The balanced six-vertex height and its Gaussian limit

Source and verification

This article explains family 225 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

Overview

The six-vertex ice rule produces a random height surface. Family 225 claims a Gaussian free field limit throughout the balanced critical range and gives the exact covariance multiplier. The height convention and the order of thermodynamic limits are essential parts of the result.

🏷️ Preliminaries: arrows become a height

Assign an arrow to every square-lattice edge, with exactly two arrows entering and two leaving each vertex. There are six allowed local configurations. In the balanced model the four configurations of types $a,b$ have weight one, and the two remaining configurations have weight $c$, with $0<c\le 2$. A finite configuration is weighted by the product of its vertex weights.

Give the face on the left of an arrow height one greater than the face on its right. Walking around a vertex has net height change zero by the ice rule, so this defines a height on the plane up to an additive constant.

The source specifies its plane state using tori of dimensions $M\times L$. The horizontal arrows through each vertical column are balanced, half left and half right. First take $M\to \infty$ at fixed even $L$, and then take $L\to \infty$. This prescription is part of the theorem; an arbitrary boundary condition or slope is not being asserted to give the same state.

🏷️ A field is tested by averages

The plane Gaussian free field $\Gamma$, modulo constants, has Green kernel

$$G(x,y)=-\frac{1}{2\pi }\log |x-y|.$$

For smooth compactly supported functions $f,g$ with integral zero, its covariance is $\iint f(x)G(x,y)g(y)dxdy$. The zero-integral condition removes the arbitrary additive constant. The field is a distribution, so convergence cannot mean convergence of finite point values to ordinary Gaussian random variables with finite variance.

Set $h_{\delta }(x)=h(x/\delta )$ using the source’s unit height jumps and face convention. No vanishing power of $\delta$ multiplies the height: the scaling occurs in space and in testing the distribution against fixed functions.

🏷️ The claimed exact limit

Balanced six-vertex height limit

For every fixed $0<c\le 2$, the prescribed plane state exists and

$$\begin{array}{c}{h}_{\delta }⟹\sigma (c)\Gamma ,\\ \sigma (c{)}^2=\frac{1}{\arcsin (c/2)}.\end{array}$$

The manuscript claims joint Gaussian limits against compactly supported, zero-mass finite-energy signed measures, and tight convergence in its stated negative-regularity function spaces, modulo constants. (OpenAI, 2026)

For a measure $\mu$, finite energy includes absolute integrability of the logarithmic singularity against $|\mu |\otimes |\mu |$. It does not permit cancelling two infinite contributions. Point masses therefore fall outside this finite-energy test class.

The formula gives the squared multiplier. Multiplying a GFF by $\sigma$ multiplies its covariance by ${\sigma }^2$, so confusing these two quantities changes every predicted fluctuation variance.

🏷️ A normalization check

Write $\Delta =1-c^2/2$. Since $c/2\in (0,1]$, the elementary double-angle identity gives

$$\begin{array}{c}\arccos \Delta =2\arcsin (c/2),\\ \sigma (c{)}^2=\frac{2}{\arccos \Delta }.\end{array}$$

For example, at $c=2$ the variance multiplier is $2/\pi$, while at $c=\sqrt{2}$ it is $4/\pi$. These numbers use the Green kernel and unit jumps fixed above. Doubling the microscopic height would multiply the variance factor by four.

For a zero-mass test $f$, the conclusion predicts

$$\begin{array}{c}\mathrm{Var}⟨h_{\delta },f⟩⟶-\frac{\sigma (c{)}^2}{2\pi }\\ \times \iint f(x)\log |x-y|f(y)dxdy,\end{array}$$

under the moment control supplied by the source. Although the kernel has a negative sign, this quadratic form is nonnegative on zero-mass tests, as expected from the inverse Laplacian.

🏷️ How the proposed proof reaches Gaussianity

Row transfer operators propagate the arrow configurations while marked operators record height increments. The manuscript first constructs their thermodynamic matrix elements with uniform tail bounds and decorrelation for widely separated parameters. It then identifies the current covariance and its logarithmic coefficient.

A covariance calculation alone would not prove a Gaussian limit. The source separately derives Wick rules for the limiting currents: higher moments factor into sums over pairings of two-point covariances. These rules identify Gaussian finite-dimensional laws. Smooth-test convergence is then extended to the finite-energy measure class using energy bounds.

Finally, wavelet estimates give tightness in negative-regularity spaces. A distributional limit needs control at every small scale, not just convergence of finitely many chosen averages. The proof treats $c<2$ and the endpoint $c=2$ separately and keeps the torus limits in the stated order.

🏷️ Key steps and proof outlines

There are three distinct identifications: constructing the prescribed plane state, proving Gaussianity, and computing its normalization. The source performs the normalization only after the covariance structure is available; it does not infer a field limit merely from the free-energy curvature (OpenAI, 2026).

Marked row limits retain the microscopic height convention

Finite row operators propagate arrow states, while marked insertions record the unit height changes. The thermodynamic construction controls words containing both rows and marks, with estimates for their tails and for separation of parameters. This is necessary because unmarked partition-function limits do not determine current correlations.

For $c<1$, the source keeps the exactly balanced sector and the prescribed order of torus limits throughout the construction. For $c\ge 1$, where an auxiliary near-balanced state is used, a separate equilibrium identification returns to the balanced law. The two axial descriptions must agree on the same local observables before the angular representation can be used. That representation supplies a constant $\kappa$ and the lattice variogram

$$\mathbb{E}(h(a)-h(0){)}^2=4\kappa \log (1\vee |a|)+O(1),$$

uniformly over lattice displacements. This fixes the coefficient later integrated against neutral tests, but does not yet establish Gaussian higher moments.

Complex analyticity forces Wick’s recursion

The source separates the limiting currents into holomorphic and antiholomorphic parts. Denote a pure holomorphic $k$-point function by $F_k$. The row limits first control it where insertion coordinates are separated in one of the two axial directions. Analytic continuation leaves candidate singularities at true collisions and at artificial grid points $p=\Re z_j+i\Im z_h$ with $j\ne h$.

A quadratic growth bound limits any such pole to order two. To remove an artificial pole, integrate ${\zeta }^{\ell }{F}_k(\zeta ,\dots )$ around a small contour and call the resulting holomorphic spectator functions $m_{\ell }$. If the pole has principal part $A/(\zeta -p{)}^2+B/(\zeta -p)$, residues give

$$m_{\ell }=B{p}^{\ell }+A\ell p^{\ell -1},m_0{m}_2-m_1^2=-A^2.$$

If $A\ne 0$, the additional identity $m_0{m}_3-m_1{m}_2=-2p{A}^2$ expresses $p$ as a quotient of holomorphic functions. That is impossible because ${\partial }_{{\bar{z}}_j}p=1/2$. Hence $A=0$; if $B\ne 0$, $p=m_1/m_0$ gives the same contradiction. The artificial puncture is removable.

At a true collision write $d=z_1-z_j$ and $\xi =(z_1+z_j)/2$. The bounded function $d^2{F}_k$ extends holomorphically across $d=0$. Exchange symmetry makes it even in $d$, and the collision calculation identifies its constant term as $-\kappa F_{k-2}$, independent of $\xi$. Thus there is no simple-pole remainder. Subtract every true double-pole term as a function of $z_1$. The remainder is entire and decays at infinity, so Liouville’s theorem gives

$$F_k(z_1,\dots ,z_k)=-\kappa \sum _{j=2}^k\frac{F_{k-2}(z_2,\dots ,{\hat{z}}_j,\dots ,z_k)}{(z_1-z_j{)}^2}.$$

Together with $F_0=1$ and $F_1=0$, this is exactly Wick’s pairing recursion. The two chiral sectors factor, so mixed contractions vanish. Their pure contractions are $-\kappa /(z-\xi {)}^2$ and its complex conjugate, the derivatives of $K_{\kappa }(z,\xi )=-2\kappa \log |z-\xi |$. This supplies the higher-moment argument missing from a covariance-only computation.

The normalization comes from a twisted seam

Below $c=1$, the source computes a stiffness directly. Insert a small boundary twist $t$ into a finite row Hamiltonian, and distribute it evenly over the links by site rotations. If $e_L(t)$ is the sector ground energy, ${\Omega }_L$ its untwisted normalized ground vector, $E_{\mathrm{exc}}$ the positive excitation operator on ${\Omega }_L^{\perp }$, and ${\mathcal{J}}_L$ the total first link derivative, second-order perturbation theory gives

$$L{e}_L^{\prime \prime }(0)=⟨D_{0,L}⟩-\frac{2}{L}{\Vert E_{\mathrm{exc}}^{-1/2}(I-|{\Omega }_L⟩⟨{\Omega }_L|){\mathcal{J}}_L{\Omega }_L\Vert }^2.$$

Here $D_{0,L}$ is the average second link derivative. Differentiating the finite root-counting equations evaluates the left side. At fixed rapidity cutoff, the terms from the implicit root derivatives combine into a total derivative, leaving an endpoint expression; only after the chain-size limit is the cutoff removed. The stiffness is $2/(\pi -\lambda )$ in the source’s angular parameter, with $\pi -\lambda =\arccos \Delta$. Its current sum rule identifies $4\pi \kappa$ with this stiffness.

For $1\le c<2$, the source instead invokes the stated conditional free-energy identification after establishing a GFF subsequential limit. At $c=2$, a separate loop argument supplies rotational invariance, and the endpoint theorem’s possible subsequential multipliers are all forced to equal $2/\pi$. These distinct routes agree on the unit-jump normalization but have different prerequisites.

Neutrality removes the divergent constant before integration

Let $V_{\delta }(x,y)$ be the lattice variogram at the rounded continuum points. The uniform variogram estimate implies

$$|V_{\delta }(x,y)-4\kappa \log {\delta }^{-1}|\le C(1+|\log |x-y||),x\ne y.$$

This remains valid when two distinct points round to the same face: then $|x-y|=O(\delta )$ absorbs the apparent divergence. A covariance of two increments is one half of a signed sum of four variograms, and the four copies of $4\kappa \log {\delta }^{-1}$ cancel exactly.

Represent a neutral signed measure using differences between its positive and negative Jordan parts. Absolute logarithmic energy makes the four-logarithm bound integrable and rules out atoms. Dominated convergence gives limiting covariances. To identify laws, mollify the measure: the same majorant proves convergence in logarithmic energy, hence

$$\underset{ϵ\downarrow 0}{\lim }\underset{\delta \downarrow 0}{\mathrm{limsup}}\Vert ⟨h_{\delta },\mu -\mu \ast {\varrho }_{ϵ}⟩{\Vert }_{L^2}=0.$$

Smooth-test Gaussian convergence at fixed $ϵ$ then transfers to the original measure. The proof never cancels two divergent signed integrals.

Wavelet tails turn test convergence into norm convergence

Pin the field by subtracting its pairing with one smooth probability density and localize it on a fixed box. A scale-$2^{-j}$ wavelet test normalized as a density has moments bounded by $C_b(1+j{)}^b$. There are $O(2^{2j})$ relevant translates. For $\alpha <\beta <0$, choose $b$ so large that $2+b\beta <0$. Then

$$\mathbb{E}\sum _{j,k}|2^{j\beta }{a}_{j,k}{|}^b\le C_b\sum _j{2}^{(2+b\beta )j}(1+j{)}^b<\infty .$$

This controls a stronger negative-regularity norm. Its high-frequency tail in the desired exponent $\alpha$ has the additional vanishing factor $2^{(\alpha -\beta )J}$ beyond level $J$. Finitely many remaining coefficients give compactness. Restricting a globally localized field uses the quotient definition of the restriction norm, so no boundary extension theorem for the open set is assumed. Separate estimates handle the source’s outward test convention and intrinsic Sobolev duals. Smooth-test identification then fixes every subsequential law and gives full-mesh convergence.

See Also

• Family 216 explains why a height-field limit alone does not identify microscopic spin insertions.
• Family 226 treats loop geometry related to height fluctuations, with a different convergence target.

🏷️ Manuscripts and verification scope

• The Gaussian free field limit of the balanced six-vertex model with variance multiplier 1/arcsin(c/2) — pinned manuscript and companion source.

No family scope page was present at lean/docs/225.md in the pinned snapshot. This does not establish that all related formalization code is absent. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

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