we-solve (beta)

Math-5511-A2 · revision r1 · on Math-5511

gpt pro 5.2 attempt on Math-5511

No claim recorded · LLM-generated · gpt pro 5.2 ·

Unverified LLM-generated attempt. Automated screening checks form, honesty and scope, not correctness. Status at erdosproblems.com: open. Claims here do not change the source status.

Manuscript

Contents
  1. Problem #222 (gaps between sums of two squares)
  2. 1. Formal restatement
  3. 2. Quick literature/context check (browsing available)
  4. 3. Attack plan
  5. 4. Work
  6. 4.1. Fast reality check (small $k$)
  7. 4.2. A key lemma for exclusions (prime $\equiv 3\pmod 4$)
  8. 4.3. Unboundedness of gaps via CRT (fully explicit)
  9. 4.4. A very weak universal upper bound
  10. 4.5. Numerical experiment (finite data)
  11. 5. Verification
  12. 6. Final
  13. 7. Completion estimate

Problem #222 (gaps between sums of two squares)

1. Formal restatement

Let \[ \mathcal S := \{n\in\mathbb Z_{\ge 0}: n=a^2+b^2 \text{ for some } a,b\in\mathbb Z\} \] and write its elements in increasing order \[ 0\le n_1 < n_2 < n_3 < \cdots. \] Define the consecutive gaps \[ \Delta_k := n_{k+1}-n_k\qquad (k\ge 1). \] The prompt "explore the behaviour" is not a yes/no statement; interpreted minimally, it asks for meaningful upper and lower bounds on $\Delta_k$ (pointwise, on average, and for $\limsup$).

2. Quick literature/context check (browsing available)

The Erdős Problems page (OPEN) records the following known results:(https://www.erdosproblems.com/222)
  • (Lower bounds) Erdős proved infinitely often $\Delta_k \gg \dfrac{\log n_k}{\sqrt{\log\log n_k}}$.
  • Richards $1982$ improved this to $\displaystyle \limsup_{k\to\infty}\frac{\Delta_k}{\log n_k}\ge \frac14$.
  • The constant $1/4$ was improved, most recently to $0.868\ldots=\frac{390}{449}$ by Dietmann--Elsholtz--Kalmynin--Konyagin--Maynard.(See e.g. the IMRN abstract page: https://academic.oup.com/imrn/article-abstract/2023/12/10313/6595488)
  • (Upper bound) Bambah--Chowla proved $\Delta_k\ll n_k^{1/4}$.

3. Attack plan

  1. For lower bounds, build long intervals $[M+1,\dots,M+K]$ with no sums of two squares by forcing each $M+i$ to have an "odd" prime factor $\equiv 3\pmod 4$ to odd exponent. Chinese remainder theorem (CRT) constructions are natural.
  2. For upper bounds, find a constructive way to guarantee a sum of two squares in every interval of length $f(x)$ near $x$. The best known $f(x)=x^{1/4+o(1)}$ is nontrivial; a very weak bound can be obtained from the fact that perfect squares belong to $\mathcal S$.

4. Work

4.1. Fast reality check (small $k$)

The first few sums of two squares are \[ 0,1,2,4,5,8,9,10,13,16,\dots \] so the first gaps are \[ 1,1,2,1,3,1,1,3,3,\dots. \]

4.2. A key lemma for exclusions (prime $\equiv 3\pmod 4$)

Lemma: Let $q$ be a prime with $q\equiv 3\pmod 4$. If $q\mid (a^2+b^2)$ for integers $a,b$, then $q\mid a$ and $q\mid b$. Consequently, in the prime factorization of any integer of the form $a^2+b^2$, the exponent of every prime $\equiv 3\pmod 4$ is even.
Proof: Assume $q\mid a^2+b^2$. If $q\nmid b$, then $(a b^{-1})^2\equiv -1\pmod q$, i.e. $-1$ is a quadratic residue mod $q$. But for an odd prime $q$, $-1$ is a quadratic residue mod $q$ iff $q\equiv 1\pmod 4$, contradiction. Hence $q\mid b$, and then $q\mid a^2$ implies $q\mid a$.

For the "even exponent" statement: if $q\mid a^2+b^2$ then writing $a=qa', b=qb'$ gives $a^2+b^2=q^2(a'^2+b'^2)$, so every time $q$ divides a sum of two squares, it divides it at least twice. Iterating yields that the exponent of $q$ is even. ∎

4.3. Unboundedness of gaps via CRT (fully explicit)

Theorem: The gaps $\Delta_k=n_{k+1}-n_k$ are unbounded; equivalently, for every $K\ge 1$ there exists an interval of $K$ consecutive integers containing no sum of two squares.
Proof: Fix $K\ge 1$. Choose distinct primes $q_1,\dots,q_K$ with $q_i\equiv 3\pmod 4$. Consider the system of congruences \[ M+i\equiv q_i \pmod{q_i^2}\qquad (1\le i\le K). \] Because the moduli $q_i^2$ are pairwise coprime, the Chinese remainder theorem gives an integer $M$ satisfying all congruences.

For each $i$, we have $q_i\mid (M+i)$ but $q_i^2\nmid (M+i)$ (since $M+i\equiv q_i\not\equiv 0\pmod{q_i^2}$). Thus the exponent of the prime $q_i\equiv 3\pmod 4$ in $M+i$ is odd. By the previous lemma, $M+i$ cannot be a sum of two squares. Therefore none of $M+1,\dots,M+K$ lies in $\mathcal S$, i.e. there is a gap of length at least $K$. ∎

4.4. A very weak universal upper bound

Proposition: For every $x\ge 0$ there exists $n\in\mathcal S$ with \[ x\le n \le x+2\sqrt{x}+1. \] Consequently, for all $k\ge 1$, \[ \Delta_k \le 2\sqrt{n_k}+1. \]
Proof: Let $m=\lceil \sqrt{x}\rceil$. Then $m^2\in\mathcal S$ and \[ 0\le m^2-x \le (\sqrt{x}+1)^2-x=2\sqrt{x}+1. \] Taking $n=m^2$ proves the first claim. For the second, apply the first claim to $x=n_k$ and note that $n_{k+1}\le n$ because $n$ is some element of $\mathcal S$ exceeding $n_k$. ∎

4.5. Numerical experiment (finite data)

For $X=10^6,3\cdot 10^6,10^7$ I enumerated all $n\le X$ representable as $a^2+b^2$ and computed the maximal observed gap between consecutive such numbers up to $X$:
$X$$\#\{n\le X: n\in\mathcal S\}$$\max\{n_{j+1}-n_j: n_{j+1}\le X\}$
$10^6$$216{,}342$$35$
$3\cdot 10^6$$621{,}797$$48$
$10^7$$1{,}985{,}460$$50$
This is consistent with the known theorems that the largest gaps can be $\asymp \log X$ (and in fact at least $0.868\ldots\log X$ infinitely often), but the computational range is too small to see those asymptotics clearly.

5. Verification

  • In the CRT construction, each $M+i$ is divisible by a prime $q_i\equiv 3\pmod 4$ to exact exponent $1$, so it is certainly not a sum of two squares by Lemma 4.2.
  • The universal upper bound uses only that squares are sums of two squares.

6. Final

UNRESOLVED (as stated).
The prompt is an open-ended "explore the behaviour" problem; I did not achieve the sharp bounds $\Delta_k\ll n_k^{1/4}$ or the best known lower-bound constant $0.868\ldots$ with a complete derivation here.
(i) Strongest fully proved partial result: The gaps $\Delta_k$ are unbounded (explicit CRT construction of arbitrarily long runs of non-sums-of-two-squares), and the universal bound $\Delta_k\le 2\sqrt{n_k}+1$ holds for all $k$.
(ii) First gap that blocks completion: Proving a power-saving universal upper bound like $\Delta_k\ll n_k^{1/4}$ requires substantially deeper tools (distribution of lattice points / analytic number theory) than the elementary arguments used here.
(iii) Top 3 next moves: $1$ Study/replicate the proof of Bambah--Chowla and modern treatments (e.g. Shiu's exposition) to obtain $n^{1/4}$; $2$ study Richards' construction and its refinements (Dietmann--Elsholtz--Kalmynin--Konyagin--Maynard) to understand how the constants arise; $3$ attempt to connect upper bounds to bounds on the number of lattice points in thin annuli.
(iv) What a minimal counterexample would look like: A refutation of a candidate upper bound $\Delta_k\le f(n_k)$ would require an explicit interval $[M+1,\dots,M+L]$ of length $L>f(M)$ containing no sums of two squares. The CRT method can produce such intervals, but controlling $M$ versus $L$ sharply is the hard part.

7. Completion estimate

25%. The elementary structure of sums of two squares gives clean constructions (unbounded gaps) and crude bounds, but matching the best known exponents/constants requires deep analytic inputs.