Six-vertex heights and the Gaussian free field across the balanced critical range

← All digests

Family 225 · Probability and statistical mechanics
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-11

1. The question

Place an arrow on every edge of the square lattice, requiring exactly two arrows to enter and two to leave each vertex. There are six allowed local arrangements, which gives the six-vertex model its name. Although the rule is local, it couples arrows over long distances.

These arrows can also be read as the slopes of an integer-valued random surface. What does that surface look like when the lattice spacing becomes very small? The expected answer in the critical regime is the Gaussian free field, the two-dimensional analogue of a Gaussian random process whose correlations are governed by a Green function.

Family 225 contains one manuscript claiming this limit for the balanced plane model with weights $a=b=1$ and every fixed $0<c\leq2$. It also gives the exact fluctuation coefficient. The choice of infinite-volume state, the units of height, and the meaning of field convergence are essential parts of the result.

2. Background and essential definitions

From arrows to heights

Assign weight $c$ to the two vertex types whose horizontal arrows both point inward or both point outward, and weight one to the other four types. The probability of a finite configuration is proportional to the product of its vertex weights. Changing $c$ changes the interaction while preserving the two-in, two-out rule.

Give the face to the left of an arrow height one more than the face to its right. The ice rule ensures that following different routes between two faces gives the same height difference. Absolute height remains arbitrary: adding a constant changes no arrows.

The manuscript selects its plane state through rectangular tori of horizontal size $M$ and even vertical size $L$. In every vertical column of horizontal edges, it conditions on exactly $L/2$ left-pointing and $L/2$ right-pointing arrows. It first sends $M\to\infty$, then even $L\to\infty$, in the sense of probabilities of every fixed finite arrow pattern. This defines the balanced plane law $\mathbb P_c$.

This is a prescribed thermodynamic construction. It should not be replaced by an arbitrary boundary condition or a simultaneous torus limit without an additional result.

At physical mesh $\delta$, the rescaled surface is

$$h_\delta(x)=h(x/\delta),$$

with the discrete height extended constantly over each face. There is no extra power of $\delta$ multiplying the height. What converges is the surface as a generalized function, tested over regions, rather than its value at each point.

What a Gaussian free field means

The plane Gaussian free field $\Gamma$, modulo additive constants, is specified by the kernel

$$G(x,y)=-\frac1{2\pi}\log|x-y|.$$

For smooth compactly supported functions $f,g$ with integral zero, its averaged values are centered jointly Gaussian and satisfy

$$\operatorname{Cov}(\langle\Gamma,f\rangle,\langle\Gamma,g\rangle) =\iint G(x,y)f(x)g(y)\,dx\,dy.$$

The zero-integral condition removes the arbitrary additive constant. One example is the difference between two smooth regional averages. The logarithmic singularity means that a point value is not an ordinary finite-variance Gaussian random variable.

The theorem also allows compactly supported signed measures $\mu$ of total mass zero when

$$\iint \log^+\!\frac1{|x-y|}\,|\mu|(dx)|\mu|(dy)<\infty,$$

where $\log^+t=\max\{\log t,0\}$ and $|\mu|$ is total variation. This finite-energy requirement controls the singularity without canceling infinite positive and negative contributions. Smooth zero-mass densities qualify; a difference of two point masses does not.

3. Development history and previous partial results

Lieb’s solution of square ice determined the exact entropy at $c=1$, and the wider exact-solution program made the six-vertex model a central example of an interacting system with accessible algebraic structure. The Coulomb-gas approach predicted Gaussian height fluctuations and their parameter-dependent coupling. The manuscript credits the work of Di Francesco, Saleur, and Zuber for the corresponding normalization after converting height units.

At the free-fermion value $c=\sqrt2$, a relation to dimers supplies determinant formulas. Kenyon’s Gaussian-free-field theorem for domino tilings established a foundational example of this kind of scaling limit. Near that special point, interacting-fermion and dimer methods gave correlation asymptotics, logarithmic variance, and, for appropriate models, Gaussian field limits.

A broader probabilistic advance was logarithmic delocalization of six-vertex heights by Duminil-Copin, Karrila, Manolescu, and Oulamara. Logarithmic variance identifies the size of fluctuations, but a non-Gaussian field can have the same covariance growth. A full GFF theorem must identify all joint fluctuations as well.

The particularly relevant predecessor is the March 2026 GFF convergence theorem of Duminil-Copin, Kozlowski, Lammers, and Manolescu, treating $\sqrt3\leq c\leq2$ in the isotropic model. It includes the same exact multiplier in the stated height convention. The new manuscript explicitly builds on this work: it uses its conditional amplitude theorem for $1\leq c<2$, and its endpoint argument after supplying the required rotation input at two.

There is also related finite-domain progress. The October 2026 paper of Avérous, Duminil-Copin, He, Lammers, and Manolescu states Dirichlet GFF convergence with piecewise constant boundary data in parameter ranges including those near $\Delta=0$ and $-1\leq\Delta\leq-1/2$. That addresses boundary-value problems distinct from this family’s full-range balanced plane statement. It prevents interpreting the plane theorem as the whole current story of six-vertex conformal limits.

4. What the new manuscripts claim

A full-range plane field with a fixed coefficient

For every fixed $c\in(0,2]$, the manuscript claims that the iterated balanced-torus plane state exists and that

$$h_\delta\Longrightarrow\sigma(c)\Gamma, \qquad \sigma(c)^2=\frac1{\arcsin(c/2)}.$$

Equivalently, with the usual parameter $\Delta=1-c^2/2$,

$$\sigma(c)^2=\frac2{\arccos\Delta}.$$

The displayed number is the variance multiplier, not the multiplier of the field itself. The latter is its positive square root. With the source’s unit height jumps and Green kernel, some examples are

Parameter	Model point	Variance multiplier
$c=1$	Square ice	$6/\pi$
$c=\sqrt2$	Free-fermion point	$4/\pi$
$c=2$	Endpoint	$2/\pi$

For any finite collection of zero-mass finite-energy measures, their integrals against $h_\delta$ converge jointly to a centered Gaussian vector. Its covariance is $\sigma(c)^2$ times the corresponding Green-kernel energy. This is a statement about the whole joint law, not only its second moments.

Convergence as a rough field

The manuscript also claims tightness and convergence in negative-regularity distribution spaces on every bounded open set. For example, it gives convergence in restriction Hölder–Besov spaces of regularity $\alpha$ for $-1<\alpha<0$, together with the stated Besov and Sobolev variants.

Negative regularity is a way to measure generalized functions too rough to evaluate pointwise. Convergence in such a space controls a large family of smoothed observations simultaneously. No regularity of the open set’s boundary is required, because the restriction spaces inherit norms from the whole plane.

Constants can be removed by choosing a smooth compactly supported density $\rho$ of integral one and replacing a field $X$ by $X-\langle X,\rho\rangle$. The same convention must be used for the discrete field and the limit. This pins an average; it does not assert a finite pointwise value of the GFF.

Separated height increments

A further claim concerns products of increments

$$\prod_{i=1}^k\bigl(h_\delta(u_i')-h_\delta(u_i)\bigr),$$

when the endpoint sets of different increments are disjoint. Their expectations converge to the Gaussian pairing rule. For two increments the limiting covariance is $\sigma(c)^2 C_{12}$, where

$$C_{ij}=G(u_i',u_j')-G(u_i',u_j) -G(u_i,u_j')+G(u_i,u_j).$$

For an even number of increments, sum the products of these covariances over all pairings; for an odd number, the limit is zero. Convergence is uniform on compact families of separated endpoint configurations.

This does not say that the variance of one fixed-endpoint increment stays finite as the mesh vanishes. Repeating the same increment would violate the separation condition. The source’s logarithmic variogram describes the growing variance, while the displayed mixed moments isolate finite correlations between distinct increments.

5. Why the advance matters

Importance: 5/5. The six-vertex model is a central interacting height model, and the claim covers the entire stated balanced critical interval rather than only an exactly determinantal point, a perturbative neighborhood, or a previously treated subinterval.

The exact multiplier matters as much as the Gaussian shape. Different microscopic interactions can lead to the same type of continuum field while retaining different fluctuation strengths. The formula specifies precisely how the parameter survives in the limit.

The field theorem also provides information used by related loop and interface problems, including the endpoint discussion in family 223. A height limit does not automatically identify all level curves, but it supplies a major ingredient for those further arguments.

The rating reflects the breadth and centrality of the claimed full-range plane result, while crediting the substantial pre-existing GFF and amplitude theory on which parts of it rely.

6. The main idea

The model has exact finite-dimensional row transfer matrices: operators that add one row of vertices and encode its weights. Bethe-ansatz identities express selected scalar products and marked-row quantities through determinants. A mark records a local height increment.

Exact finite formulas alone do not give a continuum theorem. The manuscript develops estimates allowing row length to grow, including control of root tails and simultaneous large parameters in determinant expressions. These estimates produce a plane representation in which angular transfer operators have a tractable spectrum.

The long-distance height fluctuations are then organized into two current components, corresponding to holomorphic and antiholomorphic behavior. Contributions outside the leading modes are damped. The surviving current correlations obey a Wick recursion: higher moments reduce to sums of pairwise contractions, which is the characteristic structure of a Gaussian field.

Identifying Gaussianity still leaves an unknown multiplicative coefficient. For $c\geq1$, the manuscript uses the earlier conditional normalization theorem after establishing its required convergence. For $c<1$, it compares two independently derived bounds: a stiffness calculation controlling the response to a small slope, and a bound from a staggered vacuum overlap. Their stated agreement fixes the coefficient in physical height units.

Moment bounds and removal of smoothing extend the current calculation to ordinary field averages, signed finite-energy measures, and rough distribution spaces. At $c=2$, the argument is separate: fair orientations of associated FK loops supply rotation information, after which the earlier endpoint GFF framework is used.

7. Scope, limitations, and open questions

The theorem concerns one balanced, zero-slope plane state and one specified order of thermodynamic limits. It does not assert uniqueness among all Gibbs states, cover every slope, or evaluate arbitrary boundary-conditioned finite domains.

The parameter is fixed before the mesh limit. The coefficient diverges as $c\downarrow0$, and the manuscript states no uniform estimate in that limit. The endpoint $c=0$ and the regime $c>2$ are outside its assertion. Anisotropic weights also require their own statement and embedding conventions.

Field convergence does not imply pointwise convergence or immediate convergence of contour ensembles. Those observables need additional control. Conversely, the separated-increment theorem is meaningful even though point evaluations of the limiting GFF are not ordinary random variables.

Natural directions include quantitative convergence rates and the relation among different slope and boundary regimes. Existing finite-domain results should be read alongside those questions, rather than treating every boundary problem as still untouched. Throughout, the manuscript’s mathematical claims are reported as stated.

8. Sources and further reading

The source’s introduction gives a detailed account of the exact-solution, dimer, delocalization, and GFF predecessors. The March and October 2026 primary papers linked above are especially relevant for understanding which parts of the landscape are already covered by related work.

For an introduction to the continuum object, see Sheffield, Gaussian free fields for mathematicians. The main normalization in this digest is fixed by the unit height jumps and Green kernel displayed in section 2.

Source manuscripts

• M1. OpenAI, The Gaussian free field limit of the balanced six-vertex model with variance multiplier 1/arcsin(c/2) (September 23, 2026).

Sources consulted on 11 October 2026 at repository revision fd4aeeb. Download the manuscript citations.