Setting
I am solving for $u(x,y,t)$ the wave equation $\partial_{tt} u - \partial_{xx} u = f$ on $(x,y)\in\Omega=[0,1]\times \mathbb{R}$ by splitting it into an equivalent first order system: $$\partial_t \begin{pmatrix} u\\ \phi \end{pmatrix}+ \begin{pmatrix}0&b\cdot\nabla\\ b\cdot\nabla&0 \end{pmatrix}\begin{pmatrix} u\\ \phi \end{pmatrix}=0 ,\qquad b=\begin{pmatrix} 1\\ 0\end{pmatrix}$$
and applying a Crank-Nicolson method to this system (which is equivalent to the Newmark scheme for the original wave equation). I use FEM (fenicsx software library) with polynomials of various degrees on a fully-structured triangle-element mesh for the space discretization.
Expected convergence rates (from FEM theory):
The error $u-u_h$ should converge of rate 2 in the $L^2$ norm, when using polynomials of order 1 and the Crank-Nicolson scheme, i.e. $$\epsilon = \Vert u-u_h\Vert_{L^2(\Omega)} \leq C( h^2 + \Delta t^2)$$
When i instead use polynomials of order 2 i expect to get $$\epsilon = \Vert u-u_h\Vert_{L^2(\Omega)} \leq C( h^3 + \Delta t^2)$$
Now i would like to measure these rates in a numerical experiment. In particular, I would like to measure $h^3$, be refining $\Delta t$ faster (so that I dont have to switch my time-discretization scheme).
Computing the convergence rate $r$ using the numerical solution
Suppose we have a stationary problem, then (C independent of $h$ or $i$) $$\epsilon_i = Ch_i^r \implies \left(\frac{h_{i+1}}{h_i}\right)^r=\frac{\epsilon_{i+1}}{\epsilon_i} \implies r=\frac{ \log (\epsilon_{i+1}/\epsilon_{i})}{\log(h_{i+1}/h_i)}$$
However my wave problem has 2 discretization parameter: $h$ and $\Delta t$. So i cant find a formula for the rates $r_h$ or $r_t$ using the above approach. But i can refine $\Delta t$ so fast (that it "vanishes" asymptotically), that I actually observe $r_h=3$ and then the above formula for $r$ is correct.
Lets fix $h_{i+1}=\frac{1}{2}h_i \implies \epsilon_{i+1}=\frac{1}{8}\epsilon_i$.
Does this mean, the refinement factor for the time scheme to reveal this fast FEM convergence is $$\frac{\Delta t_{i+1}}{\Delta t_i}=\frac {1}{\sqrt 8}$$
so that we end up with $$\frac{\epsilon_{i+1}}{\epsilon_i} \leq \frac{C}{C}\frac{h_{i+1}^3 + \Delta t_{i+1}^2}{h_i^3 + \Delta t_i^2} =\frac{(h_i\frac{1}{2})^3 + (\Delta t_i \frac{1}{\sqrt 8})^2}{h_i^3 + \Delta t_i^2} = \frac{1}{8}\frac{h_i^3 + \Delta t_i^2}{h_i^3 + \Delta t_i^2}=\frac{1}{8}?$$
Is this derivation correct?
A numerical experiment reveals, that I can not observe the fast rate $r_h=3$ with the $\sqrt 8$ time-refinement.
I would like to understand if
- either my theory is fine and my code is wrong
- or my theory is wrong