Why weakly perturbed Ising models would keep the same large-scale geometry

← All digests

Family 218 · Probability and statistical mechanics
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-11

1. The question

The critical square-lattice Ising model has an exceptionally rich mathematical description. Its spin correlations transform predictably under conformal maps, and the boundary between plus and minus regions approaches a particular random curve. But how much of that structure survives if the microscopic model is changed?

There are two natural changes. One can add small, fixed interactions among several nearby spins. Or one can randomly vary the strength of each bond, then hold that environment fixed while sampling the spins. Both changes destroy the special homogeneous nearest-neighbor formulas.

This family contains six manuscripts claiming that much of the critical behavior survives. They cover mixed spin-and-energy correlations, full spin-interface laws, and the slow residual effect of random bonds on correlation fluctuations. A technical companion supplies comparisons that remain valid after an interface exploration reveals an irregular boundary.

The common theme is universality: different microscopic interactions can produce the same macroscopic fields and curves after the correct temperature and field normalizations are chosen.

2. Background and essential definitions

An Ising spin is $\sigma_x\in\{-1,1\}$. The basic ferromagnetic model gives configurations weight

$$\exp\!\left(\beta\sum_{e=\{x,y\}}J_e\sigma_x\sigma_y\right), \qquad J_e>0.$$

Positive bonds favor agreement, while $\beta$ is inverse temperature. In the homogeneous model with $J_e=1$, the critical coupling is

$$K_0=\frac12\log(1+\sqrt2).$$

A spin correlation averages products of spins. A local energy observable is a neighboring product $\sigma_x\sigma_y$, centered by subtracting its expectation. At criticality, their scaling dimensions are respectively $1/8$ and one: mesh-normalized fields use factors $\delta^{-1/8}$ and $\delta^{-1}$.

A conformal map preserves angles locally. Conformal covariance of a correlation means that mapping its domain changes it by explicit powers of the map’s local scale factor $|\phi'|$. It is more information than knowing a power-law decay in the plane.

Fix plus spins along one boundary arc and minus spins along the other. These Dobrushin boundary conditions force an interface joining the two arc endpoints. Its proposed limit is chordal Schramm–Loewner evolution with parameter three, abbreviated SLE3. In the upper half-plane it is generated by

$$\partial_t g_t(z)=\frac{2}{g_t(z)-\sqrt3 B_t},$$

where $B_t$ is ordinary one-dimensional Brownian motion. The maps $g_t$ remove a growing curve; conformal mapping gives the law in other simply connected domains. The parameter three specifies the randomness of the driving motion.

For random bonds, quenched means that the Gibbs law is normalized separately for each fixed environment $\omega$. Only afterward may one average those laws over environments. The environment is not weighted by its partition function. Thus a quenched correlation is itself a random variable depending on the sampled bonds.

3. Development history and previous partial results

Onsager’s exact free energy and Yang’s spontaneous magnetization made the nearest-neighbor model a foundation of critical statistical mechanics. Conformal field theory organized its spin and energy fields, and discrete complex analysis later established rigorous continuum limits. Chelkak, Hongler, and Izyurov proved conformal convergence of bulk spin correlations and developed the theory of mixed primary fields. Spin-correlation convergence

For interfaces, Smirnov’s fermionic observables and the work of Chelkak and Smirnov supplied the analytic ingredients. Chelkak, Duminil-Copin, Hongler, Kemppainen, and Smirnov obtained convergence of critical spin interfaces to SLE3. The associated random-cluster interface has a different limit, SLE with parameter $16/3$. The distinction matters when comparing later random-bond results. Interface convergence

Universality beyond the exactly solvable model also has substantial predecessors. Giuliani, Greenblatt, and Mastropietro proved energy-correlation limits for weak finite-range pair perturbations. Antinucci, Giuliani, and Greenblatt extended energy universality to cylinders with even multispin interactions of either sign. Cava, Giuliani, and Greenblatt treated boundary spins in the free half-plane. Thus signed interactions and actual-state energy centering were already features of earlier work; the broader mixed bulk correlation problem remained. Plane energy universality, Cylinder energy universality

Greenblatt and Peltola constructed an interface martingale for non-integrable models and formulated estimates that would lead to SLE3. Their work supplies a route to the problem, rather than already proving the full interface limit claimed here. Non-integrable spin interfaces

Random bonds introduce a different difficulty. Harris’s criterion leaves two-dimensional Ising disorder at a borderline case. The analyses of Dotsenko and Dotsenko, Shalaev, Shankar, and Ludwig developed the prediction of unchanged leading exponents with logarithmic corrections. Earlier rigorous random-bond interface results cited by the manuscripts concern disorder that vanishes with the mesh and random-cluster interfaces. Here the disorder strength stays fixed and the claimed limit concerns spin interfaces.

4. What the new manuscripts claim

Mixed bulk correlations under deterministic interactions

M4 starts with nearest-neighbor Ising and adds a fixed, sufficiently small finite-range potential on even sets of spins. “Even” means invariance under flipping every spin. The potential is translation invariant and invariant under square-lattice rotations and reflections; its coefficients may have either sign.

The manuscript constructs one temperature branch $\beta_c(\lambda)$ and two nonzero factors $Z_\sigma(\lambda)$, $Z_\epsilon(\lambda)$. For any fixed number of distinct interior insertions, the mixed correlation limit is

$$Z_\sigma(\lambda)^n Z_\epsilon(\lambda)^m \mathcal C_D^b(x_1,\ldots,x_n;y_1,\ldots,y_m),$$

where $\mathcal C_D^b$ is the corresponding critical nearest-neighbor correlation with $n$ spin and $m$ energy insertions. The same branch and factors work in all bounded simply connected domains with twice continuously differentiable Jordan boundary, with free, plus, or minus boundary conditions, and in the periodic thermodynamic plane state.

Each energy is centered in its actual interacting state and domain. Under a conformal map $\phi$, the limit acquires factors

$$\prod_{i=1}^n|\phi'(x_i)|^{1/8} \prod_{j=1}^m|\phi'(y_j)|.$$

The theorem concerns separated bulk points, not coincident or boundary insertions.

Deterministically perturbed interfaces

M5 considers sufficiently weak, finite-range, square-symmetric perturbations of the contour energy, allowing many-edge interactions of either sign. It claims a domain-independent temperature at which the full oriented spin interface converges to SLE3.

Here domains need only have Jordan boundaries: simple closed boundary curves, without a smoothness assumption. Their lattice boundaries must approximate the limiting boundary uniformly through homeomorphisms. The specified interactions across the boundary are retained, and admissible exterior contours may vary with the mesh.

The convergence keeps the order of traversal, modulo increasing changes of speed. It is stronger than convergence of the curve’s image as an unparameterized compact set. For the signed deterministic models, the critical branch is established through the stated massless or interface behavior; one should not replace it by a magnetization-threshold definition without an additional argument.

Fixed weak random bonds

M2 treats independent fair bonds $J_e=1\pm\varepsilon$. M1 extends the interface claim to

$$J_e=1+\varepsilon\xi_e,$$

where the $\xi_e$ have any fixed bounded, nondegenerate, mean-zero independent identical law. After rescaling the bound, one may take their support in $[-1,1]$. The positive strength $\varepsilon$ must be sufficiently small for that law and remains fixed as the mesh shrinks.

The temperature is the physical threshold where the disorder-averaged plus-state magnetization becomes positive. For deterministic Jordan-domain approximations, the quenched oriented interface law converges in probability over environments to the deterministic SLE3 law. This is not an almost-sure assertion for every environment or permission to choose the domains after observing the disorder.

For fair binary bonds, the critical point is identified by

$$\sinh\!\bigl(2\beta_c(1-\varepsilon)\bigr) \sinh\!\bigl(2\beta_c(1+\varepsilon)\bigr)=1.$$

For general mean-zero bounded laws, M1 gives $\beta_c=K_0+O(\varepsilon^2)$, without an analogous closed self-dual formula.

Logarithmic fluctuations remain visible

M3 assumes the specified deterministic reference estimates and treats the fair binary disorder. Let $C_\omega(r)$ be the critical spin correlation in a fixed environment, after the free-box thermodynamic limit. Its relative second moment is

$$R_\varepsilon(r)= \frac{\mathbb E_\omega[C_\omega(r)^2]} {\mathbb E_\omega[C_\omega(r)]^2}.$$

The manuscript states

$$R_\varepsilon(r)=(\log r)^{1/4+o(1)}.$$

Since $R_\varepsilon-1$ is the variance divided by the squared mean, this detects sample-to-sample variability even while the macroscopic interface law becomes universal. It is an exponent statement, not a positive-amplitude equivalent. At zero disorder the ratio is identically one; the theorem fixes a positive disorder strength before sending $r$ to infinity.

A comparison theorem supporting the family

M6 gives likelihood-ratio bounds for finite ferromagnetic Ising systems with arbitrary common pinned spins. It also claims that, after exploring across a band of fixed positive width, finitely many regular signed barriers approximate the good conditional laws on a separated target in total variation. Total variation controls the probability of every target event simultaneously, so the test observable may change with the mesh.

This is a statement about the pure reference model and its conditioned laws. It supplies tools used by the other manuscripts, rather than being another random-bond scaling theorem.

5. Why the advance matters

Importance: 5/5. The family addresses universality for central observables that lie beyond the reach of the original exact solution: mixed bulk fields, whole interfaces, and fixed weak disorder.

The deterministic results would connect a broad class of microscopic perturbations to the same conformal theory using only one temperature adjustment and two field factors. The random results would show that a typical fixed environment loses its macroscopic directional irregularities in the interface law, while retaining measurable logarithmic effects in another observable.

The strength of that combination is precisely that universality is not presented as complete disappearance of all microscopic information. Amplitudes, the critical temperature, and disorder moments can change even when the limiting geometric law remains SLE3.

6. The main idea

Keep the important fields and shrink the rest

A change of scale replaces a local interaction by its conditional average in the pure critical model. Most resulting directions contract. The spin and energy fields are the important low-dimensional components, and changing temperature controls the expanding thermal direction.

The difficulty is uniformity. Repeated averaging creates functions of many microscopic spins; convergence of a fixed list of spin correlations does not control all such functions. M6’s buffered comparisons and finite descriptions of explored boundaries make it possible to approximate their effect on distant regions uniformly over bounded tests.

M4 and M5 then construct exact transformations of the local interactions. The temperature is chosen so the thermal component does not grow. Marked spin and energy insertions retain their leading components with model-dependent factors, while the remaining interactions become negligible at a fixed physical separation.

Convert local errors into geometric control

A local change deep inside a region can alter a macroscopic crossing only if several long interface strands reach it. For the pure spin model, the four-arm exponent is $21/8>2$: the rarity of such configurations beats the two-dimensional count of possible change locations. Separate boundary and corner estimates complete the comparison.

The signed perturbations need not preserve the usual ferromagnetic ordering for the physical spins. The manuscript therefore uses restricted boundary comparisons with auxiliary variables, retaining the actual cross-boundary interaction. Further passage estimates prevent a curve from tracing the right set in the wrong order.

Track slowly decaying random variance

For random bonds, a supremum over all environments would be too strong. The construction tracks environmental moments and treats rare concentrations of large local effects through positive conditional weights. It also records which original random bonds each effective interaction depends on.

After tuning the mean thermal component, the leading disorder variance $g_j$ follows a marginal recursion of the form

$$g_{j+1}=g_j-cg_j^2+\text{smaller terms}, \qquad c>0.$$

Thus $g_j$ decays like $1/j$, rather than exponentially. Its accumulated effect can produce powers of a logarithm. In M3, each distant spin insertion contributes half of the relative-second-moment exponent; the local dependence records allow their leading factors to separate.

Finally, the tuned temperature must be identified with the physical transition. Binary disorder permits a self-dual comparison. For a general law, M1 tunes primal and dual descriptions separately and uses finite-torus duality and monotonicity to show that the temperatures agree. Crossing and magnetization arguments then identify the threshold used in the theorem.

7. Scope, limitations, and open questions

All perturbations are sufficiently weak and fixed before the continuum limit. The deterministic potentials have finite range and square symmetries. The random bonds are independent, bounded, and ferromagnetic. Strong disorder, correlated environments, and general anisotropic interactions are outside these claims.

The random interface conclusion is convergence in probability of quenched laws, not uniform convergence over environments. The logarithmic correlation-moment result is stated for fair binary disorder and with its deterministic companion assumptions; M1’s broader interface theorem does not automatically broaden that moment theorem.

Domain regularity and boundary prescriptions differ between the correlation and interface results. M6 also requires positive geometric separation for its buffered comparisons; its constants need not remain controlled as that separation vanishes.

The family does not assert every boundary-field limit, every multiply connected geometry, or an entire loop-ensemble theorem for each perturbation. Extending the methods to those observables, and resolving further disorder-dependent logarithmic factors, remain separate questions.

8. Sources and further reading

M4 covers deterministic bulk fields, M5 deterministic interfaces, and M6 the pure conditional-comparison tools. M2 establishes the binary random-bond interface result; M1 extends its environment law; M3 studies a distinct random-correlation observable with explicit reference assumptions.

Useful predecessors include the pure interface theorem, energy universality under weak interactions, and Greenblatt–Peltola’s interface framework. The source introductions give the primary references for the historical logarithmic-disorder predictions and the earlier vanishing-disorder results.

Source manuscripts

• M1. OpenAI, Quenched SLE₃ limits for general weak random-bond Ising models (October 5, 2026).
• M2. OpenAI, Quenched SLE₃ Universality for the Weak Random-Bond Ising Model (October 5, 2026).
• M3. OpenAI, Logarithmic Relative Fluctuations in the Weakly Disordered Planar Ising Model (October 5, 2026).
• M4. OpenAI, Conformal universality of bulk Ising correlations under weak interactions (September 23, 2026).
• M5. OpenAI, $\mathrm{SLE}_3$ universality for weak finite-range Ising interactions (September 23, 2026).
• M6. OpenAI, Buffered comparison and stopping-band resolution in critical Ising (September 23, 2026).

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