Diagonalization of skew-Hermitian dual quaternion matrices and \phi-Hermitian dual quaternion matrices

Note that for N\in\mathbb{C}^{n\times n}, \boldsymbol{j}N=\overline{N}\boldsymbol{j}. The fact that \boldsymbol{j}N is skew-Hermitian is equivalent to N^{\ast}\boldsymbol{j}=\boldsymbol{j}N(=\overline{N}\boldsymbol{j}). Then, \boldsymbol{j}N is skew-Hermitian if and only if N^{T}=N, i.e., N is symmetric complex.

If a=0, then p=(b\boldsymbol{i}+c\boldsymbol{j})\epsilon. By Lemma 3.5, the infinitesimal part b\boldsymbol{i}+c\boldsymbol{j} is similar to the complex number \sqrt{b^{2}+c^{2}}\boldsymbol{i}\in\mathbb{C}\subseteq\mathbb{DC}. That is, there is a nonzero quaternion s such that

In particular, if we take s to have unit norm, then the similarity becomes a unitary equivalence.

If a\neq 0, let u=1+\frac{c\boldsymbol{k}}{2a}\epsilon and v=\frac{u}{|u|}. Then v^{\ast}v=vv^{\ast}=1 and

Therefore, whether a is zero or not, p is unitarily equivalent to a dual complex number.

Denote A=A_{st}+A_{\mathcal{I}}\epsilon\in\mathbb{DQ}^{n\times n}, where A_{st},A_{\mathcal{I}}\in\mathbb{Q}^{n\times n}. Then A is a skew-Hermitian dual quaternion matrix if and only if both A_{st} and A_{\mathcal{I}} are skew-Hermitian quaternion matrices. It follows from Lemma 3.2 that there is a unitary quaternion matrix U\in\mathbb{Q}^{n\times n} and a diagonal matrix \Lambda with the form in (3.2) such that

Then

where each C_{ij} is a quaternion matrix of matching dimension and each C_{ii} is a skew-Hermitian quaternion matrix. Let C_{ij}=M_{ij}+\boldsymbol{j}N_{ij}, where M_{ij} and N_{ij} are complex matrices with the same dimensions as C_{ij}. Define

where for 1\leq s,t\leq r, each P_{st} is defined as follows,

1. (1). if s=t, P_{st}=I_{k_{s}}=I_{k_{t}};

2. (2). if s<t, P_{st}=\frac{M_{st}}{r_{s}\boldsymbol{i}-r_{t}\boldsymbol{i}}\epsilon+(r_{s}\boldsymbol{i}+r_{t}\boldsymbol{i})^{-1}\boldsymbol{j}N_{st}\epsilon;

3. (3). if s>t, P_{st}=-(P_{ts})^{\ast}= \frac{M_{ts}^{\ast}}{r_{t}\boldsymbol{i}-r_{s}\boldsymbol{i}}\epsilon-N_{ts}^{\ast}\boldsymbol{j}(r_{t}\boldsymbol{i}+r_{s}\boldsymbol{i})^{-1}\epsilon.

Since the matrix P is the sum of the identity matrix I_n and a skew-Hermitian matrix, it follows from Lemma 3.7 that P is unitary and

Notice that each M_{ii} is a skew-Hermitian complex matrix and each N_{ii} is a symmetric complex matrix by Lemma 3.1. Thus we can find unitary complex matrices V_{11},\ldots,V_{rr} such that each

has the same block form as the matrix \Lambda in Lemma 3.2. That is,

where \rho_{i1}<\cdots<\rho_{id_{i}},\ l_{i1}+\cdots+l_{id_{i}}=k_{i}. For each i such that r_i>0, set

Denote V=\mathrm{diag}(V_{11},\ldots,V_{rr}). Then V is also a unitary complex matrix in \mathbb{C}^{n\times n} and

is also skew-Hermitian. Therefore, each E^{ii} is symmetric complex. We partition the matrix E^{ii} according to the same block form as matrix \Lambda_{ii} in (?). That is,

where E^{ii}_{st}\in\mathbb C^{l_{is}\times l_{it}} for 1\leq s,t\leq d_i.

We now focus on block-diagonalizing A_{3} under unitary equivalence, where the block structure follows that of matrix \Lambda_{ii}. One thing to note here is that, during the calculation, we consider the following two cases depending on whether r_{1} is zero or not.

1. (1). r_{1}\neq 0. For integers 1\leq h\leq r, let

   Q_{h}=\begin{pmatrix} I_{l_{h1}} & (2r_{h}\boldsymbol{i})^{-1}\boldsymbol{j} E^{hh}_{12}\epsilon & \cdots & (2r_{h}\boldsymbol{i})^{-1}\boldsymbol{j} E^{hh}_{1d_{h}}\epsilon \\ -(E^{hh}_{12})^{\ast}\boldsymbol{j}(2r_{h}\boldsymbol{i})^{-1}\epsilon& I_{l_{h2}} & \cdots & (2r_{h}\boldsymbol{i})^{-1}\boldsymbol{j}E^{hh}_{2d_{h}}\epsilon \\ \vdots & \vdots & \ddots & \vdots \\ -(E^{hh}_{1d_{h}})^{\ast}\boldsymbol{j}(2r_{h}\boldsymbol{i})^{-1}\epsilon & -(E^{hh}_{2d_{h}})^{\ast}\boldsymbol{j}(2r_{h}\boldsymbol{i})^{-1}\epsilon & \cdots & I_{l_{hd_{h}}} \end{pmatrix}.

   It follows from the form of Q_h and Lemma 3.7 that each Q_{h} is unitary. Denote

   Q=\begin{pmatrix} Q_{1}& & \\ & \ddots & \\ & & Q_{r} \end{pmatrix},

   then Q is a unitary matrix in \mathbb{DQ}^{n\times n}. And furthermore,

   A_{4}=QA_{3}Q^{\ast}=\begin{pmatrix} D_{1}& & \\ & \ddots & \\ & & D_{r} \end{pmatrix},

   where for 1\leq p\leq r, the block matrix

   D_{p}=\begin{pmatrix} r_{p}I_{l_{p1}}\boldsymbol{i}+\rho_{p1}I_{l_{p1}}\boldsymbol{i}\epsilon+\boldsymbol{j}E^{pp}_{11}\epsilon & & \\ & \ddots & \\ & & r_{p}I_{l_{pd_{p}}}\boldsymbol{i}+\rho_{pd_{p}}I_{l_{pd_{p}}}\boldsymbol{i}\epsilon+\boldsymbol{j}E^{pp}_{d_{p}d_{p}}\epsilon \end{pmatrix}.

   Clearly, A_{4} is a block diagonal matrix.

2. (2). r_{1}= 0. For 2\leq h\leq r, we repeat the process in the case r_1\neq0. Now, the only block that remains to be handled is the one containing r_1. This block becomes \Lambda_{11}\epsilon+\boldsymbol{j}E^{11}\epsilon. Notice that \Lambda_{11}+\boldsymbol{j}E^{11} is a skew-Hermitian quaternion matrix. Then, by Lemma 3.2, there exists a unitary quaternion matrix Q_1\in\mathbb{Q}^{k_1\times k_1} such that

   Q_1(\Lambda_{11}+jE^{11})Q_1^* =\operatorname{diag}(y_1\boldsymbol{i},\ldots,y_{k_1}\boldsymbol{i}), \qquad 0\leq y_1\leq\cdots\leq y_{k_1}.

   Together with the matrices Q_h, 2\leq h\leq r, constructed as in case (1), let

   Q=\operatorname{diag}(Q_1,Q_2,\ldots,Q_r).

   Then A_4=QA_3Q^* is block diagonal, and its first diagonal block is

   D_1=\operatorname{diag} (y_1\boldsymbol{i}\epsilon,\ldots,y_{k_1}\boldsymbol{i}\epsilon).

   In this case, set

   R_{1\ell}=y_\ell,\qquad \ell=1,\ldots,k_1.

Therefore, in either case, we have successfully transformed A_{3} into the block-diagonal matrix A_{4}. Now for A_{4}, since each E^{ii}_{qq} is symmetric complex, we may find a unitary complex matrix Z^{i}_{q}\in\mathbb{C}^{l_{iq}\times l_{iq}} such that each

is nonnegative diagonal by Lemma 3.3. Here 1\leq i\leq r when r_1\neq0, whereas 2\leq i\leq r when r_1=0. If r_1\neq0, let

If r_1=0, let

If r=1 and r_1=0, we simply take Z=I_{k_1}. Then Z is a unitary matrix in \mathbb{C}^{n\times n}. It then follows from Lemma 3.4 that,

If r_1=0, the first block is already diagonal:

All the remaining blocks, as well as all the blocks in the case r_1\neq0, have the form

Here 2\leq t\leq r if r_1=0, whereas 1\leq t\leq r if r_1\neq0.

If we set X=Z^{\ast}QVPU, then X is a unitary matrix in \mathbb{DQ}^{n\times n} and A_{5}=XAX^{\ast}.

So far, we have completed the task of diagonalizing the original matrix A under unitary equivalence. To make the diagonal entries of A_{5} look more orderly, we now proceed with the following discussion. By Lemma 3.6, each diagonal element of matrix A_{5} can be unitarily equivalent to a dual complex number. If we place in order along the diagonal all those unitary elements that convert the diagonal entries of A_{5} into dual complex numbers, the resulting matrix is unitary, we denote it by Y. Then furthermore,

Let W=YX. Then \Sigma=WAW^{\ast}. Therefore, under unitary equivalence, the matrix A is finally transformed into the desired form.

By Lemma 3.6, when r_i>0, each diagonal entry

is unitarily equivalent to

When r_1=0, the first block is already

Hence \Sigma has exactly the form in (3.8), with

Let \sigma_m be the m-th diagonal entry of \Sigma, and let e_m be the m-th standard basis vector. Since

and \Sigma=WAW^*, we have

Moreover, W^*e_m is appreciable because W is unitary. Therefore, every diagonal entry \sigma_m of \Sigma is a right eigenvalue of A.

Suppose that B is \phi-Hermitian. This is equivalent to the fact that \phi B is skew-Hermitian by Lemma 4.8. Then we can find a unitary matrix X\in\mathbb{DQ}^{n\times n} such that

where

By Lemma 3.5, \boldsymbol{i} and \phi are unitarily equivalent. Therefore, the diagonal matrix \Sigma can be transformed into

through a unitary transformation. We may denote this process by Y\Sigma Y^{\ast}=\Xi, where Y is unitary. Now we have

which is equivalent to

Denote \Gamma=\overline{\phi}\Xi, then \Gamma possesses the desired form in (4.9). By Lemma 4.2, we may write V=\overline{(YX)}^{\phi}. Then \overline{V}^{\phi}=YX and furthermore V^{*{\phi}}=(YX)^{\ast}. Thus, finally

Recall that for a matrix A=A_{st}+A_{\mathcal{I}}\epsilon\in\mathbb{DQ}^{n\times n}, A is \phi-Hermitian if and only if both A_{st} and A_{\mathcal{I}} are \phi-Hermitian. Then both A_{st} and A_{\mathcal{I}} are in the set \mathbb{C}_{I_{\phi}}^{n\times n} by Lemma 4.11. Therefore,

where

Hence A is a Hermitian matrix in \mathbb{C}_{I_{\phi}}^{n\times n}+ \mathbb{C}_{I_{\phi}}^{n\times n}\epsilon by Lemma 4.12. Following from the construction of the unitary matrix in the proof of Lemma 4.6 ([QL, Theorem 4.1]), if we replace every \mathbb{Q}^{n\times n} in that process by \mathbb{C}_{I_{\phi}}^{n\times n}, then we obtain that there exists a unitary matrix W\in \mathbb{C}_{I_{\phi}}^{n\times n}+ \mathbb{C}_{I_{\phi}}^{n\times n}\epsilon \subseteq \mathbb{DQ}^{n\times n} such that

where \Lambda is a diagonal matrix and all the diagonal entries of \Lambda are dual numbers. Furthermore, by W\in \mathbb{C}_{I_{\phi}}^{n\times n}+ \mathbb{C}_{I_{\phi}}^{n\times n}\epsilon, we have