We establish long-range order for the hard-core model on a finite, regular bipartite graph above a threshold fugacity given in terms of expansion parameters of the graph. The result applies to the $d$-dimensional hypercube graph and, more generally, to $d$-dimensional discrete tori of fixed side length, proving long-range order at fugacities $λ\geΩ(\frac{\log d}{d})$. Furthermore, we use reflection positivity to transfer the result to the lattice $\mathbb{Z}^{d}$, verifying the long-standing belief that its critical fugacity is of the form $d^{-1+o(1)}$ as $d\to\infty$.
Alex: Welcome to another episode of ResearchPod. Sam, I've been reading about some work on particle models in high dimensions—what's this paper we're diving into today?
Sam: It's a paper by Daniel Hadas and Ron Peled called "On the critical fugacity of the hard-core model on regular bipartite graphs." They study particles placed on a graph's vertices, but no two can be on connected spots—like tokens on a board game grid that can't touch neighbors. The key question: at what fugacity λ do typical setups show long-range order, mostly filling one side of a two-part board?
Alex: So it's asking when these setups shift from random and sparse to organized, favoring one half of the board?
Sam: Yes. Bipartite graphs split into two equal sides with links only across, like a checkerboard where black squares connect only to white. Low λ scatters particles evenly; high λ clusters them on one side because that's the maximum packing without neighbors touching. The paper proves order happens at λ greater than about log d over d in high dimensions d.
Alex: Trees predict order later than lattices—why the difference?
Sam: Trees overestimate sparseness because they branch without cycles, so particles stay spread out longer. Lattices like Z^d, a d-dimensional grid, organize sooner around 1/d because local crowding forces choices. Their proof shows balanced setups—particles split evenly between sides—are unlikely above log d over d.
Alex: They use a free energy argument—like a cost-benefit for arrangements?
Sam: Precisely. Free energy combines entropy, which measures how random an arrangement is, with a reward for more particles scaled by log λ. Likely setups sit at the peak. Balanced measures have free energy below half the max reward minus a penalty from interface energy—boundaries where sides mix unevenly, like edges of scattered islands.
Alex: How do they bound that penalty in high d?
Sam: Global expansion means it's hard to cut the graph into uneven parts—small sets on one side have many links to the other. For the bound, they sample a thin layer of sites on the low-density side. This reveals rough structure: roughness creates surface tension, a gain from local reward drops, while tree-like paths encode details efficiently, keeping the entropy loss small. The penalty exceeds log d over d.
Alex: Like peeking just enough to spot costly boundaries. They check on tori, these wrapped grids?
Sam: Yes, Z^d_L tori with good expansion. For λ above constant times log d over d, disorder probability drops exponentially. Binomial tails control particle count deviations, proving order with magnetization away from zero. Reflection positivity lifts this to infinite Z^d.
Alex: How does sparse exposure create surface tension?
Sam: Sample thin A on the low-density even side and denser B on odd. Free energy splits: conditioning even on odd, odd in B given A's coarse map—empty neighbors or majority—plus entropy cost of the map. The first two gain where the map varies across edges; local reward drops add up to penalize total roughness.
Alex: And tree paths for entropy?
Sam: Random tree-like subgraphs connect neighborhoods sparsely. Paths encode map differences—each flip reveals a bit. Entropy stays order of expected roughness times logs. Gain outweighs loss above log d over d. Global expansion ties balance to large roughness via Cheeger constant.
Alex: For infinite grids, they handle rare bad events?
Sam: Tori confirm expansion via dominating trees—guards covering all spots minimally. Parameters bound disorder exponentially. Chessboard Peierls contours around 3^d cubes where order flips, like mismatched patches. Reflection positivity bounds joint probabilities if average bad-event cost is below half max free energy. Tori order makes it tiny; Peierls bounds distant disagreements. This yields two distinct Gibbs measures.
Alex: Core result?
Sam: Long-range order at λ at least constant times log d over d, pinning the upper threshold to d^{-1 + o(1)}—matching tree predictions up to logs, closing a 20-year gap.
Alex: Solid step clarifying high-d order. Open questions?
Sam: Relies on expansion; log d above conjectured 1/d. Doesn't settle if thresholds coincide or exact constant.
Alex: Thanks, Sam.
Sam: My pleasure, Alex.