Facets of Propositional Abduction

Updated 6 July 2026

The paper defines facets as hypothesis variables that appear in some but not all subset‐minimal explanations, striking a balance between necessity and dispensability.
It employs formal abduction settings over Boolean constraint languages, using subset-minimality and explanation distance to capture diversity in explanation spaces.
The analysis links facets to computational complexity, showing tractable cases in selected fragments and highlighting harder scenarios when reasoning with diverse minimal explanations.

Facets of propositional abduction are variable-level descriptors of explanatory variability. In the positive propositional setting, an explanation is a subset of hypotheses whose addition to a knowledge base both preserves satisfiability and entails the manifestations; a facet is a hypothesis variable that occurs in some subset-minimal explanation but not in all of them. This places facets strictly between existence-style reasoning and full counting or enumeration: they are intended to expose heterogeneity among explanations without requiring complete output of the explanation space (Schmidt et al., 20 Jul 2025).

The faceted setting is formulated for positive propositional abduction over a Boolean constraint language $\Gamma$ . An instance of $ABD(\Gamma)$ is a tuple

$I=(KB,H,M)$

where $KB$ is the knowledge base, $H \subseteq \var(KB)$ is the set of hypotheses, and $M \subseteq \var(KB)$ is the set of manifestations. A positive explanation is a set

$E \subseteq H$

such that $KB \land E$ is satisfiable and $KB \land E \models M$ . An explanation is subset-minimal if no proper subset of it is again an explanation (Schmidt et al., 20 Jul 2025).

Within propositional abduction more broadly, this is a specialization of the standard logic-based pattern in which one seeks a hypothesis $E$ or $ABD(\Gamma)$ 0 such that the background theory together with that hypothesis is satisfiable and entails the observation. Classical formulations often allow explanations to be conjunctions of literals over a set of abducibles, as in $ABD(\Gamma)$ 1 or $ABD(\Gamma)$ 2, whereas the faceted framework fixes attention on positive explanations $ABD(\Gamma)$ 3 and on subset-minimality as the operative explanation criterion (Zanuttini, 2011).

A variable $ABD(\Gamma)$ 4 is relevant if it belongs to some subset-minimal explanation, and necessary if it belongs to all subset-minimal explanations. A variable is a facet exactly when it is relevant but not necessary: $ABD(\Gamma)$ 5 Equivalently,

$ABD(\Gamma)$ 6

The paper also clarifies that “dispensable” is used in the ordinary sense of “not necessary”: a dispensable variable can be omitted from at least one minimal explanation (Schmidt et al., 20 Jul 2025).

Facethood isolates a middle explanatory status. Relevance asks whether a variable can occur in some explanation; necessity asks whether it must occur in every explanation; a facet is a variable whose status is contingent across the minimal explanation space. In this sense, facets encode explanatory flexibility rather than mere explanatory availability (Schmidt et al., 20 Jul 2025).

This distinction eliminates two common conflations. First, a facet is not simply a relevant variable: a necessary variable is relevant, but it is not a facet. Second, a facet is not simply a dispensable variable: dispensability alone does not suffice, because the variable must still occur in at least one subset-minimal explanation. The facet concept therefore factors explanationhood into three mutually informative regimes: absent from all minimal explanations, present in all minimal explanations, and present in some but not all minimal explanations (Schmidt et al., 20 Jul 2025).

The emphasis on subset-minimal explanations connects facets to a long-standing preference discipline in logic-based abduction. In one standard formulation, a “best explanation” is explicitly defined as a subset-minimal explanation (Zanuttini, 2011). In only-knowing-based modal abduction, subset-minimality also appears as a selection method, and under suitable conditions it coincides with preferential consequence; cardinality-minimality and prioritization-based selection are treated as alternative criteria (Molick et al., 7 Jan 2026). This suggests that facethood is not an absolute notion detached from semantics: it is relative to whichever explanation-selection relation defines the admissible explanation space.

3. Distance, diversity, and heterogeneity of explanation spaces

The faceted framework also introduces an explicit distance between explanations. For $ABD(\Gamma)$ 7,

$ABD(\Gamma)$ 8

The maximum possible distance is $ABD(\Gamma)$ 9. Two explanations are called $I=(KB,H,M)$ 0-diverse if

$I=(KB,H,M)$ 1

The associated decision problem $I=(KB,H,M)$ 2 asks whether an instance has two $I=(KB,H,M)$ 3-diverse explanations (Schmidt et al., 20 Jul 2025).

The key structural connection is immediate: if $I=(KB,H,M)$ 4 and $I=(KB,H,M)$ 5 are subset-minimal explanations, then every variable in $I=(KB,H,M)$ 6 is a facet. Hence facets are the variable-level support of explanation diversity. Diversity measures how far explanations can separate globally; facets identify the coordinates on which that separation is realized (Schmidt et al., 20 Jul 2025).

This division of labor is significant because it avoids treating all non-uniqueness alike. Two instances may both admit multiple explanations, yet in one case the differences may be confined to a small set of facets, while in another the symmetric difference may be large. Facets therefore support a finer analysis of heterogeneity than a bare multiplicity statement, while remaining less demanding than counting or enumerating all explanations. That positioning is explicit in the motivation for the framework: counting and enumeration are described as computationally highly challenging, and facets are introduced to reason “between decisions and counting” (Schmidt et al., 20 Jul 2025).

4. Complexity of facet reasoning in Post’s framework

Facet reasoning is analyzed systematically under restricted Boolean constraint languages using Post’s lattice and co-clone terminology. The classification is expressed over fragments including CNF, Horn, dualHorn, EN, EP, affine, 2-CNF, 2-affine, implicative, and IHS-B variants. For $I=(KB,H,M)$ 7, the paper gives an almost complete characterization, with only two open cases remaining: affine equations of even length without unit clauses, and the same with unit clauses (Schmidt et al., 20 Jul 2025).

A central upper-bound lemma states: $I=(KB,H,M)$ 8 This is particularly relevant for Schaefer-type fragments whose satisfiability problem is polynomial-time. The paper also proves several explicit tractability results and shows that the same classification applies to the classical relevance problem as a corollary (Schmidt et al., 20 Jul 2025).

Fragment or problem	Result	Note
$I=(KB,H,M)$ 9	in $KB$ 0	Implicative fragment
$KB$ 1	in $KB$ 2	Via reduction to a unit-clause-free variant
$KB$ 3	in $KB$ 4	Uses equivalence classes and clusters
$KB$ 5	in $KB$ 6	Essentially negative fragment
$KB$ 7	in $KB$ 8	Diversity tractable
$KB$ 9	in $H \subseteq \var(KB)$0	Diversity tractable

The lower-bound picture shows that facet reasoning can be strictly harder than plain abduction. When equality is available,

$H \subseteq \var(KB)$1

A further simulation result shows

$H \subseteq \var(KB)$2

which is used to derive hardness jumps. The overall classification includes NP-hard, coNP-hard, and $H \subseteq \var(KB)$3-hard cases. The paper’s explicit conclusion is that some fragments are “not much harder” than abduction, whereas others become significantly harder (Schmidt et al., 20 Jul 2025).

For diversity, hardness can appear even in small implicative settings: $H \subseteq \var(KB)$4 Thus diversity is often harder than facet checking, despite the tight conceptual link between the two notions (Schmidt et al., 20 Jul 2025).

5. Position within the general complexity and algorithmic theory of propositional abduction

Facets were introduced against a mature complexity background. General propositional abduction has long been known to be hard: in Post’s framework, deciding whether an explanation exists is $H \subseteq \var(KB)$5-complete in general, with refined classifications into $H \subseteq \var(KB)$6, NP, coNP, $H \subseteq \var(KB)$7, and $H \subseteq \var(KB)$8 depending on the Boolean basis and the manifestation type (Creignou et al., 2010). Logic-based abduction also admits tractable islands; a projection-based algorithm yields polynomial classes for affine knowledge bases and for several DNF-based fragments, with projection identified as the main algorithmic bottleneck (Zanuttini, 2011).

The algorithmic side is correspondingly diverse. Structural parameterization by strong Horn or Krom backdoor sets yields fixed-parameter tractable transformations from abduction to SAT, with CNF encodings of size $H \subseteq \var(KB)$9 for backdoor size $M \subseteq \var(KB)$0 (Pfandler et al., 2013). Minimum-cost propositional abduction has also been attacked by implicit hitting-set methods: the Hyper algorithm integrates the background theory and manifestations directly into the hitting-set computation and is shown to reduce SAT-oracle calls by an exponential factor in the worst case relative to earlier AbHS-style methods (Ignatiev et al., 2016).

Recent fine-grained analysis adds another layer. With $M \subseteq \var(KB)$1 the number of variables, brute-force bounds of $M \subseteq \var(KB)$2 for $M \subseteq \var(KB)$3 and $M \subseteq \var(KB)$4 for $M \subseteq \var(KB)$5 can sometimes be improved: sparse model enumeration yields $M \subseteq \var(KB)$6 algorithms for certain fragments, and $M \subseteq \var(KB)$7 can be solved in $M \subseteq \var(KB)$8 time for any constraint language $M \subseteq \var(KB)$9 (Lagerkvist et al., 15 May 2025). Against that background, facets occupy a deliberately intermediate position: they are more informative than mere existence, less demanding than counting or full enumeration, and closely aligned with the variability structure of minimal explanations (Schmidt et al., 20 Jul 2025).

The facet notion is defined in a specifically propositional, positive, subset-minimal setting. Other abductive formalisms organize explanations differently. In only-knowing modal logic, abduction is represented by a derived modality $E \subseteq H$ 0, with explanations constrained by only-known background content; a preferential extension $E \subseteq H$ 1 introduces transitive and connected plausibility orderings and supports preferential, subset-minimal, cardinality-minimal, and prioritization-based explanation selection (Molick et al., 7 Jan 2026). In morphological abduction, explanations are induced by erosion operators that isolate the most central surviving part of $E \subseteq H$ 2 or of $E \subseteq H$ 3 consistent with $E \subseteq H$ 4, yielding semantic minimality through a centrality preorder rather than through subset inclusion alone (Bloch et al., 2018).

Argumentation-based abduction reorganizes the search space even more radically. In abductive argumentation frameworks $E \subseteq H$ 5, hypotheses are entire alternative argumentation frameworks $E \subseteq H$ 6, not propositional subsets; skeptical and credulous explanation problems are characterized by sound and complete dialogue procedures, and the framework instantiates abductive logic programming under partial stable semantics (Booth et al., 2014). Intuitionistic theorem-synthesis abduction shifts from model-theoretic explanation to weakest-premise synthesis: one searches for assumptions $E \subseteq H$ 7 such that $E \subseteq H$ 8 is provable in intuitionistic logic, with minimality ordered by implication rather than set inclusion (Tarau, 2022).

These contrasts matter for the interpretation of facets. As introduced, facets are defined over variables in positive subset-minimal explanations. This suggests that any generalization of facethood to other abductive settings would have to specify its underlying explanation semantics explicitly: preferential minimality in modal only-knowing systems, centrality under morphological erosion, framework variation in argumentation, or proof-theoretic weakness in intuitionistic synthesis. The current notion is therefore both precise and deliberately local: it captures heterogeneity in one important, but not universal, conception of propositional explanation (Schmidt et al., 20 Jul 2025).