A new approximate matrix factorization for implicit time integration in air pollution modeling (Q1405200): Difference between revisions

In air pollution modeling the semidiscrete system of ordinary differential equations \(y'= Vy+ r(y)\), \(y\in \mathbb{R}^N\) arises, where \(V\) is a matrix and \(r(y)\) is a nonlinear term. Implicit schemes applied to the above system result in the linear system \((I-\tau J)x= b\), \(x,b\in\mathbb{R}^N\), where \(J= V+ R\), \(\tau= \gamma\Delta t\), \(\gamma\) is the parameter of the Rosenbrock scheme and \(R\) is the Jacobian matrix of \(r(y)\) evaluated at a certain point. If the solution of this linear system is expansive, one can use the approximate matrix factorization (AMF) \(I-\tau J\approx(I- \tau R)(I-\tau V)\). The error of such approximation is \(\tau^2 RV\), which can be significant in the norm. The authors propose a new AMF, called AMF+, by the following approximation to \(I- \tau J: I-\tau J\approx (L_V-\tau R)U_V\), \(L_V U_V= I-\tau V\), where \(L_V\) and \(U_V\) are the LU factors of \(I-\tau V\). Under the assumption that \((-V)\) is a columnwise weakly dominant \(M\)-matrix the norm of the error to the AMF+ can be estimated for large \(\tau\) by \(\tau\|R\|\), while for small \(\tau\) the AMF+ error has the order \(O(\tau^2)\).