Forgetting in Answer Set Programming – A Survey

Forgetting is part of the human nature, often with negative connotations, and investigation for its causes started in Cognitive Psychology in the late 19th century [20]. Studies show that forgetting can be caused, among others, by the interference in between new and previously learned information, by decay over time, or because certain links, called cues, used to retrieve a specific memory are not (any longer) available [75]. In fact, recent experiments in Neuroscience reveal that part of this forgetting appears to be done actively, in the sense that the human brain seems to be able to remove information no longer deemed necessary [16]. This enables humans to distinguish important memories among the huge amount of memories they acquire, while abstracting irrelevant details, which allows for making better decisions under varying conditions [69].

In organizations, intentional forgetting plays a similar role [47]. As increasing amounts of information may become difficult to process, forgetting allows one to simplify the interpretation of the acquired knowledge which contributes to future success of the organization. The methods of forgetting employed build on ideas of forgetting in Cognitive Psychology, for example, rather than unlearning organizational routines, the retrieval cues are forgotten [47].

In the same vein, huge problems appear in Computer Science, where the vastly increasing amounts of data challenge the limits of space in terms of physical storage and processing speed, in particular in situations where the exponential worst-case complexity becomes prohibitive. At the same time, forgetting has become increasingly important to properly deal with legal and privacy issues, such as, for example, the elimination of illegally acquired information to implement a court order, or enforcing the right to be forgotten, i.e., the right of the eliminination of private data on a person’s request, following the EU General Data Protection Regulation [28].

Thus, forgetting has received increasing attention in Computer Science aiming at the elimination of irrelevant information for improved decision processes and tools [6]. In Artificial Intelligence and Knowledge Representation and Reasoning in particular, the ideas of forgetting can be traced back to Boole’s variable marginalization \shortciteBoole54, also called the eliminination of middle terms. These notions raised interest when Lin and Reiter \shortciteLR94 considered forgetting about a fact or a relation in a first-order logic for Cognitive Robotics. This subsequently triggered research into forgetting and closely related notion such as uniform interpolation [83], variable elimination [53] or ignorance [5], being extended for example to Description Logics [35, 88, 64], Planning [27], and Modal Logic [95]. Commonly two kinds of forgetting can be found in the literature [25]. One of them eliminates some formula from a given knowledge base, in a way that closely resembles the operation of contraction in belief revision. The other, more common one, views forgetting as an operation that omits part of the vocabulary of the knowledge base, including possible adjustments on formulas to accommodate some notion of preservation of information (for the remaining vocabulary). This is the kind of forgetting we consider here and particularly useful when we want to eliminate elements representing auxiliary concepts, with the aim to simplify a knowledge base or improve its declarativity, or related to data protection issues. The importance of forgetting in Knowledge Representation and Reasoning is also witnessed by applications in cognitive robotics [62, 63, 68], for resolving conflicts [53, 93, 26, 55, 19], and ontology abstraction and comparison [88, 52, 49, 50].

While forgetting has been extensively studied in the context of classical logic [9, 53, 56, 57, 65, 66, 89, 30], it only recently gathered wider attention in Answer Set Programming (cf. the overview by Eiter and Kern-Isberner \shortciteEiterK19). Answer Set Programming (ASP) [33, 34] provides a declarative rule-based language for knowledge representation and reasoning accompanied with efficient implementations such as CLASP [32], DLV [58, 4], and Smodels [76]. Its non-monotonic rule-based nature required the development of specific methods and techniques – similar to what happened with other belief change operations such as revision and update [3, 22, 70, 77, 78, 18, 79]. The result is a significant number of different approaches on forgetting [93, 26, 91, 86, 85, 48, 87, 19, 41, 43, 37, 8, 7, 36], some presenting single operators, others semantic characterizations of classes of operators, usually with different sets of properties deemed desirable, some adapted from the literature on classical forgetting [95, 86, 87], others introduced for the case of ASP [26, 91, 86, 85, 48, 19, 43, 37], and often defined for different classes of answer set programs (details on the individual motivations for introducing these approaches and their distinct characteristics can be found in Section 5).

The result is a complex landscape filled with classes of operators and properties, with very little effort put into drawing a map that could help us to better understand the relationships between properties and operators. Whereas, in principle, having an operator that obeys all of the properties would be desirable, it turns out that any such operator cannot be defined for any class that includes the standard class of normal logic programs [46]. In fact, one of the properties arguably best captures the essence of forgetting, however, there cannot exist an operator that always satisfies it [41, 38]. This strengthens the idea that there cannot be a one-size-fits-all forgetting operator for ASP, but rather a variety of approaches, each obeying a specific set of properties. The choice of operator will then depend on which properties are deemed more important for the specific application in hand, for which it is important to understand: a) the relationships between different properties, b) which properties are obeyed by which (classes of) operators, and even c) whether some sets of properties make more sense than others. To this end, we present a systematic, comprehensive and thorough survey of forgetting in ASP, including many novel results and insights that help answering the raised questions and provide guidelines to users which operators are most suited in which applications.

After a brief section with preliminaries and a section where the common notion of forgetting in ASP is defined, the paper is divided into three main sections. The first one contains a comprehensive account of the properties found in the literature, together with an investigation into the relationships between them, including several novel results. The subsequent section is devoted to describing the operators defined in the literature, and establishing some results on their relationships, including an equivalence between two of these operators, as well as considerations on proposed concrete operators, and the computational complexity of problems related to forgetting. Then, we devote our final major section to present a comprehensive account of which properties are satisfied by which operators, some of the results to be found scattered in the literature, but more than half being novel. This section also provides a detailed discussion of the suitability of these classes with respect to their characteristics, and establishes precisely the relationship to uniform interpolation. Finally, the suitability of the classes of operators w.r.t. several applications considered in the context of forgetting is discussed.

This paper is a considerable extension of a previous conference publication [40]. The material has been extended to incorporate novel approaches in the literature and revised to accommodate these new results and provide a more complete picture of the existing properties and classes of forgetting operators, and their relations. The new detailed material on concrete forgetting operators, computational complexity, the relation to uniform interpolation, and suitability of classes of operators for certain applications help shed further light on these relations, and provide more guidance for the reader. This is complemented with an appendix, available as supplementary material to this paper, that contains all the proofs of the novel results, as well as those in the previous publication, and pointers for all those results spread out over the literature.

In this section, we recall the necessary notions and notation on logic programs under the answer set semantics.

We assume a propositional signature $\mathcal{A}$, i.e., a finite set of (propositional) atoms, also termed propositional variables synonymously. An (extended) logic program $P$ over $\mathcal{A}$ is a finite set of (extended) rules of the form

where all $a_{1},\ldots,a_{k},b_{1},\ldots,b_{l},c_{1},\ldots,c_{m}$, and $d_{1},\ldots,d_{n}$ are atoms of $\mathcal{A}$. Such rules $r$ are also commonly written in a more succinct way as

where $\mathit{A}=\{a_{1},\ldots,a_{k}\}$, $\mathit{B}=\{b_{1},\ldots,b_{l}\}$, $\mathit{C}=\{c_{1},\ldots,c_{m}\}$, and $\mathit{D}=\{d_{1},\ldots,d_{n}\}$, and we use both forms interchangeably. For each rule $r$, we distinguish the head, $\mathit{head}(r)=A$, and its body, $\mathit{body}(r)=\mathit{B}\cup not\,\mathit{C}\cup not\,not\,\mathit{D}$, where $not\,C$ and $not\,not\,D$ represent $\{not\,p\mid p\in C\}$ and $\{not\,not\,p\mid p\in D\}$, respectively. We write the set of atoms appearing in $P$ as $\mathcal{A}(P)$ and the class of (extended) logic programs as $\mathcal{C}_{e}$. Such extended logic programs are actually equivalent to an even more expressive syntax [60], but the more concise syntax used here suffices. Also note that double negation is commonly required in the context of forgetting in ASP and commonly supported by answer set solvers such as clingo [31].

The class of extended programs includes a number of special kinds of rules $r$: if $n=0$, then we call $r$ disjunctive; if, in addition, $k\leq 1$, then $r$ is normal; if on top of that $m=0$, then we call $r$ Horn; if moreover $l=0$, then we call $r$ a fact. We also admit constraints, which are (extended) rules where $k=0$. The classes of disjunctive, normal, and Horn programs, $\mathcal{C}_{d}$, $\mathcal{C}_{n}$, and $\mathcal{C}_{H}$, are defined as a finite set of disjunctive, normal, and Horn rules, respectively. We have $\mathcal{C}_{H}\subset\mathcal{C}_{n}\subset\mathcal{C}_{d}\subset\mathcal{C}_{e}$.

We now define the answer sets [34] of a program, i.e., its models, and we recall this notion based on HT-models as defined in the context of the logic of here-and-there [44], the monotonic logic underpinning ASP [67]. We start with the notion of reduct [34] of a program $P$ with respect to a set $I$ of atoms:

An HT-interpretation is a pair $\langle X,Y\rangle$ s.t. $X\subseteq Y\subseteq\mathcal{A}$. Note that we follow a common convention and usually abbreviate sets in HT-interpretations such as $\{a,b\}$ with the sequence of its elements, $ab$. Given a program $P$, an HT-interpretation $\langle X,Y\rangle$ is an HT-model of $P$ if $Y\models P$ and $X\models P^{Y}$, where $\models$ is the standard consequence relation for classical logic and where programs are interpreted as conjunctions of classical implications

corresponding to its rules $r$ of the form (1) where $\neg$, $\wedge$, and $\vee$ denote classical negation, conjunction and disjunction, respectively. We occasionally admit for the sake of readability that the HT-models of a program $P$ are restricted to $\mathcal{A}(P)$ even if $\mathcal{A}(P)\subset\mathcal{A}$. The set of all HT-models of $P$ is written $\mathcal{HT}(P)$. Given two programs $P_{1}$ and $P_{2}$, if $\mathcal{HT}(P_{1})\subseteq\mathcal{HT}(P_{2})$, then $P_{1}$ entails $P_{2}$ in HT, written $P_{1}\models_{\sf HT}P_{2}$. Also, $P_{1}$ and $P_{2}$ are HT-equivalent, written $P_{1}\equiv_{\sf HT}P_{2}$, if $\mathcal{HT}(P_{1})=\mathcal{HT}(P_{2})$.

A set of atoms $Y$ is an answer set of $P$ if $\langle Y,Y\rangle\in\mathcal{HT}(P)$, and there is no $X\subset Y$ such that $\langle X,Y\rangle\in\mathcal{HT}(P)$. We denote by $\mathcal{AS}(P)$ the set of all answer sets of $P$. For $\mathcal{C}_{d}$ and its subclasses, all $I\in\mathcal{AS}(P)$ are pairwise incomparable. A program $P$ is consistent if $\mathcal{AS}(P)$ is not empty.

Different notions of equivalence between programs have been established [24], essentially based on comparing their answer sets in different ways. Namely, we say that two programs $P_{1},P_{2}$ are equivalent, written $P_{1}\equiv P_{2}$, if $\mathcal{AS}(P_{1})=\mathcal{AS}(P_{2})$, i.e., their answer sets coincide. Uniform equivalence strengthens this condition by imposing that the equality of answer sets should hold even if a set of facts is added to both programs. Formally, two programs $P_{1},P_{2}$ are uniformly equivalent, written $P_{1}\equiv_{u}P_{2}$, if $\mathcal{AS}(P_{1}\cup R)=\mathcal{AS}(P_{2}\cup R)$ for every set $R$ of facts. Strong equivalence further strengthens the condition by imposing equality of answer sets under the addition of any set of rules. Formally, two programs $P_{1},P_{2}$ are strongly equivalent if $\mathcal{AS}(P_{1}\cup R)=\mathcal{AS}(P_{2}\cup R)$ for every $R\in\mathcal{C}_{e}$. It is well-known that two programs $P_{1},P_{2}$ are strongly equivalent exactly when $P_{1}\equiv_{\sf HT}P_{2}$ [59], and we therefore use $\equiv_{\sf HT}$ to denote both equivalences. Relativized equivalence relaxes the condition of strong equivalence by allowing to vary the language of the additional programs. Formally, two programs $P_{1},P_{2}$ are relativized equivalent w.r.t. $V\subseteq\mathcal{A}$, written $P_{1}\equiv_{V}P_{2}$, if $\mathcal{AS}(P_{1}\cup R)=\mathcal{AS}(P_{2}\cup R)$ for every $R\in\mathcal{C}_{e}$ s.t. $\mathcal{A}(R)\subseteq\mathcal{A}\text{\textbackslash}V$. Thus, strong equivalence and equivalence can be considered special cases of relativized equivalence, where the considered set $V$ is empty and identical to $\mathcal{A}$, respectively.

Occasionally, we want to omit certain elements of the signature from sets of interpretations and HT-models. Given a set of atoms $V$, the $V$-exclusion of a set of answer sets (resp. a set of HT-interpretations) $\mathcal{M}$, written $\mathcal{M}_{\parallel V}$, is $\{X\text{\textbackslash}V\mid X\in\mathcal{M}\}$ (resp. $\{\langle X\text{\textbackslash}V,Y\text{\textbackslash}V\rangle\mid\langle X,Y\rangle\in\mathcal{M}\}$). Also, given two sets of atoms $X,X^{\prime}\subseteq\mathcal{A}$, we write $X\sim_{V}X^{\prime}$ whenever $X\text{\textbackslash}V=X^{\prime}\text{\textbackslash}V$. For $HT$-interpretations $\langle H,T\rangle$ and $\langle X,Y\rangle$, $\langle H,T\rangle\sim_{V}\langle X,Y\rangle$ denotes that $H\sim_{V}X$ and $T\sim_{V}Y$. Then, for a set $\mathcal{M}$ of $HT$-interpretations, $\mathcal{M}_{\dagger V}$ denotes the set $\{\langle X,Y\rangle\mid\langle H,T\rangle\in\mathcal{M}$ and $\langle X,Y\rangle\sim_{V}\langle H,T\rangle\}$.

In this section, we formally introduce the notion of forgetting in answer set programming. More precisely, we define operators of forgetting and classes of these in a general way that aligns with all the different approaches presented in the literature. We note that our account is on forgetting propositional atoms, just as all the literature on forgetting in answer set programming. This also allows us to capture forgetting atoms from ground programs obtained from programs built over predicate symbols, constants, and variables, simply by considering a one-to-one mapping of such ground atoms to a propositional alphabet. To keep the exposition simple, we here consider forgetting in the propositional setting.

The principal idea of forgetting in ASP is to remove certain atoms from a given program, while preserving its semantics for the remaining atoms.

Thus, forgetting can be viewed as returning a set of programs, which are equivalent in some way, e.g., according to one of the notions presented in the previous section, that only mention the remaining atoms and preserve the semantics of the given program over these remaining atoms. In the literature, concrete operators have been defined that, similar to a function, provide one unique such representative for each program $P$ and set of atoms $V$ to be forgotten. We formalize this central idea with the notion of a forgetting operator.

Our definition establishes that forgetting can be understood as a reduction of the language preserving the semantic relations for the remaining atoms in $P$. The latter is in line with the argument in Example 3 and requires that these semantic relations be established based on some semantic notion such as answer sets or one of the established equivalence notions, though, for the sake of generality, no concrete notion is fixed. This condition allows us to exclude nonsensical functions, such as, for example simply always deleting the entire program or replacing it by arbitrary rules over the remaining atoms, but we do not specify precisely how these semantic relations are established. In the literature, more precise notions for such semantic relations have been defined. For example, Delgrande \shortciteDelgrande17 defined for arbitrary logics that the result of forgetting corresponds to all the consequences of the given formulas over the remaining language. As we will see in Sect. 5, this aligns with some of the approaches in the literature of forgetting in ASP, but many others rather rely on for example preserving some notion of models (answer sets or HT-models) in some way, which in fact often turns out to be closely connected to the properties of forgetting in ASP we present in Sect. 4. For this reason, we abstain here from specifying how these semantic relations are determined specifically and refer to Sect. 5 for the details for each of the existing approaches.

We point out that the notion of closed operators allows us to indicate whether an operator, when applied to a (sub-)class of programs for which it is defined, does return a program in that same class. This is important, since based on this we are able to answer the question whether such operator can be iterated on such (sub-)class. Naturally, by Def. 1, any forgetting operator is closed for the most general class for which it is defined.

It is worth noting that some notions of forgetting do not explicitly require that atoms to be forgotten be absent from the result of forgetting, but instead that they be irrelevant, i.e., the result of forgetting is strongly equivalent to a program that does not mention the atoms to be forgotten. In our view, this is not aligned with the conceptual idea of forgetting itself. However, since such (irrelevant) occurrences of atoms in a result of forgetting are commonly assumed to be not occurring in the result, the required strong equivalence naturally holds. Hence, requiring that forgetting operators yield programs without the atoms to be forgotten does not prevent coverage of these particular approaches.

Now, as Ex. 3 indicates, preserving the semantics for the remaining atoms is not necessarily tied to one unique program. In fact, in the literature, usually, a representative up to some notion of equivalence between programs is considered, which is also why, in this case, we often refer to a result of forgetting (in indefinite terms) as opposed to the result. In this sense, many notions of forgetting for logic programs are defined semantically, i.e., they introduce a class of operators that satisfy a certain semantic characterization. To capture this, we introduce the notion of a class of forgetting operators.

The subset inclusion for the domain of operators in the previous definition is justified by the fact that there are classes of operators in the literature that include concrete operators only defined for a subclass of the programs considered by the class of operators.

Finally, with respect to uniform interpolation, we note that it is indeed closely connected to the concept of forgetting [29]. However, it does not exactly correspond to the notion of forgetting in ASP in the broad sense as considered here. We will discuss this in more detail in Sect. 6.4 and establish a precise relationship after we have properly presented properties of forgetting, classes of forgetting operators and their relations.

In the literature of forgetting in answer set programming, commonly one central focus has been the investigation of guiding principles that would provide desirable characteristics of classes of operators of forgetting, often called properties of forgetting. In this section, we recall these properties found in the literature and investigate existing relations between them.

In the course of this presentation, we opt for following a historical order without any considerations on their importance or preference among each other. In terms of technical notation, unless otherwise stated, $\mathsf{F}$ represents a class of forgetting operators, $\mathcal{C}(\mathsf{f})$ the class of programs over $\mathcal{A}$ of a given $\mathsf{f}\in\mathsf{F}$, and whenever we write that a single operator ${\mathsf{f}}$ obeys some property, we mean that the singleton class composed of that operator, $\{{\mathsf{f}}\}$, obeys such property.

The first three properties, named strengthened Consequence, weak Equivalence, and Strong Equivalence, were proposed by Eiter and Wang \shortciteEiterW08, though not formally introduced as such. The first two were in fact guiding principles for defining their notion of forgetting, and formalized later [40], while the third was frequently considered in the literature in terms of formal results, but only formalized as a property by Wang et al. \shortciteWangWZ13.

Strengthened Consequence requires that the answer sets of a result of forgetting be answer sets of the original program, ignoring the atoms to be forgotten.

• (sC)

  $\mathsf{F}$ satisfies strengthened Consequence if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, we have $\mathcal{AS}(\mathsf{f}(P,V))\subseteq\mathcal{AS}(P)_{\parallel V}$.

In other words, forgetting does not admit the introduction of new answer sets, it may only remove some in the course of forgetting.

The other two properties focus on the preservation of some notion of equivalence during forgetting, i.e., if two programs are equivalent (w.r.t. some notion of equivalence of programs), then the respective results are as well.

• (wE)

  $\mathsf{F}$ satisfies weak Equivalence if, for each $\mathsf{f}\in\mathsf{F}$, $P,P^{\prime}\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$: if $P\equiv P^{\prime}$, then $\mathsf{f}(P,V)\equiv\mathsf{f}(P^{\prime},V)$.

• (SE)

  $\mathsf{F}$ satisfies Strong Equivalence if, for each $\mathsf{f}\in\mathsf{F}$, $P,P^{\prime}\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$: if $P\equiv_{\sf HT}P^{\prime}$, then $\mathsf{f}(P,V)\equiv_{\sf HT}\mathsf{f}(P^{\prime},V)$.

Weak Equivalence and Strong Equivalence require that forgetting preserves equivalence and strong equivalence of programs, respectively. For other notions of equivalence, no corresponding property has been considered in the literature.

The next four properties, called Irrelevance, Weakening, Positive Persistence, and Negative Persistence, were introduced by Zhang and Zhou \shortciteZZ-AI09 as postulates of knowledge forgetting in the context of modal logics, and later adopted by Wang et al. \shortciteWangZZZ12,WangZZZ14 for forgetting in ASP.

Irrelevance requires that a result of forgetting be strongly equivalent to a program that does not mention the atoms to be forgotten.

• (IR)

  $\mathsf{F}$ satisfies Irrelevance if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, $\mathsf{f}(P,V)\equiv_{\sf HT}P^{\prime}$ for some $P^{\prime}$ not containing any $v\in V$.

This property corresponds to the concept of being irrelevant discussed in the end of the previous section. Thus, satisfaction of this property is an integral part of the definitions of forgetting operators and classes (cf. Defs. 1 and 2).

The other three properties focus on HT-consequences. Namely, Weakening requires that the HT-models of $P$ also be HT-models of a result of forgetting.

• (W)

  $\mathsf{F}$ satisfies Weakening if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, we have $P\models_{\sf HT}\mathsf{f}(P,V)$.

This means that a result of forgetting, $\mathsf{f}(P,V)$, has at most the same consequences as the program $P$ itself.

Positive and negative persistence concern preserving the HT-consequences of $P$ and not introducing new ones, respectively.

• (PP)

  $\mathsf{F}$ satisfies Positive Persistence if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$: if $P\models_{\sf HT}P^{\prime}$, with $P^{\prime}\in\mathcal{C}(\mathsf{f})$ and $\mathcal{A}(P^{\prime})\subseteq\mathcal{A}\text{\textbackslash}V$, then $\mathsf{f}(P,V)\models_{\sf HT}P^{\prime}$.

• (NP)

  $\mathsf{F}$ satisfies Negative Persistence if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$: if $P\not\models_{\sf HT}P^{\prime}$, with $P^{\prime}\in\mathcal{C}(\mathsf{f})$ and $\mathcal{A}(P^{\prime})\subseteq\mathcal{A}\setminus V$, then $\mathsf{f}(P,V)\not\models_{\sf HT}P^{\prime}$.

Thus, Positive Persistence requires that the HT-consequences of $P$ not containing atoms to be forgotten be preserved in a result of forgetting, while Negative Persistence requires that a program not containing atoms to be forgotten not be an HT-consequence of $\mathsf{f}(P,V)$, unless it was already a HT-consequence of $P$.

Essentially in parallel to the appearance of the previous set of properties, Wong introduced a set of properties in his PhD dissertation \shortciteWong09. They were defined for forgetting a single atom from a given disjunctive program, and did not gather much attention in the literature, possibly due to the form of publication. These properties were generalized to extended programs and to forgetting sets of atoms and assigned a more descriptive name (as Wong simply used alphanumeric identifiers) [40, 39, 42]. In the course of this generalization, it turned out that two of the resulting properties would precisely coincide with (SE) and (PP). Thus, in the following, we will present those generalizations of Wong’s properties that are distinct from the previous ones, using the descriptive names for the ease of readability. Namely, we recall, Strong Invariance, Strong Consequence, Rule Consequence, Non-contradictory Consequence and Permutation Invariance.

Strong Invariance requires that it be (strongly) equivalent to add a program without the atoms to be forgotten before or after forgetting.

• (SI)

  $\mathsf{F}$ satisfies Strong (addition) Invariance if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, we have $\mathsf{f}(P,V)\cup R\equiv_{\sf HT}\mathsf{f}(P\cup R,V)$ for all programs $R\in\mathcal{C}(\mathsf{f})$ with $\mathcal{A}(R)\subseteq\mathcal{A}\text{\textbackslash}V$.

In particular, this means that when computing a forgetting result using an operator from a class that satisfies this property, we can ignore the rules that do not mention the atoms to be forgotten while forgetting, and only add them in the end to this result.

The next three properties are related to HT-consequences. Namely, Strong Consequence requires that HT-consequences be preserved when forgetting.

• (SC)

  $\mathsf{F}$ satisfies Strong Consequence if, for each $\mathsf{f}\in\mathsf{F}$, $P,P^{\prime}\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, if $P\models_{\sf HT}P^{\prime}$, then $\mathsf{f}(P,V)\models_{\sf HT}\mathsf{f}(P^{\prime},V)$.

This is similar in spirit to properties (wE) and (SE), only here the HT-consequence is preserved while forgetting.

Rule Consequence requires that any rule which is a consequence of a result of forgetting about $V$ from $P$ be a consequence of a result of forgetting about $V$ from a single rule among the HT-consequences of $P$.

• (RC)

  $\mathsf{F}$ satisfies Rule Consequence if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, if $\mathsf{f}(P,V)\models_{\sf HT}r$, then $\mathsf{f}(\{r^{\prime}\},V)\models_{\sf HT}r$ for some rule $r^{\prime}$ such that $\{r^{\prime}\}\in\mathcal{C}(\mathsf{f})$, and $P\models_{\sf HT}r^{\prime}$.

In particular, since any rule $r$ in a result of forgetting is an HT-consequence of it, for each such rule, there is a rule $r^{\prime}$ in the original program that gives rise to $r$ (in terms of HT-consequence).

Non-contradictory Consequence requires that whenever a rule $r$ is an HT-consequence of a forgetting result $\mathsf{f}(P,V)$, then the rule obtained by adding the default negation of the atoms of $V$ to the body of $r$ should be an HT-consequence of $P$.

• (NC)

  $\mathsf{F}$ satisfies Non-contradictory Consequence if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, if $\mathsf{f}(P,V)\models_{\sf HT}\mathit{A}\leftarrow\mathit{B}\cup not\,\mathit{C}\cup not\,not\,\mathit{D}$, then $P\models_{\sf HT}\mathit{A}\leftarrow\mathit{B}\cup not\,\mathit{C}\cup not\,V\cup not\,not\,\mathit{D}$.

Again, since any rule $r$ in a result of forgetting is an HT-consequence of it, the original program $P$ has such a non-contradictory HT-consequence.

The final property by Wong, Permutation Invariance, requires that the order not be relevant when sequentially forgetting atoms.

• (PI)

  $\mathsf{F}$ satisfies Permutation Invariance if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$, and $V\subseteq\mathcal{A}$, we have that $\mathsf{f}(P,V)\equiv_{\sf HT}\mathsf{f}(...\mathsf{f}(P,V_{1}),...,V_{n})$ for every partition $\{V_{1},...,V_{n}\}$ of $V$.

In fact, such iteration of forgetting had been considered before in a similar manner in a result by Eiter and Wang 2008, but based on an already given sequence of atoms to be forgotten instead of a set. Thus, (PI) is slightly more general Gonçalves et al. (2017). Also, independently, a variant of (PI) was introduced by Wang et al. 2013, but shown to be equivalent Gonçalves et al. (2017). Therefore, in the following, we only use (PI) as a representative of these properties of invariance for different orders of forgetting a set of atoms.

The next property, called Existence, was first discussed by Wang et al. Wang et al. (2012) and formalized by Wang et al. Wang et al. (2013). It requires that a result of forgetting for $P$ in $\mathcal{C}$ be again in the class $\mathcal{C}$. This is important to determine if class of forgetting operators is suitably well-defined for the class of programs it is intended for, as well for iteration on subclasses of programs for which it is defined. We follow the notation introduced in Gonçalves et al. (2016b) which formalizes this property s.t. it be explicitly tied to a class $\mathcal{C}$, thus allowing to speak about a class of forgetting operators $\mathsf{F}$ being closed for different classes $\mathcal{C}$.

• (E_C)

  $\mathsf{F}$ satisfies Existence for $\mathcal{C}$, i.e., $\mathsf{F}$ is closed for a class of programs $\mathcal{C}$ if there exists ${\mathsf{f}}\in\mathsf{F}$ s.t. ${\mathsf{f}}$ is closed for $\mathcal{C}$.

Thus, class $\mathsf{F}$ being closed for some $\mathcal{C}$ requires that there exist some “witness in favor of it”. This also means that if a class $\mathsf{F}$ of operators does not satisfy this property for some class $\mathcal{C}$ of programs, no operator of $\mathsf{F}$ can be found that is closed for that class of programs.

The next property, called Consequence Persistence, was introduced by Wang et al. Wang et al. (2013) building on the ideas behind (sC) by Eiter and Wang Eiter and Wang (2008). Consequence persistence requires that the answer sets of a result of forgetting correspond exactly to the answer sets of the original program, ignoring the atoms to be forgotten.

• (CP)

  $\mathsf{F}$ satisfies Consequence Persistence if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, we have $\mathcal{AS}(\mathsf{f}(P,V))=\mathcal{AS}(P)_{\parallel V}$.

In other words, forgetting cannot introduce new answer sets nor remove existing ones.

The following property, called Strong Persistence, was introduced by Knorr and Alferes Knorr and Alferes (2014) with the aim of imposing the preservation of all dependencies contained in the original program building on ideas of strong equivalence between the original program and a result of forgetting (modulo the atoms to be forgotten).

• (SP)

  $\mathsf{F}$ satisfies Strong Persistence if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, we have $\mathcal{AS}(\mathsf{f}(P,V)\cup R)=\mathcal{AS}(P\cup R)_{\parallel V}$, for all programs $R\in\mathcal{C}(\mathsf{f})$ with $\mathcal{A}(R)\subseteq\mathcal{A}\text{\textbackslash}V$.

This strengthens (CP) considerably by imposing that the correspondence between answer sets of the result of forgetting and those of the original program be preserved in the presence of any additional set of rules not containing the atoms to be forgotten.

A further property, Weakened Consequence, is due to results by Delgrande and Wang Delgrande and Wang (2015). It requires that the answer sets of the original program be preserved while forgetting, ignoring the atoms to be forgotten.

• (wC)

  $\mathsf{F}$ satisfies weakened Consequence if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, we have $\mathcal{AS}(P)_{\parallel V}\subseteq\mathcal{AS}(\mathsf{f}(P,V))$.

This property can thus also be understood as being the counterpart to (sC): one does prevent the introduction of new answer sets, the other the loss of existing ones. One can also observe that they correspond to the two inclusions of (CP).

The following two properties, weakened and strengthened Strong Persistence, are similiar in spirit, as they are generalizations of (wC) and (sC) that correspond to the two inclusions of (SP) Gonçalves et al. (2017).

• (wSP)

  $\mathsf{F}$ satisfies weakened Strong Persistence if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, we have $\mathcal{AS}(P\cup R)_{\parallel V}\subseteq\mathcal{AS}(\mathsf{f}(P,V)\cup R)$, for all $R\in\mathcal{C}(\mathsf{f})$ with $\mathcal{A}(R)\subseteq\mathcal{A}\text{\textbackslash}V$.

• (sSP)

  $\mathsf{F}$ satisfies strengthened Strong Persistence if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, we have $\mathcal{AS}(\mathsf{f}(P,V)\cup R)\subseteq\mathcal{AS}(P\cup R)_{\parallel V}$, for all $R\in\mathcal{C}(\mathsf{f})$ with $\mathcal{A}(R)\subseteq\mathcal{A}\text{\textbackslash}V$.

Weakened Strong Persistence guarantees that all answer sets of $P$ are preserved when forgetting, no matter which rules $R$ over $\mathcal{A}\text{\textbackslash}V$ are added to $P$, while strengthened Strong Persistence ensures that all answer sets of a result of forgetting indeed correspond to answer sets of $P$, independently of the added set of rules $R$.

Finally, Uniform Persistence was introduced by Gonçalves et al. Gonçalves et al. (2019) in the context of forgetting in modular answer set programming. Uniform Persistence requires that the correspondence between answer sets of a result of forgetting and those of the original program be preserved in the presence of any additional set of facts not containing the atoms to be forgotten.

• (UP)

  $\mathsf{F}$ satisfies Uniform Persistence if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, we have $\mathcal{AS}(\mathsf{f}(P,V)\cup R)=\mathcal{AS}(P\cup R)_{\parallel V}$, for all sets of facts $R$ with $\mathcal{A}(R)\subseteq\mathcal{A}\text{\textbackslash}V$.

Hence, (UP) can be seen as a variant of (SP) under uniform equivalence. Since this is the only property in the literature related to uniform equivalence, we complement with one further property to complete the picture in that regard. It is a variant of a property already presented, namely (SI), but directed towards uniform equivalence.

• (UI)

  $\mathsf{F}$ satisfies Uniform (addition) Invariance if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, we have $\mathsf{f}(P,V)\cup R\equiv_{\sf HT}\mathsf{f}(P\cup R,V)$ for all sets of facts $R$ with $\mathcal{A}(R)\subseteq\mathcal{A}\text{\textbackslash}V$.

Thus, Uniform Invariance requires that it be (strongly) equivalent to add a set of facts without the atoms to be forgotten before or after forgetting.

Recently, a different relaxation has of (SI) has been introduced by Gonçalves et al. Gonçalves et al. (2021) when investigating syntactic operators under uniform equivalence.

• ${\bf(SI_{u})}$

  $\mathsf{F}$ satisfies Strong Invariance with respect to uniform equivalence if, for each $\mathsf{f}\in\mathsf{F}$, $P\in\mathcal{C}(\mathsf{f})$ and $V\subseteq\mathcal{A}$, we have $\mathsf{f}(P,V)\cup R\equiv_{u}\mathsf{f}(P\cup R,V)$, for all programs $R\in\mathcal{C}(\mathsf{f})$ with $\mathcal{A}(R)\subseteq\mathcal{A}\text{\textbackslash}V$.

This property relaxes strong invariance by allowing that rules not mentioning the atoms to be forgotten can be ignored and be added to the result still preserving uniform equivalence (and not strong equivalence as (SI)). The difference to (UI) is that (UI) allows us to ignore facts over the remaining language while forgetting (under strong equivalence) whereas ${\bf(SI_{u})}$ allows us to ignore rules over the remaining language under uniform equivalence. The latter is arguably more useful when forgetting from a program that contains general rules not mentioning the atoms to be forgotten.

All these properties are not orthogonal to one another, and in what follows we join the results established in the literature on the relations that exist between them. In particular, we opt for presenting these results in a concise way, trying to avoid repetition of results that are implicitly obtained from others.

To ease the reading, for a property (P), we represent with “(P)” that “$\mathsf{F}$ satisfies (P)”.

Note first, that 1., as proven in by Ji et al.Ji et al. (2015), also relies on (IR) in its original formulation, in particular, (W) is equivalent to (NP) and (IR). However, as (IR) is an integral part of our definition of forgetting operators, this reliance is ensured implicitly. This means that, by 1., the four properties, originally proposed by Zhang and Zhou Zhang and Zhou (2009), in the context of forgetting in ASP actually reduce to two distinct ones, namely (W) and (PP).

The following results, 2.–9., show that (SP) is an expressive property, where 2. was shown by Knorr and Alferes Knorr and Alferes (2014) 3.–6. by Gonçalves et al.  Gonçalves et al. (2016b), and 7.–9. are new. In fact, 3. provides a non-trivial decomposition of (SP) into (SI) and (CP). These two are themselves expressive, as witnessed by other results. Namely, 4. shows that (CP) in turn is the combination of (sC) and (wC), and 5. that it implies preservation of equivalence, while 6. provides the non-trivial result that Strong Equivalence and Strong Invariance imply Positive Persistence. Then, 7. provides another decomposition of (SP) into (wSP) and (sSP) which in turn, by 8. and 9., are implied by properties used in the decompositions 3. and 4., respectively. Thus, (SP) implies (SE), (CP), (SI), (sC), (wC), (wE), (PP), (wSP)  and (sSP), where the result for (PP) has been shown directly Ji et al. Ji et al. (2015).

The next three results, 10.-12., were shown by Gonçalves et al. Gonçalves et al. (2017) and clarify the relation between the properties originally proposed by Zhang and Zhou Zhang and Zhou (2009) and those by Wong Wong (2009). Namely, the two distinct properties (W) and (PP) introduced by Zhang and Zhou Zhang and Zhou (2009) together imply (SC) which in turn implies (SE), where the combination of these two results, i.e., that (W) and (PP) together imply (SE), has been shown directly by Gonçalves et al. Gonçalves et al. (2016b). In addition, (W) alone implies (NC), which indicates that among Wong’s properties, (SC) is stronger than (NC).

The following two results establish incompatibility results for classes of operators defined over (at least) normal programs. Both results are new, though 13. is a revised result by Wang et al. Wang et al. (2013) where incompatibility of (W) with (CP) is established. In this sense, this new result makes the incompatibility more precise. In comparison to 14., we can also observe that, though no formal relation exists between (W) and (RC), (W) is stronger than (RC) in the sense that it is incompatible with (wC), whereas, in the case of (RC), an additional property is necessary, namely (UI), to establish incompatibility with (wC).

The two new results 15.–16., establish relations between the new properties introduced here w.r.t. uniform equivalence and their correspondents based on strong equivalence. In fact, 15. establishes that Strong Invariance implies Uniform Invariance. The second result provides a decomposition of (UP), and it is interesting as it, together with 3., allows us to trace the difference between (UP) and (SP) to the different considered form of invariance. This strengthens previous results by Gonçalves et al. Gonçalves et al. (2019) that positioned (UP) in between (CP) and (SP). However, unlike (SP), (UP) is incompatible with any property that allows to ignore the remaining arbitrary rules (beyond facts). This is naturally the case for (SI), as (SI) relies on strong equivalence whereas (UP) relies on uniform equivalence, but, by 18., even the relaxation of (SI) to uniform equivalence, ${\bf(SI_{u})}$ (17.), is incompatible with (UP) Gonçalves et al. (2021).

To gain a better overview on the results in Thm. 1, Fig. 1 summarizes the positive results (all but the incompatibility results) in graphical form, representing equivalences and implications between (compositions of) properties. We can observe that (SP) is indeed important as the vast majority of properties is implied by it, further strengthening our view that this property is central as it arguably best captures preserving all the relations between the remaining atoms. Though (SP) can in general not be satisfied for classes of programs containing normal programs Gonçalves et al. (2016c); Gonçalves et al. (2020), in our view, it ideally corresponds to what forgetting in ASP should amount to whenever possible. Now, among the pairs of properties equivalent to (SP), (SI) and (CP) are arguably more important: the former allows one to focus on the rules containing the atoms to be forgotten when forgetting, which is beneficial for computation and amenable to syntactic forgetting, and the latter captures (the baseline in comparison to (SP)) that the answer sets be preserved. While (SP) is in general not satisfiable, its relaxation to uniform equivalence, (UP), is Gonçalves et al. (2019). In fact, (UP) is the strongest relaxation of (SP) w.r.t. programs $R$ such that there is a forgetting operator over a class including normal programs that satisfies it Gonçalves et al. (2021). This makes this property an important alternative, well-suited in the setting of ASP, where problem solutions are often encoded as rules with varying sets of facts. There are also two properties not present in Fig. 1 that we deem important, (PI) and (E_C). The former allows forgetting atoms in any order, which facilitates the usage of forgetting and its implementation in concrete operators that may be defined forgetting one atom at a time. The latter is crucial for guaranteeing that operators (and classes of these) are indeed well-defined as well as permitting the iteration of these. Finally, we note that among properties that are variants of each other with respect to equivalence and strong equivalence, such as (wE) and (SE), we consider those based on strong equivalence more important as it is well-known that equivalence does not preserve the structure of rules, and this applies also in the context of forgetting. We revisit these observations in more detail taking into consideration also the existing approaches in the literature and which properties these satisfy (as presented in the following sections).

The different properties presented in the literature have often been the driving motivation for the definition of a variety of different approaches on forgetting in answer set programming. We now turn our attention to these classes of operators of forgetting, by first reviewing the approaches found in the literature and then establishing non-trivial relations between them. In the course of the exhibition, we follow a chronological order, similar in spirit to the presentation of the properties in the previous section.