Graded Modal Logic (GML) Overview

• Graded Modal Logic is the extension of classical modal logic that uses counting modalities to specify cardinality constraints on accessible worlds.
• It exhibits diverse complexity profiles across frame conditions, with satisfiability ranging from PSPACE and NP to NEXPTIME-complete and even undecidable cases.
• Its framework bridges modal theory with automata and graph neural networks, leveraging graded bisimulation and neighbourhood semantics for practical applications.

Searching arXiv for graded modal logic and closely related papers. {"query":"all:graded modal logic satisfiability transitive Euclidean expressiveness bisimulation neighborhood graph neural networks", "max_results": 10} {"query":"ti:\"A Note on the Complexity of the Satisfiability Problem for Graded Modal Logics\" OR all:\"graded modal logic\" ", "max_results": 10} Graded modal logic (GML) is the extension of ordinary propositional modal logic by modal operators indexed with cardinality constraints, so that modalities can state not merely that some accessible world satisfies a formula, but that at least or at most a specified number of accessible worlds do so. Under standard possible-worlds semantics, this yields readings such as “it is true at no fewer than $15$ accessible worlds that $\dots$” and “it is true at no more than $2$ accessible worlds that $\dots$.” The subject has developed along several axes: the classical satisfiability theory over standard frame classes, bisimulation and first-order characterizations, alternative semantics such as neighbourhood semantics, expressiveness results on trees and forests, and recent interactions with automata theory and graph representation learning (0905.3108).

1. Formal language and Kripke semantics

In the standard presentation, the language $\mathcal{GM}$ is the smallest set containing all propositional letters, closed under the Boolean connectives, and such that whenever $\phi\in\mathcal{GM}$, the formulas $\Diamond_{\leq C}\phi$ and $\Diamond_{\geq C}\phi$ are again in $\mathcal{GM}$. A notable complexity-theoretic feature is that numerical subscripts are taken in binary coding, so the size of $\Diamond_{\leq C}$ or $\dots$0 is about $\dots$1, not $\dots$2 (0905.3108).

Over a Kripke structure $\dots$3, the graded modalities are interpreted as counting operators: $\dots$4

$\dots$5

Ordinary modal operators are recovered by abbreviation: $\dots$6 Accordingly, $\dots$7 expresses that $\dots$8 holds at at least $\dots$9 accessible worlds, while $2$0 expresses that $2$1 holds at at most $2$2 accessible worlds (0905.3108).

Later work uses equivalent notational variants. In one standard notation, $2$3 means that a state has at least $2$4 successors satisfying $2$5,

$2$6

while in graded multimodal logic one writes

$2$7

thereby indexing accessibility by relation labels as well as counts (Chen et al., 2021).

2. Frame conditions and the satisfiability landscape

A central line of work studies satisfiability over classes of frames determined by combinations of reflexivity, seriality, symmetry, transitivity, and the Euclidean property. These conditions are given by the standard first-order clauses

$2$8

$2$9

and symmetry together with transitivity implies Euclideanity (0905.3108).

For standard GML over all combinations of these five properties, satisfiability admits a complete trichotomy:

Frame assumptions	Complexity	Characterization
No $\dots$0, no $\dots$1	PSPACE-complete	Any subset of $\dots$2
$\dots$3, or $\dots$4	NP-complete	Euclidean, or symmetric-transitive
$\dots$5, $\dots$6, $\dots$7	NEXPTIME-complete	Transitive without symmetry/Euclideanity

These cases are exhaustive for all $\dots$8, and the transitive case remains NEXPTIME-hard even when all numerical subscripts are at most $\dots$9 (0905.3108).

The conceptual reason for the transitive jump is that graded modal logic lacks the tree-model property on transitive frames. The canonical witness used in the analysis is

$\mathcal{GM}$0

This formula is satisfiable over a transitive frame, but not over any tree-shaped transitive frame, because the conjunct $\mathcal{GM}$1 forces branch merging. The corresponding upper bound is obtained by normalizing formulas into a Scott-like form and proving an exponentially bounded transitive model property (0905.3108).

The interaction of counting with converse modalities changes the picture further. For graded modal languages with forward and backward graded modalities, local and global satisfiability over Euclidean and serial Euclidean frames become NExpTime-complete, which is harder than the corresponding language without graded modalities or without converse modalities. By contrast, over transitive Euclidean frames the same problems are NP-complete. An intermediate language that allows graded forward modalities but only basic converse modalities has the finite model property and decidable local and global satisfiability over transitive, reflexive-transitive, and serial-transitive frames, whereas the full two-way graded language on transitive frames is undecidable (Bednarczyk et al., 2018).

3. Bisimulation, first-order characterization, and neighbourhood semantics

Ordinary modal logic is invariant under ordinary bisimulation; GML requires a stronger matching condition that preserves multiplicities. In graded bisimulation, the forth and back clauses quantify over finite sets of pairwise distinct successors rather than individual successors. This leads to graded bisimulation games, $\mathcal{GM}$2-round and $\mathcal{GM}$3-bounded approximants, and characteristic formulas for the corresponding finite fragments. The resulting van Benthem-style characterization states that graded modal logic is exactly the fragment of first-order logic invariant under graded or counting bisimulation, both over all pointed Kripke structures and over finite pointed Kripke structures (Otto, 2019).

A complementary development recasts GML in neighbourhood terms. A graded neighbourhood frame is a monotonic neighbourhood frame

$\mathcal{GM}$4

such that for every $\mathcal{GM}$5 there exists $\mathcal{GM}$6 with

$\mathcal{GM}$7

Kripke frames embed by sending $\mathcal{GM}$8 to $\mathcal{GM}$9 with $\phi\in\mathcal{GM}$0, and this embedding preserves truth of all graded modal formulas. On this basis, the standard graded logic $\phi\in\mathcal{GM}$1 is sound and strongly complete with respect to the class of graded neighbourhood frames. Model-theoretically, that class is first-order definable in a suitable two-sorted language, but it is not modally definable because it is not closed under bounded morphic images (Chen et al., 2021).

The same neighbourhood perspective yields a reformulation of graded bisimulation for Kripke models by specializing monotonic bisimulation to neighbourhoods generated by successor sets. This new definition preserves graded modal equivalence and is shown to be equivalent to de Rijke’s graded tuple bisimulation. The significance is methodological: neighbourhood semantics imports monotonic-modal tools into the graded setting while remaining extensionally faithful to standard Kripke semantics (Chen et al., 2021).

4. Expressiveness on forests, trees, and fixpoint extensions

On finite forests, GML occupies a precise point in an expressiveness hierarchy involving modal logics with submodel-composition operators: $\phi\in\mathcal{GM}$2 Here $\phi\in\mathcal{GM}$3 uses an ambient-style composition operator that splits the forest at the children of the current world, whereas $\phi\in\mathcal{GM}$4 uses a separation-style conjunction based on arbitrary splitting of the edge relation. The equivalence $\phi\in\mathcal{GM}$5 is proved by translating composition formulas into Presburger arithmetic constraints on the numbers of children satisfying selected subformulas, eliminating quantifiers, and translating the resulting arithmetic constraints back into graded formulas. By contrast, $\phi\in\mathcal{GM}$6 is strictly weaker than GML, even though it is strictly stronger than ordinary modal logic (Bednarczyk et al., 2020).

The same work gives sharp complexity consequences on finite forests and trees. Satisfiability for $\phi\in\mathcal{GM}$7 is AExp-complete, while satisfiability for $\phi\in\mathcal{GM}$8 is tower-complete. This is one of the clearest instances in which expressiveness and satisfiability complexity diverge: the less expressive system can be computationally harder (Bednarczyk et al., 2020).

Fixpoint extensions preserve the centrality of counting. In the graded modal $\phi\in\mathcal{GM}$9-calculus, the modalities

$\Diamond_{\leq C}\phi$0

express, respectively, that at least $\Diamond_{\leq C}\phi$1 successors satisfy $\Diamond_{\leq C}\phi$2 and that all but at most $\Diamond_{\leq C}\phi$3 successors satisfy $\Diamond_{\leq C}\phi$4. A notable graded formula,

$\Diamond_{\leq C}\phi$5

defines the class of trees with outdegree bounded by $\Diamond_{\leq C}\phi$6. This makes bounded branching directly expressible in the logic and drives the complexity of separability and definability problems. When separators may use graded modalities, both $\Diamond_{\leq C}\phi$7-definability and $\Diamond_{\leq C}\phi$8-separability of $\Diamond_{\leq C}\phi$9-formulae are ExpTime-complete. When the input formulae are graded but separators are restricted to ordinary modal logic, definability remains ExpTime-complete but separability rises to TwoExpTime-complete (Jung et al., 29 Sep 2025).

5. Automata, local types, and graph-learning interfaces

Graded multimodal logic has a close automata-theoretic counterpart. Counting multichannel message passing automata (CMMPAs) operate over pointed Kripke models by propagating multisets of neighbour states separately along each modality channel. Their behavior is coordinated by graded multimodal types, which canonically summarize the local structure of a pointed model up to bounded modal depth and counting thresholds. The main equivalence states that a class of finite pointed models is definable by a countable disjunction of graded multimodal formulas iff it is recognizable by a CMMPA; the equivalence also preserves recursive enumerability. The same paper gives a first-order fixed-point account of Weisfeiler–Leman refinement using Härtig’s quantifier and greatest fixed points (Ahvonen et al., 2024).

These logical tools connect directly to graph representation learning. One exact correspondence identifies bounded aggregate-combine GNNs with graded modal logic: $\Diamond_{\geq C}\phi$0 message-passing layers match modal depth $\Diamond_{\geq C}\phi$1, and $\Diamond_{\geq C}\phi$2-bounded aggregation matches counting rank $\Diamond_{\geq C}\phi$3. In this sense, bounded local message passing is precisely GML in neural form (Grau et al., 12 May 2025). A related expressiveness picture for recurrent models states that fixed-depth GNNs correspond to GML, while, over undirected graphs, converging RGNNs and graded-bisimulation-invariant halting RGNNs have the same expressive power; combined with earlier results, this yields $\Diamond_{\geq C}\phi$4GML as the logical benchmark for the MSO-definable fragment of converging RGNNs (Bollen et al., 28 Apr 2026).

Transformer-style graph architectures broaden the modal vocabulary by adding global modalities. Relative to first-order definable vertex properties, real-valued GPS-networks have the same expressive power as $\Diamond_{\geq C}\phi$5, where $\Diamond_{\geq C}\phi$6 is the non-counting global modality. With floating-point arithmetic, the correct logic is $\Diamond_{\geq C}\phi$7, where $\Diamond_{\geq C}\phi$8 means that at least $\Diamond_{\geq C}\phi$9 vertices in the whole graph satisfy $\mathcal{GM}$0; the float-based characterization is absolute rather than FO-relative. Pure graph transformers correspond similarly to propositional fragments with global or counting-global modalities (Ahvonen et al., 1 Aug 2025).

6. Related uses of “graded” and terminological boundaries

A persistent source of ambiguity is that several neighbouring literatures use the adjective “graded” for modal systems that are not standard successor-counting GML. In lattice-based graded logic, modalities $\mathcal{GM}$1 are indexed by symbolic certainty grades from a distributive lattice, with $\mathcal{GM}$2 read as “$\mathcal{GM}$3 is known with at least grade $\mathcal{GM}$4.” The grading is qualitative and order-theoretic, not cardinality-based (Chatalic et al., 2013). In many-valued preference logics, the family $\mathcal{GM}$5 arises from level cuts of a fuzzy preference relation over a finite residuated lattice, and the semantics tracks preference degrees rather than numbers of successors (Vidal et al., 2019).

Other systems are closer in spirit than in syntax. Paraconsistent Gödel modal logic is presented as a logic for graded, incomplete, and inconsistent information using two-valued coordinates of positive and negative support, but it does not introduce standard counting modalities of the form “at least $\mathcal{GM}$6 successors” (Bílková et al., 2022). Graded concurrent PDL generalizes concurrent dynamic logic by evaluating both formulas and state-to-set reachability in a finite Łukasiewicz chain, again using graded truth degrees rather than successor counts (Lin, 16 Jan 2025).

A different abstraction appears in graded monads and graded logics for coalgebraic semantics. There, “graded” refers to depth-indexed semantics $\mathcal{GM}$7 and to the extraction of characteristic modal logics for trace-like behaviors; the framework subsumes many equivalences from the linear-time–branching-time spectrum and extends smoothly to Eilenberg–Moore coalgebras and quantale-valued behavioral distances (Dorsch et al., 2018). Resource-sensitive graded modalities form another separate line: mixed linear and graded logics use modalities such as $\mathcal{GM}$8 or decompositions $\mathcal{GM}$9, with grades drawn from a preorder-enriched semiring and interpreted as usage bounds rather than branching multiplicities (Vollmer et al., 2024). Dependent type systems combining grades with adjoint logic continue this resource-oriented tradition, where grades compose by semiring operations and modal operators transport terms between modes (Hanukaev et al., 2023).

Standard GML, in the narrow and most widely used sense, therefore remains the cardinality-indexed extension of modal logic interpreted over accessible worlds by counting how many successors satisfy a formula. Its distinctive theory is the one organized around successor multiplicities, graded bisimulation, frame-class complexity, and the expressive correspondences that link modal counting to automata, trees, and modern graph-learning architectures (0905.3108).