Wiki
Wiki

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

Updated


Source. Thomas F. Bloom, On a density conjecture about unit fractions, arXiv:2112.03726v2 (12 October 2023). Printed and PDF page numbers agree.

Use R(A)R(A), AqA_q, QAQ_A and R(A;q)R(A;q) as defined in Lemma 6; QAQ_A consists of exact prime powers, and ω(n)\omega(n) counts distinct prime divisors. Unqualified sums over qq are sums over prime powers.

Statement (Lemma 5, pp. 13-15). There is an absolute constant c>0c>0 with the following property. Let N≥M≥N1/2N\geq M\geq N^{1/2}, with NN sufficiently large, and let kk satisfy

1≤k≤clog⁡log⁡N.1\leq k\leq c\log\log N.

Suppose that A⊆[M,N]A\subseteq[M,N] is a set of integers for which

ω(n)≤(log⁡N)1/k(n∈A).\omega(n)\leq(\log N)^{1/k}\qquad(n\in A).

For every q∈QAq\in Q_A such that

R(A;q)≥(log⁡N)−1/2,R(A;q)\geq(\log N)^{-1/2},

there is an integer dd satisfying

qd>Mexp⁡(−(log⁡N)1−1/k),qd>M\exp\bigl(-(\log N)^{1-1/k}\bigr), ω(d)≤5log⁡klog⁡log⁡N,\omega(d)\leq\frac5{\log k}\log\log N,

and

∑n∈Aqqd∣n(qd,n/qd)=1qdn≫R(A;q)(log⁡N)2/k.\sum_{\substack{n\in A_q\\qd\mid n\\(qd,n/qd)=1}} \frac{qd}{n} \gg\frac{R(A;q)}{(\log N)^{2/k}}.

The endpoint k=1k=1 in the printed statement is undefined; see the proof scope immediately below.

Rewritten proof for 1<k≤clog⁡log⁡N1<k\le c\log\log N

The printed k=1k=1 endpoint has no defined 1/log⁡k1/\log k bound and is not asserted proved here. The subsequent Proposition 3 uses k→∞k\to\infty. Fix an eligible prime power qq, and define

y=exp⁡((log⁡N)1−2/k).y=\exp\bigl((\log N)^{1-2/k}\bigr).

Let DD consist of those positive integers dd for which both of the following hold:

pr∥d ⟹ pr>y,p^r\parallel d\ \Longrightarrow\ p^r>y,

and

qd∈(Mexp⁡(−(log⁡N)1−1/k), N].qd\in\left(M\exp\bigl(-(\log N)^{1-1/k}\bigr),\,N\right].

For every n∈Aqn\in A_q, begin with n/qn/q and remove all exact prime-power components pr∥n/qp^r\parallel n/q having pr≤yp^r\leq y. Since qq has already been separated, at most ω(n)−1\omega(n)-1 components are removed. Their product is strictly less than

yω(n)≤exp⁡((log⁡N)1−1/k).y^{\omega(n)} \leq\exp\bigl((\log N)^{1-1/k}\bigr).

The product of the components left behind is therefore some d∈Dd\in D. Moreover qd∣nqd\mid n and (qd,n/qd)=1(qd,n/qd)=1, because qq and all retained components are exact prime-power components of nn. It follows, after assigning to each nn such a dd, that

R(A;q)≤∑d∈D1d∑n∈Aqqd∣n(qd,n/qd)=1qdn.(5.1)R(A;q) \leq\sum_{d\in D}\frac1d \sum_{\substack{n\in A_q\\qd\mid n\\(qd,n/qd)=1}} \frac{qd}{n}. \tag{5.1}

Put

ω0=5log⁡klog⁡log⁡N.\omega_0=\frac5{\log k}\log\log N.

For fixed dd, discarding both the restriction n∈Aqn\in A_q and the coprimality condition gives

∑n∈Aqqd∣n(qd,n/qd)=1qdn≤∑n≤Nqd∣nqdn≪log⁡N.\sum_{\substack{n\in A_q\\qd\mid n\\(qd,n/qd)=1}}\frac{qd}{n} \leq\sum_{\substack{n\leq N\\qd\mid n}}\frac{qd}{n} \ll\log N.

Using $1_{\omega(d)\geq\omega_0}\leq k^{\omega(d)-\omega_0}$ and an Euler-product majorant, the portion of the right side of (5.1) with ω(d)>ω0\omega(d)>\omega_0 is at most

∑d∈Dω(d)>ω01d∑n∈Aqqd∣n(qd,n/qd)=1qdn≪log⁡N∑d: pr∥d⇒y<pr≤Nω(d)≥ω01d≪k−ω0log⁡N∑d: pr∥d⇒y<pr≤Nkω(d)d≪C1kk−ω0log⁡N∏y<p≤N(1+kp−1)≤k−ω0log⁡N(C2log⁡Nlog⁡y)k≤C2kk−ω0(log⁡N)3≤1log⁡N(5.2)\begin{aligned} &\sum_{\substack{d\in D\\\omega(d)>\omega_0}}\frac1d \sum_{\substack{n\in A_q\\qd\mid n\\(qd,n/qd)=1}}\frac{qd}{n}\\ &\quad\ll \log N \sum_{\substack{d:\ p^r\parallel d\Rightarrow y<p^r\leq N\\ \omega(d)\geq\omega_0}}\frac1d\\ &\quad\ll k^{-\omega_0}\log N \sum_{\substack{d:\ p^r\parallel d\Rightarrow y<p^r\leq N}} \frac{k^{\omega(d)}}d\\ &\quad\ll C_1^k k^{-\omega_0}\log N \prod_{y<p\leq N}\left(1+\frac{k}{p-1}\right)\\ &\quad\leq k^{-\omega_0}\log N \left(C_2\frac{\log N}{\log y}\right)^k\\ &\quad\leq C_2^k k^{-\omega_0}(\log N)^3 \leq\frac1{\log N} \end{aligned} \tag{5.2}

for absolute constants C1,C2>0C_1,C_2>0, provided cc is sufficiently small and NN sufficiently large. For prime bases p≤yp\le y, every allowed exponent is at least two. Their Euler factors have product at most exp⁡(k∑p∑a≥2p−a)≤ek\exp(k\sum_p\sum_{a\ge2}p^{-a})\le e^k. For p>yp>y, the factor is at most 1+k/(p−1)1+k/(p-1), which is at most (1−1/p)−k(1-1/p)^{-k} by Bernoulli's inequality for k>1k>1. Mertens' product estimate therefore gives the displayed bound; if 1<y<21<y<2, its logarithmic ratio is only larger, so the estimate still holds with an absolute constant. The last line uses

k−ω0=(log⁡N)−5,(log⁡Nlog⁡y)k=(log⁡N)2,k^{-\omega_0}=(\log N)^{-5}, \qquad \left(\frac{\log N}{\log y}\right)^k=(\log N)^2,

and k≤clog⁡log⁡Nk\leq c\log\log N.

Since R(A;q)≥(log⁡N)−1/2R(A;q)\geq(\log N)^{-1/2}, the last quantity in (5.2) is at most R(A;q)/2R(A;q)/2 once NN is large. Thus (5.1) gives

12R(A;q)≤∑d∈Dω(d)≤ω01d∑n∈Aqqd∣n(qd,n/qd)=1qdn.(5.3)\frac12R(A;q) \leq\sum_{\substack{d\in D\\\omega(d)\leq\omega_0}}\frac1d \sum_{\substack{n\in A_q\\qd\mid n\\(qd,n/qd)=1}} \frac{qd}{n}. \tag{5.3}

Another Euler-product estimate gives

∑d∈D1d≤∑d: pr∥d⇒y<pr≤N1d≪∏y<p≤N(1−1p)−1≪log⁡Nlog⁡y≪(log⁡N)2/k.(5.4)\begin{aligned} \sum_{d\in D}\frac1d &\leq \sum_{\substack{d:\ p^r\parallel d\Rightarrow y<p^r\leq N}}\frac1d\\ &\ll \prod_{y<p\leq N}\left(1-\frac1p\right)^{-1}\\ &\ll\frac{\log N}{\log y} \ll(\log N)^{2/k}. \end{aligned} \tag{5.4}

Comparing (5.3) and (5.4), at least one d∈Dd\in D with ω(d)≤ω0\omega(d)\leq\omega_0 has

∑n∈Aqqd∣n(qd,n/qd)=1qdn≫R(A;q)(log⁡N)2/k.\sum_{\substack{n\in A_q\\qd\mid n\\(qd,n/qd)=1}} \frac{qd}{n} \gg\frac{R(A;q)}{(\log N)^{2/k}}.

Membership in DD supplies the required lower bound for qdqd, completing the proof.

Dependencies and source details

Mertens' estimates, equations (1)–(2), p. 3, are external inputs, as cited by Bloom to Montgomery and Vaughan, Chapter 2. The Euler product in the unweighted estimate is written above in its ordinary form ∏(1−1/p)−1\prod(1-1/p)^{-1}; the paper's displayed majorant ∏(1−1/(p−1))−1\prod(1-1/(p-1))^{-1} is unnecessary and is undefined at p=2p=2, which lies in its range y<py<p whenever y<2y<2. The ordinary form handles the entire proof range k>1k>1.

Bears on