For a prime p>3, let F(p) be the largest cardinality of a subset of Fp2 containing no equilateral triangle with nonzero squared side length. Alex Iosevich asked for the exact value of F(p). By constructing such subsets of Fp2 from cap sets in F3n, we obtain the uniform lower bounds F(p)≫{p2γ,p1+γ,p≡11(mod12),p≡1(mod12). Here γ=log3(2.2202), so the two exponents are 1.452008666… and 1.726004333…, respectively. We also apply a quantitative corners theorem to show that F(p)=o(p2) as p→∞ through either congruence class.
Introduction
For a prime p>3, let Fp denote the field with p elements, and let F(p) be the largest cardinality of a subset of Fp2 containing no equilateral triangle with nonzero squared side length. For w=(a,b)∈Fp2, write ∥w∥2=a2+b2. The squared distance between w=(a,b) and w′=(a′,b′) is
∥w−w′∥2=(a−a′)2+(b−b′)2,
and an equilateral triangle consists of three distinct points X,Y,Z satisfying
∥X−Y∥2=∥Y−Z∥2=∥Z−X∥2=0.(1)
The problem of determining F(p) was raised by Alex Iosevich.
One of the motivations behind this question comes from the finite-field simplex problem.
The finite-field simplex problem asks how large a subset of a finite vector space must be to determine many, or all, congruence classes of simplices. Two configurations are congruent if a translation followed by an orthogonal linear map takes one to the other. This problem has been studied by Fourier analysis, incidence estimates, group actions, and spectral graph methods [1, 2, 11, 12].
Bennett, Iosevich, and Pakianathan [2] used the Elekes–Sharir method to show that, for every prime p≡3(mod4), any set E⊆Fp2 with ∣E∣≫p7/4 determines ≫p3 congruence classes of triangles, a positive proportion of all such classes. Bennett, Hart, Iosevich, Pakianathan, and Rudnev [1 Theorem 1.6] showed that, for every prime p>3, the same conclusion holds whenever ∣E∣≫p8/5. For p≡3(mod4), Pham [8 Theorem 1.2] proved a stronger conclusion under the same size condition, namely, E determines ≫p2 triangle congruence classes with any prescribed nonzero squared side length.
A set may determine many triangle congruence classes and still miss a particular one. Suppose that t1,t2,t3 are the squared side lengths of a noncollinear triangle Δ in Fp2, and that all three are nonzero. How large must E⊆Fp2 be to contain a triangle with these three squared side lengths?
This question is better understood in higher dimensions. View Δ in Fpd by appending zero coordinates. For d≥4, results of Vinh [11] and Iosevich and Parshall [4] show that every set E⊆Fpd with ∣E∣≥Cp(d+3)/2 contains a congruent copy of Δ. Parshall’s simplex theorem [7] gives the same conclusion for d≥3 under the hypothesis ∣E∣≥Cp2(d+1)/3. This last exponent equals 2 at d=2, although the theorem requires d≥3. In the plane, we know of no ε>0 for which ∣E∣≥Cp2−ε suffices for every such choice of t1,t2,t3.
Consider first the case t1=t2=t3=1. How large can E be if it contains no unit equilateral triangle? We can also allow the common squared side length to vary and ask how large E can be if it contains no equilateral triangle of any nonzero squared side length. This is Iosevich’s question about F(p).
Bennett, Iosevich, and Pakianathan [2, Theorem 1.4 and Corollary 4.2] showed that an equilateral triangle with nonzero squared side length exists in Fp2 if and only if 3 is a square in Fp. For p>3, this is equivalent to p≡1,11(mod12). Thus, F(p)=p2 when p≡5,7(mod12). We therefore focus on p≡1,11(mod12).
The restriction to nonzero squared side length matters when −1 is a square. In that case, the plane is partitioned into p parallel isotropic lines, on each of which all squared distances are zero. If triples of distinct points with all three squared distances zero were also forbidden, each isotropic line would contain at most two points of the set. The set would then have at most 2p points.
Main results
Polansky [9, Sections 3–4], in his honors thesis, constructed twelve points modulo seven in triangular-lattice coordinates and combined copies by base-seven expansions. His lattice construction gives
F(p)≫plog712,log712=1.276989…,
for the problem with nonzero squared side length considered here.
Using cap sets in F3n, we improve this lower bound in both residue classes.1
Theorem 1.1. Set γ=log3(2.2202). There is an absolute constant c>0 such that, for every prime p≡1,11(mod12),
F(p)≥c{p2γ,p1+γ,p≡11(mod12),p≡1(mod12).
Recall that a cap set in F3n is a set C⊆F3n for which
ξ+η+ζ=0,ξ,η,ζ∈C⟹ξ=η=ζ.
Equivalently, C contains no nonconstant three-term arithmetic progression. The following proposition gives a lower bound for F(p) in terms of n and ∣C∣, which is the key tool in proving Theorem 1.1.
Proposition 1.2. Let n≥1, and let C⊆F3n be a nonempty cap set. For every prime p≡1,11(mod12),
The exponents in Theorem 1.1 are 2γ=1.452008666… and 1+γ=1.726004333…, respectively. Since 1+γ>8/5, our construction gives, for every sufficiently large prime p≡1(mod12), a set that determines ≫p3 triangle congruence classes by [1 Theorem 1.6] yet contains no equilateral triangle of nonzero squared side length.
For the upper bound, we use a linear change of coordinates that sends each corner with nonzero step to an equilateral triangle with nonzero squared side length. Recall that a corner in Fp2 is a triple (s,t),(s+h,t),(s,t+h) with h=0. The quantitative corners theorem of Jaber, Liu, Lovett, Ostuni, and Sawhney [5] then gives the following bound. In particular, F(p)=o(p2) as p→∞ through primes p≡1,11(mod12).
Theorem 1.3. There is an absolute constant c>0 such that, for every prime p>3 with p≡1,11(mod12),
F(p)≤p2exp[−c(logp)1/600].
Corollary 1.4. Let c>0 be as in Theorem 1.3, and let p>3 be a prime with p≡1,11(mod12). Every set E⊆Fp2 satisfying
∣E∣>p2exp[−c(logp)1/600]
contains an equilateral triangle with nonzero squared side length.
In the continuous setting, Maga [6] constructed a compact set E⊆R2 with dimH(E)=2 containing no equilateral triangle. For d≥4, Iosevich and Liu [3] proved that compact subsets of Rd satisfying a Frostman bound with exponent sufficiently close to d contain an equilateral triangle.
Notation. We write f≪g or g≫f when f≤c0g for a constant c0>0. The implied constants are independent of the field size. We indicate any other dependence explicitly. We use log for the natural logarithm.
Proof of the key proposition
To prove Proposition 1.2, we first describe the coordinate change and linear form used to construct equilateral-triangle-free sets. We then construct integer point sets from cap sets and transfer them to Fp2.
Finite-field geometry
Let p>3 be a prime. Let S={e1=(1,0),e2=(0,1)} be the standard basis of Fp2, and write (a,b) for the coordinates of a point with respect to S. These are the coordinates used in Section 1, so the squared length of a point is ∥(a,b)∥2=a2+b2, and the dot product is (a,b)⋅(a′,b′)=aa′+bb′.
For the rest of this section, we assume that p≡1,11(mod12). By quadratic reciprocity, this is exactly the condition that 3 is a square in Fp, and we fix r∈Fp with r2=3. Let
B={v1=(1,0),v2=(21,2r)}.
be a second ordered basis of Fp2, written in the coordinates (a,b), and write (s,t) for the coordinates of a point with respect to B. The map
Φ:Fp2⟶Fp2,Φ(s,t)=(s+2t,2rt),
converts (s,t)-coordinates into (a,b)-coordinates:
(a,b)=Φ(s,t)=sv1+tv2.
It is invertible, since its determinant is r/2=0. To distinguish the two coordinate systems, we write
Fp2=(Fp2,(s,t)),(Fp2)′=(Fp2,(a,b)),
so that Φ:Fp2→(Fp2)′, and (Fp2)′ is the plane of Section 1 in which F(p) is defined.
We recall the criterion for equilateral triangles from [2] and verify the identities needed below. Define
The formulas for R and Q have integer coefficients. We use the same formulas to define R:Z2→Z2 and Q:Z2→Z, and also over R. Over R, the form Q is positive definite because Q(s,t)=(s+t/2)2+3t2/4.
Notations. Throughout, for a prime p, we identify Fp={0ˉ,…,p−1}. For m≥1 and the rings F=Fp,Z, the map Im:Fm→Fm denotes the identity map. The map π:Zm→Fpm denotes the natural projection; and θ:Fpm→Zm denotes a set-theoretic section of π, i.e., π∘θ=Im on Fpm, such that Im(θ)={0,…,p−1}m. For a subset B of Fpm, let B:=θ(B), which is referred to as a lift of B (along π). For v∈Zm, we denote vˉ=π(v).
Conventions. The letters X,Y,Z denote points of (Fp2)′, and the letters x,y,z denote points of Fp2. The letters ξ,η,ζ denote points of Fpn, (Fp2)n, or Z2n, including integer points of Z2. Scalars in Fp obtained from the linear form ℓ of Lemma 2.1(iv) are denoted u,u1,u2,u3.
Lemma 2.1. Let p>3 be a prime. Then the following hold.
(i) An equilateral triangle with nonzero squared side length exists in (Fp2)′ if and only if p≡1,11(mod12).
Suppose that p≡1,11(mod12), and let r, Φ, R, Q, and T be as above.
(ii) Three points X,Y,Z∈(Fp2)′ form an equilateral triangle with nonzero squared side length if and only if, after possibly interchanging X and Y,
Z−X=T(Y−X)and∥Y−X∥2=0.
(iii) For x,y,z∈Fp2, the points Φ(x),Φ(y),Φ(z)∈(Fp2)′ form an equilateral triangle with nonzero squared side length if and only if, after possibly interchanging x and y,
z−x=R(y−x)andQ(y−x)=0.(2)
(iv) Suppose additionally that p≡1(mod12). There exists λ∈Fp satisfying λ2−λ+1=0. The linear form
ℓ:Fp2⟶Fp,ℓ(s,t)=s+λt,
satisfies
ℓ(R(s,t))Q(s,t)=λℓ(s,t),=ℓ(s,t)(s+λ−1t)(3)
for every (s,t)∈Fp2.
Parts (ii) and (iii) characterize equilateral triangles with nonzero squared side length by linear equations. The map T preserves squared lengths and has determinant 1, so we call it a rotation; over R, taking r=3 gives rotation through 60∘. We call (2) the rotation equation, which is a main tool in our argument.
Part (iv) provides a linear form whose fibers ℓ−1(u), with u∈Fp, are parallel affine lines in Fp2. If x,y∈Fp2 satisfy ℓ(x)=ℓ(y), then (3) gives
Q(y−x)=0,∥Φ(y)−Φ(x)∥2=0.
Thus, Φ sends these fibers to parallel isotropic lines in (Fp2)′. Note that the map ℓ is defined only in the case p≡1(mod12). Using this map, we will obtain a better bound in Proposition 1.2 in this case
Proof. Define J(a,b)=(−b,a). Suppose that X,Y,Z form an equilateral triangle with nonzero squared side length τ. Put W=Y−X and W′=Z−X. Then
∥W∥2=∥W′∥2=τ,W⋅W′=τ/2.
Since det(W,JW)=τ=0, the vectors W,JW form a basis. Using W⋅JW=0, we can write
W′=21W+2ρJW
for some ρ∈Fp. Direct computation show that
∥W′∥2=41+ρ2∥W∥2.
Thus we have
ρ=41+ρ2ρ.
so ρ2=3. By quadratic reciprocity, this implies p≡1,11(mod12). Conversely, if r2=3, then (0,0),(2,0),(1,r) form an equilateral triangle of nonzero squared side length. This proves (i).
Since ρ=±r, the calculation in (i) shows that either Z−X=T(Y−X) or Z−X=T−1(Y−X). In the latter case, T−1=I2−T gives
Z−Y=T(X−Y).
Thus, after possibly interchanging X and Y, we obtain the required equation. Conversely, if Z−X=T(Y−X) and ∥Y−X∥2=0, then using (4) with w=Y−X, we can show that X,Y,Z form an equilateral triangle with nonzero squared side length. This proves (ii).
For (iii), apply (ii) to Φ(x),Φ(y),Φ(z). Since Φ is a linear bijection satisfying ΦR=TΦ and ∥Φ(w)∥2=Q(w), the conditions in (ii) are equivalent to
z−x=R(y−x),Q(y−x)=0,
after possibly interchanging x and y. This proves (iii).
Finally, suppose that p≡1(mod12). Choose ι∈Fp with ι2=−1, and set λ=(1+ιr)/2. Then
λ(1−λ)=41+r2=1,
so λ2−λ+1=0 and λ−1=1−λ. For ℓ(s,t)=s+λt, we obtain
ℓ(R(s,t))=−t+λ(s+t)=λ(s+λt)=λℓ(s,t),
and
(s+λt)(s+λ−1t)=s2+st+t2=Q(s,t).
This proves (iv), completing the proof.
Encoding a cap set by integer points
We construct integer point sets for which the rotation equation forces the three points to coincide. The construction has three steps, and we state each as a separate lemma.
For n≥1, let R act on (F32)n and on Z2n by applying the formula for R to each coordinate pair:
R((sj,tj)j=0n−1)=((−tj,sj+tj))j=0n−1.
For n=1, this is the map R itself. We say that a subset S of (F32)n or of Z2n is rotation-free if
ζ=(I2n−R)ξ+Rη,ξ,η,ζ∈S⟹ξ=η=ζ.
with arithmetic in F3 or in Z, respectively. Lemma 2.2 produces a rotation-free subset of (F32)n from a cap set. Lemma 2.3 lifts it to integer vectors by base-three expansions, and Lemma 2.4 packs the integer vectors into the plane by expansions in a larger base. Proposition 2.5 combines the three steps.
Lemma 2.2. Let n≥1, and let C⊆F3n be a cap set. Define
By the definition of a cap set, we have αξ=αη=αζ, and hence βξ=βη=βζ. Since σ is injective, ξ=η=ζ. Thus, D is rotation-free, as desired.
Let B be any subset of (F3)n, so that B is a subset of (Z2)n. For each k≥1, define Γ:(B)k→Z2n by
Γ(δ0,…,δk−1)=i=0∑k−13iδi.
Let Ek be the image of Γ. Then, for each component νj of an element ν of Ek, we have 0≤νj≤3k−1. Thus, Ek⊆{0,…,3k−1}2n.
For ξ∈Ek, δξ,j denotes δj∈B for j=0,…,k−1, where ξ=Γ(δ0,…,δk−1). Note that ξ≡δξ,0 (mod 3) by the definition of ξ.
Lemma 2.3. Let n≥1, let B be a subset of (F3)n and let Ek be as above for k≥1. Then the following hold.
(i) For each k≥1, Γ is injective, and so ∣Ek∣=∣B∣k.
(ii) If B is rotation-free, then so is Ek in Z2n for every k≥1.
Proof. (i): Injectivity of Γ follows from uniqueness of base-three expansions. Since Γ is injective, we have ∣Ek∣=∣B∣k.
(ii): We proceed by induction on k. Suppose that ξ,η,ζ∈Ek satisfy
ζ=(I2n−R)ξ+Rη.(5)
Since ξ≡δξ,0, η≡δη,0, ζ≡δζ,0 (mod 3), the equation (5) reduces to
δζ,0≡(I2n−R)δξ,0+Rδη,0(mod3).(6)
Let δξ,0,δη,0,δζ,0∈B be images under π of the elements δξ,0,δη,0,δζ,0∈B. By (6), we have, over F3,
δζ,0=(I2n−R)δξ,0+Rδη,0.(7)
Since B is rotation-free, we have
δξ,0=δη,0=δζ,0.
Applying θ to this equality, we obtain
δξ,0=δη,0=δζ,0,
since θ(π(δ))=δ for every δ∈B. If k=1, which is the base case for the induction, then E1=B and
δξ,0=ξ,δη,0=η,δζ,0=ζ.
Thus, ξ=η=ζ, and so E1 is rotation-free.
Now assume k≥2. Write δ0:=δξ,0=δη,0=δζ,0. Define ξ′,η′,ζ′ by
ξ=δ0+3ξ′,η=δ0+3η′,ζ=δ0+3ζ′.
Substitute these into (5):
δ0+3ζ′=(I2n−R)(δ0+3ξ′)+R(δ0+3η′).
Then since (I2n−R)+R=I2n, we obtain
ζ′=(I2n−R)ξ′+Rη′.
Note that ξ′,η′,ζ′ are elements of Ek−1. By the induction hypothesis that Ek−1 is rotation-free, we obtain ξ′=η′=ζ′. Thus, ξ=η=ζ, and so Ek is rotation-free.
For ξ=((sj,tj))j=0n−1, let ξ(j) be the j-th coordinate pair, i.e., ξ(j)=(sj,tj).
Lemma 2.4. Let n≥1 and fix an integer M≥0. Let G be a subset of Z2n with G⊆{0,…,M}2n.
(i) If N≥max{2,2M+1}, then Ψ is injective on G, and
Ψ(G)⊆{0,…,MN−1Nn−1}2.
(ii) If N≥max{2,2M+1} and G is rotation-free, then so is Ψ(G) in Z2.
Proof. (i): Since 0≤sj,tj≤M<N for ξ∈G, injectivity of Ψ on G follows from uniqueness of base-N expansions. Both coordinates of Ψ(ξ) lie between 0 and
Mj=0∑n−1Nj=MN−1Nn−1.
(ii): To prove (ii), assume that for ξ,η,ζ∈G
Ψ(ζ)=(I2−R)Ψ(ξ)+RΨ(η).(8)
Let us show that Ψ(ξ)=Ψ(η)=Ψ(ζ). For each j=0,…,n−1, let e(j)∈Z2 be the j-th error term defined by
e(j)=ζ(j)−(I2−R)ξ(j)−Rη(j).
Using the fact that Ψ∘R=R∘Ψ, and R,R are linear, from (8) we can derive
j=0∑n−1Nje(j)=0.(9)
By rewriting the equality (9) as
e(0)=(e1(0),e2(0))=−Nj=1∑n−1Nj−1e(j),
we can easily see that each component ei(0) of e(0) (i=1,2) is divisible by N in Z.
On the other hand, each component ei(j) of e(j) is, by definition, a sum of two entries minus two entries for i=1,2, and each entry lies between 0 and M, so for every j=0,…,n−1,
∣e1(j)∣,∣e2(j)∣≤2M≤N−1.
For j=0, since ei(0) is a multiple of N and ∣ei(0)∣≤N−1 for i=1,2 we have ei(0)=0, so e(0)=(0,0).
Removing this zero term from (9) gives ∑j=1n−1Nje(j)=0. After dividing by N, the same argument gives e(1)=(0,0) if n≥2. Continuing this process, we obtain
e(0)=e(1)=⋯=e(n−1)=(0,0).
Thus, ξ,η,ζ∈G satisfy
ζ=(I2n−R)ξ+Rη.
Since G is rotation-free, we have ξ=η=ζ. Therefore Ψ(ξ)=Ψ(η)=Ψ(ζ). Thus, the subset Ψ(G)⊂Z2 is rotation-free.
We now combine the three lemmas.
Proposition 2.5. Let n≥1 and C⊆F3n be a nonempty cap set. For every integer k≥0, define
Nk=2⋅3k−1.
There is an explicit set
Sk⊂{0,…,2Nkn−1}2,∣Sk∣=∣C∣2k,
such that
ζ=(I2−R)ξ+Rη,ξ,η,ζ∈Sk⟹ξ=η=ζ.
Proof. For k=0, take S0={(0,0)}. Assume henceforth that k≥1. Let D be the rotation-free set of Lemma 2.2, so that ∣D∣=∣C∣2. Next, we apply Lemma 2.3 to B=D. Then we obtain the rotation-free subset Ek⊆{0,…,3k−1}2n such that ∣Ek∣=∣D∣k=∣C∣2k. Lastly, apply Lemma 2.4 with M=3k−1 and N=Nk=2M+1, and set Sk=Ψ(Ek). Then ∣Sk∣=∣Ek∣=∣C∣2k, the set Sk is rotation-free in Z2, and its coordinates lie between 0 and
MNk−1Nkn−1=2Nkn−1,
since Nk−1=2M. This proves the proposition.
From integer points to the finite plane
We now use the integer sets from Proposition 2.5 to construct equilateral-triangle-free subsets of Fp2.
Proof of Proposition 1.2. We use direct reduction modulo p when p≡11(mod12). When p≡1(mod12), we first construct a scalar set and then use parallel isotropic lines to gain a factor of p. We finish by choosing the largest admissible scale in each case.
Let Sk be the set from Proposition 2.5, with Nk=2⋅3k−1. Thus, ∣Sk∣=∣C∣2k, and
ζ=(I2−R)ξ+Rη in Z2,ξ,η,ζ∈Sk⟹ξ=η=ζ.
Let π:Z2→Fp2 be reduction modulo p, and write ξˉ=π(ξ) for ξ∈Z2. Since R has integer coefficients, π(Rξ)=Rξˉ. We will obtain equations modulo p and use the coordinate bounds to recover the exact integer equation above. For ξ,η,ζ∈Sk, define the integer vector e=ζ−(I2−R)ξ−Rη. Then
Note that each of e1, e2, e1+e2 is expressed as a sum of two entries minus two entries. Since every entry lies between 0 and (Nkn−1)/2, we obtain
∣e1∣,∣e2∣,∣e1+e2∣≤Nkn−1.(10)
The case p≡11(mod12). Choose k with Nkn≤p. Define
E=Φ(π(Sk))⊆(Fp2)′.
Reduction modulo p is injective on Sk, since its coordinates lie between 0 and (p−1)/2. As Φ is also invertible, ∣E∣=∣Sk∣=∣C∣2k.
Suppose that X,Y,Z∈E form an equilateral triangle with nonzero squared side length. Choose the unique ξ,η,ζ∈Sk such that
X=Φ(ξˉ),Y=Φ(ηˉ),Z=Φ(ζˉ).
By (2), after possibly interchanging X,Y and the corresponding ξ,η, we have
ζˉ=(I2−R)ξˉ+Rηˉin Fp2.
Thus, the integer error e=ζ−(I2−R)ξ−Rη satisfies π(e)=0, so p∣e1,e2. On the other hand, (10) gives
∣e1∣,∣e2∣≤Nkn−1<p.
Hence, e=0, and we have the exact equality
ζ=(I2−R)ξ+Rηin Z2.
Proposition 2.5 now gives ξ=η=ζ, so X=Y=Z, a contradiction. Therefore,
F(p)≥∣C∣2kwhenever Nkn≤p.(11)
The case p≡1(mod12). The argument made so far in the case p≡11(mod12) applies verbatim when p≡1(mod12). However, the case p≡1(mod12) is more favorable since there are isotropic lines in (Fp2)′. Using this advantage we shall find a better bound in this case.
Choose k with Nkn≤p. Let λ and ℓ be as in Lemma 2.1(iv). We first show that
ℓ(ζˉ)=(1−λ)ℓ(ξˉ)+λℓ(ηˉ),ξ,η,ζ∈Sk⟹ξ=η=ζ.(12)
Suppose that the scalar equation holds, and set e=ζ−(I2−R)ξ−Rη∈Z2. The first identity in (3) gives
ℓ(eˉ)=ℓ(ζˉ)−(1−λ)ℓ(ξˉ)−λℓ(ηˉ)=0.
The second identity then yields Q(eˉ)=0 in Fp. Equivalently, the integer Q(e) is divisible by p. We will show that 0≤Q(e)<p, which forces Q(e)=0 over the integers.
For real s,t,
Q(s,t)≤max{s2,t2,(s+t)2}.(13)
Indeed, if st≥0, then Q(s,t)=(s+t)2−st≤(s+t)2. If st<0 and ∣s∣≥∣t∣, then st+t2≤0, so Q(s,t)≤s2; the other case is symmetric. Combining (10) and (13), we obtain
0≤Q(e)≤(Nkn−1)2<p.
Since p∣Q(e), we have Q(e)=0 in Z. Positive definiteness of Q over the integers gives e=0. Thus, ζ=(I2−R)ξ+Rη in Z2, and Proposition 2.5 yields ξ=η=ζ. This proves (12).
The map ξ↦ℓ(ξˉ) is also injective on Sk. Indeed, if ℓ(ξˉ)=ℓ(ηˉ), then
ℓ(ηˉ)=(1−λ)ℓ(ξˉ)+λℓ(ξˉ).
Then (12) applied to (ξ,ξ,η) gives ξ=η. This proves the injectivity. Consequently, the set
U={ℓ(ξˉ):ξ∈Sk}⊆Fp
has cardinality ∣U∣=∣Sk∣=∣C∣2k. Moreover, we claim that for u1,u2,u3∈U
u3=(1−λ)u1+λu2u1,u2,u3∈U⟹u1=u2=u3.(14)
Indeed, choosing representatives ξ,η,ζ∈Sk so that ℓ(ξˉ)=u1,ℓ(ηˉ)=u2,ℓ(ζˉ)=u3, and again applying (12), we obtain u1=u2=u3. Define
A=ℓ−1(U)={x∈Fp2:ℓ(x)∈U},E=Φ(A)⊆(Fp2)′.
For each u∈U, the fiber ℓ−1(u) is the line {(s,t)∈Fp2:s+λt=u}, which contains exactly p points. The fibers are disjoint, and Φ is invertible, so ∣E∣=∣A∣=p∣U∣=p∣Sk∣=p∣C∣2k.
Suppose that E⊆(Fp2)′ contains an equilateral triangle X,Y,Z with nonzero squared side length. Assume that Φ(x)=X,Φ(y)=Y,Φ(z)=Z for x,y,z∈A⊆Fp2. (Note that x,y,z are already points of Fp2, not elements of Sk⊆Z2.) Reordering X,Y,Z, if necessary, by (2) we obtain
z=(I2−R)x+Ry,Q(y−x)=0.
Applying ℓ gives
ℓ(z)=(1−λ)ℓ(x)+λℓ(y).
Since ℓ(x),ℓ(y),ℓ(z)∈U, by (14) we obtain
ℓ(x)=ℓ(y)=ℓ(z).
In particular, ℓ(y−x)=0, so (3) gives Q(y−x)=0, a contradiction. Therefore,
F(p)≥p∣C∣2kwhenever Nkn≤p.(15)
Let L=p in (11) and L=p in (15). Choose the largest k≥0 with Nkn≤L. Such a k exists because N0=1. Since Nk=2⋅3k−1, the maximal choice of k gives
L<(2⋅3k+1−1)n<6n3kn,3k>L1/n/6.
If ∣C∣>1, raising the latter inequality to the power 2log3∣C∣ gives
∣C∣2k≥6−2log3∣C∣L2log3∣C∣/n.(16)
If ∣C∣=1, then both sides of (16) equal 1, so it trivially holds. Substituting (16) into (11) and (15), we obtain both bounds in the proposition.
Proofs of Theorems 1.1 and 1.3
We combine Proposition 1.2 with the cap set construction of Romera-Paredes et al. [10] to obtain the exponents in Theorem 1.1.
Proof of Theorem 1.1. The construction of Romera-Paredes et al. [10 p. 472 and Supplementary Information, Figure C.9] gives an integer n≥1 and a cap set C⊆F3n with ∣C∣≥2.2202n=3nγ. Fix such n and C. Since log3∣C∣/n≥γ, Proposition 1.2 gives the required bounds with c=6−2log3∣C∣>0. Since n and C are fixed independently of p, the constant c is absolute. This proves Theorem 1.1.
We use Φ to map each corner to an equilateral triangle with nonzero squared side length. The quantitative corners theorem then bounds the size of an equilateral-triangle-free set.
Proof of Theorem 1.3. Let E⊆(Fp2)′ contain no equilateral triangle with nonzero squared side length. Let r and Φ be as in § 2, so that Φ−1(E)⊆Fp2 is the same set in the coordinates (s,t). We claim that Φ−1(E) doesn’t contain any corners. Suppose, to the contrary, that Φ−1(E) contains a corner
(s,t),(s+h,t),(s,t+h),h=0.
By linearity and the definition of Φ, the differences between these image points are
Φ(s+h,t)−Φ(s,t)=Φ(h,0)=(h,0),
Φ(s,t+h)−Φ(s,t)=Φ(0,h)=(2h,2rh),
Φ(s,t+h)−Φ(s+h,t)=Φ(−h,h)=(−2h,2rh).
The first vector has squared length h2. Since r2=3, the other two vectors have squared lengths
(2h)2+(2rh)2=(−2h)2+(2rh)2=41+r2h2=h2.
Since h=0, the image points Φ(s,t),Φ(s+h,t),Φ(s,t+h) in (Fp2)′ form an equilateral triangle with nonzero squared side length, contrary to the choice of E. Thus, Φ−1(E) doesn’t contain any corner. Applying the corners theorem of Jaber, Liu, Lovett, Ostuni, and Sawhney [5 Theorem 1.1, full version] to the additive group Fp gives
[1]M. Bennett, D. Hart, A. Iosevich, J. Pakianathan, and M. Rudnev, Group actions and geometric combinatorics in F_q^d, Forum Mathematicum, 29(1) (2017), 91–110.
[2]M. Bennett, A. Iosevich, and J. Pakianathan, Three-point configurations determined by subsets of F_q^2 via the Elekes–Sharir paradigm, Combinatorica, 34(6) (2014), 689–706.arxiv.org/abs/1201.5039
[3]A. Iosevich and B. Liu, Equilateral triangles in subsets of R^d of large Hausdorff dimension, Israel Journal of Mathematics, 231(1) (2019), 123–137.
[4]A. Iosevich and H. Parshall, Embedding distance graphs in finite field vector spaces, Journal of the Korean Mathematical Society, 56(6) (2019), 1515–1528.arxiv.org/abs/1802.06460
[5]M. Jaber, Y. P. Liu, S. Lovett, A. Ostuni, and M. Sawhney, Quasipolynomial bounds for the corners theorem, in Proceedings of the 2025 IEEE 66th Annual Symposium on Foundations of Computer Science, IEEE, 2025, 2027–2036; full version, arXiv:2504.07006.arxiv.org/abs/2504.07006
[6]P. Maga, Full dimensional sets without given patterns, Real Analysis Exchange, 36(1) (2010/11), 79–90.DOI
[7]H. Parshall, Simplices over finite fields, Proceedings of the American Mathematical Society, 145(6) (2017), 2323–2334.arxiv.org/abs/1606.04040
[8]T. Pham, Triangles with one fixed side-length, a Furstenberg-type problem, and incidences in finite vector spaces, Forum Mathematicum, 37(2) (2025), 359–372.arxiv.org/abs/2304.10897
[10]B. Romera-Paredes, M. Barekatain, A. Novikov, M. Balog, M. P. Kumar, E. Dupont, F. J. R. Ruiz, J. S. Ellenberg, P. Wang, O. Fawzi, P. Kohli, and A. Fawzi, Mathematical discoveries from program search with large language models, Nature, 625(7995) (2024), 468–475.
[11]L. A. Vinh, On kaleidoscopic pseudo-randomness of finite Euclidean graphs, Discussiones Mathematicae Graph Theory, 32(2) (2012), 279–287.arxiv.org/abs/0807.2689
[12]L. A. Vinh, The solvability of norm, bilinear and quadratic equations over finite fields via spectra of graphs, Forum Mathematicum, 26(1) (2014), 141–175.