Wiki
Wiki

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

Updated


Statement

Let

λ(N)=max⁡{∑n∈A1n:A⊆{1,…,N}, no S⊆A has ∑n∈S1n=1}.\lambda(N)=\max\left\{\sum_{n\in A}\frac1n: A\subseteq\{1,\ldots,N\},\ \text{no }S\subseteq A\text{ has }\sum_{n\in S}\frac1n=1\right\}.

Then λ(N)≫(log⁡log⁡N)2\lambda(N)\gg(\log\log N)^2.

Source. Bloom, arXiv:2112.03726v2, Theorem 4, Appendix A, p. 19. Bloom attributes the construction to Pomerance, via personal communication reported in Croot's 2000 thesis. This is a distinct obstruction method, not an alternative proof of the positive-density theorem.

Rewritten proof

Fix a large absolute constant CC. Let AA consist of those n≤Nn\le N whose largest prime divisor pp satisfies plog⁡p>Cnp\log p>Cn. For every sufficiently large prime p≤N/log⁡Np\le N/\log N, all n=pmn=pm with 1≤m<log⁡p/C1\le m<\log p/C lie in AA: here m<pm<p, so pp is the largest prime factor, and pm<Npm<N. Different choices of largest prime produce disjoint sets. Hence

R(A)≥∑p≤N/log⁡N1p∑1≤m<log⁡p/C1m≫∑p≤N/log⁡Nlog⁡log⁡pp≫(log⁡log⁡N)2.R(A)\ge\sum_{p\le N/\log N}\frac1p \sum_{1\le m<\log p/C}\frac1m \gg\sum_{p\le N/\log N}\frac{\log\log p}{p} \gg(\log\log N)^2.

The finitely many small primes can be omitted. For the last estimate, partial summation of Mertens' ∑p≤t1/p=log⁡log⁡t+O(1)\sum_{p\le t}1/p=\log\log t+O(1) gives $\sum_{p\le X}(\log\log p)/p =\frac12(\log\log X)^2+O(\log\log X)$.

Suppose distinct n1,…,nk∈An_1,\ldots,n_k\in A have reciprocals summing to one. Choose the largest prime pp dividing any denominator, and label those divisible by pp as pm1<⋯<pmrpm_1<\cdots<pm_r. For each such term, pp is its largest prime divisor, so mj<log⁡p/C<pm_j<\log p/C<p. In particular none of the mjm_j is divisible by pp. We have

1p∑j=1r1mj=1−∑j=r+1k1nj,\frac1p\sum_{j=1}^r\frac1{m_j} =1-\sum_{j=r+1}^k\frac1{n_j},

whose right side has reduced denominator coprime to pp. If D=lcm⁡(m1,…,mr)D=\operatorname{lcm}(m_1,\ldots,m_r) and T=D∑j1/mjT=D\sum_j1/m_j, then TT is a positive integer and p∤Dp\nmid D. The denominator condition forces p∣Tp\mid T, whence p≤Tp\le T. Put m=mrm=m_r. Then

p≤T≤lcm⁡(1,…,m)∑j=1m1j≤eC0mp\le T\le\operatorname{lcm}(1,\ldots,m)\sum_{j=1}^m\frac1j \le e^{C_0m}

for an absolute C0C_0, using Chebyshev's estimate log⁡lcm⁡(1,…,m)=O(m)\log\operatorname{lcm}(1,\ldots,m)=O(m) and absorbing the harmonic factor. Thus log⁡p≤C0m\log p\le C_0m, contradicting log⁡p>Cm\log p>Cm if C>C0C>C_0. Therefore AA has no unit subsum, proving the bound.

Dependencies and method

External Mertens and Chebyshev estimates. The transferable mechanism is to isolate the largest prime among a proposed representation and bound the positive integer it must divide. The complementary upper bound in this source is Theorem 3. No claim is made here that these are the best bounds in all subsequent literature.

Bears on

  • Problem 47 (the construction shows that Erdős's speculated threshold (log⁡log⁡N)2(\log\log N)^2 would be best possible)
  • Problem 298
  • Problem 299