Math-5895 · mathematics · Mathematics

Math-5895

Erdős Problem #785

Status Solved · erdosproblems.comLLM status Solved from source status

Erdős problems openaigpt 6 astra ultra 100% cataloguedSolvedSolved

Statement

Let infinite $A,B\subseteq\mathbb N$ satisfy $A+B\supseteq[N_0,\infty)\cap\mathbb N$ and $A(x)B(x)\sim x$. Then $A(x)B(x)-x\longrightarrow\infty$. We reconstruct Ruzsa's stronger bound \[ A(x)B(x)-x\geq(1-o(1))\frac{a^*(x)}{A(x)},\qquad a^*(x)=\max(A\cap[1,x]), \tag{1} \] after interchanging $A,B$ if necessary. Source: I. Z. Ruzsa, Exact additive complements, https://arxiv.org/abs/1510.00812, Section 2. See https://www.erdosproblems.com/785. The external input is Narkiewicz's dichotomy, stated as Theorem 1.1 of that paper: \[ A(2x)/A(x)\to1,\qquad B(2x)/B(x)\to2 \tag{2} \] in one fixed ordering of the two sets.

Restatement from Math-5895-A1 r1, with the source manuscript’s qualifications.

Original problem on erdosproblems.com

Statement licence: CC-BY-4.0 · Source file

Contributions 1

Math-5895-A1 · No claim recorded · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Math-5895

Erdős Problem 785: the surplus of exact additive complements

Statement and claim.
Let infinite $A,B\subseteq\mathbb N$ satisfy $A+B\supseteq[N_0,\infty)\cap\mathbb N$ and $A(x)B(x)\sim x$. Then $A(x)B(x)-x\longrightarrow\infty$. We reconstruct Ruzsa's stronger bound \[ A(x)B(x)-x\geq(1-o(1))\frac{a^*(x)}{A(x)},\qquad a^*(x)=\max(A\cap[1,x]), \tag{1} \] after interchanging $A,B$ if necessary. Source: I. Z. Ruzsa, Exact additive complements, https://arxiv.org/abs/1510.00812, Section 2. See https://www.erdosproblems.com/785. The external input is Narkiewicz's dichotomy, stated as Theorem 1.1 of that paper: \[ A(2x)/A(x)\to1,\qquad B(2x)/B(x)\to2 \tag{2} \] in one fixed ordering of the two sets.
Consequences of the dichotomy.
By iteration and monotonicity, for every fixed $c>0$, $A(cx)/A(x)\to1$, uniformly when $c$ stays in a compact subinterval of $(0,\infty)$. Together with exactness, this gives $B(cx)/B(x)\to c$, uniformly for $0\leq c\leq1$; uniformity near zero follows by monotonicity and then letting a fixed small $c$ tend to zero. Iteration of $A(2x)\leq2^\varepsilon A(x)$, valid for sufficiently large $x$, also gives $A(x)=x^{o(1)}$. Finally \[ \sum_{\substack{a\in A\\a\leq t}}a=o(tA(t)). \tag{3} \] Indeed, the terms below $\eta t$ contribute at most $\eta tA(t)$, and the others at most $t(A(t)-A(\eta t))=o(tA(t))$; then let $\eta\to0$.
Comparing repeated sums and differences.
For finite sets $U,V$, let $\sigma(s)=|\{(u,v)\in U\times V:u+v=s\}|$ and $\delta(d)=|\{(u,v)\in U\times V:v-u=d\}|$. Counting equal-sum quadruples in two ways gives \[ \sum_s\sigma(s)^2=\sum_d\delta(d)^2,\qquad \sum_s\sigma(s)=\sum_d\delta(d)=|U||V|. \] Since $\sigma(s)\leq |U|$, \[ \sum_d(\delta(d)-1)_+ \leq \sum_d\delta(d)(\delta(d)-1) \leq |U|\sum_s(\sigma(s)-1)_+. \tag{4} \]
The surplus decomposition.
It suffices initially to consider integer $x$. Put $a=A(x)$, $t=a^*(x)$, $U=A\cap[1,x]$, and $V=B\cap[1,x]$. Let $r$ be the number of positive integers at most $x$ missing from $A+B$, so $r=O(1)$. Let $y=\sum_s(\sigma(s)-1)_+$ and $z=|(U+V)\cap(x,\infty)|$. Positivity of the summands gives \[ E:=A(x)B(x)-x=y+z-r. \tag{5} \] The sums $t+b$, with $x-t<b\leq x$, are distinct and exceed $x$, so $z\geq B(x)-B(x-t)$.

If $t\geq x/2$, uniform regular variation of $B$ gives \[ z\geq(t/x-o(1))B(x)=(1-o(1))t/a. \] Since $a=A(t)=t^{o(1)}$, the quantity $t/a$ tends to infinity; the bounded term $r$ in $5$ is negligible.

The case $t<x/2$.
Set $V'=B\cap[1,x-t]$, and use $\sigma',\delta'$ for its sum and difference counts with $U$. Because there is no element of $A$ in $(t,x]$, \[ A(a'+t)=A(t)=a\qquad(a'\in U). \] Uniformly for arguments at least $t$, exactness gives $B(a'+t)=(1+o(1))(a'+t)/a$, as $t\to\infty$. By $3$, \[ \#\{(a',b)\in U\times V':b-a'\leq t\} \leq\sum_{a'\in U}B(a'+t)=(1+o(1))t. \] All remaining differences lie in the interval $[t+1,x-t-1]$, of length less than $x-2t$. For any fixed $\varepsilon>0$ and sufficiently large $x$, their repeated multiplicities therefore total at least \[ aB(x-t)-x+(1-\varepsilon)t. \] Apply $4$ to $U,V'$, and note that its repeated sums contribute to $y$. We get \[ y\geq B(x-t)-\frac{x-(1-\varepsilon)t}{a}. \] Adding the bound for $z$ and using $5$ yields \[ E+r\geq \frac{E+(1-\varepsilon)t}{a}, \qquad E\geq\frac{(1-\varepsilon)t-ar}{a-1}. \tag{6} \] Both $a$ and $t$ tend to infinity, while $a=t^{o(1)}$ and $r=O(1)$. Thus $ar=o(t)$, and $6$, followed by $\varepsilon\to0$, proves $1$ in this case as well.
Conclusion.
For real $x$, replace $x$ by $\lfloor x\rfloor$; the counting functions and $a^*(x)$ are unchanged and the surplus changes by less than one. Since $t/A(t)\to\infty$, equation $1$ proves the asserted divergence for all real $x\to\infty$. This is a full reconstruction relative to Narkiewicz's stated dichotomy, not a new proof of that external theorem.
Completion estimate.
COMPLETION: 100%

Go to contribution →