Conic programming to understand sums of squares of eigenvalues of graphs

TL;DR

This paper proves a conjecture relating eigenvalues of graphs to the vector chromatic number using conic programming, introduces a strengthened Cauchy-Schwarz inequality, and discusses computational complexity of related semidefinite programs.

Contribution

It establishes a new bound on eigenvalues of graphs based on the vector chromatic number and proves NP-hardness of certain rank-constrained semidefinite programs.

Findings

Proved a conjecture bounding eigenvalues using the vector chromatic number.

Strengthened the Cauchy-Schwarz inequality for Hermitian matrices.

Showed NP-hardness of computing rank-constrained solutions to the vector chromatic number SDP.

Abstract

In this paper we prove a conjecture by Wocjan, Elphick and Anekstein (2018) which upper bounds the sum of the squares of the positive (or negative) eigenvalues of the adjacency matrix of a graph by an expression that behaves monotonically in terms of the vector chromatic number. One of our lemmas is a strengthening of the Cauchy-Schwarz inequality for Hermitian matrices when one of the matrices is positive semidefinite. A related conjecture due to Bollob\'as and Nikiforov (2007) replaces the vector chromatic number by the clique number and sums over the first two eigenvalues only. We prove a version of this conjecture with weaker constants. An important consequence of our work is a proof that for any fixed $r$, computing a rank $r$ optimum solution to the vector chromatic number semidefinite programming is NP-hard. We also present a vertex weighted version of some of our results, and we show how it leads quite naturally to the known vertex-weighted version of the Motzkin-Straus quadratic optimization formulation for the clique number.