Title: Generalized R-squared for Detecting Non-independence

Abstract: Detecting non-independence between two random variables is a fundamental problem in statistics and machine learning. Although the celebrated Pearson correlation is effective for capturing linear dependency, it can be entirely powerless for detecting nonlinear and/or heteroscedastic patterns. We introduce a new measure, G-squared, as a generalization of the classic R-squared statistic to test whether two univariate random variables are mutually independent and to measure the strength of their relationship. The G-squared has an intuitive meaning of the piece-wise R-squared between the two variables. It is thus almost identical to the R-squared for linear relationships with constant error variances, and is particularly effective in handling nonlinearity and heteroscedastic errors. We propose two statistics to estimate the population G-squared and show that they are both consistent. Through intensive simulation studies, we demonstrate that the G-squared statistics are among the most powerful performers compared with several state-of-art methods.