Abstract
Structural equation models (SEMs) are widely used in sciences, ranging fromeconomics to psychology, to uncover causal relationships underlying a complexsystem under consideration and estimate structural parameters of interest. Westudy estimation in a class of generalized SEMs where the object of interest isdefined as the solution to a linear operator equation. We formulate the linearoperator equation as a min-max game, where both players are parameterized byneural networks (NNs), and learn the parameters of these neural networks usingthe stochastic gradient descent. We consider both 2-layer and multi-layer NNswith ReLU activation functions and prove global convergence in anoverparametrized regime, where the number of neurons is diverging. The resultsare established using techniques from online learning and local linearizationof NNs, and improve in several aspects the current state-of-the-art. For thefirst time we provide a tractable estimation procedure for SEMs based on NNswith provable convergence and without the need for sample splitting.