Wiki
Wiki

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

Updated


Notation and statement

Write R(A)=∑n∈A1/nR(A)=\sum_{n\in A}1/n, Aq={n∈A:q∣n, gcd⁡(q,n/q)=1}A_q=\{n\in A:q\mid n,\ \gcd(q,n/q)=1\}, QA={q=pa:Aq≠∅}\mathcal Q_A=\{q=p^a:A_q\ne\varnothing\} and R(A;q)=∑n∈Aqq/nR(A;q)=\sum_{n\in A_q}q/n. Thus AqA_q uses the exact prime power in nn, not every prime power divisor.

For sufficiently large NN and A⊆[1,N]∩NA\subseteq[1,N]\cap\mathbb N, there is B⊆AB\subseteq A such that

R(B)≥R(A)−(log⁡N)−1/200,R(B;q)≥2(log⁡N)−1/100(q∈QB).R(B)\ge R(A)-(\log N)^{-1/200},\qquad R(B;q)\ge2(\log N)^{-1/100}\quad(q\in\mathcal Q_B).

Source. Bloom, arXiv:2112.03726v2, Lemma 6, p. 17.

Rewritten proof

Start with A0=AA_0=A. If qi∈QAiq_i\in\mathcal Q_{A_i} has R(Ai;qi)<2(log⁡N)−1/100R(A_i;q_i)<2(\log N)^{-1/100}, delete the entire fiber: Ai+1=Ai∖(Ai)qiA_{i+1}=A_i\setminus(A_i)_{q_i}. Otherwise stop. A deletion removes at least one element, so the process stops at some BB satisfying the required fiber inequalities.

The loss at step ii is R(Ai;qi)/qiR(A_i;q_i)/q_i, less than 2/[qi(log⁡N)1/100]2/[q_i(\log N)^{1/100}]. No qiq_i can recur: after deletion no remaining integer has that exact prime power, and subsequent steps only remove integers. Every such qiq_i is at most NN. Consequently

R(A)−R(B)≤2(log⁡N)−1/100∑q≤N1q≪(log⁡N)−1/100log⁡log⁡N≤(log⁡N)−1/200R(A)-R(B) \le2(\log N)^{-1/100}\sum_{q\le N}\frac1q \ll(\log N)^{-1/100}\log\log N \le(\log N)^{-1/200}

for all sufficiently large NN.

Dependencies

The external Mertens prime-power estimate ∑q≤x1/q=log⁡log⁡x+c+O(1/log⁡x)\sum_{q\le x}1/q=\log\log x+c+O(1/\log x), equation (1), p. 3. The proof refines the pruning approach of Croot's Proposition 2.

Bears on