Abstract
Shape constraints, motivated by either application-specific assumptions or existing theory, can be imposed during model estimation to restrict the feasible region of the parameters. Although such restrictions may not provide any benefits in an asymptotic analysis, they often improve finite sample performance of statistical estimators and the computational efficiency of finding near-optimal control policies. This paper reviews an illustrative set of research utilizing shape constraints in the economics and operations research literature, highlighting methodological innovations in utility functions, production economics, and sequential decision-making.
Introduction
In empirical economics and operations research, the tension between model flexibility and theoretical consistency is a recurring challenge. While nonparametric models offer the freedom to capture complex patterns without assuming specific functional forms, they often struggle with noisy or limited data. In a perfect world with infinite data, a sufficiently flexible machine learning model would eventually learn the underlying laws of physics or economics. In the real world, data is noisy and scarce. A standard nonparametric model might "learn" that adding more fertilizer reduces crop yields simply because of a few outliers in a small sample, violating basic economic theory.
Shape constraints—mathematical restrictions such as monotonicity, convexity, or concavity—provide a robust solution by embedding fundamental economic principles directly into the estimation process. By baking theory into the model, researchers can achieve better performance on small datasets and solve complex optimization problems that would otherwise be computationally impossible.
Methodological Pillars
The literature on shape constraints generally falls into two categories: regression-based estimation and sequential decision-making.
1. Regression-Based Estimators
Nonparametric Least Squares (LSE) serves as the gold standard for multivariate convex regression. A key innovation in this area is the application of Afriat inequalities, a set of linear constraints that ensure a function is convex and consistent with utility theory.
Historically, the computational burden of these models was high, as the number of constraints grew quadratically ($O(n^2)$) with the sample size. However, recent innovations in iterative algorithms—which selectively add only the most "violated" constraints—have enabled these methods to scale from a few hundred data points to several thousand. To handle high-dimensional data, Single Index Models project predictors onto a single dimension before applying a shape-constrained link function, effectively balancing flexibility with structural stability.
2. Sequential Decision Making (ADP/RL)
In Reinforcement Learning and Approximate Dynamic Programming (ADP), shape constraints are not just about accuracy; they are about speed. Stochastic Dual Dynamic Programming (SDDP) leverages the fact that many "value functions" in operations research (such as the cost of managing a power grid) are convex. By approximating these functions with a series of linear "cuts" (hyperplanes), SDDP can solve massive problems that standard dynamic programming cannot. Other methods, such as SPAR (Separable, Projective Approximation Routine), handle concave functions with integer-valued steps, which is particularly useful in inventory management.
Empirical Applications and Insights
As a review of the field, this research synthesizes decades of applied work across various sectors:
- Agriculture: Early studies on crop output versus nitrogen fertilizer used shape constraints to ensure models respected the law of diminishing returns.
- Industrial Planning: SDDP has moved from theory to practice, successfully managing complex hydrothermal power systems with up to 69 interconnected reservoirs.
- Consumer Behavior: Revealed preference tests using Afriat inequalities have become the standard for checking if consumer behavior is "rational" without assuming a specific functional form for utility.
One of the most critical insights is the finite sample gain. While shape constraints may not improve "asymptotic" performance (the behavior of the model as data approaches infinity), they drastically improve performance on the small, messy datasets typical of real-world economics. Furthermore, the authors discuss the "S-Curve" reality—the transition from increasing returns to scale (specialization) to diminishing returns (scarcity)—and highlight new kernel-weighted estimators (SCKLS) designed to capture these nuances.
Limitations and Future Directions
Despite the progress in the field, several hurdles remain for future research:
- Endogeneity: In many economic models, inputs are not independent of the noise (e.g., a manager might increase fertilizer because they see favorable weather). Combining shape constraints with instrumental variables is an active area of study.
- Global vs. Local Constraints: Many current models only impose constraints at specific data points. Ensuring a model is globally consistent (e.g., convex everywhere) without excessive computational cost remains a challenge.
- Automated Bandwidth Selection: For kernel-based methods, choosing the right "smoothing" parameter while maintaining shape constraints is still more of an art than a science.
Conclusion
Shape constraints represent a powerful bridge between flexible data-driven modeling and rigid theoretical assumptions. By restricting the search space to functions that make "economic sense," researchers can produce more robust estimators and more efficient control policies, even in the presence of high-dimensional or noisy data.
Summary updated