On Tue, Jul 17, 2012 at 2:07 PM, <[email protected]> wrote:
> My roots-of-unity verb was designed to get higher accuracy than simpler
> approaches. If you just take numbers from 0 to pi you in effect lose an ULP
> or two of significance. So, I calculated over a smaller & more accurate
> interval & used reflection and sign-change. The results of FFT using the
> more accurate method were noticeably better. There is Forum traffic about it.
Can you point at the relevant parts of this traffic?
If we define:
rouwiki=: [: (, j.) [: (, (j.~%:0.5) , |."1&.:+.@|.@}.) [: ^@o.@:j.
i.@(%&8) % -:
rouforum=: [: (* * |)&.+. [: ^ 0j2p1 * % * i.
I would like to compare fft results generated using (rou=: -: {.
rouforum) with fft results from (rou=: rouwiki).
(I had earlier suggested something like using rou=: rouforum@-: but
that was wrong.)
I feel that having the previous problem cases in hand would be more
meaningful than testing cases which might not be relevant.
Thanks,
--
Raul
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
