Overcoming Conservatism in Region of Attraction Estimation

53 concepts43 papers1 proposal

Imagine a drone caught in a sudden, violent gust of wind, or a power grid experiencing a sudden localized surge. Will the system safely recover and return to its steady state, or will it spin out of control?

In control theory, the set of all initial states from which a system is guaranteed to converge back to its stable equilibrium is known as the Region of Attraction (ROA) or Domain of Attraction (DoA). For safety-critical systems, accurately mapping this region is paramount. However, as Korda et al. point out, determining the exact ROA for nonlinear systems is computationally intractable due to its highly complex, non-convex geometry. Because overestimating this region could lead to catastrophic failure, engineers must rely on inner approximations—subsets of the state space that are mathematically proven to be safe.

The core challenge in modern robust control is that these certified inner approximations are often deeply conservative. They capture only a tiny fraction of the states from which the system can actually recover, severely restricting the system's true operational envelope. A rapidly growing body of research addresses this problem by combining classical control theory with machine learning and advanced optimization to shrink this conservatism gap.

The Root of Conservatism: Geometric Rigidity

To understand why traditional stability certificates are so pessimistic, we have to look at the mathematical tools historically used to find them. To prove stability, engineers look for a Lyapunov function—a mathematical "energy" function that strictly decreases along the system's trajectories. If you can find a closed boundary (a level set) where the energy always decreases, everything inside that boundary is guaranteed to be stable.

To make the optimization math tractable, traditional methods restrict these Lyapunov functions to simple mathematical forms, such as low-degree polynomials or quadratic forms. As Valmórbida et al. explain, this forces the estimated safe zones to conform to rigid geometric primitives like ellipsoids or spheres.

However, true stability boundaries for complex nonlinear systems are rarely symmetrical or smooth; they can be highly irregular, lopsided, or even unbounded, as noted by Li et al.. Trying to fit a perfect ellipsoid inside an irregularly shaped stability zone is like trying to fit a rigid beach ball inside a narrow, winding cave. As highlighted by Leisenberger et al. and Shen, this geometric mismatch leaves vast portions of the safe state space uncertified and unusable.

   True Irregular ROA Boundary
  ┌──────────────────────────┐
  │      .─▀▀▀▀▀─.           │
  │    .▀         ▀.         │   ◄── Conservative Ellipsoidal Inner Bound
  │   .   ┌─────┐   .        │       (Leaves large safe areas uncertified)
  │  .    │  █  │    .       │
  │  █    └─────┘    █       │   ◄── True, Complex Stability Boundary
  │  ▀.             .▀       │
  │    ▀.         .▀         │
  │      ▀─▄▄▄▄▄─▀           │
  └──────────────────────────┘

Mapping the True Boundary via Zubov's Equation

To overcome this geometric rigidity, researchers have revitalized a classic mathematical tool: Zubov's Partial Differential Equation (PDE). Zubov's theorem provides a way to characterize the exact boundary of the region of attraction, mapping the true, non-convex topology of the stable zone without requiring arbitrary geometric shapes.

The challenge with Zubov's equation is that solving it analytically is impossible for most real-world nonlinear systems. Modern approaches solve this by training Physics-Informed Neural Networks (PINNs) to learn the solution to Zubov's PDE directly as a loss function.

For instance, Wang demonstrates that using Zubov's PDE as a physics-informed loss allows an actor-critic framework to precisely learn the stability boundary, yielding certified regions 2 to 4 times larger than previous methods without needing manual shape-tuning regularizers. Similarly, Liu et al. combine PINNs with formal Satisfiability Modulo Theories (SMT) solvers to rigorously verify much larger domains of attraction. By leveraging Zubov's theorem to guide data sampling, Li et al. achieved verified ROAs up to 150,000 times larger than traditional baselines, demonstrating the massive potential of physics-informed neural learning.

Neural Lyapunov Functions and the Learner-Verifier Loop

If rigid polynomials cause conservatism, why not use neural networks as the Lyapunov functions themselves? Neural networks are universal function approximators capable of representing highly complex, non-convex energy landscapes. A Neural Lyapunov Function can warp and bend to match the true, irregular shape of a system's stable region.

However, substituting a neural network into control loops introduces a major hazard: neural networks are notoriously difficult to trust blindly. To maintain the mathematical certainty required for safety-critical systems, researchers employ a learner-verifier loop.

 ┌────────────────┐       Candidate Certificate       ┌──────────────────┐
 │    Learner     ├──────────────────────────────────►│     Verifier     │
 │ (Neural Net)   │◄──────────────────────────────────┤  (SMT/MIP/SOS)   │
 └────────────────┘          Counterexamples          └──────────────────┘

In this framework, a "learner" network proposes a controller and a candidate Neural Lyapunov Function. A formal "verifier" then searches the state space for counterexamples—states where the Lyapunov conditions fail (i.e., where the system's energy increases instead of decreases). If a counterexample is found, it is fed back to the learner to refine the function.

To make this verification rigorous, the authors leverage formal optimization tools. Dai et al. and Chen et al. utilize Mixed-Integer Linear Programming (MIP) to formally verify the piecewise-linear boundaries of ReLU-based neural networks, guaranteeing that if a valid stability certificate exists within the search space, the algorithm will find it in a finite number of steps. In high-stakes applications like power grids, Huang et al. combine these neural certificates with SMT solvers to unlock significantly larger and more accurate stability margins than traditional quadratic methods allow.

Tightening Bounds via Advanced Optimization Constraints

While neural networks offer extreme flexibility, other researchers focus on reducing conservatism within more traditional optimization frameworks by refining how nonlinearities are bounded.

Instead of treating a nonlinear system's complex interactions as completely unknown or worst-case disturbances—which leads to highly pessimistic bounds—researchers can bound them using specialized Quadratic Constraints (QCs). This approach embeds specific algebraic inequalities directly into the stability analysis, capturing the mathematical structure of the nonlinearity without requiring full-scale simulation.

Liao et al. demonstrated that applying specialized quadratic constraints can nearly double the certified radius of a stable region in complex systems like fluid dynamics and robotics. By bridging the gap between overly simplistic linear models and computationally exhausting simulations, as shown by Kalur et al., quadratic constraints allow engineers to safely shrink the inner approximation gap, certifying larger operational envelopes while maintaining absolute mathematical rigor.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_conservative_stability_estimation_404567af,
  title = {conservative stability estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/conservative-stability-estimation/rdr_404567af64824c3e80f821473d7960cc},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_conservative_stability_estimation_404567af_v1,
  title = {conservative stability estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/conservative-stability-estimation/rdr_404567af64824c3e80f821473d7960cc/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake