Solutions to the 77th William Lowell Putnam Mathematical Competition

Saturday, December 3, 2016

Kiran Kedlaya and Lenny Ng

A1 The answer is j = 8. First suppose that j satisfies the We conclude that M(n) is the greatest integer strictly
given condition. For p(x) = x j , we have p( j) (x) = j! less than α(n), and thus that α(n) − 1 ≤ M(n) < α(n).
and thus j! is divisible by 2016. Since 2016 is divisible Now
by 25 and 7! is not, it follows that j ≥ 8. Conversely, we
√
q
claim that j = 8 works. Indeed, let p(x) = ∑nm=0 am xm α(n) 3 − 1
n + 5 − 2n + n12 3+ 5
be a polynomial with integer coefficients; then if k is lim = lim =
n→∞ n n→∞ 2 2
any integer,
√
3+ 5
n and similarly limn→∞ α(n)−1
n = 2 √, and so by the
p(8) (k) = ∑ m(m − 1) · · · (m − 7)am km−8 sandwich theorem, M(n)
limn→∞ n = 3+2 5 .
m=8
n  
A3 The given functional equation, along with the same
m
= ∑ 8!am km−8 equation but with x replaced by x−1 1
m=8 8 x and 1−x respec-
tively, yields:
is divisible by 8! = 20 · 2016, and so p(8) (k) is divisible 
1

by 2016. f (x) + f 1 − = tan−1 (x)
x
Remark: By the same reasoning, if one replaces 2016      
x−1 1 −1 x − 1
in the problem by a general integer N, then the min- f +f = tan
imum value of j is the smallest one for which N di- x 1−x x
   
vides j!. This can be deduced from Pólya’s observation 1 −1 1
f + f (x) = tan .
that the set of integer-valued polynomials is the free
Z- 1−x 1−x
module generated by the binomial polynomials nx for
n = 0, 1, . . . . That statement can be extended to polyno- Adding the first and third equations and subtracting the
mials evaluated on a subset of a Dedekind domain using second gives:
Bhargava’s method of P-orderings; we do not know if    
this generalization can be adapted to the analogue of −1 −1 1 −1 x − 1
2 f (x) = tan (x) + tan − tan .
this problem, where one considers polynomials whose 1−x x
j-th derivatives take integral values on a prescribed sub-
set. Now tan−1 (t) + tan−1 (1/t) is equal to π/2 if t > 0 and
√ −π/2 if t < 0; it follows that for x ∈ (0, 1),
3+ 5
A2 The answer is Note that for m > n + 1, both bi-
2 .
2( f (x) + f (1 − x)) = tan−1 (x) + tan−1 (1/x)


nomial coefficients are nonzero and their ratio is
  
1
   
m m−1 m!n!(m − n − 1)! −1
+ tan (1 − x) + tan −1
/ = 1−x
n−1 n (m − 1)!(n − 1)!(m − n + 1)!
mn
    
= . −1 x − 1 −1 x
− tan + tan
(m − n + 1)(m − n) x x−1
π π π
Thus the condition n−1m

> m−1


is equivalent to (m − = + +
n 2 2 2
n + 1)(m − n) − mn < 0. The left hand side of this last 3π
inequality is a quadratic function of m with roots = .
2
√
3n − 1 + 5n2 − 2n + 1 Thus
α(n) = ,
√2
Z 1 Z 1
3π
3n − 1 − 5n2 − 2n + 1 4 f (x) dx = 2 ( f (x) + f (1 − x))dx =
β (n) = , 0 0 2
2
R1 3π
and finally 0 f (x) dx = 8 .
both of which are real since 5n2 − 2n + 1 = 4n2 + (n −
1)2 > 0; it follows that m satisfies the given inequality Remark: Once one has the formula for f (x), one can
if and only if β (n) <√
m < α(n). (Note in particular that also (with some effort) directly evaluate the integral of
since α(n)−β (n) = 5n2 − 2n + 1 > 1, there is always each summand over [0, 1] to obtain the same result. A
some integer m between β (n) and α(n).)
 2

much cleaner variant of this approach (suggested on values m1 , n1 , . . . , mr , nr ∈ {−1, 1} for which
AoPS, user henrikjb) is to write
Z y s = gm1 hn1 · · · gmr hnr .
1
tan−1 (x) = dy We first check that every s ∈ G admits at least one rep-
0 1 + y2
resentation of some length; this is equivalent to saying
and do a change of variable on the resulting double in- that the set S of s ∈ G which admit representations of
tegral. some length is equal to G itself. Since S is closed un-
der the group operation and G is finite, S is also closed
A4 The minimum number of tiles is mn. To see that this under formation of inverses and contains the identity
many are required, label the squares (i, j) with 1 ≤ i ≤ element; that is, S is a subgroup of G. In particular, S
2m − 1 and 1 ≤ j ≤ 2n − 1, and for each square with i, j contains not only gh but also its inverse h−1 g−1 ; since
both odd, color the square red; then no tile can cover S also contains g−1 h, we deduce that S contains g−2 .
more than one red square, and there are mn red squares. Since g is of odd order in G, g−2 is also a generator of
It remains to show that we can cover any (2m − 1) × the cyclic subgroup containing g; it follows that g ∈ S
(2n − 1) rectangle with mn tiles when m, n ≥ 4. First and hence h ∈ S. Since we assumed that g, h generate
note that we can tile any 2 × (2k − 1) rectangle with G, we now conclude that S = G, as claimed.
k ≥ 3 by k tiles: one of the first type, then k − 2 of the To complete the proof, we must now check that for each
second type, and finally one of the first type. Thus if we s ∈ G, the smallest possible length of a representation of
can cover a 7×7 square with 16 tiles, then we can do the s cannot exceed |G|. Suppose the contrary, and let
general (2m − 1) × (2n − 1) rectangle, by decomposing
this rectangle into a 7×7 square in the lower left corner, s = gm1 hn1 · · · gmr hnr
along with m − 4 (2 × 7) rectangles to the right of the
square, and n − 4 ((2m − 1) × 2) rectangles above, and be a representation of the smallest possible length. Set
tiling each of these rectangles separately, for a total of
16 + 4(m − 4) + m(n − 4) = mn tiles. si = gm1 hn1 · · · gmi hni (i = 0, . . . , r − 1),
To cover the 7 × 7 square, note that the tiling must con-
interpreting s0 as e; since r > |G| by hypothesis, by the
sist of 15 tiles of the first type and 1 of the second type,
pigeonhole principle there must exist indices 0 ≤ i <
and that any 2 × 3 rectangle can be covered using 2 tiles
j ≤ r − 1 such that si = s j . Then
of the first type. We may thus construct a suitable cov-
ering by covering all but the center square with eight s = gm1 hn1 · · · gmi hni gm j+1 hn j+1 · · · gmr hnr
2 × 3 rectangles, in such a way that we can attach the
center square to one of these rectangles to get a shape is another representation of s of length strictly less than
that can be covered by two tiles. An example of such r, a contradiction.
a covering, with the remaining 2 × 3 rectangles left in-
tact for visual clarity, is depicted below. (Many other Remark: If one considers s1 , . . . , sr instead of
solutions are possible.) s0 , . . . , sr−1 , then the case s = e must be handled sep-
arately: otherwise, one might end up with a represen-
tation of length 0 which is disallowed by the problem
statement.
Reinterpretation: Note that the elements
gh, gh−1 , g−1 h, g−1 h−1 generate gh(g−1 h)−1 = g2
and hence all of G (again using the hypothesis that g
has odd order, as above). Form the Cayley digraph on
the set G, i.e., the directed graph with an edge from s1 to
s2 whenever s2 = s1 ∗ for ∗ ∈ {gh, gh−1 , g−1 h, g−1 h−1 }.
Since G is finite, this digraph is strongly connected:
there exists at least one path from any vertex to any
other vertex (traveling all edges in the correct direc-
tion). The shortest such path cannot repeat any vertices
(except the starting and ending vertices in case they
coincide), and so has length at most |G|.
Second solution: For r a positive integer, let Sr be the
set of s ∈ G which admit a representation of length at
most r (terminology as in the first solution); obviously
Sr ⊆ Sr+1 . We will show that Sr 6= Sr+1 unless Sr = G;
this will imply by induction on r that #Sr ≥ min{r, |G|}
A5 First solution: For s ∈ G and r a positive integer, define
and hence that Sr = G for some r ≤ |G|.
a representation of s of length r to be a sequence of
 3

has this property. It is apparent that 01 P(x) dx = 5/6.

R
Suppose that Sr = Sr+1 . Then the map s 7→ sgh de-
fines an injective map Sr → Sr+1 = Sr , so Sr is closed Since P0 (x) = (2x − 1)(6x − 5) and
under right multiplication by gh. By the same to-    
ken, Sr is closed under right multiplication by each of 1 5 25
P(0) = 0, P = 1, P = , P(1) = 1,
gh−1 , g−1 h, g−1 h−1 . Since gh, gh−1 , g−1 h, g−1 h−1 gen- 2 6 27
erate G as in the first solution, it follows that Sr = G as
[Link] r + 1 ≤ |G|, we are done in this case also. it follows that P increases from 0 at x = 0 to 1 at x =
1/2, then decreases to a positive value at x = 5/6, then
Remark: The condition on the order of g is needed
increases to 1 at x = 1. Hence P has the desired form.
to rule out the case where G admits a (necessarily nor-
mal) subgroup H of index 2 not containing either g or Remark: Here is some conceptual motivation for the
h; in this case, all products of the indicated form be- preceding solution. Let V be the set of polynomials
long to H. On the other hand, if one assumes that both of degree at most 3 vanishing at 0, viewed as a three-
g and h have odd order, then one can say a bit more: dimensional vector space over R. Let S be the subset
there exists some positive integer r with 1 ≤ r ≤ |G| of V consisting of those polynomials P(x) for which
such that every element of G has a representation of 0 ≤ P(x) ≤ 1 for all x ∈ [0, 1]; this set is convex and
length exactly r. (Namely, the set of such elements for compact. We may then compute the minimal C as the
maximum value of 01 P(x) dx over all P ∈ S, provided
R
a given r strictly increases in size until it is stable un-
der right multiplication by both gh(g−1 h)−1 = g2 and that the maximum is achieved for some polynomial of
gh(gh−1 )−1 = gh2 g−1 , but under the present hypothe- degree exactly 3. (Note that any extremal polynomial
ses these generate G.) must satisfy maxx∈[0,1] P(x) = 1, as otherwise we could
multiply it by some constant c > 1 so as to increase
A6 We prove that the smallest such value of C is 5/6. R1
0 P(x) dx.)
First solution: (based on a suggestion of Daniel Kane)
Let f : V → R be the function taking P(x) to 01 P(x) dx.
R
We first reduce to the case where P is nonnegative in This function is linear, so we can characterize its ex-
[0, 1] and P(0) = 0. To achieve this reduction, suppose trema on S easily: there exist exactly two level surfaces
that a given value C obeys the inequality for such P. for f which are supporting planes for S, and the inter-
For P general, divide the interval [0, 1] into subintervals sections of these two planes with S are the minima and
I1 , . . . , Ik at the roots of P. Write `(Ii ) for the length of the maxima. It is obvious that the unique minimum is
the interval Ii ; since each interval is bounded by a root achieved by the zero polynomial, so this accounts for
of P, we may make a linear change of variable to see one of the planes.
that
It thus suffices to exhibit a single polynomial P(x) ∈ S
such that the level plane of f through P is a support-
Z
|P(x)| dx ≤ C`(Ii ) max |P(x)| (i = 1, . . . , k).
Ii x∈Ii ing plane for S. The calculation made in the solution
amounts to verifying that
Summing over i yields the desired inequality.
Suppose now that P takes nonnegative values on [0, 1], P(x) = 4x3 − 8x2 + 5x
P(0) = 0, and maxx∈[0,1] P(x) = 1. Write P(x) = ax3 +
has this property, by interpolating between the con-
bx2 + cx for some a, b, c ∈ R; then straints P(1/2) ≤ 1 and P(1) ≤ 1.
Z 1
1 1 1 This still leaves the matter of correctly guessing the op-
P(x) dx = a + b + c timal polynomial. If one supposes that it should be
0 4 3 2
  extremized both at x = 1 and at an interval value of
2 1 1 1 1
= a + b + c + (a + b + c) the disc, it is forced to have the form P(x) = 1 + (x −
3 8 4 2 6 1)(cx − 1)2 for some c > 0; the interpolation property
 
2 1 1 then pins down c uniquely.
= P + P(1)
3 2 6 Second solution: (by James Merryfield, via AoPS) As
2 1 5 in the first solution, we may assume that P is nonnega-
≤ + = .
3 6 6 tive on [0, 1] and P(0) = 0. Since P has degree at most 3,
Simpson’s rule for approximating 01 P(x) dx is an exact
R
Consequently, the originally claimed inequality holds
formula:
with C = 5/6. To prove that this value is best possi-
ble, it suffices to exhibit a polynomial P as above with Z 1  
1 1
R1 P(x) dx = (P(0) + 4P + P(1)).
0 P(x) dx = 5/6; we will verify that 0 6 2

P(x) = 4x3 − 8x2 + 5x This immediately yields the claimed inequality for C =
5/6. Again as in the first solution, we obtain an example
showing that this value is best possible.
 4

B1 Note that the function ex − x is strictly increasing for Remark: John Rickert points out that when N = n4 ,
x > 0 (because its derivative is ex − 1, which is positive one can turn the previous estimates into exact calcula-
because ex is strictly increasing), and its value at 0 is 1. tions to obtain the formula
By induction on n, we see that xn > 0 for all n.
4  3 n  4 3/4 2 1/4
By exponentiating the equation defining xn+1 , we ob- S(N) = n + = N + N .
3 2 3 3
tain the expression
For general N, one can then use the estimates
xn = exn − exn+1 .
4 2
(N − 1)3/4 + (N − 1)1/4 ≤ S(bN 1/4 c4 )
We use this equation repeatedly to acquire increasingly 3 3
precise information about the sequence {xn }. ≤ S(N)
– Since xn > 0, we have exn > exn+1 , so xn > xn+1 . ≤ S(dN 1/4 e4 )
– Since the sequence {xn } is decreasing and 4 2
≤ (N + 1)3/4 + (N + 1)1/4
bounded below by 0, it converges to some limit 3 3
L.
to obtain the desired limit.
– Taking limits in the equation yields L = eL − eL ,
whence L = 0. B3 Since S is finite, we can choose three points A, B,C in
– Since L = 0, the sequence {exn } converges to 1. S so as to maximize the area of the triangle ABC. Let
A0 , B0 ,C0 be the points in the plane such that A, B,C are
We now have a telescoping sum: the midpoints of the segments B0C0 ,C0 A0 , A0 B0 ; the tri-
angle A0 B0C0 is similar to ABC with sides twice as long,
x0 + · · · + xn = (ex0 − ex1 ) + · · · + (exn − exn+1 ) so its area is 4 times that of ABC and hence no greater
= ex0 − exn+1 = e − exn+1 . than 4.
We claim that this triangle has the desired effect; that
By taking limits, we see that the sum x0 + x1 + · · · con- is, every point P of S is contained within the triangle
verges to the value e − 1. A0 B0C0 . (To be precise, the problem statement requires
a triangle of area exactly 4, which need not be the case
B2 We prove that the limit exists for α = 43 , β = 43 .
for A0 B0C0 , but this is trivially resolved by scaling up by
For any given positive integer n, the integers which are a homothety.) To see this, note that since the area of the
closer to n2 than to any other perfect square are the ones triangle PBC is no more than that of ABC, P must lie
in the interval [n2 − n − 1, n2 + n]. The √
number of squar-
√ in the half-plane bounded by B0C0 containing B and C.
ish numbers in this interval is 1 + b n − 1c + b nc. Similarly, P must lie in the half-plane bounded by C0 A0
Roughly speaking, this means that containing C and A, and the half-plane bounded by A0 B0
Z √N containing A and B. These three half-planes intersect
√ 4 precisely in the region bounded by the triangle A0 B0C0 ,
S(N) ∼ 2 x dx = N 3/4 .
0 3 proving the claim.
To make this precise, we use the bounds x−1 ≤ bxc ≤ x, B4 The expected value equals
and the upper
√ and lower Riemann sum estimates for the
integral of x, to derive upper and lower bounds on (2n)!
.
S(N): 4n n!
√ First solution:
b Nc−1 √
S(N) ≥ ∑ (2 n − 1 − 1) Write the determinant of A − At as the sum over permu-
n=1 tations σ of {1, . . . , 2n} of the product
Z b√Nc−2 √
√
≥ 2 x dx − N 2n 2n
0 sgn(σ ) ∏(A − At )iσ (i) = sgn(σ ) ∏(Aiσ (i) − Aσ (i)i );
4 √ √ i=1 i=1
≥ ( N − 3)3/2 − N
3
then the expected value of the determinant is the sum
over σ of the expected value of this product, which we
√ denote by Eσ .
d Ne
√
S(N) ≤ ∑ (2 n + 1) Note that if we partition {1, . . . , 2n} into orbits for the
n=1 action of σ , then partition the factors of the product
Z d√Ne+1 √ accordingly, then no entry of A appears in more than
√
≤ 2 x dx + N + 1 one of these factors; consequently, these factors are in-
0
dependent random variables. This means that we can
4 √ √
≤ ( N + 2)3/2 + N + 1.
3
 5

compute Eσ as the product of the expected values of Taking M = A − At , so that Mi j = Ai j − A ji , we wish

the individual factors. to find the expected value of (2); again, this is the sum
It is obvious that any orbit of size 1 gives rise to the of the expected values of each summand in (2). Note
zero product, and hence the expected value of the cor- that each Mi j with i < j is an independent random vari-
responding factor is zero. For an orbit of size m ≥ 3, able taking the values −1, 0, 1 with probabilities 41 , 12 , 41 ,
the corresponding factor contains 2m distinct matrix en- respectively.
tries, so again we may compute the expected value of Consider first a summand in (2) with α 6= β . Then some
the factor as the product of the expected values of the factor Mi j occurs with exponent 1; since the distribution
individual terms Aiσ (i) − Aσ (i)i . However, the distribu- of Mi j is symmetric about 0, any such summand has
tion of this term is symmetric about 0, so its expected expected value 0.
value is 0.
Consider next a summand in (2) with α = β . This sum-
We conclude that Eσ = 0 unless σ acts with n orbits of mand is a product of distinct factors of the form Mi2j ;
size 2. To compute Eσ in this case, assume without loss from the distributions of the Mi j , we see that the ex-
of generality that the orbits of σ are {1, 2}, . . . , {2n − pected value of each of these terms is 1/2n .
1, 2n}; note that sgn(σ ) = (−1)n . Then Eσ is the ex-
Since the total number of perfect matchings α is
pected value of ∏ni=1 −(A(2i−1)2i − A2i(2i−1) )2 , which is
(2n)!/(2n n!), the expected value of (2) is therefore
(−1)n times the n-th power of the expected value of
(2n)!/(2n n!) · 1/2n = (2n)!/(4n n!), as desired.
(A12 − A21 )2 . Since A12 − A21 takes the values −1, 0, 1
with probabilities 41 , 12 , 41 , its square takes the values 0, 1 B5 It is obvious that for any c > 0, the function f (x) = xc
with probabilities 21 , 12 ; we conclude that has the desired property; we will prove that conversely,
any function with the desired property has this form for
Eσ = 2−n . some c.
Define the function g : (0, ∞) → (0, ∞) given by g(x) =
The permutations σ of this form correspond to un-
log f (ex ); this function has the property that if x, y ∈
ordered partitions of {1, . . . , 2n} into n sets of size 2,
(0, ∞) and 2x ≤ y ≤ 3x, then 2g(x) ≤ g(y) ≤ 3g(x). It
so there are
will suffice to show that there exists c > 0 such that
(2n)! g(x) = cx for all x > 0.
n!(2!)n Similarly, define the function h : R → R given by h(x) =
log g(ex ); this function has the property that if x, y ∈ R
such permutations. Putting this all together yields the and x + log 2 ≤ y ≤ x + log 3, then h(x) + log 2 ≤ h(y) ≤
claimed result. h(x) + log 3. It will suffice to show that there exists c >
Second solution: (by Manjul Bhargava) Note that the 0 such that h(x) = x + c for all x ∈ R (as then h(x) = ec x
matrix A − At is skew-symmetric: for all x > 0).
By interchanging the roles of x and y, we may restate the
(A − At )t = At − A = −(A − At ). condition on h as follows: if x − log 3 ≤ y ≤ x − log 2,
then h(x)−log 3 ≤ h(y) ≤ h(x)−log 2. This gives us the
The determinant of a 2n × 2n skew-symmetric matrix
cases a + b = 0, 1 of the following statement, which we
M is the square of the Pfaffian of M, which is a polyno-
will establish in full by induction on a + b, we deduce
mial of degree n in the entries of M defined as follows.
the following: for any nonnegative integers a, b, for all
Define a perfect matching of {1, . . . , 2n} to be a per-
x, y ∈ R such that
mutation of {1, . . . , 2n} that is the product of n disjoint
transpositions. Then the Pfaffian of M is given by x + a log 2 − b log 3 ≤ y ≤ x + a log 3 − b log 2,
∑ sgn(α)Mi1 , j1 · · · Min , jn (1) we have
α

where the sum is over perfect matchings α = h(x) + a log 2 − b log 3 ≤ h(y) ≤ h(x) + a log 3 − b log 2.
(i1 , j1 ) · · · (in ,jn ), and sgn(α) denotes the sign of the
 To this end, suppose that a + b > 0 and that the claim
permutation i1 j2 i3 j4 ··· (2n−1) 2n
. The determinant is known for all smaller values of a + b. In particular,
1 1 2 2 ··· in jn
of M is then the square of (1), i.e., either a > 0 or b > 0; the two cases are similar, so we
treat only the first one. Define the function
det(M) = ∑ sgn(α) sgn(β )Mi1 , j1 · · · Min , jn Mi01 , j10 · · · Mi0n , jn0
α,β (a + b − 1)t − b(log 2 + log 3)
(2) j(t) = ,
a+b
where the sum is now over ordered pairs
so that
(α = (i1 , j1 ) · · · (in , jn ), β = (i01 , j10 ) · · · (i0n , jn0 ))
j(a log 2 − b log 3) = a log 2 − b log 3,
of perfect matchings. j(a log 3 − b log 2) = (a − 1) log 3 − b log 2,
log 2 ≤ t ≤ log 3 (t ∈ [a log 2 − b log 3, a log 3 − b log 2]).
 6

For t ∈ [a log 2 − b log 3, a log 3 − b log 2] and y = x + t, converges to 2 ln 2 − 1 as N → ∞, and so S1 = 2 ln 2 − 1.

we then have 1 1 1
To compute S2 , write k2n +1 = k2n · 1+1/(k2n ) as the geo-
(−1) m
(a − 1) log 2 − b log 3 ≤ h(x + j(t)) − h(x) ≤ (a − 1) log 3 − b log 2 metric series ∑∞
m=0 km+1 2mn+n , whence
log 2 ≤ h(y) − h(x + j(t)) ≤ log 3
∞ ∞
(−1)k+m−1
∞
S2 = ∑∑∑ .
and thus the desired inequalities. k=1 n=1 m=0 k
m+2 2mn+n

Now fix two values x, y ∈ R with x ≤ y. Since log 2

and log 3 are linearly independent over Q, the fractional (This step requires n ≥ 1, as otherwise the geometric
parts of the nonnegative integer multiples of log 3/ log 2 series would not converge for k = 0.) Now note that
are dense in [0, 1). (This result is due to Kronecker; this triple sum converges absolutely: we have
a stronger result of Weyl shows that the fractional parts ∞
1 1 1
are uniformly distributed in [0, 1).) In particular, for any ∑ = 2 n·
ε > 0 and any N > 0, we can find integers a, b > N such m=0 k m+2 2mn+n k 2 1 − k21n
that 1 1
= n
≤ 2 n−1
k(k2 − 1) k 2
y − x < a log 3 − b log 2 < y − x + ε.
and so
By writing
∞ ∞ ∞
∞ ∞
1 1
log 2 ≤
∑ ∑ ∑ km+2 2mn+n ∑ ∑ k2 2n−1
a log 2 − b log 3 = (a log 3 − b log 2) k=1 n=1 m=0 k=1 n=1
log 3 ∞
2
(log 3)2 − (log 2)2 = ∑ k2 < ∞.
−b , k=1
log 3
we see that this quantity tends to −∞ as N → ∞; in par- Thus we can rearrange the sum to get
ticular, for N sufficiently large we have that a log 2 − ! !
∞ ∞ ∞ k−1
b log 3 < y − x. We thus have h(y) ≤ h(x) + a log 2 − 1 (−1)
b log 3 < y − x + ε; since ε > 0 was chosen arbitrarily, S2 = ∑ (−1)m ∑ mn+n ∑ m+2 .
m=0 n=1 2 k=1 k
we deduce that h(y) − h(x) ≤ y − x. A similar argu-
ment shows that h(y) − h(x) ≥ y − x; we deduce that The sum in n is the geometric series
h(y) − h(x) = y − x, or equivalently h(y) − y = h(x) − x.
In other words, the function x 7→ h(x) − x is constant, as 1 1
= .
desired. 2m+1 (1 − 1
) 2m+1 − 1
2m+1
B6 Let S denote the desired sum. We will prove that S = 1. If we write the sum in k as S3 , then note that
First solution: Write
∞ ∞
1 1 1 ∞ 1
∞
1 1 ∞
1 ∑ km+2 = S3 + 2 ∑ m+2
= S3 + ∑ km+2
∑ k2n + 1 = k + 1 + ∑ k2n + 1 ; k=1 k=1 (2k) 2m+1 k=1
n=0 n=1
(where we can rearrange terms in the first equality be-
then we may write S = S1 + S2 where cause all of the series converge absolutely), and so
∞
(−1)k−1
  ∞
1 1
S1 = ∑ k(k + 1) S3 = 1 − m+1 ∑ m+2 .
k=1 2 k=1 k
∞
(−1)k−1 ∞ 1
S2 = ∑ ∑ k2n + 1 . It follows that
k=1 k n=1
(−1)m ∞ 1
∞
The rearrangement is valid because both S1 and S2 con- S2 = ∑ m+1 ∑ km+2
m=0 2 k=1
verge absolutely in k, by comparison to ∑ 1/k2 . ∞ 
1 m
∞ 
1
To compute S1 , note that =∑ 2 ∑ −
k=1 2k m=0 2k
N N
(−1)k−1
  ∞
1 1 1
∑ k(k + 1) = ∑ (−1)k−1 −
k k+1
= ∑ k(2k + 1)
k=1 k=1 k=1
N ∞  
(−1)N (−1)k−1 1 1
= −1 + +2 ∑ =2∑ − = 2(1 − ln 2).
N +1 k=1 k k=1 2k 2k + 1
 7

Finally, we have S = S1 + S2 = 1. where k runs over all positive integers for which m =
Second solution: (by Tewodros Amdeberhan) Since k2n +1 for some n. If we write e for the 2-adic valuation
R1 t 1 of m − 1 and j = (m − 1)2−e for the odd part of m − 1,
0 x dx = 1+t for any t ≥ 1, we also have then the values of k are j2i for i = 0, . . . , e. The inner
sum can thus be evaluated as
(−1)k−1
∞ ∞ Z 1
n
S= ∑∑ xk2 dx. e
k=1 n=0 k 0 1 1 1 1
−∑ i = e = .
j i=1 2 j 2 j m − 1
Again by absolute convergence, we are free to permute
the integral and the sums: We thus have
∞ ∞  
(−1)k−1 k2n 1 1 1
Z 1 ∞ ∞
S= dx ∑ ∑ x S= ∑ = ∑ − = 1.
0 n=0 k=1 k m=2 m(m − 1) m=2 m − 1 m
Z 1 ∞
n
=−
0
dx ∑ log(1 + x2 ). Fourth solution: (by Liang Xiao) Let S0 and S1 be the
n=0 sums ∑k 1k ∑∞ 1
n=0 k2n +1 with k running over all odd and
Due to the uniqueness of binary expansions of nonneg- all even positive integers, respectively, so that
ative integers, we have the identity of formal power se-
S = S0 − S1 .
ries
In S1 , we may write k = 2` to obtain
∞
1 n
= ∏ (1 + x2 ); ∞
1 ∞
1
1 − x n=0 S1 = ∑ 2` ∑ `2n+1 + 1
`=1 n=0
the product converges absolutely for 0 ≤ x < 1. We thus 1 ∞
1
have = (S0 + S1 ) − ∑
2 `=1 2`(` + 1)
Z 1
1 1
S=− log(1 − x) dx = (S0 + S1 ) −
0 2 2
= ((1 − x) log(1 − x) − (1 − x))10 because the last sum telescopes; this immediately yields
= 1. S = 1.

Third solution: (by Serin Hong) Again using absolute

convergence, we may write
∞
1 (−1)k−1
S= ∑ m∑
m=2 k k