Page 1 of 1

Can convergence criterion be specified?

PostPosted: Wed Mar 21, 2012 2:02 am
by phoebe
In CULASparseReferenceManual, "Unless otherwise specified, all of CULA Sparse’s iteration calculations are done with regards to a residual relative to
the norm of the right hand side of the linear system."
I want to know that can users specify the convergence criterion? For example, can I change this criterion to that the maximum absolute value of the change of x is less than a predefined value and how? Many thanks!

Re: Can convergence criterion be specified?

PostPosted: Wed Mar 21, 2012 7:22 am
by kyle
CULA Sparse has built in stagnation checking where if the residual doesn't change within in a relative window, a "stagnation" result is returned. This is currently a fixed ratio based off the machine precision and various norm.

However, we will consider allowing for a user specified stagnation condition for a future release.

Would a typical usage scenario simply exit after 'n' consecutive residuals of less than 'threshold' change? Would n always be 2 or should it be specified?

Re: Can convergence criterion be specified?

PostPosted: Wed Mar 21, 2012 8:02 pm
by phoebe
we will consider allowing for a user specified stagnation condition for a future release

Thanks for the consideration.

Would a typical usage scenario simply exit after 'n' consecutive residuals of less than 'threshold' change? Would n always be 2 or should it be specified?

In fact, in our scenario there are two convergence criteria. The first is that the maximum absolute value of the differcence of x between the two successive teration is less than a predefined value epsilon1. The second is that the maximum absolute value of the residual for that iteration is less than another predefined value epsilon2. Epsilon1 and epsilon2 should be specified. It will be nice if the criteria can be added. Thanks!

Re: Can convergence criterion be specified?

PostPosted: Wed Mar 21, 2012 8:03 pm
by phoebe
we will consider allowing for a user specified stagnation condition for a future release

Thanks for the consideration.

Would a typical usage scenario simply exit after 'n' consecutive residuals of less than 'threshold' change? Would n always be 2 or should it be specified?

In fact, in our scenario there are two convergence criteria. The first is that the maximum absolute value of the differcence of x between the two successive teration is less than a predefined value epsilon1. The second is that the maximum absolute value of the residual for that iteration is less than another predefined value epsilon2. Epsilon1 and epsilon2 should be specified. It will be nice if the criteria could be added. Thanks!