Reverse Mechanism Design

Nima Haghpanah, Jason D. Hartline

TL;DR

This paper introduces a novel "reverse engineering" approach to multi-dimensional mechanism design. By deriving specific virtual value functions, the authors identify natural conditions—such as positive correlation between a consumer's base value and their sensitivity to quality—under which simple pricing strategies (like selling only the highest-quality version of a product) are mathematically optimal.

What Problem Does This Paper Address?

In the world of monopoly pricing, selling a single item is easy: you pick a price, and anyone who values it more than that price buys it. This is the "posted price" mechanism, proven optimal for single items by Roger Myerson in 1981. However, when a seller offers products with multiple quality levels (second-degree price discrimination) or multiple different items (bundling), the "optimal" mechanism often becomes incredibly complex, sometimes requiring the seller to offer "lotteries" or confusing menus of options.

The core challenge is that in multi-dimensional settings, the "virtual values"—a mathematical tool used to transform a buyer's true value into a metric that represents the seller's marginal revenue—are difficult to define. In one dimension, there is only one "path" to calculate these values (integration by parts). In multiple dimensions, there are infinite paths, and most of them lead to mechanisms that aren't "incentive compatible" (meaning buyers might lie about their preferences to get a better deal).

Methods

The authors flip the standard problem on its head. Instead of starting with a distribution of buyer types and trying to find the optimal mechanism, they:

  1. Start with a simple family of mechanisms (like uniform pricing or grand bundling).
  2. Impose necessary conditions for these simple mechanisms to be optimal.
  3. Solve for the "paths" of integration that satisfy these conditions.
  4. Identify the probability distributions that make these paths work.

They use a concept called Amortization, which provides an upper bound on the expected revenue of any incentive-compatible mechanism. By finding a "tight" amortization where the upper bound equals the actual revenue of a simple mechanism, they prove that simple mechanism is the best possible choice. A key technical tool they introduce is the equi-quantile curve, which helps define the direction of the paths for integration by parts based on the specific distribution of buyer types.

Datasets

Not discussed in the paper. This is a theoretical work in microeconomics and algorithmic game theory; it relies on mathematical proofs and closed-form examples rather than empirical datasets.

Key Insights and Ideas

  • The "Sensitivity" Rule: Price discrimination is only effective if high-value consumers are more sensitive to quality differences than low-value consumers. If high-value consumers are less sensitive (positive correlation between value and quality discount factor), the seller is better off just selling the high-quality item at a single price.
  • Path Variability: Previous researchers tried to solve this by using fixed paths (like straight lines from the origin). This paper shows that the "correct" paths for integration depend entirely on the distribution of the buyers.
  • Robustness: The authors show that previous famous results (like those by Armstrong or Stokey) were actually "boundary cases." Their new conditions show that simple pricing remains optimal even when those previous models are slightly perturbed, provided the correlation remains positive.
  • Grand Bundling for Additive Preferences: For sellers with two different items, selling them only as a "grand bundle" is optimal when a higher total value for the bundle is negatively correlated with the ratio of the values of the individual items (i.e., when high-value buyers have more heterogeneous tastes).

Key Results

  • Theorem 6: Establishes that for "max-symmetric" distributions, uniform pricing is optimal if the favorite-quality version has a monotone virtual value and the ratio of values follows a specific stochastic dominance pattern.
  • Theorem 12: Provides the multi-dimensional "ironing" conditions (sweeping) that guarantee the optimality of grand bundling for additive preferences.
  • Generalization: The framework successfully unifies and extends several decades of research into a single methodology for "reducing" multi-dimensional problems into solvable single-dimensional projections.

Limitations and Future Work

  • Deterministic Focus: The authors' method primarily identifies when deterministic mechanisms (fixed prices) are optimal. However, in some weird market distributions, randomized mechanisms (lotteries) are known to be better. The current framework doesn't easily capture those cases.
  • Linearity: The results assume linear values and costs. While the authors note that virtual surplus optimization can handle some non-linearities, the specific applications in the paper stick to the linear case.
  • Complexity of "Sweeping": While they provide a method for "ironing" (handling non-regular distributions) in multiple dimensions, it is currently limited to relatively simple sweeping arguments. Expanding this to more complex, non-symmetric distributions remains an open challenge.

Summary updated

Abstract

We consider a multi-dimensional screening problem of selling a product with multiple quality levels and design virtual value functions to derive conditions that imply optimality of only selling highest quality. A challenge of designing virtual values for multi-dimensional agents is that a mechanism that pointwise optimizes virtual values resulting from a general application of integration by parts is not incentive compatible, and no general methodology is known for selecting the right paths for integration by parts. We resolve this issue by first uniquely solving for paths that satisfy certain necessary conditions that the pointwise optimality of the mechanism imposes on virtual values, and then identifying distributions that ensure the resulting virtual surplus is indeed pointwise optimized by the mechanism. Our method of solving for virtual values is general, and as a second application we use it to derive conditions of optimality for selling only the grand bundle of items to an agent with additive preferences.
BibTeX
@inproceedings{haghpanah2015reverse,
  author = {Nima Haghpanah and Jason D. Hartline},
  title = {{Reverse Mechanism Design}},
  booktitle = {EC},
  year = {2015},
  doi = {10.1145/2764468.2764498},
}
Figure