Contact-SSM: Geometric Contact State-Space Models for Long-Horizon Dynamical Sequence Modeling
$\mathbf{Contact-SSM}$ introduces a geometric continuous-time state-space architecture for long-horizon dynamical sequence forecasting. While standard SSMs (e.g., S4, Mamba) enforce strictly dissipative contractive transitions ($\mathrm{Re}(\lambda) < 0$) that rapidly damp oscillatory modes, Contact-SSM formulates latent dynamics on odd-dimensional Contact Hamiltonian manifolds $\mathbf{z} = (\mathbf{q}, \mathbf{p}, s) \in \mathbb{R}^{2d+1}$. Parameterized on the Lie algebra direct sum $\mathfrak{sp}(2d, \mathbb{R}) \oplus \mathbb{R}$ and discretized via the Generalized Symplectic Cayley Transform, it guarantees spectral stability on the complex unit disk ($|\lambda| \le 1$) with exact symplectic invariance ($|\det(\mathbf{\Phi}) - 1.0| \le 4.44 \times 10^{-16}$) while supporting $\mathcal{O}(\log L)$ parallel associative GPU prefix scans. Contact-SSM achieves a $30.6\%$ error reduction on 14-day rollouts ($H=336$) and preserves strange attractor measures under chaotic Lorenz-96 turbulence.
Authors
- Kartikey Singh (ORCID: https://orcid.org/0009-0009-1973-7532)
Institutions
- University of Delhi (IN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24377513
- Primary Topic
- Model Reduction and Neural Networks
- Type
- preprint