Finite-Time Safety Probability Analysis

• Finite-Time Safety Probability is a measure that evaluates the likelihood a stochastic system remains within a designated safe state over a fixed time interval under uncertainty.
• It employs mathematical tools such as barrier certificates, stochastic control barrier functions, and dynamic programming to derive explicit probabilistic safety bounds.
• These frameworks inform safety-critical controller synthesis, providing noise-aware and tighter guarantees in applications ranging from robotics to infrastructure systems.

A finite-time safety probability quantifies the likelihood that a stochastic system, subject to random disturbances and potentially controlled inputs, remains within a designated safe region of its state space over a fixed time horizon. This probabilistic guarantee is a central object in modern safety-verification and safe control design for stochastic and safety-critical systems, as it reflects the system's risk profile under uncertainty over operationally relevant time intervals. Finite-time safety analysis is formalized through various frameworks, including barrier certificates, stochastic control barrier functions, dynamic programming, and partial differential equation methods, each tailored to discrete- or continuous-time dynamics, with or without control synthesis objectives.

1. Mathematical Formulation of Finite-Time Safety Probability

Consider a stochastic system evolving in discrete or continuous time. For continuous-time systems, the state $x_t \in \mathbb{R}^n$ typically evolves according to an Itô SDE: $dx_t = f(x_t)\,dt + g(x_t)\,dW_t + h(x_t)\,u_t\,dt$ where $f$ is the drift, $g$ the diffusion matrix, $W_t$ is a Wiener process, $h$ the control input matrix, and $u_t$ the control input. For discrete-time systems: $x_{k+1} = f(x_k, w_k)$ with $w_k$ an i.i.d. disturbance.

Given a safe set $S \subset \mathbb{R}^n$, the finite-time safety probability is defined as

$dx_t = f(x_t)\,dt + g(x_t)\,dW_t + h(x_t)\,u_t\,dt$0

where, for discrete time, $dx_t = f(x_t)\,dt + g(x_t)\,dW_t + h(x_t)\,u_t\,dt$1 is replaced by $dx_t = f(x_t)\,dt + g(x_t)\,dW_t + h(x_t)\,u_t\,dt$2 and $dx_t = f(x_t)\,dt + g(x_t)\,dW_t + h(x_t)\,u_t\,dt$3 by $dx_t = f(x_t)\,dt + g(x_t)\,dW_t + h(x_t)\,u_t\,dt$4. Other formulations (e.g., recovery or reach-avoid probability) are analogous but adjust the safe/target sets and temporal conditions (Santoyo et al., 2019, Black et al., 2023, Xue, 2024, Mestres et al., 1 Oct 2025).

The event of interest is the invariance: that the state never leaves $dx_t = f(x_t)\,dt + g(x_t)\,dW_t + h(x_t)\,u_t\,dt$5 over the finite horizon. The probability is often equivalently characterized as $dx_t = f(x_t)\,dt + g(x_t)\,dW_t + h(x_t)\,u_t\,dt$6 with $dx_t = f(x_t)\,dt + g(x_t)\,dW_t + h(x_t)\,u_t\,dt$7 the first-exit time from the safe set.

2. Barrier Certificate Approaches for Probabilistic Guarantees

Barrier functions offer a systematic machinery for stochastic safety analysis. Let $dx_t = f(x_t)\,dt + g(x_t)\,dW_t + h(x_t)\,u_t\,dt$8 be a twice-differentiable function with level sets encoding the safe and unsafe regions (e.g., $dx_t = f(x_t)\,dt + g(x_t)\,dW_t + h(x_t)\,u_t\,dt$9). The main concept is to relate the evolution of $f$0 to the likelihood of safety-violation events.

For continuous-time systems, the infinitesimal generator $f$1 applied to $f$2 at point $f$3 is

$f$4

A stochastic barrier function $f$5 serves as a certificate for finite-time safety if, inside the safe set, it satisfies: $f$6 with $f$7, alongside boundary conditions:

• $f$8 (nonnegativity)
• $f$9 on initial states $g$0 (small sublevel)
• $g$1 on the unsafe boundary $g$2.

Analogously, in discrete time, for $g$3: $g$4 with similar boundary conditions (Xue, 2024, Mestres et al., 1 Oct 2025). The resulting certificates are often constructed as polynomials using sum-of-squares (SOS) optimization subject to polynomial set descriptions, yielding computationally tractable SDP formulations.

3. Quantitative Safety Bounds and Tightness

Given the above certificate conditions, explicit upper bounds on the finite-time safety violation probability are derived via martingale inequalities and dynamic programming arguments. For the basic continuous-time formulation with barrier function $g$5, three cases are prominent (Santoyo et al., 2019, Black et al., 2023):

• $g$6, $g$7:

$g$8

• $g$9, $W_t$0:

$W_t$1

• $W_t$2:

$W_t$3

Recent developments have introduced risk-aware control barrier functions (RA-CBF) exploiting stochastic level-crossing results to produce even tighter, noise-aware upper bounds, such as

$W_t$4

where $W_t$5 is a bound on the noise-gain, rendering the bound explicit in the noise strength and provably tighter than classical bounds in many regimes (Black et al., 2023).

For discrete time, the bound for a candidate barrier $W_t$6 and $W_t$7 steps, with $W_t$8 and $W_t$9, is: $h$0 (Xue, 2024).

Extensions to lower bounds for safety, and both upper and lower bounds for reach-avoid problems, are established through sign-reversing variations of the barrier conditions (Xue, 2024, Xue et al., 23 Sep 2025).

4. Computational and Optimization Techniques

The search for suitable barrier certificates is typically cast as a nonconvex optimization over polynomial functions constrained by sum-of-squares positivity on semialgebraic sets. For continuous-time SDEs, the problem takes the form:

• Find $h$1 polynomial, minimize (proxy for safety violation probability), subject to:
  • $h$2 on $h$3
  • $h$4 on initial set
  • $h$5 on unsafe set
  • $h$6 on $h$7 (Santoyo et al., 2019, Santoyo et al., 2019)
• For discrete time, analogous conditions hold with appropriate expectation operators (Xue, 2024, Mestres et al., 1 Oct 2025).

Key practical limitations include scalability to high-dimensional state spaces (the degree and the number of variables in the SDP), and the trade-off between tightness of bounds and computational tractability. Recent advancements have removed the boundedness requirement on the auxiliary barrier function, thus enabling the use of unbounded polynomial templates over unbounded safe sets and expanding applicability (Xue et al., 23 Sep 2025).

Alternative numerical methods employ partial differential equations (Fokker–Planck or backward Kolmogorov equations) or dynamic programming recursions for finite-horizon verification, recovery, and risk-sensitive (e.g., CVaR) safety analysis (Chern et al., 2021, Chapman et al., 2019).

5. Control Synthesis for Prescribed Safety Probabilities

Synthesizing controllers that ensure finite-time safety with prescribed risk levels includes:

• Joint optimization over the barrier function and the feedback law for systems affine in control. Typical parameterizations are polynomial or quadratic-in-monomials feedbacks, with regularization to avoid unbounded control (Santoyo et al., 2019, Santoyo et al., 2019).
• Convex quadratic-program (QP)-based formulations, particularly with RA-CBF constraints or probabilistic control barrier functions (pCBF), enable real-time implementation. The control law is selected to satisfy the barrier constraint at each time step, optionally minimizing a nominal cost subject to the risk-level requirement (Black et al., 2023, Mestres et al., 1 Oct 2025, Wang et al., 2021).

For discrete-time settings, pCBFs guarantee that by ensuring a per-step safety condition with failure rate $h$8, the long-horizon safety is at least $h$9 over $u_t$0 steps; the system designer can solve for $u_t$1 to meet a global reliability requirement (Mestres et al., 1 Oct 2025). Methods include distribution-aware, moment-based, and data-driven conditions (e.g., scenario optimization, conformal prediction, concentration inequalities), making the constraints computationally tractable for realistic scenarios.

6. Extensions, Applications, and Empirical Benchmarks

Finite-time safety probability analysis underpins the design and verification of safety-critical autonomous systems, robotics, and infrastructure control. Methods are directly applied in areas such as:

• Mobile robot safety and highway merging under noise (Black et al., 2023)
• Population growth, oscillators, and ecological models (Feng et al., 2020)
• Discrete-time reach-avoid problems for risk-sensitive design (via CVaR metrics) (Chapman et al., 2019)
• Learning algorithms with safety guarantees in unknown environments, e.g., parallel sequential probability ratio test methods for multi-armed bandits (Castellano et al., 2020).

Empirical benchmarks, including high-dimensional systems and real hardware (e.g., quadruped robots traversing stochastic terrain), validate the tightness of theoretical bounds. Case studies serve as testbeds for comparisons between classical martingale, barrier-based, and risk-aware methodologies, revealing that the latter consistently yield less conservative, noise-adaptive finite-time safety probabilities (Black et al., 2023, Mestres et al., 1 Oct 2025).

7. Theoretical and Practical Implications

Finite-time safety probability frameworks have clarified the limitations of myopic, stepwise safety analysis, showing the necessity of global (finite-horizon) approaches to prevent risk accumulation. The theoretical infrastructure provided by barrier certificates, RA-CBFs, and their discrete-time generalizations forms the backbone of certified stochastic safe control (Santoyo et al., 2019, Xue, 2024).

Removal of boundedness assumptions on barrier certificates (Xue et al., 23 Sep 2025) and development of refined dynamic-programming recursions for risk-sensitive reachability (Chapman et al., 2019) have enhanced both the expressive power and computational feasibility of the methodology.

A plausible implication is that as system complexity and uncertainty dimensions increase, scalable, noise-aware, and optimally-tuned finite-time safety certification—incorporating both tight upper and lower probabilistic bounds—will become integral to the deployment of safety-critical stochastic systems in practice.