WorksMathematics
Bregman primal-dual methods for conic linear optimization via exact-penalty and self-dual formulations
Abstract
We study Bregman primal–dual proximal methods for conic linear programs (LPs). For problems involving non-polyhedral cones, Bregman distances lead to generalized projections that are often simpler than Euclidean projections. In this paper, we reformulate conic LPs as equivalent exact-penalty problems over truncated cones, and use self-dual embeddings to construct such formulations systematically. This approach yields Bregman proximal methods with simpler subproblems and convergence guarantees over bounded domains. We also study line searches for the Bregman primal–dual hybrid gradient and extragradient methods applied to self-dual problems. For logarithmically homogeneous kernels, the Bregman distances from the starting points are unbounded on the truncated cone, and the classical ergodic bounds do not apply. For these kernels, we derive convergence rates for the exact-penalty objective at the averaged iterates of both methods. Numerical experiments with nonnegative trigonometric polynomial optimization and sparse semidefinite optimization indicate the potential of this framework as a general-purpose approach for incorporating efficient Bregman projections in conic linear optimization.