Wiki
Wiki

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

Updated


Statement

The paper gives its theorem no number. The abstract (printed p. 264) states it: "Let nn be a positive integer and let A={a1,…,as}A=\{a_1,\ldots,a_s\}, B={b1,…,bt}B=\{b_1,\ldots,b_t\} be two sets of positive integers such that the product set consists of stst distinct numbers. Then, for a certain positive constant cc, st≤c n2/log⁡nst\le c\,n^2/\log n, establishing a conjecture made by P. Erdös." The body (p. 264) states the problem with the hypothesis the abstract omits, "Let 1≤a1<⋯<ak≤n1\le a_1<\cdots<a_k\le n and b1<⋯<bl≤nb_1<\cdots<b_l\le n be two sequences of integers so that the products aibja_ib_j are all distinct", and the conjecture as display (2),

kl<C(n2/log⁡n),kl<C(n^2/\log n),

"In this paper we wish to prove (2)." The closing line of the proof (p. 269) is the theorem in the paper's final form: "Thus, our result is that for n≥2n\ge2 we have ∣A∣⋅∣B∣<C(n2/log⁡n)|A|\cdot|B|<C(n^2/\log n), where C=4max⁡{c22c3c5, 8c1−1c22c5c6log⁡22, 8c1−1c22c3c4−1c5c6log⁡22}C=4\max\{c_2^2c_3c_5,\,8c_1^{-1}c_2^2c_5c_6\log^22,\,8c_1^{-1}c_2^2c_3c_4^{-1}c_5c_6\log^22\}", reduced in three printed steps to 16c1−1c22c3c4−1c5c616c_1^{-1}c_2^2c_3c_4^{-1}c_5c_6. Here c1=(1/2)(∑p1/plog⁡p)−1c_1=(1/2)(\sum_p1/p\log p)^{-1} (Lemma 1, p. 265), c2c_2 is the constant of the Brun-sieve Lemma 2 (p. 266), c3c_3 and c4c_4 are the Mertens constants with c4/log⁡n≤∏p≤n(1−p−1)≤c3/log⁡nc_4/\log n\le\prod_{p\le n}(1-p^{-1})\le c_3/\log n for n≥2n\ge2 (p. 267), c5=∏k=1∞(1−2−k)−2k/2c_5=\prod_{k=1}^\infty(1-2^{-k})^{-2^{k/2}} (p. 267) and c6=max⁡k≥1(k+1)22−k/4c_6=\max_{k\ge1}(k+1)^22^{-k/4} (p. 268). The paper's footnote fixes "integers" to mean the natural numbers. In the notation of the problem page: for A,B⊆{1,…,N}A,B\subseteq\{1,\ldots,N\} with the products abab, a∈Aa\in A, b∈Bb\in B, all distinct, ∣A∣∣B∣<CN2/log⁡N|A||B|<CN^2/\log N for all N≥2N\ge2.

Source. E. Szemerédi, On a Problem of P. Erdös, Journal of Number Theory 8 (1976), no. 3, 264--270; the abstract and display (2) on printed p. 264 (PDF p. 1 of the publisher scan), the closing statement with the constant on p. 269 (PDF p. 6), the proof on pp. 265--269 (PDF pp. 2--6), read on the page images. The edition is identified in the source digest.

Read depth. Claims checked: the abstract, the statement of the problem with (1) and (2), Lemmas 1--3, the two cases and the closing statement with CC were read clause by clause on the page images. The proof (pp. 265--269) was read in full on the page images for its structure, and no step was checked; the reduction of CC to its final form was not checked. Nothing here is independently reviewed.

Proof pointer

Pages 265--269. For a set SS and a prime pp, SpS_p is the set of members of SS divisible by pp and p−1Spp^{-1}S_p the set of their quotients by pp. Lemma 1 (p. 265) passes to A∗⊂AA^*\subset A, B∗⊂BB^*\subset B with ∣A∗∣∣B∗∣>14∣A∣∣B∣|A^*||B^*|>\frac14|A||B| in which every prime pp that divides some member divides many: ∣Ap∗∣>c1∣A∗∣/(plog⁡p)|A^*_p|>c_1|A^*|/(p\log p), likewise for B∗B^*; the paper then assumes this of AA and BB. Lemma 2 (p. 266, Brun's method, cited to Halberstam and Roth): for n≥1n\ge1, the integers up to nn coprime to a set PP of primes ≤n\le n number at most c2n∏p∈P(1−p−1)c_2n\prod_{p\in P}(1-p^{-1}), with c2c_2 absolute. Lemma 3 (p. 266): if p−1Ap∩q−1Aq≠∅p^{-1}A_p\cap q^{-1}A_q\ne\varnothing for primes p≠qp\ne q then p−1Bp∩q−1Bq=∅p^{-1}B_p\cap q^{-1}B_q=\varnothing, since px,qx∈Apx,qx\in A and py,qy∈Bpy,qy\in B give the equal products px⋅qy=qx⋅pypx\cdot qy=qx\cdot py; this is the only use of the distinct-products hypothesis. With L(k)={p:2k≤p<2k+1,Ap≠∅,Bp≠∅}L(k)=\{p:2^k\le p<2^{k+1},A_p\ne\varnothing,B_p\ne\varnothing\}, Case I (p. 267), ∣L(k)∣≤2k/2|L(k)|\le2^{k/2} for every kk: Lemma 2 applied to AA with the primes not dividing any member of AA, and to BB, together with the Mertens bounds and ∏p∈⋃kL(k)(1−p−1)−1<c5\prod_{p\in\bigcup_kL(k)}(1-p^{-1})^{-1}<c_5, gives ∣A∣∣B∣<c22c3c5n2/log⁡n|A||B|<c_2^2c_3c_5n^2/\log n. Case II (pp. 267--269), ∣L(k)∣>2k/2|L(k)|>2^{k/2} for a largest such kk: the quotients in p−1App^{-1}A_p, p∈L(k)p\in L(k), are at most n2−kn2^{-k}, so Lemma 2 bounds ∣⋃p∈L(k)p−1Ap∣|\bigcup_{p\in L(k)}p^{-1}A_p| and the same for BB; if ∣A∣|A| or, by symmetry, ∣B∣|B| is below a threshold of order (k+1)n2−k/4(k+1)n2^{-k/4} times a sieve product, the bound Cn2/log⁡nC n^2/\log n follows directly (two subcases on p. 268 according to whether the sieve product over n2−k≥p>2k+1n2^{-k}\ge p>2^{k+1} is empty); otherwise Lemma 1 makes ∑p∈L(k)∣p−1Ap∣\sum_{p\in L(k)}|p^{-1}A_p| at least 2(k/4)+12^{(k/4)+1} times the size of the union, so some L∗⊂L(k)L^*\subset L(k) with ∣L∗∣≥2(k/4)+1|L^*|\ge2^{(k/4)+1} has a common element in all p−1App^{-1}A_p, p∈L∗p\in L^*, and then ∑p∈L∗∣p−1Bp∣≥4∣⋃p∈L∗p−1Bp∣\sum_{p\in L^*}|p^{-1}B_p|\ge4|\bigcup_{p\in L^*}p^{-1}B_p| forces two primes p1,p2∈L∗p_1,p_2\in L^* with p1−1Bp1∩p2−1Bp2≠∅p_1^{-1}B_{p_1}\cap p_2^{-1}B_{p_2}\ne\varnothing, against Lemma 3 (p. 269). The constant CC collects the three bounds and the factor 44 of Lemma 1.

Dependencies

Brun's sieve in the form of Lemma 2, cited to Halberstam and Roth, Sequences I (1966), not held; the Mertens product bounds c4/log⁡n≤∏p≤n(1−p−1)≤c3/log⁡nc_4/\log n\le\prod_{p\le n}(1-p^{-1})\le c_3/\log n, quoted as well known; the convergence of ∑p1/(plog⁡p)\sum_p1/(p\log p) and of ∏k(1−2−k)−2k/2\prod_k(1-2^{-k})^{-2^{k/2}}. External premises are taken at statement level; none was checked here. The second proof, Theorem 1 of the 1976 Erdős--Szemerédi paper, has the same outline (its "associated" primes, dyadic blocks, distinct quotient pairs and Brun--Mertens count answer to Lemma 1, the sets L(k)L(k), Lemma 3 and Lemma 2 here) and describes itself as "a simpler proof of (4), which nevertheless uses many of the ideas of the original proof" (p. 420).

Bears on

  • Problem 490: the theorem is the problem's inequality ∣A∣∣B∣≪N2/log⁡N|A||B|\ll N^2/\log N for A,B⊆{1,…,N}A,B\subseteq\{1,\ldots,N\} with all products abab distinct, in the paper the site names as its proof; the constant is explicit in terms of the Brun and Mertens constants but not evaluated. On the limit of max⁡∣A∣∣B∣log⁡N/N2\max|A||B|\log N/N^2 the paper only records the hope to investigate whether the bound holds "for avery [sic] c>1c>1 if n>n0(c)n>n_0(c)" (p. 265), and its construction (1) (p. 264, the primes in (n/log⁡n,n)(n/\log n,n) against the integers up to nn not divisible by any of them) gives kl>(1+o(1))n2/log⁡nkl>(1+o(1))n^2/\log n.