# Calogero Spectral Barrier for Recurrent Dynamics

- ID: 2966
- Canonical URL: https://synthcore.org/idea/2966/calogero-spectral-barrier-for-recurrent-dynamics
- API JSON: https://synthcore.org/api/idea/2966.json
- API Markdown: https://synthcore.org/api/idea/2966.md
- Verification status: unverified
- Source: [arXiv:2609.00164](https://arxiv.org/abs/2609.00164)
- Category: dynamics
- Solves: stability, accuracy, generalization
- ML areas: rnn, ssm, optimizer, training-dynamics
- Math tags: statistical-mechanics, random-matrix, spectral-theory, dynamical-systems
- Ratings: usefulness 6/10; difficulty 5/10; novelty 8/10

## Idea description

Apply an inverse-square Calogero barrier to the eigenvalues of a recurrent or state-space transition Jacobian, discouraging unstable eigenvalues and pathological eigenvalue collisions without forcing the matrix to be Hermitian. The paper's non-Hermitian scattering picture motivates treating the spectrum as correlated rather than assuming an ordinary pairwise Coulomb gas; the inverse-square term is used as a local, computable surrogate for that mechanism.

## Mathematical statement

For a non-Hermitian matrix with eigenvalues $z_i$, the paper identifies the spectral density, after a Vandermonde factor, with a Calogero scattering state. The underlying Calogero operator is $\mathcal{H}_{C}=-\sum_{i=1}^{N}\partial_{x_i}^{2}+g\sum_{i<j}(x_i-x_j)^{-2}$, where $x_i$ are particle coordinates and $g$ is the symmetry-dependent inverse-square coupling. For a neural transition map $h_{t+1}=F_\theta(h_t,u_t)$, let $J_t=\partial F_\theta/\partial h_t$ and let $z_i(J_t)$ be its eigenvalues. Use the adapted regularizer $R_C(J_t)=\sum_{i<j}g/(|z_i-z_j|^2+\epsilon^2)$, with softening $\epsilon>0$, together with a contraction penalty $R_\rho=[\max(0,\rho(J_t)-(1-\delta))]^2$, where $\rho(J_t)=\max_i|z_i|$ and $\delta>0$ is the desired stability margin. The discrete-time stability prediction is that perturbations decay when $\rho(J_t)\leq 1-\delta$ and grow exponentially when the effective spectral radius exceeds one; the Calogero term predicts suppression of near-collision events.

## Key formulas

- $$\mathcal{H}_{C}=-\sum_{i=1}^{N}\frac{\partial^2}{\partial x_i^2}+g\sum_{i<j}\frac{1}{(x_i-x_j)^2},$$
- $$R_C(J_t)=\sum_{i<j}\frac{g}{|z_i(J_t)-z_j(J_t)|^2+\epsilon^2},\qquad R_\rho(J_t)=\left[\max\left(0,\rho(J_t)-(1-\delta)\right)\right]^2,$$
- $$h_{t+k}-\tilde h_{t+k}\approx J_{t+k-1}\cdots J_t(h_t-\tilde h_t),\qquad \|h_{t+k}-\tilde h_{t+k}\|\lesssim (1-\delta)^k\|h_t-\tilde h_t\|\ \text{if}\ \rho(J_s)\leq 1-\delta.$$

## Implementation notes

Integrate the method into a small recurrent network, linear state-space model, or neural ODE discretization. At every K training steps, collect representative hidden states from a minibatch and compute the transition Jacobian J with automatic differentiation; for a linear state-space model this is simply the learned transition matrix A. Compute its eigenvalues z_i exactly when the hidden dimension is below roughly 256. For larger systems, estimate leading eigenvalues with Arnoldi iteration and use a random subset of eigenvalue pairs. Add $\lambda_C R_C+\lambda_\rho R_\rho$ to the task loss, detach the spectrum between updates if eigenvalue gradients are unstable, and clip the total gradient. Pseudocode is: sample states; form J; eigvals = eig(J); collision = sum over i<j of g/(abs(eigvals[i]-eigvals[j])^2+eps^2); radius = max(abs(eigvals)); loss = task_loss + lambda_C*collision + lambda_rho*relu(radius-(1-delta))^2; backpropagate. The Calogero operator and inverse-square mechanism come from the paper; the Jacobian penalty, softening, and contraction term are engineering adaptations. First test on copy/addition and sequential MNIST with equal parameter, wall-clock, and optimizer budgets against the same model with only spectral-radius control, plus an ablation removing $R_C$. Pre-register that the measured divergence boundary occurs at spectral radius 1 within 20%, hidden perturbation norms decay with slope at least $\log(1-\delta)$ over long rollouts, and the number of eigenvalue pairs with spacing below a fixed threshold $s$ is reduced by at least 30% at fixed radius. Reject the transfer if the Calogero penalty does not reduce near-collision frequency, or if it improves neither rollout decay nor validation loss while costing more than 10% training time.

## Disclaimer

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