We propose a simple framework for nonlinear instrumental variable (IV) regression based on a kernelized conditional moment restriction (CMR) known as a maximum moment restriction (MMR). The MMR is formulated by maximizing the interaction between the residual and functions of IVs that belong to a unit ball of reproducing kernel Hilbert space (RKHS). This allows us to tackle the IV regression as an empirical risk minimization where the risk depends on the reproducing kernel on the instrument and can be estimated by a U-statistic or V-statistic. This simplification not only enables us to derive elegant theoretical analyses in both parametric and non-parametric settings, but also results in easy-to-use algorithms with a justified hyper-parameter selection procedure. We demonstrate the advantages of our framework over existing ones using experiments on both synthetic and real-world data.

| Link | BibTex | Poster | Slide | Video | Code |