RSA — Latent-aware unbundled operators

@cite{lassiter-goodman-2017} @cite{kao-etal-2014-hyperbole} @cite{nouwen-2024}

Latent-variable extensions to the unbundled Operators.lean API. RSA models with latent variables (thresholds, QUDs, lexicons) parameterize their L0/S1 on a latent type L; the L1 marginalizes over L against a latent prior.

The audit in project_rsa_v2_functional.md flagged that 87% of RSAConfig consumers use Latent ≠ Unit — so latent support must be first-class in any unbundled v2 API, not an afterthought.

Operators

• L0LassiterGoodman — L&G L0 with prior baked in: P.reweight meaning. The paper-name alias for the standard PMF.reweight operation specialized to the L&G "prior in literal listener" pattern (eq 70/71 of @cite{lassiter-goodman-2017}).
• marginalizeKernel — marginalize a latent-parameterized kernel against a latent prior. Just PMF.bind of the latent prior into the kernel slice at each world. This is the operator that turns a Latent → W → PMF U family into a single W → PMF U for use as the speaker in L1.

(Latent-parameterized *Latent wrappers and composeStages were deleted in 0.230.373 — they added no operator content over partial application or function composition. See HISTORY at end.)

L&G "two priors" pattern

For the L&G family (used by @cite{lassiter-goodman-2017}, @cite{nouwen-2024}, intensifier studies), the same prior P enters BOTH:

• L0 via P.reweight meaning (the literal listener is Bayesian with prior P)
• L1 via PMF.posterior speaker P (the pragmatic listener is Bayesian with prior P)

This double-occurrence is faithful to the paper specification, not double-counting. With this file's operators, the pattern is straightforward: construct L0 via L0LassiterGoodman P meaning, build the speaker on top, then close with PMF.posterior speaker P. The same P shows up in both function calls — the structure makes it explicit.

Sequential composition

Sequential between-stage composition (Nouwen 2024 eq 73: Π = stage 1's L1 becomes stage 2's prior) is just function composition: pass the previous stage's PMF.posterior output as the prior argument to the next stage's construction. No new operator needed — the existing PMF infrastructure is sufficient.

The composeStages helper below is a thin alias documenting this idiom.

Relationship to Phase 5 migration

This file is part of Phase 1.5 of the RSA → mathlib-PMF migration (project_rsa_pmf_migration.md). It extends the unbundled operator suite to cover the latent and L&G patterns identified in the consumer audit (6 canonical patterns, 47 files). Phase 5 study migrations build on these.

A Nouwen2024PMF pilot demonstrating the L&G two-prior + latent + sequential pattern end-to-end is the natural next step — see commented sketch in L0LassiterGoodman and marginalizeKernel docstrings for how to use them.

L0 with L&G "prior in literal listener" pattern

noncomputable def RSA.L0LassiterGoodman {U : Type u_1} {W : Type u_2} (P : PMF W) (meaning : U → W → Bool) (u : U) (h_pos : (∑' (w : W), P w * if meaning u w = true then 1 else 0) ≠ 0) : PMF W

Lassiter-Goodman literal listener (eq 71 of @cite{lassiter-goodman-2017}): the prior P is part of the literal listener's Bayesian update. For a Boolean-valued meaning function meaning : U → W → Bool, this returns P reweighted by the indicator function of the meaning's extension.

Mathematically: L0_LG(w | u) = P(w) · ⟦u⟧(w) / Σ_{w'} P(w') · ⟦u⟧(w')

This is exactly PMF.reweight P (fun w => indicator (meaning u w)).

Equations

• RSA.L0LassiterGoodman P meaning u h_pos = P.reweight (fun (w : W) => if meaning u w = true then 1 else 0) h_pos ⋯

theorem RSA.L0LassiterGoodman_apply {U : Type u_1} {W : Type u_2} (P : PMF W) (meaning : U → W → Bool) (u : U) (h_pos : (∑' (w : W), P w * if meaning u w = true then 1 else 0) ≠ 0) (w : W) : (L0LassiterGoodman P meaning u h_pos) w = (P w * if meaning u w = true then 1 else 0) * (∑' (w' : W), P w' * if meaning u w' = true then 1 else 0)⁻¹

theorem RSA.mem_support_L0LassiterGoodman_iff {U : Type u_1} {W : Type u_2} (P : PMF W) (meaning : U → W → Bool) (u : U) (h_pos : (∑' (w : W), P w * if meaning u w = true then 1 else 0) ≠ 0) (w : W) : w ∈ (L0LassiterGoodman P meaning u h_pos).support ↔ P w ≠ 0 ∧ meaning u w = true

theorem RSA.L0LassiterGoodman_apply_of_meaning_true {U : Type u_1} {W : Type u_2} (P : PMF W) (meaning : U → W → Bool) (u : U) (h_meaning : ∀ (w : W), meaning u w = true) (h_pos : (∑' (w : W), P w * if meaning u w = true then 1 else 0) ≠ 0) (w : W) : (L0LassiterGoodman P meaning u h_pos) w = P w

Vacuous-utterance reduction: when meaning u is identically true (e.g. .silent), the L&G L0 at u reduces to the prior unchanged.

This is the building block for many positivity discharges in consumer code: "the silent utterance has L0 = prior (positive everywhere by assumption), so its contribution to any normalization sum is non-zero."

theorem RSA.L0LassiterGoodman_apply_lt_iff {U : Type u_1} {W : Type u_2} (P : PMF W) (meaning : U → W → Bool) (u : U) (h_pos : (∑' (w : W), P w * if meaning u w = true then 1 else 0) ≠ 0) (w₁ w₂ : W) : (L0LassiterGoodman P meaning u h_pos) w₁ < (L0LassiterGoodman P meaning u h_pos) w₂ ↔ (P w₁ * if meaning u w₁ = true then 1 else 0) < P w₂ * if meaning u w₂ = true then 1 else 0

Inequality decomposition for L&G L0: at a fixed utterance, comparing two worlds' literal-listener probabilities reduces to comparing their unnormalised scores P(w) · indicator(meaning u w). The partition function Z(u) depends on the utterance but not the world, so it cancels.

Direct lift from PMF.reweight_lt_iff_lt.

theorem RSA.L0LassiterGoodman_apply_le_iff {U : Type u_1} {W : Type u_2} (P : PMF W) (meaning : U → W → Bool) (u : U) (h_pos : (∑' (w : W), P w * if meaning u w = true then 1 else 0) ≠ 0) (w₁ w₂ : W) : (L0LassiterGoodman P meaning u h_pos) w₁ ≤ (L0LassiterGoodman P meaning u h_pos) w₂ ↔ (P w₁ * if meaning u w₁ = true then 1 else 0) ≤ P w₂ * if meaning u w₂ = true then 1 else 0

The ≤ companion of L0LassiterGoodman_apply_lt_iff.

theorem RSA.L0LassiterGoodman_apply_lt_iff_prior_lt {U : Type u_1} {W : Type u_2} (P : PMF W) (meaning : U → W → Bool) (u : U) (h_pos : (∑' (w : W), P w * if meaning u w = true then 1 else 0) ≠ 0) (w₁ w₂ : W) (h₁ : meaning u w₁ = true) (h₂ : meaning u w₂ = true) : (L0LassiterGoodman P meaning u h_pos) w₁ < (L0LassiterGoodman P meaning u h_pos) w₂ ↔ P w₁ < P w₂

Convenience corollary: when the utterance's meaning is true at both worlds, the L0 comparison collapses to a prior comparison. The indicator is 1 on both sides, the partition function cancels — only P(w) is left.

theorem RSA.L0LassiterGoodman_apply_le_iff_prior_le {U : Type u_1} {W : Type u_2} (P : PMF W) (meaning : U → W → Bool) (u : U) (h_pos : (∑' (w : W), P w * if meaning u w = true then 1 else 0) ≠ 0) (w₁ w₂ : W) (h₁ : meaning u w₁ = true) (h₂ : meaning u w₂ = true) : (L0LassiterGoodman P meaning u h_pos) w₁ ≤ (L0LassiterGoodman P meaning u h_pos) w₂ ↔ P w₁ ≤ P w₂

The ≤ companion of L0LassiterGoodman_apply_lt_iff_prior_lt.

Marginalization over latents

The latent-parameterized "wrappers" L0OfMeaningLatent, L0LassiterGoodmanLatent, S1BeliefLatent were deleted in 0.230.373 (mathlib hygiene audit). They added no operator content over partial application: e.g. L0OfMeaningLatent meaning l u h0 hTop was just L0OfMeaning (meaning l) u h0 hTop. Mathlib doesn't define Function.compLatent. Consumers should write the partial application directly.

noncomputable def RSA.marginalizeKernel {L : Type u_3} {α : Type u_4} {β : Type u_5} (latentPrior : PMF L) (κ : L → α → PMF β) (a : α) : PMF β

Marginalize a latent-parameterized kernel against a latent prior: marginalize prior κ a = ∑_l prior(l) · κ(l, a) as a PMF over β.

This is PMF.bind of the latent prior into the kernel slice at each a. The operator that turns a L → α → PMF β family into a single α → PMF β suitable for use as the speaker kernel in L1 = PMF.posterior.

For Nouwen2024-style models where the latent is a Threshold marginalized in L1, the typical pattern is:

def speakerLatent : Threshold → Height → PMF EvalUtterance := S1BeliefLatent ...
def marginalizedSpeaker : Height → PMF EvalUtterance :=
  marginalizeKernel (PMF.uniformOfFintype Threshold) speakerLatent
def L1 : EvalUtterance → PMF Height :=
  fun u => PMF.posterior marginalizedSpeaker heightPrior u (h_marginal_ne_zero u)

Equations

• RSA.marginalizeKernel latentPrior κ a = latentPrior.bind fun (l : L) => κ l a

@[simp]

theorem RSA.marginalizeKernel_apply {L : Type u_3} {α : Type u_4} {β : Type u_5} (latentPrior : PMF L) (κ : L → α → PMF β) (a : α) (b : β) : (marginalizeKernel latentPrior κ a) b = ∑' (l : L), latentPrior l * (κ l a) b

theorem RSA.mem_support_marginalizeKernel_iff {L : Type u_3} {α : Type u_4} {β : Type u_5} (latentPrior : PMF L) (κ : L → α → PMF β) (a : α) (b : β) : b ∈ (marginalizeKernel latentPrior κ a).support ↔ ∃ (l : L), latentPrior l ≠ 0 ∧ b ∈ (κ l a).support

theorem RSA.marginalizeKernel_lt_marginalizeKernel {L : Type u_3} {α : Type u_4} {β : Type u_5} (μ : PMF L) (κ : L → α → PMF β) (b : β) {a₁ a₂ : α} {l₀ : L} (hμ : μ l₀ ≠ 0) (h : ∀ (l : L), (κ l a₁) b ≤ (κ l a₂) b) (hi : (κ l₀ a₁) b < (κ l₀ a₂) b) : (marginalizeKernel μ κ a₁) b < (marginalizeKernel μ κ a₂) b

Inequality decomposition for marginalizeKernel (cross-a, same b): when the kernel slice at a₂ dominates the slice at a₁ pointwise — and is strict at some latent l₀ with positive prior mass — the marginalized kernel inherits the strict inequality.

Direct lift of PMF.marginal_lt_marginal after unfolding bind_apply.

Use case: "the marginalized speaker assigns higher probability to utterance u at world w₂ than at world w₁" reduces to "for every latent threshold, the per-latent speaker prefers u at w₂, with strict preference at some threshold."

theorem RSA.marginalizeKernel_le_marginalizeKernel {L : Type u_3} {α : Type u_4} {β : Type u_5} (μ : PMF L) (κ : L → α → PMF β) (b : β) {a₁ a₂ : α} (h : ∀ (l : L), (κ l a₁) b ≤ (κ l a₂) b) : (marginalizeKernel μ κ a₁) b ≤ (marginalizeKernel μ κ a₂) b

The ≤ companion of marginalizeKernel_lt_marginalizeKernel.

Sequential between-stage composition (Nouwen eq 73)

Two successive ρ updates, where stage 2's prior is stage 1's L1 output at a specific stage-1 utterance. This is just function composition — pass the previous-stage L1 as the prior argument to the next stage's L1 construction.

For example, given:

• evalL1 : EvalUtterance → PMF Height — stage 1 L1
• adjL1Builder : PMF Height → AdjUtterance → PMF Height — stage 2 L1 parameterized by its prior

The sequential L1 at the specific stage-1 observation u_eval = .eval_pos is: seqAdjL1 u₂ := adjL1Builder (evalL1 .eval_pos) u₂

No operator needed. (The composeStages alias was deleted in 0.230.373 — function-application doesn't need a name.)