The joint Ashkin–Teller height and current-cluster limit

Source and verification

This article explains family 233 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

Family 233 claims a joint continuum limit for the critical Ashkin–Teller height and two coupled current-cluster collections. The clusters are identified as recursive local sets of the same Gaussian field. Marginal field convergence would not determine this joint coupling.

🏷️ Preliminaries: two coupled planar systems

The Ashkin–Teller model couples two Ising-type systems. The manuscript fixes parameters

$$\begin{array}{c}J>0,U\le J,\\ \sinh (2J)=e^{-2U}.\end{array}$$

Its planar representation uses a primal spin field and a dual spin field. A primal and a dual spin disagreement cannot cross. Write $t=(1+e^{4U}{)}^{-1/2}$, so $1/2\le t<1$. Admissible spin pairs receive weight $t$ for each disagreement, with fixed primal boundary spins and free dual spins.

Conditional on the spins, a dual disagreement forces the corresponding primal current edge open, and a primal disagreement forces the dual edge open. If neither disagreement occurs, the primal-only, dual-only, and both-closed choices have probabilities $1-t,1-t,2t-1$. The two current supports therefore never cross, but they need not be complementary.

Their connected components are current clusters. The associated signed nesting height records height changes across the nested configuration, with the boundary convention fixed in the source. Both the complete supports and the distinguished cluster touching the wired boundary are retained.

🏷️ Gaussian local sets

Let $h$ be a zero-boundary GFF with covariance

$$G_D=2\pi (-{\Delta }_D{)}^{-1}.$$

This convention has logarithmic singularity $-\log |z-w|$, differing by $2\pi$ from the inverse-Laplacian convention in some other posts.

A local set is a random closed set coupled to the field so that, after revealing it and its harmonic data, the remaining field is an independent zero-boundary GFF in each complementary component. A two-valued local set has harmonic data taking two prescribed constant values. The canonical sets used here are measurable functions of the field, not independently sampled decorations.

Put

$$\begin{array}{c}c=t^{-1},g=\frac{8}{\pi }\arcsin (c/2),\\ a=\frac{\pi \sqrt{g}}{2}.\end{array}$$

The wired boundary cluster is identified with the symmetric two-valued local set with labels $-a,+a$. The free-cluster procedure first exposes CLE₄ holes and then performs the source’s asymmetric two-valued splits recursively. Both procedures use the same $h$.

🏷️ The claimed joint limit

Critical-line field and cluster convergence

For every fixed permitted parameter pair, every bounded Jordan domain, and every admissible polygonal approximation, the manuscript claims joint convergence of the height, both complete cluster collections, and the wired boundary cluster. The height limit, in the source’s normalization, is

$$H_{\delta }⟹\frac{h}{\pi \sqrt{g}}$$

in local Sobolev spaces of order $s<-1$. The cluster collections converge in the stated positive-diameter Hausdorff matching topology, retaining all nesting depths. (OpenAI, 2026)

This includes the four-state Potts endpoint $J=U=(\log 3)/4$, where $g=4$. At $U=0$, $g=2$. No uniformity toward the limiting parameter $t=1$ is claimed.

The cluster topology concerns complete compact supports, not parametrized boundary curves. It permits the source’s noninjective positive-diameter matching. The theorem does not assert convergence of cluster area measures, magnetization fields, or every spin-domain-wall curve.

🏷️ Joint convergence versus marginal convergence

Knowing separately that a field converges to a GFF and a random set converges to a candidate set law does not specify their dependence. Even two fair signs can be either identical or independent while having the same marginals.

Here the limit makes a much stronger assertion: the two cluster systems are canonical functions of a common field. Their conditional laws after exploration must agree with the GFF domain decomposition. This is why the argument tracks both spins and current information in each stopped domain.

🏷️ Structure of the proof

Exact finite cut identities describe the residual model behind a revealed monochromatic perimeter. Circuit comparisons give macroscopic count bounds and control excursions close to rough explored boundaries. The exploration must retain the actual boundary connections; two regions are not wired together merely because their closures touch.

For the height, reflection frames in axial and diagonal directions give spectral representations. Their angular identities force the rescaled covariance onto the Green-kernel form without assuming a nonexistent forty-five-degree symmetry of the lattice law. Screening then makes separated limiting moments harmonic. Collision estimates determine their logarithmic singularities and Gaussian moment recursion.

The final geometric step identifies which discrete holes belong to one current cluster and excludes macroscopic support invisible to those holes. Conditional field limits are required even when one discrete chamber splits into several limiting components; the residual GFFs in those components must become independent after the appropriate constants are removed.

Only after these estimates does uniqueness of canonical local sets identify the joint limit. All spatial buffers and diameter thresholds are fixed before the mesh limit and refined afterward.

🏷️ Key steps and proof outlines

The proof must carry conditional information through every exploration. Its key analytic output is a Gaussian remainder law valid under the retained stopped data. Its key geometric output excludes a cluster support that leaves no contribution to the conditional mean (OpenAI, 2026).

Exact cuts give signed cluster sums

Behind a revealed monochromatic perimeter, the finite product weights give independent flat chamber laws after the appropriate height constants are removed. The retained state includes the actual spin and current connections along each resolved boundary. Geometric contact of closures alone does not identify two boundary ports.

For a neutral insertion with lattice coefficients $b_v$ summing exactly to zero, the cut representation has the form

$$K_{\delta }(f)=R_{\delta }+2\sum _C{ϵ}_C{q}_C,q_C=\sum _{v\in O_C}{b}_v.$$

Here $O_C$ is the union of odd holes belonging to cluster $C$, the signs ${ϵ}_C$ are independent and fair under the specified conditional data, and $R_{\delta }$ accounts for a bounded parity-transfer remainder. A fixed wired contribution is handled separately. For a cluster of diameter $r$ in the buffer, $|q_C|\le C(r+\delta {)}^2$. A cluster contributing nonzero mass must meet a connected neighborhood of the test, since otherwise its odd-hole indicator is constant there and neutrality kills the contribution.

Assign such clusters to $O(r^{-2})$ boxes at scale $r$. Each assigned cluster forces a relative crossing with separate boundary ports; distinct global clusters give distinct restricted components. Crossing-count bounds combined with the signed sum control insertion moments uniformly, even under arbitrarily rare separated block patterns. These estimates precede the Gaussian limit and therefore do not assume the conclusion they support.

Screening promotes an unconditional identity to conditional harmonicity

For $X_{\delta }=K_{\delta }(\Delta \varphi )$, reflection methods first provide an unconditional plane identity. The necessary stronger assertion is

$${\Vert \mathbb{E}[X_{\delta }\mid {\mathcal{E}}_{\delta }(Q)]\Vert }_1\to 0,$$

where ${\mathcal{E}}_{\delta }(Q)$ contains both exterior spins and the specified current decorations. Screening uses separated block patterns, the finite colors that can transmit information through their buffers, and short wires protected by circuits.

The error bookkeeping is explicit. Truncate $X_{\delta }$ at height $T$, with uniform truncation error $r(T)\to 0$. If the coupling mismatch is controlled by $\eta (q)\to 0$ and $M$ trials each have success probability bounded below by $p(q)>0$, the source obtains

$$\underset{\delta \downarrow 0}{\mathrm{limsup}}\Vert \mathbb{E}[X_{\delta }\mid {\mathcal{E}}_{\delta }(Q)]{\Vert }_1\le 2r(T)+16T\eta (q)+8T(1-p(q){)}^M.$$

Choose $T$ first, then $q$, then $M$, then a fixed spatial layout, and finally take the mesh limit. The truncation makes coupling errors usable without assuming moment bounds for an artificially pinned law. A fixed spin ring supplies the same interior conditional kernel in plane and flat domains; contiguity transfers small conditional means, and uniform integrability upgrades probability convergence to $L^1$. This proves separate harmonicity of limiting moment kernels under the appropriate exterior tests.

Neutral collisions determine all Gaussian moments

Crossing estimates give logarithmic moment bounds and boundary decay. A shared cluster joining distance $r$ from a connected wall to distance $R$ away costs at most $C(r/R{)}^{\eta }$ with buffers fixed. These bounds give zero boundary traces for the limiting moment kernels.

To identify a collision coefficient, choose two disjoint neutral smooth tests $\alpha ,\beta$ with

$$J=\iint \alpha (u)\beta (v)\log \frac{1}{|u-v|}dudv\ne 0.$$

For example, differences of equal small radial unit bumps centered at $0,1$ and at $3,4$ along an axis give $J=\log (8/9)$ by harmonic averaging. Insert these tests in a ball of radius $r$ around the collision, with other insertions in fixed separated disks. Local mixing identifies the limit as $\kappa J$ times the spectator moment. In the local logarithmic expansion, the term proportional to $\log (1/r)$ vanishes exactly by neutrality before any continuity estimate. A derivative bound makes the regular remainder $O(r)$. Since $J\ne 0$, the logarithmic coefficient is $\kappa$ times the lower-order moment.

Subtract the corresponding Green-kernel contractions from a $j$-point moment. As a function of its first variable, the remainder is harmonic, has removable collision singularities, and has zero boundary trace. The maximum principle gives

$$T_j(x_1,\dots ,x_j)=\kappa \sum _{i=2}^j{G}_U(x_1,x_i)T_{j-2}(x_2,\dots ,{\hat{x}}_i,\dots ,x_j).$$

This is Wick’s recursion. Moment bounds exclude extra diagonal mass when smearing, and Gaussian moment growth determines the law. The separate normalization argument gives $\kappa =1/a^2$. Since the Green kernel vanishes between distinct components, their limiting remainders are independent even when a single discrete chamber pinches into several components.

Conditional characteristic functions preserve the coupling

A deterministic flat-domain limit is first made uniform over chamber configurations whose test supports have a fixed positive clearance from the complement. Otherwise a contradicting sequence would have a Hausdorff-convergent subsequence of complements; the deterministic theorem and Green-kernel convergence would force the claimed limit on that subsequence.

Conditional on an exact stopping state, the remaining law is precisely one of these deterministic product kernels. Thus the conditional characteristic function differs in $L^1$ from

$$\exp \left(-\frac{1}{2a^2}\sum _{i,j}{u}_i{u}_j{G}_{U_{\delta }}(f_i,f_j)\right)$$

on the fixed-clearance event. Multiplication by any bounded variable measurable at that stop preserves the estimate. Passing to the joint limit and then exhausting by countably many compact buffers proves the required conditional GFF law. Later revelation data cannot be silently added to the conditioning. A further thinness argument extends the residual law to zero-extended tests meeting the explored set, which is needed in the final support identification.

A flat conditional mean cannot hide a macroscopic explored set

For this lemma use the normalized field $h$ with logarithmic singularity $-\log |x-y|$, rescaling the unit-height limit $K=h/a$ by $a$. Let $h$ be the zero-boundary GFF in a bounded simply connected domain $O$. The decisive continuum lemma says: if a boundary-connected explored set $A$ has simply connected complementary components and a full zero-extended conditional GFF remainder, and its conditional mean is constant on a ball $B⋐O$, then $A$ misses $B$ almost surely.

Its proof uses oscillatory Gaussian tests. For a mollified field $h_u$, a nonnegative smooth $b$ supported in $B$, and ${\beta }^2>2$, put

$$X_u=u^{-1}\int b(x)e^{i\beta h_u(x)}dx.$$

The GFF covariance gives a uniform $L^2$ bound and weak convergence $X_u\rightharpoonup 0$. If $A$ meets a smaller ball, boundary-connectedness forces its $u$-neighborhood in $B$ to have area at least $cu$. Near $A$, the conditional mollified variance is bounded uniformly: the killed Green kernel is bounded by a constant plus ${\log }^{+}(Cu/|z-w|)$, whose mollifier average is finite. After removing the constant phase of the conditional mean, the conditional expectation of $X_u$ is therefore a nonnegative integral bounded below by a fixed positive number. Testing weak convergence against that exploration-measurable phase and event gives a contradiction.

At a fixed sweep stop, the discovered mean is a finite signed sum of visible canonical odd-hole contributions on each localization. It is constant on balls missing those supports. The lemma therefore excludes any invisible piece of an explored cluster there. A connected Hausdorff limit is consequently contained in a countable disjoint family of canonical supports and the initial wall; the continuum partition theorem forces it into one member. Separate collar estimates exclude a macroscopic cluster collapsing onto the wall. Visible precursor containment then gives equality with the appropriate canonical cluster. This identifies both complete cluster families jointly with the same field, rather than just matching their separate distributions.

See Also

• Family 232 develops related reflection, screening, and collision mechanisms for triangular-lattice heights.
• Family 226 emphasizes the extra geometric estimates needed beyond topological loop observations.

🏷️ Manuscripts and verification scope

• The joint scaling limit of critical Ashkin–Teller currents — pinned manuscript and companion source.

No family scope page was present at lean/docs/233.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 joint scaling limit of critical Ashkin–Teller currents.