DSYEVD fails to converge

[duncanpo]

Hi,

We have run into a case where DSYEVD fails to converge for a 202x202 matrix (returns INFO=15529).
DSYEV converges for the same matrix successfully.

I attempted to upload the repro program and test matrix but failed because there is a 256KB limit on size and my zip file is 410KB. Please let me know how I can get the files to you.

Thanks.

Duncan Po
MathWorks

[Julien Langou]

Hi Duncan,

Thanks for providing us with a matrix that beaks DYSEVD! (Do I really mean “thanks” here? .)

I did not do anything fancy, just compiled and ran, and yep INFO = 15529 on my end as well.
Either with reference netlib LAPACK (trunk version) or vecLib. I guess we’ll have a look at this soon. Thanks.

lapackers: The matrix in the lapack svn repository under “./lapack/bugs/bug115/“

Cheers,
Julien.

[osni]

The bug has been traced to dlaed6, which computes the root closest to the origin of a secular equation and is used in the D&C tridiagonal eigensolver (dstedc) and D&C least squares solver (dgelsd). The bug was due to the use of a too stringent tolerance in the test for convergence in line 390, which has been modified to

390 IF( ( ABS( F ).LE.FOUR*EPS*ERRETM ) .OR.
391 $ ( (UBD-LBD).LE.FOUR*EPS*ABS(TAU) ) )
392 $ GO TO 60

Osni