First-order SDSOS-convex semi-algebraic optimization and exact SOCP relaxations

Abstract: In this paper, we define a new type of nonsmooth convex function, called {\em first-order SDSOS-convex semi-algebraic function}, which is an extension of the previously proposed first-order SDSOS-convex polynomials (Chuong et al. in J Global Optim 75:885--919, 2019). This class of nonsmooth convex functions contains many well-known functions, such as the Euclidean norm, the $\ell_1$-norm commonly used in compressed sensing and sparse optimization, and the least squares function frequently employed in machine learning and regression analysis. We show that, under suitable assumptions, the optimal value and optimal solutions of first-order SDSOS-convex semi-algebraic programs can be found by exactly solving an associated second-order cone programming problem. Finally, an application to robust optimization with first-order SDSOS-convex polynomials is discussed.