[petsc-dev] strangness in Chebyshev estimate of eigenvalues

[Mark Adams]

>
>
> I agree that we should be using a random starting vector for spectral
> estimates.  I'm not overly concerned whether it's a hash of the global
> ID or a deterministic seed, but let's use something that has been
> studied.


I really think that it does not matter.  Just as long as it is noisy.  IMO,
but we have no data, and I'm not sure how we could study it.  What is the
chance that your generator will generate a low frequency vector on some god
knows what grid?  I can not imaging this happening ... and I'm not sure how
I could even generate data to 'prove' it.


> Your suggested "hash function" is way hokey.  I'm fine with
> guaranteeing drand48 and a deterministic seed.
>

I like the idea of being deterministic in parallel, which we can do with an
ID hash, but if you prefer drand48, I can do that.

Mark
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20150826/11f90ff5/attachment.html>