Composite Lyapunov Analysis

Updated 15 June 2026

Composite Lyapunov Analysis is a systematic framework that constructs composite functions using additive, piecewise, or aggregate structures to certify stability in complex dynamical systems.
It reduces conservatism inherent in classical methods by enabling scalable verification for high-dimensional, switched, and distributed systems.
The approach leverages modern tools such as sum-of-squares optimization, neural network approximations, and graph-based techniques to ensure tractable and robust performance certification.

Composite Lyapunov Analysis provides a systematic framework for stability assessment and performance certification of complex dynamical systems by constructing Lyapunov functions with additive, piecewise, or aggregate structure. This approach is essential in scalable analysis and synthesis for high-dimensional, interconnected, distributed, switched, and nonlinear systems where classical (single-form) Lyapunov techniques are often conservative or inapplicable. The composite methodology enables greater flexibility, reduced conservatism, and algorithmic tractability, particularly when paired with modern tools such as sum-of-squares optimization, neural networks, satisfiability modulo theories (SMT) solvers, and structural graph theory.

1. Composite Lyapunov Function Structures

Composite Lyapunov functions arise in several architectures, distinguished by how they aggregate local or partial Lyapunov candidates and how these pieces interact with the dynamical decomposition of the target system.

Additive and Weighted Sums: In singular perturbation and multi-timescale systems, the composite function typically takes the form $V_c(x, z) = \alpha V_s(x) + \beta V_f(z)$ , where $V_s$ and $V_f$ are Lyapunov functions certifying (respectively) the stability of the slow- and fast-scale subsystems (Tang et al., 2024). The weights are tuned to balance interconnection terms and residuals due to subsystem nonidealities.
Piecewise (Path-Complete/Multiple) Lyapunov Functions: For switched and hybrid systems, a family $\{V_i(x)\}$ is assigned to modes, nodes of a graph, or memory windows. Composite stability is ensured if there exists a systematic rule—e.g., a path-complete graph or symbolic language—that guarantees monotonic decrease along every admissible trajectory (Philippe et al., 2017, Angeli et al., 2016, Rossa et al., 2023, Jongeneel et al., 23 Mar 2025).
Learned and Parametric Composites: In recent data-driven settings, composite Lyapunov candidates merge quadratic forms with expressive neural network components, e.g., $V_\theta(x) = x^T P_\theta x + \|\varphi_\theta(x) - \varphi_\theta(0)\|^2$ , allowing adaptation to regions of the state space poorly captured by analytic forms (Bechihi et al., 14 Nov 2025).

The common principle is that the composite function encapsulates both the geometric and algebraic features of the underlying system, enabling conditions for stability to be reduced to verifiable inequalities, often on lower-dimensional subspaces or on convenient computational templates.

2. Composite Lyapunov Criteria in Singular Perturbation

Singularly perturbed systems, characterized by the presence of both slow and fast dynamics parameterized by a small $\varepsilon$ , challenge direct Lyapunov constructions due to timescale separation. The composite Lyapunov approach constructs $V_c(x, y) = \zeta V(x) + (1-\zeta) W(x, y)$ , combining Lyapunov functions for the reduced (slow) and boundary-layer (fast) dynamics (Tang et al., 2024, Tang et al., 2024, Poveda et al., 28 Dec 2025).

Deterministic ODEs: The main composite theorem asserts that, if $V(x)$ and $W(x, y)$ satisfy fixed-time (or asymptotic) decrease rates for their respective isolated dynamics and cross-system interconnection terms are suitably bounded (e.g., quadratic or power-type), then for sufficiently small $\varepsilon$ , the full system exhibits fixed-time (or global asymptotic) stability. Explicit settling-time bounds and robustness to input disturbances (ISS) can be established (Tang et al., 2024, Tang et al., 2024).
Stochastic and Hybrid Systems: In stochastic hybrid inclusions, composite Lyapunov–Foster functions are constructed to verify uniform global asymptotic stability in probability (UGASp) or uniform global recurrence (UGR) by balancing flow and jump interconnection terms, controlling cross-derivatives, and leveraging expectation drift inequalities (Poveda et al., 28 Dec 2025).
Weight Tuning: Weights (e.g., $V_s$ 0) are selected to cancel or dominate interconnection cross-terms, a direct analog to small-gain arguments in deterministic theory.

This paradigm allows scalable certification for fixed-time (FxT), input-to-state (ISS), and recurrence properties in multi-timescale and stochastic networked systems.

3. Distributed, Vector, and Partial Composite Lyapunov Methods

Distributed stability analysis for large-scale nonlinear networks leverages composite Lyapunov functions aggregated from subsystem or local (partial) candidates. Two major methodologies are prominent:

Multiple Comparison Systems: The network is decomposed into overlapping or interacting subsystems, each with its Lyapunov function $V_s$ 1. By constructing (possibly time-varying) comparison systems $V_s$ 2, where $V_s$ 3 is Metzler and Hurwitz, and systematically updating envelopes (level sets), one certifies exponential (or asymptotic) stability and obtains provable inner-approximations of regions of attraction (ROAs) (Kundu et al., 2016).
Partial Lyapunov + Sum-of-Squares (SOS): For high-dimensional polynomial networks, local SOS programs generate partial Lyapunov functions $V_s$ 4 on projected subspaces. Their convex aggregation (e.g., $V_s$ 5) yields a global composite function. The augmented comparison lemma ensures that the full system’s stability is certified if the collection of partial certificates and their interconnection residues satisfy joint SOS constraints (Wang et al., 25 Jun 2025).

Table: Distributed Composite Lyapunov Approaches for Nonlinear Networks

Method	Key Features	References
Multiple Comparison Sys.	Overlapping subsystem Lyapunov candidates, iterative comparison envelopes, scalable to large networks	(Kundu et al., 2016)
Partial SOS Composite	Parallel low-dimensional SOS programs, convex aggregation, improved ROA shape and volume	(Wang et al., 25 Jun 2025)

These formulations address scalability in both computation and memory, maintain distributed implementation, and better capture non-convex or “tentacle-like” ROA structures compared to centralized quadratic functions.

4. Composite Lyapunov Functions in Switched and Hybrid Systems

For systems with switching, hybrid, or symbolic dynamics, composite Lyapunov analysis is fundamentally tied to algebraic and language-theoretic structure:

Path-Complete Lyapunov Functions (PCLF): A finite set of candidate functions $V_s$ 6 is organized via a labeled graph $V_s$ 7. Path-complete structure ensures every possible finite switching word can be “monitored” by traversing a sequence of nodes and Lyapunov inequalities. The existence of a feasible PCLF for $V_s$ 8 yields stability, and, crucially, every PCLF induces a common Lyapunov function of min–max composition type (Philippe et al., 2017, Angeli et al., 2016, Jongeneel et al., 23 Mar 2025).
Memory and Prediction Windows: Recent symbolic dynamics frameworks generalize PCLF by encoding switching constraints as sofic shifts and constructing composite (memory-based) Lyapunov functions $V_s$ 9 accounting for finite sequences (memory windows) or even predictions (future windows), reducing conservatism and capturing richer stability properties (Rossa et al., 2023).
Combinatorics of Graph Lifts: Composition lifts and transitive closure refine and order PCLF templates, providing algorithmic handles for certificate comparison and tightening via iterative graph augmentation (Jongeneel et al., 23 Mar 2025).

The affine equivalence between multiple Lyapunov criteria and composite/memory-based certificates is established: any valid family of mode-wise functions induces an equivalent composite Lyapunov function on an extended (state × word) space.

5. Composite Lyapunov Analysis in Adaptive, Learning, and Data-Driven Control

Composite Lyapunov functions underpin modern learning-adaptive systems, where adaptation and performance guarantees must be made rigorous in the presence of rich, nonparametric function approximation or historical data collection.

Neural-Net-Based Composite Functions: By parametrizing $V_f$ 0 (with $V_f$ 1 a neural network), the composite candidate combines analytic positivity with data-driven flexibility. A homogeneous, cardinality-based loss trains the candidate to maximize the verified ROA; formal verification is performed via SMT solvers, and strict decrease near the origin is analytically guaranteed (Bechihi et al., 14 Nov 2025).
Composite Learning and Adaptive Control: For parameter estimation in uncertain systems, composite Lyapunov functions aggregate tracking error and parameter error components. Notably, composite learning schemes exploit excitation history without persistent excitation conditions by partitioning parameter subspaces (excited/unexcited) and showing exponential decay on the excited component using time-varying, adaptively updated Lyapunov–RLS structures (Shen et al., 2024, Zengin et al., 2020).
Adaptive DNN Controllers: Composite Lyapunov candidates simultaneously account for tracking, observer, and parameter errors. By including prediction-error-based dissipations, uniform ultimate boundedness of all closed-loop errors is ensured, with faster convergence rates especially under persistence-of-excitation through the DNN’s Jacobian (Patil et al., 2023).

This composite logic bridges classical stability theory and contemporary data-driven, learning-adaptive architectures, ensuring both performance and certification in increasingly complex environments.

6. Quantitative, Structural, and Algorithmic Properties

The composite Lyapunov methodology provides tighter convergence bounds, less conservative ROA approximations, and well-defined structural properties.

Strict Decay Criteria and Rates: Additive composite Lyapunov inequalities of the form

$V_f$ 2

yield explicit integral estimates, $V_f$ 3 convergence rates, and pointwise or exponential stability under provable small-gain conditions. This yields convergence to nontrivial critical sets or establishes semistability properties (Saoud, 9 Oct 2025).

Volume and Shape Approximations: By aggregating partial Lyapunov candidates via convex combinations, composite functions can more accurately approximate ROA volume and minimize shape dissimilarity relative to isotropic quadratic forms (Wang et al., 25 Jun 2025).
Scalability and Parallelization: Distributed and multi-comparison-composite frameworks enable decomposition of high-dimensional analysis into tractable subproblems, leveraging sum-of-squares and local computation (Kundu et al., 2016, Wang et al., 25 Jun 2025).
Formal Verification and Certification: SMT-based post-processing, analytic decrease near singularities, and LP-driven certificate comparison (for path-complete graphs) afford rigorous, non-probabilistic guarantees even for learned or high-degree composite certificates (Bechihi et al., 14 Nov 2025, Philippe et al., 2017, Jongeneel et al., 23 Mar 2025).

The unifying aspect is that composite constructions are not merely heuristic but are supported by theorems guaranteeing equivalence with single Lyapunov functions formed by algebraic min–max or convex combinations, ensuring that composite methods yield genuine certificates with provable quantitative properties.

7. Limitations and Outlook

Composite Lyapunov analysis, despite its breadth and power, encounters certain limitations:

Conservatism from Weight/Balancing Choices: Weighting and interconnection bounds (especially in singular perturbations or distributed certificates) can yield conservative parameter restrictions or small-gain conditions, potentially reducing the size of the certifiable ROA (Tang et al., 2024, Tang et al., 2024).
Computational Overhead: The dimension of the extended state space, size of covering graphs, or the collection of local certificates (e.g., in memory window or path-complete graph frameworks) may grow combinatorially, necessitating tradeoffs between certificate richness and tractability (Rossa et al., 2023, Jongeneel et al., 23 Mar 2025).
Template Closure and Generalization: The class of Lyapunov function templates (quadratic, polynomial, neural, etc.) must satisfy closure properties under composition or aggregation for certain theoretical results to hold; for general nonlinear or non-smooth systems, existence and verification of suitable partial certificates remain nontrivial.
Robustness to Numerical Artifacts: Composite certificates learned from data-driven paradigms may require post-hoc analytic guarantees or specialized verification (e.g., SMT solvers, analytic expansion near the origin) to avoid regions of numerical indeterminacy (Bechihi et al., 14 Nov 2025).

Ongoing work seeks to further reduce conservatism by refining interconnection bounds, enhancing graph-based certificate flexibility via new combinatorial lifts, and integrating structural priors from underlying network or mode-logic, with particular focus on certifying safety and stability in large-scale, nonlinear, and learning-augmented control systems.