While machine learning techniques are very powerful, they have some weaknesses, like iterative optimization with many local minimums, large freedom of parameters, lack of their interpretability and accuracy control. From the other side we have classical statistics based on moments not having these issues, but providing only a rough description. I will talk about approach which combines their advantages: with MSE-optimal moment-like coefficients, but designed such that we can directly translate them into probability density. For multivariate case such basis of mixed