Quote:
Originally Posted by jyb
Code:
anorm: 1.30e+037, rnorm: 1.27e+049

These are a bit small, especially anorm; the code uses a = 1e6*sqrt(skew), b=1e6/sqrt(skew) which might work well for small jobs but for numbers this large 1e6 is a big underestimate.
Really it should be something like a = sqrt(2^(2I1)*Q*skew)*const, b = sqrt(2^(2I1)*Q/skew)*const where Q is a realistic average specialq; smaller a and b give more relations so const should maybe be around 0.2? Not sure how it would be best to code this, given that the scaled difficulty informs the choice of I and Q. Perhaps use some function of difficulty that's a reasonable approximation of sqrt(2^(2I1)*Q)?
And really anorm ought to be multiplied by exp(alpha) to account for root properties of the algebraic polynomial, but if you're doing a job large enough that this matters, then you ought to be testsieving anyway.