Introduction
A core question in historiography (as in many other academic and non academic fields) is: how much of a text (or of a story as is told in a text or other medium) can we believe? In his analysis of the story of King David as it is told in the Book of Samuel and 2 Kings [11 ] , the biblical historiographer Baruch Halpern suggests (cf. [11 Part III, Ch. 5] ) that this text1 can be analysed using the same basic interpretive strategy which applies to the Assyrian annals of Tiglath-Pileser, in which “the torching of a grain field is the conquest of a whole territory beyond it. A looting raid becomes a claim of perpetual sovereignty. But this does not mean that campaigns can be confected. The technique is that of putting extreme spin on real events.” Halpern continues (ibid.):
Interpreting such literature demands only a simple rule, the Tiglath-Pileser principle. The question is, what is the minimum the king might have done to lay claim to the achievements he
publishes? Looting a town? He shoplifted a toothbrush from the local drug store. Ravaging the countryside? Perhaps he trampled crops near a farmstead. Receiving submission from distant kings in lands one hasn’t invaded? A delegation arrived to inaugurate diplomatic relations. Each small mark of prestige becomes the evidence for a grand triumph. [...] The point is, such figures cannot make claims without any basis in fact [...] But it is always a temptation to paint one’s achievements in the best of all possible lights. To correct for this tendency, one has merely to imagine what insignificant action would produce the claim that “I conquered Egypt” (a raid on a border station in the Sinai). The same political hermeneutic applies reasonably well today.
Thus, the Tiglath-Pileser principle is grounded in the assumption that the given narration is anchored in truth (“this does not mean that campaigns can be confected”, “such figures cannot make claims without any basis in fact”), because the narrator must take into account what knowledge his original audience has that cannot be contradicted. Moreover, the Tiglath-Pileser principle provides a way of attaining a more accurate approximation of the truth, by taking into account both the original audience’s knowledge of the actual circumstances (which constrain the narrator to not make claims that would be immediately disbelieved), and the narrator’s agenda, and thus correcting for the tendency of this agenda to ‘spin the truth’. Building upon this ‘simple rule’ enables Halpern to reach the following surprising conclusions in respect to King David’s story [11 Afterword, p. 479] :
[...] the only murder in the books of Samuel of which [David] was probably innocent is the one murder of which he stands accused in the apology. [...] At the end of his life, David was certainly betrayed by one of his harem women, Bathsheba, and his foreign mercenary bodyguard. Instead of his heir, Adonijah, they installed Solomon, hitherto known as the son not of David but of Uriah the Hittite, on the throne. [...] this second king of Judah quickly cleaned house, erasing his rivals, and probably all of David’s sons, from the scene.
This paper presents a framework which, inter alia , can formally represent and trace the steps from the Tiglath-Pileser principle to Halpern’s conclusions. This framework builds on the theory developed by Rubinstein and Glazer in a series of papers [8 , 9 , 10 ] in which several types of situations are analysed in which one or more agents (the Speakers) wish to persuade another agent, the Listener (who plays the principal’s role),2 ^2 2 to make a decision according to their desiderata, e.g. to accept a particular request, to award a prize to them, or to believe a certain claim they make to be true. The Listener’s decision depends on what they conclude about a certain state of affairs, which can either concern the (particular) Speaker (e.g. whether the Speaker satisfies certain requirements) or some other facts about the world. Each Speaker knows the true state of affairs, whereas the Listener knows only the set S S S of the possible states of affairs (and possibly some information about the Speaker’s goals and strategies). Each Speaker wishes to persuade the Listener regardless of the true state of affairs. Hence, the Speakers can lie, and will do so whenever declaring the true state of affairs would not favour them. However, the key feature of the scenarios in [8 , 10 ] is that the Speakers’ lies are constrained by a particularly strong bounded rationality assumption that anchors their lies to the truth. Hence, in certain cases, the Listener can engineer the enquiry in such a way that they can approximate the true state of affairs to varying degrees of accuracy. To formally represent this idea, Rubinstein and Glazer introduce the notion of codex , a finite set L L L of constraints, expressed as formulas of classical propositional logic, which the Speaker’s declared state is expected to satisfy. The specific way in which the Speakers’ lying ability is constrained is formalized by the guidance relation → L ⊆ S × S \to_{L} \subseteq S \times S → L ⊆ S × S associated with L L L , where s → L s ′ s \to_{L} s' s → L s ′ if, whenever s s s is the true state, the Speaker could declare s ′ s' s ′ in an attempt to satisfy codex L L L .
Guidance relations are perspicuous formal representations of the Tiglath-Pileser principle in action: by knowing the Speaker’s guidance relation, the Listener is able to conclude that the true state must belong to the preimage of s ′ s' s ′ under → L \to_{L} → L , whenever the Speaker declares s ′ s' s ′ as the ‘true state’. Depending on the properties of the guidance relation, the Listener is able to gauge the information needed for their decision.
We introduce a class of relations, referred to as Tiglath-Pileser relations , which generalize guidance relations, and need not arise from codices. We pursue an axiomatic development of the properties of Tiglath-Pileser relations required for the successful application of the Tiglath-Pileser principle. We introduce a logic-based framework for the analysis of scenarios in which the Listener’s approximations of the Speaker’s Tiglath-Pileser relation are dynamically formed over the course of extended interaction, and apply it to two case studies.
The framework introduced in the present paper differs from that of [8 , 9 , 10 ] in important respects. Firstly, while in [8 , 10 ] , the constraints on the Speakers’ ability to lie are purely a result of their bounded rationality, we explore situations in which the background knowledge of the Listener imposes further constraints. Secondly, we are interested in exploring more open-ended scenarios than those considered in [8 , 9 , 10 ] , where the latter mostly concern the structuring of selection procedures designed to elicit the truth in situations which can be modelled as one-off interactions, and the goals of the Speaker(s) are common knowledge among all agents. In contrast, we are interested in accounting for possibly extended interactions among agents, in which the goals of the Speaker(s) might not be explicitly known, and in fact, the Speaker(s) might have an interest in concealing them both from their primary intended audience, and from any agent who, playing the “historian’s role”, tries to interpret the whole communication. Thirdly, in contrast to [8 , 9 , 10 ] , we take a dynamic perspective on how guidance relations arise in the Listener’s mind, and how this process can be manipulated by strategic Speakers to feed the Listener a false relation.3 Fourthly, in [8 , 10 ] , the Speakers are assumed to “not think strategically , i.e. not [to] take into consideration the other party’s actions”, but rather act as “problem solvers”, e.g. by (only) making sure that their declarations verify the codices they have been assigned; in contrast, we structure the analysis of the present paper according to a hierarchy of levels of strategic/cognitive [15 , 13 , 4 ] of the agents involved, the most basic level of which is that of the “non-strategically thinking” agents considered in [8 , 10 ] . While, like [8 , 10 ] , we avoid an explicit game-theoretic treatment, this approach sets the stage for a game-theoretic development of signalling games with uncertain information and players acting at different levels of the strategic/cognitive hierarchy. In this perspective, this paper contributes novel insights about the formal modelling of trust based on inferring the motivations and persuasion strategy of other agents, rather than purely on their track-record of truth-telling in previous rounds of a game [12 , 2 ] .
The framework developed here has direct applications to detecting and managing disinformation. In adversarial information environments, strategic actors rarely invent entirely false narratives; instead, they often exploit, selectively amplify, and distort fragments of existing information to create misleading impressions [16 ] . By modelling propagandists as strategic Speakers constrained by the information their audience possess, our framework enables analysts, algorithms, and fact-checkers (Listeners) to attain a reliable approximation of the truth embedded within manipulated narratives. The present treatment of extended interactions and higher-order reasoning also models how sophisticated campaigns shape perceptions of both events and source motives.
Another area of application concerns AI safety and the governance of agentic systems. As AI agents become more capable, risks such as sycophancy and deceptive alignment become increasingly important [1 ] . This framework makes it possible to model AI systems as strategic Speakers interacting with human Listeners, and hence it can be used by supervisors to audit, probe, and constrain potentially deceptive behaviour. Such an approach aligns with emerging work on scalable oversight, interpretability, and evaluations designed to detect undesirable behaviours in advanced AI systems [3 , 5 ] .
Structure of the paper. In § 2 , we give a detailed summary and analysis of the formalisms introduced in [8 , 10 ] ; in § 3 , we further develop and refine this formal framework, with a particular focus on the relationship between codices and guidance relations; in § 4 , we recast this formalism in a purely set-theoretic way; in this formal environment, the analysis is structured along a hierarchy of levels of strategic/cognitive awareness of the agents involved, successive levels of which correspond to deeper levels of sophistication in the agents’ reasoning about other agents’ motivations and Tiglath-Pileser relations; in § 5 , we introduce a logic-based framework which accounts for how codices and their associated Tiglath-Pileser relations are dynamically generated in the mind of the Listener in the context of a narration which unfolds over several episodes, how these formal structures govern the process of formation of reliable beliefs in the Listener’s mind, how this process can be manipulated by strategic Speakers, and how the Listener can out-manoeuvre the Speaker in these manipulative attempts. This framework is applied to case studies in § 6 and § 7 , the latter section collecting the formal account of Baruch Halpern’s analysis of King David’s story. We conclude in § 8 .
Preliminaries
In this section, we review the formalisms introduced in [8 , 10 ] , where they are embedded in extensive motivating case studies of persuasion scenarios. Our presentation abstracts from the latter, while highlighting and developing the mathematical and formal aspects, preparing the ground for the introduction of the generalization to Tiglath-Pileser relations in § 3 .
Basic formalism of `A model of persuasion'
For any (finite) set P r o p = { v 1 , … , v K } \mathrm{Prop} = \{v_1, \ldots, v_K\} Prop = { v 1 , … , v K } of propositional variables4 with K ≥ 2 K \ge2 K ≥ 2 , a profile , or state 5 , is a map s : P r o p → 2 = { 0 , 1 } s : \mathrm{Prop} \to2 = \{0,1\} s : Prop → 2 = { 0 , 1 } ; we let S S S denote the set of states and write s ⊨ v s \models v s ⊨ v iff s ( v ) = 1 s(v) = 1 s ( v ) = 1 .
In [8 ] , leader-follower scenarios are analysed in which a Listener announces and commits to a codex L L L , which is a logically consistent set of propositional formulas in the language of V V V that the profile presented by a Speaker must satisfy in order for the Speaker’s request to be accepted. Then, the Speaker (who knows the true profile) announces a profile that may or may not be the true one.
Each formula (rule) φ \varphi φ in the codex is required to have the following syntactic shape:
⋀ y ∈ Y φ y → φ x \bigwedge_{y \in Y} \varphi_y \to\varphi_x y ∈ Y ⋀ φ y → φ x
where Y Y Y is a nonempty subset of P r o p \mathrm{Prop} Prop , x ∈ P r o p ∖ Y x \in\mathrm{Prop} \setminus Y x ∈ Prop ∖ Y , and each φ v \varphi_v φ v is either v v v or ¬ v \neg v ¬ v (the negation of v v v ) for any v ∈ P r o p v \in\mathrm{Prop} v ∈ Prop . For example, the proposition v 4 ∧ ¬ v 1 → v 3 v_4 \land\neg v_1 \to v_3 v 4 ∧ ¬ v 1 → v 3 can be a rule in a codex, but v 1 → ¬ v 1 v_1 \to\neg v_1 v 1 → ¬ v 1 cannot.
For every formula φ \varphi φ and every set of formulas Γ \Gamma Γ , let Var ( φ ) \operatorname{Var}(\varphi) Var ( φ ) denote the set of variables occurring in φ \varphi φ and let Var ( Γ ) = ⋃ φ ∈ Γ Var ( φ ) \operatorname{Var}(\Gamma) = \bigcup_{\varphi\in\Gamma} \operatorname{Var}(\varphi) Var ( Γ ) = ⋃ φ ∈ Γ Var ( φ ) . A rule φ \varphi φ is complete if Var ( φ ) = P r o p \operatorname{Var}(\varphi) = \mathrm{Prop} Var ( φ ) = Prop .
For any rule φ = ⋀ y ∈ Y φ y → φ x \varphi= \bigwedge_{y \in Y} \varphi_y \to\varphi_x φ = ⋀ y ∈ Y φ y → φ x , we denote its antecedent and consequent as follows:
a ( φ ) = ⋀ y ∈ Y φ y and z ( φ ) = φ x . a(\varphi) = \bigwedge_{y \in Y} \varphi_y \qquad\text{and} \qquad z(\varphi) = \varphi_x. a ( φ ) = y ∈ Y ⋀ φ y and z ( φ ) = φ x .
Codices are required to be not only satisfiable but also coherent , i.e. for all φ , ψ ∈ L \varphi, \psi\in L φ , ψ ∈ L , if a ( φ ) a(\varphi) a ( φ ) and a ( ψ ) a(\psi) a ( ψ ) are jointly satisfiable, then so are z ( φ ) z(\varphi) z ( φ ) and z ( ψ ) z(\psi) z ( ψ ) .
Definition 2.1 (Guidance relation). Let S = 2 K S = 2^K S = 2 K . The guidance relation → L \to_L → L associated with any codex L L L is a binary relation → L ⊆ S × S \to_L \subseteq S \times S → L ⊆ S × S defined as follows: s → L s ′ s \to_L s' s → L s ′ iff for every variable v ∈ P r o p v \in\mathrm{Prop} v ∈ Prop such that s ′ ( v ) ≠ s ( v ) s'(v) \ne s(v) s ′ ( v ) = s ( v ) , a rule φ ∈ L \varphi\in L φ ∈ L exists such that:
z ( φ ) ∈ { v , ¬ v } z(\varphi) \in\{v, \neg v\} z ( φ ) ∈ { v , ¬ v } ;
s ⊨ a ( φ ) s \models a(\varphi) s ⊨ a ( φ ) and s ′ ⊨ a ( φ ) s' \models a(\varphi) s ′ ⊨ a ( φ ) , and
s ′ ⊨ z ( φ ) s' \models z(\varphi) s ′ ⊨ z ( φ ) .
Hence, if s → L s ′ s \to_L s' s → L s ′ , then there exists at least one rule φ ∈ L \varphi\in L φ ∈ L s.t. s ⊭ φ s \not\models\varphi s ⊨ φ and s ′ ⊨ φ s' \models\varphi s ′ ⊨ φ , where the latter satisfaction is non-trivial , since s ′ ⊨ a ( φ ) s' \models a(\varphi) s ′ ⊨ a ( φ ) . Codices L L L consisting only of complete rules force their associated guidance relations to be subsets of the neighbourhood relation N ⊆ S × S N \subseteq S \times S N ⊆ S × S s.t. s N s ′ sNs' s N s ′ iff some x ∈ V x \in V x ∈ V exists s.t. s ( v ) = s ′ ( v ) s(v) = s'(v) s ( v ) = s ′ ( v ) for any v ∈ V ∖ { x } v \in V \setminus\{x\} v ∈ V ∖ { x } . For any two such neighbors s s s and s ′ s' s ′ , consider the following complete rule
φ ( s , s ′ ) : = ⋀ v ∈ Prop ∖ { x } φ v → φ x . (1) \varphi(s,s') := \bigwedge_{v \in\operatorname{Prop} \setminus\{x\}} \varphi_v \to\varphi_x. \tag*{(1)} φ ( s , s ′ ) := v ∈ Prop ∖ { x } ⋀ φ v → φ x . (1)
Thus by construction, s s s is the only profile which falsifies φ ( s , s ′ ) \varphi(s,s') φ ( s , s ′ ) , since every other profile would either verify its conclusion or falsify its premises, and s ′ s' s ′ is the unique profile which nontrivially satisfies φ ( s , s ′ ) \varphi(s,s') φ ( s , s ′ ) , since s ′ ⊨ a ( φ ( s , s ′ ) ) s' \models a(\varphi(s,s')) s ′ ⊨ a ( φ ( s , s ′ )) . Moreover, s → L s ′ s \to_L s' s → L s ′ for any codex L L L that contains φ ( s , s ′ ) \varphi(s,s') φ ( s , s ′ ) . For any codex L L L , let
T ( L ) : = { s ∈ S ∣ s ⊨ φ for all φ ∈ L } and P ( L ) = { s ∈ S ∣ s → L s ′ for some s ′ ∈ T ( L ) } . T(L) := \{s \in S \mid s \models\varphi\text{ for all } \varphi\in L\}
\quad\text{and}\quad
P(L) = \{s \in S \mid s \to_L s' \text{ for some } s' \in T(L)\}. T ( L ) := { s ∈ S ∣ s ⊨ φ for all φ ∈ L } and P ( L ) = { s ∈ S ∣ s → L s ′ for some s ′ ∈ T ( L )} .
That is, T ( L ) T(L) T ( L ) is the set of profiles that, whenever declared, the Listener would accept, and P ( L ) P(L) P ( L ) is the set of profiles from which the Speaker can be ‘guided’ by L L L to some state which can persuade the Listener. By definition, T ( L ) ⊆ P ( L ) T(L) \subseteq P(L) T ( L ) ⊆ P ( L ) .
For any set A A A of states, A A A is implementable if A = P ( L ) A = P(L) A = P ( L ) for some codex L L L . The set A A A is truthfully implementable if A = T ( L ) = P ( L ) A = T(L) = P(L) A = T ( L ) = P ( L ) for some codex L L L .
Let T ( → L ) = { s ∈ S ∣ ∀ t ( s → L t ⇒ s = t ) } T(\to_L) = \{s \in S \mid\forall t(s \to_L t \Rightarrow s = t)\} T ( → L ) = { s ∈ S ∣ ∀ t ( s → L t ⇒ s = t )} .
Lemma 2.2 ([8 ] , Lemma 2). For any codex L L L , T ( L ) = T ( → L ) T(L) = T(\to_L) T ( L ) = T ( → L ) .
Proof. If s ∉ T ( L ) s \notin T(L) s ∈ / T ( L ) , then s ⊭ φ s \not\models\varphi s ⊨ φ for some φ ∈ L \varphi\in L φ ∈ L ; that is, s ⊨ a ( φ ) s \models a(\varphi) s ⊨ a ( φ ) and s ⊭ z ( φ ) = φ x s \not\models z(\varphi) = \varphi_x s ⊨ z ( φ ) = φ x for some x ∈ Prop x \in\operatorname{Prop} x ∈ Prop . Thus, s → L s ′ s \to_L s' s → L s ′ where s ′ s' s ′ is the profile that differs from s s s only in the truth value of the variable x x x , which shows that s ∉ T ( → L ) s \notin T(\to_L) s ∈ / T ( → L ) . Conversely, let s ∉ T ( → L ) s \notin T(\to_L) s ∈ / T ( → L ) , i.e. s → L t s \to_L t s → L t for some profile t ≠ s t \ne s t = s . By definition of → L \to_L → L , this implies that s ⊭ φ s \not\models\varphi s ⊨ φ for some φ ∈ L \varphi\in L φ ∈ L ; therefore, s ∉ T ( L ) s \notin T(L) s ∈ / T ( L ) . □
Definition 2.3 (Canonical and quasi-canonical codices). A codex L L L is canonical if:
L L L consists of complete rules .
For every s s s , there is at most one t ≠ s t \ne s t = s such that s → L t s \to_L t s → L t .
For every s ∈ P ( L ) ∖ T ( L ) s \in P(L) \setminus T(L) s ∈ P ( L ) ∖ T ( L ) , there is some r ∈ S ∖ P ( L ) r \in S \setminus P(L) r ∈ S ∖ P ( L ) such that r → L s r \to_L s r → L s .
A codex L L L is quasi-canonical if it satisfies the conditions in the first two items above.6
Proposition 2.4 ([8 ] , Proposition 1). If A ⊆ S A \subseteq S A ⊆ S is truthfully implementable, then it is truthfully implementable by a canonical codex.
We conclude this subsection by reporting on some examples discussed in [8 ] . Let Prop = { v 1 , v 2 , v 3 } \operatorname{Prop}=\{v_1,v_2,v_3\} Prop = { v 1 , v 2 , v 3 } . The codex
L = { v 1 ∧ v 2 → v 3 , v 2 ∧ ¬ v 3 → v 1 , ¬ v 1 ∧ ¬ v 2 → ¬ v 3 , ¬ v 1 ∧ v 2 → ¬ v 3 , ¬ v 2 ∧ v 3 → ¬ v 1 , v 1 ∧ v 3 → ¬ v 2 } L=\{v_1\wedge v_2\to v_3,\ v_2\wedge\neg v_3\to v_1,\ \neg v_1\wedge\neg v_2\to\neg v_3,\ \neg v_1\wedge v_2\to\neg v_3,\ \neg v_2\wedge v_3\to\neg v_1,\ v_1\wedge v_3\to\neg v_2\} L = { v 1 ∧ v 2 → v 3 , v 2 ∧ ¬ v 3 → v 1 , ¬ v 1 ∧ ¬ v 2 → ¬ v 3 , ¬ v 1 ∧ v 2 → ¬ v 3 , ¬ v 2 ∧ v 3 → ¬ v 1 , v 1 ∧ v 3 → ¬ v 2 }
(cf. [8 Section VIII] ) is canonical and s.t. P ( L ) = { ( 0 , 0 , 0 ) , ( 1 , 0 , 0 ) , ( 0 , 0 , 1 ) } P(L)=\{(0,0,0),(1,0,0),(0,0,1)\} P ( L ) = {( 0 , 0 , 0 ) , ( 1 , 0 , 0 ) , ( 0 , 0 , 1 )} , and T ( L ) = { ( 0 , 0 , 0 ) , ( 1 , 0 , 0 ) } T(L)=\{(0,0,0),(1,0,0)\} T ( L ) = {( 0 , 0 , 0 ) , ( 1 , 0 , 0 )} . Below, a diagram of the non reflexive edges of its corresponding guidance relation → L \to_L → L ; in the picture, T ( L ) T(L) T ( L ) is shaded in blue and P ( L ) P(L) P ( L ) in red.
Figure 2.
The following codices (cf. [8 Example 1, Section III] ) have the same T ( − ) T(-) T ( − ) -set but different P ( − ) P(-) P ( − ) -sets:
L 1 = { v 1 → v 2 , ¬ v 1 → ¬ v 2 , v 1 → v 3 , ¬ v 1 → ¬ v 3 } L 2 = { v 1 → v 2 , ¬ v 1 → ¬ v 2 , v 2 → v 3 , ¬ v 2 → ¬ v 3 } L 3 = { ¬ v i ∧ v i + 1 → v i + 2 , v i ∧ ¬ v i + 1 → ¬ v i + 2 ∣ i ∈ Z 3 } \begin{aligned}
L_1&=\{v_1\to v_2,\ \neg v_1\to\neg v_2,\ v_1\to v_3,\ \neg v_1\to\neg v_3\}\\
L_2&=\{v_1\to v_2,\ \neg v_1\to\neg v_2,\ v_2\to v_3,\ \neg v_2\to\neg v_3\}\\
L_3&=\{\neg v_i\wedge v_{i+1}\to v_{i+2},\ v_i\wedge\neg v_{i+1}\to\neg v_{i+2}\mid i\in\mathbb{Z}_3\}
\end{aligned} L 1 L 2 L 3 = { v 1 → v 2 , ¬ v 1 → ¬ v 2 , v 1 → v 3 , ¬ v 1 → ¬ v 3 } = { v 1 → v 2 , ¬ v 1 → ¬ v 2 , v 2 → v 3 , ¬ v 2 → ¬ v 3 } = { ¬ v i ∧ v i + 1 → v i + 2 , v i ∧ ¬ v i + 1 → ¬ v i + 2 ∣ i ∈ Z 3 }
Below, the diagrams of the non reflexive edges of their corresponding guidance relations:
Figure 3.
The set A = { ( 1 , 0 , 0 ) , ( 0 , 1 , 0 ) , ( 0 , 0 , 1 ) } A = \{(1,0,0),(0,1,0),(0,0,1)\} A = {( 1 , 0 , 0 ) , ( 0 , 1 , 0 ) , ( 0 , 0 , 1 )} , represented below together with the non-reflexive edges of the neighbouring relation, is not implementable (cf. [8 Example 2, Section III] ):
Figure 4.
Indeed, assuming for contradiction that A = P ( L ) A = P(L) A = P ( L ) for some codex L L L ,
Case 1: if T ( L ) = P ( L ) = A T(L) = P(L) = A T ( L ) = P ( L ) = A , i.e. A A A is truthfully implementable, then by Proposition 2.4 we can assume w.l.o.g. that L L L is canonical, and since s 0 = ( 0 , 0 , 0 ) ∉ A = T ( L ) s_0 = (0,0,0) \notin A = T(L) s 0 = ( 0 , 0 , 0 ) ∈ / A = T ( L ) , there is a complete rule φ ∈ L \varphi\in L φ ∈ L s.t. s 0 ⊭ φ s_0 \not\models\varphi s 0 ⊨ φ ; this violation implies that s 0 → L s s_0 \to_L s s 0 → L s for some s ≠ s 0 s \ne s_0 s = s 0 and, since the rules in L L L are complete and hence → L \to_L → L is contained in the neighbourhood relation N N N , s ∈ A s \in A s ∈ A . Hence, s 0 ∈ P ( L ) = A s_0 \in P(L) = A s 0 ∈ P ( L ) = A , which contradicts s 0 ∉ A s_0 \notin A s 0 ∈ / A .
Case 2: if T ( L ) ⊆ A = P ( L ) T(L) \subseteq A = P(L) T ( L ) ⊆ A = P ( L ) , then one of the profiles in A = P ( L ) A = P(L) A = P ( L ) , w.l.o.g. s = ( 0 , 0 , 1 ) s = (0,0,1) s = ( 0 , 0 , 1 ) , is not in T ( L ) T(L) T ( L ) . Since ( 0 , 0 , 1 ) ∈ A = P ( L ) (0,0,1) \in A = P(L) ( 0 , 0 , 1 ) ∈ A = P ( L ) , there must be another profile in A A A , w.l.o.g. s ′ = ( 0 , 1 , 0 ) s' = (0,1,0) s ′ = ( 0 , 1 , 0 ) , such that ( 0 , 0 , 1 ) → L ( 0 , 1 , 0 ) (0,0,1) \to_L (0,1,0) ( 0 , 0 , 1 ) → L ( 0 , 1 , 0 ) . This requires that ¬ v 1 → v 2 \neg v_1 \to v_2 ¬ v 1 → v 2 (and ¬ v 1 → ¬ v 3 \neg v_1 \to\neg v_3 ¬ v 1 → ¬ v 3 ) be in the codex; indeed, v 1 v_1 v 1 is the only variable on which s s s and s ′ s' s ′ agree. However, in that case, ( 0 , 0 , 0 ) → L ( 0 , 1 , 0 ) ∈ T ( L ) (0,0,0) \to_L (0,1,0) \in T(L) ( 0 , 0 , 0 ) → L ( 0 , 1 , 0 ) ∈ T ( L ) and therefore ( 0 , 0 , 0 ) ∈ P ( L ) = A (0,0,0) \in P(L) = A ( 0 , 0 , 0 ) ∈ P ( L ) = A although ( 0 , 0 , 0 ) ∉ A (0,0,0) \notin A ( 0 , 0 , 0 ) ∈ / A , a contradiction.
Although the above set A A A is not implementable, its complement is. Let A c = S ∖ A A^c = S \setminus A A c = S ∖ A . Consider the codex L c L^c L c that consists of the three rules v i → v j v_i \to v_j v i → v j , where j ≠ i + 1 j \ne i+1 j = i + 1 ; i.e.:
L c = { v 1 → v 3 , v 2 → v 1 , v 3 → v 2 } L^c = \{v_1 \to v_3, v_2 \to v_1, v_3 \to v_2\} L c = { v 1 → v 3 , v 2 → v 1 , v 3 → v 2 }
The picture below represents the diagram of the non-reflexive edges of the guidance relation → L \to_{L} → L ; the sets T ( L c ) T(L^c) T ( L c ) and P ( L c ) P(L^c) P ( L c ) are shaded in blue and red, respectively.
Figure 1 (PDF p. 10).
Basic formalism in `Magical implementation'
In [10 ] , persuasion situations are analyzed in which (at least) two Speakers are in a dispute over a single indivisible prize, which the Listener will assign. The set S S S (usually S = 2 K S = 2^K S = 2 K with K ≥ 3 K \ge3 K ≥ 3 ) is partitioned into subsets W i W_i W i where W i = { s ∈ S ∣ i is assigned the prize if i declares s } W_i = \{s \in S \mid i\text{ is assigned the prize if }i\text{ declares }s\} W i = { s ∈ S ∣ i is assigned the prize if i declares s } denotes the winning set of Speaker i i i .
Definition 2.5. An implementation problem is a tuple ( S , { W i ∣ i ∈ I } ) (S, \{W_i \mid i \in I\}) ( S , { W i ∣ i ∈ I }) such that S S S is a set and { W i ∣ i ∈ I } \{W_i \mid i \in I\} { W i ∣ i ∈ I } is a partition of S S S .
Since { W i ∣ i ∈ I } \{W_i \mid i \in I\} { W i ∣ i ∈ I } is a partition of S S S , the Listener can assume that exactly one Speaker is truthful, namely the unique i i i such that the true state t ∈ W i t \in W_i t ∈ W i . To identify which Speaker is truthful, the Listener constructs a codex L i L_i L i for each Speaker i i i ; this codex encodes the circumstances in which i i i deserves to receive the prize. Speaker i i i declares a state s ∈ T ( L i ) s \in T(L_i) s ∈ T ( L i ) , truthfully or not (hence, P ( L i ) = S P(L_i) = S P ( L i ) = S ). It is assumed that, if i i i ’s declaration is not truthful, the state declared by i i i is ‘suggested’ by → L i \to_{L_i} → L i , i.e. i i i declares some s ∈ S s \in S s ∈ S s.t. t → L i s t \to_{L_i} s t → L i s .
Definition 2.6. The codices { L i ∣ i ∈ I } \{L_i \mid i \in I\} { L i ∣ i ∈ I } magically implement ( S , { W i ∣ i ∈ I } ) (S, \{W_i \mid i \in I\}) ( S , { W i ∣ i ∈ I }) iff:
T ( L i ) = W i T(L_i) = W_i T ( L i ) = W i for each i i i ;
For all s , t ∈ S s,t \in S s , t ∈ S , if s → L i t s \to_{L_i} t s → L i t then t ̸ → L j s t \not\to_{L_j} s t → L j s for all j ≠ i j \ne i j = i .
To illustrate the application of the one-way cheating principle in action, in [10 ] , the implementation problem ( S , W P r o , W C o n ) (S, W_{Pro}, W_{Con}) ( S , W P r o , W C o n ) is discussed, in which S = 2 P r o p S = 2^{Prop} S = 2 P r o p for P r o p = { v 1 , v 2 , v 3 } Prop = \{v_1, v_2, v_3\} P r o p = { v 1 , v 2 , v 3 } , and W P r o = { ( 1 , 0 , 0 ) , ( 0 , 1 , 0 ) , ( 0 , 0 , 1 ) } 7 W_{Pro} = \{(1,0,0), (0,1,0), (0,0,1)\}^{7} W P r o = {( 1 , 0 , 0 ) , ( 0 , 1 , 0 ) , ( 0 , 0 , 1 ) } 7 and W C o n = S ∖ W P r o W_{Con} = S \setminus W_{Pro} W C o n = S ∖ W P r o . It is easy to verify that the codices in the table below magically implement ( S , W P r o , W C o n ) (S, W_{Pro}, W_{Con}) ( S , W P r o , W C o n ) .7
Table 1.
Notice that each codex is quasi-canonical but not canonical (cf. Definition 2.3). The following diagrams represent (the non reflexive edges of) the guidance relation associated with each codex; W P r o W_{Pro} W P r o is shaded in red and W C o n W_{Con} W C o n in blue.
Figure 5 (PDF p. 11).
It is assumed that the Listener knows the cheating procedure applied by each Speaker i i i . If the true state t ∉ W i t \notin W_i t ∈ / W i , then i i i will search for a state s ∈ T ( L i ) s \in T(L_i) s ∈ T ( L i ) by advancing along a path which starts at t t t following the guidance relation → L i \to_{L_i} → L i . For any Speaker i i i , each rule in L i L_i L i has two effects: it excludes one state in S ∖ W i S \setminus W_i S ∖ W i and directs Speaker i i i to a different (unique) state. For example, the rule v 1 ∧ ¬ v 2 → ¬ v 3 v_1 \land\neg v_2 \to\neg v_3 v 1 ∧ ¬ v 2 → ¬ v 3 on P r o Pro P r o ’s list is violated by 101, and if 101 happens to be the true state, it guides P r o Pro P r o to declare 100.
To see the one-way cheating principle at work, assume, for example, that the true state is t ∈ { 111 , 110 , 101 } ⊆ W C o n t \in\{111, 110, 101\} \subseteq W_{Con} t ∈ { 111 , 110 , 101 } ⊆ W C o n . Since t ∈ W C o n t \in W_{Con} t ∈ W C o n , agent C o n Con C o n will truthfully declare t t t , while, being iteratively guided by → L P r o \to_{L_{Pro}} → L P r o , agent P r o Pro P r o declares 100. Hence, upon receiving e.g. 110 from C o n Con C o n and 100 from P r o Pro P r o , the Listener concludes that Con is the truthful one, since 100 is the state Pro would declare in case the true state was 110, while, if 100 were the true state, Con would have declared 000 instead of 110.
Another example of implementation problem mentioned in [10 Section 3] is ( S , W Y , W Z ) (S, W_Y, W_Z) ( S , W Y , W Z ) with S = W Y ∪ W Z S = W_Y \cup W_Z S = W Y ∪ W Z and W Z = { z 1 , z 2 , z 3 } W_Z = \{z_1,z_2,z_3\} W Z = { z 1 , z 2 , z 3 } and W Y = { y 1 , … , y 5 } W_Y = \{y_1,\ldots,y_5\} W Y = { y 1 , … , y 5 } . Interestingly, this example does not involve codices, and demonstrates that problems such as ( S , W Y , W Z ) (S, W_Y, W_Z) ( S , W Y , W Z ) can still be magically implemented by pairs ( → Y , → Z ) (\to_Y,\to_Z) ( → Y , → Z ) of binary relations on S S S , provided these relations satisfy conditions analogous to those of Definition 2.6 , i.e.:^8
T ( → U ) : = { s ∈ S ∣ ∀ t ( s → U t ⇒ s = t ) } = W U T(\to_U) := \{s \in S \mid\forall t(s \to_U t \Rightarrow s = t)\} = W_U T ( → U ) := { s ∈ S ∣ ∀ t ( s → U t ⇒ s = t )} = W U for any U ∈ { Y , Z } U \in\{Y,Z\} U ∈ { Y , Z } , and
for all U , V ∈ { Y , Z } U,V \in\{Y,Z\} U , V ∈ { Y , Z } s.t. U ≠ V U \ne V U = V and all s , t ∈ S s,t \in S s , t ∈ S , s → U t s \to_U t s → U t implies t ̸ → V s t \not\to_V s t → V s .
In the case of this example, it can readily be verified by inspection that the relations → Y \to_Y → Y and → Z \to_Z → Z represented below do satisfy these conditions.9 Hence, if Y Y Y and Z Z Z are guided by these relations, the Listener can establish who the truthful Speaker is whichever state in S S S is the true one. For instance, if Y Y Y declares z 1 z_1 z 1 and Z Z Z declares y 2 y_2 y 2 , then the Listener concludes that Y Y Y is the truthful one, since y 2 y_2 y 2 is the state Z Z Z would declare if t = z 1 t = z_1 t = z 1 , while if t = y 2 t = y_2 t = y 2 , then Y Y Y would have declared z 2 z_2 z 2 , and similarly for the remaining situations.
The one-way cheating principle can be understood as a particular instance of the Tiglath-Pileser principle discussed in the introduction. Indeed, the latter principle provides a general epistemic criterion for evaluating untrustworthy declarations by reconstructing the set of states that could have generated them under constrained distortion. The one-way cheating principle adapts and applies this criterion to a multi-agent setting in which the distortion mechanisms (guidance relations) are sufficiently structured so as to allow, under favourable circumstances, to uniquely identify the true state and the truthful speaker. In § 4.2 , we will suitably generalize the magical implementation framework to formally account for a wider-scoped applicability of the Tiglath-Pileser principle.
Figure 6 (PDF p. 13).
More on codices and guidance relations
In the present section, we further investigate the two-way relationship between codices and (guidance) relations. Specifically, in Section 3.1 we intrinsically characterize the class of guidance relations associated with quasi-canonical codices; in Section 3.2, we introduce and study a class of codices which will be relevant to the analysis of Section 5; in Section 3.3, we discuss a construction which allows us to recast a problem magically implementable via relations (cf. discussion at the end of Section 2.2) as a problem magically implementable via codices.
Quasi-canonical codices and guiding relations
Definition 3.1. Let S = 2 K S = 2^K S = 2 K . A reflexive and antisymmetric binary relation → \to → on S S S is a guiding relation if it satisfies the following conditions:
→ ⊆ N \to\subseteq N →⊆ N ;
for every s ∈ S s \in S s ∈ S , there is at most one s ′ ∈ S s' \in S s ′ ∈ S s.t. s ≠ s ′ s \ne s' s = s ′ and s → s ′ s \to s' s → s ′ .
We let T ( → ) : = { s ∈ S ∣ ∀ t ( s → t ⇒ s = t ) } T(\to) := \{s \in S \mid\forall t(s \to t \Rightarrow s = t)\} T ( → ) := { s ∈ S ∣ ∀ t ( s → t ⇒ s = t )} , and for any s ∈ S ∖ T ( → ) s \in S \setminus T(\to) s ∈ S ∖ T ( → ) , we let s → ∈ S s_{\to} \in S s → ∈ S denote the unique profile s ′ s' s ′ satisfying item 2 above.
As we will see next, the notion of guiding relation is a characterization of the guidance relations associated with quasi-canonical codices, formulated without any reference to codices. Examples of guiding relations occur in every magical implementation scenario discussed in Section 2.2.
Lemma 3.2. If L L L is a quasi-canonical codex, then → L \to_L → L is a guiding relation.
Proof. By construction, → L \to_L → L is reflexive; item 1 of Definition 3.1 follows from the rules of L L L being complete; item 2 immediately follows from item 2 of the definition of pre-canonical codex. To complete the proof, let us show contrapositively that the coherence of L L L implies that → L \to_L → L is antisymmetric. Let s , t ∈ S s,t \in S s , t ∈ S s.t. s ≠ t s \ne t s = t and s → L t s \to_L t s → L t and t → L s t \to_L s t → L s ; then, s , t ∉ T ( L ) s,t \notin T(L) s , t ∈ / T ( L ) and s N t sNt s N t ; hence, some φ , ψ ∈ L \varphi,\psi\in L φ , ψ ∈ L exist s.t.
s ⊨ a ( φ ) , t ⊨ a ( φ ) , t ⊨ z ( φ ) , s ⊭ z ( φ ) s \models a(\varphi),\ t \models a(\varphi),\ t \models z(\varphi),\ s \not\models z(\varphi) s ⊨ a ( φ ) , t ⊨ a ( φ ) , t ⊨ z ( φ ) , s ⊨ z ( φ )
t ⊨ a ( ψ ) , s ⊨ a ( ψ ) , s ⊨ z ( ψ ) , t ⊭ z ( ψ ) . t \models a(\psi),\ s \models a(\psi),\ s \models z(\psi),\ t \not\models z(\psi). t ⊨ a ( ψ ) , s ⊨ a ( ψ ) , s ⊨ z ( ψ ) , t ⊨ z ( ψ ) .
Hence, φ \varphi φ and ψ \psi ψ are s.t. a ( φ ) a(\varphi) a ( φ ) and a ( ψ ) a(\psi) a ( ψ ) are jointly satisfiable, and since they are complete rules, from the conditions above it also follows that z ( φ ) z(\varphi) z ( φ ) and z ( ψ ) z(\psi) z ( ψ ) share the one variable occurring in them, and are hence not jointly satisfiable, which proves that L L L is not coherent. □
As discussed above, Definition 3.1 implies that s → s_{\to} s → exists and s N s → sNs_{\to} s N s → for any s ∈ S ∖ T ( → ) s \in S \setminus T(\to) s ∈ S ∖ T ( → ) , and hence φ ( s , s → ) \varphi(s,s_{\to}) φ ( s , s → ) (see (1) ) is well defined. Let
L → : = { φ ( s , s → ) ∣ s ∈ S ∖ T ( → ) } . (2) L_{\to}:=\{\varphi(s,s_{\to}) \mid s \in S \setminus T(\to)\}. \tag*{(2)} L → := { φ ( s , s → ) ∣ s ∈ S ∖ T ( → )} . (2)
Lemma 3.3. If → \to → is a guiding relation, then L → L_{\to} L → is a quasi-canonical codex.
Proof. By construction, L → L_{\to} L → consists of complete rules, and for every s ∈ S s \in S s ∈ S there is at most one t ≠ s t \ne s t = s such that s → L → t s \to_{L_{\to}} t s → L → t , which shows that L → L_{\to} L → is pre-canonical. To finish the proof, arguing contrapositively, let us assume that L → L_{\to} L → is not coherent, and let φ , ψ ∈ L → \varphi,\psi\in L_{\to} φ , ψ ∈ L → s.t. a ( φ ) a(\varphi) a ( φ ) and a ( ψ ) a(\psi) a ( ψ ) are jointly satisfiable while z ( φ ) z(\varphi) z ( φ ) and z ( ψ ) z(\psi) z ( ψ ) are not. Let s , s ′ ∈ S ∖ T ( → ) s,s' \in S \setminus T(\to) s , s ′ ∈ S ∖ T ( → ) s.t. φ = φ ( s , s → ) \varphi=\varphi(s,s_{\to}) φ = φ ( s , s → ) and ψ = φ ( s ′ , s → ′ ) \psi=\varphi(s',s'_{\to}) ψ = φ ( s ′ , s → ′ ) , with
s ⊨ a ( φ ) , s → ⊨ a ( φ ) , s → ⊨ z ( φ ) , s ⊭ z ( φ ) s \models a(\varphi),\ s_{\to} \models a(\varphi),\ s_{\to} \models z(\varphi),\ s \not\models z(\varphi) s ⊨ a ( φ ) , s → ⊨ a ( φ ) , s → ⊨ z ( φ ) , s ⊨ z ( φ ) s ′ ⊨ a ( ψ ) , s → ′ ⊨ a ( ψ ) , s → ′ ⊨ z ( ψ ) , s ′ ⊭ z ( ψ ) . s' \models a(\psi),\ s'_{\to} \models a(\psi),\ s'_{\to} \models z(\psi),\ s' \not\models z(\psi). s ′ ⊨ a ( ψ ) , s → ′ ⊨ a ( ψ ) , s → ′ ⊨ z ( ψ ) , s ′ ⊨ z ( ψ ) . Since φ \varphi φ and ψ \psi ψ are complete rules, a ( φ ) a(\varphi) a ( φ ) and a ( ψ ) a(\psi) a ( ψ ) each contain K − 1 K-1 K − 1 variables; since z ( φ ) z(\varphi) z ( φ ) and z ( ψ ) z(\psi) z ( ψ ) are not jointly satisfiable, it follows w.l.o.g. that z ( φ ) = v z(\varphi)=v z ( φ ) = v and z ( ψ ) = ¬ v z(\psi)=\neg v z ( ψ ) = ¬ v for some v ∈ P r o p v \in\mathrm{Prop} v ∈ Prop , hence a ( φ ) a(\varphi) a ( φ ) and a ( ψ ) a(\psi) a ( ψ ) have exactly the same variables, and hence the assumption that they are jointly satisfiable implies that a ( φ ) = a ( ψ ) a(\varphi)=a(\psi) a ( φ ) = a ( ψ ) . Moreover, s ⊭ z ( φ ) s \not\models z(\varphi) s ⊨ z ( φ ) iff s ⊨ z ( ψ ) s \models z(\psi) s ⊨ z ( ψ ) and s ′ ⊭ z ( ψ ) s' \not\models z(\psi) s ′ ⊨ z ( ψ ) iff s ′ ⊨ z ( φ ) s' \models z(\varphi) s ′ ⊨ z ( φ ) , which imply that s = s → ′ s=s'_{\to} s = s → ′ and s ′ = s → s'=s_{\to} s ′ = s → , and hence s → s ′ s \to s' s → s ′ and s ′ → s s' \to s s ′ → s and s ≠ s ′ s \ne s' s = s ′ , showing that → \to → is not antisymmetric. □
Recall that T ( L ) = T ( → L ) T(L)=T(\to_L) T ( L ) = T ( → L ) for any codex L L L (cf. Lemma 2.2). The next proposition embeds this identity into a perfect correspondence between quasi-canonical codices and guiding relations.
Proposition 3.4. For any guiding relation → \to → and any quasi-canonical codex L L L ,
T ( → ) = T ( L → ) T(\to)=T(L_{\to}) T ( → ) = T ( L → ) and → L → = → \to_{L_{\to}}=\to → L → =→ ;
L = L → L L=L_{\to_L} L = L → L .
Proof. 1. By construction, s ⊭ φ ( s , s → ) s \not\models\varphi(s,s_{\to}) s ⊨ φ ( s , s → ) for every s ∈ S ∖ T ( → ) s \in S \setminus T(\to) s ∈ S ∖ T ( → ) , which shows contrapositively that T ( L → ) ⊆ T ( → ) T(L_{\to}) \subseteq T(\to) T ( L → ) ⊆ T ( → ) . Conversely, if t ∈ T ( → ) t \in T(\to) t ∈ T ( → ) , then t ≠ s t \ne s t = s for every s ∈ S ∖ T ( → ) s \in S \setminus T(\to) s ∈ S ∖ T ( → ) , and so, fixing one such s s s , s ( v ) ≠ t ( v ) s(v) \ne t(v) s ( v ) = t ( v ) for at least one v ∈ V v \in V v ∈ V . Then either t ⊭ a ( φ ( s , s → ) ) t \not\models a(\varphi(s,s_{\to})) t ⊨ a ( φ ( s , s → )) , and hence t ⊨ φ ( s , s → ) t \models\varphi(s,s_{\to}) t ⊨ φ ( s , s → ) , or t ⊨ a ( φ ( s , s → ) ) t \models a(\varphi(s,s_{\to})) t ⊨ a ( φ ( s , s → )) , and hence t = s → ⊨ φ ( s , s → ) t = s_{\to} \models\varphi(s,s_{\to}) t = s → ⊨ φ ( s , s → ) . This proves the first identity. As to the second identity, since the two relations are reflexive, it is enough to show that if s , t ∈ S s,t \in S s , t ∈ S and s ≠ t s \ne t s = t , then s → L → t s \to_{L_{\to}} t s → L → t iff s → t s \to t s → t . Let us assume that s → L → t s \to_{L_{\to}} t s → L → t ; since s ≠ t s \ne t s = t , to show that s → t s \to t s → t , it is enough to show that t = s → t = s_{\to} t = s → . Let v ∈ V v \in V v ∈ V s.t. s ( v ) ≠ t ( v ) s(v) \ne t(v) s ( v ) = t ( v ) , and let φ \varphi φ be the witness rule for v v v , i.e. s ⊨ a ( φ ) s \models a(\varphi) s ⊨ a ( φ ) , t ⊨ a ( φ ) t \models a(\varphi) t ⊨ a ( φ ) and t ⊨ z ( φ ) ∈ { v , ¬ v } t \models z(\varphi) \in\{v,\neg v\} t ⊨ z ( φ ) ∈ { v , ¬ v } . Hence, s ( v ) ≠ t ( v ) s(v) \ne t(v) s ( v ) = t ( v ) implies that s ⊭ φ s \not\models\varphi s ⊨ φ , and since, by construction, each rule in L → L_{\to} L → is of the form φ ( s ′ , s → ′ ) \varphi(s',s'_{\to}) φ ( s ′ , s → ′ ) for some s ′ ∈ S ∖ T ( → ) s' \in S \setminus T(\to) s ′ ∈ S ∖ T ( → ) and can only be falsified at s ′ s' s ′ , it follows that φ = φ ( s , s → ) \varphi= \varphi(s,s_{\to}) φ = φ ( s , s → ) . Moreover, since φ ( s , s → ) \varphi(s,s_{\to}) φ ( s , s → ) is complete and s ⊨ a ( φ ) s \models a(\varphi) s ⊨ a ( φ ) , and t ⊨ a ( φ ) t \models a(\varphi) t ⊨ a ( φ ) , then v v v is the only variable on which s s s and t t t disagree, and moreover, t ( v ) = s → ( v ) t(v) = s_{\to}(v) t ( v ) = s → ( v ) , which implies that t = s → t = s_{\to} t = s → , as required. Conversely, if s → t s \to t s → t and s ≠ t s \ne t s = t , then s ∈ S ∖ T ( → ) s \in S \setminus T(\to) s ∈ S ∖ T ( → ) and t = s → t = s_{\to} t = s → , hence the rule φ ( s , s → ) \varphi(s,s_{\to}) φ ( s , s → ) witnesses that s → L → t s \to_{L_{\to}} t s → L → t for the unique variable on which s s s and t t t differ.
By definition and Lemma 2.2, L → L = { φ ( s , s → L ) ∣ s ∈ S ∖ T ( → L ) } = { φ ( s , s → L ) ∣ s ∈ S ∖ T ( L ) } L_{\to_L} = \{\varphi(s,s_{\to_L}) \mid s \in S \setminus T(\to_L)\} = \{\varphi(s,s_{\to_L}) \mid s \in S \setminus T(L)\} L → L = { φ ( s , s → L ) ∣ s ∈ S ∖ T ( → L )} = { φ ( s , s → L ) ∣ s ∈ S ∖ T ( L )} .
Notice preliminarily that any complete rule φ \varphi φ coincides with the rule φ ( s , s ′ ) \varphi(s,s') φ ( s , s ′ ) (see (1)) for exactly two states s , s ′ ∈ S s,s' \in S s , s ′ ∈ S s.t. s ≠ s ′ s \ne s' s = s ′ and s N s ′ sNs' s N s ′ .
Let φ ∈ L \varphi\in L φ ∈ L ; the assumption that L L L is quasi-canonical implies that φ \varphi φ is a complete rule; therefore, let s , s ′ ∈ S s,s' \in S s , s ′ ∈ S be the unique states s.t. s ≠ s ′ s \ne s' s = s ′ , s N s ′ sNs' s N s ′ , and φ = φ ( s , s ′ ) \varphi= \varphi(s,s') φ = φ ( s , s ′ ) . Then s ⊨ a ( φ ) s \models a(\varphi) s ⊨ a ( φ ) , s ′ ⊨ a ( φ ) s' \models a(\varphi) s ′ ⊨ a ( φ ) , s ′ ⊨ z ( φ ) s' \models z(\varphi) s ′ ⊨ z ( φ ) and s ⊭ z ( φ ) s \not\models z(\varphi) s ⊨ z ( φ ) , which shows that s ∈ S ∖ T ( L ) s \in S \setminus T(L) s ∈ S ∖ T ( L ) , and s ′ = s → L s' = s_{\to_L} s ′ = s → L , completing the proof that φ ∈ L → L \varphi\in L_{\to_L} φ ∈ L → L . For the converse inclusion, let s ∈ S ∖ T ( → L ) = S ∖ T ( L ) s \in S \setminus T(\to_L) = S \setminus T(L) s ∈ S ∖ T ( → L ) = S ∖ T ( L ) ; hence, s ⊭ φ s \not\models\varphi s ⊨ φ for some complete rule φ ∈ L \varphi\in L φ ∈ L which, by the preliminary observation, coincides with φ ( s , s ′ ) \varphi(s,s') φ ( s , s ′ ) for a unique state s ′ ∈ S s' \in S s ′ ∈ S s.t. s ≠ s ′ s \ne s' s = s ′ and s N s ′ sNs' s N s ′ . Hence, s ′ = s → L s' = s_{\to_L} s ′ = s → L .
Although the codices mentioned in the definition of magical implementation (cf. Definition 2.6 ) are not required to be quasi-canonical in [10 ] , the defining properties of quasi-canonical codices are key to the identification of the truthful Speaker via the application of the one-way cheating principle in all the concrete examples discussed in [10 ] . Furthermore, as mentioned when discussing the last example of § 2.2 (see page 12), also in the context of the implementation problem ( S , W Y , W Z ) (S,W_Y,W_Z) ( S , W Y , W Z ) , in which the one-way cheating principle is discussed without reference to codices, the relations which allow for the applicability of the one-way cheating principle satisfy the defining properties10 of guiding relations which, as we have just seen, provide an intrinsic characterization of the guidance relations associated with quasi-canonical codices . In § 4.2 , we generalize the magical implementation framework to relations on arbitrary sets S S S (that is, beyond the context of sets S = 2 K S = 2^K S = 2 K of profiles).
Separated codices and their guidance relations
In the present section, we introduce a subclass of codices which will become useful in the developments of § 5 .
Definition 3.5 (Separated codex). A codex L L L is separated if two disjoint subsets A , Z ⊆ Prop A, Z \subseteq\operatorname{Prop} A , Z ⊆ Prop exist such that Var ( a ( φ ) ) ⊆ A \operatorname{Var}(a(\varphi)) \subseteq A Var ( a ( φ )) ⊆ A and Var ( z ( φ ) ) ⊆ Z \operatorname{Var}(z(\varphi)) \subseteq Z Var ( z ( φ )) ⊆ Z for all φ ∈ L \varphi\in L φ ∈ L .
The subclass of separated codices will play an important role in the analysis of § 5 , where a distinction between verifiable and non-verifiable variables is made in order to capture the fact that the speaker is constrained in their statements by the (partial) knowledge of the listener.
Proposition 3.6. If L L L is a separated codex, then → L \to_L → L is transitive.
Proof. Notice preliminarily that, by definition of → L \to_L → L , if s , s ′ ∈ S = 2 K s, s' \in S = 2^K s , s ′ ∈ S = 2 K and s → L s ′ s \to_L s' s → L s ′ then s s s and s ′ s' s ′ can only differ on variables occurring in consequents of rules of L L L , and since L L L is separated, the variables on which s s s and s ′ s' s ′ differ cannot occur in antecedents of rules, hence s s s and s ′ s' s ′ have to agree on a ( φ ) a(\varphi) a ( φ ) for all φ ∈ L \varphi\in L φ ∈ L . Let s , t , u ∈ S s, t, u \in S s , t , u ∈ S such that s → L t s \to_L t s → L t and t → L u t \to_L u t → L u . To show that s → L u s \to_L u s → L u , let v ∈ Prop v \in\operatorname{Prop} v ∈ Prop such that s ( v ) ≠ u ( v ) s(v) \ne u(v) s ( v ) = u ( v ) . Then either s ( v ) ≠ t ( v ) = u ( v ) s(v) \ne t(v) = u(v) s ( v ) = t ( v ) = u ( v ) or s ( v ) = t ( v ) ≠ u ( v ) s(v) = t(v) \ne u(v) s ( v ) = t ( v ) = u ( v ) . If s ( v ) ≠ t ( v ) = u ( v ) s(v) \ne t(v) = u(v) s ( v ) = t ( v ) = u ( v ) , then the assumption that s → L t s \to_L t s → L t implies that some φ ∈ L \varphi\in L φ ∈ L exists such that z ( φ ) ∈ { v , ¬ v } z(\varphi) \in\{v, \neg v\} z ( φ ) ∈ { v , ¬ v } , and s ⊨ a ( φ ) s \models a(\varphi) s ⊨ a ( φ ) , t ⊨ a ( φ ) t \models a(\varphi) t ⊨ a ( φ ) , s ⊭ z ( φ ) s \not\models z(\varphi) s ⊨ z ( φ ) , t ⊨ z ( φ ) t \models z(\varphi) t ⊨ z ( φ ) . As noted preliminarily, t → L u t \to_L u t → L u implies that u ⊨ a ( φ ) u \models a(\varphi) u ⊨ a ( φ ) . Since t ( v ) = u ( v ) t(v) = u(v) t ( v ) = u ( v ) , it follows that u ⊨ z ( φ ) u \models z(\varphi) u ⊨ z ( φ ) , as required.
If s ( v ) = t ( v ) ≠ u ( v ) s(v) = t(v) \ne u(v) s ( v ) = t ( v ) = u ( v ) , then the assumption that t → L u t \to_L u t → L u implies that some φ ∈ L \varphi\in L φ ∈ L exists such that z ( φ ) ∈ { v , ¬ v } z(\varphi) \in\{v, \neg v\} z ( φ ) ∈ { v , ¬ v } , and t ⊨ a ( φ ) t \models a(\varphi) t ⊨ a ( φ ) , u ⊨ a ( φ ) u \models a(\varphi) u ⊨ a ( φ ) , t ⊭ z ( φ ) t \not\models z(\varphi) t ⊨ z ( φ ) , u ⊨ z ( φ ) u \models z(\varphi) u ⊨ z ( φ ) . As noted preliminarily, s → L t s \to_L t s → L t implies that s ⊨ a ( φ ) s \models a(\varphi) s ⊨ a ( φ ) . Since s ( v ) = t ( v ) s(v) = t(v) s ( v ) = t ( v ) , it follows that s ⊭ z ( φ ) s \not\models z(\varphi) s ⊨ z ( φ ) , as required. □
From relations to codices
At the end of § 2.2 , we discussed a problem ( S , W Y , W Z ) (S, W_Y, W_Z) ( S , W Y , W Z ) which is magically implementable via relations rather than via codices. In this context, we also noticed (cf. Footnote 9) that the definition of magical implementation of [10 ] (cf. Definition 2.6) can equivalently be restated purely in terms of relations rather than in terms of codices. In this subsection, we illustrate, in the context of this same example, how ( S , W Y , W Z ) (S, W_Y, W_Z) ( S , W Y , W Z ) can be recast as a problem ( S P , W Y P , W Z P ) (S^P, W_Y^P, W_Z^P) ( S P , W Y P , W Z P ) implementable by means of codices, simply by letting the set S = { y 1 , … , y 5 , z 1 , z 2 , z 3 } S = \{y_1, \ldots, y_5, z_1, z_2, z_3\} S = { y 1 , … , y 5 , z 1 , z 2 , z 3 } , the domain of the original problem, be the set Prop \operatorname{Prop} Prop of proposition variables.
Under this stipulation, the relations → Y \to_Y → Y and → Z \to_Z → Z represented in the last picture of § 2.2 give rise to the following codices:
L Y : = { ¬ y j ∣ 1 ≤ j ≤ 5 } ∪ { y j → z i ∣ y j → Y z i } L Z : = { ¬ z j ∣ 1 ≤ j ≤ 3 } ∪ { z j → y i ∣ z j → Z y i } . L_Y := \{\neg y_j \mid1 \le j \le5\} \cup\{y_j \to z_i \mid y_j \to_Y z_i\} \qquad L_Z := \{\neg z_j \mid1 \le j \le3\} \cup\{z_j \to y_i \mid z_j \to_Z y_i\}. L Y := { ¬ y j ∣ 1 ≤ j ≤ 5 } ∪ { y j → z i ∣ y j → Y z i } L Z := { ¬ z j ∣ 1 ≤ j ≤ 3 } ∪ { z j → y i ∣ z j → Z y i } .
Let S P : = 2 Prop ≃ 2 8 S^P := 2^{\operatorname{Prop}} \simeq2^8 S P := 2 Prop ≃ 2 8 be the set of profiles. We assume w.l.o.g. that each s ∈ S P s \in S^P s ∈ S P is of the form y ‾ z ‾ \overline{y}\overline{z} y z , where y ‾ ∈ 2 5 \overline{y} \in2^5 y ∈ 2 5 (resp. z ‾ ∈ 2 3 \overline{z} \in2^3 z ∈ 2 3 ) is the tuple corresponding to the truth values of the proposition variables y 1 , … , y 5 y_1, \ldots, y_5 y 1 , … , y 5 (resp. $z_1, \ldots, z_3). Then
W Y P : = T ( L Y ) = { 0 ‾ Z ∣ z ‾ ∈ 2 3 } and W Z P : = T ( L Z ) = { y ‾ 0 ‾ ∣ y ‾ ∈ 2 5 } . W_Y^P := T(L_Y) = \{\overline{0}_Z \mid\overline{z} \in2^3\} \text{ and } W_Z^P := T(L_Z) = \{\overline{y}\overline{0} \mid\overline{y} \in2^5\}. W Y P := T ( L Y ) = { 0 Z ∣ z ∈ 2 3 } and W Z P := T ( L Z ) = { y 0 ∣ y ∈ 2 5 } .
Notice that 0 ‾ ∈ W Y P ∩ W Z P ≠ ∅ \overline{0} \in W_Y^P \cap W_Z^P \ne\varnothing 0 ∈ W Y P ∩ W Z P = ∅ and 1 ‾ ∉ W Y P ∪ W Z P ≠ S \overline{1} \notin W_Y^P \cup W_Z^P \ne S 1 ∈ / W Y P ∪ W Z P = S , hence ( S P , W Y P , W Z P ) (S^P, W_Y^P, W_Z^P) ( S P , W Y P , W Z P ) is not an implementation problem as per Definition 2.5 . We propose the following extension of the definition of magical implementation to structures ( S , { W i ∣ i ∈ I } ) (S, \{W_i \mid i \in I\}) ( S , { W i ∣ i ∈ I }) s.t. { W i ∣ i ∈ I } \{W_i \mid i \in I\} { W i ∣ i ∈ I } is not necessarily a partition of S S S .11
Definition 3.7. The codices { L i ∣ i ∈ I } \{L_i \mid i \in I\} { L i ∣ i ∈ I } magically implement ( S , { W i ∣ i ∈ I } ) (S, \{W_i \mid i \in I\}) ( S , { W i ∣ i ∈ I }) if, for all i ∈ I i \in I i ∈ I and s , t ∈ S s,t \in S s , t ∈ S ,
T ( L i ) = W i T(L_i)=W_i T ( L i ) = W i ,
if s ∉ ⋂ i ∈ I W i s \notin\bigcap_{i\in I} W_i s ∈ / ⋂ i ∈ I W i and s → L i t s \to_{L_i} t s → L i t then t ̸ → L j s t \not\to_{L_j} s t → L j s for all j ≠ i j \ne i j = i .
As to the guidance relations → L Y \to_{L_Y} → L Y and → L Z \to_{L_Z} → L Z , for every s ∈ S s \in S s ∈ S , s → L Y s ′ s \to_{L_Y} s' s → L Y s ′ and s → L Z s ′ ′ s \to_{L_Z} s'' s → L Z s ′′ , where
s ′ : = { s if y ˉ s = 0 ˉ 0 z ˉ ′ if y ˉ s ≠ 0 ˉ s ′ ′ : = { s if z ˉ s = 0 ˉ y ˉ ′ ′ 0 if z ˉ s ≠ 0 ˉ s' := \begin{cases}
s & \text{if } \bar{y}_s=\bar{0}\\
0\bar{z}' & \text{if } \bar{y}_s\ne\bar{0}
\end{cases}
\qquad
s'' := \begin{cases}
s & \text{if } \bar{z}_s=\bar{0}\\
\bar{y}''0 & \text{if } \bar{z}_s\ne\bar{0}
\end{cases} s ′ := { s 0 z ˉ ′ if y ˉ s = 0 ˉ if y ˉ s = 0 ˉ s ′′ := { s y ˉ ′′ 0 if z ˉ s = 0 ˉ if z ˉ s = 0 ˉ
and, for every 1 ≤ i ≤ 3 1 \le i \le3 1 ≤ i ≤ 3 , z ′ ( i ) = 1 z'(i)=1 z ′ ( i ) = 1 iff y s ( j ) = 1 y_s(j)=1 y s ( j ) = 1 and y j → Y z i y_j \to_Y z_i y j → Y z i for some 1 ≤ j ≤ 5 1 \le j \le5 1 ≤ j ≤ 5 , and for every 1 ≤ i ≤ 5 1 \le i \le5 1 ≤ i ≤ 5 , y ′ ′ ( i ) = 1 y''(i)=1 y ′′ ( i ) = 1 iff z s ( j ) = 1 z_s(j)=1 z s ( j ) = 1 and z j → Z y i z_j \to_Z y_i z j → Z y i for some 1 ≤ j ≤ 3 1 \le j \le3 1 ≤ j ≤ 3 .
Hence, → L Y \to_{L_Y} → L Y guides any profile s ∈ S P s \in S^P s ∈ S P to itself if the value of all its y y y -coordinates is 0 (i.e. if s ∈ W Y P s \in W_Y^P s ∈ W Y P ), and otherwise, to an element s ′ ∈ W Y P s' \in W_Y^P s ′ ∈ W Y P whose z z z -coordinates take the values determined by which y y y -coordinate of s s s is non-zero, and to which z z z -variable the y y y -variable corresponding to that coordinate is mapped by the original relation → Y \to_Y → Y , and symmetrically for → L Z \to_{L_Z} → L Z . For instance, if s = y ˉ s z ˉ s s=\bar{y}_s\bar{z}_s s = y ˉ s z ˉ s with y ˉ s = ( 1 , 1 , 0 , 0 , 1 ) \bar{y}_s=(1,1,0,0,1) y ˉ s = ( 1 , 1 , 0 , 0 , 1 ) and z ˉ s = 0 ˉ \bar{z}_s=\bar{0} z ˉ s = 0 ˉ , then s ′ = ( 0 , 0 , 0 , 0 , 0 , 1 , 1 , 0 ) s'=(0,0,0,0,0,1,1,0) s ′ = ( 0 , 0 , 0 , 0 , 0 , 1 , 1 , 0 ) , since y 1 → Y z 1 y_1\to_Y z_1 y 1 → Y z 1 , y 2 → Y z 2 y_2\to_Y z_2 y 2 → Y z 2 , and y 5 → Y z 1 y_5\to_Y z_1 y 5 → Y z 1 .
Proposition 3.8. The codices ( L Y , L Z ) (L_Y,L_Z) ( L Y , L Z ) magically implement ( S P , W Y P , W Z P ) (S^P,W_Y^P,W_Z^P) ( S P , W Y P , W Z P ) .
Proof. By construction, W U P : = T ( L U ) W_U^P:=T(L_U) W U P := T ( L U ) for each U ∈ { Y , Z } U\in\{Y,Z\} U ∈ { Y , Z } . To finish the proof, we need to show that, for any s ∈ S P s\in S^P s ∈ S P , s ≠ 0 s\ne0 s = 0 , if s → L Y s ′ s\to_{L_Y}s' s → L Y s ′ , then s ′ ̸ → L Z s s'\not\to_{L_Z}s s ′ → L Z s and if s → L Z s ′ ′ s\to_{L_Z}s'' s → L Z s ′′ , then s ′ ′ ̸ → L Y s s''\not\to_{L_Y}s s ′′ → L Y s . By assumption, s s s has at least one non-zero coordinate. We proceed by cases. If y ˉ s = 0 ˉ \bar{y}_s=\bar{0} y ˉ s = 0 ˉ , then s ′ = s s'=s s ′ = s , and the non-zero coordinates in s ′ = s s'=s s ′ = s are among the z z z -coordinates; however, by definition, the element s ′ ′ ∈ S s''\in S s ′′ ∈ S s.t. s ′ → L Z s ′ ′ s'\to_{L_Z}s'' s ′ → L Z s ′′ is such that z ˉ s ′ ′ = 0 ˉ \bar{z}_{s''}=\bar{0} z ˉ s ′′ = 0 ˉ , hence s ′ ′ ≠ s s''\ne s s ′′ = s , and so s ′ ̸ → L Z s s'\not\to_{L_Z}s s ′ → L Z s . If y ˉ s ≠ 0 ˉ \bar{y}_s\ne\bar{0} y ˉ s = 0 ˉ , then s ′ : = 0 z ˉ ′ s':=0\bar{z}' s ′ := 0 z ˉ ′ , where z ˉ ′ ≠ 0 ˉ \bar{z}'\ne\bar{0} z ˉ ′ = 0 ˉ . To finish the proof, we need to show that if s ′ ′ ∈ S s''\in S s ′′ ∈ S and s ′ → L Y s ′ ′ s'\to_{L_Y}s'' s ′ → L Y s ′′ , then s ′ ′ ≠ s s''\ne s s ′′ = s . Since z ˉ s ′ = z ˉ ′ ≠ 0 ˉ \bar{z}_{s'}=\bar{z}'\ne\bar{0} z ˉ s ′ = z ˉ ′ = 0 ˉ , by definition, s ′ ′ = y ˉ ′ ′ 0 s''=\bar{y}''0 s ′′ = y ˉ ′′ 0 . Hence, if z ˉ s ≠ 0 ˉ \bar{z}_s\ne\bar{0} z ˉ s = 0 ˉ , we can conclude that s ′ ′ ≠ s s''\ne s s ′′ = s . Let us assume that z ˉ s = 0 ˉ \bar{z}_s=\bar{0} z ˉ s = 0 ˉ . Then, the only non-zero coordinates in z ˉ s ′ \bar{z}_{s'} z ˉ s ′ are those that correspond to the variables z i z_i z i such that y j → Y z i y_j\to_Y z_i y j → Y z i for some non-zero y y y -coordinate y j y_j y j in s s s . As discussed at the end of § 2.2 , the relations → Y \to_Y → Y and → Z \to_Z → Z magically implement the original implementation problem ( S , W Y , W Z ) (S,W_Y,W_Z) ( S , W Y , W Z ) ; hence, y j → Y z i y_j\to_Y z_i y j → Y z i implies z i ̸ → Z y j z_i\not\to_Z y_j z i → Z y j for any non-zero coordinate of z ˉ s ′ \bar{z}_{s'} z ˉ s ′ . This implies that the j j j th coordinate of y ˉ s ′ ′ = y ˉ ′ ′ \bar{y}_{s''}=\bar{y}'' y ˉ s ′′ = y ˉ ′′ is zero, and hence s ′ ′ ≠ s s''\ne s s ′′ = s , as required. □
Notice that, although the construction presented above has been introduced in the context of the example discussed at the end of § 2.2 , the definitions and the proof above do not depend on the specific example, and hence have general validity.
Tiglath-Pileser relations and game-theoretic analyses
In the present section, we revisit the formal frameworks of [8 , 10 ] , reported on in § 2 , by framing them in a purely set-theoretic way. Specifically, in § 4.1 , we isolate a class of binary relations (which we refer to as Tiglath-Pileser relations ) that can function as guidance relations, and use them in the context. of a game-theoretic analysis of the persuasion situations of [8 ] ; in Section 4.2, we pursue an axiomatic analysis on Tiglath-Pileser relations that can function as ‘magical implementers’. A common feature of these developments is that the analysis is structured along a hierarchy of levels of strategic/cognitive awareness of the agents involved.
Models of persuasion via Tiglath-Pileser relations
In this section, as in [8 ] , we consider persuasion situations with a Speaker A A A and Listener B B B , in which A A A makes declarations aimed at persuading B B B of a certain state of affairs, irrespective of the true state of affairs, whereas B B B tries to guess the true state of affairs using A A A ’s message. We represent these persuasion situations as tuples ( S , W A , R A ) (S, W_A, R_A) ( S , W A , R A ) such that S S S is the set of states, W A ⊆ S W_A \subseteq S W A ⊆ S is the winning set of A A A and R A ⊆ S × S R_A \subseteq S \times S R A ⊆ S × S is a Tiglath-Pileser relation over S S S towards W A W_A W A :
Definition 4.1. For any tuple ( S , W ) (S, W) ( S , W ) such that S S S is a nonempty set and W ⊆ S W \subseteq S W ⊆ S , a Tiglath-Pileser relation over S S S towards W W W is a relation R ⊆ S × S R \subseteq S \times S R ⊆ S × S verifying the following conditions:
R 1. R − 1 [ S ] : = { s ∈ S ∣ ∃ s ′ ( s ′ ∈ S & R ( s , s ′ ) ) } = S R1.\quad R^{-1}[S] := \{s \in S \mid\exists s'(s' \in S \mathbin{\&} R(s, s'))\} = S R 1. R − 1 [ S ] := { s ∈ S ∣ ∃ s ′ ( s ′ ∈ S & R ( s , s ′ ))} = S ;
R 2. R ∣ W × S = Δ W R2.\quad R|_{W \times S} = \Delta_W R 2. R ∣ W × S = Δ W ;
R 3. ∀ s ( s ∉ W ⇒ R [ s ] ⊆ W or R [ s ] = { s } ) R3.\quad\forall s(s \notin W \Rightarrow R[s] \subseteq W \text{ or } R[s] = \{s\}) R 3. ∀ s ( s ∈ / W ⇒ R [ s ] ⊆ W or R [ s ] = { s }) .
As in Definition 3.1, for each such relation, we let T ( R ) = { s ∈ S ∣ ∀ t ( R ( s , t ) ⇒ s = t ) } T(R) = \{s \in S \mid\forall t(R(s,t) \Rightarrow s = t)\} T ( R ) = { s ∈ S ∣ ∀ t ( R ( s , t ) ⇒ s = t )} . We let TP ( S , W ) \operatorname{TP}(S, W) TP ( S , W ) denote the set of Tiglath-Pileser relations over S S S towards W W W .
Tiglath-Pileser relations are understood as generalizations of the guidance relations of Definition 2.1, and fulfill a similar function in modelling persuasion scenarios. Condition R1 states that every s ∈ S s \in S s ∈ S is R R R -related to one or more states; hence, in every possible true state, the Tiglath-Pileser relation guides the speaker to some state; by R2, when the true state is in their own winning set, according to their Tiglath-Pileser relation, the agent is to declare it truthfully; by R3, when the true state is not in their winning set, the Tiglath-Pileser relation will either let the agent declare some state in their winning set, or the true state.
The game proceeds as follows: the Speaker announces a nonempty S ′ ⊆ S S' \subseteq S S ′ ⊆ S , and the Listener concludes a subset X ⊆ S X \subseteq S X ⊆ S . The payoff of the Speaker is 1 if X ⊆ W A X \subseteq W_A X ⊆ W A , and 0 otherwise.12 The payoff of the Listener is ∣ { s } ∩ X ∣ ÷ ∣ X ∣ |\{s\} \cap X| \div|X| ∣ { s } ∩ X ∣ ÷ ∣ X ∣ , where s ∈ S s \in S s ∈ S denotes the true state.
We analyse situations in which the Speaker and the Listener display various degrees of strategic/cognitive awareness (see [4 ] ); specifically, we assume that a non-strategic A A A always follows their Tiglath-Pileser relation and announces some nonempty S ′ ⊆ R A [ s ] S' \subseteq R_A[s] S ′ ⊆ R A [ s ] , whereas a non-strategic B B B believes A A A to be honest. We also assume that this is known by a strategic B B B who furthermore knows the R A R_A R A and W A W_A W A associated with a non-strategic A A A , and that all this is known by a strategic A A A who knows that B B B is strategic.
We consider the following cases, corresponding to five levels of analysis, which we express in the language of epistemic logic, in which, for any i ∈ { A , B } i \in\{A, B\} i ∈ { A , B } , the proposition p i p_i p i stands for ‘agent i i i is strategic’, and the epistemic operator □ i \Box_i □ i stands for ‘agent i i i believes that’.
Case 0 ¬ p A , ¬ p B Case 1 p A , ¬ p B , □ A ¬ p B Case 2 ¬ p A , p B , □ B ¬ p A Case 3 p A , p B , □ B ¬ p A , □ A p B , □ A □ B ¬ p A Case 4 p A , p B , □ B p A , □ A p B , □ B □ A p B , □ A □ B ¬ p A , □ B □ A □ B ¬ p A \begin{aligned}
\text{Case 0} &\quad\neg p_A, \neg p_B \\
\text{Case 1} &\quad p_A, \neg p_B, \Box_A\neg p_B \\
\text{Case 2} &\quad\neg p_A, p_B, \Box_B\neg p_A \\
\text{Case 3} &\quad p_A, p_B, \Box_B\neg p_A, \Box_A p_B, \Box_A\Box_B\neg p_A \\
\text{Case 4} &\quad p_A, p_B, \Box_B p_A, \Box_A p_B, \Box_B\Box_A p_B, \Box_A\Box_B\neg p_A, \Box_B\Box_A\Box_B\neg p_A
\end{aligned} Case 0 Case 1 Case 2 Case 3 Case 4 ¬ p A , ¬ p B p A , ¬ p B , □ A ¬ p B ¬ p A , p B , □ B ¬ p A p A , p B , □ B ¬ p A , □ A p B , □ A □ B ¬ p A p A , p B , □ B p A , □ A p B , □ B □ A p B , □ A □ B ¬ p A , □ B □ A □ B ¬ p A
Case 0. If R A [ s ] ∩ W A ≠ ∅ R_A[s] \cap W_A \ne\varnothing R A [ s ] ∩ W A = ∅ , then A A A announces S ′ : = R A [ s ] ∩ W A S' := R_A[s] \cap W_A S ′ := R A [ s ] ∩ W A . Since B B B is non-strategic, they take A A A to be honest and thus believe the true state lies in S ′ S' S ′ . As S ′ ⊆ W A S' \subseteq W_A S ′ ⊆ W A , A A A successfully persuades B B B . If s ∉ W A s \notin W_A s ∈ / W A , then s ∉ S ′ s \notin S' s ∈ / S ′ , and hence B B B ’s payoff is 0 and A A A ’s payoff is 1. If s ∈ W A s \in W_A s ∈ W A , then, by Definition 4.1.2, R A [ s ] = S ′ = { s } R_A[s] = S' = \{s\} R A [ s ] = S ′ = { s } , which B B B believes; thus the payoff of both players is 1. If R A [ s ] ∩ W A = ∅ R_A[s] \cap W_A = \varnothing R A [ s ] ∩ W A = ∅ , then by Definition 4.1.3, R A [ s ] = { s } R_A[s] = \{s\} R A [ s ] = { s } . Hence, A A A is forced to announce S ′ = { s } S' = \{s\} S ′ = { s } , which results in A A A ’s payoff being 0 while B B B ’s payoff is 1.
Case 1. Since A A A is strategic and knows that B B B is non-strategic, A A A can announce any nonempty S ′ ⊆ W A S' \subseteq W_A S ′ ⊆ W A (independently of R A R_A R A ), and as B B B is non-strategic, they believe A A A who then gets a payoff of 1. If s ∉ S ′ s \notin S' s ∈ / S ′ , then B B B ’s payoff is 0; if s ∈ S ′ s \in S' s ∈ S ′ , then B B B ’s payoff is 1 / ∣ S ′ ∣ > 0 1/|S'| > 0 1/∣ S ′ ∣ > 0 .
Case 2. In this case, A A A is non-strategic, while B B B is strategic, knows that A A A is non-strategic, and knows R A R_A R A . Then A A A announces S ′ : = R A [ s ] ∩ W A S' := R_A[s] \cap W_A S ′ := R A [ s ] ∩ W A if R A [ s ] ∩ W A ≠ ∅ R_A[s] \cap W_A \ne\varnothing R A [ s ] ∩ W A = ∅ , and S ′ : = R A [ s ] = { s } S' := R_A[s] = \{s\} S ′ := R A [ s ] = { s } otherwise, and B B B concludes X : = R A − 1 [ S ′ ] X := R_A^{-1}[S'] X := R A − 1 [ S ′ ] ; by construction, s ∈ X s \in X s ∈ X , which guarantees a positive payoff for B B B , while A A A ’s graded payoff is nonzero only if X ∩ W A = R A − 1 [ S ′ ] ∩ W A ≠ ∅ X \cap W_A = R_A^{-1}[S'] \cap W_A \ne\varnothing X ∩ W A = R A − 1 [ S ′ ] ∩ W A = ∅ iff S ′ ∩ R A [ W A ] = S ′ ∩ W A ≠ ∅ S' \cap R_A[W_A] = S' \cap W_A \ne\varnothing S ′ ∩ R A [ W A ] = S ′ ∩ W A = ∅ iff R A [ s ] ∩ W A ≠ ∅ R_A[s] \cap W_A \ne\varnothing R A [ s ] ∩ W A = ∅ . If R A [ s ] ∩ W A = ∅ R_A[s] \cap W_A = \varnothing R A [ s ] ∩ W A = ∅ then B B B ’s payoff is 1 and A A A ’s payoff is 0.
Case 3. In this case, A A A knows that for any nonempty S ′ ⊆ S S' \subseteq S S ′ ⊆ S they announce, B B B will conclude X : = R A − 1 [ S ′ ] X := R_A^{-1}[S'] X := R A − 1 [ S ′ ] as in Case 2. Hence, A A A declares a nonempty S ′ ⊆ S S' \subseteq S S ′ ⊆ S that maximises A A A ’s graded payoff, which is ∣ R A − 1 [ S ′ ] ∩ W A ∣ / ∣ R A − 1 [ S ′ ] ∣ |R_A^{-1}[S'] \cap W_A|/|R_A^{-1}[S']| ∣ R A − 1 [ S ′ ] ∩ W A ∣/∣ R A − 1 [ S ′ ] ∣ . If some S ′ S' S ′ exists s.t. R A − 1 [ S ′ ] ⊆ W A R_A^{-1}[S'] \subseteq W_A R A − 1 [ S ′ ] ⊆ W A , then A A A can achieve a payoff of 1 by declaring such an S ′ S' S ′ . Otherwise, since W A ⊆ R A − 1 [ R A [ W A ] ] W_A \subseteq R_A^{-1}[R_A[W_A]] W A ⊆ R A − 1 [ R A [ W A ]] , A A A can secure a graded payoff of ∣ W A ∣ / ∣ R A − 1 [ R A [ W A ] ] ∣ |W_A|/|R_A^{-1}[R_A[W_A]]| ∣ W A ∣/∣ R A − 1 [ R A [ W A ]] ∣ by declaring S ′ : = W A S' := W_A S ′ := W A .
Case 4. As in the previous case, if some S ′ S' S ′ exists s.t. R A − 1 [ S ′ ] ∩ W A ≠ ∅ R_A^{-1}[S'] \cap W_A \ne\varnothing R A − 1 [ S ′ ] ∩ W A = ∅ A A A announces one that maximizes ∣ R A − 1 [ S ′ ] ∩ W A ∣ / ∣ R A − 1 [ S ′ ] ∣ |R_A^{-1}[S'] \cap W_A|/|R_A^{-1}[S']| ∣ R A − 1 [ S ′ ] ∩ W A ∣/∣ R A − 1 [ S ′ ] ∣ ; otherwise, A A A announces S ′ : = { s } S' := \{s\} S ′ := { s } . Here, B B B can infer the true state only if no such S ′ S' S ′ exists; then S ′ = { s } S' = \{s\} S ′ = { s } and B B B trivially identifies s s s . If such an S ′ S' S ′ exists, A A A may announce it regardless of the true state, and B B B knows this; hence B B B cannot infer anything from the message. Thus, A A A cannot persuade B B B , while B B B gains no information about the true state.
Magical implementation via relations
As in [8 , 10 ] , we consider situations in which one or more Speakers make declarations aimed at persuading a Listener of a certain state of affairs, irrespective of the true state of affairs. We assume that the Speakers’ declarations are not arbitrary, but are always anchored to the true state of affairs, because the Speakers are either limited in their ability to lie by the Listener’s knowledge and/or their own bounded rationality. We represent these persuasion situations as tuples ( S , { W i } i ∈ I , { R i } i ∈ I ) (S, \{W_i\}_{i\in I}, \{R_i\}_{i\in I}) ( S , { W i } i ∈ I , { R i } i ∈ I ) s.t. S S S is a nonempty domain of states, I I I is a nonempty finite set of agents, W i ⊆ S W_i \subseteq S W i ⊆ S , and each R i R_i R i is a Tiglath-Pileser relation on S S S towards W i W_i W i for all i ∈ I i \in I i ∈ I (cf. Definition 4.1). We do not insist on requiring that { W i } i ∈ I \{W_i\}_{i\in I} { W i } i ∈ I be a partition of S S S ; hence, we allow for more than one Speaker to be truthful, or none. Intuitively, agent i i i wishes to persuade the Listener that the true state belongs to W i W_i W i (the winning set of i i i ). Hence, if t ∈ S t \in S t ∈ S is the true state, agent i i i will declare some s i ∈ R i [ t ] : = { s ∈ S ∣ R i ( t , s ) } s_i \in R_i[t] := \{s \in S \mid R_i(t,s)\} s i ∈ R i [ t ] := { s ∈ S ∣ R i ( t , s )} . Therefore, upon knowing R i R_i R i , the Listener can apply the Tiglath-Pileser principle to conclude that the true state is in R i − 1 [ s i ] : = { s ∈ S ∣ R i ( s , s i ) } R_i^{-1}[s_i] := \{s \in S \mid R_i(s,s_i)\} R i − 1 [ s i ] := { s ∈ S ∣ R i ( s , s i )} when i i i declares s i s_i s i . Likewise, upon receiving the declarations s i s_i s i of each Speaker i ∈ I i \in I i ∈ I , the Listener will conclude that t ∈ ⋂ i ∈ I R i − 1 [ s i ] t \in\bigcap_{i\in I} R_i^{-1}[s_i] t ∈ ⋂ i ∈ I R i − 1 [ s i ] . In most situations, we will assume that Tiglath-Pileser relations satisfy a stronger version of R3 in Definition 4.1, namely R i [ S ] ⊆ W i R_i[S] \subseteq W_i R i [ S ] ⊆ W i . In this case, T ( R i ) = W i T(R_i) = W_i T ( R i ) = W i . The next definition is the set-theoretic reformulation of Definition 3.7.
Definition 4.2. The set of relations { R i ∣ i ∈ I } \{R_i \mid i \in I\} { R i ∣ i ∈ I } magically implements ( S , { W i ∣ i ∈ I } ) (S, \{W_i \mid i \in I\}) ( S , { W i ∣ i ∈ I }) if:
R i [ S ] ⊆ W i R_i[S] \subseteq W_i R i [ S ] ⊆ W i and R i ∣ W i × W i = Δ W i R_i|_{W_i \times W_i} = \Delta_{W_i} R i ∣ W i × W i = Δ W i for all i ∈ I i \in I i ∈ I ;
for all s , t ∈ S s,t \in S s , t ∈ S , if s ∉ ⋂ i ∈ I W i s \notin\bigcap_{i\in I} W_i s ∈ / ⋂ i ∈ I W i and t ∈ R i [ s ] t \in R_i[s] t ∈ R i [ s ] , then s ∉ R j [ t ] s \notin R_j[t] s ∈ / R j [ t ] for all j ≠ i j \ne i j = i .
For simplicity of presentation, we consider only two agents, A A A and B B B , although the same analysis and constructions can be carried out for any finite number of agents.
For a persuasion situation ( S , W A , W B , R A , R B ) (S, W_A, W_B, R_A, R_B) ( S , W A , W B , R A , R B ) as above, in which R A R_A R A and R B R_B R B satisfy R1, R2 and the stronger version of R3, let R ⊆ S × ( W A × W B ) R \subseteq S \times(W_A \times W_B) R ⊆ S × ( W A × W B ) be the product relation defined as follows: for all s ∈ S s \in S s ∈ S , s ′ ∈ W A s' \in W_A s ′ ∈ W A , and s ′ ′ ∈ W B s'' \in W_B s ′′ ∈ W B ,
R ( s , ( s ′ , s ′ ′ ) ) iff R A ( s , s ′ ) and R B ( s , s ′ ′ ) . (3) R(s,(s',s'')) \quad\text{iff} \quad R_A(s,s') \text{ and } R_B(s,s''). \tag*{(3)} R ( s , ( s ′ , s ′′ )) iff R A ( s , s ′ ) and R B ( s , s ′′ ) . (3)
Hence, by definition, for all s ∈ S s \in S s ∈ S , s ′ ∈ W A s' \in W_A s ′ ∈ W A and s ′ ′ ∈ W B s'' \in W_B s ′′ ∈ W B ,
R [ s ] = R A [ s ] × R B [ s ] and R − 1 [ ( s ′ , s ′ ′ ) ] = R A − 1 [ s ′ ] ∩ R B − 1 [ s ′ ′ ] . R[s] = R_A[s] \times R_B[s] \quad\text{and} \quad R^{-1}[(s',s'')] = R_A^{-1}[s'] \cap R_B^{-1}[s'']. R [ s ] = R A [ s ] × R B [ s ] and R − 1 [( s ′ , s ′′ )] = R A − 1 [ s ′ ] ∩ R B − 1 [ s ′′ ] .
Assumption R1 (R A − 1 [ W A ] = S = R B − 1 [ W B ] R_A^{-1}[W_A] = S = R_B^{-1}[W_B] R A − 1 [ W A ] = S = R B − 1 [ W B ] ) guarantees that, whenever A A A and B B B are not strategic (in the sense that neither takes the other parties’ actions and knowledge into consideration), for every “declaration” ( s ′ , s ′ ′ ) ∈ W A × W B (s',s'') \in W_A \times W_B ( s ′ , s ′′ ) ∈ W A × W B from A A A and B B B , the principal can always get back to the set of states s ∈ R − 1 [ ( s ′ , s ′ ′ ) ] s \in R^{-1}[(s',s'')] s ∈ R − 1 [( s ′ , s ′′ )] from which the declarations possibly originated (this formally represents the ‘anchor to the truth’). This is because, when both A A A and B B B are not strategic and t ∈ S t \in S t ∈ S is the true state, t ∈ R A − 1 [ s ′ ] ∩ R B − 1 [ s ′ ′ ] = R − 1 [ ( s ′ , s ′ ′ ) ] t \in R_A^{-1}[s'] \cap R_B^{-1}[s''] = R^{-1}[(s',s'')] t ∈ R A − 1 [ s ′ ] ∩ R B − 1 [ s ′′ ] = R − 1 [( s ′ , s ′′ )] . Furthermore, if
R4. R R R is relation-injective , i.e. s ≠ t s \ne t s = t implies R [ s ] ∩ R [ t ] = ∅ R[s] \cap R[t] = \varnothing R [ s ] ∩ R [ t ] = ∅ for all s , t ∈ S s,t \in S s , t ∈ S .13 then for all s ′ ∈ W A s' \in W_A s ′ ∈ W A and s ′ ′ ∈ W B s'' \in W_B s ′′ ∈ W B , the set R − 1 [ ( s ′ , s ′ ′ ) ] R^{-1}[(s',s'')] R − 1 [( s ′ , s ′′ )] has at most one element (for if R ( s , ( s ′ , s ′ ′ ) ) R(s,(s',s'')) R ( s , ( s ′ , s ′′ )) and R ( t , ( s ′ , s ′ ′ ) ) R(t,(s',s'')) R ( t , ( s ′ , s ′′ )) , then ( s ′ , s ′ ′ ) ∈ R [ s ] ∩ R [ t ] ≠ ∅ (s',s'') \in R[s] \cap R[t] \ne\varnothing ( s ′ , s ′′ ) ∈ R [ s ] ∩ R [ t ] = ∅ ), and hence, the Listener can always get back to the true state.
As we do not assume that W A ∩ W B = ∅ W_A \cap W_B = \varnothing W A ∩ W B = ∅ , if t t t belongs to both winning sets, then both agents would truthfully declare it, and in fact, they would agree only on states in the intersection of their winning sets.
Proposition 4.3. The following are equivalent for any ( S , W A , W B , R A , R B ) (S,W_A,W_B,R_A,R_B) ( S , W A , W B , R A , R B ) satisfying the condition of Definition 4.2.1 and such that W A ∪ W B = S W_A \cup W_B = S W A ∪ W B = S :
{ R A , R B } \{R_A,R_B\} { R A , R B } magically implements ( S , W A , W B ) (S,W_A,W_B) ( S , W A , W B ) ;
the product relation R R R defined as in (3) is relation-injective.
Proof. It is enough to show that Definition 4.2.2 is equivalent to R R R being relation-injective while assuming the condition of Definition 4.2.1. Arguing contrapositively, let us assume that R R R is not relation-injective, i.e. some s 1 , s 2 , s , t ∈ S s_1,s_2,s,t \in S s 1 , s 2 , s , t ∈ S exist such that s 1 ≠ s 2 s_1 \ne s_2 s 1 = s 2 , R A ( s i , s ) R_A(s_i,s) R A ( s i , s ) and R B ( s i , t ) R_B(s_i,t) R B ( s i , t ) for 1 ≤ i ≤ 2 1 \le i \le2 1 ≤ i ≤ 2 . Hence, s ∈ W A s \in W_A s ∈ W A and t ∈ W B t \in W_B t ∈ W B by Definition 4.2.1. Since W A ∪ W B = S W_A \cup W_B = S W A ∪ W B = S , we can assume without loss of generality that s 1 ∈ W A s_1 \in W_A s 1 ∈ W A and s 2 ∉ W A s_2 \notin W_A s 2 ∈ / W A (hence, s 2 ∈ W B s_2 \in W_B s 2 ∈ W B ); this is because, if e.g. s 1 , s 2 ∈ W A s_1,s_2 \in W_A s 1 , s 2 ∈ W A , then s 1 = s = s 2 s_1 = s = s_2 s 1 = s = s 2 by Definition 4.2.1, against the assumption that s 1 ≠ s 2 s_1 \ne s_2 s 1 = s 2 , and a similar argument excludes s 1 , s 2 ∈ W B s_1,s_2 \in W_B s 1 , s 2 ∈ W B . Therefore, R A ( s 1 , s ) R_A(s_1,s) R A ( s 1 , s ) and R B ( s 2 , t ) R_B(s_2,t) R B ( s 2 , t ) respectively imply s 1 = s s_1 = s s 1 = s and s 2 = t s_2 = t s 2 = t by Definition 4.2.1. Then, s 2 ∉ W A ∩ W B s_2 \notin W_A \cap W_B s 2 ∈ / W A ∩ W B , s 1 ∈ R A [ s 2 ] s_1 \in R_A[s_2] s 1 ∈ R A [ s 2 ] and s 2 ∈ R B [ s 1 ] s_2 \in R_B[s_1] s 2 ∈ R B [ s 1 ] , which violates Definition 4.2.2.
Conversely, if s ∉ W B s \notin W_B s ∈ / W B and s ∈ R A [ t ] s \in R_A[t] s ∈ R A [ t ] and t ∈ R B [ s ] t \in R_B[s] t ∈ R B [ s ] , then t ∈ W B t \in W_B t ∈ W B , so t ≠ s t \ne s t = s . Moreover, the assumptions and Definition 4.2.1 imply that R ( s , ( s , t ) ) R(s,(s,t)) R ( s , ( s , t )) and R ( t , ( s , t ) ) R(t,(s,t)) R ( t , ( s , t )) , i.e. R R R is not relation-injective. □
Example 4.4. The last example discussed at the end of § 2.2 gives rise to the persuasion situation ( S , W Y , W Z , → Y , → Z ) (S,W_Y,W_Z,\to_Y,\to_Z) ( S , W Y , W Z , → Y , → Z ) represented below, which satisfies all the assumptions of Proposition 4.3 . Hence, an equivalent way of understanding why → Y \to_Y → Y and → Z \to_Z → Z magically implement ( S , W Y , W Z ) (S,W_Y,W_Z) ( S , W Y , W Z ) is that, whichever state is the true one, the principal can consider the inverse image of the tuple declared by Y Y Y and Z Z Z according to their product relation R ⊆ S × ( W Y × W Z ) R \subseteq S \times(W_Y \times W_Z) R ⊆ S × ( W Y × W Z ) , defined as in (3) and represented in the third column below. Since R R R is relation-injective, the inverse image of each tuple ( s ′ , s ′ ′ ) ∈ W Y × W Z (s',s'') \in W_Y \times W_Z ( s ′ , s ′′ ) ∈ W Y × W Z has at most one element; when such an element exists, under the assumption that neither Speaker is strategic, the principal can conclude that it is the true state.
Figure 7 (PDF p. 22).
Table 2 (PDF p. 22).
Example 4.5. The present example shows that the magical implementation mechanism can work also under more general assumptions than those of [10 ] ; for instance, the winning sets do not need to form a partition. Compared to the previous example, the present one also shows that the individual Tiglath-Pileser relations do not need to be relation-injective in order for their product to be. In this more general case, the pivot is not anymore Definition 4.2, but rather conditions R1-R4. Let S = { a 1 , a 2 , a 3 , b 1 , b 2 , b 3 , s } S = \{a_1, a_2, a_3, b_1, b_2, b_3, s\} S = { a 1 , a 2 , a 3 , b 1 , b 2 , b 3 , s } , W A = { a 1 , a 2 , a 3 } W_A = \{a_1, a_2, a_3\} W A = { a 1 , a 2 , a 3 } , W B = { b 1 , b 2 , b 3 } W_B = \{b_1, b_2, b_3\} W B = { b 1 , b 2 , b 3 } (hence, W A ∪ W B ≠ S W_A \cup W_B \ne S W A ∪ W B = S ), and let R A R_A R A and R B R_B R B be as in the pictures below (except that R A ∣ W A × S = Δ W A R_A|_{W_A \times S} = \Delta_{W_A} R A ∣ W A × S = Δ W A and R B ∣ W B × S = Δ W B R_B|_{W_B \times S} = \Delta_{W_B} R B ∣ W B × S = Δ W B are not represented in the pictures). The ensuing relation R ⊆ S × ( W A × W B ) R \subseteq S \times(W_A \times W_B) R ⊆ S × ( W A × W B ) is represented in the third column below.
Figure 8.
Table 3.
The relations R A R_A R A and R B R_B R B satisfy R1, R2, and the stronger version of R3, and R R R satisfies R4. Moreover, R R R is onto W A × W B W_A \times W_B W A × W B and is not functional, since ( b 3 , ( a 1 , b 3 ) ) , ( b 3 , ( a 3 , b 3 ) ) ∈ R (b_3,(a_1,b_3)),(b_3,(a_3,b_3)) \in R ( b 3 , ( a 1 , b 3 )) , ( b 3 , ( a 3 , b 3 )) ∈ R . Removing ( a 2 , b 3 ) (a_2,b_3) ( a 2 , b 3 ) from R B R_B R B would result in removing ( a 2 , b 3 ) (a_2,b_3) ( a 2 , b 3 ) from the image of R R R , which would not be onto W A × W B W_A \times W_B W A × W B any longer, while R B R_B R B and R R R would retain all the remaining properties.
Notice that, for R4 to be satisfied, the pigeonhole principle requires that ∣ S ∣ ≤ ∣ W A × W B ∣ |S| \leq|W_A \times W_B| ∣ S ∣ ≤ ∣ W A × W B ∣ , which may be impossible if the size of W A W_A W A and/or W B W_B W B is small relative to S S S . For example, if S = { s 1 , s 2 , s 3 , s 4 } S = \{s_1, s_2, s_3, s_4\} S = { s 1 , s 2 , s 3 , s 4 } and W A = { s 1 } W_A = \{s_1\} W A = { s 1 } and W B = { s 2 , s 3 , s 4 } W_B = \{s_2, s_3, s_4\} W B = { s 2 , s 3 , s 4 } , then ∣ S ∣ = 4 |S| = 4 ∣ S ∣ = 4 but ∣ W A × W B ∣ = 3 |W_A \times W_B| = 3 ∣ W A × W B ∣ = 3 .
Strategic/cognitive hierarchy. As done in Section 4.1, we analyse situations in which agents display various degrees of strategic/cognitive awareness (see [4 ] ) which allows them to depart from the basic assumption that their declarations will adhere to what their Tiglath-Pileser relation prescribes; in other words, their strategies may take into account the expectations of the other agents and of the principal revolving around their Tiglath-Pileser relations. We discuss the following cases; in each of them, the principal’s approximation of the true state obtains as the R R R -inverse image of the tuple of states declared by A A A and B B B .14
Neither A A A nor B B B are strategic; they declare what their associated Tiglath-Pileser relation indicates;
Agent A A A is not strategic and declares what their associated Tiglath-Pileser relation indicates, while B B B is strategic, assumes A A A is not, and knows A A A ’s Tiglath-Pileser relation;
Both A A A and B B B are strategic, assume the other agent is not, and know the other’s Tiglath-Pileser relation;
Both agents are strategic and know the other’s Tiglath-Pileser relation, but while A A A knows that B B B is strategic, B B B assumes that A A A is not strategic;
Both A A A and B B B are strategic, assume the other agent is also strategic, and know the other’s Tiglath-Pileser relation.
In the first case, the principal always succeeds in restricting the search for the true state to R − 1 [ ( s ′ , s ′ ′ ) ] R^{-1}[(s',s'')] R − 1 [( s ′ , s ′′ )] , and if R R R is relation-injective (R4), the principal is able to establish the true state, and the agent Y ∈ { A , B } Y \in\{A,B\} Y ∈ { A , B } s.t. s ∈ W Y s \in W_Y s ∈ W Y will also win.
In the second case, if
R5. R A ∣ ( S ∖ ( W A ∪ W B ) ) × W A R_A|_{(S\setminus(W_A\cup W_B))\times W_A} R A ∣ ( S ∖ ( W A ∪ W B )) × W A is functional, and symmetrically R B ∣ ( S ∖ ( W A ∪ W B ) ) × W B R_B|_{(S\setminus(W_A\cup W_B))\times W_B} R B ∣ ( S ∖ ( W A ∪ W B )) × W B is functional (notice that this condition is trivially satisfied when W A ∪ W B = S W_A \cup W_B = S W A ∪ W B = S )
R6. R A − 1 [ s ] ∩ W B ≠ ∅ R_A^{-1}[s] \cap W_B \ne\varnothing R A − 1 [ s ] ∩ W B = ∅ for all s ∈ W A s \in W_A s ∈ W A , and R B − 1 [ s ] ∩ W A ≠ ∅ R_B^{-1}[s] \cap W_A \ne\varnothing R B − 1 [ s ] ∩ W A = ∅ for all s ∈ W B s \in W_B s ∈ W B , i.e. R A ∣ W B × W A R_A|_{W_B\times W_A} R A ∣ W B × W A is onto W A W_A W A , and R B ∣ W A × W B R_B|_{W_A\times W_B} R B ∣ W A × W B is onto W B W_B W B .
then B B B always wins: if t ∈ W B t \in W_B t ∈ W B , then the pair of agents declares ( s ′ , t ) (s',t) ( s ′ , t ) with s ′ ∈ R A [ t ] s' \in R_A[t] s ′ ∈ R A [ t ] , from which the principal correctly concludes that the true state t t t is in R − 1 [ ( s ′ , t ) ] R^{-1}[(s',t)] R − 1 [( s ′ , t )] ; if R R R is relation-injective, the principal will identify t t t as the only element in R − 1 [ ( s ′ , t ) ] R^{-1}[(s',t)] R − 1 [( s ′ , t )] . If t ∉ W B t \notin W_B t ∈ / W B , then, by R5, B B B knows that A A A will declare the unique s ′ ∈ R A [ t ] s' \in R_A[t] s ′ ∈ R A [ t ] (notice that, by R2 and R5, s ′ s' s ′ is unique both if t ∈ W A t \in W_A t ∈ W A and if t ∉ W A t \notin W_A t ∈ / W A ); then B B B declares some s ′ ′ ∈ W B ∩ R A − 1 [ s ′ ] s'' \in W_B \cap R_A^{-1}[s'] s ′′ ∈ W B ∩ R A − 1 [ s ′ ] which exists since, by R6, W B ∩ R A − 1 [ s ′ ] ≠ ∅ W_B \cap R_A^{-1}[s'] \ne\varnothing W B ∩ R A − 1 [ s ′ ] = ∅ . Then the principal will consider the set R − 1 [ ( s ′ , s ′ ′ ) ] = R A − 1 [ s ′ ] ∩ R B − 1 [ s ′ ′ ] R^{-1}[(s',s'')] = R_A^{-1}[s']\cap R_B^{-1}[s''] R − 1 [( s ′ , s ′′ )] = R A − 1 [ s ′ ] ∩ R B − 1 [ s ′′ ] . By construction, s ′ ′ ∈ R A − 1 [ s ′ ] ∩ R B − 1 [ s ′ ′ ] s'' \in R_A^{-1}[s']\cap R_B^{-1}[s''] s ′′ ∈ R A − 1 [ s ′ ] ∩ R B − 1 [ s ′′ ] , and since R R R is relation-injective, s ′ ′ s'' s ′′ is the only element of R − 1 [ ( s ′ , s ′ ′ ) ] R^{-1}[(s',s'')] R − 1 [( s ′ , s ′′ )] . Hence, in this case, B B B successfully deceives the principal into concluding that the true state is s ′ ′ s'' s ′′ .
Remark 4.6. Conditions R5 and R6 are satisfied by the relations R A R_A R A and R B R_B R B described in Example 4.5. However, removing the pair ( a 2 , b 3 ) (a_2,b_3) ( a 2 , b 3 ) from R B R_B R B , as discussed there, causes R B R_B R B no longer to satisfy R6. Hence, there might be a tension between the non-surjectivity of R R R and R6.
Remark 4.7. In case R5 does not hold, knowing R A R_A R A is in general not sufficient for agent B B B to know what agent A A A will declare in the case that s ∉ W A s \notin W_A s ∈ / W A , which therefore makes it impossible to reliably pick the correct value W B ∩ R A − 1 [ s ] W_B \cap R_A^{-1}[s] W B ∩ R A − 1 [ s ] to ensure that the principal concludes in B B B ’s favour. However, the following alternative assumption ensures that there is an element in W B W_B W B that covers every possible declaration A A A could make, thanks to which we can still make sure that B B B wins using it and R6.
R5′. for all s ∈ S ∖ ( W A ∪ W B ) s \in S \setminus(W_A \cup W_B) s ∈ S ∖ ( W A ∪ W B ) some s ′ ∈ W A s' \in W_A s ′ ∈ W A exists such that R B [ s ] ⊆ R B [ s ′ ] R_B[s] \subseteq R_B[s'] R B [ s ] ⊆ R B [ s ′ ] and, symmetrically, some s ′ ′ ∈ W B s'' \in W_B s ′′ ∈ W B exists such that R A [ s ] ⊆ R A [ s ′ ′ ] R_A[s] \subseteq R_A[s''] R A [ s ] ⊆ R A [ s ′′ ] .
Indeed, if s ∈ W A s \in W_A s ∈ W A , then we proceed as indicated above, using R6 and the fact that B B B knows that A A A will announce s s s ; if s ∈ S ∖ ( W A ∪ W B ) s \in S \setminus(W_A \cup W_B) s ∈ S ∖ ( W A ∪ W B ) , then, by R5′, agent B B B can announce any s ′ ′ ∈ W B s'' \in W_B s ′′ ∈ W B such that R A [ s ] ⊆ R A [ s ′ ′ ] R_A[s] \subseteq R_A[s''] R A [ s ] ⊆ R A [ s ′′ ] , while agent A A A announces some s ′ ∈ R A [ s ] s' \in R_A[s] s ′ ∈ R A [ s ] . The principal considers R − 1 [ ( s ′ , s ′ ′ ) ] = R A − 1 [ s ′ ] ∩ R B − 1 [ s ′ ′ ] R^{-1}[(s',s'')] = R_A^{-1}[s'] \cap R_B^{-1}[s''] R − 1 [( s ′ , s ′′ )] = R A − 1 [ s ′ ] ∩ R B − 1 [ s ′′ ] . From s ′ ′ ∈ W B s'' \in W_B s ′′ ∈ W B , it follows that s ′ ′ ∈ R B − 1 [ s ′ ′ ] s'' \in R_B^{-1}[s''] s ′′ ∈ R B − 1 [ s ′′ ] . From s ′ ∈ R A [ s ] ⊆ R A [ s ′ ′ ] s' \in R_A[s] \subseteq R_A[s''] s ′ ∈ R A [ s ] ⊆ R A [ s ′′ ] it follows that s ′ ′ ∈ R A − 1 [ s ′ ] s'' \in R_A^{-1}[s'] s ′′ ∈ R A − 1 [ s ′ ] . Hence, s ′ ′ ∈ R − 1 [ ( s ′ , s ′ ′ ) ] s'' \in R^{-1}[(s',s'')] s ′′ ∈ R − 1 [( s ′ , s ′′ )] , and since R R R is relationally injective, the principal wrongly concludes that s ′ ′ s'' s ′′ is the “true” state, and agent B B B wins.
In the third case (recall that we have assumed that the true state s ∉ W A ∩ W B s \notin W_A \cap W_B s ∈ / W A ∩ W B ), we need to distinguish several subcases: if s ∈ W A ∪ W B s \in W_A \cup W_B s ∈ W A ∪ W B , then the agent Y ∈ { A , B } Y \in\{A,B\} Y ∈ { A , B } s.t. s ∉ W Y s \notin W_Y s ∈ / W Y will win, since the two agents will behave as in the two subcases of the second case (each of the two agents will play it strategically but assuming that the other would not). The principal is always deceived by relying on R R R .
If s ∈ S ∖ ( W A ∪ W B ) s \in S \setminus(W_A \cup W_B) s ∈ S ∖ ( W A ∪ W B ) , then assuming, according to R5′, that some s ′ ∈ W A s' \in W_A s ′ ∈ W A exists s.t. R B [ s ] ⊆ R B [ s ′ ] R_B[s] \subseteq R_B[s'] R B [ s ] ⊆ R B [ s ′ ] and some s ′ ′ ∈ W B s'' \in W_B s ′′ ∈ W B exists s.t. R A [ s ] ⊆ R A [ s ′ ′ ] R_A[s] \subseteq R_A[s''] R A [ s ] ⊆ R A [ s ′′ ] , we consider the following sub-subcases:
If (a), then, despite their being strategic, A A A and B B B declare what they would have declared if they were not. Hence, by considering R − 1 [ ( s ′ , s ′ ′ ) ] R^{-1}[(s',s'')] R − 1 [( s ′ , s ′′ )] , the principal is able to retrieve the true state and wins. Similarly, if (b) or (c), then we would fall to the case in which one agent is strategic and the other is not, that is, the strategic agent wins and the principal is tricked. If (d), then more subcases:
Figure 1 (PDF p. 25). An illustration of the case considered in Remark 4.8.
(d2) s ′ ∈ R A [ s ′ ′ ] ∖ R A [ s ] s' \in R_A[s''] \setminus R_A[s] s ′ ∈ R A [ s ′′ ] ∖ R A [ s ] and s ′ ′ ∈ W B ∖ R B [ s ′ ] s'' \in W_B \setminus R_B[s'] s ′′ ∈ W B ∖ R B [ s ′ ] ;
(d3) s ′ ∈ W A ∖ R A [ s ′ ′ ] s' \in W_A \setminus R_A[s''] s ′ ∈ W A ∖ R A [ s ′′ ] and s ′ ′ ∈ R B [ s ′ ] ∖ R B [ s ] s'' \in R_B[s'] \setminus R_B[s] s ′′ ∈ R B [ s ′ ] ∖ R B [ s ] ;
(d4) s ′ ∈ W A ∖ R A [ s ′ ′ ] s' \in W_A \setminus R_A[s''] s ′ ∈ W A ∖ R A [ s ′′ ] and s ′ ′ ∈ W B ∖ R B [ s ′ ] s'' \in W_B \setminus R_B[s'] s ′′ ∈ W B ∖ R B [ s ′ ] .
Case (d1) cannot occur if s ′ ≠ s ′ ′ s' \ne s'' s ′ = s ′′ . Indeed, the assumptions s ′ ∈ R A [ s ′ ′ ] s' \in R_A[s''] s ′ ∈ R A [ s ′′ ] and s ′ ′ ∈ R B [ s ′ ] s'' \in R_B[s'] s ′′ ∈ R B [ s ′ ] imply s ′ ′ ∈ R A − 1 [ s ′ ] s'' \in R_A^{-1}[s'] s ′′ ∈ R A − 1 [ s ′ ] and s ′ ∈ R B − 1 [ s ′ ′ ] s' \in R_B^{-1}[s''] s ′ ∈ R B − 1 [ s ′′ ] , and since assumption R2, i.e. that R A ∣ W A × W A = Δ W A R_A|_{W_A \times W_A} = \Delta_{W_A} R A ∣ W A × W A = Δ W A and R B ∣ W B × W B = Δ W B R_B|_{W_B \times W_B} = \Delta_{W_B} R B ∣ W B × W B = Δ W B , implies that s ′ ∈ R A − 1 [ s ′ ] s' \in R_A^{-1}[s'] s ′ ∈ R A − 1 [ s ′ ] and s ′ ′ ∈ R B − 1 [ s ′ ′ ] s'' \in R_B^{-1}[s''] s ′′ ∈ R B − 1 [ s ′′ ] , it follows that s ′ , s ′ ′ ∈ R A − 1 [ s ′ ] ∩ R B − 1 [ s ′ ′ ] = R − 1 [ ( s ′ , s ′ ′ ) ] s',s'' \in R_A^{-1}[s'] \cap R_B^{-1}[s''] = R^{-1}[(s',s'')] s ′ , s ′′ ∈ R A − 1 [ s ′ ] ∩ R B − 1 [ s ′′ ] = R − 1 [( s ′ , s ′′ )] , against the relation-injectivity of R R R .
If (d2), then the unique element in R − 1 [ ( s ′ , s ′ ′ ) ] = R A − 1 [ s ′ ] ∩ R B − 1 [ s ′ ′ ] R^{-1}[(s',s'')] = R_A^{-1}[s'] \cap R_B^{-1}[s''] R − 1 [( s ′ , s ′′ )] = R A − 1 [ s ′ ] ∩ R B − 1 [ s ′′ ] coincides with s ′ ′ s'' s ′′ , since s ′ ∈ R A [ s ′ ′ ] s' \in R_A[s''] s ′ ∈ R A [ s ′′ ] iff s ′ ′ ∈ R A − 1 [ s ′ ] s'' \in R_A^{-1}[s'] s ′′ ∈ R A − 1 [ s ′ ] . Hence, B B B wins, and the principal and A A A lose. The case of (d3) is symmetric, with A A A winning, and B B B and the principal losing.
If (d4), then, by relation-injectivity (R4), R − 1 [ ( s ′ , s ′ ′ ) ] = R A − 1 [ s ′ ] ∩ R B − 1 [ s ′ ′ ] R^{-1}[(s',s'')] = R_A^{-1}[s'] \cap R_B^{-1}[s''] R − 1 [( s ′ , s ′′ )] = R A − 1 [ s ′ ] ∩ R B − 1 [ s ′′ ] contains at most one element. If this element exists, it cannot be s ′ ′ s'' s ′′ , since, by assumption, s ′ ∉ R A [ s ′ ′ ] s' \notin R_A[s''] s ′ ∈ / R A [ s ′′ ] i.e. s ′ ′ ∉ R A − 1 [ s ′ ] s'' \notin R_A^{-1}[s'] s ′′ ∈ / R A − 1 [ s ′ ] , nor can it be s ′ s' s ′ , since, by assumption, s ′ ′ ∉ R B [ s ′ ] s'' \notin R_B[s'] s ′′ ∈ / R B [ s ′ ] i.e. s ′ ∉ R B − 1 [ s ′ ′ ] s' \notin R_B^{-1}[s''] s ′ ∈ / R B − 1 [ s ′′ ] . So both A A A and B B B lose, but the principal does not know that they are strategic. If R − 1 [ ( s ′ , s ′ ′ ) ] = ∅ R^{-1}[(s',s'')] = \varnothing R − 1 [( s ′ , s ′′ )] = ∅ , then both A A A and B B B lose, and the principal knows that A A A and B B B deviated from their response relations.
Taking stock. For the principal, the best opportunity for detecting strategic play is if the range of R R R is small relative to W A × W B W_A \times W_B W A × W B , since this would restrict the deviations that the agents could make from their response relations. Trying to minimize the range of R R R is limited by the requirement that R R R satisfy R4. Condition R4 requires there to be at least as many pairs in r a n g e ( R ) range(R) r an g e ( R ) as there are elements in S S S . Hence, the number of pairs from W A × W B W_A \times W_B W A × W B not in the range of R R R is at most ∣ W A × W B ∣ − ∣ S ∣ |W_A \times W_B| - |S| ∣ W A × W B ∣ − ∣ S ∣ . Therefore, for the principal, the optimal requirements for R R R are R4 together with the following condition
R7. the number of pairs from W A × W B W_A \times W_B W A × W B not in the range of R R R is exactly ∣ W A × W B ∣ − ∣ S ∣ |W_A \times W_B| - |S| ∣ W A × W B ∣ − ∣ S ∣ .
The best situation for the principal is therefore when W A W_A W A and W B W_B W B form a partition of S S S into equally sized parts, if S S S is even, or differing in size by at most one if S S S is odd.
In the fourth case, we consider various subcases: if s ∈ W B s \in W_B s ∈ W B , then A wins by applying the same strategy as in the third case, namely, B, who believes that A is non strategic, will announce s s s truthfully, hence A will win by announcing an s ′ s' s ′ ∈ R B − 1 [ s ] ∩ W A \in R_B^{-1}[s] \cap W_A ∈ R B − 1 [ s ] ∩ W A , and the principal will erroneously conclude that the true state is s ′ s' s ′ .
If s ∈ W A s \in W_A s ∈ W A , then B, who believes that A will truthfully announce s s s since A (as B believes) is non strategic, will announce some s ′ ′ ∈ R A − 1 [ s ] ∩ W B s'' \in R_A^{-1}[s] \cap W_B s ′′ ∈ R A − 1 [ s ] ∩ W B ; if
R8. R A − 1 [ s ] ∩ W B = { s ′ ′ } R_A^{-1}[s] \cap W_B = \{s''\} R A − 1 [ s ] ∩ W B = { s ′′ } for all s ∈ W A s \in W_A s ∈ W A , and R B − 1 [ s ] ∩ W A = { s ′ } R_B^{-1}[s] \cap W_A = \{s'\} R B − 1 [ s ] ∩ W A = { s ′ } for all s ∈ W B s \in W_B s ∈ W B .
then the state s ′ ′ s'' s ′′ announced by B is unique, and hence A, by R6, can announce some s ′ ∈ R B − 1 [ s ′ ′ ] ∩ W A s' \in R_B^{-1}[s''] \cap W_A s ′ ∈ R B − 1 [ s ′′ ] ∩ W A and wins, while the principal erroneously concludes that s ′ s' s ′ is the true state.
If s ∈ S ∖ ( W A ∩ W B ) s \in S \setminus(W_A \cap W_B) s ∈ S ∖ ( W A ∩ W B ) , agent B, who believes that A is not strategic, believes that A will announce some s ′ ∈ R A [ s ] s' \in R_A[s] s ′ ∈ R A [ s ] , hence by R5′, agent B will announce some s ′ ′ ∈ W B s'' \in W_B s ′′ ∈ W B such that R A [ s ] ⊆ R A [ s ′ ′ ] R_A[s] \subseteq R_A[s''] R A [ s ] ⊆ R A [ s ′′ ] . If we further assume that the following strengthened version of R5′ holds:
R5′′. for all s ∈ S ∖ ( W A ∪ W B ) s \in S \setminus(W_A \cup W_B) s ∈ S ∖ ( W A ∪ W B ) some unique s ′ ∈ W A s' \in W_A s ′ ∈ W A exists such that R B [ s ] ⊆ R B [ s ′ ] R_B[s] \subseteq R_B[s'] R B [ s ] ⊆ R B [ s ′ ] and, symmetrically, some unique s ′ ′ ∈ W B s'' \in W_B s ′′ ∈ W B exists such that R A [ s ] ⊆ R A [ s ′ ′ ] R_A[s] \subseteq R_A[s''] R A [ s ] ⊆ R A [ s ′′ ] .
then A announces some s ′ ′ ′ ∈ R B − 1 [ s ′ ′ ] ∩ W A s''' \in R_B^{-1}[s''] \cap W_A s ′′′ ∈ R B − 1 [ s ′′ ] ∩ W A and win, while the principal erroneously concludes that s ′ ′ ′ s''' s ′′′ is the true state.
In the fifth case, the agents are not constrained by the true state, and hence neither by their response relations, and the analysis reduces to a classic game-theoretic situation, which can be analyzed similarly to the following example.
Example 4.9. The following table is the payoff matrix of the scenario in Example 4.5:
We observe that neither A nor B has a dominant strategy: for instance, if B plays b 1 b_1 b 1 then A should play a 3 a_3 a 3 , but if B plays b 2 b_2 b 2 then A should play either a 1 a_1 a 1 or a 2 a_2 a 2 , and similarly, if A chooses a 1 a_1 a 1 , then B must choose b 3 b_3 b 3 , while if A chooses a 2 a_2 a 2 , then B must choose b 1 b_1 b 1 . Moreover, no outcome is a Nash equilibrium.
Let x = ( x 1 , x 2 , x 3 ) x=(x_1,x_2,x_3) x = ( x 1 , x 2 , x 3 ) be the mixed strategy for Player A, and let v v v be the value of the game. Then the linear program for Player A is:
maximize v subject to 0 x 1 − x 2 + x 3 ≥ v ( against b 1 ) x 1 + x 2 − x 3 ≥ v ( against b 2 ) − x 1 + x 2 − x 3 ≥ v ( against b 3 ) x 1 + x 2 + x 3 = 1 x 1 , x 2 , x 3 ≥ 0 \begin{aligned}
\operatorname{maximize}\quad&v\\
\text{subject to}\quad&0x_1-x_2+x_3\ge v \quad(\text{against }b_1)\\
&x_1+x_2-x_3\ge v \quad(\text{against }b_2)\\
&-x_1+x_2-x_3\ge v \quad(\text{against }b_3)\\
&x_1+x_2+x_3=1\\
&x_1,x_2,x_3\ge0
\end{aligned} maximize subject to v 0 x 1 − x 2 + x 3 ≥ v ( against b 1 ) x 1 + x 2 − x 3 ≥ v ( against b 2 ) − x 1 + x 2 − x 3 ≥ v ( against b 3 ) x 1 + x 2 + x 3 = 1 x 1 , x 2 , x 3 ≥ 0
Let y = ( y 1 , y 2 , y 3 ) y=(y_1,y_2,y_3) y = ( y 1 , y 2 , y 3 ) be the mixed strategy for Player B, and let w w w be the value of the game. Then the dual linear program is:
minimize w subject to 0 y 1 + y 2 − y 3 ≤ w ( for a 1 ) − y 1 + y 2 + y 3 ≤ w ( for a 2 ) y 1 − y 2 − y 3 ≤ w ( for a 3 ) y 1 + y 2 + y 3 = 1 y 1 , y 2 , y 3 ≥ 0 \begin{aligned}
\operatorname{minimize}\quad&w\\
\text{subject to}\quad&0y_1+y_2-y_3\le w \quad(\text{for }a_1)\\
&-y_1+y_2+y_3\le w \quad(\text{for }a_2)\\
&y_1-y_2-y_3\le w \quad(\text{for }a_3)\\
&y_1+y_2+y_3=1\\
&y_1,y_2,y_3\ge0
\end{aligned} minimize subject to w 0 y 1 + y 2 − y 3 ≤ w ( for a 1 ) − y 1 + y 2 + y 3 ≤ w ( for a 2 ) y 1 − y 2 − y 3 ≤ w ( for a 3 ) y 1 + y 2 + y 3 = 1 y 1 , y 2 , y 3 ≥ 0
By solving the above linear programs (or checking directly), we find that the optimal strategies are:
x = ( 0 , 1 2 , 1 2 ) , y = ( 1 2 , 0 , 1 2 ) , v = 0 x=\left(0,\frac{1}{2},\frac{1}{2}\right),\qquad y=\left(\frac{1}{2},0,\frac{1}{2}\right),\qquad v=0 x = ( 0 , 2 1 , 2 1 ) , y = ( 2 1 , 0 , 2 1 ) , v = 0
That is, Player A should mix evenly between a 2 a_2 a 2 and a 3 a_3 a 3 , while Player B should mix evenly between b 1 b_1 b 1 and b 3 b_3 b 3 . The value of the game is zero, indicating that with optimal play, neither player gains an advantage.
The analysis above assumes that the agents are not penalized for announcing pairs of states not in range(R R R ). If they are penalized, then the payoff matrix changes and the strategies might change accordingly.
Dynamically formed Tiglath-Pileser relations
In this section, we discuss how the analysis of the previous section can be refined to formally model persuasion situations which dynamically unfold over several episodes. Such situations allow us to account for the genesis of the Tiglath-Pileser relations, as they are learned by the principal, based on the their observations of what the Speaker declares. This formal analysis is applied in the two case studies of Sections 6 and 7.
The intention is to generate a sequence ( W n ∣ n ∈ N ) (W_n \mid n \in\mathbb{N}) ( W n ∣ n ∈ N ) of approximations of the Speaker’s winning set associated with codices which the principal extracts from the episodes of the narration. The guidance relations of these codices will generate the sequence of approximations of the winning set which can be used to confirm or update the principal’s beliefs about which claims from the narration are reliable.
Let us represent a story as a finite sequence e 1 , … , e k e_1,\ldots,e_k e 1 , … , e k of episodes (each of which is a consistent set of formulas).15 Let Prop be a non-empty set of propositional variables. Let P r o p V , P r o p D ⊆ P r o p \mathrm{Prop}^{V},\mathrm{Prop}^{D} \subseteq\mathrm{Prop} Prop V , Prop D ⊆ Prop such that P r o p V ∩ P r o p D = ∅ \mathrm{Prop}^{V} \cap\mathrm{Prop}^{D} = \varnothing Prop V ∩ Prop D = ∅ . Intuitively, P r o p V \mathrm{Prop}^{V} Prop V and P r o p D \mathrm{Prop}^{D} Prop D denote the atomic facts that are, respectively, verifiable and non-verifiable (or doubtful) by the principal.16 Then, if S S S denotes the set of *+ states of the world, episode e i e_i e i can be associated with the set17
E i ′ = { s ∈ S ∣ s ⊨ e i } . E_i' = \{s \in S \mid s \models e_i\}. E i ′ = { s ∈ S ∣ s ⊨ e i } .
Since we assume that each e i e_i e i is consistent, E i ′ ≠ ∅ E_i' \ne\varnothing E i ′ = ∅ . At each episode e i e_i e i , the principal learns the separated (but not complete and hence not pre-canonical) codex
L i ′ : = { ⋀ y ∈ Y φ y → φ x ∣ Y ⊆ P r o p V and x ∈ P r o p D and E i ′ ⊨ ⋀ y ∈ Y φ y → φ x and E i ′ ⊨ ⋀ y ∈ Y φ y } (4) L_i' := \left\{ \bigwedge_{y \in Y} \varphi_y \to\varphi_x \mid Y \subseteq\mathrm{Prop}^{V} \text{ and } x \in\mathrm{Prop}^{D} \text{ and } E_i' \models\bigwedge_{y \in Y} \varphi_y \to\varphi_x \text{ and } E_i' \models\bigwedge_{y \in Y} \varphi_y \right\} \tag*{(4)} L i ′ := ⎩ ⎨ ⎧ y ∈ Y ⋀ φ y → φ x ∣ Y ⊆ Prop V and x ∈ Prop D and E i ′ ⊨ y ∈ Y ⋀ φ y → φ x and E i ′ ⊨ y ∈ Y ⋀ φ y ⎭ ⎬ ⎫ (4)
where each φ z ∈ { z , ¬ z } \varphi_z \in\{z,\neg z\} φ z ∈ { z , ¬ z } for any z ∈ Y ∪ { x } z \in Y \cup\{x\} z ∈ Y ∪ { x } . Since E i ′ ≠ ∅ E_i' \ne\varnothing E i ′ = ∅ , it follows that L i ′ L_i' L i ′ is consistent. In what follows we write φ x ∂ \varphi_x^{\partial} φ x ∂ for ¬ x \neg x ¬ x if φ x = x \varphi_x = x φ x = x and x x x if φ x = ¬ x \varphi_x = \neg x φ x = ¬ x . We allow the case when Y = ∅ Y = \varnothing Y = ∅ , in which case ⋀ y ∈ Y φ y → φ x \bigwedge_{y \in Y}\varphi_y \to\varphi_x ⋀ y ∈ Y φ y → φ x reduces to φ x \varphi_x φ x .
We also assume the existence of a background theory Γ \Gamma Γ for the principal/reader, which is a set of formulas, the variables of which intersect the variables of the narration but add some new (background ) variables, namely the elements of var ( Γ ) ∖ P r o p = P r o p B \operatorname{var}(\Gamma) \setminus\mathrm{Prop} = \mathrm{Prop}_{B} var ( Γ ) ∖ Prop = Prop B . Moreover, every formula in Γ \Gamma Γ is of the form ⋀ x ∈ X ψ x → ψ y \bigwedge_{x \in X}\psi_x \to\psi_y ⋀ x ∈ X ψ x → ψ y where X ⊆ P r o p X \subseteq\mathrm{Prop} X ⊆ Prop , y ∈ P r o p ∪ P r o p B y \in\mathrm{Prop} \cup\mathrm{Prop}_{B} y ∈ Prop ∪ Prop B and ψ z ∈ { z , ¬ z , ◊ z , ◊ ¬ z } \psi_z \in\{z,\neg z,\Diamond z,\Diamond\neg z\} ψ z ∈ { z , ¬ z , ◊ z , ◊ ¬ z } for each z ∈ X ∪ { y } z \in X \cup\{y\} z ∈ X ∪ { y } . We read ◊ φ z \Diamond\varphi_z ◊ φ z as ‘it is possible that φ z \varphi_z φ z ’. Intuitively, the background theory encodes the general knowledge of the principal/reader which allows them to interpret the narrative. In particular, the variables in P r o p B \mathrm{Prop}_{B} Prop B represent meta facts or judgements about the story informed by e.g. social norms, psychological insights, historical facts etc.
At each episode e i e_i e i we form the set
B i ′ = { p ∣ p ∈ Prop B and e i , Γ ⊨ p } ∪ { ¬ p ∣ p ∈ Prop B and e i , Γ ⊨ ¬ p } , B'_i = \{p \mid p \in\operatorname{Prop}_B \text{ and } e_i,\Gamma\models p\} \cup\{\neg p \mid p \in\operatorname{Prop}_B \text{ and } e_i,\Gamma\models\neg p\}, B i ′ = { p ∣ p ∈ Prop B and e i , Γ ⊨ p } ∪ { ¬ p ∣ p ∈ Prop B and e i , Γ ⊨ ¬ p } ,
consisting of all literals in background variables logically entailed by the formulas describing the episode together the background theory. Intuitively, the sets B i ′ B'_i B i ′ represent the meta-facts that the principal deduces about the story and its characters in the current episode.
So far, the sets L i ′ L'_i L i ′ and B i ′ B'_i B i ′ encode information at the level of each episode; let us now define their cumulative counterparts: the sets B 1 , … , B k B_1,\ldots,B_k B 1 , … , B k are defined inductively as follows: B 1 = B 1 ′ B_1 = B'_1 B 1 = B 1 ′ . Assuming B i − 1 B_{i-1} B i − 1 has already been defined,
B i = { B i − 1 ∪ B i ′ if B i − 1 ∪ B i ′ is consistent; ( B i − 1 ∪ B i ′ ) ∖ { φ α ∈ B i − 1 ∪ B i ′ ∣ φ α ∈ B i − 1 ∪ B i ′ } otherwise. B_i =
\begin{cases}
B_{i-1} \cup B'_i & \text{if } B_{i-1} \cup B'_i \text{ is consistent;}\\
(B_{i-1} \cup B'_i) \setminus\{\varphi^\alpha\in B_{i-1} \cup B'_i \mid\varphi^\alpha\in B_{i-1} \cup B'_i\} & \text{otherwise.}
\end{cases} B i = { B i − 1 ∪ B i ′ ( B i − 1 ∪ B i ′ ) ∖ { φ α ∈ B i − 1 ∪ B i ′ ∣ φ α ∈ B i − 1 ∪ B i ′ } if B i − 1 ∪ B i ′ is consistent; otherwise.
Intuitively, the sets B i B_i B i collect together the meta-facts deduced by the principal up to that point in the story, as long as these are consistent. These cumulative meta facts allow the reader to form conclusions about the aims of the narrator. Any inconsistencies that arise witness and flag to the principal a break in the overall narrative continuity and aims of the narrator. Note that inconsistencies at the level of the background variables are far more consequential than inter-episode inconsistencies between non-background variables which might refer to circumstances which change as the story proceeds. In contrast, inconsistencies arising in the update of the B i B_i B i function as a switch between different modes of updating the cumulative codices L i L_i L i , inductively defined below.
Each codex L i L_i L i gives rise to a winning set W i : = { s ∈ S ∣ s ⊨ L i } = T ( L i ) W_i := \{s \in S \mid s \models L_i\} = T(L_i) W i := { s ∈ S ∣ s ⊨ L i } = T ( L i ) , and an associated guidance relation → L i \to_{L_i} → L i (cf. Definition 2.1 ). Intuitively, this guidance relation represents the principal’s approximation, up to this point in the narrative, of the narrator’s persuasion strategy, as formally captured by a Tiglath-Pileser relation. By Lemma 2.2 , T ( L i ) = T ( → L i ) T(L_i) = T(\to_{L_i}) T ( L i ) = T ( → L i ) , i.e. → L i ∣ W i × S = Δ W i \to_{L_i}|_{W_i \times S} = \Delta_{W_i} → L i ∣ W i × S = Δ W i , i.e. condition R2 holds for → L i \to_{L_i} → L i ; moreover, since the codex is separated, every s ∈ S s \in S s ∈ S is guided to some state in W i W_i W i , i.e. → L i − 1 [ W i ] = S \to_{L_i}^{-1}[W_i] = S → L i − 1 [ W i ] = S as it is always possible to change the truth values of consequents of rules without affecting the truth values of any antecedents. In particular, condition R1 holds. In general, condition R3 would not hold for these approximations.
Let us define L 1 , … , L k L_1,\ldots,L_k L 1 , … , L k inductively as follows; L 1 : = L 1 ′ L_1 := L'_1 L 1 := L 1 ′ . Assuming L i − 1 L_{i-1} L i − 1 has already been defined, there are two cases to consider:
If B i − 1 ∪ B i ′ B_{i-1} \cup B'_i B i − 1 ∪ B i ′ is consistent, then L i : = ( L i − 1 ∪ L i ′ ) ∖ Contr i L_i := (L_{i-1} \cup L'_i) \setminus\operatorname{Contr}_i L i := ( L i − 1 ∪ L i ′ ) ∖ Contr i where
Contr i = { ⋀ y ∈ Y φ y → φ x ∈ L i − 1 , ⋀ y ′ ∈ Y ′ φ y ′ → φ x α ∈ L i ′ | { φ y } y ∈ Y ⊆ { φ y ′ } y ′ ∈ Y ′ or { φ y ′ } y ′ ∈ Y ′ ⊆ { φ y } y ∈ Y } . \operatorname{Contr}_i =
\left\{
\bigwedge_{y\in Y}\varphi_y \to\varphi_x \in L_{i-1},\
\bigwedge_{y'\in Y'}\varphi_{y'} \to\varphi_x^\alpha\in L'_i
\;\middle|\;
\{\varphi_y\}_{y\in Y} \subseteq\{\varphi_{y'}\}_{y'\in Y'}
\text{ or }
\{\varphi_{y'}\}_{y'\in Y'} \subseteq\{\varphi_y\}_{y\in Y}
\right\}. Contr i = ⎩ ⎨ ⎧ y ∈ Y ⋀ φ y → φ x ∈ L i − 1 , y ′ ∈ Y ′ ⋀ φ y ′ → φ x α ∈ L i ′ { φ y } y ∈ Y ⊆ { φ y ′ } y ′ ∈ Y ′ or { φ y ′ } y ′ ∈ Y ′ ⊆ { φ y } y ∈ Y ⎭ ⎬ ⎫ .
As mentioned above, if B i − 1 ∪ B i ′ B_{i-1} \cup B'_i B i − 1 ∪ B i ′ is inconsistent, then this signals to the reader an apparent anomaly in the aims of the writer, insofar as the reader has inferred them up to this point. This leads the reader to conclude that the episode e i e_i e i does not provide information that could be used to update the current
approximation of the writer’s persuasion strategy, as represented by the guidance relation associated with codex L i − 1 L_{i-1} L i − 1 , nor the current approximation of their winning set W i − 1 W_{i-1} W i − 1 . Therefore, we set L i : = L i − 1 L_i := L_{i-1} L i := L i − 1 .
This allows us to define a sequence of sets of formulas T 1 ′ , … , T k ′ T'_1,\ldots,T'_k T 1 ′ , … , T k ′ , representing what the reader considers reliable information related to each episode.
T i ′ = { { φ x ∣ x ∈ Prop i V ∪ Prop i D and E i ′ ⊨ φ x } if B i − 1 ∪ B i ′ inconsistent and E i ′ ∩ W i − 1 = ∅ { φ x , ◊ φ y , ◊ φ y ∂ ∣ x ∈ Prop i V , y ∈ Prop i D , E i ′ ⊨ φ x , E i ′ ⊨ φ y } otherwise T'_i =
\begin{cases}
\{\varphi_x \mid x \in\operatorname{Prop}^{V}_i \cup\operatorname{Prop}^{D}_i \text{ and } E'_i \models\varphi_x\} & \text{if } B_{i-1} \cup B'_i \text{ inconsistent and } E'_i \cap W_{i-1} = \varnothing\\
\{\varphi_x,\Diamond\varphi_y,\Diamond\varphi^{\partial}_y \mid x \in\operatorname{Prop}^{V}_i, y \in\operatorname{Prop}^{D}_i, E'_i \models\varphi_x, E'_i \models\varphi_y\} & \text{otherwise}
\end{cases} T i ′ = { { φ x ∣ x ∈ Prop i V ∪ Prop i D and E i ′ ⊨ φ x } { φ x , ◊ φ y , ◊ φ y ∂ ∣ x ∈ Prop i V , y ∈ Prop i D , E i ′ ⊨ φ x , E i ′ ⊨ φ y } if B i − 1 ∪ B i ′ inconsistent and E i ′ ∩ W i − 1 = ∅ otherwise
Informally, if B i − 1 ∪ B i ′ B_{i-1} \cup B'_i B i − 1 ∪ B i ′ is inconsistent and E i ′ ∩ W i − 1 = ∅ E'_i \cap W_{i-1} = \varnothing E i ′ ∩ W i − 1 = ∅ , then in episode e i e_i e i the writer’s claims go against the current approximation W i − 1 W_{i-1} W i − 1 of their winning set and overarching aims as surmised by the principal up to this point in the narration and captured in B i − 1 B_{i-1} B i − 1 . Under the assumption that the narrator’s persuasion strategy is captured by a Tiglath-Pileser relation (i.e. satisfying R3), the principal is led to conclude that the narrator has been forced to truthfully declare something outside their winning set. In other words, the principal can deduce that the epistemic constraints under which the writer operates make it impossible for them to pursue their aims irrespective of the truth in the present episode; this allows the principal to regard as reliable information also the unverified claims at this point in the narration which fall outside the writer’s winning set.
Now, analogously to the sets B 1 , … , B k B_1,\ldots,B_k B 1 , … , B k , we define the sets T 1 , … , T k T_1,\ldots,T_k T 1 , … , T k , each of which representing what the reader considers reliable information up to that point in the narration. Preliminarily, for any set { ψ x ∣ x ∈ X } \{\psi_x \mid x \in X\} { ψ x ∣ x ∈ X } , we define the direct consequences of { ψ x ∣ x ∈ X } \{\psi_x \mid x \in X\} { ψ x ∣ x ∈ X } via Γ \Gamma Γ to be
D C Γ ( { ψ x ∣ x ∈ X } ) : = { ψ y ∣ ⋀ x ∈ X ′ ψ x → ψ y ∈ Γ and { ψ x ∣ x ∈ X ′ } ⊆ { ψ x ∣ x ∈ X } } . DC_{\Gamma}(\{\psi_x \mid x \in X\}) := \{\psi_y \mid\bigwedge_{x \in X'} \psi_x \to\psi_y \in\Gamma\text{ and } \{\psi_x \mid x \in X'\} \subseteq\{\psi_x \mid x \in X\}\}. D C Γ ({ ψ x ∣ x ∈ X }) := { ψ y ∣ x ∈ X ′ ⋀ ψ x → ψ y ∈ Γ and { ψ x ∣ x ∈ X ′ } ⊆ { ψ x ∣ x ∈ X }} .
Let T 1 : = T 1 ′ T_1 := T'_1 T 1 := T 1 ′ . Assuming T i − 1 T_{i-1} T i − 1 has already been defined,
T i = ( ( T i − 1 ∪ T i ′ ) ∪ D C Γ ( T i − 1 ∪ T i ′ ) ) ∖ T_i = ((T_{i-1} \cup T'_i) \cup DC_{\Gamma}(T_{i-1} \cup T'_i)) \setminus T i = (( T i − 1 ∪ T i ′ ) ∪ D C Γ ( T i − 1 ∪ T i ′ )) ∖
( { φ x ∈ D C Γ ( T i − 1 ∪ T i ′ ) ∣ φ x ∂ ∈ D C Γ ( T i − 1 ∪ T i ′ ) } ∪ { ◊ φ x , ◊ φ x ∂ ∣ φ x ∈ D C Γ ( T i − 1 ∪ T i ′ ) or φ x ∂ ∈ D C Γ ( T i − 1 ∪ T i ′ ) } ) (\{\varphi_x \in DC_{\Gamma}(T_{i-1} \cup T'_i) \mid\varphi^{\partial}_x \in DC_{\Gamma}(T_{i-1} \cup T'_i)\} \cup\{\Diamond\varphi_x,\Diamond\varphi^{\partial}_x \mid\varphi_x \in DC_{\Gamma}(T_{i-1} \cup T'_i) \text{ or } \varphi^{\partial}_x \in DC_{\Gamma}(T_{i-1} \cup T'_i)\}) ({ φ x ∈ D C Γ ( T i − 1 ∪ T i ′ ) ∣ φ x ∂ ∈ D C Γ ( T i − 1 ∪ T i ′ )} ∪ { ◊ φ x , ◊ φ x ∂ ∣ φ x ∈ D C Γ ( T i − 1 ∪ T i ′ ) or φ x ∂ ∈ D C Γ ( T i − 1 ∪ T i ′ )})
Note that, as suggested by the notation in the above definition, the only possible sources of inconsistency are the propositional literals, since formulas of the form ◊ φ x \Diamond\varphi_x ◊ φ x never occur negated in the elements of Γ \Gamma Γ . Hence, intuitively, the reliable information acquired by the reader up to episode e i e_i e i consists of the direct consequences acquired up to e i e_i e i where logical inconsistencies are removed, together with ‘doubts’, whenever ‘doubts’ are superseded by reliable factual information.
The procedure described above accounts for the formal representation of the ‘writer’s spin’ as understood by the principal/reader, in the form of the Tiglath-Pileser relation (i.e. the guidance relation associated with the codices gauged by the principal up to each episode in the narration), as well as the formation of the reader’s beliefs as to which pieces of information in the narration are reliable. However, note that the definition of the codices L i L_i L i leaves the reader vulnerable to the possibility that a strategic writer who understands this procedure could engineer the narration so as to manipulate the listener into believing unverified information. Therefore, a reader who is aware of this possibility might update the information they consider reliable in a different way, which we present below.
Update protocol for the strategic listener/reader. From the discussion above, it emerges that the way in which the listener can be manipulated into believing unverified information involves creating a series of episodes which first allow the reader to build some fundamental convictions about the aims of the narrative, only to create a narrative fracture (the fracture point being the episode at which B i − 1 ∪ B i ′ B_{i-1} \cup B_i' B i − 1 ∪ B i ′ is inconsistent) in which what is claimed is at odds with the general aims of the narrator as they are understood by the reader up to that point (as represented by E i ′ ∩ W i − 1 = ∅ E_i' \cap W_{i-1} = \varnothing E i ′ ∩ W i − 1 = ∅ ), and the reader’s fundamental convictions formed up to that point are contradicted. Engineering such a moment is expensive , as it not only requires significantly greater complications in weaving the story, but also has detrimental effects on achieving some of the ostensible narrative goals, e.g. concerning the reputation of some of the characters in the story, and revealing certain details about which the writer would be reasonably expected to rather keep quiet. As discussed above, the ordinary reader might be tempted to apply Ockham’s razor and assume that, if the writer got to a point in the story in which he had to make claims which are detrimental to his aims, then these claims could not have been avoided, softened or ‘given a spin’ (and this is why they can be trusted). However, the strategic reader might wonder whether the significant costs incurred by the writer could actually be worth their while ; that is, whether the benefits of being able to plant certain unverifiable pieces of information in the mind of the reader would justify the costs of this expensive maneuver. If these costs are justified, then the reader should better believe the opposite of what the writer claims concerning these variables. This is captured in the following, alternative update rule for T i ′ T_i' T i ′ :
T i ′ = { { φ x , φ y d ∣ x ∈ Prop i V , y ∈ Prop i D , E i ′ ⊨ φ x , E i ′ ⊨ φ y } if B i − 1 ∪ B i ′ inconsistent and E i ′ ∩ W i − 1 = ∅ { φ x , ◊ φ y , ◊ φ y d ∣ x ∈ Prop i V , y ∈ Prop i D , E i ′ ⊨ φ x , E i ′ ⊨ φ y } otherwise T_i' =
\begin{cases}
\{\varphi_x,\varphi_y^{d} \mid x \in\operatorname{Prop}_i^{V}, y \in\operatorname{Prop}_i^{D}, E_i' \models\varphi_x, E_i' \models\varphi_y\} & \text{if } B_{i-1} \cup B_i' \text{ inconsistent and } E_i' \cap W_{i-1} = \varnothing\\
\{\varphi_x,\Diamond\varphi_y,\Diamond\varphi_y^{d} \mid x \in\operatorname{Prop}_i^{V}, y \in\operatorname{Prop}_i^{D}, E_i' \models\varphi_x, E_i' \models\varphi_y\} & \text{otherwise}
\end{cases} T i ′ = { { φ x , φ y d ∣ x ∈ Prop i V , y ∈ Prop i D , E i ′ ⊨ φ x , E i ′ ⊨ φ y } { φ x , ◊ φ y , ◊ φ y d ∣ x ∈ Prop i V , y ∈ Prop i D , E i ′ ⊨ φ x , E i ′ ⊨ φ y } if B i − 1 ∪ B i ′ inconsistent and E i ′ ∩ W i − 1 = ∅ otherwise
At this juncture, a more sophisticated update protocol might handle different doubtful variables differently (e.g., negate, prefix with modalities, or remove them) depending on their relevance to the context of the episode. This will not be pursued here.
Modelling the story of the small businessman and the banker
The present section focuses on a case study from [6 ] :
Around 1950, there was a small businessman in Amsterdam whose main business was to sell the products of foreign textile manufacturers to Dutch clothing firms. Like most small businessmen in Amsterdam at the time, he banked with the Amsterdamsche Bank (which later became the present ABN AMRO). One day, he received an invitation to lunch with one of the directors of that bank. This invitation puzzled him a great deal, because he did not know this director personally, and a small businessman like him usually only dealt with bank employees at much lower levels. However, he accepted the invitation and showed up for the lunch at the top floor of the bank’s headquarters in the city centre. During the copious lunch, the bank director talked about all kinds of general subjects and asked the businessman’s opinion about the economic climate in Amsterdam. Rather than being flattered, he found it hard to imagine that he was invited to provide opinions about matters the bank knew better than he. When the dessert was served, the banker mentioned aside some other matter the name of a certain Amsterdam firm, which was an important client of the businessman. This firm, the bank director said, was doing very well under the present solid leadership. The small businessman realized that this must have been the point of the whole lunch. And if this large bank went to so much effort to increase the confidence of one small businessman in this firm, it must have been very important to the bank that the businessman believed that the firm was doing well.
The small businessman said he wanted to wash his hands, although coffee still needed to be served, but instead of walking to the bathroom he ran down the stairs and onto the street to find a telephone booth and call the office to stop all deliveries to that firm and also claim back any supplies that had already been delivered. Two weeks later, that firm went bankrupt and it turned out that the bank not only was its major creditor but also had preferential right to sell off any stocks in the possession of the firm to pay back the debt to the bank before other creditors would be satisfied.
In the following subsections, we present static and dynamic formal analyses of this case study as a persuasion scenario in which the banker is the Speaker A and the businessman is the Listener B B B .
Static analysis
Our account of the narration includes the following variables:
Table 1 (PDF p. 32).
The table above also reports the verifiable/doubtful/background status of each variable, in preparation for the discussion of § 6.2 . The average businessman’s background theory is given by:
Γ = { k ∧ p → n , q → ¬ n } . \Gamma= \{k \land p \to n, q \to\lnot n\}. Γ = { k ∧ p → n , q → ¬ n } .
In other words, if the bank is knowledgeable about the current business conditions and solves its clients’ problems then individual clients need the bank more than the bank needs each of them, and if the bank asks questions about the economic climate then it is not the case that individual clients need it more than it needs each of them. Let S = 2 P r o p S = 2^{\mathrm{Prop}} S = 2 Prop be the set of profiles for P r o p = { k , q , p , n , f } \mathrm{Prop} = \{k,q,p,n,f\} Prop = { k , q , p , n , f } . Let s 0 = ( 1 , 1 , 1 , 1 , 0 ) s_0 = (1,1,1,1,0) s 0 = ( 1 , 1 , 1 , 1 , 0 ) be true state. The interaction between an ordinary businessman and the banker corresponds to case 3 of the hierarchy discussed in § 4.1 . That is, the businessman B B B is strategic but believes that the banker A A A is non-strategic, while A A A is strategic, believes that B B B is strategic and that B B B believes A A A to be non-strategic. The ordinary businessman believes that the bank tries to appear knowledgeable whenever they can. This leads them to believe that the speaker’s winning set is [ [ k ] ] = { s ∈ S ∣ s ( k ) = 1 } [\![k]\!] = \{s \in S \mid s(k) = 1\} [ [ k ] ] = { s ∈ S ∣ s ( k ) = 1 } and that the speaker’s guidance relation R A ⊆ S × S R_A \subseteq S \times S R A ⊆ S × S is such that for any s ∈ S s \in S s ∈ S ,
if s ( k ) = 1 s(k) = 1 s ( k ) = 1 , then R A [ s ] = { s } R_A[s] = \{s\} R A [ s ] = { s } ,
if s ( k ) = 0 s(k) = 0 s ( k ) = 0 and s ′ ( y ) = s ( y ) s'(y) = s(y) s ′ ( y ) = s ( y ) for all y ∈ P r o p ∖ { k } y \in\mathrm{Prop} \setminus\{k\} y ∈ Prop ∖ { k } and s ′ ( k ) = 1 s'(k) = 1 s ′ ( k ) = 1 and T h ( s ′ ) ∪ Γ Th(s') \cup\Gamma T h ( s ′ ) ∪ Γ is consistent, then R A [ s ] = { s ′ } R_A[s] = \{s'\} R A [ s ] = { s ′ } ,
if s ( k ) = 0 s(k) = 0 s ( k ) = 0 and s ′ ⊭ Γ s' \not\models\Gamma s ′ ⊨ Γ , where s ′ s' s ′ is the k k k -variant of s s s s.t. s ′ ( k ) = 1 s'(k) = 1 s ′ ( k ) = 1 , then R A [ s ] = { s } R_A[s] = \{s\} R A [ s ] = { s } .
Note that R A ∈ T P ( S , [ [ k ] ] ) R_A \in\mathrm{TP}(S, [\![k]\!]) R A ∈ TP ( S , [ [ k ] ]) . In this particular case, the speaker claims ¬ k \neg k ¬ k , p p p , q q q , f f f . This is formalized as the speaker announcing the set of states S ′ = [ [ ¬ k ∧ q ∧ p ∧ f ] ] = { ( 0 , 1 , 1 , 0 , 1 ) , ( 0 , 1 , 1 , 1 , 1 ) } S' = [\![\neg k \land q \land p \land f]\!] = \{(0,1,1,0,1),(0,1,1,1,1)\} S ′ = [ [ ¬ k ∧ q ∧ p ∧ f ] ] = {( 0 , 1 , 1 , 0 , 1 ) , ( 0 , 1 , 1 , 1 , 1 )} . As per the discussion in § 4.1 , the ordinary businessman guesses that the true state is in R A − 1 [ S ′ ] = S ′ ⊆ [ [ f ] ] R_A^{-1}[S'] = S' \subseteq[\![f]\!] R A − 1 [ S ′ ] = S ′ ⊆ [ [ f ] ] . Hence, the banker succeeds in persuading the ordinary businessman that the firm is solid.
Unlike the ordinary businessman, the small businessman described in the story above believes that the banker is strategic. So, as in case 4 of § 4.1 , he cannot draw any conclusions regarding the true state without making further behavioural assumptions about the banker. The small businessman notices that, in this case, R A − 1 [ S ′ ] ∩ W A = ∅ R_A^{-1}[S'] \cap W_A = \varnothing R A − 1 [ S ′ ] ∩ W A = ∅ ; that is, the banker’s claims are inconsistent with the businessman’s assumptions about the banker’s winning set. However, since the businessman assumes that the banker is strategic, they do not come to same conclusion as the ordinary businessman, and look for a different explanation. Among the claims made by the banker, the only unverifiable one concerns the solidity of the firm. Hence, the businessman concludes that the banker’s actual winning set must have been [ [ f ] ] [\![f]\!] [ [ f ] ] . However, this static framework does not allow for making updates about the winning set of the speaker or their associated Tiglath-Pileser relations.
Dynamic analysis
For the sake of the dynamic analysis, we define the following two episodes, e 1 e_1 e 1 and e 2 e_2 e 2 :
e 1 e_1 e 1 previous relationship with the bank (the bank solves problems and sets the terms)
e 2 e_2 e 2 lunch
In the successive episodes the following propositions are asserted:
e 1 ∣ k ∧ p e 2 ∣ ¬ k ∧ q ∧ f \begin{array}{rcl}
e_1 &\mid& k \land p \\
e_2 &\mid& \neg k \land q \land f
\end{array} e 1 e 2 ∣ ∣ k ∧ p ¬ k ∧ q ∧ f
The codices learned from each of the episodes are:
L 1 ′ = { k , p → k } L 2 ′ = { ¬ k , f , q → ¬ k , q → f } L'_1 = \{k,p \to k\} \qquad L'_2 = \{\neg k,f,q \to\neg k,q \to f\} L 1 ′ = { k , p → k } L 2 ′ = { ¬ k , f , q → ¬ k , q → f }
Notice that, as we allow for the conjunction in the premise of each rule to be empty, the sets above include also the doubtful literals satisfied in the given episode. The sets B i ′ B'_i B i ′ for i = 1 , 2 i = 1, 2 i = 1 , 2 are computed as follows:
B 1 ′ = { n } B 2 ′ = { ¬ n } B'_1 = \{n\} \qquad B'_2 = \{\neg n\} B 1 ′ = { n } B 2 ′ = { ¬ n }
Since B 1 ∪ B 2 ′ = { n , ¬ n } B_1 \cup B'_2 = \{n, \neg n\} B 1 ∪ B 2 ′ = { n , ¬ n } is inconsistent, the sets B i B_i B i and L i L_i L i are computed as follows:
B 1 = { n } B 2 = ∅ L 1 = { k , p → k } L 2 = L 1 B_1 = \{n\} \qquad B_2 = \varnothing\qquad L_1 = \{k, p \to k\} \qquad L_2 = L_1 B 1 = { n } B 2 = ∅ L 1 = { k , p → k } L 2 = L 1
From the codex L 1 L_1 L 1 , the Listener extracts the approximation W 1 = [ [ k ] ] = { s ∈ 2 P r o p ∣ s ( k ) = 1 } W_1 = [\![k]\!] = \{s \in2^{\mathrm{Prop}} \mid s(k) = 1\} W 1 = [ [ k ] ] = { s ∈ 2 Prop ∣ s ( k ) = 1 } of the Speaker’s winning set. The set T 1 ′ T'_1 T 1 ′ consists of the verifiable literal p p p which is true at e 1 e_1 e 1 , together with the assertions of the possibility of the doubtful literals k k k and ¬ k \neg k ¬ k .
T 1 ′ = { p , ◊ k , ◊ ¬ k } T'_1 = \{p, \Diamond k, \Diamond\neg k\} T 1 ′ = { p , ◊ k , ◊ ¬ k }
In episode e 2 e_2 e 2 , the conditions indicating an epistemic fracture are verified: indeed, in addition to B 1 ∪ B 2 ′ = { n , ¬ n } B_1 \cup B'_2 = \{n, \neg n\} B 1 ∪ B 2 ′ = { n , ¬ n } being inconsistent, in episode e 2 e_2 e 2 , the banker makes declarations the truth-set of which is disjoint from W 1 W_1 W 1 . These two conditions trigger an extraordinary update of the set T 2 ′ T'_2 T 2 ′ , which can be performed in two possible ways:
T 2 ′ = { q , ¬ k , f } T 2 ′ = { q , k , ¬ f } T'_2 = \{q, \neg k, f\} \qquad T'_2 = \{q, k, \neg f\} T 2 ′ = { q , ¬ k , f } T 2 ′ = { q , k , ¬ f }
The set T 2 ′ T'_2 T 2 ′ above on the left reflects the update of a non-strategic Listener, as expected by the banker in the story; this Listener detects the epistemic fracture and believes that the banker’s claims in e 2 e_2 e 2 can be trusted precisely because they contradict the image that the bank had projected up to that point. In contrast, the T 2 ′ T'_2 T 2 ′ above on the right reflects the update of a strategic Listener, who recognizes the costs (in terms of time, effort, and projected image) involved in generating the epistemic fracture, and understands that the stakes involved in making the Listener believe the unverified claims must be high enough to justify this investment. While T 1 : = T 1 ′ T_1 := T'_1 T 1 := T 1 ′ , the two versions of T 2 ′ T'_2 T 2 ′ above give rise to the following two versions of the set T 2 T_2 T 2 , respectively:
T 2 = { p , ¬ k , q , f } T 2 = { p , k , q , ¬ f } T_2 = \{p, \neg k, q, f\} \qquad T_2 = \{p, k, q, \neg f\} T 2 = { p , ¬ k , q , f } T 2 = { p , k , q , ¬ f }
These two sets have been obtained by taking the union of T 1 T_1 T 1 and T 2 ′ T'_2 T 2 ′ , adding their direct consequences, and removing the doubts whenever they are superseded by reliable information; however, the two different updates of T 2 ′ T'_2 T 2 ′ make a crucial difference: by e 2 e_2 e 2 , the conclusion that the firm is solid has been planted in the mind of the non-strategic businessman, while the strategic businessman has concluded the exact opposite.
Modelling Baruch Halpern’s account of King David’s story
We now turn our attention to the analysis of Baruch Halpern’s account of the King David story, as already discussed in the introduction. In § 7.1 , we provide an informal account of the main episodes of this story, together with Halpern’s analysis. This is formalized in § 7.2 using the framework introduced in § 5 .
The informal analysis
In the present section, we summarize some of the main episodes of King David’s story as related in 1 Samuel, 2 Samuel, and the opening of 1 Kings, in the light of Baruch Halpern’s analysis [11 ] .
Ancient royal historiography is often imagined as straightforward propaganda: kings win, enemies fall, disasters disappear from the record. The classic model according to Halpern is the Assyrian court chronicle, the sort of inscription associated with rulers like Tiglath-Pileser III, which can be interpreted through the Tiglath-Pileser principle, discussed in the introduction. The governing rule of these chronicles is simple: if something happened and it was not a disaster, narrate it as a resounding triumph; if it was a disaster, pass over it in silence as far as possible without completely compromising the overall credibility of the report given what the purported original audience knows. The result is what we may term the “Nineveh Times” style of news-reporting: observable 18 facts are reported, but selectively, and always favourably. These observables might be embellished with unobservables, e.g. claims of overwhelming victory or divine favour. In this mode of writing, readers are trained in a negative deduction procedure, which allows them to infer or approximate the truth both from the ‘spin’ with which the writer seems to present the narration and from the writer’s omissions. Since the audience is aware that unfavorable news is not recorded, the propagandistic effect of the narration is weakened, as readers of the “Nineveh Times” may wonder and worry about what they are not told.
The narrative of David in 1 Samuel, 2 Samuel, and the opening of 1 Kings adopts a more sophisticated strategy. At first glance, it appears to abandon the “Nineveh Times” style of news-reporting, in so far as it includes potentially unfavourable information the “Nineveh Times” might have preferred to omit: civil war, court intrigue, dynastic rebellion, adultery, arranged death are displayed in broad daylight. The historian seems almost reckless in his willingness to expose the king to suspicion. What we may call the “Jerusalem Times” style of news-reporting operates differently. It creates the impression that all newsworthy observables are reported, including those that could damage the royal image: the narrator not only acknowledges these facts, but also supplies additional (often unverifiable) details , especially at precisely those points where the reader might expect less than full disclosure. Because the narrator demonstrably anticipates and flags potential suspicions, the reader feels no need to search for hidden scandals beyond those identified in the text. Suspicion is not erased, it is managed.
This creates a new dynamic. Unlike the Assyrian reader, who discounts good news and assumes suppression of bad, the reader of Samuel is trained to believe that the unpleasant facts have already been disclosed. The narrator flatters the reader’s intelligence by openly wrestling with embarrassing material and by going to elaborate lengths to explain, contextualize, or morally frame it. Even when the explanations are not wholly convincing, the gesture itself constrains the field of suspicion, as the reader’s doubt is thereby focused on the issues the narrator signals, and nowhere else.
The episode of Bathsheba will prove pivotal in this more sophisticated narrative strategy. Up to this point, the writer has portrayed David in as positive a light as possible. Although the narrator has described many events which could have put the king in an unfavourable light, he has gone to great lengths to explain away anything in the narrative that could have suggested that the king was at fault; for instance, when an enemy dies in suspicious circumstances, it is explained that David had not ordered his death and was very distraught about it. However, now it appears that the writer gratuitously sacrifices this carefully built image, telling a story that he could have narrated far more favourably. But, according to Halpern [11 ] , once the larger narrative strategy is understood, this “sacrifice” can be seen as functional: it reinforces the impression of total candour, and closes off more dangerous lines of doubt. The writer does not suppress incriminating observables, and he even adds seemingly “gratuitous” unobservables, such as the secret letter from David to Joab. By doing so, he creates a far more resilient form of royal apology than any Assyrian victory inscription could achieve.
What follows, then, is not just a retelling of key episodes in David’s life, but an exploration of how their narration differs from the Assyrian model, and how this difference calls for a more sophisticated application of the Tiglath-Pileser principle, leading to a new understanding of the narrator’s real agenda, which may be less about glorifying David, and more about strengthening the legitimacy of the House of David as a royal lineage, especially from Solomon onwards.
Episode 1. David first appears in the narrative when the prophet Samuel anoints him in secret while Saul still reigns (1 Samuel 16). The act is quiet but politically explosive: the divine spirit departs from Saul and comes upon David. From that moment, tension is unavoidable. The narrative does not conceal the enmity that follows. Saul’s jealousy, his attempts on David’s life, David’s life as a fugitive—these are all laid out in full view. A court chronicle in the “Nineveh Times” style would likely have erased such a dynastic fracture. Here, however, the reader sees it all. Yet at every critical juncture David is shown refusing to harm “the LORD’s anointed.” Twice he spares Saul when he could have killed him (1 Samuel 24, 26). The embarrassing reality of civil hostility is fully exposed, but the interpretation is steered: David is not an opportunist; he is the reluctant heir, waiting for God’s timing. Because the text does not hide the tension, its exculpation of David feels earned.
Episode 2. The episode of Nabal (1 Samuel 25) sharpens this pattern. After protecting the flocks of the wealthy farmer Nabal, David demands a payoff. When Nabal refuses, David leads a detachment of his men to massacre Nabal and his people. The narrative does not hide the basic structure of a protection racket David is organizing. Violence is only prevented when Nabal’s wife Abigail intercepts David with gifts of food and drink. After her return, Nabal dies struck down by God, and David marries his widow Abigail. On the face of it, this narrative could have been troubling: a small-scale gang leader threatens a wealthy landowner who soon conveniently dies. Yet, the story’s willingness to expose David’s protection racket works as a token of trust, which—even if the reader was suspicious of David’s involvement in Nabal’s death—builds confidence that nothing worse is being hidden.
Episode 3. When Saul and his three sons die in a battle against the Philistines (1 Samuel 31, 1 Chronicles 10), the political beneficiary is obviously David. The text does not minimize this, nor does it hide the fact that, at the time of this battle, David and his warriors were under the protection and in the service of the Philistine king Achish. However, the narrator provides an elaborate storyline that explains why David and his forces were not there on the day of the battle. Three days later, an Amalekite from the Israelite army camp brings David news of the defeat and death of Saul and his sons, and claims to have delivered the final blow to Saul (2 Samuel 1). David executes him for daring to strike the LORD’s anointed, and composes a lament of extraordinary beauty over Saul and Jonathan. Again, the suspicious circumstance that David profited from his rival’s death is fully visible. The narrative counters, not by denying the advantage, but by emphasizing David’s grief and his principled refusal to legitimize regicide. The narrator expends considerable effort here, almost ostentatiously, to distance David from any whiff of complicity. In Halpern’s terms, this is one of the early “sacrifices” of narrative energy: the narrator works hard to defend the king where suspicion is plausible. This narrative strategy habituates the skeptical reader to expect David to be elaborately exculpated whenever the facts allow for it. Then comes the Bathsheba episode.
Episode 4. In 2 Samuel 11, while David’s army is laying siege to the capital of the Ammonites, David seduces Bathsheba, wife of Uriah the Hittite, an elite soldier in David’s army. Bathsheba falls pregnant and David attempts to cover up his adultery by summoning Uriah home from the front. The narrative lingers over the details: Uriah refuses to go down to his own house and sleep with his wife; he sleeps at the palace entrance; David sends him back to battle carrying the letter to Joab in which he orders Joab to organize an attack on the city walls where Uriah is likely to be killed. A host of incriminating unobservables (first and foremost the secret letter to Joab) are placed in full view of the reader.
Up to this point in the larger story, the reader has been trained to expect that such dangers to David’s reputation will be offset. But here no attempt at exculpation is made; in fact, the story emphasises David’s failings and punishment: the prophet Nathan confronts David (1 Samuel 12) and prophesies violence against David’s house. The first child born to David and Bathsheba dies after the confrontation with Nathan. Only after all this is over do they conceive Solomon, who in this way is born at many removes from “sin”.
This is the second and far more dramatic “ostensible sacrifice”. The narrative that has hitherto so carefully shielded David now exposes him. Because the narrator has demonstrated his willingness to tell the worst possible story about a king betraying a loyal officer in his own army, the reader is disinclined to suspect that something even worse lies concealed as, apparently, the worst has already been confessed.
Here, Halpern asks a pointed question. The narrator did not need to tell the story this way. He could have portrayed Uriah as a loyal hero who fell bravely in battle, and David as compassionately marrying his widow. Instead, he chose to blacken David’s character with adultery and murder. Why make such a sacrifice?
Halpern’s suggestion is that this “ostensible sacrifice” of David’s image serves a larger purpose. By narrating the sin and its punishment, the historian seals off a more dangerous suspicion; namely, that Solomon might actually be Uriah’s son. The narrator’s trustworthiness seems to be guaranteed precisely by his display of this plethora of damning unobservable details without any attempt to place them in a better light. The sequence is emphatic: conception in sin, exposure, judgment, death of the first son; then a second son, Solomon, is born after the reckoning. This creates a narrative buffer to protect Solomon’s paternity from doubt. Because the narrator was willing to humiliate David, the reader is relieved of having to imagine worse scenarios.
Episode 5. The consequences of David’s sin unfold in the tragedy of Absalom. A fraught relationship between David and Absalom (1 Samuel 13 - 14) culminates in Absalom rebelling and usurping the throne (1 Samuel 15). Absalom’s popularity is not hidden by the narration: the Israelites side with him, and David has to flee Jerusalem in humiliation. David regains the throne only thanks to his professional army of foreign mercenaries. The revolt nearly tears the kingdom apart. When Absalom is killed by Joab despite David’s express and public command not to harm him, the king’s grief overwhelms him. Here we see the old pattern of exculpation resumed and reinforced in the mind of the reader: David clearly benefits from Absalom’s death, but the details of the narrative are arranged to clear him of any complicity. The civil war is not minimized, but it is framed as the consequences of David’s own sin, in accordance with Nathan’s prophecy. Hence, the narrative frames civil war as a punishment for moral turpitude; the instability of David’s reign is acknowledged, yet contained within a moral explanation.
Episode 6. In 1 Kings 1, the reader finds king David very advanced in years and in failing health. His son Adonijah emerges as his natural heir, as the narrative makes clear: he is David’s son, older than Solomon, and his public celebrations and broad support (also by Joab, the head of the army) suggest that he is precisely the successor everyone would expect. However, Bathsheba and the prophet Nathan conspire to extract from the apparently senile David a declaration in favour of Solomon’s succession and instructions to have him publicly anointed as king. This is carried out, and Adonijah, Joab, and their followers are killed.
Here the earlier Bathsheba narrative does its quiet work. The succession struggle is framed as a contest between two sons of the same king. Because the Bathsheba episode has already exhausted the reader’s suspicion by exposing adultery, conspiracy, and punishment, the possibility that Solomon might be the spearhead of a coup organized by David’s personal guard does not rise on the reader’s horizon. The dark story that seemed so damaging to David has, paradoxically, made Solomon’s legitimacy narratively secure.
In this light, the Bathsheba episode is not gratuitous confession, but rather, strategic genius. The narrator sacrifices David’s moral image in order to fortify the dynasty’s continuity. The reader, having seen the apparent worst, feels no need to search for something worse still, and that is precisely why the most dangerous question of all—the question of Solomon’s paternity—never quite occurs.
The formalization
In line with the discussion of § 7.1 , we consider the following episodes:
e 1 e_1 e 1 Samuel anoints David.
e 2 e_2 e 2 Death of Nabal.
e 3 e_3 e 3 Death of Saul and Jonathan.
e 4 e_4 e 4 Death of Uriah. e 5 e_{5} e 5 Death of Absalom.
e 6 e_{6} e 6 Succession of Solomon.
Our account of the narration includes the following variables:19
where the labels V, D, B stand for ‘verifiable’, ‘doubtful’, ‘background’, respectively. The literals corresponding to these variables will be evaluated in the successive episodes of the narrative, but the ultimate intention is to interpret them over the narrative as a whole. Therefore if a literal is concluded at the end of the narrative, we read that as meaning that it held stably true in the last stretch of the narration, which could include that it held throughout.
In the successive episodes the following propositions are asserted:
e 1 g ∧ a e 2 a ∧ a ′ ′ ∧ ¬ i ∧ ¬ e ∧ o 1 e 3 a ∧ ¬ a ′ ′ ∧ ¬ i ∧ e ∧ o 2 e 4 a ∧ i ∧ ¬ e ∧ o 2 ∧ s e 5 a ∧ ¬ a ′ ′ ∧ ¬ i ∧ e ∧ o 2 e 6 h ∧ p ∧ j ∧ k ∧ t ∧ m ∧ c ∧ ¬ d \begin{aligned}
e_{1} &\quad g \land a \\
e_{2} &\quad a \land a'' \land\lnot i \land\lnot e \land o_{1} \\
e_{3} &\quad a \land\lnot a'' \land\lnot i \land e \land o_{2} \\
e_{4} &\quad a \land i \land\lnot e \land o_{2} \land s \\
e_{5} &\quad a \land\lnot a'' \land\lnot i \land e \land o_{2} \\
e_{6} &\quad h \land p \land j \land k \land t \land m \land c \land\lnot d
\end{aligned} e 1 e 2 e 3 e 4 e 5 e 6 g ∧ a a ∧ a ′′ ∧ ¬ i ∧ ¬ e ∧ o 1 a ∧ ¬ a ′′ ∧ ¬ i ∧ e ∧ o 2 a ∧ i ∧ ¬ e ∧ o 2 ∧ s a ∧ ¬ a ′′ ∧ ¬ i ∧ e ∧ o 2 h ∧ p ∧ j ∧ k ∧ t ∧ m ∧ c ∧ ¬ d
The set of profiles is therefore S = 2 Prop ′ S = 2^{\operatorname{Prop}^{\prime}} S = 2 Prop ′ , where Prop ′ \operatorname{Prop}^{\prime} Prop ′ is the set of variables in the table above, i.e., Prop ′ = Prop ∪ Prop V ∪ Prop D \operatorname{Prop}^{\prime} = \operatorname{Prop} \cup\operatorname{Prop}_{V} \cup\operatorname{Prop}_{D} Prop ′ = Prop ∪ Prop V ∪ Prop D .
As described above, we associate the following sequence of sets of profiles with the sequence of episodes:
E 1 ′ = { s ∈ S ∣ s ( g ) = 1 , s ( a ′ ) = 1 } E^{\prime}_{1} = \{s \in S \mid s(g) = 1, s(a^{\prime}) = 1\} E 1 ′ = { s ∈ S ∣ s ( g ) = 1 , s ( a ′ ) = 1 }
E 2 ′ = { s ∈ S ∣ s ( a ) = 1 , s ( a ′ ′ ) = 1 , s ( i ) = 0 , s ( e ) = 0 , s ( o 1 ) = 1 } E^{\prime}_{2} = \{s \in S \mid s(a) = 1, s(a^{\prime\prime}) = 1, s(i) = 0, s(e) = 0, s(o_{1}) = 1\} E 2 ′ = { s ∈ S ∣ s ( a ) = 1 , s ( a ′′ ) = 1 , s ( i ) = 0 , s ( e ) = 0 , s ( o 1 ) = 1 }
E 3 ′ = { s ∈ S ∣ s ( a ) = 1 , s ( a ′ ′ ) = 0 , s ( i ) = 0 , s ( e ) = 1 , s ( o 2 ) = 1 } E^{\prime}_{3} = \{s \in S \mid s(a) = 1, s(a^{\prime\prime}) = 0, s(i) = 0, s(e) = 1, s(o_{2}) = 1\} E 3 ′ = { s ∈ S ∣ s ( a ) = 1 , s ( a ′′ ) = 0 , s ( i ) = 0 , s ( e ) = 1 , s ( o 2 ) = 1 }
E 4 ′ = { s ∈ S ∣ s ( a ) = 1 , s ( i ) = 1 , s ( e ) = 0 , s ( o 2 ) = 1 , s ( s ) = 1 } E^{\prime}_{4} = \{s \in S \mid s(a) = 1, s(i) = 1, s(e) = 0, s(o_{2}) = 1, s(s) = 1\} E 4 ′ = { s ∈ S ∣ s ( a ) = 1 , s ( i ) = 1 , s ( e ) = 0 , s ( o 2 ) = 1 , s ( s ) = 1 }
E 5 ′ = { s ∈ S ∣ s ( a ) = 1 , s ( a ′ ′ ) = 0 , s ( i ) = 0 , s ( e ) = 1 , s ( o 2 ) = 1 } E^{\prime}_{5} = \{s \in S \mid s(a) = 1, s(a^{\prime\prime}) = 0, s(i) = 0, s(e) = 1, s(o_{2}) = 1\} E 5 ′ = { s ∈ S ∣ s ( a ) = 1 , s ( a ′′ ) = 0 , s ( i ) = 0 , s ( e ) = 1 , s ( o 2 ) = 1 }
E 6 ′ = { s ∈ S ∣ s ( h ) = s ( p ) = s ( j ) = s ( k ) = s ( t ) = s ( m ) = s ( c ) = 1 , s ( d ) = 0 } E^{\prime}_{6} = \{s \in S \mid s(h) = s(p) = s(j) = s(k) = s(t) = s(m) = s(c) = 1, s(d) = 0\} E 6 ′ = { s ∈ S ∣ s ( h ) = s ( p ) = s ( j ) = s ( k ) = s ( t ) = s ( m ) = s ( c ) = 1 , s ( d ) = 0 }
From this the following episode-related codices are generated:
L 1 ′ = { g , a ′ } L^{\prime}_{1} = \{g,a^{\prime}\} L 1 ′ = { g , a ′ }
L 2 ′ = { u , a → u , ¬ e → u , a ∧ ¬ e → u ∣ u ∈ { a ′ ′ , ¬ i , o 1 } } L^{\prime}_{2} = \{u, a \to u, \neg e \to u, a \land\neg e \to u \mid u \in\{a^{\prime\prime}, \neg i, o_{1}\}\} L 2 ′ = { u , a → u , ¬ e → u , a ∧ ¬ e → u ∣ u ∈ { a ′′ , ¬ i , o 1 }}
L 3 ′ = { v , a → v , e → v , a ∧ e → v ∣ v ∈ { ¬ a ′ ′ , ¬ i , o 2 } } L^{\prime}_{3} = \{v, a \to v, e \to v, a \land e \to v \mid v \in\{\neg a^{\prime\prime}, \neg i, o_{2}\}\} L 3 ′ = { v , a → v , e → v , a ∧ e → v ∣ v ∈ { ¬ a ′′ , ¬ i , o 2 }}
L 4 ′ = { ⋀ Y → x ∣ Y ⊆ { a , ¬ e , o 2 } , x ∈ { i , s } } L^{\prime}_{4} = \left\{\bigwedge Y \to x \mid Y \subseteq\{a,\neg e,o_{2}\}, x \in\{i,s\}\right\} L 4 ′ = { ⋀ Y → x ∣ Y ⊆ { a , ¬ e , o 2 } , x ∈ { i , s } }
L 5 ′ = L 3 ′ L^{\prime}_{5} = L^{\prime}_{3} L 5 ′ = L 3 ′
L 6 ′ = { ⋀ W → v ′ ∣ W ⊆ { p , j , k , t , m } and v ′ ∈ { h , c , ¬ d } } L^{\prime}_{6} = \left\{\bigwedge W \to v^{\prime} \mid W \subseteq\{p,j,k,t,m\}\text{ and }v^{\prime} \in\{h,c,\neg d\}\right\} L 6 ′ = { ⋀ W → v ′ ∣ W ⊆ { p , j , k , t , m } and v ′ ∈ { h , c , ¬ d } }
We assume that the background theory of the reader is given by
Γ = { o 2 → ¬ o 1 , g ∧ a ′ → b , a ∧ i → ¬ b , ◊ ¬ s ∧ ◊ ¬ h → ◊ f , ¬ s ∧ ◊ ¬ h → ◊ f } \Gamma= \{o_{2} \to\neg o_{1}, g \land a^{\prime} \to b, a \land i \to\neg b, \Diamond\neg s \land\Diamond\neg h \to\Diamond f, \neg s \land\Diamond\neg h \to\Diamond f\} Γ = { o 2 → ¬ o 1 , g ∧ a ′ → b , a ∧ i → ¬ b , ◊ ¬ s ∧ ◊ ¬ h → ◊ f , ¬ s ∧ ◊ ¬ h → ◊ f }
We therefore get B 1 ′ = { b } B^{\prime}_{1} = \{b\} B 1 ′ = { b } , B 2 ′ = B 3 ′ = B 5 ′ = B 6 ′ = ∅ B^{\prime}_{2} = B^{\prime}_{3} = B^{\prime}_{5} = B^{\prime}_{6} = \varnothing B 2 ′ = B 3 ′ = B 5 ′ = B 6 ′ = ∅ , B 4 ′ = { ¬ b } B^{\prime}_{4} = \{\neg b\} B 4 ′ = { ¬ b } . And hence B 1 = B 2 = B 3 = { b } B_{1} = B_{2} = B_{3} = \{b\} B 1 = B 2 = B 3 = { b } , then B 4 = B 5 = B 6 = ∅ B_{4} = B_{5} = B_{6} = \varnothing B 4 = B 5 = B 6 = ∅ .
We are now ready to calculate the codices L i L_{i} L i :
L 1 = { g , a ′ } L_{1} = \{g,a^{\prime}\} L 1 = { g , a ′ }
L 2 = { g , a ′ ′ u , a → u , ¬ e → u , a ∧ ¬ e → u ∣ u ∈ { a ′ ′ , ¬ i , o 1 } } L_{2} = \{g,a^{\prime\prime}u, a \to u, \neg e \to u, a \land\neg e \to u \mid u \in\{a^{\prime\prime},\neg i,o_{1}\}\} L 2 = { g , a ′′ u , a → u , ¬ e → u , a ∧ ¬ e → u ∣ u ∈ { a ′′ , ¬ i , o 1 }}
L 3 = { g , a ′ ′ , ¬ i , o 1 , o 2 } ∪ { a → u ′ ∣ u ′ ∈ { ¬ i , o 1 , o 2 } } ∪ { y ′ → u ∣ y ′ ∈ { e , e ∧ a } , u ∈ { a ′ ′ , ¬ i , o 1 } } ∪ { y → v ∣ y ∈ { ¬ e , ¬ e ∧ a } , v ∈ { ¬ a ′ ′ , ¬ i , o 2 } } \begin{aligned}
L_{3} ={}& \{g,a^{\prime\prime},\neg i,o_{1},o_{2}\} \cup\{a \to u^{\prime} \mid u^{\prime} \in\{\neg i,o_{1},o_{2}\}\}\\
&\cup\{y^{\prime} \to u \mid y^{\prime} \in\{e,e \land a\},u \in\{a^{\prime\prime},\neg i,o_{1}\}\}\\
&\cup\{y \to v \mid y \in\{\neg e,\neg e \land a\},v \in\{\neg a^{\prime\prime},\neg i,o_{2}\}\}
\end{aligned} L 3 = { g , a ′′ , ¬ i , o 1 , o 2 } ∪ { a → u ′ ∣ u ′ ∈ { ¬ i , o 1 , o 2 }} ∪ { y ′ → u ∣ y ′ ∈ { e , e ∧ a } , u ∈ { a ′′ , ¬ i , o 1 }} ∪ { y → v ∣ y ∈ { ¬ e , ¬ e ∧ a } , v ∈ { ¬ a ′′ , ¬ i , o 2 }}
L 4 = L 3 L_{4} = L_{3} L 4 = L 3
L 5 = L 4 ∪ L 5 ′ = L 3 ∪ L 3 ′ = L 3 L_{5} = L_{4} \cup L^{\prime}_{5} = L_{3} \cup L^{\prime}_{3} = L_{3} L 5 = L 4 ∪ L 5 ′ = L 3 ∪ L 3 ′ = L 3
L 6 = L 5 ∪ L 6 ′ . L_{6} = L_{5} \cup L^{\prime}_{6}. L 6 = L 5 ∪ L 6 ′ .
From the codices above, we can derive their corresponding winning sets W i : = T ( L i ) W_i := T(L_i) W i := T ( L i ) .
Let us compute the sets T i ′ T'_i T i ′ : T 1 ′ = { ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } T'_1 = \{\lozenge g,\lozenge\neg g,\lozenge a',\lozenge\neg a'\} T 1 ′ = { ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } , since there are no verifiable variables and there are no inconsistencies in B 1 B_1 B 1 ; T 2 ′ = { a , ¬ e , ◊ i , ◊ ¬ i , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ o 1 , ◊ ¬ o 1 } T'_2 = \{a,\neg e,\lozenge i,\lozenge\neg i,\lozenge a'',\lozenge\neg a'',\lozenge o_1,\lozenge\neg o_1\} T 2 ′ = { a , ¬ e , ◊ i , ◊ ¬ i , ◊ a ′′ , ◊ ¬ a ′′ , ◊ o 1 , ◊ ¬ o 1 } ; T 3 ′ = { a , e , o 2 , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ i , ◊ ¬ i } T'_3 = \{a,e,o_2,\lozenge a'',\lozenge\neg a'',\lozenge i,\lozenge\neg i\} T 3 ′ = { a , e , o 2 , ◊ a ′′ , ◊ ¬ a ′′ , ◊ i , ◊ ¬ i } ; the interesting one is T 4 ′ = { a , i , ¬ e , o 2 , s } T'_4 = \{a,i,\neg e,o_2,s\} T 4 ′ = { a , i , ¬ e , o 2 , s } ; notice that i , s i,s i , s are ‘non verifiable’ literals; however, the reader considers their truth reliable since W 3 ⊆ { s ∈ S ∣ s ⊨ ¬ i } W_3 \subseteq\{s \in S \mid s \models\neg i\} W 3 ⊆ { s ∈ S ∣ s ⊨ ¬ i } and E 4 ⊆ { s ∈ S ∣ s ⊨ i } E_4 \subseteq\{s \in S \mid s \models i\} E 4 ⊆ { s ∈ S ∣ s ⊨ i } , hence W 3 ∩ E 4 = ∅ W_3 \cap E_4 = \varnothing W 3 ∩ E 4 = ∅ , and moreover B 3 ∪ B 4 ′ = { b , ¬ b } B_3 \cup B'_4 = \{b,\neg b\} B 3 ∪ B 4 ′ = { b , ¬ b } is inconsistent. Then T 5 ′ = { a , e , o 2 , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ i , ◊ ¬ i } T'_5 = \{a,e,o_2,\lozenge a'',\lozenge\neg a'',\lozenge i,\lozenge\neg i\} T 5 ′ = { a , e , o 2 , ◊ a ′′ , ◊ ¬ a ′′ , ◊ i , ◊ ¬ i } and T 6 ′ = { p , j , k , t , m , ◊ c , ◊ ¬ c , ◊ h , ◊ ¬ h , ◊ d , ◊ ¬ d } T'_6 = \{p,j,k,t,m,\lozenge c,\lozenge\neg c,\lozenge h,\lozenge\neg h,\lozenge d,\lozenge\neg d\} T 6 ′ = { p , j , k , t , m , ◊ c , ◊ ¬ c , ◊ h , ◊ ¬ h , ◊ d , ◊ ¬ d } .
We now calculate the sets T i T_i T i as follows:
T 1 = { ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } T_1 = \{\lozenge g,\lozenge\neg g,\lozenge a',\lozenge\neg a'\} T 1 = { ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ }
T 2 = { a , ¬ e , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ i , ◊ ¬ i , ◊ o 1 , ◊ ¬ o 1 , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } T_2 = \{a,\neg e,\lozenge a'',\lozenge\neg a'',\lozenge i,\lozenge\neg i,\lozenge o_1,\lozenge\neg o_1,\lozenge g,\lozenge\neg g,\lozenge a',\lozenge\neg a'\} T 2 = { a , ¬ e , ◊ a ′′ , ◊ ¬ a ′′ , ◊ i , ◊ ¬ i , ◊ o 1 , ◊ ¬ o 1 , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ }
T 3 = { a , o 2 , ¬ o 1 , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ i , ◊ ¬ i , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } T_3 = \{a,o_2,\neg o_1,\lozenge a'',\lozenge\neg a'',\lozenge i,\lozenge\neg i,\lozenge g,\lozenge\neg g,\lozenge a',\lozenge\neg a'\} T 3 = { a , o 2 , ¬ o 1 , ◊ a ′′ , ◊ ¬ a ′′ , ◊ i , ◊ ¬ i , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ }
T 4 = { a , ¬ e , i , o 2 , ¬ o 1 , s , ¬ b , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } T_4 = \{a,\neg e,i,o_2,\neg o_1,s,\neg b,\lozenge a'',\lozenge\neg a'',\lozenge g,\lozenge\neg g,\lozenge a',\lozenge\neg a'\} T 4 = { a , ¬ e , i , o 2 , ¬ o 1 , s , ¬ b , ◊ a ′′ , ◊ ¬ a ′′ , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ }
T 5 = { a , i , o 2 , ¬ o 1 , s , ¬ b , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } T_5 = \{a,i,o_2,\neg o_1,s,\neg b,\lozenge a'',\lozenge\neg a'',\lozenge g,\lozenge\neg g,\lozenge a',\lozenge\neg a'\} T 5 = { a , i , o 2 , ¬ o 1 , s , ¬ b , ◊ a ′′ , ◊ ¬ a ′′ , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ }
T 6 = { a , i , o 2 , ¬ o 1 , s , ¬ b , p , j , k , t , m , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ , ◊ c , ◊ ¬ c , ◊ h , ◊ ¬ h , ◊ d , ◊ ¬ d } T_6 = \{a,i,o_2,\neg o_1,s,\neg b,p,j,k,t,m,\lozenge a'',\lozenge\neg a'',\lozenge g,\lozenge\neg g,\lozenge a',\lozenge\neg a',\lozenge c,\lozenge\neg c,\lozenge h,\lozenge\neg h,\lozenge d,\lozenge\neg d\} T 6 = { a , i , o 2 , ¬ o 1 , s , ¬ b , p , j , k , t , m , ◊ a ′′ , ◊ ¬ a ′′ , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ , ◊ c , ◊ ¬ c , ◊ h , ◊ ¬ h , ◊ d , ◊ ¬ d }
Note that the story successfully establishes the unverifiable claim that Solomon is the son of David in the mind of the reader. Given the fact that the reader also entertains the possibility that Solomon might not be David’s designated heir (in the face of the multiple verifiable facts around Adonijah’s claim), this planted piece of knowledge is the key factor which blocks the introduction of ◊ ¬ s \lozenge\neg s ◊ ¬ s in T 4 ′ T'_4 T 4 ′ , and hence prevents the inclusion of the formula ◊ f \lozenge f ◊ f in T 6 T_6 T 6 , via the implication ◊ ¬ s ∧ ◊ ¬ h → ◊ f \lozenge\neg s \land\lozenge\neg h \to\lozenge f ◊ ¬ s ∧ ◊ ¬ h → ◊ f , which would lead the reader to entertain the possibility that Solomon was part of a foreign coup.
Indeed, if the episode of Uriah’s death was told in a more sanitized way (e.g. ‘Uriah was a valiant soldier who regrettably died in a very bloody battle, and David married his widow out of pity’), the variables a a a and ¬ i \neg i ¬ i rather than a a a and i i i would have been part of e 4 e_4 e 4 , and a a a and ¬ i \neg i ¬ i would not have triggered the application of the ‘exception rule’ in the update from B 3 B_3 B 3 to B 4 B_4 B 4 , and hence the unverified variable s s s would have been stored in the form of ◊ s \lozenge s ◊ s and ◊ ¬ s \lozenge\neg s ◊ ¬ s in T 4 ′ T'_4 T 4 ′ , and hence ◊ s \lozenge s ◊ s and ◊ ¬ s \lozenge\neg s ◊ ¬ s would have been added to T 4 T_4 T 4 and persisted to T 6 T_6 T 6 , where also ◊ ¬ h \lozenge\neg h ◊ ¬ h would have occurred, leaving the reader free to conclude, via the background formula ◊ ¬ s ∧ ◊ ¬ h → ◊ f \lozenge\neg s \land\lozenge\neg h \to\lozenge f ◊ ¬ s ∧ ◊ ¬ h → ◊ f , that ◊ f \lozenge f ◊ f is the case, i.e. that a foreign coup might have occurred.
Modelling Halpern’s analysis using the update protocol for the strategic reader. The modelling above captures the conclusions drawn by the audience in the way engineered by the scribe. However, to capture a more critical reader who reasons in the way suggested by Halpern, we apply the update protocol of the strategic listener, as set out in § 5 . The set L i ′ L'_i L i ′ , L i L_i L i , B i ′ B'_i B i ′ and B i B_i B i are not affected, nor are T i ′ T'_i T i ′ and T i T_i T i for 1 ≤ i ≤ 3 1 \le i \le3 1 ≤ i ≤ 3 , but from T 4 ′ T'_4 T 4 ′ onwards these sets are different.
Let us compute the sets T i ′ T'_i T i ′ . As before, T 1 ′ = { ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } T'_1 = \{\Diamond g, \Diamond\neg g, \Diamond a', \Diamond\neg a'\} T 1 ′ = { ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } , since there are no verifiable variables and there are no inconsistencies in B 1 B_1 B 1 ; T 2 ′ = { a , ¬ e , ◊ i , ◊ ¬ i , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ o 1 , ◊ ¬ o 1 } T'_2 = \{a, \neg e, \Diamond i, \Diamond\neg i, \Diamond a'', \Diamond\neg a'', \Diamond o_1, \Diamond\neg o_1\} T 2 ′ = { a , ¬ e , ◊ i , ◊ ¬ i , ◊ a ′′ , ◊ ¬ a ′′ , ◊ o 1 , ◊ ¬ o 1 } ; T 3 ′ = { a , e , o 2 , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ i , ◊ ¬ i } T'_3 = \{a, e, o_2, \Diamond a'', \Diamond\neg a'', \Diamond i, \Diamond\neg i\} T 3 ′ = { a , e , o 2 , ◊ a ′′ , ◊ ¬ a ′′ , ◊ i , ◊ ¬ i } ; the difference is T 4 ′ = { a , ¬ i , ¬ e , o 2 , ¬ s } T'_4 = \{a, \neg i, \neg e, o_2, \neg s\} T 4 ′ = { a , ¬ i , ¬ e , o 2 , ¬ s } ; notice that we have reversed the polarity of the ‘non verifiable’ literals i i i , s s s . This is in accordance with the idea that the narrative costs paid by the writer through the harm to David’s reputation are justified by persuading the reader that Salomon is David’s biological son. The reader who realizes this, and recognizes this mechanism, should then believe that the non-verifiable claims made in this episode are most probably false. Then, T 5 ′ = { a , e , o 2 , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ i , ◊ ¬ i } T'_5 = \{a, e, o_2, \Diamond a'', \Diamond\neg a'', \Diamond i, \Diamond\neg i\} T 5 ′ = { a , e , o 2 , ◊ a ′′ , ◊ ¬ a ′′ , ◊ i , ◊ ¬ i } and T 6 ′ = { p , j , k , t , m , ◊ c , ◊ ¬ c , ◊ h , ◊ ¬ h , ◊ d , ◊ ¬ d } T'_6 = \{p, j, k, t, m, \Diamond c, \Diamond\neg c, \Diamond h, \Diamond\neg h, \Diamond d, \Diamond\neg d\} T 6 ′ = { p , j , k , t , m , ◊ c , ◊ ¬ c , ◊ h , ◊ ¬ h , ◊ d , ◊ ¬ d } . We now calculate the sets T i T_i T i as follows:
T 1 = { ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } T 2 = { a , ¬ e , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ i , ◊ ¬ i , ◊ o 1 , ◊ ¬ o 1 , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } T 3 = { a , o 2 , ¬ o 1 , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ i , ◊ ¬ i , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } T 4 = { a , ¬ e , ¬ i , o 2 , ¬ o 1 , ¬ s , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } T 5 = { a , ¬ i , o 2 , ¬ o 1 , ¬ s , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } T 6 = { a , ¬ i , o 2 , ¬ o 1 , ¬ s , p , j , k , t , m , ◊ a ′ ′ , ◊ ¬ a ′ ′ , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ , ◊ c , ◊ ¬ c , ◊ h , ◊ ¬ h , ◊ d , ◊ ¬ d , ◊ f } \begin{aligned}
T_1 &= \{\Diamond g, \Diamond\neg g, \Diamond a', \Diamond\neg a'\}\\
T_2 &= \{a, \neg e, \Diamond a'', \Diamond\neg a'', \Diamond i, \Diamond\neg i, \Diamond o_1, \Diamond\neg o_1, \Diamond g, \Diamond\neg g, \Diamond a', \Diamond\neg a'\}\\
T_3 &= \{a, o_2, \neg o_1, \Diamond a'', \Diamond\neg a'', \Diamond i, \Diamond\neg i, \Diamond g, \Diamond\neg g, \Diamond a', \Diamond\neg a'\}\\
T_4 &= \{a, \neg e, \neg i, o_2, \neg o_1, \neg s, \Diamond a'', \Diamond\neg a'', \Diamond g, \Diamond\neg g, \Diamond a', \Diamond\neg a'\}\\
T_5 &= \{a, \neg i, o_2, \neg o_1, \neg s, \Diamond a'', \Diamond\neg a'', \Diamond g, \Diamond\neg g, \Diamond a', \Diamond\neg a'\}\\
T_6 &= \{a, \neg i, o_2, \neg o_1, \neg s, p, j, k, t, m, \Diamond a'', \Diamond\neg a'', \Diamond g, \Diamond\neg g, \Diamond a', \Diamond\neg a', \Diamond c, \Diamond\neg c, \Diamond h, \Diamond\neg h, \Diamond d, \Diamond\neg d, \Diamond f\}
\end{aligned} T 1 T 2 T 3 T 4 T 5 T 6 = { ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } = { a , ¬ e , ◊ a ′′ , ◊ ¬ a ′′ , ◊ i , ◊ ¬ i , ◊ o 1 , ◊ ¬ o 1 , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } = { a , o 2 , ¬ o 1 , ◊ a ′′ , ◊ ¬ a ′′ , ◊ i , ◊ ¬ i , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } = { a , ¬ e , ¬ i , o 2 , ¬ o 1 , ¬ s , ◊ a ′′ , ◊ ¬ a ′′ , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } = { a , ¬ i , o 2 , ¬ o 1 , ¬ s , ◊ a ′′ , ◊ ¬ a ′′ , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ } = { a , ¬ i , o 2 , ¬ o 1 , ¬ s , p , j , k , t , m , ◊ a ′′ , ◊ ¬ a ′′ , ◊ g , ◊ ¬ g , ◊ a ′ , ◊ ¬ a ′ , ◊ c , ◊ ¬ c , ◊ h , ◊ ¬ h , ◊ d , ◊ ¬ d , ◊ f }
Notice now that having ¬ s \neg s ¬ s in T 4 T_4 T 4 ensures that it persists into T 6 T_6 T 6 and allows the reader to conclude ◊ f \Diamond f ◊ f in T 6 T_6 T 6 , leaving open the very damaging possibility, inimical to the writer’s aims, that Salomon rose to power through a foreign coup.
Conclusions and further directions
Main contributions. In this paper, we have developed a general framework for modelling the extraction of information from strategically biased communication, grounded in the Tiglath-Pileser principle. Building on the codex-based models of persuasion introduced by Rubinstein and Glazer, we shifted the focus from codices to their associated guidance relations, and further generalized these to the broader class of Tiglath-Pileser relations. We provided a set-theoretic and axiomatic analysis of these relations, identifying the structural conditions under which they support successful applications of the Tiglath-Pileser principle in persuasion settings. Within this generalized framework, we reformulated Rubinstein and Glazer’s models of persuasion and magical implementation in relational terms, thereby extending their scope beyond one-shot interactions with common knowledge of agents’ goals. Crucially, we introduced a dynamic perspective which accounts for how Tiglath-Pileser relations are formed, revised, and strategically manipulated over the course of interaction, giving rise to higher-order reasoning about agents’ strategies and reliability. Finally, we applied this framework to the case studies of the ‘small businessman and the banker’, and Baruch Halpern’s historiographical analysis of King David’s apology, showing how the proposed machinery can trace the inferential steps from biased narratives to plausible reconstructions of underlying states of affairs. Several directions for further research naturally emerge from the present framework, which are listed below.
A dynamic account of magical implementation. In this paper, we developed the dynamic account of persuasion scenarios similar to those treated in [8 ] but we did not develop this approach for magical implementation [10 ] to the same extent. Many real world situations concerning a principal and two or more agents with (partially) conflicting goals, such as judicial trials, proceed in repeated rounds and evolve dynamically. Formalising such situations would require the parties involved to deduce the motivations and strategies of others in real time, which would call for a dynamic version of the magical implementation approach. This is an ongoing line of research.
Equilibrium analysis. Another line of investigation concerns equilibrium analysis. In the present work, Tiglath-Pileser relations are taken as objects that agents form and revise, but their emergence is not yet embedded in a fully strategic model. Integrating the framework into a game-theoretic setting in which agents choose communication strategies while anticipating how their messages will be interpreted via evolving Tiglath-Pileser relations would allow one to study equilibrium notions in which both communication and interpretation co-evolve. This could also be relevant to information extraction and reasoning in environments with competing or adversarial sources.
Probabilistic approaches. In connection with the game-theoretic approach mentioned above, it would be of interest to develop frameworks in which players’ information about the scenario and other players’ motivations and strategies can be probabilistic. This would also allow for a quantitative account of how players deal with uncertainty while aggregating information extracted from biased reports, and under which conditions such aggregation converges to accurate beliefs.
Inferring Tiglath-Pileser relations. In the present paper we proposed mechanisms for inferring the Tiglath-Pileser relations of other agents from repeated observations of their behaviour. In situations where large amounts of data on agents’ behaviour are available (e.g. books or articles of authors, manifestos of political parties, rulings in legal cases), it would be of interest to develop data-analytic methods to extract (approximations of) Tiglath-Pileser relations automatically. This line of inquiry would also be relevant to a research programme seeking to automate the identification of bias or deliberate disinformation and the extraction of reliable information.