Generalized higher-order moments in independent component analysis
J.O. Coleman · 2002
In independent component analysis (ICA), random-variable independence is often equated with factorization of the joint moments, expectations of products of powers. This paper shows that many nonpower functions are equally useful: if E[f(X)g(Y)] factors into E[f(X)]E[g(Y)] for every f and g from an independence class, then random variables X and Y are independent. Examples of and sufficient conditions for independence classes are presented for bounded random variables.