[Scipy-tickets] [SciPy] #728: maxiter broken for Scipy.sparse.linalg gmres, in addition to no ability to track convergence

SciPy scipy-tickets@scipy....
Sun Aug 24 19:33:15 CDT 2008


#728: maxiter broken for Scipy.sparse.linalg  gmres, in addition to no ability to
track convergence
----------------------+-----------------------------------------------------
 Reporter:  jschrod3  |       Owner:  somebody
     Type:  defect    |      Status:  new     
 Priority:  normal    |   Milestone:          
Component:  Other     |     Version:          
 Severity:  normal    |    Keywords:          
----------------------+-----------------------------------------------------
 I'm using scipy version
 Out[4]: scipy.version.version
 Out[5]: '0.7.0.dev4561'

 When I use the gmres routine from scipy.sparse.linalg, I cannot control
 its maximum number of iterations through the maxiter variable, nor does
 the "callback" routine ever seem to get called, so I can't track
 convergence.

 I found this ticket,
 http://www.scipy.org/scipy/scipy/attachment/ticket/360/new_iterative_gmres.py
 which I was able to adapt to the current gmres (see attached modified
 iterative.py) so that I can get access to residuals in gmres and track
 convergence.  However since maxiter is broken, gmres will sometimes
 iterate an insane number of times before returning and my residual history
 can be about 10,000 values long for a 744x744 matrix.  This also takes a
 long time, so that doing automated numerical tests is not at all
 efficient.  (note the 10,000 gmres iterations are what my research aims to
 fix.  I'm well aware that a well-preconditioned gmres call will terminate
 much sooner.)

 Could gmres be extended so that it
 (1) uses maxiter
 (2) executes "callback(resid)" every full gmres iteration so that I can
 track convergence?

 These functionalities are critical if I'm to use gmres at all.  I would
 imagine that others are in my boat.

 Thanks

-- 
Ticket URL: <http://scipy.org/scipy/scipy/ticket/728>
SciPy <http://www.scipy.org/>
SciPy is open-source software for mathematics, science, and engineering.


More information about the Scipy-tickets mailing list