The Hardness of Conditional Independence Testing and the Generalised Covariance Measure (Q118262)

After proving general results on the testability of conditional independence, this paper proposes a testing procedure. This test procedure is based on evaluating conditional independence of $X$ and $Y$ conditional on $Z$ for the case where $X$ and $Y$ are univariate by nonlinearly regressing $X$ on $Z$ and $Y$ on $Z$, and deriving a test statistic, named the generalised covariance measure (GCM), from the sample covariance between the regression residuals. The validity of this test procedure is shown to be related to the regression procedures being able to estimate the conditional means at a suitable rate. The methodology is extended to handle settings where $X$ and $Y$ may be multivariate. In this case, the proof of the validity of the test additionally requires the errors to obey certain moment restrictions and slightly faster rates of convergence for the prediction errors. These theoretical guarantees are developed for kernel ridge regression. A simulation evaluation is presented. The code is available in the R package GeneralisedCovarianceMeasure.