Wiki
Wiki

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

Updated


Source. Theorem 1.2 and the paragraph after it, p. 298, with Lemma 4.1 (p. 304) and the proof of Section 4 (pp. 304--305), of N. Alon and G. Freiman, On sums of subsets of a set of integers, Combinatorica 8 (4) (1988), 297--306, doi:10.1007/BF02189086; the edition read is named on the source card.

Setting

Write N={1,2,…,n}N=\{1,2,\ldots,n\} and, for A⊆NA\subseteq N, let A∗A^* be the set of sums of subsets of AA, A∗={∑b∈Bb:B⊆A}A^*=\{\sum_{b\in B}b: B\subseteq A\} (p. 297). For m≥1m\ge1, f(n,m)f(n,m) is the largest size of a set A⊆NA\subseteq N with m∉A∗m\notin A^*, and snd⁡(m)\operatorname{snd}(m) is the smallest integer that does not divide mm (p. 298). The multiples of snd⁡(m)\operatorname{snd}(m) in NN avoid mm as a subset sum, so f(n,m)≥⌊n/snd⁡(m)⌋f(n,m)\ge\lfloor n/\operatorname{snd}(m)\rfloor (p. 298).

Statement

Theorem 1.2 (p. 298). For every ε>0\varepsilon>0, every n>n(ε)n>n(\varepsilon) and every mm with

3n5/3+ε<m<n220log⁡2n,3n^{5/3+\varepsilon}<m<\frac{n^2}{20\log^2 n}, f(n,m)=⌊nsnd⁡(m)⌋+snd⁡(m)−2.f(n,m)=\Bigl\lfloor\frac{n}{\operatorname{snd}(m)}\Bigr\rfloor+ \operatorname{snd}(m)-2 .

The print writes the upper end of the range as n2/20log⁡2nn^2/20\log^2 n; the proof (p. 305) uses it as n2/(20log⁡2n)n^2/(20\log^2 n).

Lemma 4.1 (p. 304), the lower bound. For every sufficiently large nn and every m≤n2m\le n^2, f(n,m)≥⌊n/snd⁡(m)⌋+snd⁡(m)−2f(n,m)\ge\lfloor n/\operatorname{snd}(m)\rfloor+\operatorname{snd}(m)-2.

Consequence (p. 298). For every nn there is an mm with f(n,m)=(1/2+o(1)) n/log⁡nf(n,m)=(1/2+o(1))\,n/\log n: the paper takes mm to be the least common multiple of all integers smaller than ss, with ss the largest integer for which this least common multiple is at most n2/20log⁡2nn^2/20\log^2 n; the prime number theorem gives s=(2+o(1))log⁡ns=(2+o(1))\log n. The paper says this verifies a conjecture of Erdős and Graham (its reference [3]), who observed that f(n,m)≥(1/2+o(1)) n/log⁡nf(n,m)\ge(1/2+o(1))\,n/\log n for all n,mn,m.

The paper also recalls (p. 298) the earlier bounds: from Alon's Subset sums (its reference [1]), $f(n,m)\le c(\varepsilon)\lfloor n/\operatorname{snd}(m)\rfloor$ for n1+ε<m<n2/log⁡2nn^{1+\varepsilon}<m<n^2/\log^2 n, and from Lipkin (its reference [4]), f(n,m)=(1+o(1)) n/snd⁡(m)f(n,m)=(1+o(1))\,n/\operatorname{snd}(m) for nlog⁡n<m<n3/2n\log n<m<n^{3/2}.

Proof pointer

Lemma 4.1 (p. 304): with s=snd⁡(m)s=\operatorname{snd}(m) and m≡i(mods)m\equiv i\pmod s, 1≤i≤s−11\le i\le s-1, the ⌊n/s⌋\lfloor n/s\rfloor multiples of ss in NN together with i−1i-1 numbers congruent to 11 and s−i−1s-i-1 numbers congruent to −1-1 modulo ss have no subset sum congruent to ii modulo ss. The upper bound (p. 305) applies Lemma 3.4 through Lemma 4.3 to find q≤sq\le s such that every multiple of qq in [2n5/3+ε,n2/(20log⁡2n)][2n^{5/3+\varepsilon},n^2/(20\log^2n)] is a sum of a subset of the elements of AA divisible by qq; when q<sq<s this already covers mm, and when q=sq=s, so that ss is a prime power pkp^k, Lemma 4.2 (for s=pks=p^k, the subset sums of any s−1s-1 non-zero elements of Zs\mathbb Z_s include every ipk−1ip^{k-1}, 1≤i≤p−11\le i\le p-1) uses s−1s-1 elements of AA not divisible by ss to correct the residue of mm. Lemma 3.4 rests on Proposition 1.3.

Read depth

Claims checked: Theorem 1.2, Lemma 4.1 and the consequence on p. 298 were read clause by clause on the page images of the print, and the proofs of Section 4 were followed. Nothing here is independently reviewed.

Bears on

  • Problem 771: the problem's f(n)f(n) is the least of the f(n,m)f(n,m) over m≥1m\ge1. The consequence on p. 298 gives, for every nn, an mm with f(n,m)=(1/2+o(1)) n/log⁡nf(n,m)=(1/2+o(1))\,n/\log n, an upper bound for f(n)f(n); with the lower bound of Erdős and Graham that the paper restates, this is the asymptotic f(n)=(1/2+o(1)) n/log⁡nf(n)=(1/2+o(1))\,n/\log n the problem asks about, and the paper says it verifies their conjecture.