Epistemic Logic of Cops and Robbers

• ELCR is a dynamic-epistemic framework for pursuit–evasion games on directed graphs that rigorously defines players’ positions, observational power, and inference under imperfect information.
• The formalism integrates a first-order static language with bespoke dynamic operators, enabling precise updates of positions and epistemic states after each move.
• ELCR achieves complete axiomatization and decidability, while laying the groundwork for extensions to multi-agent scenarios, probabilistic variants, and reactive synthesis.

Searching arXiv for the specified ELCR and related papers to ground the article in current research. Epistemic Logic of Cops and Robbers (ELCR) is a dynamic-epistemic logic for pursuit–evasion on directed graphs under imperfect information. In its canonical formulation, ELCR combines a first-order–flavored static language for graph structure and player positions, epistemic operators for knowledge of values and formulas, and a bespoke dynamic operator that updates both positions and information states after moves. The framework was introduced to make precise such notions as players’ positions, observational power, inference, and information update in Cops and Robbers / Hide and Seek, and it is accompanied by axiomatization and decidability results (Li et al., 9 Jul 2025). A closely related line of work recasts generalized multi-agent Cops and Robbers as a reactive synthesis problem using LTL, LTLt, and KLTL, thereby providing a complementary algorithmic perspective on imperfect information, memory, and communication in pursuit–evasion (Fishell et al., 14 Mar 2025).

1. Game-theoretic setting and epistemic motivation

ELCR studies a turn-based game on a finite, directed, serial graph $(D,R)$, where seriality means that every vertex has at least one outgoing edge: $\forall s\in D\;\exists t\in D:(s,t)\in R$. The core two-player setting has Cop $X$ and Robber $Y$, represented by variables $x$ and $y$. A situation is a pair of vertices indicating the current positions of $x$ and $y$. In the main setting, each round consists of a move by the cop followed by a move by the robber, and the turn order is common knowledge. A legal move takes the current player from vertex $s$ to some $t$ with $\forall s\in D\;\exists t\in D:(s,t)\in R$0 (Li et al., 9 Jul 2025).

The central departure from classical perfect-information formulations is imperfect information governed by a $\forall s\in D\;\exists t\in D:(s,t)\in R$1-sight model. Players always know their own position. A player $\forall s\in D\;\exists t\in D:(s,t)\in R$2 can see another player $\forall s\in D\;\exists t\in D:(s,t)\in R$3 iff the distance, measured in the undirected sense via the symmetric closure of $\forall s\in D\;\exists t\in D:(s,t)\in R$4, is at most $\forall s\in D\;\exists t\in D:(s,t)\in R$5. The paper defines neighborhoods inductively by

$\forall s\in D\;\exists t\in D:(s,t)\in R$6

Hence $\forall s\in D\;\exists t\in D:(s,t)\in R$7 means that $\forall s\in D\;\exists t\in D:(s,t)\in R$8 is within $\forall s\in D\;\exists t\in D:(s,t)\in R$9 undirected steps of $X$0. The $X$1-sight condition guarantees what players can know at least by direct observation, but the framework also allows them to know more by reasoning from graph structure and prior information (Li et al., 9 Jul 2025).

This informational shift changes the winning condition. Instead of taking capture alone as primitive, the ELCR paper fixes a natural number $X$2 and says that Cop $X$3 wins iff the position of robber $X$4 is known by $X$5 within $X$6 rounds. The point is not that physical capture disappears from the theory, but that knowledge of location becomes a formal objective in its own right. This distinguishes ELCR from graph-theoretic treatments in which all players see the full state and winning is defined only by co-location (Li et al., 9 Jul 2025).

The precursor synthesis formalization exhibits the same motivation from another angle. It starts from the classical graph arena $X$7 with positions for cops and robbers, discrete synchronous rounds, non-collision and no-switching constraints, and then adds an explicitly imperfect-information version in which “players see only their adjacent neighbors.” There, knowledge constraints are written in KLTL as part of a baseline specification $X$8, including

$X$9

which states that if cop and robber are not adjacent, then neither side knows the other’s position (Fishell et al., 14 Mar 2025). This is not yet ELCR proper, but it establishes the same epistemic problem space.

2. Language, vocabulary, and model structure

ELCR is built over a vocabulary $Y$0, where $Y$1 is a set of predicate symbols including a distinguished binary symbol $Y$2 for graph edges, $Y$3 is a finite non-empty set of constants naming vertices, and $Y$4 are the player-variables. Terms are $Y$5. Thus the current position of a player is treated as the value of a variable, while graph vertices may also be referred to by constant names (Li et al., 9 Jul 2025).

The language is stratified into three levels. The Boolean fragment $Y$6 contains atomic predicates and equality: $Y$7 The dynamic extension $Y$8 adds formulas of the form $Y$9, read as “after any move of player $x$0, $x$1 holds.” The full language $x$2 adds two epistemic forms: $x$3 read respectively as “player $x$4 knows the value of term $x$5” and “player $x$6 knows that $x$7” (Li et al., 9 Jul 2025).

Semantically, an ELCR model is a tuple

$x$8

Here $x$9 is a non-empty finite domain of vertices, $y$0 interprets predicate symbols and constants, $y$1 is a non-empty set of situations, and for each player $y$2, $y$3 is an equivalence relation on $y$4 encoding epistemic indistinguishability. Name completeness is imposed: for each $y$5 there exists $y$6 with $y$7. This ensures that the domain is fully named by constants (Li et al., 9 Jul 2025).

A $y$8-sight model is then a model satisfying the visibility constraint: if $y$9 and player $x$0 is within $x$1 in situation $x$2, then $x$3. In other words, whenever another player is within sight radius $x$4, that player’s position is constant across all situations that $x$5 considers possible (Li et al., 9 Jul 2025).

The truth clauses for the epistemic operators are standardly relational: $x$6 and

$x$7

The distinction between $x$8 and $x$9 is significant. ELCR makes “knowing the value of a term” primitive rather than reducing it entirely to formula knowledge; this aligns it with logics of epistemic dependence while keeping the game-theoretic reading explicit (Li et al., 9 Jul 2025).

The language also contains object-level distance formulas. Define

$y$0

and

$y$1

Then $y$2 expresses that the value of $y$3 is within sight radius $y$4 of player $y$5. A key validity is

$y$6

which internalizes the $y$7-sight assumption into the object language (Li et al., 9 Jul 2025).

3. Dynamic update and information change

The most distinctive feature of ELCR is the dynamic operator $y$8, intended to mean “after any legal move of player $y$9, $s$0 holds.” Its semantics is not a generic temporal successor clause, but a model update that simultaneously accounts for movement, visibility, and inference (Li et al., 9 Jul 2025).

For each player $s$1, ELCR defines a movement relation $s$2 on assignments: $s$3 iff $s$4 and the other player’s value is unchanged. For a set of situations $s$5, $s$6 is the set of all one-step successors reachable by a move of $s$7. The update also uses the “relevant” subset

$s$8

that is, the situations compatible with at least one player’s current epistemic state (Li et al., 9 Jul 2025).

The truth condition for $s$9 at $t$0 is: $t$1 where the updated model $t$2 depends on whether the players are in sight after the move. If they are in sight, then

$t$3

everyone learns the exact situation. If they are not in sight, then

$t$4

The updated epistemic relations are regenerated by own-position identity: $t$5 Thus, after each move, each player knows exactly her own position, while uncertainty about the other is shaped by visibility and by elimination of impossible alternatives (Li et al., 9 Jul 2025).

Two points are structurally important. First, the use of $t$6 is designed to prevent pathological forgetting; the paper proves that if a player knows the other’s position, moving does not erase that knowledge: $t$7 Second, sequential and simultaneous movement are not interdefinable in the naive way. A simultaneous operator $t$8 can be defined, but

$t$9

is not valid for $\forall s\in D\;\exists t\in D:(s,t)\in R$00, because sequential updating yields more informational refinement than a single simultaneous step (Li et al., 9 Jul 2025).

The worked example in the ELCR paper shows how this update mechanism captures reasoning beyond direct sight. In a graph with vertices $\forall s\in D\;\exists t\in D:(s,t)\in R$01–$\forall s\in D\;\exists t\in D:(s,t)\in R$02 and sight $\forall s\in D\;\exists t\in D:(s,t)\in R$03, the cop begins at $\forall s\in D\;\exists t\in D:(s,t)\in R$04 and the robber at $\forall s\in D\;\exists t\in D:(s,t)\in R$05. Initially, the cop can exclude positions visible from $\forall s\in D\;\exists t\in D:(s,t)\in R$06 and infer that the robber is at $\forall s\in D\;\exists t\in D:(s,t)\in R$07, $\forall s\in D\;\exists t\in D:(s,t)\in R$08, or $\forall s\in D\;\exists t\in D:(s,t)\in R$09. After a move from $\forall s\in D\;\exists t\in D:(s,t)\in R$10 to $\forall s\in D\;\exists t\in D:(s,t)\in R$11, one candidate situation is eliminated because it would have brought the robber into sight; after the robber’s subsequent move and a second cop move, the updated possibility set collapses to a singleton. At that point the cop knows the robber’s exact location even though their graph distance is $\forall s\in D\;\exists t\in D:(s,t)\in R$12 and they are not in one another’s sight (Li et al., 9 Jul 2025). ELCR therefore treats non-observation as informative.

4. Axiomatization, reduction, and decidability

The static fragment without dynamic operators is denoted $\forall s\in D\;\exists t\in D:(s,t)\in R$13. Its Hilbert system contains four blocks: propositional and equality axioms; axioms for game structure and sight such as seriality, “at-some-where,” and the $\forall s\in D\;\exists t\in D:(s,t)\in R$14-sight axiom $\forall s\in D\;\exists t\in D:(s,t)\in R$15; epistemic axioms and rules for $\forall s\in D\;\exists t\in D:(s,t)\in R$16 where $\forall s\in D\;\exists t\in D:(s,t)\in R$17; and interaction principles connecting value knowledge to propositional knowledge, such as

$\forall s\in D\;\exists t\in D:(s,t)\in R$18

when the agent knows the values of all arguments of a true atomic fact (Li et al., 9 Jul 2025).

The paper proves that $\forall s\in D\;\exists t\in D:(s,t)\in R$19 is sound for $\forall s\in D\;\exists t\in D:(s,t)\in R$20-sight models. It then constructs a canonical model from maximally consistent sets, establishes that distinct maximally consistent sets correspond to distinct situations, proves a Truth Lemma, and derives completeness: every consistent $\forall s\in D\;\exists t\in D:(s,t)\in R$21 is satisfiable in a $\forall s\in D\;\exists t\in D:(s,t)\in R$22-sight model. Compactness follows, and because satisfiable formulas have finite canonical models and the system is finitely axiomatized, $\forall s\in D\;\exists t\in D:(s,t)\in R$23 is decidable (Li et al., 9 Jul 2025).

For the full dynamic logic, the axiomatization proceeds by reduction. The paper gives reduction axioms $\forall s\in D\;\exists t\in D:(s,t)\in R$24–$\forall s\in D\;\exists t\in D:(s,t)\in R$25 eliminating $\forall s\in D\;\exists t\in D:(s,t)\in R$26 in favor of static formulas. At the Boolean level,

$\forall s\in D\;\exists t\in D:(s,t)\in R$27

together with the expected clauses for negation and conjunction. More specialized reductions handle knowledge after movement, including $\forall s\in D\;\exists t\in D:(s,t)\in R$28, $\forall s\in D\;\exists t\in D:(s,t)\in R$29, $\forall s\in D\;\exists t\in D:(s,t)\in R$30, and $\forall s\in D\;\exists t\in D:(s,t)\in R$31 under the various cases in which sight or prior uncertainty determine what can be inferred after a move (Li et al., 9 Jul 2025).

These reduction axioms yield a complete dynamic system $\forall s\in D\;\exists t\in D:(s,t)\in R$32. The paper proves soundness by verifying the reduction principles against the update semantics, then proves completeness by reducing every consistent dynamic theory to an equivalent static one and invoking completeness of $\forall s\in D\;\exists t\in D:(s,t)\in R$33. As a consequence, ELCR is compact and decidable (Li et al., 9 Jul 2025).

The established metatheory is exact but deliberately limited in scope. Higher-order knowledge such as $\forall s\in D\;\exists t\in D:(s,t)\in R$34 is excluded from the object language. The basic framework is two-player, turn-based, and uses fixed move order. Sight is deterministic and symmetric, and the object language does not include probabilistic or noisy observations. The paper explicitly lists possible extensions: multiple cops with differing sights $\forall s\in D\;\exists t\in D:(s,t)\in R$35, distributed knowledge, simultaneous-move axiomatization, alternative winning conditions, probabilistic variants, and limited-memory enrichments. It also states that, beyond decidability, further complexity analysis remains to be done: static model checking is $\forall s\in D\;\exists t\in D:(s,t)\in R$36-complete, whereas the dynamic case remains open (Li et al., 9 Jul 2025).

5. Relation to dynamic epistemic logic, strategy logics, and reactive synthesis

ELCR is closely related to dynamic epistemic logic (DEL), but it does not merely import a standard product update. The paper compares its built-in operator $\forall s\in D\;\exists t\in D:(s,t)\in R$37 with a DEL treatment in which moves are represented by event models whose events are edges $\forall s\in D\;\exists t\in D:(s,t)\in R$38, with preconditions saying that the moving player is at $\forall s\in D\;\exists t\in D:(s,t)\in R$39 and postconditions updating that player’s position to $\forall s\in D\;\exists t\in D:(s,t)\in R$40. The comparison shows that DEL product update can reproduce the same relevant epistemic states, but ELCR achieves this more directly because it updates only the positional and visibility structure needed for the pursuit–evasion setting (Li et al., 9 Jul 2025). A common misconception is therefore that $\forall s\in D\;\exists t\in D:(s,t)\in R$41 is just syntactic sugar for “next”; in ELCR it is a domain-specific dynamic operator with an endogenous epistemic update.

The framework also sits near earlier logics of graph games. The ELCR paper presents it as a successor to Hide and Seek logics and to relation-changing and sabotage-style graph logics, but distinguishes it by explicitly modeling partial information induced by sight. Its static component is further connected to logics of epistemic dependence, since $\forall s\in D\;\exists t\in D:(s,t)\in R$42 behaves as a “knowing the value” operator specialized to player positions (Li et al., 9 Jul 2025).

At the level of strategy reasoning, a useful comparison point is Epistemic Strategy Logic (ESL), which extends Strategy Logic with knowledge operators and strategy quantifiers over epistemic concurrent game models. ESL formulas include atomic propositions, temporal operators $\forall s\in D\;\exists t\in D:(s,t)\in R$43 and $\forall s\in D\;\exists t\in D:(s,t)\in R$44, epistemic operators $\forall s\in D\;\exists t\in D:(s,t)\in R$45, and existential strategy quantification $\forall s\in D\;\exists t\in D:(s,t)\in R$46. It supports de dicto and de re distinctions such as $\forall s\in D\;\exists t\in D:(s,t)\in R$47 versus $\forall s\in D\;\exists t\in D:(s,t)\in R$48, and its model-checking complexity is stated to be PTIME-complete in the size of the model, NON-ELEMENTARYTIME in the size of the formula, with $\forall s\in D\;\exists t\in D:(s,t)\in R$49-EXPSPACE-hardness for bounded alternation depth (Belardinelli, 2014). ESL is not ELCR, but it provides a general framework for reasoning about knowledge of strategies; this suggests one route for extending ELCR from knowledge of positions to knowledge of strategic choices.

A second adjacent line comes from reactive synthesis. The 2025 synthesis paper formalizes generalized multi-agent Cops and Robbers using LTL, LTLt, and KLTL, treating the game as a standard turn-based input/output synthesis problem rather than as a concurrent game structure. The generalized imperfect-information version states that “players see only their adjacent neighbors,” and uses KLTL to encode knowledge constraints inside a baseline specification $\forall s\in D\;\exists t\in D:(s,t)\in R$50. Variants then conjoin additional temporal goals such as safe-zone alternation or information-sharing clauses like

$\forall s\in D\;\exists t\in D:(s,t)\in R$51

and memory clauses such as

$\forall s\in D\;\exists t\in D:(s,t)\in R$52

The same paper proposes coordination synthesis with agents and a manager modeled as CSP processes, separating Public Actions and Private Actions, where Public Actions “represent the information that is shared throughout the system” and Private Actions encode more subtle local choices (Fishell et al., 14 Mar 2025). This suggests a close structural affinity between ELCR’s information update semantics and synthesis-based knowledge-aware controller construction.

6. Variants, limitations, and broader research directions

The ELCR paper is explicit that its object language excludes higher-order knowledge and that its main semantics is tailored to two players with fixed turn order. Simultaneous moves are discussed only in a sketched extension, multiple cops and robbers are proposed but not fully developed, and sight is deterministic and symmetric. These are not accidental omissions but delimit a decidable core. A second misconception is therefore that ELCR already provides a full multi-agent epistemic game logic for pursuit–evasion. What it provides, more precisely, is a first formal and decidable dynamic-epistemic core for imperfect-information Cops and Robbers (Li et al., 9 Jul 2025).

The synthesis literature points to one major direction of expansion. The reactive-synthesis formulation introduces finite and infinite grids, multiple cops and robbers, safe zones, wall memory, information sharing, and LTL/LTLt specifications over graph movement. It also emphasizes state explosion: direct propositional encoding of an $\forall s\in D\;\exists t\in D:(s,t)\in R$53 grid with $\forall s\in D\;\exists t\in D:(s,t)\in R$54 cops and $\forall s\in D\;\exists t\in D:(s,t)\in R$55 robbers yields $\forall s\in D\;\exists t\in D:(s,t)\in R$56 states, with the example of a $\forall s\in D\;\exists t\in D:(s,t)\in R$57 grid with $\forall s\in D\;\exists t\in D:(s,t)\in R$58 cops and $\forall s\in D\;\exists t\in D:(s,t)\in R$59 robber producing $\forall s\in D\;\exists t\in D:(s,t)\in R$60 states, motivating LTLt and Boolean abstraction (Fishell et al., 14 Mar 2025). A plausible implication is that any large-scale ELCR for synthesis or verification will need similarly symbolic encodings of positions, observations, and epistemic indistinguishability.

Another direction is asymptotic and structural. In large random graphs, a separate line of work studies winning possibilities through first-order definability and zero–one laws. There the underlying game is classical and perfect-information: players are assumed to have perfect information, and several winning conditions are analyzed through extension-axiom-type first-order formulas. The paper proves, for example, that in $\forall s\in D\;\exists t\in D:(s,t)\in R$61 with constant $\forall s\in D\;\exists t\in D:(s,t)\in R$62, a robber almost always has a winning strategy against any fixed finite number of cops, while a pair of tandem-cops almost always has a winning strategy; it also analyzes sparse regimes and threshold behavior for certain extension statements (Chakraborty et al., 27 Nov 2025). These results are not epistemic in themselves, but they identify graph-structural conditions that would remain relevant if partial observation were added. This suggests a broader ELCR research program in which structural randomness and epistemic uncertainty are studied together rather than separately.

Within the ELCR paper itself, the open directions are more immediate: distributed knowledge among multiple cops, simultaneous-move logics, probabilistic variants, higher-order knowledge, and bounded or limited memory (Li et al., 9 Jul 2025). In combination with ESL-style knowledge of strategies and synthesis-style KLTL specifications, these indicate that ELCR occupies a middle position between domain-specific dynamic epistemic logic and more general strategic reasoning formalisms. Its distinctive contribution is to make movement, observation, and inference cohere in one update semantics tailored to pursuit–evasion, thereby turning “reasoning under uncertainty in Cops and Robbers” into a formally axiomatized and decidable subject (Li et al., 9 Jul 2025).