How to Compose a Problem

Hung-Hsun Hans Yu
February 16, 2023

Each year, there is endless need for quality and novel problems in TSTs and contests
all around the world. For example, the entire Taiwan team selection process now requires
47 problems a year (it was 36 before). Even if, say, 25 problems are chosen from the
ISL, there are still 22 slots to fill in. On the other hand, coming up with a good problem
is not that easy, and very often the problem selection committee faces a lack of quality
problems. I thus think it is important for people done with contests to learn a bit and
compose some problems.
In this note, I will discuss some ways that one could compose a problem in with the
thought process behind the problems I proposed before. The purpose of this note is
mainly to share my two cents about problem-making to inspire people who are interested
in doing so. The hope is that people who are done with contests can propose some
interesting problems, and people who are still doing competitive math can also come up
with some interesting problems to think about themselves.

1 Ask a Simple Question

The first way to come up with a problem is pretty simple—ask something that you would
naturally ask! This works especially well for extremal combinatorics, as all you need to
do is to come up with some model, and ask how big or how small a set can be.
The advantage of this method is that the problem would look elegant as long as the
model used is simple enough. However, decent problem solving skills are needed as you
would need to solve the problem by yourself without knowing how hard the problem is.
Sometimes you might also need to realize that the problem is too hard for Olympiads!
In the following, each subsection discusses what motivates the question I ask for each
problem I proposed.

1.1 2021 TMO P2

For two subsets A, B of , we define
A + B = {a + b | a ∈ A, b ∈ B},
A · B = {a · b | a ∈ A, b ∈ B}.
We can easily show that |A + B| ≥ |A| + |B| − 1 and |A · B| ≥ |A| + |B| − 1. For the first
inequality, the equality holds when A, B are arithmetic progressions of the same common
difference. For the second one, the equality holds when A, B are geometric progressions
of the same common ratio. The two scenario looks strikingly difference, and it thus makes
sense to ask if they can both be small. The sum-product phenomenon basically states
that it is impossible to both have small |A + B| and small |A · B|. In fact, Solymosi
proved the following:

Hung-Hsun Hans Yu 1
Theorem 1.1 (Solymosi). For any finite set A, B ⊆, we have

|A|4/3
max(|A + B|, |A · B|) ≥ .
2⌈log2 |A|⌉1/3

The Erdős-Szemerédi conjecture states that the inequality still holds when the right
hand side is replaced by Ω(|A|2−ε ) for any ε > 0, but this is still wide open.
We can also ask this over Z/nZ. We know that there is a proper subset S of Z/nZ
such that S · S = S—we can simply take S to be all the quadratic residues. Then our
intuition from our experience with the sum-product problem in reals tells us that S is
unlikely an arithmetic progression. Is this actually the case? Can we prove this only
using tools available in Olympiads? The answer is (mostly) yes if S includes 0, and this
is 2021 TMO P2.
Problem 1.1 (2021 TMO P2). Find all integers n = 2k + 1 > 1 so that there exists a
permutation a0 , a1 , . . . , ak of 0, 1, . . . , k such that

a21 − a20 ≡ a22 − a21 ≡ · · · ≡ a2k − a2k−1 (mod n).

1.2 2021 Taiwan TST Round 1 Mock Exam P6

In additive combinatorics, something like the following is usually asked: Given a subset
of integers (or a given abelian group), does this subset contain a constellation of some
shapes? Here, a shape is some finite subset V of integers, and a constellation of a shape
V is a set of the form {a + bv|v ∈ V } for some integer a and nonzero integer b. For
example, Roth’s theorem states that any set of positive integers of positve upper density
contains a nonzero 3-term arithmetic progression. This can in fact be rephrased as the
following.
Theorem 1.2 (Roth). Any subset of N of positive upper density contains a constellation
of shape {0, 1, 2}.

We could also ask if the set contains constellations of other shapes. In fact, Szemerédi
shows the following.
Theorem 1.3 (Szemerédi). Any subset of N of positive upper density contains a con-
stellation of any given finite shape.

The famous Green–Tao theorem can also be restated as the following.

Theorem 1.4 (Green–Tao). The set of primes contains a constellation of any given
finite shape.

Well, we could also consider those questions over Z/N Z, and see what we can say
about the size of a subset of Z/N Z that contains a constellation of any shapes of size k.
To make the discussion simpler, I instead consider “ordered shapes” (v1 , . . . , vn ) and k
copies of Z/N Z, and I require a + bvi lies in the chosen subset of the i-th copy. I also set
b = 1 to make the question simpler. Then a simple counting argument shows that the
sum of the sizes of the subsets is at least kN (k−1)/k . It thus makes sense to set N to be a
k-th power and see if this is the optimal bound. After coming up with the construction
meeting this bound, we are done composing the problem!

Hung-Hsun Hans Yu 2
Problem 1.2 (2021 TST Round 1 Mock P6). Let n be a positive integer and N = n2021 .
There are 2021 concentric circles centered at O, and N equally-spaced rays are emitted
from point O. Among the 2021N intersections of the circles and the rays, some are
painted red while the others remain unpainted.
It is known that, no matter how one intersection point from each circle is chosen,
there is an angle θ such that after a rotation of θ with respect to O, all chosen points are
moved to red points. Prove that the minimum number of red points is 2021n2020 .

1.3 2021 Taiwan TST Round 3 Independent Study I-C

The model we start with is the following: for any n points on a plane, we can also view
them as vertices of a directed graph, where for each point we draw an edge to the closest
point among the others. To make the setting simpler, assume that there is one unique
closest point for any point. We can then ask: what properties does this graph have?
First, notice that if the graph is not connected, then we can always divide the con-
figuration into two smaller configurations. Therefore, we might want to assume that the
graph is connected. Now, it is clear that the graph has n directed edges, showing that
there must be at least one directed cycle. A simple argument shows that this directed
cycle has to be of length two.
Everything up to this point is pretty trivial. Now, we can start to consider some
quantities associated to this graph and see if we can say anything clever about it. The
quantity I considered was the number k of vertices whose in-degree is positive. Clearly
there can be at most n − 1 such vertices, so I turned to consider the minimum of the
number.
Since the number of edges is fixed, the most naïve thing to try is to bound the in-
degree of a vertex. With a simple argument, we can show that the in-degree is bounded
by 5. This shows that 5k ≥ n and so k ≥ n/5. However, there was not a matching
example to this bound. The best we could get is k = (1/4 + o(1))n where each city
with positive in-degree has in-degree 4 (well, except for one of them). What’s going on
here is that in fact, we can show that when viewed as a simple undirected graph, the
graph we are considering has maximum degree bounded by 5. Then we have a matching
upper bound and lower bound up to some smaller order in n. This leads to the following
problem.
Problem 1.3 (2021 TST Round 3 Independent I-C). A city is a point on the plane.
Suppose there are n ≥ 2 cities. Suppose that for each city X, there is another city
N (X) that is strictly closer to X than all the other cities. The government builds a road
connecting each city X and its N (X); no other roads have been built. Suppose we know
that, starting from any city, we can reach any other city through a series of road.
We call a city Y [i]suburban[/i] if it is N (X) for some city X. Show that there are at
least (n − 2)/4 suburban cities.

1.4 2022 TMO P3

In geometry, or in first year linear algebra classes, we ask: what are all the maps from
R2 to R2 that preserve lengths or inner products? This kind of maps is usually called
isometries. If this question is raised in a linear algebra class, then usually we also assume
that the map is affine, and we deduce that all isometries of the 2-dimensional Euclidean
space are translations, rotations and sliding reflections. Well, the affine assumption is

Hung-Hsun Hans Yu 3
actually not needed if we just view this as a geometry problem, where we notice that
knowing where the isometry sends two points to already almost determines the entire
isometry.
The usual inner product is not the only way we measure the length of a vector, or the
distance between two points. In fact, one measurement that also appears a lot in math
contests is the Manhattan distance, or equivalently the L1 norm. The simple question we
ask is thus: what are all the maps from R2 to R2 that preserves the Manhattan distance
between two points? It turns out that the same geometric insight solves this immediately,
but if we write it in terms of functional equations, then this would actually increase the
difficulty by a lot for students with little exposure to those kinds of natural questions
asked by mathematicians.
Problem 1.4 (2022 TMO P3). Find all functions f, g : R2 → R satisfying that

|f (a, b) − f (c, d)| + |g(a, b) − g(c, d)| = |a − c| + |b − d|

for all real numbers a, b, c, d.

I was told by Yi Kung later that in fact there is a result showing that any endormor-
phism of a normed space that preserves norms is an affine map. This is the Mazur–Ulam
theorem. Well, I do not know anything in functional analysis and thought that people
would always assume affineness in these settings, but it seems that I am totally wrong
about this.

1.5 2022 Taiwan TST Round 1 Mock Exam P5

To be completely honest, I’ve already forgot what I was thinking about precisely. This
is my issue with geometry problems—some of them are really not that memorable (es-
pecially the ones that are less good). I think the entire motivation is the following: let
X, Y, Z be points on BC, CA, AB, then by Miquel’s theorem we know that AY Z, XBZ, XY C
intersect at a point M . If D, E, F are points on AY Z, XBZ, XY C, respectively, what is
the condition for M DEF to be concyclic?
After playing a bit around with GGB, I found that the answer is that this happens
when AD, BE, CF intersect at a common point P on the circumcircle of DEF . This can
be proved by a simple angle chase. The goal now is then try to hide this diagram in some
other diagram. I ended up choosing P to be the orthocenter to hide the condition that
DEF P are concyclic—this now becomes AHEF are concyclic in the proposed version.
So my old D degenerates to point A on AH, and my old P is H, and the problem is just
the specialization of the observation above.
Problem 1.5 (2022 TST Round 1 Mock P5). Let H be the orthocenter of a given
triangle ABC. Let BH and AC meet at a point E, and CH and AB meet at F .
Suppose that X is a point on the line BC. Also suppose that the circumcircle of triangle
BEX and the line AB intersect again at Y , and the circumcircle of triangle CF X and
the line AC intersect again at Z. Show that the circumcircle of triangle AY Z is tangent
to the line AH.

I don’t know about other people, but for me, it seems that this way of composing a
problem usually leads to a simple problem when it comes to geometry.

Hung-Hsun Hans Yu 4
1.6 2022 Taiwan TST Round 3 Mock Exam P6
The motivation question of this is easy: how “competitive” can a “streaming algorithm”
be for determining the median of a sequence? Here, the competitivity is determined by
its distance it outputs from the actual median, and it is streaming in the sense that the
sequence is given term by term, and we can choose a term to be our answer only after
it is immediately given. One other way to see this is that: if we have really limited
memory for the algorithm, so limited that we can only store constantly many integers in
the memory, how well can we do in determining the median?
Well, in order for any algorithm of this kind to work, we need to know beforehand
the range of the numbers n and the length of the sequence 2k + 1. It is also pretty easy
to see that when k is fixed and n is sufficiently large, we cannot do much better than
n/2; and in order to achieve n/2, we can simply cut the interval in half, and choose an
element when the half it belongs to already has k + 1 elements. Therefore, I decided to
see if I can actually determine the exact answer. It turned out pretty surprisingly that
it is indeed possible to determine the answer completely, and that the answer turned out
to be piecewise “linear” (in quotation marks as we need to round it to integers).
Problem 1.6 (2022 TST Round 3 Mock P6). Positive integers n and k satisfying
n ≥ 2k + 1 are known to Alice. There are n cards with numbers from 1 to n, randomly
shuffled as a deck, face down. On her turn, she does the following in order:
(i) She first flips over the top card of the deck, and puts it face up on the table. (ii)
Then, if Alice has not signed any card, she can sign the newest card now.
The game ends after 2k + 1 turns, and Alice must have signed on some card. Let A
be the number on the signed cards, and M be the (k + 1)st largest number among all
2k + 1 face-up cards. Alice’s score is |M − A|, and she wants the score to be as close to
zero as possible.
For each (n, k), find the smallest integer d = d(n, k) such that Alice has a strategy to
guarantee her score no greater than d.

1.7 2023 TMO P5

This problem came out when I encountered the following vague problem during research:
given the average of the first few powers of some real numbers, what are the possible
positive integers n so that there actually exist those n real numbers with the prescribed
“moments”.
If we want them to be non-negative, then the actual answer when the first three
moments are given is already complicated enough. A much cleaner question is to ask for
the “equality cases”: if now we allow the average to be weighted (so we are thinking about
distributions with finite support), when do the first few moments uniquely determine the
distributions? Now this becomes much easier: one can show that this happens precisely
when one can show that E[(x − a1 )2 · · · (x − an )2 ] = 0 using the given moments, showing
that the distribution has to be supported on a1 , . . . , an and this would in tern uniquely
determine the distribution. Note that when we are only given the first two moments, this
basically says that Cauchy inequality is the only constraint on the moments.
Well, Cauchy inequality is sort of boring, so I decided to go to the fourth moment. I
want to find an interesting distribution supported on two reals so that E[x], E[x2 ], E[x3 ]
and E[x4 ] are all not too ugly. I decided that my definition of interesting is that the
distribution is a with probability n/7 and b with probability 1 − n/7 for some a, b ∈ R

Hung-Hsun Hans Yu 5
and n = 1, . . . , 6. I also wanted that E[x], E[x2 ], E[x3 ], E[x4 ] to be integers and possibly
small. Playing with the numbers a bit gives me a = 3, b = −4, n = 5. This leads to the
following problem.
Problem 1.7 (2023 TMO P5). Let m be a positive integer, and real numbers a1 , a2 , . . . , am
satisfy
1 ∑
m
ai = 1,
m i=1

1 ∑ 2
m
a = 11,
m i=1 i

1 ∑ 3
m
a = 1,
m i=1 i
and
1 ∑ 4
m
a = 131.
m i=1 i
Prove that m is a multiple of 7.

2 Start From a Trick or a Property

One other way is to start from some known fact or trick and build up a problem from
there. Usually, this consists of some of the following.

1. Try to wrap the fact up with some fancy stuff so it is not so obvious how the
problem is composed.

2. Take a special case of the problem to make it harder to see how the problem is
composed.

3. Test-solve it once again to see if there are other solutions that appear after morphing
the problem.

This works especially well for geometry problems, as this makes discovering a true and
interesting statement way easier, and wrapping geometry problems up is in general easier.
The advantage of this method is that you don’t need to worry about coming up with
a solution—you already have one in mind when you are composing it. The drawback of
this is that the problem might not look as elegant and natural, and sometimes there are
some unexpected solutions that totally mess up the difficulty (so make sure that you find
other people to test-solve).

2.1 2021 Taiwan TST Round 2 Independent Study I-G

For every triangle with unequal sides, there is a special line associated to it called the Euler
line. An intuitive question is: is Euler line the locus of some moving point? Suppose
that the moving point P is given by the intersection of BQ and CR, where Q, R are
some other moving points. We know that O, G, H are points on the Euler line. Note
that O = BO ∩ CO and G = BMb ∩ CMc . This motivates us to constrain Q, R on

Hung-Hsun Hans Yu 6
the perpendicular bisector ℓb , ℓc of AC, AB, respectively. Then we can also see that
H = B∞ℓb ∩ C∞ℓc .
Now we see that BQ∩CR does lie on the Euler line if (Q, R) = (O, O), (Mb , Mc ), (∞ℓb , ∞ℓc ).
The last question to ask here is what constraint is needed to put on Q and R. It is not
hard to see that we just need QR parallel to BC by the moving point technique. Thus we
have our first version of the problem: Given a triangle ABC and a line ℓ parallel to BC.
If Q, R are the intersections ℓ ∩ ℓb , ℓ ∩ ℓc respectively (where ℓb , ℓc are the perpendicular
bisectors of AC, AB, respectively), then BQ ∩ CR lies on the Euler line of ABC.
Well, this version kind of exposes the moving-point nature too much. To hide it a
bit, we can think of some ways to determine Q and R such that they are still points on
the perpendicular bisectors that determine the line parallel to BC. Since they are on
the perpendiculars, it is natural to set Q, R as circumcenters of ACD, ABD for some
point D. The condition that QR is parallel to BC can then be replaced by that BC is
perpendicular to AD. Note that in this case, due to symmetry, we know that the Euler
line of BCD also passes through BQ ∩ CR. After relabeling, we have the version that
appears on the test!
Problem 2.1 (2021 TST Round 2 Independent I-G). Let ABCD be a convex quadri-
lateral with pairwise distinct side lengths such that AC ⊥ BD. Let O1 , O2 be the
circumcenters of ∆ABD, ∆CBD, respectively. Show that AO2 , CO1 , the Euler line of
∆ABC and the Euler line of ∆ADC are concurrent.

In fact, we can hide this configuration in any configuration that has four points
A, B, C, D with AC ⊥ BD. However, I did not find a good configuration to hide this
property in, so I decided to propose this version.

2.2 2021 Taiwan TST Round 2 Independent I-N

I was kind of intrigued by the trick used in 2020 IMO P5, so I was wondering if I could
come up with something similar to that.
Some important structures that are used in that problem are the homogeneity of the
conditions, some inequalities that can be cleverly used to limit the choice of the cards,
and simple application of the p-adic analysis. To mimic those structures, I also consider
some positive integers, allow one to choose two numbers a, b out of it, and assert that
there exists a subset of the given positive integers that meet some sort of divisibility
condition. The simplest one I came up with that is homogeneous and not boring is the
condition
∃c s.t. c2 | a(a + b).
Being inspired by the solution to IMO P5, I try to utilize some inequalities to constrain
the choice of c. It seems that to do so, we would want to choose a, b small. However, if
a = b then we can always choose c = a. Moreover, if a | b, then we can also choose c = a.
Therefore this becomes hard when the smallest number divides all the other numbers.
In fact, if this is the case, then we can always choose c to be the smallest number! I
conjectured that the smallest number always has to divide the other numbers given the
condition, and it turned out to be the case.
Problem 2.2 (2021 TST Round 2 Independent I-N). Let S be a set of positive integers
such that for every a, b ∈ S, there always exists c ∈ S such that c2 divides a(a + b). Show
that there exists an a ∈ S such that a divides every element of S.

Hung-Hsun Hans Yu 7
 Because of the discussion above, to prove it, one natural approach is to take a to be
the smallest number and b to be the smallest number not divisible by a. Then we must
have c ≥ b and so a(a + b) = c2 or 2c2 . This is not the contradiction yet, but it is not
hard to go from here to show that such b must not exist, and we are done composing yet
another problem once again!

2.3 2021 Taiwan TST Round 3 Independent Study II-G

I wanted to compose a problem that can be solved using the butterfly theorem. Recall
that the theorem asserts that if XY is a chord in a circle with midpoint M and ABCD
are four points on the circle so that P = AB ∩ XY, Q = CD ∩ XY are symmetric w.r.t.
M , then AD ∩ XY, BC ∩ XY are also aymmetric w.r.t. M . To hide this, we can take
B ′ , C ′ to be the reflections of B, C w.r.t. the perpendicular bisector of XY . Then we have
that AD, B ′ C ′ , XY are concurrent. If we choose specifically that D = A,√then we can
actually see that B ′ , C ′ are the images of P, Q under the composition of the bc-inversion
and the reflection about the angle bisector of ∠BAC. We shall then choose P, Q on XY
such that P M = M Q and that B ′ , C ′ are some special points. If we choose P, Q to be the
projections of the incenter and the A-excenter of AXY , then B ′ , C ′ are the points where
the two mixtilinear circles tangent to the circumcircle. We thus get that AA ∩ XY, B ′ , C ′
are collinear.
Now note that B ′ , C ′ are the center of the homotheties sending the circumcircle to
the two mixtilinear circles. Therefore, on B ′ C ′ there is also the center of one of the
homotheties sending one mixtilinear circle to the other one. We can then construct the
homothety center S using some other methods and ask them to show that the line passing
through S and AA ∩ XY also passes through B ′ , C ′ .
Problem 2.3 (2021 TST Round 3 Independent II-G). Let ABC be a triangle with
AB < AC, and let Ia be its A-excenter. Let D be the projection of Ia to BC. Let X be
the intersection of AIa and BC, and let Y, Z be the points on AC, AB, respectively, such
that X, Y, Z are on a line perpendicular to AIa . Let the circumcircle of AY Z intersect
AIa again at U . Suppose that the tangent of the circumcircle of ABC at A intersects
BC at T , and the segment T U intersects the circumcircle of ABC at V . Show that
∠BAV = ∠DAC.

It turns out that the butterfly

√ theorem is not needed at all—everything can be killed
easily once one considers the bc-inversion. It does not matter that much though. After
all, the solution does not have to be exactly like the motivating idea. It is fine as long as
those alternative solutions are discovered during test-solving sessions.

2.4 2021 Taiwan TST Round 3 Independent Study II-N

The motivation comes from a research question I encountered. A toy problem of the
question I encountered is the following.
Theorem 2.1. Let V ⊆ C[x, y] be a finite dimensional vector subspace so that there
exists (x, y) ∈ C2 such that f (x, y) = 0 for all f ∈ V . If any nonzero polynomial in V is
not irreducible, then there exists a nontrivial common factor among all the elements in
V.

Hung-Hsun Hans Yu 8
 The proof basically goes like this. We first focus on the case where V is of dimension 2.
Then we need to show that if f, g ̸= 0, V (f ) ∩ V (g) ̸= ∅ and af + bg is not irreducible for
all [a : b] ∈ P1 , then gcd(f, g) is non-constant. This follows from some algebrogeometric
fact. Once we have this, we can actually use it to show the general case. To see this, we
pick an arbitrary nonzero f ∈ V , and look at all the prime divisors of f . If for any prime
divisor p of f I can find gp ∈ V such that p ∤ gp , then I can find some linear combination
∑
g= ap gp
p prime divisor of f

such that gcd(f, g) = 1. By the contrapositive of the 2-dimensional case, we know that
there exists [a : b] ∈ P1 such that af + bg is irreducible. Thus
∑
af + bap gp
p prime divisor of f

is irreducible and lies in V , which is a contradiction.

Well, the first step is definitely too hard for an Olympiad-level problem. However, the
central ideal of the second step is pretty elementary. To reduce the general case to the
2-dimensional case, all we need is to find a coprime linear combination and that “linear
coefficient is multiplicative.” Though we are currently in a linear algebra setting, we can
extract this and put it in somewhere that has similar structure.
To do so, we first think about what we should replace C[x, y] with. Since we need
the structure of unique factorization, the object that is most reasonable to consider here
is simply the set of (positive) integers. Now, we need to think about what we should
replace the “linear coefficients” with. Of course we could simply consider all the integral
linear combination, but that’s too much freedom provided. In fact, all we need is a
multiplicative set with some additional properties to guarantee the existence of coprime
combination. It seems most natural to pick the set of perfect n-th powers. Then one can
see that the argument in the reduction still works fine, and we just need to come up with
some statement regarding axn + by n for two given positive integers a, b.
To differentiate the case gcd(a, b) = 1 and gcd(a, b) has lots of prime divisors, it seems
more natural do consider the following statement: if gcd(a, b) = 1, can we find positive
n n
integers x, y so∑that ax + by has few prime divisors? I tried to estimate the density,
but it fails as 1/p diverges. There also does not seem to ∑be any other elementary tools
available to deal with this kind of problems. However, 1/p2 does converge, so I can
instead require axn + by n has few prime divisors p satisfying p2 |axn + by n . Equivalently,
if there exists too many such primes for any choice of x, y, then a, b has to have some
common factors. The general statement thus looks like the following: if a1 , . . . , ak are
positive integers such that for any x1 , . . . , xk , there exist a lot of primes p such that
p2 |a1 xn1 + · · · + ak xnk ,
then x1 , . . . , xk have gcd larger than 1. This is the originally proposed version of the
problem. To make it easier, the actual version appearing in the test becomes the following.
Problem 2.4 (2021 TST Round 3 Independent II-N). Let n be a given positive integer.
We say that a positive integer m is n-good if and only if there are at most 2n distinct
primes p satisfying p2 | m.
(a) Show that if two positive integers a, b are coprime, then there exist positive integers
x, y so that axn + by n is n-good.
(b) Show that for any k positive integers a1 , . . . , ak satisfying gcd(a1 , . . . , ak ) = 1,
there exist positive integers x1 , . . . , xk so that a1 xn1 + a2 xn2 + · · · + ak xnk is n-good.

Hung-Hsun Hans Yu 9
2.5 2022 Taiwan TST Round 1 Independent Study I-A
This functional equation was composed a long time ago using some idea I had that I
thought would be really really innovative if it actually worked. Unfortunately, it was
really hard to make it work, and the only sensible problem I produced was this one. The
readers are encouraged to try playing around with this trick and see what interesting
problems they ended up with. I would be interested in learning about them.
The trick I had in mind originated from test-solving some random functional equation
that Pang-Cheng Wu wrote. We did not end up solving that functional equation, but the
functional equation led me to think about the set

S = {(x − y, f (x) − f (y)) | x, y ∈ Z}.

There, one can show if (a, b) ∈ S, then (b, a − b) ∈ S or something like that. I thus
started to try if I can use this idea to make up any interesting functional equation. My
original plan was to somehow compose the problem so that the conditions allow us to do
“Euclidean algorithm”. But there are two issues with this: the first is that this usually
requires two kinds of operations: (a, b) ∈ S implies (b, a) ∈ S and (a, b) ∈ S implies
(a, b − a) ∈ S. I do not see how one can compress those two operations into one equation.
Another more fatal issue is that the set S is destined to contain (0, d) for some nonzero
d, which would prove the non-existence of the functions we are to solve for.
This issue might be annoying, but unfortunately the only interesting thing we can say
about S is that if (0, d) ∈ S then d = 0. What this entails is that if we ever succeed in
solving the functional equation using only the set S, then we need to say something like:
if S is not of a particular form, then starting from some pair (a, b) in S, after finitely many
operations we end up at (0, d) for d ̸= 0, which is then a contradiction. The only one
operation I found that seemed reasonable to me was: if (a, b) ∈ S then (⌊ 2b ⌋, −⌈ 2b ⌉) ∈ S.
With this, one can see that if b is ever allowed to be nonzero, then we can eventually
produce (0, d) ∈ S for some d ̸= 0. This would then prove that f has to be the constant
function. Now it just remains to write down a condition that shows that if (a, b) ∈ S
then (⌊ 2b ⌋, −⌈ 2b ⌉) ∈ S. This is how I ended up with the proposed functional equation.
Problem 2.5 (2022 TST Round 1 Independent I-A). Find all f : Z → Z such that
(õ û) õ û
f (x) + f (y) f (x) + f (y)
f + f (x) = f (f (y)) +
2 2

holds for all x, y ∈ Z.

Me four years ago dreamt about the imaginary world where the consideration of the
set S becomes trendy in Olympiad math, but me now just thinks it’s impossible. I think
it is still worth trying for those passionate about writing problems though, and do let me
know if you end up with anything interesting.

2.6 2022 Taiwan TST Round 3 Independent Study I-N

One day, when I was thinking about some research problem, I had to manipulate large
expressions consisting of gcd’s and multiplications. After struggling a little bit, I realized
that the property
gcd(a, b) gcd(c, d) = gcd(ac, ad, bc, bd)

Hung-Hsun Hans Yu 10
would help a lot in my computation. This motivates the following question: if we restrict,
say, the number of indeterminates and the degree of each term, how much can we always
simplify such expressions by? It turns out that the identity above together with some p-
adic analysis can turn the entire problem into one about lattice points in convex positions.
This gives the proposed problem.
Problem 2.6 (2022 TST Round 3 Independent I-N). Denote the set of all positive
integers by N, and the set of all ordered positive integers by N2 . For all non-negative
integers k, define good functions of order k recursively for all non-negative integers k,
among all functions from N2 to N as follows:
(i) The functions f (a, b) = a and f (a, b) = b are both good functions of order 0.
(ii) If f (a, b) and g(a, b) are good functions of orders p and q, respectively, then
gcd(f (a, b), g(a, b)) is a good function of order p+q, while f (a, b)g(a, b) is a good function
of order p + q + 1. (n)
Prove that, if f (a, b) is a good function of (order
) k ≤ 3
for some positive integer
n ≥ 3, then there exist a positive integer t ≤ n2 and t pairs of non-negative integers
(x1 , y1 ), . . . , (xn , yn ) such that

f (a, b) = gcd(ax1 by1 , . . . , axt byt )

holds for all positive integers a and b.

To be honest, I am not particularly proud of this problem. At least it looks... unique

enough.

2.7 2022 Taiwan TST Round 3 Mock Exam P2

The idea came from a conversation with Yufei Zhao in one meeting. He told me about a
cool proof of a well-known result saying that the number of lines is at least the number of
points unless they are all collinear. In fact, the proof is independent of the metric on R2
and only uses the fact that two points determine a line and two lines determine a point.
To see this, if n points on the plane determine more than 1 line but less than n lines,
let v1 , . . . , vn ∈ Rk be n vectors such that vij records whether the i-th point lies on j-th
′
point. Then ∑ vi · vi′ = 1 for any i ̸= i , and since n > k we can find a1 , . . . , an not all zero
such that ai vi = 0. This shows that
(∑ ) (∑ ) (∑ ) (∑ )
0= ai vi · ai vi > vi · vi ≥ 0,

which is a contradiction.
In fact, this proof is exactly the proof of the proposed problem when s = 1, As = [n].
Generalizing this proof gives the this problem as stated in the exam.
Problem 2.7 (2022 TST Round 3 Mock P2). Let n, s, t be three positive integers,
and let A1 , . . . , As , B1 , . . . , Bt be non-necessarily distinct subsets of {1, 2, . . . , n}. For any
subset S of {1, . . . , n}, define f (S) to be the number of i ∈ {1, . . . , s} with S ⊆ Ai and
g(S) to be the number of j ∈ {1, . . . , t} with S ⊆ Bj . Assume that for any 1 ≤ x < y ≤ n,
we have f ({x, y}) = g({x, y}). Show that if t < n, then there exists some 1 ≤ x ≤ n so
that f ({x}) ≥ g({x}).

Hung-Hsun Hans Yu 11
2.8 2023 TMO P3
To cover the details of the motivation behind this problem, I think it makes more sense
to introduce formal sum that is introduced by Sean Lee (at the time of writing this,
it has already been somewhat popular in Taiwan). The formal sum is essentially the
free abelian group on the lines direct sum R/πZ with the extra relations ℓ2 = ℓ1 + θ if
∡(ℓ1 , ℓ2 ) = π. Note that R/πZ survives (i.e. there is no relation θ = 0 generated for any
θ ̸= 0) as ∡(ℓ1 , ℓ2 ) + ∡(ℓ2 , ℓ3 ) = ∡(ℓ1 , ℓ3 ), and so this would behave more or less as we
expect.
When we are given a circle, we can even consider a slightly larger abelian group, by
taking the direct sum of the original group with the free abelian group on the points lying
on the circle with the extra relations A + B − AB = 0 for every A, B on the circle. The
original group would be a proper subgroup of the new group as AB +CD −AC −BD = 0
and so the new relations would not kill any elements in the original group. There are
more things to check but this is not the focus of this note so I will leave it for the readers
to check.
The advantage of these notations is that angle chasing becomes really easy (and
conceptual too!). For example, if ABC is a triangle, ℓ is a line and D is the isogonal
conjugate of ∞ℓ , then we know that A, B, C, D are on the same circle. The condition
that D is the isogonal conjugate of ∞ℓ is precisely AD + ∞ℓ = AB + AC (note that
there is no more asymmetry here, which is a benefit of using formal sums). Equivalently,
we have A + D + ∞ℓ = 2A + B + C, and so D + ∞ℓ = A + B + C.
This identity is how the problem is motivated. If we want A, B, C to be symmetric
while making the identity above holds, the most obvious choice is to make ABC an
equilateral triangle. In this case, the center O of ABC becomes a really special point,
and we still have a degree of freedom for where ℓ is placed, so let just make ℓ passes
through O. Then as ℓ passes through the orthocenter and also the incenter of ABC,
we see that ℓ∗ is a rectangular hyperbola that is tangent to ℓ at O. We also know that
ℓ∗ passes through D. Lastly, we know that the midpoint of OD is the center of the
rectangular hyperbola (one can show that this holds if and only if O is the midpoint of
the chord cut out by ℓ, which is true). So we could ask the converse: given ℓ through O,
a circle Γ centered at O and a point D on Γ, if h is a rectangular hyperbola with OD
being one of its diameter, do the other three intersections h∩Γ always form an equilateral
triangle?
The answer is indeed yes. To see this, formal sums come in handy again. As OD
is its diamter, we know that OP + DP is constant when P is moving on h (another
place where formal sum is convenient!). By taking P = O, we see that this constant is
ℓ + OD = ℓ + 2D + π/2. Now if P is the intersection of h and Γ that is not D, then
OP + DP = D + 3P + π/2 (why does it not necessarily hold when P = D?), and so we
get 3P = D + ℓ, showing that the other three intersections indeed form an equilateral
triangle.
Of course, this cannot be the statement of the problem. In general, to hide any
conic, we can use Pascal’s theorem. Let ∞1 and ∞2 be the two intersections of h with
the infinity line. Then we may apply Pascal’s theorem on OOD∞1 P ∞2 to get that
X = ℓ ∩ P ∞1 , Y = OD ∩ P ∞2 and Z = D∞1 ∩ O∞2 . Also note that ∞1 , ∞2 are
exactly the directions that the angular bisectors of OD and ℓ point in, and so we exactly
know what they are. This also shows that Z is determined. We could thus construct P
backward: let X be a point on ℓ, Y be a point on OD such that XY Z are collinear, we

Hung-Hsun Hans Yu 12
can let P be the intersection of Y ∞2 and X∞1 and assume that P is on the circle Γ.
We also need to hide the formal sum identity 3P = D + ℓ. The best way to hide this
is to construct two points differently and say that they are actually the same, but then
this identity has the “wrong order”: both sides have “order three” (i.e. 3P = 3N + θ
for some reference point N and some θ ∈ R/πZ). The most natural way to move things
around so that we have the right order is that 2P − D = ℓ − P . Now if Q = 2P − D,
then D + Q = 2P and so DQ is the line through D parallel to the tangent of P at Γ, or
the line perpendicular to the line OP . If Q′ = ℓ − P , then P + Q′ = ℓ and so P Q′ is the
line through P parallel to ℓ. Thus 3P = D + ℓ is equivalent to that the line through D
perpenciular to OP intersects the line through P parallel to ℓ at some point on Γ. After
relabelling, we get the proposed problem.
Problem 2.8 (2023 TMO P3). Let O be the center of a circle Γ, and let A, B be two
distinct points such that O, A, B are not collinear. Let M be the midpoint of AB, and
let P, Q be points on OA, OB such that P M Q are collinear. Let X be the intersection of
the line through P parallel to AB and the line through Q parallel to OM . Let Y be the
intersection of the line through X parallel to OA and the line through B perpendicular
to OX.
Show that if X is on Γ, then Y is on Γ.

The last thing to do is to find a proof that does not use any hyperbola nonsense. It
turns out that an easy (but interesting) trig-bash kills it. Saen Lee also found a solution
that is completely geometric without using hyperbola at all. This is thus a fair problem
to put in high school contest!

3 Modify the Prototype

Quite often, the very first version of the problem is not that suitable for proposal. If
one starts with the first method, it is quite often for them to generate a problem that
is either too easy or too hard. If one starts with the second method, it is sometimes
hard to look for a good way to hide the trick. I think the key in those scenarios is to
realize that we are composing problems, not solving problems. Unlike contestants, we
are allowed to modify the conditions at our will. Thus, if the question is too easy, ask
some deeper/harder question or get rid of some constraints; if the question is too hard,
add some more constraint or consider some other quantities or objects that are more in
control; if the trick or property considered is too exposed, maybe switch up a bit in how
it is hidden or simply choose another related trick to hide instead. This entire process is
more like an art work, and you should never feel constrained or limit the possibilities of
the problem!

3.1 2021 Taiwan TST Round 1 Mock Exam P4

The trick I started with for this problem is the FKG inequality. The version I used is the
following.
Theorem 3.1 (Fortuin–Kasteleyn–Ginibre). Let X1 , X2 , . . . , Xk be some discrete inde-
pendent random variables taking values in . Let f, g be two functions on X1 , . . . , Xk that

Hung-Hsun Hans Yu 13
are simultaneously increasing or decreasing in the variables. Then

E[f g] ≥ E[f ]E[g].

The proof of this inequality can be obtained by induction on k and do Chebyshev’s sum
inequality on each inductive step. This inequality is an important correlation inequality
as it basically says that “the product of two positively correlated random variables has
a higher expected value than the product of the individual expected values”, which is
intuitively correct but hard to verify in many cases. It thus finds lots of applications in
statistical mechanics and probabilistic combinatorics.
For simplicity, I took k = 2. Let X1 take values in {a1 , . . . , an } ⊆≥0 and X2 take values
in {b1 , . . . , bn } ⊆≥0 . If f, g are decreasing in X1 , X2 , then the equallity in FKG holds when
X1 = X2 with probability 1. Therefore, we have a way to compose an inequality with
multiple equality cases here: we can let P(X1 = ai ) proportional to ai and P(X2 = bi )
proportional to bi . Then the equality would hold when all nonzero ai ’s and bi ’s are equal
to each other.
Now I oopsed a bit and thought that f = X1 /(X1 X2 + 1) and g = X2 /(X1 X2 + 1)
were both increasing in X1 , X2 . Assuming that this were the case, then by the FKG
inequaltiy,
( )( )( ) ( )( )
∑ ∑ ∑ a2i b2j ∑ ai b j ∑ ai b j
ai bj 2
≥ ai · bj · .
i j i,j
(a i b j + 1) i,j
a i b j + 1 i,j
a i b j + 1
∑ ∑
This is kind of ugly, so let us set S = i ai = j bj . Note that ai bj /(ai bj + 1)2 ≤ 1/4,
so the left hand side is upper bounded by S 4 /4. We still have some of the equality cases
remaining when S is some positive integer at most n, and we get that
( )( )
∑ ai b j ∑ ai b j 1
ai · bj · ≤ S 4.
i,j
ai b j + 1 i,j
ai b j + 1 4

Calvin Hsu verified that despite we arrived at this point with a totally wrong proof, this
inequality is still true when S = n/2 with the help of computer. Obviously though, this
cannot be an actual problem if we∑ do not have an accessible proof. What I noticed is that,
if mA is the
∑ minimum value of j ai bj /(ai bj + 1) for nonzero ai , and mB is the minimum
value of i ai bj /(ai bj + 1) for nonzero bj , then the left hand side can be bounded by
S 2 mA mB . Assuming that S ̸= 0, then we get that
√ 1
min(mA , mB ) ≤ mA mB ≤ S.
2
Setting S = n/2 and doubling n, we get the proposed problem. After checking that there
are actually plenty of easy ways to solve the problem, I decided to propose this version
of the problem.
Problem 3.1 (2021 TST Round 1 Mock P4). Let n be a positive integer. ∑2n For∑each 4n-
tuple of nonnegative real numbers a1 , . . . , a2n , b1 , . . . , b2n that satisfy i=1 ai = 2n j=1 bj =
n, define the sets
{ 2n }
∑ ai b j ∑2n
ai b j
A := : i ∈ {1, . . . , 2n} s.t. ̸= 0 ,
j=1
a i b j + 1 j=1
a i b j + 1

Hung-Hsun Hans Yu 14
 { 2n }
∑ ai b j ∑
2n
ai b j
B := : j ∈ {1, . . . , 2n} s.t. ̸= 0 .
i=1
ai b j + 1 i=1
ai b j + 1
Let m be the minimum element of A ∪ B. Determine the maximum value of m among
those derived from all such 4n-tuples a1 , . . . , a2n , b1 , . . . , b2n .

3.2 2021 Taiwan TST Round 2 Mock Exam P5

I started with a simple idea: I wanted to compose a FE of the form f : N → N and
f (g(x)) (x) = x for all x ∈ N where g : N → Z≥0 is some function that depends on f . Well,
we can see that f (x) = x is obviously a solution. If it is possible for g(x) = 0 holds for
all x ∈ N, then this is also a solution. I therefore set g(x) = |f (x) − x| and wanted to see
if this forces f (x) ≡ x.
I soon realized that this problem actually concerns the existence of a sequence a1 , . . . , ak
that satisfies ak+1 = a1 and k|ai+1 − ai for all i = 1, . . . , k. With this, one can easily
construct solutions other than f (x) ≡ x that still satisfies f (|f (x)−x|) (x) = x. The form of
the solution makes this version not suitable for an Olympiad test.
To fix it, I observe that (a2 − a1 ) + · · · + (ak − ak−1 ) + (a1 − ak ) = 0. Therefore, if I can
somehow construct a function h such that h(x) = ±x or ±x−1, and set g(x) = h(f (x)−x),
then we must have that either h(ai − ai−1 ) = ±x for all i, or h(ai − ai−1 ) = ±x − 1 for
all i. Therefore, if we choose h such that this mostly does not happen, then we will have
a lot of control over what the sequence can be. The choice h(x) = (|x| + |x − 1| − 1)/2
satisfies this, which gives rise to this problem.
Problem 3.2 (2021 TST Round 2 Mock P5). Let ∥x∥∗ = (|x| + |x − 1| − 1)/2. Find
all f : N → N such that
f (∥f (x)−x∥∗ ) (x) = x, ∀x ∈ N.
Here f (0) (x) = x and f (n) (x) = f (f (n−1) (x)) for all n ∈ N.

3.3 2021 Taiwan TST Round 2 Mock Exam P6

This problem was composed with Cheng-Ying Chang. The idea we started with is, given
a graph, what can we know about it if we can test whether two chosen vertices have
distance smaller than a specific value k. Well, if k is larger than the number of vertices
n, then what we can know from a query is if the chosen two vertices are in the same
component. Therefore, the natural question to ask is: how many queries do we need to
know if the graph is connected or not?
When k is 1, we need Θ(n2 ), and when k is n, we can do it in (n−1) queries. Therefore,
it is natural to guess that we can always do it in O(n2 /k) queries. However, it is quite
hard to come up with an efficient algorithm when k is in between. One of the difficulty is
that when k is small, even if the graph is connected, we might need a lot of queries until
we finally find two vertices that are within distance k from each other. This motivates us
to ask a natural question: if S is a set of vertices where any two vertices are not within
distance k from each other, how big can S be? With a simple argument using Euler tour
on the spanning tree, we can see that the size of S is bounded by O(n/k). This in turn
shows that if G is connected, then we can find a pair of vertices with distance at most k
using O(n2 /k 2 ) queries.
This, however, does not solve the problem. The issue is that in order to know that the
graph is connected, we need to “expand our knowledge to the entire graph”. Knowing that

Hung-Hsun Hans Yu 15
two vertices are close enough is not quite sufficient—we might end up choosing vertices
that are either very close to each other or very far from each other, and thus making no
progress. To fix this, I decided to change the problem so that we will also know it if the
two vertices have distance exactly k. This gives us some way to guarantee that we are
always making some progress in expanding our knowledge about the graph. It is then
not hard to use the observation in the previous paragraph to show that 2n2 /k queries
suffice.
Problem 3.3 (2021 Round 2 Mock P6). Let k ≤ n be two positive integers. IMO-
nation has n villages, some of which are connected by a road. For any two villages, the
distance between them is the minimum number of toads that one needs to travel from
one of the villages to the other, if the traveling is impossible, then the distance is set as
infinite.
Alice, who just arrived IMO-nation, is doing her quarantine in some place, so she does
not know the configuration of roads, but she knows n and k. She wants to know whether
the furthest two villages have finite distance. To do so, for every phone call she dials
to the IMO office, she can choose two villages, and ask the office whether the distance
between them is larger than, equal to, or smaller than k. The office answers faithfully
(infinite distance is larger than k). Prove that Alice can know whether the furthest two
villages have finite distance between them in at most 2n2 /k calls.

At this point, I still don’t know if O(n2 /k) is the best possible answer. I also don’t
know if we can do as good without knowing if two vertices have distance exactly k. Inter-
ested readers are welcome to think about those two problems, and I would be interested
in knowing the answer.

3.4 2022 TMO P1

This easy problem took me surprisingly long to compute. To explain the motivation of
this problem is actually going to be a long journey.
The initial driving force I had was that I wanted to be cool and composed a problem
in the intersection of geometry and number theory. The simple question I came up with
was the following: what kind of lattice triangles have some of their centers also as lattice
points?
The problem is pretty boring if we are looking for lattice triangles with their centroids
as lattice points. We could also easily construct lattice triangles with orthocenters as
lattice points by simply drawing a circle centered at a lattice point, passing through
another lattic point. Because of Euler line, I would also imagine that anything asked
regarding orthocenters would either be trivial or extremely hard. Therefore the next up
in the list to try for me was incenters.
One powerful tool regarding lattice points is the Pick’s theorem, and incenters interact
well with areas. This shows us that incenters might be the way to go. By Heron’s formula,
we see that one necessary condition
√ for this to hold is that if a, b, c are side lengths
a
of the triangle, then 2(a+b+c) (a + b + c)(a + b − c)(a − b + c)(−a + b + c) and similar
quantities for b, c all have to be integers. From here, we can already see that a, b, c have to
live in the same quadratic number field. To simplify the discussion a bit, I decided to take
a, b, c as positive integers with gcd(a, b, c) = 1. I also take my, and probably everyone’s,
favorite substitution x = (−a + b + c)/2, y = (a − b + c)/2, z = (a + b −√c)/2 and assume
that they are positive integers. Then what we have is that 2(x+y +z) | xyz(x + y + z).

Hung-Hsun Hans Yu 16
This implies x + y + z | xyz, and you can finally start to see why this is relevant to 2022
TMO P1 at all.
I was a bit stuck, so I decided to look for examples of those triangles. Clearly those
triangles have to be Heronian triangles as discussed above, so I tested some of them and
found that, actually for any Pythagorean triples (a, b, c), the triangle (0, 0), (a, 0), (0, b)
not only have their incenters at lattice points, but also somehow have all three excenters
at lattice points!
Now if we run through the argument above for excenters too, we get x | yz(x+y+z), y |
xz(x + y + z) and z | xy(x + y + z)! In fact besides the four divisbility conditions, the
only other one left is that xyz(x + y + z) is an even perfect square, but as we all know
now, this conditions is automatically satisfied under some extra assumptions.
My focus soon turns to deal with the four conditions x+y+z | xyz, x | yz(x+y+z), y |
xz(x + y + z) and z | xy(x + y + z). Note however that for any x, y, z, we can always
scale them up so that the four conditions hold. This makes it boring, so a natural
condition to add is gcd(x, y, z) = 1. A standard substitution x = atu, y = bsu, z = cst,
where a, b, c are mutually coprime, gives us that x + y + z = abc, and we notice that
xyz(x + y + z) = (abcstu)2 is a perfect square!
Problem 3.4 (2022 TMO P1). Let x, y, z be three positive integers with gcd(x, y, z) =
1. If
x | yz(x + y + z),
y | xz(x + y + z),
z | xy(x + y + z),
and
x + y + z | xyz,
show that xyz(x + y + z) is a perfect square.

At this point we have already composed 2022 TMO P1, but I did not stop here. I
really wanted to add some constraint so that only the right triangles survive. If we look
back at the Pythagorean triple example and compute the corresponding x, y, z, we will
see that gcd(y, z) = 1 and x | yz. This corresponds to having a = s = 1, and now we have
x(x+y+z) = bctu = yz. This last equation is equivalent to (x+y)2 +(x+z)2 −(y+z)2 = 0,
showing that the triangle must have a right angle! This corresponds to the two alternative
versions that I proposed, and they can be found on AoPS.

3.5 2022 USAMO P3

The original trick I had in mind is something that appears in Galois theory (here I am
being really sloppy): if F (·) is the fixed field of some Galois group, and G(·) is the
Galois element that fixes some underlying field, then we have F (G(F (H))) = F (H) for
any Galois group H. This is obtained by H ⊆ G(F (H)), K ⊆ F (G(K)) and the fact
that F is inclusion-reversing, i.e. if H ′ ⊆ H then F (H) ⊆ F (H ′ ). This is because
F (G(F (H))) ⊆ F (H), but when we plug in K = F (H) to K ⊆ F (G(K)), we get
F (H) ⊆ F (G(F (H))) and so F (H) = F (G(F (H))).
This was a rather new idea to me five years ago when I first learned about Galois
theory. As I was such an FE enthusiast back then, I then started to think if I do something

Hung-Hsun Hans Yu 17
similar for FE. This came to my mind pretty quickly: if f is strictly decreasing with
f (f (x)) ≥ x, then we have
f (x) ≤ f (3) (x) ≤ f (x)
and so f (x) = f (3) (x) (the first inequality comes from plugging in f (x) to the inequality,
and the second inequality comes from the first inequality together with the fact that f
is strictly decreasing). In order to hide the inequality in the problem, it is natural to
consider f : R+ → R+ . Since the function f we are going for is strictly decreasing and
also an involution (i.e. f (f (x)) = x), it is natural to look for FE with f (x) = 1/x being
a solution. With the common trick, the inequality f (f (x)) ≥ x can be hidden by setting
f (x) ̸= f (f (2) (x) + y) for any x, y > 0. Now note that if f actually satisfies f (2) (x) = x
and it is strictly decreasing, then we have to have f (x) > f (f (2) (x) + y). The very very
last thing we need to do is to encode the assumption that f is strictly decreasing to the
problem statement. Note that this is usually too strong of an assumption to just throw
in. Therefore I decided to introduce a new function g with the hope that maybe it would
make the problem not dumb. This is what I ended up with:

First version. Find all f, g : R+ → R+ with g strictly increasing and

( )
xf (x) = (x + g(y)) · f f (2) (x) + y .

As a by-product, f (x) = k/x also becomes a solution. Quite unfortunately, it turns

out that there is a bashing solution that doesn’t even use the condition that g is strictly
increasing (credit to Wei-Jiun Kao). Basically, one plugs in (x, f (2) (y)) and exchange
x, y to get something useful that ends up solving the entire FE. This reemphasizes the
importance of test-solving if you write a problem with some trick in mind—you really
want to make sure that there is not some unexpectedly sad/dumb solution that you
overlook.
Though it was a bit sad, I decided to see if I can solve the FE without g being
strictly increasing, in a way that I prefer more. After a long struggle, I finally discovered
the actual main trick of the problem: if f (x) > f (f (2) (x) + y) for any x, y > 0, then
f (4) (x) = f (2) (x)! At this point, the “Galois trick” is already all gone—or is it? To me, it
seems that the proof of this claim still pertains some characteristic of the original trick.
Anyways, having discovered this, I started to play around with this trick and see what
FE I can write out. After a really long trial-and-error, I found two versions that I think
are more reasonable.
( )
Second version. xf (x) = xf f (2) (x) + y + yf (x + y).
( )
Third version. f (x) = f f (2) (x) + y + yf (x)f (x + y).
Then... it’s time to test-solve! It turns out that there is another solution for the second
version (due to Ting-Wei William Chao) that proceeds by analyzing limx→0 xf (x). On
the other hand, there does not seem to be any reasonably short alternative solution for
the third version, so I was already pretty satisfied with it. One unfortunate charactersitic
of it is that the y variable is pretty exposed, and one can get injectivity basically for free
because of that. I tried really hard to wrap everything up with f ’s, but this more or
less makes proving injectivity really hard, and, as a consequence, making it hard going
from f (4) (x) = f (2) (x) to f (2) (x) = x. After another long trial-and-error (this time with

Hung-Hsun Hans Yu 18
Pang-Cheng Wu), I finally ended with the proposed version (and the proof of injectivity
is due to him)!
Problem 3.5 (2022 USAMO P3). Let R>0 be the set of all positive real numbers. Find
all functions f : R>0 → R>0 such that for all x, y ∈ R>0 we have

f (x) = f (f (f (x)) + y) + f (xf (y))f (x + y).

I had this idea first in early 2017, and the final version came up in (I think) mid 2018,
so it was a long time to refine and crystallize all the ideas. I am glad that after a long
wait, I can finally share the FE I will be most proud of with other people (the second
being 2016 Taiwan TST Round 2 Quiz 1 P2).

3.6 2022 Taiwan TST Round 1 Independent Study I-A

The inequality in the statement is actually something that appears in a note I was reading
for research. It was from Lecture Notes 6 in https://www.math.ucla.edu/~tao/254b.
1.99s/, where one shows that the weaker formulation of Kakeya’s conjecture is actually
equivalent to the stronger formulation. There, we have a sequence of non-negative real
numbers c0 , c1 , . . . that are eventually zero with

ck ≤ ck−1 + Cα−k c2k

for all k ≥ 1, where C > 0 and α > 1 are some constants. We need to show that there is
some constant M > 0 and k0 ∈ N depending only on C and α such that ck < M maxi≤k0 ci
for all k. What was done in the note is that we take dk = (1 + C ′ α−k )ck for some C ′ > 0
chosen sufficiently large so that

dk ≤ dk−1 + Cα−k [A(d2k − dk ) + (dk−1 − dk )]

for some A ≥ 0 and sufficiently large k. In fact, this holds as long as

(1 + C ′ α−k+1 )(1 + Cα−k ) ≥ (1 + C ′ α−k )(1 + (A + 1)Cα−k )

and
(1 + C ′ α−2k )A ≥ (1 + C ′ α−k )(1 + (A + 1)Cα−k ).
If A > 1 and (α − 1)C ′ > AC, then both hold for sufficiently large k. Therefore for some
k0 , we have
dk ≤ dk−1 + Cα−k [A(d2k − dk ) + (dk−1 − dk )]
holds for all k ≥ k0 . Since dk is eventually zero, we can choose kmax that maximizes dk
when k ≥ k0 . If kmax > k0 , then the inequality above leads to an immediate contradiction.
Therefore kmax = k0 and so dk ≤ dk0 for all k ≥ k0 , showing that ck ≤ (1 + C ′ α−k0 )ck0 for
all k ≥ k0 .
I thought that this substitution is hard to motivate, so I tried to simplify the setting
and set a question based on it. If we take α = 2, C = 1, and ask ck to be eventually 1
instead of 0, then we now actually get a lower√bound on ck0 as long as k0√is sufficiently
large. For example, if we take A = 9(7 − 21)/14 and C ′ = 20 − 4 21, then the
inequalities hold for k ≥ 3. This shows that
1
c3 ≥ √ ≈ 0.827.
1 + (20 − 4 21) · 2−3

Hung-Hsun Hans Yu 19
Then, I gave it to Cheng-Ying, who then told me that this bound is weaker than what
he could get by a simple argument considering the largest element. His method also does
not depend on the assumption that each term is non-negative. This is when I realized
that Tao probably took on an unnecessarily harder approach (which explains why the
substitution seems so unmotivated), and probably Cheng-Ying’s method is the way it
should have been done. Anyways, the new method makes both the problem statement
and the solution look way cleaner, and this becomes the proposed problem.
Problem 3.6 (2022 TST Round 1 Independent I-A). Let a1 , a2 , a3 , . . . be a sequence
of reals such that there exists N ∈ N so that an = 1 for all n ≥ N , and for all n ≥ 2 we
have
an ≤ an−1 + 2−n a2n .
Show that ak > 1 − 2−k for all k ∈ N.

Hung-Hsun Hans Yu 20