Stratified Abductive Logic Programs

• Stratified abductive logic programs are structured frameworks that integrate abduction, integrity constraints, and causal intervention by enforcing acyclic dependencies among predicates.
• They guarantee a unique stable-model semantics, and quantify explanation simplicity via a degree of arbitrariness, ensuring constrained, justified conclusions.
• The formalism bridges logic programming with Pearl-style causal models, facilitating rigorous analysis of interventions and decision problems with clear complexity bounds.

A stratified abductive logic program is a structured framework that extends classic logic programming to support abductive reasoning under stratification constraints. Stratification imposes an acyclic dependency among predicates, guaranteeing unique stable-model semantics and enabling principled reasoning about explanations, their complexity, and their causal interpretation. This formalism is central in declarative knowledge representation, enabling systematic modeling of abduction, integrity constraints, and interventions.

1. Formal Structure and Stratification

Let $\mathcal{P} = (P, \mathfrak{A}, IC)$ denote an abductive logic program (Rückschloß et al., 7 Jul 2025). Here $P$ is a finite set of definite logic programming clauses over a propositional alphabet $\mathfrak{P}$, $\mathfrak{A} \subseteq \mathfrak{P}$ designates a finite set of abducible predicate symbols (none defined by $P$), and $IC$ is a finite set of integrity constraints in the form $\bot \leftarrow b_1 \wedge \cdots \wedge b_n$. A program $P$ is stratified if its dependency graph — with edges induced positively ($p\xrightarrow{+}q$ when $p$ appears positively in the body of a rule for $P$0) or negatively ($P$1 for negative occurrences) — is acyclic with respect to negative edges, i.e., there is no cycle involving a negative dependency. This property ensures that for every instantiation of the abducibles and facts, there exists a unique stable model coinciding with the well-founded model (Caroprese et al., 2014).

2. Abductive Explanations: Structure and Simplicity

Given an observation $P$2 consisting of ground atoms outside the set of abducibles, an abductive explanation is a pair $P$3 where $P$4, $P$5 are disjoint, finite sets of abducible atoms with $P$6. The updated program $P$7 must entail $P$8 in its unique stable model while satisfying all $P$9 (Caroprese et al., 2014). Explanations are classified by the degree of arbitrariness, which quantifies to what extent the constants in $\mathfrak{P}$0 could be consistently and independently replaced without affecting the explanatory status with respect to $\mathfrak{P}$1 and $\mathfrak{P}$2.

Let $\mathfrak{P}$3 appear in $\mathfrak{P}$4. For every non-empty subset of its occurrences, a replacement function $\mathfrak{P}$5 specifies a uniform renaming. The degree of arbitrariness $\mathfrak{P}$6 is the largest $\mathfrak{P}$7 for which there exist $\mathfrak{P}$8 pairwise-independent such replacement functions, each yielding an explanation. An explanation is called constrained (simplicity-optimal) if $\mathfrak{P}$9; otherwise, it is arbitrary. Constrained explanations represent the most semantically justified answers to abduction queries (Caroprese et al., 2014).

3. Stable-Model and Causal Semantics

The stable model of $\mathfrak{A} \subseteq \mathfrak{P}$0 (or $\mathfrak{A} \subseteq \mathfrak{P}$1) is defined via the absence of unfounded subsets in a candidate structure $\mathfrak{A} \subseteq \mathfrak{P}$2. For stratified programs, this model is unique (Rückschloß et al., 7 Jul 2025). A foundational result is that stratified abductive logic programs can be canonically translated to Pearl-style structural causal models. Each clause $\mathfrak{A} \subseteq \mathfrak{P}$3 is reified as a deterministic equation $\mathfrak{A} \subseteq \mathfrak{P}$4, with $\mathfrak{A} \subseteq \mathfrak{P}$5 acting as an exogenous (abducible) variable. The resulting system $\mathfrak{A} \subseteq \mathfrak{P}$6 provides a causal interpretation: abducible selection corresponds to external interventions or unobserved causes.

The process of forced intervention ($\mathfrak{A} \subseteq \mathfrak{P}$7) is realized by replacing clauses for $\mathfrak{A} \subseteq \mathfrak{P}$8 in $\mathfrak{A} \subseteq \mathfrak{P}$9 with facts setting $P$0 to $P$1. The stable model of the modified program corresponds exactly to the solution of the intervened structural equations, aligning the logic program’s semantics with that of causal models (Rückschloß et al., 7 Jul 2025).

4. Complexity and Algorithmic Properties

Let $P$2 be a fixed stratified background program without abducible heads, $P$3 a set of ground abducibles, and $P$4 the observation. The paper (Caroprese et al., 2014) analyzes two key decision problems:

1. Given $P$5, decide whether $P$6 is a constrained explanation of $P$7;
2. Decide whether there exists any constrained explanation of $P$8.

For arbitrary stratified programs, the first problem is in coNP and the second in $P$9. These upper bounds are tight: each source of complexity — negation, recursion, integrity constraints — suffices for completeness at these levels (Theorems 4-6 in (Caroprese et al., 2014)). In the absence of negation, recursion, and integrity constraints (i.e., non-recursive Horn), both problems are decidable in polynomial time (Theorem 7).

The procedures can be outlined as:

• For verifying constrainedness: compute the unique stable model $IC$0 of $IC$1, check $IC$2 and $IC$3, then, for each constant and occurrence-subset, systematically test all possible independent replacements, checking whether arbitrariness is present.
• For finding a constrained explanation: nondeterministically guess $IC$4 up to the bounded size dictated by $IC$5 (as no fresh constants appear in constrained explanations for stratified $IC$6), test constrainedness via the above method (Caroprese et al., 2014).

5. Philosophical Principles: Causation and Explanation

Stratified abductive logic programs under stable model semantics conform to several foundational principles in the philosophy of causation (Rückschloß et al., 7 Jul 2025). These include:

• Natural necessity (Aquinas): If the cause(s) occur, so must the effect.
• Sufficient causation (Leibniz): Every effect must have a sufficient cause in the abducibles or program rules.
• Irrelevance of unobserved effects (Williamson): Modifying only descendants in the dependency graph cannot alter any assignment to non-descendants.

Formally, stratification ensures the “irrelevance” property: interventions restricted to a stratum’s descendants cannot falsify any fixed pre-assignment to the stratum itself. Main theorems (Theorems 3.8, 3.13, 3.14 in (Rückschloß et al., 7 Jul 2025)) establish that for stratified programs, these properties are satisfied by construction; stable models correspond to causally founded worlds, and interventions on sets of variables preserve non-descendants’ values in stable models.

6. Illustrative Examples

The security-breach scenario and “sprinkler” causal system exemplify arbitrariness, explanation simplicity, and causal interpretation (Caroprese et al., 2014, Rückschloß et al., 7 Jul 2025).

Security-breach Example: For an observation $IC$7, an explanation using a fixed staff member (e.g., tom) is constrained, while explanations parameterized by a variable $IC$8 (free constant) are arbitrary; substituting $IC$9 by a fresh constant does not alter the explanatory power if not otherwise justified.

Sprinkler Example: Rules such as $\bot \leftarrow b_1 \wedge \cdots \wedge b_n$0, $\bot \leftarrow b_1 \wedge \cdots \wedge b_n$1, $\bot \leftarrow b_1 \wedge \cdots \wedge b_n$2, $\bot \leftarrow b_1 \wedge \cdots \wedge b_n$3, $\bot \leftarrow b_1 \wedge \cdots \wedge b_n$4 with abducible $\bot \leftarrow b_1 \wedge \cdots \wedge b_n$5 admit exactly two supported stable models, each with a unique causal account via structural equations. Forcing the sprinkler $\bot \leftarrow b_1 \wedge \cdots \wedge b_n$6 to be on via intervention eliminates dependence on the abnormality $\bot \leftarrow b_1 \wedge \cdots \wedge b_n$7, and the resulting model’s predictions agree with the corresponding modified causal system (Rückschloß et al., 7 Jul 2025).

7. Significance and Implications

Stratified abductive logic programs provide a declarative, non-cyclic platform for abduction with principled complexity, naturally supporting model-based explanation, simplicity via arbitrariness minimization, and causal interpretation consistent with established philosophical accounts. Stratification’s enforcement of acyclicity in negative dependencies guarantees uniqueness of stable models and enables strong algorithmic and conceptual properties. The translation to causal structural models and precise complexity-theoretic characterizations render stratified abduction a robust formalism for inference, explanation, and intervention in knowledge representation and causal reasoning (Caroprese et al., 2014, Rückschloß et al., 7 Jul 2025).