A quadratic ancestor inequality for spanning trees

Impact Assessment: "A quadratic ancestor inequality for spanning trees" (G. Cai)

Core Contribution

The paper proves a clean, general inequality for weighted spanning trees on finite graphs. For any root r and any set B, the matrix H_B H_B^T − J_B is positive semidefinite. Here H_B(x) is the probability that the root path of x hits B, and J_B(x,y) is the probability that x, y, r lie in three distinct components of T∖B. A corollary is P(LCA(X,Y)∈B) ≤ P(γ_X∩B≠∅)², with optimal constant 1. The proof uses complete log-concavity of matroid polynomials (Anari–Liu–Oveis Gharan–Vinzant; Brändén–Huh). The author builds a polynomial Q_B whose Hessian of log at 0 is exactly −(H H^T − J), and the construction (direct sum of a graphic and a free matroid, dualization, coefficient extraction) is clever and short.

Two applications follow:

1. Benjamini–Kozma conjecture. For any finite transitive graph with n vertices, E d_T(X,Y) ≥ √n − 1 in the UST. This resolves a conjecture from 2005 in full generality, in contrast to earlier √n bounds that needed high-dimensional or model-specific assumptions (Peres–Revelle, Schweinsberg). The reduction uses the massive spanning forest, transitivity (LCA is uniform), and m→0.

2. 3D LERW growth exponent β ≥ 3/2. This improves the long-standing β > 1 (and the recent β > 1.0001 of Liu) to a substantial, plausibly near-sharp value. The numerical value of β is believed to be about 1.62. The argument combines the inequality, a second-moment local connectivity estimate (via Wilson's algorithm and the Lyons–Peres–Schramm inequality), determinantal monotone coupling, and the known one-point estimates.

Further consequences include:

Methodological Rigor

The proofs appear correct and tightly argued. The key steps are verifiable:

The applications rely on a series of standard but nontrivial tools: Wilson's algorithm, heat-kernel bounds, determinantal domination, and Li–Shiraishi/Shiraishi estimates. The 3D LERW argument is elegant. The inequality gives a|B_R|² ≤ |B_KR|·A_m², so A ≳ R^{3/2}, and this works because the local connectivity at scale R is bounded below uniformly. Some details are terse (for example the infinite-volume passage and the coupling), but they are standard. The paper is a pure theory paper with no numerics. The sharpness example for the constant is given. The AI-assistance statement is unusual, but the author takes responsibility, and the proofs are checkable.

Potential Impact

Timeliness & Relevance

LERW exponent bounds in 3D have been stuck near trivial for years. Liu's 2026 paper only reached 1.0001, so a jump to 3/2 is remarkable. The log-concave polynomial methods are a hot area, and applying them to spanning tree geometry is timely. The AI-assisted proof development is also of metascientific interest, and this paper may be cited as an example.

Strengths

Limitations

Overall

This is a high-impact probability paper. A strikingly compact new inequality yields a long-standing conjecture and a notable exponent improvement. It is likely to be widely cited in the UST and LERW literature and to inspire further use of Lorentzian polynomials in probability.