You are currently browsing the tag archive for the ‘operator splitting’ tag.
The summer continues to allow some clearing of the backlog of projects accumulated during the academic year: Helge Holden, Kenneth Karlsen, Nils Risebro, and myself have uploaded to the arXiv the paper “Operator splitting for the KdV equation“, submitted to Math. Comp.. This paper is concerned with accurate numerical schemes for solving initial value problems for the Korteweg-de Vries equation
though the analysis here would be valid for a wide range of other semilinear dispersive models as well. In principle, these equations, which are completely integrable, can be solved exactly by the inverse scattering method, but fast and accurate implementations of this method are still not as satisfactory as one would like. On the other hand, the linear Korteweg-de Vries equation
can be solved exactly (with accurate and fast numerics) via the (fast) Fourier transform, while the (inviscid) Burgers equation
can also be solved exactly (and quickly) by the method of characteristics. Since the KdV equation is in some sense a combination of the equations (2) and (3), it is then reasonable to hope that some combination of the solution schemes for (2) and (3) can be used to solve (1), at least in some approximate sense.
One way to do this is by the method of operator splitting. Observe from the formal approximation (where should be thought of as small, and is some matrix or linear operator), that one has
[we do not assume A and B to commute here] and thus we formally have
if for some fixed time T (thus ). As a consequence, if one wants to solve the linear ODE
for time , one can achieve an approximate solution (accurate to order ) by alternating times between evolving the ODE
for time , and evolving the ODE
for time , starting with the initial data .
It turns out that this scheme can be formalised, and furthermore generalised to nonlinear settings such as those for the KdV equation (1). More precisely, we show that if for some , then one can solve (1) to accuracy in norm for any fixed time by alternating between evolving (2) and (3) for times (this scheme is known as Godunov splitting).
Actually, one can obtain faster convergence by modifying the scheme, at the cost of requiring higher regularity on the data; the situation is similar to that of numerical integration (or quadrature), in which the midpoint rule or Simpson’s rule provide more accuracy than the Riemann integral if the integrand is smooth. For instance, one has the variant
of (5), which can be seen by expansion to second order in (or by playing around with the Baker-Campbell-Hausdorff formula). For KdV, we can rigorously show that the analogous scheme (known as Strang splitting) involving the indicated combination of evolutions of (2) and (3) will also converge to accuracy in norm, provided that and .