Polytopic LPV Systems: Analysis & Control

• Polytopic LPV systems are control frameworks where state-space matrices vary affinely with a scheduling parameter within a compact polytope, facilitating convex analysis.
• Lyapunov-based techniques (CQLF and Poly-QLF) and conic sector methods provide vertex-wise, less conservative stability and performance guarantees via LMIs.
• The approach enables robust controller and observer design, invariant set computation, and adaptive MPC across applications like automotive, process control, and distributed systems.

A polytopic linear parameter-varying (LPV) system is a class of control system whose state-space matrices depend affinely on a scheduling parameter that takes values in a known compact polytope. Central to the polytopic LPV framework is the representation of the system as a convex combination of a finite set of "vertex" systems, enabling rigorous analysis and synthesis through linear matrix inequalities (LMIs), robust invariant sets, and parameter-dependent Lyapunov functions. Polytopic LPV systems facilitate tractable robust control across diverse applications such as model predictive control (MPC), observer and controller design, and the abstraction of nonlinear or distributed systems.

1. Mathematical Structure and Parameterization

A continuous-time polytopic LPV system has the form

$$\dot{x}(t) = \sum_{i=1}^N s_i(t) A_i x(t) + \sum_{i=1}^N s_i(t) B_i u(t), \quad y(t) = \sum_{i=1}^N s_i(t) C_i x(t)$$

where the scheduling weights $s_i(t)$ satisfy $s_i \geq 0$, $\sum_{i=1}^N s_i = 1$ for all $t$, capturing the trajectory of the exogenous scheduling signal in the polytope defined by the system vertices $\{A_i,B_i,C_i\}$ (Walsh et al., 2019). Discrete-time and hybrid extensions are standard, substituting $$x_{k+1} = \sum_i s_i(k) A_i x_k + \sum_i s_i(k) B_i u_k$$.

A key conceptual property is that for any admissible scheduling trajectory, the system's matrices remain within the convex hull of the vertices. The affine dependence is exploited in both analysis and synthesis, enabling the derivation of robust stability and performance conditions via vertex-wise or parameter-dependent LMIs. Extensions to switched systems (with mode-dependent polytopic structure) and nonlinear or distributed parameter embeddings are constructed by similar convex hull techniques (Lacerda et al., 2020).

2. Lyapunov Analysis and Conic Sector Conditions

Lyapunov-based methods provide necessary and sufficient conditions for stability or robust performance in polytopic LPV systems. Two dominant approaches are:

Common Quadratic Lyapunov Functions (CQLF): Require a single $P \succ 0$ satisfying LMIs at all vertices, yielding strong (but sometimes conservative) certificates of uniform stability (Gießler et al., 28 May 2025). For exponential stability under arbitrarily fast parameter variations, such a $P$ is sought so that, for all vertices $i$,

$(A_i + B_i K)^\top P + P (A_i + B_i K) \prec 0.$

Poly-Quadratic Lyapunov Functions (Poly-QLF): Generalize to parameter-dependent $P(p) = \sum_{i=1}^N \xi_i(p) P_i$, with $P_i \succ 0$, where $p$ is the scheduling parameter and $\xi_i(p)$ its barycentric weights. Necessary and sufficient conditions for stabilizability or detectability require, for all vertex pairs $(i,j)$,

$P_i - A_i^\top P_j A_i + C^\top C \succ 0,$

and similarly, for state-feedback stabilizability,

$S_j - A_i S_i A_i^\top + B B^\top \succ 0,$

with $P_i = S_i^{-1}$ (Meijer et al., 1 Feb 2026).

Conic-Sector and Small-Gain Methods: The conic sector theorem unifies gain and passivity analysis for polytopic LPV systems, certifying sector bounds through vertex-wise LMI tests involving slack matrices and a common $P$ (Walsh et al., 2019). This approach allows relaxation or generalization beyond symmetric gain or passivity conditions.

3. Robust Invariant Sets and Constraints

Robust control invariant (RCI) sets are polytopic sets that, under a parameter-dependent control law, ensure that trajectories remain within prescribed constraints for all admissible scheduling and disturbance realizations. For a polytopic LPV system $x_{k+1} = A(\rho_k)x_k + B(\rho_k) u_k + w_k$, with $\rho_k \in$ convex polytope and $w_k$ in a disturbance set, the RCI set $S$ satisfies

$$A_i x + B_i u + w \in S, \quad \forall\, x \in S,\, u \in \mathcal{U},\, w \in \mathcal{W},\, \forall i$$

with vertex-based LMIs ensuring robust invariance (Mejari et al., 2023, Mejari et al., 2023). Modern data-driven methods generate these sets and the associated gain-scheduled feedback control law directly from a persistently exciting trajectory, using polytopic parameterization of the feasible model set, thereby bypassing explicit model identification (Mejari et al., 2023).

Tube-based MPC relies on offline computation of an RPI (robust positively invariant) tube around the nominal trajectory, tightening input and state constraints to account for model and disturbance uncertainty (Ismail et al., 2020). The tube cross-section is typically constructed via vertex-wise Minkowski iteration or support function optimization.

4. Synthesis of Controllers and Observers

Controller synthesis for polytopic LPV systems exploits the convex structure for tractable optimization:

• H-infinity and Conic-LMI Synthesis: The H-infinity approach constructs controllers guaranteeing closed-loop induced norm bounds by solving vertex-wise LMIs, optionally embedding passivity or more general conic-sector constraints for less conservatism (Zhang, 2017, Walsh et al., 2019).
• Parameter-Dependent Dynamic Controllers: Recent approaches formulate projected-gradient-flow laws for dynamic state-feedback controllers whose gains adapt online, constrained within a polytopic hyperrectangle, with convex optimization certifying quadratic stability—even under arbitrary parameter variation rates (Gießler et al., 28 May 2025).
• Observer Design: Poly-QLF-based observer synthesis ensures uniform detectability by vertex-wise sample-and-hold observer gains, again confirmed by coupled LMIs (Meijer et al., 1 Feb 2026).

Adaptive MPC schemes (e.g., indirect-adaptive MPC) employ recursive parameter estimation to track the true scheduling vector, update the parameter-dependent terminal costs and sets, and use polytopically defined invariance sets to enforce robust feasibility and constraint satisfaction (Cairano, 2015, Chen et al., 2022). Reinforcement learning has been combined with polytopic MPC through policy gradients, directly optimizing the controller and estimator parameters in the face of modeling mismatch and parametric uncertainty (Esfahani et al., 2022).

5. Applications and Extensions

Polytopic LPV system frameworks underpin robust control for complex, high-dimensional, or nonlinear systems by providing systematic abstractions:

• Distributed Parameter Systems: Nonlinear PDEs—such as flexible stacker cranes or fluid dynamics—are approximated as low-dimensional polytopic LPV surrogates, enabling efficient online MPC via robust invariant tube and convex QP techniques (Ismail et al., 2020, Heiland et al., 2024).
• Automotive and Process Control: Vapor compression cycles in automotive HVAC and other processes are represented by tensor-product convex polytopic LPV models, with controllers synthesized for quadratic stability and H-infinity performance (Zhang, 2017).
• Data-driven Synthesis: Direct data-driven SDPs and LPs synthesize polytopic RCI sets and controllers from open-loop measurements, assuming persistency of excitation and suitable polytope parameterization (Mejari et al., 2023, Mejari et al., 2023).

Deep polytopic autoencoders create low-dimensional LPV approximations of large-scale nonlinear systems, enabling series-expansion-based nonlinear feedback synthesis that outperforms conventional LQR in robustness and performance (Heiland et al., 2024).

6. Computational and Theoretical Considerations

The computational tractability of polytopic LPV analysis and synthesis derives from vertex-wise decomposition: analysis and synthesis conditions are enforced at all vertices, and the number of LMIs scales quadratically (poly-QLF) or linearly (CQLF) in vertex count. Most standard problems, including robust MPC, conic sector certification, and invariant set computation, reduce to convex (semi)definite programming solvable with modern SDP solvers for typical vertex counts ($N \leq 10$) (Walsh et al., 2019, Gießler et al., 28 May 2025, Meijer et al., 1 Feb 2026).

Although CQLF is sometimes conservative, poly-QLF offers less conservatism but greater computational complexity. Novel methods leveraging parameter-dependent or "switched" Lyapunov functions further reduce conservatism for switched or rapidly varying parameter systems (Lacerda et al., 2020). Relaxing the requirement of a common Lyapunov matrix across the polytope is a focus of current research, with parameter-dependent or piecewise quadratic certificates enabling improved performance and broader applicability (Walsh et al., 2019, Meijer et al., 1 Feb 2026).

7. Performance, Robustness, and Comparative Outcomes

Polytopic LPV analysis and synthesis typically yield less conservative and more robust closed-loop performance compared to fixed-gain or nominal controllers. For instance, conic controller synthesis reduces sector conservatism and achieves improved temperature regulation and reduced RMS tracking error compared to standard LPV H-infinity methods in heat-exchanger control (Walsh et al., 2019).

Data-driven polytopic controllers and invariant sets can approach the size and performance of model-based solutions, even with moderate data volumes, provided excitation conditions hold (Mejari et al., 2023, Mejari et al., 2023). Tube-based MPC, system-level synthesis, and adaptive or RL-tuned approaches all demonstrate high robust feasibility–domain coverage and reduced computational burden, with feasible run-times for MPC QPs well under 10 ms for moderate system dimensions (Ismail et al., 2020, Chen et al., 2022, Esfahani et al., 2022).

Overall, the polytopic LPV paradigm underpins a rigorous, scalable, and application-agnostic framework for robust analysis, synthesis, and implementation of control systems under structured, convex parameter uncertainty.