Smoothness of multinomial dimer limit shapes in dimensions $d>2$ (Kenyon-Wolfram, Question 10.2)

For the multinomial dimer model -- the large-$N$ limit of the dimer model on $N$-fold blow-up graphs introduced by Kenyon and Pohoata -- Kenyon and Wolfram proved a variational principle in every dimension $d$: as the multiplicity $N$ and then the graph size grow, random $N$-dimer covers concentrate on a limit shape, a divergence-free flow minimizing a surface-tension functional, equivalently encoded by a limiting gauge function $H$ solving the degenerate-elliptic gauge PDE $\mathrm{div}(\nabla F(\nabla H))=0$. They proved that for flexible boundary data these limit shapes have no facets in any dimension, and that in two dimensions they are smooth, by combining the no-facets theorem with elliptic regularity for gradient variational problems. The two-dimensional argument exploits a duality special to the plane, and the authors pose as an open question (Question 10.2) whether multinomial limit shapes are smooth in any dimension; a positive answer would remove the $C^1$ hypothesis of their critical-gauge scaling-limit theorem in all dimensions. It remains open at the source's full generality, for every $d\ge 3$, every periodic bipartite lattice, and every flexible boundary condition.

Background

The dimer model is the statistical mechanics of perfect matchings (pairings of adjacent vertices, called dimers) of a graph; on a periodic bipartite lattice $\Lambda\subset\mathbb{R}^d$ it represents random tilings of a region by lozenges or dominoes. It is exactly solvable in two dimensions via Kasteleyn determinants (the signed adjacency-matrix determinant that counts the perfect matchings of a planar graph), but in dimensions $d\ge 3$ almost nothing explicit is known.

An $N$-dimer cover of a bipartite graph $G$ is a multiset of edges in which every vertex is contained in exactly $N$ edges (counted with multiplicity); equivalently, it is a single perfect matching of the $N$-fold blow-up graph $G_N$, obtained by replacing every vertex by $N$ copies and every edge by a complete bipartite graph $K_{N,N}$. The multinomial dimer measure of Kenyon and Pohoata is the probability measure on $N$-dimer covers obtained by projecting the uniform measure on perfect matchings of $G_N$: each $N$-dimer cover is weighted by the number of ways it refines into $N$ single dimer covers. This introduces a multiplicity parameter $N$, in analogy with the large-$N$ limits of lattice gauge theory.

Kenyon and Wolfram (2025) study the iterated limit $N\to\infty$ (then graph size $n\to\infty$) of this model on periodic bipartite graphs in any dimension $d$, and prove a variational principle: random samples concentrate on a limit shape. In the continuum, the boundary data is a boundary asymptotic flow $b$ on $\partial R$ -- the flux through the boundary of a divergence-free flow defined in a neighborhood of $R$ -- an asymptotic flow is a measurable vector field $\omega$ on $R$ that is divergence-free as a distribution, takes values in the Newton polytope $\mathcal{N}(\Lambda)$ (the compact convex set of allowed local tile slopes) and is compatible with $b$, and the limit shape is the unique minimizer of the functional $\int_R \sigma(\omega(x))\,dx$, where the surface tension $\sigma$ is a strictly convex function on $\mathcal{N}(\Lambda)$. For any lattice $\Lambda$ the Legendre dual $F$ of $\sigma$ has an explicit closed form -- for $\Lambda=\mathbb{Z}^3$, $F(\alpha)=\log\sum_{i=1}^{3}(e^{\alpha_i/2}+e^{-\alpha_i/2})$ -- making this one of the first statistical-mechanics models in $d\ge 3$ whose limit shapes can be computed explicitly (for example for the Aztec cuboid in the body-centered cubic lattice, the three-dimensional analogue of the Aztec diamond, the classical random-domino-tiling region of $\mathbb{Z}^2$).

A facet of a limit shape is a region where $\omega$ takes values on the boundary $\partial\mathcal{N}(\Lambda)$ of the allowed slopes, so the tiling is frozen there. For the standard ($N=1$) dimer model, facets are typical in two dimensions and expected in three, where even their existence is conjectural. The pair $(R,b)$ is called flexible if around every point of $R^\circ$ some asymptotic flow extending $b$ takes values in the interior of $\mathcal{N}(\Lambda)$ almost everywhere on a neighborhood -- by a lemma of the paper, a single such flow working on all of $R^\circ$ suffices -- a mild condition ruling out boundary data that force saturation. Kenyon and Wolfram prove a no-facets theorem (Theorem 7.8): in any dimension, for any $\Lambda$ and any flexible $(R,b)$, the $N\to\infty$ multinomial limit shape has no facets. In two dimensions they combine this with classical elliptic regularity for gradient variational problems (variational problems of the form $u\mapsto\int f(\nabla u)\,dx$ over scalar potentials $u$) to conclude (Corollary 7.9) that the limit shape height function $h$ (the two-dimensional potential with $\nabla h=(-\omega_2,\omega_1)$) is smooth, hence that $\omega$ is a smooth vector field.

Two further structures make the higher-dimensional regularity question central. First, the Euler-Lagrange equations for the minimizer can be rewritten as a gradient variational problem: the gauge flow $\alpha=\nabla\sigma(\omega)$ is itself a gradient $\nabla H$ of a limiting gauge function $H$, which is a weak solution of the gauge PDE

$\mathrm{div}(\nabla F(\nabla H)) = 0$

with suitable boundary data; $H$ determines $\omega$ and vice versa. Unlike the two-dimensional height function, which is Lipschitz, $H$ has no a priori regularity in general dimension. Second, $H$ arises as the scaling limit of the critical gauge: on a fixed finite graph the large-$N$ behavior is captured by critical gauge functions $f,g$ -- positive weights on the white and black vertices whose associated edge weights $c(e)=f(u)g(v)$ sum to one at every vertex (the critical gauge equations); they are unique up to scaling and computable by the Sinkhorn matrix-scaling iteration, which alternately rescales rows and columns of the bipartite adjacency matrix until it is doubly stochastic -- and Theorem 8.1 shows that if $H$ is $C^1$ and solves the gauge PDE, then $(1/n)\log g_n\to H$ uniformly; the paper's perturbative estimates then make $H$ real analytic.

The obstruction in $d\ge 3$ is that $F$ is strictly convex but not uniformly convex on all of $\mathbb{R}^d$, so the classical De Giorgi-Nash-Moser elliptic regularity theory (the standard interior-regularity machinery for uniformly elliptic equations) does not apply to the gauge PDE. The two-dimensional smoothness proof lifts to $H$ only because in the plane the dual equation (the height-function gradient variational problem) is itself a gradient variational problem -- a duality special to two dimensions -- and in $d\ge 3$ the flow $\omega$ is a priori merely a measurable divergence-free vector field, since no height function exists. Whether the $C^1$ condition, and hence smoothness (in fact real analyticity), holds in higher dimensions is Question 10.2 of the paper, and the authors state that they expect a positive answer.

Problem Statement

Is it true that multinomial dimer limit shapes are smooth in any dimension? Concretely, for a periodic bipartite lattice $\Lambda\subset\mathbb{R}^d$ with $d\ge 3$ and a flexible region-with-boundary pair $(R,b)$, is the $N\to\infty$ multinomial dimer limit shape -- the unique minimizer of $\int_R\sigma(\omega)\,dx$ over divergence-free asymptotic flows extending $b$ -- smooth on the interior of $R$? Equivalently: is the limit shape flow $\omega$ continuous on $R^\circ$, and is the limiting gauge function $H$ (the weak solution of $\mathrm{div}(\nabla F(\nabla H))=0$ determining $\omega$) of class $C^1$, so that the $C^1$ hypothesis of the critical-gauge scaling-limit theorem (Theorem 8.1 of Kenyon-Wolfram) is satisfied in any dimension? (For this model the source shows these formulations are equivalent, and that $C^1$ of $H$ upgrades to real analyticity by perturbative estimates.) The question is posed at the source's full generality -- every $d\ge 3$, every periodic bipartite lattice, every flexible $(R,b)$ -- and the authors expect a positive answer. The two-dimensional case is resolved affirmatively by the no-facets theorem plus elliptic regularity (Corollary 7.9). A complete answer is either a proof of smoothness in all dimensions, or an explicit flexible $(R,b)$ in some dimension $d\ge 3$ whose multinomial limit shape fails to be smooth.

References

1. Richard Kenyon and Catherine Wolfram, "The multinomial dimer model" (v1 June 2025, v3 February 2026), arXiv:2506.12171, DOI 10.48550/arxiv.2506.12171, https://arxiv.org/abs/2506.12171
2. Richard Kenyon and Cosmin Pohoata, "The multinomial tiling model" (2021), arXiv:2104.03205; The Annals of Probability 50(5), 2022, DOI 10.1214/22-aop1575, https://arxiv.org/abs/2104.03205
3. Henry Cohn, Richard Kenyon, and James Propp, "A variational principle for domino tilings", Journal of the American Mathematical Society 14(2):297-346, 2001, DOI 10.1090/S0894-0347-00-00355-6, https://doi.org/10.1090/S0894-0347-00-00355-6
4. Richard Kenyon and István Prause, "Gradient variational problems in R^2" (v1 2020), arXiv:2006.01219; Duke Mathematical Journal 171(14), 2022, DOI 10.1215/00127094-2022-0036, https://arxiv.org/abs/2006.01219
5. Nishant Chandgotia, Scott Sheffield, and Catherine Wolfram, "Large deviations for the 3D dimer model" (v1 April 2023, v3 May 2026), arXiv:2304.08468, https://arxiv.org/abs/2304.08468
6. Lorenzo Taggi and Wei Wu, "Mermin-Wagner theorem for dimers, monomer double-dimers, and spatial random permutations", Communications in Mathematical Physics 407(8), 2026, DOI 10.1007/s00220-026-05626-0, arXiv:2312.16911, https://arxiv.org/abs/2312.16911
7. Daniel C. Douglas, Richard Kenyon, Nicholas Ovenhouse, Samuel Panitch, and Sri Tata, "A quantum N-dimer model" (v1 October 2025, v3 June 2026), arXiv:2510.07543, https://arxiv.org/abs/2510.07543
8. Christina Meng, "Asymptotics of the partial n-fold dimer model" (v1 December 2024, v2 September 2025), arXiv:2412.04607, https://arxiv.org/abs/2412.04607
9. Tomas Berggren, Alexei Borodin, and Terrence George, "Dimer models on astroidal zig-zag graphs" (May 2026), arXiv:2605.03896, https://arxiv.org/abs/2605.03896
10. Jan de Gier, Richard Kenyon, Michael Wheeler, and Yuyang Zhou, "The asymmetric five vertex model on a rectangle" (June 2026), arXiv:2606.22004, https://arxiv.org/abs/2606.22004
11. Nikolai Kuchumov, "Limit shapes and harmonic tricks" (v1 March 2026, v2 May 2026), arXiv:2603.21255, https://arxiv.org/abs/2603.21255