Unverified 2026

Fourier equilibrium projection for tensor fields

Implementation & benchmark of arXiv:2608.12364 — Single-axis high-energy X-ray diffraction tomography for elastic residual strain: uniqueness and stability of solutions in the presence of equilibrium constraints

Usefulness6/10

Difficulty5/10

Novelty6/10

Geometry Stability Accuracy Generalization Ssm Mlp Loss Regularization Pde Linear algebra Harmonic analysis

Source paper: Single-axis high-energy X-ray diffraction tomography for elastic residual strain: uniqueness and stability of solutions in the presence of equilibrium constraints arXiv:2608.12364 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Insert a differentiable Fourier-domain layer after a network predicts a symmetric strain field, projecting every frequency onto the subspace satisfying isotropic mechanical equilibrium. The projection is a closed-form least-squares correction, so the network cannot spend capacity representing large equilibrium violations and the resulting field is physically admissible by construction.

Formulas

$$\sigma=2\mu\epsilon+\lambda\operatorname{tr}(\epsilon)I,\qquad \nabla\cdot\sigma=0,$$

$$B(y)e=\begin{bmatrix}(2\mu+\lambda)y_1&\lambda y_1&\lambda y_1&2\mu y_2&2\mu y_3&0\\ \lambda y_2&(2\mu+\lambda)y_2&\lambda y_2&2\mu y_1&0&2\mu y_3\\ \lambda y_3&\lambda y_3&(2\mu+\lambda)y_3&0&2\mu y_1&2\mu y_2\end{bmatrix}e=0,$$

$$P(y)=I-B(y)^T\bigl(B(y)B(y)^T+\tau^2I_3\bigr)^{-1}B(y),\qquad \widetilde e(y)=P(y)e_\theta(y),$$

$$\det A(y)=4\alpha y_1y_2y_3^2(y_1^2-y_2^2).$$

Mathematical statement

The paper combines a symmetric strain tensor $\epsilon(x)\in\mathbb{R}^{3\times3}$ with isotropic Hooke law $\sigma=2\mu\epsilon+\lambda\operatorname{tr}(\epsilon)I$, where $\sigma$ is stress, $\mu>0$ and $\lambda$ are Lamé constants, and imposes equilibrium $\nabla\cdot\sigma=0$. For a Fourier frequency $y\in\mathbb{R}^3$, differentiation becomes multiplication by $iy$, giving the linear constraint $B(y)e(y)=0$, where $e=(\epsilon_{11},\epsilon_{22},\epsilon_{33},\epsilon_{12},\epsilon_{13},\epsilon_{23})^T$. Explicitly, $B(y)$ is the matrix encoding $(2\mu\epsilon+\lambda\operatorname{tr}(\epsilon)I)y=0$. The exact orthogonal projection onto this nullspace is $P(y)=I-B(y)^T(B(y)B(y)^T)^\dagger B(y)$, with ${}^\dagger$ the Moore-Penrose pseudoinverse. In implementation, a ridge $\tau^2I$ replaces the pseudoinverse. The paper also obtains $\det A(y)=4\alpha y_1y_2y_3^2(y_1^2-y_2^2)$, where $\alpha\neq0$ is an elastic parameter; its zero set identifies characteristic planes.

Implementation notes

(1) Integration point: let a neural operator, 3D CNN, or coordinate MLP predict six channels representing the symmetric strain components on a regular 3D grid. Apply a 3D FFT to the six channels, pack each Fourier coefficient into $e_\theta(y)$, apply the projection independently at every frequency, and inverse FFT before computing the task loss. Use periodic padding or a padded computational box; handle the zero-frequency coefficient separately because equilibrium imposes no constraint there. (2) Pseudocode: eps_hat = fft3(net(inputs)); for each frequency y: construct B(y) from mu, lambda; P = I - B.T @ inverse(B @ B.T + tau**2 * I3) @ B; eps_hat_proj[y] = P @ eps_hat[y]; eps_eq = ifft3(eps_hat_proj); loss = task_loss(eps_eq, target). (3) The matrix $B(y)$, projection, and elastic constants come from the mathematics. The ridge $\tau$, padding, and optional blend output = beta*eps_eq + (1-beta)*eps_raw are empirical choices. Measure the equilibrium residual $R=\|\mathcal{F}^{-1}[B(y)\widetilde e(y)]\|_2/(\|\widetilde e\|_2+10^{-8})$. (4) First experiment: train a small Fourier neural operator on synthetic 3D strain fields generated from random equilibrated stress potentials, with noisy targets. Compare an unconstrained model, an equilibrium-penalty model, and this projection at equal parameter count and FLOPs. Success means at least a tenfold lower equilibrium residual and lower held-out reconstruction error without slower loss descent.

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.