Helmholtz equation Neumann and divergence

[Heimdall]

"Helmholtz equation" Neumann and divergence

Hello, I'm trying to solve the following elliptic problem :

S = B - \mu\nabla^2 B

Where S(x,y) and B(x,y) are 3 component vectors.

I have \nabla\cdot S = 0 and I want B such that \nabla\cdot B = 0 everywhere.

I'm using finite differences on a grid with nx+1 in the x direction and ny+1 points in the y direction. The x=0 and x=nx boundaries are periodic, so we have :

Bx(0,y) = Bx(nx,y)
By(0,y) = By(nx,y)
Bz(0,y) = Bz(nx,y)I thought that maybe the following boundary conditions would ensure \nabla\cdot B=0 : Bx(x,0) = B_1
Bx(x,ny) = B_2

\frac{\partial B_y}{\partial y}\left(x,ny\right)= \frac{\partial B_y}{\partial y}\left(x,0\right) = 0

(this make the divergence of B equal to zero on the y=0 and y=ny boundaries)

and homogenous dirichlet conditions for Bz at y=0 and y=ny.

Do you so far agree with that ?

I'm using centered second order scheme to discretize my equation (standard 5 point laplacien). And for the Neumann BC I'm doing :

By(x,-1) = By(x,1) for the y=0 border
By(x,ny+1) = By(x,ny-1) for the y=ny border.

This is supposed to be second order first derivative. Thanks to this, I can replace the "ghost" point in my Laplacian when I'm on the top or bottom border.

But I have a problem, when I look at \nabla\cdot B, it is 0 in the middle of my domain but on a small length from the y=constant borders, the divergence of B is starting to raise anormally.

example :
http://nico.aunai.free.fr/divB.png

another one (del dot B versus y-direction) :
http://nico.aunai.free.fr/divb.png

you can see that there is no problem at all on the periodic boundaries :-sWhen I'm solving the equation for an analytical source term for which I know the analytical solution, I can notice that there is a small error (but definitely bigger than everywhere else in the domain) on the Y boundary regarding to the Neumann BC.

Please, would you know where I should look at to fix this problem ?
Is my boundary conditions are bad to satisfy \nabla\cdot B=0 ?
Is my discretisation not correct ? I've checked the local truncation error which seems to be second order consistent, and eigenvalues of my linear operator looks pretty much the same that those of the Laplacian (1- L), and if I'm correct it should be stable and so converge towards the solution with second order accuracy everywhere, no ?

I can post my gauss-seidel routine if needed.

Thanks a lot !
Please tell me if something's not clear.

 

[jubrack]

When \nabla \cdot B =0 your equation can be written,

\mu\nabla\times\nabla\times B + B = S.

The correct boundary conditions for this equation are discussed in http://dx.doi.org" , reference number doi:10.1016/j.jcp.2008.03.046.

There it is shown that with \hat{n} \times B=0 on the boundary,
a compatibility condition,

\int_{\partial D} \hat{n} \cdot [ B-S] \mathrm{d} \mathcal S=0

must be satisfied,
or with

\hat{n} \times \nabla \times B=0

on the boundary, the boundary condition,

\hat{n} \cdot [B-S]=0

must be imposed.