Lyapunov Norm Optimization

• Lyapunov norm optimization is a method that systematically optimizes over general classes of Lyapunov functions beyond the quadratic norm to enhance stability analysis.
• It employs convex programming techniques such as SDP, LP, and SOS relaxations to design optimal norms for improved contraction rates, decay, and robust control synthesis.
• This approach enables sharper analysis for nonlinear, constrained, and infinite-dimensional systems, providing practical insights for both control design and optimization algorithm tuning.

Lyapunov norm optimization refers to a class of methodologies in systems theory, control, and optimization that systematically search over families of norms or Lyapunov functions to maximize (or otherwise optimize) performance, stability rates, or robustness guarantees of dynamical or control systems. Instead of fixing a specific quadratic or ℓ₂-type Lyapunov function, Lyapunov norm optimization explicitly optimizes over a broader class of candidate norms—polyhedral, homogeneous, weighted ℓ₁/ℓ∞, or even parameterized non-quadratic functions—often via convex programming (SDP, LP, or SOS relaxations). This approach is now central to the verification, synthesis, and stability analysis of complex, constrained, or high-dimensional dynamics, as well as in optimization algorithm analysis and learning stabilizing controllers.

1. Fundamental Concepts and Classes of Lyapunov Norms

A Lyapunov function is a scalar functional $V(x)$ that certifies stability or contraction by satisfying appropriate decrease conditions along system trajectories. Classical approaches typically use quadratic Lyapunov functions, $V(x) = x^\top P x$ with $P \succ 0$, corresponding to the Euclidean norm. However, more general norms admit sharper or less conservative analysis, accommodate constraints, and enable explicit trade-offs between various performance metrics.

Key classes of Lyapunov norms used in Lyapunov norm optimization include:

• Polyhedral Norms: $V(x) = \max_{1 \leq i \leq m} h_i^\top x$ where $\{h_i\}$ define a polyhedral unit ball. These can arbitrarily approximate any norm and are particularly suitable for systems with non-Euclidean geometry or when modeling invariance with polyhedral sets (Kousoulidis et al., 2021, Proskurnikov et al., 2024).
• Weighted ℓ₁ and ℓ∞ Norms: $V(x) = \sum_i w_i |x_i|$ or $V(x) = \max_i w_i |x_i|$, for positive weights $w_i$. These allow alignment of the norm with the system's dominant directions or monotonicity structure (Coogan, 2017, Proskurnikov et al., 2024, 2207.14579).
• Homogeneous and Higher-degree Lyapunov Functions: Polynomial (homogeneous) functions, e.g., $V(x) = x^{\otimes d} P x^{\otimes d}$, which reduce conservatism in norm bounds or reachability for nonlinear or lifted systems (Abdelraouf et al., 2022).
• Nonquadratic and Regular Pairings: Use of general norms accompanied by compatible "pairings" (generalized semi-inner products) satisfying the curve-norm-derivative and Lumer's inequality, enabling general stability and contraction results outside the quadratic norm setting (Proskurnikov et al., 2024, 2207.14579).

2. Optimization Methodologies for Lyapunov Norm Design

The search for an "optimal" Lyapunov norm is typically cast as a convex or meta-convex optimization problem that maximizes performance criteria, subject to Lyapunov decrease conditions encoded as LMIs, LPs, or SOS constraints.

• Semidefinite Programming over Quadratic/Homogeneous Functions: For quadratic (or lifted homogeneous) Lyapunov functions, decay conditions are encoded as LMIs (e.g., for LTI/LTI-PDEs), with the goal to maximize decay rate, minimize induced norms, or minimize state-space volume subject to invariance (Dobson et al., 2023, Abdelraouf et al., 2022, Thomas et al., 3 Apr 2026).
• Linear Programming for Polyhedral Norms: Optimization over the facet orientation/weights $(H)$ of polyhedral Lyapunov functions is recast as LPs to quantify and minimize the contraction "gap" on each facet, enabling fixed-complexity iteration toward a contractive polyhedral gauge (Kousoulidis et al., 2021, Proskurnikov et al., 2024).
• Sum-of-Squares (SOS) Programming for Polynomial Systems: Weights in separable Lyapunov functions or state-dependent metrics are synthesized with SOS programming, enforcing contraction via polynomial/nonlinear inequalities (Coogan, 2017).
• Weighted ℓ₁/ℓ∞ Norm Optimization: For monotone and positive systems, optimal choice of weights in diagonal norms reduces to log-convex or geometric programming (Coogan, 2017, Proskurnikov et al., 2024).
• Non-quadratic S-Lemma and Duality: For absolute stability or contraction in non-Euclidean norms, primal-dual programs and the non-polynomial S-Lemma yield tractable certificates of (optimal) contraction rates in weighted ℓₚ norms (2207.14579).

Table: Optimization Problem Types by Lyapunov Norm Class

Lyapunov Norm Class	Problem Type	Typical Solver Approach
Quadratic/homogeneous	LMI/SDP	Semidefinite programming
Polyhedral	LP	Linear programming
Weighted ℓ₁/ℓ∞ (diagonal)	LP/geometric	Linear/geometric programming
Polynomial/separable	SOS	Sum-of-squares programming
Non-Euclidean S-Lemma	Convex duality	LP/SDP (for p=1,2,∞, diag W)

3. Contraction, Decay Rate, and Region-of-Attraction Optimization

Optimizing the Lyapunov norm directly impacts the contraction rate, region of attraction, and robustness properties that can be established for a system:

• Maximal Contraction/Decay Rate: The optimal norm minimizes the maximal logarithmic norm $V(x) = x^\top P x$0 of the system generator over a family, corresponding to the sharpest possible exponential decay/certified stability (Proskurnikov et al., 2024, 2207.14579). For PDEs or infinite-dimensional systems, operator inequalities in lifted spaces can be solved via SDP to bound or minimize induced norms (Thomas et al., 3 Apr 2026).
• Region-of-Attraction Maximization: For nonlinear and constrained systems, Lyapunov norm optimization can maximize the largest sublevel set of the Lyapunov function guaranteed to be invariant—often by offline SDPs over CPA norms on triangulations (Lavaei et al., 2021).
• Input-Constrained and Minimum-Norm Control Design: The norm structure enables minimum-norm QP-based feedback synthesis, which is optimal with respect to the chosen norm (e.g., $V(x) = x^\top P x$1 or polyhedral) and preserves invariance and decay properties (Lavaei et al., 2021, Westenbroek et al., 2020).

4. Algorithmic and Computational Frameworks

Implementation of Lyapunov norm optimization exploits problem structure:

• Iterative Tangent/Gap Reduction for Polyhedral Norms: A fixed-facet polyhedral Lyapunov search is conducted by iteratively reducing the "contraction gap" via LP-derived sensitivity updates, with each iteration solving small LPs in the facet parameters (Kousoulidis et al., 2021).
• Primal/Dual SDP and Bisection: For quadratic or operator-valued Lyapunov functions, a bisection on the decay rate or norm bound is performed where each guess is checked for feasibility via an SDP (Dobson et al., 2023, Thomas et al., 3 Apr 2026, Abdelraouf et al., 2022).
• SOS Relaxations for Polynomial Norms: Weight functions or state-dependent metrics are parameterized as polynomials, and contraction inequalities are imposed via SOS constraints, typically yielding small SDPs for moderate system sizes (Coogan, 2017).
• Convex Parameterization for Minimum-Norm Control Learning: Explicit convexity enforcement in controller parametrization and penalty-based relaxation ensures global optimality in learned min-norm controllers (Westenbroek et al., 2020).
• Alternating Minimization in Polyhedral Norm Design: When both the norm-defining matrix and the induced log-norm are variables, alternating minimization is used, often converging rapidly in practice (Proskurnikov et al., 2024).

5. Applications Across Domains

Lyapunov norm optimization has been deployed across a range of application domains:

• Nonlinear and Constrained Control Synthesis: Construction of continuous piecewise-affine (CPA) Lyapunov functions and controllers facilitates stabilization of nonlinear, input-saturated, and state-constrained systems with explicit invariant-region guarantees (Lavaei et al., 2021).
• PDE and Infinite-Dimensional Control: Lyapunov-operator optimization in the PIE framework enables computation of impulse-to-peak output norms and optimal state-feedback synthesis for general PDE systems (Thomas et al., 3 Apr 2026).
• Optimization Algorithm Analysis: Phase-space Lyapunov methods and high-resolution inequalities characterize norm-minimization rates of ISTA, FISTA, and Nesterov’s method under composite or non-Euclidean objectives, linking algorithm tuning to Lyapunov norm selection (Li et al., 2022, Dobson et al., 2023).
• Positive and Monotone System Analysis: Weighted $V(x) = x^\top P x$2/$V(x) = x^\top P x$3 Lyapunov functions, optimized via LP or geometric programming, yield tight contraction and separability guarantees in monotone systems and cooperative networked dynamics (Coogan, 2017, 2207.14579).
• Robustness and Reachability: Homogeneous and polyhedral Lyapunov norm optimization leads to tighter estimates on $V(x) = x^\top P x$4-induced and peak-to-peak norms, and less conservative reachability sets for LTI and hybrid systems (Abdelraouf et al., 2022).
• Learning Stablizing Controllers: Min-norm Lyapunov-based penalty metrics embedded in model-free reinforcement learning enable sample-efficient stabilization of systems with unknown dynamics (Westenbroek et al., 2020).

6. Generalizations, Limitations, and Theoretical Insights

• General Regular Pairings: Recent work unifies the theory of semi-inner products (pairings) for norms and develops computational tools for log-norm estimation and contraction analysis of non-quadratic Lyapunov functions, enhancing tractability for polyhedral and general norms (Proskurnikov et al., 2024).
• Non-Euclidean S-Lemma: The extension of the classical S-Lemma to non-Euclidean norms (p=1/inf/2, diagonal W) enables convex dual criteria for contraction, but the joint optimization over $V(x) = x^\top P x$5 and $V(x) = x^\top P x$6 becomes intractable for general norms (2207.14579).
• Complexity Considerations: For homogeneous Lyapunov functions, computational cost grows rapidly with state dimension and degree, restricting practical implementation to moderate system dimensions and low-degree lifts (Abdelraouf et al., 2022, Kousoulidis et al., 2021).
• Losslessness and Duality: For quadratic and $V(x) = x^\top P x$7 (Metzler) norms, the dual program achieves zero duality gap; in general cases, duality may be conservative, depending on the regularity of the pairing and system structure (2207.14579, Proskurnikov et al., 2024).
• Trade-offs in Norm Choice: Optimizing the Lyapunov norm reflects trade-offs between conservatism, computational feasibility, and the desired region-of-attraction, contraction rate, or robustness margin.

7. Representative Examples

Select archetypal instances from the literature:

• Polyhedral Lyapunov for Linear Systems: Iterative LP minimizing contraction gap over $V(x) = x^\top P x$8-facet polyhedra for (potentially time-varying) LTI dynamics, with sensitivity-based updates and applications to control synthesis (Kousoulidis et al., 2021).
• Optimal ℓ₁/ℓ∞-Weighted Norms via Geometric Programming: Minimization of the induced log-norm via log-convex optimization in the weights for Metzler or monotone systems (Coogan, 2017, Proskurnikov et al., 2024, 2207.14579).
• CPA Lyapunov Synthesis for Constrained Nonlinear Systems: Joint construction of Lipschitz CLFs and minimum-norm controllers via SDP on triangulated state spaces for control-affine systems with explicit decay-rate optimization (Lavaei et al., 2021).
• Homogeneous Lyapunov for $V(x) = x^\top P x$9-Gain Bounds: Kronecker lifting to degree-$P \succ 0$0 homogeneous Lyapunov functions sharpens bounds for reachability and peak-to-peak norms of LTI systems, implemented via high-dimensional SDP (Abdelraouf et al., 2022).
• Lyapunov Operator Design in PDEs: Induced-norm minimization and optimal feedback design for PDEs are realized by primal-dual SDP over Lyapunov PI operators (Thomas et al., 3 Apr 2026).
• Lyapunov-based Metric Design for Optimization Algorithms: Phase-space Lyapunov analysis, with tuning of the matrix weights, parameters, and meta-optimization via LMI/SDP, yields provably optimal rates for first-order algorithms (Li et al., 2022, Dobson et al., 2023).

---

Lyapunov norm optimization is foundational in modern control, stability analysis, robust optimization, and algorithmic design, allowing for systematic, computationally tractable, and often provably optimal synthesis of Lyapunov functions beyond the classical quadratic paradigm. Its continued development is driven by the requirements of nonlinear, constrained, infinite-dimensional, and data-driven systems.