We establish non-asymptotic minimax separation rates for univariate function selection in sparse additive models, and discuss the difference between optimal estimation and selection.