Beckner spectral logit regularizer
Implementation & benchmark of arXiv:2608.11126 — Sharp Beckner's Inequalities for Axially Symmetric Functions on $\mathbb{S}^N$
Source paper: Sharp Beckner's Inequalities for Axially Symmetric Functions on $\mathbb{S}^N$ arXiv:2608.11126 ⓘ · analyzed Sep 1, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Represent an axially symmetric neural field on the sphere as a scalar function of latitude and regularize it with the paper's Paneitz energy together with its exponential log-partition term. Enforce a center-of-mass condition on the normalized exponential density so that the regularizer cannot be reduced by simply translating the field toward a first spherical-harmonic mode.
Formulas
Mathematical statement
For normalized surface measure $dw$ on the unit sphere $\mathbb{S}^{N}$, the paper defines $J_{\alpha,N}(u)=\frac{\alpha}{2}\int_{\mathbb{S}^{N}}u(P_Nu)\,dw+(N-1)!\int_{\mathbb{S}^{N}}u\,dw-\frac{(N-1)!}{N}\ln\int_{\mathbb{S}^{N}}e^{Nu}\,dw$. Here $u\in H^{N/2}(\mathbb{S}^{N})$, $N\geq3$, and $\alpha\geq\frac12$. The conformally invariant operator is $P_N=\prod_{k=0}^{(N-2)/2}(-\Delta+k(N-k-1))$ for even $N$, while for odd $N$, $P_N=\left(-\Delta+((N-1)/2)^2\right)^{1/2}\prod_{k=0}^{(N-3)/2}(-\Delta+k(N-k-1))$; $\Delta$ is the spherical Laplace–Beltrami operator. The theorem states $J_{\alpha,N}(u)\geq0$ for axially symmetric $u$ whose center of mass is at the origin. In the adaptation, $u_\theta$ is a network-produced axial logit field, $\mathcal{E}_P(u)=\langle u,P_Nu\rangle$ is evaluated spectrally, and $\mathcal{R}_{\mathrm{B}}(u)=J_{\alpha,N}(u)$ is added to the task loss. The center-of-mass constraint is implemented using the exponential density $\rho_u(x)=e^{Nu(x)}/\int e^{Nu}\,dw$ and penalty $\|\int x\rho_u(x)\,dw\|_2^2$, which is the practical version of the theorem's geometric admissibility condition.
Implementation notes
Integrate this at the output-field loss, not inside the network weights. Choose $N=3$ first and let the network produce an axial field $u_\theta(t)$ with $t=x_{N+1}\in[-1,1]$, so the output is independent of azimuth. Use a Gauss--Gegenbauer quadrature grid $t_i$ with normalized weights $q_i$ to approximate spherical integrals. Precompute the first $K$ Gegenbauer basis functions and the diagonal eigenvalues of $P_N$ in that basis, or precompute a matrix representation $P_{\mathrm{mat}}$ on the quadrature grid. At every step evaluate $u_i=u_\theta(t_i)$, compute $\log Z=\operatorname{logsumexp}_i(\log q_i+Nu_i)$, set $\rho_i=\exp(Nu_i-\log Z)$, transform $u$ to coefficients $a_k$, and compute $E_P=\sum_{k=0}^{K}\mu_k a_k^2$ where $\mu_k$ are the operator eigenvalues. Then calculate $B=\frac{\alpha}{2}E_P+(N-1)!\sum_iq_iu_i-\frac{(N-1)!}{N}\log Z$. Compute the center penalty $C=\|\sum_iq_i\rho_i x_i\|_2^2$ and optimize $L=L_{\mathrm{task}}+\lambda_BB+\lambda_cC$. The paper supplies the operator and the inequality target; quadrature resolution, truncation $K$, and coefficients $\lambda_B,\lambda_c$ are empirical. Start with spherical MNIST or directional regression, comparing this regularizer against weight decay and a standard Laplacian penalty. Record validation error, calibration, rotation robustness, gradient variance, and the percentage of batches with numerically negative $B$. Success means improved validation or calibration at equal parameter count without collapse to a constant field; failure is no improvement over the Laplacian baseline or excessive smoothing.
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.