Unverified 2026

Ward-Residual Model Selection

Implementation & benchmark of arXiv:2608.25103 — Convergence of the conformal Ward identity in the derivative expansion approximation

Usefulness5/10

Difficulty5/10

Novelty7/10

Regularization Stability Generalization Accuracy Training Regularization Optimizer Mlp Dynamical systems Renormalization Geometry Optimization

Source paper: Convergence of the conformal Ward identity in the derivative expansion approximation arXiv:2608.25103 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Train a neural approximation to a scale-dependent effective action, energy functional, or field while penalizing the residual of a known continuous-symmetry Ward identity. Select the regulator, smoothing scale, or architecture hyperparameter at the minimum Ward residual, and require that the residual decreases when model capacity or derivative-expansion order increases.

Formulas

$$\delta_{\mathrm{conf}}f(x)=\epsilon_\mu\left(x^2\partial_\mu-2x_\mu x_\nu\partial_\nu-2x_\mu D_f\right)f(x),$$

$$\mathcal W_\mu\Gamma_k=\Delta_{k,\mu},$$

$$R(\theta,\lambda)=\frac{\mathbb E_{x,\mu}\left[\left|\mathcal W_\mu\Gamma_\theta(x;\lambda)-\widehat{\Delta}_{k,\mu}(x;\lambda)\right|^2\right]}{\mathbb E_{x,\mu}\left[\left|\mathcal W_\mu\Gamma_\theta(x;\lambda)\right|^2+\epsilon\right]},$$

$$\lambda^*=\arg\min_\lambda R(\theta^*(\lambda),\lambda),\qquad \theta^*(\lambda)=\arg\min_\theta\left[L_{\mathrm{data}}(\theta)+\alpha R(\theta,\lambda)\right].$$

Mathematical statement

For a scalar function f(x) with scaling dimension D_f, the special conformal generator used in the paper is $\mathcal K_\mu=x^2\partial_\mu-2x_\mu x_\nu\partial_\nu-2x_\mu D_f$. An exactly invariant effective action satisfies $\mathcal W_\mu\Gamma=0$; at finite RG cutoff k and finite truncation, the regulator generates a breaking term, $\mathcal W_\mu\Gamma_k=\Delta_{k,\mu}$. For a neural approximation $\Gamma_\theta$, define a normalized residual $R$ by comparing its computed Ward operator with the known or estimated breaking term. Here $x$ is the input coordinate or field configuration, $\mu$ indexes the symmetry generator, $\theta$ are network parameters, $\lambda$ is a regulator or smoothing parameter, and $\epsilon$ prevents division by zero. The transferable criterion is to choose $\lambda$ where R and the sensitivity of a downstream observable are jointly small, then verify that R falls with increasing truncation capacity.

Implementation notes

1. Exact integration point: represent a scalar effective action, energy, log-density, or coarse-grained field functional with an MLP or neural operator $\Gamma_\theta(x;\lambda)$. Let $\lambda$ be a Gaussian smoothing width, Fourier cutoff, spectral damping coefficient, or injected-noise scale. Sample coordinates or field configurations x from each training batch and use automatic differentiation to obtain the derivatives in the Ward generator. 2. Pseudocode: compute gamma = net(x, lambda); compute grad = jacobian(gamma, x); calculate x2 = sum(x*x) and xdotgrad = sum(x*grad); form Kmu = x2*grad_mu - 2*x_mu*xdotgrad - 2*x_mu*D; set ward = Kmu - Delta_hat; and optimize loss = data_loss + alpha*mean(ward2)/(mean(Kmu2)+eps). 3. Computed versus estimated: the generator and derivatives are computed by autodiff. The cutoff-breaking term $\widehat{\Delta}$ should be derived analytically when the discretization or regulator is known; otherwise use zero only when exact symmetry is the target, and report this approximation explicitly. Estimate observable sensitivity by finite differences, $S_Q=|Q(\lambda+\delta)-Q(\lambda-\delta)|/(2\delta)$. 4. First cheap experiment: fit a small MLP to synthetic two-dimensional functions constructed to obey scaling or conformal constraints. Compare ordinary MSE against MSE plus Ward loss over ten smoothing widths and widths 32, 64, and 128. The predicted signature is a U-shaped residual curve $R(\lambda)$ with a stable minimum, reduced observable sensitivity near that minimum, and at least a 20% reduction in R when capacity is increased. If R has no reproducible minimum or does not decrease with capacity, reject the mechanism for the task.

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.