The upper numerical range of a quaternionic matrix is not a complex numerical range (Q677895)

The paper deals with the study of convex properties of the numerical range of quaternionic matrices. If \(\mathbb{H}\) denotes the set of quaternions and \(Q \in \mathbb{H}^{n\times n}\), the numerical range of \(Q\) is defined by: \[ W(Q)={x^*Qx: x \in \mathbb{H}^{n\times n},\;x^*x=1}. \] If \(\mathbb{H}\) is replaced by \(\mathbb{C}\) then \(W(C)\) is the numerical range of a complex matrix \(C\), known to be a convex subset of \(\mathbb{C}\). For the quaternionic case however, \(W(Q)\) is not generally convex. If \(W(Q)\) is intersected with the upper complex plane \(\mathbb{C}_+\) one obtains the so calIed upper numerical range, denoted by \(B_+(Q)\). It happens that there exists a rotation group that applied to \(B_+(Q)\) fully specifies the numerical range of \(Q\). Using this fact one concentrates on properties of \(B_+(Q)\). It has been proved that \(B_+(Q)\) is convex although the known proofs of this fact are rather lengthy and complicated. In order to simplify those proofs it was noted by \textit{W. So} and \textit{R. C. Thompson} [Linear Multilinear Algebra 41, No. 4, 303-365 (1996; Zbl 0876.15022)] that if \(B_+(Q)\) was equal to the intersection of the upper complex plane with the numerical range of a complex matrix \(C\), then the convexity of \(B_+(Q)\) would follow immediatly from the convexity of \(W(C)\). The main result in this paper is to show that there exists \(Q\in \mathbb{H}^{n\times n}\) such that \(B_+(Q)\neq \mathbb{C}_+ \cap W(C)\), \(\forall C\in \mathbb{C}^{n\times n}\), which invalidates the attempts to have convexity of \(B_+(Q)\) using that of \(W(C)\). The key step to show that \(B_+(Q)\neq \mathbb{C}_+\cap W(C)\) for a particular choice of \(C\in \mathbb{C}^{n\times n}\) is the following result about the numerical range of a complex matrix \(C\): If the boundary of \(W(C)\) contains an elliptical arc, then the full ellipse and its interior lie in \(W(C)\).