# Linearized Observability Regularizer

- ID: 1933
- Canonical URL: https://synthcore.org/idea/1933/linearized-observability-regularizer
- API JSON: https://synthcore.org/api/idea/1933.json
- API Markdown: https://synthcore.org/api/idea/1933.md
- Verification status: unverified
- Source: [arXiv:2608.07177](https://arxiv.org/abs/2608.07177)
- Category: regularization
- Solves: accuracy, stability, sample-efficiency
- ML areas: world-model, loss, regularization
- Math tags: pde, functional-analysis, control-theory, optimization
- Ratings: usefulness 6/10; difficulty 6/10; novelty 7/10

## Idea description

Train a neural coefficient-recovery model with an additional loss that rewards observation sensitivity in every learnable coefficient direction. Instead of only minimizing the reconstruction error of the observed trajectory, explicitly discourage a nearly singular parameter-to-observation Jacobian, which should reduce ambiguous reconstructions and improve robustness to noise.

## Mathematical statement

The paper starts from the controlled evolution equation \(\mathcal{D}_{t}u(t)+(\mathcal{A}+\mathcal{B}(t)+\mathcal{H}(\theta))u(t)=f(t)\), with initial state \(u(0)=u_{0}(\theta)\), and observations \(y(t)=\mathcal{C}u(t)\). Here \(u(t)\) is the state, \(\theta\) is a space-dependent coefficient, \(\mathcal{D}_{t}\) is a time derivative operator, \(\mathcal{A}\) is the spatial operator, \(\mathcal{B}(t)\) is a time-dependent control operator, \(\mathcal{H}(\theta)\) is the coefficient-dependent operator, and \(\mathcal{C}\) is a possibly non-local observation operator. Linearizing around \(\theta_{\mathrm{ref}}\), a perturbation \(\delta u\) generated by coefficient perturbation \(\delta\theta\) satisfies \(\mathcal{D}_{t}\delta u+(\mathcal{A}+\mathcal{B}(t)+\mathcal{H}(\theta_{\mathrm{ref}}))\delta u=-\mathcal{H}'(\theta_{\mathrm{ref}})[\delta\theta]u\), with \(\delta u(0)=u_{0}'(\theta_{\mathrm{ref}})\delta\theta\), and produces \(\delta y(t)=\mathcal{C}\delta u(t)\). If the only perturbation producing zero observations is \(\delta\theta=0\), the linearized inverse map is injective. For a finite-dimensional neural parameter vector \(\phi\), define \(J(t)=\partial y(t;\phi)/\partial\phi\) and the empirical observability Gramian \(G=\int_{0}^{T}J(t)^{\mathsf T}WJ(t)\,dt+\lambda I\), where \(W\succeq0\) weights observation channels and \(\lambda>0\) stabilizes the inverse. Add \(L_{\mathrm{obs}}=-\log\det(G)\), or the cheaper surrogate \(L_{\mathrm{obs}}=\operatorname{tr}(G^{-1})\), to the reconstruction loss.

## Key formulas

- $$\mathcal{D}_{t}u(t)+(\mathcal{A}+\mathcal{B}(t)+\mathcal{H}(\theta))u(t)=f(t),\qquad u(0)=u_{0}(\theta),\qquad y(t)=\mathcal{C}u(t).$$
- $$\mathcal{D}_{t}\delta u(t)+(\mathcal{A}+\mathcal{B}(t)+\mathcal{H}(\theta_{\mathrm{ref}}))\delta u(t)=-\mathcal{H}'(\theta_{\mathrm{ref}})[\delta\theta]u(t),\qquad \delta u(0)=u_{0}'(\theta_{\mathrm{ref}})\delta\theta,\qquad \delta y(t)=\mathcal{C}\delta u(t).$$
- $$G(\phi)=\sum_{k=1}^{K}\Delta t_k\,J(t_k;\phi)^{\mathsf T}WJ(t_k;\phi)+\lambda I,\qquad J(t_k;\phi)=\frac{\partial y(t_k;\phi)}{\partial\phi},$$
- $$L(\phi)=L_{\mathrm{data}}(\phi)+\beta\bigl[-\log\det G(\phi)\bigr]\quad\text{or}\quad L(\phi)=L_{\mathrm{data}}(\phi)+\beta\,\operatorname{tr}\bigl(G(\phi)^{-1}\bigr).$$

## Implementation notes

1. Integration point: use this as a training-time regularizer for a neural inverse model \(\phi=g_\psi(x_{\mathrm{obs}})\), where \(\phi\) parameterizes a spatial coefficient field, such as a CNN-decoded diffusion coefficient or a small latent vector controlling a differentiable PDE/ODE simulator. Run the simulator at the predicted \(\phi\), collect observations \(y(t_k)\), and differentiate those observations with respect to \(\phi\). Do not apply the term to the network weights directly; it should condition the physical coefficient coordinates being reconstructed. 2. Pseudocode: `phi = inverse_net(y_obs); y_hat = simulator(phi, control=B); J = jacobian(y_hat.reshape(K,-1), phi); G = lambda*I; for k in range(K): G += dt[k] * J[k].T @ W @ J[k]; loss = mse(y_hat,y_obs) + beta*(-slogdet(G)); loss.backward(); optimizer.step()`. For large \(\phi\), use 8--16 random probe vectors and estimate \(\operatorname{tr}(G^{-1})\) with conjugate-gradient solves instead of forming \(G\). 3. The paper supplies the state equation, its linearized sensitivity equation, and the injectivity interpretation. The Jacobian, Gramian, damping \(\lambda\), and coefficient parameterization are estimated or chosen computationally; use autodiff through the simulator or tangent-linear forward solves for \(J\). Normalize columns of \(J\) so the regularizer does not simply favor coefficients with large units. 4. First experiment: use a 1D diffusion inverse problem on a 64-cell grid, with \(\theta(x)\) represented by 8 Fourier or piecewise-constant coefficients, one fixed control and 8 noisy time observations. Compare an MLP/CNN inverse model trained with observation MSE alone against MSE plus the log-determinant term, at equal training steps. Evaluate coefficient relative error, error under increased observation noise, the smallest eigenvalue and condition number of \(G\), and performance from multiple random initializations. Success is lower coefficient error and a larger minimum Gramian eigenvalue without a substantial increase in forward-simulation cost.

## Disclaimer

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