Issue 667: dneupd.f (and similar) gives different eigenvalue orderings for rvec = .false. and rvec = .true.

Reported by Sylvestre Ledru, Feb 21, 2012

reported by Tim Mitchell
- When rvec = .false., the Ritz values are outputted according to 
their dsortc() order
- When rvec = .true., the eigenvectors also need to be rearranged so 
dneupd() calls dtrsen() from LAPACK to reorder both the Ritz values 
and their associated vectors
- Unfortunately, dtrsen() only uses a boolean select array to do the 
regrouping; it is not written to take advantage of any ordering info
- The easiest fix might be to create a modified dtrsen which can use 
an ordering (which is already computed previously in dneupd)

Comment 1 by Danie Ludick, May 14, 2014

Good day,

I also get this issue, i.e., that when rvec=.true. the Ritz values 
are not sorted - even though I am specifying 'LM' as the ordering 
parameter. Is there any updates / progress regarding the modified 
version of dtrsen() - or perhaps any other suggestion.

For reference, I have attached a file containing the Ritz values and 
relative residuals I get when solving a 4248 x 4248 double precision 

Created: 6 years 9 months ago by Sylvestre Ledru

Updated: 4 years 7 months ago

Status: New

Followed by: 1 person