The function being minimized, namely
probability of finding a 50-digit factor on one curve
-
time per curve
is flat near its minimum. Implementation and platform differences
can obviously affect the denominator
Hi all,
I'm working on version 19 of prime95 and I need your help.
In the past, the exponents at which a larger FFT is used was picked
rather haphazardly. I simply picked a few exponents trying to find
one that could run a thousand or so iterations without the convolution
error greatly
George,
You indicate that that the error distribution looks "like a Bell curve".
There is reasonable theoretical basis for the errors to follow a Bell
curve. The sum of many random plusses and minuses combine as in the
classic "random walk" problem to give a Gaussian probability distribution.