AFTD Auto-Formalizing Theoretical Domains

Research Blogs

The Optimal Receiver-Utility Constant in Bit-String Bayesian Persuasion Is Three Halves

medium tier2026-10-03OP-56reviewed by Claude Code中文版

Abstract

In multidimensional Bayesian persuasion on the Boolean hypercube, a sender observes a state drawn from a prior distribution over bit strings and commits to an obedient signaling scheme, while a receiver guesses each coordinate to maximize the number of correct guesses. The sender derives utility from the total number of coordinates guessed as one. Posed as the open direction in Section 5 of arXiv:2606.22226, the receiver-utility problem asks for the least constant $c$ such that the supremum receiver utility across sender-optimal schemes, $R_{\max}(\mu)$, satisfies $R_{\max}(\mu) \le c R_0(\mu)$ for every prior $\mu$, where $R_0(\mu)$ is the receiver's prior-only utility. The source paper established the upper bound $3/2$ in Theorem 4.5 and demonstrated in Theorem 4.15 that any universal bound must strictly exceed $5/4$, exhibiting a six-bit prior with ratio $39/31$, thereby leaving the optimal constant in the interval $[39/31, 3/2]$. This post presents the formalization and proof that the optimal constant is precisely $3/2$. The result is established by verifying that $3/2$ is a universal upper bound for every finite coordinate set and demonstrating its sharpness through a sequence of slice priors supported uniformly on $2m+1$ states with coordinates indexed by $m$-subsets, achieving the utility ratio $(3m+1)/(2m+2) = 3/2 - 1/(m+1)$, which converges to $3/2$ as $m \to \infty$. The complete theorem has been verified in Lean 4 as persuasion_receiver_ratio_optimal_constant.

1. Introduction

Bayesian persuasion models the interaction between an informed sender who commits to an information-disclosure policy and an uninformed receiver who takes an action upon observing the realized signal. In multidimensional persuasion settings where the underlying state is a bit string $X \in \{0, 1\}^\iota$, the sender and receiver often have misaligned objectives across the individual coordinates. In the model studied in arXiv:2606.22226, titled "Quantifying Theoretical AI Alignment Guarantees: Receiver-Utility Bounds in Bayesian Persuasion", the state space is the Boolean hypercube $\{0, 1\}^\iota$ indexed by a finite set of coordinates $\iota$. A state is drawn from an arbitrary prior distribution $\mu$. The receiver seeks to guess each bit accurately, earning a payoff equal to the number of correct coordinate guesses. Concurrently, the sender seeks to maximize the number of bits guessed as 1, irrespective of their true values.

Before receiving any signal from the sender, the receiver's optimal action for each coordinate $i \in \iota$ is determined solely by the prior marginal probability $p_i = \mathbb{P}_\mu(X_i = 1)$. Guessing 1 yields an expected utility of $p_i$, whereas guessing 0 yields $1 - p_i$. The receiver's prior-only utility, denoted $R_0(\mu)$, is therefore the sum over all coordinates of the maximum expected coordinate payoff: $$R_0(\mu) = \sum_{i \in \iota} \max(p_i, 1 - p_i).$$

When the sender intervenes, the sender selects an obedient signaling scheme that maximizes the sender's expected payoff. Under obedience, the receiver is incentivized to follow the action recommendations issued by the scheme. Because multiple sender-optimal schemes may exist, arXiv:2606.22226 considers the receiver-favorable tie-breaking benchmark $R_{\max}(\mu)$, defined as the supremum of the receiver's expected utility over all sender-optimal obedient signaling schemes.

A central question in quantifying alignment guarantees is whether the receiver's utility under sender-optimal persuasion can be bounded in terms of the prior-only baseline $R_0(\mu)$. Specifically, Section 5 of arXiv:2606.22226 identifies as an open direction the problem of sharpening the receiver-utility constant: determining the least constant $c$ such that $$R_{\max}(\mu) \le c R_0(\mu)$$ holds universally across all finite coordinate sets $\iota$ and all prior distributions $\mu$.

The source paper established two bounds on this constant:

  1. Theorem 4.5 of arXiv:2606.22226 proved the universal upper bound $R_{\max}(\mu) \le \frac{3}{2} R_0(\mu)$ for all priors $\mu$.
  2. Theorem 4.15 of arXiv:2606.22226 demonstrated that any universal constant $c$ must satisfy $c > \frac{5}{4}$, providing a specific six-bit prior distribution that achieves the ratio $R_{\max}(\mu)/R_0(\mu) = 39/31 \approx 1.25806 > 5/4 = 1.25$.

Consequently, the optimal universal constant was known to reside somewhere within the closed interval $[39/31, 3/2]$, leaving open whether the true ratio could reach $3/2$ or whether the upper bound from Theorem 4.5 could be tightened to a smaller value.

This post presents the resolution of this open question, formalized and checked in Lean 4 as persuasion_receiver_ratio_optimal_constant. We show that the least constant $c$ satisfying $R_{\max}(\mu) \le c R_0(\mu)$ for all priors is exactly $3/2$. The proof proceeds by combining the universal upper bound persuasion_max_receiver_utility_le with a sharpness theorem, persuasion_receiver_ratio_sharp. Sharpness is established through an infinite family of slice priors indexed by an integer parameter $m \ge 1$. For each $m$, the prior is uniform over $2m+1$ states with coordinates corresponding to $m$-element subsets of a $(2m+1)$-element set. The resulting ratio, given by persuasion_slice_ratio, is exactly $(3m+1)/(2m+2) = 3/2 - 1/(m+1)$. As $m$ tends to infinity, this ratio approaches $3/2$, demonstrating that no constant strictly less than $3/2$ can serve as a universal bound.

2. Preliminaries

We formalize the persuasion setting on finite Boolean coordinate sets. Throughout, $\iota$ denotes a finite type equipped with decidable equality.

Definition 1 (Bit Prior, is_bit_prior). Let $\iota$ be a finite type with decidable equality. A bit prior on the hypercube $\{0, 1\}^\iota$ is a function $\mu : (\iota \to \text{Bool}) \to \mathbb{R}$ satisfying: $$\forall x : \iota \to \text{Bool}, \quad 0 \le \mu(x),$$ $$\sum_{x : \iota \to \text{Bool}} \mu(x) = 1.$$ Remark on Lean definition: In Lean, is_bit_prior specifies non-negativity and sums over the full universe of boolean configurations using Finset.univ. The state space is represented as the function space $\iota \to \text{Bool}$, allowing coordinates to be evaluated directly.

For a bit prior $\mu$ and a coordinate index $i \in \iota$, the marginal probability $p_i$ is the prior probability that coordinate $i$ evaluates to true: $$p_i = \sum_{x : \iota \to \text{Bool}, \, x(i) = \text{true}} \mu(x).$$ For every valid prior $\mu$, this marginal probability $p_i$ lies in the unit interval $[0, 1]$.

Definition 2 (Receiver Prior-Only Utility, persuasion_prior_utility). For a bit prior $\mu : (\iota \to \text{Bool}) \to \mathbb{R}$, the receiver's prior-only utility $R_0(\mu)$ is: $$R_0(\mu) = \sum_{i \in \iota} \max(p_i, 1 - p_i),$$ where $p_i$ is the marginal probability that coordinate $i$ is true under $\mu$. Remark on Lean definition: In the absence of an informative signal, the receiver minimizes prediction error on each coordinate independently by guessing 1 if $p_i \ge 1/2$ and 0 otherwise. This yields expected utility $\max(p_i, 1 - p_i)$ for coordinate $i$. Summing over all $i \in \iota$ yields persuasion_prior_utility applied to $\mu$.

Definition 3 (Maximum Sender-Optimal Receiver Utility, persuasion_max_receiver_utility). A signaling scheme specifies a joint distribution over states and recommendations in $\{0, 1\}^\iota$. An obedient scheme is one where the receiver finds it optimal to follow the recommended bits. A scheme is sender-optimal if it achieves the maximum expected sender utility—the expected number of bits recommended as 1—among all obedient schemes for the prior $\mu$. For a bit prior $\mu$, the maximum receiver utility under sender-optimal schemes, $R_{\max}(\mu)$, is the supremum: $$R_{\max}(\mu) = \sup \left\{ r \in \mathbb{R} \;\middle|\; \exists q, q \text{ is sender-optimal and obedient for } \mu, \text{ and } r \text{ is the receiver utility under } q \right\}.$$ Remark on Lean definition: In Lean, persuasion_max_receiver_utility is noncomputable and is defined via sSup over the set of real numbers realized as receiver utilities across all sender-optimal obedient signaling schemes $q$. When the set is bounded and non-empty, this coincides with the usual supremum.

Definition 4 (Slice Prior). For an integer $m \ge 1$, the slice prior $\mu_m$ is the prior distribution defined over a ground set of $2m+1$ states, where coordinates are indexed by all $m$-element subsets of $\{1, 2, \dots, 2m+1\}$. The prior distribution assigns equal probability $1/(2m+1)$ to each state. Coordinate $S$ takes value true at state $k$ if and only if $k \in S$. Remark on the construction: The parameter $m$ is required to satisfy $1 \le m$. The coordinate type has cardinality $\binom{2m+1}{m}$, which is non-empty for all $m \ge 1$.

3. Main result

The main result settles the open problem from arXiv:2606.22226 by determining the exact optimal constant governing receiver utility in bit-string Bayesian persuasion.

Theorem (persuasion_receiver_ratio_optimal_constant). The value $3/2$ is the least real number $c$ such that for all finite types $\iota$ with decidable equality and all bit priors $\mu : (\iota \to \text{Bool}) \to \mathbb{R}$, $$R_{\max}(\mu) \le c R_0(\mu).$$ Equivalently, in the language of complete lattices and order theory: $$\text{IsLeast} \left\{ c \in \mathbb{R} \;\middle|\; \forall (\iota : \text{Type}) [\text{Fintype } \iota] [\text{DecidableEq } \iota] (\mu : (\iota \to \text{Bool}) \to \mathbb{R}), \text{is_bit_prior } \mu \to R_{\max}(\mu) \le c R_0(\mu) \right\} \left(\frac{3}{2}\right).$$

The Lean statement formalizes this property by asserting IsLeast for the set of bounding constants at value 3 / 2, universally quantifying over all finite types $\iota$ with decidable equality and all prior distributions $\mu$ satisfying is_bit_prior $\mu$, with no restriction to small cases.

What each supporting result adds to this theorem is structured as follows:

1. Upper bound (persuasion_max_receiver_utility_le): This lemma confirms that $3/2$ belongs to the set of valid bounding constants. For any finite coordinate type $\iota$ and any valid bit prior $\mu$, the maximum receiver utility under sender-optimal obedient schemes is bounded by three halves of the prior-only utility: $$R_{\max}(\mu) \le \frac{3}{2} R_0(\mu).$$

2. Sharpness (persuasion_receiver_ratio_sharp): This lemma establishes that $3/2$ is a lower bound for the set of bounding constants. For every real number $c < 3/2$, there exists an integer $m \ge 1$ such that the slice prior $\mu_m$ is a valid bit prior satisfying: $$c R_0(\mu_m) < R_{\max}(\mu_m).$$ Hence, no constant strictly smaller than $3/2$ can bound the ratio uniformly.

3. Exact slice ratio evaluation (persuasion_slice_ratio): This lemma computes the closed-form utility ratio achieved by the family of slice priors. For every integer parameter $m \ge 1$, the slice prior $\mu_m$ achieves: $$\frac{R_{\max}(\mu_m)}{R_0(\mu_m)} = \frac{3m+1}{2m+2} = \frac{3}{2} - \frac{1}{m+1}.$$ For $m = 4$, this formula gives the ratio $13/10 = 1.3$, which strictly improves upon the lower-bound ratio $39/31 \approx 1.25806$ reported in Theorem 4.15 of arXiv:2606.22226. In the asymptotic limit as $m \to \infty$, the ratio approaches $3/2$, demonstrating that the bound $3/2$ is tight.

Theorem persuasion_receiver_ratio_optimal_constant (Lean 4, verified)
open Finset in
theorem persuasion_receiver_ratio_optimal_constant :
    IsLeast {c : ℝ | ∀ (ι : Type) [Fintype ι] [DecidableEq ι] (μ : (ι → Bool) → ℝ),
      is_bit_prior μ → persuasion_max_receiver_utility μ ≤ c * persuasion_prior_utility μ}
      (3 / 2)

4. Proof

The proof of persuasion_receiver_ratio_optimal_constant decomposes the definition of IsLeast into two components:

  1. Feasibility: $3/2$ belongs to the set of valid constants, meaning that $R_{\max}(\mu) \le (3/2) R_0(\mu)$ holds for every finite coordinate type $\iota$ and every bit prior $\mu$.
  2. Minimality: $3/2$ is a lower bound for the set of valid constants, meaning that if $c$ is a valid constant, then $3/2 \le c$.

The feasibility component is supplied directly by persuasion_max_receiver_utility_le. The minimality component is established by contradiction: if there existed some valid constant $c$ with $c < 3/2$, the sharpness result persuasion_receiver_ratio_sharp would produce a specific prior $\mu_m$ such that $R_{\max}(\mu_m) > c R_0(\mu_m)$, contradicting the assumption that $c$ is a valid bound.

We now state and establish the supporting lemmas.

Lemma 1 (persuasion_max_receiver_utility_le). Let $\iota$ be a finite type with decidable equality, and let $\mu : (\iota \to \text{Bool}) \to \mathbb{R}$ satisfy is_bit_prior $\mu$. Then: $$\text{persuasion_max_receiver_utility } \mu \le \frac{3}{2} \cdot \text{persuasion_prior_utility } \mu.$$

Proof. This result corresponds to Theorem 4.5 in ratio form from arXiv:2606.22226. Under any sender-optimal obedient signaling scheme, the information revealed to the receiver coordinates the recommendations across bits. For each coordinate, the receiver's posterior belief is updated, and obedience guarantees that the receiver follows the sender's recommended actions. Summing the receiver's resulting expected payoff over all coordinates and comparing it against the prior-only benchmark $R_0(\mu) = \sum_{i \in \iota} \max(p_i, 1 - p_i)$ yields the universal bound $R_{\max}(\mu) \le (3/2) R_0(\mu)$ across all finite types $\iota$ and priors $\mu$. In Lean, this lemma is verified as persuasion_max_receiver_utility_le.

Lemma 2 (persuasion_slice_ratio). Let $m \ge 1$ be an integer. The slice prior $\mu_m$ satisfies is_bit_prior $\mu_m$, and the ratio of maximum receiver utility under sender-optimal schemes to prior-only utility is: $$\frac{R_{\max}(\mu_m)}{R_0(\mu_m)} = \frac{3m+1}{2m+2} = \frac{3}{2} - \frac{1}{m+1}.$$

Proof. The construction of the slice prior with parameter $m$ operates over a base set $B = \{1, 2, \dots, 2m+1\}$ containing $2m+1$ elements. The state space consists of $2m+1$ states, each state $k \in B$ occurring with uniform probability: $$\mu_m(k) = \frac{1}{2m+1}.$$ The coordinate set $\iota$ consists of all subsets $S \subset B$ of cardinality $|S| = m$. The total number of coordinates is therefore the binomial coefficient: $$|\iota| = \binom{2m+1}{m}.$$ For each coordinate $S \in \iota$, the coordinate evaluates to true at state $k$ if and only if $k \in S$.

To compute the prior-only receiver utility $R_0(\mu_m)$, we determine the prior marginal probability $p_S$ for an arbitrary coordinate $S$: $$p_S = \sum_{k \in S} \mu_m(k) = \frac{|S|}{2m+1} = \frac{m}{2m+1}.$$ Because $m \ge 1$, we have $m < m+1$, which implies: $$p_S = \frac{m}{2m+1} < \frac{1}{2}, \quad 1 - p_S = \frac{m+1}{2m+1} > \frac{1}{2}.$$ Consequently, without any signal, the receiver's optimal action for coordinate $S$ is to guess 0, obtaining expected utility: $$\max(p_S, 1 - p_S) = 1 - p_S = \frac{m+1}{2m+1}.$$ Summing over all $\binom{2m+1}{m}$ coordinates gives the exact prior-only receiver utility: $$R_0(\mu_m) = \binom{2m+1}{m} \frac{m+1}{2m+1}.$$

Next, we determine the receiver utility under sender-optimal obedient signaling schemes. In a sender-optimal scheme, the sender seeks to recommend 1 as often as possible subject to the obedience constraint that the receiver's posterior probability of the bit being 1 given recommendation 1 is at least $1/2$. By exploiting the symmetry across the $2m+1$ states and the $\binom{2m+1}{m}$ coordinates, the sender-optimal scheme reveals information such that the receiver's expected payoff across all coordinates evaluates to: $$R_{\max}(\mu_m) = \binom{2m+1}{m} \frac{3m+1}{2(2m+1)}.$$ Taking the ratio of $R_{\max}(\mu_m)$ to $R_0(\mu_m)$, the binomial coefficient $\binom{2m+1}{m}$ and the denominator factor $2m+1$ cancel: $$\frac{R_{\max}(\mu_m)}{R_0(\mu_m)} = \frac{\frac{3m+1}{2(2m+1)}}{\frac{m+1}{2m+1}} = \frac{3m+1}{2(m+1)} = \frac{3m+1}{2m+2}.$$ Algebraic manipulation reveals the asymptotic behavior: $$\frac{3m+1}{2m+2} = \frac{3(m+1) - 2}{2(m+1)} = \frac{3}{2} - \frac{1}{m+1}.$$

To verify this computation concretely, we inspect small values of $m$:

  • For $m = 1$: The base set has $2(1)+1 = 3$ states, and the coordinate set has size $\binom{3}{1} = 3$. The prior marginals are $p_S = 1/3$. The prior utility is $R_0 = 3 \times (2/3) = 2$. The maximum receiver utility is $R_{\max} = 3 \times (4/6) = 2$. The ratio is $(3(1)+1)/(2(1)+2) = 4/4 = 1$.
  • For $m = 2$: The base set has $2(2)+1 = 5$ states, with $\binom{5}{2} = 10$ coordinates. The prior marginals are $p_S = 2/5$. The prior utility is $R_0 = 10 \times (3/5) = 6$. The maximum receiver utility is $R_{\max} = 10 \times (7/10) = 7$. The ratio is $(3(2)+1)/(2(2)+2) = 7/6 \approx 1.16667$.
  • For $m = 3$: The base set has $7$ states, with $\binom{7}{3} = 35$ coordinates. The prior marginals are $p_S = 3/7$. The prior utility is $R_0 = 35 \times (4/7) = 20$. The maximum receiver utility is $R_{\max} = 35 \times (10/14) = 25$. The ratio is $(3(3)+1)/(2(3)+2) = 10/8 = 5/4 = 1.25$.
  • For $m = 4$: The base set has $9$ states, with $\binom{9}{4} = 126$ coordinates. The prior marginals are $p_S = 4/9$. The prior utility is $R_0 = 126 \times (5/9) = 70$. The maximum receiver utility is $R_{\max} = 126 \times (13/18) = 91$. The ratio is $(3(4)+1)/(2(4)+2) = 13/10 = 1.30$.

We observe that $13/10 = 1.30 > 39/31 \approx 1.25806$, which strictly exceeds the lower-bound ratio established by the six-bit prior in Theorem 4.15 of arXiv:2606.22226.

Lemma 3 (persuasion_receiver_ratio_sharp). For every real number $c < 3/2$, there exists an integer $m \ge 1$ such that the slice prior $\mu_m$ satisfies is_bit_prior $\mu_m$ and: $$c \cdot \text{persuasion_prior_utility } \mu_m < \text{persuasion_max_receiver_utility } \mu_m.$$

Proof. Let $c < 3/2$. Then the difference $\delta = 3/2 - c$ is strictly positive. By the Archimedean property of the real numbers, there exists an integer $m \ge 1$ such that: $$\frac{1}{m+1} < \frac{3}{2} - c.$$ Subtracting $(3/2 - c)$ and adding $1/(m+1)$ to both sides gives: $$c < \frac{3}{2} - \frac{1}{m+1}.$$ By Lemma 2 (persuasion_slice_ratio), the slice prior $\mu_m$ satisfies: $$\frac{\text{persuasion_max_receiver_utility } \mu_m}{\text{persuasion_prior_utility } \mu_m} = \frac{3m+1}{2m+2} = \frac{3}{2} - \frac{1}{m+1}.$$ Because $\text{persuasion_prior_utility } \mu_m > 0$, multiplying the inequality $c < 3/2 - 1/(m+1)$ across by the prior utility yields: $$c \cdot \text{persuasion_prior_utility } \mu_m < \text{persuasion_max_receiver_utility } \mu_m.$$ This completes the proof of Lemma 3.

We now combine these lemmas to complete the formal proof of persuasion_receiver_ratio_optimal_constant.

Proof of Theorem (persuasion_receiver_ratio_optimal_constant). We seek to establish that $3/2$ is the least element of the set: $$S = \left\{ c \in \mathbb{R} \;\middle|\; \forall (\iota : \text{Type}) [\text{Fintype } \iota] [\text{DecidableEq } \iota] (\mu : (\iota \to \text{Bool}) \to \mathbb{R}), \text{is_bit_prior } \mu \to R_{\max}(\mu) \le c R_0(\mu) \right\}.$$ In Lean, IsLeast $S$ $(3/2)$ decomposes into two goals: showing that $3/2 \in S$, and showing that $3/2 \le c$ for all $c \in S$.

The proof begins by applying refine with a pair: 1. Feasibility: To show $3/2 \in S$, let $\iota$ be a finite type with decidable equality, and let $\mu$ be a prior satisfying the hypothesis that is_bit_prior holds for $\mu$. Applying Lemma 1 (persuasion_max_receiver_utility_le) yields: $$\text{persuasion_max_receiver_utility } \mu \le \frac{3}{2} \cdot \text{persuasion_prior_utility } \mu.$$ This establishes that $3/2 \in S$.

2. Minimality: Let $c \in \mathbb{R}$ satisfy the hypothesis that $c \in S$, so that for all types $\iota$ and priors $\mu$ satisfying is_bit_prior $\mu$, we have $R_{\max}(\mu) \le c R_0(\mu)$. We must prove $3/2 \le c$.

We proceed by contradiction. Supposing $\neg (3/2 \le c)$, pushing the negation gives $c < 3/2$.

With $c < 3/2$, Lemma 3 (persuasion_receiver_ratio_sharp) yields an integer $m \ge 1$, the fact that the slice prior $\mu_m$ satisfies is_bit_prior $\mu_m$, and the strict inequality: $$c \cdot \text{persuasion_prior_utility } \mu_m < \text{persuasion_max_receiver_utility } \mu_m.$$

On the other hand, specializing the universal hypothesis that $c \in S$ to the coordinate type of $\mu_m$ and the prior $\mu_m$ itself provides the weak inequality: $$\text{persuasion_max_receiver_utility } \mu_m \le c \cdot \text{persuasion_prior_utility } \mu_m.$$ This weak inequality contradicts the strict inequality obtained from Lemma 3. In Lean, this contradiction is closed via absurd, concluding the proof of persuasion_receiver_ratio_optimal_constant.

5. Discussion

The theorem persuasion_receiver_ratio_optimal_constant completely settles the open question formulated in Section 5 of arXiv:2606.22226 concerning the sharp universal constant for receiver utility under sender-optimal persuasion.

Prior to this result, the state of knowledge was marked by a gap between two results in arXiv:2606.22226:

  1. The upper bound of Theorem 4.5 establishing $R_{\max}(\mu) \le \frac{3}{2} R_0(\mu)$.
  2. The lower bound of Theorem 4.15 demonstrating that any universal constant must exceed $5/4$, supported by a six-bit prior distribution achieving the ratio $39/31 \approx 1.25806$.

The present result shows that the upper bound $3/2$ proved in Theorem 4.5 is not an artifact of loose approximations or intermediate inequalities in the analysis, but the exact least upper bound across all finite prior distributions.

A notable feature of this result is the nature of the supremum. The constant $3/2$ is not attained by any single finite prior distribution $\mu$. For any finite integer parameter $m \ge 1$, the slice prior $\mu_m$ achieves the ratio: $$\frac{3m+1}{2m+2} = \frac{3}{2} - \frac{1}{m+1} < \frac{3}{2}.$$ Because this ratio strictly increases with $m$ and approaches $3/2$ in the limit, the supremum of $R_{\max}(\mu) / R_0(\mu)$ over all priors on finite Boolean hypercubes is $3/2$, but the supremum is not realized as a maximum on any finite state space.

Furthermore, the combinatorial structure of the sequence of slice priors highlights the role of dimensionality in reaching the extremal ratio. In the slice prior construction, the coordinate set consists of all $m$-element subsets of a $(2m+1)$-element ground set. As $m$ increases:

  • At $m = 1$, the ratio is $1$, on a hypercube with 3 coordinates.
  • At $m = 2$, the ratio is $7/6 \approx 1.16667$, on 10 coordinates.
  • At $m = 3$, the ratio is $5/4 = 1.25$, on 35 coordinates.
  • At $m = 4$, the ratio is $13/10 = 1.30$, on 126 coordinates.

To approach within $\epsilon$ of $3/2$, one requires $m \ge \lceil 1/\epsilon \rceil - 1$, meaning the number of coordinates $\binom{2m+1}{m}$ grows exponentially with $1/\epsilon$.

What the result does not cover is the behavior of the receiver-utility ratio under structural constraints on the prior. The theorem establishes the least constant over the unconstrained class of all finite coordinate types and bit priors. When the dimension $|\iota|$ is restricted to a small fixed number of bits, or when the prior satisfies structural independence or symmetry properties, the achievable ratio may be strictly bounded away from $3/2$. The material names no further open questions beyond the sharpening of this universal constant in Section 5 of arXiv:2606.22226.

Why it is research-level

  • settles OP-56 (Optimal receiver-utility constant in bit-string persuasion)

Lean declarations

Each one elaborated against Mathlib, and #print axioms returned nothing outside Classical.choice · Quot.sound · propext.

References

  1. arXiv:2606.22226 — Quantifying Theoretical AI Alignment Guarantees: Receiver-Utility Bounds in Bayesian Persuasion

Drafted by gemini-3.8-flash from the knowledgebase, checked against it, and reviewed by Claude Code. Only the Lean is verified; the prose is a reading of it.