No label defined (Q4036846)

[For the entire collection see Zbl 0755.00007.] The authors introduce a class of preconsistent \(q\)-processor \(m\)-stage \(k\)-value general linear methods, which can be expressed as \(Z=C^ Tz_{n-1}+hLf(Z)\), \(z_ n=Az_{n-1}+hBf(Z)\), where \(z_ n\) is a set of \(k\) values needed for future steps and \(C=(C_ 1,C_ 2,\dots,C_ m)\). The methods are used for the numerical integration of the general problem \(y'=f(y)\), \(y(0)=y_ 0\). They include most existing sequential and parallel methods. For a serial explicit method, \(L\) is strictly lower triangular, and for a parallel explicit method \(L\) is block strictly lower triangular. The preconsistency conditions are \(As=s\), \(c^ T_{ij}s=1\), where \(s\) is a constant vector which satisfies \(z_ n=sy(t_ n)+O(h)\). Theorem 1 shows that, if the preconsistent \(q\)-processor general linear method has \(\text{rank} B=r<q\), then there exists a preconsistent \(r\)-processor general linear method having the same stability matrix \(A(\mu)\), where \(A(\mu)=A+\mu B(I-\mu L)^{-1}C^ T\). Consequently, more than one processor offer no advantage for one-step and straight multistep methods and there is no advantage in having more processors than the number of saved values. The degree coefficients \(\alpha_ j(\mu)\) of the characteristic polynomial \(\text{det}[\zeta I- A(\mu)]=\zeta^ k+\alpha_ 1(\mu)\zeta^{k-1}+\cdots+\alpha_ k(\mu)\), are limited by the number of stages and the number of processors: \(\deg\alpha_ j\leq\min\{jm,qm\}\). Higher degree coefficients provide more degrees of freedom in creating the approximation. In conclusion, parallelism may have more to offer for methods with more stages.