Weighted sequential splittings and their analysis (Q815224): Difference between revisions

The necessity of splitting occurs frequently when an evolution problem of a partial differential equation of the following form \(u_t=Au,\) \(A=\sum_{k=1}^rA_k,\) where the operators \(A_k,\) \(k=1,2,\dots,r\) are of very different character, has to be solved numerically. Splitting means to approximate the solution \(u\) by a sequential numerical resolution of simpler problems \(u^{(k)}_t=A_ju^{(k)}\) of well determined character. The authors recall the ``classical'' splitting procedures for the equation \(w_t=(A_1+A_2+\cdots)w,\;\;w(0)=w^{(0)}\). We present three of them in form of the following diagrams \[ \begin{matrix} &1&&2\cr \tau&\uparrow&\searrow&\uparrow\cr &1&\leftarrow&2\cr \tau&\uparrow&\searrow&\uparrow&\cr &1&\leftarrow&2&\cr &&&w^{(0)}\end{matrix}\tag{1} \] \[ \begin{matrix} {\tau\over2}&\uparrow&\searrow&\uparrow&&\cdot&\cr &1&\leftarrow&2&\leftarrow&3&\cr {\tau\over2}&\cdot&&\uparrow&\searrow&\uparrow&\cr &1&&2&&3&\cr {\tau\over2}&\uparrow&\searrow&\uparrow&&\cdot&\cr &1&\leftarrow&2&&3\cr &&&w^{(0)}&&\end{matrix}\tag{2} \] \[ \begin{matrix} &\vdots&&\vdots&\cr &1&+&2&\cr \tau&\uparrow&&\uparrow&\cr &1&+&2&\cr \tau&\uparrow&&\uparrow&\cr &1&&2&\cr &w^{(0)}&&w^{(0)}&\end{matrix}\tag{3} \] The numbers 1, 2, \(\dots\) represent the splitted processes \(w^{(j)}_t=A_jw^{(j)}\), the vertical arrows show the actions of processes during a short time step \(\tau\), the horizontal and diagonal arrows represent the transfer of initial values. Let us explain the first diagram. For homogenity of any step we begin with the lower right angle of the diagram, going to the left for the transfer of the initial value \(w^{(0)}\). Then the first process starts followed by second one, and so on. In the splitting process (3) \(w\) is computed at any step as \(w=w^{(1)}+w^{(2)}\) and then \(w\) enters as initial condition for the next step for both processes. The authors give an analysis of the above mentioned processes. In particular they show that in the case of non-commutativity the local error at any step of (1) and (3) is of the order 1 (the error is proportional to \(\tau^2\) ) while the process (2) (the Strang process) is of the order 2 with respect to \(\tau\). It is worth to mention that the process (3) is easily parallelisable. The hit of this paper is the weighted sequential splitting process, \[ \begin{matrix} &&&\Theta&+&1-\Theta&\cr &1&&2&&1&\cr \tau&\uparrow&\searrow&\uparrow&\searrow&\uparrow&\cr &1&\leftarrow&2&&1&\cr &&&w^{(0)}&&&\end{matrix} \] where after each step \(w\) is computed as weighted mean with \(\Theta\in[0,1]\) of the result of the process 2 from the first half of the step and the result of the process 1 from the second half of the step (see at the diagram). The detailed analysis given by the authors show that for \(\Theta={1\over2}\) the process is of the order 2. Moreover under additional conditions the process is of the order 3 or 4. The paper contains test examples and a consistency analysis for a system of ordinary differential equations. It is proven that for \(\Theta={1\over2}\) the weighted splitting method is second order consistent if for any sub-process a second order consistent numerical method is used.