Eigenvalue sorting in GEVP wrong below t0
Closed this issue · 19 comments
for t < t0 we have to reverse the sorting order, I think this is currently not done.
This might explain why principal correllator plots look as they do and also, why the pc model is not working as expected.
It does not always work. Below is a plot for D15.48 with t0 = 8 where you see the intended effect with thermal state removal but not without:
There is, however no consistent behavior depending on the thermal state treatment.
how did this look like before the patch?
how do you sort?
Here is a before picture:
For sorting I use the default parameters:
sort.type = "vectors", sort.t0 = TRUE
Any chance that you could produce these plots with slightly smaller symbols, vertical facets and more stringent lower y plot limit? It's really quite hard to see what's going on.
vertical facets
and those should of course be tall enough to see everything...
thats really strange, are you sure you used the new code everywhere?
I will rerun the complete analysis over night to completely rule out the possibility of inconsistent data but I checked the creation dates ... I can't think of anything in the rho analysis that would erroneously cause such a behavior.
I will rerun the complete analysis over night to completely rule out the possibility of inconsistent data but I checked the creation dates ... I can't think of anything in the rho analysis that would erroneously cause such a behavior.
Would it be a good idea to test this on an MWE instead?
Would it be a good idea to test this on an MWE instead?
What's a MWE?
Minimal Working Example
Could you set sort.t0 = FALSE? From the discussion @ #157 and a quick reading of the logic, it seems clear that when sort.t0 == TRUE, sorting is done using the eigenvectors at t0 + 2 only. When sort.t0 == FALSE, sorting is attempted for each point.
That's even explained in ?gevp ...
yes, of course. I know what it does
sorry, I was not saying that you didn't, I was just stating that it was indeed documented
but don't remember why it was
t0 + 2...
my hunch would be that some of the sorting methods would fail when done at t0 directly
The new version seems to work on first glance. Please find below the plot from before, with fix156 merged into the master.
From briefly going through the plots it looks like the eigenvalue are consistent before and after t0 now.
excellent!