[petsc-dev] bug with TSInterpolate_Theta and TSInterpolate_Euler in master?

[Andrew Spott]

Looking at the code for TSInterpolate_Theta  (X is the output vector):



1    ierr = VecCopy(ts->vec_sol,th->X);CHKERRQ(ierr);
2    if (th->endpoint) alpha *= th->Theta;
3    ierr = VecWAXPY(X,alpha,th->Xdot,th->X);CHKERRQ(ierr);


I can’t figure out why line 1 exists.  And, since this appears to just be using a simple taylor series:  U(t + dt) ~= U(t) + Udot(t) * dt, I’m not sure why th->endpoint has any effect on this.  Is th->Xdot scaled by th->Theta at some point?


Attempting to see if something I didn’t understand was going on, I also took a look at TSInterpolate_Euler (X is still the output vector).


PetscReal alpha = (ts->ptime - t)/ts->time_step;

ierr = VecAXPBY(ts->vec_sol,1.0-alpha,alpha,X);CHKERRQ(ierr);



Which doesn’t even modify X (and does modify the solution…).  I assume that this is an attempt to find U(t-δt) ~= U(t) (1-δt/Δt) + U(t-Δt)(δt/Δt).  However, I’m not sure where U(t-Δt) is supposed to come from…


TSInterpolate_Theta is easy enough to fix, but TSInterpolate_Euler seems to need a retained solution, which isn’t supplied.  If a retained solution WERE supplied, then the TSInterpolate_Euler version could work for every time step method, and should be a default.


I’ll fix these if I’m right, but since I’m still learning the internals of Petsc, I wanted to check and make sure these are problems first.


-Andrew
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20150401/222eebac/attachment.html>