I'm working through the paper where the Finite Difference method is employed to solve the PDE

$\displaystyle \frac{\partial u(x,t)}{\partial t} = a \cdot \frac{\partial^2 u(x,t)}{\partial x^2} + b \cdot \frac{\partial u(x,t)} {\partial x} + c \cdot u(t,x) $

on $\mathbb{R}^{+} \times [0, T]$, but corresponding boundary conditions aren't known. Introducing the $x$-mesh on $[0, B]$, $x_i = i \cdot h$, where $h = \dfrac{B}{m}$, $i \in \{0, ..., m\}, m \in \mathbb{N}$, derivatives at points $x = x_0$, $x = x_m$ are thus approximated using forward and backward differences for each time level, respectively (e.g. $\dfrac{\partial u}{\partial x_o} \approx \dfrac{u_1 - u_0}{h}$ and so on). The resulting system matrix is thus of size $(m+1) \times (m+1)$.

My question is which stability/convergence results do we get here compared to the case with boundary conditions imposed? When implementing the method, I encountered an issue that values of approximated solution with $x$ close to $B$ vary heavily for different $B$'s. For example, taking $B= 40.4$ and $B = 22000$ I get:

approximation for different B's

I wonder whether this an implementation issue or consequence of no boundary.


Usually, when no boundary conditions are provided, this kind of problem has not a unique solution. There are some cases where a "no boundary condition" is used, usually for only one part of the boundary of the domain where an outflow condition is required, but these cases are rare and this technique still no well understood AFAIK.

Did you check whether the matrix of your linear system is singular or not?

If it is singular (or nearly), this could explain the high sensibility that you see.

Another possibility is that you impose something without really noticing it. You should show the exact equations that you use for the discretization, in particular how you treat the case close to the boundary of the domain.

EDIT: After your precision, I can see two points that can make your solution depend on $B$:

  • I'm still not sure whether the continuous problem has a unique solution without boundary conditions. I know some cases (e.g. Poisson problem on a square with one "open" side) where the continuous problem without boundary conditions is ill-posed while the discrete problem yields a non-singular linear system to solve (in a FEM context). So, you can get a discrete solution, but you do not know which solution is approximated. Different $B$ can then yield different solutions.
  • "No boundary condition" means "I don't know at all what is after this point", however, when you increase your domain size, you actually give more information to your solution, since you provide an initial value for the solution after the point $x=44$, and this information will change the solution in the domain $(0,44)$ due to the diffusion.

The fact that the solution differs more on the right side than on the left side seems to indicate that the second point is the reason that you were looking for. To confirm this, try to decrease the diffusivity $a$ or increase the transport ($b>0$) and see if the two solutions get closer.

  • $\begingroup$ Thanks for the answer. I (actually authors) check for regularity of the matrix (which is conditional on some relations of $a$, $b$, $c$, $h$, $\theta$). As for equations, the $\theta$-method is used. I.e., $\dfrac{\partial u}{\partial t}$ is approximated via forward differences; for inner points: $\dfrac{\partial u}{\partial x}$ via central differences, $\dfrac{\partial^2 u}{\partial x^2}$ via second (central) differences. Derivatives for points on the boundary: at $x_0$ via first and second forward differences, at $x_m$ via first and second backward differences. $\endgroup$ – Dello Jun 28 '13 at 12:35
  • $\begingroup$ @Victor Thanks for the precision. Answer updated. $\endgroup$ – Dr_Sam Jun 28 '13 at 13:24
  • $\begingroup$ Thanks again. With the decreased diffusion solutions indeed get closer to each other. In the related literature I have found the following statement: "The numerical analysis revealed that there is only small influence of the boundary condition chosen, except very close to the boundary. Therefore, the boundaries should lie sufficiently far away from the region of interest". However, it isn't clear whether this is the case for some particular PDEs or in general. $\endgroup$ – Dello Jun 28 '13 at 13:59
  • $\begingroup$ @Victor In my opinion, this is related to the PDE. If it is dominated by the diffusion effects, the information from the boundary will spread quickly in the whole domain and affect the whole solution. For transport dominated cases, the effect of the boundary condition might indeed be limited to a small region close to outflow boundaries. $\endgroup$ – Dr_Sam Jun 29 '13 at 10:28
  • $\begingroup$ Then I don't get how this is reflected in stability/ convergence results. For example, if we take the heat equation, and use forward/backward differences on the boundary instead of boundary conditions, will it influence stability/convergence results? Or do we get the different solution in this case and FDM scheme converges to this new solution? $\endgroup$ – Dello Jun 29 '13 at 15:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.