Homotopy Methods for Convex Optimization (2403.02095v1)

Abstract: Convex optimization encompasses a wide range of optimization problems, containing many efficiently solvable subclasses. Interior point methods are currently the state-of-the-art approach for solving such problems, particularly effective for classes like semidefinite programming, quadratic programming, and geometric programming. However, their success hinges on the construction of self-concordant barrier functions for the feasible sets. In this work, we introduce an alternative method for tackling convex optimization problems, employing a homotopy. With this technique, the feasible set of a trivial optimization problem is continuously transformed into the target one, while tracking the solutions. We conduct an analysis of this approach, focusing on its application to semidefinite programs, hyperbolic programs, and convex optimization problems with a single convexity constraint. Moreover, we demonstrate that our approach numerically outperforms state-of-the-art methods in several interesting cases.