Continuous-Time Signal Temporal Logic

• Continuous-Time Signal Temporal Logic (CT-STL) is a dense-time formalism that evaluates temporal properties of continuous trajectories using real-time semantics.
• It supports diverse applications such as monitoring, control synthesis, and robustness analysis through quantitative measures like min/max and AGIM robustness.
• CT-STL methodologies balance semantic fidelity and computational scalability with techniques including barrier certificates, timed automata, and scenario-based verification.

Searching arXiv for recent and foundational papers on continuous-time STL and closely related topics. Continuous-Time Signal Temporal Logic (CT-STL) denotes the use of Signal Temporal Logic over dense-time trajectories $x:\mathbb{R}_{\ge 0}\to \mathbb{R}^n$, with temporal operators interpreted over real-valued intervals and satisfaction defined on continuous signals rather than only on sampled traces. In the literature, STL was introduced for monitoring temporal properties of continuous-time signals, and several later frameworks treat dense time as intrinsic, whether for monitoring, deductive verification, motion planning, control synthesis, robustness analysis, or uncertainty-aware verification (Dluhoš et al., 2012, Ahmad et al., 2021). This suggests that CT-STL is best understood not as a single distinct logic separate from STL, but as STL and its variants under explicitly continuous-time semantics.

1. Dense-time semantics and logical structure

A standard presentation of CT-STL uses formulas

$$\varphi ::= \top \mid \mu \mid \neg \varphi \mid \varphi_1 \wedge \varphi_2 \mid \varphi_1 \bm U_{[a,b]} \varphi_2,$$

with derived operators

$$\bm F_{[a,b]}\varphi \equiv \top \bm U_{[a,b]} \varphi,\qquad \bm G_{[a,b]}\varphi \equiv \neg \bm F_{[a,b]} \neg \varphi.$$

An atomic predicate $\mu$ is induced by a real-valued function $g:\mathbb R^n\to\mathbb R$ through $g(x(t))\ge 0$. Satisfaction is evaluated at real time $t$: predicates are checked pointwise on the continuous trajectory, $\bm G_{[a,b]}$ quantifies universally over all $t'\in[t+a,t+b]$, $\bm F_{[a,b]}$ existentially over the same interval, and $$\varphi ::= \top \mid \mu \mid \neg \varphi \mid \varphi_1 \wedge \varphi_2 \mid \varphi_1 \bm U_{[a,b]} \varphi_2,$$0 requires a witness time in $$\varphi ::= \top \mid \mu \mid \neg \varphi \mid \varphi_1 \wedge \varphi_2 \mid \varphi_1 \bm U_{[a,b]} \varphi_2,$$1 together with prefix satisfaction up to that witness (Uzun et al., 5 Apr 2026).

Dense time is not merely a notational choice. Several later developments rely on the fact that the signal domain is $$\varphi ::= \top \mid \mu \mid \neg \varphi \mid \varphi_1 \wedge \varphi_2 \mid \varphi_1 \bm U_{[a,b]} \varphi_2,$$2, not $$\varphi ::= \top \mid \mu \mid \neg \varphi \mid \varphi_1 \wedge \varphi_2 \mid \varphi_1 \bm U_{[a,b]} \varphi_2,$$3. For example, STdL interprets temporal formulas over hybrid traces whose subtraces are placed on a global real-time axis, and its semantics explicitly quantify over all real time points in bounded intervals (Ahmad et al., 2021). Likewise, STL$$\varphi ::= \top \mid \mu \mid \neg \varphi \mid \varphi_1 \wedge \varphi_2 \mid \varphi_1 \bm U_{[a,b]} \varphi_2,$$4 is formulated over finite-length dense-time signals $$\varphi ::= \top \mid \mu \mid \neg \varphi \mid \varphi_1 \wedge \varphi_2 \mid \varphi_1 \bm U_{[a,b]} \varphi_2,$$5, with semantics given over triples $$\varphi ::= \top \mid \mu \mid \neg \varphi \mid \varphi_1 \wedge \varphi_2 \mid \varphi_1 \bm U_{[a,b]} \varphi_2,$$6 so that the freeze operator can compare values at two continuous times (Dluhoš et al., 2012).

A recurrent misconception is that CT-STL has one canonical syntax and one canonical proof theory. The literature instead contains multiple dense-time formalisms. Some works use full bounded until syntax, while others deliberately restrict the temporal fragment to bounded $$\varphi ::= \top \mid \mu \mid \neg \varphi \mid \varphi_1 \wedge \varphi_2 \mid \varphi_1 \bm U_{[a,b]} \varphi_2,$$7 and $$\varphi ::= \top \mid \mu \mid \neg \varphi \mid \varphi_1 \wedge \varphi_2 \mid \varphi_1 \bm U_{[a,b]} \varphi_2,$$8, exclude nested temporal operators, or distinguish state formulas from trace formulas for deductive purposes (Ahmad et al., 2021). Some extensions, such as STL$$\varphi ::= \top \mid \mu \mid \neg \varphi \mid \varphi_1 \wedge \varphi_2 \mid \varphi_1 \bm U_{[a,b]} \varphi_2,$$9, add a freeze operator precisely because standard STL is not expressive enough to distinguish detailed oscillatory properties on continuous trajectories (Dluhoš et al., 2012).

2. Quantitative semantics and robustness notions

The standard quantitative semantics of CT-STL assign each formula a real-valued robustness $$\bm F_{[a,b]}\varphi \equiv \top \bm U_{[a,b]} \varphi,\qquad \bm G_{[a,b]}\varphi \equiv \neg \bm F_{[a,b]} \neg \varphi.$$0. For predicates,

$$\bm F_{[a,b]}\varphi \equiv \top \bm U_{[a,b]} \varphi,\qquad \bm G_{[a,b]}\varphi \equiv \neg \bm F_{[a,b]} \neg \varphi.$$1

negation changes sign; conjunction and disjunction use $$\bm F_{[a,b]}\varphi \equiv \top \bm U_{[a,b]} \varphi,\qquad \bm G_{[a,b]}\varphi \equiv \neg \bm F_{[a,b]} \neg \varphi.$$2 and $$\bm F_{[a,b]}\varphi \equiv \top \bm U_{[a,b]} \varphi,\qquad \bm G_{[a,b]}\varphi \equiv \neg \bm F_{[a,b]} \neg \varphi.$$3; and temporal operators lift these extrema over real-time windows, e.g.

$$\bm F_{[a,b]}\varphi \equiv \top \bm U_{[a,b]} \varphi,\qquad \bm G_{[a,b]}\varphi \equiv \neg \bm F_{[a,b]} \neg \varphi.$$4

Its sign matches Boolean satisfaction: positive robustness implies satisfaction and negative robustness implies violation (Das et al., 16 Apr 2026).

For CT-STL over continuous trajectories, an important regularity property is that if predicate functions are Lipschitz, then the STL robustness map is Lipschitz in the signal argument with respect to the sup norm: $$\bm F_{[a,b]}\varphi \equiv \top \bm U_{[a,b]} \varphi,\qquad \bm G_{[a,b]}\varphi \equiv \neg \bm F_{[a,b]} \neg \varphi.$$5 This property is central in later resilience and scenario-based analyses because it permits lifting finite sampled checks to neighborhoods of unsampled continuous disturbances (Das et al., 16 Apr 2026).

Standard min/max robustness is not the only quantitative semantics in dense time. AGIM robustness replaces worst-case aggregation by operator-sensitive averaging: arithmetic integrals for existential-style satisfaction and geometric product-integral means for universal-style satisfaction. It is normalized to $$\bm F_{[a,b]}\varphi \equiv \top \bm U_{[a,b]} \varphi,\qquad \bm G_{[a,b]}\varphi \equiv \neg \bm F_{[a,b]} \neg \varphi.$$6, is claimed to preserve Boolean soundness, and was introduced to mitigate the masking effect and locality of classical robustness in falsification and control synthesis (Mehdipour et al., 2019).

A different line of work develops smooth dense-time parameterizations that remain sign-exact. A recent trajectory-optimization formulation defines $$\bm F_{[a,b]}\varphi \equiv \top \bm U_{[a,b]} \varphi,\qquad \bm G_{[a,b]}\varphi \equiv \neg \bm F_{[a,b]} \neg \varphi.$$7 on the dense numerical trajectory induced by the integration scheme and proves

$$\bm F_{[a,b]}\varphi \equiv \top \bm U_{[a,b]} \varphi,\qquad \bm G_{[a,b]}\varphi \equiv \neg \bm F_{[a,b]} \neg \varphi.$$8

The point is not equality with classical robustness values, but exact preservation of the truth set together with a smoother optimization landscape for dense-time $$\bm F_{[a,b]}\varphi \equiv \top \bm U_{[a,b]} \varphi,\qquad \bm G_{[a,b]}\varphi \equiv \neg \bm F_{[a,b]} \neg \varphi.$$9, $\mu$0, and $\mu$1 constraints (Uzun et al., 5 Apr 2026).

CT-STL robustness is also not limited to spatial perturbations. In continuous-time multi-agent motion planning, asynchronous temporal robustness quantifies the largest independent time shifts of agents’ trajectories that preserve satisfaction of a joint STL specification. The resulting recursive robustness uses continuous-time temporal perturbations rather than signal-value perturbations, and the implemented optimization relies on a segment-based underapproximation over Bézier trajectory segments (Verhagen et al., 2023).

3. Monitoring, theorem proving, and uncertainty-aware verification

Monitoring is historically one of the primary CT-STL use cases. STL$\mu$2 extends dense-time STL with a freeze operator that stores the current time for later comparison, enabling formulas about local maxima, minima, damped oscillation, and phase relations. Monitoring then becomes a geometric computation over a two-dimensional satisfaction set

$\mu$3

because both current time and frozen time matter. For piecewise-linear signals, the prototype algorithm represents these sets as unions of polygons, with reported complexity $\mu$4 for a formula of size $\mu$5 over a signal partitioned into $\mu$6 linear pieces (Dluhoš et al., 2012).

Deductive verification introduces a different perspective. STdL combines a bounded fragment of STL with differential dynamic logic, distinguishing state formulas, trace formulas, and program modalities such as $\mu$7. Its semantics are dense-time and length-aware: a trace formula is only required when the execution is long enough, which preserves modal duality and avoids imposing nonlocal time obligations on zero-duration traces. The proof theory uses timing instrumentation with a fresh clock variable $\mu$8 to shift intervals through sequential composition, thereby reducing bounded temporal obligations to dL-style reasoning about hybrid programs (Ahmad et al., 2021).

Reachability-based verification adds a third layer, namely uncertainty internal to the verification procedure itself. Incremental verification over nonlinear continuous-time systems under uncertainty evaluates STL on reachable tubes rather than on single trajectories. The method introduces Boolean interval arithmetic so that predicate truth values can become $\mu$9 when a reachable enclosure intersects both satisfying and violating regions. These uncertainty markers propagate through the satisfaction tree, including nested formulas, and only the reachable sets responsible for indeterminacy are refined. The framework is designed for both offline and online operation, with the explicit goal of reducing satisfaction ambiguity without recomputing unaffected parts of the tube (Besset et al., 18 Nov 2025).

Taken together, these strands show that CT-STL verification spans at least three regimes: dense-time signal monitoring, deductive proof over hybrid executions, and interval-valued reasoning over uncertain reachable tubes. Their semantics are not interchangeable, but all are explicitly continuous-time.

4. Control synthesis and motion planning

A central CT-STL challenge is that continuous-time satisfaction is stronger than sampled-time satisfaction. Under zero-order hold control, a trajectory may satisfy a predicate at all optimization nodes and still violate it between nodes. This motivates synthesis methods that certify inter-sample behavior, especially for $g:\mathbb R^n\to\mathbb R$0-type path constraints. One offline MIQP framework for linear systems uses standard mixed-integer STL encodings at selected time instants together with control barrier function constraints that guarantee continuous satisfaction between updates, including predicates with higher relative degree (Yang et al., 2019). A related non-uniform sampled formulation for linear continuous-time systems with ZOH control builds mixed-integer plans from active temporal instants and derives CBF lower bounds for continuous-time $g:\mathbb R^n\to\mathbb R$1 together with predicate lower bounds for continuous-time $g:\mathbb R^n\to\mathbb R$2, thereby certifying satisfaction over the full inter-sample intervals rather than only at nodes (Yang et al., 2020).

Barrier-based synthesis extends beyond linear sampled-data planning. One CBF framework for single-integrator systems with bounded speed translates a broad STL fragment into feasible subtask sequences, uses a primary CBF for the current target, and a secondary CBF to preserve future feasibility under actuation constraints and overlapping deadlines (Buyukkocak et al., 2022). For nested temporal operators, a more recent approach introduces the signal temporal logic tree (sTLT), which compiles a nested STL formula into a hierarchy of timed set constraints. Its semantics are exact for formulas in a designated normal form and under-approximate otherwise; the resulting tree then directly guides the design of time-varying CBFs and an online CBF-constrained controller (Yu et al., 2023). Neural synthesis pushes the same idea into learned certificates: time-varying neural control barrier functions $g:\mathbb R^n\to\mathbb R$3 and bounded neural controllers $g:\mathbb R^n\to\mathbb R$4 are trained for conjunctions of bounded $g:\mathbb R^n\to\mathbb R$5 and $g:\mathbb R^n\to\mathbb R$6 formulas, together with a Lipschitz-based validity condition that lifts sampled barrier inequalities to the continuous state-time domain (Jagabathula et al., 15 Dec 2025).

Automaton-guided synthesis offers a different route. For non-nested STL over continuous-time nonlinear systems with polynomial predicates, a time-partitioned timed NFA can be constructed by separating the formula along temporal boundaries, converting the pieces to finite automata, and reassembling them into a timed product with a predicate-invariant partition of the continuous state space. This product automaton then guides a sampling-based motion tree both spatially and temporally, and the resulting algorithm is proved correct and probabilistically complete (Ho et al., 2022).

Continuous-time implementations also raise execution-level questions. An event-triggered feedback framework replaces a previously developed continuous-time STL controller with a sampled piecewise-constant control law, where triggering is based on state deviation and a maximum inter-event time. The analysis shows that if the input approximation error remains bounded, then the transformed prescribed-performance error remains bounded as well, preserving the STL guarantee and excluding Zeno behavior (Lindemann et al., 2020).

5. Disturbance tolerance, robustness, and resilience

CT-STL robustness is often interpreted as the margin of a single trajectory, under a fixed disturbance realization, with respect to a given formula. A recent resilience framework for continuous-time systems formalizes a different question: for a given initial state $g:\mathbb R^n\to\mathbb R$7, what is the largest disturbance radius $g:\mathbb R^n\to\mathbb R$8 such that every disturbance signal $g:\mathbb R^n\to\mathbb R$9 satisfying $g(x(t))\ge 0$0 preserves STL satisfaction? The disturbed trajectory family is

$g(x(t))\ge 0$1

and the resilience function is

$g(x(t))\ge 0$2

provided the nominal trajectory satisfies $g(x(t))\ge 0$3. Robustness is therefore a signal-level quantity for one execution; resilience is a disturbance-admissibility margin at the system/specification/initial-state level (Das et al., 16 Apr 2026).

For linear disturbed systems $g(x(t))\ge 0$4, the framework derives a time-varying trajectory envelope rather than a full reachable set. With $g(x(t))\ge 0$5 and a real-part Jordan surrogate $g(x(t))\ge 0$6, every disturbed trajectory satisfies

$g(x(t))\ge 0$7

This leads to a family of bounding trajectories $g(x(t))\ge 0$8, indexed by $g(x(t))\ge 0$9, and to the lower bound

$t$0

The result is explicitly sufficient rather than exact, because the envelope over-approximates the true disturbed trajectory set and then requires all envelope trajectories to satisfy the formula (Das et al., 16 Apr 2026).

To make the optimization tractable, the same work samples finitely many envelope parameters $t$1, solves a scenario program, and lifts the result to all disturbances using a Lipschitz bound on robustness over the envelope. If $t$2 is the sampled optimum and the covering radius is $t$3, then

$t$4

is sufficient to guarantee positive robustness over the entire disturbance ball. The nonlinear extension is local: it linearizes around an equilibrium, bounds the Taylor remainder using the Hessian, and subtracts the resulting $t$5 margin from the linearized resilience lower bound. The framework is therefore sound, but conservative in several identifiable ways (Das et al., 16 Apr 2026).

6. Conceptual fault lines and research directions

The literature makes clear that CT-STL is not synonymous with one uniform methodology. Some works use purely Boolean dense-time semantics and theorem proving over hybrid traces (Ahmad et al., 2021); some use quantitative semantics centered on robustness or temporal robustness (Mehdipour et al., 2019, Verhagen et al., 2023); some enforce continuous-time satisfaction exactly on a dense numerical trajectory representation (Uzun et al., 5 Apr 2026); and others use conservative underapproximations or barrier-based sufficient conditions for particular fragments (Yu et al., 2023, Jagabathula et al., 15 Dec 2025). This suggests that CT-STL is best viewed as a family of dense-time semantic and algorithmic regimes rather than a single mature toolbox.

A second fault line concerns the difference between sampled enforcement and continuous-time enforcement. Several planning and trajectory-optimization papers are explicit that nodewise satisfaction does not imply satisfaction of the underlying continuous-time trajectory, especially for always-type safety clauses. This is why CBF-based inter-sample certificates, dense-time predicate lower bounds, and dense integration-grid evaluation have become central design patterns in CT-STL synthesis (Yang et al., 2019, Yang et al., 2020, Uzun et al., 5 Apr 2026).

A third issue is conservativeness versus tractability. Exact dense-time semantics, nonlinear reachability, nested temporal operators, and uncertainty-aware verification all increase computational burden. The published responses include time partitioning and automata for non-nested formulas, under-approximating tree semantics for nested formulas, Jordan-envelope lower bounds for resilience, and neural TVCBFs with Lipschitz validity certificates (Ho et al., 2022, Yu et al., 2023, Das et al., 16 Apr 2026, Jagabathula et al., 15 Dec 2025). A plausible implication is that future progress in CT-STL will continue to depend on intermediate representations that are neither purely logical nor purely numerical: timed automata, tree semantics, barrier-certified time-varying sets, and interval-valued satisfaction structures all play that role.

Several open directions are explicitly identified. Resilience analysis points to integration with interval-based STL monitoring and to reducing conservativeness through piecewise linearization or nonlinear reachability such as zonotopes (Das et al., 16 Apr 2026). Smooth dense-time trajectory optimization points to free-final-time formulations and alternatives to explicitly storing dense trajectories (Uzun et al., 5 Apr 2026). Neural barrier synthesis remains limited to a bounded fragment and depends on nontrivial validity and training assumptions (Jagabathula et al., 15 Dec 2025). More broadly, the persistent tension between semantic fidelity, computational scalability, and support for full nested STL remains one of the defining structural problems of CT-STL research.