JOS Index |
JOS Pubs |
JOS Home |
Multi-step methods can be treated in a very similar way. An explicit -step method is defined by
for constant coefficients
contained in subsets
. Taking the Fourier transform of this recursion gives
A simple way of examining (4) is to look for solutions of the form
. This gives the amplification polynomial equation
the solutions of which,
must be bounded by unity for stability (though in general, this is not sufficient, as we will show presently for a special case).
A particular form of the amplification polynomial equation which will appear frequently in our subsequent treatment of finite difference schemes for the wave equation is that of a simple two-step centered difference approximation, namely
for some real function
. This expression has solutions
which will be bounded by (and in fact equal to) unity in magnitude if we have
for all . Furthermore, if
for some , then we will necessarily have an amplification factor with magnitude greater than one at that frequency. For any for which
are not equal, we can write
are the spatial frequency spectra of the two grid functions (at time steps and ) used to initialize the two-step method. It is easy to show that the norm of
can be bounded in terms of the norms of the initial conditions if the spectral amplification factors are distinct and bounded by 1 in magnitude at all wavenumbers.
It is important to realize, however, that the condition that these
be bounded by
unity is necessary, but not sufficient to ensure no growth in the
norm of the solution; this point has not been addressed in the
finite difference treatment of waveguide meshes. In fact, as shown in
, the simple centered difference approximation
to the wave equation admits linearly growing solutions.
This behavior can be examined in the spectral domain as we will now show, as per . Notice that the solutions (6) of the amplification polynomial equation for the two-step scheme can coincide if, and only if at some frequency
, in which case we have
. The evolution of the particular spatial frequency component at frequency
can be written as
We can thus expect some linear growth at any such frequency
if we do not properly initialize the algorithm, so as to cancel the linearly growing part of the solution. It also follows that in employing such a method, one may need to be particularly careful when applying an excitation which contains such frequency components, and that nonlinear signal quantization may pump energy into such modes, even if none is originally present there.
Strikwerda does not classify such linear growth as unstable, because the wave equation itself admits, in addition to traveling wave solutions, a solution which grows linearly with time2. For the physical modelling of musical instruments and acoustic spaces, however (the problems to which finite difference schemes of the form to be discussed shortly are usually applied), such solutions are nonphysical and definitely not acceptable. These comments concerning this mild linear instability apply to schemes in unbounded domains; when boundary conditions are present, further analysis will be required.
In order to simplify the analysis of these schemes, we mention that for difference schemes for the wave equation, it is often possible to write
is independent of . In this case, the stability condition can be rewritten as
This new condition on
is easier to analyze: we first require
and if (8) holds, we get a further bound on , namely
Thus the stability of these schemes can be simply analyzed in terms of the global maximum and minimum of
For certain schemes (in particular, the interpolated schemes to be discussed in §3.2 and §4.3), the function
depends on several parameters. Condition (8) tells us the the range of parameters over which our scheme is stable, and over the stability region, condition (9) gives us a maximum time step , in terms of the grid spacing .
JOS Index |
JOS Pubs |
JOS Home |