rtol + initial guess nonzero

[Barry Smith]

    Yes, in fact it can be pretty complicated. The default convergence  
test is KSPDefaultConverged() in src/ksp/ksp/interface/iterativ.c

    In the manual page for KSPDefaultConverged() you'll find the  
cryptic When initial guess is non-zero you
           can call KSPDefaultConvergedSetUIRNorm() to use the norm of  
(b - A*(initial guess))
           as the starting point for relative norm convergence testing.

    In the manual page for that function you'll see the option - 
ksp_converged_use_initial_residual_norm you'll also see mention of the  
manual
page for KSPDefaultConvergedSetUMIRNorm() and that manual page  
mentions the option -ksp_converged_use_min_initial_residual_norm

    Thus there are basically three choices to use for the initial  
residual norm.

    Barry


On Feb 19, 2009, at 2:58 PM, tribur at vision.ee.ethz.ch wrote:

> Dear petsc-people,
>
> can it be possible that the initial residuum norm used in rtol is  
> always calculated by inserting zero, even if the initial guess is  
> nonzero (using KSPSetInitialGuessNonzero)?
>