V. V. Pozdyayev - Ph.D. (Phys.-Math.), Associate Professor, Arzamas Polytechnic Institute (branch NSTU). E-mail: vitalybolnokin@yandex.ru; d.i.mutin@mail.ru

Optimization problems with polynomial objective function and polynomial scalar or matrix inequality (PI/PMI) constraints are considered. A framework has been proposed that further transforms the existing moment theory based method of PI/PMI solution, and allows to construct equivalent solution algorithms working in the augmented original search space instead of the moment space. The equivalence here implies that the new optimization method is to find an r-atomic measure whose support set is the set of global optima, hence the name "atomic optimization". The new optimization method provides a way of solving PI/PMI problems that has much lower computational cost than the original moment theory based method while keeping its key benefits for the given class of problems. In this paper, optimization problems with polynomial objective function and inequality constraints are considered with the purpose of extending the atomic optimization method by enabling it to solve primal-dual forms of said problems - linear matrix inequality (LMI) relaxations. We show how the primal-dual form of a linear matrix inequality relaxation can be modified in a way that preserves its compatibility with the atomic optimization framework. By applying the search space transformation and eliminating all references to the intermediate (moment) space we obtain a system ready for application of various interior-point methods. For 1D problems and same-dimension search space transformations, an atom space based equivalent of a primal-dual LMI relaxation is constructed. This equivalent problem can be directly solved within the existing atomic optimization framework by using primal-dual semidefinite programming methods. This study provides a starting point for further research of atomic primal-dual forms.