Blockwise perturbation theory for nearly uncoupled Markov chains and its application (Q5932210)

The goal of this note is to analyze the sensitivity of each aggregate distribution \(\pi^T_i\) to small relative blockwise perturbations in the transition matrix \(P\). Under the assumption that \(\|F_{ij}\|\leq \eta\|P_{ij}\|\), \(i,j=1,2,\dots, k\), the author proves that \[ {\|\pi^T_i-\widehat\pi^T_i\|\over \|\pi^T_i\|}\leq 2k\overline f(\varepsilon, \eta)\eta+ O(\eta^2),\qquad i= 1,2,\dots, k,\tag{1} \] where \(\overline f(\varepsilon, \eta)\) is usually of moderate size. The error bound shows that small relative blockwise perturbations in \(P\) induce small relative errors in each aggregate distribution \(\pi^T_i\). The note contains six sections: In Section 2 the author introduces some notations and lemmas, especially a special decomposition of nonnegative matrices. In Section 3 he uses this decomposition to define the quantities involved in \(\overline f(\varepsilon,\eta)\), and analyzes these quantities through the spectral analysis of \(P_{ii}\). In Section 4 the structure of each block of the inverse of the matrix \(I-P_i\) is considered, where \(P_i\) is the principal submatrix of \(P\) with the \(i\)th row and column of blocks removed. This structure will be presented in Section 5 to get the error bound (1). In Section 6 the application in iterative aggregation/disaggregation methods is analyzed.