I wrote:

> rather, one first has to use the corresponding modular
> output to correct the bottom log2(q)/2 bits of the
> floating output

If one has an approximate floating output and the same 
number modulo q (exactly) in hand, the one can obviously
use the latter to correct up to the log2(q) least-
significant bits of the former, i.e. there should be
no divide by 2 in there. What I intended to say was
that this allows one to use up to log2(q)/2 more bits
per INPUT to the convolution (since outputs scale
quadratically with input word size).

For example, if
one does the modular transform modulo the Mersenne
prime M61, then floating outputs may be in error by as
much as 2^60 (we lose one bit since floating outputs are
signed) and still be exactly reconstructed, and this
allows us to use roughly 30 more bits per input word.

-Ernst


_________________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
Mersenne Prime FAQ      -- http://www.tasam.com/~lrwiman/FAQ-mers

Reply via email to