Wiki
Wiki

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

Updated


Statement

For a kk-partition (kk-coloring) of N={1,2,…}\mathcal N=\{1,2,\ldots\} let CC be the set of integers with a monochromatic representation

n=a1+a2witha1≠a2n=a_1+a_2\quad\text{with}\quad a_1\ne a_2

(display (2), p. 47), let C2C^2 be the set of even integers in CC, and put CM=C∩[1,M]C_M=C\cap[1,M], CM2=C2∩[1,M]C^2_M=C^2\cap[1,M]. Roth conjectured (display (3), p. 48) that there is an absolute constant c>0c>0 such that ∣CM∣>cM|C_M|>cM for an arbitrary kk-partition; "(Note that if also a1=a2a_1=a_2 is allowed, then this is trivial.)" The paper proves the conjecture "in a sharper and more general form".

Theorem 1 (p. 48). (i) For each k≥2k\ge2 there is a threshold M0(k)M_0(k) such that every kk-partition of N\mathcal N satisfies

∣CM2∣>M2−3M1−2−k−1whenever M>M0(k).(4)|C^2_M|>\frac M2-3M^{1-2^{-k-1}}\qquad\text{whenever } M>M_0(k). \tag{4}

(ii) Every 22-partition satisfies

∣CM2∣>M2−(log⁡1+52)−1log⁡M.(5)|C^2_M|>\frac M2-\Bigl(\log\frac{1+\sqrt5}{2}\Bigr)^{-1}\log M. \tag{5}

(iii) Some 22-partition has 2n∉C22^n\notin C^2 for every n∈Nn\in\mathcal N (6).

The exponent in (4) is 1−2−k−11-2^{-k-1}; the proof (p. 50) applies Lemma 1 with d=k+1d=k+1. Statement (ii) says a 22-partition misses at most about log⁡M/log⁡φ\log M/\log\varphi of the even integers up to MM, φ\varphi the golden ratio, so the even monochromatic sums have full density; (iii) shows that infinitely many even integers can be missed.

Source. P. Erdős, A. Sárközy and V. T. Sós, On a conjecture of Roth and some related problems I, in Irregularities of Partitions (Springer, 1989), 47--59; Theorem 1 on printed p. 48 (PDF p. 2), Lemma 1 on p. 48, proof on pp. 49--51 (PDF pp. 3--5). The copy read is a scan whose text layer garbles formulas; the statements were read on the page images.

Read depth. Claims checked: Theorem 1 (i)--(iii), Lemma 1 and the definitions (pp. 47--48) were read clause by clause on the page images. The deduction of (i) from Lemma 1 (p. 50) and the proofs of (ii) and (iii) (p. 51) were read for structure; the proof of Lemma 1 (pp. 49--50) was not checked.

Proof sketch

Lemma 1 (p. 48), a density version of Hilbert's cube lemma: for d∈Nd\in\mathcal N and M>M0(d)M>M_0(d), every set B⊆[1,M]B\subseteq[1,M] with ∣B∣>3M1−2−d|B|>3M^{1-2^{-d}} contains a dd-dimensional cube: a positive integer uu and pairwise distinct positive integers v1,…,vdv_1,\ldots,v_d for which each of the 2d2^d sums u+∑i=1dεiviu+\sum_{i=1}^d\varepsilon_iv_i (εi∈{0,1}\varepsilon_i\in\{0,1\}) lies in BB.

(i) (p. 50). Suppose more than 3M1−2−k−13M^{1-2^{-k-1}} even integers not exceeding MM have no monochromatic representation, and let BB be their set. Lemma 1 with d=k+1d=k+1 gives u,v1,…,vk+1u,v_1,\ldots,v_{k+1} with all the sums in BB; in particular u∈Bu\in B is even, u=2zu=2z. The k+1k+1 distinct integers z+v1,…,z+vk+1z+v_1,\ldots,z+v_{k+1} fall into kk classes, so two of them, z+viz+v_i and z+vjz+v_j with i<ji<j, share a class, and their sum (z+vi)+(z+vj)=u+vi+vj∈B(z+v_i)+(z+v_j)=u+v_i+v_j\in B is a monochromatic representation with distinct summands, contradicting the definition of BB.

(ii) (p. 51). Let b1<b2<⋯<btb_1<b_2<\cdots<b_t be the even integers not exceeding 2M2M without a monochromatic representation. If bj+2<bj+bj+1b_{j+2}<b_j+b_{j+1} for some jj, the system x+y=bjx+y=b_j, x+z=bj+1x+z=b_{j+1}, y+z=bj+2y+z=b_{j+2} has positive integer solutions, two of x,y,zx,y,z share a class, and one of the bb's is a monochromatic sum; hence bj+2≥bj+bj+1b_{j+2}\ge b_j+b_{j+1} for every jj, which forces Fibonacci-type growth and proves (ii).

(iii) (p. 51). Define A1A_1 recursively: 1∈A11\in A_1; once A1∩[1,2k−1]A_1\cap[1,2^{k-1}] is defined, put 2k∈A12^k\in A_1 and, for 2k−1<n<2k2^{k-1}<n<2^k, n∈A1n\in A_1 iff 2k−n∉A12^k-n\notin A_1; let A2=N∖A1A_2=\mathcal N\setminus A_1. Then 2n∉C2^n\notin C for every nn.

Dependencies

Lemma 1 of the paper (proved there on pp. 49--50, not checked here).

Bears on

  • Problem 484: (i) gives ∣CM∣≥∣CM2∣>M/2−3M1−2−k−1≥cM|C_M|\ge|C^2_M|>M/2-3M^{1-2^{-k-1}}\ge cM for any fixed c<1/2c<1/2 once MM is large in terms of kk and cc, the absolute constant the problem asks for; (ii) and (iii) are the site's remarks for k=2k=2.