We develop an asymptotic theory of adversarial estimators (`A-estimators'). Like maximum-likelihood-type estimators (`M-estimators'), both the estimator and estimand are defined as the critical points of a sample and population average respectively. A-estimators generalize M-estimators, as their objective is maximized by one set of parameters and minimized by another. The continuous-updating Generalized Method of Moments estimator, popular in econometrics and causal inference, is among the earliest members of this class which distinctly falls outside the M-estimation framework. Since the recent success of Generative Adversarial Networks, A-estimators received considerable attention in both machine learning and causal inference contexts, where a flexible adversary can remove the need for researchers to manually specify which features of a problem are important. We present general results characterizing the convergence rates of A-estimators under both point-wise and partial identification, and derive the asymptotic root-n normality for plug-in estimates of smooth functionals of their parameters. All unknown parameters may contain functions which are approximated via sieves. While the results apply generally, we provide easily verifiable, low-level conditions for the case where the sieves correspond to (deep) neural networks. Our theory also yields the asymptotic normality of general functionals of neural network M-estimators (as a special case), overcoming technical issues previously identified by the literature. We examine a variety of A-estimators proposed across econometrics and machine learning and use our theory to derive novel statistical results for each of them. Embedding distinct A-estimators into the same framework, we notice interesting connections among them, providing intuition and formal justification for their recent success in practical applications.
View on arXiv