gpt pro 5.2 attempt on Math-929
mathematicsErdős problemscontributionManuscript
Contents
- Problem 773: Sidon subsets of the first $N$ squares
- 1) Formal restatement
- 2) Quick literature/context check (no more than 8 lines)
- 3) Attack plan
- 4) Work
- Upper bound $\boldsymbol{f(N)\ll N(\log N)^{-1/4}}$.
- A general extraction lemma.
- Lower bound for squares.
- 5) Verification
- 6) Final (exactly one label and one sub-label)
- 7) Completion estimate
Problem 773: Sidon subsets of the first $N$ squares
1) Formal restatement
Let $S_N=\{1^2,2^2,\dots,N^2\}$. A set $A\subseteq\mathbb{Z}$ is Sidon (a $B_2$-set) if \[ a+b=c+d\ (a,b,c,d\in A)\quad\Longrightarrow\quad \{a,b\}=\{c,d\}. \] Equivalently, all sums $a+b$ with $a<b$ are distinct. Define \[ f(N):=\max\{|A|:A\subseteq S_N \text{ is Sidon}\}. \] The problem asks for the best possible asymptotic bounds on $f(N)$.2) Quick literature/context check (no more than 8 lines)
- Landau's theorem on sums of two squares yields $f(N)\ll N(\log N)^{-1/4}$.
- A graph/hypergraph extraction lemma of Alon--Erdős implies $f(N)\gg_\varepsilon N^{2/3-\varepsilon}$ for every $\varepsilon>0$.
- Lefmann--Thiele improved the lower bound to $f(N)\gg N^{2/3}$.
- The true order of $f(N)$ remains open: there is still a large gap between exponents $2/3$ and $1$ (up to logs).
3) Attack plan
- Upper bound: count the number of distinct pair-sums $a+a'$ forced by the Sidon property and compare with Landau's count of integers that are sums of two squares.
- Lower bound: prove a general lemma that any $B(2k)$ sequence of length $n$ contains a Sidon subsequence of size $\gg k^{-1/3}n^{2/3}$, then apply to squares using a divisor-function bound on the number of representations $m=i^2+j^2$.
4) Work
Upper bound $\boldsymbol{f(N)\ll N(\log N)^{-1/4}}$.
Let $A\subseteq S_N$ be Sidon and let \[ \Sigma:=\{a+a': a,a'\in A,\ a<a'\}. \] Sidon implies that each unordered pair $\{a,a'\}$ gives a distinct sum, so $|\Sigma|=\binom{|A|}{2}$. Each element of $\Sigma$ is a sum of two squares, and $\Sigma\subseteq [2,2N^2]$.Let $R(X)$ be the number of integers $m\le X$ representable as a sum of two squares. Landau proved $R(X)\ll X/\sqrt{\log X}$, so \[ \binom{|A|}{2}=|\Sigma|\le R(2N^2)\ll \frac{N^2}{\sqrt{\log N}}. \] Hence $|A|^2\ll N^2(\log N)^{-1/2}$, i.e. \[ |A|\ll \frac{N}{(\log N)^{1/4}}. \] Therefore $f(N)\ll N(\log N)^{-1/4}$.
A general extraction lemma.
Let $r_m$ be the number of pairs $i<j$ with $a_i+a_j=m$. Then $\sum_m r_m=\binom{n}{2}$ and $r_m\le k$ for all $m$. Each fiber of size $r_m$ contributes at most $\binom{r_m}{2}$ edges, so \[ |E(H)|\le \sum_m \binom{r_m}{2}\le \frac{\binom{n}{2}}{k}\binom{k}{2}\le \frac{(k-1)n^2}{4}. \]
Choose each vertex independently with probability $p=\alpha k^{-1/3}n^{-1/3}$, where $\alpha\in(0,1)$ is a fixed small constant. Then $\mathbb{E}|X|=pn=\alpha k^{-1/3}n^{2/3}$. Also $\mathbb{E}|E(H[X])|=p^4|E(H)|\ll \alpha^4 k^{-1/3}n^{2/3}$. Deleting one vertex from each edge of $H[X]$ yields an independent set $I$ with $\mathbb{E}|I|\ge \mathbb{E}|X|-\mathbb{E}|E(H[X])|\ge (\alpha-C\alpha^4)k^{-1/3}n^{2/3}$. Choosing $\alpha$ small so that $\alpha-C\alpha^4>0$ gives $|I|\ge c k^{-1/3}n^{2/3}$ for some absolute $c>0$. ∎
Lower bound for squares.
Apply Lemma Lemma to $a_i=i^2$ for $i=1,\dots,N$ (so $n=N$). We need an upper bound on the number of representations of $m\le 2N^2$ as $m=i^2+j^2$ with $1\le i<j\le N$.Let $r_2(m)$ denote the number of integer solutions $(x,y)\in\mathbb{Z}^2$ to $x^2+y^2=m$. A classical formula implies $r_2(m)\le 4d(m)$, where $d(m)$ is the divisor function. Hence the number of pairs $(i,j)$ with $1\le i<j\le N$ and $i^2+j^2=m$ is $O(d(m))$.
Fix $\varepsilon>0$. It is standard that $d(m)\le m^{\varepsilon}$ for all sufficiently large $m$. Since $m\le 2N^2$, for $N$ large we have $d(m)\ll_\varepsilon N^{2\varepsilon}$, and thus \[ \#\{(i,j): i^2+j^2=m,\ 1\le i<j\le N\}\le N^{3\varepsilon} \] after increasing $N$ if needed (absorbing constants). Therefore $\{1^2,\dots,N^2\}$ is a $B(2k)$-sequence with $k=N^{3\varepsilon}$, and Lemma Lemma gives a Sidon subset of size \[ \gg k^{-1/3}N^{2/3} = N^{2/3-\varepsilon}. \] So $f(N)\gg_\varepsilon N^{2/3-\varepsilon}$ for every $\varepsilon>0$. (A stronger bound $f(N)\gg N^{2/3}$ is known by Lefmann--Thiele.)
5) Verification
- Upper bound: $|\Sigma|=\binom{|A|}{2}$ is exact for Sidon; Landau's bound applies to all sums $a+a'\le 2N^2$.
- Lemma Lemma: the hypergraph edge count $\sum_m\binom{r_m}{2}$ is maximized by packing fibers to size $k$, giving $|E(H)|\le (k-1)n^2/4$.
- Divisor bound step: only needs the qualitative fact $d(m)\le m^\varepsilon$ for large $m$, hence $k=N^{O(\varepsilon)}$.
6) Final (exactly one label and one sub-label)
LABEL: UNRESOLVED.SUBLABEL: Proved the Landau upper bound and an $N^{2/3-\varepsilon}$ lower bound (with known improvement to $N^{2/3}$); the true order of $f(N)$ is open.