Wiki
Wiki

Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.

Updated

Second order construction


Statement and scope

There is an unbounded sequence of integers NN for which

M(N)≥23N+(229 31/4−o(1))N1/4.(1)M(N)\ge \frac{2}{\sqrt3}\sqrt N+ \left(\frac{2\sqrt2}{9\,3^{1/4}}-o(1)\right)N^{1/4}. \tag{1}

Equivalently,

lim sup⁡N→∞M(N)−(2/3)NN1/4≥229 31/4.\limsup_{N\to\infty} \frac{M(N)-(2/\sqrt3)\sqrt N}{N^{1/4}} \ge \frac{2\sqrt2}{9\,3^{1/4}}.

The construction includes diagonal pair sums and has only one possible exceptional integer sum. This does not improve the leading constant or settle the proposed upper bound.

The idea is to compress the reflected construction by an amount of order N3/4N^{3/4}, while deleting only order N1/4N^{1/4} points to eliminate the new collisions. Averaging over the compression parameter reduces the required relation count to an ordinary Cartesian-product count; no sparse four-variable counting lemma is assumed.

A cyclic Sidon set and its translate variance

For a prime power qq, put v=q2+q+1v=q^2+q+1. There is a set D0⊂ZvD_0\subset\mathbb Z_v, of size q+1q+1, in which every nonzero ordered difference occurs exactly once.

Here is a self-contained verification of the standard Singer construction used below. Let K=Fq3K=\mathbb F_{q^3}, and let UU be a two-dimensional Fq\mathbb F_q-subspace of KK. The projective points of UU give q+1q+1 elements of the cyclic group K∗/Fq∗K^*/\mathbb F_q^*, which has order vv. For x∈K∗∖Fq∗x\in K^*\setminus\mathbb F_q^*, the subspaces UU and xUxU intersect in exactly one dimension. They cannot coincide: otherwise for 0≠u∈U0\ne u\in U, the three elements u,xu,x2uu,xu,x^2u would belong to UU, contradicting the degree-three minimal polynomial of xx. Their unique common projective point gives exactly one representation of the quotient class of xx as a quotient of two points of D0D_0. Under a cyclic identification, this is the required difference property. It implies ordinary modular Sidonicity, including diagonal pairs.

For comparison, see §3.4 and entry [4] of source record. The proof above supplies all properties used here.

For any set I⊂ZvI\subset\mathbb Z_v of ℓ\ell residues, and a uniformly random translate τ\tau, let

XI(τ)=∣(D0+τ)∩I∣.X_I(\tau)=|(D_0+\tau)\cap I|.

The perfect-difference property gives

∑τXI(τ)=(q+1)ℓ,∑τXI(τ)(XI(τ)−1)=ℓ(ℓ−1).\sum_\tau X_I(\tau)=(q+1)\ell,\qquad \sum_\tau X_I(\tau)(X_I(\tau)-1)=\ell(\ell-1).

Since (q+1)2=v+q(q+1)^2=v+q, it follows that

EXI=(q+1)ℓv,Var⁡XI=qℓ(v−ℓ)v2.(2)\mathbb E X_I=\frac{(q+1)\ell}{v},\qquad \operatorname{Var}X_I=\frac{q\ell(v-\ell)}{v^2}. \tag{2}

Simultaneous distribution in two short endpoint intervals

Fix c>0c>0, and put W=⌊cq3/2⌋W=\lfloor c q^{3/2}\rfloor. For every sufficiently large qq, one can choose a translate of D0D_0, represented by D⊂[0,v−1]D\subset[0,v-1], so that both measures

μL=qW∑d∈D0≤d≤Wδd/W,μR=qW∑d∈Dv−1−W≤d≤v−1δ(v−1−d)/W(3)\mu_L=\frac qW\sum_{\substack{d\in D\\0\le d\le W}}\delta_{d/W}, \qquad \mu_R=\frac qW\sum_{\substack{d\in D\\v-1-W\le d\le v-1}} \delta_{(v-1-d)/W} \tag{3}

converge weakly to Lebesgue measure on [0,1][0,1], as qq tends to infinity through prime powers.

For completeness, apply (2) to the endpoint prefixes of relative length j/Jj/J, where J=⌊q1/8⌋J=\lfloor q^{1/8}\rfloor. Their normalized counts have expectation j/J+o(1)j/J+o(1), uniformly in jj, and variance O(q/W)=O(q−1/2)O(q/W)=O(q^{-1/2}). Chebyshev's inequality, with normalized tolerance q−1/8q^{-1/8}, bounds each failure probability by O(q−1/4)O(q^{-1/4}). There are O(J)O(J) prefixes in both windows, so the probability of any failure is O(q−1/8)O(q^{-1/8}). A good translate exists. Monotonicity fills the gaps between grid points and gives simultaneous uniform convergence of the two distribution functions. In particular, both endpoint intervals contain O(W/q)O(W/q) elements of DD.

No independence between the two endpoint measures is asserted or needed. For this deterministic good translate, their Cartesian products converge to the corresponding products of Lebesgue measure.

Averaging the new forbidden relations

Fix also a∈(0,1)a\in(0,1). For each integer ⌈aW⌉≤w≤W\lceil aW\rceil\le w\le W, let EwE_w count pairs ({d1,d2,d3},d4)(\{d_1,d_2,d_3\},d_4), with an unordered multiset of three inputs and di∈Dd_i\in D, satisfying

d1+d2+d3=d4−(v−1)+w.(4)d_1+d_2+d_3=d_4-(v-1)+w. \tag{4}

Every input belongs to [0,W][0,W], and every output to [v−1−W,v−1][v-1-W,v-1]. On writing

xi=di/W(1≤i≤3),x4=(v−1−d4)/W,x_i=d_i/W\quad(1\le i\le3),\qquad x_4=(v-1-d_4)/W,

equation (4) is equivalent to ∑ixi=w/W\sum_i x_i=w/W. Thus the sum over these ww of the ordered relation counts equals the number of endpoint quadruples in the simplex shell

a≤x1+x2+x3+x4≤1,a\le x_1+x_2+x_3+x_4\le1,

with the lower boundary adjusted by O(W−1)O(W^{-1}) for rounding. The limiting product measure assigns the boundary measure zero. By (3), this count is

(1−a424+o(1))(Wq)4.(5)\left(\frac{1-a^4}{24}+o(1)\right)\left(\frac Wq\right)^4. \tag{5}

Indeed the simplex of nonnegative quadruples with sum at most t≤1t\le1 has volume t4/24t^4/24.

Quadruples with a repeated input contribute only

O ⁣((W/q)3)=O(q3/2):O\!\left((W/q)^3\right)=O(q^{3/2}):

choose which two input positions coincide, their common value, the remaining input, and the output. For all other tuples there are exactly six input orderings. Consequently

∑w=⌈aW⌉WEw=(1−a4144+o(1))(Wq)4+O(q3/2).\sum_{w=\lceil aW\rceil}^{W}E_w =\left(\frac{1-a^4}{144}+o(1)\right)\left(\frac Wq\right)^4 +O(q^{3/2}).

There are (1−a)W+O(1)(1-a)W+O(1) parameters. With Sa=1+a+a2+a3S_a=1+a+a^2+a^3, some ww therefore satisfies

Ew≤(Sac3144+o(1))q.(6)E_w\le \left(\frac{S_a c^3}{144}+o(1)\right)\sqrt q. \tag{6}

The repeated-input error becomes O(1)O(1) after averaging. Here and below all errors tend to zero with qq for fixed a,ca,c.

Deletion, reflection, and the unique exception

Delete from DD every output d4d_4 participating in (4), and call the remaining set D′D'. At most EwE_w elements are deleted. Put

t=v−1−w2,B=t+D′.t=\frac{v-1-w}{2},\qquad B=t+D'.

For large qq, t>0t>0. The set BB is ordinary Sidon because D′D' is, and (4) is precisely the condition that an element of BB be a sum of three elements of BB. Hence

B∩(B+B+B)=∅.B\cap(B+B+B)=\varnothing.

The positive-half characterization now shows that ±B\pm B has unique unordered pair sums away from zero. Explicitly, positive sums are either sums of two elements of BB or positive differences of two elements of BB. Each type is unique by Sidonicity; a collision between the types would give the excluded three-to-one relation. Negative sums follow by reflection. This checks diagonal pairs as well.

A translation gives the integer set

A=(v−D′) ∪ (2v−1−w+D′)⊂[1,N],N=3v−2−w.(7)A=(v-D')\ \cup\ (2v-1-w+D') \subset[1,N],\qquad N=3v-2-w. \tag{7}

This avoids any parity restriction on ww. The two copies are disjoint for large qq, and

∣A∣≥2q+2−2Ew.|A|\ge2q+2-2E_w.

Their exceptional sum is 3v−1−w=N+13v-1-w=N+1; it is the only integer sum that can repeat.

Optimizing the gain

Uniformly for w=O(q3/2)w=O(q^{3/2}),

233v−2−w=2q−w3q+O(1).\frac2{\sqrt3}\sqrt{3v-2-w} =2q-\frac{w}{3q}+O(1).

Using w≥aWw\ge aW and (6) in (7) yields

∣A∣−23N≥(ac3−Sac372+o(1))q.(8)|A|-\frac2{\sqrt3}\sqrt N \ge \left(\frac{ac}{3}-\frac{S_a c^3}{72}+o(1)\right)\sqrt q. \tag{8}

For fixed aa, this expression is maximized by

c2=8aSa,c^2=\frac{8a}{S_a},

and its coefficient becomes

F(a)=42 a3/29Sa⟶229(a↑1).F(a)=\frac{4\sqrt2\,a^{3/2}}{9\sqrt{S_a}} \longrightarrow\frac{2\sqrt2}{9} \qquad(a\uparrow1).

Choose a sequence aj↑1a_j\uparrow1, take the corresponding cjc_j, and then choose increasing prime powers qjq_j large enough that the normalized error in (8) is at most 1/j1/j. Since Nj∼3qj2N_j\sim3q_j^2, this proves (1).

The diagonal choice proves an unbounded sequence statement, not the same second-order inequality for every sufficiently large NN. The established leading bounds remain those on the bounds page.