The Sylvester equation and approximate balanced reduction (Q1611928): Difference between revisions

Let \(\dot x(t)=Ax(t)+ Bu(t)\), \(y(t)=Cx(t) +Dy(t)\) be a linear time-invariant system, where \(A,B,C\), and \(D\) are real matrices of appropriate sizes. Assuming the system is reachable, observable, and stable (i.e., \(A\) has all its eigenvalues in the open left half-plane), the linear matrix equation \[ \left[ \begin{matrix} A & 0\\ 0 & A^T\end{matrix} \right] \left[ \begin{matrix} P & X\\ X^T & Q \end{matrix}\right] +\left[ \begin{matrix} P & X\\ X^T & Q\end{matrix} \right] \left[ \begin{matrix} A & 0\\ 0&A^T \end{matrix} \right] +\left[ \begin{matrix} B\\ C^T\end{matrix} \right] [B^TC]=0 \] has a unique solution, where \(P\), the reachability gramian, and \(Q\), the observability gramian, are positive definite. The matrix \(X\) is called the cross gramian. A part of the paper contains useful tutorial information on the gramians, their properties, especially as applied to balanced realizations, Sylvester equations, and connections with Löwner matrices and rational interpolation. Error bounds in the \(H^2\) norm are derived for truncated systems in terms of the gramians. An implicitly restarting algorithm is described to achieve approximate balancing through low rank approximation of the cross gramian. The main results are formulated for SISO systems and for symmetric MIMO systems, i.e., such that all Markov parameters \(CA^kB\), \(k=0,1,\dots\), are symmetric matrices. To deal with non-symmetric MIMO systems, an embedding into a symmetric MIMO system is proposed. Experimental results implemented via Matlab codes and described in the paper suggest that both singular values and eigenvalues of the cross gramian typically decay very rapidly, thus making it possible to approximate well the cross gramian by matrices of very low rank.