Unverified 2026

Anisotropic Anharmonic Diffusion Layer

Implementation & benchmark of arXiv:2607.17580 — Sharp Time-Decay Estimates for Fractional Heat Semigroups Associated with Polynomial Anharmonic Oscillators

Usefulness6/10

Difficulty5/10

Novelty7/10

Architecture Stability Accuracy Regularization Cnn Ssm Regularization Training dynamics Pde Spectral theory Harmonic analysis Numerical analysis

Source paper: Sharp Time-Decay Estimates for Fractional Heat Semigroups Associated with Polynomial Anharmonic Oscillators arXiv:2607.17580 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Insert a learnable semigroup layer that evolves features according to a positive operator combining frequency damping and spatially varying confinement. Unlike isotropic Gaussian smoothing, the layer can damp selected frequencies differently along different axes and can suppress activations in learned spatial regions, while the positive-semigroup construction prevents amplification.

Formulas

$$\mathcal{H}_{P,Q}:=Q(D)+P(x),\qquad Q\in\mathcal{P}_{2\ell},\quad P\in\mathcal{P}_{2k},$$

$$\liminf_{|x|\rightarrow\infty}\frac{P(x)}{|x|^{2k}}>0,$$

$$e^{-t\mathcal{H}_{P,Q}^{s}},\qquad \|e^{-t\mathcal{H}_{P,Q}^{s}}\|\ \text{has large-time factor}\ e^{-t\lambda_{0}^{s}},$$

$$e^{-t(Q(D)+P(x))}\approx\left[e^{-\Delta t Q(D)}e^{-\Delta t P(x)}\right]^M,\qquad \Delta t=t/M.$$

Mathematical statement

The paper studies the Hamiltonian $\mathcal{H}_{P,Q}=Q(D)+P(x)$, where $x\in\mathbb{R}^{n}$ is the spatial coordinate, $D=-i\nabla_x$, $P(x)$ is a real polynomial satisfying $\liminf_{|x|\to\infty}P(x)/|x|^{2k}>0$, and $Q(\xi)$ is a real polynomial satisfying $\liminf_{|\xi|\to\infty}Q(\xi)/|\xi|^{2\ell}>0$. After adding a constant if needed, $\mathcal{H}_{P,Q}$ is positive and self-adjoint. Its fractional heat semigroup is $e^{-t\mathcal{H}_{P,Q}^{s}}$, with diffusion time $t>0$ and exponent $s>0$. The paper states that large-time behavior contains the factor $e^{-t\lambda_0^s}$, where $\lambda_0$ is the smallest eigenvalue. The ML layer uses the same positive operator: $Q(D)$ becomes a Fourier multiplier $Q(\xi)$, while $P(x)$ becomes pointwise feature damping. For a computable first version, use $s=1$ and the Lie-Trotter approximation $e^{-t(Q(D)+P(x))}\approx[e^{-\Delta t Q(D)}e^{-\Delta t P(x)}]^M$, with $\Delta t=t/M$. Positivity of $Q$ and $P$ makes each factor non-expansive in the corresponding discretized norm; the learned minimum eigenvalue or an explicit shift controls long-time contraction.

Implementation notes

Integrate the layer after a convolution block or between transformer/SSM blocks, operating on a feature tensor $X\in\mathbb{R}^{B\times C\times H\times W}$; a 1D version applies to token or time sequences. Parameterize a nonnegative anisotropic kinetic symbol, for example $Q(\xi_x,\xi_y)=\epsilon+\sum_{j=1}^{2}a_j\xi_j^2+b_j\xi_j^4$, with $a_j,b_j\geq0$ enforced by softplus and $\epsilon>0$. Parameterize a nonnegative spatial potential $P(x,y)=\epsilon_p+c_x(x-c_x^0)^2+c_y(y-c_y^0)^2+d(x^2+y^2)^2$, also using softplus coefficients; share it across channels initially, then test channel-wise gates. For each forward pass, compute the real FFT of every channel, multiply by $\exp[-\Delta t Q(\xi)]$, inverse FFT, and multiply pointwise by $\exp[-\Delta t P(x)]$; repeat $M$ times. Use a gated residual, $Y=X+\alpha(\operatorname{Diffuse}(X)-X)$, with $\alpha$ initialized near zero. Pseudocode: Z=X; for m in range(M): Z=ifft(exp(-dt*Q(freq))*fft(Z)); Z=exp(-dt*P(grid))*Z; Y=X+alpha*(Z-X). The paper supplies the Hamiltonian decomposition, coercivity constraints, fractional semigroup, and eigenvalue-controlled contraction; FFT discretization and polynomial parameterization are adaptations. Estimate the smallest discrete eigenvalue with 5-10 Lanczos iterations on the composed operator, or monitor maximum feature amplification on validation batches. Start on CIFAR-10 with a small ResNet-18, comparing against Gaussian blur, depthwise convolution, and an identity residual at equal wall-clock budget. Measure clean accuracy, accuracy under additive high-frequency noise, gradient norm variance, and failed or divergent training runs. Success means improved noisy-image accuracy or equal accuracy with more stable optimization and no feature-norm growth; also test whether increasing diffusion time $t$ degrades accuracy more slowly than Gaussian 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.