Arctan is non-convex (convex for negative arguments, concave for positive arguments), so can not be used in CVX. The only thing you can do with arctan in CVX is some type of convex (or concave, depending on usage) approximation, such as Taylor series, with only convex or concave terms, expanded about an appropriate point. This may not work well, even with successive approximation to refine the expansion point, but you’re welcome to try.