Convexification of Multi-period Quadratic Programs with Indicators

TL;DR

This paper develops a convexification approach for multi-period quadratic programs with indicator variables, providing a tight second-order cone formulation and a polynomial-time algorithm, with applications in learning and control.

Contribution

It introduces a novel convex hull representation and a second-order cone programming formulation for multi-period quadratic problems with indicators, along with an efficient shortest path algorithm.

Findings

Derived a closed-form inverse for block-structured cost matrices.

Established a tight convex hull representation of the quadratic epigraph.

Proposed a polynomial-time shortest path algorithm for the reformulated problem.

Abstract

We study a multi-period convex quadratic optimization problem, where the state evolves dynamically as an affine function of the state, control, and indicator variables in each period. We begin by projecting out the state variables using linear dynamics, resulting in a mixed-integer quadratic optimization problem with a (block-) factorizable cost matrix. We discuss the properties of these matrices and derive a closed-form expression for their inverses. Employing this expression, we construct a closed convex hull representation of the epigraph of the quadratic cost over the feasible region in an extended space. Subsequently, we establish a tight second-order cone programming formulation with $\mathcal{O}(n^2)$ conic constraints. We further propose a polynomial-time algorithm based on a reformulation of the problem as a shortest path problem on a directed acyclic graph. To illustrate the applicability of our results across diverse domains, we present case studies in statistical learning and hybrid system control.