Critical-Coupling Noise Controller
Implementation & benchmark of arXiv:2609.02957 — Decoherence-controlled collective criticality in a two-dimensional quantum Stag Hunt
Source paper: Decoherence-controlled collective criticality in a two-dimensional quantum Stag Hunt arXiv:2609.02957 ⓘ · analyzed Sep 4, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Add calibrated multiplicative noise to a binary energy-based model, Hopfield network, or discrete recurrent attractor system so that its effective coupling follows a chosen distance from the collective critical point. Unlike ordinary temperature annealing, the controller distinguishes microscopic channel noise from Gibbs or sampling noise and predicts when ordered attractors should disappear through a measurable susceptibility peak.
Formulas
Mathematical statement
Represent binary states by $s_i\in\{-1,+1\}$ and the effective population energy by $\mathcal{H}(s)=-J_{\mathrm{eff}}\sum_{\langle i,j\rangle}s_i s_j-H_{\mathrm{eff}}\sum_i s_i$, where $\langle i,j\rangle$ denotes neighboring or selected interacting units, $J_{\mathrm{eff}}$ is the collective coupling, and $H_{\mathrm{eff}}$ is a field or bias. The Gibbs distribution is $P(s)\propto\exp[-\beta\mathcal{H}(s)]$, with inverse sampling temperature $\beta$. For depolarization, the paper's channel-to-Ising mapping gives $J_{\mathrm{eff}}(p)=(1-p)^2J_0$ while the neutrality branch keeps $H_{\mathrm{eff}}=0$. The dimensionless coupling is $K(p)=\beta J_{\mathrm{eff}}(p)$, and the square-lattice critical value is $K_c=\frac12\log(1+\sqrt{2})\approx0.4406868$. Thus the predicted critical channel strength is $p_*=1-\sqrt{K_c/(\beta J_0)}$, provided $\beta J_0>K_c$. In a neural model, replace $J_0$ by an empirically estimated dominant collective coupling $J_{\mathrm{NN}}$, such as the largest eigenvalue or average coupling after normalization.
Implementation notes
1. Exact integration point: use this controller in a binary energy-based model, Hopfield-style recurrent layer, or discrete latent-variable module. Let the pre-noise update be $h_i=\sum_jW_{ij}s_j+b_i$, and attenuate pair interactions with $W_t^{\mathrm{eff}}=(1-p_t)^2W_t$; keep Gibbs sampling temperature $\beta_t$ separate from channel noise $p_t$. 2. Pseudocode: initialize $W$ and choose $K_{target}$; at each epoch estimate $\widehat J_t$ from $\lambda_{max}((W_t+W_t^T)/2)$, or from the mean coupling on a fixed interaction graph; compute $p_t=\operatorname{clip}(1-\sqrt{K_{target}/(\beta_t\widehat J_t+\epsilon)},0,1)$; use $h_i=\sum_j(1-p_t)^2W_{ij}s_j+b_i$ in Gibbs or recurrent updates; record magnetization, susceptibility, and state autocorrelation. 3. Computed from the paper: the $(1-p)^2$ renormalization law, the neutrality condition, and the square-lattice critical value $K_c$. Estimated empirically: the neural interaction scale $\widehat J_t$, finite-size susceptibility, and whether the learned graph has a comparable effective coordination number. 4. First cheap experiment: train a $32\times32$ binary Hopfield or energy model on binarized MNIST patches, comparing fixed noise, ordinary temperature annealing, and this controller at $K_{target}=0.35,0.44,0.55$. Sweep $p$ after training. The falsifiable prediction is a magnetization loss and susceptibility peak when $\beta(1-p)^2\widehat J$ crosses $K_c$; on a square-grid model, the measured transition should be within approximately 20% of $p_*=1-\sqrt{K_c/(\beta\widehat J)}$. Above the boundary, ordered attractors should disappear and autocorrelation times should peak near the transition, indicating critical slowing down rather than a smooth monotonic noise effect.
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.