Sound and complete axiomatizations of coalgebraic language equivalence

Coalgebras provide a uniform framework to study dynamical systems, including several types of automata. In this paper, we make use of the coalgebraic view on systems to investigate, in a uniform way, under which conditions calculi that are sound and complete with respect to behavioral equivalence can be extended to a coarser coalgebraic language equivalence, which arises from a generalised powerset construction that determinises coalgebras. We show that soundness and completeness are established by proving that expressions modulo axioms of a calculus form the rational fixpoint of the given type functor. Our main result is that the rational fixpoint of the functor $FT$, where $T$ is a monad describing the branching of the systems (e.g. non-determinism, weights, probability etc.), has as a quotient the rational fixpoint of the “determinised” type functor $\bar F$, a lifting of $F$ to the category of $T$-algebras. We apply our framework to the concrete example of weighted automata, for which we present a new sound and complete calculus for weighted language equivalence. As a special case, we obtain non-deterministic automata, where we recover Rabinovich’s sound and complete calculus for language equivalence.

💡 Research Summary

**
The paper develops a uniform, category‑theoretic methodology for extending sound and complete calculi that characterize behavioral equivalence of coalgebras to a coarser notion called coalgebraic language equivalence. The authors work in the setting where systems are modeled as coalgebras for a composite functor (FT), where (F) describes the observable transition structure and (T) is a monad encoding branching effects such as nondeterminism, probabilities, or weights. By applying a generalized powerset construction, an (FT)-coalgebra can be “determinised’’ into a coalgebra for a lifted functor (\bar F) that lives in the Eilenberg‑Moore category of (T)‑algebras.

The central technical contribution is the relationship between four canonical fixpoints: the final coalgebras (\nu(FT)) and (\nu F), and the rational (i.e., finitely generated) fixpoints (\rho(FT)) and (\rho(\bar F)). The authors prove that, under mild assumptions (finite‑generated (T)‑algebras closed under kernel pairs, (T) finitary, and (F) finitary preserving weak pullbacks and admitting a lifting), the rational fixpoint of (FT) maps onto the rational fixpoint of (\bar F) via a quotient. Dually, the final coalgebra (\nu F) is a quotient of (\nu(FT)). Consequently, (FT)-behavioural equivalence (bisimilarity) implies coalgebraic language equivalence (the behaviour observed after determinisation).

The authors then connect these categorical results to equational reasoning. They show that a set of syntactic expressions together with a collection of axioms and inference rules forms the rational fixpoint (\rho(\bar F)) precisely when the calculus is sound and complete for coalgebraic language equivalence. This yields an abstract Kleene theorem: every state of a finite (FT)-coalgebra can be represented by an expression modulo the axioms.

To demonstrate the framework, the paper focuses on weighted automata. Here the branching monad (V) is the free semimodule monad over a Noetherian semiring (S) (e.g., integers or reals), and the functor is (F X = S \times X^{A}) for a finite alphabet (A). Expressions are built from variables, the zero constant, binary sum (\oplus), scalar weights, action‑prefixed weighted terms (a.(r\bullet E)), and a least‑fixpoint operator (\mu). The authors start from a known sound and complete calculus for weighted bisimilarity and augment it with three additional axioms that distribute scalar multiplication over sum, collapse nested scalars, and commute scalar multiplication with action prefixes. They prove that the extended calculus is sound and complete for weighted language equivalence.

Specialising the construction to the Boolean semiring recovers nondeterministic automata. In this case the expression language simplifies (constants become (0,1), action prefix becomes (a.E)), and only two extra axioms are needed: (a.(E_1\oplus E_2) \equiv a.E_1 \oplus a.E_2) and (a.0 \equiv 0). This yields precisely the calculus of Rabinovich for trace equivalence of finite labelled transition systems.

Compared with earlier work that used a Kleene star operator and often required an infinite axiom scheme, the present approach relies on a single (\mu) operator and a finite set of equational axioms, making the proof of completeness more modular and conceptually clearer.

In summary, the paper establishes a robust categorical foundation linking bisimilarity, determinisation, and language equivalence, and translates this foundation into concrete, usable equational calculi for weighted and nondeterministic automata. The results have immediate implications for formal verification, model checking, and quantitative reasoning about systems, providing a systematic way to reason about language‑level behaviour using algebraic expressions.