Upwind finite difference: Matrix Implementation

Question

I want to implement the upwind finite difference scheme for the 1D linear advection equation using a finite difference matrix in python:

$$ A =\begin{pmatrix} 1-a\cfrac{\Delta t}{\Delta x} & 0 & 0 & \cdots & a\cfrac{\Delta t}{\Delta x}\\ a\cfrac{\Delta t}{\Delta x} & 1-a\cfrac{\Delta t}{\Delta x} & 0 & \cdots & 0 \\ \vdots & \vdots& \vdots & \ddots & \vdots \\ 0 & 0 & 0 & \cdots & 1-a\cfrac{\Delta t}{\Delta x} \end{pmatrix}$$

with $1 -a\cfrac{\Delta t}{\Delta x}$ on the diagonal and $a\cfrac{\Delta t}{\Delta x}$ on the lower subdiagonal and periodic boundary conditions as indicated by the top right entry of the matrix ($a$ being constant). To generate the velocities at the next time step $n$ I used the following equation:

$$\ U_{n+1} = AU_n $$

My idea is that given inital conditions $ U_1 $, I could apply $ A^{n+1} $ to $U_1$ to get the $U_{n+1}$:

$$\ U_{n+1} = A^{n+1} U_1 $$

I am not entirely sure if this is a valid approach as I am complete novice to the subject. Any help and hints would be greatly appreciated.

Answer 1

Your idea of taking matrix powers is not a good one.

You should code it for step by step updating $$ u_0 = \textrm{initial condition} $$ $$ u_0 \rightarrow u_1 \rightarrow u_2 \rightarrow $$ so at any time you at most store two levels of solution $u_n, u_{n+1}$.

Even when updating $u_n \rightarrow u_{n+1}$ I would avoid using a matrix since you are multiplying with a lot of zeroes, unless you use a sparse matrix, which is possible in Python.