Abstract
In parametric, nonlinear structural models, a classical sufficient condition for local identification, like Fisher (1966) and Rothenberg (1971), is that the vector of moment conditions is differentiable at the true parameter with full rank derivative matrix. We derive an analogous result for the nonparametric, nonlinear structural models, establishing conditions under which an infinite dimensional analog of the full rank condition is sufficient for local identification. Importantly, we show that additional conditions are often needed in nonlinear, nonparametric models to avoid nonlinearities overwhelming linear effects. We give restrictions on a neighborhood of the true value that are sufficient for local identification. We apply these results to obtain new, primitive identification conditions in several important models, including nonseparable quantile instrumental variable (IV) models and semiparametric consumption-based asset pricing models.
Introduction: The Parametric Legacy
In econometrics, identification is the cornerstone of structural modeling. For parametric models, the conditions for local identification are well-established. Following the seminal work of Fisher (1966) and Rothenberg (1971), a model is locally identified if the vector of moment conditions is differentiable at the true parameter with a full-rank derivative matrix. This "full rank" condition ensures that the mapping from parameters to observables is locally injective.
However, as economic theory increasingly relies on nonparametric and semiparametric specifications—where the "parameter" of interest is an entire function—the classical parametric intuition often fails. This paper bridges that gap, establishing a rigorous framework for local identification in infinite-dimensional spaces.
The Nonparametric Challenge: Ill-Posedness and Nonlinearity
The transition from finite-dimensional vectors to infinite-dimensional functions introduces the "ill-posed inverse problem." In nonparametric models, a derivative operator might have a full rank (or be injective) without having a continuous inverse. This mathematical nuance has profound implications: small nonlinear "wiggles" in a function can grow large enough to cancel out the linear signal, making it impossible to distinguish the true function from nearby alternatives.
In such settings, the linear approximation (the derivative) does not necessarily dominate the nonlinear remainder in a standard neighborhood. Consequently, a full-rank derivative is a necessary but insufficient condition for local identification in the nonparametric world.
Theoretical Framework: Fréchet Derivatives and Restricted Neighborhoods
The authors utilize functional analysis to solve this problem, treating models as mappings between Banach or Hilbert spaces. The core of their contribution lies in several key methodological innovations:
1. Fréchet Differentiability
The model is analyzed through its Fréchet derivative, providing a linear approximation of the structural equations in function space. This allows the authors to treat the infinite-dimensional problem with the same rigor as the finite-dimensional case.
2. The Nonlinearity Bound (Assumption 2)
To prevent nonlinearities from overwhelming the linear signal, the authors introduce a critical bound on the approximation error. They require the remainder term of the Taylor expansion to be controlled by the distance from the truth raised to a power $r$. This ensures that as we move closer to the true parameter, the linear part of the model becomes the dominant force.
3. Restricted Neighborhoods
Rather than seeking identification in a simple ball around the truth, the paper defines specific "restricted neighborhoods"—often ellipsoids in Hilbert spaces. Within these sets, the linear signal is guaranteed to outweigh the nonlinear noise, ensuring local identification even when the inverse problem is ill-posed.
4. Partialling Out
For semiparametric models involving both a finite-dimensional parameter $\beta$ and a functional nuisance component $g$, the authors develop a method to "partial out" the infinite-dimensional part. This allows for the identification of the parameters of interest by isolating their effects from the functional noise.
Applications: From Quantile IV to Asset Pricing
The theoretical framework is applied to several high-stakes econometric models, providing new, primitive identification conditions:
- Nonseparable Quantile Instrumental Variable (IV) Models: The paper provides the first primitive identification conditions for these models with continuous regressors. While discrete cases were previously understood, the continuous case required the new "restricted neighborhood" approach to handle the infinite-dimensional functional parameter.
- Semiparametric Consumption-Based Asset Pricing (CCAPM): The authors derive new results for identifying discount factors and marginal utility functions. By applying the Krein-Rutman theorem (an infinite-dimensional version of the Perron-Frobenius theorem), they establish conditions under which solutions are unique, leading to global identification in complex, habit-based models.
- Single-Index IV Models: The research shows that "index" structures—which reduce high-dimensional data to a single index—significantly relax the requirements for instrumental variables, making identification more robust than in fully nonparametric settings.
Conclusion
This paper provides a theoretical safety net for researchers using complex structural models. By proving a "genericity" result, the authors show that in Hilbert spaces, the rank condition is satisfied for "most" data-generating processes. While the focus remains on local identification, the framework establishes the necessary mathematical boundaries—specifically the trade-off between model curvature and the size of the identification set—required to move toward a more complete understanding of nonparametric structural estimation.
Summary updated