The title says it all! When using orthogonal series expansions like the Gram-Charlier expansion to approximate probability density function, a big problem (making this approach less usefull and less mused than it could have been otherwise), is that, the approximation might be negative in some intervals, while a probability density function cannot be negative.
A quick thought assures us that two positive functions cannot be orthogonal, at lest not for the usual inner products of type $\int f(x) g(x) w(x) d\;x$.
So a natural idea is to ask for (non-orthogonal) expansions where each term in the expansion is non-negative. Does anybody know of any work in this direction?