LTLf Synthesis with Multiple Properties

Updated 22 January 2026

LTLf synthesis with multiple properties is a method for creating reactive strategies that guarantee the maximal achievement of finite-trace temporal objectives under conflicting conditions.
The approach utilizes symbolic fixed-point algorithms and Boolean goal variables to compute maximal realizable subsets of LTLf specifications efficiently.
Empirical results demonstrate that BDD-based methods significantly outperform enumeration techniques, scaling to large state spaces and complex objectives.

Linear Temporal Logic over finite traces (LTLf) synthesis with multiple properties addresses the synthesis of reactive strategies guaranteed to satisfy as many objectives as possible out of a given set—where the objectives, typically expressed as LTLf formulas, may not all be simultaneously realizable. The central challenge is to efficiently compute the maximal subsets of realizable properties and to synthesize a strategy that ensures, from any given system state, the satisfaction of exactly the maximal set possible. Symbolic and compositional methods sidestep the exponential blowup inherent in naïve enumeration of all property subsets, yielding scalable and practically effective solutions.

1. Formalization of Multi-Property LTLf Synthesis

Let $\mathrm{AP}$ be a finite set of atomic propositions, partitioned into environment-controlled variables $X$ and system-controlled outputs $Y$ . Each LTLf specification $\varphi_i$ in a goal set $G=\{\varphi_1,\dots,\varphi_n\}$ is interpreted over finite traces on $X\cup Y$ . A system strategy is a function $g:(2^X)^*\rightarrow 2^Y$ , mapping input histories to outputs; its execution against an infinite environment input sequence produces a finite trace

$\rho_k(g,\xi) = (g(\epsilon)\cup X_0)\,(g(X_0)\cup X_1)\cdots(g(X_0\cdots X_{k-1}) \cup X_k)$

for some $k\geq 0$ .

A subset $C\subseteq G$ is realizable if there exists a strategy $g$ such that, for every infinite environment play, at least one prefix $\rho_k(g,\xi)$ satisfies all $\varphi_i\in C$ . The multi-property synthesis problem is to decide, for all $C\subseteq G$ , which subsets are realizable and to construct a strategy $g$ that—upon invocation with any realizable $C$ —guarantees exactly the objectives from $C$ are satisfied.

2. Product Construction and Symbolic Encoding

Each $\varphi_i$ is compiled into a DFA $A_i=(2^{X\cup Y},S_i,q^0_i,\delta_i,F_i)$ accepting traces that satisfy $\varphi_i$ . The synchronous product automaton $A^\times = (2^{X\cup Y}, S^\times, s_0^\times, \delta^\times)$ is constructed as

$S^\times = S_1 \times \dots \times S_n, \quad s_0^\times = (q^0_1, ..., q^0_n)$

with

$\delta^\times\bigl((q_1, \ldots, q_n), a\bigr) = (\delta_1(q_1, a), \ldots, \delta_n(q_n, a)).$

A product state $s=(q_1, ..., q_n)$ satisfies a goal set $C$ iff $q_i \in F_i$ for all $\varphi_i \in C$ .

To represent subsets symbolically, Boolean goal variables $k_1, ..., k_n$ are introduced: an assignment $K \subseteq \{k_1,\dots, k_n\}$ encodes the subset $C = \{\varphi_i \mid k_i \in K\}$ . The base winning condition is given as

$w_0(Z, K) = \bigwedge_{i=1}^n (k_i \Rightarrow f_i(Z_i))$

where $f_i(Z_i)$ characterizes acceptance in DFA $A_i$ .

3. Symbolic Fixed-Point Algorithm

The core solution relies on a fixed-point characterization over $(s, C)$ pairs using the controllable multi-property predecessor operator: $PreMC(\mathcal{E}) = \{ (s, C) \mid \exists Y \in 2^Y . \forall X \in 2^X: (\delta^\times(s, Y \cup X), C) \in \mathcal{E}\}.$ The least fixed-point is computed via

$Win^M_0 = \{(s, C) \mid s \models C\}$

$Win^M_{i+1} = Win^M_i \cup PreMC(Win^M_i)$

until convergence. In $\mu$ -calculus notation,

$Win^M = \mu Z. \{(s, C) \mid s \models C\} \cup \{(s, C) \mid \exists Y. \forall X. (\delta^\times(s, Y \cup X), C) \in Z\}$

Exact realizability can then be determined by the inclusion $(s, C) \in Win^M$ , and maximal sets are extracted by downward closure.

The symbolic algorithm operates on BDD representations of states, inputs, outputs, and goal bits, iterating quantifier-eliminated BDD operations efficiently. The following pseudocode summarizes the symbolic fixed-point iteration:

// Initialization
w := w0(Z,K)
t := w
repeat
  t' := t ∨ [¬w(Z,K) ∧ ∀X. w(η(X,Y,Z),K)]
  w' := ∃Y. t'
  if w' == w then break else (w, t) := (w', t')
end
// At fixpoint: w(Z,K) encodes Win^M, t(Z,Y,K) encodes which Y to pick

Universal quantification over inputs and existential quantification over outputs are implemented via standard BDD quantification.

4. Theoretical Properties and Complexity

Constructing each DFA $A_i$ is $2^{2^{O(|\varphi_i|)}}$ in the worst case; the synchronous product has at most $\prod|S_i|$ states. The explicit fixed-point computation over $S^\times \times 2^n$ is therefore $2$EXPTIME in formula size and number of goals. Symbolic implementation, however, often compresses exponentially many subsets into compact BDDs, offering much better scalability in practice.

Soundness and completeness are guaranteed:

$(s, C) \in Win^M$ iff $C$ is realizable from $s$ (by induction and Knaster-Tarski).
Extractable strategies (transducers) can be built from the fixed-point solution: for any $C$ , the system selects outputs $Y$ according to $t(Z, Y, K)$ so as to realize exactly $C$ .
Maintaining only maximal $(s, C)$ pairs ensures that only those sets are kept for which no strict superset is also realizable, exploiting the monotonicity property.

5. Comparison with Alternative Approaches

In contrast to classical enumeration, which analyzes $2^n$ property subsets separately, the symbolic approach uses Boolean variables and monotonicity to represent and reason about exponentially many combinations jointly. This yields a dramatic improvement in practical scalability:

Enumeration-based baselines must construct, solve, and prune one single-property game per subset.
Compositional approaches process subformulas incrementally, applying on-the-fly composition of minimized DFAs or symbolic representations, with trade-offs between early unrealizability detection and minimization overhead (Li et al., 6 Aug 2025).
For DECLARE-style conjunctive specifications, pastification and symbolic DFA construction reduce complexity to single-exponential (or lower in symbolic size) in the number of patterns (Geatti et al., 2022).

A further extension, adaptive multi-tier synthesis, computes an adaptive strategy that, dynamically at runtime, enforces the strongest currently enforceable objective and tracks environment cooperation to opportunistically achieve stricter goals. This is achieved with only quadratic overhead in the number of objectives (Giacomo et al., 29 Apr 2025).

6. Empirical Results and Practical Effectiveness

The fully symbolic approach has been implemented in MPSynth (built atop LydiaSyft), and benchmarked against enumeration-based solvers. Across benchmarks including parametric chains (chain, until, next), conflict-ridden counters (counter), and complex navigation scenarios (robotnav), MPSynth routinely outperforms enumeration by up to two orders of magnitude—reducing runtimes from, e.g., hundreds of seconds to single-digit seconds for instances with product arenas up to $10^8$ states and up to $17$ properties (Weinhuber et al., 15 Jan 2026).

Empirically, the advantages stem from:

Compact BDD-based representation of all $2^n$ property subsets.
On-the-fly fixed-point iteration exploiting monotonicity.
Avoidance of redundant computation by on-demand pruning of strictly dominated property sets in maximal synthesis.

Multi-property LTLf synthesis generalizes to scenarios with environment assumptions, unreliable or partially observable inputs, and quantitative or best-effort/permissive synthesis objectives (Aminof et al., 2023, Hagemeier et al., 2024). For LTLf under environment reachability or safety constraints, the synthesis procedure consists of reduction to suitable combinations of DFA-product arenas, reachability/safety games, and is $2$EXPTIME-complete in general.

Alternative logics, such as LTLf+ $/$ PPLTL $^+$ , utilize expressive automata and game-based techniques (Emerson-Lei and Manna-Pnueli games) to efficiently solve multi-objective finite-trace synthesis with guarantee and safety or more complex combinations, taking advantage of nested fixpoint and DAG decomposition algorithms. These approaches are, in most cases, competitive and can be implemented with symbolic automata (BDD-) representations for performance (Hausmann et al., 20 Aug 2025).

Summary Table: Algorithmic Approaches for Multi-Property LTLf Synthesis and Variants

Approach	Symbolic/Explicit	Best Complexity	Scalability Features
Symbolic Fixed-Point (Weinhuber et al., 15 Jan 2026)	Symbolic	2EXPTIME	BDD compression, monotonicity
Enumeration (baseline)	Explicit	2EXPTIME	None (infeasible for $n>8$ )
DECLARE symbolic (Geatti et al., 2022)	Symbolic	EXPTIME in $n$	Pure-past reduction, SDFAs
Adaptive tiered (Giacomo et al., 29 Apr 2025)	Explicit	Quadratic in $n$	Dynamic adaptive strategies
Compositional (Li et al., 6 Aug 2025)	Both	2EXPTIME	On-the-fly pruning, minimization

All current approaches for general LTLf multi-property synthesis are $2$EXPTIME-complete in the number and size of formulas, but symbolic and compositional techniques unlock practical performance in many scenarios of interest.