Imel, Guo and Steinert-Threlkeld (2026): An Efficient Communication Analysis of Modal Typology #
This file formalizes the efficient-communication model of [IGST26].
A modal meaning is a subset of the six-point space space, weak and strong force by epistemic,
deontic and circumstantial flavor, and a language is a multiset of meanings. The complexity of
a meaning, Equation (1), is the least number of atoms of a formula of a Boolean Language of
Thought denoting it, computed here from the denotations of formulas with a given number of
atoms (Formula.dens, complexity). The informativeness of a language, Equations (2) and (3),
is the expected utility of literal communication under a communicative need distribution
(a MeasureTheory.Measure on the space): the speaker picks uniformly among the modals
expressing the intended point (speak), the listener guesses uniformly among the points of the
modal heard (ProbabilityTheory.uniformOn) and earns half credit for each axis guessed right
(listen, informativeness). Languages are compared by Pareto dominance on complexity and
communicative cost (Dominates), and naturalness is the fraction of a vocabulary satisfying
the Independence of Force and Flavor universal of [STIG23], the
substrate's Modality.ForceFlavorIndependent.
The paper's three results, that every Pareto-optimal system consists of IFF modals, that naturalness correlates with optimality, and that the attested inventories are more optimal than the sampled ones, are computational and are not formalized. What is proved are the properties of the measures the results rest on. Table 2's complexities are recomputed over the experiment's space, a modal for each point conveys every point exactly (the informativeness is the need of the space) where a single all-purpose modal earns five twelfths of it, a product meaning's listener utility depends only on its two axis sizes, and a language with synonyms is dominated by, and so never Pareto-optimal against, the same language without them.
Implementation notes #
- Table 2 illustrates complexity over a four-flavor space with a teleological flavor the library folds into circumstantial. Over the experiment's six-point space may is denoted by w ∧ ¬c and has complexity two, not the table's three, so the table's language has total complexity seven rather than eight.
- may and St'át'imcets k'a are read off the paper's examples (2) and (3), whose rows record a force and a flavor each.
- The paper's two counterexamples to the Single Axis of Variability universal, Washo -eʔ and Koryak ivək, are not in its sample of 27 languages. ivək's doxastic and assertive flavors ([MA19]) lie outside the six-point space.
- The probabilistic layer is measure-theoretic: the need distribution is a measure, the
listener is
ProbabilityTheory.uniformOnand the speaker a sum of Dirac measures over the multiset of applicable modals, so synonyms keep their multiplicity, and the statements areℝ≥0∞-valued;costis a truncated subtraction, zero if a non-probability need drove the informativeness above one.
References #
The forces of the experiment's space, weak and strong.
Equations
Instances For
The flavors of the experiment's space.
Equations
Instances For
The meaning space P, six force-flavor pairs.
Equations
Instances For
A modal meaning, the force-flavor pairs a modal can express.
Equations
Instances For
The paper's examples #
The force-flavor pair a row's annotation records.
Equations
- One or more equations did not get rendered due to their size.
Instances For
English may, the pairs of examples (2a) and (2b).
Equations
- One or more equations did not get rendered due to their size.
Instances For
St'át'imcets k'a, the pairs of examples (3a) and (3b), after [RMD08].
Equations
- One or more equations did not get rendered due to their size.
Instances For
Table 1: may varies in flavor and k'a in force, each on a single axis.
The Language of Thought and complexity #
Equations
- ImelGuoSteinertThrelkeld2026.instDecidableEqAtom x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯
Equations
- One or more equations did not get rendered due to their size.
The points at which an atom holds, a row or a column of the space.
Equations
- ImelGuoSteinertThrelkeld2026.Atom.weak.den = {Modality.ModalForce.possibility} ×ˢ ImelGuoSteinertThrelkeld2026.flavors
- ImelGuoSteinertThrelkeld2026.Atom.strong.den = {Modality.ModalForce.necessity} ×ˢ ImelGuoSteinertThrelkeld2026.flavors
- ImelGuoSteinertThrelkeld2026.Atom.epistemic.den = ImelGuoSteinertThrelkeld2026.forces ×ˢ {Modality.ModalFlavor.epistemic}
- ImelGuoSteinertThrelkeld2026.Atom.deontic.den = ImelGuoSteinertThrelkeld2026.forces ×ˢ {Modality.ModalFlavor.deontic}
- ImelGuoSteinertThrelkeld2026.Atom.circumstantial.den = ImelGuoSteinertThrelkeld2026.forces ×ˢ {Modality.ModalFlavor.circumstantial}
Instances For
The atom for a force of the space.
Equations
Instances For
The atom for a flavor of the space.
Equations
- ImelGuoSteinertThrelkeld2026.Atom.ofFlavor Modality.ModalFlavor.epistemic = ImelGuoSteinertThrelkeld2026.Atom.epistemic
- ImelGuoSteinertThrelkeld2026.Atom.ofFlavor Modality.ModalFlavor.deontic = ImelGuoSteinertThrelkeld2026.Atom.deontic
- ImelGuoSteinertThrelkeld2026.Atom.ofFlavor x✝ = ImelGuoSteinertThrelkeld2026.Atom.circumstantial
Instances For
The number of atom occurrences, the length in literals of the complexity measure.
Equations
Instances For
The points of the space at which a formula holds.
Equations
Instances For
The conjunction of a force atom and a flavor atom denoting one point of the space.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Denotations by number of atoms #
dens n is the set of denotations of formulas with exactly n atoms: the literals for one
atom, and for more the meets and joins of denotations of formulas whose atoms add up to n.
The recursion is fuelled so that it computes.
The denotations of literals, an atom's points or their complement in the space.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The meets and joins of a member of each set.
Equations
- One or more equations did not get rendered due to their size.
Instances For
densAux fuel n is dens n once n ≤ fuel.
Equations
- One or more equations did not get rendered due to their size.
- ImelGuoSteinertThrelkeld2026.Formula.densAux 0 x✝ = ∅
Instances For
The denotations of formulas with exactly n atoms.
Equations
Instances For
Equation (1) for one modal, the least number of atoms of a formula denoting it.
Equations
- ImelGuoSteinertThrelkeld2026.complexity m = if h : m ⊆ ImelGuoSteinertThrelkeld2026.space then Nat.find ⋯ else 0
Instances For
The hypothetical mought of Table 2, weak epistemic or strong deontic.
Equations
Instances For
The hypothetical notcirc of Table 2, everything but circumstantial.
Equations
Instances For
Over the six-point space may is w ∧ ¬c, two atoms, where Table 2's four-flavor space needs three.
mought needs four atoms, (w ∧ e) ∨ (s ∧ d).
notcirc is the single literal ¬c.
Complexity of a language, the sum over its modals, Equation (1).
Equations
- ImelGuoSteinertThrelkeld2026.totalComplexity L = (Multiset.map ImelGuoSteinertThrelkeld2026.complexity L).sum
Instances For
Table 2's language costs seven atoms over the experiment's space, one fewer than the paper's eight over its illustration space.
Informativeness #
Equation (3), half credit for each axis of the intended point guessed right.
Equations
- ImelGuoSteinertThrelkeld2026.utility p q = (if p.force = q.force then 1 / 2 else 0) + if p.flavor = q.flavor then 1 / 2 else 0
Instances For
The expected utility when the speaker intends p and a literal listener guesses uniformly
among the points a modal expresses.
Equations
- ImelGuoSteinertThrelkeld2026.listen m p = ∫⁻ (q : Modality.ForceFlavor), ImelGuoSteinertThrelkeld2026.utility p q ∂ProbabilityTheory.uniformOn ↑m
Instances For
The expected utility is the mean of the probabilities that the listener's guess matches each axis of the intended point.
The axis-match probabilities counted on the modal's finset.
A listener hearing a product modal earns half the reciprocal of each axis size, so the utility of an IFF modal depends only on how many forces and how many flavors it leaves open.
Table 2's may and mought have two points each; the IFF one is the more informative, sharing an axis with any guess.
The modals of a language a literal speaker chooses among to express p.
Equations
- ImelGuoSteinertThrelkeld2026.speakers L p = Multiset.filter (fun (x : ImelGuoSteinertThrelkeld2026.Meaning) => p ∈ x) L
Instances For
The literal speaker's choice among the modals of L expressing p, uniform over the
multiset so that synonyms keep their multiplicity.
Equations
- ImelGuoSteinertThrelkeld2026.speak L p = (↑(ImelGuoSteinertThrelkeld2026.speakers L p).card)⁻¹ • (Multiset.map MeasureTheory.Measure.dirac (ImelGuoSteinertThrelkeld2026.speakers L p)).sum
Instances For
Equation (2), the expected utility of literal communication under a communicative need distribution.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Communicative cost, the inverse of informativeness.
Equations
- ImelGuoSteinertThrelkeld2026.cost need L = 1 - ImelGuoSteinertThrelkeld2026.informativeness need L
Instances For
Table 5, the communicative need distribution estimated from the corpus.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The corpus need distribution sums to one.
A language with a modal for each point of the space.
Equations
- ImelGuoSteinertThrelkeld2026.singletons = Multiset.map (fun (x : Modality.ForceFlavor) => {x}) ImelGuoSteinertThrelkeld2026.space.val
Instances For
The language of one modal expressing every point.
Instances For
A modal for each point is maximally informative: every point of the space is conveyed exactly, so the informativeness is the need of the space, whatever the need.
One all-purpose modal conveys five twelfths of a point on average, half of a half plus a third, whatever the need.
Synonymy and dominance #
L dominates L' on the trade-off: no worse on either measure, better on one.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Pareto optimality within a pool of languages.
Equations
- ImelGuoSteinertThrelkeld2026.ParetoOptimal need pool L = (L ∈ pool ∧ ∀ L' ∈ pool, ¬ImelGuoSteinertThrelkeld2026.Dominates need L' L)
Instances For
The speaker of a language of copies of one modal is the speaker of the modal alone.
Copies of one modal are as informative as the modal alone.
Synonymy hurts the trade-off: copies of a modal cost the same and add complexity.
A language of copies of one modal is not Pareto-optimal in any pool containing the modal alone.
The universals and naturalness #
Naturalness, the fraction of an inventory satisfying the IFF universal.
Equations
- ImelGuoSteinertThrelkeld2026.naturalness L = ↑(List.countP (fun (x : Modality.ModalItem) => decide (Modality.ForceFlavorIndependent x.meaning)) L) / ↑L.length
Instances For
Washo -eʔ varies on both axes, against the Single Axis of Variability universal of [Nau08], and satisfies IFF, its meaning being every force-flavor pair.
Koryak ivək varies on both axes too, in force and between the doxastic and assertive flavors of [MA19], even without the existential 'say' they could not confirm.
The meaning the universal rules out, epistemic necessity with circumstantial possibility.
Naturalness is graded: Modern Greek, one of the sampled languages, has one IFF modal in three.