Unverified 2026

Kac-Ward Criticality Controller

Implementation & benchmark of arXiv:2608.16949 — Critical couplings of two dimensional Ising model on various lattices

Usefulness6/10

Difficulty6/10

Novelty7/10

Dynamics Stability Memory Accuracy Graph nn Rnn Ssm Training dynamics Regularization Statistical mechanics Spectral theory Linear algebra Dynamical systems

Source paper: Critical couplings of two dimensional Ising model on various lattices arXiv:2608.16949 ⓘ · analyzed Sep 1, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Replace an unconstrained message-passing or recurrent propagation matrix by a directed-edge operator with non-backtracking connectivity and orientation-dependent turning phases, inspired by the Kac–Ward construction. During training, monitor and control the zero-momentum spectral gap of $\mathcal A(0)=I-K(0)$, keeping the model near but on the stable side of the critical surface to obtain long memory without uncontrolled amplification.

Formulas

$$K_{e,e'}(\mathbf{k})=\mathbf{1}_{\{\operatorname{head}(e)=\operatorname{tail}(e'),\ e'\neq\bar e\}}\tanh(J_e)\,e^{i\theta(e,e')/2}e^{i\mathbf{k}\cdot\delta(e,e')},\qquad \mathcal A(\mathbf{k})=I-K(\mathbf{k}).$$

$$\det \mathcal A(\mathbf{0})=0\quad\Longleftrightarrow\quad\text{critical zero mode},\qquad s_{\min}(\mathcal A(\mathbf{0}))\approx0.$$

$$\det\mathcal A(\mathbf{k})\simeq c\left(m^2+\mathbf{k}^{\mathsf T}V\mathbf{k}\right),\qquad \|K^n\|\sim e^{-nm}\ \text{for }m>0.$$

$$L_{\mathrm{gap}}=\lambda\left(s_{\min}(\mathcal A(0))-g\right)^2,\qquad g>0,$$

Mathematical statement

For a periodic directed graph, let $e,e'$ be oriented edges, $\bar e$ the reversal of $e$, $J_e$ a trainable edge coupling, and $\theta(e,e')$ the signed turning angle from $e$ to $e'$. Define the Kac–Ward-like propagation matrix $K_{e,e'}(\mathbf{k})=\tanh(J_e)\exp[i\theta(e,e')/2]\exp[i\mathbf{k}\cdot\delta(e,e')]$ when the head of $e$ equals the tail of $e'$ and $e'\neq\bar e$, and zero otherwise. Here $\delta(e,e')$ is the cell displacement induced by the transition. Set $\mathcal A(\mathbf{k})=I-K(\mathbf{k})$. The paper's criticality criterion is $\det\mathcal A(0)=0$; in implementation, use the robust diagnostic $s_{\min}(\mathcal A(0))=0$, where $s_{\min}$ is the smallest singular value. A low-momentum expansion has the form $\det\mathcal A(\mathbf{k})\approx c[m^2+\mathbf{k}^{\mathsf T}V\mathbf{k}]$, with mass parameter $m$, positive semidefinite velocity tensor $V$, and nonzero constant $c$. Positive mass predicts exponentially decaying responses; the zero-mode surface predicts divergent correlation length and gradient amplification.

Implementation notes

Use this as a propagation module for a graph neural network, cellular recurrent network, or state-space model. Represent node features on directed incident edges. For each layer or recurrent step, construct sparse $K(0)$ from non-backtracking transitions; parameterize $J_e=\operatorname{softplus}(u_e)$, optionally multiplied by a learned bounded sign, and fix $\theta(e,e')$ from the graph embedding orientation. Propagate with $h_{t+1}=K(0)h_t+Bx_t$, then aggregate directed-edge states back to nodes. Pseudocode: assemble sparse $K$; form $A=I-K$; estimate $s_{\min}(A)$ using 5--10 Lanczos iterations on $A^{\mathsf T}A$; add $L_{\mathrm{gap}}=\lambda(s_{\min}(A)-g)^2$ to the task loss; and reduce $g$ only when validation loss improves while hidden-state gains remain bounded. The graph connectivity, turning phases, and zero-mode criterion are taken from the paper; the positive target gap and penalty are ML adaptations. Estimate the low-frequency tensor by evaluating $A(k)$ at $k=\pm\Delta e_j$ and fitting the quadratic variation of the smallest singular value. First experiment: compare a 20-step directed-edge recurrent GNN with a standard message-passing GNN on synthetic long-range graph diffusion and a sequence prediction task. Sweep a coupling multiplier $\alpha$ in $J_e\leftarrow\alpha J_e$. The predicted signature is a sharp transition near measured $s_{\min}(I-K(0))=0$: below it, response norms decay exponentially with depth; near it, memory length scales as $1/m$; beyond it, hidden states and gradients grow. The empirical transition should match the zero of the smallest singular value within 20 percent, and fitted decay rates should be approximately linear in estimated mass $m$.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.