Null spaces of correlation matrices (Q1399237)

It is an interesting problem to determine those \(n\times n\) correlation matrices that annihilate a given \(k\)-dimensional subspace of \(\mathbb{R}^n\). (A correlation matrix is real positive definite symmetric with all diagonal elements equal to 1). The solution is known when \(k=n-1\) or \(n-2\). The authors give the complete solution when \(k=n-3\) and \(n=4\) or 5. Those \(4 \times 4\) correlation matrices (\(a_{ij})\) that annihilate an (appropriately normalized) vector \(v\in\mathbb{R}^1\) are parametrized by the corresponding points \((a_{24}, a_{34})\) in \(\mathbb{R}^2\). The latter from a convex subset \(B_v\) of the closed square with vertices \((\pm 1,\pm 1)\), whose boundary \(B_v\) is the intersection with \(B\) of an explicitly given cubic curve. The solution for \(n=5\) is derived from that for \(n=4\). A number of worked examples illustrate the intriguing geometrical figures that can arise.