INMO 2023

Official Solutions

Problem 1. Let S be a finite set of positive integers. Assume that there are precisely
2023 ordered pairs (x, y) in S × S so that the product xy is a perfect square. Prove that
one can find at least four distinct elements in S so that none of their pairwise products is
a perfect square.

Note: As an example, if S = {1, 2, 4}, there are exactly five such ordered pairs: (1, 1), (1, 4),
(2, 2), (4, 1), and (4, 4).

Solution 1. Consider the graph whose vertices are elements of S, with an edge between
x and y if and only if xy is a perfect square. We claim every connected component is a
clique.
Indeed, take any two vertices corresponding to x, y in S in the same connected com-
ponent. It suffices to show they are adjacent. By assumption, there is a path between
them; so there is a sequence x = a1 , a2 , . . . , an−1 , an = y so that ai ai+1 is a perfect square
for 1 ≤ i < n. Therefore
(a1 a2 )(a2 a3 ) · · · (an−1 an )
xy = a1 an = ,
a22 · · · a2n−1

is a perfect square as well. This proves our claim.

Now suppose first there are at most 3 connected components, with sizes a, b, c (possibly
zero). Note that for (x, y) ∈ S × S , xy is a perfect square if and only if x, y are in the same
component, which can be chosen is a2 + b2 + c2 ways. Thus

a2 + b2 + c2 = 2023.

But since squares can only be 0, 1 or 4 mod 8, and 2023 is 7 mod 8, the above equation is
impossible. Thus our graph must have at least four components. Picking a number from
each component, we can now satisfy the requirements of the problem.

Solution 2. For a in S, let Sa = {x ∈ S|ax is a square}. Let a, b be elements of S. Suppose

that x is in Sa ∩ Sb . Then ax and bx are squares and hence ab = ax·bx x2 is a square. Then
for any y in S such that ay is a square it follows that by is a square, so Sa = Sb . Hence for
two elements a, b in S, either Sa = Sb or Sa ∩ Sb = ∅.
Now, S = ∪Sa where the union runs over elements of S (since a ∈ Sa for any a ∈ S).
Let S = Sa1 ∪ Sa2 ∪ · · · ∪ San for some elements a1 , a2 , . . . , an of S such that Sai ∩ Saj = ∅ for
1 ≤ i < j ≤ n. Then the number of distinct pairs (x, y) of S × S such that xy is a square
is precisely |Sa1 |2 + |Sa2 |2 + · · · + |San |2 . Since 2023 ≡ 7 (mod 8) it follows that n > 3 as in
the previous solution. Thus we have four elements a1 , a2 , a3 , a4 none of whose pairwise
products is a square.

For those familiar with the language of linear algebra and finite fields, the above argu-
ment can be reformulated as follows:

Solution 3. Let p1 < p2 < . . . be the sequence of prime numbers. Denote by Fω 2 the F2
vector space of all binary sequences (a1 , a2 , . . . ) with entries in F2 . Consider the set map
Φ : N → Fω
2 defined by
Φ(n) := (vpi (n) (mod 2))i≥1
for all n ∈ N. It is clear that for x, y ∈ N, the product xy is a perfect square if and
only if Φ(x) = Φ(y). So we want |Φ(S)| ≥ 4. Indeed, decompose S as a union of fibres,
S = ∪a∈Φ(S) Φ−1 (a). Each fibre with size r accounts for 2 2r + r = r2 pairs towards the

count, so if Φ(S) has at most three elements, then a2 + b2 + c2 = 2023 has a solution in
non-negative integers. This is a contradiction mod 8, by simply checking that no triple
formed using one of {0, 1, 4} can add to 7 mod 8.

1
Problem 2. Suppose a0 , . . . , a100 are positive reals. Consider the following polynomial for
each k in {0, 1, . . . , 100}:

a100+k x100 + 100a99+k x99 + a98+k x98 + a97+k x97 + · · · + a2+k x2 + a1+k x + ak ,

where indices are taken modulo 101, i.e., a100+i = ai−1 for any i in {1, 2, . . . , 100}. Show
that it is impossible that each of these 101 polynomials has all its roots real.

Solution 1. Let n = 50. For the sake of contradiction, assume that each of these poly-
nomials has all real roots; these roots must be negative. Let

−α1,k , −α2,k , . . . , −α2n,k

be the roots of the polynomial

a2n+k x2n + 2na2n−1+k x2n−1 + a2n−2+k x2n−2 + a2n−3+k x2n−3 + · · · + a2+k x2 + a1+k x + ak .

Indices are taken modulo 2n + 1, so a2n+k = ak−1 and a2n−1+k = ak−2 . Then
2n   2n
X ak−2 Y ak
αj,k = 2n. ; αj,k = .
j=1
ak−1 j=1
ak−1

Since the αj,k ’s are positive, AM-GM inequality can be applied and by virtue of it we are
led to  2n
ak−2 ak
≥
ak−1 ak−1
for each k. As both sides of the inequalities are positive, multiplying them we obtain
2n  2n 2n
Y ak−2 Y ak
≥ .
ak−1 ak−1
k=0 k=0

But both sides are equal to 1. Therefore all the 2n + 1 A.M-G.M inequalities are equalities
implying that for each k,
ak−2
α1,k = α2,k = . . . = α2n,k = .
ak−1
Since n ≥ 2, using Vieta’s relations gives
  2
ak−3 X 2n ak−2
= αi,k αj,k = .
ak−1 2 ak−1
1≤i<j≤2n

Simplifying leads  
2n 2
ak−2 = ak−1 ak−3
2
for each k. Multiplying all these equations yields
 2n+1 ! 2n
!2
2n Y
−1 ak = 0,
2
k=0

which shows that at least one ak = 0, a contradiction.

Solution 2. As above, one proves that

ak−2
α1,k = α2,k = . . . = α2n,k = .
ak−1
This implies
 2n
2n 2n−1 ak−2
a2n+k x + 2na2n−1+k x + · · · + a1+k x + ak = a2n+k x+ .
ak−1

For n ≥ 2, comparing coefficients of x0 and x1 , we see that

 2n  2n−1
ak−2 ak−2
ak = a2n+k , ak+1 = a2n+k · 2n ,
ak−1 ak−1

2
whence we obtain
ak+1 ak−1
= 2n · .
ak ak−2
ak+1
This must hold for all k. However, if we pick k is such that ak is minimal, we must
necessarily have
ak+1 ak−1 ak−1
≤ < 2n · ,
ak ak−2 ak−2
a contradiction.

Several other beautiful solutions were pointed out to us; we include a few of them.

Solution 3. As usual, assume all of these polynomials have all real roots. We consider
the polynomial obtained by writing the coefficients of the given polynomials in reverse
order:
ak x100 + ak+1 x99 + ak+2 x98 + · · · + 100ak+99 x + ak+100 .
This also has all its roots real: in fact, its roots are reciprocals of the original polynomial.
By Rolle’s theorem, its derivative must also have all roots real. Repeating this argument,
we see that the polynomial obtained by differentiating this 98 times
100!
ak x2 + 99!ak+1 x + 98!ak+2
2!
also have real roots. Therefore this must have nonnegative discriminant:

99!2 a2k+1 ≥ 2 · 98! · 100!ak ak+2

which simplifies to
99a2k+1 ≥ 200ak ak+2 .
This holds for all k, so multiplying these as k varies, we obtain
100
Y 100
Y
99101 a2i ≥ 200101 a2i ,
i=0 i=0

which is impossible since 99 < 200.

Solution 4. Choose k so that ak+1 is minimal. As before, consider the polynomial

ak x100 + ak+1 x99 + ak+2 x98 + · · · + 100ak+99 x + ak+100

and suppose its roots are β1 , · · · , β100 . We have, by Vieta’s relations,

100
X ak+1 X ak+2
βi = − , βi βj = .
i=1
ak ak
1≤i<j≤n

However, we note that

100
!2 100
X X X X
βi = βi2 + 2 βi βj ≥ 2 βi βj ,
i=1 i=1 1≤i<j≤n 1≤i<j≤n

which yields
 2
ak+1 2ak+2
− ≥ .
ak ak
This simplifies to a2k+1 ≥ 2ak ak+2 . However, by the choice of k, we have

a2k+1 ≤ ak ak+2 < 2ak ak+2 ,

a contradiction.

Problem 3. Let N denote the set of all positive integers. Find all real numbers c for which
there exists a function f : N → N satisfying:
f (x+a)−f (x)
(a) for any x, a ∈ N, the quantity a is an integer if and only if a = 1;
(b) for all x ∈ N, we have |f (x) − cx| < 2023.

3
Solution 1. We claim that the only possible values of c are k + 12 for some non-negative
integer
k. The fact that  these values are possible is seen from the function f (x) =
k + 21 x + 1 = kx + x2 + 1. Indeed, if you have any x, a ∈ N, then


        
f (x + a) − f (x) 1 x+a x 1 x+a x
= ka + − =k+ − .
a a 2 2 a 2 2
This is clearly an integer for a = 1. But for a ≥ 2, we have
       
x+a x x+2 x
− ≥ − = 1.
2 2 2 2
If a = 2k, then    
x+a x
− = k < 2k = a,
2 2
and if a = 2k + 1 for k ≥ 1, then
       
x+a x x + 2k + 2 x
− ≤ − = k + 1 < 2k + 1 = a.
2 2 2 2

So in either case, the quantity x+a

  x
2 − 2 is strictly between 0 and a, and thus cannot be
divisible by a. Thus condition (a) holds; condition (b) is obviously true.

Now let us show these are the only possible values, under the weaker assumption that
there exists some d ∈ N so that |f (x) − cx| < d. It is clear that c ≥ 0: if −d < f (x) − cx < d
and c < 0, then for large x the range [cx − d, cx + d] consists only of negative numbers and
cannot contain f (x).
Now we claim that c ≥ 21 . Indeed, suppose that 0 ≤ c < 12 , and that d > 0 is such that
2d
|f (x) − cx| ≤ d. Pick N > 1−2c so that 2(cN + d) < N . Then the N values {f (1), . . . , f (N )}
must be all be in the range {1, . . . , cN + d}, and by pigeonhole principle, some three values
f (i), f (j), f (k) must be equal. Some two of i, j, k are not consecutive: suppose WLOG
i > j + 1. Then f (i)−f
i−j
(j)
= 0, which contradicts condition (a) for x = j and a = i − j.
Now for the general case, suppose c = k + λ, where k ∈ Z and λ ∈ [0, 1). Let d ∈ N be
such that −d ≤ f (x) − cx ≤ d. Consider the functions

g1 (x) = f (x) − kx + d + 1, g2 (x) = x − f (x) + kx + d + 1.

Note that

g1 (x) ≥ cx − d − kx + d + 1 = λx + 1 ≥ 1,
g2 (x) ≥ x − (cx + d) + kx + d + 1 = (1 − λ)x + 1 ≥ 1

so that these are also functions from N to N. They also satisfy condition (a) for f :
g1 (x + a) − g1 (x) f (x + a) − k(x + a) + d − f (x) + kx − d f (x + a) − f (x)
= = −k
a a a
is an integer if and only if f (x+a)−f
a
(x)
is, which happens if and only if a = 1. A similar
argument holds for g2 .
Now note that g1 (x) − λx = f (x) − cx + d + 1 is bounded, and so is g2 (x) − (1 − λ)x =
cx − f (x) + d + 1. So they satisfy the weaker form of condition (b) as well. Thus applying
the reasoning in the second paragraph, we see that λ ≥ 12 and 1 − λ ≥ 12 . This forces λ = 12 ,
which finishes our proof.

Solution 2. We will show that for any such c, we have c > 0 and {c} = 12 . Also 2023 can
be replaced by any fixed d ≥ 1 in condition (b) which we assume now.
d
Clearly c ≥ 0 else for c < 0 and x > |c| , cx−d < f (x) < cx+d < 0 which is a contradiction.
1
Suppose {c} =6 2 . Put r = bcc and λ = min({c}, 1 − {c}) and define
(
f (x) − rx if {c} < 12
g(x) =
x + rx − f (x) if {c} > 12
1
so that |f (x) − cx| = |g(x) − λx| and g(x) ∈ Z for all x ∈ N. Here 0 ≤ λ < 2. Take
N > 2(λN + 2d). Then from |g(x) − λx| = |f (x) − cx| < d, we get

−d ≤ λn − d < g(n) < λn + d ≤ λN + d

4
for all 1 ≤ n ≤ N . That is N integers g(n), 1 ≤ n ≤ N can take at most λN + 2d values.
Since N > 2(λN + 2d), by pigeonhole principle, there are 3 positive integers i < j < k such
that g(i) = g(j) = g(k). Then k − i ≥ 2 and
(
g(k) + rk − (g(i) + ri) = r(k − i) if {c} < 21
f (k) − f (i) =
(1 + r)k − g(k) − ((1 + r)i − g(i)) = (1 + r)(k − i) if {c} > 21

so that f (k)−f
k−i
(i)
is an integer. This contradicts the condition (a). Also for each c = k + 21 ,
the function f (x) = b(k + 12 )xc satisfy the conditions (a) and (b).

Solution 3. We give a different proof that {c} = 1/2. Let us first prove a claim:

Claim. For any k ≥ 1 and any x, f (x + 2k ) − f (x) is divisible by 2k−1 but not 2k .
Proof. We prove this via induction on k. For k = 1, the claim is trivial. Now assume the
statement is true for some k, and note that f (x + 2k ) − f (x) = 2k−1 y1 and f (x + 2k + 2k ) −
f (x + 2k ) = 2k−1 y2 for some odd integers y1 , y2 . Adding these, we see that

f (x + 2k+1 ) − f (x) = 2k−1 (y1 + y2 )

which is divisible by 2k because y1 + y2 is even. The fact that this is not divisible by 2k+1
follows from the condition on f .
Now using this claim, we see that for any k ≥ 1, f (1 + 2k ) = f (1) + 2k−1 (2yk + 1) for some
integer yk , which means
 
k k k 1
f (1 + 2 ) − c(1 + 2 ) = f (1) − c + 2 yk + − c .
2

Thus 2k (yk + 12 − c) is bounded. But yk + 21 − c has the same fractional part 1

 1 as 2 −
 c, so if

this quantity is never zero, its absolute value must be at least m = min 2 − c , c − 21
and thus we have
1
2k yk + − c ≥ 2k m,
2
1
contradicting boundedness. Thus we must have yk + 2 − c = 0 for some k. Since yk is an
integer, so that {c} = 12 .

A more rigorous treatment is given below.

Obtain  
k k k 1
f (1 + 2 ) − c(1 + 2 ) = f (1) − c − 2 yk + − c
2
as before. We obtain that 2k |yk + 21 − c| ≤ M for some M > 0 by condition (b). Suppose that
6 21 . Writing yk + 12 − c = mk + δ with mk ∈ Z and 0 ≤ δ < 1, we have 0 < δ < 1. Then
{c} =
there exists ` > 1 such that min(δ, 1 − δ) ≥ 21` . Hence
(
1 δ ≥ 21` if mk ≥ 0
|yk + − c| = |mk + δ| ≥ 1
2 −mk − δ ≥ 1 − δ ≥ 2` if mk < 0

1
implying M ≥ 2k |yk + 2 − c| ≥ 2k−` which is a contradiction for large k. Thus {c} = 21 .

Problem 4. Let k ≥ 1 and N > 1 be two integers. On a circle are placed 2N + 1 coins
all showing heads. Calvin and Hobbes play the following game. Calvin starts and on
his move can turn any coin from heads to tails. Hobbes on his move can turn at most
one coin that is next to the coin that Calvin turned just now from tails to heads. Calvin
wins if at any moment there are k coins showing tails after Hobbes has made his move.
Determine all values of k for which Calvin wins the game.

5
Solution 1. Calvin wins if k ∈ {1, 2, . . . , N + 1} and Hobbes wins otherwise.

Label the coins 1, 2, . . . , 2N + 1. Note that if k ≥ N + 2 then Hobbes wins as follows: he

pairs the coins 2i − 1 and 2i for 1 ≤ i ≤ N . If Calvin in a move makes both coins in a pair
tails, Hobbes in that move turns the one which was tails prior to Calvin’s move back to
heads. Thus, he can ensure that after his move, no pair has more than one tails. So, the
number of tails after his move is ≤ 1 + (2N +1)−1 2 = N + 1, hence Hobbes wins. If k ≤ N ,
then Calvin wins by simply turning coins 2i for 1 ≤ i ≤ N . Now let k = N + 1.

Let N = 2m + ε where ε ∈ {0, 1}. Now consider m arcs on the circle with the ith arc
containing {4i + 1, 4i + 2, 4i + 3, 4i + 4}, for all 0 ≤ i < m. Calvin makes 3m moves as follows:
on move 3i+1, 3i+2, and 3i+3, he turns coins numbered 4i+2, 4i+4, and 4i+3, respectively,
to tails, for all 0 ≤ i < m. Thus, no matter what Hobbes does, each arc will have either
{4i + 2, 4i + 3} tails (type c
23), or {4i + 3, 4i + 4} tails (type c
34), or all of {4i + 2, 4i + 3, 4i + 4}
tails (type 234),
d by the end of these 3m moves. We now split into cases:

Case 1. ε = 0 In this case, if we have any arc of type 234, d we get that there are
≥ 3 + 2(m − 1) = N + 1 tails at the end and the game is won. Assume all arcs are of
type c 23 or c34; hence we currently have 2m tails. Now, if 4m + 1 has no tails neighbours,
Calvin turns it to win. So assume the arc {4m − 3, 4m − 2, 4m − 1, 4m} is of type c 34. If
{1, 2, 3, 4} is also of type 34, Calvin can turn 1 to tails to win as it has no tails neighbours.
c
If it is of type c23, then we must have an 0 ≤ i < m − 1 such that the ith arc is of type 23 c
but the (i + 1)th arc is of type 34, which means that Calvin can turn 4i + 1 to tails to win,
c
as it has no tails neighbours.

Case 2. ε = 1 Again, if we have any arc of type 234, d Calvin turns 4m + 2 and we end up
with ≥ 3 + 2(m − 1) + 1 = N + 1 tails at the end and the game is won. Assume all arcs are
of type c 23 or c34; hence we currently have 2m tails. If Calvin can turn 4m + 1, then he wins,
by turning 4m + 3 next move; so assume the arc {4m − 3, 4m − 2, 4m − 1, 4m} is of type c 34.
If {1, 2, 3, 4} is of type c
34, then Calvin can turn 1 and 4m + 2 to secure his win; so assume
{1, 2, 3, 4} is of type c23. Thus, there exists 0 ≤ i < m − 1 such that the ith arc is of type c
23
and the (i + 1)th arc is of type c 34, hence Calvin wins by turning 4m + 2 and 4i + 1, in the
next two moves.

In conclusion, Calvin wins if k = N + 1, completing the proof.

Again, we include some remarkable alternative solutions that were brought to our
attention.

Solution 2. As before, we can prove that Hobbes wins for k ≥ N + 2. It remains to show
that Calvin can win for k = N + 1.

Let Calvin turn the N coins labeled 1, 3, · · · , 2N − 1 in order. Note that no two coins
showing tails are adjacent at any point till now, so Hobbes cannot change any of them
back to heads. On the next move, let Calvin turn coin 2 to tails. Now if Hobbes turns coin
1 to heads, Calvin can turn coin 2N + 1 to tails and win. If Hobbes turns coin 3 to heads
instead, let Calvin turn the even-numbered coins 4, 6, · · · , 2N − 2 in order. After Calvin
flips the coin 2k, Hobbes must respond by flipping 2k + 1 after each move to avoid losing.
Therefore after these moves, the N coins 1, 2, 4, · · · , 2N − 2 are showing tails, so now Calvin
can turn 2N and win.

Solution 3. Again we show a procedure for Calvin to win for k = N + 1. Let Calvin turn
coins 1 and 3 tails, and then coin 2 to tails. Hobbes must respond by turning one of 1 or
3 back to heads. In any case, Calvin secures two consecutive tails at the end of Hobbes’
turn. He removes these two coins and their neighbours; and from the (2N + 1) − 4 = 2N − 3
remaining coins, he turns 12 (2N − 3 + 1) = N − 1 of them tails by alternately flipping them.
Hobbes never gets to make a move to stop this sequence, so Calvin wins.

6
Problem 5. Euler marks n different points in the Euclidean plane. For each pair of
marked points, Gauss writes down the number blog2 dc where d is the distance between
the two points. Prove that Gauss writes down less than 2n distinct values.

Note: For any d > 0, blog2 dc is the unique integer k such that 2k ≤ d < 2k+1 .

Solution 1. We first prove that the Gauss writes down at most n even numbers.
For each even number 2k that Gauss writes down, choose a single pair of points whose
distance d satisfies 22k ≤ d < 22k+1 . Connect these points with a red edge. We claim there
cannot be a cycle {of length m, 3 ≤ m ≤ n}: indeed, if the edges corresponding to the
distinct even integers 2k1 , . . . , 2km , 2km+1 form a cycle in that order, then assume without
loss of generality that 2km+1 is the largest among these, and 2ki is the largest among the
rest. The sum of distances for the first m edges is at most

22ki +1
22k1 +1 + · · · + 22km +1 ≤ 22ki +1 (1 + 2−2 + 2−4 + · · · ) ≤ < 22ki +2 ≤ 22km+1 ,
1 − 212

i.e., less than the distance corresponding to the last edge: a contradiction to triangle
inequality. So there are at most n − 1 red edges.
This implies that Gauss only writes at most n − 1 even numbers, and similarly at most
n − 1 odd numbers. Thus, Gauss writes down at most 2n − 2 numbers in total.

Solution 2. Given a pair of points with distance d, we say their log-distance is blog2 dc.
By dilating by a suitable power of 2, we can assume all log-distances are positive and the
smallest one of them is 1. Define a sequence of graphs G0 , G1 , G2 , · · · as follows: Gk has
the n points as vertices, and two points are joined by an edge in it if and only if their
log-distance is at most k. In particular, G0 is a graph with n vertices and no edges.
Now we define a second sequence of graphs H0 , H1 , H2 , · · · with Hi being a subgraph
of Gi inductively as follows. We let H0 be the same as G0 and H1 by adding an edge with
log-distance 1. Now once we have defined Hk−1 , Hk is obtained from it as follows:
• If Hk−1 is a disjoint union of cliques, choose a pair of points with log-distance k,
and add the edge between them to Hk−1 to form Hk : such an index k will be called
clique-destroyer. If no such pair exists, let Hk = Hk−1 : in this case, the index k will
be called empty.
• If not, then Hk−1 was obtained by adding one edge between cliques C1 and C2 in
Hk−2 ⊆ Gk−2 . We claim that all edges between C1 and C2 are in Gk . Indeed, if
suppose the edge was added between points A1 ∈ C1 and A2 ∈ C2 . Take any point
B1 ∈ C1 and B2 ∈ C2 . By assumption, B1 A1 and B2 A2 are edges in Gk−2 and A1 A2
is an edge in Gk−1 , so |B1 A1 | < 2k−1 , |B2 A2 | < 2k−1 , and |A1 A2 | < 2k , so by triangle
inequality,
|B1 B2 | < 2k−1 + 2k−1 + 2k = 2k+1 ,
which proves our claim. Now to form Hk , we take Hk−1 and add all edges between
C1 and C2 : this yields a disjoint union of cliques. This index k will be called clique-
restorer.
Now for every clique-destroyer index k, either Hk is again a disjoint union of cliques with
one fewer clique than Hk−1 , or it is followed by a clique-restorer index k + 1 (as we add all
the missing edges between some two disjoint cliques C1 and C2 connected by an edge),
and Hk+1 is a disjoint union of cliques with one fewer clique than Hk−1 . Initially, H0 has n
cliques and the chain stops once Hk becomes a single clique on n vertices. There can be
at most 2(n − 1) indices that are one of the above two types. However, the empty indices
correspond to values of k that do not occur as a log-distance: therefore the number of
distinct values of log-distance is at most the number of non-empty indices, i.e., at most
2n − 2 as desired.

Problem 6. Euclid has a tool called cyclos which allows him to do the following:
• Given three non-collinear marked points, draw the circle passing through them.
• Given two marked points, draw the circle with them as endpoints of a diameter.
• Mark any intersection points of two drawn circles or mark a new point on a drawn
circle.

7
 Show that given two marked points, Euclid can draw a circle centered at one of them
and passing through the other, using only the cyclos.

Solution 1. We begin by proving a series of lemmas.

Lemma 1. Given a non-right angled triangle ABC, we can draw the nine-point circle
and mark the orthocentre H using only a cyclos.

Proof. Draw circles (BC), (CA), (AB) and mark their intersections to get the three feet of
altitudes D, E, F opposite A, B, C. Now draw the circle (DEF ) to get the nine-point circle.
Draw (BDF ), (CDE), (AEF ) and they meet at H, which we can also mark.

Lemma 2. Given points A, B, we can mark the midpoint M of AB using only a cyclos.

Proof. Draw the circle (AB) and choose a point X on it. Draw circles (XA), (XB) and
mark their intersection Y . Now mark a point Z on the circle (XA) apart from the marked
points. Clearly, Z does not lie on AB nor on (AB), hence we can draw (AZB). Mark five
points Z1 , . . . , Z5 on this circle, each different from all previous points and verify if either
A lies on (Zi B) or B lies on (Zi A) for each 1 ≤ i ≤ 5 before marking the new point. By
pigeonhole principle, for some three indices i, j, k, the three triangles AZi B are non-right
angled, hence we can draw their ninepoint circles by Lemma 1. All of them pass through
M , and their centres are not collinear, else homothety at the centre of (ABZ) implies the
orthocentres of the three triangles are collinear; but they all lie on the reflection of (AZB)
in AB, a contradiction! Thus, these three nine-point circles meet at only M , and we mark
this point.

Lemma 3. Given points A, B, C, D on the plane in general position, we can mark the
intersection point E of lines AB and CD using only a cyclos.

Proof. Draw (AB) and mark five points on it, all different from previously marked points.
For each marked point X; draw (CX) and (DX) and check whether they have an inter-
section apart from X (i.e., if they are tangent, or if X lies on CD). We can find three
points X1 , X2 , X3 among them not lying on CD. Denote by Yi the second intersection of
(AXi ), (BXi ) and by Zi the second intersection of (CXi ), (DXi ) and mark them, for each
1 ≤ i ≤ 3. Draw the circles (Xi Yi Zi ) and note that they all pass through E and have
diameters EXi for all i; so they are not coaxial as X1 , X2 , X3 are not collinear; all lying on
(AB). Thus we mark E as the unique point common to them all. (Note: if C, D lie on (AB),
we can pick a point T on it other than these four, then a point X on (AT ), and continue
the same argument again, avoiding all edge cases.)

Lemma 4. Given a circle Γ, we can mark the centre of Γ using only a cyclos.

Proof. Mark points A, B, C ∈ Γ and mark the midpoints of BC, CA, AB to get A1 , B1 , C1
according to Lemma 2. Draw the circles (AB1 C1 ), (BA1 C1 ), (CB1 A1 ) and mark the inter-
section to get the centre of Γ.

Lemma 5. Given a circle Γ and point A on Γ, we can mark a point K such that line AK
is tangent to Γ using only a cyclos.

Proof. Mark points B1 , B2 , B3 and C on Γ. By Lemma 4, mark the point O, the centre of
Γ. Draw (Bi O) and (AOC) and mark the intersection denoted Fi ; there exists an index j
for which Fj 6= O; mark the point K which is the intersection of Bj Fj and OM where M
is the midpoint of AC (which we mark by Lemma 2) by Lemma 3. Clearly, K lies on A
tangent to Γ. Note that we can do this again to get multiple such points K by choosing
different C each time.

Lemma 6. Given a circle Γ and a point A on Γ, and a point B not on Γ, we can mark the
point C which is the second intersection of line AB and Γ using only a cyclos.

8
Proof. Mark the foot of perpendicular M from O onto line AB as done in Lemma 1. Mark
the intersection of line OM and AK by Lemma 3, where K is a point on the A-tangent to
Γ as constructed in Lemma 5. Draw (OAK) and mark the second intersection with Γ to
obtain C.

Lemma 7. Given points A, B, C not all on a line, we can draw the reflection of A in BC
using only a cyclos.

Proof. Draw (ABC) and mark the orthocentre H of ABC by Lemma 1. Mark the in-
tersection A0 of line AH with (BHC) using Lemma 6, which is the A-reflection in line
BC.

Lemma 8. Given points A, B, we can mark the point C which is the reflection of A in B
using only a cyclos.

Proof. Draw (AB), and by Lemma 6, mark two points K1 , K2 such that BKi is tangent
to (AB) for i ∈ {1, 2}. By Lemma 7, mark the reflection C of A in line K1 K2 as desired.

Thus, a cyclos can do everything a compass can: to draw a circle with given centre
A and given radius B, we use Lemma 8 to mark the reflection C of B in A and use the
cyclos to draw (BC) which has centre A and passes through B.

Solution 2. After deriving the first two lemmas in the previous solution, one can proceed
as follows: Let C denote the point such that A is the midpoint of CB.
Choose a generic point X. We can get the midpoint M of BX and the foot of perpendic-
ular D from X to AB. Draw the circle passing through A, D and M . This is the nine-point
of circle of triangle CBX. Intersect this circle with the circle whose diameter is BX. The
intersection point other than D is the foot of perpendicular E from B to CX. Note that
|AE| = |AB|. Similarly, we can construct another point F such that |AF | = |AB|. The
circle through B, E and F is the required circle.