ScaLAPACK Archives

[Scalapack] complex pivoted qr factorisation

Dear friends,
the deep spike or more erratic behavior of the LINPACK QRCP is likely
to happen when the scaled condition of A (cond(A with equilibrated columns))
exceeds 1/sqrt(machine precision). In other words, if scaled condition 
number
is taken as a realistic sensitivity measure (and for the QRF it is), 
then LINPACK-style QRCP causes
problems even if we could get half of the digits in the answer.
Zvonimir and myself did not go into fixing this in the ScaLAPACK, we 
just assumed
that ScaLAPACK developers will do that.  (How far this bag can go, see 
an example
in the attachment.  That report contains some discussion related to our 
ACM TOMS paper
on this.)
I should point out here that the fix we provided for LAPACK is not the best
we can -- a truly reliable implementation requires extra N scalars in 
the workspace,
and that we could not afford because of backward compatibility. Best fix
should get different name, because of this extra N locations in the 
workspace.
Please let me know if you are interested in having this in (Sca)LAPACK.

Best regards
Zlatko


On 20.10.2011. 15:26, James Demmel wrote:
On 10/20/11 6:01 AM, Sven Hammarling wrote:
Dear Jim,

Unless I misunderstand (quite likely!), I think that Zlatko and 
Zvonimir showed a similar example in Figure 1 of LAWN 176 - it 
certainly has a deep spike.

Once ScaLAPACK gets the LAPACK fix it will be interesting to see the 
result on David's example.

Best wishes,

Sven.
In that figure, the "true" answer is around macheps * norm(A), and 
R(i,i) drops
down way below that. So everything is tiny. I think this is the only 
way you can
get such completely out-of-order R(i,i), when everything is 
O(macheps*norm(A))
and so possibly totally losing all leading digits.
In the example from David Wang below, R(3,3) is O(norm(A)) = O(1), and 
appears after
R(2,2) =  O(macheps), which can't be explained by the bug Zlatko and 
Zvonimir found:

R(0,0) -0.707107 + 0i
R(1,1) -0.707107 + 0i
R(2,2) 8.16273e-17 + 0i
R(3,3) -0.243173 + 0i
R(4,4) 4.82605e-17 + 0i

Jim




-----
No virus found in this message.
Checked by AVG - www.avg.com
Version: 2012.0.1831 / Virus Database: 2092/4563 - Release Date: 10/20/11

-------------- next part --------------
A non-text attachment was scrubbed...
Name: sb.pdf
Type: application/pdf
Size: 1109485 bytes
Desc: not available
Url : 
http://lists.eecs.utk.edu/mailman/private/scalapack/attachments/20111020/50199983/sb-0001.pdf
 

<Prev in Thread] Current Thread [Next in Thread>


For additional information you may use the LAPACK/ScaLAPACK Forum.
Or one of the mailing lists, or