Order-Invariant MSO is Stronger than Counting MSO in the Finite

We compare the expressiveness of two extensions of monadic second-order logic (MSO) over the class of finite structures. The first, counting monadic second-order logic (CMSO), extends MSO with first-order modulo-counting quantifiers, allowing the expression of queries like ``the number of elements in the structure is even’’. The second extension allows the use of an additional binary predicate, not contained in the signature of the queried structure, that must be interpreted as an arbitrary linear order on its universe, obtaining order-invariant MSO. While it is straightforward that every CMSO formula can be translated into an equivalent order-invariant MSO formula, the converse had not yet been settled. Courcelle showed that for restricted classes of structures both order-invariant MSO and CMSO are equally expressive, but conjectured that, in general, order-invariant MSO is stronger than CMSO. We affirm this conjecture by presenting a class of structures that is order-invariantly definable in MSO but not definable in CMSO.

💡 Research Summary

The paper investigates the expressive power of two extensions of monadic second‑order logic (MSO) on finite structures: counting MSO (CMSO) and order‑invariant MSO (OI‑MSO). CMSO augments MSO with first‑order modulo‑counting quantifiers (∃^{≡k mod m}), enabling statements such as “the number of elements is even” or “the size of the universe is congruent to 3 modulo 5”. OI‑MSO, by contrast, adds a fresh binary predicate interpreted as an arbitrary linear order on the domain; a formula is considered order‑invariant if its truth value does not depend on which linear order is chosen. It is immediate that every CMSO formula can be simulated in OI‑MSO by ignoring the order, so CMSO ⊆ OI‑MSO. The open question, originally raised by Courcelle, is whether the converse inclusion holds in general. Courcelle proved equivalence on restricted classes (e.g., bounded‑tree‑width graphs) and conjectured that OI‑MSO is strictly stronger on unrestricted finite structures.

The authors settle the conjecture positively by constructing a concrete class 𝒞 of finite structures that is definable in OI‑MSO but not in CMSO. Each member Aₙ of 𝒞 is an n × n grid graph whose vertices are naturally identified with coordinate pairs (i, j). The signature of Aₙ contains only the adjacency relation; the order predicate < is external. For any linear order <_π on the vertex set, the order is required to be either “row‑major” (lexicographic by i then j) or its reverse, or “column‑major” (lexicographic by j then i) or its reverse. In other words, <_π must respect one of the two natural traversals of the grid.

Using the order, an OI‑MSO formula can express the following invariant property: “For every possible linear order, the vertex set can be partitioned into two disjoint consecutive intervals X and Y such that X consists exactly of the vertices of a set of whole rows and Y consists exactly of the vertices of a set of whole columns.” The formula quantifies over sets X and Y, uses the order to enforce consecutiveness, and asserts that the partition respects the grid’s two‑dimensional structure. Because the definition quantifies over all linear orders, it holds regardless of whether the order is row‑major or column‑major, thereby satisfying the invariance requirement. Consequently, the class 𝒞 is order‑invariantly definable.

The core of the separation argument shows that no CMSO sentence can capture 𝒞. The authors assume, for contradiction, that a CMSO sentence ψ defines 𝒞. CMSO sentences can be transformed into a normal form where the only numerical information they can test is congruence of the size of certain definable sets modulo a fixed integer m (the product of all moduli appearing in ψ). By analyzing ψ on the grid Aₙ, one can show that ψ’s truth value depends only on the residue of n modulo m. Hence ψ can distinguish at most m different grid sizes. However, the class 𝒞 contains grids of arbitrarily many sizes, including infinitely many pairwise incongruent values modulo any fixed m (e.g., all prime numbers). Therefore ψ cannot correctly accept all members of 𝒞 and reject all non‑members, yielding a contradiction. This argument establishes that 𝒞 is not CMSO‑definable.

The paper concludes that order‑invariant MSO is strictly more expressive than counting MSO on unrestricted finite structures. The result clarifies the landscape of logical extensions: while counting quantifiers add quantitative power, the ability to refer to an arbitrary linear order (even under the invariance restriction) provides a qualitatively different, stronger tool for encoding two‑dimensional and global structural information. The authors discuss implications for descriptive complexity (e.g., the relationship between OI‑MSO and complexity classes such as AC⁰