[Pw_forum] Problems running pwcond.x : error "problems with the linear system"

[Andrea Dal Corso]

I suggest that you first try to get the complex band structure correctly
and then the transmission. You need to converge the parameters ewind,
epsproj and nz1 in order to have correct band structures before
computing transmission. Ewind and epsproj control the number of basis
functions used in the calculation (more details in the original paper of
Choi and Ihm PRB 59, 2267 (1999)). 
A very accurate calculation is done with a large ewind (4.0 or 5.0) and
a small epsproj (10^-8, 10^-9).
The actual basis set reduction is written in the output. For instance in
the example of the alwire:

 ngper, n2d =          177          37

where ngper gives the number of G_perpendicular vectors (that is the
basis size for a fully converged calculation) and the actual number n2d
of basis functions used.
You have to find a reasonable compromise between accuracy and cpu time
decreasing ewind and increasing epsproj in such a way to reduce the
basis until the complex bands remain accurate.

Also the number of points used to make the integrals in each slab nz1
could be small in your case. The default for nz1 is 11.  

Finally sometimes the routine for solving the generalized eigenvalue
problem does not find all the complex eigenvalues and this reflects in
the impossibility to solve the linear system in the transmission
calculation.
If this is the case you can try to improve the situation adding a very
small number on the diagonal of the A and B matrices of the generalized
eigenvalue problem. This is done with the variable delgep. A value
10^-9, 10^-8 usually solves the problem without changing the
eigenvalues.

Hope this helps,

Andrea


On Sun, 2006-08-27 at 12:37 -0400, stewart at cnf.cornell.edu wrote:
>    ewind=1.d0
>    epsproj=1.d-3
>    nz1=4
-- 
Andrea Dal Corso                    Tel. 0039-040-3787428
SISSA, Via Beirut 2/4               Fax. 0039-040-3787528
34014 Trieste (Italy)               e-mail: dalcorso at sissa.it