Example BAND cnbdr2.f

[wo80]

This is not an issue, I'm just looking for confirmation.

The BAND/cnbdr2.f example exits with

Error with _naupd info = -9 Check the documentation of _naupd

This means the initial residual was 0, which seems to make sense, since the example uses mode 3,

arpack-ng/EXAMPLES/BAND/cnbdr2.f

Line 172 in 569a385

mode = 3

which accesses the M matrix

arpack-ng/EXAMPLES/BAND/cnband.f

Lines 492 to 503 in 569a385

	else if ( mode .eq. 3 ) then
	c
	c %-----------------------------------------%
	c | Perform y <-- OP*x |
	c | = inv[A-SIGMA*M]*M* x
	c | to force the starting vector into the |
	c | range of OP. |
	c %-----------------------------------------%
	c
	call cgbmv('Notranspose', n, n, kl, ku, one, mb(itop,1),
	& lda, workd(ipntr(1)), 1, zero,
	& workd(ipntr(2)), 1)

but that is never set (just filled with zeros).

I'm running the examples as tests, so I just need someone to confirm that this is expected behavior. All the other examples run successfully (EDIT: well, znbdr2.f also fails).

[wo80]

So, since this example seems to solve a standard eigenvalue problem, setting M to the identity matrix should be enough (cnbdr2.f lines 199-209):

c     %---------------%
c     | Main diagonal |
c     %---------------%
c
      h = one / cmplx(nxi+1)
      h2 = h*h
c
      idiag = kl+ku+1
      do 30 j = 1, n
         a(idiag,j) = (4.0E+0, 0.0E+0)  / h2
         m(idiag,j) = (1.0E+0, 0.0E+0)
  30  continue

Same for znbdr2.f. Error message gone...

[fghoussen]

On which OS do you see the info = -9 error? Would make sense to PR this fix.

[wo80]

I checked on Debian 12 (with FlexiBLAS and OpenBLAS) and on Windows (msys2 with OpenBLAS).

[wo80]

I'd be glad to do a PR with the fix, but can someone else here confirm the issue? Would like to know if this is expected or if there's something wrong with my setup....

[wo80]

@fghoussen If the examples run correctly for you, could you please post their output, so I can compare with my results (with above fix applied).

cnbdr2

_NBDR2
======

 The size of the matrix is 100
 Number of eigenvalue requested is 4
 The number of Arnoldi vectors generated (NCV) is 10
 The number of converged Ritz values is 4
 What portion of the spectrum LM
 The number of Implicit Arnoldi update iterations taken is 9
 The number of OP*x is 53
 The convergence tolerance is 5.960464e-08


 Ritz values (Real,Imag) and relative residuals
 ----------------------------------------------
                Col    0        Col    1        Col    2
  Row    0:      2.51803e+02    1.52711e+02    6.55662e-07
  Row    1:      2.51803e+02   -1.52711e+02    1.07034e-06
  Row    2:      2.80417e+02    1.52711e+02    9.81432e-07
  Row    3:      2.80417e+02   -1.52711e+02    1.23237e-06

znbdr2

_NBDR2
======

 The size of the matrix is 100
 Number of eigenvalue requested is 4
 The number of Arnoldi vectors generated (NCV) is 10
 The number of converged Ritz values is 4
 What portion of the spectrum LM
 The number of Implicit Arnoldi update iterations taken is 19
 The number of OP*x is 105
 The convergence tolerance is 1.110223e-16


 Ritz values (Real,Imag) and relative residuals
 ----------------------------------------------
                Col    0        Col    1        Col    2
  Row    0:      2.51803e+02    1.52711e+02    2.35566e-15
  Row    1:      2.51803e+02   -1.52711e+02    2.80388e-15
  Row    2:      2.80417e+02    1.52711e+02    2.13067e-15
  Row    3:      2.80417e+02   -1.52711e+02    4.04126e-15

[fghoussen]

PR the fix. @sylvestre may add the cmake/autotools part on top of it before merge.

[fghoussen]

call exit(1) (do not use stop which returns 0) is missing here

arpack-ng/EXAMPLES/BAND/cnbdr2.f

Line 315 in 569a385

print *, ' '

This is likely true for all examples.

[wo80]

This is likely true for all examples.

I guess. So lets not clutter the PR with +100 call exit(1) additions, right?

[fghoussen]

Ideally one PR for each. Rebase the 2d (correction) on the 1st (exit call that will break CI)

[wo80]

I'll leave the PR as is. The exit code is probably a subject for a separate discussion and changes can be done after the merge. If you'd want to do it right, there should also be some check for the number of converged eigenvalues, like

if ( nconv .lt. nev ) then
      call exit(1)
end if