betaFeedback

WorksMathematics

Recurring differences from overlapping arithmetic progressions

Abstract

We give sufficient conditions for the signed differences occurring infinitely often in a set of natural numbers to have bounded gaps. The hypotheses use pairs of finite arithmetic progressions with coprime step quotients, positive normalized overlap, and bounded mutual distance. Bounded perturbations of the progressions are allowed. A finite circular-gap criterion gives the explicit bound 2gH+4C2gH+4C, including equality at the overlap threshold. Exact centered pairs yield a finite-intersection theorem. Separated constructions realize these conditions at zero upper Banach density and give examples with no recurrent three-point pattern, no infinite arithmetic progression of recurrent differences, and infinitely many pairwise almost disjoint translates. Further examples describe the rational limiting threshold and the order of the quantitative gap bound.

Introduction and main results

For A⊆N={1,2,…}A \subseteq\mathbb{N} = \{1,2,\ldots\}, define its signed recurrent difference set by

D(A)={d∈Z:infinitely many (a,b)∈A2 satisfy a−b=d}.(1)D(A) = \{d \in\mathbb{Z} : \text{infinitely many } (a,b) \in A^2 \text{ satisfy } a-b=d\}. \tag*{(1)}

A set S⊆ZS \subseteq\mathbb{Z} is syndetic if some positive integer MM satisfies

S∩[z,z+M)≠∅(z∈Z).(2)S \cap[z,z+M) \ne\varnothing\qquad(z \in\mathbb{Z}). \tag*{(2)}

We call such an MM a gap bound. Erdős and Graham asked for conditions ensuring this property for D(A)D(A), and also considered finite intersections of recurrent difference sets [1 p. 50]. Positive upper Banach density is a classical sufficient condition, including for finite intersections [5, 4]. Sparse sufficient conditions arise in the work of Leth and Goldbring–Leth [3, 2]. The conditions below concern the steps, lengths, overlap, and relative positions of finite progression pairs.

Fix g∈Ng \in\mathbb{N}. For each n≥1n \ge1, let

Pn=an+gpn{0,1,…,Ln},Qn=bn+gqn{0,1,…,Rn},(3)\begin{aligned} P_n &= a_n + gp_n\{0,1,\ldots,L_n\},\\ Q_n &= b_n + gq_n\{0,1,\ldots,R_n\}, \tag*{(3)} \end{aligned}

where an,bn∈Za_n,b_n \in\mathbb{Z}, Ln,Rn∈Z≥0L_n,R_n \in\mathbb{Z}_{\ge0}, and

pn,qn≥2,gcd⁡(pn,qn)=1,pn,qn⟶∞,min⁡(an,bn)⟶∞.(4)p_n,q_n \ge2,\qquad\gcd(p_n,q_n)=1,\qquad p_n,q_n \longrightarrow\infty,\qquad\min(a_n,b_n) \longrightarrow\infty. \tag*{(4)}

Write InI_n and JnJ_n for the real convex hulls of PnP_n and QnQ_n, and put Mn=gpnqnM_n = gp_nq_n. We assume that, for a fixed integer C≥0C \ge0,

dist⁡(t,A)≤C(t∈Pn∪Qn, n≥1).(5)\operatorname{dist}(t,A) \le C \qquad(t \in P_n \cup Q_n,\ n \ge1). \tag*{(5)}

For θ∈T=R/Z\theta\in\mathbb{T} = \mathbb{R}/\mathbb{Z} and H≥1H \ge1, let ΔH(θ)\Delta_H(\theta) be the largest circular gap between the distinct points of

{0,θ,…,(H−1)θ}⊆T.(6)\{0,\theta,\ldots,(H-1)\theta\} \subseteq\mathbb{T}. \tag*{(6)}

The gap of a singleton is 1.

Theorem 1.1. Assume (3)–(5). Each of the following conditions makes D(A)D(A) syndetic.

(1) There are δ>0\delta> 0 and B<∞B < \infty such that, for infinitely many nn,

length⁡(In∩Jn)≥δMn,min⁡u∈Pn, v∈Qn∣u−v∣≤B.(7)\operatorname{length}(I_n \cap J_n) \ge\delta M_n,\qquad\min_{u \in P_n,\,v \in Q_n} |u-v| \le B. \tag*{(7)}

(2) Fix H≥1H \ge1. Let rn∈{0,…,g−1}r_n \in\{0,\ldots,g-1\} be congruent to an−bna_n-b_n modulo gg, and let vn∈{0,…,qn−1}v_n \in\{0,\ldots,q_n-1\} satisfy pnvn≡1(modqn)p_nv_n \equiv1 \pmod{q_n}. Put αn=vn/qn\alpha_n = v_n/q_n. For infinitely many nn, assume

length⁡(In∩(Jn+rn))Mn≥ΔH(αn).(8)\frac{\operatorname{length}(I_n \cap(J_n+r_n))}{M_n} \ge\Delta_H(\alpha_n). \tag*{(8)}

Then 2gH+4C2gH+4C is a gap bound for D(A)\mathcal{D}(A).

The second condition is non-strict. In particular, its conclusion includes models whose normalized overlap equals the largest circular gap. The first condition also permits contacts at endpoints of the overlapping hulls.

Theorem 1.2. Let A1,…,Am⊆NA_1,\ldots,A_m \subseteq\mathbb{N}. For each ii, suppose AiA_i contains the exact progression pairs

ci,n+gipi,n{−Li,n,…,Li,n},ci,n+giqi,n{−Ri,n,…,Ri,n}.(9)c_{i,n}+g_ip_{i,n}\{-L_{i,n},\ldots,L_{i,n}\}, \qquad c_{i,n}+g_iq_{i,n}\{-R_{i,n},\ldots,R_{i,n}\}. \tag*{(9)}

Assume their minimum points tend to infinity, gi∈Ng_i \in\mathbb{N} is fixed, and the coprime integers pi,n,qi,np_{i,n},q_{i,n} tend to infinity. Suppose that, for some δi>0\delta_i>0,

Li,nqi,n≥δi,Ri,npi,n≥δi.(10)\frac{L_{i,n}}{q_{i,n}} \ge\delta_i,\qquad\frac{R_{i,n}}{p_{i,n}} \ge\delta_i. \tag*{(10)}

Then ⋂i=1mD(Ai)\bigcap_{i=1}^{m}\mathcal{D}(A_i) is syndetic. Its positive part has positive lower asymptotic density and divergent reciprocal sum.

Recurrence, perturbations, and rotations

For a fixed dd, infinitely many pairs in (1) are equivalent to pairs with both entries arbitrarily large. Indeed, each first entry determines at most one second entry. In particular, D(A)\mathcal{D}(A) is symmetric, and 0∈D(A)0 \in\mathcal{D}(A) when AA is infinite.

For the models (3), define

S={s∈Z:s∈(Pn∪Qn)−(Pn∪Qn) for infinitely many n}.(11)S=\{s\in\mathbb{Z}:s\in(P_n\cup Q_n)-(P_n\cup Q_n)\text{ for infinitely many }n\}. \tag*{(11)}

This set is symmetric as well.

Lemma 2.1 (Transfer under bounded perturbations). For every s∈Ss \in S there is d∈D(A)d \in\mathcal{D}(A) with ∣d−s∣≤2C|d-s| \le2C. If NN is a gap bound for SS, then N+4CN+4C is a gap bound for D(A)\mathcal{D}(A).

Proof. Choose infinitely many modeled pairs un−wn=su_n-w_n=s. By (5), choose an′,bn′∈Aa'_n,b'_n\in A satisfying ∣an′−un∣,∣bn′−wn∣≤C|a'_n-u_n|,|b'_n-w_n|\le C. Hence

∣(an′−bn′)−s∣≤2C.|(a'_n-b'_n)-s|\le2C.

Both entries tend to infinity along the chosen indices. The difference takes values in a fixed finite set of integers, so one value dd occurs infinitely often and belongs to D(A)\mathcal{D}(A). Given z∈Zz\in\mathbb{Z}, choose s∈S∩[z+2C,z+2C+N)s\in S\cap[z+2C,z+2C+N). The corresponding dd lies in [z,z+N+4C)[z,z+N+4C).

We use the circle norm ∥x∥T=min⁡j∈Z∣x−j∣\|x\|_{\mathbb{T}}=\min_{j\in\mathbb{Z}}|x-j|.

Lemma 2.2 (Compact rotation lemma). Let γ=(γ1,…,γm)∈Tm\gamma=(\gamma_1,\ldots,\gamma_m)\in\mathbb{T}^m and let εi>0\varepsilon_i>0. Then

R={k∈Z:∥kγi∥T<εi for every i}R=\{k\in\mathbb{Z}:\|k\gamma_i\|_{\mathbb{T}}<\varepsilon_i\text{ for every }i\}

is syndetic. If an irrational rotation of T\mathbb{T} is given, its visits to any nonempty open arc are syndetic.

Proof. Let KK be the closure of {kγ:k∈Z}\{k\gamma: k \in\mathbb{Z}\} and let U={x∈Tm:∥xi∥T<ϵi for every i}U = \{x \in\mathbb{T}^{m} : \|x_i\|_{\mathbb{T}} < \epsilon_i \text{ for every } i\}. The set KK is a compact subgroup. Since the orbit is dense in KK and UU is symmetric, the sets fγ+Uf\gamma+ U, f∈Zf \in\mathbb{Z}, cover KK. Choose a finite subcover indexed by F⊆ZF \subseteq\mathbb{Z}. For every integer kk, some f∈Ff \in F satisfies (k−f)γ∈U(k-f)\gamma\in U. Thus Z=R+F\mathbb{Z} = R + F. If a=min⁡Fa = \min F and b=max⁡Fb = \max F, apply this identity to k=z+bk = z+b to obtain an element of RR in [z,z+b−a][z,z+b-a].

For the last assertion, the orbit closure of an irrational circle rotation is the whole circle. Translates of a nonempty open arc by the orbit therefore cover T\mathbb{T}. The same finite-subcover argument applies.

We recall briefly the density assertion used here. An infinite subgroup of T\mathbb{T} has nonzero elements arbitrarily close to zero, by compactness and taking differences. Multiples of an element with sufficiently small nonzero representative approximate every point of the circle. Its closure is consequently T\mathbb{T}. The subgroup generated by an irrational element is infinite. The pigeonhole principle gives positive integers qj→∞q_j \to\infty with qjα→0q_j\alpha\to0. Positive iterates therefore approximate each negative iterate, so the forward orbit is dense as well. Deleting finitely many iterates preserves this density.

The integer congruence and its interval

Suppress the subscript nn. Let r∈Zr \in\mathbb{Z} satisfy r≡a−b(modg)r \equiv a-b \pmod{g}, and put

x=b−a+rg∈Z,pv≡1(modq).x = \frac{b-a+r}{g} \in\mathbb{Z}, \qquad pv \equiv1 \pmod{q}.

The target r+gkr+gk occurs in P−QP-Q exactly when

pi−qj=x+k,0≤i≤L,0≤j≤R.(12)pi-qj=x+k,\qquad0\leq i\leq L,\qquad0\leq j\leq R. \tag*{(12)}

Define

ℓ(k)=max⁡(0,x+kp),u(k)=min⁡(L,x+k+qRp).(13)\ell(k)=\max\left(0,\frac{x+k}{p}\right),\qquad u(k)=\min\left(L,\frac{x+k+qR}{p}\right). \tag*{(13)}

Lemma 3.1. The equation (12) has an admissible solution if and only if there is an integer ii such that

i≡v(x+k)(modq),ℓ(k)≤i≤u(k).(14)i\equiv v(x+k)\pmod{q},\qquad\ell(k)\leq i\leq u(k). \tag*{(14)}

Moreover,

∣ℓ(k)−ℓ(k′)∣≤∣k−k′∣p,∣u(k)−u(k′)∣≤∣k−k′∣p,(15)\lvert\ell(k)-\ell(k')\rvert\leq\frac{\lvert k-k'\rvert}{p},\qquad\lvert u(k)-u(k')\rvert\leq\frac{\lvert k-k'\rvert}{p}, \tag*{(15)}

and

length⁡(I∩(J+r))=gpmax⁡{0,u(0)−ℓ(0)}.(16)\operatorname{length}(I\cap(J+r))=gp\max\{0,u(0)-\ell(0)\}. \tag*{(16)}

Proof. The congruence in (14) is equivalent to q∣pi−x−kq\mid pi-x-k. It makes j=(pi−x−k)/qj=(pi-x-k)/q an integer. The conditions 0≤j≤R0\leq j\leq R are precisely x+k≤pi≤x+k+qRx+k\leq pi\leq x+k+qR, giving (3.2) together with the bounds on $i$. Taking a minimum or maximum with a fixed real number preserves the Lipschitz bound of the other argument, which gives (3.4. Finally, in the coordinate $t=a+gpi$, the hulls $I$ and $J+r$ correspond respectively to $[0,L]$ and $[x/p,(x+qR)/p]$. This proves {xref}(3.5).

In particular, every integer ii in the common interior

[ℓ(0)+∣k∣p, u(0)−∣k∣p](17)\left[\ell(0)+\frac{\lvert k\rvert}{p},\,u(0)-\frac{\lvert k\rvert}{p}\right] \tag*{(17)}

that has the required residue class is admissible for the target r+gkr+gk.

Bounded contact

Proof of Theorem 1.1(1). There are only finitely many integer differences of absolute value at most BB. Pass to an infinite subsequence on which one fixed signed difference d∗d_* occurs in Pn−QnP_n-Q_n and the overlap bound holds. For this subsequence set

xn=bn−an+d∗g,x_n=\frac{b_n-a_n+d_*}{g},

and use (13) with r=d∗r=d_*. Choose an anchor in∈[ℓn(0),un(0)]i_n\in[\ell_n(0),u_n(0)] representing d∗d_*.

Translating one interval by d∗d_* changes its intersection length with a fixed interval by at most 2∣d∗∣2|d_*|. Consequently

un(0)−ℓn(0)≥δqn−2∣d∗∣gpn.u_n(0)-\ell_n(0)\ge\delta q_n-\frac{2|d_*|}{g p_n}.

After discarding finitely many indices, there is a fixed η>0\eta>0 such that

un(0)−ℓn(0)≥ηqn.(18)u_n(0)-\ell_n(0)\ge\eta q_n. \tag*{(18)}

Pass to a further subsequence on which αn=vn/qn\alpha_n=v_n/q_n converges in T\mathbb{T} to α\alpha.

Fix h∈Z∖{0}h\in\mathbb{Z}\setminus\{0\} with ∥hα∥T<η/16\lVert h\alpha\rVert_{\mathbb{T}}<\eta/16. Let tnt_n be a least absolute residue of vnhv_nh modulo qnq_n. Then

∣tn∣qn=∥hαn∥T<η8(19)\frac{|t_n|}{q_n}=\lVert h\alpha_n\rVert_{\mathbb{T}}<\frac{\eta}{8} \tag*{(19)}

eventually. For qn>∣h∣q_n>|h|, the residue tnt_n is nonzero. Otherwise the inverse congruence would give qn∣hq_n\mid h. Thus ∣tn∣≥1|t_n|\ge1.

Put En=∣h∣/pnE_n=|h|/p_n. For all sufficiently large nn, we have En<1≤∣tn∣E_n<1\le|t_n| and, by (18)–(19),

∣tn∣+En≤un(0)−ℓn(0)2.(20)|t_n|+E_n\le\frac{u_n(0)-\ell_n(0)}{2}. \tag*{(20)}

If ini_n lies in the lower half of its interval, choose a sign σn∈{1,−1}\sigma_n\in\{1,-1\} so that σntn=∣tn∣\sigma_nt_n=|t_n|. If it lies in the upper half, choose the sign so that σntn=−∣tn∣\sigma_nt_n=-|t_n|. The near endpoint has margin at least ∣tn∣≥En|t_n|\ge E_n, and (20) gives the far-end margin. Therefore

in+σntn∈[ℓn(0)+En,un(0)−En].i_n+\sigma_nt_n\in[\ell_n(0)+E_n,u_n(0)-E_n].

Its residue is vn(xn+σnh)v_n(x_n+\sigma_nh) modulo qnq_n. Lemma 3.1 proves that d∗+σngh∈Pn−Qnd_*+\sigma_ngh\in P_n-Q_n.

For this fixed hh, one sign occurs infinitely often. Symmetry of SS then gives an sh∈Ss_h\in S satisfying

∣sh−gh∣≤∣d∗∣.(21)|s_h-gh|\le|d_*|. \tag*{(21)}

For h=0h=0, the anchor itself gives the same conclusion. By Lemma 2.2, the set of integers hh with ∥hα∥T<η/16\lVert h\alpha\rVert_{\mathbb{T}}<\eta/16 is syndetic. Its dilation by gg is syndetic, and (21) implies that SS is syndetic. Lemma 2.1 completes the proof. □

The finite circular-gap criterion

The following elementary interpretation will be useful. If pv≡1(modq)pv\equiv1\pmod q, then

G=qΔH(v/q)(22)G=q\Delta_H(v/q) \tag*{(22)}

is a positive integer. Every block of GG consecutive integers meets

{vj+qz:0≤j<H, z∈Z}.\{vj+qz:0\le j<H,\ z\in\mathbb{Z}\}.

Indeed, the largest gap between consecutive elements of this periodic integer set is GG. Translating the phase set preserves this property.

Lemma 5.1 (Two-window lemma). Use the notation of Section 3 and let GG be given by (22). Suppose u(0)−ℓ(0)≥Gu(0)-\ell(0)\ge G. For an integer k0k_0, assume

p>2(∣k0∣+2H),q>2H.(23)p>2(|k_0|+2H),\qquad q>2H. \tag*{(23)}

Then some k∈{k0,…,k0+2H−1}k\in\{k_0,\ldots,k_0+2H-1\} satisfies r+gk∈P−Qr+gk\in P-Q.

Proof. Put E=(∣k0∣+2H)/p<1/2E=(|k_0|+2H)/p<1/2 and m=⌊ℓ(0)+E⌋+1m=\lfloor\ell(0)+E\rfloor+1. Every integer ii with

m≤i≤m+G−2(24)m\le i\le m+G-2 \tag*{(24)}

lies in [ℓ(0)+E,u(0)−E][\ell(0)+E,u(0)-E]. The lower bound follows from the definition of mm. For the upper bound, use

m≤ℓ(0)+E+1,ℓ(0)+G≤u(0),2E<1.m\le\ell(0)+E+1,\qquad\ell(0)+G\le u(0),\qquad2E<1.

The resulting common interval is contained in [ℓ(k),u(k)][\ell(k),u(k)] for every kk in the asserted 2H2H-window, by (15).

Consider separately the two disjoint windows

W0={k0,…,k0+H−1},W1={k0+H,…,k0+2H−1}.W_0=\{k_0,\ldots,k_0+H-1\},\qquad W_1=\{k_0+H,\ldots,k_0+2H-1\}.

For a window beginning at ss, apply the periodic-gap property at m−v(x+s)m-v(x+s) and then translate by v(x+s)v(x+s). Thus, for each ν∈{0,1}\nu\in\{0,1\}, there are hν∈Wνh_\nu\in W_\nu and an integer iν∈[m,m+G−1]i_\nu\in[m,m+G-1] with

iν≡v(x+hν)(modq).i_\nu\equiv v(x+h_\nu)\pmod q.

If either iνi_\nu satisfies (24), Lemma 3.1 gives the required target. If neither does, both equal m+G−1m+G-1. Multiplying their congruences by pp then gives h1−h0≡0(modq)h_1-h_0\equiv0\pmod q. But the two chosen indices satisfy 0<h1−h0<2H<q0<h_1-h_0<2H<q, a contradiction. □

Proof of Theorem 1.1[2]. Restrict to the infinite set of indices satisfying (8), and then to a subsequence with rn=rr_n=r fixed. For each such model use xn=(bn−an+r)/gx_n=(b_n-a_n+r)/g. By (16), the finite test is exactly

un(0)−ℓn(0)≥qnΔH(vn/qn).u_n(0)-\ell_n(0)\ge q_n\Delta_H(v_n/q_n).

For each fixed integer k0k_0, condition (23) holds for all sufficiently large nn. Lemma 5.1 produces a difference r+gkr+gk with kk in the fixed 2H2H-window. One of these finitely many differences occurs for infinitely many indices, and hence belongs to SS.

Given z∈Zz\in\mathbb{Z}, take k0=⌈(z−r)/g⌉k_0=\lceil(z-r)/g\rceil. Then z≤r+gk0<z+gz\le r+gk_0<z+g, and every r+gkr+gk from this 2H2H-window lies in [z,z+2gH)[z,z+2gH). Thus 2gH2gH is a gap bound for SS. Lemma 2.1 gives 2gH+4C2gH+4C for D(A)D(A). □

Centered finite families

Proof of Theorem 1.2. First consider one set and suppress ii. Pass to a subsequence on which the normalized inverse αn=vn/qn\alpha_n=v_n/q_n converges to α∈T\alpha\in\mathbb{T}. Fix k∈Zk\in\mathbb{Z} with ∥kα∥T<δ/2\lVert k\alpha\rVert_{\mathbb{T}}<\delta/2. Choose a least absolute residue unu_n of vnkv_nk modulo qnq_n. For all sufficiently large nn,

∣un∣qn=∥kαn∥T<3δ4,∣k∣pnqn<δ4.\frac{|u_n|}{q_n}=\lVert k\alpha_n\rVert_{\mathbb{T}}<\frac{3\delta}{4},\qquad\frac{|k|}{p_nq_n}<\frac{\delta}{4}.

The number wn=(pnun−k)/qnw_n=(p_nu_n-k)/q_n is an integer, and

∣wn∣pn≤∣un∣qn+∣k∣pnqn<δ.\frac{|w_n|}{p_n}\le\frac{|u_n|}{q_n}+\frac{|k|}{p_nq_n}<\delta.

Thus ∣un∣≤Ln|u_n|\le L_n and ∣wn∣≤Rn|w_n|\le R_n. The two modeled points with these indices have difference gkgk, and their entries escape to infinity. We have proved

{gk:k∈Z, ∥kα∥T<δ/2}⊆D(A).(25)\{gk:k\in\mathbb{Z},\ \lVert k\alpha\rVert_{\mathbb{T}}<\delta/2\}\subseteq D(A). \tag*{(25)}

Apply (25) separately to each AiA_i, obtaining αi∈T\alpha_i \in\mathbb{T}. Let G=lcm⁡(g1,…,gm)G = \operatorname{lcm}(g_1,\ldots,g_m) and set γi=(G/gi)αi\gamma_i = (G/g_i)\alpha_i. Lemma 2.2 shows that

R={k∈Z:∥kγi∥T<δi/2 for every i}\mathcal{R} = \{k \in\mathbb{Z} : \lVert k\gamma_i\rVert_{\mathbb{T}} < \delta_i/2 \text{ for every } i\}

is syndetic. For k∈Rk \in\mathcal{R}, formula (25) gives Gk∈D(Ai)Gk \in D(A_i) for every ii. The dilation GRG\mathcal{R} is syndetic, proving the intersection assertion.

If MM is a gap bound for the intersection, each interval [jM,(j+1)M)[jM,(j+1)M), j≥1j \ge1, contains a positive element. These disjoint intervals give lower asymptotic density at least 1/M1/M. Choosing one element from each also gives a reciprocal-sum lower bound

∑j≥11(j+1)M=∞.\sum_{j\ge1}\frac{1}{(j+1)M}=\infty.

□\square

Separated models and zero-density examples

Separated-pair realizations are discussed in [6], and recurring-template conditions in [7].

Proposition 7.1. Suppose A=⋃n(Pn∪Qn)⊆NA = \bigcup_n(P_n \cup Q_n) \subseteq\mathbb{N} is an exact union, the block hulls are ordered with successive gaps tending to infinity, and (4) holds. Suppose also that length⁡(In∩Jn)≥δMn\operatorname{length}(I_n \cap J_n) \ge\delta M_n for every sufficiently large nn, for some δ>0\delta> 0. Then

min⁡u∈Pn,v∈Qn∣u−v∣⟶∞⟹D(A)={0}.\min_{u\in P_n,v\in Q_n}|u-v| \longrightarrow\infty\quad\Longrightarrow\quad D(A)=\{0\}.

If these minimum distances do not tend to infinity, then D(A)D(A) is syndetic.

Proof. Fix a nonzero integer dd. Sufficiently late distinct blocks are separated by more than ∣d∣|d|, and sufficiently late row steps exceed ∣d∣|d|. Thus a pair with difference dd in the late blocks must use different rows of the same block. If their minimum mutual distance tends to infinity, these pairs are eventually impossible as well. The finitely many earlier points yield only finitely many pairs with difference dd. Since AA is infinite, D(A)={0}D(A)=\{0\}.

Failure of convergence of the minimum distances to infinity gives a finite BB and infinitely many bounded contacts. Theorem 1.1(1), with C=0C=0, applies. □\square

For the explicit examples, choose xn∈Z≥0x_n \in\mathbb{Z}_{\ge0} and place the rows by

a1=1,bn=an+gxn,an+1=an+gmax⁡(pnLn,xn+qnRn)+n+1.(26)\begin{aligned} a_1 &= 1,\qquad b_n = a_n + g x_n,\\ a_{n+1} &= a_n + g\max(p_nL_n,x_n+q_nR_n)+n+1. \tag*{(26)} \end{aligned}

The gap between consecutive block hulls is exactly n+1n+1. We use the upper Banach density

d∗(A)=lim sup⁡N→∞sup⁡t∈Z∣A∩[t,t+N)∣N,d^*(A)=\limsup_{N\to\infty}\sup_{t\in\mathbb{Z}}\frac{|A\cap[t,t+N)|}{N},

where NN ranges over the positive integers.

Lemma 7.2. Any union placed by (26), with both row steps tending to infinity, has zero upper Banach density. Every fixed three-element set has only finitely many translates contained in the union.

Proof. Fix a positive integer TT. Beyond finitely many blocks, every hull gap and both row steps exceed TT. An interval of diameter at most TT meets at most one of these blocks and contains at most one point from each of its rows. It therefore contains at most two late points. Let FF be the finite set of earlier points. Any interval of NN consecutive integers contains at most

2⌈NT⌉+∣F∣2\left\lceil\frac{N}{T}\right\rceil+|F|

points of AA. Dividing by NN, taking the supremum over interval positions and then the limit superior as N→∞N \to\infty gives d∗(A)≤2/Td^*(A) \le2/T. Since TT is arbitrary, d∗(A)=0d^*(A) = 0. Taking TT larger than the diameter of a fixed three-element set excludes every translate sufficiently far out. Only finitely many earlier translates remain. □

Proposition 7.3 (An exact irrational example). For every irrational α∈(0,1)\alpha\in(0,1) there is an exact separated union satisfying both conditions of Theorem 1.1, with g=1g=1 and C=0C=0, such that

D(A)={0}∪{d∈Z:∥αd∥T>1/4}.(27)D(A)=\{0\}\cup\{d\in\mathbb{Z}:\lVert\alpha d\rVert_{\mathbb{T}}>1/4\}. \tag*{(27)}

The set AA has zero upper Banach density. The set D(A)D(A) contains no infinite arithmetic progression of nonzero step, and AA has infinitely many pairwise almost disjoint translates.

Proof. Take an increasing sequence of primes qnq_n large enough that vn=⌊αqn⌋v_n=\lfloor\alpha q_n\rfloor lies in {1,…,qn−1}\{1,\ldots,q_n-1\}. Choose wn∈{1,…,qn−1}w_n\in\{1,\ldots,q_n-1\} with vnwn≡1(modqn)v_nw_n\equiv1\pmod{q_n}, and put

pn=qn+wn,xn=pn⌊qn/2⌋(modqn),Ln=⌊qn/4⌋,Rn=⌊pn/4⌋.(28)\begin{aligned} p_n&=q_n+w_n, & x_n&=p_n\lfloor q_n/2\rfloor\pmod{q_n},\\ L_n&=\lfloor q_n/4\rfloor, & R_n&=\lfloor p_n/4\rfloor. \tag*{(28)} \end{aligned}

using the least nonnegative residue for xnx_n. Then gcd⁡(pn,qn)=1\gcd(p_n,q_n)=1, pnvn≡1(modqn)p_nv_n\equiv1\pmod{q_n}, and vn/qn→αv_n/q_n\to\alpha. Place the blocks by (26).

For any fixed dd, the normalized admissible interval [ℓn(d)/qn,un(d)/qn][\ell_n(d)/q_n,u_n(d)/q_n] tends to [0,1/4][0,1/4], and the fractional residue of an admissible i/qni/q_n tends to

{12+αd}.\left\{\frac{1}{2}+\alpha d\right\}.

because vnxn≡⌊qn/2⌋(modqn)v_nx_n\equiv\lfloor q_n/2\rfloor\pmod{q_n}. For d≠0d\ne0, irrationality ensures that this limit is neither endpoint. The interval eventually has length less than one and lies arbitrarily close to [0,1/4][0,1/4]. Lemma 3.1 therefore gives

d∈Pn−Qn eventually⟺{12+αd}∈(0,1/4).d\in P_n-Q_n\text{ eventually}\quad\Longleftrightarrow\quad\left\{\frac{1}{2}+\alpha d\right\}\in(0,1/4).

The reversed orientation gives the reflected arc. The separated-block argument excludes all other recurrent nonzero differences, proving (27).

The normalized overlap tends to 1/41/4. Density of the irrational orbit provides a fixed dd whose phase lies in the interior arc, giving bounded contact. It also gives an HH with ΔH(α)<1/4\Delta_H(\alpha)<1/4. The largest circular gap of a finite point set is continuous under movement of its points: moving each point by at most ρ\rho changes that gap by at most 2ρ2\rho, as follows by trimming ρ\rho from each end of an empty arc and reversing the roles of the sets. Hence (8) holds eventually. Lemma 7.2 proves zero upper Banach density.

On an arithmetic progression d0+hkd_0+hk with h≠0h\ne0, the phases α(d0+hk)\alpha(d_0+hk) are dense in T\mathbb{T}. Infinitely many have circle norm less than 1/41/4 and give nonzero integers outside (27). Thus no such infinite progression is contained in D(A)D(A).

Finally choose infinitely many distinct positive integers ff with {αf}∈(0,1/8)\{\alpha f\}\in(0,1/8). For distinct choices f,f′f,f', we have ∥α(f−f′)∥T<1/8\lVert\alpha(f-f')\rVert_{\mathbb{T}}<1/8, so f−f′∉D(A)f-f'\notin D(A). An infinite intersection (A−f)∩(A−f′)(A-f)\cap(A-f') would give infinitely many pairs in AA with this difference. Every such intersection is therefore finite. □

Remark 7.4. Lemma 7.2 also separates these examples from the supra-SIM class. Goldbring–Leth [2 Corollary 4.7] show that every supra-SIM set contains B+CB+C with BB infinite and CC of any prescribed finite cardinality. Taking ∣C∣=3|C|=3 contradicts the lemma.

Boundary and order-of-magnitude examples

Proposition 8.1. For every rational a/b∈[0,1)a/b\in[0,1) in lowest terms, including a=0,b=1a=0,b=1, there is an exact separated union with normalized inverse tending to a/ba/b, normalized overlap tending to 1/b1/b, and recurrent difference set {0}\{0\}.

Proof. Take increasing integers qn≥(8b)2q_n \ge(8b)^2 satisfying aqn≡−1(modb)aq_n \equiv-1 \pmod{b}, and set

pn=qn+b,vn=(aqn+1)/b,mn=⌈qn⌉,τn=1/b−2mn/qn,xn=pn(qn−mn)(modqn),Ln=⌊τnqn⌋,Rn=⌊τnpn⌋.(29)\begin{aligned} p_n &= q_n+b, & v_n &= (aq_n+1)/b, & m_n &= \lceil\sqrt{q_n}\rceil,\\ \tau_n &= 1/b-2m_n/q_n, & x_n &= p_n(q_n-m_n) \pmod{q_n},\\ L_n &= \lfloor\tau_nq_n\rfloor, & R_n &= \lfloor\tau_np_n\rfloor. \tag*{(29)} \end{aligned}

The congruence is automatic when b=1b=1. The parameters have τn>0\tau_n>0, gcd⁡(pn,qn)=1\gcd(p_n,q_n)=1, and pnvn≡1(modqn)p_nv_n\equiv1\pmod{q_n}. Use the least nonnegative xnx_n and the placement (26), with g=1g=1.

Since 0≤xn<qn0\le x_n<q_n, the normalized overlap tends to 1/b1/b. For a fixed dd, the fractional class in the admissibility test is

adb−mnqn+dbqn(mod1).(30)\frac{ad}{b}-\frac{m_n}{q_n}+\frac{d}{bq_n}\pmod{1}. \tag*{(30)}

If ad/bad/b is an integer, its representative is eventually near 1−mn/qn1-m_n/q_n. Otherwise that representative is at least 1/b−mn/qn+Od(1/qn)1/b-m_n/q_n+O_d(1/q_n). In both cases it eventually exceeds τn\tau_n, because mn→∞m_n\to\infty. Every allowed i/qni/q_n lies in [0,Ln/qn]⊆[0,τn][0,L_n/q_n]\subseteq[0,\tau_n]. Thus every fixed cross-row difference is eventually absent. The same holds after reversal, while same-row and cross-block differences are excluded as in Proposition 7.1. Hence D(A)={0}D(A)=\{0\}.

The construction shows that equality at a limiting rational overlap does not supply the finite inequalities in (8). A zero limiting overlap likewise allows failure with an irrational limiting inverse. To see this, retain the steps and offsets in (28), and replace both normalized row lengths by a positive sequence tending to zero. Every fixed nonzero target has a limiting phase different from zero and is eventually absent.

Proposition 8.2. For each g,H≥1g,H\ge1 there is an exact separated union satisfying the finite test eventually and having D(A)=gHZD(A)=gH\mathbb{Z}. Thus the order gHgH of the gap bound in Theorem 1.1(2) is necessary.

Proof. For H≥2H\ge2, take increasing qn≥2q_n\ge2 with qn≡−1(modH)q_n\equiv-1\pmod{H}, let pn=qn+Hp_n=q_n+H and vn=(qn+1)/Hv_n=(q_n+1)/H, and set

λ=32H,β=34H,xn=pn⌊βqn⌋(modqn),\lambda=\frac{3}{2H},\qquad\beta=\frac{3}{4H},\qquad x_n=p_n\lfloor\beta q_n\rfloor\pmod{q_n},
Ln=⌊λqn⌋,Rn=⌊λpn⌋.L_n=\lfloor\lambda q_n\rfloor,\qquad R_n=\lfloor\lambda p_n\rfloor.

Place the rows by (26) with the chosen gg. All within-block differences are multiples of gg. For a fixed target gkgk, its limiting fractional phase is {β+k/H}\{\beta+k/H\} and the normalized admissible interval tends to [0,λ][0,\lambda]. Among the HH possible phases, exactly the one with k≡0(modH)k\equiv0\pmod{H} is in (0,λ)(0,\lambda), and none is an endpoint. The reflected orientation gives the same congruence class. Thus the separated-block argument gives D(A)=gHZD(A)=gH\mathbb{Z}.

The largest circular gap of the first HH multiples of vn/qnv_n/q_n tends to 1/H<λ1/H<\lambda, so the finite test holds eventually. For H=1H=1, take qn=n+1q_n=n+1, pn=qn+1p_n=q_n+1, xn=0x_n=0, and Ln=Rn=2qnL_n=R_n=2q_n. Each fixed multiple gkgk recurs by using equal indices in the two rows and, when necessary, reversing the orientation. Every within-block difference is a multiple of gg, so D(A)=gZD(A)=g\mathbb{Z}. The normalized overlap tends to 22, and the singleton gap is 11.

The smallest gap bound of gHZgH\mathbb{Z} is gHgH, which proves the assertion.

For a centered zero-density example in Theorem 1.2, take

qn=n+1,pn=qn+1,Ln=qn,Rn=pn,cn=an+gpnqn,q_n=n+1,\qquad p_n=q_n+1,\qquad L_n=q_n,\qquad R_n=p_n,\qquad c_n=a_n+gp_nq_n,

where a1=1a_1=1 and an+1=an+2gpnqn+n+1a_{n+1}=a_n+2gp_nq_n+n+1. The centered block hull is [an,an+2gpnqn][a_n,a_n+2gp_nq_n]. Equal centered indices give every fixed multiple of gg eventually, so D(A)=gZD(A)=g\mathbb{Z}. Lemma 7.2 gives zero upper Banach density.

The uncentered hypotheses alone do not ensure the finite-intersection conclusion. Use the irrational construction with common α\alpha, normalized row length 1/201/20, and phases β=4/5\beta=4/5 and β=1/2\beta=1/2 in place of 1/21/2 in (28). The respective recurrent nonzero phases {αd}\{\alpha d\} lie in

(1/5,1/4)∪(3/4,4/5)and(1/2,11/20)∪(9/20,1/2).(1/5,1/4)\cup(3/4,4/5)\quad\text{and}\quad(1/2,11/20)\cup(9/20,1/2).

These open sets are disjoint. The two recurrent difference sets are individually syndetic and have intersection {0}\{0\}.

Reciprocal divergence

Reciprocal divergence by itself does not force a nonzero recurrent difference. An elementary example, also recorded in the note of Plasma AI [8], is

am=m(1+⌊log⁡2m⌋),m≥1.(31)a_m=m(1+\lfloor\log_2 m\rfloor),\qquad m\geq1. \tag*{(31)}

For 2j−1≤m<2j2^{j-1}\leq m<2^j, we have am=jma_m=jm. Consecutive gaps within this block equal jj, and the transition to the next block has a still larger gap. Thus the gaps tend to infinity and D({am:m≥1})={0}D(\{a_m:m\geq1\})=\{0\}. On the other hand,

∑m=2j−12j−11am≥12j,\sum_{m=2^{j-1}}^{2^j-1}\frac{1}{a_m}\geq\frac{1}{2j},

so the reciprocal sum diverges.

The rational construction of Proposition 8.1 can also be made reciprocally divergent. After its nnth row block, whose last point is ene_n, insert the progression

cn+(n+1){0,…,Tn},cn=en+n+1,Tn=cn22(n+1).c_n+(n+1)\{0,\ldots,T_n\},\qquad c_n=e_n+n+1,\qquad T_n=c_n2^{2(n+1)}.

Start the next row block at cn+(n+1)Tn+n+1c_n+(n+1)T_n+n+1. Each inserted progression contributes more than 11 to the reciprocal sum, since

∑j=0Tn1cn+(n+1)j≥∫0Tn+1dtcn+(n+1)t=1n+1log⁡(1+(n+1)(Tn+1)cn)>2log⁡2>1.\begin{aligned} \sum_{j=0}^{T_n}\frac{1}{c_n+(n+1)j}\geq\int_0^{T_n+1}\frac{dt}{c_n+(n+1)t} \\ &=\frac{1}{n+1}\log\left(1+\frac{(n+1)(T_n+1)}{c_n}\right)>2\log2>1. \end{aligned}

Every inserted step and every gap between the enlarged pieces tends to infinity. No new fixed nonzero difference recurs, and the argument of Lemma 7.2 still gives zero upper Banach density. The resulting set therefore has divergent reciprocal sum and recurrent difference set {0}\{0\}.

Acknowledgments

The authors acknowledge AI assistance from OpenAI’s ChatGPT in developing and reviewing the arguments and preparing the mathematical exposition.

References

  1. [1]P. Erdős and R. L. Graham, Old and New Problems and Results in Combinatorial Number Theory, Monographies de L’Enseignement Mathématique, no. 28, Université de Genève, Geneva, 1980.
  2. [2]J. Goldbring and S. Leth, On supra-SIM sets of natural numbers, arXiv:1805.05933, 2018.arxiv.org/abs/1805.05933
  3. [3]S. Leth, Some nonstandard methods in combinatorial number theory, Studia Logica 47 (1988), no. 3, 265–278. doi:10.1007/BF00370556.DOI
  4. [4]I. Z. Ruzsa, On difference sets, Studia Scientiarum Mathematicarum Hungarica 13 (1978), 319–326.
  5. [5]C. L. Stewart and R. T. Tidemann, On infinite-difference sets, Canadian Journal of Mathematics 31 (1979), no. 5, 897–910. doi:10.4153/CJM-1979-085-6.DOI
  6. [6]Erdős Problem a Day, Erdős problem #332: live audit and a sharp upper-Banach-density boundary, working report, 28 July 2026. Sections 3–4. erdosproblemd​ay.com/report/332.
  7. [7]jeremy-math-332-worker, Fixed-template discussion for Erdős #332, Botnet, 29 September 2026. botnet.com, Erdős #332 discussion.
  8. [8]Plasma AI, Recurrent difference sets, research note, accessed 11 November 2026. erdosproblems.ai, Recurrent difference sets.

Paper details

vlnc:W_zhy825jx
Version
v1 · latest
Published
Fresh — published within the last 30 days
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

OpenAI’s ChatGPT

View statement
Looking for related papers…
Contents