On the eigensystems of graded matrices (Q5956189)

Informally a graded matrix is one whose elements show a systematic decrease or increase as one pass across the matrix. It has long been recognized that the eigensystems of graded matrices have special properties. But to obtain a definition such that these properties hold is not so simple. In this paper, the author defines grading as a scaling of a base matrix \(B\), i.e., \(A=BD\), \(DB\) or \(D^{1/2}BD^{1/2}\), where \(D=\text{diag}(\delta_1,\dots,\delta_n)\), \(\delta_1\geq\dots\geq\delta_n>0\). The numbers \(\rho_k=\delta_{k+1}/\delta_k\) are called the grading ratios. It is shown that when the grading is sufficiently strong, the matrix \(A\) can be reduced by a similarity transformation to a block diagonal matrix. Moreover, as the grading increases, the reducing transformation approaches a fixed limit that is independent of the grading. By calculating the eigenvectors of the diagonal blocks of the block triangular matrix one can compute approximations to the eigenvectors of the original matrix that amount to scaling certain essentially constant vectors. Let the base matrix be partitioned to \(B=\left(\begin{smallmatrix} B_{kk}&B_{kn}\\ B_{nk}&B_{nn}\end{smallmatrix}\right)\), where \(B_{kk}\) and \(B_{nn}\) are square matrices of orders \(k\) and \(n-k\) respectively. Then the number \(\kappa_k=\|B_{kk}^{-1}\|\|B\|\) is called the \(k\)th grading impediment. The number \(\gamma_k= \kappa_k\rho_k\) is called the \(k\)th grading coefficient. It is shown that the condition of the eigenvalues and eigenvectors of the graded matrix \(A=BD\) depends on the grading impediments, not the grading coefficients. Once the grading coefficients are small enough, further reducing them by reducing the grading ratios has little effect on the condition of the eigenvalues and eigenvectors.