gpt 6 astra pro attempt on Phys-5

Conformal Field Theory Limit of the Critical Three-Dimensional Ising Model

Definitions and mathematical statement

For Ising spins $\sigma_x\in\{-1,1\}$ on a three-dimensional lattice, the zero-field Gibbs weight is proportional to $\exp(\beta\sum_{\langle x,y\rangle}\sigma_x\sigma_y)$. At critical inverse temperature $\beta_c$, rescale the lattice spacing to $a\downarrow0$ and normalize the spin field as a random distribution, schematically \[ \Phi_a(f)=c_a\sum_{x\in a{\mathbb{Z}}^3}f(x)\sigma_x. \] The goal is a nontrivial limiting field with conformally covariant correlation functions and non-Gaussian statistics. For example, a centered Gaussian limit would satisfy $\langle\Phi_1\Phi_2\Phi_3\Phi_4\rangle=\sum_{\rm pairings}\langle\Phi_i\Phi_j\rangle\langle\Phi_k\Phi_l\rangle$. The selected target requires failure of such Wick factorization. Normalization and convergence topology are part of a rigorous scaling-limit construction, not specified numerically by the catalogue.

Formulation and concept references: [S1].

Short English statement

Does the critical three-dimensional Ising model converge to a genuinely interacting conformally invariant field theory?

Research attempt

Current frontier.

The requested conclusion contains three distinct tasks: construction of a scaling limit, conformal covariance of that limit, and survival of a connected four-point function. The survey in [S1], Sections 6.3 and 8.4, explains both the random-current description of non-Gaussianity and the outstanding three-dimensional conformal-limit problem. The two-point estimates of Duminil-Copin--Panis imply divergence of the critical bubble diagram in three and four dimensions; their bound $\eta\le 1/2$ in dimension three is conditional on existence of that critical exponent, not a construction of it [S3], [E1]. Bootstrap calculations of a three-dimensional four-point function are informative within their conformal-field-theory framework but do not prove convergence of the lattice model [E2]. These inputs are kept separate below.

Attack route.

One route is to normalize by an actual lattice variance rather than assume an unproved critical exponent. This makes compactness relatively accessible, but leaves uniqueness and the scale dependence of the normalization. A second route is to use switching of random currents to quantify exactly what must persist for a non-Gaussian limit. A third, genuinely additional, step would establish conformal covariance; neither compactness nor a nonzero cumulant supplies that step. We pursue the first two routes and make the missing probabilistic estimate explicit.

Attempt: a normalization and a subsequential construction.

Work in a translation-invariant, spin-flip-symmetric, zero-field infinite-volume critical Ising state obtained by the usual ferromagnetic limiting procedure. Write $G(u,v)=\langle\sigma_u\sigma_v\rangle\ge0$ on ${\mathbb{Z}}^3$. The positivity and four-point inequality used here are the standard ferromagnetic inequalities and the switching identity reviewed in [S1], Sections 6.1--6.3. The following compactness argument is an auxiliary deduction; it is not claimed as a historically new scaling-limit theorem.

For $0<a\le1/2$, set $L=\lfloor a^{-1}\rfloor$ and \[ B_L=\{0,\ldots,L-1\}^3,\quad M_L=\sum_{u\in B_L}\sigma_u, \quad b_L^2=\langle M_L^2\rangle,\quad \Phi_a(f)=b_L^{-1}\sum_{u\in{\mathbb{Z}}^3}f(au)\sigma_u. \] Since all covariances are nonnegative and $\sigma_u^2=1$, $b_L^2\ge L^3$. In particular the normalization is defined and uses no assumed power law. Tile ${\mathbb{Z}}^3$ by translates of $B_L$. A fixed compact physical set $K$ intersects at most $N_K$ such blocks, uniformly in small $a$, because $aL\ge1/2$. Translation invariance gives $L^2$ norm one to every normalized block sum. For any possibly complex $f$ supported in $K$, positivity of $G$ and then the triangle inequality in $L^2$ give \begin{align*} \langle |\Phi_a(f)|^2\rangle &\le \|f\|_\infty^2 b_L^{-2} \left\langle\left(\sum_{u\text{ in the covering blocks}} \sigma_u\right)^2\right\rangle\\ &\le N_K^2\|f\|_\infty^2. \tag{42.1} \end{align*} There is no assumption of independence between blocks in this estimate.

Here is a topology in which $42.1$ gives tightness. Put a smooth cutoff $\chi$ inside a fixed larger box, identify that box with a torus, and let $e_k$ be its Fourier basis. Its functions have uniformly bounded sup norm. For $s>3/2$, $42.1$ and the definition of the negative Sobolev norm imply \[ \sup_a{\mathbb{E}}\|\chi\Phi_a\|_{H^{-s}}^2 \le C_\chi\sum_{k\in{\mathbb{Z}}^3}(1+|k|^2)^{-s}<\infty. \tag{42.2} \] Fix $s'>s$. Bounded sets in $H^{-s}$ are relatively compact in $H^{-s'}$: the high-frequency tail of their squared $H^{-s'}$ norm is bounded by $(1+R^2)^{-(s'-s)}$ times their squared $H^{-s}$ norm, and the remaining Fourier coordinates form a finite-dimensional bounded set. Markov's inequality applied to $42.2$ therefore gives compact containment. Using a countable family of cutoffs exhausting ${\mathbb{R}}^3$, with failure probabilities summing to an arbitrarily small number, gives tightness in $H^{-s'}_{\rm loc}({\mathbb{R}}^3)$. Consequently every sequence $a_j\downarrow0$ has a subsequence converging in law in this space. This proves only a subsequential statement with the displayed normalization.

The subsequential limits are not identically zero. Choose smooth, nonnegative, compactly supported $f_0$ with $f_0\ge1$ on $[0,1]^3$. Covariance positivity gives ${\mathbb{E}}\Phi_a(f_0)^2\ge1$. The four-spin inequality implies, for nonnegative weights, \[ {\mathbb{E}}\Phi_a(f_0)^4\le3\left({\mathbb{E}}\Phi_a(f_0)^2\right)^2\le C_{f_0}. \tag{42.3} \] For distinct sites this inequality follows from the nonpositive connected four-point function. Coincident indices cause no problem: for example, $\langle\sigma_i^2\sigma_j\sigma_k\rangle=G(j,k)$ is at most $G(j,k)+2G(i,j)G(i,k)$. Summing the inequalities proves $42.3$. Thus the squares are uniformly integrable. Evaluation at $f_0$ is continuous in the chosen local distribution topology, so passage along any convergent subsequence preserves the lower bound on its second moment. This establishes nontriviality, not non-Gaussianity.

Attempt: an exact non-Gaussianity certificate.

For four distinct sites, let $U_4$ denote the four-point function minus its three Wick pairings. The switching identity in [S1], equation $6.20$, is \[ \begin{aligned} U_4(u_1,u_2,u_3,u_4) &=-2G(u_1,u_2)G(u_3,u_4)\\ &\quad\times \left(\mathbb P^{\{u_1,u_2\}}\!\otimes \mathbb P^{\{u_3,u_4\}}\right)[\mathcal C], \end{aligned} \tag{42.4} \] where $\mathcal C$ is the event that all four sites are connected in $\mathbf n_1+\mathbf n_2$. The probabilities here are source-conditioned random-current probabilities, not independent random-walk probabilities. One may first use the finite-volume identity and pass to the chosen state; equivalently the bounded connection factor is the corresponding limiting ratio where the denominator is positive.

Choose four nonnegative smooth test functions with pairwise separated compact supports. Put $X_i^{(a)}=\Phi_a(f_i)$ and $B_{ij}^{(a)}={\mathbb{E}} X_i^{(a)}X_j^{(a)}$. When $B_{12}^{(a)}B_{34}^{(a)}>0$, define $p_a$ to be the weighted average of the connection probabilities in $42.4$, with weights \[ b_L^{-4}\prod_{i=1}^4 f_i(au_i)G(u_1,u_2)G(u_3,u_4), \] divided by $B_{12}^{(a)}B_{34}^{(a)}$. All sums are finite, all weights are nonnegative, and their sum equals that denominator. Therefore $0\le p_a\le1$ and summing $42.4$ proves the exact identity \[ \kappa_4(X_1^{(a)},X_2^{(a)},X_3^{(a)},X_4^{(a)}) =-2B_{12}^{(a)}B_{34}^{(a)}p_a. \tag{42.5} \] This isolates a useful conditional criterion. Suppose, along a convergent subsequence, that $B_{12}^{(a)}\to b_{12}>0$, $B_{34}^{(a)}\to b_{34}>0$, and $\liminf p_a\ge p_*>0$. Assume also uniform eighth moments for these four smeared fields. This last assumption is stated explicitly here rather than inferred from $42.3$. Hölder's inequality then gives uniform integrability of their mixed fourth products. Joint convergence and moment convergence in $42.5$ imply \[ \kappa_4(X_1,X_2,X_3,X_4)\le-2b_{12}b_{34}p_*<0. \tag{42.6} \] The limiting field consequently fails Wick factorization. The outstanding substantive input for this route is a uniform lower bound for the weighted, source-conditioned connection probability, together with nondegeneracy of the separated covariances. The unit-block normalization alone establishes neither.

Stress test.

Finite-lattice non-Gaussianity does not survive automatically. For $N$ independent symmetric spins, $Y_N=N^{-1/2}\sum_{i=1}^N\sigma_i$ has ${\mathbb{E}} Y_N^2=1$ and \[ {\mathbb{E}} Y_N^4=3-2/N,\qquad \kappa_4(Y_N)=-2/N\longrightarrow0. \] This exact calculation is checked by enumeration for $N\le8$ in the accompanying script. It is a counterexample to a proposed inference, not to the interacting critical Ising statement. Similarly, divergence of a bubble sum is not itself a lower bound for $p_a$: the cited bubble result also covers dimension four, where Gaussian scaling behavior is established for the relevant Ising limits [S1], Section 6.3; [S3]. Source conditioning and the normalization in $42.5$ cannot be dropped in a random-walk intersection analogy.

The compactness argument deliberately leaves major information undetermined. Different subsequences may give different fields; $b_L$ has not been shown to have regular variation; the limit has not been shown to be rotation, dilation, or inversion covariant. Even translation covariance inherited in a limit would not establish the full conformal group. Neither a chosen normalization nor a nonzero cumulant supplies a conformal transformation law. These omissions are part of the target, not optional refinements.

Outcome.

Outcome: Exploratory approach with unresolved gap. A direct variance-normalized tightness argument gives nontrivial subsequential random-distribution limits, and an exact random-current identity reduces persistence of a specified non-Gaussian four-point function to a positive weighted connection probability. The conditional criterion and the elementary compactness proof are not assertions of historical novelty. Uniqueness across scales, the required connection estimate, and conformal covariance remain unproved here. Thus the full scaling-limit and conformal-invariance assertion has not been established.

Sources

• S1 100 Years of the (Critical) Ising Model on the Hypercubic Lattice (Hugo Duminil-Copin). https://arxiv.org/html/2208.00864. Catalogue formulation source.
• S2 The Ising model: highlights and perspectives. https://link.springer.com/article/10.1007/s11040-025-09515-1. Catalogue research/status reference; not a proof endorsement.
• S3 New lower bounds for the (near) critical Ising and phi4 models' two-point functions. https://www.unige.ch/~duminil/publi/2024_lower_bound_Ising.pdf. Catalogue research/status reference; not a proof endorsement.
• S4 Certified reduced-basis emulation of conformal field theories on the fuzzy sphere. https://arxiv.org/abs/2609.11675. Catalogue research/status reference; not a proof endorsement.
• E1 Hugo Duminil-Copin and Romain Panis, New lower bounds for the (near) critical Ising and $\varphi^4$ models' two-point functions, arXiv:2404.05700v1 $2024$, Theorems 1.5 and 1.8; author PDF dated 9 April 2024. https://arxiv.org/abs/2404.05700. Identification of the manuscript retained as [S3]; exponent qualification and bubble divergence. Consulted 27 September 2026.
• E2 Slava Rychkov, David Simmons-Duffin and Bernardo Zan, Non-gaussianity of the critical 3d Ising model, arXiv:1612.02436 $2016$, abstract and four-point-function analysis. https://arxiv.org/abs/1612.02436. Conformal-bootstrap analysis, not an unconditional lattice scaling-limit construction. Consulted 27 September 2026.