The Geometry of Heterogeneous Extremes: Optimal Transport and Entropic Design

1. Problem Statement and Motivation

The paper addresses a fundamental limitation in standard Extreme Value Theory (EVT): the assumption of deterministic sample sizes (i.i.d. sequences). In many economic environments (e.g., labor markets, network search, firm sourcing), agents face heterogeneous opportunity sets. Different agents sample different numbers of opportunities, and this draw count is itself random.

The core problem is to characterize the distribution of the maximum (extreme outcome) when the number of draws $N$ follows a Mixed-Poisson distribution, where the intensity parameter is random. Specifically, if an agent's draw intensity is $X \sim F$ (with mean 1), and $N|X \sim \text{Poisson}(\theta X)$, the paper seeks to understand the operator $T: F \mapsto H_{\gamma, F}$, which maps the distribution of heterogeneity $F$ to the limiting distribution of normalized extremes $H_{\gamma, F}$.

The author aims to move beyond pointwise comparisons to establish geometric stability (how perturbations in $F$ affect the entire law of extremes) and to solve normative design problems (how to optimally reallocate opportunities to improve tail outcomes).

2. Methodology and Framework

A. The Mixed-Poisson Framework

The model assumes:

1. Underlying Offers: A sequence of i.i.d. offers $Y_j$ with CDF $G$ in the domain of attraction of a Generalized Extreme Value (GEV) distribution $H_\gamma$ with tail index $\gamma$.
2. Heterogeneous Search: Agents draw $N(\theta)$ offers, where $N(\theta)|X \sim \text{Poisson}(\theta X)$ and $X \sim F$ is the agent's type (intensity).
3. Limit Law: The normalized maximum $Z_\theta = (M_\theta - b_\theta)/a_\theta$ converges to a limit distribution:
   $$H_{\gamma, F}(x) = P_0(v_\gamma(x))$$
   where $P_0(z) = \mathbb{E}[e^{-zX}]$ is the Laplace transform of $F$, and $v_\gamma(x) = -\log H_\gamma(x)$ is the tail transform.

B. Geometric Approach: Optimal Transport

The paper treats the mapping $F \mapsto H_{\gamma, F}$ as an operator on the space of probability measures. Instead of relying solely on Laplace analytic properties, the author employs Optimal Transport (OT) theory:

• Canonical Coupling: A key insight is the representation of the limit variable $Z_{\gamma, F}$ as a function of the type $X$ and an independent exponential shock $E \sim \text{Exp}(1)$:
  $$Z_{\gamma, F} = w_\gamma(E/X)$$
  where $w_\gamma$ is the inverse of the tail transform $v_\gamma$.
• Adapted Metrics: The author defines a metric $d_{\gamma, p}$ on the type space $F$ that is "adapted" to the tail index $\gamma$. This metric is the standard Wasserstein distance $W_p$ applied to the transformed types $s_\gamma(X)$, where:
  $$s_\gamma(x) = \begin{cases} x^\gamma & \gamma \neq 0 \\ \log x & \gamma = 0 \end{cases}$$

C. Normative Approach: Entropic Design

The paper also studies a planner's problem: choosing a new distribution $F$ to maximize a tail-sensitive objective (e.g., expected utility of extremes) subject to a mean constraint ($\mathbb{E}[X]=1$) and a Kullback-Leibler (KL) penalty relative to a baseline $F_0$. This is framed as a one-marginal entropy projection.

3. Key Contributions and Results

I. Structural Decomposition and Order Properties

• Convex Order: The paper establishes that if $F_2$ is a mean-preserving spread of $F_1$ (i.e., $F_2 \succeq_{cx} F_1$), then the resulting extreme distribution $H_{\gamma, F_2}$ is stochastically smaller than $H_{\gamma, F_1}$.
• Implication: Heterogeneity in opportunity access (inequality) systematically lowers the distribution of maximum outcomes for a randomly drawn agent compared to a homogeneous economy, even if the mean number of opportunities is unchanged.

II. Wasserstein Stability (The Main Geometric Result)

Theorem 1 provides the paper's central quantitative contribution. It proves that the operator $F \mapsto H_{\gamma, F}$ is Lipschitz continuous with respect to the adapted metric $d_{\gamma, p}$ and the Wasserstein distance on the extreme laws.

• Bound: $W_p(H_{\gamma, F_1}, H_{\gamma, F_2}) \leq C_{\gamma, p} \cdot d_{\gamma, p}(F_1, F_2)$.
• Significance: This allows for whole-law stability. It bounds the error in the entire distribution of extremes (including the full quantile schedule) given an error in the heterogeneity distribution $F$.
• Regime Specifics:
  • For $\gamma > 0$ (Fréchet), the constant depends on moments of $E^{-\gamma}$, requiring $p\gamma < 1$.
  • For $\gamma = 0$ (Gumbel), the map is a simple shift in log-space.
  • Bridge to Raw Space: The paper explicitly quantifies the gap between the adapted metric $d_{\gamma, p}$ and the standard Wasserstein distance on the raw type space. In Fréchet regimes, this bridge is often Hölder continuous (order $\gamma$) rather than linear, highlighting that estimation error in raw types propagates non-linearly to tail predictions.

III. Second-Order Expansion

Theorem 2 provides a second-order expansion for the pre-limit CDF. It cleanly separates the approximation error of the EVT limit (driven by the underlying offer distribution $G$) from the sensitivity to the heterogeneity kernel. This allows researchers to distinguish between "EVT approximation error" and "heterogeneity estimation error."

IV. Entropic Design and Exponential Tilting

Theorem 3 solves the planner's problem. For objectives linear in $F$ (including expected utility of normalized extremes), the optimal distribution $F^*$ is an exponential tilt of the baseline $F_0$:
$$\frac{dF^*}{dF_0}(x) \propto \exp\left( \frac{\psi(x) + \eta^* x}{\lambda} \right)$$
where $\psi(x)$ is the marginal score derived from the heterogeneous-EVT kernel, $\eta^*$ is the Lagrange multiplier for the mean constraint, and $\lambda$ is the cost of deviation.

• Novelty: The paper identifies that the same Laplace kernel governing the positive analysis ($F \mapsto H_{\gamma, F}$) also generates the score functions for the normative design problem.

4. Application: Labor Market Networks

The framework is applied to labor markets where network position determines job offer arrival rates.

• Mechanism: Network degree (or centrality) maps to the intensity $X$.
• Findings:
  • Inequality: Increased network inequality (convex order spread) reduces the top wage distribution for the average worker.
  • Robustness: The stability results provide explicit error bounds for predicting top wages when network heterogeneity is measured with error.
  • Policy: The entropic design framework suggests that policies (e.g., referral programs) can be modeled as reweighting the distribution of access. The optimal policy involves an exponential tilt that balances tail improvement against implementation frictions (entropy cost).

5. Significance and Implications

1. Geometric Rigor in EVT: The paper moves EVT from pointwise asymptotic statements to a geometric framework using Optimal Transport. This allows for robust comparative statics on entire distributions rather than just tail probabilities.
2. Robustness to Misspecification: By establishing Lipschitz bounds in the adapted metric, the paper provides a rigorous way to quantify how errors in estimating the distribution of opportunity access ($F$) propagate to errors in predicting extreme outcomes (e.g., top wages, firm productivity).
3. Policy Design: It offers a tractable method for designing policies that target the "tail" of economic outcomes. The solution (exponential tilt) is explicit and depends on the specific tail index $\gamma$ and the cost of reallocation.
4. Identification and Estimation: The paper clarifies the statistical challenges. While the Laplace transform of $F$ is identified from count data, recovering $F$ is an ill-posed inverse problem. The results emphasize that stability depends heavily on the metric used; in heavy-tailed regimes, raw space errors may lead to large tail prediction errors due to the Hölder nature of the bridge.

In summary, Buhai provides a unified mathematical toolkit that links the geometry of heterogeneity to the geometry of extremes, offering both robust stability guarantees for empirical work and explicit solutions for normative policy design in environments with random, heterogeneous sampling.