Exponential multistep methods of Adams-type (Q657891)

This paper is concerned with a general analysis of the error behaviour of the so-called exponential methods of Adams-type when applied to stiff initial value problems (IVPs) of type \[ u'(t)= - A u(t) + g(t, u(t) ), \quad t \in [0,T], \quad u(0)= u_0\tag{1} \] where the stiffness arises from the linear term \( - A u \) whereas \( g(t,u)\) satisfies a Lipschitz condition with respect to \(u\) with a moderately sized Lipschitz constant. From the variation of constants formula the exact solution of (1) at \( t_{n+1} = t_n + h\) satisfies \[ u(t_{n+1}) = \exp [ - h A] \; u(t_n) + \int_0^h \exp [ - (h- \tau) A] \; g(t_n+ \tau, u(t_n+ \tau)) d \tau . \] In the exponential methods of Adams type the function \( g( t_n+ \tau, u(t_n+ \tau))\) is substituted by its interpolation polynomial at the back points \( (t_{n-j}, u_{n-j})\), \(j=1, \dots ,k\), so that we arrive at explicit methods of type \[ u_{n+1} = \exp [ - h A] \; u_n + h \sum_{j=1}^k b_j (hA) g_{n-j} \] where the coefficients in the right hand side are matrix valued functions of \(( - h A)\). In this context an error analysis of these fixed step-size multistep methods is carried out in the general framework of linear semigroups in a Banach space \(V\) proving that under some assumptions error bounds of type \( \| u_n - u(t_n) \|_V \leq C \; h^k\), for all \(n\) with \( 0 \leq n h \leq T\), hold with a constant independent of the stiffness. This theory is also extended to non-linear IVPs \[ u' = F(t,u), \quad t \in [0,T], \quad u(0) = u_0 \] that admit a continuous linearization along the numerical solution so that at each \((t_n, u_n)\), the vector field can be written as \( F(t, u(t)) = J_n u(t)+ d_n t + g_n(t, u(t))\) with suitable assumptions on the above terms. The paper includes a number of remarks on the practical implementation of these methods and numerical experiments that confirm the order of convergence with test problems that arise from the spatial discretization of some parabolic problems. Finally, an application is proposed to partitioned problems where the stiff and non stiff components can be appropriately separated.