IMO 2024 Solution Notes

Evan Chen《陳誼廷》
12 November 2024

This is a compilation of solutions for the 2024 IMO. The ideas of the
solution are a mix of my own work, the solutions provided by the competition
organizers, and solutions found by the community. However, all the writing
is maintained by me.
These notes will tend to be a bit more advanced and terse than the “official”
solutions from the organizers. In particular, if a theorem or technique is not
known to beginners but is still considered “standard”, then I often prefer to
use this theory anyways, rather than try to work around or conceal it. For
example, in geometry problems I typically use directed angles without further
comment, rather than awkwardly work around configuration issues. Similarly,
sentences like “let R denote the set of real numbers” are typically omitted
entirely.
Corrections and comments are welcome!

Contents
0 Problems 2

1 Solutions to Day 1 3
1.1 IMO 2024/1, proposed by Santiago Rodriguez (COL) . . . . . . . . . . . . 3
1.2 IMO 2024/2, proposed by Valentino Iverson (IDN) . . . . . . . . . . . . . 4
1.3 IMO 2024/3, proposed by William Steinberg (AUS) . . . . . . . . . . . . 6

2 Solutions to Day 2 10
2.1 IMO 2024/4, proposed by Dominik Burek (POL) . . . . . . . . . . . . . . 10
2.2 IMO 2024/5, proposed by Chu Cheuk Hei (HKG) . . . . . . . . . . . . . . 12
2.3 IMO 2024/6, proposed by Japan . . . . . . . . . . . . . . . . . . . . . . . 17

1
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

§0 Problems
1. Find all real numbers α so that, for every positive integer n, the integer

bαc + b2αc + b3αc + · · · + bnαc

is divisible by n.

2. For which pairs of positive integers (a, b) is the sequence

gcd(an + b, bn + a) n = 1, 2, . . .

eventually constant?

3. Let a1 , a2 , a3 , . . . be an infinite sequence of positive integers, and let N be a

positive integer. Suppose that, for each n > N , the number an is equal to the
number of times an−1 appears in the list (a1 , a2 , . . . , an−1 ). Prove that at least one
of the sequences a1 , a3 , a5 , . . . and a2 , a4 , a6 , . . . is eventually periodic.

4. Let triangle ABC with incenter I satisfying AB < AC < BC. Let X be a point
on line BC, different from C, such that the line through X and parallel to AC is
tangent to the incircle. Similarly, let Y be a point on line BC, different from B,
such that the line through Y and parallel to AB is tangent to the incircle. Line
AI intersects the circumcircle of triangle ABC again at P . Let K and L be the
midpoints of AC and AB, respectively. Prove that ∠KIL + ∠Y P X = 180◦ .

5. Turbo the snail is in the top row of a grid with 2024 rows and 2023 columns and
wants to get to the bottom row. However, there are 2022 hidden monsters, one in
every row except the first and last, with no two monsters in the same column.
Turbo makes a series of attempts to go from the first row to the last row. On each
attempt, he chooses to start on any cell in the first row, then repeatedly moves
to an orthogonal neighbor. (He is allowed to return to a previously visited cell.)
If Turbo reaches a cell with a monster, his attempt ends and he is transported
back to the first row to start a new attempt. The monsters do not move between
attempts, and Turbo remembers whether or not each cell he has visited contains a
monster. If he reaches any cell in the last row, his attempt ends and Turbo wins.
Find the smallest integer n such that Turbo has a strategy which guarantees being
able to reach the bottom row in at most n attempts, regardless of how the monsters
are placed.

6. A function f : Q → Q is called aquaesulian if the following property holds: for every

x, y ∈ Q,
f (x + f (y)) = f (x) + y or f (f (x) + y) = x + f (y).
Show that there exists an integer c such that for any aquaesulian function f there
are at most c different rational numbers of the form f (r) + f (−r) for some rational
number r, and find the smallest possible value of c.

2
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

§1 Solutions to Day 1
§1.1 IMO 2024/1, proposed by Santiago Rodriguez (COL)
Available online at https://aops.com/community/p31205921.

Problem statement

Find all real numbers α so that, for every positive integer n, the integer

bαc + b2αc + b3αc + · · · + bnαc

is divisible by n.

The answer is that α must be an even integer. Let S(n, α) denote the sum in question.

¶ Analysis for α an integer. If α is an integer, then the sum equals

n(n + 1)
S(n, α) = (1 + 2 + · · · + n)α = ·α
2
which is obviously a multiple of n if 2 | α; meanwhile, if α is an odd integer then n = 2
gives a counterexample.

¶ Main case. Suppose α is not an integer; we show the desired condition can never be
true. Note that replacing α with α ± 2 changes by

S(n ± 2, α) − S(n, α) = 2(1 + 2 + · · · + n) = n(n + 1) ≡ 0 (mod n)

for every n. Thus, by shifting appropriately we may assume −1 < α < 1 and α ∈
/ Z.

• If 0 < α < 1, then let m ≥ 2 be the smallest integer such that mα ≥ 1. Then

S(m, α) = 0 + · · · + 0 +1 = 1
| {z }
m−1 terms

is not a multiple of m.

• If −1 < α < 0, then let m ≥ 2 be the smallest integer such that mα ≤ −1. Then

S(m, α) = (−1) + · · · + (−1) +0 = −(m − 1)

| {z }
m−1 terms

is not a multiple of m.

3
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

§1.2 IMO 2024/2, proposed by Valentino Iverson (IDN)

Available online at https://aops.com/community/p31205957.

Problem statement

For which pairs of positive integers (a, b) is the sequence

gcd(an + b, bn + a) n = 1, 2, . . .

eventually constant?

The answer is (a, b) = (1, 1) only, which obviously works since the sequence is always 2.
Conversely, assume the sequence
xn := gcd(an + b, bn + a)
is eventually constant. The main crux of the other direction is to consider
M := ab + 1.

Remark (Motivation). The reason to consider the number is the same technique used in
IMO 2005/4, namely the idea to consider “n = −1”. The point is that the two rational
numbers
1 ab + 1 1 ab + 1
+b= , +a=
a b b a
have a large common factor: we could write “x−1 = ab + 1”, loosely speaking.
Now, the sequence is really only defined for n ≥ 1, so one should instead take n ≡ −1
(mod ϕ(M )) — and this is exactly what we do.

Obviously gcd(a, M ) = gcd(b, M ) = 1. Let n be a sufficiently large multiple of ϕ(M ) so

that
xn−1 = xn = xn+1 = · · · .
We consider the first three terms; the first one is the “key” one that gets the bulk of the
work, and the rest is bookkeeping and extraction.
• Consider xn−1 . Note that
a(an−1 + b) = an + ab ≡ 1 + (−1) ≡ 0 (mod M )
and similarly b(bn + a) ≡ 0 (mod M ). Hence M | xn−1 .
• Consider xn , which is now known to be divisible by M . Note that
0 ≡ an + b ≡ 1 + b (mod M )
0 ≡ bn + a ≡ 1 + a (mod M ).
So a ≡ b ≡ −1 (mod M ).
• Consider xn+1 , which is now known to be divisible by M . Note that
0 ≡ an+1 + b ≡ bn+1 + a ≡ a + b (mod M ).
We knew a ≡ b ≡ −1 (mod M ), hence this means 0 ≡ 2 (mod M ), so M = 2.
From M = 2 we then conclude a = b = 1, as desired.

4
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

Remark (No alternate solutions known). At the time nobody seems to know any solution
not depending critically on M = ab + 1 (or prime numbers dividing M , etc.). They vary in
execution once some term of the form xkϕ(n)−1 is taken, but avoiding the key idea altogether
does not currently seem possible.
A good example to consider for ruling out candidate ideas is (a, b) = (18, 9).

5
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

§1.3 IMO 2024/3, proposed by William Steinberg (AUS)

Available online at https://aops.com/community/p31206050.

Problem statement

Let a1 , a2 , a3 , . . . be an infinite sequence of positive integers, and let N be a positive

integer. Suppose that, for each n > N , the number an is equal to the number
of times an−1 appears in the list (a1 , a2 , . . . , an−1 ). Prove that at least one of the
sequences a1 , a3 , a5 , . . . and a2 , a4 , a6 , . . . is eventually periodic.

We present the solution from “gigamilkmen’tgeg” in https://aops.com/community/

p31224483, with some adaptation from the first shortlist official solution as well. Set
M := max(a1 , . . . , aN ).

¶ Setup. We will visualize the entire process as follows. We draw a stack of towers
labeled 1, 2, . . . , each initially empty. For i = 1, 2, . . . , we imagine the term ai as adding
a block Bi to tower ai .
Then there are N initial blocks placed, colored red. The rest of the blocks are colored
yellow: if the last block Bi was added to a tower that then reaches height ai+1 , the next
block Bi+1 is added to tower ai+1 . We’ll say Bi contributes to the tower containing Bi+1 .
In other words, the yellow blocks Bi for i > N are given coordinates Bi = (ai , ai+1 )
for i > N . Note in particular that in towers M + 1, M + 2, . . . , the blocks are all yellow.

28

24

18

26 14

22 10

20

16

12

9 13 17 23 27

N 21 11 15 19 25
1 2 3 4 5 6 7 8 9
M

We let h` denote the height of the `th tower at a given time n. (This is an abuse of
notation and we should write h` (n) at time n, but n will always be clear from context.)

¶ Up to alternating up and down. We start with two independent easy observations:

the set of numbers that occur infinitely often is downwards closed, and consecutive terms
cannot both be huge.

6
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

36

32

28

34 24

30 18

26 14

22 10

20

16

12

9 13 17 23 27 31 35

N 21 11 15 19 25 29 33
1 2 3 4 5 6 7 8 9 10 11
L M

Claim — If the (k + 1)st tower grows arbitrarily high, so does tower k. In fact,
there exists a constant C such that hk ≥ hk+1 − C at all times.

Proof. Suppose Bn is a yellow block in tower k + 1. Then with at most finitely many
exceptions, Bn−1 is a yellow block at height k + 1, and the block Br right below Bn−1 is
also yellow; then Br+1 is in tower k. Hence, with at most finitely many exceptions, the
map
Bn 7→ Bn−1 7→ Br 7→ Br+1
provides an injective map taking each yellow block in tower k + 1 to a yellow block in
tower k. (The figure above shows B32 → B31 → B19 → B20 as an example.)

Claim — If an > M then an+1 ≤ M .

Proof. Assume for contradiction there’s a first moment where an > M and an+1 > M ,
meaning the block Bn was added to an all-yellow tower past M that has height exceeding
M . (This is the X’ed out region in the figure above.) In Bn ’s tower, every (yellow)
block (including Bn ) was contributed by a block placed in different towers at height
an > M . So before Bn , there were already an+1 > M towers of height more than M .
This contradicts minimality of n.

It follows that the set of indices with an ≤ M has arithmetic density at least half, so
certainly at least some of the numbers must occur infinitely often. Of the numbers in
{1, 2, . . . , M }, define L such that towers 1 through L grow unbounded but towers L + 1
through M do not. Then we can pick a larger threshold N 0 > N such that

7
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

• Towers 1 through L have height greater than (M, N );

• Towers L + 1 through M will receive no further blocks;
• aN 0 ≤ L.
After this threshold, the following statement is true:

Claim (Alternating small and big) — The terms aN 0 , aN 0 +2 , aN 0 +4 , . . . are all at

most L while the terms aN 0 +1 , aN 0 +3 , aN 0 +5 , . . . are all greater than M .

¶ Automaton for n ≡ N 0 (mod 2). From now on we always assume n > N 0 . When
n ≡ N 0 (mod 2), i.e., when an is small, we define the state
S(n) = (h1 , h2 , . . . , hL ; an ).
For example, in the figure below, we illustrate how
S(34) = (9, 11; a34 = 1) −→ S(36) = (9, 12; a36 = 2)

36

32

28

34 24

30 18

26 14

22 10

20

16

12

9 13 17 23 27 31 35

N 21 11 15 19 25 29 33
1 2 3 4 5 6 7 8 9 10 11
L M

The final element an simply reminds us which tower was most recently incremented.
At this point we can give a complete description of how to move from S(n) to S(n + 2):
• The intermediate block Bn+1 is placed in the tower corresponding to the height
an+1 of Bn ;
• That tower will have height an+2 equal to the number of towers with height at
least an+1 ; that is, it equals the cardinality of the set
{i : hi ≥ han }

8
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

• We increment han+2 by 1 and update an .

For example, the illustrated S(34) → S(36) corresponds to the block B34 at height h1
in tower 1 giving the block B35 at height 2 in tower h1 , then block B36 at height h2 + 1
being placed in tower 2.

¶ Pigeonhole periodicity argument. Because only the relative heights matter in the
automata above, if we instead define

T (n) = (h1 − h2 , h2 − h3 , . . . , hL−1 − hL ; an ).

then T (n + 2) can be determined from just T (n).

So it would be sufficient to show T (n) only takes on finitely many values to show that
T (n) (and hence an ) is eventually periodic.
Since we have the bound hk+1 ≤ hk + C, we are done upon proving the following lower
bound:

Claim — For every 1 ≤ ` < L and n > N 0 , we have h` ≤ h`+1 + C · (L − 1).

Proof. Assume for contradiction that there is some moment n > N 0 such that

h` > h`+1 + C · (L − 1)

and WLOG assume that h` was just updated at the moment n. Together with hk+1 ≤
hk + C for all k and triangle inequality, we conclude

min(h1 , . . . , h` ) > q := max(h`+1 , . . . , hL ).

We find that the blocks now in fact alternate between being placed among the first `
towers and in towers with indices greater than q thereafter. Hence the heights h`+1 , . . . ,
hL never grow after this moment. This contradicts the definition of L.

Remark. In fact, it can be shown that the period is actually exactly L, meaning the periodic
part will be exactly a permutation of (1, 2, . . . , L). For any L, it turns out there is indeed a
permutation achieving that periodic part.

9
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

§2 Solutions to Day 2
§2.1 IMO 2024/4, proposed by Dominik Burek (POL)
Available online at https://aops.com/community/p31218657.

Problem statement

Let triangle ABC with incenter I satisfying AB < AC < BC. Let X be a point
on line BC, different from C, such that the line through X and parallel to AC is
tangent to the incircle. Similarly, let Y be a point on line BC, different from B,
such that the line through Y and parallel to AB is tangent to the incircle. Line
AI intersects the circumcircle of triangle ABC again at P . Let K and L be the
midpoints of AC and AB, respectively. Prove that ∠KIL + ∠Y P X = 180◦ .

Let T be the reflection of A over I, the most important point to add since it gets rid of
K and L as follows.

Claim — We have ∠KIL = ∠BT C, and lines T X and T Y are tangent to the
incircle.

Proof. The first part is true since 4BT C is the image of 4KIL under a homothety of
ratio 2. The second part is true because lines AB, AC, T X, T Y determine a rhombus
with center I.

We thus delete K and L from the picture altogether; they aren’t needed anymore.

K L

B C
X Y

10
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

Claim — We have BXP T and CY P T are cyclic.

Proof. ]T Y C = ]T Y B = ]ABC = ]AP C = ]T P C and similarly. (Some people call

this Reim’s theorem.)

To finish, observe that

]CT B = ]CT P + ]P T B = ]CY P + ]P XB = ]XY P + ]XY P = ]XP Y

as desired. (The length conditions AC > AB > BC ensure that B, X, Y , C are collinear
in that order, and that T lies on the opposite side of BC as A. Hence the directed
equality ]CT B = ]XP Y translates to the undirected ∠BT C + ∠XP Y = 180◦ .)

11
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

§2.2 IMO 2024/5, proposed by Chu Cheuk Hei (HKG)

Available online at https://aops.com/community/p31218774.

Problem statement

Turbo the snail is in the top row of a grid with 2024 rows and 2023 columns and
wants to get to the bottom row. However, there are 2022 hidden monsters, one in
every row except the first and last, with no two monsters in the same column.
Turbo makes a series of attempts to go from the first row to the last row. On
each attempt, he chooses to start on any cell in the first row, then repeatedly moves
to an orthogonal neighbor. (He is allowed to return to a previously visited cell.) If
Turbo reaches a cell with a monster, his attempt ends and he is transported back to
the first row to start a new attempt. The monsters do not move between attempts,
and Turbo remembers whether or not each cell he has visited contains a monster. If
he reaches any cell in the last row, his attempt ends and Turbo wins.
Find the smallest integer n such that Turbo has a strategy which guarantees being
able to reach the bottom row in at most n attempts, regardless of how the monsters
are placed.

Surprisingly the answer is n = 3 for any grid size s × (s − 1) when s ≥ 4. We prove this
in that generality.

¶ Proof that at least three attempts are needed. When Turbo first moves into the
second row, Turbo could encounter a monster M1 right away. Then on the next attempt,
Turbo must enter the third row in different column as M1 , and again could encounter a
monster M2 right after doing so. This means no strategy can guarantee fewer than three
attempts.

¶ Strategy with three attempts. On the first attempt, we have Turbo walk through
the entire second row until he finds the monster M1 in it. Then we get two possible cases.

Case where M1 is not on the edge. In the first case, if that monster M1 is not on
the edge of the row, then Turbo can trace two paths below it as shown below. At least
one of these paths works, hence three attempts is sufficient.

Starting row

M1

Goal row

Case where M1 is on the edge. WLOG, M1 is in the leftmost cell. Then Turbo
follows the green staircase pattern shown in the left figure below. If the staircase is free

12
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

of monsters, then Turbo wins on the second attempt. Otherwise, if a monster M2 is

encountered on the staircase, Turbo has found a safe path to the left of M2 ; then Turbo
can use this to reach the column M1 is in, and escape from there. This is shown in purple
in the center and right figure (there are two slightly different cases depending on whether
M2 was encountered going east or south).
Starting row Starting row Starting row
M1 M1 M1

M2
M2

Goal row Goal row Goal row

Thus the problem is solved in three attempts, as promised.

¶ Extended remark: all working strategies look similar to this. As far as we know, all
working strategies are variations of the above. In fact, we will try to give a description of
the space of possible strategies, although this needs a bit of notation.
Definition. For simplicity, we use s even only in the figures below. We define the happy
triangle as the following cells:
• All s − 1 cells in the first row (which has no monsters).
• The center s − 3 cells in the second row.
• The center s − 5 cells in the third row.
• ...
s th
• The center cell in the 2 row.
For s = 12, the happy triangle is the region shaded in the thick border below.
Starting row

Goal row

13
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

Definition. Given a cell, define a shoulder to be the cell directly northwest or northeast
of it. Hence there are two shoulders of cells outside the first and last column, and one
shoulder otherwise.
Then solutions roughly must distinguish between these two cases:
• Inside happy triangle: If the first monster M1 is found in the happy triangle,
and there is a safe path found by Turbo to the two shoulders (marked F in the
figure), then one can finish in two more moves by considering the two paths from F
that cut under the monster M1 ; one of them must work. This slightly generalizes
the easier case in the solution above (which focuses only on the case where M1 is
in the first row).
Starting row
⋆ ⋆

M1

Goal row

• Outside happy triangle: Now suppose the first monster M1 is outside the happy
triangle. Of the two shoulders, take the one closer to the center (if in the center
column, either one works; if only one shoulder, use it). If there is a safe path to
that shoulder, then one can take a staircase pattern towards the center, as shown in
the figure. In that case, the choice of shoulder and position guarantees the staircase
reaches the bottom row, so that if no monster is along this path, the algorithm
ends. Otherwise, if one encounters a second monster along the staircase, then one
can use the third trial to cut under the monster M1 .
Starting row Starting row

⋆ ⋆

M1 M1

M2

Goal row Goal row

We now prove the following proposition: in any valid strategy for Turbo, in the case
where Turbo first encounters a monster upon leaving the happy triangle, the second path
must outline the same staircase shape.

14
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

The monsters pre-commit to choosing their pattern to be either a NW-SE diagonal or

NE-SW diagonal, with a single one-column gap; see figure below for an example. Note
that this forces any valid path for Turbo to pass through the particular gap.

Starting row

M
M
M
M
M
M
M
M
M
M
Goal row

We may assume without loss of generality that Turbo first encounters a monster M1
when Turbo first leaves the happy triangle, and that this forces an NW-SE configuration.

Starting row
(M )

(M )

(M )

M1

X ♣
♣
♣
♣
♣
♣
M2

Goal row

Then the following is true:

15
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

Proposition
The strategy of Turbo on the second path must visit every cell in “slightly raised
diagonal” marked with ♣ in the figure above in order from top to bottom, until
it encounters a second Monster M2 (or reaches the bottom row and wins anyway).
It’s both okay and irrelevant if Turbo visits other cells above this diagonal, but the
marked cells must be visited from top to bottom in that order.

Proof. If Turbo tries to sidestep by visiting the cell southeast of M1 (marked X in the
Figure), then Turbo clearly cannot finish after this (for s large enough). Meanwhile,
suppose Turbo tries to “skip” one of the ♣, say in column C, then the gap could equally
well be in the column to the left of C. This proves the proposition.

Remark (Memories of safe cells are important, not just monster cells). Here is one additional
observation that one can deduce from this. We say a set S of revealed monsters is called
obviously winnable if, based on only the positions of the monsters (and not the moves or
algorithm that were used to obtain them), one can identify a guaranteed winning path
for Turbo using only S. For example, two monsters in adjacent columns which are not
diagonally adjacent is obviously winnable.
Then no strategy can guarantee obtaining an obviously winnable set in 2 moves (or even
k moves for any constant k, if s is large enough in terms of k). So any valid strategy must
also use the memory of identified safe cells that do not follow just from the revealed monster
positions.

16
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

§2.3 IMO 2024/6, proposed by Japan

Available online at https://aops.com/community/p31218720.

Problem statement

A function f : Q → Q is called aquaesulian if the following property holds: for every

x, y ∈ Q,
f (x + f (y)) = f (x) + y or f (f (x) + y) = x + f (y).
Show that there exists an integer c such that for any aquaesulian function f there
are at most c different rational numbers of the form f (r) + f (−r) for some rational
number r, and find the smallest possible value of c.

We will prove that

{f (x) + f (−x) | x ∈ Q}
contains at most 2 elements and give an example where there are indeed 2 elements.
We fix the notation x → y to mean that f (x + f (y)) = f (x) + y. So the problem
statement means that either x → y or y → x for all x, y. In particular, we always have
x → x, and hence
f (x + f (x)) = x + f (x)
for every x.

¶ Construction. The function

f (x) = b2xc − x

can be seen to satisfy the problem conditions. Moreover, f (0) + f (0) = 0 but f (1/3) +
f (−1/3) = −1.

Remark. Here is how I (Evan) found the construction. Let h(x) := x + f (x), and let
S := h(Q) = {h(x) | x ∈ Q}. Hence f is the identity on all of S. If we rewrite the problem
condition in terms of h instead of f , it asserts that at least one of the equations

h(x + h(y) − y) = h(x) + h(y)

h(y + h(x) − x) = h(x) + h(y)

is true. In particular, S is closed under addition.

Now, the two trivial solutions for h are h(x) = 2x and h(x) = 0. To get a nontrivial
construction, we must also have S 6= {0} and S 6= Q. So a natural guess is to take S = Z.
And indeed h(x) = b2xc works fine.

Remark. This construction is far from unique. For example, f (x) = 2 bxc − x = bxc − {x}
seems to have been more popular to find.

¶ Proof (communicated by Abel George Mathew). We start by proving:

Claim — f is injective.

17
IMO 2024 Solution Notes web.evanchen.cc, updated 12 November 2024

Proof. Suppose f (a) = f (b). WLOG a → b. Then

f (a) + a = f (a + f (a)) = f (a + f (b)) = f (a) + b =⇒ a = b. .

Claim — Suppose s → r. Then either f (r) + f (−r) = 0 or f (f (s)) = s + f (r) +

f (−r).

Proof. Take the given statement with x = s + f (r) and y = −r; then

x + f (y) = s + f (r) + f (−r)

y + f (x) = f (s + f (r)) − r = f (s).

Because f is injective, if x → y then f (r) + f (−r) = 0. Meanwhile, if y → x then indeed

f (f (s)) = s + f (r) + f (−r).

Finally, suppose a and b are different numbers for which f (a) + f (−a) and f (b) + f (−b)
are both nonzero. Again, WLOG a → b. Then
a→a a→b
f (a) + f (−a) = f (f (a)) − a = f (b) + f (−b).

This shows at most two values can occur.

Remark. The above solution works equally well for f : R → R. But the choice of Q permits
some additional alternate solutions.

Remark. After showing f injective, a common lemma proved is that −f (−f (x)) = x, i.e.
f is an involution. This provides some alternative paths for solutions.

18