Probabilistic Signal Temporal Logic (PSTL)

• PSTL is a formal language that extends Signal Temporal Logic by incorporating probabilistic operators to handle stochastic uncertainty in system trajectories.
• It enables both verification and synthesis tasks by quantifying the probability that system behaviors satisfy specified temporal constraints.
• PSTL is applied in cyber-physical systems, robotics, and adaptive control, using methods like convex optimization and sampling-based techniques for scalable analysis.

Probabilistic Signal Temporal Logic (PSTL) is a family of formal specification languages extending classical Signal Temporal Logic (STL) to rigorously reason about temporal properties of systems under stochastic uncertainty. PSTL applies when system dynamics, environment, or specification satisfaction are subject to randomness, measurement noise, or epistemic uncertainty in model parameters. Instead of strictly Boolean satisfaction, PSTL enables specifying and analyzing the probability that a system trajectory satisfies a temporal logic formula, admitting both verification and synthesis tasks with quantified probabilistic guarantees. This article presents the core definitions, semantic frameworks, algorithmic techniques, and methodological variants of PSTL as developed in recent literature.

1. Core Syntax and Semantics of PSTL

PSTL extends STL formulas—which are interpreted over real-valued, discrete-time trajectories $x(0), x(1), \dots \in \mathbb{R}^n$—by introducing probabilistic operators. A canonical form expresses a “chance constraint”:

$P_{\geq 1-\delta}[\psi]$

which asserts that “the probability that trajectory $\xi$ satisfies the STL subformula $\psi$ is at least $1-\delta$” (Salamati et al., 2020). In formal terms, given a stochastic system model $M(\theta)$ parameterized by $\theta \in \Theta$, the chance-constraint semantics are:

$\mathrm{Pr}[M(\theta) \models \psi] \geq 1-\delta$

with the binary satisfaction function $f_\psi^\delta(\theta) = 1$ if the above holds, 0 otherwise.

Atomic predicates have affine form $\alpha(x) \ge 0$; Boolean connectives ($P_{\geq 1-\delta}[\psi]$0, $P_{\geq 1-\delta}[\psi]$1), and temporal operators (“bounded until” $P_{\geq 1-\delta}[\psi]$2) retain standard STL structure. Robustness semantics $P_{\geq 1-\delta}[\psi]$3 return real-valued scores, further enabling a quantitative notion of satisfaction.

Related, alternative PSTL grammars—such as those of (Yoo et al., 2015) and (Lee et al., 2021)—allow two-layer formulations, distinguishing between a Boolean-valued “instance” layer and an inner probabilistic “event” layer, and generalizing atomic predicates to probabilistic or statistical events (e.g., Bernoulli random fields or beliefs over latent states).

2. Probabilistic Reasoning Models and Inference

PSTL admits several frameworks for reasoning about the probability of formula satisfaction:

• Two-noise-layer models: System behavior may be driven by both intrinsic dynamical noise (process noise $P_{\geq 1-\delta}[\psi]$4) and extrinsic measurement noise ($P_{\geq 1-\delta}[\psi]$5), requiring a hierarchical approach to reasoning over both (Salamati et al., 2020). The system model is specified as

$P_{\geq 1-\delta}[\psi]$6

with both $P_{\geq 1-\delta}[\psi]$7 and $P_{\geq 1-\delta}[\psi]$8 i.i.d. Gaussian.

• Belief-space temporal logic: Probabilistic predicates reference beliefs over unobserved states, as in partially observable Markov decision processes (POMDPs), allowing explicit encoding of chance constraints on belief updates computed via recursive Bayesian inference (Yoo et al., 2015).
• Adaptive, data-driven models: When system parameters $P_{\geq 1-\delta}[\psi]$9 are uncertain and to be inferred from data, a prior $\xi$0 is updated to a posterior $\xi$1 using likelihoods derived from trajectory observations, typically via Gaussian models (Salamati et al., 2020, Sadigh et al., 2015). The confidence that the true system satisfies the property is then

$\xi$2

providing a scalar between 0 and 1 representing the posterior probability of satisfaction, given collected data.

• Random predicate semantics: Satisfaction itself is a random variable. The satisfaction probability is estimated by propagating predicate-level Bernoulli distributions through the temporal logic formula using conditional-independence approximations (Lee et al., 2021).

3. Algorithmic Approaches for Verification and Synthesis

Multiple algorithmic templates have been proposed for PSTL model checking, verification, and controller synthesis:

• Chance-constrained reduction: The probability constraint $\xi$3 is structurally reduced to a family of finite, probabilistic constraints on atomic predicates, typically in the form:

$\xi$4

For linear-Gaussian models, this is equivalent to an affine constraint involving the quantile of the normal distribution (Salamati et al., 2020, Sadigh et al., 2015):

$\xi$5

• Convex and mixed-integer programming: Chance constraints can be encoded as second-order cone constraints or via mixed-integer encodings (when temporal operators or logic connectives are involved), enabling efficient verification or receding-horizon controller synthesis using SOCP or MISOCP solvers (Sadigh et al., 2015).
• Robust polytope construction: Satisfaction sets in parameter space (e.g., the set $\xi$6) are approximated by polytopes using LP-based techniques (e.g., via Farkas’ lemma), with piecewise-affine refinement for nonlinearity in quantiles (Salamati et al., 2020).
• Monte Carlo and sampling-based confidence estimation: Confidence integrals are computed by rejection sampling in parameter space (for $\xi$7) or direct sampling of trajectories, with statistical error controlled by concentration inequalities (Salamati et al., 2020, Sadigh et al., 2015, Scher et al., 2022).
• Elliptical Slice Sampling (ESS) and Multilevel Splitting: Efficient, rejection-free MCMC sampling from Gaussian (or Gaussian mixture) trajectory distributions restricted to satisfying a robustness constraint enables scalable estimation of rare event probabilities under STL constraints, explicitly avoiding over-counting or conservative approximations (Scher et al., 2022).
• Gradient-based MAP inference: By recasting PSTL-constrained synthesis as MAP inference for control trajectories, differentiable semantics (log-odds propagations) enable optimization with gradient ascent and efficient GPU parallelism (Lee et al., 2021).

4. Specialized PSTL Variants and Comparison

Distinct PSTL formulation variants have been developed:

Variant	Main Feature	Reference
“Chance-constraint” PSTL	Global probability threshold $\xi$8	(Salamati et al., 2020)
Belief-space PSTL	Predicates over time-evolving agent beliefs	(Yoo et al., 2015)
PrSTL (Probabilistic STL)	Adaptive logic, chance constraints on parametric predicates, update with data	(Sadigh et al., 2015)
Random STL (RSTL)	Randomized predicate satisfaction, smooth probability propagation	(Lee et al., 2021)

PSTL reduces to standard STL when all randomness is degenerate (Dirac delta), beliefs are deterministic, or predicates are exact. The logic recovers temporal logic for POMDPs (when beliefs are updated via Bayesian inference) in the appropriate limit (Yoo et al., 2015, Lee et al., 2021).

A notable alternative to probability-centric PSTL is STL “erosion” (Ma et al., 11 Feb 2025), which reduces the probabilistic model checking problem to deterministic STL verification over a tightened (eroded) specification, using high-probability bounds on stochastic perturbations. This approach avoids explicit chance constraints but guarantees correctness (but not completeness) by design.

5. Illustrative Numerical Examples

Numerical studies provide significant insights:

• Data-driven PSTL with two-state model: For

$\xi$9

and output $\psi$0, the PSTL property $\psi$1 is checked by LP over $\psi$2-space and post-data Bayesian updating (Salamati et al., 2020). Prior to data, confidence is low ($\psi$3 2.7%). With $\psi$4 experiments, posterior confidence rises sharply ($\psi$5 90%).

• Quadrotor safety under PrSTL: A quadrotor navigates under an uncertain ceiling with probabilistic classifier boundaries updated from sensor data, enforcing $\psi$6. Online SOCP ensures safety as knowledge is refined (Sadigh et al., 2015).
• Probabilistic verification by STL erosion: For a double integrator with a safety reach-avoid specification, erosion with appropriate radii (from high probability bounds) yields eroded obstacles and goals on which deterministic STL verification guarantees original probabilistic satisfaction (Ma et al., 11 Feb 2025).
• Sampling-based verification: Elliptical Slice Sampling with multilevel splitting enables efficient estimation of small satisfaction probabilities (down to $\psi$7), even for high-dimensional (100+) trajectory spaces or rare events (Scher et al., 2022).
• Controller synthesis under belief-space PSTL: Receding-horizon forward search with beam pruning solves high-dimensional belief-space surveillance and search tasks for UAVs, scaling tractably to horizon lengths $\psi$8–$\psi$9 (Yoo et al., 2015).

6. Theoretical Guarantees, Limitations, and Computational Complexity

Theoretical properties of PSTL frameworks vary:

• Soundness: All approaches provide probabilistic soundness with respect to the model assumptions and noise structure, i.e., computed confidence is a lower bound on true satisfaction probability (Salamati et al., 2020, Ma et al., 11 Feb 2025, Sadigh et al., 2015).
• Completeness: For convex/decomposable formulations (e.g., PWA or SOCP relaxations), exactness is achieved in the limit (as cell partitioning/refinements increase) (Salamati et al., 2020, Sadigh et al., 2015).
• Scalability: Sampling-based and gradient approaches—especially when GPU-accelerated—allow scaling synthesis and verification to long horizons and high-dimensional state spaces (Lee et al., 2021, Scher et al., 2022).
• Adaptivity: Bayesian and data-driven PSTL updates satisfaction probabilities in real time as more data is acquired, allowing the logic to adapt to evolving knowledge (Sadigh et al., 2015, Salamati et al., 2020).
• Limitations: Most tractable methods require affine or linear-Gaussian models; extension to nonlinear or non-Gaussian dynamics often requires additional approximation, linearization, or conservative bounding (Sadigh et al., 2015, Scher et al., 2022). Nonconvexity precludes global optimality for gradient or sampling algorithms (Lee et al., 2021).

7. Significance and Applications

PSTL provides a rigorous, compositional, and quantitatively meaningful language for specifying, verifying, and synthesizing controllers for systems in which uncertainty is fundamental. Domains include learning-based or data-driven CPS, robotic planning under sensor noise, formal safety analysis of autonomy in stochastic environments, and adaptive mission planning with belief-space reasoning (Salamati et al., 2020, Sadigh et al., 2015, Yoo et al., 2015, Scher et al., 2022).

Key contributions include seamless integration of probabilistic inference (Bayesian and frequentist), systematic connection to robust and chance-constrained optimization (SOCP, MISOCP), and establishment of scalable, efficient algorithms for real-time control and verification. The abstraction of temporal logic specifications to soft, probabilistically meaningful constraints positions PSTL as a central framework for formal reasoning about complex, uncertain cyber-physical systems.