Time-frequency tilings which best expose the non-Gaussian behavior of a stochastic process

We develop a new representation of non-Gaussian stochastic processes. We search a library of orthogonal bases for the basis in which the process looks the least Gaussian. When the library is a library of time-frequency atoms this has the interpretation given in the title. We give examples showing that the new representation can be more satisfactory than the classical Karhunen-Loeve expansion.