Unverified 2026

Hodge-factorized neural vector field

Implementation & benchmark of arXiv:2608.15789 — The Hodge structure of Berry-phase transport: topology, geometry, and noise

Usefulness6/10

Difficulty5/10

Novelty6/10

Geometry Stability Generalization Accuracy Diffusion Diffusion sampling World model Training dynamics Differential geometry Geometry Topology Harmonic analysis Functional analysis

Source paper: The Hodge structure of Berry-phase transport: topology, geometry, and noise arXiv:2608.15789 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Parameterize a periodic neural vector field as the sum of a harmonic global drift, an exact gradient field, and a co-exact divergence-free field. This gives separate control over conservative attraction/repulsion, rotational transport, and domain-wide drift, potentially preventing one unconstrained MLP from entangling incompatible dynamics.

Formulas

$$\Omega=\underbrace{\gamma}_{\text{harmonic}}+\underbrace{\mathrm{d}\mathcal{A}}_{\text{exact}}+\underbrace{\delta\mathcal{B}}_{\text{co-exact}},\qquad \Delta\gamma=0.$$

$$\langle\alpha,\beta\rangle=\int_{\mathbb{T}^{d}}\alpha\wedge\star\beta,\qquad \langle\gamma,\mathrm{d}\mathcal{A}\rangle=\langle\gamma,\delta\mathcal{B}\rangle=\langle\mathrm{d}\mathcal{A},\delta\mathcal{B}\rangle=0.$$

$$\hat v_{\mathrm{harm}}(0)=\hat v(0),\qquad \hat v_{\mathrm{exact}}(k)=\frac{kk^{\mathsf T}}{\|k\|^{2}}\hat v(k),\qquad \hat v_{\mathrm{coexact}}(k)=\left(I-\frac{kk^{\mathsf T}}{\|k\|^{2}}\right)\hat v(k),\quad k\neq0.$$

$$v(x)=h+\nabla\phi(x)+w(x),\qquad h=\frac{1}{|\mathbb{T}^{d}|}\int_{\mathbb{T}^{d}}v(x)\,dx,\qquad \nabla\cdot w=0,\quad \int_{\mathbb{T}^{d}}\phi(x)\,dx=0.$$

Mathematical statement

The paper uses the Hodge decomposition on the closed oriented flat torus $\mathbb{T}^{d}$: $\Omega=\gamma+\mathrm{d}\mathcal{A}+\delta\mathcal{B}$, where $\Omega$ is a differential form, $\gamma$ is harmonic with $\Delta\gamma=0$, $\mathcal{A}$ is a lower-degree form, $\mathcal{B}$ is a higher-degree form, $\mathrm{d}$ is the exterior derivative, and $\delta$ is the co-derivative. The three terms are mutually orthogonal under $\langle\alpha,\beta\rangle=\int_{\mathbb{T}^{d}}\alpha\wedge\star\beta$, and the harmonic term represents the cohomology class and the globally nonlocal component. For an adapted periodic vector field $v:\mathbb{T}^{d}\to\mathbb{R}^{d}$, use the equivalent 1-form decomposition $v=h+\nabla\phi+w$, where $h$ is constant, $\phi$ is a zero-mean scalar potential, and $w$ is divergence-free and zero-mean. In Fourier space, for wavevector $k\neq0$, the exact projector is $P_{\mathrm{ex}}(k)=kk^{\mathsf T}/\|k\|^{2}$, the co-exact projector is $P_{\mathrm{co}}(k)=I-P_{\mathrm{ex}}(k)$, and the harmonic component is the zero-frequency coefficient $\hat v(0)$. Orthogonality yields $\|v\|_{2}^{2}=\|h\|_{2}^{2}+\|\nabla\phi\|_{2}^{2}+\|w\|_{2}^{2}$.

Implementation notes

Integrate this into the vector field of a periodic neural ODE, flow-matching model, or continuous-time diffusion sampler. The model should receive a periodic state $x\in[0,1]^{d}$ and time $t$, and output three quantities: a learned constant vector $h(t)$, a scalar potential $\phi_\theta(x,t)$, and an unconstrained vector pre-field $u_\theta(x,t)$. Construct the exact branch as $e_\theta=\nabla_x\phi_\theta$, using autograd or spectral derivatives, and construct the co-exact branch by projecting $u_\theta$ with a periodic FFT. The pseudocode is: phi,u = net(x,t); e = grad_x(phi); U = fftn(u); for k != 0: W_hat[k] = (I - outer(k,k)/dot(k,k)) @ U[k]; W_hat[0] = 0; w = ifftn(W_hat).real; v = h(t)+e+w. Set the zero Fourier mode of $\phi$ to zero to remove gauge ambiguity. Train with the original vector-field or flow-matching loss on $v$, optionally adding $\lambda_{\mathrm{div}}\|\nabla\cdot w\|_{2}^{2}$, $\lambda_h\|h\|^{2}$, or sector-specific target losses. The mathematics determines the exact Fourier projectors and orthogonality; the implementation estimates derivatives and FFT coefficients on minibatch spatial points or a fixed periodic grid. Start with a 2D toy mixture on the torus whose ground-truth field is $v_*(x)=h_*+\nabla\phi_*(x)+J\nabla\psi_*(x)$, where $J(a,b)=(-b,a)$. Compare an unconstrained MLP, a gradient-only model, and this factorized model at equal parameter count. Measure flow-matching loss, divergence and curl reconstruction error, long-horizon ODE energy drift, solver function evaluations, and generated-sample quality. The first success signal is lower long-horizon trajectory error or fewer solver steps at equal one-step training loss; on a mixture with known rotational motion, the co-exact branch should reduce trajectory damping and mode collapse relative to a gradient-only baseline.

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.