PhyxMamba: Chaotic System Reconstruction from Short Context Observations with Generative State-Space Models
Chang Liu, Bohao Zhao, Jingtao Ding, Huandong Wang, Yong Li · 2025
Read PhyxMamba: Chaotic System Reconstruction from Short Context Observations with Generative State-Space Models by Chang Liu, Bohao Zhao, Jingtao Ding, Huandong Wang, Yong Li (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-23863.txt.
Introduction
Chang Liu and four co-authors present PhyxMamba, a new method to rebuild the long-run path of a chaotic system from short clips of data. Their paper appeared in twenty twenty-five. Chaotic systems, like some weather and brain signals, are hard to test by direct trial. Past tools need long clean records, while short records are what labs most often have.
The method pairs a fast sequence model with rules drawn from physics. It uses time-delay views to rebuild the shape the system moves on. Tests on fake and real chaotic data beat the best past method by a wide margin on the Lorenz96 test case. The code is shared in full for others to try.
Catalog record
| CREATOR | Chang Liu, Bohao Zhao, Jingtao Ding, Huandong Wang, Yong Li |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 108 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-23863
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
PhyxMamba: Chaotic System Reconstruction from Short Context Observations with Generative State-Space Models Chang Liu, Bohao Zhao, Jingtao Ding, Huandong Wang, Yong Li Abstract Understanding chaotic dynamics is a fundamental problem across scientific disciplines, including climate science, neuroscience, and fluid dynamics, yet direct experimentation and intervention in such systems are often infeasible. Chaotic system reconstruction aims to identify a surrogate dynamical model that preserves a system’s invariant geometric and long-term temporal signatures from observed time series, thereby providing a controllable foundation for probing its mechanisms through systematic perturbation and analysis. However, faithful system reconstruction is hampered by high observational costs, which often restrict data to short, discontinuous sequences spanning only limited timescales. Conventional approaches such as reservoir computing struggle in this data-scarce regime since they typically require long-term synchronization windows to localize states on the attractor. Similarly, while deep learning-based time-series forecasting models effectively fit local trajectories, they often fail to capture global invariants, leading to a collapse of long-term dynamical integrity. Here, we propose PhyxMamba, a framework that synergizes Mamba-based state-space models with physics-informed principles. By leveraging time-delay embeddings to reconstruct the attractor manifold and employing a generative training scheme with geometry-aware regularization, PhyxMamba effectively captures both fine-grained local evolution and global physical constraints. Extensive experiments on simulated and real-world chaotic systems demonstrate that PhyxMamba achieves superior reconstruction performance, outperforming the strongest baseline by over 44% in prediction accuracy and 8% in topological fidelity on the Lorenz96 system, while exhibiting strong robustness against partial observations and noise. Codes are available at https://github.com/changliu01/PhyxMamba . I Introduction Chaotic systems, defined by their deterministic nature and extreme sensitivity to initial conditions, are pervasive across a wide range of scientific and engineering disciplines, including climate science [ 1 ] , epidemiology [ 2 ] , neuroscience [ 3 ] , and financial modeling [ 4 ] . However, rigorous inquiry into these complex dynamics is often impeded by the infeasibility of direct experimentation; for instance, conducting controlled trials on climate systems or monitoring sensitive biological networks without disrupting them is infeasible. Consequently, learning a faithful surrogate model from observational data, known as chaotic system reconstruction, is essential. Unlike forecasting that prioritizes point-wise accuracy, reconstruction seeks to preserve global topological invariants (e.g., fractal dimensions, state-space distribution) to enable valid long-term simulations. However, in many applications, models must infer the current dynamical state and initiate autonomous rollouts from short context observations available at inference time or within each training segment. Such contexts may span only limited characteristic timescales (e.g., Lyapunov times), making it difficult to recover the global attractor geometry required for faithful reconstruction. Existing paradigms for chaotic system reconstruction range from classical analytical models defined by differential equations [ 5 , 6 , 7 ] to contemporary data-driven approaches [ 8 , 9 , 10 ] . Analytical models are rendered impractical in real-world reconstruction scenarios where the governing equations are unknown or the system parameters are unobservable. Conversely, data-driven methods seek to infer dynamics directly from observations. Recurrent sequence models, such as Reservoir Computing (RC) [ 8 , 11 , 12 ] and Piecewise-Linear Recurrent Neural Network (PLRNN) [ 9 , 10 ] , rely heavily on the synchronization of internal reservoir states or specialized latent structures that necessitate extended continuous observation windows. Deep time-series predictors based on MLPs and Transformers [ 13 , 14 , 15 , 16 , 17 , 18 ] primarily optimize point-wise trajectory accuracy, but they do not explicitly preserve a chaotic system’s invariant measure. More recently, state-space models represented by Mamba [ 19 ] have emerged as promising candidates because their continuous-time latent states provide a favorable inductive bias for physical evolution. Yet using state-space models as generic forecasters is still insufficient for chaotic reconstruction: the model should be generative after a short context initialization and must preserve invariant statistics under long autonomous rollouts. How to build such generative state-space models for chaotic system reconstruction from short context observations remains largely underexplored. The key challenge lies in the intrinsic tension between limited observational context and chaotic complexity, which gives rise to three critical issues: (i) Extracting complex dynamical information from short context windows. It requires the model to infer the governing laws of the system by deciphering the latent topological information embedded within brief evolutionary segments. (ii) Mitigating error amplification during autoregressive evolution. Driven by positive Lyapunov exponents, microscopic approximation errors accumulate exponentially over time. An effective generative modeling process is essential to suppress this intrinsic divergence and support stable autonomous rollout. (iii) Enforcing ergodic invariance and geometric fidelity. It lies in constraining the model evolution within the manifold of the strange attractor, thereby precluding dynamical collapse into trivial states. To address these challenges, we propose PhyxMamba, a framework that unifies a Mamba-based architecture with geometric regularization principles to reconstruct the underlying dynamics of chaotic systems. First, we introduce a physics-informed patch representation via time-delay embedding. It synthesizes coupled multivariate dynamics to reconstruct local phase-space states from short contexts, thereby equipping the model with essential geometric priors. Second, we design a multi-step aware generative training strategy based on the Mamba backbone [ 19 , 20 ] . By incorporating multi-horizon perception into the generative process, it mitigates error accumulation and improves autoregressive stability. Third, we implement an invariant measure alignment stage. Utilizing a student-forcing strategy regularized by Maximum Mean Discrepancy (MMD) [ 21 ] , we enforce statistical consistency between generated trajectories and the true system dynamics, preventing geometrical collapse. Our contributions are summarized as follows: • A generative state-space framework for chaotic system reconstruction. We propose PhyxMamba, a framework that integrates the modeling capabilities of the Mamba backbone with physics-informed representations and regularizations. It enables robust reconstruction of chaotic systems from short context observations. • Tailored architectural designs for reconstruction fidelity. The model components unify representation, evolution, and regularization. It infers local phase-space structure from short context windows, stabilizes long-horizon generative rollouts via internalized multi-step evolution guidance, and simultaneously enforces explicit constraints on the system’s invariant measure. These designs ensure that the learned model evolves strictly on the correct attractor manifold. • Comprehensive empirical validation of performance and robustness. Extensive evaluations across diverse simulated and real-world datasets demonstrate that PhyxMamba consistently achieves state-of-the-art reconstruction quality. Crucially, our framework remains robust to non-ideal conditions, maintaining effectiveness even under partial observations, varying sampling resolutions, and observational noise. II Preliminaries and Problem Formulation In this section, we start with the theoretical preliminaries and notations, then formally define the chaotic system reconstruction problem. II-A Chaotic System A chaotic system is characterized by deterministic evolution exhibiting extreme sensitivity to initial conditions, where infinitesimal perturbations lead to exponentially diverging trajectories in phase space. Formally, consider a continuous-time dynamical system governed by the autonomous ordinary differential equation (ODE): \frac{d\bm{x}(t)}{dt}=\bm{f}(\bm{x}(t)), (1) where \bm{x}(t)\in\mathbb{R}^{V} represents the V -dimensional system state at time t , and \bm{f}:\mathbb{R}^{V}\rightarrow\mathbb{R}^{V} denotes a smooth nonlinear vector field driving the dynamics. A defining hallmark of such systems is the existence of a strange attractor, a compact invariant set \mathcal{A}\in\mathbb{R}^{V} to which trajectories asymptotically converge. Unlike simple fixed points or limit cycles, the strange attractor possesses a complex fractal geometry [ 22 ] and is supported by an invariant measure [ 23 , 24 ] , describing the system’s ergodic statistical properties. II-B