Labels

Technical Summary: "Labels" by Mark Whitmeyer

Problem Statement
The paper addresses the design of public labels—such as grades, credentials, scores, and ratings—when these labels serve a dual function: they inform receivers about an agent's quality, and they act as prizes that agents compete to obtain. Unlike standard information design problems where the goal is purely to reveal information, this framework recognizes that the creation of a distinction (a "rung" on a ladder) induces costly self-selection by agents. Agents expend resources to cross boundaries between categories to secure higher posterior means (and thus higher private benefits). The central question is: Which distinctions are worth building? Specifically, the paper investigates the optimal classification rule when labels must be earned through costly actions, asking whether full transparency (certifying every type exactly) is efficient or if coarser, "noisy" communication can improve welfare.

Methodology and Framework
The author models a signaling environment involving three parties: a sender (with a private type $q \in [0,1]$ representing a quantile of quality), a receiver, and an information designer.

• Sender: Chooses a costly action $a$. The cost is $aS(x(q))$, where $x(q)$ is the quality and $S(\cdot)$ is a strictly decreasing cost function. The sender's payoff is $b(m) - \text{cost}$, where $m$ is the receiver's posterior mean of quality.
• Receiver: Obtains value $r(m)$ based on the posterior mean.
• Designer: Commits to a deterministic certification technology $\gamma$ mapping actions to public certificates.

The analysis proceeds by reducing the design problem to the choice of an admissible signal $M(q)$, representing the posterior mean induced for type $q$. This signal must be non-decreasing and Bayes-plausible ($M = E[x|M]$).

• Key Reduction: The author derives a "cost-of-distinctions" formula (Proposition 3.1). The minimum expected implementation cost $K(M)$ for a given signal $M$ is expressed as an integral involving the mimicry pressure $L(q)$ and the jump in sender benefits across category boundaries:
  $$K(M) = \int_{(0,1]} L(q) \, d[b(M(q))]$$
  where $L(q) = \frac{\int_q^1 S(x(t))dt}{S(x(q))}$. This formula equates the cost of a distinction to the "price" of the boundary, defined as the mimicry pressure multiplied by the private prize jump.

Key Contributions and Results

1. Inefficiency of Full Certification:
   The paper demonstrates that full certification (where every type is perfectly distinguished, $M(q)=x(q)$) is generally suboptimal.

   • Theorem 4.1: Under smoothness assumptions and positive entry stakes, pooling a small bottom interval $[0, \epsilon]$ and fully certifying all types above $\epsilon$ yields higher welfare than full certification.
   • Mechanism: Pooling the bottom types saves first-order signaling costs because it eliminates the first rung of the certification ladder, relaxing action requirements for a positive mass of higher types. The loss in receiver value is only higher-order because the types being pooled are close in quality, making the averaging error negligible compared to the signaling savings.

2. Optimality of Lower Censorship:
   The paper identifies conditions under which this bottom-pooling strategy is not just an improvement, but the globally optimal design.

   • Proposition 4.2 & Corollary 4.3: If mimicry pressure $L(q)$ is convex in quantile, receiver value $r(m)$ is convex in posterior mean, and sender stakes $b(m)$ are weakly concave (diminishing returns), then lower censorship is optimal. This design consists of a free default label for the bottom interval $[0, c^*]$ followed by full certification for all types above $c^*$.
   • Intuition: Under these conditions, any non-initial pooled interval is inferior to full certification. The only pooling that survives is the initial one, which acts as a "free default" to avoid the high marginal cost of distinguishing the lowest types.

3. Finite Optimality:

   • Theorem 6.4: The paper provides sufficient conditions (Assumption 6.1) under which every optimal classification uses a finite number of categories. If binary splits on sufficiently small intervals are locally wasteful (the receiver's gain from finer precision is outweighed by the sender's cost of sustaining the new distinction), then any optimal signal must have a finite essential range. The number of categories is bounded by $1 + 2/\delta$.

4. Robustness and Extensions:

   • General Preferences (Section 5): The result that full certification is suboptimal holds even if the receiver's payoff is not a function of the posterior mean but is a continuous function of the posterior distribution (Proposition 5.1).
   • Signaling Tournaments (Section 7): The insights extend to winner-take-all tournaments (Proposition 7.1). Even when the prize is generated by a tournament among $n$ candidates, full certification remains suboptimal; pooling the bottom types improves welfare because the social value of distinguishing the worst types is less than the cost of the distinctions.

Significance and Claims
The paper claims to offer a "costly-certification interpretation" of optimal information design. It argues that the standard intuition—that more information is always better—fails when information design induces costly behavior. The primary contribution is the identification of a robust, structural improvement over full transparency: erasing or blurring the bottom rung of the ladder.

The author positions this work not as a rediscovery that coarse communication can be valuable (a point already made in literature regarding intermediaries, manipulation, and rank disclosures), but as a specific characterization of which distinctions should be eliminated. The paper asserts that when labels are prizes, the institution should not merely "reveal" information but should strategically choose which rungs to build. The finding that a "free default" followed by full certification is often optimal provides a theoretical justification for practices where low-quality agents are grouped into a generic "fail" or "unverified" category, while higher-quality agents are fully distinguished.

The paper concludes that the "juice is not worth the squeeze" for the worst types: the cost of distinguishing them via self-selection exceeds the value of the precision gained by the receiver.