On the Existence of Quadratic Control Lyapunov Functions for Koopman-Operator based Bilinear Systems

Sami Leon Noel Aziz Hanna, Nicolas Hoischen, Sandra Hirche, Armin LedererarXiv 2026 · arXiv

TL;DR

While quadratic Control Lyapunov Functions (CLFs) are the go-to tool for stabilizing high-dimensional bilinear systems derived from Koopman operator theory, this paper reveals a major catch. The authors prove that for single-input and many high-dimensional multi-input systems, a quadratic CLF exists if and only if the system can be stabilized by a single constant control input. Because complex nonlinear systems are rarely stabilizable by a fixed, time-invariant control, this finding exposes a fundamental scalability limitation in current data-driven control workflows.

What Problem Does This Paper Address?

Data-driven control often relies on Koopman operator theory to lift complex, unknown nonlinear systems into a higher-dimensional space where the dynamics look linear or bilinear. While this "lifting" simplifies the system structure, it drastically increases the state dimension ($N \gg n$).

To design stabilizing controllers for these high-dimensional bilinear models, researchers frequently turn to Control Lyapunov Functions (CLFs). Specifically, they choose quadratic CLFs because finding them reduces to a convex optimization problem (like Semidefinite Programming) that is easy to solve computationally.

However, prior to this paper, it was poorly understood exactly when a bilinear system actually admits a quadratic CLF. When the optimization failed, it was unclear whether the system itself was unstabilizable or if the quadratic function class was simply too restrictive. This paper addresses this gap by mapping out the exact existence conditions of quadratic CLFs in high-dimensional spaces.

Methods

The authors establish their results through a rigorous optimization-based framework:

  1. QCQP Formulation: They first show that verifying the existence of a quadratic CLF is equivalent to checking if the optimal value of a specific Quadratically Constrained Quadratic Program (QCQP) is strictly negative.
  2. Semidefinite Relaxation (SDP): Since QCQPs are generally NP-hard, they relax the problem into a convex Semidefinite Program (SDP) by dropping a rank-1 constraint.
  3. Proving SDP-Exactness: Using Pataki’s Lemma, they prove that for single-input systems ($m=1$), this SDP relaxation is mathematically exact (the dropped rank constraint holds automatically). For multi-input systems, they hypothesize and empirically demonstrate that when the state dimension is much larger than the control dimension ($N \gg m$), the relaxation remains exact.
  4. Farkas' Lemma & Constant Control: By applying the semidefinite version of Farkas' Lemma to the exact SDP, they prove that the existence of a quadratic CLF logically implies the existence of a constant stabilizing control law ($u_i = -c_i$).

The overall logical flow of their proof strategy is illustrated below:

Logical relationship between the SDP, QCQP, quadratic CLFs, and constant control stabilization.

Datasets

The paper does not use real-world datasets. Instead, to evaluate their hypothesis regarding multi-input systems, the authors conducted a comprehensive numerical simulation using synthetically generated bilinear systems. They randomly sampled system matrices ($A, B_i$) across varying state dimensions ($N \in \{10, \dots, 100\}$) and control dimensions ($m \in \{2, 5, 8\}$), running 20 trials for each configuration to evaluate the rank of the optimization solutions.

Key Insights and Ideas

  • The High-Dimensional Trap: Counterintuitively, higher dimensions make quadratic CLFs less viable. In low dimensions, the SDP relaxation is loose, meaning a quadratic CLF might still exist even if constant control fails. But as the dictionary size $N$ grows (which is standard practice in Koopman learning to minimize reconstruction error), the SDP relaxation becomes tight (exact).
  • The Geometry of Relaxation: As shown in Figure 2, the SDP expands the non-convex QCQP domain into a convex cone. When "SDP-exactness" occurs, the optimal solution lands precisely on the rank-1 boundary where the two problems align.
Conceptual visualization of the feasible domains for the SDP and the QCQP.
  • A Fatal Flaw for Koopman Control: Because Koopman models are approximations of underlying nonlinear systems, requiring them to be stabilizable by a constant input is an incredibly high bar that real-world systems (like robotics or fluid dynamics) almost never meet. Therefore, searching for a quadratic CLF in high dimensions is often a mathematical dead end.

Key Results

  • Theorem 1 (Main Result): Under the assumption of SDP-exactness, a valid quadratic CLF exists if and only if there exists a constant stabilizing controller.
  • Exactness in Single-Input Systems: The authors analytically prove that the condition is always exact for $m=1$.
  • Empirical Proof for Multi-Input Systems: The numerical simulations reveal that as the state dimension $N$ increases, the mean rank of the SDP solution ($Z^*$) converges rapidly to 1. For $N > 80$, the rank was exactly 1 across all test cases, confirming that SDP-exactness consistently holds in high-dimensional settings.
Averaged ranks of recovered SDP solutions showing convergence to rank 1 as state dimension N increases.

Limitations and Future Work

The primary limitation is that this paper acts as a diagnostic tool rather than a cure: it thoroughly explains why current quadratic CLF methods fail in high dimensions, but it does not provide an immediate alternative control synthesis method. Additionally, while SDP-exactness is proven analytically for single-input systems, its extension to multi-input systems relies on empirical validation rather than a universal closed-form proof.

The authors suggest that future research must move away from pure quadratic CLFs for high-dimensional Koopman models. Instead, researchers should either explore more complex, non-quadratic CLF architectures (such as neural CLFs) or explicitly embed structural constraints during the Koopman learning phase to guarantee that the learned bilinear model is compatible with constant-control stabilization.

Summary updated

Abstract

Koopman operator-based methods enable data-driven bilinear representations of unknown nonlinear control systems. Accurate representations often demand significantly higher dimensions than the original system, making control design challenging. Control Lyapunov Functions (CLFs) are widely used for controller synthesis, with quadratic CLF candidates being the most common due to their simplicity. Yet, we show that this class is highly restrictive, especially when the state dimension is large: under mild conditions, their existence implies stabilizability of the bilinear system by a constant input -- that is, the control remains fixed over time. We establish this result by formulating a quadratically constrained quadratic program (QCQP) that exactly characterizes valid CLFs. Since QCQPs are NP-hard, we propose a convex semidefinite relaxation that offers a sufficient validity condition. For single-input systems, we prove that a quadratic CLF requires constant control stabilizability, and empirically demonstrate that this extends to high-dimensional multi-input systems in many cases.
BibTeX
@misc{hanna2026existence,
  author = {Sami Leon Noel Aziz Hanna and Nicolas Hoischen and Sandra Hirche and Armin Lederer},
  title = {{On the Existence of Quadratic Control Lyapunov Functions for Koopman-Operator based Bilinear Systems}},
  year = {2026},
  eprint = {2604.09267},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.09267},
}
Figure