The Bravais-Pearson linear correlation coefficient and the Sarmanov maximal coefficient are well known statistical tools that permit to measure, respectively, correlation (also called linear dependence) and stochastic dependence of two suitable random variables X 1 and X 2 defined on a probability space \((\Omega ,\mathcal {A},P)\) . Since these coefficients just are the first canonical coefficients obtained from linear and nonlinear canonical analysis, respectively, it is relevant to improve them by using all the canonical coefficients. In order to give an unified framework for these notions, we introduce the canonical analysis (CA) of two closed subspaces H 1 and H 2 of a Hilbert space H. Then, a class of measures of association that admits the aforementioned coefficients as particular cases can be constructed.