Research map / Probabilistic and scientific modeling / Dynamical systems
Beyond Unit-Circle Eigenvalues: Invariant Bases for Stable State Space Dynamics
Jiaqian Zhu, Yang Zhang, Junhua Ding, Xiaowei Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes invariant-basis parameterizations for state space models to stabilize long-horizon dynamics beyond merely constraining eigenvalues to the unit circle.
Problem
Diagonal state space models become unstable over long horizons, where small spectral deviations compound into information decay or uncontrolled growth, especially for conservation-preserving or non-dissipative dynamics.
Approach
Argues that constraining eigenvalues to the unit circle is insufficient for long-horizon stability; proposes invariant bases beyond unit-circle eigenvalue constraints to stabilize state space dynamics.
Key results
- Unit-circle eigenvalue constraint yields +3.31 dB gain at step 10 over short horizons, per the abstract.
- Abstract claims unit-circle constraint alone is not sufficient for long-horizon stability; full results not available in the provided text.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| constrain eigenvalues to unit modulus only (level B) instead of full Kronecker-Laplacian assembly (ablation) | diagonal recurrence · Visual N-body | momentum drift (%): 141 percent momentum drift with norm held at 4e-6 | Hurt |
| add Laplacian zero-row-sum condition (L1=0) on top of unit modulus (ablation) | KC · Visual N-body perturbation test | momentum drift (%): 110 -> 3.5e-5 | Improved |
| Kronecker structure with Laplacian vs other norm-preserving operators (Haar, dense Cayley, diagonal unitary) | operator level · 2000-step rollout | momentum drift (%): 170.7 -> 4.0e-3 | Improved |
| tie decoder to the conserved quantity instead of an unconstrained decoder (ablation) | KC · Dam-break | mass drift: 115 -> 8.3e-3 | Improved |
| assemble KC instead of any diagonal recurrence | diagonal SSM · Allen-Cahn | mass drift: 3.28 -> 2.0e-5 | Improved |
| apply output mass corrector (CPFNO) to neural operator | CPFNO · Dam-break, 100-step rollout | stability: four of five seeds diverge between steps 48 and 83 | Mixed |
| add fixed uniform mass corrector on FNO | FNO · Dam-break | stability: diverges on every seed between steps 58 and 97 | Hurt |
| add momentum conservation penalty to a free orthogonal basis (ablation) | free modal basis · Visual N-body | momentum drift: drift 44 to 932 percent on 14 of 16 seeds | Hurt |
| increase training set size of a Hamiltonian network (ablation) | Hamiltonian network · Visual N-body | momentum drift (%): 16.9 -> 1.49 | Mixed |
| unitary KC step applied to dissipative dynamics | KC · Stiff reaction network | state fit accuracy | Hurt |
| KC on N-body systems vs diagonal baselines | KC · N-body (charged, gravity, spring) | linear momentum drift: 0.5 to 8 -> 2e-6 | Improved |
Surprising findings
- Unit-circle eigenvalue constraint is not enough for long-horizon stability, despite large short-horizon gains.
Limitations
- Full text unavailable; limitations not stated in the abstract.
Experimental setup
Metrics: dB gain at step 10
Baselines: unit-circle eigenvalue constraint
Builds on: diagonal state space models
Closest papers by meaning
- Finding Koopman Invariant Subspaces via Personalized PageRank (NeurIPS 2026 · similarity 0.84)
- Uncovering the Spectral Bias in Diagonal State Space Models (NeurIPS 2025 · similarity 0.84)
- Noether's Razor: Learning Conserved Quantities (NeurIPS 2024 · similarity 0.84)
- Autocorrelation Matters: Understanding the Role of Initialization Schemes for State Space Models (ICLR 2025 · similarity 0.84)
- SpectraLDS: Provable Distillation for Linear Dynamical Systems (NeurIPS 2025 · similarity 0.84)
- Structured Sparse Transition Matrices to Enable State Tracking in State-Space Models (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
