GPT 6 Astra Ultra attempt on Math-5732

Erdős Problem #551: restrictions on a smallest cycle--clique counterexample

Date: 2026-09-23. Partial research attempt; no claim of a new Ramsey bound.

1) TARGET AND SOURCE AUDIT

For $k\ge n\ge3$, excluding $k=n=3$, the target is \[ R(C_k,K_n)=(k-1)(n-1)+1. \] Equivalently, every graph on the displayed number of vertices should contain either a cycle of length exactly $k$ or an independent set of size $n$. Cycles and copies are not required to be induced. The live page labels the problem DECIDABLE and records a theorem covering all sufficiently large clique parameters. The finite remainder is not automatically settled by that theorem.

The objective here is to obtain exact necessary conditions on a smallest remaining counterexample and to verify a small instance with a reproducible exhaustive search.

2) WORK

The lower bound.

The disjoint union of $n-1$ copies of $K_{k-1}$ has $(k-1)(n-1)$ vertices. Every cycle stays in a component, so it has no $C_k$. An independent set meets each component in at most one vertex, so it has no independent $n$-set. Hence the stated expression is always a lower bound. The entire unresolved content is the matching upper bound.

Independent sets in a minimal failure must expand.

Fix $k\ge4$, and suppose $n\in\{3,\ldots,k\}$ is the smallest clique parameter for which the upper bound fails at this $k$. Let $G$ be a counterexample with \[ N=(k-1)(n-1)+1 \] vertices. If $S$ is a nonempty independent set of size $s$, then \[ |N_G(S)|\ge(k-2)s+1, \tag{1} \] where $N_G(S)$ means vertices outside $S$ adjacent to at least one vertex of $S$.

To prove this, let $W$ consist of vertices outside $S$ with no neighbour in $S$. The graph $G[W]$ is $C_k$-free and cannot contain an independent set of size $n-s$, because such a set together with $S$ would contradict $\alpha(G)<n$. Minimality of $n$ therefore gives \[ |W|\le(k-1)(n-s-1). \tag{2} \] The smaller parameters $n-s=2$ and $n-s=1$ cause no gap: $R(C_k,K_2)=k$, since a graph with no independent pair is complete, and $R(C_k,K_1)=1$. Also $s\le n-1$. Subtracting $2$ and $|S|$ from $N$ proves $1$.

In particular $\delta(G)\ge k-1$, by taking $s=1$. More generally, an independent set cannot be separated from the rest of the graph by at most $(k-2)|S|$ vertices. This is a stronger necessary condition than the minimum-degree condition and can be used to reject candidate graphs in a finite search.

Why the minimum-degree conclusion is insufficient.

A large minimum degree does not by itself force a cycle of a specified length. For example, if $k$ is odd then $K_{k-1,k-1}$ has minimum degree $k-1$ but has no odd cycle. This example is not a counterexample to the conjecture because it has a large independent set. It demonstrates precisely why an argument retaining only $\delta(G)\ge k-1$ has discarded crucial information.

An additional restriction away from the diagonal.

Choose the smallest $k$ for which any failure occurs, and then the smallest $n$ for that $k$. If $k>n$, the established statement for $(k-1,n)$ forces $G$ to contain a cycle $C$ of length $k-1$: the graph has at least $(k-2)(n-1)+1$ vertices and no independent $n$-set. The potential exception $(k,n)=(4,3)$ is ruled out by the finite verification below, so this argument never invokes the excluded pair $(3,3)$.

No vertex outside $C$ can be adjacent to two consecutive vertices of $C$. Otherwise replacing that cycle edge by the two-edge path through the outside vertex gives a cycle of length exactly $k$. Consequently, \[ |N_G(v)\cap V(C)|\le\left\lfloor\frac{k-1}{2}\right\rfloor \quad(v\notin V(C)). \] Combining this with $1$ for singletons gives \[ \delta\left(G-V(C)\right)\ge\left\lceil\frac{k-1}{2}\right\rceil. \tag{3} \] The restriction does not assume that $C$ is induced. It does require $k>n$: applying the conjecture at $(k-1,n)$ on the diagonal would leave the permitted parameter range. Also $3$ does not by itself force a suitable cycle or independent set; it only sharpens the description of a possible minimal failure.

A completed finite check: $R(C_4,K_3)=7$.

The accompanying program verification/add_graphs_checks.py recursively generates every labelled triangle-free graph $B$ on seven vertices. An edge is added only when its endpoints have no common neighbour among edges already present. Since every prefix of a triangle-free graph is triangle-free, this pruning omits none of the target graphs. Each of the $133501$ generated graphs is checked separately.

Set $G=\overline B$. The absence of a triangle in $B$ is exactly the absence of an independent triple in $G$. A graph contains a $C_4$ if and only if some two distinct vertices have at least two common neighbours. The program checks this criterion, including cycles with chords, and finds a $C_4$ in every complement. Thus the upper bound is verified. The disjoint union of two triangles proves the lower bound on six vertices. This is a verification of a known small instance, not an improvement over existing results or a finite check of all outstanding pairs.

3) ADVERSARIAL VERIFICATION

Minimality is taken first in $n$ at fixed $k$ for $1$, and first in $k$ and then $n$ for $3$; these hypotheses have been kept explicit. Formula $2$ concerns the common nonneighbour set, not the union of the individual nonneighbour sets. The finite program uses labelled graphs without a potentially unsafe isomorphism reduction. Its search covers seven vertices only, and there is no extrapolation to larger parameters.

4) FINAL

UNRESOLVED. The lower bound, the independent-set expansion requirement, the off-diagonal cycle restriction, and the $(k,n)=(4,3)$ check are complete. No contradiction has been obtained from the necessary conditions for arbitrary remaining pairs. These observations are not claimed to be new to the literature.

5) SOURCES CHECKED

The statement and current status were fetched on 2026-09-23 from https://www.erdosproblems.com/551. The bibliographic results recorded there supply context only; the arguments above are self-contained apart from the explicitly conditional use of minimality. The reproducible computation is in verification/add_graphs_checks.py, with results in its adjacent JSON file.

6) COMPLETION ESTIMATE

Only one small parameter pair is computationally settled here; the structural restrictions have not resolved the remaining finite range. The estimate is subjective and is neither a probability nor a guaranteed fraction of the mathematics completed. COMPLETION: 4%