STL Decomposition Techniques

• STL decomposition is a method for partitioning complex signal temporal logic formulas into smaller, manageable sub-problems with disjoint time intervals.
• It employs strategies like time-interval and multi-agent decomposition to enable staged control synthesis and maintain global satisfaction guarantees.
• The approach reduces combinatorial complexity and enhances computational efficiency through MILP and convex optimization techniques in real-world systems.

Signal Temporal Logic (STL) decomposition refers to a collection of algorithmic strategies for partitioning complex STL specifications—typically encoding spatiotemporal goals for systems such as robots or multi-agent collectives—into smaller, logically coupled sub-problems. These sub-problems are distributed over subsystems, subteams, time intervals, or progress milestones, providing computational advantages, scalability, and tractable controller synthesis while retaining correctness guarantees. STL decomposition frameworks leverage structural properties of formula syntax, multi-agent architectures, task graphs, and temporal progress, making STL decomposition a central paradigm in the formal synthesis, verification, and model predictive control (MPC) of cyber-physical systems under rich logical constraints.

1. Syntactic and Temporal Decomposition of STL Formulas

A recurring principle in STL decomposition is the syntactic partitioning of a global temporal formula $\varphi$ into a sequence or conjunction of sub-formulae $\varphi_i$ with pairwise disjoint time intervals. Given a bounded-horizon specification $\varphi^{[S_\varphi,T_\varphi]}$, one constructs a representation of the form

$\varphi = \bigwedge_{i=1}^N \varphi_i^{[S_i, T_i]}$

with $S_1 = S_\varphi$, $T_N = T_\varphi$, and $T_i \le S_{i+1}$ for all $i$ (Yu et al., 2022).

This structure enables the division of long-horizon STL tasks into independent or weakly coupled temporal tasks, supporting staged control synthesis and reducing worst-case combinatorial complexity. In the context of model predictive control for STL, shrinking-horizon MPC is sequentially applied to each sub-formula, with each optimization window limited to the active interval $[S_i, T_i]$.

A semantically similar approach underpins the reachability/invariance flattening techniques developed for temporal decomposition, where every bounded-time STL formula can be expressed as a set of reachability constraints (enforcing the existence of a satisfaction point for a predicate in a given interval) and invariance constraints (requiring continuous satisfaction of a predicate over an interval). These progress conditions can be systematically derived by recursive traversal of formula syntax, giving a strict decomposition of nested and conjunctive STL into atomic scheduling constraints for controller synthesis (Kapoor et al., 2024, Liu et al., 20 Apr 2026).

2. Decomposition in Multi-Agent and Decentralized Contexts

When STL specifications couple multiple agents, decomposition becomes both a logical and structural problem. A principal scenario involves partitioning global STL formulas, defined over the joint state space of a multi-agent system, into local STL specifications for spatially or semantically disjoint subteams. This is formalized by identifying agent subteams $\mathcal V_1, \dots, \mathcal V_v$ and constructing local formulas $\varphi_i$0 over subteam-specific variables $\varphi_i$1, such that satisfaction of all $\varphi_i$2 implies satisfaction of the global task $\varphi_i$3 (Charitidou et al., 2021).

Predicate functions associated with global tasks are parameterized as axis-aligned hypercubes (or more generally, hyper-rectangles), with the parameters optimized (e.g., via convex programming) to guarantee the under-approximation of the global feasible set by the intersection of local feasible sets. This yields decentralized controller design, often with provable soundness: any trajectory ensemble satisfying all local STL tasks is guaranteed to satisfy the original global STL formula.

Extensions to communication-constrained systems refine this methodology by further decomposing collaborative predicates (those dependent on multi-agent relative states) into subtasks associated with 1-hop neighborhoods, according to the sparsity pattern of the communication graph. This enables decentralized STL task realization subject to communication constraints, handled via structured convex programming and explicit conflict exclusion constraints (Marchesini et al., 2024).

3. Computational and Algorithmic Foundations

STL decomposition frameworks employ a range of algorithmic tools to support tractable synthesis:

• Mixed-Integer Linear Programming (MILP): In time-interval decomposition for MPC, each subtask's satisfaction is encoded via standard big-M MILP formulations for STL temporal operators (e.g., "eventually," "always," "until"), subject to system dynamics and actuation constraints (Yu et al., 2022).
• Convex Optimization and QCQP: For multi-agent decomposition, the search for optimal local predicate parameterizations (e.g., hypercube center and radius) is cast as a convex quadratic-constrained quadratic program, leveraging the concavity of global predicate functions to guarantee convex feasible sets (Charitidou et al., 2021, Marchesini et al., 2024).
• Scheduling and Incremental Planning: Task decomposition into reachability/invariance constraints enables sequential planning via partial orders, time-variable instantiation, and incremental controller synthesis. Each atomic planning subproblem involves the satisfaction of a reduced set of STL constraints over a narrow interval, reducing overall binary variable count and enabling scalable synthesis even in the presence of deep temporal nesting (Kapoor et al., 2024, Liu et al., 20 Apr 2026).
• Conflict Detection/Exclusion: For collaborative and decentralized scenarios, formal procedures for detecting and excluding conflicting conjunctions of local tasks (e.g., conflicting invariance requirements over intersecting intervals) are realized through convex set inclusion constraints and supporting lemmas, ensuring the feasibility of the decomposed program (Marchesini et al., 2024).

4. Satisfaction Guarantees and Theoretical Properties

STL decomposition methodologies are sound by construction: the global satisfaction of the original specification is implied by the satisfaction of all subtasks, under the prescribed parameterizations and constraints. This is formalized in several frameworks:

• In time-interval decomposition MPC, satisfaction is guaranteed by imposing terminal constraints that encode inner-approximations of the set of states from which all remaining subgoals are feasible. Recursive feasibility is ensured via backward induction over the decomposition stages, as long as the initial stage is feasible and all transitions satisfy the terminal set constraints (Yu et al., 2022).
• For local-predicate (hypercube) parameterization in multi-agent settings, satisfaction of all local STL tasks provides a robust under-approximation of the global feasible set, provably enforcing overall STL satisfaction (Charitidou et al., 2021).
• In reachability/invariance flattening schemes, correctness is certified by lemmas ensuring that satisfaction of all progress conditions (for some feasible assignment of symbolic timing variables) is strictly equivalent to satisfaction of the original STL formula (Liu et al., 20 Apr 2026, Kapoor et al., 2024). Soundness of timing-allocation procedures and DAG-based scheduling is similarly established by construction.

5. Representative Case Studies and Performance Impact

Empirical results consistently demonstrate that STL decomposition provides large computational benefits compared to monolithic or full-horizon formulations:

• MPC with Time-Interval Decomposition: On a double-integrator robot task involving three temporal goals, time-interval decomposition (two-stage shrinking-horizon MPC) was over $\varphi_i$4 faster than non-decomposed MILP MPC, with only a small cost in solution optimality and guaranteed satisfaction of all temporal objectives. The method meets real-time sampling constraints that monolithic approaches cannot (Yu et al., 2022).
• Decentralized Multi-Agent Synthesis: For five-agent 2D formation and coverage tasks, per-agent convex programs were solved in $\varphi_i$5 seconds, providing robust satisfaction of the global formula through decentralized controller execution (Charitidou et al., 2021). Communication-constrained decomposition in eight-agent systems required $\varphi_i$6 seconds and enabled all tasks to be realized with only 1-hop communication (Marchesini et al., 2024).
• Incremental Decomposition in Trajectory Synthesis: STLInc decomposition (Kapoor et al., 2024) and DAG-STL (Liu et al., 20 Apr 2026) provide orders-of-magnitude speedup in planning for deeply nested and long-horizon STL goals in both linear and nonlinear systems, with tasks that caused monolithic encodings to time out successfully solved in sub-second to seconds wall time. Robustness scores of the decomposed approaches were competitive with or exceeded those of full-horizon optimization.

6. Hierarchical and Data-Driven Extensions

Recent work on data-driven, model-agnostic planning under STL (e.g., "DAG-STL") shows that formula decomposition enables the separation of logical reasoning (progress condition DAGs) from the realization of trajectories in state space, including when explicit models are unavailable (Liu et al., 20 Apr 2026). This supports zero-shot generalization to unseen STL tasks and tasks spanning large horizons, by focusing synthesis on short segments between progress milestones, using learned or generative models for trajectory completion conditioned on these logical waypoints. Computational complexity is effectively mitigated through pruning, merging of progress conditions, and anytime search, operationally enabling high-dimensional STL planning infeasible via direct optimization.

7. Limitations, Open Questions, and Extensions

STL decomposition is restricted by several factors:

• The expressivity of decomposition is limited by the available STL syntax fragment (e.g., disjunction-free forms or limitations on nesting).
• Decomposition is sound but typically not complete: not every global STL formula is amenable to a fully parallel or local decomposition compatible with arbitrary agent or communication graph structures.
• Robust satisfaction margins depend on the tightness of local/global feasible set approximations; parameter choices may be conservative due to vertex or Minkowski sum inclusion constraints.
• Complexity reduction for arbitrary formulas with extensive disjunction or interleaved nested intervals remains a challenge; further research addresses DAG-homomorphic or symbolic flattening as a remedy (Liu et al., 20 Apr 2026).

Ongoing work explores integrating learning-based prediction with symbolic progress scheduling, extending decomposition methods to richer logics, and closing the gap between global planning-level correctness and execution-level feasibility in uncertain, dynamic environments.