Flux-Frequency Homogeneity Regularizer
Source paper: An Almgren-type formula for planar $p$-harmonic functions arXiv:2608.30847 ⓘ · analyzed Sep 1, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Add a differentiable penalty that encourages a neural implicit field to have a controlled local homogeneity degree across concentric spatial scales. The penalty compares the flux-normalized frequency at adjacent radii, optionally targeting a desired degree k, so the network is discouraged from producing scale-inconsistent or oscillatory local geometry.
Formulas
Mathematical statement
The paper defines the planar p-harmonic equation div(|Du|^{p-2}Du)=0 for 1<p<infty, where u is a scalar field, Du is its spatial gradient, and div is divergence. For a center x_0 and radius r, it defines the flux-normalized frequency N_*(x_0,r)=r E_p(x_0,r)/H_p(x_0,r), with E_p(x_0,r)=integral over B_r(x_0) of |Du(x)|^p dx and H_p(x_0,r)=integral over the boundary partial B_r(x_0) of |Du(x)|^{p-2}(u(x)-u(x_0))^2 dS. Here B_r(x_0) is the disk or ball of radius r, partial B_r(x_0) is its boundary, dS is surface measure, and p is a tunable exponent. The paper states that N_* exactly recovers the degree of homogeneity for homogeneous p-harmonic profiles. We use this scale behavior as a regularization signal: when the desired degree k is known, penalize (N_*(x_0,r)-k)^2; when it is unknown, penalize log-scale variation [N_*(x_0,2r)-N_*(x_0,r)]^2. Neural fields are not automatically p-harmonic, so this is an inductive bias rather than a guaranteed monotonicity theorem.
Implementation notes
Integrate the regularizer into the loss of a coordinate MLP f_theta(x), neural SDF, or other implicit field with spatial automatic differentiation. At each training step, sample centers x_0 from valid domain locations and radii r_j=r_min*2^j that remain inside the domain. For every radius, draw M approximately uniform interior points y_m in the disk using y_m=x_0+r_j*sqrt(U_m)*[cos(theta_m),sin(theta_m)], and Q approximately uniform boundary points z_q=x_0+r_j*[cos(phi_q),sin(phi_q)]. Evaluate u=f_theta(x), retain u_0=f_theta(x_0), and obtain Du through automatic differentiation. Estimate the interior energy by E_hat=(pi*r_j^2/M)*sum_m ||Du(y_m)||^p. Estimate the boundary term by H_hat=(2*pi*r_j/Q)*sum_q (||Du(z_q)||+epsilon)^(p-2)*(u(z_q)-u_0)^2. Form N_hat[j]=r_j*E_hat/(H_hat+epsilon), where epsilon is a small positive stabilizer such as 1e-8 times a running energy scale. Compute L_freq as the mean squared difference between adjacent N_hat values, optionally plus lambda_k times the squared deviation from a known degree k. Optimize L_task+lambda_freq*L_freq. The paper's mathematical object is N_*; Monte Carlo quadrature, epsilon, p, and loss weights are engineering choices. Begin with p=2 or p=3 and gradient clipping because small boundary denominators can create large derivatives. The first experiment should fit a 2D coordinate MLP to analytic homogeneous fields such as u(x,y)=x^2-y^2 or u=r^k cos(k theta), comparing vanilla MSE with MSE plus the regularizer at equal model size and FLOPs. Evaluate interpolation and extrapolation error across radii, variance of N_hat across scales, gradient smoothness, and optimization stability. Then test a neural SDF on simple shapes under sparse point supervision; success means lower unseen-point error and fewer small-scale geometric artifacts without worse zero-level-set accuracy.
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.