Kozma–Nitzan inequalities: proofs and counterexamples

ChatGPT 5.6 Sol and Claude Fable 5.1, prompted by Ahmed Bou-Rabee

Last updated 5 September 2026

Kozma and Nitzan (arXiv:2401.12397) reduced the statement \(\theta(p_c) = 0\) for Bernoulli bond percolation on \(\mathbb{Z}^d\), \(d \ge 2\), to an inequality on finite graphs, their Conjecture 3, and posed four further conjectures and four questions about connection probabilities in finite graphs: Conjectures 1, 2, 4 and 6 and Questions 5, 7, 8 and 9. Anthropic's Lean 4 / Mathlib proof of \(\theta(p_c) = 0\) for Bernoulli bond percolation on \(\mathbb{Z}^d\) in all dimensions \(d \ge 2\) (written by Claude under the direction of Justin Leder) proves Conjecture 3. It does not claim Conjecture 1, but Conjecture 1 is a relatively easy consequence of its intermediate first-contact inequality (GEN). That same result gives the stronger version retaining \(o\leftrightarrow A\) on the left, as explained below.

Here the remaining conjectures and questions of the paper are proved on top of that Lean proof: Conjectures 1, 2, 4 and 6 and Questions 5, 7 and 9, each in a stronger form than stated. Question 8 fails under the every-minimizer reading; the distinction matters when minimizers tie. The proofs are in the summary (PDF); the Lean files are in KNConjectures/ in the separate public repository.

In a companion page we prove, and check in Lean, that \(\theta(p_c) = 0\) for site percolation on \(\mathbb{Z}^d\) in every dimension \(d \ge 3\), from a single inequality about a finite family of independent hyperedges that yields bond percolation as well.

Notation

Throughout, \(G\) is a finite graph in which each edge \(e\) is open independently with probability \(p_e \in [0,1]\). For vertices \(x\) and \(y\), \(x \leftrightarrow y\) means that \(x\) and \(y\) are joined by a path of open edges, and \(C_x\) is the set of vertices joined to \(x\), the open cluster of \(x\). For a set \(A\) of vertices, \(o \leftrightarrow A\) means that \(o \leftrightarrow x\) for some \(x \in A\), and \(o \nleftrightarrow A\) is its negation. A comma inside a probability means intersection, \(\mathbb{P}(E, F) = \mathbb{P}(E \cap F)\), and \(\mathbb{E}[X;\, E]\) is the expectation of \(X\) on the event \(E\). Increasing functions \(F\) of vertex sets are those with \(F(K) \le F(L)\) whenever \(K \subseteq L\). All minima and convex combinations below use a nonempty set \(A\). Page numbers refer to the paper.

Conjecture 1 from the Lean proof of Conjecture 3

Fix \(A\), \(o\) and \(b\), and order the vertices of \(A\) so that \(\mathbb{P}(x \leftrightarrow b)\) is nondecreasing. Inequality (GEN) in the preceding proof, applied to \(F(K)=\mathbf1_{\{b\in K\}}\), gives \[ \sum_{x \in A} \mathbb{P}\bigl(o \leftrightarrow x \text{ and } o \nleftrightarrow y \text{ for every } y \in A \text{ before } x\bigr)\, \mathbb{P}(x \leftrightarrow b) \;\le\; \mathbb{P}(o \leftrightarrow b,\ o \leftrightarrow A). \] The events on the left partition \(\{o \leftrightarrow A\}\), and every factor \(\mathbb{P}(x \leftrightarrow b)\) is at least the minimum over \(A\), so \[ \mathbb{P}(o \leftrightarrow A)\, \min_{x \in A} \mathbb{P}(x \leftrightarrow b) \;\le\; \mathbb{P}(o \leftrightarrow b,\ o \leftrightarrow A). \] Dropping \(o\leftrightarrow A\) on the right gives Conjecture 1 (p. 3). Thus the stronger joint-event inequality already follows from the preceding bond proof, for every finite graph and every choice of edge probabilities.

The results

In the paper, Conjectures 2 and 4 are proved when \(A\) has two vertices (Theorems 1 and 7), when \(A\) has three vertices under extra hypotheses (Theorems 2, 3 and 11), when \(o\) is isolated in \(G \setminus A\) or joined only to \(A\) and to one further vertex (Theorems 4 and 5), and for the functions \(f = \mathbf{1}\{|C| \ge k\}\) with \(k \le 4\) on every graph (Theorem 9).

The method

The proofs run the argument of the Lean proof of Conjecture 3 with two changes. First, the clusters of the vertices of \(A\) are conditioned on missing a given set \(Y\) of vertices: if the vertices of \(A\) are ordered so that \(\mathbb{E}[F(C_x) \mid x \nleftrightarrow Y]\) is nondecreasing, then \[ \sum_{x \in A} \mathbb{P}\bigl(o \nleftrightarrow Y,\; o \leftrightarrow x,\; o \nleftrightarrow y \text{ for every } y \in A \text{ before } x\bigr)\, \mathbb{E}[F(C_x) \mid x \nleftrightarrow Y] \;\le\; \mathbb{E}[F(C_o);\; o \nleftrightarrow Y,\; o \leftrightarrow A]. \] Second, \(F(K)\) is replaced by \(F(K)\) minus the expectation of \(F\) at the cluster of \(a\) in the graph with every edge at a vertex of \(K\) removed; this new function is increasing in \(K\). Taking \(Y = \{a\}\) for the vertex \(a \in A\) minimising \(\mathbb{E}\,F(C_x)\) makes every conditional expectation on the left nonnegative, and the inequality becomes Conjecture 4 with the minimiser specified. The other conjectures and questions follow from that statement. For Conjecture 6 and Question 9 the cluster of \(o\) is replaced by the union of the clusters of a set of vertices: the union of the clusters of \(v\) and \(w\) does not depend on whether the edge \(\{v, w\}\) is open, and the cluster of \(o\) is the union of the clusters of \(o\) and of its open neighbours in the graph with the edges at \(o\) removed.

Implications

Files

kn_summary.pdf The summary: statements and proofs of all the results above.
KN-Conjectures The Lean modules, build instructions, and verification results. The principal declarations use only propext, Classical.choice, and Quot.sound, with no sorry.

How to check

git clone https://github.com/nitromannitol/percolation-after-anthropic.git
cd percolation-after-anthropic
bash scripts/bootstrap.sh
lake env lean scripts/Audit.lean

The repository pins Lean 4.32.0, Mathlib, and Anthropic's original proof. The default Lake build checks both parts, including all their retained Lean dependencies. See the README for the additional declaration checks and regression tests.

All of this was done autonomously by ChatGPT 5.6 Sol and Claude Fable 5.1, prompted by Ahmed Bou-Rabee, with minimal human intervention. Released under the Apache License 2.0, the licence of Anthropic's Lean proof.