In the quadratic minimum spanning tree problem (QMSTP) one wants to find the minimizer of a quadratic function over all possible spanning trees of a graph. We give two formulations of the QMSTP as mixed-integer semidefinite programs exploiting the algebraic connectivity of a graph. Based on these formulations, we derive a doubly nonnegative relaxation for the QMSTP and investigate classes of valid inequalities to strengthen the relaxation using the Chvátal-Gomory procedure for mixed-integer conic programming.
Solving the resulting relaxations is out of reach for off-the-shelf software. We therefore develop and implement a version of the Peaceman-Rachford splitting method that allows to compute the new bounds for graphs from the literature. The numerical results demonstrate that our bounds significantly improve over existing bounds from the literature in both quality and computation time, in particular for graphs with more than 30 vertices.
This work is further evidence that semidefinite programming is a valuable tool to obtain high-quality bounds for problems in combinatorial optimization, in particular for those that can be modelled as a quadratic problem.