Cheap talk (a formal model)
毫无代价沟通(形式模型)
A formal model of cheap talk was first given in Crawford, Sobel 1982 in which sender (“S”) sends the recipient (“R”) a message (“m”) about the state of the world (“𝜽”). “R” cannot observe “𝜽” directly, but updates his belief about “𝜽” given “m” (i.e., “Pr(𝜽 | m)” [Note 1]), then takes action “a”. The payoffs are based on the action taken (“a”) and state of the world (“𝜽”), and represented by utility functions:
payoff to S: u(a, 𝜽, b);
payoff to R: v(a, 𝜽),
where the scalar “b” measures how much R’s and S’s interests diverge.
Each party maximises their respective payoffs, taking into account the other party’s behaviour. “S” chooses “m” to maximise u(a, 𝜽, b); “R” chooses “a” to maximise v(a, 𝜽). The outcome is taken to be an equilibrium where no party has reason to divert from their behaviour (a Nash equilibrium).
In Crawford, Sobel 1982, “𝜽” is a continuum, and “m” are discrete messages corresponding to partitions of “𝜽”. For example:
“m1” —>“𝜽” lies between 𝜽0-𝜽1,
“m2” —>“𝜽” lies between 𝜽1-𝜽2,
…,
“mN” —> “𝜽” lies between 𝜽N-1- 𝜽N,
where “N” is the number of partitions of “𝜽”.
The number of partitions depends on how far the parties’ interests diverge:
as “b —> ∞” (i.e., full divergence), “N —> 1” (i.e., “𝜽” is coarsely partitioned) so that “m” reveals no information (i.e., babbling equilibrium);
as “b —> 0” (i.e., full alignment), “N —> ∞” (i.e., “𝜽” is finely partitioned) so that “m” reveals “𝜽” ( full information disclosure).
Crawford, Sobel 1982 shows that:
if the parties’ interests are fully aligned (b = 0), “m” reveals “𝜽” (“full information disclosure”), and “a” maximises both parties’ payoffs;
if the parties’ interests are widely divergent (b = ∞), “m” is reveals no information (“babbling equilibrium”), and “a” can be any one of the set of possible actions.
******************
Reference: Crawford, Sobel 1982, “Strategic Information Transmission”. The presentation of their model in this Substack post is based on the one described in the Wikipedia page on cheap talk.
Note 1: Pr(𝜽 | m) = 0.5 suggests the recipient has no prior view on the matter.
毫无代价沟通的 形式模型 最由 Crawford, Sobel 1982 提出,其中 发送者(“S”)向接收者(“R”) 发送 一条关于世界状态(“𝜽”)的 消息(“m”)。“R”无法 直接观察“𝜽”,但会 在给定“m”的 情况下 更新 其对“𝜽”的 信念(即“Pr(𝜽 | m)”[注1]),然后 采取行动“a”。收益 基于所采取的行动(“a”) 和 世界状态(“𝜽”),并 由效用函数 表示:
S的收益:u(a, 𝜽, b);
R的收益:v(a, 𝜽),
其中 标量“b”衡量 “R” 和 “S” 的 利益分歧程度。
每一方 都会最大化自己的 收益,同时考虑 对方的 行为。“S”选择“m” 来最大化 u(a, 𝜽, b);“R”选择“a” 来最大化 v(a, 𝜽)。结果 被假设是 一种均衡,即 任何一方 都没有理由改变其行为(就是 呐什 均衡)。
在Crawford, Sobel 1982,“𝜽” 是 一个 连续统,“m”是 对应于“𝜽”分割的 离散消息。例如:
“m1” —>“𝜽”位于𝜽0-𝜽1之间;
“m2” —>“𝜽”位于𝜽1-𝜽2之间;
… ;
“mN” —>“𝜽”位于𝜽N1—𝜽N之间。
其中“N”是“𝜽”的 分割数。
区分的 次数 取决于 各方利益的 分歧程度:
当“b —> ∞”(即 完全分歧)时,“N —> 1”(即 “𝜽” 被粗略分割),因此“m” 不透露 任何信息(即 胡言乱语均衡);
当“b —> 0”(即 完全一致)时,“N —> ∞”(即 “𝜽” 被精细分割),因此“m”透露“𝜽”(完全 信息披露)。
Crawford, Sobel 1982 表明:
如果 各方利益 完全一致(b = 0),则“m”透露“𝜽”(“完全信息透露”),并且“a” 使双方受益 最大化;
如果 双方利益 相差悬殊(b = ∞),则“m” 不透露 任何信息(“胡言乱语均衡”),“a”可以是 任意 一种可能的行为。
*************
参考:Crawford, Sobel 1982 《战略 信息 传输》。在这篇Substack文章中 提出的 模型 基于 在维基百科“毫无代价沟通”页面中描述的 模型。
注1: Pr(𝜽 | m) = 0.5 表示 接收者 对此事 没有先前的 看法。

