Duality for Non Convex Composite Functions via the Fenchel Rockafellar Perturbation Framework (2510.05741v1)

Abstract: We examine the duality theory for a class of non-convex functions obtained by composing a convex function with a continuous one. Using Fenchel duality, we derive a dual problem that satisfies weak duality under general assumptions. To better understand this duality, we compare it with classical Lagrange duality by analyzing a related, yet more complex, constrained problem. We demonstrate that the newly derived stationarity conditions are equivalent to the classical Lagrange stationarity conditions for the constrained problem, as expected by the close relationship between Fenchel and Lagrange dualities. We introduce two non-convex optimization problems and prove strong duality results with their respective duals. The second problem is a constrained optimization problem whose dual is obtained through the concurrent use of the duality theory introduced in this paper and classical Lagrange duality for constrained optimization. We also report numerical tests where we solve randomly generated instances of the presented problems using an ad-hoc primal-dual potential reduction interior point method that directly exploits the global optimality conditions established in this paper. The results include a comparison with a well-known conic programming solver applied to the convex duals of the analyzed problems. The interior point method successfully reduces the duality gap close to zero, validating the proposed duality framework. The theory presented in this paper can be applied to various non-convex problems and serves as a valuable tool in the field of hidden convex optimization.