Sensitivity analysis of semi-simple eigenvalues of regular quadratic eigenvalue problems (Q2355368)

Consider the following quadratic eigenvalue problem: \[ [\lambda^2(p)M(p)+\lambda(p)C(p)+K(p)]u(p)=0, \;\;\lambda(p)\in \mathbb{C}, \;u(p)\in \mathbb{C}^n, \;p\in \mathcal{N}(p^*), \] where \(\mathcal{N}(p^*)\) is a neighborhood of \(p^*\in \mathbb{C}^n\), and \(M(p), C(p), K(p)\in \mathbb{C}^{n\times n}\) are analytic matrix-valued functions on \(\mathcal{N}(p^*)\). \textit{I. Gohberg} et al. [Adv. Appl. Math. 7, 253--281 (1986; Zbl 0623.15003)] and \textit{H. Langer} et al. [SIAM J. Matrix Anal. Appl. 13, No. 2, 474--489 (1992; Zbl 0752.15009)] studied the sensitivity analysis of these problems dependent on a single parameter. It is known [\textit{A. L. Andrew} et al., SIAM J. Matrix Anal. Appl. 14, No. 4, 903--926 (1993; Zbl 0786.15011)] that the simple eigenvalues and corresponding eigenvectors of these problems that are analytically dependent on several parameters are analytical. \textit{H. Q. Xie} and \textit{H. Dai} [Math. Numer. Sin. 28, No. 1, 75--88 (2006)] gave some results on the sensitivity analysis of multiple eigenvalues of these problems that are analytically dependent on several parameters, when \(M(p^*)\) is invertible. Let \(M, C, K\in \mathbb{C}^{n\times n}\). Consider the problem \((\lambda^2M+\lambda C+K)u=0\). Then, the matrices \(M, C, K\) corresponding to this problem are called matrix triple, denoted by \(\{M,C,K\}\). If \(\det(\lambda^2M+\lambda C+K)\) is nonzero for \(\lambda\in \mathbb{C}\), then the problem is called regular quadratic eigenvalue problem and \(\{M,C,K\}\) is called regular matrix triple. Let \(\{M,C,K\}\) be a regular matrix triple and \(\lambda_1, \dots, \lambda_r\) (\(r<+\infty\)) be \(r\) eigenvalues of this problem \((\lambda^2M+\lambda C+K)u=0\). Suppose \(\{\lambda_1, \dots, \lambda_r\}\) are eigenvalues of \(T_1\) and \(U_1\in \mathbb{C}^{n\times r}\), \(T_1\in \mathbb{C}^{r\times r}\) satisfy \(\mathrm{rank}\left[\begin{matrix} U_1 \\ U_1T_1 \end{matrix} \right]=r\) and \(MU_1T_1^2+CU_1T_1+KU_1=0\). Then, \(U_1\) is called eigenvector matrix of \(\{M,C,K\}\) corresponding to \(\lambda_1, \dots, \lambda_r\) and \((U_1, T_1)\) is called eigenmatrix pair of \(\{M,C,K\}\) corresponding to \(\lambda_1, \dots, \lambda_r\). In addition, \((V_1, T_1)\) is called left eigenmatrix pair of \(\{M,C,K\}\) corresponding to \(\lambda_1, \dots, \lambda_r\) if \((V_1, T_1^T)\) is an eigen-matrix pair of \(\{M^T,C^T,K^T\}\) corresponding to \(\lambda_1, \dots, \lambda_r\). In this case, when \(V_1^TMU_1T_1+T_1V_1^TMU_1+V_1^TCU_1=I_r\), \(\{U_1, V_1, T_1\}\) is called eigen-matrix triple of \(\{M,C,K\}\) corresponding to \(\lambda_1, \dots, \lambda_r\). In this paper, the author proves that the average of semisimple eigenvalues and their corresponding eigenmatrix triples of the quadratic eigenvalue problem are analytic and provides the explicit formulae for the partial derivatives of the semisimple eigenvalues. In addition, the discussion of the sensitivities of semisimple eigenvalues and their corresponding eigenvector matrices of the quadratic eigenvalue problems is given.