Stavros Macrakis wrote:
> On 1/3/07, Michel Van den Bergh <michel.vandenbergh at uhasselt.be> wrote:
>> Does anybody know if a bfloat number can overflow or underflow?
>
> No, it cannot. The exponent is a Lisp bignum, so is limited only by
> memory size. The precision is also only limited by memory size (and
> calculation time...).
>
> -s
Good! The formula for the levin transform divides by a large integer
factor in numerator
and denominator to avoid floating point overflow. So I can just cancel
that factor:-)
For the record: I can now reliably compute the propagation of
(b)floating point errors
in the levin transform. So I can now do
fpprec:50;
blevin_u_sum(1/n-log(1+1/n),n,1);
and it will compute the sum with a precision of 50 digits.
Michel