Epistemic Separability Principle (ESP)

• ESP is a foundational concept in epistemic logic programs that extends ASP's splitting theorem by enabling a modular, two-phase evaluation of bottom and top layers.
• It ensures that world views are computed by first solving a bottom module and then reducing the top layer via subjective literals, guaranteeing constraint monotonicity and stratification uniqueness.
• ESP underpins practical applications like conformant planning, supporting scalable program design and validated by frameworks such as G91 semantics and Founded Autoepistemic Equilibrium Logic.

The Epistemic Separability Principle (ESP)—more precisely, the Epistemic Splitting Property—is a foundational concept in the semantics of epistemic logic programs. It extends the celebrated splitting theorem of classical Answer Set Programming (ASP) to logic programs that incorporate subjective (epistemic) literals, enabling modular reasoning about knowledge and belief within non-monotonic frameworks. ESP characterizes those semantics under which a program, given suitable structural decomposability, can be solved in a strictly two-phase fashion: first by computing world views for a “bottom” layer, and then by evaluating a “top” layer that queries the bottom only via epistemic operators. This property underpins desirable results such as uniqueness in stratified cases, monotonicity of subjective constraints, and correctness of modular design patterns for applications such as conformant planning (Cabalar et al., 2018).

1. Formal Definition of the Epistemic Splitting Property

Let $\mathcal{A}$ be a finite set of propositional atoms, and let an epistemic logic program $\Pi$ over $\mathcal{A}$ be a finite set of rules of the form

$$a_1 \vee \dots \vee a_n \;\leftarrow\; L_1, \dots, L_m$$

where each $a_i \in \mathcal{A}$ and each $L_j$ is either an objective literal or a subjective literal ($K\ell$, $M\ell$, $\neg K\ell$, $\neg M\ell$, with $\Pi$0 an objective literal).

A semantics $\Pi$1 assigns to $\Pi$2 a (possibly empty) set of world views, with each world view $\Pi$3. The formalism introduces the crucial notion of an epistemic splitting set:

Epistemic Splitting Set: $\Pi$4 is an epistemic splitting set for $\Pi$5 if every rule $\Pi$6 is such that either (i) all atoms in $\Pi$7 are in $\Pi$8, or (ii) none of the atoms in the head or objective body of $\Pi$9 are in $\mathcal{A}$0.

Given such $\mathcal{A}$1, the program splits as $\mathcal{A}$2 (“bottom”) and $\mathcal{A}$3 (“top”), as per conditions (i) and (ii), respectively. The top is permitted to mention atoms from $\mathcal{A}$4 only inside subjective literals.

World views of the full program are built in two phases:

1. Compute all $\mathcal{A}$5-world views $\mathcal{A}$6 of $\mathcal{A}$7;
2. For each $\mathcal{A}$8, form the subjective reduct $\mathcal{A}$9 of $$a_1 \vee \dots \vee a_n \;\leftarrow\; L_1, \dots, L_m$$0 by replacing subjective literals with $$a_1 \vee \dots \vee a_n \;\leftarrow\; L_1, \dots, L_m$$1 or $$a_1 \vee \dots \vee a_n \;\leftarrow\; L_1, \dots, L_m$$2 according to satisfaction in $$a_1 \vee \dots \vee a_n \;\leftarrow\; L_1, \dots, L_m$$3. Compute all $$a_1 \vee \dots \vee a_n \;\leftarrow\; L_1, \dots, L_m$$4-world views $$a_1 \vee \dots \vee a_n \;\leftarrow\; L_1, \dots, L_m$$5 of $$a_1 \vee \dots \vee a_n \;\leftarrow\; L_1, \dots, L_m$$6.

These are then combined pointwise: $$a_1 \vee \dots \vee a_n \;\leftarrow\; L_1, \dots, L_m$$7.

ESP: $$a_1 \vee \dots \vee a_n \;\leftarrow\; L_1, \dots, L_m$$8 satisfies epistemic splitting if, for every splitting set $$a_1 \vee \dots \vee a_n \;\leftarrow\; L_1, \dots, L_m$$9 of $a_i \in \mathcal{A}$0, the $a_i \in \mathcal{A}$1-world views of $a_i \in \mathcal{A}$2 are precisely all $a_i \in \mathcal{A}$3 with $a_i \in \mathcal{A}$4 and $a_i \in \mathcal{A}$5 as above.

Mathematically,

$a_i \in \mathcal{A}$6

2. Intuitive Foundation and Relation to Classical Splitting

ESP is the epistemic extension of the classical ASP splitting theorem (Lifschitz–Turner, 1994). In standard splitting, a bottom module generates base facts, and the top builds on these facts without recursion. In the epistemic setting, the top is statically barred from deriving bottom atoms but is permitted to inspect and reason about them strictly via epistemic (subjective) queries. Consequently, the solution process is decomposed: the bottom is entirely evaluated first, and the top is then solved with only the epistemic summary (truth of $a_i \in \mathcal{A}$7/$a_i \in \mathcal{A}$8 queries) about the bottom.

This modular, two-phase procedure is the heart of ESP: every admissible world view must result from this controlled, stepwise construction, and every such constructible world view must be admitted by the semantics.

3. Principal Theorems and Semantic Characterizations

Gelfond’s 1991 Semantics and Epistemic Splitting

A central result is that Gelfond's 1991 semantics (G91) for epistemic logic programs, as well as its propositional formula extensions (G05, G11), do satisfy ESP. The proof strategy observes that the G91 world-view reduct preserves the splitting of the program; thus, the classical ASP splitting theorem applies to each reduct, and the modular construction yields precisely the set of world views.

The theorem (informally): For every program $a_i \in \mathcal{A}$9 and epistemic splitting set $L_j$0, G91 and variants satisfy ESP.

Uniqueness Under Stratification

ESP implies, by analogy with stratified classical ASP, that epistemically stratified programs have at most one world view under any semantics satisfying both ESP and supra-ASP. An assignment of layers to atoms (with proper monotonicity in the dependency graph) enforces acyclicity, thereby precluding ambiguity.

Subjective Constraint Monotonicity

Another consequence is the monotonic behavior of subjective constraints. If logic program $L_j$1 is augmented with a constraint $L_j$2 whose body is purely subjective, the set of world views can only decrease: those failing to satisfy $L_j$3 are removed, but no new world views are introduced. This is enforced by selecting the splitting set $L_j$4 and observing that $L_j$5 becomes a top-only rule.

Table: Properties of Selected Semantics

Semantics	Supra-S5	Supra-ASP	Subjective Constraint Monotonicity	ESP	Foundedness
G91	✓	✓	✓	✓
G11	✓	✓	✓
F15	✓	✓
K15	✓	✓
S17	✓	✓
C19	✓	✓	✓	✓	✓

C19 refers to Founded Autoepistemic Equilibrium Logic [Cabalar & Fandinno 2019], which adds a foundedness requirement to G91 and still enjoys ESP.

4. Applications and Major Consequences

Modular Reasoning and Program Design

ESP supports a divide-and-conquer approach for epistemic ASP: if the program exhibits the syntactic structure required for splitting, automated reasoning can proceed by first generating candidate world views for the bottom module, simplifying the top accordingly, and then composing the results. This underpins modular development, facilitates understanding, and supports scalable implementation.

Conformant Planning

A canonical application is conformant planning—generating action sequences that guarantee a goal despite nondeterministic initial conditions. Programs can be decomposed into:

• Generate: non-deterministically produce action choices, with rules like $L_j$6.
• Define: specify transition logic with objective rules.
• Test: enforce goals or constraints with subjective constraint rules like $L_j$7.

ESP ensures that each phase is well-posed: generate candidate actions, simulate state evolution, then test subjective constraints, with each world view originating from this stepwise procedure.

Agreement on Stratified Structures

For epistemically stratified (acyclic) programs, all semantics satisfying ESP yield unique, coinciding world views. This consensus is robust to the choice of epistemic semantics, provided ESP is respected.

5. Failure of ESP in Most Alternative Semantics

Recent proposals designed to address self-supported or unfounded world views in G91 typically fail to preserve ESP. Standard counterexamples illustrate this failure.

• G11 (Gelfond 2011): Modifies the reduct by preserving fulfilled $L_j$8 as objective literal $L_j$9 while unmet $K\ell$0 becomes $K\ell$1. In the counterexample programs involving rules like $K\ell$2, $K\ell$3, and subjective constraints $K\ell$4, G11 admits world views that violate modular construction—re-activating rules in the top that ESP would block.
• K15 / S17: Similar failures occur, as the manipulation of reducts for subjective literals enables constructions of world views that do not arise via splitting.
• F15: Equilibrium logic and S5-based approaches (Fariñas-Del-Cerro & Herzig 2015) admit erroneous world views by not preserving subjective constraint monotonicity.

6. Illustrative Examples

Example 1: Basic Split

$K\ell$5. Here, $K\ell$6 is a splitting set. Bottom world views: $K\ell$7. For each, the top reduces $K\ell$8 to $K\ell$9 or $M\ell$0 depending on $M\ell$1's truth in the bottom. The resulting (G91) world views are $M\ell$2 and $M\ell$3.

Example 2: Subjective Constraint Monotonicity

For $M\ell$4, world views are $M\ell$5. Adding subjective constraint $M\ell$6 does not eliminate any world view (since $M\ell$7 fails), but adding $M\ell$8 would eliminate all.

Example 3: Conformant Planning Sketch

• Initial uncertainty: $M\ell$9, $\neg K\ell$0.
• Light after toggling: $\neg K\ell$1, with a constraint $\neg K\ell$2.
• Goal as constraint: $\neg K\ell$3.
• Actions: $\neg K\ell$4. ESP supports first generating all action plans ($\neg K\ell$5), then evaluating transitions and constraints per plan ($\neg K\ell$6), yielding the final set of conformant plans.

7. Synthesis and Theoretical Significance

The Epistemic Splitting Property provides an essential tool for modular reasoning in epistemic logic programs, ensuring the program can be solved “bottom-up” and “top-down” only when the required structural separation is present. Its preservation of classical ASP intuitions in the epistemic domain distinguishes those semantics that are suitable for structured, scalable, and analyzable knowledge-intensive applications. Only Gelfond’s 1991 semantics, its propositional formula extensions, and Founded Autoepistemic Equilibrium Logic achieve ESP, while most other recent alternatives fail basic counterexamples and do not exhibit constraint monotonicity or agreement on stratified inputs. The property thus forms a crucial benchmark for the evaluation and design of epistemic ASP semantics (Cabalar et al., 2018).