Research map / Probabilistic and scientific modeling / Dynamical systems
Provable State Estimation with Recurrent Models
Vincent Andrieu, Pauline Bernard, Lucas Brivadis, Laurent Praly, Christian Wolf
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.
Analyzes recurrent neural networks as time discretizations of Neural ODEs through the lens of nonlinear observer theory to obtain formal convergence and robustness guarantees for state estimation.
Problem
Recurrent models used for estimating the state of dynamical systems from observations lack formal guarantees on convergence and robustness.
Approach
Frames RNNs as time-discretized continuous latent dynamical systems (Neural ODEs) and applies nonlinear observer theory to derive provable convergence and robustness properties for state estimation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| enforce contraction in the GRU (Contracting GRU with spectral normalization) instead of a vanilla GRU | 2-layer GRU pose-correction model · Real mobile-robot pose correction (accuracy vs relative error) | pose accuracy | Improved |
| enforce contraction, measured as variability over 5 training runs | GRU · Real mobile-robot pose correction | training stability across seeds | Improved |
| train an unconstrained GRU with no contraction constraint | GRU · Real mobile-robot pose correction | state-distance contraction | Improved |
| increase latent gain kappa beyond about 2 (ablation) | RNN-MLP observer · Simulated nonlinear oscillator, every decoder width | steady-state RMS error | Hurt |
| increase latent gain kappa from 0.5 up to about 2 (ablation) | RNN-MLP observer · Simulated nonlinear oscillator | steady-state RMS error | Improved |
| widen the MLP decoder (ablation) | RNN-MLP observer · Simulated nonlinear oscillator | absolute error | Improved |
| raise gain kappa to 20 with widest decoder (4096) (ablation) | RNN-MLP observer · Simulated nonlinear oscillator, reconstruction error over time | transient speed and steady-state error | Mixed |
Experimental setup
Builds on: Neural ODE
Closest papers by meaning
- Neural Contractive Dynamical Systems (ICLR 2024 · similarity 0.88)
- Identifying latent state transitions in non-linear dynamical systems (ICLR 2025 · similarity 0.87)
- Learning Neural Contracting Dynamics: Extended Linearization and Global Guarantees (NeurIPS 2024 · similarity 0.86)
- Markovian Dynamics Enforcer: Feasibility Preserving Correction on Learned Dyanmics Manifolds (NeurIPS 2026 · similarity 0.86)
- Almost-Linear RNNs Yield Highly Interpretable Symbolic Codes in Dynamical Systems Reconstruction (NeurIPS 2024 · similarity 0.86)
- Learning long range dependencies through time reversal symmetry breaking (NeurIPS 2025 · similarity 0.86)
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.
