betaFeedback

WorksEngineering

Data-Driven Sparse Identification and Robust Tube-Based Predictive Control for Multivariable Chaotic Systems

Abstract

This paper addresses the data-driven synchronization and robust control of multivariable chaotic systems subject to strict actuation constraints, modeling uncertainties and bounded input disturbances. Using empirical data collected from a four-dimensional chaotic Rossler response system, Sequentially Thresholded Least Squares (STLSQ) sparse regression is employed to recover a nominal control-affine model, achieving an identification accuracy of R^2 = 1.0000 across all channels with root-mean-square error within [4.15e-6, 8.33e-4]. To establish a mathematically rigorous bridge between data-driven identification and robust synthesis, all structural discrepancies in the drift vector field and input matrix are unified into a single lumped uncertainty field delta(x,u). A computable scalar bound delta_bar is established based on library truncation error, regression residuals, input bounds, and Lipschitz constants (Lemma 1). We subsequently prove that input-output feedback linearization maps delta(x,u) into a state- and control-dependent additive disturbance d(z,nu) in Brunovsky canonical coordinates, bounded by ||d(z,nu)||_2 <= alpha_1 delta_bar + alpha_2 delta_bar ||nu||_2 with closed-form expressions for alpha_1 and alpha_2 (Theorem 1). This certified bound, together with a bounded input disturbance and a zero-order-hold margin, yields a compact disturbance set that seeds a robust tube-based model predictive controller. We prove that the resulting tightened quadratic program is recursively feasible, that the physical actuator limit (|u_i| <= 1100) remains inactive under an explicit sufficient condition, and that the closed-loop tracking error is Input-to-State Stable (ISS) with an asymptotic gain that is linear in delta_bar plus a term due to the exogenous input disturbance (Theorems 2 and 3).

Introduction

Motivation

Chaotic oscillators are attractive building blocks for secure communication, encryption and sensing, and every such application rests on one capability: making a response system follow a drive system whose trajectory is aperiodic and exponentially sensitive to initial conditions [1, 2, 3]. In practice the problem is rarely posed in its textbook form. The actuator saturates, the input is corrupted by disturbances, the controller runs at a finite sampling rate, and the equations of the response system are known only through measurements. Each of these departures is manageable on its own. Together they make a controller that was designed on paper behave differently once it runs on hardware, and a guarantee derived for the nominal model says little about the system that was actually built.

We study this situation on a deliberately demanding testbed, a four-dimensional hyperchaotic Rössler drive–response pair, in which the response is driven through two input channels limited to ±1100\pm1100 and perturbed by a bounded disturbance. The question is not only whether synchronization can be achieved, but what can be certified about it when the model of the response has been learned from a single short record. Our answer combines three ingredients that are usually treated separately: sparse regression to recover the model, input–output feedback linearization to reduce the synchronization error to a linear system, and tube-based predictive control to enforce the constraints. The contribution of the paper is the bridge between them.

Related work

Synchronization of chaotic systems.

The synchronization literature is large and mature. Classical constructions rely on Lyapunov functions or on conditional Lyapunov exponents. Yao et al. propose a different mechanism, in which two systems synchronize when their energy variation rates agree [4], and Butusov et al. show that a time-reversible integration scheme generalizes the Pecora–Carroll approach and synchronizes systems from sparse or noisy data of a single state variable without any controller, although the outcome depends strongly on the initial conditions [5]. Controller-based schemes cover a wide range of settings: generalized synchronization of integer-order, fractional-order and delayed systems with feedback and finite-time controllers [6], adaptive laws derived from Lyapunov’s second method for a new three-dimensional chaotic system [3], adaptive sliding mode control with parameter identification for multi-wing attractors [7], and control and synchronization of a discrete fractional-order model [8]. Intelligent controllers have joined this toolbox, among them a preset-time fuzzy zeroing neural network validated on an FPGA [1] and a fuzzy Z-number controller for secure communication [2]. These works are mainly concerned with the convergence of the synchronization error under a known model structure, or one that is uncertain only through a few parameters. Hard actuation limits and a quantitative description of the disturbance are not the central object of their analysis, and that is precisely where the present paper starts.

Identification of chaotic dynamics from data.

When the model is not available, identification enters the picture. Peng and Li identify the parameters of the Lorenz system with observer/Kalman-filter and bilinear-transform schemes and use the estimate to initialize an adaptive synchronization law; they also observe that extremely small parameter deviations can produce dramatic mismatches in the predicted output of a chaotic model [9]. This observation is the core difficulty of any data-driven design. For a chaotic system, trajectory-level accuracy is a poor proxy for the quality of a model, and what a controller needs is a bound on the error of the vector field. Learning-based methods have widened the class of systems that can be modelled. Wu et al. estimate the parameters of discrete fractional chaotic systems with neural networks and report robustness to noise [10]. Reduced-order models built from autoencoders and neural ordinary differential equations capture the spatiotemporal chaos of the Kuramoto–Sivashinsky equation and have been used to train a reinforcement-learning controller [11, 12]. Spectral-submanifold models reconstruct regular and chaotic flapping of an inverted flag directly from video [13], and Kolmogorov–Arnold networks discover models that need not be sparse, with the caveat that many different approximate models generate the same invariant set and the same statistics [14]. A different line avoids models altogether: Wei et al. design a chaos controller for a vibro-impact system from input–output data through a compact-form dynamic linearization [15]. All of these succeed at what they measure, be it predictive accuracy, statistical fidelity or control performance in simulation, but none of them reports a bound on the model error that a robust controller could consume. The nonuniqueness found in [14] makes the point sharply: statistical agreement between models cannot by itself serve as a certificate for control.

Sparse identification for control.

Sparse regression offers a different trade. Brunton et al. extended the sparse identification of nonlinear dynamics (SINDy) to systems with inputs [16], which is the form of regression used in this paper. When the true dynamics are sparse in the chosen library, the identified model is a closed-form control-affine expression that can be passed directly to feedback linearization, and the identification error becomes a structured, bounded quantity rather than the opaque residual of a network. Two recent papers are closest to ours in architecture. Razani et al. combine SINDy, feedback linearization, a neural-network extended state observer and decentralized linear model predictive control to synchronize heterogeneous multi-agent systems, with the observer estimating the lumped uncertainty online [17]; in a second paper they combine SINDy with a sliding mode observer and a feedback-linearization plus LQR controller to detect bias attacks on a cyber-physical system [18]. In both, the uncertainty left by identification is handled by online estimation or by the robustness of the linear loop, and the bound that it induces on the linearized dynamics is not derived. Neither treats a chaotic drive–response pair, and, as far as we can tell from the published descriptions, neither carries an actuator limit through the analysis.

Tube-based predictive control.

On the control side, tube-based model predictive control is the natural tool once the disturbance set is known. In the linear setting a local feedback keeps the disturbed trajectory inside a tube around a nominal trajectory, the cross-section of the tube is a robust positively invariant set, and the nominal problem is solved with tightened constraints [19]; Mayne and Kerrigan extend the procedure to constrained nonlinear systems through an ancillary problem [20]. Invariant sets for such designs can be computed or approximated [21, 22], and the stability of the resulting loop is conveniently expressed as input-to-state stability [23]. What a tube design requires, and what the works above do not supply for a learned model, is the disturbance set itself. Feedback linearization is useful twice here. It replaces the nonlinear synchronization error by a linear one, so that the tube machinery applies in its classical form, and it provides a closed-form map that carries the identification error into that linear system as an additive disturbance. The present paper makes this map explicit.

Gap.

In summary, the individual pieces exist: sparse identification with inputs, feedback linearization, and tube-based predictive control. What is missing, to our knowledge, is the quantitative link between them for chaotic synchronization: a computable bound on the identification error, its propagation through the linearizing coordinates, and a controller whose constraint tightening is built on that bound and respects a hard input limit. This is the gap we address.

Contributions

This paper makes the following contributions.

First, we define one lumped uncertainty field δ(x,u)=[f(x)−f^(x)]+[g(x)−g^(x)]u\delta(x,u) = [f(x) - \hat{f}(x)] + [g(x) - \hat{g}(x)]u that carries every identification error, library truncation included, and prove a computable bound δˉ\bar{\delta} on X×U\mathcal{X} \times\mathcal{U} in terms of the truncation error, the regression residual, the input range and Lipschitz constants Lemma 1. We also show how δˉ\bar{\delta} is calibrated in practice, from validation residuals and, for a benchmark with known structure, from the coefficient errors directly.

Second, we prove that the image of δ\delta in Brunovsky coordinates is an additive disturbance d(z,ν)d(z,\nu) with ∥d∥2≤α1δˉ+α2∥ν∥2\lVert d\rVert_2 \le\alpha_1\bar{\delta} + \alpha_2\lVert\nu\rVert_2, and give α1,α2\alpha_1,\alpha_2 in closed form for general MIMO control-affine systems and explicitly for the Rössler model Theorem 1.

Third, we derive from this bound, combined with a bounded input disturbance and a zero-order-hold margin, a compact disturbance set that enters the tube design without further tuning (Corollary 1 and § 5).

Fourth, we show that the tightened quadratic program built on this set is recursively feasible, and that the saturation limit ∣ui∣≤1100\lvert u_i\rvert\le1100 stays inactive whenever the explicit condition (νmax⁡+βˉ)/σmin⁡≤umax⁡(\nu_{\max}+\bar{\beta})/\sigma_{\min} \le u_{\max} on the virtual-input bound and the cancellation mismatch holds, and is verified on the closed loop otherwise (Theorem 2).

Fifth, we prove that the closed loop is input-to-state stable, that the part of the ISS gain due to the identification error is linear in δˉ\bar{\delta} and vanishes as the identification improves, and that the remaining part is fixed by the bounded input disturbance (Theorem 3).

Sixth, we validate the pipeline on a hyperchaotic Rössler drive–response pair. From 1500 samples, STLSQ recovers the eleven nonzero coefficients of the model with a coefficient error below 10−410^{-4}, which gives δˉ=1.8×10−3\bar{\delta} = 1.8 \times10^{-3}. With the input limited to ±1100\pm1100 and a bounded input disturbance, the tube-based controller keeps all 20000 controlled steps inside the tube, with no failure of the quadratic program and a peak input of 70% of the limit. We also state where the guarantees stop: for this plant the worst-case saturation bound does not close, so the inactivity of the saturation is verified on the closed loop rather than proved (§ 6).

Notation and organization

The plant alone is written in the classical form x˙=f(x)+g(x)u\dot{x}=f(x)+g(x)u. Everything else follows one scheme. The state region is X⊂Rn\mathcal{X}\subset\mathbb{R}^{n}, the admissible input set is U={u:∣ui∣≤umax⁡}\mathcal{U}=\{u:|u_i|\leq u_{\max}\} with umax⁡=1100u_{\max}=1100, and Z:=T(X)\mathcal{Z}:=T(\mathcal{X}) is the image of X\mathcal{X} under the linearizing map TT. The true fields are f,gf,g; the identified ones carry a hat. The lumped uncertainty is δ\delta, with scalar bound δˉ\bar{\delta}, and W={w:∥w∥2≤wˉ}\mathcal{W}=\{w:\lVert w\rVert_2\leq\bar{w}\} is the disturbance set generated by δ\delta, with wˉ\bar{w} from Corollary 1. The exogenous input disturbance is dud_u with amplitude dampd_{\mathrm{amp}}, and wmax⁡∈Rmw_{\max}\in\mathbb{R}^{m} collects the per-channel disturbance bounds of § 2.5. The output map is hh, the relative degrees are r1,…,rmr_1,\ldots,r_m, the decoupling matrix is A(x^)A(\hat{x}) with pseudo-inverse A†(x^)A^{\dagger}(\hat{x}), and z=T(x^)z=T(\hat{x}) is the Brunovsky state with virtual control v=vd+νv=v_d+\nu, where ν\nu is the increment computed by the predictive controller and νmax⁡\nu_{\max} is its bound. The tube-MPC objects are the ancillary pole-placement gain KK with the LQR terminal feedback KfK_f, the nominal feedback κ\kappa, the deviation ee between actual and nominal trajectories, and the operators ⊕\oplus and ⊖\ominus for the Minkowski sum and the Pontryagin difference. The norms are ∥⋅∥2\lVert\cdot\rVert_2 (Euclidean and induced) and ∥⋅∥F\lVert\cdot\rVert_F.

Four symbols of the notation scheme would otherwise carry two meanings, and we separate them. The image set Z=T(X)\mathcal{Z}=T(\mathcal{X}), the invariant tube E\mathcal{E} (a robust positively invariant set), and the constraint sets in tracking coordinates (Zc\mathcal{Z}_c and its tightening Zˉ=Zc⊖E\bar{\mathcal{Z}}=\mathcal{Z}_c\ominus\mathcal{E}) are three different objects; likewise the physical input set U\mathcal{U} and the virtual input set V\mathcal{V} with tightening Vˉ=V⊖KE\bar{\mathcal{V}}=\mathcal{V}\ominus K\mathcal{E}. The simulation horizon is written tf=250 st_f=250\ \mathrm{s} so that TT is reserved for the coordinate map, and the control sampling period is τ=0.005 s\tau=0.005\ \mathrm{s}.

§ 2 and § 3 fix the synchronization problem, the assumptions and the identification layer, and prove the bound on δ\delta. § 4 carries δ\delta through the feedback linearization and proves the flagship result. § 5 builds the tube-based controller and proves feasibility and stability. § 6 reports the case study, and § 8 concludes and lists the open questions.

Problem Formulation and Preliminaries

Drive–response synchronization problem

The drive is the uncontrolled four-dimensional system

x˙1=−(x2+x3),x˙2=x1+0.25x2+x4,x˙3=3.0+x1x3,x˙4=0.05x4−0.5x3.(1)\begin{aligned} \dot{x}_1&=-(x_2+x_3), & \dot{x}_2&=x_1+0.25x_2+x_4,\\ \dot{x}_3&=3.0+x_1x_3, & \dot{x}_4&=0.05x_4-0.5x_3. \tag*{(1)} \end{aligned}

and the response is the same vector field with two actuated channels,

x^˙1=−(x^2+x^3),x^˙2=x^1+0.25x^2+x^4+u1,x^˙3=3.0+x^1x^3+u2,x^˙4=0.05x^4−0.5x^3.(2)\begin{aligned} \dot{\hat{x}}_1&=-(\hat{x}_2+\hat{x}_3), & \dot{\hat{x}}_2&=\hat{x}_1+0.25\hat{x}_2+\hat{x}_4+u_1,\\ \dot{\hat{x}}_3&=3.0+\hat{x}_1\hat{x}_3+u_2, & \dot{\hat{x}}_4&=0.05\hat{x}_4-0.5\hat{x}_3. \tag*{(2)} \end{aligned}

Written in control-affine form, (2) reads

x˙=f(x)+g(x)u,x∈X⊂Rn,u∈U⊂Rm,(3)\dot{x}=f(x)+g(x)u,\qquad x\in\mathcal{X}\subset\mathbb{R}^{n},\quad u\in\mathcal{U}\subset\mathbb{R}^{m}, \tag*{(3)}

with n=4n=4, m=2m=2 and g=[00100100]g=\begin{bmatrix}0&0\\1&0\\0&1\\0&0\end{bmatrix}. We display ff and gg because they define the benchmark; during identification they are unknown, and only the response is identified. The drive is a known exogenous signal generator. In the closed-loop study the plant is additionally affected by a bounded input disturbance (§ 2.5).

The synchronization outputs are

e1=x^1−x1,e2=x^4−x4,(4)e_1=\hat{x}_1-x_1,\qquad e_2=\hat{x}_4-x_4, \tag*{(4)}

References

  1. [1]Xiao, L., Zhao, L., Jin, J.: Preset-time convergence fuzzy zeroing neural network for chaotic system synchronization: Fpga validation and secure communication applications. Sensors 25(17), 5394 (2025)DOI
  2. [2]Babanli, K.M., Kabaoglu, R.O.: Synchronization of fuzzy-chaotic systems with z-controller in secure communication. Information sciences 657, 119988 (2024)DOI
  3. [3]Saeed, N.A., Saleh, H.A., El-Ganaini, W.A., Kamel, M., Mohamed, M.S.: On a new three-dimensional chaotic system with adaptive control and chaos synchronization. Shock and Vibration 2023(1), 1969500 (2023)DOI
  4. [4]Yao, Z., Sun, K., He, S.: Energy variation rate synchronization for coupled chaotic systems. Chaos, Solitons & Fractals 184, 114970 (2024)DOI
  5. [5]Butusov, D., Rybin, V., Karimov, A.: Fast time-reversible synchronization of chaotic systems. Physical Review E 111(1), 014213 (2025)DOI
  6. [6]Shukla, V.K., Joshi, M.C., Mishra, P.K., Xu, C.: Generalized synchronization between coupled integer, fractional and delay chaotic systems. Evolving Systems 16(4), 132 (2025)DOI
  7. [7]Wu, X., Yan, S.: Adaptive sliding mode synchronization of a controllable grid-multi-wing chaotic system with parameter identification. Applied Mathematical Modelling, 116351 (2025)DOI
  8. [8]Kahouli, O., Zouak, I., Hammad, M.A., Ouannas, A.: Chaos, control and synchronization in discrete time computer virus system with fractional orders. AIMS Math 10(6), 13594–13621 (2025)DOI
  9. [9]Peng, C.-C., Li, Y.-R.: Parameters identification of nonlinear lorenz chaotic system for high-precision model reference synchronization. Nonlinear Dynamics 108(2), 1733–1754 (2022)DOI
  10. [10]Wu, G.-C., Wu, Z.-Q., Zhu, W.: Data-driven discrete fractional chaotic systems, new numerical schemes and deep learning. Chaos: An Interdisciplinary Journal of Nonlinear Science 34(9) (2024)DOI
  11. [11]Linot, A.J., Graham, M.D.: Data-driven reduced-order modeling of spatiotemporal chaos with neural ordinary differential equations. Chaos: An Interdisciplinary Journal of Nonlinear Science 32(7) (2022)DOI
  12. [12]Zeng, K., Linot, A.J., Graham, M.D.: Data-driven control of spatiotemporal chaos with reduced-order neural ode-based models and reinforcement learning. Proceedings: Mathematical, Physical and Engineering Sciences 478(2267), 1–20 (2022)DOI
  13. [13]Xu, Z., Kaszás, B., Cenedese, M., Berti, G., Coletti, F., Haller, G.: Data-driven modelling of the regular and chaotic dynamics of an inverted flag from experiments. Journal of Fluid Mechanics 987, 7 (2024)DOI
  14. [14]Panahi, S., Moradi, M., Bollt, E.M., Lai, Y.-C.: Data-driven model discovery with kolmogorov-arnold networks. Physical Review Research 7(2), 023037 (2025)arxiv.org/abs/2409.15167
  15. [15]Wei, X.-j., Li, N.-z., Ding, W.-c.: Chaotic motion control of a vibro-impact system based on data-driven control method. Journal of Vibration and Control 29(1-2), 148–157 (2023)
  16. [16]Brunton, S.L., Proctor, J.L., Kutz, J.N.: Sparse identification of nonlinear dynamics with control (sindyc). IFAC-PapersOnLine 49(18), 710–715 (2016)arxiv.org/abs/1605.06682
  17. [17]Razani, S., Tavakoli-Kakhki, M., Kalhor, A.: Robust data-driven synchronization of heterogeneous nonlinear multi-agent systems. Nonlinear Dynamics 114(5), 372 (2026)DOI
  18. [18]Razani, S., Zarei, M., Dehghan, S.M.: Indirect data-driven resilient control against bias attacks and disturbance using sparse identification of nonlinear dynamics. IEEE Access (2025)DOI
  19. [19]Mayne, D.Q., Rawlings, J.B., Rao, C.V., Sсokaert, P.O.M.: Constrained model predictive control: Stability and optimality. Automatica 36(6), 789–814 (2000)DOI
  20. [20]Mayne, D.Q., Kerrigan, E.C.: Tube-based robust nonlinear model predictive control. IFAC Proceedings Volumes 40(12), 36–41 (2007) https://doi.org/10.3182/20070822-3-ZA-2920.00006 . 7th IFAC Symposium on Nonlinear Control Systems
  21. [21]Raković, S.V., Kerrigan, E.C., Kouramas, K.I., Mayne, D.Q.: Invariant approximations of the minimal robust positively invariant set. IEEE Transactions on Automatic Control 50(3), 406–410 (2005)DOI
  22. [22]Kolmanovsky, I., Gilbert, E.G.: Theory and computation of disturbance invariant sets for discrete-time linear systems. Mathematical Problems in Engineering 4(4), 317–367 (1998)
  23. [23]Limon, D., Alamo, T., Raimondo, D.M., Peña, D., Bravo, J.M., Ferramosca, A., Camacho, E.F.: Input-to-state stability: a unifying framework for robust model predictive control. In: Nonlinear Model Predictive Control: Towards New Challenging Applications. Lecture Notes in Control and Information Sciences, vol. 384, pp. 1–26. Springer, ??? (2009)DOI
  24. [24]Sontag, E.D.: Smooth stabilization implies coprime factorization. IEEE Transactions on Automatic Control 34(4), 435–443 (1989)DOI

Paper details

vlnc:W_su62y8bj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC0-1.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents