In the first part of this article, we looked at the one-dimensional linear second-order wave equation
We saw that for a discretization with grid size and timestep , a central difference approximation is convergent if the CFL condition
is satisfied. The numerical method is also stable in this case because the PDE is linear.
Of course, in the above equation we have chosen the units of space and time such that the wave speed is 1. In general, it is more convenient to work with quantities for which the wave speed is not 1, and the 1D wave equation has the form
where, for linear elasticity, is the bulk modulus and is the mass-density. In that case the CFL condition will be scaled accordingly by the wave speed.
Note that the above equation is equivalent to the following first-order wave equations:
Linear PDEs with constant coefficients of this form have plane wave solutions:
where is the (real) wave number and is the (complex) frequency. The relation between the wave number and the frequency, , is called a dispersion relation.
The von Neumann approach to stability analysis
A numerical algorithm for solving a PDE is said to be stable if the solution at a given time does not blow up as the time step decreases, i.e., approximation errors do not increase with time.
The von Neumann technique, proposed by John von Neumann in the 1950s expanding upon earlier work by Crank and Nicolson, is the most popular approach for finding stability conditions for linear PDEs with periodic boundary conditions. No general approach is for stability analysis is yet known for nonlinear PDEs. However, for many practical nonlinear PDEs, the linearized form is amenable to the same type of analysis.
The von Neumann approach uses the observation that linear PDEs admit plane wave solutions to find a bound on numerical errors produced by explicit one-step finite differences. Let us examine the technique in a bit more detail for the first-order wave equation with forward differences. We have
We can write the above as
Since the PDE admits plane wave solutions, let us examine what happens when we insert solutions of that form into the discretized equation, i.e.,
We then get
Therefore the expected numerical solution at is of the form
However, due to cumulative numerical rounding errors, we get a sequence of solutions
and the actual numerical solution that we get is
If we assume that a Taylor series expansion of the function is possible, we can make the first-order approximation
The quantity is called the amplification factor. For the first-order wave equation with forward differences, from the plane wave solution we have
Clearly, for the solution to be stable, the amplification factor must satisfy the requirement that
which implies that errors do not grow in an unbounded manner with time. For the first-order wave equation,
Since is arbitrary, the above condition is satisfied only if
You can easily show that this is equivalent to the CFL condition for this particular backward propagating one-dimensional wave propagation problem.
For the forward propagating second-order wave equation, the equivalent stability condition is, (from both CFL analysis and the von Neumann analysis)
The CFL procedure is inadequate for closed form stability bounds for many PDEs even though convergence results can be obtained. The von Neumann approach solves some of these issues for linear PDEs.
We will see how these ideas are applied to discrete element methods in the next part of this series.
If you have questions/comments/corrections, please contact banerjee at parresianz dot com dot zen (without the dot zen).
📅 14.04.2017 📁 DEM