + 1/23n+1 . This
a=1 d|2015 last expression tends to 4/7 as n → ∞, and so no num-
ber greater than 4/7 can be a lower bound for f (x) for
We factor 2015 = pqr with p = 5, q = 13, and r = 31,
all x ∈ [0, 1).
and calculate
A5 First solution: By inclusion-exclusion, we have
∑ d · φ (pqr/d)
 
d|pqr bq/4c
Nq = ∑ µ(d)
= 1 · (p − 1)(q − 1)(r − 1) + p · (q − 1)(r − 1) d|q
d
+ q · (p − 1)(r − 1) + r · (p − 1)(q − 1) + pq · (r − 1) 
q/d

+ pr · (q − 1) + qr · (p − 1) + pqr · 1 = ∑ µ(d)
d|q
4
= (2p − 1)(2q − 1)(2r − 1).  
q/d
≡ ∑ (mod 2),
When (p, q, r) = (5, 13, 31), this is equal to 13725. d|q squarefree
4
Remark: Noam Elkies suggests the following similar
but shorter derivation of the equality ∏nb=1
1
(1+ω b ) = 2: where µ is the Möbius function. Now
write   (
q/d 0 (mod 2) if q/d ≡ 1, 3 (mod 8)
n1 −1 n1 −1
∏b=1 (1 − ω 2b ) ≡
b 4 1 (mod 2) if q/d ≡ 5, 7 (mod 8).
∏ (1 + ω )= n1 −1
(1 − ω b )
b=1 ∏b=1
 3

So Nq is odd if and only if q has an odd number of If q = 1, then Nq is even. If q has more than one
squarefree factors q/d congruent to 5 or 7 (mod 8). prime factor, then the group (Z/qZ)× has exponent di-
If q has a prime factor p congruent to 1 or 3 (mod 8), viding φ (q)/2, so (−1)Nq ≡ (−2)φ (q)/2 ≡ 1 (mod q),
then the squarefree factors d of q occur in pairs c, pc, and thus Nq must be even in this case as well. Fi-
which are either both 1 or 3 (mod 8) or both 5 or 7 nally, suppose that q is a prime power pk with p odd
(mod 8). Hence q must have an even number of factors and k positive. Since (Z/qZ)× is a cyclic group of
that are congruent to 5 or 7 (mod 8), and so Nq is even order φ (q) = pk−1 (p − 1), in which the only square
in this case. roots of unity are ±1, it follows that (−2)φ (q)/2 ≡ ±1
If q has two prime factors p1 and p2 , each congruent to (mod q) in accordance with whether (−2)(p−1)/2 ≡ ±1
either 5 or 7 (mod 8), then the squarefree factors d of q (mod p), i.e., whether −2 is a quadratic residue or non-
occur in quadruples d, p1 d, q1 d, p1 q1 d, which are then residue. But recall that −2 is a quadratic residue mod-
congruent respectively to some permutation of 1,3,5,7 ulo p if and only if p ≡ 1, 3 (mod 8). Thus Nq is odd
(mod 8) (if p1 and p2 are distinct mod 8) or are congru- in this case if and only if p ≡ 5 or 7 (mod 8).
ent respectively to d, p1 d, p1 d, d (mod 8). Either way, We conclude that for any odd integer q ≥ 1, the quantity
we see that exactly two of the four residues are congru- Nq is odd if and only if q = pk with k positive and p a
ent to 5 or 7 (mod 8). Thus again q must have an even prime that is 5 or 7 (mod 8).
number of factors that are 5 or 7 (mod 8), and so Nq is Remark: The combination of the two solutions recov-
even in this case as well. ers Gauss’s criterion for when −2 is a quadratic residue
If q = 1, then Nq = 0 is even. The only case that remains modulo p, with essentially the original proof.
is that q = pk is a positive power of a prime p congruent
to 5 or 7 (mod 8). In this case, q has two squarefree A6 First solution: (by Noam Elkies) Using row and col-
factors, 1 and p, of which exactly one is congruent to 5 umn operations, we may construct invertible matrices
or 7 (mod 8). We conclude that Nq is odd in this case, U,V such that U −1 MV is a block diagonal matrix of
as desired. the form
 
Second solution: I 0
.
Consider the set S of all integers in {1, . . . , q − 1} that 0 0
are even and relatively prime to q. Then the product of
all elements in S is Put A0 = U −1 AU, M 0 = U −1 MV, B0 = V −1 BV , X 0 =
V −1 XU, so that A0 M 0 = M 0 B0 , det(A − MX) =
2φ (q)/2 ∏ a. det(U −1 (A − MX)U) = det(A0 − M 0 X 0 ), and det(B −
1≤a≤(q−1)/2 XM) = det(V −1 (B − XM)V ) = det(B0 − X 0 M 0 ). Form
(a,q)=1
the corresponding block decompositions
On the other hand, we can rewrite the set of ele-      
A11 A12 B11 B12 X11 X12
ments in S (mod q) as a set T of residues in the in- A0 = , B0 = ,X0 = .
A21 A22 B21 B22 X21 X22
terval [−(q − 1)/2, (q − 1)/2]. Then for each 1 ≤ a ≤
(q − 1)/2 with (a, q) = 1, T contains exactly one el- We then have
ement from {a, −a}: if −2r ≡ 2s (mod q) for some
r, s ∈ {1, . . . , (q − 1)/2}, then r ≡ −s (mod q), which is
   
A11 0 B11 B12
impossible given the ranges of r and s. Thus the product A0 M 0 = , M 0 B0 = ,
A21 0 0 0
of all elements in T is
so we must have A11 = B11 and A21 = B12 = 0; in partic-
(−1)n ∏ a, ular, the characteristic polynomial of A is the product of
1≤a≤(q−1)/2
(a,q)=1
the characteristic polynomials of A11 and A22 , and the
characteristic polynomial of B is the product of the char-
where n denotes the number of elements of S greater acteristic polynomials of B11 and B22 . Since A11 = B11 ,
than (q − 1)/2. We conclude that (−1)n ≡ 2φ (q)/2 it follows that A22 and B22 have the same characteristic
(mod q). polynomial. Since
However, note that the number of elements of S less    
X11 0 X11 X12
than (q − 1)/2 is equal to Nq , since dividing these num- X 0M0 = , M0X 0 = ,
X21 0 0 0
bers by 2 gives exactly the numbers counted by Nq .
Hence the total cardinality of S is Nq + n; however,
this cardinality also equals φ (q)/2 because the num-
bers in {1, . . . , q − 1} relatively prime to q come in pairs
{a, q − a} in each of which exactly one member is even.
We thus obtain

(−1)Nq = (−1)φ (q)/2+n

≡ (−1)φ (q)/2 2φ (q)/2 = (−2)φ (q)/2 (mod q).
 4

we conclude that To establish this equality, first apply the remark follow-
ing the previous solution to write
det(A − MX) = det(A0 − M 0 X 0 )

A11 − X11 A12 − X12
 Trace(Ai1 MXAi2 MX · · · Aim−1 MXAim )
= det
0 A22 = Trace(Aim +i1 MXAi2 MX · · · Aim−1 MX).
= det(A11 − X11 ) det(A22 )
Then apply the relation AM = MB repeatedly to com-
= det(B11 − X11 ) det(B22 ) mute M past A, to obtain
 
B − X11 0
= det 11 Trace(MBim +i1 XMBi2 XM · · · XMBim−1 X).
B21 − X21 B22
= det(B0 − X 0 M 0 ) Finally, apply the remark again to shift MBim from the
= det(B − XM), left end to the right end.
Remark: The conclusion holds with R replaced by an
as desired. (By similar arguments, A − MX and B − XM arbitrary field. In the second solution, one must reduce
have the same characteristic polynomial.) to the case of an infinite field, e.g., by replacing the orig-
Second solution: We prove directly that A − MX and inal field with an algebraic closure. The third solution
B − XM have the same characteristic polynomial, i.e., only applies to fields of characteristic 0 or positive char-
for any t ∈ R, writing At = A − tI, Bt = B − tI, we have acteristic greater than n.
Remark: It is tempting to try to reduce to the case
det(At − MX) = det(Bt − XM).
where M is invertible, as in this case A − MX and
For fixed A, B, M, the stated result is a polynomial iden- B − XM are in fact similar. However, it is not clear how
tity in t and the entries of X. It thus suffices to check to make such an argument work.
it assuming that At , Bt , X are all invertible. Since AM = B1 Let g(x) = ex/2 f (x). Then g has at least 5 distinct real
MB, we also have At M = MBt , so At MBt−1 = M. Since zeroes, and by repeated applications of Rolle’s theorem,
det(At ) = det(Bt ) by hypothesis, g0 , g00 , g000 have at least 4, 3, 2 distinct real zeroes, respec-
tively. But
det(At − MX) = det(At − At MBt−1 X)
= det(At ) det(1 − MBt−1 X) 1
g000 (x) = ex/2 ( f (x) + 6 f 0 (x) + 12 f 00 (x) + 8 f 000 (x))
8
= det(At ) det(X) det(Bt )−1 det(X −1 Bt − M)
= det(X) det(X −1 Bt − M) and ex/2 is never zero, so we obtain the desired result.
= det(Bt − XM). B2 We will prove that 42015 is such a number in the se-
quence. Label the sequence of sums s0 , s1 , . . . , and let
Remark: One can also assert directly that det(1 − an , bn , cn be the summands of sn in ascending order. We
MBt−1 X) = det(1 − XMBt−1 ) using the fact that for any prove the following two statements for each nonnega-
square matrices U and V , UV and VU have the same tive integer n:
characteristic polynomial; the latter is again proved by
reducing to the case where one of the two matrices is (a)n The sequence
invertible, in which case the two matrices are similar.
a3n , b3n , c3n , a3n+1 , b3n+1 , c3n+1 , a3n+2 , b3n+2 , c3n+2
Third solution: (by Lev Borisov) We will check that
for each positive integer k, is obtained from the sequence 10n + 1, . . . , 10n +
10 by removing one of 10n + 5, 10n + 6, 10n + 7.
Trace((A − MX)k ) = Trace((B − XM)k ).
(b)n We have
This will imply that A − MX and B − XM have the same
s3n = 30n + 6,
characteristic polynomial, yielding the desired result.
s3n+1 ∈ {30n + 15, 30n + 16, 30n + 17},
We establish the claim by expanding both sides
and comparing individual terms. By hypothesis, Ak s3n+2 = 30n + 27.
and Bk have the same characteristic polynomial, so
Trace(Ak ) = Trace(Bk ). To compare the other terms, These statements follow by induction from the follow-
it suffices to check that for any sequence i1 , i2 , . . . , im of ing simple observations:
nonnegative integers, – by computing the table of values
Trace(Ai1 MXAi2 MX · · · Aim−1 MXAim ) n an bn cn sn
= Trace(Bi1 XMBi2 XM · · · Bim−1 XMBim ). 0 1 2 3 6
1 4 5 7 16
2 8 9 10 27

we see that (a)0 holds;

 5

– (a)n implies (b)n ; Second solution: If a, b, c, d are in arithmetic progres-

– (a)n and (b)1 , . . . , (b)n together imply (a)n+1 . sion, then we may write

To produce a value of n for which sn ≡ 2015 a = r − 3s, b = r − s, c = r + s, d = r + 3s

(mod 10000), we take n = 3m + 1 for some nonnega-
tive integer m for which s3m+1 = 30m + 15. We must for some r, s. If s = 0, then clearly all powers of M are
also have 30m ≡ 2000 (mod 10000), or equivalently in xS. Also, if r = 0, then one easily checks that M 3 is
m ≡ 400 (mod 1000). By taking m = 1400, we ensure in S.
that m ≡ 2 (mod 3), so sm = 10m + 7; this ensures that We now assume rs 6= 0, and show that in that case M
sn does indeed equal 30m + 15 = 42015, as desired. cannot be in S. First, note that the characteristic polyno-
Remark: With a bit more work, we can give a complete mial of M is x2 − 2rx − 8s2 , and since M is nonsingular
description of sn , and in particular find the first term in (as s 6= 0), this is also the minimal polynomial of M by
the sequence whose decimal expansion ends in 2015. the Cayley-Hamilton theorem. By repeatedly using the
Define the function on nonnegative integers relation M 2 = 2rM + 8s2 I, we see that for each positive
integer, we have M k = tk M + uk I for unique real con-
f (n) = s3n+1 − (30n + 16) stants tk , uk (uniqueness follows from the independence
of M and I). Since M is in S, we see that M k lies in S
which takes values in {−1, 0, 1}; we then have only if uk = 0.
 On the other hand, we claim that if k > 1, then rtk > 0
0
 n ≡ 0 (mod 3) and uk > 0 if k is even, and tk > 0 and ruk > 0 if k is
f (n) = − f ((n − 1)/3) n ≡ 1 (mod 3) odd (in particular, uk can never be zero). The claim is


−1 n ≡ 2 (mod 3). true for k = 2 by the relation M 2 = 2rM + 8s2 I. Assum-
ing the claim for k, and multiplying both sides of the
Consequently, if we write n in base 3, then f (n) = 0 relation M k = tk M + uk I by M, yields
unless the expansion ends with 2 followed by a string
M k+1 = tk (2rM + 8s2 I) + uk M = (2rtk + uk )M + 8s2tk I,
of 1s of length k ≥ 0, in which case f (n) = (−1)k+1 .
In this notation, we have sn ≡ 2015 (mod 10000) if and implying the claim for k + 1.
only if n = 3m + 1 for some nonnegative integer m for Remark: (from artofproblemsolving.com, user
which m ≡ 400 (mod 1000) and f (m) = −1. Since hoeij) Once one has uk = 0, one can also finish using
400 = 112211(3) , the first such term in the sequence is the relation M · M k = M k · M.
in fact s1201 = 12015.
B4 First solution: The answer is 17/21. For fixed b, c,
B3 First solution: Any element be
written as M =

of S can−1 there is a triangle of side lengths a, b, c if and only if
αA + β B, where A = 11 11 , B = −3 1 3
, and α, β ∈ R. |b − c| < a < b + c. It follows that the desired sum is
Note that A2 = 44 44 and B3 = −24 −8


8 24 are both in S,
and so any matrix of the form αA or β B, α, β ∈ R,
!
1 b+c−1
2b+c − 2|b−c|+1
satisfies the given condition. S=∑ b c ∑ 2a = ∑ .
b,c 3 5 a=|b−c|+1 b,c 3b 5c
We claim that these are also the only matrices in S
satisfying the given condition. Indeed,  suppose
√ 
M= We write this as S = S1 +S2 where S1 sums over positive
1 1/ 2
αA + β B where α, β 6= 0. Let C = √ with in- integers b, c with b ≤ c and S2 sums over b > c. Then
  −1 1/ 2
1/2 −1/2
verse C−1 = 1/√2 1/√2 . If we define D = C−1 MC, ∞ ∞
2b+c − 2c−b+1
√ S1 =
 
then D = 2α 0γ 1γ where γ = − β α 2 . Now suppose
∑∑ 3b 5c
b=1 c=b
!  !
that M k is in S with k ≥ 2. Since ( 1 −1 ) A −1 1
2 b 2 2 c


= ∞   ∞
1 =∑ − b ∑
= 0, we have ( 1 −1 ) M k −1 1



( 1 −1 ) B −1 = 0, and b=1 3 6 c=b 5
so the upper left entry of C−1 M kC = Dk is 0. On the b
!  
2 5 2 b
∞  
other hand, from the expression for 2
 2D, an easy induc- =∑ − b
k
tion on k shows that D = (2α) k γ pk−1 γ pk 3 6 3 5
γ pk pk+1 , where
b=1
  !
pk is defined inductively by p0 = 0, p1 = 1, pk+2 = 5 4 b 10 1 b
∞  
γ 2 pk + pk+1 . In particular, it follows from the inductive =∑ −
b=1 3 15 3 15
definition that pk > 0 when k ≥ 1, whence the upper left
entry of Dk is nonzero when k ≥ 2, a contradiction. 85
= .
Remark: A variant of this solution can be obtained by 231
diagonalizing the matrix M.
 6

Similarly, be the number which has n − 1, n consecutively in that

order, but not at the beginning or end. It is clear that
∞
2b+c − 2b−c+1
∞
every permutation π counted by Pn either lies in exactly
S2 = ∑ ∑
c=1 b=c+1 3b 5c one of the sets counted by Un ,Vn ,Wn , Tn , Sn , or is the
 c  ∞  b ! reverse of such a permutation. Therefore
∞
2 2 2
=∑ − c ∑ 3
5 10 Pn = 2(Un +Vn +Wn + Tn + Sn ).
c=1 b=c+1
∞  c   c+1
2 2 2 By examining how each of the elements in the sets
=∑ − c 3
5 10 3 counted by Un+1 ,Vn+1 ,Wn+1 , Tn+1 , Sn+1 can be ob-
c=1
c tained from a (unique) element in one of the sets
1 c
∞      
4 counted by Un ,Vn ,Wn , Tn , Sn by suitably inserting the el-
=∑ 2 −4
c=1 15 15 ement n + 1, we obtain the recurrence relations
34
= . Un+1 = Un +Wn + Tn ,
77
Vn+1 = Un ,
17
We conclude that S = S1 + S2 = 21 . Wn+1 = Wn ,
Second solution: Recall that the real numbers a, b, c Tn+1 = Vn ,
form the side lengths of a triangle if and only if
Sn+1 = Sn +Vn .
a+b+c
s − a, s − b, s − c > 0 s= , Also, it is clear that Wn = 1 for all n.
2
So far we have assumed n ≥ 3, but it is straightforward
and that if we put x = 2(s−a), y = 2(s−b), z = 2(s−c), to extrapolate the sequences Pn ,Un ,Vn ,Wn , Tn , Sn back
to n = 2 to preserve the preceding identities. Hence for
y+z z+x x+y
a= ,b = ,c = . all n ≥ 2,
2 2 2
To generate all integer triples (a, b, c) which form the Pn+5 = 2(Un+5 +Vn+5 +Wn+5 + Tn+5 + Sn+5 )
side lengths of a triangle, we must also assume that = 2((Un+4 +Wn+4 + Tn+4 ) +Un+4
x, y, z are either all even or all odd. We may therefore +Wn+4 +Vn+4 + (Sn+4 +Vn+4 ))
write the original sum as = Pn+4 + 2(Un+4 +Wn+4 +Vn+4 )
2(y+z)/2 2(y+z)/2 = Pn+4 + 2((Un+3 +Wn+3 + Tn+3 ) +Wn+3 +Un+3 )
∑ + ∑ . = Pn+4 + Pn+3 + 2(Un+3 −Vn+3 +Wn+3 − Sn+3 )
x,y,z>0 odd
3(z+x)/2 5(x+y)/2 x,y,z>0 even 3(z+x)/2 5(x+y)/2
= Pn+4 + Pn+3 + 2((Un+2 +Wn+2 + Tn+2 ) −Un+2
To unify the two sums, we substitute in the first case +Wn+2 − (Sn+2 −Vn+2 ))
x = 2u + 1, y = 2v + 1, z = 2w + 1 and in the second = Pn+4 + Pn+3 + 2(2Wn+2 + Tn+2 − Sn+2 −Vn+2 )
case x = 2u + 2, y = 2v + 2, z = 2w + 2 to obtain
= Pn+4 + Pn+3 + 2(2Wn+1 +Vn+1
2a 2v+w 2−1
∞  
− (Sn+1 +Vn+1 ) −Un+1 )
=
∑ 3b 5c ∑ 3w+u 5u+v 1 +
(a,b,c)∈T u,v,w=1 3−1 5−1 = Pn+4 + Pn+3 + 2(2Wn +Un − (Sn +Vn ) −Un
17 ∞
 u ∞  v ∞  w
1 2 2 − (Un +Wn + Tn ))
= ∑ 15 ∑ 5 ∑ 3 = Pn+4 + Pn+3 − Pn + 4,
2 u=1 v=1 w=1
17 1/15 2/5 2/3 as desired.
=
2 1 − 1/15 1 − 2/5 1 − 2/3
Remark: There are many possible variants of the
17 above solution obtained by dividing the permutations
= .
21 up according to different features. For example, Karl
Mahlburg suggests writing
B5 The answer is 4.
Assume n ≥ 3 for the moment. We write Pn = 2Pn0 , Pn0 = Q0n + R0n
the permutations π counted by Pn as sequences
π(1), π(2), . . . , π(n). Let Un be the number of permuta- where Pn0 counts those permutations counted by Pn for
tions counted by Pn that end with n − 1, n; let Vn be the which 1 occurs before 2, and Q0n counts those permuta-
number ending in n, n − 1; let Wn be the number starting tions counted by Pn0 for which π(1) = 1. One then has
with n − 1 and ending in n − 2, n; let Tn be the number the recursion
ending in n − 2, n but not starting with n − 1; and let Sn
Q0n = Q0n−1 + Q0n−3 + 1
 7
√ √ √
corresponding to the cases where π(1), π(2) = 1, 2; of summands is at most n (since n(2√ n − 1) ≥ n),
where π(1), π(2), π(3) = 1, 3, 2; and the unique case and so the total is bounded above by 1/ n. Hence the
1, 3, 5, . . . , 6, 4, 2. Meanwhile, one has difference converges to zero as n → ∞; that is, S1 con-
verges and equals S2 .
R0n = R0n−1 + Q0n−2
We may thus focus hereafter on computing S2 . We be-
corresponding to the cases containing 3, 1, 2, 4 (where gin by writing
removing 1 and reversing gives a permutation counted ∞ ∞ Z 1
1
by R0n−1 ); and where 4 occurs before 3, 1, 2 (where re- S2 = ∑ 2` − 1 ∑ (−1)m−1 t m−1 dt.
moving 1, 2 and reversing gives a permutation counted `=1 m=` 0
by Q0n−2 ).
Our next step will be to interchange the inner sum and
Remark: The permutations counted by Pn are known as the integral, but again this requires some justification.
key permutations, and have been studied by E.S. Page,
Systematic generation of ordered sequences using re- Lemma 1. Let f0 , f1 , . . . be a sequence of continuous func-
currence relations, The Computer Journal 14 (1971), tions on [0, 1] such that for each x ∈ [0, 1], we have
no. 2, 150–153. We have used the same notation for
consistency with the literature. The sequence of the Pn f0 (x) ≥ f1 (x) ≥ · · · ≥ 0.
also appears as entry A003274 in the On-line Encyclo-
pedia of Integer Sequences (http://oeis.org). Then
Z 1 Z 1
!
∞ ∞
B6 (from artofproblemsolving.com) We will prove n n
that the sum converges to π 2 /16. Note first that the ∑ (−1) 0
fn (t) dt =
0
∑ (−1) fn (t) dt
n=0 n=0
sum does not converge absolutely, so we are not free to
rearrange it arbitrarily. For that matter, the standard al- provided that both sums converge.
ternating sum test does not apply because the absolute
values of the terms does not decrease to 0, so even the Proof. Put gn (t) = f2n (t) − f2n+1 (t) ≥ 0; we may then rewrite
convergence of the sum must be established by hand. the desired equality as
Setting these issues aside momentarily, note that the el- !
∞ Z 1 Z 1 ∞
ements of the set counted by A(k) are those odd posi-
∑ gn (t) dt = ∑ gn (t) dt,
tive integers
√ d for which m = k/d is also an integer and n=0 0 0 n=0
d < 2dm; if we write d = 2` − 1, then the condition
on m reduces to m ≥ `. In other words, the original sum which is a case of the Lebesgue monotone convergence theo-
equals rem.
∞
(−1)m−1 By Lemma 1, we have
S1 := ∑ ∑ ,
k=1 `≥1,m≥` m(2` − 1) !
∞ Z 1 ∞
k=m(2`−1) 1 m−1 m−1
S2 = ∑ ∑ (−1) t dt
and we would like to rearrange this to `=1 2` − 1 0 m=`
(−t)`−1
∞ Z 1
1
∞
1 ∞
(−1)m−1 = ∑ 2` − 1 dt.
S2 := ∑ ∑ , `=1 0 1+t
`=1 2` − 1 m=` m
Since the outer sum is absolutely convergent, we may
in which both sums converge by the alternating sum freely interchange it with the integral:
test. In fact a bit more is true: we have
!
1 (−t)`−1
  Z 1 ∞
 ∞ (−1)m−1  1
S2 = ∑ dt
< , `=1 2` − 1 1 + t
 
∑ 0
m=` m  ` !
(−1)`−1t `−1/2
Z 1 ∞
1
so the outer sum converges absolutely. In particular, S2 = √ ∑ 2` − 1 dt
0 t(1 + t) `=1
is the limit of the truncated sums
Z 1
1 √
1 ∞
(−1)m−1 = √ arctan( t) dt
S2,n = ∑ ∑ . 0 t(1 + t)
`(2`−1)≤n
2` − 1 m=` m Z 1 √
2
= 2
arctan(u) du (u = t)
0 1+u
To see that S1 converges to the same value as S2 , write
π2
n = arctan(1)2 − arctan(0)2 = .
A(k) 1 ∞
(−1)m−1 16
S2,n − ∑ (−1)k−1 = ∑ ∑ .
k=1 k `(2`−1)≤n
2` − 1 m=b n +1c
m
2`−1
The expression on the right is bounded above in abso-
lute value by the sum ∑`(2`−1)≤n n1 , in which the number