Let $\mathbf{\Sigma}$ be a $N \times N$ Hermitian and positive definite complex matrix. Let us define the matrix $\mathbf{A}$ as \begin{equation} \mathbf{A} \triangleq \mathbf{\Sigma}^{-T} \otimes \mathbf{\Sigma}^{-1} - \frac{1}{N} \mathrm{vec}(\mathbf{\Sigma}^{-1})\mathrm{vec}(\mathbf{\Sigma}^{-1})^H, \end{equation} where $(\cdot)^T$ and $(\cdot)^H$ are the transpose and the Hermitian operators, respectively. Moreover $\mathrm{vec}$ is the operator stacking the columns of a matrix on top of each other and $\otimes$ indicates the Kronecker product.

Is it possible to write the matrix $\mathbf{C}$ such that $\mathbf{A}=\mathbf{C}^H\mathbf{C}$, explicitly as function of $\mathbf{\Sigma}$ (or $\mathbf{\Sigma}^{-1}$)?


  • $\begingroup$ By $\Sigma^{-T}$ you mean $(\Sigma^T)^{-1}$? $\endgroup$ – jobe Mar 9 '18 at 13:53
  • $\begingroup$ Yes, exactly. _ $\endgroup$ – Vuk Mar 9 '18 at 13:57

Yes. Let $P=\Sigma^{-1/2}$. Then $P^T\otimes P$ is Hermitian (because $P$ is) and $\operatorname{vec}(\Sigma^{-1})=(P^T\otimes P)\operatorname{vec}(I)$. Therefore \begin{align} A&=(P^T\otimes P)^2-\frac1N(P^T\otimes P)\operatorname{vec}(I)\operatorname{vec}(I)^H(P^T\otimes P)\\ &=(P^T\otimes P)\left(I-\frac1N\operatorname{vec}(I)\operatorname{vec}(I)^T\right)(P^T\otimes P) \end{align} Note that $\Pi=I-\frac1N\operatorname{vec}(I)\operatorname{vec}(I)^T$ is the orthogonal projection on the linear space of (vectorised) traceless matrices. Therefore $\Pi^H\Pi=\Pi^2=\Pi$. Consequently, $A=C^HC$ when $C=\Pi(P^T\otimes P)$. As every positive definite matrix has a unique positive definite square root, $P$ and $C$ are indeed functions of $\Sigma$.

  • $\begingroup$ Can we go a little further? For example, the matrix $\mathbf{I} - \frac{1}{N}\mathrm{vec}(\mathbf{I})\mathrm{vec}(\mathbf{I})^T$ can be interpreted as the orthogonal projection matrix on the orthogonal complement of the $\mathrm{span}(\mathrm{vec}(\mathbf{I}))$, i.e. $\Pi^{\perp}_{\mathrm{vec}(\mathbf{I})}$, right? Can this help us to simplify the expression? $\endgroup$ – Vuk Mar 14 '18 at 9:08
  • $\begingroup$ @Vuk Yes, as you said, it is an orthogonal projection, so you may take itself as its positive semidefinite square root. $\endgroup$ – user1551 Mar 14 '18 at 10:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.