-----BEGIN PGP SIGNED MESSAGE-----

Hash: SHA1

While I do agree that such testing would be reassuring and consume few

resources, it is my experience that coding an FFT for the 1 million digit

range is rather trivial. That's less than 4194304 bits or 2^22 bits. Assuming

radix 2^16, that would imply a transform size of 2^18, if the routine employs

the right-angle DWT, which cuts the transform length in half by using the

imaginary part of the transform elements to store useful data. Dominique

Delande's pi calculation program does length 2^25 transforms (that's seven

additional bits required for the pyramid) with radix 10^4, that's only less

than 3 bits smaller. SB's routine should have a lot of headroom if correctly

done and using all the usual tricks -- balanced representation, fast and

accurate trig identities for computing primitive roots of C, etc.

So, barring a stupid mistake, I am not worried that such a routine would fail

at this range.

Décio

On Monday 31 March 2003 23:04, David Broadhurst wrote:

> As no doubt everyone noticed, I foolishly confused

> Proths with Riesels. But the principle stands:

> it there a way of setting a decent limit on FFT failures?

> Even the rather smaller Cosgrave Proth 3*2^2145353+1

> would be some reassurance.

> David (foolish, though trying to be helpful)

-----BEGIN PGP SIGNATURE-----

Version: GnuPG v1.2.1 (GNU/Linux)

iD8DBQE+iX4zce3VljctsGsRAmjTAJ9bnytktF1lmA1s3iQG0lzmUxc7xACgo/1g

Q+VCtNkKjBdwA8yT1mnBJzc=

=fV+w

-----END PGP SIGNATURE-----