Shortlisted Problems

20th ELMO
Pittsburgh, PA, 2018
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

Note of Confidentiality
The shortlisted problems should be kept strictly confidential until disclosed pub-
licly by the committee on the ELMO.

Contributing Students
The Problem Selection Committee for ELMO 2018 thanks the following proposers
for contributing 90 problems to this year’s Competition:

Adam Ardeishar, Andrew Gu, Ankan Bhattacharya, Bran-

don Wang, Carl Schildkraut, Daniel Hu, Daniel Liu, Eric
Gan, Kevin Ren, Krit Boonsiriseth, Luke Robitaille, Michael
Kural, Michael Ma, Michael Ren, Sam Ferguson, Tristan
Shin, Vincent Bian, Vincent Huang, Zack Chroman

Problem Selection Committee

The Problem Selection Committee for ELMO 2018 was led by Evan Chen and
consisted of:

• Andrew Gu

• Daniel Liu

• James Lin

• Michael Ma

• Michael Ren

• Mihir Singhal

• Zack Chroman

2
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

Contents
1 Problems 4
1.1 Problem A1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Problem A2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Problem A3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Problem A4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Problem C1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.6 Problem C2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.7 Problem C3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.8 Problem G1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.9 Problem G2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.10 Problem G3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.11 Problem G4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.12 Problem G5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.13 Problem N1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.14 Problem N2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.15 Problem N3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.16 Problem N4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Solutions 7
2.1 Solution A1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Solution A2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Solution A3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Solution A4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5 Solution C1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6 Solution C2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.7 Solution C3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.8 Solution G1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.9 Solution G2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.10 Solution G3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.11 Solution G4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.12 Solution G5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.13 Solution N1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.14 Solution N2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.15 Solution N3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.16 Solution N4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

3
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

Problems
A1. Let f : R → R be a bijective function. Does there always exist an infinite number
of functions g : R → R such that f (g(x)) = g(f (x)) for all x ∈ R?

(Daniel Liu)

A2. Let a1 , a2 , . . . , am be a finite sequence of positive integers. Prove that there exist
nonnegative integers b, c, and N such that
$m %
X√ j√ k
n + ai = bn + c
i=1

holds for all integers n > N .

(Carl Schildkraut)

A3. Let a, b, c, x, y, z be positive reals such that 1

x + 1
y + 1
z = 1. Prove that

4abcxyz
ax + by + cz ≥ .
(x + y + z − 3)2

(Daniel Liu)

A4. Elmo calls a monic polynomial with real coefficients tasty if all of its coefficients
are in [−1, 1]. A monic polynomial P with real coefficients and complex roots χ1 , . . . , χm
(counted with multiplicity) is given to Elmo, and he discovers that there does not exist
a monic polynomial Q with real coefficients such that P · Q is tasty. Find all possible
values of max (|χ1 |, . . . , |χm |).

(Carl Schildkraut)

C1. Let n be a positive integer. There are 2018n + 1 cities in the Kingdom of Sellke
Arabia. King Mark wants to build two-way roads that connect certain pairs of cities
such that for each city C and integer 1 ≤ i ≤ 2018, there are exactly n cities that are a
distance i away from C. (The distance between two cities is the least number of roads
on any path between the two cities.)
For which n is it possible for Mark to achieve this?

(Michael Ren)

C2. We say that a positive integer n is m-expressible if one can write a expression
evaluating to n in base 10, where the expression consists only of

• exactly m numbers from the set {0, 1, . . . , 9}

• the six operations +, −, ×, ÷, exponentiation ∧ , concatenation ⊕, and

• some number (possibly zero) of left and right parentheses.

4
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

For example, 5625 is 3-expressible (in two ways), as 5625 = 5 ⊕ (5∧ 4) = (7 ⊕ 5)∧ 2, say.
Does there exist a positive integer A such that all positive integers with A digits are
(A − 1)-expressible?

(Krit Boonsiriseth)

C3. A windmill in the plane consists of a line segment of unit length with a distinguished
endpoint, the pivot. Geoff has a finite set of windmills,
√ such that no two windmills intersect,
and any two pivots are distance more than 2 apart. In an operation, Geoff can choose
a windmill and rotate it about its pivot, either clockwise or counterclockwise and by any
amount, as long as no two windmills intersect during or after the rotation. Show that
Geoff can, in finitely many operations, rotate the windmills so that they all point in the
same direction.

(Michael Ren)

G1. Let ABC be an acute triangle with orthocenter H, and let P be a point on the
nine-point circle of ABC. Lines BH, CH meet the opposite sides AC, AB at E, F ,
respectively. Suppose that the circumcircles of 4EHP and 4F HP intersect lines CH,
BH a second time at Q, R, respectively. Show that as P varies along the nine-point
circle of ABC, the line QR passes through a fixed point.

(Brandon Wang)

G2. Let ABC be a scalene triangle with orthocenter H and circumcenter O. Let P be
the midpoint of AH and let T be on line BC with ∠T AO = 90◦ . Let X be the foot of
the altitude from O onto line P T . Prove that the midpoint of P X lies on the nine-point
circle of 4ABC.

(Zack Chroman)

G3. Let A be a point in the plane, and ` a line not passing through A. Evan doesn’t
have a straightedge, but instead has a special compass which has the ability to draw a
circle through three distinct noncollinear points. (The center of the circle is not marked
in this process.) Additionally, Evan can mark the intersections between two objects
drawn, and can mark an arbitrary point on a given object or on the plane.

(i) Can Evan construct the reflection of A over `?

(ii) Can Evan construct the foot of the altitude from A to `?

(Zack Chroman)

G4. Let ABCDEF be a convex hexagon inscribed in a circle Ω such that triangles
ACE and BDF have the same orthocenter. Suppose that BD and DF intersect CE
at X and Y , respectively. Show that there is a point common to Ω, the circumcircle of
DXY , and the line through A perpendicular to CE.

(Michael Ren and Vincent Huang)

5
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

G5. Let scalene triangle ABC have altitudes AD, BE, CF and circumcenter O. The
circumcircles of 4ABC and 4ADO meet at P 6= A. The circumcircle of 4ABC meets
lines P E at X 6= P and P F at Y 6= P . Prove that XY k BC.

(Daniel Hu)

N1. Determine all nonempty finite sets S = {a1 , . . . , an } of n distinct positive integers
such that a1 · · · an divides (x + a1 ) · · · (x + an ) for every positive integer x.

(Ankan Bhattacharya)

N2. Call a number n good if it can be expressed in the form 2x + y 2 where x and y are
nonnegative integers.

(a) Prove that there exist infinitely many sets of 4 consecutive good numbers.

(b) Find all sets of 5 consecutive good numbers.

(Michael Ma)

N3. Let a1 , a2 , . . . be an infinite sequence of positive integers satisfying a1 = 1 and

an | ak + ak+1 + · · · + ak+n−1

for all positive integers k and n. Find the maximum possible value of a2018 .

(Krit Boonsiriseth)

N4. Fix a positive integer n > 1. We say a nonempty subset S of {0, 1, . . . , n − 1} is

d-coverable if there exists a polynomial P with integer coefficients and degree at most d,
such that S is exactly the set of residues modulo n that P attains as it ranges over the
integers.
For each n, determine the smallest d such that any nonempty subset of {0, . . . , n − 1}
is d-coverable, or prove that no such d exists.

(Carl Schildkraut)

6
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

Solutions
A1. Let f : R → R be a bijective function. Does there always exist an infinite number
of functions g : R → R such that f (g(x)) = g(f (x)) for all x ∈ R?

(Daniel Liu)

Yes. It’s clear f 0 , f 1 , f 2 , . . . all commute with f . If f doesn’t have finite order this
collection is infinite and valid.
Else, suppose that f n = id, where n is minimal. If n = 1 the problem is clear, so suppose
n > 1. Then f is composed of some cycles; some cycle length d | n appears infinitely
many times. Let a countable number of these cycles be xr,1 → xr,2 → · · · → xr,d → xr,1
for r ∈ Z.
Then for every integer s, create a new function hs fixing everything except the xk,` ,
and send every xr,a → xr+s,a . It is clear that every hs commutes with f .
This gives infinitely many g, unless all but finitely many of the cycles have length
1. In that case, we can find more g by swapping any two fixed points of f and leaving
everything else intact.

7
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

A2. Let a1 , a2 , . . . , am be a finite sequence of positive integers. Prove that there exist
nonnegative integers b, c, and N such that
$m %
X√ j√ k
n + ai = bn + c
i=1
holds for all integers n > N .
(Carl Schildkraut)

P √ √
If all the ai are equal, then m i=1 n + ai = m2 n + m2 a1 and so (b, c) = (m2 , m2 a1 )
works fine.
Let us assume this is not the case. Instead, will take b = m2 and c = m(a1 +· · ·+am )−1
and claim it works for N large enough.
On the one hand,
m r
X √ a1 + · · · + am
n + ai < m · n +
m
i=1
p lp m jp k
= m2 · n + c + 1 ≤ m2 · n + c + 1 ≤ m2 · n + c + 1.
c
On the other hand, let λ = < 12 . We use the following estimate.
2(c+1)
√ √ λa
Claim. If n is large enough in terms of (a1 , . . . , an ) then n + ai ≥ n + √i.
n
λ2 a2i
Proof. Squaring both sides, it’s equivalent to ai ≥ 2λ · ai + n , which holds for n big
enough as 2λ < 1.
Now,
m m 
√ √

X X λai
n + ai ≥ n+ √
n
i=1 i=1
√ λ · (a1 + · · · + an )
≥m n+ √
n
√ λ · (c + 1)
=m n+ √
m n
√ c p jp k
=m n+ √ > m2 · n + c ≥ m2 n + c .
2m n
This finishes the problem.
Remark. Obviously, b = m2 for asymptotic reasons (by taking n large). As for possible
values of c:
• If a1 = · · · = am , then one can show c = m(a1 + · · · + am ) is the only valid choice.
2
Indeed, taking n of the form n = k 2 − a and n = km−1 2 − a is enough to see this.

• But if not all ai are equal, the natural guess of taking c = m(a1 + · · · + an ) is not
valid in general. For example, we have that
√
n ∈ {t2 − 1 | t = 2, 3, . . . }.
√  √ 
n + n + 2 6= 4n + 4
I think one can actually figure out exactly which c are valid, though the answer will
depend on some quadratic residues, and we do not pursue this line of thought here.
So any correct solutions must distinguish these two cases.

8
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

A3. Let a, b, c, x, y, z be positive reals such that 1

x + 1
y + 1
z = 1. Prove that

4abcxyz
ax + by + cz ≥ .
(x + y + z − 3)2

(Daniel Liu)

We present three solutions.

First solution, proof without words (by proposer)

ax ay az
ax + by + cz = yz · + zx · + xy ·
yz zx xy
 x yz  y zx  z xy  1
a b c xy+yz+zx
≥ (xy + yz + zx)
yz zx xy
xyz
(abc) xy+yz+zx
= (xy + yz + zx) · xy+zx yz+xy zx+yz
x xy+yz+zx y xy+yz+zx z xy+yz+zx
xyz
(abc) xy+yz+zx
≥ (xy + yz + zx) ·  xy+zx yz+xy zx+yz
2
x· xy+yz+zx +y· xy+yz+zx +z· xy+yz+zx
2
xyz
4(abc) xy+yz+zx
= (xy + yz + zx) · P  2
yz
cyc x · 1 − xy+yz+zx
4abc(xy + yz + zx)
= xyz
(x + y + z − 3 xy+yz+zx )2
4abcxyz
= .
(x + y + z − 3)2

Second solution, by weighted AM-GM (Andrew Gu) By weighted AM-GM,

1 1 1 1 1 1
· xax + · yby + · zcz ≥ x x y y z z abc.
x y z
Hence it suffices to show
1 1 1 4xyz
xx y y z z ≥ .
(x + y + z − 3)2
By weighted AM-GM,
1 1 1
(1− y1 ) 1 1 1
2x 2 (1− x ) y 2 z 2 (1− z ) ≤ 2 · (x − 1 + y − 1 + z − 1) = x + y + z − 3.
2
Squaring both sides and rearranging proves the required inequality.

Third solution, by Hölder and Schur/Muirhead (Evan Chen) By Hölder and weighted
AM-GM we have
s 
1 1 1 1 1 1
2
+ 2 + 2 (ax + by + cz ) ≥ · ax/2 + · by/2 + · cz/2 ≥ (abc)1/2 .
x y z x y z

9
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

Hence, it suffices to prove that

1 1 1
(x + y + z − 3)2 ≥ 4xyz 1/x2 + 1/y 2 + 1/z 2


∀ + + =1
x y z
which is a 3-variable symmetric inequality. It also happens to be is MOP 2011, K4.1,
done in my SOS handout. We give a proof below (with a = 1/x, etc).

Claim (Black MOP 2011, Test 4, Problem 1). If a, b, c > 0 then

2
(a + b + c)(a2 + b2 + c2 )
    
1 1 1
(a + b + c) + + −3 ≥4
a b c abc

Proof. Expanding and clearing denominators it’s just

X X X X
a4 b2 + a3 b3 + 6a2 b2 c2 ≥ 2 a4 bc + 2 a3 b2 c
sym cyc cyc sym

which can also be written as

0
0 0
1 -2 1
2 -2 -2 2
1 -2 6 -2 1
0 -2 -2 -2 -2 0
0 0 1 2 1 0 0

in Chinese dumbass notation. This rewrites as

X X
a4 (b − c)2 + 2 ab(ab − bc)(ab − ac) ≥ 0
cyc cyc

which is evident (the latter sum is “upsidedown triangle Schur”).

10
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

A4. Elmo calls a monic polynomial with real coefficients tasty if all of its coefficients
are in [−1, 1]. A monic polynomial P with real coefficients and complex roots χ1 , . . . , χm
(counted with multiplicity) is given to Elmo, and he discovers that there does not exist
a monic polynomial Q with real coefficients such that P · Q is tasty. Find all possible
values of max (|χ1 |, . . . , |χm |).

(Carl Schildkraut)

We claim the answer is r > 1.

Part 1: Any value of r > 1 can be achieved. To prove this, we will show that the
polynomial
P (x) = xn − rn
has no tasty multiples if rn ≥ 2 (such an n exists because r > 1). Set M = rn . Assume
we have a polynomial
XN
R(x) = ai xi
i=0

so that −1 ≤ ai ≤ 1 for all i (aN = 1) and P |R. Taking R modulo P , we get that, with
N = bn + c and 0 ≤ c < n (setting ak = 0 if k > N ),
n−1 b n−1
" b #
XX X X
R(x) = akn+j xkn+j ≡ xj akn+j Rk .
j=0 k=0 j=0 k=0

We have this must be the zero polynomial (since P |R); specifically, taking j = c,
b
X
ank+c Rk = 0
k=0
b−1
X
(−ank+c )Rk = abn+c Rb
k=0
b−1
X
|ank+c |Rk ≥ Rb
k=0

(since abn+c = aN = 1). However, since |ank+c | ≤ 1, we then have

b−1
X
Rk ≥ Rb
k=0
Rb − 1
≥ Rb
R−1
Rb − 1 ≥ Rb+1 − Rb
Rb (2 − R) ≥ 1.

However, as R ≥ 2, this is false.

11
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

Part 2: Any polynomial with r ≤ 1 has a tasty multiple. Define the sparsity of a
polynomial to be the greatest common divisor of the exponents m for which the coefficient
of xm in P is not zero. Equivalently, it is the largest integer d so that P (x) = Q(xd ) for
some polynomial Q.
We prove the following theorem:
Theorem. Given any complex number z for which |z| ≤ 1, there exist tasty polynomials
with z as a root that have arbitrarily large sparsities.
Proof. Let z = reiθ . If θ is a rational multiple of π (say, θ = aπ/b), then we take the
polynomial xbn −rbn for any integer n; this has sparsity bn and is tasty (as r ≤ 1, rbn ≤ 1).
So, it suffices to prove this in the case where θ is not a rational multiple of π, and we
henceforth assume this.
We claim that, for infinitely many n, the polynomial

x2n − 2 cos (nθ) rn xn + r2n

is tasty (note that this polynomial has sparsity n and as such the theorem is implied by
this claim). First note that this polynomial reduces to
 n
xn = rn e±niθ = re±iθ ,

which is true at x = reiθ = z, so z is in fact a root.

We recall the following lemma:
Lemma. For any real number φ which is not a rational multiple of π, the sequence
an = cos (nφ) has infinitely many terms in the range [−1/2, 1/2].
Indeed, let {x} be the fractional part of x, and consider the sequence
 
nφ
αn = .
2π
We see that −1/2 ≤ an ≤ 1/2 iff 1/6 ≤ αn ≤ 1/3 or 2/3 ≤ αn ≤ 5/6. It is well known
that the sequence xn = {nx} is dense in [0, 1] for any irrational x, so this is true. Thus,
for infinitely many n, as θ has been assumed not to be a rational multiple of π, the
coefficients of P are bounded above in absolute value by rn and r2n for infinitely many
n, both of which are ≤ 1 as r ≤ 1.

We now provide a second lemma.

Lemma. If P (x) and Q(x) are both tasty polynomials and the sparsity D of P is greater
than the degree d of Q, then the product R(x) = P (x)Q(x) is also tasty.
Proof. Write
s
X d
X
P (x) = aj xDj , Q(x) = bk xk .
j=0 k=0

Then,
s X
X d
P (x)Q(x) = aj bk xDj+k .
j=0 k=0

As D > d, none of these terms interfere with one another (for each integer n, there is at
most one choice of 0 ≤ j ≤ s, 0 ≤ k ≤ d so that Dj + k = s), so the coefficients of R(x)
are just the values of aj bk as j and k range over the desired range; as each aj and bk are
of magnitude ≤ 1, each pairwise product is as well, finishing the proof.

12
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

Given a polynomial P with roots χ1 , . . . , χm in C (possibly with duplicates), we will

inductively construct the polynomial R(x) that is tasty and that P divides. We define a
sequence of polynomials R0 , . . . , Rm so that R0 (x) = 1, and for each 0 < k ≤ m, we take
a tasty polynomial Pk (x) with root χk and sparsity greater than the degree of Rk−1 , and
take Rk (x) = Rk−1 (x)Pk (x). Such a Pk (x) is guaranteed to exist by our theorem, and
the product Rk−1 (x)Pk (x) is guaranteed to be tasty by our lemma. Thus, we may take
R = Rm , finishing the proof.

Remark. A polynomial P that has a tasty multiple exists for all r < 2: We have upon
fixing r < 2 that for large enough n, we know rn − rn−1 − · · · − r − 1 ≤ 0. If n is minimal,
rn − rn−1 − · · · − r > 0, and we can thus take some value 0 ≤ c ≤ 1 for the constant term
by the intermediate value theorem so that P (x) = xn − xn−1 − · · · − x − c has a root at
r. If r ≥ 2, then n = 1 can be taken in Part 1 and thus no tasty multiples exist.

13
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

C1. Let n be a positive integer. There are 2018n + 1 cities in the Kingdom of Sellke
Arabia. King Mark wants to build two-way roads that connect certain pairs of cities
such that for each city C and integer 1 ≤ i ≤ 2018, there are exactly n cities that are a
distance i away from C. (The distance between two cities is the least number of roads
on any path between the two cities.)
For which n is it possible for Mark to achieve this?

(Michael Ren)

The answer is n even.

To see that n odd fails, note that by taking i = 1 we see the graph is n-regular; since
it has an odd number of vertices we need n to be even.
On the other hand, if n is even, then consider the graph formed by taking the vertices
of a regular (2018n + 1)-gon and drawing edges between vertices which are at most n/2
apart. Then this works.

14
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

C2. We say that a positive integer n is m-expressible if one can write a expression
evaluating to n in base 10, where the expression consists only of

• exactly m numbers from the set {0, 1, . . . , 9}

• the six operations +, −, ×, ÷, exponentiation ∧ , concatenation ⊕, and

• some number (possibly zero) of left and right parentheses.

For example, 5625 is 3-expressible (in two ways), as 5625 = 5 ⊕ (5∧ 4) = (7 ⊕ 5)∧ 2, say.
Does there exist a positive integer A such that all positive integers with A digits are
(A − 1)-expressible?

(Krit Boonsiriseth)

Here is a solution by Evan Chen achieving A = 6 · 106 , and reprising the joke “six
consecutive zeros”.
We will replace “exactly m numbers” with “at most m numbers”, since this is the
same. Suppose we group the digits of N into base 1000000, so that we have

N = s1 s2 s3 . . . sm

where each sm is a group of six digits (s1 padded with leading zeros, if needed, but
s1 6= 000000). We consider two cases.

• Suppose some group is zero; then we find that N has six consecutive zeros in its
decimal representations. Thus N has the form

N = X ⊕ (b · (1 ⊕ 0)∧ 6) ⊕ Y

for some strings X and Y (possibly empty), which are formed by repeated concate-
nation.

• Otherwise, note that m ≥ 106 . By a classical pigeonhole argument there exist

1
indices i < j such that si + · · · + sj ≡ 0 (mod 999999). Let n = 999999 si . . . sj .
Then we can write
N = X ⊕ ((1 ⊕ 0)∧ 6 − 1) · n ⊕ Y
 

for strings X = s1 . . . si and Y = sj+1 . . . sn .

Remark (Possible motivational remarks). Ankan Bhattacharya says: I knew that the
answer had to be yes — the obvious counting argument to show answer no doesn’t work,
and the given elements are unrelated enough that proving a no answer would be very
difficult.
Evan says: I think you really do have to use exponentiation, since otherwise the
numbers aren’t big enough; but exponentiation is really painful to deal with, so I tried
to find a way to use it only once. This is less daunting than it seems because you can
concatenate digits “for free” from a size perspective; thus you just need a substring that
you can “save space” on. After a bit of guesswork I came upon the idea of taking modulo
106 − 1 = 999999 (which saves about two digits) and from there I had it.

15
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

Remark (Possible motivational remarks). Ankan Bhattacharya points out that if we fix
all N − 2 operations, then there are only 10N −1 choices, compared to 9 · 10N −1 numbers
we need to obtain. Thus we need to use different operations to reach different numbers.
This suggests that all solutions are likely to use some amount of casework.
Unlike Ankan, I did not find the case split to be a substantial part of the problem. It
came up naturally because I had an edge case where six consecutive zeros might appear
in my argument, and the first case was patch-only in that situation.

16
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

C3. A windmill in the plane consists of a line segment of unit length with a distinguished
endpoint, the pivot. Geoff has a finite set of windmills,
√ such that no two windmills intersect,
and any two pivots are distance more than 2 apart. In an operation, Geoff can choose
a windmill and rotate it about its pivot, either clockwise or counterclockwise and by any
amount, as long as no two windmills intersect during or after the rotation. Show that
Geoff can, in finitely many operations, rotate the windmills so that they all point in the
same direction.
(Michael Ren)

Throughout the solution we will general denote pivots by P , Q, R, . . . and non-pivots

by A, B, C, . . .
We say that a configuration of windmills around S is admissible if no two windmills
intersect. The problem is equivalent to showing one can reach any admissible configuration
from any other (and the final position with the windmills pointing the same direction is
just one example of a clearly admissible configuration).
Draw a red line segment between any two pivots which have distance at most 2 (thus
these windmills could intersect). This naturally gives us a graph G.
√
Lemma. For c ≥ 2, the graph G is planar.
Proof. Indeed, if P A and QB intersect, we can consider convex quadrilateral P QAB, one
of whose angles is at least 90◦ . WLOG it is ∠P QA, in which case P A2 ≥ P Q2 + QA2 >
2 + 2 = 4, so P A should not be red.

Clearly, we can ignore any isolated vertices. We can also ignore any leaves in G; indeed
suppose P is a pivot with P Q the only red edge. Then we can rotate the windmill at P
to point away from Q and it will never obstruct other windmills since c ≥ 1, so we can
delete the pivot P from consideration (and use induction on the number of pivots, say).
Thus, we may assume G is a finite planar graph with no leaves. Thus it makes sense
to speak of the faces of planar graph G, consisting of several polygons.
Lemma. A windmill with pivot P can never intersect a red edge other than those touching
P.
Proof. Suppose windmill P A intersects red edge QR. Then the altitude from P H
to QR has length at most 1. WLOG that QH < RH, so QH < 12 QR = 1. Then
P Q2 < QH 2 + HP 2 < 1 + 1 = 2, contradiction.

17
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

From now on, a windmill P A is said to hug a red edge P Q if the angle ∠QP A < ε
for some sufficiently small ε in terms of G; each red edge P Q has at most two windmills
hugging it (namely the windmills with pivots P and Q; if this happens, the windmills
are on opposite sides of P Q). Call a windmill configuration cuddly if every windmill is
hugging an edge.

Claim. We can reach some cuddly configuration from any admissible one.

Proof. Indeed, consider a windmill P A not hugging any edge, and an edge P Q, and such
that ∠AP Q = θ is minimal among all such pairs. Let ∠RP Q be the corresponding angle
of the face containing P A, and let QB, RC be windmills.
If QB is hugging P Q, we perturb it slightly so that A and B are on opposite sides of
P Q; thus QB is no longer in the way.
We rotate P A towards P Q now. Because we assumed θ = ∠AP Q was minimal, it is
impossible for the body of the windmill to collide with the points B or C. So the only
way it can be obstructed is if the point A collides with the interior of QB or RC.
R

C
B

θ
P Q

Suppose that A collided with QB. At the moment of collision, we would have to have
∠P AQ ≤ 90◦ . (This is because just before the collision P A was still disjoint from QB,
and if ∠P AQ ≥ 90◦ just before then it would remain disjoint as P A rotated.) But then
P Q2 ≤ P A2 + AQ2 ≤ 2, contradiction. A similar proof works for RC.
Thus we can rotate the windmills one by one so they hug the edges, as desired.

It remains to show any two cuddly configurations can be reached from each other. For
this, we make two observations.

• Suppose P A and QB both hug P Q. We show we can interchange the two. Assume
∠RP Q is the angle of a face containing A, and ∠T P Q, ∠P QS are the angles of
the face containing B.

A Q

P B S

18
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

Rotate P A so it hugs P R (possibly perturbing the windmill at R), and then rotate
QB so it hugs QS (possibly perturbing the windmill at S). Then rotate P A so it
hugs P T , then move QB back so it hugs P Q from the other side, and rotate P A
back.

• Now suppose P A hugs P Q, and ∠RP Q is the angle of a face containing A. Then
we can rotate it so that P A hugs P R (here P A could be blocked by QB initially,
but then we perform the switching operation above).

Together these two observations finish the problem.

√
Remark (Michael Ren). Here is a solution achieving just c = 3.
Draw a disk of radius 1 +  around
√ every point in S such that the distance between
any two points in S is more than 3(1 + ) for some  > 0 that clearly exists. Note that
no three disks can intersect. Indeed, if disks centered at A, B, and C intersected, then
the circumradius of√ABC is at most 1 + , which means that some two of A, B, C are at
most a distance of 3(1 + ) apart. In light of this, for any two points A and B in S that
are a distance of at most 2 apart, draw a rhombus AP BQ of length 1 + . By our work
before, all such rhombi are distinct. Furthermore, windmill collisions only happen inside
these rhombi by definition. Now, have Geoff move each of his windmills one by one to
Sasha’s windmills. If a windmill collision happens, have Geoff move the other windmill
out of the way inside the rhombus before moving the windmill by and then restore the
position of the other windmill. Hence, he can always get his windmills to coincide, as
desired.

19
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

G1. Let ABC be an acute triangle with orthocenter H, and let P be a point on the
nine-point circle of ABC. Lines BH, CH meet the opposite sides AC, AB at E, F ,
respectively. Suppose that the circumcircles of 4EHP and 4F HP intersect lines CH,
BH a second time at Q, R, respectively. Show that as P varies along the nine-point
circle of ABC, the line QR passes through a fixed point.

(Brandon Wang)

Let D denote the foot of the A-altitude, and M the midpoint of BC. We claim that
R and Q both lie on line P M . That will solve the problem (M is the fixed point).

P
E

R
F
H
Q

B D M C

By angle chasing, it is not hard to show that

]F HE = ]F EM.

Now,
]F P R = ]F HR = ]F HE = ]F EM = ]F P M
as desired so P , R, M are collinear. Similarly, P , Q, M are collinear, as desired.

20
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

G2. Let ABC be a scalene triangle with orthocenter H and circumcenter O. Let P be
the midpoint of AH and let T be on line BC with ∠T AO = 90◦ . Let X be the foot of
the altitude from O onto line P T . Prove that the midpoint of P X lies on the nine-point
circle of 4ABC.

(Zack Chroman)

We present two solutions, one synthetic and by complex numbers.

A

X
P

O
H
T B C
M

First solution (Zack Chroman) Let M be the midpoint of BC. Note that since
AP ⊥ BC and AT ⊥ AO k P M , we find that P is the orthocenter of 4AT M . Thus
Y = T P ∩ AM satisfies ∠P Y M = 90, so it lies on the 9-point circle.
It then suffices to note that the reflection X 0 of P over Y lies on the circumcircle of
(AM T ) = (T O), so ∠T X 0 O = 90 =⇒ X = X 0 .

Second solution (complex numbers, Evan Chen) Let Q denote the reflection of P
over M , the midpoint of BC.

Claim. We have QO ⊥ P T .

Proof. By complex numbers. We have

aa(b + c) − bc(a + a) a2 (b + c) − 2abc

t= =
aa − bc a2 − bc
2
 
a (b + c) − 2abc b+c
t−p= − a+
a2 − bc 2
a ( 2 b + 2 c − a) + (−a + 2 b + 12 c)bc
2 1 1 1
=
a2 − bc
2
b + c − 2a a + bc
= · 2
2 a − bc
b+c b + c − 2a
q =2· −p=
2 2
a2 +bc
Since a2 −bc
∈ iR, the claim is proven.

Thus, QOX are collinear. By considering right triangle 4P QX with midpoint M , we

conclude that M X = M P . Since the nine-point circle is the circle with diameter P M , it
passes through the midpoint of P X.

21
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

G3. Let A be a point in the plane, and ` a line not passing through A. Evan doesn’t
have a straightedge, but instead has a special compass which has the ability to draw a
circle through three distinct noncollinear points. (The center of the circle is not marked
in this process.) Additionally, Evan can mark the intersections between two objects
drawn, and can mark an arbitrary point on a given object or on the plane.

(i) Can Evan construct the reflection of A over `?

(ii) Can Evan construct the foot of the altitude from A to `?

(Zack Chroman)

The trick is to invert the figure around a circle centered at A of arbitrary radius. We
let ω = `∗ denote the image of ` under this inversion. Then, under the inversion, Evan’s
compass has the following behavior:

• Evan can draw a line through two points other than A; or

• Evan can draw a circle through three points other than A.

In other words, the point A is “invisible” to Evan, but Evan otherwise has a straightedge
and the same compass.
It is clear then that the answer to (ii) is no.
Part (i) is equivalent to showing that Evan can construct the center of ω; we give one
construction here anyways. Take any cyclic quadrilateral W XY Z inscribed in ω, and
let P = W Z ∩ XY . Then the circumcircles of 4P W X and 4P Y Z meet again at the
Miquel point M , and the second intersection of (M XZ) and (M W Y ) is the center of ω.

Remark. The proof of (ii) implies that it’s actually more or less impossible in this
context to construct any point other than the reflection of A, as a function of A and `.
An alternative proof of (ii) is possible by inverting around a generic point P on ` with
radius P A; this necessarily preserves the entire construction, but the foot from A to ` is
not fixed by this inversion.

22
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

G4. Let ABCDEF be a convex hexagon inscribed in a circle Ω such that triangles
ACE and BDF have the same orthocenter. Suppose that BD and DF intersect CE
at X and Y , respectively. Show that there is a point common to Ω, the circumcircle of
DXY , and the line through A perpendicular to CE.

(Michael Ren and Vincent Huang)

We present many, many solutions. In all of them, we let H denote the common
orthocenter.
A

C X Y E

First solution by Simson lines (Vincent Huang) Let AH meet CE and Ω again at M
and A1 , respectively, and P and Q be the projections of A1 onto BD and DF , respectively.
Note that P Q is the Simson line of A1 with respect to BDF . It is well known that this
Simson line bisects the segment between A1 and H. Hence, M lies on P Q. But P , M ,
and Q are respectively the projections of A1 onto DX, XY , and Y D, so A1 must lie on
the circumcircle of DXY , as desired.

Second solution by dual Desargue involution (Michael Ren) Let O and r be the
center and radius of Ω, respectively. Let E be the ellipse with foci O and H consisting
of the set of points P such that OP + HP = r. Note that as the reflections of H over
AC, CE, EA, BD, DF, F B lie on Ω, E is tangent to the sides of ACE and BDF . Let
E and AD meet CE at P and Q, respectively. By the dual of Desargue’s involution
theorem on quadrilateral ACP E with inscribed conic E, D(CE; XY ; P Q) is an involution.
Hence, the circumcircles of DCE, DXY , and DP Q are coaxial, so it suffices to show
that A1 DP Q is cyclic, where A1 is the second intersection of AH and Ω. But note that
A1 lies on OP , so ∠QDA1 = ∠ADA1 = π2 − ∠OA1 A = π2 − ∠P A1 A, which is the angle
between P A1 and P Q by the perpendicularity of AA1 and CE, as desired.

Third solution by angle chasing (Mihir Singhal) Let A1 be the reflection of H over
CE. Note A1 is on Ω so it suffices to show that DA1 XY is cyclic. Let M be the foot of
the altitude from A to CE. Note that M is the midpoint of HA1 so since A1 is on Ω, M
must be on the nine-point circle of DBF . Let R and S be the feet of the altitudes from
F and B in DBF .

23
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

Note M XRH and M Y SH are cyclic. Moreover, M lies on the nine-point circle of
4BDF , and hence ]SM R = 2]SDR. Then

]XHY = ]XHM + ]M HY
= ]XRM + ]M SY = ]DRM + ]M SD
= −(]RM S + ]SDR) = ]SM R + ]RDS
= 2]SDR + ]RDS = ]SDR = ]Y DX.

Thus ]XA1 Y = −]XHY = ]RDS = ]XDY , as needed.

Fourth solution by inversion (James Lin) Let K be the second intersection of Ω and
the perpendicular from A to CE. We want to show DKXY is cyclic. We invert about
H. It’s clear that now, A0 C 0 E 0 and B 0 D0 F 0 share the same circumcircle Ω0 and incenter
H. Note that K maps to the midpoint MA0 of the arc C 0 E 0 on Ω0 not containing
A0 . Also note that X 0 is the intersection of circles (HB 0 D0 ) and (HC 0 E 0 ), which are
centered at midpoint MF 0 of the arc B 0 D0 on Ω0 not containing F 0 and the midpoint
MD0 of the arc B 0 F 0 on Ω0 not containing D0 , respectively. Thus, X 0 is the reflection
of H over MA0 MF 0 . Similarly, Y 0 is the reflection of H over MA0 MB 0 . Then, note that
MA0 X = MA0 H = MA0 Y . Now we reformulate the problem by erasing A0 , C 0 and E 0 , as
the rest of the problem can be defined without them. The reformulated statement is that
if we fix B, D, F, H and vary MA0 along Ω0 , then D0 MA0 X 0 Y 0 is always cyclic.
We proceed with directed angles. Note that ]X 0 D0 MA0 = ]X 0 D0 H + ]HD0 MA0 =
]MA0 MF 0 F + ]MD0 MF 0 MA0 = ]MD0 MF 0 F . Similarly, ]Y 0 D0 MA0 = MD0 MB 0 B =
−]MD0 MF 0 F = −]X 0 DMA0 , so it follows that MA0 lies on an angle bisector of ]X 0 D0 Y 0 .
Assume that D0 MA0 and X 0 Y 0 are not perpendicular. Then from MA0 X 0 = MA0 Y 0 , it
follows that D0 MA0 X 0 and D0 MA0 Y 0 have the same circumradius, and if they don’t have
the same circumcircle, then D0 MA0 and X 0 Y 0 must be perpendicular, a contradiction. So
D0 X 0 MA0 Y 0 is cyclic. Hf D0 MA0 and X 0 Y 0 are perpendicular, then use the new problem
formulation (without A, C and E and just varying MA0 ) to move MA0 by a miniscule
amount. Then D0 MA0 and X 0 Y 0 will not be perpendicular, so D0 X 0 MA0 Y 0 is cyclic both
after and before moving MA0 by continuity. We are done.

Fifth solution, by complex numbers (Carl Schildkraut) Let Ω be the unit circle, and
let A = a, etc. We have that

c+e 1 a(h − a)
c + e = h − a =⇒ = h̄ − =⇒ ce = .
ce a ah̄ − 1
Let T be the second intersection of the line through A perpendicular to CE and Ω. We
see that
ce h−a
t=− =− .
a ah̄ − 1

24
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

We endeavor to show that DT XY is a cyclic quadrilateral. We have that

ce(b + d) − bd(c + e)
x=
ce − bd
a(b+d)(h−a)
ah̄−1
− bd(h − a)
= a(h−a)
ah̄−1
− bd
 
a(b + d) − bd(ah̄ − 1)
= (h − a)
a(h − a) − bd(ah̄ − 1)
abd
!
ab + ad − ab − ad − f + bd
= (h − a) abd
ab + ad + af − a2 − ab − ad − f + bd
 
bd(f − a)
= (h − a)
(af + bd)(f − a)
bd(h − a)
= .
af + bd
Similarly
bf (h − a)
y= .
ab + df
So, we want to show that

h − a bd(h − a) bf (h − a)
d, − , ,
ah̄ − 1 af + bd ab + df
are concyclic. This is equivalent to, dividing each by h − a and reciprocating,
h−a af ab
, 1 − ah̄, 1 + ,1 +
d bd df
being concyclic. This is equivalent to, subtracting 1 and multiplying by bdf ,

f (b + f − a), −a(bd + bf + df ), ab2 , af 2

being concyclic. This is equivalent to, adding abf and dividing by b + f ,

f, −ad, ab, af

being concyclic. However, all of these points lie on the unit circle, finishing the proof.

Sixth solution by complex numbers (Evan Chen) As usual let Ω denote the unit circle.
We immediately have

c+e=b+d+f −a
1 1 c+e 1 1 1 1
and thus + = = + + −
c e ce b d f a
b+f +d−a
=⇒ ce = 1 1 1 1 .
b + f + d − a

These two equations let us eliminate c and e, leaving only a, b, d, f .

25
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

Now consider the point p = − ce

a on the circumcircle. We compute

x−p x + ce
a
=
b−p b + ce
a
bd(c+e)−ce(b+d) ce
bd−ce + a
=
b + ce
a
abcd + abde − abce − adce + bdce − (ce)2
=
(ab + ce)(bd − ce)
abcde(1/a + 1/e + 1/c − 1/d − 1/b) − (ce)2
=
(ab + ce)(bd − ce)
abcde(1/f ) − (ce)2 (ce)(abd − cef )
= =
(ab + ce)(bd − ce) f (ab + ce)(bd − ce)
Now, we write
ab( 1b + 1
f
1
+ 1
d − a ) + (b + f + d − a)
ab + ce = 1 1 1 1
b + f + d − a
ab( d1 + f1 ) + d + f 1
df (d + f )(ab + df )
= 1 1 1 1 = 1 1 1 1
b + f + d − a b + f + d − a
bd( 1b + d1 + f1 − a1 ) − (b + f + d − a)
bd − ce = 1 1 1 1
b + f + d − a
bd( f1 − a1 ) + (a − f ) 1
af (a − f )(bd + af )
= 1 1 1 1 = 1 1 1 1
b + f + d − a b + f + d − a
f (b + f + d − a)
abd − cef = abd − 1 1 1 1
b + f + d − a
abd( 1b + 1
f + 1 1
d − a ) − f (b + f + d − a)
= 1 1 1 1
b + f + d − a
(b + f )( abd
bf − f ) + b(a − d) + f (a − d)
= 1 1 1 1
b + f + d − a
(b + f )( ad
f − f + (a − d))
= 1 1 1 1
b + f + d − a
1
f (b + f )(a − f )(f + d)
= 1 1 1 1 .
b + f + d − a

Putting that all together gives

x−p ce · adf (b + f )( 1b + f1 + d1 − a1 )
=
b−p (ab + df )(bd + af )

which is symmetric in d and f , so the analogous calculation with fy−p

−p yields the same
result. Consequently, P is the center of the spiral similarity sending Y F to BX, as
desired.
Remark. Philosophical point: it’s necessary to use both a + c + e = b + d + f and its
conjugate, to capture two degrees of freedom.

26
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

G5. Let scalene triangle ABC have altitudes AD, BE, CF and circumcenter O. The
circumcircles of 4ABC and 4ADO meet at P 6= A. The circumcircle of 4ABC meets
lines P E at X 6= P and P F at Y 6= P . Prove that XY k BC.

(Daniel Hu)

Denote by Ω and H the circumcircle and orthocenter of 4ABC. Let T lie on Ω such
that AT k BC. Let 4ABC have orthocenter H.

A
Y X

F
O

B D C
P

First solution, synthetic First we prove a lemma.

Claim. The points H, P, T are collinear.

Proof. Let HT meet Ω at P ∗ 6= T . Let AD meet Ω at K 6= A. By homothety at K,

HT k DO. By angle chasing, ∠P ∗ AD = ∠P ∗ AK = ∠P ∗ T K = ∠P ∗ T O = ∠OP ∗ T =
∠P ∗ OD, so P ∗ lies on the circumcircle of 4AOD. Therefore, P ≡ P ∗ as desired.

We now provide two finishes.

• First finish: By DDIT on AEHF , the pairs of lines (P A, P H), (P B, P C), (P E, P F )

are part of a single involution, so AT , BC, XY are concurrent. Since AT k BC,
this implies that XY k BC as desired.

• Second finish: Let Q = AP ∩ EF . By inversion at A, BF P Q, CEP Q, DHP Q are

all cyclic. By the lemma, this implies that ∠ABC + ∠ACB = ∠AP T = ∠AP H =
∠QP H = ∠QDH = ∠QAH, so DQ ⊥ EF .
Let G = EF ∩ BC; since (G, D; B, C) = −1, ∠BQD = ∠DQC. Thus ∠BAY =
∠BP Y = ∠BP F = ∠BQF = ∠CQE = ∠CP E = ∠CP X = ∠CAX, so XY k
BC as desired.

27
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

Second solution by complex numbers (Adam Ardeishar) Let ABC be the complex
unit circle. Then D = 21 (a + b + c − bc
a ), and we know

p−a d−o
· ∈R
p−o d−a
bc 1 1 1 1 1 a
p−a a+b+c− a p − a a + b + c − bc
· = 1 ·
p b+c−a− bc
a p − a1 + 1b + 1c − bc
a

bc 1 1 1 a
1 a+b+c− a −1 a + b + c − bc
· bc
= · 1 1 1 a
p b+c−a− a
a b + c − a − bc

−a a2 + ab + ac − bc bc + ab + ab − a2
· =
p ab + ac − a2 − bc ab + ac − bc − a2
a2 + ab + ac − bc
p=a·
a2 − ab − ac − bc
p−e
Now note that p + x = e + pxe, so x = pe−1 But we compute that

a2 + ab + ac − bc 1 ac
p−e=a· − (a + b + c − )
a2 − ab − ac − bc 2 b

a3 b + a3 + 2a2 b2 + a2 bc + ab3 + ab2 c + b3 c + b2 c2 − a2 c2

=
2b(a2 − ab − ac − bc)
(a + b)(b + c)(a2 + ab − ac + bc)
=
2b(a2 − ab − ac − bc)
And also compute

a2 + ab + ac − bc 1 1 1 1 b
pe − 1 = a · 2
· ( + + − )−1
a − ab − ac − bc 2 a b c ac

a3 b + a3 c + a2 bc + a2 c2 + ab2 c + 2abc2 + b3 c + b2 c2 − ab3

=
2bc(a2 − ab − ac − bc)
(a + b)(b + c)(a2 + ac + bc − ab)
=
2bc(a2 − ab − ac − bc)
So
(a+b)(b+c)(a2 +ab−ac+bc)
2b(a2 −ab−ac−bc) a2 + ab + bc − ac
x= (a+b)(b+c)(a2 +ac+bc−ab)
=c·
a2 + ac + bc − ab
2bc(a2 −ab−ac−bc)

By symmetry,
a2 + ac + bc − ab
y =b·
a2 + ab + bc − ac
Now note that xy = bc to finish.

28
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

N1. Determine all nonempty finite sets S = {a1 , . . . , an } of n distinct positive integers
such that a1 · · · an divides (x + a1 ) · · · (x + an ) for every positive integer x.

(Ankan Bhattacharya)

Answer: {a1 . . . , an } = {1, . . . , n}. This works since

 
(x + n) . . . (x + 1) x+n
= ∈Z
n! n

so we now show that it is the only possibility. There are two approaches.

First solution Let P (x) = (x + a1 ) . . . (x + an ). Then, a1 . . . an should divide the nth

finite difference of P , which is n!. But

a1 . . . an | n! =⇒ {a1 . . . , an } = {1, . . . , n}

for size reasons.

Second solution (Kevin Sun) Let s + 1 be the smallest positive integer not in our set
A and denote B = A \ {1, . . . , s}.
It’s clear that the divisibility holds for negative x as well. Set x = −s − 1 to obtain
1 Y
Z3 (x + a)
a1 . . . an
a∈A
Y x
= 1+
a
a∈A
  Y 
Y s+1 s+1
= 1− · 1−
a b
a∈{1,...,s} b∈B
   
Y a − (s + 1) Y s+1
= · 1−
a b
a∈{1,...,s} b∈B
 
(−s)(−(s − 1)) . . . (−1) Y s+1
= · 1−
1 · 2 · ··· · s b
b∈B
 
Y s+1
= (−1)|A| 1− .
b
b∈B

If B is nonempty this has magnitude strictly between 0 and 1, (since min B > s + 1 and
thus each term is in (0, 1)). Thus B is empty and A = {1, . . . , s}.

29
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

N2. Call a number n good if it can be expressed in the form 2x + y 2 where x and y are
nonnegative integers.

(a) Prove that there exist infinitely many sets of 4 consecutive good numbers.

(b) Find all sets of 5 consecutive good numbers.

(Michael Ma)

For (a), note that for any t, the numbers t2 + 1, t2 + 2, t2 + 4 are good. So it suffices
to show t2 + 3 is good infinitely often, that is, t2 + 3 = 2x + y 2 has infinitely many
nonnegative integer solutions (since for fixed t there are finitely many (x, y)). But this
rearranges t2 − y 2 = 2x − 3 which has a solution for every x.
We now turn to the laborious task of (b), determining all sets of five consecutive good
numbers. The answers are the six tuples {1, 2, 3, 4, 5}, {2, 3, 4, 5, 6}, {8, 9, 10, 11, 12},
{9, 10, 11, 12, 13}, {288, 289, 290, 291, 292}, {289, 290, 291, 292, 293}. These all work since

1 = 20 + 02 , 2 = 20 + 12 , 3 = 21 + 12 ,
4 = 22 + 02 , 5 = 22 + 12 , 6 = 21 + 22
8 = 23 + 02 , 9 = 23 + 12 , 10 = 20 + 32 ,
11 = 21 + 32 , 12 = 23 + 22 , 13 = 22 + 32 ,
288 = 25 + 162 , 289 = 26 + 152 , 290 = 20 + 172 ,
291 = 21 + 172 , 292 = 28 + 62 , 293 = 22 + 172 .

We now show they are the only ones. First, consider the following table which shows
2x + y 2 (mod 8):

x=0 x=1 x=2 x≥3

y≡1 (mod 2) 2 3 5 1
y≡0 (mod 4) 1 2 4 0
y≡2 (mod 4) 5 6 0 4

Note that from this table, no good number is 7 (mod 8). Thus any five good numbers
must have a 3 (mod 8) number. By table can only occur if that good number is of the
form t2 + 21 = t2 + 2 for an odd integer t.
We now have several cases.
Case 1: Suppose the five good numbers are {t2 + 1, t2 + 2, t2 + 3, t2 + 4, t2 + 5}.
Note that t2 +5 ≡ 6 (mod 8), and by table, this can only occur if t2 +5 = s2 +22 = s2 +4
for some integer s; hence t2 − s2 = 1, so t = 1 and s = 0. This gives the solution set
{2, 3, 4, 5, 6}.
Case 2: Suppose the five good numbers are {t2 , t2 + 1, t2 + 2, t2 + 3, t2 + 4}.
Since t2 is good, we have t2 = 2w +z 2 for some w and z, which we write as (t−z)(t+z) =
w
2 .
We now split into cases.

• Subcase 2.1: We handle the situation where w < 4.

– If w = 0, then we get t = 1, which gives the solution {1, 2, 3, 4, 5}.
– If w = 1, then there are no solutions by taking mod 4.

30
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

– If w = 2, then t2 = 4 + z 2 which implies t = 2, but t was odd.

– If w = 3, we get t2 = 8 + z 2 which implies t = 3, which gives {9, 10, 11, 12, 13}.
– If w = 4, we get t2 = 16 + z 2 which together with t odd implies t = 5, which
gives {25, 26, 27, 28, 29}. However, the number 28 is not good, so this is not a
solution.

• Subcase 2.2: Suppose w ≥ 5. As gcd(t
− z, t + z) | 2t we must have t − z = 2,

t + z = 2w−1 , and thus t = 12 2 + 2w−1 = 2w−2 + 1. Since t was odd, we actually
have w ≥ 3.
But t2 + 3 is also good, so write

t2 + 3 = 2x + y 2 .

So we split into cases again.

– Subcase 2.2.1: We handle the case x < 3.
∗ If x = 0, we get t2 + 2 = y 2 which has no solutions.
∗ If x = 1, we get t2 + 1 = y 2 which implies t = 0, but t is supposed to be
odd.
∗ If x = 2, then we get t2 = y 2 + 1 which implies t = 1, which was an earlier
solution.
– Subcase 2.2.2: Otherwise, assume x ≥ 3.

2x + y 2 = t2 + 3

2
=⇒ 2x + y 2 = 2w−2 + 1 +3
2w−4 w−1
=2 +2 +4
2w−6 w−3 x−2 2
=⇒ 2 +2 +1=2 + (y/2)

since y is clearly even; the last line implies y/2 is odd, since 2w − 6 > 0,
w − 3 > 0, x − 2 > 0.
Let c = w − 3 ≥ 2, a = x − 2 ≥ 1, b = y/2 ≥ 1 for brevity; then the equation
rewrites as
22c + 2c + 1 = 2a + b2 .
We rewrite this as

(2c + 1 − b)(2c + 1 + b) = (2c + 1)2 − b2 = 2a + 2c ≥ 0.

In light of this, we have 2a + 2c ≥ (2c + 1)2 − 22c > 2c+1 , so 2a > 2c , ergo
a > c. Thus we may further write

(2c + 1 − b)(2c + 1 + b) = 2c (2a−c + 1).

The factors on the left-hand side are nonnegative and have gcd dividing 2b,
hence one of them has at most one factor of 2. So one of the factors must be
divisible by 2c−1 . Thus, b ≡ ±1 (mod 2c−1 ).
But, b < 2c + 1. So we have four possibilities:
∗ Subcase 2.2.2.1: suppose b = 1. Then we get 22c + 2c = 2a , which is
impossible.

31
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

∗ Subcase 2.2.2.2: suppose b = 2c−1 − 1. Then we get (2c−1 + 2)(2c +

2c−1 ) = 2c (2a−c + 1) and hence 3 · 2c−2 = 2a−c − 2. This implies a − c = 3
and c − 2 = 1, so c = 3, or w = 6, hence t = 2w−2 + 1 = 17.
This gives {289, 290, 291, 292, 293} which indeed works.
∗ Subcase 2.2.2.3: suppose b = 2c−1 +1. Then we get 2c−1 (2c +2c−1 +2) =
2c (2a−c + 1), or 2c−1 + 2c−2 + 1 = 2a−c + 1, which is impossible.
∗ Subcase 2.2.2.4: suppose b = 2c − 1. This gives 2 · 2c+1 = 2c (2a−c + 1),
which is impossible.

Case 3: Suppose the five good numbers are {t2 − 1, t2 , t2 + 1, t2 + 2, t2 + 3}.

In that case, {t2 , t2 + 1, t2 + 2, t2 + 3, t2 + 4} is also a set of five consecutive good
numbers. Using case 2, the new candidate this now gives are {8, 9, 10, 11, 12} and
{288, 289, 290, 291, 292}, which work.

32
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

N3. Let a1 , a2 , . . . be an infinite sequence of positive integers satisfying a1 = 1 and

an | ak + ak+1 + · · · + ak+n−1

for all positive integers k and n. Find the maximum possible value of a2018 .
(Krit Boonsiriseth)

The answer is a2018 ≤ 21009 − 1. To see this is attainable, consider the sequence
(
1 n odd
an = n/2
2 − 1 n even.

This can be checked to work, so we prove it’s optimal.

We have a2 | a1 + a2 = 1 + a2 =⇒ a2 = 1.
Now consider an integer n, and let s = sn = a1 + · · · + an . Then

an+1 | s
an+2 | s + an+1
an+2 ≡ 1 (mod an+1 ).
s+an+1
Thus, gcd(an+2 , an+1 ) = 1. So an+2 ≤ an+1 , and thus
s
an+1 + an+2 ≤ 1 + an+1 + ≤ s + 2.
an+2
So, we have

a1 + a2 = 2
a3 + a4 ≤ 2 + 2 = 4
a5 + a6 ≤ (2 + 4) + 2 = 8
a7 + a8 ≤ (2 + 4 + 8) + 2 = 16
..
.
a2017 + a2018 ≤ 21009 .

Thus a2018 ≤ 21009 − a2017 ≤ 21009 − 1.

Remark (Motivational notes). It’s very quick to notice an+1 | a1 +· · ·+an , which already
means that given the first n terms of the sequence there are finitely many possibilities
for the next one. Thus it’s possible to play with “small cases” by drawing a large tree.
When doing so, one might hope that somehow an = a1 + · · · + an−1 is achievable, but
quickly notices in such a tree that if an is the sum of all previous terms, then an+1 = 1 is
forced. This gives the idea to try to look at the terms in pairs, rather than one at a time,
and this gives the correct bound.
As for extracting the equality case from this argument, there are actually two natural
curves to try. We have a3 | 1 + 1 = 2. If we have a3 = 2 we get a4 = 1, a5 ≤ 5, but then a6
actually gets stuck. But if we have a3 = 1 instead, we get a4 = 3, a5 = 1, a6 = 7, and so
on; pushing this gives the equality case above, seen to work. I think it’s quite unnatural
to guess the correct construction before having the corresponding s + 2 estimate.

33
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

N4. Fix a positive integer n > 1. We say a nonempty subset S of {0, 1, . . . , n − 1} is

d-coverable if there exists a polynomial P with integer coefficients and degree at most d,
such that S is exactly the set of residues modulo n that P attains as it ranges over the
integers.
For each n, determine the smallest d such that any nonempty subset of {0, . . . , n − 1}
is d-coverable, or prove that no such d exists.

(Carl Schildkraut)

This is possible for n = 4 or n prime, in which case d = n − 1 is best possible. Let

P (Z/n) denote the range of a polynomial modulo n.

• We first note that if n = q1 . . . qk is the product of k ≥ 2 distinct prime powers,

then
Yk
|P (Z/n)| = |P (Z/qi )|.
i=1

Hence any subset S with size n − 1 is not coverable.

• If n = pe is a prime power with other than 4 with e ≥ 2, consider the set

S = {0, 1, . . . , p − 1, p}. We claim it is not coverable.
Indeed, if P covers it, WLOG P (0) = 0. Now, P is surjective modulo p, hence
bijective, and thus P (x) ≡ 0 (mod p) ⇐⇒ x ≡ 0 (mod p). Now we can write

P (x) = a1 x + a2 x2 + . . . .

– If a1 ≡ 0 (mod p), then x ≡ 0 (mod p) =⇒ P (x) ≡ 0 (mod p2 ), so p does

not appear in the image.
– If a1 6≡ 0 (mod p), then p, 2p, . . . all appear in the miage, which is wrong for
n > 4.

• Let n = 4, and consider S (mod 4).

– If S = {k} take P (x) = k.
– If S = {k, k + 1} take P (x) = x2 + k.
– If S = {k, k + 2} take P (x) = 2x2 + k.
– If S = {k − 1, k, k + 1} take P (x) = x3 + k.
We claim also the example S = {−1, 0, 1} is not 2-coverable. Indeed, WLOG
P (0) = 0 so P (x) = x(x + c). Then P (2) ≡ 0 (mod 4), meaning c is even. But
then P (1) ≡ c + 1 (mod 4) and P (−1) ≡ 1 − c (mod 4), so P (1) ≡ P (−1).
– If S = {0, 1, 2, 3} take P (x) = x.

• Let n = 2.
– If S = {k} take P (x) = k.
– If S = {0, 1} take P (x) = x. This is obviously not 0-coverable.

34
20th ELMO 2018 (Pittsburgh, PA) Shortlisted Problems (Evan Chen et al)

• If n = p is an odd prime, we claim S = {1, . . . , p − 1} is not (p − 2)-coverable.

Indeed, suppose P (x) = ap−2 xp−2 + · · · + a0 covered it. Then
X X X
P (x) ≡ ak xk ≡ 0 (mod p).
x k x

However, if P (Z/p) = {1, . . . , p − 1} then some element appears

P twice and the
others appear once. If k is the repeated element though, then x P (x) = (1 + · · · +
(p − 1)) + k ≡ k 6≡ 0 (mod p).

35