The hard-core model places particles on a graph with no two neighbors occupied. The partition function counts valid configurations weighted by activity. Understanding where this function has zeros in the complex activity plane controls everything: approximation algorithms, phase transitions, sampling efficiency.
Peters, Reppekus, and Regts connect two seemingly separate properties. Correlation decay (very strong spatial mixing, VSSM) describes how fixing a distant vertex's state has vanishing influence on a target vertex's occupation probability — a local, probabilistic property. Zero-freeness of the partition function is a global, analytic property in the complex plane. They prove VSSM at a real parameter implies the partition function has no zeros nearby in the complex plane.
The bridge is a non-autonomous dynamical system of Mobius transformations. Each vertex in the graph's computation tree contributes a Mobius map, and the composition of these maps tracks how boundary conditions propagate inward. VSSM means the composed maps contract — orbits converge regardless of starting point. This contraction in the real dynamics extends to a neighborhood in the complex plane, and the extension is what keeps the partition function away from zero.
The structural point: a local mixing property (correlations decay) provides a global analytic guarantee (no zeros). The connection runs through dynamics — the same contraction that makes sampling efficient makes the partition function well-behaved. But the paper also proves the converse fails: a weaker mixing variant does not protect against zeros. The shield is specific to very strong mixing; weaker forms leave gaps.
Note (2026-11-11): checked against the full text of the paper (Peters, Reppekus and Regts, arXiv 2603.17858). The main theorem (very strong spatial mixing at a real parameter for a family of graphs implies zero-freeness of the partition function near it for each graph), the Mobius-transformation dynamical system, and the statement that a closely related variant of very strong spatial mixing does not imply zero-freeness are as in the paper. Three points. (1) The essay defines very strong spatial mixing as 'fixing a distant vertex's state has vanishing influence on a target vertex's occupation probability'; that is the ordinary notion of (strong) spatial mixing. The paper introduces very strong spatial mixing as a strengthening of strong spatial mixing, defined informally by weak spatial mixing of the family of self-avoiding-walk trees of the graphs in a family closed under induced subgraphs, and says that on the infinite regular tree it is equivalent to strong spatial mixing. The essay's 'each vertex in the graph's computation tree' is that self-avoiding-walk tree. (2) The essay calls the failure for the weaker variant 'the converse'; the paper's reverse direction is a different statement, earlier work of the last author showing that zero-freeness near a parameter implies strong spatial mixing, and this paper studies the reverse implication of that. (3) The essay omits the paper's consequence that very strong spatial mixing implies spectral independence, which bears on the mixing time of Glauber dynamics.