Unverified 2026

Potential-Steered Observable Wave Layer

Implementation & benchmark of arXiv:2607.11702 — Uniform controllability for the wave equation with large potential

Usefulness5/10

Difficulty6/10

Novelty8/10

Architecture Accuracy Stability Sample efficiency Ssm Cnn Training dynamics Regularization Pde Dynamical systems Geometry Spectral theory

Source paper: Uniform controllability for the wave equation with large potential arXiv:2607.11702 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Replace homogeneous feature propagation with a discretized wave equation containing a positive, spatially varying learnable potential. The potential changes Hamiltonian trajectories so that feature energy reaches the layer's readout or sensor region instead of remaining in dynamically hidden modes. Train the potential jointly with the task objective and an empirical observability penalty.

Formulas

$$\partial_t^2v-\Delta_gv+\mathsf{V}v=b_\omega f.$$

$$\mathcal{P}_h=-h^2\Delta_g+V,\qquad p(x,\xi)=q(x,\xi)^2,\qquad q(x,\xi)=\sqrt{|\xi|_g^2+V(x)}.$$

$$\mathfrak{C}_{\mathrm{obs}}^{\mathrm{opt}}(\lambda,V,T)=\sup_{(u_0,u_1)\ne(0,0)}\frac{\mathscr{E}_{\lambda V}(u_0,u_1)}{\int_0^T\|b_\omega\partial_tu_\lambda(t)\|_{L^2(M)}^2\,dt}.$$

$$\mathcal{L}_{\mathrm{obs}}(V)=\frac{1}{B}\sum_{j=1}^{B}\frac{E_j}{O_j+\varepsilon},\qquad E_j=\|\nabla_hu_0^{(j)}\|_2^2+\|w_0^{(j)}\|_2^2+\lambda\|\sqrt{V_\theta}u_0^{(j)}\|_2^2,\qquad O_j=\Delta t\sum_{n=0}^{N}\|b_\omega\odot w_n^{(j)}\|_2^2.$$

Mathematical statement

The paper studies the controlled wave equation $\partial_t^2v-\Delta_gv+\mathsf{V}v=b_\omega f$, where $v(t,x)$ is the state, $\Delta_g$ is the Laplace-Beltrami operator, $\mathsf{V}=\lambda V$ is a positive spatial potential, $b_\omega(x)\ge 0$ is an observation mask, and $f$ is a control. For the uncontrolled solution $u_\lambda$, the optimal observability cost is $\mathfrak{C}_{\mathrm{obs}}^{\mathrm{opt}}(\lambda,V,T)=\sup_{(u_0,u_1)\ne(0,0)}\mathscr{E}_{\lambda V}(u_0,u_1)/\int_0^T\|b_\omega\partial_tu_\lambda(t)\|_{L^2(M)}^2\,dt$, where $\mathscr{E}_{\lambda V}$ is the initial wave energy. The semiclassical operator $\mathcal{P}_h=-h^2\Delta_g+V$ has principal symbol $p=q^2$, with $q(x,\xi)=\sqrt{|\xi|_g^2+V(x)}$, where $h$ is the semiclassical scale, $x$ is position, and $\xi$ is momentum. Its Hamiltonian rays satisfy $\dot{x}=\partial_\xi q=\xi/q$ and $\dot{\xi}=-\partial_xq=-\nabla V/(2q)$. The paper proves that the potential-dependent geometric control condition GCC_V^+ is necessary and sufficient for a uniform observability cost as the potential scale grows. The neural adaptation uses the same energy-to-observation ratio as a minibatch regularizer and uses sampled Hamiltonian rays as a practical surrogate for GCC_V^+.

Implementation notes

1. Integration point: insert the module between CNN stages or use it as a spatial state-space block for a feature map $z\in\mathbb{R}^{B\times C\times H\times W}$. Maintain displacement $u_n$ and velocity $w_n$. Parameterize a positive potential by $V_\theta=V_{\min}+\operatorname{softplus}(a_\theta)$, using one spatial map shared across channels or a low-rank channel mixture. Define a fixed observation mask $b_\omega$, such as a central crop, sparse sensor locations, or the locations consumed by the next stage.

2. Pseudocode: set $u_0=z$, $w_0=0$. For $n=0,\ldots,N-1$, perform leapfrog integration: $w_{n+1/2}=w_n+(\Delta t/2)(\Delta_hu_n-V_\theta\odot u_n)$; $u_{n+1}=u_n+\Delta t w_{n+1/2}$; $w_{n+1}=w_{n+1/2}+(\Delta t/2)(\Delta_hu_{n+1}-V_\theta\odot u_{n+1})$. Return $u_N$, or concatenate $u_N$ and $w_N$. For random probe initial states, compute $E_j$ and $O_j$ from the formula above and add $\alpha\mathcal{L}_{\mathrm{obs}}$ to the task loss. Optionally sample rays using $q=\sqrt{\|\xi\|^2+V_\theta(x)}$, $x_{n+1}=x_n+\Delta t\xi_n/q_n$, and $\xi_{n+1}=\xi_n-\Delta t\nabla V_\theta(x_n)/(2q_n)$, penalizing rays with low average $b_\omega(x_n)$.

3. Computed versus estimated: the finite-difference wave update and the discrete energy/observation ratio are computed exactly for the chosen grid. GCC_V^+ is not evaluated symbolically; estimate it with 8--32 random probes per minibatch or with random Hamiltonian-ray rollouts. For a stronger worst-case estimate, apply 3--5 power iterations to the linear map from initial states to masked velocity observations. Choose $\Delta t$ below the CFL limit inferred from the grid spacing and $\max V_\theta$, and clip the potential range during the first experiments.

4. First cheap experiment: train CIFAR-10 with a small ResNet-18 or four-layer CNN. Compare a standard convolutional block, a homogeneous wave block with constant $V$, and the learned-potential wave block at matched parameter count and FLOPs. Use a fixed 25% observation mask and compare training with and without $\mathcal{L}_{\mathrm{obs}}$. Record validation accuracy, loss descent at equal FLOPs, gradient-norm variance, the minimum probe ratio $O_j/E_j$, and robustness after shifting the mask or reducing it to 10%. A positive result is faster loss descent or higher accuracy, at least a 2x increase in the worst-probe observability ratio, and fewer exploding-gradient events.

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.