Decision-Theoretic Question Semantics #
@cite{van-rooy-2003} @cite{van-rooy-safarova-2003} @cite{blackwell-1953} @cite{chierchia-2013}
Topical home for question semantics extended with utility / cost structure. The "value of a question" is defined relative to a decision problem; relevance, polar profiles, and Blackwell dominance are all derived notions over the inquisitive substrate.
Layout #
- Utility-based relevance (
@cite{van-rooy-2003}Op(P)).IsDecisionRelevant Q σ dpholds when resolvingQfrom stateσstrictly changes the optimal action indp. - Polar utility profiles (
@cite{van-rooy-safarova-2003}).IsPPQ,IsNPQ,IsAltQclassify a polar question's utility-asymmetry pattern. - Blackwell dominance (
@cite{blackwell-1953}). The inquisitive setting splits the partition notion of "refinement" into two independent directions, both inCore.Question.Relevance:questionEntails P Q— every P-alt ⊆ some Q-alt (P-alts are contained in Q-alts).CoversAltsOf P Q— every nonempty Q-alt is covered by some nonempty P-alt (the dual). For partition questions the two coincide. Decision-relevance preservation requires the dual; theCoversAltsOf.preserves_decisionRelevanttheorem (proven below) is the @cite{van-rooy-2003} bridge.
- Question entropy / NPI strength (
@cite{chierchia-2013}). An NPI-licensing context is one whose alternatives carry high Shannon entropy. The information-theoretic NPI account is @cite{chierchia-2013} (and earlier @cite{israel-1996} for the scalar-strengthening variant); @cite{krifka-1995a} Linguistic Analysis is the foundational lexical-semantic NPI paper but does not invoke entropy directly.
Substrate split #
Question denotations are inquisitive (Core.Question W) — Set-based
alternatives. Decision problems live in Core.Agent.DecisionTheory
with Finset-based action and cell sets. The bridge requires
[Fintype W] plus per-alternative decidability; consumers that don't
need finite computation work directly with the Set-shaped
predicates here.
Decision-relevance (@cite{van-rooy-2003} Op(P)) #
A question is decision-relevant if learning its answer can change
which action maximises expected utility. The notion specialises
@cite{van-rooy-2003}'s Op(P) to the inquisitive substrate: rather
than computing EU(after Q) − EU(before Q) numerically, we ask
whether some alternative of Q would shift the optimal action.
Q is decision-relevant to dp: there exist two non-empty
alternatives p, p' ∈ alt Q and two actions a, a' ∈ acts such
that a strictly dominates a' on p while a' strictly
dominates a on p'. Then learning which alternative obtains
shifts the optimal choice — the operational content of
@cite{van-rooy-2003} Op(P) > 0.
The non-emptiness clauses rule out the degenerate case where
Q = ⊥ (only ∅ in alt): with empty witness propositions
the dominance conditions hold vacuously without genuinely
distinguishing any worlds. Mathlib-shape: a substantive
decision-relevance claim requires substantive (nonempty) witnesses.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Bridge to Core.DecisionTheory.IsResolved #
@cite{van-rooy-2003} §3.1 connection: a IsDecisionRelevant
witness (p, p', a, a') exhibits two Q-alts whose two-action
restriction is IsResolved to different witnesses. On p,
action a resolves the restricted DP (dp, {a, a'}); on p',
action a' does. So IsDecisionRelevant is precisely "two
Q-alts resolve to distinct actions" in the substrate's
IsResolved vocabulary.
The trivial question (single alternative) is not decision-relevant: one substantive alternative cannot witness the action-reversal.
Polar utility profiles (@cite{van-rooy-safarova-2003}) #
A polar question Q? = {p, ¬p} has a utility profile determined by
the asymmetry of action-payoffs between the p-cell and the ¬p-cell.
- PPQ (positive polar): the
p-alternative is utility-preferred. Speaker bias towardp(e.g., Did John leave? in a context where John leaving is good news). - NPQ (negative polar): the
¬p-alternative is preferred. Speaker bias toward¬p(Didn't John leave?). - Alt (alternative question): both alternatives carry comparable utility — neither dominates.
Q is a PPQ for action set acts and polar split (p, ¬p):
one specific alternative p ∈ alt Q strictly dominates every
other alternative across every action. The "positive" cell is
where the speaker's preferred outcome lies. @cite{van-rooy-safarova-2003}.
Equations
- Semantics.Questions.DecisionTheoretic.IsPPQ Q dp acts p = (p ∈ Q.alt ∧ ∃ a ∈ acts, ∀ p' ∈ Q.alt, p' ≠ p → ∀ w ∈ p, ∀ w' ∈ p', dp.utility w a > dp.utility w' a)
Instances For
Q is an NPQ for action set acts and polar split (p, ¬p):
a strictly non-p alternative is the utility-preferred one.
@cite{van-rooy-safarova-2003}.
Equations
- Semantics.Questions.DecisionTheoretic.IsNPQ Q dp acts p = (p ∈ Q.alt ∧ ∃ p' ∈ Q.alt, p' ≠ p ∧ Semantics.Questions.DecisionTheoretic.IsPPQ Q dp acts p')
Instances For
Q is an alternative-utility question for action set acts:
no single alternative strictly dominates. The utility-balanced
case. @cite{van-rooy-safarova-2003}.
Equations
- Semantics.Questions.DecisionTheoretic.IsAltQ Q dp acts = ∀ p ∈ Q.alt, ¬Semantics.Questions.DecisionTheoretic.IsPPQ Q dp acts p
Instances For
A PPQ rules out the alt-utility classification.
Blackwell informativeness (@cite{blackwell-1953}) #
The Blackwell informativeness order on questions is
Core.Question.Relevance.questionEntails P Q: every alternative of
P is contained in some alternative of Q (P is finer than Q).
The dual condition — every nonempty Q-alt is covered by a nonempty
P-alt — is Core.Question.Relevance.CoversAltsOf P Q. On partition
questions the two coincide; on general inquisitive Question W they
are distinct. The decision-relevance preservation theorem requires
the dual form.
The decision-relevance preservation theorem itself
(CoversAltsOf.preserves_decisionRelevant) is declared at the bottom
of this file, outside the topic namespace, under
Core.Question.CoversAltsOf so dot notation works at consumer
sites.
Question entropy and NPI strength (@cite{chierchia-2013}) #
@cite{chierchia-2013} relates NPI distribution to the informativeness of the alternative set raised by a sentence (the information-theoretic strand earlier traced through @cite{israel-1996} on lexical-semantic polarity sensitivity). A question whose alternatives are equiprobable under the prior carries maximal Shannon entropy; NPIs prefer such high-entropy contexts. The earlier @cite{krifka-1995a} Linguistic Analysis polarity-items paper is the foundational lexical-semantic NPI account but does not formulate licensing in entropy terms — the Chierchia/Israel synthesis is the right citation here.
The numeric API requires real-valued logarithms over the prior;
deferred to a follow-up that lifts Core.Probability's PMF API to
expose Shannon entropy. The Prop-level placeholder
HasMultipleSubstantiveAlternatives captures the load-bearing
condition that the question is informative — necessary for any
NPI-licensing notion.
Q has multiple substantive alternatives: some pair of distinct
alternatives are both nonempty. Necessary for any NPI-licensing
formulation (a single-alt question carries no informational
asymmetry). @cite{chierchia-2013} prerequisite.
Equations
- Semantics.Questions.DecisionTheoretic.HasMultipleSubstantiveAlternatives Q = ∃ p ∈ Q.alt, ∃ q ∈ Q.alt, p ≠ q ∧ p.Nonempty ∧ q.Nonempty
Instances For
A decision-relevant question necessarily has multiple substantive
(nonempty, distinct) alternatives. The substrate-level
contrapositive of IsDecisionRelevant.singleton_alt_false:
decision-relevance requires informational asymmetry, which is
@cite{chierchia-2013}'s prerequisite for NPI licensing. So any
decision-relevant question is a candidate NPI-licensing context.
CoversAltsOf API extensions #
Theorem placed under Core.Question.CoversAltsOf namespace (not the
file's Semantics.Questions.DecisionTheoretic topic namespace) so
dot-notation hCover.preserves_decisionRelevant resolves at consumer
sites. Mathlib pattern: declarations follow the type's namespace, not
the file's topic namespace.
@cite{blackwell-1953} ↔ @cite{van-rooy-2003}: when every nonempty
alt of Q is covered by a nonempty alt of P, decision-relevance
is preserved from Q to P.