Some schemes, such as those with upwind FDs, are intrinsically dissipative, with a fixed “amount” of dissipation for a given resolution. Another approach is to add to the discretization a dissipative operator with a tunable strength factor ,

The operator has derivatives of higher order than the ones in the principal part of the equation, mimicking dissipative physical systems and/or parabolic equations but in such a way that as . Furthermore, is usually chosen so that scales with the gridspacing as the highest FD approximation (so that the amplification factor depends only on ). For example, for first-order-in-space systems, FDs scale as and is usually chosen to scale in the same way. More precisely, in the absence of boundaries the standard way to add numerical dissipation to a first-order-in-space system is through Kreiss–Oliger dissipation where denotes the application of times and denote forward and backward one-sided FDs, respectively, as defined in Eq. (7.70). Thus, scales with the gridspacing as , like Eq. (8.52). If the accuracy order of the scheme is not higher than in the absence of dissipation, it is not decreased by the addition of numerical dissipation of the form (8.53).The main property that sometimes allows numerical dissipation to stabilize otherwise unstable schemes is when they strictly carry away energy (as in the energy definitions involved in well-posedness or numerical-stability analysis) from the system. For example, the operators (8.53) are semi-negative definite

with respect to the trivial scalar product , under which centered differences satisfy SBP.In the presence of boundaries, it is standard to simply set the operators (8.53) to zero near them. The result is, in general, not semi-negative definite as in (8.54), which cannot only not help resolve instabilities but also trigger them. Many times this is not the case in practice if the boundary is an outer one, where the solution is weak, but not for inter-domain boundaries (see Section 10). For example, for a discretization of the standard wave equation on a multi-domain, curvilinear grid setting, using the SBP operator with Kreiss–Oliger dissipation set to zero near interpatch boundaries does not lead to stability while the more elaborate construction below does [141].

For SBP-based schemes, adding artificial dissipation may lead to an unstable scheme unless the dissipation operator is semi-negative under the SBP scalar product. In addition, the dissipation operator should ideally be non-vanishing all the way up to the boundary and preserve the accuracy of the scheme everywhere (which is more difficult in the SBP case, as it is non-uniform). In [303], a prescription for operators satisfying both conditions for arbitrary–high-order SBP scalar products, is presented. A compatible dissipation operator is constructed as

where is the SBP scalar product, is a consistent approximation of with minimal bandwidth (other choices are presumably possible), and is called the boundary operator. The latter has to be positive semi-definite and its role is to allow boundary points to be treated differently from interior points. cannot be chosen freely, but has to follow certain restrictions (which become somewhat involved in the non-diagonal SBP case) based on preserving the accuracy of the schemes near and at boundaries; see [303] for more details.
Living Rev. Relativity 15, (2012), 9
http://www.livingreviews.org/lrr-2012-9 |
This work is licensed under a Creative Commons License. E-mail us: |