Wiki
Wiki

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

Updated

Graham 1977 extremal density theorems linear forms

../

equation_12: For the forms x, 2x, 3x, Graham, Witsenhausen and Spencer show that the largest subsets of [1, N] with no triple n, 2n, 3n have a limiting density, equal to one third of the sum of 1/d_k over the indices k at which the extremal count on the first k 3-smooth numbers grows; they ask whether this density is irrational, the question of Problem 168.

theorem_1: Graham, Witsenhausen and Spencer's extremal theorem for augmented arithmetic progressions: every subset of the first N integers with more than N - [N/n] elements contains integers x, y with x + ky in it for all 0 <= k < n, so the critical density of that system is 1 - 1/n.

theorem_2: Graham, Witsenhausen and Spencer's formula for the critical density of a system of linear forms in one variable: the product of 1 - 1/q over the primes q dividing the coefficients, times the sum of 1/d_k over the indices where the extremal count on the first k such smooth numbers grows; stated with the remark that the Section 4 arguments prove it.


R. L. Graham, H. S. Witsenhausen and J. H. Spencer, On extremal density theorems for linear forms, in Number Theory and Algebra, Academic Press, New York, 1977, pp. 103--109. The offprint prints the authors in this order; the problem page lists them as Graham, Spencer and Witsenhausen.

The copy read for this card is an image-only scan of the offprint (seven pages, distilled in 2005, no text layer; physical PDF p. nn is printed p. 102+n102+n), headed "REPRINTED FROM NUMBER THEORY AND ALGEBRA © 1977 ACADEMIC PRESS, INC.". Its identity was confirmed on the page images from the title, the authors and the printed page numbers 103--109, and every statement below was read on the page images. Provenance: downloaded in September 2026; the download URL was not recorded; 226,338 bytes. Read status: claims checked. The offprint prints "© 1977 ACADEMIC PRESS, INC." in the head of p. 103, read on the page image, every other right reserved.

Contents

  • Sections 1--2 (p. 104): for a set L={Li(x1,…,xm)=∑jaijxj:1≤i≤n}\mathscr L=\{L_i(x_1,\dots,x_m)=\sum_ja_{ij}x_j:1\le i\le n\} of linear forms with integer coefficients, R⊆[1,N]R\subseteq[1,N] is L\mathscr L-free if for every choice of positive integers t1,…,tmt_1,\dots,t_m at least one value Li(t1,…,tm)L_i(t_1,\dots,t_m) is not in RR; otherwise L\mathscr L hits RR. SL(N)S_{\mathscr L}(N) is the largest size of an L\mathscr L-free subset of [1,N][1,N], and the critical density is δ(L)=lim inf⁡NSL(N)/N\delta(\mathscr L)=\liminf_NS_{\mathscr L}(N)/N. For Ln={x1+kx2:0≤k<n}\mathscr L_n=\{x_1+kx_2:0\le k<n\} (nn-term progressions), Szemerédi's theorem gives δ(Ln)=0\delta(\mathscr L_n)=0.
  • Section 3 (pp. 104--106), augmented progressions Ln∗=Ln∪{x2}\mathscr L_n^*=\mathscr L_n\cup\{x_2\}: Examples 1 and 2 give Ln∗\mathscr L_n^*-free sets {x>[N/n]}\{x>[N/n]\} and, for nn prime, {x≢0(modn)}\{x\not\equiv0\pmod n\}, of size N−[N/n]N-[N/n]. Theorem 1 (p. 105): if R⊆[1,N]R\subseteq[1,N] has ∣R∣>N−[N/n]|R|>N-[N/n] then Ln∗\mathscr L_n^* hits RR (proof pp. 105--106, a double count over the progressions Ti={i+kΔ}T_i=\{i+k\Delta\} with Δ\Delta the least element of RR). Hence SLn∗(N)=N−[N/n]S_{\mathscr L_n^*}(N)=N-[N/n] (8) and δ(Ln∗)=1−n−1\delta(\mathscr L_n^*)=1-n^{-1}.
  • Section 4 (pp. 106--108), the special case L={x,2x,3x}\mathscr L=\{x,2x,3x\}: with D={d1<d2<⋯ }D=\{d_1<d_2<\cdots\} the integers 2a3b2^a3^b, C(t)=[1,N]∩tDC(t)=[1,N]\cap tD for (t,6)=1(t,6)=1, and f(r)f(r) the size of the largest L\mathscr L-free subset of {d1,…,dr}\{d_1,\dots,d_r\}, a set RR is L\mathscr L-free if and only if each R∩C(t)R\cap C(t) is, so maximal L\mathscr L-free sets RNR_N satisfy lim⁡N∣RN∣/N=13∑r≥1f(r)(1/dr−1/dr+1)\lim_N|R_N|/N=\tfrac13\sum_{r\ge1}f(r)(1/d_r-1/d_{r+1}) (11); since f(r+1)−f(r)≤1f(r+1)-f(r)\le1, with K(L)={k:f(k)>f(k−1)}K(\mathscr L)=\{k:f(k)>f(k-1)\}, δ(L)=13∑k∈K(L)1/dk\delta(\mathscr L)=\tfrac13\sum_{k\in K(\mathscr L)}1/d_k (12). Table 1 lists f(k)f(k) for k≤36k\le36 and (13) lists K(L)={1,2,4,5,6,8,9,11,13,14,15,17,18,20,22,23,24,26,28,29,31,32,34,35,36,… }K(\mathscr L)=\{1,2,4,5,6,8,9,11,13,14,15,17,18,20,22,23,24,26,28,29, 31,32,34,35,36,\dots\}. The authors see no simple way to determine K(L)K(\mathscr L), suggest that f(k)=1+[2k/3]f(k)=1+[2k/3] when k≢0(mod3)k\not\equiv0\pmod3 and that perhaps there is always a maximal L\mathscr L-free set Rk={2ai3bi}⊆{d1,…,dk}R_k=\{2^{a_i}3^{b_i}\}\subseteq\{d_1,\dots,d_k\} with all ai−bia_i-b_i congruent modulo 33, and write (p. 108) "It would also be interesting to know if δ(L)\delta(\mathscr L) is irrational."
  • Section 5 (pp. 108--109), forms in one variable L={a1x,…,anx}\mathscr L=\{a_1x,\dots,a_nx\}: with q1,…,qrq_1,\dots,q_r the primes dividing the aia_i, d1<d2<⋯d_1<d_2<\cdots the integers composed of them, and f(k)f(k), K(L)K(\mathscr L) defined as before, Theorem 2 states δ(L)=∏j=1r(1−qj−1)∑k∈K(L)dk−1\delta(\mathscr L)=\prod_{j=1}^r(1-q_j^{-1})\sum_{k\in K(\mathscr L)} d_k^{-1} (14), which the paper says can be proved by essentially the arguments of Section 4 (p. 109); no separate proof.
  • Section 6 (p. 109): δ(L(1,p,…,pm−1))=(pm−p)/(pm−1)\delta(\mathscr L(1,p,\dots,p^{m-1}))=(p^m-p)/(p^m-1) for prime pp, δ(L(1,n))=n/(n+1)\delta(\mathscr L(1,n))=n/(n+1), δ(L(2,3))=3/4\delta(\mathscr L(2,3))=3/4, δ(L(1,2,8))=57/62\delta(\mathscr L(1,2,8))=57/62 (arguments omitted), and the closing sentence "It seems quite likely that almost all systems L\mathscr L have δ(L)\delta(\mathscr L) irrational although not even one such L\mathscr L is known at present!" The references are Harlambis's 1973 dissertation and Szemerédi's 1975 Acta Arith. paper.

Compiled scope

Every statement above was read on the page images of the seven pages. The proof of Theorem 1 and the derivation of (11) and (12) were read for their structure, not checked line by line; Theorem 2 and the values of Section 6 are asserted in the paper without proof. Nothing has been independently reviewed.

Bears on. #168: a set with no triple {n,2n,3n}\{n,2n,3n\} is exactly an {x,2x,3x}\{x,2x,3x\}-free set, so equations (11) and (12) (pp. 107--108) show that the problem's limit exists and equal it to δ({x,2x,3x})\delta(\{x,2x,3x\}), given as the series (12) over the set K(L)K(\mathscr L) determined by the extremal counts f(k)f(k) on the 33-smooth numbers; the paper tabulates f(k)f(k) for k≤36k\le36 and raises the irrationality question that the problem repeats. It does not determine K(L)K(\mathscr L) or evaluate the limit, and it leaves the irrationality question open. Theorem 2 (p. 109) contains this case and adds nothing to it.

Results. Labels and pages are those of the printed volume.

  • Theorem 1 (p. 105; proof pp. 105--106): a subset of [1,N][1,N] with more than N−[N/n]N-[N/n] elements is hit by Ln∗\mathscr L_n^*; with Example 1 this gives SLn∗(N)=N−[N/n]S_{\mathscr L_n^*}(N)=N-[N/n] and δ(Ln∗)=1−n−1\delta(\mathscr L_n^*)=1-n^{-1}.
  • Equations (11) and (12) (pp. 107--108): for {x,2x,3x}\{x,2x,3x\} the extremal density exists and equals 13∑k∈K(L)1/dk\tfrac13\sum_{k\in K(\mathscr L)}1/d_k; the page also records the list (13), the suggestions and the irrationality question of p. 108.
  • Theorem 2 (p. 109; no proof printed): for forms a1x,…,anxa_1x,\ldots,a_nx, δ(L)=∏j(1−qj−1)∑k∈K(L)dk−1\delta(\mathscr L)=\prod_j(1-q_j^{-1})\sum_{k\in K(\mathscr L)}d_k^{-1} over the integers built from the primes qjq_j dividing the aia_i; the page also records the values of Section 6.

No file of this source is held: no license on record permits its redistribution, and the card cites the edition it names above.