USAMO 2024 Solution Notes

Evan Chen《陳誼廷》
15 December 2024

This is a compilation of solutions for the 2024 USAMO. 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 USAMO 2024/1, proposed by Luke Robitaille . . . . . . . . . . . . . . . . 3
1.2 USAMO 2024/2, proposed by Rishabh Das . . . . . . . . . . . . . . . . . 4
1.3 USAMO 2024/3, proposed by Krit Boonsiriseth . . . . . . . . . . . . . . . 7

2 Solutions to Day 2 10
2.1 USAMO 2024/4, proposed by Rishabh Das . . . . . . . . . . . . . . . . . 10
2.2 USAMO 2024/5, proposed by Anton Trygub . . . . . . . . . . . . . . . . 12
2.3 USAMO 2024/6, proposed by Titu Andreescu and Gabriel Dospinescu . . 17

1
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

§0 Problems
1. Find all integers n ≥ 3 such that the following property holds: if we list the divisors
of n! in increasing order as 1 = d1 < d2 < · · · < dk = n!, then we have

d2 − d1 ≤ d3 − d2 ≤ · · · ≤ dk − dk−1 .

2. Let S1 , S2 , . . . , S100 be finite sets of integers whose intersection is not empty. For
each non-empty T ⊆ {S1 , S2 , . . . , S100 }, the size of the intersection of the sets in T
is a multiple of |T |. What is the smallest possible number of elements which are in
at least 50 sets?

3. Let (m, n) be positive integers with n ≥ 3 and draw a regular n-gon. We wish to
triangulate this n-gon into n − 2 triangles, each colored one of m colors, so that
each color has an equal sum of areas. For which (m, n) is such a triangulation and
coloring possible?

4. Let m and n be positive integers. A circular necklace contains mn beads, each

either red or blue. It turned out that no matter how the necklace was cut into
m blocks of n consecutive beads, each block had a distinct number of red beads.
Determine, with proof, all possible values of the ordered pair (m, n).

5. Point D is selected inside acute triangle ABC so that ∠DAC = ∠ACB and
∠BDC = 90◦ + ∠BAC. Point E is chosen on ray BD so that AE = EC. Let
M be the midpoint of BC. Show that line AB is tangent to the circumcircle of
triangle BEM .

6. Let n > 2 be an integer and let ` ∈ {1, 2, . . . , n}. A collection A1 , . . . , Ak of

(not necessarily distinct) subsets of {1, 2, . . . , n} is called `-large if |Ai | ≥ ` for all
1 ≤ i ≤ k. Find, in terms of n and `, the largest real number c such that the
inequality
k X k k
!2
X |Ai ∩ Aj |2 X
xi xj ≥c xi
|Ai | · |Aj |
i=1 j=1 i=1

holds for all positive integer k, all nonnegative real numbers x1 , x2 , . . . , xk , and all
`-large collections A1 , A2 , . . . , Ak of subsets of {1, 2, . . . , n}.

2
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

§1 Solutions to Day 1
§1.1 USAMO 2024/1, proposed by Luke Robitaille
Available online at https://aops.com/community/p30216459.

Problem statement

Find all integers n ≥ 3 such that the following property holds: if we list the divisors
of n! in increasing order as 1 = d1 < d2 < · · · < dk = n!, then we have

d2 − d1 ≤ d3 − d2 ≤ · · · ≤ dk − dk−1 .

The answer is n ∈ {3, 4}. These can be checked by listing all the divisors:

• For n = 3 we have (1, 2, 3, 6).

• For n = 4 we have (1, 2, 3, 4, 6, 8, 12, 24).

We prove these are the only ones.

The numbers 5 ≤ n ≤ 12 all fail because:

• For n = 5 we have 20 − 15 > 24 − 20.

• For n = 6 we have 18 − 15 > 20 − 18.

• For 7 ≤ n ≤ 12 we have because 14 − 12 > 15 − 14 (and 13 - n!).

Now assume n ≥ 13. In that case, we have

j n k2
− 1 ≥ 2n.
2
So by Bertrand postulate, we can find a prime p such that
j n k2
n<p< − 1.
2
However, note that
j n k2 j n k  j n k  j n k2
−1= −1 +1 ,
2 2 2 2
are consecutive integers both dividing n! (the latter number divides ). So
n  n 

2 · 2 2
the prime p causes the desired contradiction.

3
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

§1.2 USAMO 2024/2, proposed by Rishabh Das

Available online at https://aops.com/community/p30216494.

Problem statement

Let S1 , S2 , . . . , S100 be finite sets of integers whose intersection is not empty. For
each non-empty T ⊆ {S1 , S2 , . . . , S100 }, the size of the intersection of the sets in T
is a multiple of |T |. What is the smallest possible number of elements which are in
at least 50 sets?

The answer is 50 100

.


50

¶ Rephrasing (cosmetic translation only, nothing happens yet). We encode with

binary strings v ∈ F100 2 of length 100. Write v ⊆ w if w has 1’s in every component v
does, and let |v| denote the number of 1’s in v.
Then for each v, we let f (v) denote the number of elements x ∈ Si such that
S
x ∈ Si ⇐⇒ vi = 1. For example,

• f (1 . . . 1) denotes | 100
1 Si |, so we know f (1 . . . 1) ≡ 0 (mod 100).
T

• f (1 . . . 10) denotes the number of elements in S1 through S99 but not S100 so we
know that f (1 . . . 1) + f (1 . . . 10) ≡ 0 (mod 99).

• . . .And so on.

So the problem condition means that f (v) translates to the statement

|u| divides
X
P (u) : f (v)
v⊇u

for any u 6= 0 . . . 0, plus one extra condition f (1 . . . 1) > 0. And the objective function is
to minimize the quantity X
A := f (v).
|v|≥50

So the problem is transformed into an system of equations over Z≥0 (it’s clear any
assignment of values of f (v) can be translated to a sequence (S1 , . . . , S100 ) in the original
notation).
Note already that:

Claim — It suffices to assign f (v) for |v| ≥ 50.

Proof. If we have found a valid assignment of values to f (v) for |v| ≥ 50, then we can
always arbitrarily assign values of f (v) for |v| < 50 by downwards induction on |v| to
satisfy the divisibility condition (without changing M ).

Thus, for the rest of the solution, we altogether ignore f (v) for |v| < 50 and only consider
P (u) for |u| ≥ 50.

4
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

¶ Construction. Consider the construction

f0 (v) = 2|v| − 100.

This construction is valid since if |u| = 100 − k for k ≤ 50 then

       
X k k k k
f0 (v) = · 100 + · 98 + · 96 + · · · + · (100 − 2k)
0 1 2 k
v⊇u

= (100 − k) · 2k = |u| · 2k

is indeed a multiple of |u|, hence P (u) is true. In that case, the objective function is
100    
X 100 100
A= (2i − 100) = 50
i 50
i=50

as needed.
Remark. This construction is the “easy” half of the problem because it coincides with what
you get from a greedy algorithm by downwards induction on |u| (equivalently, induction on
k = 100 − |u| ≥ 0). To spell out the first three steps,
• We know f (1 . . . 1) is a nonzero multiple of 100, so it makes sense to guess f (1 . . . 1) =
100.
• Then we have f (1 . . . 10) + 100 ≡ 0 (mod 99), and the smallest multiple of 99 which
is at least 100 is 198. So it makes sense to guess f (1 . . . 10) = 98, and similarly guess
f (v) = 98 whenever |v| = 99.
• Now when we consider, say v = 1 . . . 100 with |v| = 98, we get

f (1 . . . 100) + f (1 . . . 101) + f (1 . . . 110) + f (1 . . . 111) ≡ 0 (mod 98).

| {z } | {z } | {z }
=98 =98 =100

we obtain f (1 . . . 100) ≡ 96 (mod 98). That makes f (1 . . . 100) = 96 a reasonable

guess.
Continuing in this way gives the construction above.

¶ Proof of bound. We are going to use a smoothing argument: if we have a general

working assignment f , we will mold it into f0 .
We define a push-down on v as the following operation:
• Pick any v such that |v| ≥ 50 and f (v) ≥ |v|.
• Decrease f (v) by |v|.
• For every w such that w ⊆ v and |w| = |v| − 1, increase f (w) by 1.

Claim — Apply a push-down preserves the main divisibility condition. Moreover,

it doesn’t change A unless |v| = 50, where it decreases A by 50 instead.

Proof. The statement P (u) is only affected when u ⊆ v: to be precise, one term on the
right-hand side of P (u) decreases by |v|, while |v| − |u| terms increase by 1, for a net
change of −|u|. So P (u) still holds.
To see A doesn’t change for |v| > 50, note |v| terms increase by 1 while one term
decreases by −|v|. When |v| = 50, only f (v) decreases by 50.

5
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

Now, given a valid assignment, we can modify it as follows:

• First apply pushdowns on 1 . . . 1 until f (1 . . . 1) = 100;

• Then we may apply pushdowns on each v with |v| = 99 until f (v) < 99;

• Then we may apply pushdowns on each v with |v| = 98 until f (v) < 98;

• . . .and so on, until we have f (v) < 50 for |v| = 50.

Hence we get f (1 . . . 1) = 100 and 0 ≤ f (v) < |v| for all 50 ≤ |v| ≤ 100. However, by
downwards induction on |v| = 99, 98, . . . , 50, we also have

f (v) ≡ f0 (v) (mod |v|) =⇒ f (v) = f0 (v)

since f0 (v) and f (v) are both strictly less than |v|. So in fact f = f0 , and we’re done.

Remark. The fact that push-downs actually don’t change A shows that the equality case
we described is far from unique: in fact, we could have made nearly arbitrary sub-optimal
decisions during the greedy algorithm and still ended up with an equality case. For a
concrete example, the construction

500
 |v| = 100
f (v) = 94 |v| = 99

100 − 2|v| 50 ≤ |v| ≤ 98


works fine as well (where we arbitrarily chose 500 at the start, then used the greedy algorithm
thereafter).

6
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

§1.3 USAMO 2024/3, proposed by Krit Boonsiriseth

Available online at https://aops.com/community/p30216513.

Problem statement

Let (m, n) be positive integers with n ≥ 3 and draw a regular n-gon. We wish to
triangulate this n-gon into n − 2 triangles, each colored one of m colors, so that
each color has an equal sum of areas. For which (m, n) is such a triangulation and
coloring possible?

The answer is if and only if m is a proper divisor of n.

Throughout this solution, we let ω = exp (2πi/n) and let the regular n-gon have
vertices 1, ω, . . . , ω n−1 . We cache the following frequent calculation:

Lemma
The triangle with vertices ω k , ω k+a , ω k+b has signed area

(ω a − 1)(ω b − 1)(ω −b − ω −a )
T (a, b) := .
4i

Proof. Rotate by ω −k to assume WLOG that k = 0. Apply complex shoelace to the

triangles with vertices 1, ω a , ω b to get
   
1 1 1 0 0 1
i i
det ω a ω −a 1 = det ω a − 1 ω −a − 1 1
4 4
ω b ω −b 1 ω b − 1 ω −b − 1 1

which equals the above.

¶ Construction. It suffices to actually just take all the diagonals from the vertex 1, and
then color the triangles with the m colors in cyclic order. For example, when n = 9 and
m = 3, a coloring with red, green, blue would be:

Green
Blue R

Red

Green R
Blue

7
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

To see this works one can just do the shoelace calculation: fix a residue r mod m
corresponding to one of the colors. Then

Area(ω 0 , ω j , ω j+1 ) =
X X
T (j, j + 1)
0≤j<n 0≤j<n
j≡r mod m j≡r mod m
X (ω j − 1)(ω j+1 − 1)(ω −(j+1) − ω −j )
=
4i
0≤j<n
j≡r mod m
ω−1 X
= (ω −j − 1)(ω j − ω −1 )
4i
0≤j<n
j≡r mod m
 
ω−1 n 1 + ω −1 +
X
(ω −1−j − ω j )


= .
4i  m
0≤j<n
j≡r mod m

(We allow degenerate triangles where j ∈ {0, m − 1} with area zero to simplify the
notation above.)
However, if m is a proper divisor of m, then j ω j = ω r (1+ω m +ω 2m +· · ·+ω n−m ) = 0.
P

For the same reason, j ω −1−j = 0. So the inner sum vanishes, and the total area of
P
this color equals

n (ω − 1)(ω −1 + 1) n ω − ω −1
Area(ω 0 , ω j , ω j+1 ) =
X
= · .
m 4i m 4i
0≤j<n
j≡r mod m

Because the right-hand side does not depend on the residue r, this shows all colors have
equal area.

¶ Proof of necessity. It’s obvious that m < n (in fact m ≤ n − 2). So we focus on just
showing m | n.
Repeating the same calculation as above, we find that if there was a valid triangulation
and coloring, the total area of each color would equal

n ω − ω −1
S := · .
m 4i
However:

Claim — The number 4i · S is not an algebraic integer when m - n.

Proof. This is easiest to see if one knows the advanced result that K := Q(ω) is a number
field whose ring of integers is known to be OK = Z[ω], in which case it follows right
away.

Remark. We spell out the details in the proof a bit more explicitly here. It’s enough to
show that ω · 4i · S = m
n 2
ω −m n
is not an algebraic integer for completeness.
Take K = Q(ω) of degree d := ϕ(n) ≥ 2; as a Q-module, it obeys K = Q⊕ωQ⊕· · ·⊕ω d−1 Q.
The theorem we are quoting is that, as Z-modules, we have OK = Z[ω] = Z⊕ωZ⊕· · ·⊕ω d−1 Z
i.e. OK contains exactly those numbers in K for which the canonical Q-coefficients happen

8
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

to be integers. (This is quite famous, but for a print reference, see Neukirch’s Algebraic
Number Theory, Section I.10.) And ω · 4i · S fails this criteria, since m
n
∈
/ Z.

However, for any a and b, the number

4i · T (a, b) = (ω a − 1)(ω b − 1)(ω −b − ω −a )

is an algebraic integer. Since a finite sum of algebraic integers is also an algebraic integer,
the sum of expressions of the form 4i · T (a, b) will never equal 4i · S.

Remark. If one wants to avoid citing the fact that OK = Z[ω], then one can instead note
that T (a, b) is actually always divisible by (ω − 1)(ω −1 + 1) = ω − ω −1 over the algebraic
integers (at least one of {ω a − 1, ω b − 1, ω −a − ω −b } is a multiple of ω + 1, by casework on
a, b mod 2). Then one using (ω−1)(ω 4i
−1 +1) as the scaling factor instead of 4i, one sees that

we actually need m to be an algebraic integer. But a rational number is an algebraic integer

n

if and only if it’s a rational integer; so m n

is an algebraic integer exactly when m divides n.

9
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

§2 Solutions to Day 2
§2.1 USAMO 2024/4, proposed by Rishabh Das
Available online at https://aops.com/community/p30227198.

Problem statement

Let m and n be positive integers. A circular necklace contains mn beads, each either
red or blue. It turned out that no matter how the necklace was cut into m blocks
of n consecutive beads, each block had a distinct number of red beads. Determine,
with proof, all possible values of the ordered pair (m, n).

The answer is m ≤ n + 1 only.

¶ Proof the task requires m ≤ n + 1. Each of the m blocks has a red bead count
between 0 and n, each of which appears at most once, so m ≤ n + 1 is needed.

¶ Construction when m = n + 1. For concreteness, here is the construction for n = 4,

which obviously generalizes. The beads are listed in reading order as an array with n + 1
rows and n columns. Four of the blue beads have been labeled B1 , . . . , Bn to make them
easier to track as they move.
 
R R R R
 R R R B1 
 
T0 =  R R B B2  .

 R B B B3 
B B B B4

To prove this construction works, it suffices to consider the n cuts T0 , T1 , T2 , . . . , Tn−1

made where Ti differs from Ti−1 by having the cuts one bead later also have the property
each row has a distinct red count:
     
R R R R R R R R R R R R
R R B1 R  R B1 R R  B1 R R B 
     
T1 = R B B2 R T2 = B B2 R B  T3 = 
   
B2 R B B  .

B B B3 B  B B3 B B  B3 B B B 
B B B4 R B B4 R R B4 R R R

We can construct a table showing for each 1 ≤ k ≤ n + 1 the number of red beads which
are in the (k + 1)st row of Ti from the bottom:

k T0 T1 T2 T3
k =4 4 4 4 4
k =3 3 3 3 2
.
k =2 2 2 1 1
k =1 1 0 0 0
k =0 0 1 2 3

10
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

This suggests following explicit formula for the entry of the (i, k)th cell which can then
be checked straightforwardly:

k
 k>i
#(red cells in kth row of Ti ) = k − 1 i ≥ k > 0

i k = 0.


And one can see for each i, the counts are all distinct (they are (i, 0, 1, . . . , k−1, k+1, . . . , k)
from bottom to top). This completes the construction.

¶ Construction when m < n + 1. Fix m. Take the construction for (m, m − 1) and
add n + 1 − m cyan beads to the start of each row; for example, if n = 7 and m = 5 then
the new construction is
 
C C C R R R R
C C C R R R B1 
 
T = C C C R R B B2  .

C C C R B B B3 
C C C B B B B4

This construction still works for the same reason (the cyan beads do nothing for the first
n + 1 − m shifts, then one reduces to the previous case). If we treat cyan as a shade of
blue, this finishes the problem.

11
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

§2.2 USAMO 2024/5, proposed by Anton Trygub

Available online at https://aops.com/community/p30227196.

Problem statement

Point D is selected inside acute triangle ABC so that ∠DAC = ∠ACB and ∠BDC =
90◦ + ∠BAC. Point E is chosen on ray BD so that AE = EC. Let M be the
midpoint of BC. Show that line AB is tangent to the circumcircle of triangle BEM .

This problem has several approaches and we showcase a collection of them.

¶ The author’s original solution. Complete isosceles trapezoid ABQC (so D ∈ AQ).
Reflect B across E to point F .
Q
B

D M

A C

Claim — We have DQCF is cyclic.

Proof. Since EA = EC, we have QF ⊥ AC as line QF is the image of the perpendicular

bisector of AC under a homothety from B with scale factor 2. Then

]F DC = −]CDB = 180◦ − (90◦ + ]CAB) = 90◦ − ]CAB

= 90◦ − ]QCA = ]F QC.

To conclude, note that

]BEM = ]BF C = ]DF C = ]DQC = ]AQC = ]ABC = ]ABM.

Remark (Motivation). Here is one possible way to come up with the construction of point
F (at least this is what led Evan to find it). If one directs all the angles in the obvious way,
there are really two points D and D0 that are possible, although one is outside the triangle;
they give corresponding points E and E 0 . The circles BEM and BE 0 M must then actually
coincide since they are both alleged to be tangent to line AB. See the figure below.

12
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

D′

F′
′
E
Q
B

D M

A C

One can already prove using angle chasing that AB is tangent to (BEE 0 ). So the point of
the problem is to show that M lies on this circle too. However, from looking at the diagram,
one may realize that in fact it seems
+
4M EE 0 ∼ 4CDD0

is going to be true from just those marked in the figure (and this would certainly imply the
desired concyclic conclusion). Since M is a midpoint, it makes sense to dilate 4EM E 0 from
B by a factor of 2 to get 4F CF 0 so that the desired similarity is actually a spiral similarity
at C. Then the spiral similarity lemma says that the desired similarity is equivalent to
requiring DD0 ∩ F F 0 = Q to lie on both (CDF ) and (CD0 F 0 ). Hence the key construction
and claim from the solution are both discovered naturally, and we find the solution above.
(The points D0 , E 0 , F 0 can then be deleted to hide the motivation.)

¶ Another short solution. Let Z be on line BDE such that ∠BAZ = 90◦ . This lets
us interpret the angle condition as follows:

Claim — Points A, D, Z, C are cyclic.

Proof. Because ]ZAC = 90◦ − A = 180◦ − ]CDB = ]ZDC.

13
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

D
M

E
O
A C
N W

Define W as the midpoint of BZ, so M W k CZ. And let O denote the center of
(ABC).

Claim — Points M , E, O, W are cyclic.

Proof. Note that

]M OE = ](OM , BC) + ](BC, AC) + ](AC, OE)

= 90◦ + ]BCA + 90◦
= ]BCA = ]CAD = ]CZD = ]M W D = ]M W E.

To finish, note

]M EB = ]M EW = ]M OW
= ](M O, BC) + ](BC, AB) + ](AB, OW )
= 90◦ + ]CBA + 90◦ = ]CBA = ]M BA.

This implies the desired tangency.

¶ A Menelaus-based approach (Kevin Ren). Let P be on BC with AP = P C. Let Y

be the point on line AB such that ∠ACY = 90◦ ; as ∠AY C = 90◦ − A it follows BDY C
is cyclic. Let K = AP ∩ CY , so 4ACK is a right triangle with P the midpoint of its
hypotenuse.

14
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

B
P
D M

A C

Claim — Triangles BP E and DY K are similar.

Proof. We have ]M P E = ]CP E = ]KCP = ]P KC and ]EBP = ]DBC =

]DY C = ]DY K.

Claim — Triangles BEM and Y DC are similar.

Proof. By Menelaus 4P CK with respect to collinear points A, B, Y that

BP Y C AK
= 1.
BC Y K AP
Since AK/AP = 2 (note that P is the midpoint of the hypotenuse of right triangle ACK)
and BC = 2BM , this simplifies to
BP YK
= .
BM YC
To finish, note that
]DBA = ]DBY = ]DCY = ]BM E
implying the desired tangency.

15
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

¶ A spiral similarity approach (Hans Yu). As in the previous solution, let Y be the
point on line AB such that ∠ACY = 90◦ ; so BDY C is cyclic. Let Γ be the circle through
B and M tangent to AB, and let Ω := (BCY D). We need to show E ∈ Γ.

Γ
B
P
D
M Ω
S
E
O
A C

Denote by S the second intersection of Γ and Ω. The main idea behind is to consider
the spiral similarity
Ψ:Ω→Γ C 7→ M and Y 7→ B
centered at S (due to the spiral similarity lemma), and show that Ψ(D) = E. The spiral
similarity lemma already promises Ψ(D) lies on line BD.

Claim — We have Ψ(A) = O, the circumcenter of ABC.

Proof. Note 4OBM ∼ 4AY C; both are right triangles with ]BAC = ]BOM .
+

Claim — Ψ maps line AD to line OP .

Proof. If we let P be on BC with AP = P C as before,

](AD, OP ) = ]AP O = ]OP C = ]Y CP = ](Y C, BM ).

As Ψ maps line Y C to line BM and Ψ(A) = O, we’re done.

Hence Ψ(D) should not only lie on BD but also line OP . This proves Ψ(D) = E, so
E ∈ Γ as needed.

16
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

§2.3 USAMO 2024/6, proposed by Titu Andreescu and Gabriel Dospinescu

Available online at https://aops.com/community/p30227192.

Problem statement

Let n > 2 be an integer and let ` ∈ {1, 2, . . . , n}. A collection A1 , . . . , Ak of

(not necessarily distinct) subsets of {1, 2, . . . , n} is called `-large if |Ai | ≥ ` for all
1 ≤ i ≤ k. Find, in terms of n and `, the largest real number c such that the
inequality
k X k k
!2
X |Ai ∩ Aj |2 X
xi xj ≥c xi
|Ai | · |Aj |
i=1 j=1 i=1

holds for all positive integer k, all nonnegative real numbers x1 , x2 , . . . , xk , and all
`-large collections A1 , A2 , . . . , Ak of subsets of {1, 2, . . . , n}.

The answer turns out to be

n + `2 − 2`
c= .
n(n − 1)
The following solution using indicator functions was suggested by the language reasoning
system developed by Thang Luong and the team at Google DeepMind (https://dpmd.
ai/imo-silver).

¶ Rewriting the sum. Abbreviate [n] := {1, . . . , n}, and WLOG assume xi = 1. To
P
prove the bound, we first rewrite the left-hand side using indicator functions and switch
the order of summation as
k X k
xi xj
LHS =
X
|Ai ∩ Aj |2
|Ai | |Aj |
i=1 j=1
 
k X k
xi xj 
1p∈Ai 1q∈Aj 
X X X
=
|Ai | |Aj |
i=1 j=1 p∈[n] q∈[n]
k X
X XX k
xi 1p∈A i
xj 1q∈Aj
= · .
|Ai | |Aj |
p∈[n] q∈[n] i=1 j=1

Hence, if we define the n2 variables

X xi
vp,q :=
|Ai |
i:p,q∈Ai

for each (p, q) ∈ [n]2 then the previous expression simplifies to just

LHS =
X X
2
vp,q .
p∈[n] q∈[n]

¶ Proof of bound. We split the sum into p = q versus p 6= q and bound by QM-AM
separately in the obvious way. For the terms p = q,
k k
X X xi X
vp,p = |Ai | · = xi = 1
|Ai |
p∈[n] i=1 i=1

17
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024
P 2
p∈[n] vp,p 1
=⇒
X
2
vp,p ≥ = .
n n
p∈[n]

When p 6= q we instead have

k   k k
X X |Ai | xi X X
vp,q = 2 · = (|Ai | − 1)xi ≥ (` − 1)xi = ` − 1
2 |Ai |
p6=q∈[n] i=1 i=1 i=1
P 2
p6=q∈[n] v p,q (` − 1)2
=⇒
X
2
vp,q ≥ ≥ .
n(n − 1) n(n − 1)
p6=q∈[n]

Combining the two bounds gives

1 (` − 1)2 n + `2 − 2`
LHS ≥ + =
n n(n − 1) n(n − 1)

which matches the answer claimed above.

¶ Construction. Equality occurs if vp,p is the same for all p and
vp,q is the same for all
p 6= q, and |Ai | = ` for all i. This occurs if one takes, say, k = n` and Ai to be all the
subsets of size ` each exactly once.

Remark. Originally, the write-up I used considered vectors in Rn where entries will be
2

indexed by ordered pairs (p, q) ∈ {1, . . . , n}2 . It was equivalent to the solution above, but
looking back, I think this makes the problem look harder than it actually is.
For comparison, here is how the notation with vectors works, using h•, •i for dot product,
and k•k for the vector norm. For i = 1, . . . , n define vi by

p ∈ Ai and q ∈ Ai
(
1
vi [p, q] := |Ai | v := xi vi .
X
0 otherwise; i

Then
|Ai ∩ Aj |2
xi xj hvi , vj i
XX XX
xi xj =
i j
|Ai ||Aj | i j

= hv, vi = kvk .
2

Define two more vectors e and 1; the vector e has 1 in the (p, q)th component if p = q, and
0 otherwise, while 1 has all-ones. In that case, note that

he, vi = xi he, vi i =
X X
xi
i i

h1, vi = xi h1, vi i =
X X
xi |Ai |.
i i

That means for any positive real constants α and β, by Cauchy-Schwarz for vectors, we
should have

kαe + β1k kvk ≥ hαe + β1, vi = α he, vi + β h1, vi

X X X
=α· xi + β · xi |Ai | ≥ (α + `β) xi .

18
USAMO 2024 Solution Notes web.evanchen.cc, updated 15 December 2024

Set w := αe + β1 for brevity. Then

α + β if p = q
(
w[p, q] = =⇒ kwk = n · (α + β)2 + (n2 − n) · β 2 .
p
β if p 6= q

Therefore, we get an lower bound

kvk α + `β
P ≥p
xi n · (α + β)2 + (n2 − n) · β 2

Letting α = n − ` and β = ` − 1 gives a proof that the constant

((n − `) + `(` − 1))2 (n + `2 − 2`)2 n + `2 − 2`

c= = =
n · (n − 1)2 + (n2 − n) · (` − 1)2 n(n − 1) (n + `2 − 2`) n(n − 1)

makes the original inequality always true. (The choice of α : β is suggested by the example
below.)
Meanwhile, for the construction (where k = n` , etc.) it would be sufficient to show that

w and v are scalar multiples, so that the above Cauchy-Schwarz is equality. However, we
can compute

n − 1 if p = q if p = q
( (
n−1

1
w[p, q] = , v[p, q] = `−1
· `
` − 1 if p 6= q n−2
`−2 · 1
` if p 6= q

which are indeed scalar multiples.

19