WorksMathematics
Many faces of integer semidefinite programming
Abstract
We first establish several properties of integer positive semidefinite matrices, extending known results for binary and ternary positive semidefinite matrices. We then show how to reformulate an integer quadratic program (IQP) as an equivalent integer semidefinite program (ISDP). We distinguish three cases, namely IQPs with two, three, or more than three distinct integer values, and provide several ISDPs for each case. Each reformulation yields either a binary ISDP, a ternary ISDP, or a rescaled version of one of these. The results show that ternary SDP formulations for problems with three distinct integer values provide SDP relaxations with fewer constraints and, in the case of equally spaced integers, also smaller matrix variables than binary-based SDP relaxations. For quadratic problems whose variables can take more than three distinct integer values, one can choose among three formulations based on the size of their matrix variables, the number of constraints, and the strength of the resulting SDP relaxations. We also show how to reformulate a mixed-integer QP as a mixed-integer SDP.