Wiki
Wiki

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

Updated


Statement

Here f(n)f(n) is the function of the Theorem (p. 307) and w=w(n)=(1/40)log⁡n(log⁡log⁡n)−1w=w(n)=(1/40)\sqrt{\log n(\log\log n)^{-1}} is the quantity of display (3.2) (p. 310), logarithms to base 2.

Remark (p. 313, unnumbered, closing the paper). The authors state that the lower-bound process of Section 3 can be iterated to give f(n)>n w2/2f(n)>\sqrt n\,w^2/2, and after kk iterations f(n)>n wk/k!f(n)>\sqrt n\,w^k/k!, which in their words "grows up as high as n1/2+ϵn^{1/2+\epsilon}". They do not carry this out, citing messy details. The remark ends: "It is conceivable that f(n)>n1−ϵf(n)>n^{1-\epsilon} for every ϵ\epsilon and n≥n0(ϵ)n\ge n_0(\epsilon)."

The final sentence is an expectation, not a proved statement, and the iterated bounds are announced without proof. Since ∑kwk/k!=ew\sum_kw^k/k!=e^w and ew=no(1)e^w=n^{o(1)} for the ww of (3.2), the announced bounds are all at most n no(1)\sqrt n\,n^{o(1)}, so the ϵ\epsilon in "as high as n1/2+ϵn^{1/2+\epsilon}" tends to 00 with nn (an observation made here). The last sentence, for every ϵ>0\epsilon>0, is equivalent to f(n)≥n1−o(1)f(n)\ge n^{1-o(1)}.

Source. S. L. G. Choi, J. Komlós and E. Szemerédi, On sum-free subsequences, Trans. Amer. Math. Soc. 212 (1975), 307--313, DOI 10.1090/S0002-9947-1975-0376594-1; the remark on printed p. 313, with (3.2) and footnote 2 on p. 310, read in the journal's printing. The paper and its edition are recorded on the source card.

Read depth. Claims checked: the remark was read clause by clause on the page image. It carries no proof.

Proof pointer

None: the paper gives no proof of the iterated bounds and none is claimed for the last sentence.

Dependencies

The lower-bound argument of the Theorem (Section 3, pp. 310--313), which the remark proposes to iterate.

Bears on

  • Problem 790: the last sentence is what the site's commentary renders as the authors' "conjecture that l(n)≥n1−o(1)l(n)\geq n^{1-o(1)}", the paper's f(n)f(n) being the problem's l(n)l(n). If true it would answer the problem's second displayed question, whether l(n)<n1−cl(n)<n^{1-c} for some c>0c>0, in the negative. The paper proves nothing toward it beyond the Theorem's lower bound.