# Connected Collision Energy Latent Dynamics

- ID: 2846
- Canonical URL: https://synthcore.org/idea/2846/connected-collision-energy-latent-dynamics
- API JSON: https://synthcore.org/api/idea/2846.json
- API Markdown: https://synthcore.org/api/idea/2846.md
- Verification status: unverified
- Source: [arXiv:2608.29694](https://arxiv.org/abs/2608.29694)
- Category: dynamics
- Solves: stability, generalization, accuracy
- ML areas: world-model, ssm, graph-nn, regularization
- Math tags: dynamical-systems, linear-algebra, graph-theory, control-theory
- Ratings: usefulness 6/10; difficulty 6/10; novelty 6/10

## Idea description

Construct a graph-based latent state whose velocities evolve through free-flight updates and pairwise elastic collision operators. Each collision operator is orthogonal, so total latent kinetic energy is exactly conserved; a connected interaction graph is intended to eliminate unwanted component-wise polynomial invariants and improve long-horizon stability.

## Mathematical statement

The paper uses free flight \(\dot q=p\), \(\dot p=0\), implying the first-integral equation \(f(q+tp,p)=f(q,p)\) whenever the segment remains in the free region. At a collision with unit normal \(n\), momentum changes by the orthogonal reflection \(R_n p=p-2\langle p,n\rangle n\), which preserves \(H(p)=\frac12\|p\|^2\). For hard balls with connected collision graph, every polynomial-in-momentum first integral is \(P(H)\), with \(P\) a scalar polynomial. Transfer this by representing node \(i\) with latent position \(q_i\in\mathbb R^d\) and velocity \(p_i\in\mathbb R^d\), using free drift \(q_i^+=q_i+\Delta t p_i\), and applying pairwise orthogonal operators \(p^+=R_{ij}p\). The conserved latent energy is \(H(z)=\frac12\sum_i\|p_i\|^2\). Exact orthogonality gives \(H(z^+)=H(z)\), while graph connectivity predicts collapse of low-degree invariant searches to polynomials in \(H\).

## Key formulas

- $$\dot q=p,\qquad \dot p=0,\qquad f(q+tp,p)=f(q,p)$$
- $$R_n p=p-2\langle p,n\rangle n,\qquad R_n^{\mathsf T}R_n=I$$
- $$H(p)=\frac12\|p\|^2,\qquad H(R_n p)=H(p)$$
- $$\text{connected collision graph}\ \Longrightarrow\ \text{polynomial first integrals}=\{P(H):P\in\mathbb R[x]\}$$

## Implementation notes

1. Integration point: use this as the transition operator of a graph neural state-space model or world model. Split the latent state into positions \(q_i\), velocities \(p_i\), and unconstrained content variables \(c_i\). Decode observations from \((q,p,c)\), but update \((q,p)\) with the structured dynamics. For each edge \((i,j)\), generate a normalized learned interaction direction \(n_{ij}=u_{ij}/(\|u_{ij}\|+\epsilon)\), and apply a Householder reflection or a two-node orthogonal rotation to the relative velocity. 2. Pseudocode: initialize \(q,p,c\); at every step set \(q\leftarrow q+\Delta t p\); compute edge directions and gates; for every active edge construct \(R_{ij}=I-2n_{ij}n_{ij}^{\mathsf T}\) in the affected relative-velocity subspace and set \(p\leftarrow R_{ij}p\); update \(c\) with an ordinary message-passing block; decode the target. Optionally add \(L_E=|H_t-H_0|^2\) only to measure numerical error, not to replace exact orthogonality. 3. Computed from the paper: free-flight invariance, reflection orthogonality, and the connected-graph rigidity prediction. Estimated empirically: interaction directions, edge activation, and graph connectivity. To test invariant collapse, collect many latent trajectories and fit every monomial in \((q,p)\) up to degree four to conserved quantities; regress each candidate against \(1,H,H^2\). 4. First cheap experiment: train a 4- or 8-node model on synthetic elastic-collision trajectories, comparing an MLP transition, an unconstrained GNN, and this orthogonal-collision model. Measure 500-step rollout error, energy drift, Jacobian spectral growth, and the residual dimension of the fitted polynomial invariant space. The quantitative prediction is machine-precision energy conservation for the structured model and bounded energy over arbitrarily long rollouts. In a connected graph, no independent degree-two invariant beyond \(H\) should remain; after removing one edge to create two components, two independent component energies should appear.

## Disclaimer

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