# Schrodinger spectral-gap regularizer for learned metrics

- ID: 146
- Canonical URL: https://synthcore.org/idea/146/schrodinger-spectral-gap-regularizer-for-learned-metrics
- API JSON: https://synthcore.org/api/idea/146.json
- API Markdown: https://synthcore.org/api/idea/146.md
- Verification status: unverified
- Source: [arXiv:2608.24853](https://arxiv.org/abs/2608.24853)
- Category: regularization
- Solves: stability, generalization
- ML areas: embedding, regularization, training-dynamics
- Math tags: differential-geometry, spectral-theory, functional-analysis, geometry
- Ratings: usefulness 5/10; difficulty 8/10; novelty 8/10

## Idea description

Equip a learned embedding with a pullback Riemannian metric and regularize the bottom eigenvalue of the operator -Δ_g+γ scal_g. The regularizer searches for localized functions with low Dirichlet energy plus curvature potential, thereby penalizing unstable regions that ordinary Jacobian-norm penalties may miss.

## Mathematical statement

The paper defines the γ-spectral constant of an oriented Riemannian m-manifold (M,g) as Λ_γ(g)=inf{∫_M(|∇u|²+γ scal_g u²)dV : u∈H_0^1(M), ∫_M u²dV=1}, where u is a compactly supported Sobolev test function, |∇u|²=g^{ij}∂_i u∂_j u is the Dirichlet energy, scal_g is scalar curvature, dV is Riemannian volume, and H_0^1(M) is the completion of compactly supported smooth functions in the Sobolev norm. Equivalently, Λ_γ is the infimum of the Rayleigh quotient of the Schrödinger operator -Δ_g+γ scal_g. The paper considers γ>(m-1)/(4m). For an encoder f_θ, use the regularized pullback metric g_θ(x)=J_f(x)^T J_f(x)+εI, where J_f is the encoder Jacobian and ε>0 ensures positive definiteness. With localized basis functions φ_k, approximate u by u_a=Σ_k a_kφ_k and estimate Λ_γ by the smallest generalized eigenvalue of A=L_g+γS_g against the mass matrix M_g: Λ̂_γ=λ_min(A,M_g). Here M_{kl}=Σ_i w_iφ_k(x_i)φ_l(x_i), L_{kl}=Σ_i w_i g^{pq}(x_i)∂_pφ_k(x_i)∂_qφ_l(x_i), and S_{kl}=Σ_i w_i scal_g(x_i)φ_k(x_i)φ_l(x_i).

## Key formulas

- $$\Lambda_{\gamma}(g)=\inf\Bigl\{\int_{M}\bigl(|\nabla u|^{2}+\gamma\,\mathrm{scal}_{g}\,u^{2}\bigr)\,dV\colon u\in H_{0}^{1}(M),\ \int_{M}u^{2}\,dV=1\Bigr\}.$$
- $$\Lambda_\gamma(g)=\inf_{u\neq 0}\frac{\int_M\left(g^{ij}\partial_i u\partial_j u+\gamma\,\mathrm{scal}_g u^2\right)dV_g}{\int_Mu^2dV_g},\qquad \gamma>\frac{m-1}{4m}.$$
- $$g_\theta(x)=J_{f_\theta}(x)^\top J_{f_\theta}(x)+\varepsilon I.$$
- $$\widehat\Lambda_\gamma=\lambda_{\min}(L_g+\gamma S_g,M_g),\quad (M_g)_{kl}=\sum_iw_i\phi_k(x_i)\phi_l(x_i),\quad (L_g)_{kl}=\sum_iw_i g^{pq}(x_i)\partial_p\phi_k(x_i)\partial_q\phi_l(x_i).$$

## Implementation notes

(1) Integration point: insert the regularizer after the encoder forward pass during training. For a minibatch x_i, compute embeddings z_i=f_θ(x_i), retain the input or local chart coordinates, and obtain the Jacobian J_f(x_i) with forward- or reverse-mode autodiff. Set g_i=J_f(x_i)^T J_f(x_i)+εI with ε=10^-4. (2) Choose K centers c_k from the current embedding minibatch and define localized Gaussian basis functions φ_k(x_i)=exp(-||z_i-c_k||²/(2σ²)). Compute their derivatives by autodiff. Use minibatch weights w_i=1/B and build M, L, and S. For the first MVP, set S=0; the exact curvature version requires differentiating the metric and estimating scalar curvature, which is substantially more expensive. For each batch, solve the K×K generalized eigenproblem (L+γS)v=λMv, or use 3–5 inverse-iteration steps. Add L_spec=softplus(τ-λ_min) to the task loss, with γ=(m-1)/(4m)+0.05 and target margin τ. (3) The paper-derived quantity is the Rayleigh quotient and its Galerkin discretization. The network Jacobian, RBF centers, bandwidth σ, quadrature weights, and curvature estimate are empirical approximations. Use gradient clipping and stop-gradient through eigenvectors if eigenderivative instability appears. (4) First experiment: train a two-layer MLP encoder/classifier on MNIST or CIFAR-10, comparing baseline, Jacobian Frobenius regularization, and L_spec at matched wall-clock and parameter budgets. Measure accuracy, Gaussian-noise and corruption accuracy, Jacobian norms, estimated minimum Rayleigh quotient, and per-step variance. The signal of success is increased λ_min and improved perturbation robustness at equal clean accuracy, with less than 20% training overhead.

## Disclaimer

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