codex 5.2 extra high attempt on Math-5373

Problem 20 (Sunflowers): fixed-window entropy ledger framework

1) FORMAL RESTATEMENT

For integers $n\ge 1$ and $k\ge 2$, a $k$-sunflower is a collection of distinct sets $S_1,\ldots,S_k$ for which there is a common kernel $K$ such that \[ S_i\cap S_j=K\qquad (i\ne j). \] Let $f(n,k)$ be the least integer such that every $n$-uniform family $\mathcal F$ with $|\mathcal F|\ge f(n,k)$ contains a $k$-sunflower. Problem 20 asks whether, for every fixed $k$, there is a constant $C_k$ such that \[ f(n,k)\le C_k^n\qquad\text{for all }n. \]

This attempt focuses on the three-petal case after the standard reduction to balanced two-collision codes. It does not claim a complete proof of the sunflower conjecture.

2) QUICK LITERATURE/CONTEXT CHECK

The classical Erdos--Rado sunflower lemma gives a superexponential-in-$n$ bound for fixed $k$. The modern line of work of Alweiss--Lovett--Wu--Zhang, Rao, Tao, and Bell--Chueluecha--Warnke gives the current logarithmic-base improvements through spread or robust-sunflower methods. Entropy arguments are therefore already standard in this problem. Collision estimates are also standard local tools, and potential/amortized accounting is a standard proof technique.

The proposed novelty of this attempt is architectural rather than a new standalone inequality. It asks whether the reusable spread invariant in ALWZ/Rao/Tao-type arguments can be replaced by a coordinate-level collision/accounting invariant.

3) ATTACK PLAN

Reduction target. Reduce the three-petal sunflower problem to bounding balanced two-collision codes: \[ \mathcal C\subseteq \prod_{i=1}^n \Omega_i, \] where every three distinct codewords have a coordinate in which exactly two of the three symbols agree. A bound $|\mathcal C|\le D^n$ with $D$ independent of the alphabets would imply a constant-base sunflower bound after the standard random-coloring reduction.

Proposed replacement for spread refinement. Instead of maintaining a spread subfamily, maintain:

• per-coordinate collision information,
• a fixed labelled hash window for low-collision coordinates,
• terminal cylinder shadows for high-mass values,
• fresh/old/scale potentials that pay for repeated use.

Audit target. The attempt succeeds only if every nonterminal transition can be assigned to a finite list of local account inequalities, with constants independent of $n$ and independent of the alphabet sizes.

4) WORK

4.1. Standard reduction to balanced two-collision codes

Pad a rank-$n$ sunflower-free family to an $n$-uniform family by adding private dummy elements. Randomly color the ground set by colors $1,\ldots,n$ and keep the colorful sets. Some coloring keeps at least $(n!/n^n)|\mathcal F|$ sets. Encoding a colorful set by its unique element of each color gives a code \[ \mathcal C\subseteq \prod_i\Omega_i . \] If three codewords have no coordinate with exactly two equal symbols, then in each color class the corresponding set-elements are either all equal or all distinct, so the three pairwise intersections are equal and the original three sets form a sunflower. Thus the resulting code is balanced two-collision.

Therefore an alphabet-free estimate $|\mathcal C|\le D^n$ implies $|\mathcal F|\le (eD)^n$.

4.2. Terminal entropy descent

Let $\mu$ be the uniform law on a balanced two-collision code $\mathcal C$. The proof tree samples one or more independent codewords from $\mu$, but the terminal object is only a projected cylinder: \[ \theta=((J_s,a_s))_{s\in S}, \] meaning $X^s_{J_s}=a_s$ for each sample label $s$. Write \[ T(\theta)=\sum_s |J_s|,\qquad R(\theta)=|S|. \] If a terminal cylinder has mass \[ \Pr[\theta]\ge \exp(-\lambda T(\theta)-cR(\theta)), \] then the usual induction on rank gives $|\mathcal C|\le D^n$ whenever $\log D\ge \lambda+c$. Indeed, each cylinder $\{x:x_{J_s}=a_s\}$ is again a balanced two-collision code after deleting the fixed coordinates.

Thus the goal is to produce a random terminal projected cylinder $\Theta$ with \[ \mathbf H(\Theta) \le \mathbb E[\lambda T(\Theta)+cR(\Theta)] \] for absolute constants $\lambda,c$.

4.3. Fixed labelled hash window

For a one-coordinate law $p$, write \[ \operatorname{Coll}(p)=\sum_a p(a)^2. \] If $X,Y,Z$ are independent samples and coordinate $i$ has collision at most $\beta$, then \[ \Pr[i\in \operatorname{Split}(X,Y,Z)]\le 3\beta. \] Hence a block $H$ of $h$ low-collision coordinates satisfies \[ \Pr\left[|\operatorname{Split}_H(X,Y,Z)|\le 6\beta h\right]\ge 1/2. \]

The proposed mechanism collects low-collision coordinates in a fixed labelled hash window. The label is one of the active sample labels. The window has a fixed transcript budget $B_{\rm win}$ and stops before exceeding that budget. High-mass values are not exposed recursively: if a coordinate develops an atom of mass at least $\beta$, the branch terminalizes by pinning that value as a projected shadow.

If a nonterminal labelled window obtains $h$ surviving coordinates, three conditioned copies under the same public transcript produce a near-sunflower block. The transcript key has Renyi-3 copy cost \[ 2\mathbf H_3(K)+O(1)\le 2\mathbf H(K)+O(1)\le O(B_{\rm win}). \] The sparse split exposure costs $O(h)$. Thus the local hash-return cost is \[ C_{\rm hash}=O(B_{\rm win}+h), \] with $B_{\rm win}$ and $h$ fixed constants.

4.4. Product-KL routing

In a close-pair node, one has a block $J$ on which two sampled codewords agree and the third differs: \[ X_J=Y_J=A_J,\qquad Z_J=B_J\ne A_J . \] The intended product-KL branch theorem scans the actual tuple-coordinate list \[ (A_{j_1},B_{j_1},A_{j_2},B_{j_2},\ldots). \] Every branch must route to exactly one of:

1. an amortized terminal prefix;
2. a one-coordinate heavy value pin;
3. a charged pruning event not inherited by descendants;
4. a genuine low-collision law of an actual sample coordinate $X_j$, $Y_j$, or $Z_j$, sent to the labelled hash window.

The critical requirement is that the low-collision law is the actual conditional sample-coordinate law, not a law on the pair alphabet $(A_j,B_j)$ and not the auxiliary reference product law.

4.5. Global account inequality

The fixed hash window gives only a one-time near-sunflower block. It does not produce a reusable spread subfamily. The proposed replacement invariant is a state with potential \[ \Phi = B_f |[n]\setminus U| +(B_o|U|-N_{\rm old}) +B_s\sum_{\text{active }J}|J|. \] Here $U$ is the old support, $N_{\rm old}$ is a weighted count of registered old incidences, and the last term pays active scale/rank bookkeeping.

The desired branch inequality is \[ \operatorname{spent}(\pi) + C_{\rm hash}q(\pi) + \Phi(S_\pi) \le \Phi(S_0)+\operatorname{terminal\ shadow\ charge}(\pi), \] where $q(\pi)$ is the number of completed hash windows on the branch.

At the one-step level, the required inequalities have the form \[ (B_f-B_o)|F| \ge \text{fresh selector cost}+\text{child deposit}+\text{possible hash token}, \] \[ \omega_{\rm old}m \ge \text{old transcript cost}+\text{child deposit}+\text{possible hash token}, \] and \[ \text{rank/scale drop} \ge \text{new active-block deposit}+\text{selector cost}. \] If all transition modes are exhaustive and these inequalities cover every selector, the entropy criterion above would close with constants independent of the alphabet.

5) VERIFICATION / ATTEMPTED BREAKS

Known verification obligations. The following are the main places where the attempt can fail:

1. The fixed-window hash return must really have per-window cost $O(B_{\rm win}+h)$, not a hidden per-branch or alphabet-dependent cost.
2. The product-KL branch theorem must output actual sample-coordinate laws.
3. Completed hash-window tokens must be matched injectively to hard exits; tokens cannot pile up along a nested chain.
4. The residual repair procedure must immediately route every returned near-sunflower block to fresh, old, scale, or terminal progress.
5. The transition system must be exhaustive; no branch may continue without a terminal shadow or a potential drop.

Falsification tests reported by the contributor. The contributor reports using executable smoke tests for the following checks. The supporting scripts are not included in this submission:

• alphabet stretching of toy balanced two-collision codes;
• multi-label hash-token matching;
• terminal cylinder potential compatibility;
• transition-exhaustiveness signatures;
• constant-hierarchy acyclicity;
• old-window queue refinement;
• hash-prefix budget handling;
• Renyi copy-tax accounting;
• collision-to-split estimates;
• high-revisit branch exhaustiveness;
• randomized ledger walks.

These tests are not formal proofs. They are only attempts to catch common bookkeeping failures.

Internal LLM audit status. Several long GPT Pro, Claude Opus, and Codex adversarial-checking passes reported that the current interface-level ledger appears closed and did not identify a remaining bookkeeping counterexample. This is recorded only as provenance for the attempt. It is not expert verification, not a formal proof, and not a claim that the sunflower conjecture is solved.

6) FINAL

PARTIAL / UNRESOLVED.

This is not a complete proof of the Erdos--Rado sunflower conjecture. The main contribution of the attempt is a proposed conditional framework: if the product-KL routing, fixed-window hash return, residual repair, and global account inequality are all correct with alphabet-free constants, then the usual terminal-cylinder entropy descent would yield a constant-base bound for balanced two-collision codes, and hence for three-petal sunflower-free families after the standard reduction.

The exact open gap is: \[ \text{prove, without ``by inspection'', that every transition branch is covered by the one-step account inequalities above.} \]

7) COMPLETION ESTIMATE

COMPLETION ESTIMATE: 20%.

This percentage reflects that the approach has a defined reduction, candidate local mechanisms, and executable falsification tests, but still lacks an independent verification of the full transition-system exhaustiveness and global charge inequality.

8) DISCLOSURE / PROVENANCE

This file is not the transcript of a single model run. It is a Codex-organized summary of a much longer LLM-assisted research attempt. The underlying framework was developed through many extended conversations with frontier LLMs, including GPT Pro, Claude Opus, and Codex. Some of those conversations became too long to open reliably, some are no longer accessible in full, and the Codex continuation itself used compressed conversation summaries. Consequently this submission should be read as a reconstructed and organized record of the LLM attempt, not as a verbatim complete transcript.

The contributor also prepared a longer draft, a short roadmap, audit notes, and smoke-test scripts as separate research materials. Those materials are not included in this submission. No part of this submission should be treated as expert-verified.

9) REFERENCES

References

• [ALWZ] R. Alweiss, S. Lovett, K. Wu, and J. Zhang, Improved bounds for the sunflower lemma, Ann. of Math. $2$ 194 $2021$, no. 3, 795--815.
• [Rao] A. Rao, Coding for sunflowers, Discrete Anal. $2020$, Paper No. 2.
• [Tao] T. Tao, The sunflower lemma via Shannon entropy, blog post, 2020.
• [BCW] T. Bell, S. Chueluecha, and L. Warnke, Note on sunflowers, Discrete Math. 344 $2021$, no. 7, 112367.
• [ER] P. Erdos and R. Rado, Intersection theorems for systems of sets, J. London Math. Soc. 35 $1960$, 85--90.