On Wed, Sep 2, 2009 at 2:38 PM, Vijay S. Mahadevan<[email protected]> wrote:
> On first look, I would say that the neumann_value is supposed to be
> -0.25 and not -0.5 at x=2. Try this change and see if it resolves the
> bug. Also like John suggested, see if your L2 and H1 errors give right
> convergence orders.

I don't see anything obviously wrong...

If the original problem is: d/dx (x * du/dx) = 2/x^2

The weak form is:

-(x u',v') + x u'(2) v(2) -  x u'(1) v(1) = (2/x^2,v)

Bringing the right endpoint bc over to the rhs and assuming the
Dirichlet bc is handled,

-(x u',v')   = (2/x^2,v) - x u'(2) v(2)

And replacing with the Neumann condition, (-x * du/dx) = 0.5   @ x = 2, we get

-(x u',v')   = (2/x^2,v) + 0.5 v(2)

Multiplying thru by neg. 1 as in the code, we get

(x u',v')   = -(2/x^2,v) - 0.5 v(2)

If the theoretical convergence rates are off, I would probably start
by looking at the endpoint integral, though there may be something
else obvious I'm missing.

-- 
John

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day 
trial. Simplify your report design, integration and deployment - and focus on 
what you do best, core application coding. Discover what's new with 
Crystal Reports now.  http://p.sf.net/sfu/bobj-july
_______________________________________________
Libmesh-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/libmesh-users

Reply via email to