Control Barrier Certificates Overview

• Control barrier certificates are a function-theoretic formalism that certifies safety in controlled dynamical systems by ensuring forward invariance of safe sets.
• They utilize optimization techniques such as SOS programming, counterexample-guided synthesis, and neural methods to construct tractable safety certificates.
• Applications span autonomous vehicles, robotics, and cyber-physical systems, with extensions to handle stochasticity, adversarial influences, and networked architectures.

A control barrier certificate (CBC) is a function-theoretic formalism used to provide constructive, often computationally tractable, certificates of safety for controlled dynamical systems. CBCs generalize the classical concept of (autonomous) barrier certificates to the controlled or controlled-stochastic setting, where safety is to be enforced in the presence of inputs, uncertainties, adversaries, or compositional network structures. The CBC formalism underpins a wide variety of recent advances in formal controller synthesis, safe learning, robust and adaptive control, and compositional verification for both continuous-time and discrete-time systems. This article surveys the mathematical structure of control barrier certificates, synthesis approaches, theoretical properties, extensions to nonconvex and high-dimensional settings, and selected application domains, drawing on formal definitions, optimization-based methods, and computational case studies across the modern barrier certificate literature.

1. Mathematical Formulation of Control Barrier Certificates

Let $\dot{x} = f(x) + g(x)u$ or, in discrete time, $x_{k+1} = f(x_k, u_k)$ define the controlled system on $x \in \mathbb{R}^n$, with admissible controls $u \in U \subseteq \mathbb{R}^m$. Given “safe” sets $S$ (e.g., $S = \{x \mid s(x) \ge 0\}$), CBCs are constructed to characterize forward invariance, that is, to certify that, for all $x(0) \in S$, there exists a control law ensuring $x(t) \in S$ for all $t$.

A canonical continuous-time CBC consists of a $C^1$ function $B : \mathbb{R}^n \rightarrow \mathbb{R}$ satisfying:

• Containment: $B(x) \ge 0$ on the initial set $I$; $B(x) < 0$ for all $x \notin S$.
• Tangent cone positivity: For $x \in \partial\mathcal{B} = \{B(x)=0\}$,

$$\sup_{u \in U} \nabla B(x) \cdot (f(x) + g(x)u) > 0.$$

This ensures viable flow directions point into the safe set.

For discrete-time systems, a CBC is a function $B : X \rightarrow \mathbb{R}_{\ge 0}$ such that (for initialization $X_0$, unsafe $X_u$):

• $B(x) \le \eta$ for $x \in X_0$, $B(x) \ge \beta$ for $x \in X_u$ (with $\beta > \eta \ge 0$).
• For all $x$, there exists $u \in U$ such that

$$\mathbb{E}[B(f(x,u, w)) \mid x, u] \leq \max\{ \kappa(B(x)), c \},$$

where $\kappa$ is a class-$\mathcal{K}_\infty$ function with $\kappa(s) < s$, and $c \ge 0$ is a slack term to accommodate stochasticity or adversarial actions (Anand et al., 2021, Jagtap et al., 2019, Nejati et al., 2020). For purely deterministic systems, the expectation is dropped.

CBCs also appear in the form of control barrier functions (CBFs) and stochastic CBCs, which implement forward-invariance via supermartingale-type conditions or difference inequalities in the controlled Markov process setting (Chalaki et al., 2022, Liu et al., 11 Aug 2025, Lavaei et al., 2022, Akbarzadeh et al., 2023).

2. Synthesis Algorithms and Optimization Formulations

CBC and CBF synthesis naturally reduce to solving functional inequalities over semi-algebraic sets of states and input constraints. Several algorithmic approaches are prevalent:

• Sum-of-squares (SOS) programming: When $f, g, S, I, U$ are polynomial/semi-algebraic, CBC conditions can be encoded as SOS constraints and relaxed to semidefinite programs (SDPs). This approach yields polynomial certificates and polynomial control laws (Wang et al., 2022, Jagtap et al., 2019, Chalaki et al., 2022, Anand et al., 2021).
• Counterexample-Guided Inductive Synthesis (CEGIS): When inputs are discrete or $f$ non-polynomial, a candidate parameterized $B(x)$ is iteratively refined by alternately solving for feasibility over samples and adversarially identifying violating states (Ravanbakhsh et al., 2015, Anand et al., 2021, Nejati et al., 2020).
• Gradient-based learning: For high-dimensional or black-box systems, neurosymbolic approaches—e.g., monotone neural CBCs—combine monotonicity constraints and gradient descent on composite loss functions encoding CBC inequalities, verifiable on only boundary points due to monotonicity structure (Nadali et al., 16 Aug 2025).
• Sample-efficient certification: For data-driven and learning-enabled systems, CBC properties can be checked using Lipschitz-based multi-slice validation with tight guarantees on sample complexity (Mulagaleti et al., 4 Sep 2025, Mazouz et al., 2024).

Controllers enforcing CBCs are constructed via quadratic programs (QPs) that minimally adjust a nominal input to guarantee satisfaction of the CBC inequality at each point, with the constraint being linear or convex in control for typical barrier forms (Chalaki et al., 2022, Wang et al., 2017, Wang et al., 2022).

3. Extensions: Stochasticity, Adversaries, and Uncertainty

CBCs have been systematically extended to address non-determinism, adversarial actions, uncertainty, and networked architectures:

• Stochastic control systems: CBCs for Markov (discrete-time) or Itô (continuous-time) systems require expected decrements of $B$ along dynamics, yielding supermartingale bounds and finite-horizon probability guarantees for safe set invariance (Jagtap et al., 2019, Akbarzadeh et al., 2023, Lavaei et al., 2022, Anand et al., 2021).
• Adversarial and game-theoretic scenarios: Secure-CBCs (S-CBCs) encode safety as a two-player (defender–adversary) zero-sum game, using

$$\inf_{u_d \in U_d} \sup_{u_a \in U_a} \mathbb{E}_w [B(f(x,u_d,u_a,w))|x] \leq B(x) + c,$$

where $u_d$ and $u_a$ are defender/adversary inputs (Ramasubramanian et al., 2019).

• Robust and adaptive CBCs: Online parameter uncertainty is handled via robust-adaptive CBCs that maintain forward-invariance under unknown but bounded parameters $\theta^*$ and process noise, with parameter estimation and safety filtering modularized (Liu et al., 11 Aug 2025).
• Chance-constrained/barrier-certificates under measurement noise: Probabilistic CBCs incorporate chance constraints by inflating the barrier inequality with an explicit uncertainty margin, e.g., by adding a Gaussian confidence radius (Zhang et al., 2023, Mazouz et al., 2024, Wang et al., 2017).
• Compositionality and networks: Subsystem-level CBCs (control sub-barrier certificates, CSBCs) are composed via small-gain or dissipativity-type matrix inequalities to yield a global safety certificate for large-scale or switched interconnected systems (Anand et al., 2021, Nejati et al., 2020, Anand et al., 2021).

4. Theoretical Guarantees and Properties

The principal theoretical results associated with CBCs concern forward-invariance, probabilistic safety, synthesis soundness, and recursive feasibility:

• Controlled forward-invariance: If there exists $u(x)$ ensuring the CBC boundary (tangent cone) condition is satisfied everywhere on the barrier, then the corresponding safe set $\mathcal{B}$ is forward-invariant under all admissible trajectories (Wang et al., 2022, Wang et al., 2017).
• Finite-horizon probability bounds: For stochastic systems, CBCs yield explicit upper bounds on the probability of violating the safe set within a finite time, e.g.,

$$\mathbb{P}\{\exists\,k \leq T: x(k) \in X_u\} \le (\eta + c T)/\beta$$

with $B(x) \leq \eta$ on $X_0$, $B(x) \geq \beta$ on $X_u$ (Akbarzadeh et al., 2023, Anand et al., 2021).

• Recursive feasibility for MPC: Terminal CBC constraints in nonlinear MPC guarantee the continued feasibility of the OCP at each time-step, and, hence, perpetual safety (Katriniok et al., 2023).
• Compositional safety: Under suitable gain or dissipativity conditions, the max or sum of subsystem-level CBCs/CSBCs provides a global certificate for the entire network (Anand et al., 2021, Anand et al., 2021, Nejati et al., 2020).

CBCs are empirically less conservative than CBFs imposing global class-$\mathcal{K}$ inequalities, since the former only require the vector field (or controlled vector field) to point strictly into the safe set at the boundary, as opposed to enforcing a strict decay everywhere (Wang et al., 2022).

5. Computational Approaches and Scalability

Efficient computational methods underpin practical CBC synthesis:

• Sum-of-squares (SOS)/SDP: Standard for polynomial systems and sets, enabling convex relaxations of CBC inequalities via the S-procedure and Gram matrix techniques (Wang et al., 2022, Jagtap et al., 2019, Chalaki et al., 2022, Anand et al., 2021, Nejati et al., 2020).
• SMT and CEGIS: Employed particularly in non-polynomial or high-dimensional scenarios with finite inputs or switched systems for counterexample-driven barrier search (Ravanbakhsh et al., 2015, Anand et al., 2021).
• ADMM-based distributed optimization: For compositional CBC/CSBC synthesis in large networks, primal-dual approaches (e.g., alternating direction method of multipliers) decouple local SOS feasibility and global dissipativity-type LMI constraints (Anand et al., 2021).
• Neural and monotone CBCs: Scalability to thousands or tens of thousands of dimensions via monotone neural architectures, exploiting order-preserving structure and only requiring verification on boundary points to guarantee global invariance (Nadali et al., 16 Aug 2025).

Practical simulation studies confirm that barrier certificate synthesis via SOS/SDP or neural approaches scales to very large systems (e.g., power grids with $13,\!000$ dimensions or $1,000$-room HVAC networks) and robustly avoids unsafe states with certified probability margins.

6. Applications and Case Studies

CBCs have been deployed across domains, including:

• Motion planning and autonomous vehicles: CBC layers enforce real-time safety (state, input, inter-vehicle gap) as convex QPs for connected and automated vehicles in intersection management (Chalaki et al., 2022, Katriniok et al., 2023).
• Safe learning and robotics: Barrier certificates integrated with GP model learning yield exploration and safety expansion in adaptive flight controllers and learning-based robotic policies (Wang et al., 2017, Mazouz et al., 2024).
• Cyber-physical systems security: CBCs are combined with fault-tolerant estimation to guarantee safety even under sensor spoofing or LiDAR attacks (Zhang et al., 2022).
• Networked and hybrid systems: Barrier certificates provide global safety in networks via compositional max-type or dissipativity conditions, applied to large-scale room-temperature regulation, switched-mode cascades, and power flow (Anand et al., 2021, Nejati et al., 2020, Nadali et al., 16 Aug 2025).
• Stochastic and adversarial systems: Secure CBCs and stochastic barrier certificates formalize safety and temporal logic enforcement under uncertainty, adversarial environments, and communication loss (Ramasubramanian et al., 2019, Akbarzadeh et al., 2023, Lavaei et al., 2022).

7. Open Directions, Limitations, and Extensions

Current CBC frameworks have some restrictions and active research directions:

• Scalability and conservatism: While SOS and neural-symbolic approaches have enabled high-dimensional synthesis, they rely on system structure (polynomiality, monotonicity, compositionality). Non-monotone or irregular unsafe sets present open challenges (Nadali et al., 16 Aug 2025).
• Extensions beyond polynomial/nonlinear dynamics: SOS-based methods are restricted to polynomial templates. Data-driven and scenario-based methods are being developed to generalize CBCs to black-box models, arbitrary uncertainty, and online adaptation (Mazouz et al., 2024, Mulagaleti et al., 4 Sep 2025).
• Robustness and adaptivity: New schemes maintain safety under bounded parameter identification error or in the presence of persistent disturbances, with separation of online estimation and barrier-based safety filtering (Liu et al., 11 Aug 2025).
• Temporal logic and hybrid systems: Automata-theoretic decompositions relate CBC synthesis for reachability to LTL/LTL$_F$ specifications, enabling pathwise probabilistic bounds for complex logic-driven tasks (Jagtap et al., 2019, Ramasubramanian et al., 2019, Anand et al., 2021, Anand et al., 2021).
• Chance constraints and stochasticity: Explicit formulations for probabilistic safety via chance constraints and GP-uncertainty margins bridge model-based and learning-based approaches, critical in robotic and CPS applications with imperfect state measurement (Zhang et al., 2023, Wang et al., 2017, Mazouz et al., 2024).

Control barrier certificates represent a synthesisable, computationally tractable, and highly extensible methodology for providing rigorous safety guarantees across a broad spectrum of controlled, stochastic, and adversarial dynamical systems.