Abstract:We investigate the relationship between conditional independence (CI) x perpendicular to y vertical bar Z and the independence of two residuals x - E(x vertical bar Z) perpendicular to y - E(y vertical bar Z), where x and y are two random variables, and Z is a set of random variables. We show that if x, y and Z are generated by following linear structural equation model and all external influences follow Gaussian distributions, then x perpendicular to y vertical bar Z if and only if x - E(x vertical bar Z) perpendicular to y - E(y vertical bar Z). That is, the test of x perpendicular to y vertical bar Z can be relaxed to a simpler unconditional independence test of x - E(x vertical bar Z) perpendicular to y - E(y vertical bar Z). Furthermore, if all these external influences follow non-Gaussian distributions and the model satisfies structural faithfulness condition, then we have x perpendicular to y vertical bar Z double left right arrow x - E(x vertical bar Z) perpendicular to y - E(y vertical bar Z).We apply the results above to the causal discovery problem, where the causal directions are generally determined by a set of V- structures and their consistent propagations, so CI test- based methods can return a set of Markov equivalence classes. We show that in linear non- Gaussian context, x - E(x vertical bar Z) perpendicular to y - E(y vertical bar Z) perpendicular to x - E(x vertical bar Z) perpendicular to z or y - E(y vertical bar Z) perpendicular to z (for all z is an element of Z) if Z is a minimal d- separator, which implies z causes x (or y) if z directly connects to x (or y). Therefore, we conclude that CIs have useful information for distinguishing Markov equivalence classes.In summary, compared with the existing discretization-based and kernel-based CI testing methods, the proposed method provides a simpler way to measure CI, which needs only one unconditional independence test and two regression operations. When being applied to causal discovery, it can find more causal relationships, which is experimentally validated.

CONDITIONAL DISTANCE CORRELATION.

Conditional Local Distance Correlation For Manifold-Valued Data

A Distribution Free Conditional Independence Test with Applications to Causal Discovery

A Conditional Distribution Function-Based Measure for Independence and K-sample Tests in Multivariate Data

ASYMPTOTIC DISTRIBUTIONS OF HIGH-DIMENSIONAL DISTANCE CORRELATION INFERENCE

Measures of conditional dependence for nonlinearity, asymmetry and beyond

Distance-based and RKHS-based dependence metrics in high dimension

Semi-Distance Correlation and Its Applications

Unbiased Multivariate Correlation Analysis.

CHARACTERISTIC FUNCTION BASED TESTING FOR CONDITIONAL INDEPENDENCE: A NONPARAMETRIC REGRESSION APPROACH

A Kernel-Based Measure for Conditional Mean Dependence

Distance Correlation Test for High-Dimensional Independence

Measuring Conditional Independence By Independent Residuals: Theoretical Results And Application In Causal Discovery

Detecting practically significant dependencies in infinite dimensional data via distance correlations

The Conditional Distance Autocovariance Function

A Ball Divergence Based Measure For Conditional Independence Testing

The Chi-Square Test of Distance Correlation

Distance Correlation in Multiple Biased Sampling Models

The affinely invariant distance correlation

Testing Independence with High-Dimensional Correlated Samples

Measuring Conditional Independence by Independent Residuals for Causal Discovery